diff --git a/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/c6118b29-9fd0-4c0e-82ff-626a5de37c61_content_list.json b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/c6118b29-9fd0-4c0e-82ff-626a5de37c61_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..169a05dc6577d9a5ccb093a4fe7ee55ce1319ed7 --- /dev/null +++ b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/c6118b29-9fd0-4c0e-82ff-626a5de37c61_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31b5a3aea4d7477472c8b4f21a8390bec87a1e9e09a76030b5d20d4b59ec200f +size 75668 diff --git a/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/c6118b29-9fd0-4c0e-82ff-626a5de37c61_model.json b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/c6118b29-9fd0-4c0e-82ff-626a5de37c61_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bd1c050d53a2194fc2d7f474ab086f2a06a83148 --- /dev/null +++ b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/c6118b29-9fd0-4c0e-82ff-626a5de37c61_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:628b6f6970097aa30f80bb9c06614b25fefb58d4e53f5e9f0c965ddfca82761d +size 92656 diff --git a/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/c6118b29-9fd0-4c0e-82ff-626a5de37c61_origin.pdf b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/c6118b29-9fd0-4c0e-82ff-626a5de37c61_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..63a87ca9434334fbaa79eea6395433a3bdf06a9e --- /dev/null +++ b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/c6118b29-9fd0-4c0e-82ff-626a5de37c61_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf3629d4aa53326bc57adba6384815d7e8ab03034ba140a61f71ecbc1956e589 +size 1568374 diff --git a/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/full.md b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9d66cbc0de4565a2286d28786b5e74bc204fba8a --- /dev/null +++ b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/full.md @@ -0,0 +1,338 @@ +# Object-aware Sound Source Localization via Audio-Visual Scene Understanding + +Sung Jin Um $^{1,*}$ Dongjin Kim $^{1,2,*}$ Sangmin Lee $^{3,\dagger}$ Jung Uk Kim $^{1,\dagger}$ $^{1}$ Kyung Hee University $^{2}$ KAIST AI $^{3}$ Sungkyunkwan University + +{sungzin1,ju.kim}@khu.ac.kr,dj_kim@kaist.ac.kr,sangmin.lee@skku.edu + +# Abstract + +Audio-visual sound source localization task aims to spatially localize sound-making objects within visual scenes by integrating visual and audio cues. However, existing methods struggle with accurately localizing sound-making objects in complex scenes, particularly when visually similar silent objects coexist. This limitation arises primarily from their reliance on simple audio-visual correspondence, which does not capture fine-grained semantic differences between sound-making and silent objects. To address these challenges, we propose a novel sound source localization framework leveraging Multimodal Large Language Models (MLLMs) to generate detailed contextual information that explicitly distinguishes between sound-making foreground objects and silent background objects. To effectively integrate this detailed information, we introduce two novel loss functions: Object-aware Contrastive Alignment (OCA) loss and Object Region Isolation (ORI) loss. Extensive experimental results on MUSIC and VGGSound datasets demonstrate the effectiveness of our approach, significantly outperforming existing methods in both single-source and multi-source localization scenarios. Code and generated detailed contextual information are available at: https://github.com/VisualAIKHU/OA-SSL. + +# 1. Introduction + +In recent years, there has been a growing emphasis on audio-visual multimodal tasks in computer vision [17, 22, 27, 30, 31]. These tasks aim to solve complex challenges by combining auditory and visual information, enabling models to interpret scenes and capture both the spatial and temporal dynamics of real-world environments. + +Among these various multimodal tasks, audio-visual sound source localization has been receiving significant attention [3, 10, 14, 15, 18, 20, 21, 23, 25, 26, 28, 31–38, 42, 47]. Inspired by human perception, which naturally integrates visual and auditory cues to identify the + +![](images/6d3b35be3054b52a0d66a503830d7e4c36e5574a83075cf2b1f8a5db6b6ae960.jpg) + +![](images/44d1c58bb420315d3a4fd8bec4ed8c1ff5cca9e63d73fefc56b31bda71e101bf.jpg) + +![](images/ea2379989c5f4f77bc69eef96b179716da4443f8a03165a211abb45f2049c9aa.jpg) + +![](images/e0ce8354be597fe636c66784ee4dfbc79cd3e100eacdf00c513e60f446eb7412.jpg) + +![](images/3009ff0fa31191541a0e19ccc670bf0fb466a151857ef72358c84e70ddf533f5.jpg) + +![](images/3dd220e2c6291a1354369a2438a80a43ece51a6ca6f8f8ea7b702d752557cf93.jpg) +(a) +Figure 1. Conceptual comparison between (a) existing methods and (b) our method. Existing methods focus on localizing the 'guitar' object through simple audio-visual correspondence, ignoring whether it is being 'played'. In contrast, our method localizes only the playing (i.e., sound-making) guitar with fine-grained audio-visual correspondence from audio-visual scene understanding. + +![](images/c2b3bd863d86b1e46e101348c8beb79c7e4a32569e3c4eafa4c18edc7a3c1cd9.jpg) +Foreground +There is a playing guitar. +Background +There are guitars hanging on the wall, drum set, and Audio-Visual +Scene Understanding +(b) + +sources of sounds, this task focuses on accurately determining the spatial locations of sound-making objects within video scenes. Previous studies have explored audio-visual correspondence using self-supervised learning techniques, including hard positive mining [32], iterative learning [20], and negative-free learning [36]. In addition, to handle more realistic and challenging settings involving multiple sound sources, generating supervisory signals [14], graph-based modeling [15], and iterative matching schemes [18] have been explored. + +Despite these advancements, existing methods still struggle to accurately localize sound sources in complex scenes, particularly when scenes include both sound-making and visually similar silent objects simultaneously. The main limitation lies in their reliance on simple audio-visual correspondence, which does not adequately capture fine-grained differences between sound-making object and silent objects. For example, as shown in Figure 1(a), existing methods mislocalize non-playing guitars by focusing solely on identifying the 'guitar' in the scene based on the sound, without considering whether the guitar is actually being 'played.' This oversimplification reduces localization accuracy, highlighting the need for a more contextu + +ally aware approach. + +To address this issue, we propose a novel approach that captures more detailed contextual information by fully leveraging audio-visual scene understanding. As shown in Figure 1(b), our approach achieves fine-grained localization by (i) identifying objects present in the visual scene (e.g., guitars and drum set) and (ii) distinguishing sound-making foreground objects (e.g., a 'playing' guitar) while differentiating silent background objects (e.g., 'non-playing' guitars and drum set), effectively localizing only the sound-making objects. To enable fine-grained localization through audio-visual scene understanding, we utilize Multimodal Large Language Models (MLLMs) to generate detailed guidance using their rich external knowledge during training. By incorporating MLLMs, we acquire detailed information about both sound-making and silent objects, enabling our model to learn fine-grained representations for precise audio-visual correspondence. + +We further introduce two specialized loss functions during model training to effectively leverage fine-grained information. First, we propose an Object-aware Contrastive Alignment (OCA) loss that enhances the ability of our model to differentiate between sound-making and silent objects. The OCA loss leverages detailed object information from audio-visual scene understanding to align audio-visual features and accurately differentiate sound-making object from silent object, even when their visual appearances are similar. Moreover, we propose Object Region Isolation (ORI) loss to address the challenge of multiple sound sources within the same scene. This loss function promotes spatial separation of distinct sound-making objects in multisource scenarios, enabling the model to isolate each sound source and minimize overlap between their localized regions, even for visually similar objects (e.g., violin and cello). By incorporating these loss functions with audio-visual scene understanding, our model effectively learns fine-grained audio-visual representations, significantly improving its ability to precisely localize and identify sound sources in complex and challenging environments. + +The major contributions of our paper are as follows: + +- We propose a novel approach that captures more detailed contextual information by fully leveraging audio-visual scene understanding. To the best of our knowledge, this is the first attempt to address fine-grained guidance about sound-making objects and silent objects through MLLMs for audio-visual sound source localization tasks. +- We propose two novel loss functions, Object-aware Contrastive loss and Object Region Isolation loss, which improve fine-grained alignment and differentiation of audiovisual features, even in complex scenes. +- Experimental results on MUSIC and VGGSound datasets demonstrate the effectiveness of the proposed method for both single- and multi-source sound localization. + +# 2. Related Work + +# 2.1. Audio-Visual Sound Source Localization + +Audio-visual sound source localization aims to identify spatial positions of sound sources within video frames, typically employing self-supervised cross-modal learning methods. Early approaches primarily focus on single-source scenarios [3, 4, 10, 20, 21, 25, 31-38, 42, 47], utilizing techniques such as dual-stream attention mechanisms [31], contrastive learning targeting challenging samples [3, 37], iterative pseudo-label refinement [20], optical-flow guided cross-attention [10, 35], enhanced semantic alignment [33], and audio-spatial integration across modalities [38]. Recent studies expand toward more realistic multi-source environments [15, 18, 26, 28], proposing methods including discriminative supervisory signal generation [14], coarse-to-fine audiovisual separation [28], graph-based modeling [15], category-specific semantic feature extraction and integration [26], iterative object identification without prior knowledge [18], and leveraging textual-audio embeddings for complex mixtures [23]. + +However, existing studies primarily rely on simple similarity calculations between visual and audio data, often utilizing the semantic information associated with sounds. This leads to ineffective localization when both sound-emitting and non-sound-emitting objects coexist. We propose a novel approach that captures detailed contextual information by leveraging audio-visual scene understanding, effectively bridging the audio-visual gap and enhancing fine-grained localization accuracy. + +# 2.2. Audio-Visual tasks with MLLMs + +Multimodal Large Language Models (MLLMs) [40, 43] combine the strengths of audio, visual, and language models to process complex tasks that involve both audio, image and text data. MLLMs are designed to handle various visual tasks that benefit from semantic understanding, such as object detection, scene recognition, and action detection, by integrating the contextual depth offered by language models with the spatial and visual recognition capabilities of vision models. This integration enables MLLMs to provide more interpretable outputs for audio-visual tasks. + +In audio-visual tasks like audio-visual scene understanding [44], video classification [29], video sound generation [39, 41], audio-visual feature alignment [8], MLLMs enable more effective cross-modal alignment by leveraging textual descriptions to enhance the understanding of visual content associated with specific sounds. Building on this, we propose a method that leverages MLLMs for audio-visual scene understanding to generate detailed information on both sound-making and silent objects. This detailed information supports fine-grained audio-visual alignment, allowing for improved sound source localization performance. + +![](images/f57adca1d703058f3563d82c0f676894ebd17633b937ce51a5957fb3fd4e2b8b.jpg) +Figure 2. Network configuration of the proposed sound source localization framework. MLLM is used only during the training process and not during inference. + +# 3. Proposed Method + +To achieve more fine-grained sound source localization, we aim to leverage detailed contextual information that enhances audio-visual scene understanding. As shown in Figure 2, during training, we leverage a comprehensive understanding of audio-visual scenes by identifying both soundmaking and silent objects using Multimodal Large Language Model (MLLM) with input image and audio class information. We encode this detailed information as the reference anchor feature for fine-grained audio-visual representation learning. + +We introduce two loss functions to guide fine-grained audio-visual alignment through detailed information leveraged to understand audio-visual scenes: Object-aware Contrastive Alignment (OCA) loss and Object Region Isolation (ORI) loss. These loss functions guide the model in identifying sound-making objects while ignoring silent ones and distinguishing each sound source in multi-source scenarios, ultimately improving sound source localization performance through enhanced audio-visual alignment. + +# 3.1. Preliminaries + +Sound-associated Region Localization. Sound source localization identifies and localizes sound-making objects + +in visual scenes using audio-visual information. Similar to previous works [3, 31], our model uses a two-stream network to extract features from both visual and audio modalities. The input image and audio spectrogram pass through each modal encoder (i.e., visual encoder and audio encoder) to generate visual feature $\mathbf{F}_v \in \mathbb{R}^{B \times w \times h \times c}$ ( $B$ indicates batch number, $w$ , $h$ , $c$ are the width, height, and channel) and audio feature $l_a \in \mathbb{R}^{B \times c}$ , respectively. To localize sound-associated regions, we compute the cosine similarity between the visual feature $\mathbf{F}_v$ and audio feature $l_a$ , resulting in the sound-associated map $\mathbf{S}_a$ . + +Object Identification. For multi-sound source localization where multiple sound-making objects are present, we adopt an iterative object identification algorithm [18]. In this algorithm, object identification is performed iteratively for $K$ iterations corresponding to the number of sound sources. + +# 3.2. Audio-Visual Scene Understanding + +To effectively distinguish sound-making objects and silent objects in complex scenes with various objects, we utilize audio-visual understanding by using the Multimodal Large Language Model (MLLM). The MLLM, trained with extensive external knowledge of diverse and complex scenes, re + +ceives an input consisting of the given image and audio class information. Through a carefully designed prompt to distinguish whether each object is making sound or silent, we guide MLLM to effectively utilize the provided image and audio class information with some examples for complex scenes (e.g., if there are multiple objects that match with audio class information, distinguish between sound-making object and silent object and describe the action of objects.) This process generates $K$ foreground captions, corresponding to the number of sound sources and a single background caption. (Please refer to supplementary material for details.) + +Each foreground caption provides detailed information about specific sound-making objects with their details, thereby enhancing the understanding of sound-related activities within the audio-visual scene. In contrast, background caption describes objects unrelated to sound production, offering explicit details about silent elements. By incorporating not only sound-making (foreground) but also silent (background) information, our method can understand which objects are really sound-making and silent. + +The generated captions are subsequently processed through a text modality encoder to extract features, resulting in a foreground reference feature $\mathbf{F}_r^p\in \mathbb{R}^{B\times K\times c}$ and a background reference feature $l_r^n\in \mathbb{R}^{B\times c}$ . These features guide the model during the fine-grained audio-visual alignment process, enabling it to effectively distinguish between sound-making and non-sound-making objects, as well as differentiate between multiple sound-making objects in multi-source scenarios. + +# 3.3. Object-aware Contrastive Alignment Loss + +Existing sound source localization methods mainly rely on contrastive learning with cosine similarity for audio-visual alignment. However, since these methods focus primarily on foreground object features rather than understanding audio-visual scenes, silent objects cannot be effectively distinguished from sound-making objects. To address this issue, we propose the Object-aware Contrastive Alignment (OCA) loss. The OCA loss has an important role in learning fine-grained audio-visual correspondence through leveraging the foreground reference feature $\mathbf{F}_r^p$ and background reference feature $l_r^n$ from audio-visual scene understanding. + +First, we compute the cosine similarity between visual feature $\mathbf{F}_v$ and audio feature $l_{a}$ to generate the sound-associated map $\mathbf{S}_a\in \mathbb{R}^{B\times w\times h}$ , defined as: + +$$ +\operatorname {S i m} (A, B) = \frac {\langle A , B \rangle}{\left| \left| A \right| \right| \left| \left| B \right| \right|}, \tag {1} +$$ + +$$ +\mathbf {S} _ {a} = \operatorname {S i m} \left(\mathbf {F} _ {v}, l _ {a}\right), +$$ + +where $\langle \cdot, \cdot \rangle$ denotes dot product. Following this, inspired by the prior work [3], we create a foreground map $\mathbf{M}^p$ and a background map $\mathbf{M}^n$ for each foreground and background + +selection, defined as: + +$$ +\mathbf {M} ^ {p} = \operatorname {s i g m o i d} \left(\left(\mathbf {S} _ {a} - \alpha_ {p}\right) / \omega\right) +$$ + +$$ +\mathbf {M} ^ {n} = \mathbf {1} - \operatorname {s i g m o i d} \left(\left(\mathbf {S} _ {a} - \alpha_ {n}\right) / \omega\right), \tag {2} +$$ + +where $\alpha_{p}$ , $\alpha_{n}$ and $\omega$ denotes the hyper-parameters [3]. + +Using $\mathbf{M}^p$ and $\mathbf{M}^n$ , we then generate the foreground visual feature $l_v^p \in \mathbb{R}^{B \times c}$ and background visual feature $l_v^n \in \mathbb{R}^{B \times c}$ , represented as: + +$$ +l _ {v} ^ {p} = \operatorname {G A P} \left(\left\langle \mathbf {F} _ {v}, \mathbf {M} ^ {p} \right\rangle\right), l _ {v} ^ {n} = \operatorname {G A P} \left(\left\langle \mathbf {F} _ {v}, \mathbf {M} ^ {n} \right\rangle\right), \tag {3} +$$ + +where GAP denotes the global average pooling. + +After that, we design OCA loss to enhance the ability of our model to differentiate between sound-making and silent objects. The OCA loss is divided into a foreground alignment loss $\mathcal{L}_{frg}$ and a background alignment loss $\mathcal{L}_{bkg}$ . The $\mathcal{L}_{frg}$ aims to encourage the alignment of the foreground visual feature $l_v^p$ with the average-pooled foreground reference feature $l_r^p \in \mathbb{R}^{B \times c}$ , while simultaneously pushing the background visual feature $l_v^n$ away from $l_r^p$ , calculated as: + +$$ +p _ {i} = \exp (S i m (l _ {v _ {i}} ^ {p}, l _ {r _ {i}} ^ {p})), +$$ + +$$ +n _ {i} ^ {\text {h a r d}} = \exp \left(S i m \left(l _ {v _ {i}} ^ {n}, l _ {r _ {i}} ^ {p}\right)\right), +$$ + +$$ +n _ {i} ^ {\text {s o f t}} = \sum_ {j \neq i} ^ {B} \mathbb {1} _ {[ \operatorname {S i m} \left(l _ {r _ {j}} ^ {p}, l _ {r _ {i}} ^ {p}\right) \leq \tau ]} \exp \left(S i m \left(l _ {v _ {j}} ^ {p}, l _ {r _ {i}} ^ {p}\right)\right), \tag {4} +$$ + +$$ +\mathcal {L} _ {f r g} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {p _ {i}}{p _ {i} + n _ {i} ^ {\text {h a r d}} + n _ {i} ^ {\text {s o f t}}}, +$$ + +where $n_i^{\mathrm{soft}}$ handles the false negative issue [9, 16, 46] inherent in batch contrastive learning, where similar sound-making objects within the batch can cause unintended negative samples. We compare average-pooled foreground reference features $l_r^p$ across different clips in the batch, applying a threshold $\tau$ to exclude samples with similarity values exceeding this threshold, as these are considered false negatives. + +To fully understand the audio-visual scene, it is crucial to capture the background information through silent objects in a visual scene. To this end, the background alignment loss $\mathcal{L}_{bkg}$ aims to encourage the background visual feature $l_v^n$ to align with the background reference feature $l_r^n$ , while pushing the foreground visual feature $l_v^p$ away from $l_r^n$ . The $\mathcal{L}_{bkg}$ is defined as: + +$$ +\hat {p} _ {i} = \exp (S i m \left(l _ {v _ {i}} ^ {n}, \mathbf {F} _ {r _ {i}} ^ {n}\right)), +$$ + +$$ +\hat {n} _ {i} ^ {\text {h a r d}} = \exp \left(S i m \left(l _ {v _ {i}} ^ {p}, \mathbf {F} _ {r _ {i}} ^ {n}\right)\right), \tag {5} +$$ + +$$ +\mathcal {L} _ {b k g} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\hat {p} _ {i}}{\hat {p} _ {i} + \hat {n} _ {i} ^ {\mathrm {h a r d}}}. +$$ + +Finally, the proposed OCA loss $\mathcal{L}_{oca}$ is defined as: + +$$ +\mathcal {L} _ {o c a} = \frac {\mathcal {L} _ {f r g} + \mathcal {L} _ {b k g}}{2}. \tag {6} +$$ + +Through $\mathcal{L}_{\mathrm{oca}}$ , our model can learn fine-grained audio-visual representation from foreground and background object information, even in complex scenarios where both sound-making and visually similar silent objects are present. + +# 3.4. Object Region Isolation Loss + +While the audio-visual alignment achieved through the OCA loss enables effective localization of sound-making objects, multi-source scenarios require more distinct separation of sound-making objects and clear differentiation between foreground and background regions. To address this need, we propose the Object Region Isolation (ORI) loss, which enhances object separation in complex scenes by enforcing spatial distinctiveness between regions associated with each sound-making object and the background. + +We generate the combined reference feature $\mathbf{F}_r \in \mathbb{R}^{B \times (K + 1) \times c}$ by concatenating $\mathbf{F}_r^p$ and $l_r^n$ . Then, using similarity calculations between the visual feature $\mathbf{F}_v$ and $\mathbf{F}_r$ , we create a reference associated map $\mathbf{S}_r = \{S_{r_k}\}_{k = 1,\dots,K + 1} \in \mathbb{R}^{B \times (K + 1) \times w \times h}$ , where each $S_{r_k}$ corresponds to a similarity map for each object or background region. + +To further separate these regions, we employ the first-order Wasserstein Distance, also known as the Earth-Mover Distance, to measure the distance between the joint probability distributions of different objects or background regions. For this calculation, we flatten $\mathbf{S}_r$ into $\bar{\mathbf{S}}_r = \{\bar{S}_{r_k}\}_{k = 1,\dots ,K + 1} \in \mathbb{R}^{B \times (K + 1) \times wh}$ . The Sinkhorn algorithm [6] is then applied to efficiently compute the optimal transport for minimizing the Wasserstein Distance. The loss function is defined as: + +$$ +\begin{array}{l} \mathbf {S} _ {r} = \operatorname {S i m} \left(\mathbf {F} _ {v}, \mathbf {F} _ {r}\right), \\ D_{W}(P,Q) = \inf_{\gamma \in \Pi (P,Q)}\mathbb{E}_{(x,y)\sim \gamma}[||x - y||], \\ \mathcal{L}_{ori} = \sum_{i = 1}^{B}\sum_{\substack{n,m = 1\\ n\neq m}}^{K + 1}D_{W}(\bar{S}_{r_{n}}^{i},1 - \bar{S}_{r_{m}}^{i}), \tag{7} \\ \end{array} +$$ + +where $D_W$ represents the first-order Wasserstein Distance. + +Through $\mathcal{L}_{ori}$ , the model is guided to avoid overlapping regions by encouraging clear separation in the maps generated from similarity with each reference feature. As a result, the model achieves improved separation of foreground and background regions, and in multi-source scenarios, it distinctly identifies each sound-making object within the foreground. This method ensures the model can handle complex scenes with multiple sound sources, enhancing both localization and object identification accuracy. + +# 3.5. Training Objective + +To train our audio-visual scene understanding-based method, we construct the total training loss function as fol + +lows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {1} \mathcal {L} _ {\text {o c a}} + \lambda_ {2} \mathcal {L} _ {\text {o r i}}, \tag {8} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ denote balancing parameters for each loss function. By optimizing $\mathcal{L}_{total}$ , the model learns a more robust and fine-grained audio-visual correspondence, leading to enhanced localization accuracy and better discrimination of sound sources in complex scenes. + +# 4. Experiment + +# 4.1. Datasets and Evaluation Metrics + +MUSIC Dataset. The MUSIC dataset [45] consists of 448 unedited YouTube videos featuring solos and duets across 11 different musical instrument categories. Following prior works [18, 23, 26], we use the same training and testing subsets to maintain consistency with previous studies. The MUSIC-Solo subset includes 358 solo videos for training and 90 solo videos for evaluation, designed for single sound source localization. Also, the MUSIC-Duet subset contains 124 duet videos for training and 17 duet videos for evaluation, aimed at multi-sound source localization tasks. + +VGG-Sound Dataset. The VGG-Sound dataset [2], referred to as VGGSound-Single, includes over 200k videos from 221 different sound categories. We use 144k image-audio pairs for training. For single sound source localization evaluation, we use VGG-Sound Source [3]. During training for multi-sound source localization, we randomly combine two video frames to generate a single input image of size $448 \times 224$ , while also mixing the corresponding audio waveforms. Following [18, 23, 26], we use the VGGSound-Duet dataset for evaluation. + +Evaluation Metrics. For a comprehensive comparison, we adopt metrics used in prior works [15, 18, 23, 26]. For single sound source localization, we use Average Precision (AP), Intersection over Union (IoU), and Area Under the Curve (AUC). For multi-sound source localization, we use Class-aware Average Precision (CAP), Class-aware IoU (CIoU), and AUC. + +# 4.2. Implementation Details + +For the visual input, we resize images to $224 \times 224$ , extracting them from the center frame of each 3-second video clip. We resample the raw 3-second audio signal to $16~\mathrm{kHz}$ for the audio input and convert it to a log-scale spectrogram. + +Following [3], we use a ResNet-18 [12] architecture for both the visual and audio feature extraction. To handle the one-channel input of the sound spectrogram, we modify the first convolutional layer of the ResNet-18 from 3 channels to 1 channel. The visual encoder is pre-trained on ImageNet [7]. Additionally, for a fair comparison with T-VSL + +
MethodBackboneMUSIC-Duet [45]VGGSound-Duet [2]
CAP(%)CloU@0.3(%)AUC(%)CAP(%)CloU@0.3(%)AUC(%)
Attention10k (CVPR'18) [31]VGG16-21.619.6-11.515.2
OTS (ECCV'18) [1]VGG1611.613.318.510.512.215.8
DMC (CVPR'19) [13]VGG16-17.521.1-13.817.1
CoarseToFIne (ECCV'20) [28]ResNet18-17.620.6-14.718.5
DSOL (NeurIPS'20) [14]ResNet18-30.122.3-22.321.1
LVS (CVPR'21) [3]ResNet18-22.520.9-17.319.5
EZ-VSL (ECCV'22) [24]ResNet18-24.321.3-20.520.2
Mix-and-Localize (CVPR'22) [15]ResNet1847.526.521.516.321.120.5
AVGN (CVPR'23) [26]ResNet1850.632.524.621.926.223.8
NoPrior (CVPR'24) [18]ResNet1852.138.630.132.546.929.2
Proposed MethodResNet1861.445.936.145.955.244.8
AudioCLIP (ICASSP'22) [11]AudioCLIP56.136.929.228.434.932.8
AVGN (CVPR'23) [26]AudioCLIP59.239.332.431.937.835.4
T-VSL (CVPR'24) [23]AudioCLIP62.943.235.935.740.137.9
Proposed MethodAudioCLIP64.146.938.447.157.747.4
+ +Table 1. Experimental results on MUSIC-Duet (left) and VGGSound-Duet (right) test sets for multi-sound source localization. Bold/underlined fonts indicate the best/second-best results. + +[23], we also use the AudioCLIP setup, which includes pretrained encoders with a ResNet-50 image encoder and ES-ResNeXt audio encoder. For the text encoder, we utilize BERT, specifically using the [CLS] token embedding from the last hidden state for sentence-level representation. We train our model using the Adam optimizer [19] with a learning rate of $10^{-4}$ and a batch size of 128, running for 100 epochs. For MLLM, we utilize the InternVL 2.0-8B model [5]. InternVL 2.0 is trained with an 8k context window, using training data that includes long texts, multiple images, and videos. For comparative analyses with alternative MLLMs, please refer to our supplementary materials. + +We use two RTX 4090 GPUs for training with the ResNet-18 backbone and two RTX A6000 GPUs for the AudioCLIP backbone. We set the weighting parameter as $\tau = 0.7$ for our Object-aware Contrastive Alignment (OCA) loss. For our total loss function $\mathcal{L}_{Total}$ , we use $\lambda_{1} = 1$ and $\lambda_{2} = 0.1$ . Following [3, 18, 23], we use the provided implementation code and maintain the same hyperparameters for other aspects of our model. + +# 4.3. Comparison to Prior Works + +Multi-Sound Source Localization. We conduct experiment to compare the performance of our method with state-of-the-art multi-sound source localization methods [1, 3, 13-15, 18, 23, 24, 26, 28, 31]. As shown in Table 1, for the ResNet18 backbone, our method significantly outperforms the existing methods on the MUSIC-Duet [45] test set in terms of CAP, CIoU@0.3, and AUC by $9.3\%$ , $7.3\%$ , and $6.0\%$ , respectively. In the VGGSound-Duet [2] test set, our approach consistently shows superior performance compared to previous methods by leveraging the audio-visual scene understanding. + +Single Sound Source Localization. Similarly, for single sound source localization, we also conduct experiments comparing our method with existing methods [1, 3, 13-15, 18, 23, 24, 26, 28, 31]. Table 2 presents performances on the MUSIC-Solo [45] and VGGSound-Single [2] datasets. Specifically, for the MUSIC-Solo test set, our method achieves improvements with the ResNet18 backbone of $2.4\%$ in AP, $9.0\%$ in IoU@0.5, and $1.5\%$ in AUC. We observe a similar tendency on the VGGSound-Single test set. + +Overall, our approach consistently outperforms the existing methods across both single and multi-sound source localization tasks. This demonstrates the effectiveness of our OCA loss $\mathcal{L}_{\mathrm{oca}}$ and ORI loss $\mathcal{L}_{\mathrm{ori}}$ in understanding scene to enhance fine-grained audio-visual correspondence for more accurate sound source localization. + +# 4.4. Ablation Study + +In this section, we conduct various ablation studies to investigate (1) the effect of our proposed losses and (2) the variation of hyper-parameter $\tau$ for $\mathcal{L}_{\mathrm{oca}}$ . All ablation studies are conducted based on the VGGSound-Duet dataset with the ResNet18 backbone. + +Effect of the Proposed Losses. We first evaluate the impact of the two introduced loss functions: OCA loss $\mathcal{L}_{oca}$ and ORI loss $\mathcal{L}_{ori}$ . Note that, without applying any of the proposed losses, we use the traditional audio-visual contrastive loss [3]. As shown in Table 3, consideration of $\mathcal{L}_{oca}$ significantly improves the performance of our method with the fine-grained audio-visual correspondence compared to the baseline without this awareness. Considering $\mathcal{L}_{ori}$ also enhances the separation of overlapping regions, resulting + +
MethodBackboneMUSIC-Solo [45]VGGSound-Single [2]
AP(%)IoU@0.5(%)AUC(%)AP(%)IoU@0.5(%)AUC(%)
Attention10k (CVPR'18) [31]VGG16-37.238.7-19.230.6
OTS (ECCV'18) [1]VGG1669.326.135.829.832.835.7
DMC (CVPR'19) [13]VGG16-29.138.0-23.927.6
CoarseToFIne (ECCV'20) [28]ResNet1870.733.639.828.229.134.8
DSOL (NeurIPS'20) [14]ResNet18-51.443.7-35.737.2
LVS (CVPR'21) [3]ResNet1870.641.940.329.634.438.2
EZ-VSL (ECCV'22) [24]ResNet1871.545.841.231.338.939.5
Mix-and-Localize (CVPR'22) [15]ResNet1868.630.540.832.536.338.9
AVGN (CVPR'23) [26]ResNet1877.258.148.535.340.842.3
NoPrior (CVPR'24) [18]ResNet1877.462.159.446.241.441.2
Proposed MethodResNet1879.871.160.951.747.344.9
AudioCLIP (ICASSP'22) [11]AudioCLIP83.863.155.742.847.448.5
AVGN (CVPR'23) [26]AudioCLIP85.465.856.944.149.649.5
T-VSL (CVPR'24) [23]AudioCLIP88.268.560.148.153.752.9
Proposed MethodAudioCLIP89.672.162.550.457.156.4
+ +Table 2. Experimental results on MUSIC-Solo (left) and VGGSound-Single (right) test sets for single sound source localization. Bold/underlined fonts indicate the best/second-best results. + +
LocaLoriCAP(%)CloU@0.3(%)AUC(%)
--28.831.426.2
✓-43.850.943.1
-✓34.349.134.8
✓✓45.955.244.8
+ +Table 3. Effect of the two proposed losses ${\mathcal{L}}_{oca}$ and ${\mathcal{L}}_{ori}$ on VGGSound-Duet test set. + +
τMUSIC-DuetVGGSound-Duet
CAP(%)CloU@0.3(%)AUC(%)CAP(%)CloU@0.3(%)AUC(%)
0.154.141.131.236.147.231.0
0.455.842.132.444.152.833.5
0.761.445.936.145.955.244.8
1.059.244.235.939.952.640.7
+ +Table 4. Experimental results on MUSIC-Duet and VGGSound-Duet test sets according to the hyper-parameter $\tau$ for ${\mathcal{L}}_{\text{oca }}$ . + +in better localization, particularly in complex multi-source scenarios. We achieved the highest performance when both losses are considered, indicating that each loss uniquely contributes to improving scene understanding for sound source localization. + +Variation of $\tau$ . We also conduct experiments to analyze the impact of the hyper-parameter $\tau$ , which controls the similarity threshold for false negatives in the OCA loss in Section 3.3. A lower $\tau$ value results in selecting fewer negative samples, which helps address the false negative issue by ensuring more thorough sampling of negative pairs. However, it may limit the proper understanding of similar but distinct objects, as true negatives might be misinterpreted. Conversely, a higher $\tau$ value leads to more negative sam + +![](images/7fc4e2cb718479b0f65bf0cf2f99f3ae6190cbe96001169ce1cdeef135654414.jpg) +Figure 3. Visualization results for VGGSound-Single test set. We compare our method with NoPrior [18]. + +plies being selected, which may reduce the effectiveness of false negative detection. Table 4 shows that we achieve the best performance with $\tau = 0.7$ , where the most balanced training is observed. + +# 4.5. Visualization Results + +Figures 3 and 4 present examples of both single-sound source localization on the VGGSound-Single test set, multi-sound source localization on the MUSIC-Duet and VGGSound-Duet test set. We compare our method with the state-of-the-art method, i.e., NoPrior [18]. These visualization results highlight the ability of our model to perform fine-grained localization achieved with OCA loss by + +![](images/f5ff2bde53b16c026f8c5ab7cb3339eff352517966cfd5010120212301297e78.jpg) +Figure 4. Visualization results for (a) MUSIC-Duet and (b) VGGSound-Duet test set. We compare our method with NoPrior [18]. + +
MethodCAP(%)CloU@0.3(%)AUC(%)
AVGN [26]18.522.721.8
NoPrior [18]29.034.029.3
Proposed Method34.344.933.7
+ +Table 5. Experimental results on VGGSound-Trio test set. Bold/underlined fonts indicate the best/second-best results. + +accurately identifying only the sound-making object. Furthermore, in multi-source scenarios, our method effectively separates sound-making objects, demonstrating the performance of the ORI loss. These results validate the effectiveness of our approach in enhancing audio-visual localization accuracy across diverse scenarios. + +# 4.6. Discussion + +Following previous works [18, 26], we conduct additional experiments to evaluate our method with flexible number of sound sources. To this end, we use the VGGSound-Trio [18] dataset, which consists of a mixture of three sources. As shown in Table 5, our method outperforms previous works across all evaluation metrics, highlighting its effectiveness in handling flexible number of sound-making objects by fully leveraging audio-visual scene understanding. + +# 5. Conclusion + +In this paper, we propose a novel sound source localization framework to address the challenges of sound source localization in complex scenes. To this end, our framework + +leverages detailed contextual information for fine-grained localization by incorporating Multimodal Large Language Models (MLLMs). By generating rich contextual details through audio-visual scene understanding, our approach enables more effective differentiation between sound-making and silent objects within various environments. Based on such contextual details, we propose two novel loss functions: Object-aware Contrastive Alignment (OCA) loss and Object Region Isolation (ORI) loss to effectively guide the model in aligning audio-visual features and ensuring distinct separation of sound sources. We believe our approach not only enhances localization accuracy but also holds diverse practical applications. + +Acknowledgements. This work was supported by the NRF grant funded by the Korea government (MSIT) (No. RS-2023-00252391, RS-2025-00563942, RS-2025-00555621), and by IITP grant funded by the Korea government (MSIT) (No. RS-2022-00155911: Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University), RS-2022-II220124: Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities, IITP-2023-RS2023-00266615: Convergence Security Core Talent Training Business Support Program, No. RS-2024-00509257: Global AI Frontier Lab, IITP-2025-RS-2023-00254129: Graduate School of Metaverse Convergence Support Program) + +# References + +[1] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In ECCV, 2018. 6, 7 +[2] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audiovisual dataset. In ICASSP, 2020. 5, 6, 7 +[3] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In CVPR, 2021. 1, 2, 3, 4, 5, 6, 7 +[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 2 +[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. 6 +[6] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. NeurIPS, 2013. 5 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5 +[8] Haoyi Duan, Yan Xia, Zhou Mingze, Li Tang, Jieming Zhu, and Zhou Zhao. Cross-modal prompts: Adapting large pre-trained models for audio-visual downstream tasks. In NeurIPS, 2024. 2 +[9] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In ICCV, 2021. 4 +[10] Dennis Fedorishin, Deen Dayal Mohan, Bhavin Jawade, Srirangaraj Setlur, and Venu Govindaraju. Hear the flow: Optical flow-based self-supervised visual sound source localization. In WACV, 2023. 1, 2 +[11] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP, 2022. 6, 7 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5 +[13] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In CVPR, 2019. 6, 7 +[14] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. In NeurIPS, 2020. 1, 2, 6, 7 +[15] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In CVPR, 2022. 1, 2, 5, 6, 7 + +[16] Tri Huynh, Simon Kornblith, Matthew R Walter, Michael Maire, and Maryam Khademi. Boosting contrastive self-supervised learning with false negative cancellation. In WACV, 2022. 4 +[17] Denis Ivanko, Dmitry Ryumin, and Alexey Karpov. A review of recent advances on deep learning methods for audio-visual speech recognition. Mathematics, 2023. 1 +[18] Dongjin Kim, Sung Jin Um, Sangmin Lee, and Jung Uk Kim. Learning to visually localize sound sources from mixtures without prior source knowledge. In CVPR, 2024. 1, 2, 3, 5, 6, 7, 8 +[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[20] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Unsupervised sound localization via iterative contrastive learning. CVIU, 2023. 1, 2 +[21] Jinxiang Liu, Chen Ju, Weidi Xie, and Ya Zhang. Exploiting transformation invariance and equivariance for self-supervised sound localisation. In ACM MM, 2022. 1, 2 +[22] Yang Liu, Dingkang Yang, Yan Wang, Jing Liu, Jun Liu, Azzedine Boukerche, Peng Sun, and Liang Song. Generalized video anomaly event detection: Systematic taxonomy and comparison of deep models. ACM Computing Surveys, 2024. 1 +[23] Tanvir Mahmud, Yapeng Tian, and Diana Marculescu. T-vsl: Text-guided visual sound source localization in mixtures. In CVPR, 2024. 1, 2, 5, 6, 7 +[24] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In ECCV, 2022. 6, 7 +[25] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. In NeurIPS, 2022. 1, 2 +[26] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In CVPR, 2023. 1, 2, 5, 6, 7, 8 +[27] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018. 1 +[28] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In ECCV, 2020. 1, 2, 6, 7 +[29] Rui Qian, Yeqing Li, Zheng Xu, Ming-Hsuan Yang, Serge Belongie, and Yin Cui. Multimodal open-vocabulary video classification via pretrained vision and language models. arXiv preprint arXiv:2207.07646, 2022. 2 +[30] Ludan Ruan, Yiyang Ma, Huan Yang, Huiguo He, Bei Liu, Jianlong Fu, Nicholas Jing Yuan, Qin Jin, and + +Baining Guo. Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In CVPR, 2023. 1 +[31] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In CVPR, 2018. 1, 2, 3, 6, 7 +[32] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In ICASSP, 2022. 1 +[33] Arda Senocak, Hyeonggon Ryu, Junsik Kim, TaeHyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In ICCV, 2023. 2 +[34] Jiayin Shi and Chao Ma. Unsupervised sounding object localization with bottom-up and top-down attention. In WACV, 2022. +[35] Rajsuryan Singh, Pablo Zinemanas, Xavier Serra, Juan Pablo Bello, and Magdalena Fuentes. Flowgrad: Using motion for visual sound source localization. In ICASSP, 2023. 2 +[36] Zengjie Song, Yuxi Wang, Junsong Fan, Tieniu Tan, and Zhaoxiang Zhang. Self-supervised predictive learning: A negative-free method for sound source localization in visual scenes. In CVPR, 2022. 1 +[37] Weixuan Sun, Jiayi Zhang, Jianyuan Wang, Zheyuan Liu, Yiran Zhong, Tianpeng Feng, Yandong Guo, Yanhao Zhang, and Nick Barnes. Learning audiovisual source localization via false negative aware contrastive learning. In CVPR, 2023. 2 +[38] Sung Jin Um, Dongjin Kim, and Jung Uk Kim. Audiovisual spatial integration and recursive attention for robust sound source localization. In ACM MM, 2023. 1, 2 +[39] Heng Wang, Jianbo Ma, Santiago Pascual, Richard Cartwright, and Weidong Cai. V2a-mapper: A lightweight solution for vision-to-audio generation by connecting foundation models. In AAAI, 2024. 2 +[40] Jiayang Wu, Wensheng Gan, Zefeng Chen, Shicheng Wan, and S Yu Philip. Multimodal large language models: A survey. In IEEE BigData, 2023. 2 +[41] Zhifeng Xie, Shengye Yu, Qile He, and Mengtian Li. Sonicvisionlm: Playing sound with vision language models. In CVPR, 2024. 2 +[42] Hanyu Xuan, Zhiliang Wu, Jian Yang, Yan Yan, and Xavier Alameda-Pineda. A proposal-based paradigm for self-supervised sound source localization in videos. In CVPR, 2022. 1, 2 +[43] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 2 + +[44] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2 +[45] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In ECCV, 2018. 5, 6, 7 +[46] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. In ICCV, 2021. 4 +[47] Xinchi Zhou, Dongzhan Zhou, Di Hu, Hang Zhou, and Wanli Ouyang. Exploiting visual context semantics for sound source localization. In WACV, 2023. 1, 2 \ No newline at end of file diff --git a/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/images.zip b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..634bb51c57e92b7c6e667640334e9281dbde843e --- /dev/null +++ b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76c7c5c199092a7ed120bcc9d46452acf71306292823c921076db4fafa8e3455 +size 822556 diff --git a/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/layout.json b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b1086d61fbaaba71dee8cae24595e577da6ce98e --- /dev/null +++ b/objectawaresoundsourcelocalizationviaaudiovisualsceneunderstanding/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b9a3f0dc16664af0c66a5b5061cf874b1ee41ed72dcf33b7505606b3560feb1 +size 383245 diff --git a/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/2b3d2731-29bf-4fe5-a556-07cbb31fd742_content_list.json b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/2b3d2731-29bf-4fe5-a556-07cbb31fd742_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f7b81caf3754ced81a5017503dce295770169030 --- /dev/null +++ b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/2b3d2731-29bf-4fe5-a556-07cbb31fd742_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00f54ec794daa8bf788714cfc791f678ba41444f002e7dc514e8dd80142e75b6 +size 77597 diff --git a/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/2b3d2731-29bf-4fe5-a556-07cbb31fd742_model.json b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/2b3d2731-29bf-4fe5-a556-07cbb31fd742_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7a03f466a70b0ef495f925f00ad82a43524a9d7 --- /dev/null +++ b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/2b3d2731-29bf-4fe5-a556-07cbb31fd742_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da80fdde87b51e16dfe2762315575e54464d01c6c6cf324a6fc4f502262c82b1 +size 98187 diff --git a/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/2b3d2731-29bf-4fe5-a556-07cbb31fd742_origin.pdf b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/2b3d2731-29bf-4fe5-a556-07cbb31fd742_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5754afbfffed83efd25b74f1c210096f07cfee6b --- /dev/null +++ b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/2b3d2731-29bf-4fe5-a556-07cbb31fd742_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:def17e806512a47b0dbcfa1c379d557a1c9d933bca481c0a1fe401dd114287ec +size 2560195 diff --git a/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/full.md b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1d7981012db271a33167712758440936cdf00c5f --- /dev/null +++ b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/full.md @@ -0,0 +1,275 @@ +# Object-Centric Prompt-Driven Vision-Language-Action Model for Robotic Manipulation + +Xiaoqi Li $^{1,2}$ , Jingyun Xu $^{1}$ , Mingxu Zhang $^{1}$ , Jiaming Liu $^{1}$ , Yan Shen $^{1,2}$ , Iaroslav Ponomarenko $^{1,2}$ , Jiahui Xu $^{1}$ , Liang Heng $^{1,2}$ , Siyuan Huang $^{1}$ , Shanghang Zhang $^{1}$ , Hao Dong $^{1,2*}$ $^{1}$ School of Computer Science, Peking University $^{2}$ PKU-Agibot Lab + +# Abstract + +In robotic, task goals can be conveyed through various modalities, such as language, goal images, and goal videos. However, natural language can be ambiguous, while images or videos may offer overly detailed specifications. To tackle these challenges, we introduce CrayonRobo that leverages comprehensive multi-modal prompts that explicitly convey both low-level actions and high-level planning in a simple manner. Specifically, for each key-frame in the task sequence, our method allows for manual or automatic generation of simple and expressive 2D visual prompts overlaid on RGB images. These prompts represent the required task goals, such as the end-effector pose and the desired movement direction after contact. We develop a training strategy that enables the model to interpret these visual-language prompts and predict the corresponding contact poses and movement directions in SE(3) space. Furthermore, by sequentially executing all key-frame steps, the model can complete long-horizon tasks. This approach not only helps the model explicitly understand the task objectives but also enhances its robustness on unseen tasks by providing easily interpretable prompts. We evaluate our method in both simulated and real-world environments, demonstrating its robust manipulation capabilities. + +# 1. Introduction + +As thoughtful helpers for humans, it is crucial for robots to understand and successfully execute their assigned tasks. Various approaches exist to convey the goals that robots should achieve, such as language descriptions, goal images, or goal videos. Language instructions [2, 23, 24, 31, 33, 38, 41, 45, 46, 56] can be ambiguous and brief, making it challenging for the robot to understand the tasks, or they can be overly detailed, increasing the difficulty for the model to follow. Goal images [6, 7, 25, 39, 61], while providing an accurate target, often contain extraneous informa + +tion irrelevant to the task, such as background elements and non-interactive objects. Some methods use human demonstration videos [10] or generated videos [6, 13, 14, 57] to outline tasks step-by-step. However, human demonstration videos are burdensome to encode, and generated videos depend heavily on their quality. To address these challenges, several works propose using visual prompts as a convenient yet expressive modality for goal specification. These visual prompts are easy for users to create and can effectively convey the precise goals that the policy model should focus on. + +Among visual prompt-conditioned approaches, methods [20, 32, 42, 47, 59] have attempted to convey task goals more effectively. The RT-Sketch [49] method emphasizes drawing the target state of the most relevant object to represent the goal. However, it only depicts the final state and overlooks the intermediate key frames that are crucial for successful task execution. In contrast, RT-Trajectory [20] (Figure1.(c)) illustrates the entire movement path of the end-effector, which helps bridge the gap between task components and enhances generalization. Despite its advantages, they provide only positional information and neglect the action directional information, which is also critical for accurate task completion. Additionally, as trajectories become longer and overlap between key-frames, they can create confusion for the model regarding overall task planning. + +This leads us to consider: is there any way to precisely and non-redundantly convey the goal while also clearly communicating the end-effector's action? Since key-frames represent important or bottleneck steps of the gripper during the task execution [18, 19, 26, 46, 58], we propose CrayonRobo, an approach that automatically generates or draws crayon-style visual prompts on images to effectively represent both low-level actions and high-level planning within key-frame sequences. As illustrated in Figure 1(a), these prompts can include contact point, end-effector z-axis direction, end-effector y-axis direction, and next moving direction after contact, conveying the objective of where and how to interact with the object. We further develop a training strategy and supervision objectives aimed at progressively enhancing + +![](images/bf864b70135829d4bbe6aaf3206d7d193f4bed34b35cbe533095d8f6c1ebe4db.jpg) +Figure 1. (a) shows our expression of different color prompts. (b) shows that we utilize a sequence of images with crayon visual prompts to express the key-frame steps, with each step illustrating the required low-level task goals, i.e., t1-pick, t2-place. Based on the input goal prompt, the model determines the 6 DoF contact pose, enabling it to interact with the object as required. When a yellow prompt is presented, the model also predicts 3D movement directions, guiding the movement after contact, e.g., picking upward in t1. For simple steps, such as t2-place, there is no need to present the post-contact moving direction. By sequentially executing each step in the key-frame sequence, the overall task is completed. (c) shows its differences with RT-trajectory [9] + +![](images/49146f2342d829b4ff70edca6b36f1866594fb14e8e725709d4e5c2ad4ce8910.jpg) + +the model's ability to comprehend each type of prompt. This allows the model to predict the corresponding SE(3) contact pose and the 3D next movement direction after contact, facilitating the completion of each sub-goal task. Furthermore, as illustrated in Figure 1(b), we use key-frame sequences as input to express the task planning procedure, with each image presented with crayon visual prompts indicating its key-frame task goal. We leverage the predicted 3D movement direction to connect each step sequentially, completing the overall long-horizon task. This approach enables us to instruct the model's task objectives while also enhancing its robustness when encountering novel tasks by providing model-understandable prompts. + +Our experimental setup includes a diverse range of manipulation tasks, both familiar and novel, where our method achieves a promising success rate in manipulation. Additional demonstrations are available at https://sites.google.com/view/crayonrobo. + +In summary, our contributions are as follows: + +- 1) We propose employing a sequence of key-frames presented with prompts to explicitly convey the task objectives in both low-level action and high-level planning. +- 2) We train a model that comprehends the visual-language prompts and predicts accurate contact poses along with moving directions in SE(3) space, ensuring the reliable completion of each key-frame task and the connection between steps to formulate a long-horizon task. +- 3) Experiments at scale demonstrate its promising performance and robustness. + +# 2. Related Work + +Pure Vision-based Manipulation. In vision-based robotic manipulation, numerous studies have employed a variety of solutions, including deep learning [8, 9, 18, 46], imitation learning [11, 27, 28, 60], and reinforcement learning [3, 12, 37, 43]. For example, in deep learning-based solutions, some methods design action policy networks [4, 15, 40, 52, 55] to calculate dense affordance maps and determine contact points and action poses. However, pure vision-based manipulation networks focus more on actionability than functionality, potentially failing when functional actions are required to meet human needs. + +Language-conditioned Manipulation. With advancements in language foundation models, language instructions are increasingly used for goal specification in goal-conditioned policy learning, as seen in [8]. Other works, such as [2, 23, 31, 33, 38, 41, 45, 46, 54, 56], utilize templated or freeform language for task specification. Building on this, Belkhale et al. [5] introduces language motions as an intermediate layer between high-level goals and low-level actions. Despite progress, challenges remain: language instructions often struggle to specify detailed actions or convey spatial objectives, frequently requiring human assistance to work effectively. + +Goal Image or Video-conditioned Manipulation. To enhance goal specificity and detail, several image-conditioned policy representations have been developed, with goal-image conditioning being one of the most prominent techniques. In goal-image conditioning, a final goal image defines the desired end state of a task [6, 7, 25, 39, 61]. This approach inputs both the initial and target states of the ob + +ject and outputs the actions needed to achieve the goal. In addition to static images, some methods use video or generated video [6, 13, 14, 57] to represent each step of the process frame by frame, offering a more detailed execution procedure. However, both solutions encode excessive information, much of which may be irrelevant to the task, such as background details or unrelated objects. + +Visual Prompt-conditioned Manipulation. To address the issue of redundant goal information, recent approaches have proposed using visual prompts as goals. Sundaresan et al. [49] suggests using goal sketches to indicate the target of the task-involved object, while Gu et al. [20] proposes drawing moving trajectories with key waypoints to specify the desired movement of the end-effector. Additionally, Nasiriany et al. [42] introduces a method that iteratively selects waypoints from a pool of potential options to form the overall trajectory, and Liu et al. [34], Stone et al. [47], Yang et al. [56] use external models to choose waypoints based on mark-based visual prompts to complete tasks. + +Different from them, we propose an innovative solution that allows manual drawing or automatic generating 2D prompts to represent both low-level action and high-level planning goals. This approach clearly conveys the objective of each key-frame step, i.e., where and how to interact with the object, as well as the required task-planning procedure. + +# 3. Method + +# 3.1. Problem Formulation + +The task is defined as follows: the model takes the visual input $I \in \mathbb{R}^{H \times W \times 3}$ of an object presented with crayon visual prompts. This prompt utilizes four colors: blue, representing the contact point; red and green, indicating the z-axis and y-axis directions of the end-effector when contacting the object; and yellow, symbolizing the next moving direction after contacting. To eliminate ambiguities caused by visual prompts overlapping each other, we specify their numerical values in text inputs as a language prompt $P$ , including the coordinates of the contact point denoted as $a_0^p \in \mathbb{R}^2$ , the gripper's 2D z-axis direction $a_0^z$ and y-axis direction $a_0^y$ , and the 2D moving direction $a_0^m$ , all represented as unit vectors in $\mathbb{R}^2$ . The objective of the model is to generate an action $a_0 = (a_0^{p'}, a_0^Z, a_0^Y, a_0^M)$ , where $a_0^{p'}$ is mapped into 3D position $a_0^P$ using the depth information and camera parameters, while $a_0^Z$ , $a_0^Y$ , and $a_0^M$ are unit vectors in $\mathbb{R}^3$ representing the end-effector's 3D z-axis, y-axis, and moving directions, respectively. $a_0^Z$ and $a_0^Y$ jointly determine the rotation for contacting the object, while $a_0^M$ controls the following movement after contact. Note that, the predicted $a_0^{p'}$ can be different from the $a_0^p$ in the text prompt input due to potential input noise. + +# 3.2. Data Collection + +In this section, we elaborate on how to obtain the 2D prompt required for training. In the simulator, building upon some effective rule-based heuristic approaches [31, 40], we interact with objects and record the success contact pose of the end-effector including 3D contact point, z-axis direction $a_0^{Z'}$ , and y-axis direction $a_0^{Y'}$ along with the moving direction $a_0^{M'}$ of the object contact point. These are then served as ground-truth to guide the training. + +As for the input prompt, 3D contact point is unprojected onto a 2D image as $a_0^p = (x,y)$ . We then adopt 3D contact point as the center and locate other 3D points along $a_0^{Z'}$ and $a_0^{Y'}$ . These points are then unprojected onto the 2D image. Subsequently, we connect these points with $a_0^p$ to calculate 2D vectors $a_0^z$ and $a_0^y$ , while obtaining red and green lines respectively. We obtain the yellow moving direction line and obtain the 2D moving direction vector $a_0^m$ in the same way utilizing $a_0^{M'}$ . These are then served as information embedded in the input language prompt. The corresponding visual prompts are generated by overlaying these elements in their respective colors onto the images. + +# 3.3. Training Strategy + +# 3.3.1. Model Architecture + +Given both the visual and language prompts, we require the model to effectively interpret inputs from both modalities. Observing the robust language understanding and visual processing capabilities of Vision Language Action Models (VLAs) and inspired by their applications in prior robotic manipulation tasks [29-31, 35], we follow the backbone architecture of an open-source VLA model [31] and adopt its parameter updating strategy which only finetunes the injected adapters [21] within LLaMa [51], alongside the multi-modal projection module, while keeping the primary parameters frozen. This approach aims to preserve the inherent existing pre-trained knowledge, particularly in sim-to-real transfer, while enhancing the model's ability to comprehend visual prompts and perform robotic manipulation in an efficient training time. Specifically, when presented with an RGB image containing visual prompts $I$ , we employ CLIP's visual encoder [44] to extract visual features. Simultaneously, text prompts $P$ are encoded into text features using LLaMa's pre-trained tokenizer [51]. The alignment of visual and text feature representations is achieved through the multi-modal projection module. + +# 3.3.2. Policy Learning + +Input pairs: To help the model better understand the meaning of various prompts, we design a series of visual-textual input pairs that progressively convey different levels of information. This gradual progression enables the model to develop a deeper understanding of the physical significance + +![](images/03cd495d0e4c0233c4092d6f84798942d7c0f7e85b53c2fa8c41418023199d48.jpg) +Figure 2. We design training pairs that convey varying levels of information to enable the model to comprehend each type of prompt and introduce loss functions to guide it in predicting accurate poses. + +of each component. Specifically, as illustrated in (a)-(d) in Figure 2 input. We formulate the problem of pose prediction as a language modeling task, directly outputting 3D directions in textual form. Thus, we generate the ground-truth text as shown in Figure 2 to guide the training. The entire language description is shown in Appendix.1. + +Training objectives: Our model's objective is to accurately predict 3D pose based on 2D visual-language prompts given different object and camera views. Therefore, we introduce the following losses to guide the policy training: + +Text Supervision Loss $\mathcal{L}_T$ : This loss ensures the effective alignment of the model's visual and linguistic input, making sure it can output the correct output pattern. We formulate pose prediction as a classification task by discretizing the continuous numbers in the normalized 3D direction vector into 100 discrete bins [-50,50], with each bin spanning 0.02, and adopt cross-entropy loss to supervise. + +Orthogonal Loss $\mathcal{L}_O$ : The orthogonality of a rotation matrix necessitates that its components maintain orthogonal relationships between each pair of directions. However, formulating direction prediction as language predictions for $a_0^Z$ and $a_0^Y$ does not explicitly consider their geometric relation. To address this, we extract $a_0^Z$ and $a_0^Y$ from the model's output text and ensure their orthogonal relationship under Gram-Schmidt loss [17]. + +Projection Loss $\mathcal{L}_P$ : 3D directional predictions should align with their corresponding 2D directional prompts when unprojected back to the 2D plane. To establish an explicit connection between the input 2D directional prompt and the output 3D directions, we introduce a projection loss designed to guide their correlation. Specifically, in the model's output, we extract the contact point $a_0^P$ and map it to 3D space using the depth map and camera parameters. Subse + +sequently, leveraging the predicted 3D direction in the $z$ -axis $a_0^Z$ , we locate 3D points along the direction with the contact point as the center. The 3D point are then unprojected back onto the 2D plane, and connecting with $a_0^P$ to generate predicted 2D directional vectors $a_0^{z*}$ . We employ a similar approach to acquire $a_0^{y*}$ and $a_0^{m*}$ , which are then supervised based on the 2D directional prompts $a_0^z$ , $a_0^y$ , and $a_0^m$ using cosine similarity loss, where $\cdot$ denotes the dot product: + +$$ +\mathcal {L} _ {P} = \left(1 - \frac {\mathbf {a} _ {\mathbf {0}} ^ {\mathbf {z} ^ {*}} \cdot \mathbf {a} _ {\mathbf {0}} ^ {\mathbf {z}}}{\| \mathbf {a} _ {\mathbf {0}} ^ {\mathbf {z} ^ {*}} \| \| \mathbf {a} _ {\mathbf {0}} ^ {\mathbf {z}} \|}\right) + \left(1 - \frac {\mathbf {a} _ {\mathbf {0}} ^ {\mathbf {y} ^ {*}} \cdot \mathbf {a} _ {\mathbf {0}} ^ {\mathbf {y}}}{\| \mathbf {a} _ {\mathbf {0}} ^ {\mathbf {y} ^ {*}} \| \| \mathbf {a} _ {\mathbf {0}} ^ {\mathbf {y}} \|}\right) + \left(1 - \frac {\mathbf {a} _ {\mathbf {0}} ^ {\mathbf {m} ^ {*}} \cdot \mathbf {a} _ {\mathbf {0}} ^ {\mathbf {m}}}{\| \mathbf {a} _ {\mathbf {0}} ^ {\mathbf {m} ^ {*}} \| \| \mathbf {a} _ {\mathbf {0}} ^ {\mathbf {m}} \|}\right) +$$ + +Note that, $\mathcal{L}_T$ and $\mathcal{L}_O$ are consistent across all input pairs, but $\mathcal{L}_P$ varies based on the input prompts. For example, in Input(c) of Figure. 2 where does not include the $a_0^m$ , we will exclude the supervision for $a_0^m$ and $a_0^{m*}$ in $\mathcal{L}_P$ . After completing the training process, the model, having been trained with dynamic input patterns, can understand each type of prompt and predict the required actions based on the provided prompts. The aforementioned losses are trained simultaneously under the total objective function: $\mathcal{L} = \lambda_1 * \mathcal{L}_T + \lambda_2 * \mathcal{L}_O + \lambda_3 * \mathcal{L}_P$ . + +# 3.4. Inference and Interaction + +# 3.4.1. Model Inference + +During the inference stage, we provide both automated and manual drawing approach to extract these visual prompts, as illustrated in Figure 3(i) and (ii). For automatic generation approach, we first use Grounded-DINO [36] to detect the object's bounding box and select its center, forming the blue circle. Then, we automatically generate 32 uniformly sampled 2D directional lines around the full 360-degree circle with the blue circle as the center or as the middle. + +![](images/9807b582f259f4802f7cffe66a70a2a872a1a6d882efe9624b82cf472d34a908.jpg) +Figure 3. Illustration of model inference with input generated in different ways. + +GPT-4 [1] is then prompted to select lines from all candidates to represent the gripper's z-axis direction, y-axis direction, and moving direction, resulting in the red, green, and yellow lines, respectively. As shown in Figure 3(ii), for manual drawing approach, we allow users to draw crayon visual prompts on the image, which then serve as the visual input. + +The contact pose is always required, while whether the movement direction after contact is needed depending on the task characteristics. Specifically, Goyal et al. [18, 19], Johns [26], Shridhar et al. [46], Yuan et al. [58] observe that long-horizon tasks usually can be decomposed into several key-frame tasks. Meanwhile, different key-frame tasks exhibit distinct motion patterns: some require a specific moving direction to guide the subsequent movement after contact, while others end when executing contact pose. For example, pick action requires moving along $a_0^M$ , which usually indicates an upward movement. Similarly, pushing and pulling actions require the robot to determine the direction in which to push or pull. In contrast, certain primitives, such as place and move, are simple motions that do not require additional movement after approaching the object. For example, given a pick-and-place task, once the object is picked up, the "preplace" pose is already established. Once the object is placed according to the contact pose prompt, the task is considered complete. Therefore, there is no need to draw a moving prompt for these primitives. For actions like rotate, such as rotating a button, the movement typically does not involve translating the position after contact. Instead, it focuses on rotating the last joint. As a result, $a_0^m$ , which specifies translation, is not needed. Instead, we provide two key-frames as prompts: one depicting the end effector pose at the moment of contact and the other after the rotation. From these + +two temporal prompts, the required rotation direction becomes evident. More visual details are demonstrated in the Appendix.4. After generating the visual prompt image, we automatically extract the coordinates of the contact point and the direction vectors based on the pixel RGB values and then incorporate them into the language prompt. Given the visual and language input, the model outputs the predicted action $a_0$ . + +Note that although 2D directions may overlap, the model does not require them to originate from the same point, as long as the direction is accurate and clearly visible in the image, the model can interpret it. Additionally, if the direction vector is perpendicular to the camera, it may be challenging to represent it as a 2D projection. In such cases, adding slight noise to the direction vector can enhance its visibility. Since the model is trained to interpret such situations, it can handle similar scenarios during inference, where the direction may be perpendicular to the camera. + +# 3.4.2. Interaction Strategy + +We illustrate how we enable the robot to interact with objects given the predicted action $a_0$ . Same with other VLA models [22, 31, 35], $a_0^p$ is projected into 3D space utilizing depth information and camera parameters. The $a_0^Z$ and $a_0^Y$ jointly contribute to determining the rotation matrix of the end-effector, facilitating the establishment of initial contact with the object. If provided with $a_0^m$ , we then follow the predicted moving direction $a_0^M$ to determine the subsequent movements after contact. + +For the long-horizon tasks, we leverage a sequence of key-frames with visual prompts to serve as high-level planning, with each frame representing a sub-goal step. By executing + +the key-frame steps sequentially, we can complete the long-horizon tasks. The benefit of this approach revolves around breaking down the complexity of long-horizon tasks and allowing us to optimize the success rate of each key-frame step to ensure the overall task's success. Meanwhile, regarding each key-frame step as compositions, we can group them into arbitrary combinations, enabling the model to handle diverse manipulation tasks. + +# 4. Experiment + +In our experiments, we mainly focus on exploring the following questions: + +- Section 4.3.1: What is the effect of different types of prompts on model performance? +- Section 4.3.2: How does the model handle noise in the input prompt, and is it user-friendly? +- Section 4.4: Can the model perform specific tasks without prompts, given certain input distributions? + +# 4.1. Setup Details + +Data Collection. Following previous work [31, 40], we utilize SAPIEN [53] along with the PartNet-Mobility dataset to construct an environment, interacting with about 1500 object shapes under various camera view randomization. The tasks involve several actions that require changes in both translation and rotation, e.g., pull the drawer, open the door, open the laptop lid, etc. The setup uses the flying Franka Panda Gripper for both the training and testing stages. We follow the procedure in Section 3.2 to collect prompts within the simulator, which takes about 6-8 hours to collect about 10,000 training samples. Details of the training and testing dataset split and camera view randomization can be found in Appendix.2. + +Evaluation Metric. We utilize the manipulation success rate to assess the effectiveness of the manipulation, calculated as the ratio of successfully manipulated samples to the total number of test samples. A successful sample is defined using a binary criterion, with success determined by thresholding the distance that the object part moves. + +# 4.2. Comparisons with Baselines + +To ensure fairness in comparison, all methods adhere to the same train and test split. + +Pure vision model Flowbot3D [15]: It predicts motion direction on the point cloud, denoting it as flow. The point with the largest flow magnitude serves as the interaction point, while the direction of the flow represents the end-effector's orientation and moving direction. + +Language conditioned model ManipLLM [31]: It takes the task description and the object's initial image, and then outputs the end-effector pose. We use its predicted contact point and rotation to contact with the object and adopt its active adaptation policy for moving. + +Vision Goal conditioned 3D model Implicit3D [61]: It develops a manipulation policy that utilizes the transporter to detect key-points for 3D objects. By providing the initial and target state point cloud of the object, key points are then used to determine the end-effector pose for manipulation. + +The above three methods use suction gripper as the end-effector, which is compared with ours(s) in Table. 1 + +Visual prompt conditioned model RT-trajectory [20]: Since the RT-Trajectory code is not publicly available at the time of this paper's submission, we replicate its method for comparisons with implementation details in Appendix.3. + +Goal video conditioned model AVDC [57]: The generative model is trained on our training split. During testing, we employ Gemini-1.5 [50] to evaluate the quality of the generated videos by assessing whether they successfully complete the given tasks, e.g., open the door. The ratio of successfully generated tasks is $10.2\%$ . Given that the performance of the generative video-based execution policy is highly dependent on the quality of the generated videos, we believe this score can somehow reflect its overall effectiveness during the following execution. + +In Table 1, both Ours(s) and Ours(f) demonstrate a substantial performance advantage over the baselines with manual generated prompts. For automatically generated prompts, the results are $0.64 / 0.62$ on seen and unseen tasks, still outperforming the baselines. It is important to note that by selecting the center of the bounding box generated by GroundingDiNO as the contact point input, there may be misalignment with the actual contact point, introducing noise into the positional prompt. However, our results demonstrate the robustness of CrayonRobo in handling such input inaccuracies. This is because the model is trained to manipulate objects, it can, to some extent, correct the noise in the prompts. Simulator visualizations are shown in the left of Figure 4, illustrating the prompt input, the robot's contact state with the object, and the final state after movement. We further conduct experiments involving tasks with multiple steps, such as pulling a door and subsequently pushing it. Our method results in satisfactory performance: 0.69 and 0.68 on seen and unseen tasks, showing its potential in handling long-horizon tasks. + +# 4.3. Ablation Study + +# 4.3.1. Analysis on The Effect of Different Types of Prompt + +In Table 2 Ex1-Ex3, since our model is able to handle various input patterns thanks to the proposed training strategy, we progressively introduce each type of prompt during testing. When no $a_0^m$ is provided, we use rule-based approach [40] to determine the movement after contact. Beginning with Ex1, where only a 2D position prompt is provided, the model achieves impressive performance with scores of 0.42/0.37. We compare our results with using AnyGrasp [16] to pre + +
MethodSeen Tasks
Flowbot0.760.860.080.670.260.050.580.290.710.350.070.230.400.630.520.04
ManipLLM0.680.640.360.770.260.620.650.610.380.520.530.400.640.710.830.64
Implicit3D0.530.580.350.550.300.660.580.510.310.410.450.340.690.540.310.43
Ours(s)0.890.850.670.910.870.500.920.850.870.850.540.840.850.920.850.75
RT-Traj0.560.580.280.450.560.500.650.420.290.560.400.490.460.530.520.27
Ours(f)0.710.790.580.810.790.430.810.810.750.780.500.750.760.790.810.73
MethodSeen TasksUnseen Tasks
AVG
Flowbot0.180.100.390.100.100.610.340.190.430.120.600.210.510.100.220.38
ManipLLM0.410.750.440.670.380.220.810.860.510.380.850.420.600.430.650.47
Implicit3D0.270.650.200.330.450.170.800.530.550.150.410.570.390.280.520.39
Ours(s)0.650.900.780.790.680.910.940.460.800.470.880.830.700.760.630.79
RT-Traj0.320.460.400.310.370.680.580.470.570.250.390.480.570.130.590.52
Ours(f)0.630.670.480.780.670.740.830.440.740.460.880.820.630.610.610.72
+ +![](images/50a2cf7fccc0b16eabf4f94108812bd5b84ac941da54077e49088c74646caf21.jpg) + +![](images/363b77cedff46e1818d97de754d2c670b0921d4803bf6bccfd349061376d9794.jpg) +(a) lift folding chair +(b) rotate faucet + +![](images/58f0c5944306cc9b3b23b3e4e06c92afc6da2f718a9ff88512d26b62ecedc501.jpg) + +t +Figure 4. Visualization results in SAPIEN simulator and real world. + +Table 1. Comparison of our method against baseline methods. (s) and (f) denote suction gripper and finger gripper, respectively. Bold text indicates the highest score within each end-effector type. + +
a0pa0zayam0SeenUnseen
Ex1✓---0.420.37
Ex2✓✓--0.550.50
Ex3✓✓✓-0.700.68
Ours✓✓✓✓0.740.72
Ours(w/o v.)✓✓✓✓0.690.68
+ +Table 2. Ablation Study. + +dict rotation given the same pixel coordinate, which results in lower scores of 0.35/0.31. This shows even without directional prompts, our model can accurately predict 3D poses, showcasing its ability to comprehend objects and predict appropriate poses based on positional prompts. Moreover, comparing Ex2 and Ex3, we observe that by introducing more prompts during testing, the model achieves more ac- + +curate predictions. + +Additionally, to investigate the differential effects of visual and language prompts, in the last row of Table 2, we enable the model to learn from images capturing only the object (without visual prompt), along with language containing instructional prompts. This configuration results in a performance drop compared to inputting prompts from both modalities. The reduced performance can be attributed to the difficulty the model has in effectively linking the given language prompt to the relevant objects. Consequently, it becomes challenging to make accurate object-centric manipulation predictions when relying solely on language prompts. We thus conclude that prompts from both modalities work together to enhance the model's understanding and improve its pose prediction capabilities. + +![](images/53912d65ed7dc485d181b90650592057568d9cb86a94cf83de76db608888c5ff.jpg) +Figure 5. Robustness analysis on the noise in prompts. + +
Open the trashcanOpen microwave*Lift lidWipe table*Heat toaster*
dra.9/107/105/58/103/5
auto.8/106/105/58/102/5
w/o P.-5/10-6/10-
+ +Table 3. Real-world success rate. * contains multiple steps. + +# 4.3.2. Tolerance Analysis of Prompt Noise + +We further introduce noise into the 2D directional prompt during testing while maintaining the accurate 2D contact point, in order to evaluate the model's tolerance to input disturbances. Random noise with a uniform distribution is added to the directional prompt, ranging from $10\%$ , $20\%$ , $30\%$ , to $40\%$ of the original directional value. The results, shown in Figure 5, indicate that with $10\%$ and $20\%$ noise, our method achieves performance levels comparable to those of the noise-free scenario. This demonstrates the model's ability to handle noisy inputs during testing. However, with $30\%$ and $40\%$ noise, performance degradation occurs as the directional values deviate significantly from their intended targets—deviations that would be unlikely to happen if drawn by a human, as they appear visibly unreasonable. Despite this, the model is still capable of successfully executing tasks, even when the directional inputs are imperfect or provided by a non-expert. This highlights the model's resilience, not only to noise but also to imperfect or non-ideal inputs, allowing it to perform effectively even when the prompts are not perfectly accurate. + +Ablation experiments regarding the effectiveness of each loss and failure case analysis are shown in Appendix.5 and Appendix.6. + +# 4.4. Real-world Experiment + +We conduct experiments involving interaction with various real-world objects without additional sim-to-real finetuning. Our setup includes a Franka Emika robotic arm equipped with its default parallel gripper, along with a RealSense 415 camera for capturing RGB images and depth maps. In our real-world experiment, we employ the workstation's built-in + +image editor to directly draw crayon visual prompts on the images or automatic generating prompts, denoted as "dra." and "auto." in Table. 3. We subsequently extract 2D vectors from the visual prompt, forming the text input. For each task, we select one or two shapes of objects and perform five trials each with a different camera view angle and object's initial pose. If the task has been completed, we consider the execution successful. To avoid a collision, we introduce a "pre-move" waypoint before reaching the predicted contact pose. This waypoint is positioned along the end-effector's z-axis, at a distance away from the predicted contact pose. Qualitative and quantitative results are presented in Table. 3 and the right part of Figure. 4. The results demonstrate that CrayonRobo possesses strong zero-shot capability on novel tasks, as it interprets task execution prompts rather than relying solely on images or language processing, which can be challenging to handle in novel scenarios. Meanwhile, it exhibits strong capability in long-horizon tasks, achieving a success rate comparable to that of single-step tasks. + +In the "w/o P." row of Table. 3, we further explore whether it is possible to fine-tune the model using prompts provided during the first set of executions for a specific task, and enable the model to perform subsequent executions independently, without requiring additional prompts. To investigate this, we selected two multi-step tasks: opening the microwave and wiping the table. The model is fine-tuned for 20 epochs using key frames and the corresponding predicted 3D poses from successful trials in the previous experiment as ground truth (GT), without providing any visual or text prompt inputs. During fine-tuning, the image input includes only the object, while the language input excludes all 2D prompts but incorporates the current robot state to ensure temporal consistency. The output format remains consistent with Figure 2. The fine-tuned model is then evaluated on these tasks, achieving success rates of 5/10 and 6/10, respectively. + +# 5. Conclusion + +We introduce a novel approach using comprehensive multimodal prompts that explicitly convey both low-level actions and high-level planning in a simple manner. We devise a training strategy that enables the model to comprehend the 2D visual-language prompt and predict accurate contact pose along with moving direction, ensuring the reliability of task execution. Such model-interpretable prompts show promising performance on both seen and unseen tasks, demonstrating its robustness. Limitation: As for the limitation, though our method can not directly avoid obstacles, we can incorporate collision-free motion planner library like curobo [48] to realize this to some extent. + +# Acknowledge + +This project was supported by National Youth Talent Support Program (8200800081) and National Natural Science Foundation of China (62376006). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5 +[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 1, 2 +[3] Boshi An, Yiran Geng, Kai Chen, Xiaqi Li, Qi Dou, and Hao Dong. Rgbmanip: Monocular image-based robotic manipulation through active object pose estimation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7748-7755. IEEE, 2024. 2 +[4] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13778-13790, 2023. 2 +[5] Suneel Belkhale, Tianli Ding, Ted Xiao, Pierre Sermanet, Quon Vuong, Jonathan Tompson, Yevgen Chebotar, Debidatta Dwibedi, and Dorsa Sadigh. Rt-h: Action hierarchies using language. arXiv preprint arXiv:2403.01823, 2024. 2 +[6] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models. arXiv preprint arXiv:2310.10639, 2023. 1, 2, 3 +[7] Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Devin, Alex X Lee, Maria Bauza, Todor Davchev, Yuxiang Zhou, Agrim Gupta, Akhil Raju, et al. Robocat: A self-improving foundation agent for robotic manipulation. arXiv preprint arXiv:2306.11706, 2023. 1, 2 +[8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2 +[9] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2 +[10] Elliot Chane-Sane, Cordelia Schmid, and Ivan Laptev. Learning video-conditioned policies for unseen manipulation tasks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 909-916. IEEE, 2023. 1 +[11] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffu + +sion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 2 +[12] Josef Dai, Xuehai Pan, Ruiyang Sun, Jiaming Ji, Xinbo Xu, Mickel Liu, Yizhou Wang, and Yaodong Yang. Safe rlhf: Safe reinforcement learning from human feedback. arXiv preprint arXiv:2310.12773, 2023. 2 +[13] Yilun Du, Mengjiao Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Brian Ichter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, et al. Video language planning. arXiv preprint arXiv:2310.10625, 2023. 1, 3 +[14] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3 +[15] Ben Eisner, Harry Zhang, and David Held. Flowbot3d: Learning 3d articulation flow to manipulate articulated objects. arXiv preprint arXiv:2205.04382, 2022. 2, 6 +[16] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 2023. 6 +[17] Luc Giraud, Julien Langou, and Miroslav Rozloznik. The loss of orthogonality in the gram-schmidt orthogonalization process. Computers & Mathematics with Applications, 50 (7):1069-1075, 2005. 4 +[18] Ankit Goyal, Jie Xu, Yijie Guo, Valts Blukis, Yu-Wei Chao, and Dieter Fox. Rvt: Robotic view transformer for 3d object manipulation. In Conference on Robot Learning, pages 694–710. PMLR, 2023. 1, 2, 5 +[19] Ankit Goyal, Valts Blukis, Jie Xu, Yijie Guo, Yu-Wei Chao, and Dieter Fox. Rvt-2: Learning precise manipulation from few demonstrations. arXiv preprint arXiv:2406.08545, 2024. 1, 5 +[20] Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, et al. Rt-trajectory: Robotic task generalization via hindsight trajectory sketches. arXiv preprint arXiv:2311.01977, 2023. 1, 3, 6 +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[22] Siyuan Huang, Haonan Chang, Yuhan Liu, Yimeng Zhu, Hao Dong, Peng Gao, Abdeslam Boullarias, and Hongsheng Li. A3vlm: Actionable articulation-aware vision language model. arXiv preprint arXiv:2406.07549, 2024. 5 +[23] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1, 2 +[24] Yueru Jia, Jiaming Liu, Sixiang Chen, Chenyang Gu, Zhilue Wang, Longzan Luo, Lily Lee, Pengwei Wang, Zhongyuan Wang, Renrui Zhang, et al. Lift3d foundation policy: Lifting 2d large-scale pretrained models for robust 3d robotic manipulation. arXiv preprint arXiv:2411.18623, 2024. 1 + +[25] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2(3):6, 2022. 1, 2 +[26] Edward Johns. Coarse-to-fine imitation learning: Robot manipulation from a single demonstration. In 2021 IEEE international conference on robotics and automation (ICRA), pages 4613-4619. IEEE, 2021. 1, 5 +[27] Yuanchen Ju, Kaizhe Hu, Guowei Zhang, Gu Zhang, Mingrun Jiang, and Huazhe Xu. Robo-abc: Affordance generalization beyond categories via semantic correspondence for robot manipulation. arXiv preprint arXiv:2401.07487, 2024. 2 +[28] Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations. arXiv preprint arXiv:2402.10885, 2024. 2 +[29] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024.3 +[30] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023. +[31] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18061-18070, 2024. 1, 2, 3, 5, 6 +[32] Yi Li, Yuquan Deng, Jesse Zhang, Joel Jang, Marius Memme, Raymond Yu, Caelan Reed Garrett, Fabio Ramos, Dieter Fox, Anqi Li, et al. Hamster: Hierarchical action models for openworld robot manipulation. arXiv preprint arXiv:2502.05485, 2025. 1 +[33] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 1, 2 +[34] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. arXiv preprint arXiv:2403.03174, 2024. 3 +[35] Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Lily Lee, Kaichen Zhou, Pengju An, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. Robomamba: Multimodal state space model for efficient robot reasoning and manipulation. arXiv preprint arXiv:2406.04339, 2024. 3, 5 +[36] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4 + +[37] Jianlan Luo, Zheyuan Hu, Charles Xu, You Liang Tan, Jacob Berg, Archit Sharma, Stefan Schaal, Chelsea Finn, Abhishek Gupta, and Sergey Levine. Serl: A software suite for sample-efficient robotic reinforcement learning. arXiv preprint arXiv:2401.16013, 2024. 2 +[38] Corey Lynch and Pierre Sermanet. Language conditioned imitation learning over unstructured data. arXiv preprint arXiv:2005.07648, 2020. 1, 2 +[39] Corey Lynch, Mohi Khansari, Ted Xiao, Vikash Kumar, Jonathan Tompson, Sergey Levine, and Pierre Sermanet. Learning latent plans from play. In Conference on robot learning, pages 1113-1132. PMLR, 2020. 1, 2 +[40] Kaichun Mo, Leonidas J Guibas, Mustafa Mukadam, Abhinav Gupta, and Shubham Tulsiani. Where2act: From pixels to actions for articulated 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6813-6823, 2021. 2, 3, 6 +[41] Suraj Nair, Eric Mitchell, Kevin Chen, Silvio Savarese, Chelsea Finn, et al. Learning language-conditioned robot behavior from offline data and crowd-sourced annotation. In Conference on Robot Learning, pages 1303-1315. PMLR, 2022. 1, 2 +[42] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv preprint arXiv:2402.07872, 2024. 1, 3 +[43] Hai Nguyen and Hung La. Review of deep reinforcement learning for robot manipulation. In 2019 Third IEEE international conference on robotic computing (IRC), pages 590-595. IEEE, 2019. 2 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[45] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on robot learning, pages 894-906. PMLR, 2022. 1, 2 +[46] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, pages 785–799. PMLR, 2023. 1, 2, 5 +[47] Austin Stone, Ted Xiao, Yao Lu, Keerthana Gopalakrishnan, Kuang-Huei Lee, Quan Vuong, Paul Wohlhart, Sean Kirmani, Brianna Zitkovich, Fei Xia, et al. Open-world object manipulation using pre-trained vision-language models. arXiv preprint arXiv:2303.00905, 2023. 1, 3 +[48] Balakumar Sundaralingam, Siva Kumar Sastry Hari, Adam Fishman, Caelan Garrett, Karl Van Wyk, Valts Blukis, Alexander Millane, Helen Oleynikova, Ankur Handa, Fabio Ramos, et al. Curobo: Parallelized collision-free robot motion generation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8112-8119. IEEE, 2023. 8 + +[49] Priya Sundaresan, Quan Vuong, Jiayuan Gu, Peng Xu, Ted Xiao, Sean Kirmani, Tianhe Yu, Michael Stark, Ajinkya Jain, Karol Hausman, et al. Rt-sketch: Goal-conditioned imitation learning from hand-drawn sketches. 2023. 1, 3 +[50] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 6 +[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[52] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 2 +[53] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6 +[54] Ted Xiao, Harris Chan, Pierre Sermanet, Ayzaan Wahid, Anthony Brohan, Karol Hausman, Sergey Levine, and Jonathan Thompson. Robotic skill acquisition via instruction augmentation with vision-language models. arXiv preprint arXiv:2211.11736, 2022. 2 +[55] Zhenjia Xu, Zhanpeng He, and Shuran Song. Universal manipulation policy network for articulated objects. IEEE Robotics and Automation Letters, 7(2):2447-2454, 2022. 2 +[56] Jiange Yang, Wenhui Tan, Chuhao Jin, Bei Liu, Jianlong Fu, Ruihua Song, and Limin Wang. Pave the way to grasp anything: Transferring foundation models for universal pickplace robots. arXiv preprint arXiv:2306.05716, 2023. 1, 2, 3 +[57] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 2023. 1, 3, 6 +[58] Wentao Yuan, Adithyavairavan Murali, Arsalan Mousavian, and Dieter Fox. M2t2: Multi-task masked transformer for object-centric pick and place. In 7th Annual Conference on Robot Learning, 2023. 1, 5 +[59] Wentao Yuan, Jiafei Duan, Valts Blukis, Wilbert Pumacay, Ranjay Krishna, Adithyavairavan Murali, Arsalan Mousavian, and Dieter Fox. Robopoint: A vision-language model for spatial affordance prediction for robotics. arXiv preprint arXiv:2406.10721, 2024. 1 +[60] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy. arXiv preprint arXiv:2403.03954, 2024. 2 +[61] Chengliang Zhong, Yuhang Zheng, Yupeng Zheng, Hao Zhao, Li Yi, Xiaodong Mu, Ling Wang, Pengfei Li, Guyue Zhou, Chao Yang, et al. 3d implicit transporter for temporally consistent keypoint discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3869-3880, 2023. 1, 2, 6 \ No newline at end of file diff --git a/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/images.zip b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c397db744a48cd1fdd1e88446af1b4b80049edc2 --- /dev/null +++ b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ce4af6440c2024bd1a6623f025c939a22e994fc638d2834ff674ed43211d6c9 +size 543607 diff --git a/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/layout.json b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4e0ffed5fb1148cd86d6978e7f158e6ec85d1bad --- /dev/null +++ b/objectcentricpromptdrivenvisionlanguageactionmodelforroboticmanipulation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6b5e83018d56ceede79cb8c92c1a5c8c3ff406caab271c547cd21d67c504024 +size 376583 diff --git a/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/1ec7f5cb-f521-4c84-b7b4-7504b783f3e2_content_list.json b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/1ec7f5cb-f521-4c84-b7b4-7504b783f3e2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..121a58f6f33147126fb4f78c892a031f51fab6f4 --- /dev/null +++ b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/1ec7f5cb-f521-4c84-b7b4-7504b783f3e2_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ec39ae69fba17f698aeb074f20df0f0f0dcbe8f23016c8dffc608ceae8afdd1 +size 78562 diff --git a/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/1ec7f5cb-f521-4c84-b7b4-7504b783f3e2_model.json b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/1ec7f5cb-f521-4c84-b7b4-7504b783f3e2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2e27841001e65e7de0ae0fbf2519f87a0d8d0245 --- /dev/null +++ b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/1ec7f5cb-f521-4c84-b7b4-7504b783f3e2_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34dc7ac2ab3a5ac52e6f98e66d478c0773f331ca30dd5f423ac77e06b4bed8a7 +size 96946 diff --git a/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/1ec7f5cb-f521-4c84-b7b4-7504b783f3e2_origin.pdf b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/1ec7f5cb-f521-4c84-b7b4-7504b783f3e2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d6e816ef03b49fd25f0561ab16e9648f2122b167 --- /dev/null +++ b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/1ec7f5cb-f521-4c84-b7b4-7504b783f3e2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:448820dfbd9c2592d44d15a10732e7bd09c583473dc1de3be96f20b49f8d5e77 +size 1996989 diff --git a/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/full.md b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b0ad09bdc76634eab2647e430b533d7f569a7892 --- /dev/null +++ b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/full.md @@ -0,0 +1,312 @@ +# Object Detection using Event Camera: A MoE Heat Conduction based Detector and A New Benchmark Dataset + +Xiao Wang $^{1}$ , Yu Jin $^{1}$ , Wentao Wu $^{2}$ , Wei Zhang $^{3}$ , Lin Zhu $^{4}$ , Bo Jiang $^{1*}$ , Yonghong Tian $^{3,5,6}$ + +1School of Computer Science and Technology, Anhui University, Hefei, China + +$^{2}$ School of Artificial Intelligence, Anhui University, Hefei, China + +$^{3}$ Peng Cheng Laboratory, Shenzhen, China + +4Beijing Institute of Technology, Beijing, China + +$^{5}$ National Key Laboratory for Multimedia Information Processing, + +School of Computer Science, Peking University, China + +$^{6}$ School of Electronic and Computer Engineering, Shenzhen Graduate School, Peking University, China + +{xiaowang,jiangbo} $@$ ahu.edu.cn，{jy0x4f,wuwentao0708} $@$ 163.com, + +zhangwei1213052@126.com, {linzhu, yhtian} @pku.edu.cn + +# Abstract + +Object detection in event streams has emerged as a cutting-edge research area, demonstrating superior performance in low-light conditions, scenarios with motion blur, and rapid movements. Current detectors leverage spiking neural networks, Transformers, or convolutional neural networks as their core architectures, each with its own set of limitations including restricted performance, high computational overhead, or limited local receptive fields. This paper introduces a novel MoE (Mixture of Experts) heat conduction-based object detection algorithm that strikingly balances accuracy and computational efficiency. Initially, we employ a stem network for event data embedding, followed by processing through our innovative MoE-HCO blocks. Each block integrates various expert modules to mimic heat conduction within event streams. Subsequently, an IoU-based query selection module is utilized for efficient token extraction, which is then channeled into a detection head for the final object detection process. Furthermore, we are pleased to introduce EvDET200K, a novel benchmark dataset for event-based object detection. Captured with a high-definition Prophesee EVK4-HD event camera, this dataset encompasses 10 distinct categories, 200,000 bounding boxes, and 10,054 samples, each spanning 2 to 5 seconds. We also provide comprehensive results from over 15 state-of-the-art detectors, offering a solid foundation for future research and comparison. The source code has been released on: https://github.com/Event-AHU/OpenEvDET + +# 1. Introduction + +Object detection aims to identify predefined target objects by delineating them with bounding boxes and assigning category labels. It stands as a cornerstone problem in computer vision and finds extensive application across fields such as intelligent video surveillance, autonomous vehicles, and industrial automation. With the advent of deep learning, a plethora of cutting-edge deep object detectors have emerged, demonstrating remarkable performance with RGB cameras. Notable examples include the RCNN variants [11, 12, 31],YOLO-based models [1, 16, 24, 33, 35, 41], and DETR-inspired detectors [3, 21, 46, 48, 52, 53]. Nonetheless, frame-based object detectors continue to struggle in demanding conditions such as low light, complex backgrounds, and rapid motion. The constraints in image quality inherent in traditional frame cameras, which capture images at a fixed frame rate (e.g., 30 FPS) and employ a uniform exposure setting, are largely to blame for the prevalence of missed and erroneous detections across various detection algorithms. + +To overcome the limitations of traditional sensors, researchers have turned to innovative technologies for object detection. Among these, bio-inspired event cameras, also known as Dynamic Vision Sensors (DVS), have garnered significant interest within the computer vision sphere. Event cameras outshine conventional RGB frame-based cameras in several aspects: high dynamic range, high temporal resolution, low energy consumption, and nearly no motion blur. These sensors have been applied to a range of tasks, from high-level applications such as event-based tracking [38], recognition [39], and captioning [37, 47]; to + +![](images/ca68061077b62551ae18b677d729a35e9f280047f5e7e85f4a7f75a7ed32918e.jpg) +Figure 1. (a). Comparison of existing datasets and our proposed EvDET200K dataset for event stream based detection, the bubble size represents the scale of the dataset; (b). Comparison of our proposed MvHeat-DET and existing SOTA detectors on the EvDET200K dataset, the bubble size represents parameters. + +![](images/03733125263e74e9e72bd5465959a995f84aaf30ca504d2888275f6812210daa.jpg) + +low-level operations including image enhancement and reconstruction [51]. In the realm of event-based object detection, DAGr [9], introduced by Daniel et al., integrates RGB frames with event streams to strike an optimal balance between latency and bandwidth, all while maintaining detection precision. SpikeYOLO [24] has made strides by incorporating the I-LIF spiking neuron, along with integer training and spike-driven inference, to minimize quantization errors in Spiking Neural Networks (SNNs). + +Despite the notable advancements made, current event-based object detectors still face the following issues: 1). Most event-based detectors rely on CNNs (Convolutional Neural Networks) or Transformers as their backbone architectures. However, CNNs are constrained by their local receptive fields, which hinders their ability to capture long-range and intricate dependencies. In contrast, Transformer-based vision models, such as ViT [6], grapple with high computational complexity, scaling as $\mathcal{O}(N^2)$ , and lack interpretability. 2). Some researchers resort to bio-inspired SNNs (Spiking Neural Networks) [24, 33] for encoding event streams, reaping benefits in terms of energy efficiency. Nonetheless, their overall performance lags significantly behind that of ANN (Artificial Neural Network)-based detectors. Consequently, the quest for an effective, efficient, and interpretable event-based object detection algorithm remains a formidable challenge. + +Recently, Wang et al. proposed a physics-inspired vision backbone model, vHeat [42], which is grounded in the principles of heat conduction. The core module of this model is the Heat Conduction Operator (HCO), which envisions image patches as heat sources and conceptualizes the determination of their correlations as the process of ther + +mal energy diffusion. By employing 2D Discrete Cosine Transform (DCT) and Inverse Discrete Cosine Transform (IDCT) operations to approximate the HCO, they achieve a lower computational complexity of $\mathcal{O}(N^{1.5})$ . When applied to object detection, the HCO outperforms both the Swin-Transformer [22] and ConvNeXt [23]. However, we argue that the 2D DCT and IDCT operators may not be the most suitable for simulating heat conduction in the context of vision models. As these operators are designed for general signal processing, they may not fully capture the spatial and temporal dynamics inherent in visual data, which are critical for accurate object detection. This motivates us to explore alternative approaches that can more accurately and efficiently model the heat conduction process for event-based detection. + +In this paper, we propose a novel backbone network for event-based detection that consists of Mixture-of-Expert (MoE) based heat conduction operators, termed MvHeatDET. We split the input event streams into multiple clips and get their feature embeddings using a stem network. Then, we pass the event embeddings into an MHCO layer, as shown in Fig. 2, which first selects a transform branch using the policy network. Then, randomly initialized features FEs are fed into a linear layer to predict the thermal diffusivity and multiplied by the transformed representations. After that, an inverse operator is adopted for complete signal transformation. The obtained tokens will be fed into an IoU-based query selection module and detection head for object localization and recognition by following [48]. + +Although several event stream object detection datasets have been proposed, e.g., Gen1 [5] and 1Mpx [29], they are still relatively scarce compared to RGB frame based detec + +tion datasets. Therefore, this paper introduces a new benchmark dataset to fill this gap, named EvDET200K. It involves ten categories of target objects, including people, cars, bicycles, electric bicycles, basketball, ping pong, goose, cats, birds, and UAVs. The dataset was captured using the Prophee see EVK4-HD event camera and comprises 10,054 samples, each ranging from 2 to 5 seconds in duration. An extensive annotation effort has yielded 200,000 high-quality bounding boxes, and we provide over 15 state-of-the-art detectors for future research to benchmark against. We are confident that the introduction of EvDET200K, along with the associated benchmark algorithms, will mark a significant stride forward in the realm of event camera-based object detection. + +To sum up, we draw the contributions of this paper as the following three aspects: + +1). We introduce a novel Mixture-of-Experts (MoE)-based heat conduction framework, named MvHeat-DET, designed for event stream object detection. This framework strikingly balances performance, efficiency, and interpretability, offering an improved trade-off in these critical areas. +2). We present EvDET200K, a new high-definition benchmark dataset for event stream object detection. Comprising 10,054 samples captured with the Prophesee EVK4-HD camera, each sample spans 2 to 5 seconds and includes 200,000 bounding boxes spanning 10 distinct object categories. +3). We have re-trained and evaluated over 15 state-of-the-art (SOTA) object detectors, including models from the YOLO, RCNN, and DETR families, on the newly introduced EvDET200K dataset. This provides a comprehensive baseline for future research to compare and build upon. + +# 2. Related Works + +- RGB Frame based Detection. Object detection using RGB images has progressed significantly in recent years, primarily due to advancements in deep learning [20, 44, 49]. These detectors can be divided into three main streams, i.e., RCNN-based detectors [11, 12, 15, 31],YOLO series, and DETR-based detection algorithms. The DETR [3] model simplifies the object detection process, effectively eliminating the need for many manually designed components, such as non-maximum suppression (NMS) or anchor generation. Based on DETR, methods like Deformable DETR [52], Adaptive Clustering Transformer [50], PnP-DETR [36], and Sparse DETR [32] reduce computational resource consumption by applying sparse processing to the transformer's attention mechanism, achieving a favorable balance between efficiency and accuracy. Furthermore, models like Conditional DETR [25], Anchor DETR [40], Efficient DETR [45], and Dynamic DETR [4] leverage spatial prior knowledge to better focus on regions of interest + +(ROIs), significantly reducing learning difficulty. Notably, [34] proposes integrating R-CNN with Transformer to promote DETR convergence and yield better results. Different from these Transformer-based detectors, in this paper, we propose a novel MoE Heat Conduction based backbone network for event stream based detection algorithm. + +- Event Stream based Detection. Event-based object detection has recently gained traction, particularly with the development of neuromorphic sensors that operate in an asynchronous, event-driven manner. Due to the specific features of event streams, current researchers usually adopt SNN (Spiking Neural Networks), Graph Neural Networks (GNN), LSTM (Long-short Term Memory) for event-based object detection. To be specific, in methods based on SNNs, SpikingYOLO [13], SpikeYOLO [24], and EMSYOLO [33] combine YOLO with SNNs. SpikingYOLO is the first object detection model implemented in deep SNNs. SpikeYOLO introduces the I-LIF spiking neuron to reduce quantization errors in SNNs. SFOD [7] achieves multi-scale feature map fusion in SNNs for the first time, improving the model's ability to detect objects of various sizes. DAG-r [8] employs an efficient asynchronous graph neural network to handle event data, achieving a trade-off between bandwidth and latency. Additionally, AED [19] is a lightweight detector with fast detection speed, better suited to the high temporal resolution of event cameras. GET [27] introduces the EDSA block, which effectively extracts features and enables feature communication in the spatial and time-polarity domains. S5-ViT [54] performs temporal aggregation using a state-space model (SSM), addressing the challenge of RNNs' limited generalization when handling inputs with varying frequencies. + +# 3. MvHeat-DET + +# 3.1. Preliminaries: Physical Heat Conduction + +In an ideal scenario, the temperature at a point with coordinates $(x,y)$ in a two-dimensional object at time $t$ , marked $u(x,y,t)$ , is governed by the following heat conduction equation in an isotropic medium: + +$$ +\frac {\partial u (x , y , t)}{\partial t} = k \left(\frac {\partial^ {2} u (x , y , t)}{\partial x ^ {2}} + \frac {\partial^ {2} u (x , y , t)}{\partial y ^ {2}}\right) = k \left(u _ {x x} + u _ {y y}\right) \tag {1} +$$ + +where $k$ is the thermal diffusivity, which measures the efficiency of heat diffusion within the material. To find the general solution of Eq. 1, we apply the Fourier transform to both sides of the equation, rewriting it as: + +$$ +\mathcal {F} \frac {\partial u (x , y , t)}{\partial t} = k \mathcal {F} \left(u _ {x x} + u _ {y y}\right) \tag {2} +$$ + +where $\mathcal{F}$ is the Fourier transform function. Let $\hat{u} (v_x,v_y,t)$ denotes the Fourier transform of $u(x,y,t)$ , and $v_{x}$ and $v_{y}$ are the frequency variables in Fourier space. This transfor + +mation converts the partial differential equation into an algebraic equation, which is simpler to solve. Therefore, we can rewrite the Eq. 2 as: + +$$ +\frac {\partial \hat {u} \left(v _ {x} , v _ {y} , t\right)}{\partial t} = - k \left(v _ {x} ^ {2} + v _ {y} ^ {2}\right) \hat {u} \left(v _ {x}, v _ {y}, t\right) \tag {3} +$$ + +where $t = 0$ represent the initial state of the object, i.e., $u(x,y,t)\mid_{t = 0}$ . For short, we use $f(x,y)$ instead, and $\hat{f} (v_x,v_y)$ denotes the FT-transformed $f(x,y)$ . By setting the initial state in Eq. 3, we can get the following solution: + +$$ +\hat {u} (v _ {x}, v _ {y}, t) = \hat {f} (v _ {x}, v _ {y}) e ^ {- k (v _ {x} ^ {2} + v _ {y} ^ {2}) t} \tag {4} +$$ + +To obtain a general solution of the heat equation in the spatial domain, we apply the inverse Fourier transform (denoted as $\mathcal{F}^{-1}$ ) to Eq. 4 and get the following expression: + +$$ +u (x, y, t) = \mathcal {F} ^ {- 1} \left(\hat {f} \left(v _ {x}, v _ {y}\right) e ^ {- k \left(v _ {x} ^ {2} + v _ {y} ^ {2}\right) t}\right) \tag {5} +$$ + +Inspired by the aforementioned process, Wang et al. propose a new vision backbone vHeat [42] which is built based on Heat Conduction Operator (HCO). They adopt the 2D discrete cosine transformation $DCT_{2D}$ and the 2D inverse discrete cosine transformation $IDCT_{2D}$ to simulate the HCO process in the visual domain. Despite good results that can be obtained, we think it can be further extended as the DCT-IDCT transformation may not be optimal for such a simulation. In the following subsections, we will introduce our MoE heat conduction based backbone network for event-based object detection. + +# 3.2. Overview + +As shown in Fig. 2, given the event streams, we first stack into event frames and get the event embeddings using a stem network. Then, we feed the event embeddings into the MoE-HCO blocks which provides multiple transform candidates. In this work, we consider DFT-IDFT, DCT-IDCT, and HT-IHT as three experts for the validation. A policy network with Gumbel Softmax is utilized for expert selection. In addition, the Frequency Embeddings (FEs) are used to predict the thermal diffusivity and multiplied by the transformed frequency representations. More importantly, we adopt an IoU-based query selection module [48] to find the key tokens for final detection. + +# 3.3. MyHeat Backbone Network + +Drawing inspiration from the physical concept of thermal conduction, we investigate the spread of image features across spatial domains through the lens of heat diffusion, culminating in the development of the MvHeat backbone network. This network harnesses the MoE (Mixture of Experts) Heat Conduction Operation, a novel framework that adapts thermal conduction principles for the processing of + +discrete visual data features. The MvHeat backbone is engineered to deliver specialized processing for unique image features, thereby enhancing their integration and analysis. + +- MoE-HCO Block. The MvHeat Encoder is structured into four stages, each comprising $L_{i}, i = \{1,2,3,4\}$ MHCO (MoE Heat Conduction Operation) Layers. As the data progresses through each stage, the spatial resolution is reduced by half via downsampling, followed by processing through multiple MHCO layers. The MHCO architecture is closely similar to Vision Transformer (ViT), with a pivotal distinction: it substitutes ViT's attention blocks with our innovative method while retaining the rest of the architectural framework. The efficacy of this structure has been validated by prior Transformer-based research, ensuring both scalability and a reduction in the computational burden associated with traditional attention mechanisms. In the MHCO module, input data is initially funneled through a selection mechanism to identify the optimal expert branch for the current feature set. Subsequently, the thermal diffusivity $k$ is learned through Feature Embeddings (FEs) and multiplied by a coefficient matrix $e^{-k(v_x^2 + v_y^2)t}$ to generate an intermediate output. This intermediate result is then multiplied by the input data transformed into the frequency domain before being reconverted to the temporal domain. + +- MHCO: MoE Heat Conduction Operator. As described in the section on Physical Heat Conduction, we design the MoE Heat Conduction Operator (MHCO) module to extract visual features by fully simulating thermal diffusion. This module enables effective visual feature extraction and facilitates the exchange of visual information between different image patches. Specifically, we first use deep convolution to extend the temperature distribution in the two-dimensional space along the channel dimension, with the resulting multi-channel features denoted as $U_{0}$ . Then, to obtain the output $U_{t}$ after thermal diffusion, we apply Eq 5: + +$$ +U _ {t} = \mathcal {F} ^ {- 1} \left(\mathcal {F} \left(U _ {0}\right) e ^ {- k \left(v _ {x} ^ {2} + v _ {y} ^ {2}\right) t}\right) \tag {6} +$$ + +Following a logical reasoning process, we utilize the Discrete Fourier Transform (DFT) to transform discrete image patch features into the frequency domain and then apply the inverse Fourier Transform (IDFT) to revert them back to the spatial domain. In the physical sciences, when considering that a medium does not occupy the entire space, a unique solution to the equation requires specifying boundary conditions for $u$ as indicated in Eq. 5. Additionally, each image patch can be viewed as a diffusion of features within a bounded space. Given that visual data is spatially constrained and semantic information does not propagate beyond the boundaries, a natural boundary condition arises. Here, we introduce a common Neumann boundary condition $D$ : + +$$ +\frac {\partial u (x , y , t)}{\partial \mathbf {n}} = 0, \forall (x, y) \in D, t > 0 \tag {7} +$$ + +![](images/9e87547d8e754490fd1e641ec58af679a1f623ad2e8a28d40e9fe20fff255e6a.jpg) +Figure 2. An overview of our proposed event-based object detection framework, termed MvHeat-DET. + +where $\mathbf{n}$ denotes the normal to the image boundary D. Additionally, because visual data is typically rectangular, this boundary condition enables us to perform reasonable transformations using the 2D Discrete Cosine Transform (DCT), 2D inverse Discrete Cosine Transform (IDCT), and Haar Transform (HT), inverse Haar Transform (IHT). Thus, Eq. 6 can be also rewritten as: + +$$ +U _ {t} = \mathcal {C} ^ {- 1} \left(\mathcal {C} \left(U _ {0}\right) e ^ {- k \left(v _ {x} ^ {2} + v _ {y} ^ {2}\right) t}\right) \tag {8} +$$ + +$$ +U _ {t} = \mathcal {H} ^ {- 1} \left(\mathcal {H} \left(U _ {0}\right) e ^ {- k \left(v _ {x} ^ {2} + v _ {y} ^ {2}\right) t}\right) \tag {9} +$$ + +where $\mathcal{C}$ denotes DCT, $\mathcal{C}^{-1}$ denotes IDCT and $\mathcal{H}$ denotes HT, $\mathcal{H}^{-1}$ denotes IHT. We attempt to construct an expert network utilizing three methods simultaneously. This approach is motivated by the characteristics of transformations: Specifically, the DCT demonstrates superior performance in detecting small targets, while HT proves more effective in complex scenarios. The DFT, on the other hand, offers more generalized applicability. Consequently, the strategic integration of these transformations can lead to enhanced flexibility and superior overall performance in event detection applications. + +In physical heat conduction, the thermal diffusivity $k$ indicates how quickly heat spreads within a material. In visual heat conduction, we assume that the most noteworthy content in the image carries more "heat", and thus, visual heat should flow toward these regions. Naturally, the thermal diffusivity parameter $k$ should be learnable and adapt to the image content, enhancing the adaptability of heat diffusion to the learning of visual representations. After applying the DCT, DFT or HT transformation, the input data $x$ is con + +verted into the frequency domain (denoted as $\hat{x}$ ). Therefore, the learnable thermal diffusivity $k$ also needs to be derived from the frequency domain information. Inspired by the positional embeddings in ViT, we randomly initialize a Frequency Embeddings (FEs) with the same shape as $\hat{x}$ , which is then fed into a linear layer to predict the thermal diffusivity $k$ . Specifically, we set a fixed value for $t$ , and the FEs is utilized across each stage of the MvHeat network to enhance convergence throughout the training process. + +# 3.4. IQS: IoU-based Query Selection + +In the DETR model, object queries are a set of learnable embeddings. To reduce the difficulty of optimizing object queries, some studies have proposed query selection schemes, which typically leverage classification scores to select the top K features from the encoder. However, due to discrepancies between the distributions of classification scores and localization confidence, some predicted boxes with high classification scores may not be close to the ground truth. This results in a bias toward selecting boxes with high classification scores but low IoU scores, while overlooking boxes with lower classification scores but higher IoU scores. Such selection biases undermine the overall performance of the detector. During training, the IoU score is incorporated into the objective function of the classification branch, encouraging the model to associate high classification scores with ground-truth boxes that have high IoU scores. Thus, we can still select the top K based on classification scores to obtain higher-quality queries. The overall loss can be expressed as follows: + +$$ +\mathcal {L} (y, \hat {y}) = \mathcal {L} _ {b b o x} (b, \hat {b}) + \mathcal {L} _ {c l s} (I o U, c, \hat {c}) \tag {10} +$$ + +where $y = \{b, c\}$ denotes the ground truth boxes and categories, and $\hat{y} = \{\hat{b}, \hat{c}\}$ represents the predicted boxes and categories. + +# 4. EvDET200K Benchmark Dataset + +# 4.1. Protocols + +We aim to provide a good platform for the training and evaluation of event-based object detection. When constructing the EvDET200K benchmark dataset, we follow the following protocols: 1). Large-scale: With the deep integration of information technology, human production and life, large-scale datasets show an increasingly important position. In our work, we collect more than 10k event sequences, totaling about 200k objects from 10 classes. 2). Diversity: During the shooting process, we anticipated potential challenges and configured certain factors in advance. These factors may affect the performance of the data captured by the sensor in object detection tasks. More in detail, Multi-view, Multi-illumination, Multi-motion, Dynamic Background, Non-detection Interference are all considered when recording these event streams. 3). Small Object: We focus on enhancing the detection capability for small objects. Since small objects are often overlooked in detection tasks, we specifically plan to capture data from multiple perspectives to ensure diversity across different scenarios. Finally, the dataset contains $51\%$ small objects, providing a sufficient number of samples for training. + +# 4.2. Statistical Analysis + +The EvDET200K dataset comprises 10,054 video streams, annotated with 10 common object categories, totally 202,260 annotations. As shown in the upper left image of Fig. 3, the most annotated class is "people", with a total of 105,265 annotations. This is followed by "goose" and "car", which have 30,960 and 25,850 annotations, respectively. Among them, 2,949 videos are taken from dense scenes. Each video has a duration ranging from 2 to 5 seconds. The dataset is randomly divided into training, validation, and test subsets using the ratio 6:1:3, which contains 6,031, 1,002, and 3,021 video streams, respectively. + +# 4.3. Benchmark Baselines + +To establish a comprehensive benchmark dataset for event-based object detection, we have selected more than 15 SOTA or representative detectors for evaluation on our proposed dataset, including: 1). CNN-based Detectors: DetectorRS [30], RED [29], Mask R-CNN [12], RetinaNet [18], and Faster R-CNN [31], all using ResNet-50 as the backbone, as well asYOLO-style detectors likeYOLO v10 [35],YOLO v6 [16]. 2). Transformer-based Detectors: S5-ViT [54], SAST [28], RVT [10], Swin-T [22], DETR [3], and vHeat [42]. 3). SNN-based Detectors: + +![](images/0309c2d8e237886c67f6e4bf2a5b66372729beee31f3d6e22a58eb4093dee3d2.jpg) + +![](images/a17b9a1ee914e949d8cb30b9cad811271ecef16cae6695b1b8ec5b50c246b872.jpg) +Figure 3. Visualization of All Annotation Information in the Dataset. Top Left (Instance Count per Class) shows the number of instances for each class in the whole dataset; Top Right (Bounding Box Size Distribution) illustrates the distribution of bounding box sizes across the dataset; Bottom Left (Object Center Distribution) shows the relative position (x, y coordinates) of object centers within the images. Bottom Right (Aspect Ratio Distribution) displays the distribution of width-to-height ratios of objects in the dataset. + +![](images/c888dfcf8c6f22fe2c660a61c3688e0b0c69db0d8ccb1e5a9ec535e54fc7c613.jpg) + +![](images/aaf0f677bcfec36bfd422dfcf92feb752f80bf74669a733b7fca970cf72d6cb2.jpg) + +Table 1. Experimental results on the N-Caltech101 dataset. + +
MethodsFormatmAP
NvS [17]Event Points34.6
YOLE [2]Event Frames39.8
Jeziorek et al. [14]Event Frames53.4
EAS-SNN [43]Event Points53.8
OursEvent Frames55.7
+ +Spiking neural network (SNN) detectors, such as spikeY-OLO [24] and EMS-YOLO [33], are also included for their event-based data processing capability. + +# 5. Experiments + +# 5.1. Dataset and Evaluation Metric + +In addition to the newly proposed EvDET200K dataset, we also conducted a comparison with several state-of-the-art detectors on the N-Caltech101 [26] dataset to validate the generalization capability of our method. The N-Caltech dataset contains 101 object categories and approximately 9,000 event streams, which are split into training and test sets in an 8:2 ratio. This dataset features complex and vari- + +Table 2. Experimental results on the newly proposed EvDET200K benchmark dataset. vHeat† means using vHeat as the encoder and Transformer as the decoder. + +
IndexAlgorithmPublishBackbonemAP@50:95mAP@50mAP@75PRParamsFLOPsFPSCode
01Faster R-CNN [31]TPAMI 2016ResNet5046.073.348.676.388.340.9M71.2G23URL
02S5-ViT [54]CVPR 2024Former+SSM42.976.344.169.766.918.2M5.6G84URL
03SAST [28]CVPR 2024Transformer27.453.625.330.129.618.5M15.9G51URL
04SpikeYOLO [24]ECCV 2024SNN41.274.839.881.668.568.8M78.1G77URL
05YOLOv10-N [35]arXiv 2024CNN42.775.142.175.368.92.3M8.2G116URL
YOLOv10-S [35]43.576.243.075.471.27.3M21.6G83
YOLOv10-M [35]44.077.542.876.671.715.4M59.1G32
YOLOv10-B [35]44.177.943.176.073.219.1M92.0G30
06RVT [10]CVPR 2023Transformer40.773.142.370.365.99.9M8.4G88URL
07EMS-YOLO [33]ICCV 2023SNN32.166.627.477.562.514.40M3.3M119URL
08YOLOv6 [16]arXiv 2022RepVGG41.375.738.950.453.817.2M44.2M70URL
09Swin-T [22]ICCV 2021Transformer49.078.452.779.488.8160M1043G26URL
10DetectoRS [30]CVPR 2021ResNet5049.178.853.578.885.8123.2M117.2G32URL
11RED [29]NeurIPS 2020ResNet5035.468.335.269.066.124.1M46.3G34URL
12DETR [3]ECCV 2020Transformer40.974.539.574.590.941M86G29URL
13Mask R-CNN [12]ICCV 2017ResNet5048.877.652.077.887.143.8M142.7G28URL
14RetinaNet [18]ICCV 2017ResNet5048.677.050.877.893.736.2M81.4G76URL
15vHeat† [42]arXiv 2024vHeat50.372.254.256.969.456.3M74.5G50URL
16Ours-MvHeat52.980.455.958.970.047.5M56.4G58-
+ +able backgrounds, which present significant challenges for detection algorithms. For evaluation metrics, we used the mean Average Precision (mAP) at different IoU thresholds, the most commonly used metric in object detection. We also report Precision and Recall to assess the accuracy of predictions and the ability to detect positive instances. Additionally, we measured the number of parameters, FLOPs, and FPS for each detector, providing a more comprehensive and accurate understanding of the models' performance. + +# 5.2. Compare With other Detectors + +Results on EvDET200K Dataset. As shown in Tab. 2, our baseline $vHeat^{\dagger}$ achieves 50.3/72.2/54.2 on mAP/mAP@50/mAP@75, meanwhile, our model MvHeat-DET achieves 52.9/80.4/55.9, which is significantly better than baseline. Obviously, our detector is also better than other SOTA detectors including R-CNN based methods, YOLO-based detectors, SNN-based detectors and Transformer-based methods. This result strongly demonstrates the effectiveness of our method. + +Results on N-Caltech Dataset. Tab. 1 presents experimental results on the N-Caltech101 dataset, comparing methods in terms of mean Average Precision (mAP). Among the methods listed, NvS and EAS-SNN (using event points) achieve 34.6/53.8, while YOLE and Jeziorek et al. (both using event frames) achieve mAPs of 39.8 and 53.4, respectively. Our model MvHeat-DET outperforms all others with a mAP of 55.7. + +# 5.3. Component Analysis + +We use DETR as the base model for component analysis. IQS denotes adding an IoU-based query selection strategy to the base model, vEnc. indicates replacing the transformer + +Table 3. Component Analysis on Our Proposed EvDET200K dataset. IQS means adding an IoU-based Query selection strategy, vEnc. means replace the encoder with vHeat Encoder, MoE means add MoE strategy to encoder. + +
IndexBaselineIQSvEnc.MoEmAP
1✓40.9
2✓✓41.6
3✓✓✓50.3
4✓✓✓✓52.9
+ +encoder with vHeat, and MoE represents using a multi-expert strategy. Tab. 3 shows that the base model achieves $40.9\mathrm{mAP}$ on the EvDET200K dataset. Adding IQS to optimize query selection raises the detection result to 41.6. Next, replacing DETR's encoder with vHeat significantly improves accuracy, reaching 50.3. Finally, the introduction of a multi-expert mechanism further boosts the model to 52.9. With each additional component, the mAP steadily increases, indicating that each component contributes to the model's performance. The notable improvements from IQS and vHeat suggest that these components significantly enhance feature extraction and information processing capabilities. The addition of MoE also further refines the model's performance. + +# 5.4. Ablation Study + +Analysis on Number of MHCO in Each Stage. Tab. 4 shows the effect of varying the number of MHCO modules on model performance, computational complexity, and parameter count. We set the input event stream resolution to $640 \times 640\mathrm{px}$ and varied the number of MHCO modules in the third stage. Specifically, increasing the number of + +Table 4. Ablation studies on Number of MHCO in Each Stage. + +
Number of MHCO(2,2,6,2)(2,2,12,2)(2,2,18,2)(2,2,24,2)
mAP52.652.952.953.4
FLOPs39.3G56.4G73.4G90.5G
Param36.2M47.5M65.1M70.1MG
+ +Table 5. Ablation studies on the number of experts. + +
Number of ExpertmAP
1 (DCT)50.3
2 (DCT+DFT)52.7
3 (DCT+DFT+HT)52.9
+ +MHCO modules had a positive impact on the experimental results (6 layers: 52.6, 12 layers: 52.9, 18 layers: 52.9, 24 layers: 53.4). While increasing the number of MHCO modules can improve model accuracy, both FLOPs and the number of parameters grow significantly, leading to a substantial rise in computational and storage demands. We choose the (2, 2, 12, 2) configuration for building the experimental model to achieve a balance between model accuracy and computational efficiency while managing resource consumption. + +Analysis on Number of Expert. In this section, we investigate the impact of varying the number of experts on the experimental results. As shown in Tab. 5, we select three transformations (DCT, DFT, and HT) as the experts. As the number of experts increases, the mAP gradually improves, rising from 50.3 with one expert (DCT) to 52.7 with two experts $(\mathrm{DCT} + \mathrm{DFT})$ , and further to 52.9 with three experts $(\mathrm{DCT} + \mathrm{DFT} + \mathrm{HT})$ . This indicates that increasing the number of experts helps enhance model performance, primarily because different types of experts provide diverse feature extraction capabilities. It also demonstrates the effectiveness of the proposed MHCO module. + +Analysis on Different Experts Combinations. Tab. 6 presents an ablation study on the performance of different frequency transformation combinations. When used individually, DFT, DCT, and HT achieve mAP scores of 48.8, 50.3 and 50.6 respectively. Notably, combining two transformations yields consistent improvements, reaching up to $52.7\mathrm{mAP}$ . The full combination of three experts achieves optimal performance with $52.9\mathrm{mAP}$ , $80.4\mathrm{mAP}@50$ , and $55.9\mathrm{mAP}@75$ , demonstrating the complementary nature of these experts. + +Analysis on thermal diffusivity $k$ . The Tab. 7 presents experimental results on different settings for the thermal diffusivity $k$ , evaluating its impact on model performance (mAP). The experiment is conducted using a model with layer configurations of (2, 2, 12, 2). When $k$ is fixed, the model achieves the mAP of 48.2. When $k$ is treated as a learnable parameter, the mAP increases to 49.1, indicating that allowing the model to automatically learn $k$ im + +Table 6. Ablation studies on different expert combinations. + +
NO.DFTDCTHTmAPmAP@50mAP@75
1✓48.871.547.6
2✓50.371.948.8
3✓50.678.852.8
4✓✓52.780.155.8
5✓✓52.480.055.1
6✓✓51.979.654.9
7✓✓✓52.980.455.9
+ +Table 7. Ablation studies on thermal diffusivity $k$ . + +
SettingsmAP
Fixed k = 548.2
k as learnable parameter49.1
Predicting k using FEs49.7
+ +proves performance to some extent. When $k$ is predicted using FEs, the mAP further increases to 49.7, achieving the best performance. Dynamic learning and prediction of thermal diffusivity $k$ , particularly through frequency embedding, significantly improve the model's accuracy and allow it to better adapt to frequency feature variations in the input data. + +[Note] More experimental results and visualizations can be found in our Supplementary Materials due to the limited space in this paper. + +# 6. Conclusion + +In this paper, we introduce a novel approach to event stream-based object detection, termed MvHeat-DET, which leverages a MoE-based heat conduction framework for efficient and interpretable feature extraction. This method balances performance, efficiency, and interoperability, offering a promising solution to the challenges faced by existing event-based detectors. Additionally, we propose a new high-definition dataset, EvDET200K, designed to advance research in this field by providing a comprehensive benchmark with diverse object categories and samples. We conclude by re-training more than 15 state-of-the-art object detectors on this dataset, paving the way for future advancements in event-based object detection. + +Acknowledgment: This work is supported by the National Natural Science Foundation of China (No. 62102205, 62302041, 62332002, 62027804, 61825101); Anhui Provincial Natural Science Foundation 2408085Y032; Natural Science Foundation of Anhui Province 2408085J037; Key Technologies R & D Program of Anhui Province (202423k09020039); Major Key Project of PCL (PCL2021A13) and the project of Peng Cheng Laboratory (PCL2023A08). The authors acknowledge the High-performance Computing Platform of Anhui University for providing computing resources. + +# References + +[1] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection, 2020. +[2] Marco Cannici, Marco Ciccone, Andrea Romanoni, and Matteo Matteucci. Asynchronous convolutional networks for object detection in neuromorphic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. +[4] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2988-2997, 2021. +[5] Pierre De Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499, 2020. +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. +[7] Yimeng Fan, Wei Zhang, Changsong Liu, Mingyang Li, and Wenrui Lu. Sfod: Spiking fusion object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17191-17200, 2024. +[8] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. arXiv preprint arXiv:2211.12324, 2022. +[9] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014):1034-1040, 2024. +[10] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13884-13893, 2023. +[11] Ross Girshick. Fast r-cnn. In 2015 IEEE International Conference on Computer Vision (ICCV), page 1440. IEEE, 2015. +[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. +[13] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. +[14] Tomasz Kryjak. Optimising graph representation for hardware implementation of graph convolutional networks for event-based vision. Design and Architectures for Signal and Image Processing, page 110. +[15] Trung-Nghia Le, Shintaro Ono, Akihiro Sugimoto, and Hiroshi Kawasaki. Attention r-cnn for accident detection. In + +2020 IEEE intelligent vehicles symposium (IV), pages 313-320. IEEE, 2020. +[16] Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, et al. Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976, 2022. +[17] Yijin Li, Han Zhou, Bangbang Yang, Ye Zhang, Zhaopeng Cui, Hujun Bao, and Guofeng Zhang. Graph-based asynchronous event processing for rapid object recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 934–943, 2021. +[18] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. +[19] Bingde Liu, Chang Xu, Wen Yang, Huai Yu, and Lei Yu. Motion robust high-speed light-weighted object detection with event camera. IEEE Transactions on Instrumentation and Measurement, 72:1-13, 2023. +[20] Chang Liu, Yongsheng Yuan, Xin Chen, Hutchuan Lu, and Dong Wang. Spatial-temporal initialization dilemma: towards realistic visual tracking. Visual Intelligence, 2(1):35, 2024. +[21] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. +[23] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. +[24] Xinhao Luo, Man Yao, Yuhong Chou, Bo Xu, and Guoqi Li. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection. arXiv preprint arXiv:2407.20708, 2024. +[25] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. +[26] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:437, 2015. +[27] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: Group event transformer for event-based vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6038-6048, 2023. +[28] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based + +object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2024. +[29] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. Advances in Neural Information Processing Systems, 33:16639-16652, 2020. +[30] Siyuan Qiao, Liang-Chieh Chen, and Alan Yuille. Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10213-10224, 2021. +[31] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. +[32] Byungseok Roh, JaeWoong Shin, Wuhyun Shin, and Saehoon Kim. Sparse detr: Efficient end-to-end object detection with learnable sparsity. arXiv preprint arXiv:2111.14330, 2021. +[33] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6555-6565, 2023. +[34] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3611-3620, 2021. +[35] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 37:107984-108011, 2024. +[36] Tao Wang, Li Yuan, Yunpeng Chen, Jiashi Feng, and Shuicheng Yan. Pnp-detr: Towards efficient visual analysis with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4661-4670, 2021. +[37] Xiao Wang, Yao Rong, Fuling Wang, Jianing Li, Lin Zhu, Bo Jiang, and Yaowei Wang. Event stream based sign language translation: A high-definition benchmark dataset and a new algorithm. arXiv preprint arXiv:2408.10488, 2024. +[38] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19248-19257, 2024. +[39] Xiao Wang, Zongzhen Wu, Bo Jiang, Zhimin Bao, Lin Zhu, Guoqi Li, Yaowei Wang, and Yonghong Tian. Hardvs: Revisiting human activity recognition with dynamic vision sensors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5615-5623, 2024. +[40] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. + +[41] Zeyu Wang, Chen Li, Huiying Xu, and Xinzhong Zhu. Mamba yolo: Ssms-based yolo for object detection. arXiv preprint arXiv:2406.05835, 2024. +[42] Zhaozhi Wang, Yue Liu, Yunfan Liu, Hongtian Yu, Yaowei Wang, Qixiang Ye, and Yunjie Tian. vheat: Building vision models upon heat conduction. arXiv preprint arXiv:2405.16555, 2024. +[43] Ziming Wang, Ziling Wang, Huaning Li, Lang Qin, Runhao Jiang, De Ma, and Huajin Tang. Eas-snn: End-to-end adaptive sampling and representation for event-based detection with recurrent spiking neural networks. arXiv preprint arXiv:2403.12574, 2024. +[44] Peilei Yan, Xuehu Liu, Pingping Zhang, and Hutchuan Lu. Learning convolutional multi-level transformers for image-based person re-identification. Visual Intelligence, 1(1):24, 2023. +[45] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. +[46] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. +[47] Pengyu Zhang, Hao Yin, Zeren Wang, Wenyue Chen, Shengming Li, Dong Wang, Huchuan Lu, and Xu Jia. Evsign: Sign language recognition and translation with streaming events. In European Conference on Computer Vision, pages 335-351. Springer, 2025. +[48] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16965-16974, 2024. +[49] Aihua Zheng, Juncong Liu, Zi Wang, Lili Huang, Chenglong Li, and Bing Yin. Visible-infrared person re-identification via specific and shared representations learning. Visual Intelligence, 1(1):29, 2023. +[50] Minghang Zheng, Peng Gao, Renrui Zhang, Kunchang Li, Xiaogang Wang, Hongsheng Li, and Hao Dong. End-to-end object detection with adaptive clustering transformer. arXiv preprint arXiv:2011.09315, 2020. +[51] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Ultra-high temporal resolution visual reconstruction from a fovea-like spike camera via spiking neuron model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1233-1249, 2022. +[52] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. +[53] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6748-6758, 2023. +[54] Nikola Zubic, Mathias Gehrig, and Davide Scaramuzza. State space models for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5819-5828, 2024. \ No newline at end of file diff --git a/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/images.zip b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab4b5e335fd88828cddb53540866bd5a4c67731c --- /dev/null +++ b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b0fe3bd8acb3aedb3f18ebd0f17f6dd523b7e3291d34d63886c926bcb1f9268 +size 488493 diff --git a/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/layout.json b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ea6489421070c6df8a38208cababb26d2478b86d --- /dev/null +++ b/objectdetectionusingeventcameraamoeheatconductionbaseddetectorandanewbenchmarkdataset/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20ad107727f81bb0d4ea1381a52768bcc6a7b25365453d34f29e6b88ce72b194 +size 359543 diff --git a/objectmovergenerativeobjectmovementwithvideoprior/30279836-e12f-449e-82ef-79ffe18be3fd_content_list.json b/objectmovergenerativeobjectmovementwithvideoprior/30279836-e12f-449e-82ef-79ffe18be3fd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a041cad9e496a5c41a79b8d72097e21d445a017 --- /dev/null +++ b/objectmovergenerativeobjectmovementwithvideoprior/30279836-e12f-449e-82ef-79ffe18be3fd_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f8b125a3cdd109e50a87066c0a2a64120a9c1d6ead6e499538335b923551305 +size 94113 diff --git a/objectmovergenerativeobjectmovementwithvideoprior/30279836-e12f-449e-82ef-79ffe18be3fd_model.json b/objectmovergenerativeobjectmovementwithvideoprior/30279836-e12f-449e-82ef-79ffe18be3fd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6292777723b06ac73ac1c9d484b31647c30e1d5d --- /dev/null +++ b/objectmovergenerativeobjectmovementwithvideoprior/30279836-e12f-449e-82ef-79ffe18be3fd_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73ffcbedeb0154a2dce6aa40697af3f710d762f2d24727281aa6b6387c4cc6ef +size 108651 diff --git a/objectmovergenerativeobjectmovementwithvideoprior/30279836-e12f-449e-82ef-79ffe18be3fd_origin.pdf b/objectmovergenerativeobjectmovementwithvideoprior/30279836-e12f-449e-82ef-79ffe18be3fd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2f5e589a127107b7381382230f30b5ee31bb8088 --- /dev/null +++ b/objectmovergenerativeobjectmovementwithvideoprior/30279836-e12f-449e-82ef-79ffe18be3fd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46c1636bdfe6f6396eebe6d72aa59f35f0960997523a10e2f50587abe96b6bdc +size 8827552 diff --git a/objectmovergenerativeobjectmovementwithvideoprior/full.md b/objectmovergenerativeobjectmovementwithvideoprior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..50f925cc260cf6e81dac277c0b649ae3a3ac1e6a --- /dev/null +++ b/objectmovergenerativeobjectmovementwithvideoprior/full.md @@ -0,0 +1,431 @@ +# ObjectMover: Generative Object Movement with Video Prior + +Xin Yu $^{1,*}$ Tianyu Wang $^{2}$ Soo Ye Kim $^{2}$ Paul Guerrero $^{2}$ Xi Chen $^{1}$ Qing Liu $^{2}$ Zhe Lin $^{2}$ Xiaojuan Qi $^{1,\dagger}$ $^{1}$ The University of Hong Kong $^{2}$ Adobe Research https://xinyu-andy.github.io/ObjMover + +![](images/fc906550d2bb9c0599dc6c69ab24a9a0fafcda152540cc7b70d6c9d0374d978f.jpg) + +![](images/acb36c1944044f3e03153c249884a8943abb4fc12d8cbe8e8804dfcb7ffbd4ee.jpg) + +![](images/a0ce9837cefc8d0a07cb5cf92f54f3f5f354e1ef1329720bffcd52ab4332b1c3.jpg) + +![](images/13bb3c7a72b4e78e0f99e1904fa292be312b91404bd76176fa9e129efc450519.jpg) +Figure 1. Results of object movement. We demonstrate the object movement capability of ObjectMover in a variety of complex and challenging scenarios. ObjectMover can well keep object identity, synchronously edit lighting and shadow effects, complete occluded parts, understand materials, adjust object perspective, and comprehend occlusion relationships to generate realistic images where the object has been moved, maintaining all other elements of the scene unchanged. + +![](images/3495e8e9593a81a3a98f5096515a2bff1f46357843e2fb933b29fee3d1635094.jpg) + +![](images/13e74eb3c0c1bfc7fa77209a961ee939921785450e42db45f165c4474f130d57.jpg) + +![](images/3013609d39dfaccb875c4db7cf251ffd03aa2754f58a46d92d8d072dae4429d4.jpg) + +![](images/57d65ecdf74b472221f5d35ceca1c14595cff331dbd57f9a7cb8890d4699de24.jpg) + +![](images/38f4ab936fe9f435950a669bca07621b6c52de159aec4f9a9fcea187155d854c.jpg) + +![](images/0b94534058cf0144098ad6d8fd3c46145cbb1f68ea45e52b94ea4e07d1d3855e.jpg) + +![](images/fd8c0fb9d4b425a90c2d43a2159c7333bc65a37b3bd03973588c0cb0200c37c0.jpg) + +![](images/0d0af646967bf76fcf714f1778dd1036daf024d41eb459831dd165b2e6150d89.jpg) + +![](images/2e05ad6dc0f960126842b1be0538656d9a3216bc9edb7b9e5a7b491026939ef4.jpg) + +![](images/aac3d0e8e19a13f20712ce5aa09a579613db6c9f82c9b33717e94f8e63d1f447.jpg) + +![](images/8c2af714f045792be5bbf2a38e46e7f72f681e914ff21b4fd1a2678a0f3a8c79.jpg) + +![](images/ab5d654fbd252bdce7459915f1703b35e2028bdaa41a1c8eec040109ab1fa03b.jpg) + +![](images/89092936081c2d01d875ac9729a55b3685b6611aa62d751e9c71449607d78d7e.jpg) + +![](images/9596775084c097c5d4e1f0b83a8ce5c70e1fe69e297e389ba3b66fd1d822988f.jpg) + +![](images/e8e7a93ed6b40f030d0a7fc84d381cb1e74cde7cb7a8d544c4ab46cabf8504fd.jpg) + +# Abstract + +Simple as it seems, moving an object to another location within an image is, in fact, a challenging image-editing task that requires re-harmonizing the lighting, adjusting the pose + +based on perspective, accurately filling occluded regions, and ensuring coherent synchronization of shadows and reflections while maintaining the object identity. In this paper, we present ObjectMover, a generative model that can perform object movement in highly challenging scenes. Our key insight is that we model this task as a sequence-to-sequence problem and fine-tune a video generation model to leverage its knowledge of consistent object generation across video frames. We show that with this approach, our model is able to adjust to complex real-world scenarios, handling extreme lighting harmonization and object effect movement. As large-scale data for object movement are unavailable, we construct a data generation pipeline using a modern game engine to synthesize high-quality data pairs. We further propose a multi-task learning strategy that enables training on real-world video data to improve the model generalization. Through extensive experiments, we demonstrate that ObjectMover achieves outstanding results and adapts well to real-world scenarios. + +# 1. Introduction + +Moving an object within an image is a fundamental task in the field of image editing, with applications spanning photography enhancement, content creation, and entertainment. This seemingly simple task is inherently complex as it needs to preserve the object's identity, seamlessly fill occluded regions, and update effects like shadows, reflections, and lighting to be consistent to the new object location (see Fig. 1). Additionally, to realistically render the object movement in a 3D scene as a 2D image, the perspective view of the object often needs to change or more generally, the pose of the object and its new surroundings may need to adapt to each other (see the last three examples in Fig. 1). + +Despite significant advancements in image editing [6, 13, 14, 20, 32, 33, 40, 42, 45], few methods can directly accomplish the task of object movement. A natural approach to solve this problem is to formulate it as two sequential image editing tasks: object removal [35, 45] at the source location and object insertion [6, 32, 33, 42] at the target location. However, for a complete removal of the object at its source location, first defining an appropriate region covering the object with its associated effects such as shadows and reflections can already be highly challenging. Moreover, as object insertion models are not explicitly trained for object movement within the same image, they will often noticeably modify the identity of the object and have inconsistent lighting or shadows compared to the object in the original location (see top row in Fig. 2 for artifacts resulting from this two-step approach). Another two-step approach involves first manually copy-and-pasting the object to a new location, followed by a model to harmonize it with the surrounding environment [1, 40]. However, this fails to account for natural + +![](images/a3c19505f76c735c43d41b64bf703826eca7fa684f048bbf3d29d5010b8af9f5.jpg) +Input + +![](images/150601fb407bf2bb2c7a93d8dd738aa3a919c01f59a2ca5887065b5e18897552.jpg) +PowerPaint (Removal) + +![](images/017e5a0cd0407ba644dbb5d55aaa6c08b8b3bc3e98ebbf92e3d9f60295674032.jpg) +Anydoor (Insertion) + +![](images/acc58a5950a605ec699474012d7005cbe837c6c42fa1343de193dc641e7f8f51.jpg) +Ours + +![](images/4f4650302b5cd5f431cc7b4ede45f504f6218b5e09f5cbbb2bc48b743b75ffa8.jpg) +Input +Figure 2. Limitations of existing methods for object movement. (Top): The approach of removing an object [45] and then re-inserting [6] it for repositioning leads to issues with identity preservation and ineffective synchronization of lighting effects. (Bottom): The copy-paste-based method [1] is unable to modify the perspective of the object. + +![](images/e32974034375d451924abb13f6e161ce5c05cc7d08985348aa6e326432cffbb4.jpg) +MagicFixup (Coarse Edit) + +![](images/b35c0ec4997cdda1b6f3af8f3f49e365608b3ce1eba70ad2b25668e059b20168.jpg) +MagicFixup (Final Result) + +![](images/f4b0aa88f25be85d718730ef5b445718baa2f9a6ff2732f80d8944b296e56de0.jpg) +Ours + +perspective changes for the object at the target location due to the naive copy-and-pasting and often fails to harmonize shadows (see Fig. 2, bottom). + +To address these limitations, we present ObjectMover, a single-stage approach for object movement based on a diffusion transformer model. Unlike conventional methods that rely on pre-trained text-to-image models for image editing [1, 6, 32, 33, 40], our approach reformulates the task as a sequence-to-sequence prediction problem and repurposes a pre-trained video diffusion model to solve the problem. By treating the input scene image, object of interest, user instructions, and target frame as a sequence of frames, our method can leverage learned video priors to capture the consistent evolution of lighting, object identity, and scene context across frames. This enables natural, consistent results, which image generative models - lacking such priors - often fail to achieve (see Fig. 3). + +One main challenge in training is the lack of paired data. To tackle this, we propose to use two complementary sources of data for training. Firstly, we propose to utilize game engines like Unreal Engine [8] to generate high-quality, aligned synthetic training data specifically designed for movement, which is crucial for transferring and adapting video prior to the target task. Another source of data is from incorporating natural videos to enhance the generalization and improve the performance in real-world scenarios. Despite not being perfectly aligned with our task due to issues like changing backgrounds and hence not directly usable for object movement, we adopt a multi-task learning strategy and train with videos on an auxiliary task of mask-based insertion, which enriches the model with real-world domain knowledge, leading to high-quality results, strong generalization ability (see Tab. 3 and Fig. 10). Meanwhile, we fully utilize the synthetic data by training on mask-free object insertion and removal, which equips our model with the ability to address multiple tasks within a single framework (see Fig. 8 and Fig. 9). + +Through extensive experiments, we demonstrate that $Ob$ - + +![](images/de58063f57bc14315cff17a624a0faeb7eb3a03e789d19824d75a4f12ffa7f89.jpg) +Input +Figure 3. Results trained with different priors. The model fine-tuned from a video prior outperforms the one from an image prior. + +![](images/3a1b1d7ae42db47e80e8174da4ef7e625f8d98254467ee5fda9dd545f8bbd9dc.jpg) +With Image Prior + +![](images/460d40b3713b587ec015e2100ab7bfd7fb12c8d112935c0272afeedde462bbf5.jpg) +With Video Prior + +jectMover achieves state-of-the-art results and outperforms existing methods by a large margin (see Fig. 7 and Tab. 1). In summary, our contributions are: + +- We introduce a novel framework that models object movement as a sequence-to-sequence problem and repurpose a video model for single-image editing tasks. +- To address data scarcity, we develop a synthetic data construction pipeline using modern game engines which produces a high-quality, high-resolution dataset suitable for object movement, removal, and insertion. +- We propose a multi-task learning pipeline that fully leverages different data sources across relevant tasks, enhancing target task performance with real-world video data and equipping the model with versatile capabilities in object removal and insertion. +- Through extensive experiments, we demonstrate that ObjectMover achieves superior results and outperforms existing methods. + +# 2. Related work + +Most related to our approach are image editing methods for generative removal, insertion, and movement. Additionally, motion-controlled video generators could implement generative movement by generating a video of the moving object. + +Generative Removal & Insertion. Inpainting methods based on diffusion models [2, 18, 19, 29, 34] are currently state-of-the-art for generative removal, where the object is removed by inpainting image region covered by the object. Similarly, generative insertion [6, 17, 32, 33, 42, 44] is currently dominated by diffusion-based inpainting that additionally guide the generated content to match a given object through identity preservation approaches. Most of these techniques maintain identity only roughly, which is particularly problematic for generative movement, where even minor discrepancies are noticeable. Furthermore, all inpainting-based insertion and removal methods fail to create or eliminate effects outside the inpainting mask, such as long shadows, reflections, or indirect lighting. Expanding the inpainting mask results in the loss of the original image's identity, as it is not maintained within the inpainting mask. To tackle this, certain methods focus on removing only shadows while preserving other scene attributes [10, 39], but they cannot eliminate + +other effects like reflections or indirect lighting, and they rely on an accurate shadow mask, which is challenging to obtain reliably. Recently, ObjectDrop [40] has been proposed as a mask-free object insertion and removal method, allowing one to address object movement in a two-step manner (i.e., first remove and then re-insertion). Nevertheless, this method primarily involves copying and pasting, treating movement as two separate editing tasks. This approach may lead to inconsistencies, unnatural perspectives, poses, or changes in lighting. Another recent insertion method [36] allows changes across the whole image when inserting objects, yet it does not support removal or movement. + +Generative Movement. Several image editing methods have recently been proposed to move an object shown in an image in a semantically plausible way. Typically, they generate an edited image by adding identity preservation and some form of control mechanism to a diffusion model. A popular control mechanism is a drag-based interface [3, 21, 22, 31], where the user defines where specific points in the image should end up. Other forms of control include motion fields [9], a coarse version of the edit [1, 40], or 3D-aware controls [4, 20, 24, 43] that also allow for user-driven changes to the 3D pose and perspective of an object. However, none of the methods have demonstrated strong context-driven changes of the object being moved, such as changes in perspective or lighting. + +Motion-Controlled Video Generators. Motion control for video generation shares the goal of moving an object in a scene. Recent methods allow for specifying the motion directions of full trajectories of objects in an image [30, 38]. However, trained on video data, these methods do not have a mechanism to ensure that unrelated parts of the scene do not change, such as lighting or secondary objects. The recent DragAnything [41] addresses this issue to some extent, but is still hard to achieve background preservation and context-driven changes compared to our approach (see Fig. 7). + +# 3. Method + +# 3.1. Overview + +There are two main goals for the object movement problem: (i) the complete removal of the object at the source location including its associated effects such as shadows or reflections, and (ii) the consistent insertion of the object at the target location. More specifically, the generated object should retain an identity that aligns with the source object while ensuring that its interactions with the surrounding environment are accurately updated. This often involves the appearance of a new shadow in the target area and its disappearance from the original location due to movement. For instance, when an object moves from sunlight to shade, it should blend with the darker environment; similarly, a person's shadow should be cast on the wall when they are closer to it. It is important to note that when there are several plausible ways + +![](images/5ca3cc43ced5721fb0f89710efe17ff6231b7d8972533d62569b4f0492a2b5a8.jpg) +Figure 4. Model architecture. (Left): Our overall sequence-to-sequence framework by leveraging an image-to-video prior for training our single-frame image editing task (Sec. 3.2). (Right): Frame formulation on different tasks for multi-task learning (Sec. 3.4). + +![](images/3e0502de65c262cb568e2f25d08a82f5c71a470aa74c3a39d93782bb67c7a5ce.jpg) + +of harmonizing an object at a new location to make the image look natural, unlike with object insertion, the harmonization in object movement should match the object at the original location to maintain consistent physical interaction between the static scene and moving objects. + +Our key insight is that object movement can be considered a special case of a two-frame video, where the object consistency and dynamic effect evolution between different frames have already been inherently learned in video generation models, as it was trained by watching multi-frame real-world events. Thus, instead of following conventional image-editing approaches that rely on an image diffusion model, we propose to leverage a video model as our foundation. Our approach consists of three primary components: First, we reformulate the task of object movement as a sequence-to-sequence prediction problem, enabling us to repurpose and fine-tune a pre-trained image-to-video model (Sec. 3.2). Second, to fine-tune the pre-trained video diffusion model for the target task, we use a game engine to construct a task-aligned synthetic dataset that is sufficiently diverse and capable of simulating instructive data pairs not typically found in video datasets (Sec. 3.3). Lastly, to fully leverage the potential of synthetic data and also real-world video data, we employ multi-task learning, incorporating related tasks that can be trained on both types of data in a unified manner, to enhance the generalization capability of our model (Sec. 3.4). + +# 3.2. Repurposing Video Diffusion Models + +We utilize a pre-trained image-to-video model based on a diffusion transformer architecture [25, 37] similar to Sora [23] and devise a sequence-to-sequence framework that is consistent with the original video model input, even though we only target at a single-frame image generation. As depicted in Fig. 4, we formulate all conditional information, + +including editing instructions, as image frames: (1) an input image $I_{1}$ ; (2) a foreground object image $I_{2}$ ; and (3) an instruction map $I_{3}$ . The instruction map $I_{3}$ is formed by layering two bounding boxes across separate channels such that $I_{3} = [M_{src}, M_{tar}, M_{tar}]$ with $M_{src}$ signifying the initial position and $M_{tar}$ the target position. Since we use a latent-diffusion [28] model, all image frames are converted into tokens within the latent space via a VAE encoder. Notationally, unless specified otherwise, the symbols $I_{1}, I_{2}, I_{3}$ , and related terms will refer to the VAE-encoded tokens rather than the original images. + +During inference, starting from pure noise $X^0\sim \mathcal{N}(0,1)$ a noise frame $X^{t}$ is introduced to combine with condition images to create a sequence $S^t = [I_1,I_2,I_3,X^t ]$ and our model $\nu_{\theta}$ iteratively denoises this sequence and finally predicts the clean image $X^{1}$ . At each denoising step, the model also produces an output sequence of equal length, $S^{t + \Delta t}:= [I_1^{t + \Delta t},I_2^{t + \Delta t},I_3^{t + \Delta t},X^{t + \Delta t}]$ . Before each subsequent denoising step, $I_1^{t + \Delta t},I_2^{t + \Delta t},I_3^{t + \Delta t}$ are replaced back with $I_{1},I_{2},I_{3}$ . + +The training employs a flow-matching loss [15, 16]. Specifically, we use linear interpolation to infuse noise into the ground-truth target, generating the noisy input as follows: + +$$ +X ^ {t} = t X ^ {1} + (1 - t) X ^ {0}, \tag {1} +$$ + +where $X^1$ is the ground-truth image and $X^0 \sim \mathcal{N}(0,1)$ represents Gaussian noise. The model is designed to predict the velocity $V^{t} = \frac{dX^{t}}{dt} = X^{t} - X^{0}$ , which can be converted to an estimated $X^{t + \Delta t}$ during inference. The loss function is formulated as: + +$$ +\mathcal {L} = \mathbb {E} _ {t, X ^ {0}, X ^ {1}} \left\| v (S ^ {t}, t; \theta) _ {[ 4 ]} - V ^ {t} \right\| ^ {2}, \tag {2} +$$ + +where $\nu_{[4]}$ is the fourth frame of the output sequence, focusing the loss calculation on the target frame. + +A key challenge, however, is the lack of access to the ground-truth image $X^1$ . In Sec. 3.3, we present a synthetic data simulation pipeline to support training, and in Sec. 3.4, we explore multi-task training on unpaired video data. + +# 3.3. Data Generation with a Game Engine + +As large-scale paired data for object movement are unavailable, we propose to construct the training data with a game engine, Unreal Engine [8]. As shown in Fig. 6, our data generation pipeline consists of three steps: (1) background scene generation; (2) movement template pre-configuration; and (3) object movement sequence generation. + +Step 1: Background Scene Generation. For the background scene, we use a collection of maps intended for gaming purposes to enhance realism, different from earlier simpler configurations [20] that depended on manually constructed scenes featuring a limited number of scene objects and simplistic environmental maps, as shown in Fig. 5. Using these maps has the following advantages: (1) scene particulars including lighting, background meshes, and assets are meticulously orchestrated by specialists to ensure high fidelity, resulting in realistic shadows, reflections, and background scenes; (2) expansive maps that allow versatile object placement for rendering across various locations within the same environment where diverse background images can be created even from a single map; (3) adjustable lighting parameters, such as intensity, temperature, and direction, for diverse lighting conditions. + +Step 2: Movement Template Pre-configuration. Manually placing objects at multiple locations within a background scene can be extremely time-consuming. To address this, we develop a variety of trajectory and camera position templates to automate object placement in $k$ intermediate locations along the trajectory, capturing intermediate frames in the process. Notably, direct interpolation and placement of objects can lead to clipping artifacts when objects penetrate the ground, violating physical placement principles. To prevent this, after determining each location, we perform ray casting to detect any intersections between the object and the ground (or other objects). We then automatically adjust the position and orientation to ensure the object's bottom surface aligns accurately with the ground. This adjustment is dynamically influenced by the map's local terrain, providing more variation and enhancing diversity. This step is a fully automated process to facilitate large-scale data generation. + +Step 3: Object Movement Sequence Generation. Finally, with different scene maps, objects and movement trajectories, we generate 18,783 distinct sequences with frame lengths of $k = 10$ or $k = 15$ , including one frame with a clean background devoid of objects for mask-free object insertion and removal training (see Sec. 3.3). This results in a total of 1,348,248 training pairs, $13 \times$ greater than the previous + +![](images/1ead13db9b5a0b3b82fbf824bd8e8b0c68a3d0564b2f64641f1bdc40034308ee.jpg) + +![](images/aa78fc8ebd0dffbf83176c818d3d94f1c46469a7b800a8175f6bb276f8aeefa1.jpg) + +![](images/4f9158ebfa29621d00c70cfac03bf13f133b057bee4b64a28f26a8e4a6f7c355.jpg) +Figure 5. Dataset comparison. Row (1-2) shows our synthetic data using a game engine; Row (3) shows existing synthetic data [20]. Our data is more realistic and has complex lighting effects. + +![](images/4ba9d5e235638dd75370e875dbdab00eda77b965a29fce41cfd1858c1a6667ca.jpg) + +![](images/28139bb9ffbb5ff41d3ae05c4c1e7b75ce350789af0d1274fea9fe6f44fff46b.jpg) + +![](images/b10817fda96d3a7616ae2026a76717c98797aea7e616f827d039c889b0d14777.jpg) + +![](images/f2d987ec46b5c2aabe78ed11e4f48de5090e2954b74e08ccd6184283bd99f1c5.jpg) + +![](images/964b8eeefcf8a8ed48db3b3cc72a1684c16aa9c88abc9bc58a49ea673b4e7550.jpg) + +![](images/79840a66b0621c3aceef5749e0c070c1153b9c42ea9cf46ab36ec9fc7c4b626f.jpg) + +![](images/ab279af4d61a0cf191f861afa13bbd538f1213aa72ac1c256332d2f13c9a6e6d.jpg) + +![](images/d4bafefdd673e16e5e8cdfe283b56b70112b88b18254719c835eb18fb3508371.jpg) +Figure 6. Data generation pipeline. The pipeline consists of three steps: (1) background scene generation; (2) movement template pre-configuration; and (3) object movement sequence generation. + +work [20]. Furthermore, the images are captured in high resolutions (1024 × 1024 and 1080 × 1920), unlike earlier datasets [20] (256 × 256), and thus supporting our high-resolution training for a more practical usage. + +# 3.4. Training with Video via Multi-Task Learning + +The benefit of the synthetic data generation pipeline described in Section 3.3 is clear: it enables the collection of diverse paired data dedicated to object movement under various lighting scenarios through intentional control of lighting parameters. Nevertheless, a limitation of the synthetic data is that it may still exhibit a domain gap from the natural image distribution and lack the real-world diversity and complexities of dynamic changes across frames. Hence, a good source of data for object movement can be natural videos with moving objects. For video data to serve our task, ideally, everything would remain static except for the moving object; however, this is not the case in most videos. Thus, we devise a multi-task learning (MTL) strategy that allows us to use both real video data and synthetic data. + +MTL on Synthetic Data. Our synthetic data includes full background images without the object, which not only supports object movement but also facilitates object removal and insertion tasks. It should be noted that existing compositing methods [6, 32] require masking out the background image, leading to significant limitations: a restricted generation scope, where the model can only generate within the masked area, thus preventing the creation of more realistic effects such as shadows and reflections; and unintended background alterations, where regenerating within the mask inadvertently alters original background regions. In contrast, our approach + +![](images/ab2780f7080256d31e3fc8db33e3723b688a0358e42d0c144c6e3c7c715a4a18.jpg) + +![](images/4d07cd4e730191a40cec2feacf51f5023a436f51a5730727aa0a89750dc517de.jpg) + +![](images/3fd828e715119a89c7d633145c80171b9df6a2bdc6e8e9a9f7abc38b68b142b3.jpg) + +![](images/d38550bae8acffd48147f15279d1f2979e17a20166b8389e920495eed5bbd457.jpg) + +![](images/da1068b7ecccb2690845804e49c9bb517cea146a3632607677988e6fcaa0f916.jpg) + +![](images/3cefdab625e352b9f25336442bda45bb5137e153d339e8e017a863564ac0b733.jpg) + +![](images/10dc52e498f32a0fffd0bdf59672ace322920f818646a79aea6dcc4604adddfa.jpg) + +![](images/5a38f87e4e12cc2e36e3f5d8521dab16b3269ef115368ccfc8518f6acbcb94b3.jpg) + +![](images/f5c3fdb8516c6af7ac9ef946c7f55a37cba30dc7406ebcd50d9cc6f22bfbe7e5.jpg) + +![](images/4280d373c5bcbcedf7500358135241ffb81bf4e1e96d3d8385c54a848e12efad.jpg) + +![](images/299da53364591ace07cd2c4bf1a04542c7cc2222dab0269387a4a3bb91071582.jpg) + +![](images/314a23780a9a076970a35ce7e6540eb2338f990149438cc7d6155f30350de15e.jpg) + +![](images/4c19c95eb7aeb8cd54b45853f8089724b0cc20b09050fe84e3f2469afb11de16.jpg) + +![](images/6d4c4bb2b1095743f24acb23c4e821328ce2d1c062d36be4fe31f727c0de8c95.jpg) + +![](images/a04876913e42b7a167c83a4713462549e367c776dc790dce87fd8ce0eb3bd649.jpg) + +![](images/58698bb01e7ce0893934903118d00f77fd5fa166bcd9cdd028735e2e36bbab4f.jpg) + +![](images/d04cda74b358b23b2be67e584ae7c619531c1b13e35e59ab98bd3744a0136aa0.jpg) + +![](images/962db823027a9e7dbdd6b9193103df693c13e37eb4ea39da457244c59f17b78a.jpg) + +![](images/70379ea70e978250032b7a37065518aef55de7715f98bea393c68fcd7d91c32b.jpg) + +![](images/38e0b6059d52b690d383a4e8856a679d680ba706d1ff239076f8c279c17ebe8f.jpg) + +![](images/2a7d87084652bd5da9e1536315c18d9210526b19a628a34fa4761dcf069d882f.jpg) + +![](images/05a66a302914919febb83a0aee0e55365c92d4eba795640541aa5c7bea63a5dc.jpg) + +![](images/e0ad4e25210a558d93a09a9d32135b34f061f5b7263ea78ad4b9745ee0eea7d2.jpg) + +![](images/8e0a33f24c6c1107a921265995eecfb318f2fde61d64d4385fa849acec505d52.jpg) + +![](images/b13fe550cdd1ba42d7eb3f385167083d7feff4420240ae4962bf6ba36d333a7b.jpg) + +![](images/f56114e39bb7de6cc9786291565d55a019831545259e06831ff83167dc7d0098.jpg) + +![](images/b70dbae215831f41a8c7b2e0c8b661e4a53c33c91d22f5e1fbcffd0f5c1622ba.jpg) + +![](images/8e952ecb087fc6180b0e4b33ef0391e7a8c302bfed4d3f3e5b3715c596b4a061.jpg) + +![](images/fca1914bf3e20bc150b35e50c6431e62051e98ecfbda46bacb69497113bd1c77.jpg) + +![](images/93fafcb34b1e0c117a6cb2ad0fe75d9c3a76cd0368d147f64c87c75f4b1be640.jpg) + +![](images/3776a533aca3940777808899a00963170232ec3b7bbe35e43c870084f731308a.jpg) + +![](images/1a602f2408df4b1ff709920c36a0ffff4261c0d81b4d36d23cfbf8bdda19c669.jpg) + +![](images/216c1edfe13eb85cf546b06bc54cb5e6e1988bce7ee01254a0e7398a4c618dbd.jpg) + +![](images/f2a15e2b2a7c824451ba8c398d78a75bcc9e905ebb7be0041c8ffc0df44e3cff.jpg) + +![](images/1f87c8bc59d06db54650cd7cef689fe034a8ea219ec62d554b064035cda1f66d.jpg) + +![](images/06169e5c750a28fc346686b5712ee305e9a438a1865d2484d0c17061181fc4cb.jpg) +Input + +![](images/4e026bd2c12886c74699900d6441ef89b863f6fc766aa018b8a17545ed2014c8.jpg) +Ours + +![](images/97fea103d16632a37b0d68e4a7f0aa724889c462306d196b89336ff8cfec91b2.jpg) +MagicFixup + +![](images/c9b5a1f5413231d93be25c0cccc94a4c00baff261e49e51f3c48844d5db9b2e8.jpg) +Anydoor + +![](images/eec8792127582c5194209a23205f518bb42050cac791d1ba6800a46c29ab0444.jpg) +Paint-by-Example +Figure 7. Qualitative comparisons on object movement. Our method consistently outperforms state-of-the-art methods in maintaining object identity, lighting consistency, and overall quality. Notably, in the last example, only our method successfully harmonizes with the target region, achieving a realistic effect with partial shadows and light. + +![](images/683c4c9eb4506395432315d3dc24da60859523ab88a4d8920c46697b637af02b.jpg) +3DiT + +![](images/71dc453246bee7b42069f732eec794483abab9e29d6e158fd0e1ce2bf6cbf699.jpg) +Drag-Anything + +does not require pre-masking the target location because full background images are available in our synthetic dataset, thereby avoiding the aforementioned limitations. On the synthetic data, our model is jointly trained for all three tasks of object removal, insertion, and movement. + +MTL on Video Data. As we lack full background frames for the video data, we incorporate an auxiliary task of mask-based insertion (i.e., masking out the background image as in conventional compositing models, wherein we replace the pixels within the mask region with black pixels) to integrate the video data into our training regimen. This enables our model to adapt to real-world content by training in natural videos and helps improve synthesize complex lighting effects. As illustrated on the right side of Fig. 4, our sequence-to-sequence framework consolidates the training of various tasks. + +In this framework, the edited image consistently corresponds to $I_{1}$ , the segmented object to $I_{2}$ , and the instruction map to $I_{3}$ . Specifically, for removal tasks, only the source location mask is employed in $I_{3}$ , while for insertion tasks, only the target location mask is utilized. For the insertion task, we use the object image from a frame different from the target image to compel the model to adapt to the new environment. Across all tasks, the noisy target input is positioned in the final frame. Besides, a unique task embedding is developed for each distinct task. + +# 4. Experiments + +Training Details. We train ObjectMover using an image-to-video model that is based on a transformer architecture and has 5B parameters. Note that the method can be applied + +![](images/8e7d0eecb2675d183ba44d7be10d23d4492a580e990b628296f61ac35ed278da.jpg) +Input + +![](images/cd215a6256e08953ef3f7e47f5cab8781e73620583f84a852214498f7a377393.jpg) +Ours + +![](images/25c689fba8c678a9d59100147cd4a76fda1509ce2f1c1123ce0f2cc4d5df0f2e.jpg) +SD-Inpaint + +![](images/e4689f05f4c5a933f8785f63a753635a5f6f0b6729e381e10063caa7fba5369d.jpg) +PowerPaint + +![](images/8e0f59b74422e041070b428294977879787d3f40aa689c3ca10e13a446a73232.jpg) +3DiT + +to any video diffusion model. We employ the AdamW [11] optimizer with a weight decay of 0.01 and set the initial learning rate to $1 \times 10^{-4}$ . The model is trained at a resolution of $512 \times 512$ pixels with a batch size of 256. EMA with a decay rate of 0.9999 is applied starting after the first 1,000 iterations to stabilize training. Training involves both synthetic and video datasets, with a 1:1 ratio. Tasks in the synthetic dataset are divided into object movement, removal, and insertion at a 6:1:1 ratio. We leave the data processing of video data in the supplementary file. Box augmentation is employed by maintaining the bounding box's center and adjusting its dimensions randomly within set ranges (i.e., 0.8-1.2 on synthetic data; 1.0-1.5 on video data for full mask-out), which improves the model's robustness. + +Evaluation Details. For evaluation, we introduce two new datasets: ObjMove-A and ObjMove-B. ObjMove-A comprises 200 image sets captured by experienced photographers. Each set includes an object placed at two distinct locations within the same scene, alongside a pure background image without the object. Object masks are subsequently generated using SAM [12]. Thus, the model performance can be quantitatively assessed between the generated image and the referenced target image on this dataset. Following [6, 33, 36], we utilize object-level metrics such as DINO-Score [5], CLIP-Score [27], and DreamSim [7] to evaluate the cropped target region. Since our task involves removing the object from the source region, we also evaluate these metrics on the cropped source region, and we show the average score on these two regions. We also assess the PSNR for the full generated image to measure overall image similarity. + +Given that our task is inherently ill-posed and reference-based evaluations cannot fully capture the realism of the generated images, we conduct human evaluations using our ObjMove-B dataset. This dataset consists of 150 image sets sourced from Pexels [26], consisting of manually annotated object masks and bounding box masks for plausible target regions. Compared with ObjMove-A, this dataset presents more diverse scenes and intricate challenges, such as occlusions, complex shadows and reflections, and complex background, thereby providing a more challenging bench + +
MethodPSNR ↑DINO-Score ↑CLIP-Score ↑DreamSim ↓User-Study ↑
Drag-Anything [41]16.3655.5684.440.4110.19%
3DiT [20]19.7245.3081.690.5140.19%
Paint-by-Example [42]20.8355.4685.230.4200.75%
Anydoor [6]21.8669.3288.950.2893.56%
MagicFixup [1]23.8278.4991.060.19831.14%
Ours25.2785.0793.160.14264.17%
+ +Table 1. Quantitative comparisons on object movement. Our method consistently outperforms state-of-the-art methods. For the evaluation metrics, only the user-study is conducted on ObjMove-B; all other evaluations are performed using ObjMove-A, as only ObjMove-A has ground-truth data. + +![](images/7b9d7698bfa6ec5e00dd897fc93c5bc8e9ac4a3b0ae8498c31af2f5cb4133de6.jpg) +Figure 8. Qualitative comparisons on object removal. Our method consistently outperforms state-of-the-art methods. + +![](images/d0d391e34ebf1effb181ce5352ddd1e92febf7b5e53c1f0a5618f74e5d38151c.jpg) +Input +Figure 9. Qualitative comparisons on object insertion. Our method consistently outperforms state-of-the-art methods. + +![](images/5fd64f0b0fc6791fc69bceb5ee2daa6012c6d3b6d13ea3a008c89656a4bab945.jpg) + +![](images/ea35972aaa15b491f2937a440c6dd279aa2627a424300ca2bdd69259366031b5.jpg) +Paint-by-Example + +![](images/debed74e4b1b63755e3a040e4873c9807a322e8c06c5c509fd7050369dc2ff8f.jpg) + +![](images/3bf05a0f329d3f4f3d89280040fabb720f35eeac44e212c7a2809392029fe619.jpg) +Anydoor + +![](images/2c74135cc130d762b21d573c17c9ee9ed749abc6d842c507701f9ded3dbeb3dc.jpg) + +![](images/7973eb88a46fc2febf1016721e685614cb06e4fe764ef5d5b27175b3574fa3f7.jpg) +Ours + +mark for assessment. To evaluate the result, we conduct a user study in which participants select the best result from different methods based on identity preservation, consistent lighting effects editing, and overall image quality. + +# 4.1. Comparison to Existing Methods + +Comparison on Object Movement. We first evaluate our model on the core task of object movement against the most relevant approaches [1, 6, 20, 41, 42]. Specifically, Any-door [6] and Paint-by-Example [42] are limited to object insertion and cannot relocate existing objects. Therefore, we employ the state-of-the-art inpainting model PowerPaint [45] to remove the object from its original position and subsequently utilize these methods to insert the object into the target location. For Drag-Anything [41], we generate a video following a linear trajectory from the source location to the target location and save the last frame as the edited result (see Supplementary for implementation details). + +Tab. 1 shows a quantitative analysis, where our approach consistently surpasses existing methods. Qualitative comparisons in Fig. 7 highlight our model's significant superiority. Specifically, in the first three examples, competing methods fail to effectively adjust object shadows or reflections to the target location. Furthermore, Anydoor and Paint-by-Example struggle with preserving object identity. In the challenging fourth example, involving completing an incomplete object and understanding glass material transparency requiring background synchronization, our method excels. Other methods either leave the object incomplete or distort the background. For the fifth image, where changing the + +
Method +Object RemovalPSNR ↑DINO-Score ↑CLIP-Score ↑DreamSim ↓
3DiT [20]20.5241.8385.020.499
PowerPaint [45]25.1562.7189.400.347
SD-Inpaint [28]24.1650.8083.910.492
LaMa [35]27.0463.8289.910.322
Ours28.9083.9494.840.143
Object InsertionPSNR ↑DINO-Score ↑CLIP-Score ↑DreamSim ↓
Paint-by-Example [42]22.6454.8081.810.457
Anydoor [6]24.1077.4988.330.223
Ours24.9985.6991.450.152
+ +Table 2. Quantitative comparisons on auxiliary tasks. Our method consistently outperforms state-of-the-art methods. + +
MethodPSNR ↑DINO-Score ↑CLIP-Score ↑DreamSim ↓
Model A22.8675.0192.190.242
Model B23.8380.0293.160.196
Model C23.8979.9793.220.197
Model D24.2482.0993.620.180
+ +Table 3. Ablation of different models. Model A indicates training from the T2I checkpoint with synthetic data only. Model B indicates training from the I2V checkpoint with synthetic data only and only on the movement task. Model C indicates training from the I2V checkpoint with synthetic data only but adding the insertion and removal tasks. Model D indicates adding video data. + +car's perspective angle to suit the new location is crucial, MagicFixup fails, and Anydoor and Paint-by-Example alter the object's identity. In the final example, only our method successfully integrates into the target region using lighting information, achieving a realistic effect of partial shadow and light. We also perform a self-evaluation game (details in the supplementary file), presenting four images: one is the input image, and the other three are generated by our model. Users are asked to identify the input image and the results reveal that users incorrectly identify the input image $70\%$ of the time, demonstrating our method's ability to generate images that obscure artifacts. + +Comparison on Auxiliary Tasks. Our model is capable of object removal and insertion, and we compare these capabilities with other methods [6, 20, 28, 35, 42, 45]. For object insertion, we use the segmented object from the source frame as a reference image and prompt the model to place the object into the clean background at the target frame's location. We evaluate object-centric metrics in the cropped target region for insertion and in the source region for removal. Tab. 2 shows quantitative comparisons, while Fig. 8 and Fig. 9 provide visual comparisons. Our results indicate that our model outperforms other approaches. Specifically, in object removal, our method effectively removes the object without introducing artifacts, eliminating shadows and reflections more successfully than others, and show stronger background completion ability. In object insertion, other methods struggle with consistent reflections, identity, and background preservation, while our approach maintains consistent reflections, preserves identity and background integrity, and blends seamlessly with the environment. + +![](images/239dce67202b1e6982e9bfa968e4c7365428eb5567beedd15a7e5dbe715bb6c3.jpg) +Figure 10. Qualitative comparison among different models. Model A performs the worst, failing to move shadows and leaving obvious artifacts. Models B and C reduce artifacts but still struggle to move shadows effectively. Model D performs the best. + +# 4.2. Ablation Studies + +We perform ablation studies to illustrate the impact of our primary design choices. Every model is trained at a resolution of $256 \times 256$ for an equal number of iterations. As shown in Tab. 3 and Fig. 10, our designs are proven to be effective. We begin by highlighting the importance of utilizing a pretrained image-to-video model. To this end, we train an alternative model based on a pretrained text-to-image model (Model A) using our synthetic data across three tasks. Notably, the architecture and parameters of the text-to-image model are the same as those of the image-to-video model, with the exception that the former is pre-trained on single-frame images. Model A performs significantly worse compared to Model C, which is trained using a pretrained image-to-video model with the identical dataset and tasks, as evidenced by both quantitative and qualitative evaluations. Furthermore, incorporating object removal and insertion tasks into the synthetic dataset maintains similar performance for our primary task while enabling inference over three tasks (Model B vs Model C). Finally, training with real-world video data improves performance (Model C vs Model D), particularly enhancing background occlusion filling (e.g., the second example) and the consistency of lighting and shadow adjustments. + +# 5. Conclusion + +In this paper, we presented ObjectMover, a novel framework for object movement that integrates a sequence-to-sequence modeling using a video diffusion model. Our method offers a unified, single-stage solution that addresses the challenges of object identity preservation, lighting consistency, and realistic harmonization into new environments. For task-specific model finetuning, we developed a data generation pipeline utilizing a game engine to produce high-quality data pairs. Additionally, we propose a multi-task learning approach that allows for training on real-world video data, enhancing model generalization. The trained model can also be used for object removal and insertion tasks. Extensive experimental results demonstrate that our model not only attains but surpasses the effectiveness of current leading methods. + +# References + +[1] Hadi Alzayer, Zhihao Xia, Xuaner Zhang, Eli Shechtman, JiaBin Huang, and Michael Gharbi. Magic fixup: Streamlining photo editing by watching dynamic videos. arXiv preprint arXiv:2403.13044, 2024. 2, 3, 7 +[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18208-18218, 2022. 3 +[3] Omri Avrahami, Rinon Gal, Gal Chechik, Ohad Fried, Dani Lischinski, Arash Vahdat, and Weili Nie. Diffuhaul: A training-free method for object dragging in images. arXiv preprint arXiv:2406.01594, 2024. 3 +[4] Shariq Farooq Bhat, Niloy J. Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning. In SIGGRAPH, 2024. 3 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7 +[6] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 6593-6602. IEEE, 2024. 2, 3, 5, 7, 8 +[7] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023. 7 +[8] Epic Games. Unreal engine 5.3, 2024. 2, 5 +[9] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. International Conference on Learning Representations, 2024. 3 +[10] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14049-14058, 2023. 3 +[11] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7 +[12] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 7 +[13] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[14] Wenbo Li, Xin Yu, Kun Zhou, Yibing Song, Zhe Lin, and Jiaya Jia. Image inpainting via iteratively decoupled probabilistic modeling. arXiv preprint arXiv:2212.02963, 2022. 2 +[15] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian + +Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 4 +[16] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 4 +[17] Shilin Lu, Yanzhu Liu, and Adams Wai-Kin Kong. Tf-icon: Diffusion-based training-free cross-domain image composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[18] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 3 +[19] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. ICLR, 2022. 3 +[20] Oscar Michel, Anand Bhattachad, Eli VanderBilt, Ranjay Krishna, Aniruddha Kembhavi, and Tanmay Gupta. Object 3dit: Language-guided 3d-aware image editing. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5, 7, 8 +[21] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. *Dragondiffusion: Enabling drag-style manipulation on diffusion models.* *ICLR*, 2024. 3 +[22] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8488-8497, 2024. 3 +[23] OpenAI. Video generation models as world simulators. https://openai.com/research/video-generation-models-as-world-simulators, 2023. Accessed: yyyy-mm-dd. 4 +[24] Karran Pandey, Paul Guerrero, Matheus Gadelha, Yannick Hold-Geoffroy, Karan Singh, and Niloy J. Mitra. Diffusion handles enabling 3d edits for diffusion models by lifting activations to 3d. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 7695-7704. IEEE, 2024. 3 +[25] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4 +[26] Pexels. Pexels - free high-quality images. https://www.pexels.com/, 2024.7 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4, 8 + +[29] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 3 +[30] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[31] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 3 +[32] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Objectstitch: Object compositing with diffusion model. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 2, 3, 5 +[33] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, He Zhang, Wei Xiong, and Daniel Aliaga. Imprint: Generative object compositing by learning identity-preserving representation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 8048-8058. IEEE, 2024. 2, 3, 7 +[34] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 3 +[35] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022. 2, 8 +[36] Gemma Canet Tarres, Zhe Lin, Zhifei Zhang, Jianming Zhang, Yizhi Song, Dan Ruta, Andrew Gilbert, John Collomosse, and Soo Ye Kim. Thinking outside thebbox: Unconstrained generative object compositing, 2024. 3, 7 +[37] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4 +[38] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. In ICML, pages 52274-52289, 2024. 3 +[39] Tianyu Wang, Xiaowei Hu, Qiong Wang, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1880–1889, 2020. 3 +[40] Daniel Winter, Matan Cohen, Shlomi Fruchtner, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectdrop: Bootstrap + +ping counterfactuals for photorealistic object removal and insertion, 2024. 2, 3 +[41] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2025. 3, 7 +[42] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 2, 3, 7, 8 +[43] Jiraphon Yenphraphai, Xichen Pan, Sainan Liu, Daniele Panozzo, and Saining Xie. Image sculpting: Precise object editing with 3d geometry control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4241-4251, 2024. 3 +[44] Bo Zhang, Yuxuan Duan, Jun Lan, Yan Hong, Huijia Zhu, Weiqiang Wang, and Li Niu. Controlcom: Controllable image composition using diffusion model. arXiv preprint arXiv:2308.10040, 2023. 3 +[45] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv preprint arXiv:2312.03594, 2023. 2, 7, 8 \ No newline at end of file diff --git a/objectmovergenerativeobjectmovementwithvideoprior/images.zip b/objectmovergenerativeobjectmovementwithvideoprior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..64ea755ea96fb8a19665021c81f063601ff185be --- /dev/null +++ b/objectmovergenerativeobjectmovementwithvideoprior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f4e9df400630a4fe218b98dba510124179050ef6b6eabe55b1dc351a481f3fa +size 994614 diff --git a/objectmovergenerativeobjectmovementwithvideoprior/layout.json b/objectmovergenerativeobjectmovementwithvideoprior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f51985044b21651b714d2d089ba28ce4936ceb66 --- /dev/null +++ b/objectmovergenerativeobjectmovementwithvideoprior/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45f6c5791e983fcd993fc473600a13e19d8cd7e0a88350b12182f6a3163cdb5b +size 484735 diff --git a/objectshotenhancedgroundingnetworkforegocentricvideo/cef57492-dc63-4ddb-9c34-f275ae7f5a12_content_list.json b/objectshotenhancedgroundingnetworkforegocentricvideo/cef57492-dc63-4ddb-9c34-f275ae7f5a12_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e124452e017dc2906b86948e3f20a590bc64ec8d --- /dev/null +++ b/objectshotenhancedgroundingnetworkforegocentricvideo/cef57492-dc63-4ddb-9c34-f275ae7f5a12_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98bc89c61b4923c115ce4f06159fd55f3051d759cad0b18a56f59d54b72b0892 +size 90361 diff --git a/objectshotenhancedgroundingnetworkforegocentricvideo/cef57492-dc63-4ddb-9c34-f275ae7f5a12_model.json b/objectshotenhancedgroundingnetworkforegocentricvideo/cef57492-dc63-4ddb-9c34-f275ae7f5a12_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dc31ba36576bc80e7bdd85625e0fc9ecd60effb5 --- /dev/null +++ b/objectshotenhancedgroundingnetworkforegocentricvideo/cef57492-dc63-4ddb-9c34-f275ae7f5a12_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7cfc43888f3ab49be007b6013bec8189dd459fff729bbf1bf5d99b304dce8dc1 +size 110824 diff --git a/objectshotenhancedgroundingnetworkforegocentricvideo/cef57492-dc63-4ddb-9c34-f275ae7f5a12_origin.pdf b/objectshotenhancedgroundingnetworkforegocentricvideo/cef57492-dc63-4ddb-9c34-f275ae7f5a12_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..60c9772dd8d770ce09e201c42cee124a545ad531 --- /dev/null +++ b/objectshotenhancedgroundingnetworkforegocentricvideo/cef57492-dc63-4ddb-9c34-f275ae7f5a12_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e43656807e41f2a3041c03c2c85c02b4886897063d9f8808468c30407344b103 +size 802868 diff --git a/objectshotenhancedgroundingnetworkforegocentricvideo/full.md b/objectshotenhancedgroundingnetworkforegocentricvideo/full.md new file mode 100644 index 0000000000000000000000000000000000000000..81b6ae0739805202662d7791a5b9958e15eb2787 --- /dev/null +++ b/objectshotenhancedgroundingnetworkforegocentricvideo/full.md @@ -0,0 +1,356 @@ +# Object-Shot Enhanced Grounding Network for Egocentric Video + +Yisen Feng $^{1}$ Haoyu Zhang $^{12}$ Meng Liu $^{3*}$ Weili Guan $^{1}$ Liqiang Nie $^{1*}$ $^{1}$ Harbin Institute of Technology (Shenzhen) $^{2}$ Pengcheng Laboratory + $^{3}$ Shandong Jianzhu University + +{yisenfeng.hit, zhang.hy.2019, mengliu.sdu, honeyguan, nieliqiang}@gmail.com + +# Abstract + +Egocentric video grounding is a crucial task for embodied intelligence applications, distinct from exocentric video moment localization. Existing methods primarily focus on the distributional differences between egocentric and exocentric videos but often neglect key characteristics of egocentric videos and the fine-grained information emphasized by question-type queries. To address these limitations, we propose OSGNet, an Object-Shot enhanced Grounding Network for egocentric video. Specifically, we extract object information from videos to enrich video representation, particularly for objects highlighted in the textual query but not directly captured in the video features. Additionally, we analyze the frequent shot movements inherent to egocentric videos, leveraging these features to extract the wearer's attention information, which enhances the model's ability to perform modality alignment. Experiments conducted on three datasets demonstrate that OSGNet achieves state-of-the-art performance, validating the effectiveness of our approach. Our code can be found at https://github.com/Yisen-Feng/OSGNet. + +# 1. Introduction + +With advancements in wearable camera technology, Ego4D [8] introduces the Natural Language Query (NLQ) task for egocentric video grounding. NLQ aims to identify the specific video moment that answers a question-type query within an untrimmed egocentric video, as shown in Figure 1(b). The dynamic and complex camera perspectives in egocentric videos [55] make video comprehension [20, 21] significantly more challenging compared to the fixed viewpoints of exocentric video grounding, as shown in Figure 1(a). Moreover, NLQ queries often focus on fine-grained details of background objects (e.g., "measuring tape" in Figure 1(b)), unlike exocentric video grounding tasks that emphasize character actions (e.g., "enters") + +Query: A man enters the kitchen, removes a frying pan from the drawer and places it on the stove. + +![](images/7f402ab75e42b6575c997c8dd04ccf5510f275e53763373db65faa0543a19f8e.jpg) + +![](images/e1d0b945ce0882a8e7b388c9850e1950901028d5833fc93c64f72abfc47f2138.jpg) + +![](images/7f268134a2882e78390c75f377fb277c456c6928bc7e671fc9ff7ff6c6ee0178.jpg) + +![](images/48ce2f54766b70db248b46fa9d5c04ae874472a07262b393b1a4274811c18527.jpg) + +![](images/9254a2a18e0998eadaca75563810cb5da2a1bb4eba24142a090a801642865a1c.jpg) + +(a) Exocentric video moment localization task from TACoS [38] + +Query: Where was the measuring tape before I picked the drill with my right hand? + +![](images/526df0cb4111a92f65041b7e83dc489aa691173c60aea0aeb191b710f9bf5bf4.jpg) + +![](images/91b7009d67afc48f18e0aefd10d83ca69cc7e20c71419eb2fe40ff2c258f5e09.jpg) + +![](images/1109ef9d18632d46f021981dd951c7aa5dce4f0629f3c77b3b6124c45583a534.jpg) + +![](images/a72c3d0b7e56b87415c129d7f0ad2fc09f8511a25d315264df82e3b2e8025d32.jpg) + +![](images/ee9d9afeda140aaf55308aac638eecf20cd1066113e40d438ed44f2b0f08cc84.jpg) + +C clamps the wood at the top with another vertical wood. + +C picks a drill driver with his right hand. + +Narration + +C presses the trigger of the drill driver with his left hand. + +(b) Natural language query task and narrations from Ego4D [8] + +Figure 1. Illustration of exocentric and egocentric video grounding, accompanied by annotated narrations for the egocentric video. Key verbs and nouns are highlighted in red. + +and "removes" in Figure 1(a)), adding complexity to mining background information from video. Consequently, despite progress in exocentric video grounding, existing methods struggle with the unique challenges of NLQ. Nonetheless, NLQ enables innovative applications [53, 54], such as smart assistant systems and memory retrieval modules for autonomous robots, underscoring the urgent need to address these challenges. + +Though existing methods [2, 22, 31, 34, 37] for NLQ have made significant advances in video-text pretraining, they fail to address the issue of video features lacking the fine-grained object information needed to answer detailed queries. As shown in Figure 1(b), egocentric videotext pretraining datasets generally provide narrations focusing on character actions involving objects. These datasets, combined with clip-narration contrastive learning, lead current video backbones to overlook fine-grained background object details in the extracted features. In contrast, NLQ emphasizes enhancing people's memory experience, often involving queries centered on background objects that are not part of active interactions, such as the "measuring tape" placed casually on the table, as illustrated in Figure 1(b). + +Moreover, existing methods fail to fully leverage the rich attention information embedded in egocentric videos. Egocentric videos feature frequent camera movements, as the camera is typically worn on the head, moving with the wearer's actions. This movement implicitly encodes head motion information, signaling shifts in the wearer's attention and focus, an aspect that is crucial for NLQ but often overlooked by current methods. As shown in Figure 1(b), initially, the individual fixes the wood, then walks to pick up a drill driver, and finally uses the drill driver to drill holes in the wood. These behaviors are independent yet interconnected. Capturing these attention shift points is vital for clarifying video structure and improving video understanding. + +Building upon these findings, we propose a novel Object-Shot enhanced Grounding Network (OSGNet) for egocentric videos, addressing the unique challenges posed by the NLQ task. To address the challenge of insufficient fine-grained object information, we introduce an object extraction process (see Figure 2(a)), which captures detailed object-level information. These features are then integrated using a multi-modal fusion mechanism that employs parallel cross-attention within the main branch. This strategy ensures that both visual and textual cues are effectively leveraged to enhance localization accuracy. To exploit the dynamics of the wearer's attention, we extract head-turning data from egocentric video, which provides insights into shifts in the wearer's focus. Based on these shifts, we segment the video into semantically distinct shots and apply contrastive learning to strengthen the model's ability to align these shots with the query. This approach enhances the model's capacity to capture and utilize attention-driven context for improved video grounding. We evaluate our model on three benchmark datasets: Ego4D-NLQ, Ego4D-Goal-Step [42], and TACoS [38]. On Ego4D-NLQ, OSGNet outperforms GroundVQA [5], achieving a $2.15\%$ improvement in Rank@1 at IoU=0.5. On Ego4D-Goal-Step, OSGNet surpasses BayesianVSLNet [33] with a $3.65\%$ increase in Rank@1 at IoU=0.3. On TACoS, OSGNet achieves a $3.32\%$ improvement in Rank@1 at IoU=0.5 over SnAG [27]. These results underscore the effectiveness of our model in enhancing video grounding. + +Contributions. 1) We integrate fine-grained object information into the egocentric video grounding task, thereby improving the accuracy of background object-related query localization. 2) We introduce a wearer movement-aware shot branch that leverages shot-level contrastive learning, collaborating with the main branch to further enhance egocentric video grounding performance. And 3) our OSGNet outperforms current state-of-the-art approaches across three benchmark datasets, setting a new standard for egocentric video grounding. + +# 2. Related Work + +# 2.1. Video Moment Localization + +Video moment localization aims to determine the start and end timestamps of a video moment in response to a natural language query. Existing approaches [23, 24, 35, 43, 48, 50] can be broadly classified into proposal-based and proposal-free methods, which primarily differ in their strategies for generating candidate moments. Proposal-based methods generally follow a two-stage process involving the generation of candidate moment features and subsequently matching them with textual queries. Sliding windows have been extensively employed for generating candidates [11, 15, 16], with approaches like [7] incorporating regression to refine window boundaries. Subsequent work [4, 47, 49] enhances candidate quality by integrating text information at an early stage, while others [56, 57] increased the density of candidate moments through exhaustive enumeration. + +On the other hand, proposal-free methods aim to directly generate features for individual moments or the entire video, leading to fewer features and increased computational efficiency. Some approaches [18, 28, 61] extract features for the entire video sequence, whereas others [6, 52] focus on predicting the probability of a particular moment being part of the desired query segment. The emergence of new datasets involving long video sequences [8, 41] has posed significant challenges for existing methods, which often struggle with performance degradation on longer videos. To address these challenges, [13] proposes a query-guided window selection strategy under the assumption that short windows are sufficient, while [27] presents a single-stage approach that aligns multiple video moments without making similar assumptions. + +As discussed in Section 1, significant differences in the distribution of videos and queries between the exocentric video moment localization task and the egocentric video NLQ task hinder the direct applicability of existing frameworks. To overcome these challenges, our model incorporates fine-grained object-level information and employs shot-level contrastive learning, thereby enhancing the accuracy of egocentric video grounding. + +# 2.2. Natural Language Query + +The NLQ task aims to accurately localize video moments that answer natural language questions in egocentric videos. Existing approaches primarily pursue two directions: 1) Video-Text Pretraining [17, 58]. Due to the differences in feature distributions between egocentric and exocentric videos, several studies [2, 22, 31, 34] have fine-tuned video backbones using egocentric videos, improving the model's robustness in tasks involving egocentric video comprehension. 2) Data Enhancement. Ramakrishnan et al. [37] constructed a large narrative dataset of 940K video-narration + +![](images/1bddc2ca424ca04e4da8e3c6718ebce84ea8990dae467fe4d9dadac5a4390955.jpg) +Figure 2. The framework of our OSGNet, which consists of four key components: (a) Object Extraction, which captures fine-grained object features; (b) Feature Extraction, where visual and textual cues are processed; (c) Main Branch, responsible for primary grounding tasks; and (d) Shot Branch, which leverages wearer movement dynamics and shot-level contrastive learning to improve localization accuracy. + +pairs to pre-training models, followed by fine-tuning on Ego4D-NLQ, yielding substantial improvements. + +However, prior methods [12, 13, 25] often treat NLQ as a general long-video localization problem [30], overlooking the need for fine-grained object information that is crucial for NLQ. Our approach addresses this gap by extracting query-relevant fine-grained object features and integrating them into the training process, thus improving localization accuracy. Additionally, existing methods tend to overlook the frequent shot transitions inherent in egocentric videos. To mitigate this, we segment egocentric videos into distinct shots and employ contrastive learning to better align the two modalities, ultimately enhancing localization performance. + +# 3. Method + +# 3.1. Overview + +Given a video consisting of $N$ frames, denoted as $\mathcal{V} = \{v_{1}, v_{2}, \dots, v_{N}\}$ , and a natural language query comprising $L$ words, denoted as $\mathcal{Q} = \{w_{1}, w_{2}, \dots, w_{L}\}$ , our grounding model aims to localize the specific moment within an untrimmed egocentric video that best answers the query, represented by the timestamp $[t_s, t_e]$ . + +As outlined in Section 1, existing video feature representations are often insufficient for NLQ, primarily due to + +the absence of fine-grained object-level information. To address this limitation, we propose a novel framework that integrates object-level details into the feature representation process, significantly enhancing the localization accuracy. Our approach involves extracting features from video, text, and objects, followed by a comprehensive feature encoding process, as described in Section 3.2. The model architecture comprises two primary components: the main branch and the shot branch. The main branch (Section 3.3) combines video, text, and object features to improve the accuracy of video localization. In contrast, the shot branch (Section 3.4) focuses exclusively on video features to generate shot representations and applies contrastive learning with text inputs, enhancing the model's ability to perform modality alignment. Finally, Section 3.5 details the training and inference process of our framework. + +# 3.2. Feature Extraction + +Video Feature. We divide the video into non-overlapping clips, denoted as $\mathcal{C} = \{\mathcal{C}_i|\mathcal{C}_i = \{v_{(i - 1)\times s + 1},\dots ,v_{(i - 1)\times s + s}\} \}_{i = 1}^T$ where $s$ denotes the size of the sliding window and $T$ denotes the number of clips in the video. Pretrained video backbones are used to extract clip-level video features, which are then projected into a feature space, yielding the video representation as + +$\mathbf{V}_{clip} \in \mathbb{R}^{T \times D}$ , where $D$ denotes the dimensionality of the feature vectors. + +Query Feature. Query features are extracted using a text backbone, resulting in word-level features $\mathbf{Q}_{word} \in \mathbb{R}^{L \times D_T}$ , where $D_T$ is the dimensionality of the word embeddings. To capture the relationships among words in the query, we employ a multi-layer transformer, termed the text encoder, outputting the query representation $\mathbf{Q}_F \in \mathbb{R}^{L \times D}$ . + +Object Feature. To effectively capture object information, we utilize the object detector Co-DETR [63], pretrained on the LVIS dataset [9], which is capable of identifying a wide range of object categories within video frames. As illustrated in Figure 2(a), we select object categories that are related to the nouns in the query and whose confidence scores exceed the threshold $\theta$ . For object representation, we encode the object categories as textual features, allowing seamless integration with the textual query. To achieve this, we employ CLIP (ViT-B/32) [36] to encode textual features of the detected object categories, resulting in $\mathbf{O}_{clip} \in \mathbb{R}^{T \times N_o \times D_o}$ , where $N_o$ is the maximum number of objects detected in a single frame and $D_o$ denotes the dimensionality of the object features. + +Different objects in the same query play distinct roles in the video grounding task. For example, in the query "How many drill bit did I remove from the drill before I moved the yellow carton?", the beginning of the target moment should be related to the "drill bit" instead of the "drill", and the end of the target moment should before the "carton" appears. Therefore, we design an object encoder to refine the query-relevant object features. The object encoder is a multi-layer transformer, processing the object features in the context of the input query. To avoid confusion between object features within the same frame, we replace the conventional self-attention mechanism of transformers with a cross-attention mechanism, where object features serve as queries and query features act as keys and values. This architecture effectively encodes object-related information concerning the query, denoted as $\mathbf{O}_F\in \mathbb{R}^T\times N_o\times D$ + +# 3.3. Main Branch + +Our main branch consists of three core components: a multimodal fusion module to integrate features from different modalities, a multi-scale network to generate candidate moment representations at multiple temporal scales, and task-specific heads to predict the temporal offsets and confidence scores for the localized video moments. + +Multi-modal Fusion. To effectively integrate fine-grained object features, we design a multi-modal fusion module that combines video, text, and object features. This module consists of multiple stacked layers, each including a bidirectional Mamba (BiMamba) block, followed by a multi-layer + +perceptron (MLP), cross-attention (CA) block, and Feedforward Neural Network (FFN) for query and object processing, culminating in a fusion block. + +To be specific, we enhance video features using a BiMamba layer [62], instead of traditional local self-attention, to better capture long-range dependencies within video data. The updated video features are computed as follows: + +$$ +\hat {\mathbf {V}} ^ {(i)} = \mathbf {V} _ {f} ^ {(i)} + M L P (B i M a m b a (\mathbf {V} _ {f} ^ {(i)})), \tag {1} +$$ + +where $\mathbf{V}_f^{(0)} = \mathbf{V}_{clip}$ and $\mathbf{V}_f^{(i)}\in \mathbb{R}^{T\times D}$ is the output of the $i$ -th layer. Next, we apply a cross-attention block and FFN to aggregate video and query information: + +$$ +\left\{ \begin{array}{l} \mathbf {V} _ {Q} ^ {(i)} = \hat {\mathbf {V}} ^ {(i)} + C A \left(\hat {\mathbf {V}} ^ {(i)}, \mathbf {Q} _ {F}, \mathbf {Q} _ {F}\right), \\ \hat {\mathbf {V}} _ {Q} ^ {(i)} = \mathbf {V} _ {Q} ^ {(i)} + F F N \left(\mathbf {V} _ {Q} ^ {(i)}\right). \end{array} \right. \tag {2} +$$ + +Similarly, we aggregate video feature $\hat{\mathbf{V}}^{(i)}$ and object features $\mathbf{O}_F$ using a parallel cross-attention block for object features. The final output $\hat{\mathbf{V}}_O^{(i)}$ is calculated as follows: + +$$ +\left\{ \begin{array}{l} \mathbf {V} _ {O} ^ {(i)} = \hat {\mathbf {V}} ^ {(i)} + C A \left(\hat {\mathbf {V}} ^ {(i)}, \mathbf {O} _ {F}, \mathbf {O} _ {F}\right), \\ \hat {\mathbf {V}} _ {O} ^ {(i)} = \mathbf {V} _ {O} ^ {(i)} + F F N \left(\mathbf {V} _ {O} ^ {(i)}\right). \end{array} \right. \tag {3} +$$ + +Finally, these features are combined using a gating mechanism: + +$$ +\left\{ \begin{array}{l} \mathbf {A} = \sigma \left(M L P \left(\hat {\mathbf {V}} _ {Q} ^ {(i)} \right| \left| \hat {\mathbf {V}} _ {O} ^ {(i)}\right)\right), \\ \mathbf {V} _ {f} ^ {(i + 1)} = \mathbf {A} \cdot \hat {\mathbf {V}} _ {Q} ^ {(i)} + (\mathbf {1} - \mathbf {A}) \cdot \hat {\mathbf {V}} _ {O} ^ {(i)}. \end{array} \right. \tag {4} +$$ + +where $\sigma$ is sigmoid fuction, $||$ is vector concatenation. + +Multi-scale Network. The multi-scale network generates a feature pyramid to facilitate the grounding of video moments at various temporal scales. This network is a multilayer transformer, with each layer including a 1D depth-wise convolution before the self-attention and FFN modules to enable sequence downsampling, as described in [12]. By feeding the output of the multi-modal fusion module, we can obtain the multi-scale candidate moment representations, denoted as $\left[\mathbf{V}_m^{(0)},\mathbf{V}_m^{(1)},\dots ,\mathbf{V}_m^{(L_s)}\right]$ , where $\mathbf{V}_m^{(0)} = \mathbf{V}_f^{(L_f)}$ . Here $L_{f}$ and $L_{s}$ are the number of layers in the multi-modal fusion and multi-scale network, respectively. The representation $\mathbf{V}_m^{(j)}\in \mathbb{R}^{T / 2^j\times D}$ represents the video features at the $j$ -th scale, with progressively reduced sequence lengths due to downsampling at each layer. + +Task Heads. The task heads are responsible for decoding the multi-scale feature pyramid into final predictions for video grounding. Specifically, we use two heads: a classification head and a regression head following previous work [12, 27, 51]. Each task head is implemented using two layers of a 1D convolutional network. The classification head predicts the confidence score for each candidate + +![](images/04bf0e338563b81614a2df728e224048a93ad4a1de509f242cb5ed00266d68bd.jpg) +Figure 3. Illustration of captions generated by LAVILA [59] describing camera movements. + +moment, while the regression head predicts the offsets of the moment boundaries relative to the anchor point. + +For each feature $\mathbf{v}_k^{(j)} = \mathbf{V}_m^{(j)}[k] \in \mathbb{R}^D$ , $\mathbf{v}_k^{(j)}$ represents the feature for the $k$ -th anchor point in the $j$ -th layer, which has $T / 2^j$ anchor points with a stride of $2^j$ . The corresponding timestamp is computed as $t_k^{(j)} = 2^j * k$ . The classification head then predicts the confidence $c_k^{(j)}$ for $\mathbf{v}_k^{(j)}$ , and the regression head predicts normalized offsets $(s_k^{(j)}, e_k^{(j)})$ for $\mathbf{v}_k^{(j)}$ . The predicted video moment for this anchor point is defined by the start and end boundaries: $(t_k^{(j)} - s_k^{(j)} \cdot 2^j, t_k^{(j)} + e_k^{(j)} \cdot 2^j)$ . + +Localization Loss. The moment localization loss, $\mathcal{L}_{ML}$ , is designed to enhance model accuracy in video moment localization and is defined as follows: + +$$ +\mathcal {L} _ {M L} = \mathcal {L} _ {c l s} + \mathcal {L} _ {r e g}. \tag {5} +$$ + +The first component $\mathcal{L}_{cls}$ is focal loss [39], which is used for classification to assess whether a proposal aligns with the query. The second component $\mathcal{L}_{reg}$ is Distance-IoU loss [60], which refines the boundaries of the localized moment by calculating the Distance-IoU between the predicted and ground truth moment boundaries. Notably, $\mathcal{L}_{reg}$ is calculated only on positive candidates. + +# 3.4. Shot Branch + +The shot branch is specifically designed to harness the attention information embedded in egocentric videos, which reflects the wearer's focus during video capture. This is achieved by segmenting the video into semantically distinct shots, with head movement serving as a key indicator for defining these boundaries. Contrastive learning is then employed to enhance the model's ability to align cross-modal features. This shot-level understanding deepens the model's comprehension of the video structure, significantly improving video grounding performance. + +Shot Segmentation. As illustrated in Figure 3, the wearer's head movements, captured by the pretrained captioning + +model LAVILA [59], are often described in video captions with expressions such as "looks around" or "turns around". These cues are used to define shot boundaries, denoted by timestamps $t_{shot} = \{(t_i, t_{i+1})\}_{i=1}^{N_S}$ , where $t_1 = 0$ represents the start of the video, $t_{N_S+1}$ represents the video's end, and $\{t_i\}_{i=2}^{N_S}$ corresponds to timestamps that capture head movement activities. Here $N_S$ is the total number of shots. + +Feature Aggregation. To perform contrastive learning effectively, it is crucial to extract contextual video features that represent the content of each shot independently of query and object information. This is achieved by employing multiple stacked layers, each consisting of a BiMamba block integrated with an MLP. These layers share parameters with the multi-modal fusion module, enabling us to extract self-interaction features from the video, denoted as $\mathbf{V}_{\text {pure }} \in \mathbb{R}^{T \times D}$ . + +To aggregate shot and query features for contrastive learning, we employ a transformer encoder with an architecture similar to the Q-Former [19]. In this setup, learnable embeddings serve as queries, while the shot and query features function as keys and values. Following this process, we apply a 1D convolution to obtain the final shot representations, denoted as $\mathbf{V}_{\text{shot}} \in \mathbb{R}^{N_S \times D}$ , and text representation, denoted as $\mathbf{q}_{\text{sent}} \in \mathbb{R}^D$ . + +Contrastive Learning. To enhance contrastive learning, we gather all shots and their associated queries within a mini-batch, denoted as $\mathbf{Q}_{batch} = \{\mathbf{q}_{sent}^i\}_{i=1}^m \in \mathbb{R}^{m \times D}$ and $\mathbf{V}_{batch} = \{\mathbf{v}_{shot}^i\}_{i=1}^n \in \mathbb{R}^{n \times D}$ , respectively, where $m$ and $n$ represent the number of queries and shots in the mini-batch. Next, the text and video features are projected into a joint semantic space using separate MLPs. These projected features are used for contrastive learning to enable the model to learn the alignment between video shots and natural language queries, thereby improving the accuracy of video moment localization. + +To align shot features with textual queries, we apply contrastive learning using the InfoNCE loss [45], which encourages positive shot-query pairs to be closer in the embedding space while pushing apart the negative pairs. The InfoNCE loss is defined as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c o n}} = \frac {1}{m} \sum_ {i = 1} ^ {m} \frac {\sum_ {\left(\mathbf {q} _ {\text {s h o t}} ^ {i} , \mathbf {v} _ {\text {s h o t}} ^ {s}\right) \in \mathcal {P}} \exp \left(S i m \left(\mathbf {q} _ {\text {s e n t}} ^ {i} , \mathbf {v} _ {\text {s h o t}} ^ {s}\right) / \tau\right)}{\sum_ {j = 1} ^ {n} \exp \left(S i m \left(\mathbf {q} _ {\text {s e n t}} ^ {i} , \mathbf {v} _ {\text {s h o t}} ^ {j}\right) / \tau\right)} \\ + \frac {1}{n} \sum_ {i = 1} ^ {n} \frac {\sum_ {\left(\mathbf {q} _ {\text {s e n t}} ^ {l} , \mathbf {v} _ {\text {s h o t}} ^ {i}\right) \in \mathcal {P}} \exp \left(S i m \left(\mathbf {v} _ {\text {s h o t}} ^ {i} , \mathbf {q} _ {\text {s e n t}} ^ {l}\right) / \tau\right)}{\sum_ {j = 1} ^ {m} \exp \left(S i m \left(\mathbf {v} _ {\text {s h o t}} ^ {i} , \mathbf {q} _ {\text {s e n t}} ^ {j}\right) / \tau\right)}, \end{array} \tag {6} +$$ + +where $Sim$ represents the cosine similarity, $\mathcal{P}$ is the set of positive query-shot pairs where the query's ground truth intersects with the corresponding shot, and $\tau$ is the temperature coefficient. + +Table 1. Performance comparison on Ego4D-NLQ. $C$ denotes the CLIP video feature. \* indicates results reproduced using the released code. refers to results that utilize the NaQ [37] pretraining strategy. + +
MethodPublishedFeatureValidationTest
R@1R@5R@1R@5
0.30.50.30.50.30.50.30.5
Ego4D-NLQ v1InternVideo [2]CVPRW 2022E+I15.6410.1724.7818.3016.4510.0622.9516.10
CONE [13]ACL 2023E14.158.1830.3318.0213.467.8423.6814.16
SnAG* [27]CVPR 2024E----14.3410.2927.7219.91
NaQ† [37]CVPR 2023E+C19.3111.5923.6215.7518.4610.7421.5013.74
RGNet* [10]ECCV 2024E16.8610.5334.4321.8416.6810.6127.9517.83
RGNet*† [10]ECCV 2024E18.6611.7236.3722.4318.2111.6929.8019.06
OSGNetE16.1311.2836.7825.6315.2310.7128.2720.32
OSGNet†E21.9715.2044.6132.9622.1315.4636.5425.70
Ego4D-NLQ v2NaQ*† [37]CVPR 2023E+I+C24.1015.0329.9020.8521.7013.6425.1216.33
ASL† [40]CVPRW 2023E+I22.6215.6446.8632.1624.1315.4634.3723.18
GroundNLQ† [12]CVPRW 2023E+I26.9818.8353.5640.0024.5017.3140.4629.17
EgoEnv† [29]NeurIPS 2024E+I+C+EgoEnv25.3715.33--23.2814.3627.2517.58
EgoVideo† [31]CVPRW 2024EgoVideo28.6519.7353.3040.4225.0717.3140.8829.67
GroundVQA*† [5]CVPR 2024E+I29.6820.2352.1737.8326.6717.6339.9427.70
OSGNet†E+I31.6322.0357.9145.1927.6019.7843.4632.77
+ +# 3.5. Training and Inference + +Training. Our training is divided into pretraining and fine-tuning phases. In pretraining, we use the dataset from [37], only the main branch without object cross-attention is trained with the moment localization loss. In fine-tuning, the final loss function combines both the moment localization loss and the contrastive learning loss, as: + +$$ +\mathcal {L} = \frac {1}{C} \left(\mathcal {L} _ {M L} + \lambda \mathcal {L} _ {\text {c o n}}\right), \tag {7} +$$ + +where $C$ is the number of positive candidate moments (those that match the query) using momentum update, and $\lambda$ is a balancing hyperparameter. + +Inference. During inference, the main branch generates the start and end boundaries, along with the predicted confidence for all anchors in the pyramid. To eliminate duplicates, we apply SoftNMS [1] for dedduplication. + +# 4. Experiment + +# 4.1. Datasets + +Since our contributions primarily target the NLQ task, the Ego4D-NLQ dataset is our main choice. To further evaluate the versatility of our method, we also incorporated the Ego4D-Goal-Step dataset, another egocentric video grounding benchmark. Additionally, we used the TACoS dataset to enable more comprehensive comparisons. + +Ego4D-NLQ. Ego4D-NLQ consists of two versions. Ego4D-NLQ v1 contains 1,659 videos with 11,279, 3,874, and 4,004 video-query pairs for training, validation, and testing. Due to noisy data, Ego4D-NLQ v2 was released with 2,018 videos and 13,847, 4,552, and 4,004 video-query + +pairs for training, validation, and testing. Both versions share the same test set. + +Ego4D-Goal-Step. Ego4D-Goal-Step provides annotations for step grounding, which involves locating a video clip based on a step description. It contains 851 videos with 31,566, 7,696, and 5,540 video-query pairs for training, validation, and testing. + +TACoS. TACoS is an exocentric video moment localization dataset with 127 cooking videos. The training, validation, and test sets consist of 10,146, 4,589, and 4,083 video-query pairs. For consistency with existing works, we followed the setup in [57], using 9,790, 4,436, and 4,001 video-query pairs for training, validation, and testing. + +# 4.2. Experimental Settings + +Implementation Details. We segmented videos into clips using a sliding window of 16 frames for both window size and stride. Features are extracted using EgoVLP [22] (E) and InternVideo [2] (I), then concatenated as in [12]. Text features are extracted using CLIP (ViT-B/32) [36]. For TACoS, we used C3D [44] for video features and Glove [32] for text features to ensure a fair comparison [27]. The object's confidence score threshold $\theta$ is 0.6. In pretraining, we used a batch size of 16, a learning rate of $8e^{-4}$ , 4 warmup epochs, and 10 total epochs. In fine-tuning, warmup and total epochs are set to 4 and 10, respectively. Components that were not pretrained are initialized with weights from other pretrained structures. For further details, please refer to our released code. + +Baselines. On Ego4D-NLQ v1, we compared our model with several strong baselines, including methods utilizing the NaQ pretraining strategy (NaQ [37] and RGNet [10]) + +Table 2. Performance comparison under R@1 on Ego4D-Goal-Step. * indicates results that leverage the order prior of the dataset. + +
MethodValidationTest
0.30.50.30.5
EgoVideo [31]28.0223.6632.9925.92
BayesianVSLNet* [33]18.158.9735.1820.48
OSGNet29.6124.9432.7725.50
OSGNet*42.6135.3838.8330.16
+ +Table 3. Performance comparison on TACoS. * indicates results that utilize the E+I video feature. + +
MethodPublishedR@1R@5
0.30.50.30.5
VSLNet [52]ACL 202029.6124.27--
2D-TAN [57]AAAI 202037.2925.3257.8145.04
Tri-MRF [46]TMM 202452.4441.4976.0163.46
DPHANet [3]TMM 202447.0134.12--
MESM [26]AAAI 202452.6939.52--
MRNet [14]ACMMM 202455.4138.5477.1864.78
SnAG [27]CVPR 202456.4444.8681.1570.66
OSGNet57.5748.1882.0272.05
OSGNet*66.4355.7787.1679.45
+ +and methods without the NaQ pretraining strategy (InternVideo [2], CONE [13], SnAG [27], and RGNet [10]), using the same validation set for a fair comparison. Note that RGNet [10] and SnAG [27] employ different validation settings and have not reported performance on the Ego4D-NLQ v1 test set, we evaluated their performance in our unified setting. Further details see the Supplementary Material. On Ego4D-NLQ v2, we compared with strong baselines (NaQ [37], ASL [40], GroundNLQ [12], EgoEnv [29], EgoVideo [31], and GroundVQA [5]) utilizing the NaQ pretraining strategy. As GroundVQA [5] and NaQ [37] did not report performance on the Ego4D-NLQ v2 validation set, we used their published checkpoints for testing. + +On Ego4D-Goal-Step, we compared with baselines, including BayesianVSLNet [33], which leverages the order prior, and EgoVideo [31], which did not use prior information. On TACoS, we compared with the classic baselines (VSLNet [52] and 2D-TAN [57]) and current baselines (TriMRF [46], DPHANet [3], MESM [26], MRNet [14], and SnAG [27]). + +Evaluation Metric. Following previous work [13, 27], we utilized the metric $\mathrm{Rank}@\mathfrak{m}$ , IoU=n (R@m, n), which evaluates the percentage of queries that contain at least one correct moment among the top m retrieved predictions. A correct moment is defined as having an IoU greater than n with the ground truth. + +# 4.3. Performance Comparison + +On Ego4D-NLQ. Table 1 presents the performance comparison on the validation and the test sets of Ego4D-NLQ. On Ego4D-NLQ v2, our model outperforms all compared + +methods across all metrics. Specifically, compared to the strong baseline GroundVQA, our model achieves an absolute improvement of over $1.5\%$ in $\mathbf{R}@\mathbf{1}$ , 0.5 on both the validation and test sets. On Ego4D-NLQ v1, our model surpasses the baseline RGNet [10] in most metrics, including the challenging $\mathbf{R}@\mathbf{1}$ , 0.5. Our carefully designed architecture results in a nearly $4.0\%$ absolute improvement in $\mathbf{R}@\mathbf{1}$ , 0.5 after pretraining. In contrast, RGNet [10] shows considerably less benefit from pretraining, likely due to its dual-branch fusion structure. + +On Ego4D-Goal-Step. Table 2 presents the performance of our model compared to existing methods on the validation and test sets of Ego4D-Goal-Step. Our OSGNet outperforms all competing methods across all metrics on the validation set. Without any prior information, our model achieves an absolute improvement of approximately $1.59\%$ in $\mathbb{R}@\mathbb{1}$ , 0.3 on the validation set, compared to the strong baseline EgoVideo. When incorporating the order prior of the steps, as done by BayesianVSLNet [33], our model further improves by around $3.65\%$ in $\mathbb{R}@\mathbb{1}$ , 0.3 on the test set, surpassing BayesianVSLNet [33]. + +On TACoS. Table 3 presents the performance comparison of our model with existing methods on the test set of TACoS. Our OSGNet outperforms all competing methods across all metrics on the test set when using the C3D feature. Compared to the strong baseline SnAG, our model achieves an absolute improvement of approximately $3.32\%$ in the challenging metric R@1, 0.5. Additionally, when concatenating the InternVideo [2] and EgoVLP [22] video features, along with the pretraining strategy, our model achieves $66.43\%$ IoU=0.3 and $55.77\%$ IoU=0.5 at R@1. + +Summary. Comparing results across different datasets, we observed that our model performs lowest on Ego4D-NLQ, followed by Ego4D-Goal-Step, and highest on TACoS under the same settings. We attributed the relatively lower performance on egocentric video grounding datasets, compared to exocentric ones, to the inherent complexity of understanding egocentric video content. Furthermore, NLQ presents additional challenges due to its emphasis on grounding fine-grained objects within the background. Despite this, our model excels in R@1, 0.5 on both Ego4D-NLQ and TACoS, demonstrating the effectiveness of our proposed approach. + +# 4.4. Ablation Study + +On Model Structure. We conducted ablation studies on the validation set of Ego4D-NLQ v2 to assess the contributions of key components in our model. Specifically, we examined the impact of removing fine-grained object features by excluding the object branch and fusion block from the multi-modal encoder, as well as the effect of removing shot-level contrastive learning by eliminating the $\mathcal{L}_{con}$ term from the total loss function $\mathcal{L}$ . Table 4 presents the results + +![](images/0cdeceaed20b4f88257aeeb68e7b4998b1dfd960d3c6ab302cd657634bec1c5e.jpg) +(a) Query:Where was the drill machine? + +![](images/023f2c3edcabf5ce85a1bf1396ed89915e2fa4d1fd8e27bc67b1b31d43c270d6.jpg) +(b) Query:In what location did I last see the car? +Figure 4. Qualitative comparison with GroundVQA on Ego4D-NLQ. + +of these ablations. Removing $\mathcal{L}_{con}$ leads to a performance drop of around $0.4\%$ across all R@1 metrics, indicating that the shot branch plays a crucial role in enhancing the model's ability to learn better video representations. Additionally, removing the object feature results in over $2.0\%$ decline in the R@1 metric, highlighting the importance of object-level information for the NLQ task. + +On Shot Segmentation. Table 5 displays the effect of our shot segmentation method on the validation set of Ego4D-NLQ. We hypothesized that camera movement, in addition to head rotation, may be relevant for shot segmentation. To test this, we selected movement-related verbs from the generated captions that occur more than 100 times, including "walks", "moves around", "rides", "runs", "cycles", "jogs", and "jumps". Based on these frequencies, we divided the verbs into three groups: head rotation verbs (R: "turns around", "looks around"), high-frequency movement verbs (HM: "walks", "moves around"), and other main movement verbs (RM: "rides", "runs", "cycles", "jogs", "jumps"). We first examined segmentation using only R and HM and observed that the model using R for segmentation typically performs better on $\mathrm{R}@\mathrm{1}$ , which aligns with the main verbs in the captions. When combining R with HM and RM, the results are comparable to using R alone, leading us to adopt R for segmentation due to its efficiency. + +On Object Features. Table 6 illustrates the impact of different types of object features on the validation set of Ego4D-NLQ. The results indicate that text-based object features outperform image-based object features by $0.65\%$ in R@1, 0.3. We suppose that text-based object features align more effectively with the video modality. + +# 4.5. Qualitative Analysis + +We compared our model with the strong baseline GroundVQA [5] on the Ego4D-NLQ validation set through qualitative analysis. As shown in Figure 4(a), OSGNet accurately grounds the video moment where the "drill machine" is lo + +Table 4. Ablation studies on model structure. + +
LconObjectR@1R@5
0.30.50.30.5
XX28.8719.6054.5341.78
✓X29.2219.9955.1241.78
X✓31.2621.6458.1945.19
✓✓31.6322.0357.9145.19
+ +Table 5. Ablation studies on verb selection for shot segmentation. + +
RHMRMR@1R@5
0.30.50.30.5
✓✓✓31.6121.9958.5244.86
✓✓X31.2221.8456.8344.44
X✓X30.9121.2058.2845.41
✓XX31.6322.0357.9145.19
+ +Table 6. Ablation studies on the object features. + +
Object FeaturesR@1R@5
0.30.50.30.5
Image30.9821.6858.2245.14
Text31.6322.0357.9145.19
+ +ated. In contrast, GroundVQA struggles with fine-grained object localization due to the limitations of its video features, resulting in less accurate predictions. In Figure 4(b), OSGNet accurately selects the moment when the car was last seen, demonstrating stronger semantic alignment. On the other hand, while GroundVQA can identify the car's appearance in the video, it fails to predict the correct moment due to its limited video understanding. + +# 5. Conclusion + +In this paper, we propose OSGNet for egocentric video grounding. Our key improvements are twofold: 1) capturing object information to improve video representation, better aligning it with text queries; and 2) capturing shot information for contrastive learning, implicitly enhancing video representation. Through comprehensive experiments, we validate the effectiveness of these improvements. + +Acknowledgements. This work is partially supported by the National Natural Science Foundation of China, No.62236003, No.62476071, No.62376140, and No.U23A20315; the Shenzhen College Stability Support Plan, No.GXWD20220817144428005; the Science and Technology Innovation Program for Distinguished Young Scholars of Shandong Province Higher Education Institutions, No.2023KJ128, and the Special Fund for Taishan Scholar Project of Shandong Province. + +# References + +[1] Navaneeth Bodla, Bharat Singh, Rama Chellappa, and Larry S Davis. Soft-NMS-improving object detection with one line of code. In Proceedings of the IEEE International Conference on Computer Vision, pages 5561-5569, 2017. 6 +[2] Guo Chen, Sen Xing, Zhe Chen, Yi Wang, Kunchang Li, Yizhuo Li, Yi Liu, Jiahao Wang, Yin-Dong Zheng, Bingkun Huang, et al. Internvideo-ego4d: A pack of champion solutions to ego4d challenges. arXiv preprint arXiv:2211.09529, pages 1-11, 2022. 1, 2, 6, 7 +[3] Ruihan Chen, Junpeng Tan, Zhijing Yang, Xiaojun Yang, Qingyun Dai, Yongqiang Cheng, and Liang Lin. DPHANet: Discriminative parallel and hierarchical attention network for natural language video localization. IEEE Transactions on Multimedia, 2024. 7 +[4] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8199-8206, 2019. 2 +[5] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 2, 6, 7, 8 +[6] Xiang Fang, Daizong Liu, Pan Zhou, and Guoshun Nan. You can ground earlier than see: An effective and efficient pipeline for temporal sentence grounding in compressed videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2448-2460, 2023. 2 +[7] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. TALL: Temporal activity localization via language query. In Proceedings of the IEEE International Conference on Computer Vision, pages 5267-5275, 2017. 2 +[8] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonio Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 1, 2 +[9] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5356-5364, 2019. 4 +[10] Tanveer Hannan, Md Mohaiminul Islam, Thomas Seidl, and Gedas Bertasius. RGNet: A unified retrieval and grounding network for long videos. In European Conference on Computer Vision, pages 1-23, 2024. 6, 7 +[11] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE International Conference on Computer Vision, pages 5803-5812, 2017. 2 +[12] Zhijian Hou, Lei Ji, Difei Gao, Wanjun Zhong, Kun Yan, Chao Li, Wing-Kwong Chan, Chong-Wah Ngo, Nan Duan, and Mike Zheng Shou. Groundnloq@ ego4d natural language queries challenge 2023. arXiv preprint arXiv:2306.15255, pages 1-5, 2023. 3, 4, 6, 7 + +[13] Zhijian Hou, Wanjun Zhong, Lei Ji, Difei Gao, Kun Yan, Wk Chan, Chong-Wah Ngo, Mike Zheng Shou, and Nan Duan. CONE: An efficient coarse-to-fine alignment framework for long video temporal grounding. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 8013-8028, 2023. 2, 3, 6, 7 +[14] Jingjing Hu, Dan Guo, Kun Li, Zhan Si, Xun Yang, and Meng Wang. Maskable retentive network for video moment retrieval. In ACM Multimedia 2024, 2024. 7 +[15] Yupeng Hu, Meng Liu, Xiaobin Su, Zan Gao, and Liqiang Nie. Video moment localization via deep cross-modal hashing. IEEE Transactions on Image Processing, 30:4667-4677, 2021. 2 +[16] Yupeng Hu, Liqiang Nie, Meng Liu, Kun Wang, Yinglong Wang, and Xian-Sheng Hua. Coarse-to-fine semantic alignment for cross-modal moment localization. IEEE Transactions on Image Processing, 30:5933-5943, 2021. 2 +[17] Kumara Kahatapitiya, Anurag Arnab, Arsha Nagrani, and Michael S Ryoo. Victr: Video-conditioned text representations for activity recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18547-18558, 2024. 2 +[18] Juncheng Li, Junlin Xie, Long Qian, Linchao Zhu, Siliang Tang, Fei Wu, Yi Yang, Yueting Zhuang, and Xin Eric Wang. Compositional temporal grounding with structured variational cross-graph correspondence learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3032-3041, 2022. 2 +[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, pages 19730–19742. PMLR, 2023. 5 +[20] Xiaojie Li, Shaowei He, Jianlong Wu, Yue Yu, Liqiang Nie, and Min Zhang. Mask again: Masked knowledge distillation for masked video modeling. In Proceedings of the ACM International Conference on Multimedia, page 2221-2232. ACM, 2023. 1 +[21] Xiaojie Li, Jianlong Wu, Shaowei He, Shuo Kang, Yue Yu, Liqiang Nie, and Min Zhang. Fine-grained key-value memory enhanced predictor for video representation learning. In Proceedings of the ACM International Conference on Multimedia, page 2264-2274. ACM, 2023. 1 +[22] Kevin Qinghong Lin, Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z Xu, Difei Gao, Rong-Cheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. Advances in Neural Information Processing Systems, 35:7575-7586, 2022. 1, 2, 6, 7 +[23] Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Baoquan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. In The 41st international ACM SIGIR conference on research & development in information retrieval, pages 15-24, 2018. 2 +[24] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In Proceedings of the 26th ACM international conference on Multimedia, pages 843-851, 2018. 2 + +[25] Naiyuan Liu, Xiaohan Wang, Xiaobo Li, Yi Yang, and Yueting Zhuang. Refer@ zju-alibaba submission to the ego4d natural language queries challenge 2022. arXiv preprint arXiv:2207.00383, 2022. 3 +[26] Zhihang Liu, Jun Li, Hongtao Xie, Pandeng Li, Jiannan Ge, Sun-Ao Liu, and Guoqing Jin. Towards balanced alignment: Modal-enhanced semantic modeling for video moment retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3855-3863, 2024. 7 +[27] Fangzhou Mu, Sicheng Mo, and Yin Li. SnAG: Scalable and accurate video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18930-18940, 2024. 2, 4, 6, 7 +[28] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 2 +[29] Tushar Nagarajan, Santhosh Kumar Ramakrishnan, Ruta Desai, James Hillis, and Kristen Grauman. EgoEnv: Humancentric environment representations from egocentric video. Advances in Neural Information Processing Systems, 36, 2024. 6, 7 +[30] Yulin Pan, Xiangteng He, Biao Gong, Yiliang Lv, Yujun Shen, Yuxin Peng, and Deli Zhao. Scanning only once: An end-to-end framework for fast temporal grounding in long videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13767-13777, 2023. 3 +[31] Baoqi Pei, Guo Chen, Jilan Xu, Yuping He, Yicheng Liu, Kanghua Pan, Yifei Huang, Yali Wang, Tong Lu, Limin Wang, et al. EgoVideo: Exploring egocentric foundation model and downstream adaptation. arXiv preprint arXiv:2406.18070, pages 1-8, 2024. 1, 2, 6, 7 +[32] Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar, 2014. Association for Computational Linguistics. 6 +[33] Carlos Plou, Lorenzo Mur-Labadia, Ruben Martinez-Cantin, and Ana C Murillo. CARLOR@ Ego4D step grounding challenge: Bayesian temporal-order priors for test time refinement. arXiv preprint arXiv:2406.09575, pages 1-4, 2024. 2, 7 +[34] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5285-5297, 2023. 1, 2 +[35] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. ChatVTG: Video temporal grounding via chat with video dialogue large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1847-1856, 2024. 2 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 4, 6 +[37] Santhosh Kumar Ramakrishnan, Ziad Al-Halah, and Kristen Grauman. NaQ: Leveraging narrations as queries to supervise episodic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6694-6703. IEEE Computer Society, 2023. 1, 2, 6, 7 +[38] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics, 1:25-36, 2013. 1, 2 +[39] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2980-2988, 2017. 5 +[40] Jiayi Shao, Xiaohan Wang, Ruijie Quan, and Yi Yang. Action sensitivity learning for the ego4d episodic memory challenge 2023. arXiv preprint arXiv:2306.09172, 2023. 6, 7 +[41] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5026-5035, 2022. 2 +[42] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. Advances in Neural Information Processing Systems, 36, 2024. 2 +[43] Chaolei Tan, Jianhuang Lai, Wei-Shi Zheng, and Jian-Fang Hu. Siamese learning with joint alignment and regression for weakly-supervised video paragraph grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13569-13580, 2024. 2 +[44] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3D convolutional networks. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 4489-4497, Santiago, Chile, 2015. IEEE. 6 +[45] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5 +[46] Di Wang, Xiantao Lu, Quan Wang, Yumin Tian, Bo Wan, and Lihuo He. Gist/content,target-oriented:a 3-level human-like framework for video moment retrieval. IEEE Transactions on Multimedia, pages 1-13, 2024. 7 +[47] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2986-2994, 2021. 2 +[48] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18709-18719, 2024. 2 +[49] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9062–9069, 2019. 2 +[50] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Helping hands: An object-aware ego-centric video recognition model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13901-13912, 2023. 2 +[51] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510. Springer, 2022. 4 +[52] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6543-6554, 2020. 2, 7 +[53] Haoyu Zhang, Meng Liu, Zan Gao, Xiaojiang Lei, Yinglong Wang, and Liqiang Nie. Multimodal dialog system: Relational graph-based context-aware question understanding. In Proceedings of the 29th ACM International Conference on Multimedia, page 695-703. Association for Computing Machinery, 2021. 1 +[54] Haoyu Zhang, Meng Liu, Yuhong Li, Ming Yan, Zan Gao, Xiaojun Chang, and Liqiang Nie. Attribute-guided collaborative learning for partial person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14144–14160, 2023. 1 +[55] Haoyu Zhang, Meng Liu, Zixin Liu, Xuemeng Song, Yaowei Wang, and Liqiang Nie. Multi-factor adaptive vision selection for egocentric video question answering. In Proceedings of the 41st International Conference on Machine Learning, pages 59310-59328. PMLR, 2024. 1 +[56] Mingxing Zhang, Yang Yang, Xinghan Chen, Yanli Ji, Xing Xu, Jingjing Li, and Heng Tao Shen. Multi-stage aggregated transformer network for temporal language localization in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12669-12678, 2021. 2 +[57] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2D temporal adjacent networks for moment localization with natural language. Proceedings of the AAAI Conference on Artificial Intelligence, 34(07):12870-12877, 2020. 2, 6, 7 +[58] Wei Zhang, Chaoqun Wan, Tongliang Liu, Xinmei Tian, Xu Shen, and Jieping Ye. Enhanced motion-text alignment for image-to-video transfer learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18504-18515, 2024. 2 +[59] Yue Zhao, Ishan Misra, Philipp Krahenbihl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 5 + +[60] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-IoU loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12993-13000, 2020. 5 +[61] Hao Zhou, Chongyang Zhang, Yan Luo, Yanjun Chen, and Chuanping Hu. Embracing uncertainty: Decoupling and de-bias for robust temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8454, 2021. 2 +[62] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In *Forty-First International Conference on Machine Learning*, 2024. 4 +[63] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6748-6758, 2023. 4 \ No newline at end of file diff --git a/objectshotenhancedgroundingnetworkforegocentricvideo/images.zip b/objectshotenhancedgroundingnetworkforegocentricvideo/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9c80212bc14b459763f052cea4d7a92692ce6dfe --- /dev/null +++ b/objectshotenhancedgroundingnetworkforegocentricvideo/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9bdb9e5be608cd01a8c918ac11bda33e9df853a918798657ac24f056107b7a8f +size 526925 diff --git a/objectshotenhancedgroundingnetworkforegocentricvideo/layout.json b/objectshotenhancedgroundingnetworkforegocentricvideo/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cbaeee26e610e1b2ebab668105c27b7697a75e39 --- /dev/null +++ b/objectshotenhancedgroundingnetworkforegocentricvideo/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7940aeb5f541cd180a05bb69ddc9ec2aa697d23e5ff0160034a37792afba4f55 +size 430682 diff --git a/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/73b16b04-6826-4e05-8352-a4b07b235461_content_list.json b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/73b16b04-6826-4e05-8352-a4b07b235461_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9bc806896bd4bdca21973b71e5858b80df4259bc --- /dev/null +++ b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/73b16b04-6826-4e05-8352-a4b07b235461_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22f8aa89981ce65f1923b4b40f550da8b435a5712d27995a5c05d2bd2ee370c2 +size 89196 diff --git a/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/73b16b04-6826-4e05-8352-a4b07b235461_model.json b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/73b16b04-6826-4e05-8352-a4b07b235461_model.json new file mode 100644 index 0000000000000000000000000000000000000000..db81efb8de2f36370d64a2137b4df5b05c50dd0e --- /dev/null +++ b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/73b16b04-6826-4e05-8352-a4b07b235461_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80820171326b947192aca49e526e43c65d5770765216446283ee082abb598f19 +size 109848 diff --git a/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/73b16b04-6826-4e05-8352-a4b07b235461_origin.pdf b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/73b16b04-6826-4e05-8352-a4b07b235461_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d84a236b5838cafa78eb0c2f15ffc3b9c7f1f776 --- /dev/null +++ b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/73b16b04-6826-4e05-8352-a4b07b235461_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f43f14978e09c2254b00c7f6d71d5360abb17b4db7a89bb38fbd705d6a0d0b75 +size 2838146 diff --git a/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/full.md b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c9945136ec74b9ae078b86e33a84482dee06bd17 --- /dev/null +++ b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/full.md @@ -0,0 +1,324 @@ +# Occlusion-aware Text-Image-Point Cloud Pretraining for Open-World 3D Object Recognition + +Khanh Nguyen, Ghulam Mubashar Hassan & Ajmal Mian +The University of Western Australia +duykhanh.nguyen@research.uwa.edu.au +{ghulam.hassan, ajmal.mian}@uwa.edu.au + +# Abstract + +Recent open-world representation learning approaches have leveraged CLIP to enable zero-shot 3D object recognition. However, performance on real point clouds with occlusions still falls short due to unrealistic pretraining settings. Additionally, these methods incur high inference costs because they rely on Transformer's attention modules. In this paper, we make two contributions to address these limitations. First, we propose occlusion-aware text-image-point cloud pretraining to reduce the training-testing domain gap. From $52K$ synthetic 3D objects, our framework generates nearly $630K$ partial point clouds for pretraining, consistently improving real-world recognition performances of existing popular 3D networks. Second, to reduce computational requirements, we introduce DuoMamba, a two-stream linear state space model tailored for point clouds. By integrating two space-filling curves with 1D convolutions, DuoMamba effectively models spatial dependencies between point tokens, offering a powerful alternative to Transformer. When pretrained with our framework, DuoMamba surpasses current state-of-the-art methods while reducing latency and FLOPs, highlighting the potential of our approach for real-world applications. Our code and data are available at ndkhanh360.github.io/project-occtip. + +# 1. Introduction + +3D understanding plays a vital role in robotics [7], virtual reality [1], and autonomous driving [51], enabled by deep-learning models that perform recognition tasks such as 3D object classification [32], object detection [28, 34], and semantic segmentation [17, 20]. However, existing 3D networks [28, 32, 33, 44, 46] are trained using closed-set annotation, constraining them to recognize only pre-defined categories and struggle with 'unseen' ones. Inspired by CLIP [36], recent open-world studies [16, 23, 48, 53, 57, 59] have extended the aligned image-text latent space to include + +3D object representations, allowing generalization beyond 'seen' categories and enabling zero-shot 3D recognition. + +Existing works in this line of research take 3D-image-text triplets as input and align the three embedding spaces using cross-modal contrastive learning. These methods represent 3D shapes either as depth maps [16, 26, 53, 59] or raw point clouds [10, 23, 48, 49, 54, 57]. Depth-based approaches must first convert point clouds into 2D depth maps and use pretrained image encoders, such as Vision Transformer (ViT) [9], for 3D feature extractions. However, their performance typically suffers from information loss during the projection and the domain gap caused by differences between RGB and depth images. On the other hand, point-based methods [10, 23, 48, 49, 54, 57] can directly exploit all intrinsic geometry in the point clouds. An example is the recent work $\mathrm{CLIP}^2$ [52], which pretrains a 3D encoder using real-scanned objects extracted from scene-level point clouds. For contrastive learning, it pairs these with cropped images and simple category-based prompts as text descriptions. However, limited caption diversity and poor cropped image quality (due to occlusion, lighting, etc.) hinder CLIP knowledge transfer, leading to suboptimal performance. + +Other works [10, 23, 48, 49, 54, 57] instead leverage synthetic 3D models1 to construct pretraining triplets. These methods uniformly sample points from the mesh surface to create full point clouds2. They also render RGB images from preset camera positions and generate diverse captions from multiple sources, allowing for control over the quality of images and texts. As a result, these methods demonstrate promising zero-shot performance on complete point cloud benchmarks such as ModelNet40 [47]. However, their performance degrades significantly on real-scanned data, leading to unsatisfactory results in practical scenarios. As shown in Figure 1a, there is a $20\%$ accuracy drop from the synthetic ModelNet40 [47] to the real ScanOb + +![](images/8246425c0414bba550a77bcc4e96d94cd3b601eaf07dd31a48d01496e40d9af5.jpg) + +![](images/76e6664629caaf1320a8673c7a5ebdac7b164a4c8e81b5cb410181c7a4fa9e87.jpg) + +![](images/8b9536a44e93bd247e3a7cfd702026b6a43392c4c8deee96b63ae2da6873fb6c.jpg) +(a) Performance gap + +![](images/9659e3c6d22add7fe9f9715fb920333dd98ae4abc62037fdec874d7a3691aa07.jpg) +(b) Accuracy comparison + +![](images/bb8cf6bb873ee9c02f86cf920fc47e0e4f06a9863a83614fe1e63763e2e8f8fc.jpg) + +![](images/35dd45cad1543600dad933776d72d54597221c8493692b62635eb831cb419023.jpg) +(c) FLOPs and latency comparison +Figure 1. Comparison to existing methods. (a) State-of-the-art approaches pretrain 3D encoders on complete point clouds, which differ significantly from occluded ones in practical scenarios (top). This leads to a substantial gap in zero-shot performance between ModelNet40 [47] benchmark with full point clouds and ScanObjectNN [41] with real-world data (bottom). (b) The proposed framework OccTIP pretrains 3D models on partial point clouds to better simulate practical conditions, leading to significant improvements on various recognition tasks, especially when combined with our DuoMamba architecture. (c) Compared to the popular PointBERT [50], DuoMamba has significantly lower FLOPs (top) and latency (bottom) during inference, making it better suited for real-world applications. + +jectNN [41], caused by the large domain gap between complete point clouds in pretraining and occluded ones encountered in real-world conditions. To address this data discrepancy, we introduce an occlusion-aware pretraining framework that leverages synthetic 3D meshes to create partial point clouds. We simulate real-world scenarios by putting a virtual camera around an object and only sample points visible from the camera position. From 52K ShapeNetCore [2] 3D models, our framework generates nearly 630K occluded point clouds for pretraining and enhances the zero-shot accuracy of SparseConv [4] and PointBERT [50] by $3.8\%$ and $5.1\%$ on ScanObjectNN [41]. Despite using only synthetic objects, our framework consistently improves recognition performance on various real-world tasks and even outperforms methods that use real-scanned data. + +Moreover, existing multi-modal pretraining approaches [10, 23, 48, 49, 57] heavily rely on Transformer-based 3D encoders due to their strong learning capacity. However, these pretrained models have high inference costs because of the attention's quadratic complexity. This poses significant challenges when we want to increase the number of point tokens in the input or use the pretrained encoder as a classification head in a 3D object detector. Inspired by Mamba [12], we introduce an efficient architecture named DuoMamba as an alternative to Transformer-based models. At the core of our network is the two-stream DuoMamba block, developed using linear-time S6 modules from Mamba [12]. Each stream processes point tokens in the order from a space-filling curve, either Hilbert [15] or its transposed variant Trans-Hilbert. Intuitively, these turn an unordered point cloud into a geometrically structured sequence where close points in 3D space stay adjacent in the sequence, facilitating S6 to capture meaningful geometric + +relationships. We also replace causal 1D convolutions commonly used in Mamba models [12, 14, 22, 24] with standard 1D convolutions to allow point tokens to aggregate information of their neighbors in both directions, enriching their spatial context. Compared to the popular Transformer-based PointBERT [50], our model achieves higher performance across several benchmarks (Figure 1b) while significantly reducing FLOPs and latency (Figure 1c). It also exhibits a better performance-computation balance than existing Mamba-based point cloud networks [14, 22]. In summary, our main contributions are: + +- We propose an occlusion-aware pretraining framework for open-world 3D recognition. By generating partial point clouds from synthetic 3D models, our approach simulates real-world conditions and removes the need for real-scanned data in pretraining. +- We demonstrate, through extensive experiments, that our framework consistently improves the performance of two popular networks: PointBERT [50] and SparseConv [4]. +- We introduce DuoMamba, a two-stream linear-time architecture integrated with space-filling curves and 1D convolutions for efficient point cloud learning. Our network achieves higher accuracy than Transformer-based methods, with reduced computation and lower latency. + +# 2. Related Work + +CLIP for 3D Representation Learning. Vision-Language Models (VLMs) such as CLIP [36] and ALIGN [18] have demonstrated impressive zero-shot capabilities through contrastive learning on large image-text corpora. These models effectively map the two modalities into a shared latent space with rich semantic and visual concepts, forming a foundation for various 2D applications [19, 37, 38, 58]. + +![](images/4d280771aa0e6b862d13a526c5a2021809249061bbadb6e8119039b5ad7fc486.jpg) +(a) Triplets Generation + +![](images/e4b6dbef53d56672d755d175b28b094cf5aa07b5ac6814d0f9641cbbc983c6cc.jpg) +(b) Text-Image-Point Cloud Pretraining +Figure 2. Overview of our OccTIP pretraining framework. (a) Given a 3D object, we generate RGB and depth images from preset camera positions, which are used to construct partial point clouds. Texts are generated from dataset metadata, image captioning models [21], and retrieved descriptions of similar photos from LION-5B [39]. (b) During pretraining, we extract multi-modal features using a learnable point cloud network and frozen CLIP [36] encoders, then align them through contrastive learning. + +![](images/1b110cef3bef8e6d529da97fcb5b8b56018253652f65d9c6b458f85fb2ee4511.jpg) + +Recently, several studies have leveraged CLIP for 3D representation learning, showing promising results in object-level zero-shot 3D recognition [10, 16, 23, 26, 48, 52-54]. + +Among them, several works [16, 26, 53, 59] project point clouds into depth maps and rely on fine-tuning CLIP image encoders for zero-shot classification. However, they often experience information loss during 3D-to-depth projections, which significantly impacts their performance. In contrast, other methods [10, 23, 48, 49, 52, 54, 57] train specialized point cloud encoders to distill CLIP knowledge, extending the image-text co-embedding space to encompass 3D representations. These approaches form text-image-point cloud triplets and utilize contrastive learning to align the latent spaces of the three modalities. For instance, $\mathrm{CLIP}^2$ [52] uses object point clouds and images from real scenes to generate pretraining triplets. However, the quality of the cropped images can vary due to lighting conditions, object size, and occlusion. Also, object descriptions are created from simple prompts, leading to suboptimal transfer of CLIP knowledge and unsatisfactory performance. Other works [10, 23, 48, 49, 54, 57] use synthetic 3D models to render RGB images and leverage metadata, image captioning models [21], and retrieved texts for diverse descriptions. However, they typically pretrain 3D encoders on complete point clouds, which greatly differ from the real ones encountered in practical conditions due to occlusion and viewpoint limitations. To address this, we propose a framework that uses synthetic 3D models to generate occluded point clouds for pretraining, reducing data discrepancies while maintaining high image quality and caption diversity for effective transfer of CLIP's knowledge. + +Deep Learning-Based Point Cloud Encoders. Leverag- + +ing deep learning, the pioneering PointNet [32] directly processes point clouds using multi-layer perceptrons applied on each point independently. Subsequent methods [33, 35, 44] introduce hierarchical structures to model local neighborhoods and geometric relationships, addressing PointNet's limitations. Alternatively, convolution-based approaches [11, 27] convert point clouds into 3D voxel grids, utilizing established 3D convolutions for feature learning. SparseConv [11] reduces the high memory requirements of 3D convolutions through sparse convolution, enhancing the voxel-based method's applicability. Since the introduction of self-attention in Transformers [43], most state-of-the-art encoders [30, 45, 46, 50] are based on this architecture, with PointBERT [50] being a representative for object-level point cloud pretraining [23, 48-50, 54]. However, the attention mechanism's quadratic complexity results in high computational costs as the input length increases. + +To overcome this, Mamba3D [14] and PointMamba [22] were developed using the linear-time S6 from Mamba [12] as alternatives to attention layers. However, these networks overlook key characteristics of point clouds. Specifically, Mamba3D [14] applies S6 to point tokens in random order due to the unstructured nature of point clouds, which is not optimal since S6 was designed for sequence data with meaningful order, such as natural language and audio. PointMamba [22] improves on this by sorting points using Hilbert and Trans-Hilbert curves [15], ensuring that spatially close points remain adjacent in the sequence. However, it simply concatenates the two resulting orders as input for S6, doubling the sequence length and computations. Moreover, both methods employ causal 1D convolution, which is beneficial for causal data like audio but subopti + +![](images/7f74c779c411713c9309c66fba5ed24679453fb133211a0c6a1e60624122ce86.jpg) +Figure 3. Overview of the proposed architecture and detailed design of our DuoMamba block. We integrate two Hilbert curves [15] and standard 1D convolutions with linear-time S6 [12] modules to efficiently model geometric dependencies and enrich spatial context. + +mal for spatial data. Therefore, we propose a new Mamba-based architecture that integrates point cloud properties into its design and leverages multi-modal pretraining to enhance model knowledge and extend its applicability. + +# 3. Preliminaries + +State Space Model represents a continuous system that maps an input $x_{t}$ to an output $y_{t}$ via an implicit latent state $h_t \in \mathbb{R}^N$ . S4 [13] introduces a discretized version for sequence-to-sequence transformation, defined as: + +$$ +h _ {t} = \bar {\boldsymbol {A}} h _ {t - 1} + \bar {\boldsymbol {B}} x _ {t}, \quad y _ {t} = \boldsymbol {C} h _ {t}, \tag {1} +$$ + +where $\overline{A}$ and $\overline{B}$ are derived from the model parameters $(A,B,C,\Delta)$ using zero-order hold discretization. As the update matrices $\overline{A},\overline{B},C$ are shared across time steps, S4 achieves linear-time computation through a convolution kernel, though its capacity to capture dynamic input sequences is limited. To improve context awareness, the Selective SSM (S6) introduced in Mamba [12] makes $B,C,\Delta$ dependent on the input. To maintain near-linear time complexity, Mamba [12] employs a hardware-aware implementation for S6, which we follow to ensure computational efficiency. For further details, please refer to [12]. + +Space-Filling Curves pass through every point in a high-dimensional space, preserving spatial proximity of the original structure. For point clouds, they can be defined as a bijective function $\Phi : \mathbb{Z}^3 \to \mathbb{Z}$ , mapping each $(x,y,z)$ coordinate to a position in a 1D sequence. Our Duom-Mamba leverages the Hilbert space-filling curve [15] and its transposed variant (Trans-Hilbert) for their strong locality-preserving properties, ensuring that points close in 3D space remain adjacent in the sequence. This is especially valuable for point cloud processing, where points are inherently unordered, making it challenging for sequence models like S6 to capture geometric relationships. By establishing a meaningful order with Hilbert curves, we enable S6 to model spatial dependencies in point clouds more effectively. + +Cross-Modal Contrastive Learning. CLIP [36] is a pioneering approach that employs cross-modal contrastive + +learning to align embeddings of the same concept across two modalities (e.g., a caption "this is a dog" and an image of a dog) by pulling their representations closer in a shared-embedding space while pushing apart those of different concepts. Formally, for a batch of $B$ paired features from two modalities $M_1$ and $M_2$ , represented as $\{(z_i^{M_1}, z_i^{M_2})\}_{i=1}^B$ , the training objective is to minimize the contrastive loss $\mathcal{L}^{M_1 \leftrightarrow M_2}$ , defined as: + +$$ +\mathcal {L} ^ {M _ {1} \leftrightarrow M _ {2}} = - \frac {1}{2} \left(l ^ {M _ {1} \rightarrow M _ {2}} + l ^ {M _ {2} \rightarrow M _ {1}}\right), \tag {2} +$$ + +with $l^{M_1 \to M_2}$ and $l^{M_2 \to M_1}$ calculated as follows: + +$$ +l ^ {M _ {1} \rightarrow M _ {2}} = \sum_ {i = 1} ^ {B} \log \frac {\exp \left(z _ {i} ^ {M _ {1}} \cdot z _ {i} ^ {M _ {2}} / \tau\right)}{\sum_ {j = 1} ^ {B} \exp \left(z _ {i} ^ {M _ {1}} \cdot z _ {j} ^ {M _ {2}} / \tau\right)}, \tag {3} +$$ + +$$ +l ^ {M _ {2} \rightarrow M _ {1}} = \sum_ {i = 1} ^ {B} \log \frac {\exp (z _ {i} ^ {M _ {2}} \cdot z _ {i} ^ {M _ {1}} / \tau)}{\sum_ {j = 1} ^ {B} \exp (z _ {i} ^ {M _ {2}} \cdot z _ {j} ^ {M _ {1}} / \tau)}, +$$ + +where $\tau$ is a temperature parameter that controls the sharpness of the Softmax distributions during training. + +# 4. Pretraining Framework + +Triplets Generation. Given a 3D model, we first center and normalize it to lie within a unit sphere. Following OpenShape [23], we select 12 camera positions uniformly distributed around the object. They lie on a sphere of radius 2, with four viewpoints above the mesh $(z > 0)$ , four at the same level $(z = 0)$ , and four below $(z < 0)$ to cover all angles. From each position, we render an RGB image and a depth map using BlenderProc [8]. We set up the scene with area light and use Blender's ray-tracing render engine 'CYCLES' for more realistic output. We then construct a partial point cloud based on the camera position, color information from the RGB image, and geometric information in the depth map. On a single RTX 4090 GPU, it takes around three days to process 52K ShapeNetCore [2] 3D models. + +For language modality, we use the captions provided by OpenShape [23], which come from three sources: (1) metadata of the dataset, (2) captions generated by BLIP [21] and + +![](images/559a01faf744d39659a25d2fc6a922467b546fe1db268b2c3af370942b7e19a7.jpg) +Figure 4. t-SNE visualization of ScanObjectNN [41] features extracted by different pretraining methods. Compared to other approaches based on complete point clouds, our method OccTIP achieves clearer class separation and significantly reduces overlap between classes. + +![](images/11df0542191b5b18fab067abeb15b70d237d0e6a6bfcd8ce9d4d39a2b2d4408c.jpg) + +![](images/6a52c142d27ccc309b5f91deddd0798a5a0e7f13ab87d50597e26daf30fd9a1a.jpg) + +Azure cognitive services, and (3) retrieved captions of visually similar images from LAION-5B dataset [39]. An illustration of the generation process is shown in Figure 2a. + +Training Pipeline. For the $i$ -th object, we randomly select the partial point cloud and RGB image corresponding to a viewpoint, along with a text from its available captions, to form a triplet $x_{i} = (T_{i},I_{i},P_{i})$ . Input at each iteration is a batch of $B$ triplets, represented as $\{(T_i,I_i,P_i)\}_{i = 1}^B$ . + +The framework trains a point cloud network $f^{P}$ to learn 3D representations that are aligned with the embedding spaces of language and images. To achieve this, we leverage pretrained text and image encoders from CLIP [3], denoted as $f^{T}$ and $f^{I}$ , to generate prior features that serve as anchors in the new co-embedding space. Since CLIP [36] was trained on a large image-text corpus and provides a well-aligned latent space, we freeze the primary CLIP [36] encoders during training. To enable flexible alignment of this shared latent space with the additional 3D modality, we introduce learnable projection heads $h^{I}$ and $h^{T}$ for image and text inputs. They will be updated jointly with the point cloud model during pretraining. Given an input batch, we extract features for each modality as follows: + +$$ +z _ {i} ^ {T} = h ^ {T} \left(f ^ {T} \left(T _ {i}\right)\right), z _ {i} ^ {I} = h ^ {I} \left(f ^ {I} \left(I _ {i}\right)\right), z _ {i} ^ {P} = f ^ {P} \left(P _ {i}\right). \tag {4} +$$ + +Inspired by MixCon3D [10], we introduce additional mixed representations to enhance contrastive learning constraints. Specifically, we compute a combined embedding from the point cloud and image features as: + +$$ +z _ {i} ^ {M} = h ^ {M} \left(\operatorname {C o n c a t} \left(z _ {i} ^ {P}, z _ {i} ^ {I}\right)\right), \tag {5} +$$ + +where $h^M$ is a learnable linear projection mapping the concatenated features to the shared latent space. + +Finally, we employ cross-modal contrastive learning to 'pull' multi-modal features together. The training objective is to minimize the following total loss: + +$$ +\mathcal {L} = \mathcal {L} ^ {P \leftrightarrow I} + \mathcal {L} ^ {P \leftrightarrow T} + \mathcal {L} ^ {I \leftrightarrow T} + \mathcal {L} ^ {M \leftrightarrow T}, \tag {6} +$$ + +where $T, I, P,$ and $M$ denote text, image, point cloud, + +and mixed modalities, with each contrastive loss defined in Equation 2. Our training pipeline is illustrated in Figure 2b. + +# 5. DuoMamba + +Overview. Given an input point cloud $P_0 \in \mathbb{R}^{N \times 3}$ (and color information $C_0 \in \mathbb{R}^{N \times 3}$ if available), we first apply Farthest Point Sampling (FPS), similar to previous works [4, 50], to obtain a set of $S$ center points, denoted as $P \in \mathbb{R}^{S \times 3}$ . Next, kNN is applied to form a local patch with $k$ points around each center. These point patches (along with points' color) are then processed by a mini-PointNet [32] to obtain point tokens $E \in \mathbb{R}^{S \times C}$ , where $C$ is the dimension of the token embedding space. An encoder composed of $L$ DuoMamba blocks is employed to propagate information across local patches and capture global features. Finally, the encoder outputs are passed through a linear layer followed by average pooling to produce a single vector $z^P \in \mathbb{R}^C$ , which can be used for cross-modal contrastive learning as described in Section 4. Figure 3 illustrates our network. + +DuoMamba Block. At the core of the proposed architecture is the two-stream DuoMamba block, which leverages Mamba's linear complexity [12] for improved efficiency over the Transformer's quadratic self-attention [30, 50, 55, 56]. We introduce two key adaptations to Mamba, originally designed for structured sequence data, to efficiently process point clouds. First, we use Hilbert and TransHilbert [15] space-filling curves to transform 3D point clouds into 1D sequences. Unlike audio or text, point clouds are essentially sets of unordered 3D coordinates, making them challenging to process with order-aware models like Mamba. By sorting point tokens along Hilbert curves, adjacent patches in the sequence correspond to nearby regions in 3D space, facilitating local information propagation compared to random ordering. Additionally, using two Hilbert variants allows us to capture more diverse spatial relationships, enriching local point interactions [46]. Second, we replace the causal 1D convolution used in previous Mamba-based models [12, 14, 22, 24] with the standard convolution. In tasks like audio and language modeling where data fol + +
MethodEncoderModelNet40-PScanObjectNN
Top 1Top 3Top 5Top 1Top 3Top 5
OpenShape [23]42.161.669.452.772.783.6
TAMM [54]45.564.873.157.975.383.1
MixCon3D [10]SparseConv [4]---54.473.983.3
MixCon3D† [10]42.159.367.556.073.282.8
OccTIP64.581.086.761.778.486.9
OpenShape [23]46.364.271.951.369.478.4
TAMM [54]45.666.274.754.874.583.3
MixCon3D [10]PointBERT [50]---52.669.978.7
MixCon3D† [10]50.369.778.655.572.881.1
OccTIP67.782.787.360.678.286.0
OpenAlign [26]ViT-H-14 [9]---59.576.883.7
OccTIPDuoMamba67.782.987.863.581.389.2
+ +ows a natural order, restricting tokens to attend only to preceding ones can be beneficial [29]. By contrast, for spatial data like point clouds, allowing patches to aggregate information bidirectionally along scanning curves enables them to consider neighbors in every direction, providing a more comprehensive spatial context. + +Figure 3 illustrates our DuoMamba block, which consists of two parallel streams that extract point features using two S6 modules [12]. In each branch, point patches are ordered along the Hilbert or Trans-Hilbert curve, then local relationships are propagated with a 1D convolution. S6 further facilitates information flow between tokens and models long-range dependencies. Finally, two sequences are reordered and combined to produce output. Specifically, the $l$ -th block transforms the output $Z_{l-1}^{\mathrm{out}}$ from the previous module as follows: + +$$ +\begin{array}{l} Z _ {l} ^ {\mathrm {i n}} = \operatorname {L a y e r N o r m} \left(Z _ {l - 1} ^ {\mathrm {o u t}}\right), \qquad \qquad Z _ {l} = \operatorname {S i L U} \left(\operatorname {L i n e a r} \left(Z _ {l} ^ {\mathrm {i n}}\right)\right), \\ H _ {l} ^ {\prime} = \operatorname {H S o r t} \left(\operatorname {L i n e a r} \left(Z _ {l} ^ {\mathrm {i n}}\right)\right), \qquad H _ {l} ^ {\prime \prime} = \operatorname {S i L U} \left(\operatorname {C o n v 1 D} \left(H _ {l} ^ {\prime}\right)\right), \\ T _ {l} ^ {\prime} = \operatorname {T H S o r t} \left(\operatorname {L i n e a r} \left(Z _ {l} ^ {\mathrm {i n}}\right)\right), \qquad T _ {l} ^ {\prime \prime} = \operatorname {S i L U} \left(\operatorname {C o n v 1 D} \left(T _ {l} ^ {\prime}\right)\right), \\ H _ {l} = \operatorname {U n s o r t} \left(\mathrm {S} 6 \left(H _ {l} ^ {\prime \prime}\right)\right) \odot Z _ {l}, \qquad T _ {l} = \operatorname {U n s o r t} \left(\mathrm {S} 6 \left(T _ {l} ^ {\prime \prime}\right)\right) \odot Z _ {l}, \\ Z _ {l} ^ {\text {o u t}} = Z _ {l - 1} ^ {\text {o u t}} + \operatorname {L i n e a r} \left(H _ {l} + T _ {l}\right), \tag {7} \\ \end{array} +$$ + +where HSort and THSort represent sorting operations based on Hilbert and Trans-Hilbert curves while Unsort is the operation that restores the original order. + +# 6. Experiments + +We generate text-image-point cloud pretraining triplets using 52,417 3D models from the ShapeNetCore [2] dataset following the procedure in Section 4. For evaluation, we create 12 partial point clouds for each ModelNet40 [47] test object, resulting in ModelNet40-P with 29,610 occluded point clouds of 40 classes. We also use three other real-scanned benchmarks: ScanObjectNN [41], ScanNetV2 [5], and SUN RGB-D [40]. For fair comparisons, we mainly + +Table 1. Zero-shot classification accuracy on ModelNet-P and ScanObjectNN [41]. ScanObjectNN results are from prior work, while ModelNet40-P results are obtained by running official pretrained models. $\dagger$ : As MixCon3D [10] weights are unavailable, we retrain 3D encoders using the authors' code, achieving higher accuracy than previously reported, which we use in all comparisons. + +
MethodEncoderScanObjectNN
1-shot2-shot4-shot8-shot16-shot
OpenShape* [23]SparseConv [4]41.749.158.163.470.0
MixCon3D† [10]42.451.159.965.471.5
TAMM* [54]43.652.859.570.973.4
OccTIP48.558.267.372.376.7
OpenShape* [23]PointBERT [50]38.450.961.367.771.5
MixCon3D† [10]39.053.961.268.071.5
TAMM* [54]46.356.766.673.477.6
OccTIP50.860.868.673.776.8
OccTIPDuoMamba52.563.070.276.079.4
+ +Table 2. Few-shot linear probing accuracy on ScanObjectNN [41]. (*: results obtained using released pretrained weights, †: results reproduced using the authors' public code.) + +evaluate our work against methods that are also pretrained on ShapeNetCore [2] objects, including OpenShape [23], MixCon3D [10], and TAMM [54]. + +Evaluation Tasks. We conduct extensive experiments in four recognition tasks with varying difficulty levels (zero-shot classification, few-shot linear probing, real-world instance recognition, and zero-shot object detection) to demonstrate the superiority of our pretraining framework OccTIP and the proposed architecture DuoMamba. The details of each experiment will be described in the following subsections. + +Implementation and Training Details. We implement our method in PyTorch [31] and conduct all experiments on a single NVIDIA RTX 4090 GPU. We sample 2,048 points per point cloud as input and train the 3D encoders for 200 epochs using AdamW [25] with a 10-epoch warmup, which takes around 1.5 days. Following prior works [10, 23, 54], we use OpenCLIP ViT-bigG-14 [3] as pretrained image-text encoders. Further details are in the supplementary material. + +# 6.1. Zero-Shot Classification + +A pretrained network can perform zero-shot classification without fine-tuning by comparing its 3D shape representations to text embeddings of candidate categories. To assess the quality of the learned latent space, we conduct zero-shot classification experiments on ModelNet40-P and ScanObjectNN [41] (OBJ_ONLY version). ScanObjectNN [41] contains 2,890 real-scanned point clouds in 15 classes, providing a more realistic benchmark than our synthetic ModelNet40-P. As summarized in Table 1, our method significantly outperforms previous approaches. For SparseConv [4] and PointBERT [50], our framework improves their performance by $19.0\%$ and $17.4\%$ on ModelNet40-P compared to the best existing results. On ScanObjectNN [41], OccTIP raises accuracy by $3.8\%$ and $5.1\%$ , reaching $61.7\%$ and $60.6\%$ , both surpassing the current state-of-the-art OpenDlign [26]. These results highlight our framework's effectiveness in bridging the training-testing domain gap for improved real-world recognition. Furthermore, when combining OccTIP with DuoMamba, accuracy increases by an additional $1.8\%$ , establishing a new state-of- + +
MethodAvg.CabBedChairSofaTablDoorWindBksfPicCntrDeskCurtFridgShwrCurtToilSinkBath
PointCLIP [53] w/ TP.26.10.055.772.85.05.11.70.077.20.00.051.70.00.040.385.349.20.0
CLIP2Point [16] w/ TP.35.211.80.045.127.610.561.52.671.90.333.629.94.711.592.486.134.072.2
\( CLIP^2 \) [52]38.567.232.669.342.318.319.14.062.61.412.752.840.19.141.071.045.559.7
OpenShape* [23] (SparseConv)39.90.059.376.861.942.357.014.271.431.10.067.720.90.00.089.743.941.9
MixCon3D† [10] (SparseConv)39.90.069.169.667.043.751.05.775.353.21.959.84.50.010.793.112.261.3
TAMM* [54] (SparseConv)43.70.567.972.572.252.951.820.277.950.025.061.47.50.00.087.936.758.1
OccTIP (SparseConv)45.31.171.680.787.645.752.35.070.156.33.964.64.51.80.096.651.077.4
OpenShape* [23] (PointBERT)40.50.560.570.567.041.750.19.272.744.13.975.67.50.00.072.454.158.1
MixCon3D† [10] (PointBERT)40.30.860.573.274.256.965.52.564.961.71.962.23.00.010.767.28.271.0
TAMM* [54] (PointBERT)38.61.956.871.166.046.965.717.767.523.47.775.60.00.00.072.444.938.7
OccTIP (PointBERT)47.811.680.373.083.554.956.816.761.072.51.953.511.921.17.191.441.871.2
OccTIP (DuoMamba)49.04.679.077.187.654.652.310.379.261.33.953.531.316.80.094.856.171.0
+ +the-art of $63.5\%$ on ScanObjectNN [41] and demonstrating DuoMamba's learning prowess in cross-modal representation learning. + +# 6.2. Few-Shot Linear Probing + +To further evaluate the learned embedding space, we conduct few-shot linear probing on ScanObjectNN [41]. Following OpenShape [23], we use a pretrained model to extract features for all test samples and train a linear classifier using only a limited number of labeled instances per class. We report classification accuracy across a range of few-shot settings, specifically with 1, 2, 4, 8, and 16 labeled samples per category. As shown in Table 2, when trained with OccTIP, PointBERT [50] and SparseConv [4] consistently achieve better results than existing approaches in nearly all few-shot settings. DuoMamba further enhances performance, attaining the highest accuracy under all configurations. This showcases the proposed network's strong learning capacity and highlights our framework's effectiveness in facilitating transferable feature learning, underscoring its applicability in label-scarce scenarios. + +# 6.3. Real-World Instance Recognition + +Following prior work [52, 54], we test the pretrained models' capability to understand complex objects with the real-world instance recognition task. In this setting, the models have to classify object instances from a scene in a zero-shot manner. Using the same setting as $\mathrm{CLIP}^2$ [52], we report results on the popular scene-level ScanNetV2 [5] dataset. We extract object instances using ground-truth instance masks and classify them with the pretrained models. Table 3 summarizes the per-class accuracy and overall class average. + +Our method significantly outperforms approaches pretrained on 1.6M real-world text-image-point cloud triplets, including PointCLIP w/TP [53], CLIP2Point w/TP [16], and $\mathrm{CLIP^2}$ [52]. Compared to other ShapeNetCore-based pretraining methods, OccTIP consistently boosts PointBERT [50] and SparseConv [4] accuracy by $7.3\%$ and $1.6\%$ , respectively. When combined with DuoMamba, the class-average accuracy rises by an additional $1.2\%$ , reaching $49.0\%$ . These results once again underscore our model's + +strong learning capacity and highlight the effectiveness of our pretraining framework for robust feature extraction in real-world 3D shape understanding. + +Table 3. Zero-shot classification accuracy on the real-world ScanNetV2 [5] instances. (*: results obtained using released pretrained weights, †: results reproduced using the authors' public code.) + +
MethodScanNetV2SUN RGB-D
\( \mathrm{mAP}_{25} \)PointCLIP [16]6.0-
PointCLIP V2 [59]19.0-
OpenShape* [23]20.418.6
MixCon3D† [10]24.118.7
TAMM* [54]23.118.9
OccTIP28.924.4
\( \mathrm{mAP}_{50} \)PointCLIP [53]4.8-
PointCLIP V2 [59]11.5-
OpenShape* [23]16.19.8
MixCon3D† [10]19.19.6
TAMM* [54]18.110.0
OccTIP22.713.0
+ +Table 4. Zero-shot 3D object detection results on ScanNetV2 [5] and SUN RGB-D [40]. For complete results, please refer to our supplementary materials. (*: results obtained using released pretrained weights, †: results reproduced using the authors' code.) + +# 6.4. Zero-Shot 3D Object Detection + +To showcase how our pretrained model can be combined with existing methods to tackle more challenging tasks, we conduct zero-shot 3D object detection experiments on ScanNetV2 [5] and SUN RGB-D [40]. Following the setup in PointCLIP V2 [59], we leverage 3DETR-m [28] detector to predict 3D bounding boxes, which enables the extraction of points corresponding to each object instance. Our pretrained 3D network is then applied to classify these object point clouds in a zero-shot manner. Based on 3DETR-m's localization and our classifier's semantic predictions, we calculate the mean Average Precision (mAP) at IoU thresholds of 0.25 and 0.5 across 18 object categories in ScanNetV2 and 10 most frequent classes in SUN RGB-D. + +As shown in Table 4, our method achieves $\mathrm{mAP}_{25}$ and $\mathrm{mAP}_{50}$ scores of $28.9\%$ and $22.7\%$ on ScanNetV2, marking significant improvements of $9.9\%$ and $11.2\%$ over the depth-based PointCLIP V2 [59]. Compared to other point-based methods, we outperform the second-best approach MixCon3D [10] by $4.8\%$ and $3.6\%$ on $\mathrm{mAP}_{25}$ and $\mathrm{mAP}_{50}$ , respectively. A similar trend is observed on the SUN RGB- + +
SettingHilbertTrans-HilbertConv1DScanObjectNN
(i)---60.6
(ii)✓-✓62.2
(iii)-✓✓61.7
(iv)✓✓-63.1
(v)✓✓✓63.5
+ +(a) Contribution of each component in our two-stream DuoMamba block. + +
Point orderScanObjectNN
FPS order61.8
Z-order and Trans-Z-order62.7
Hilbert and Z-order62.4
Hilbert and Trans-Hilbert63.5
+ +(b) Effect of different sorting strategies on DuoMamba's performance. + +
ModelParam. (M) ↓FLOPs (G) ↓ScanObjectNN ↑
Mamba3D [14]29.96.860.7
PointMamba [22]21.410.362.6
DuoMamba29.27.163.5
+ +(c) Mamba-based encoders comparison. Our DuoMamba architecture achieves a better computation-performance trade-off than previous methods. + +Table 5. Ablation studies to validate the design of our proposed network and comparisons with existing Mamba-based point cloud models. + +D benchmark, where our approach achieves the highest $\mathrm{mAP}_{25}$ and $\mathrm{mAP}_{50}$ scores of $24.4\%$ and $13.0\%$ . The results again confirm the superiority of our method in learning robust features for recognizing noisy 3D objects in complex scenes, highlighting its strong potential for general 3D open-world learning. + +# 6.5. Visualization of the Embedding Space + +We further compare the latent spaces of our model with those of existing works. Specifically, we use the pretrained encoders to extract features of ScanObjectNN [41] test instances and employ t-SNE [42] for dimensionality reduction. As shown in Figure 4, our method exhibits superior separation and clustering of object classes compared to previous approaches. Note that the pretrained model did not encounter any of these samples during training, yet it successfully captures the characteristics of each category and minimizes overlap between them. This separation indicates that our method's feature representations are more robust, leading to better real-world zero-shot performance as demonstrated in previous experiments. + +# 6.6. Ablation Study + +We conduct ablation studies and report the zero-shot classification accuracy on ScanObjectNN [41] to validate the design of DuoMamba. We also compare with two existing Mamba-based models to demonstrate the advantages of our proposed architecture. + +Component Contribution. We analyze the impact of different components in our DuoMamba block, with results summarized in Table 5a. In the baseline setting (i), applying the original Mamba [12] to FPS-based ordered sequences yields the lowest accuracy of $60.6\%$ . Replacing causal 1D convolutions with the standard ones and using either Hilbert (ii) or Trans-Hilbert (iii) ordering consistently improves the performance, with higher accuracy from Hilbert order. Combining both curves with standard 1D convolution as in DuoMamba (v) leads to the best accuracy of $63.5\%$ . Without the standard 1D convolutions as in (iv), accuracy drops $0.4\%$ to $63.1\%$ . These findings emphasize the importance of integrating geometric structures from both Hilbert curves with standard 1D convolutions for optimal information propagation. + +The Effect of Scanning Routines. We further explore the impact of scanning patterns for serializing point clouds and + +report the results in Table 5b. We compare the performance of the default FPS order with three combinations of the widely used Z-order and Hilbert curves [15]. Our results show that combining two variants of Z-order outperforms FPS, and using two Hilbert curves achieves the highest accuracy. This improvement is attributed to the fact that space-filling curves better preserve spatial relationships between point patches, enhancing information flow among nearby tokens in the sequence. Moreover, the superior locality-preserving properties of Hilbert curves over Z-order [46] contribute to a performance boost when used for processing point cloud sequences, as implemented in our DuoMamba block. + +Mamba-Based Encoder Comparison. To justify the significance of our new architecture, we compare it with two existing Mamba-based models: Mamba3D [14] and PointMamba [22]. Table 5c shows that DuoMamba surpasses both models on the ScanObjectNN [41] benchmark, outperforming Mamba3D [14] by a significant margin of $2.8\%$ in accuracy. Although DuoMamba has more parameters than PointMamba [22], it achieves better performance while also maintaining a lower FLOPs count3. Overall, the proposed architecture demonstrates a better computation-performance balance than both existing networks. + +# 7. Conclusion + +In this paper, we propose an occlusion-aware multi-modal pretraining framework for open-world 3D shape recognition. Our method uses synthetic 3D models to generate partial point clouds for pretraining, effectively reducing the training-testing domain gap and enhancing real-world recognition performance. Moreover, we introduce a Mamba-based architecture for point cloud processing, offering better performance with lower FLOPs and latency than Transformer-based networks. We hope our paper paves the way for future research on more realistic pretraining and computationally efficient models. + +Limitations. Due to resource constraints, we have not been able to leverage Objaverse [6] - the largest dataset with nearly 800K 3D objects - for pretraining, which we believe could further enrich the learned latent space and enhance recognition performance. + +Acknowledgment. The authors sincerely thank Mr. Trong-Thang Pham for his valuable feedback during the preparation of this paper. This research was supported by the Australian Research Council (ARC) under discovery grant project #240101926. Professor Ajmal Mian is the recipient of an ARC Future Fellowship Award (project #FT210100268) funded by the Australian Government. Mr. Khanh Nguyen is supported by the tuition fee scholarship from the University of Western Australia and a stipend from the ARC project #FT210100268. + +# References + +[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 1 +[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 4, 6 +[3] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2829, 2023. 5, 6 +[4] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 2, 5, 6, 7 +[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6, 7 +[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 8 +[7] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1778-1787, 2021. 1 +[8] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 4 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 6 +[10] Yipeng Gao, Zeyu Wang, Wei-Shi Zheng, Cihang Xie, and Yuyin Zhou. Sculpting holistic 3d representation in contrastive language-image-3d pre-training. In CVPR, 2024. 1, 2, 3, 5, 6, 7 +[11] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 3 +[12] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. In First Conference on Language Modeling, 2024. 2, 3, 4, 5, 6, 8 +[13] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations, 2022. 4 +[14] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In ACM Multimedia 2024, 2024. 2, 3, 5, 8 +[15] David Hilbert. Dritter Band: Analysis·Grundlagen der Mathematik·Physik Verschiedenes: Nebst Einer Lebensgeschichte. Springer-Verlag, 2013. 2, 3, 4, 5, 8 +[16] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22157-22167, 2023. 1, 3, 7 +[17] Muhammad Ibrahim, Naveed Akhtar, Saeed Anwar, and Ajmal Mian. Sat3d: Slot attention transformer for 3d point cloud semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 24(5):5456-5466, 2023. 1 +[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2 +[20] Huan Lei, Naveed Akhtar, and Ajmal Mian. Octree guided cnn with spherical kernels for 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9631-9640, 2019. 1 +[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 3, 4 + +[22] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In Advances in Neural Information Processing Systems, 2024. 2, 3, 5, 8 +[23] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 6, 7 +[24] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model, 2024. 2, 5 +[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6 +[26] Ye Mao, Junpeng Jing, and Krystian Mikołajczyk. Opendlign: Open-world point cloud understanding with depth-aligned images. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 3, 6 +[27] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 922-928. IEEE, 2015. 3 +[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 1, 7 +[29] Aaron van den Oord. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016. 6 +[30] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 3, 5 +[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 3, 5 +[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 3 +[34] Charles R. Qi, Or Litany, Kaiming He, and Leonidas J. Guibas. Deep hough voting for 3d object detection in point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1 +[35] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved + +training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 3 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5 +[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 2 +[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2 +[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, 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. 3, 5 +[40] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, pages 567-576, 2015. 6, 7 +[41] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 2, 5, 6, 7, 8 +[42] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8 +[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[44] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 1, 3 +[45] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 3 +[46] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 1, 3, 5, 8 +[47] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d + +shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 1, 2, 6 +[48] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1179-1189, 2023. 1, 2, 3 +[49] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27091-27101, 2024. 1, 2, 3 +[50] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19313-19322, 2022. 2, 3, 5, 6, 7 +[51] Yiming Zeng, Yu Hu, Shice Liu, Jing Ye, Yinhe Han, Xiaowei Li, and Ninghui Sun. Rt3d: Real-time 3-d vehicle detection in lidar point cloud for autonomous driving. IEEE Robotics and Automation Letters, 3(4):3434-3440, 2018. 1 +[52] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingqiu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15244-15253, 2023. 1, 3, 7 +[53] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 1, 3, 7 +[54] Zhihao Zhang, Shengcao Cao, and Yu-Xiong Wang. Tamm: Triadapter multi-modal learning for 3d shape understanding. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21413-21423, 2024. 1, 3, 6, 7 +[55] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 5 +[56] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 5 +[57] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3 +[58] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Confer- + +ence on Computer Vision, pages 350-368. Springer, 2022. 2 +[59] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2639-2650, 2023. 1, 3, 7 \ No newline at end of file diff --git a/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/images.zip b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f783279cad27bdb4b8c5ae4048b5cadc2a919056 --- /dev/null +++ b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9d6db099920a6a4f04f2480cb28032781f12750400b39b9eb8be397de36d45c +size 560814 diff --git a/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/layout.json b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2237474eecfb34e5a92671359f1ed7978cadee54 --- /dev/null +++ b/occlusionawaretextimagepointcloudpretrainingforopenworld3dobjectrecognition/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a90eef63dcca7e31d7ab0af9a08c1e22c34cbbb410aa0e7a0e149eb4030cf946 +size 406365 diff --git a/occmambasemanticoccupancypredictionwithstatespacemodels/6cdd5490-e841-47b9-81f6-b9aa357887a1_content_list.json b/occmambasemanticoccupancypredictionwithstatespacemodels/6cdd5490-e841-47b9-81f6-b9aa357887a1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ef77029098cd44639690b51f024e13b90dfb1e9d --- /dev/null +++ b/occmambasemanticoccupancypredictionwithstatespacemodels/6cdd5490-e841-47b9-81f6-b9aa357887a1_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ed6747c22e810c434b8fe64b03500ba0f05efda8be8dc8cb2ad6291cc2d82b1 +size 81744 diff --git a/occmambasemanticoccupancypredictionwithstatespacemodels/6cdd5490-e841-47b9-81f6-b9aa357887a1_model.json b/occmambasemanticoccupancypredictionwithstatespacemodels/6cdd5490-e841-47b9-81f6-b9aa357887a1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fa5cb6c1ad38f460a7a06f616e43d219787c6c67 --- /dev/null +++ b/occmambasemanticoccupancypredictionwithstatespacemodels/6cdd5490-e841-47b9-81f6-b9aa357887a1_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fb5a2ea3bcda77713855b94ddc34250545cc4fd9b4c6f23938af26931fc8842 +size 101624 diff --git a/occmambasemanticoccupancypredictionwithstatespacemodels/6cdd5490-e841-47b9-81f6-b9aa357887a1_origin.pdf b/occmambasemanticoccupancypredictionwithstatespacemodels/6cdd5490-e841-47b9-81f6-b9aa357887a1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f9c67f069687705b033d92fd9c95af9ba8fd1ab0 --- /dev/null +++ b/occmambasemanticoccupancypredictionwithstatespacemodels/6cdd5490-e841-47b9-81f6-b9aa357887a1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5669e773dfb8a997f7d0ccace8c2655b0ba716cb9406567730e79954af9b9673 +size 1978215 diff --git a/occmambasemanticoccupancypredictionwithstatespacemodels/full.md b/occmambasemanticoccupancypredictionwithstatespacemodels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..87631e8bfd9a40d459026a8ddd4b412d0bb026d8 --- /dev/null +++ b/occmambasemanticoccupancypredictionwithstatespacemodels/full.md @@ -0,0 +1,326 @@ +# OccMamba: Semantic Occupancy Prediction with State Space Models + +Heng Li $^{1}$ , Yuenan Hou $^{2*}$ , Xiaohan Xing $^{3}$ , Yuexin Ma $^{4}$ , Xiao Sun $^{2}$ , Yanyong Zhang $^{15*}$ $^{1}$ University of Science and Technology of China + +$^{2}$ Shanghai AI Laboratory $^{3}$ Stanford University $^{4}$ ShanghaiTech University $^{5}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + +liheng@mail.ustc.edu.cn, houyuanan@pjlab.org.cn, xhxing@stanford.edu, mayuexin@shanghaitech.edu.cn, sunxiao@pjlab.org.cn, yanyongz@ustc.edu.cn + +# Abstract + +Training deep learning models for semantic occupancy prediction is challenging due to factors such as a large number of occupancy cells, severe occlusion, limited visual cues, complicated driving scenarios, etc. Recent methods often adopt transformer-based architectures given their strong capability in learning input-conditioned weights and long-range relationships. However, transformer-based networks are notorious for their quadratic computation complexity, seriously undermining their efficacy and deployment in semantic occupancy prediction. Inspired by the global modeling and linear computation complexity of the Mamba architecture, we present the first Mamba-based network for semantic occupancy prediction, termed OccMamba. Specifically, we first design the hierarchical Mamba module and local context processor to better aggregate global and local contextual information, respectively. Besides, to relieve the inherent domain gap between the linguistic and 3D domains, we present a simple yet effective 3D-to-1D reordering scheme, i.e., height-prioritized 2D Hilbert expansion. It can maximally retain the spatial structure of 3D voxels as well as facilitate the processing of Mamba blocks. Endowed with the aforementioned designs, our OccMamba is capable of directly and efficiently processing large volumes of dense scene grids, achieving state-of-the-art performance across three prevalent occupancy prediction benchmarks, including OpenOccupancy, SemanticKITTI, and SemanticPOSS. Notably, on OpenOccupancy, our OccMamba outperforms the previous state-of-the-art Co-Occ by $5.1\%$ IoU and $4.3\%$ mIoU, respectively. Our implementation is open-sourced and available at: https://github.com/USTCLH/OccMamba. + +# 1. Introduction + +Semantic occupancy prediction is becoming an indispensable component in autonomous driving, augmented reality, robotics, etc., which estimates the occupancy and categorical labels of the surrounding environment [36]. It faces many challenges, such as the enormous number of occupancy grids, severe occlusion, limited visual clues as well as complex driving scenarios [53]. + +Recent attempts, such as MonoScene [4] and JS3C-Net [49], have made progress in addressing these challenges, but limitations remain due to their reliance on uni-modal inputs. Multi-modal approaches, such as FusionOcc [51] and M-CONet [45], offer improvements, yet they struggle to capture global information due to the inherent deficiency of CNN architectures. Despite the success of transformer-based models [20, 40, 52], they suffer from high computational complexity, particularly when processing a large number of voxel grids. + +Mamba [9], which is an important variant of state space models, emerges as a promising next-generation structure for replacing the transformer architecture. In semantic occupancy prediction, we anticipate leveraging its global modeling capabilities to better handle the complex scenarios and overcome limited visual cues, while its linear computation complexity helps manage the enormous number of occupancy grids efficiently. However, it is primarily designed for language modeling, with 1D input data, whereas the input data for semantic occupancy prediction is 3D voxels. The use of dense voxels in large scenes makes the deployment of Mamba even more challenging. When converting 3D data into a 1D format, the inherent spatial relationships between adjacent voxels in the 3D space are missing, causing neighboring voxels to become far apart in the 1D sequence. This spatial separation undermines Mamba's ability to effectively understand local and global scene contexts, hampering the prediction performance. To address this issue, developing an effective reordering strategy is crucial. + +![](images/0d3604474deecf745b2fb4dc065ab80c9891d1e3a200be1226ce69e27f2ed426.jpg) +(a) A large number of voxel grids + +![](images/caa8a73435306d85920a7258c7056b1538d04bbefe1b7dbd62f3b7d31b03a89a.jpg) +(b) GPU memory consumption +Figure 1. (a) Challenges in semantic occupancy prediction, (b) comparison in GPU memory consumption, and (c) performance comparison on OpenOccupancy validation set. Our OccMamba demonstrates high efficiency in handling a large number of voxel grids, outperforming all other semantic occupancy predictors, such as Co-Occ [31] and M-CONet [45]. + +![](images/23153a9a0f91c73ed3d57e8aad2b5ae65264d097786a4dd902333278287d22fd.jpg) +(c) Performance comparison + +cial to preserving spatial proximity in the transformation from 3D to 1D. Several attempts have been made to achieve this, for instance, Point Mamba [21] rearranges the point clouds according to the 3D Hilbert curve [12] and concatenates the features of the reordered points. While these advancements have shown potential in the point cloud classification and segmentation field, the exploration of Mamba-based architectures is still in its infancy in outdoor semantic occupancy prediction tasks. + +In this work, we present the first Mamba-based network for semantic occupancy prediction, which we refer to as OccMamba. Owing to the global modeling and linear computation complexity of Mamba, our OccMamba can efficiently process a large number of voxel grids given limited computation resources, as shown in Fig. 1(a)-(b). To effectively utilize the global and local information hidden in the input voxel grids, we design the hierarchical Mamba module and local context processor. To facilitate the processing of Mamba blocks as well as preserve the spatial structure of the 3D input data, we present a simple yet effective 3D-to-1D reordering policy, i.e., height-prioritized 2D Hilbert expansion. The designed policy sufficiently utilizes the categorical clues in the height information as well as the spatial prior in the XY plane. In this way, OccMamba effectively exploits LiDAR and camera cues, enabling effective fusion and processing of voxel features extracted from these sources without compression. This uncompressed voxel feature processing method, combined with Mamba's global modeling capabilities, enhances OccMamba with the occlusion reasoning capability, particularly in complex driving scenarios. We perform extensive experiments on three semantic occupancy prediction benchmarks, i.e., OpenOccupancy [45], SemanticKITTI [1] and SemanticPOSS [33]. Our OccMamba consistently outperforms state-of-the-art algorithms in all benchmarks. It is noteworthy that on OpenOccupancy, partly shown in Fig. 1(c), our OccMamba surpasses the previous state-of-the-art CoOcc [31] by $5.1\%$ IoU and $4.3\%$ mIoU, respectively. + +The contributions are summarized as follows: + +- To our knowledge, we design the first Mamba-based net + +work for outdoor semantic occupancy prediction. It possesses global modeling capability with linear computation complexity, which is crucial to the processing of a large number of voxel grids. We design the hierarchical Mamba module and local context processor to better aggregate global and local contextual information, respectively. + +- To facilitate the processing of Mamba blocks as well as maximally retain the original spatial structure of 3d voxels, we design a simple yet effective reordering policy that projects the point clouds into 1D sequences. +- Our OccMamba achieves the best performance on three popular semantic occupancy prediction benchmarks. + +# 2. Related work + +Semantic occupancy prediction. The objective of semantic occupancy prediction is to estimate the occupancy and semantic labels of 3D spaces using various types of input signals [13, 36]. Given that LiDAR and RGB-D camera can provide accurate spatial measurement, many studies relied on such detailed geometric data like LiDAR points [34, 35, 48, 49] and RGB-D images [5, 8, 15, 16]. Meanwhile, image-based methods, increasingly popular due to camera accessibility, such as MonoScene [4] and TPVFormer [14], estimate occupancy of the environment using RGB images. However, inaccurate depth estimation in RGB images results in lower performance than LiDAR-based models. In response, multi-modal fusion combining both modalities has attracted significant attention. CONet [45] and CoOcc [31] demonstrate the advantages of combining modalities, enhancing precision and reliability. Furthermore, recent methods have shifted from CNNs to integrating transformers, as seen in OccFormer's dual-path transformer [52] and OccNet's cascade refinement [40] reduce the transformer's computational demands. + +Multi-modal fusion. Multi-modal fusion aims to make the strength of different input modalities, thus making more accurate and robust perception [7, 15, 17, 18, 25, 27, 43]. For instance, AICNet [15] integrates RGB and depth data using anisotropic convolutional networks to enhance the accuracy + +and completeness of semantic occupancy prediction. PointPainting [42] performs semantic segmentation on RGB images and attaches semantic probabilities to LiDAR points, thereby enriching the point cloud with rich semantic information. For semantic occupancy prediction, images and point clouds are two prevalent input signals. Recent trends in the occupancy prediction field favour the multi-modal fusion as it utilizes the strength of both signals. + +State space models. The transformer [41] has reshaped the computer vision field but suffers from the quadratic computation complexity [22]. To relieve this, more efficient operators like linear attention [44], flash attention [6] have been proposed. State Space Models (SSMs) such as Mamba [9], S4 [10], and S4nd [30] are gaining prominence, with Mamba being notable for integrating selective mechanisms, which can effectively capture long-range dependencies and process large-scale data in linear time. This innovation has extended into computer vision domain through variants like VMamba [26], which includes a cross-scan module, and Vision Mamba [55], which utilizes a bidirectional SSM. In point cloud processing, PointMamba [21] improves the global modeling of point clouds by rearranging input patches based on 3D Hilbert curve, while Point Mamba [24] employs an octree-based ordering for efficient spatial relationship capture. They demonstrate Mamba's proficiency in processing large scale 3D data. Building on these advancements, we attempt to design a Mamba-based network tailored for semantic occupancy prediction. + +# 3. Methodology + +To efficiently and effectively process a large number of voxel grids in semantic occupancy prediction, we propose to take Mamba [9] as the basic building block that enjoys the benefit of both global modeling and linear computation complexity. To facilitate the processing of Mamba blocks, we design a simple yet effective reordering scheme, dubbed height-prioritized 2D Hilbert expansion, along with OccMamba Encoder. In the following sections, we first provide a brief review of state space models and Mamba in Sec. 3.1. Then, we present the framework overview of OccMamba in Sec. 3.2. Thereafter, we provide a detailed explanation of the reordering scheme, the hierarchical Mamba module and local context processor in Sec. 3.3. Eventually, the training objective is presented in Sec. 3.4. + +# 3.1. Preliminaries + +State space models. Before introducing the Mamba module, we first have a brief review of the state space models (SSMs). SSMs are inspired by the control theory and maps the system input to the system output through the hidden state. This approach allows for effectively handling sequences of information. When the input is discrete, the mathematical representation of SSMs is given as follows: + +$$ +h _ {k} = \bar {A} h _ {k - 1} + \bar {B} x _ {k}, \tag {1} +$$ + +$$ +y _ {k} = \bar {C} h _ {k}, \tag {2} +$$ + +where $k$ is the sequence number, $\overline{A}$ , $\overline{B}$ , $\overline{C}$ are matrices that represent the discretized parameters of the model, which involve the sampling step $\Delta$ . The $x_{k}$ , $y_{k}$ , and $h_{k}$ denote the input, output, and hidden state of the system, respectively. As an improvement, Structured State Space Sequence Models (S4) [10] optimize traditional SSMs by introducing structured matrices, which allow the system dynamics, involving matrices $\overline{A}$ , $\overline{B}$ , and $\overline{C}$ , to be parameterized in a way that significantly improves computational efficiency and scalability for large sequences. By leveraging these structured matrices, S4 achieves reduced computational complexity without sacrificing accuracy. + +Mamba Module. Mamba introduces an adaptation to the S4 models [10], where it makes the matrices $\overline{B}$ and $\overline{C}$ , as well as the sampling size $\Delta$ , dependent on the input. This dependency arises from incorporating the sequence length and batch size of the input, enabling dynamic adjustments of these matrices for each input token. By this way, Mamba allows $\overline{B}$ and $\overline{C}$ to dynamically influence the state transition conditioned on the input, enhancing the model's content-awareness. Additionally, Mamba incorporates optimizations for the scan operation and a hardware-aware algorithm, enabling efficient parallel computation. + +# 3.2. Framework overview + +Figure 2 depicts the pipeline of our OccMamba, which is built upon M-CONet [45]. + +Multi-modal visual encoders. Taking both point cloud and multi-view images as input, OccMamba processes each modality with respective visual encoder. Specifically, for the LiDAR branch, we first voxelize the input point clouds $\mathbf{P}$ [54] and then employ sparse-convolution-based LiDAR encoder $\mathbf{E}_{\mathcal{L}}$ [50] to generate LiDAR voxel features $\mathbf{V}_{\mathcal{L}} \in \mathbb{R}^{B \times W_{\mathcal{L}} \times H_{\mathcal{L}} \times D_{\mathcal{L}} \times C_{\mathcal{L}}}$ . For the image branch, we feed the multi-view images $\mathbf{I}$ to ResNet-based image encoder [11] which utilizes FPN [23] to aggregate multi-scale features, and then utilize the 2D-to-3D view transformer [27] to produce image voxel features $\mathbf{V}_{\mathcal{C}} \in \mathbb{R}^{B \times W_{\mathcal{C}} \times H_{\mathcal{C}} \times D_{\mathcal{C}} \times C_{\mathcal{C}}}$ . Here, $B$ is the batch size, $W_{\mathcal{L}}, W_{\mathcal{C}}, H_{\mathcal{L}}, H_{\mathcal{C}}, D_{\mathcal{L}}$ and $D_{\mathcal{C}}$ are the spatial dimensions of the voxel features. $C_{\mathcal{L}}$ and $C_{\mathcal{C}}$ are the channel dimension of these voxel features. We ensure that $\mathbf{V}_{\mathcal{L}}$ and $\mathbf{V}_{\mathcal{C}}$ are identical in the spatial dimensions. Thereafter, these voxel features are concatenated along the channel dimensions: + +$$ +\mathbf {V} _ {\mathcal {F}} = \operatorname {c o n c a t} \left(\mathbf {V} _ {\mathcal {L}}, \mathbf {V} _ {\mathcal {C}}\right). \tag {3} +$$ + +OccMamba encoder. To process features extracted from the multi-modal encoders, we designed the hierarchical + +![](images/5aaab49c69a045949edcead5b2b560dce85ccd407dc4d6e828e5860b9d5b7f75.jpg) +Figure 2. Schematic overview of our OccMamba. Given surround-view images and LiDAR point clouds, we first employ the 2D encoder and 3D encoder to process them, obtaining camera features and LiDAR voxel features, respectively. View transformer is utilized to project camera features to camera voxel features. The camera and LiDAR voxel features are then fused and sent to the hierarchical Mamba module, where the proposed height-prioritized 2D Hilbert expansion reordering is used to maximally utilize the spatial clues of voxels. Besides, local context processor is designed to divide the Mamba features into multiple windows along the XY plane, further enhancing the local semantic information. Eventually, the Mamba features are fed to the occupancy head, producing semantic occupancy predictions. + +Mamba module and local context processor. The former includes the Mamba encoder $\mathbf{E}_{\mathcal{M}}$ and Mamba decoder $\mathbf{D}_{\mathcal{M}}$ while the latter is denoted as $\mathbf{P}_{\mathcal{L}}$ . In the process, fusion voxel features $\mathbf{V}_{\mathcal{F}}$ are fed into them, which will be detailed in Sec. 3.3. The transformation can be described by + +$$ +\mathbf {V} _ {\mathcal {M}} = \mathbf {D} _ {\mathcal {M}} \left(\mathbf {E} _ {\mathcal {M}} \left(\mathbf {V} _ {\mathcal {F}}\right)\right), \tag {4} +$$ + +$$ +\mathbf {V} _ {\mathcal {P}} = \mathbf {P} _ {\mathcal {L}} (\mathbf {V} _ {\mathcal {M}}), \tag {5} +$$ + +where $\mathbf{V}_{\mathcal{M}}$ and $\mathbf{V}_{\mathcal{P}}$ retains the same size as $\mathbf{V}_{\mathcal{F}}$ . + +Occupancy head. We feed the processed features of OccMamba Encoder, i.e., $\mathbf{V}_{\mathcal{P}}$ , to the coarse-to-fine module $\mathbf{F}_{c\rightarrow f}$ [45], which interpolates the dimensions to match the size of the ground truth, and then employ a MLP to predict the category of each voxel grid. The process is presented in the following equation: + +$$ +\mathbf {O} _ {\mathrm {o c c}} = \operatorname {M L P} \left(\mathbf {F} _ {c \rightarrow f} (\mathbf {V} _ {\mathcal {P}})\right). \tag {6} +$$ + +# 3.3. OccMamba encoder with height-prioritized reordering + +Semantic occupancy prediction is challenging due to the high dimensionality and density of voxel grids, often involving millions of voxels. Previous methods [14, 28] resort to projection techniques but suffer from information loss. In contrast, Mamba's linear computational complexity enables direct processing of large voxel features, avoiding the limitations of deformable attention and convolution, such as the need for key point selection or limited receptive fields. + +Height-prioritized reordering scheme. Before feeding 3D voxel features into the Mamba block, it's necessary to + +reorder them into 1D sequences. A poor reordering strategy can disrupt the intrinsic spatial relationships between adjacent voxels in the 3D space, especially when dealing with a large number of voxel grids, thereby impacting Mamba's performance. Inspired by the Hilbert curve [12], we propose a height-prioritized 2D Hilbert expansion, tailored to the flat spatial structure typical of semantic occupancy prediction tasks, where height information provides valuable categorical clues, as it often correlates with object categories, reveals terrain features, and distinguishes different spatial regions in the scene. Specifically, we divide the "xyz" coordinates into the "xy" plane and a "z" dimension. Starting on the xy-plane at $z = 0$ , the process extends vertically along the $z$ -axis, creating vertical sequences. These sequences are then ordered and interconnected following the 2D Hilbert curve on the xy plane, forming the height-prioritized 2D Hilbert expansion, as illustrated in Fig. 3(d). For comparison, Fig. 3(a)-(c) illustrate the results of reordering according to the 'xyz' sequence, the 'zxy' sequence, and the 3D Hilbert curve, respectively. Our reordering strategy prioritizes $z$ -axis spatial information while maintaining strong spatial proximity in the xy plane. By this way, we can better leverage Mamba's contextual modeling. + +As a result, with the spatial size $W$ , $H$ , $D$ and 1D data length $L$ , the input $\mathbf{V} \in \mathbb{R}^{B \times W \times H \times D \times C}$ and the output $\mathbf{V}' \in \mathbb{R}^{B \times L \times C}$ in the above reordering scheme $\mathcal{R}_{3D \to 1D}$ and $\mathcal{R}_{1D \to 3D}$ can be expressed as: + +$$ +\mathbf {V} = \mathcal {R} _ {1 D \rightarrow 3 D} \left(\mathbf {V} ^ {\prime}\right) = \mathcal {R} _ {1 D \rightarrow 3 D} \left(\mathcal {R} _ {3 D \rightarrow 1 D} (\mathbf {V})\right) \tag {7} +$$ + +Hierarchical mamba module. To effectively utilize the + +![](images/ef874c360238229dd946a4aca87baadb79509016b1a172545a57e0ce15bf8718.jpg) +(a) XYZ sequence + +![](images/1ea5142293f02da09fd4bd2feff4bf81a394a4be31fe9714f0c0f941bd5e0c68.jpg) +(b) ZXY sequence + +![](images/161019b82c9f482a191c809b74975af62c3ab9b699f86e0c53d05d487c04f367.jpg) +(c) 3D Hilbert curve +Figure 3. Comparison between different reordering schemes. (a) XYZ sequence, (b) ZXY sequence, (c) 3D Hilbert curve, and (d) our height-prioritized 2D Hilbert expansion. Small cubes of varying colors, ranging from red to purple, represent different voxels. The corresponding colored edges indicate the adjacency of voxels that become neighbors after reordering. In each sub-figure, its top half represents the isometric view, and the bottom half represents the top view. Our method maximizes z-axis proximity while also striving to preserve xy-axis proximity. + +![](images/c5da11fc0cde4527e57dce0d0f75c7fda053ca90ab2bb2d08e4e03d6cd22aca8.jpg) +(d) Height-prioritized 2D Hilbert expansion + +contextual information across spatial resolutions, we introduce the hierarchical Mamba module, which includes an encoder and decoder. The encoder comprises multiple groups, each with two Mamba blocks and a downsampling operation between groups for multi-scale voxel features. Correspondingly, the Mamba decoder mirrors this structure, utilizing paired Mamba blocks and upsampling operations. The upsampling incorporates skip connection from the corresponding layer in the Mamba encoder, ensuring consistency in feature dimensions across scales. Crucially, the reordering scheme is applied both before and after each Mamba blocks group, maintaining voxel representation during both downsampling and upsampling phases. + +As to a single Mamba block, it consists of Layer-Norm(LN), linear layers, 1D convolution, Silu activation, Selective SSM, and residual connections. Given an input $\mathbf{V} \in \mathbb{R}^{B \times L \times C}$ , where $B$ , $L$ and $C$ denote the batch size, the length of 1D data, and the feature dimension, respectively, the output $\mathbf{V}' \in \mathbb{R}^{B \times L \times C}$ can be computed as below: + +$$ +\mathbf {V} _ {\mathbf {1}} ^ {\prime} = \operatorname {L N} (\mathbf {V}), \tag {8} +$$ + +$$ +\mathbf {V} _ {2} ^ {\prime} = \text {S e l e c t i v e S S M (S i l u (C o n v 1 d (L i n e a r (\mathbf {V} _ {1} ^ {\prime}))))}, \tag {9} +$$ + +$$ +\mathbf {V} ^ {\prime} = \operatorname {L i n e a r} \left(\mathbf {V} _ {2} ^ {\prime} \odot \operatorname {S i l u} \left(\operatorname {L i n e a r} \left(\mathbf {V} _ {1} ^ {\prime}\right)\right)\right), \tag {10} +$$ + +where $\odot$ represents the dot product. For convenience, we denote Eq.(8)-Eq.10) as $\mathcal{M}$ . Then, Given the input $\mathbf{V}\in \mathbb{R}^{B\times W\times H\times D\times C}$ and output $\mathbf{V}'\in \mathbb{R}^{B\times W\times H\times D\times C}$ the computation for a single Mamba block with our reordering scheme $\mathcal{R}_{3D\to 1D}$ and $\mathcal{R}_{1D\to 3D}$ can be represented as + +$$ +\mathbf {V} ^ {\prime} = \mathcal {R} _ {1 D \rightarrow 3 D} \left(\mathcal {M} \left(\mathcal {R} _ {3 D \rightarrow 1 D} (\mathbf {V})\right)\right) \tag {11} +$$ + +Local context processor. Although the hierarchical Mamba module has improved the network's ability to process multi-scale global information, handling more local in + +formation can further enhance the accuracy of semantic occupancy prediction. To this end, we design a lightweight local context processor. It divides the Mamba features $\mathbf{V}_{\mathcal{M}} \in \mathcal{R}^{B \times W \times H \times D \times C}$ along the XY plane into patches $\mathcal{V} = \{\mathcal{V}_{s_i, w_i} | s_i \in S, w_i \in W\}$ , using a list of window sizes $W = \{w_i\}$ and a corresponding list of sliding strides $S = \{s_i\}$ to scan the plane. For each item $\mathbf{V}_{p,q} \in \mathcal{V}_{s_i, w_i}$ , we have the following equations: + +$$ +\mathbf {V} _ {p, q} = \mathbf {V} _ {\mathcal {M}} [,: p \cdot s _ {i}: p \cdot s _ {i} + w _ {i}, q \cdot s _ {i}: q \cdot s _ {i} + w _ {i}, :, : ], +$$ + +$$ +p = 0, 1, \dots \left\lfloor \frac {W - w _ {i}}{s _ {i}} \right\rfloor , q = 0, 1, \dots \left\lfloor \frac {H - w _ {i}}{s _ {i}} \right\rfloor . \tag {12} +$$ + +Next, we reorder $\mathbf{V}_{p,q} \in \mathcal{R}^{B \times w_i \times w_i \times D \times C}$ to $\mathcal{R}^{B \times L \times C}$ , and then stack $\nu_{s_i,w_i}$ to $\mathcal{R}^{(B \times N) \times L \times C}$ . Each $\nu_{s_i,w_i}$ is processed by the dedicated two-layer Mamba blocks, and the outputs are reshaped to $\nu_{s_i,w_i}' = \{\mathbf{V}_{p,q}' \in \mathcal{R}^{B \times w_i \times w_i \times D \times C}\}$ . Finally, $\nu' = \{\nu_{s_i,w_i}'\}$ patches are reassembled into the original spatial size and concatenated along the channel dimension to produce $\mathbf{V}'_{\mathcal{L}} \in \mathcal{R}^{B \times W \times H \times D \times C'}$ . A 3D convolution with a kernel size of $1 \times 1 \times 1$ is then applied along the channel dimension to reduce it, resulting in $\mathbf{V}_{\mathcal{L}} \in \mathcal{R}^{B \times W \times H \times D \times C}$ . + +$$ +\mathcal {V} _ {s i, w _ {i}} ^ {\prime} = \operatorname {R e s h a p e} (\mathcal {M} (\operatorname {S t a c k} (\operatorname {R e o r d e r} (\mathcal {V} _ {s i, w _ {i}})))) \tag {13} +$$ + +$$ +V _ {\mathcal {L}} = \operatorname {C o n v 3 d} \left(\text {R e a s s e m b l e} \_ \text {P a t c h e s} \left(\mathcal {V} ^ {\prime}\right)\right) \tag {14} +$$ + +Endowed with the hierarchical Mamba module and local context processor, our OccMamba utilizes both global and local information from dense scene grids, thereby capturing broad context while preserving local details. This integrated approach enhances spatial understanding and mitigates potential performance degradation with large data volumes, by effectively balancing global awareness with local precision. + +
MethodInput ModalityIoUmIoUbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive surf.other flatsidewalkterrainmanmadevegetation
MonoScene [4]C18.46.97.13.99.37.25.63.05.94.44.94.214.96.37.97.410.07.6
TPVFormer [14]C15.37.89.34.111.310.15.24.35.95.36.86.513.69.08.38.09.28.2
SparseOcc [39]C21.814.116.19.315.118.67.39.411.29.47.213.031.821.720.718.86.110.6
3DSketch [5]C&D25.610.712.05.110.712.46.54.05.06.38.07.221.814.813.011.812.021.2
AICNet [15]C&D23.810.611.54.011.812.35.13.86.26.08.27.524.113.012.811.511.620.2
LMSCNet [35]L27.311.512.44.212.812.16.24.76.26.38.87.224.212.316.614.113.922.2
JS3C-Net [49]L30.212.514.23.413.612.07.24.37.36.89.29.127.915.314.916.214.024.9
M-CONet [45]C&L29.520.123.313.321.224.315.315.918.013.315.320.733.221.022.521.519.623.2
Co-Occ [31]C&L30.621.926.516.822.327.010.120.920.714.516.421.636.923.55.523.720.523.5
OccMamba-128 (ours)C&L34.725.229.119.125.528.518.124.723.419.819.324.537.025.425.425.428.129.9
OccMamba-384 (ours)C&L35.726.230.220.526.529.518.826.023.719.920.625.438.426.527.026.628.930.5
+ +Table 1. Quantitative comparisons on OpenOccupancy validation set with v0.0 annotations. C, D, L denote camera, depth and LiDAR, respectively. OccMamba-384 means OccMamba with mamba feature dimension being 384. The best and second-best are in bold and underlined, respectively. + +
MethodInput ModalitymIoU
MonoScene [4]C11.1
SurroundOcc [46]C11.9
OccFormer [52]C12.3
RenderOcc [32]C12.8
LMSCNet [35]L17.0
JS3C-Net [49]L23.8
SSC-RS [29]L24.2
Co-Occ [31]C&L24.4
M-CONet [45]C&L20.4
OccMamba-128 (ours)C&L24.6
+ +Table 2. Performance on SemanticKITTI test set. The best and second-best are in bold and underlined, respectively. + +
MethodInput ModalitymIoU
SSCNet [38]L15.2
LMSCNet [35]L16.5
MotionSC [47]L17.6
JS3C-Net [49]L22.7
OccMamba-128 (ours)L23.4
+ +Table 3. Comparisons on SemanticPOSS validation set. The best is in bold, and the second-best is underlined. + +# 3.4. Training objective + +Our ultimate training objective is composed of five terms, including the cross-entropy loss $\mathcal{L}_{\mathrm{CE}}$ , the lovasz-softmax loss $\mathcal{L}_{\mathrm{iou}}$ [2], the geometric and semantic affinity loss $\mathcal{L}_{\mathrm{geo}}$ and $\mathcal{L}_{\mathrm{sem}}$ [4], and the depth supervision loss $\mathcal{L}_{\mathrm{depth}}$ [19]. These terms are crucial for optimizing our model's performance in various aspects: $\mathcal{L}_{\mathrm{CE}}$ ensures accurate classifica + +tion, $\mathcal{L}_{\mathrm{iou}}$ enhances semantic segmentation, $\mathcal{L}_{\mathrm{geo}}$ improves spatial alignment, $\mathcal{L}_{\mathrm{sem}}$ refines semantic understanding, and $\mathcal{L}_{\mathrm{depth}}$ guides depth estimation and spatial relationships. The ultimate training objective is presented as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {C E}} + \lambda_ {1} \mathcal {L} _ {\mathrm {i o u}} + \lambda_ {2} \mathcal {L} _ {\mathrm {g e o}} + \lambda_ {3} \mathcal {L} _ {\mathrm {s e m}} + \lambda_ {4} \mathcal {L} _ {\mathrm {d e p t h}}, \tag {15} +$$ + +where $\lambda_1 \sim \lambda_4$ are the loss coefficients to balance the effect of each loss item on the final performance. They are all empirically set to 1. + +# 4. Experiments + +# 4.1. Experimental setup + +Benchmarks. We conduct experiments on OpenOccupancy [45], SemanticKITTI [1] and SemanticPOSS [33]. OpenOccupancy is built upon the nuScenes dataset [3], inheriting the data format of nuScenes. It comprises 700 training sequences and 150 validation sequences, with annotations for 17 classes. The occupancy annotations are represented in a $512 \times 512 \times 40$ voxel grid, with each voxel sized at 0.2 meters. Notably, each frame in OpenOccupancy has a data range four times larger than the other datasets, imposing a heavier computational burden. In SemanticKITTI, sequences 00-10 (excluding 08), 08, and 11-21 are allocated for training, validation, and testing, respectively. For the occupancy annotations, a $256 \times 256 \times 32$ grid is used, with voxels measuring 0.2 meters each. After pre-processing, a total of 19 classes are utilized for training and evaluation. SemanticPOSS, which closely resembles SemanticKITTI, employs sequences 00-05 and 02 as training and validation sets, respectively, with annotations for 11 classes. + +Evaluation metrics. We adopt the official evaluation metrics, i.e., Intersection-over-Union (IoU) and mean Intersection-over-Union (mIoU). + +![](images/89f1d2ef7869aac7a1e6f27897de0e6770c782e827c402f11712a4660625731c.jpg) +Figure 4. Visual comparison between OccMamba and M-CONet. From top to bottom: ground-truth, predictions of M-CONet and OccMamba. Our OccMamba makes preciser predictions than M-CONet especially in regions highlighted by red ellipses. + +Implementation details. We use ResNet-50 [11] as the image backbone. Both the Mamba Encoder and Mamba Decoder in the hierarchical mamba module consist of four groups and each group contains two Mamba blocks. The local context processor utilizes window sizes of [3, 5, 7], with two Mamba blocks implemented. For the OpenOccupancy dataset, we maintain identical setting to M-CONet [45]: for each frame, we utilize six surround view camera images as input, coupled with a fusion of ten frames of LiDAR points spanning the range of $[-51.2\mathrm{m}, 51.2\mathrm{m}]$ along the X and Y axes, and $[-2.0\mathrm{m}, 6.0\mathrm{m}]$ along the Z axis. Within the Mamba block, experiments are conducted with mamba feature dimension set to 128 and 384, respectively. On the SemanticKITTI dataset, our inputs consist of forward-facing stereo camera images alongside a single-frame LiDAR points, with their spatial extents defined as $[0\mathrm{m}, -25.6\mathrm{m}, -2\mathrm{m}, 51.2\mathrm{m}, 25.6\mathrm{m}, 4.4\mathrm{m}]$ . We set the mamba feature dimension to 128 and employ Test Time Augmentation (TTA) to boost the performance. The used augmentations include the flipping of the X and Y axes, and four augmented inputs (including the original input) are sent to the model. Finally, for the SemanticPOSS, we exclusively leverage single-frame LiDAR points like SemanticKITTI. + +# 4.2. Experimental results + +OpenOccupancy. As shown in Table 1, we compare OccMamba on OpenOccupancy with v0.0 annotation to previous methods which use the settings of camera-only [4, 5, 14, 15], LiDAR-only [35, 49], and multi-modal fusion [31, 45]. Our OccMamba achieves the best performance. Particularly, OccMamba-384 outperforms Co-Occ [31] by $5.1\%$ in IoU and $4.3\%$ in mIoU, respectively. Additionally, OccMamba-128 also surpasses Co-Occ by $4.1\%$ in IoU and $3.3\%$ in mIoU. These results underscore the efficacy of our approach in handling large-scale scenarios. + +SemanticKITTI & SemanticPOSS. As evident from Table 2, our OccMamba-128 still outperforms the second-best method, i.e., Co-Occ [31], by $0.2\%$ mIoU in SemanticKITTI test set. Note that we do not compare our OccMamba with SCPNet [48] and Occfiner [37] as they employ tricks such as label rectification, knowledge distillation or multi-frame concatenation. Strong performance is also observed in SemanticPOSS, as shown in Table 3. These results sufficiently demonstrate the superiority of our approach. The detailed classwise performance is put in the supplementary material. + +Visual comparisons. From Fig. 4, our OccMamba-384 produces more accurate predictions than M-CONet [45] on OpenOccupancy validation set. Our OccMamba not only + +
Reordering SchemesmIoU
XYZ sequence24.5
ZXY sequence24.7
3D Hilbert24.8
Height-prioritized 2D Hilbert expansion25.2
+ +Table 4. Ablation study on different reordering schemes. + +
MethodCNNHierarchical Mamba ModuleLocal Context ProcessormIoU
M-CONet [45]✓20.1
OccMamba-128✓25.0
OccMamba-128✓✓25.2
OccMamba-384✓✓26.2
+ +provides a more detailed representation of object shapes and semantics, but also performs preciser predictions on occluded objects and surfaces. + +# 4.3. Ablation study + +The reported results are on the OpenOccupancy validation set with v0.0 annotations unless otherwise specified. + +Reordering schemes. To accelerate the training process, we set the mamba feature dimension to 128 and then compare our method with three alternative methods: ordering by XYZ sequence, ordering by ZXY sequence, and ordering using the 3D Hilbert curve. As shown in Table 4, our method achieves the highest mIoU of $25.2\%$ . Additionally, it is apparent that reordering schemes that prioritize the z-axis outperform the other methods. This could be attributed to the fact that in semantic occupancy prediction, the overall shape of the scenes typically manifests as a flattened cuboid. Therefore, reordering schemes that prioritize the z-axis allow Mamba to better capture spatial proximity information, leading to improved prediction performance. + +OccMamba encoder. Consistent with previous settings, we conduct ablation studies on mamba feature dimension and two modules within our OccMamba encoder. As shown in Table 5, by replacing the CNN-based occupancy encoder from M-CONet [45] with our hierarchical Mamba module using a mamba feature dimension of 128, the mIoU improves from $20.1\%$ to $25.0\%$ . The incorporation of our local context processor further enhances the performance, achieving $25.2\%$ mIoU. When the Mamba feature dimension is increased from 128 to 384, the mIoU improves further, reaching $26.2\%$ . + +Generalization. We apply our hierarchical Mamba module to the MonoScene[4] framework. Specifically, we replace the 3D voxel decoder in this framework with our implementation, using a mamba feature dimension of 128, and train on $25\%$ of the SemanticKITTI training set with the training + +Table 5. Ablation study on OccMamba Encoder. + +
MethodCNNHierarchical Mamba ModuleLocal Context ProcessormIoU
MonoScene [4]✓11.1
MonoScene [4]✓11.7
MonoScene [4]✓✓11.9
+ +Table 6. Ablation study on MonoScene framework. + +
MethodMemory Usage(GiB)Inference Time(ms)mIoU
TrainingInference
M-CONet [45]37.312.1416.020.1
OccMamba-38437.79.2448.326.2
OccMamba-12823.16.8269.325.2
+ +Table 7. Comparison of memory usage and inference time. + +strategy unchanged. As shown in Table 6, our OccMamba encoder can improve the performance of MonoScene from $11.1\%$ mIoU to $11.9\%$ mIoU on SemanticKITTI val set. + +Memory usage and inference time. In Table 7, we present a comparative analysis of memory usage and inference time between OccMamba and M-CONet during both training and inference phases. The networks are trained on 8 A40 GPUs and perform inference on a single RTX 4090 GPU. When the mamba feature dimension is set to 384, OccMamba shows similar memory usage to M-CONet during training but reduces inference memory usage by about $24\%$ . When the mamba feature dimension is lowered to 128, the memory savings are more pronounced, with a reduction of approximately $38\%$ during training and $44\%$ during inference compared to M-CONet. Additionally, OccMamba-128 significantly improves inference speed, taking about $35\%$ of the time required by M-CONet, which highlights its efficiency in resource-constrained scenarios. In terms of performance, OccMamba-384 achieves the highest mIoU, while OccMamba-128 configurations also surpass M-CONet in performance, providing a balance between efficiency and accuracy. + +# 5. Conclusion + +In this paper, we present the first Mamba-based network, termed OccMamba, for semantic occupancy prediction. To facilitate the processing of Mamba blocks and maximally retain the 3D spatial relationship, we design a novel reordering scheme and OccMamba encoder. Endowed with these designs, our OccMamba is capable of directly and efficiently processing large volumes of dense scene grids, surpassing the previous state-of-the-art algorithms on three prevalent occupancy prediction benchmarks. As an innovative approach in this field, our OccMamba provides an effective means for handling large-scale voxels directly, and we believe it will inspire new advancements. + +# References + +[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9297-9307, 2019. 2, 6 +[2] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, pages 4413-4421, 2018. 6 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6 +[4] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2, 6, 7, 8 +[5] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4193-4202, 2020. 2, 6, 7 +[6] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 3 +[7] Jing Gao, Peng Li, Zhikui Chen, and Jianing Zhang. A survey on deep learning for multimodal data fusion. Neural Computation, 32(5):829-864, 2020. 2 +[8] Martin Garbade, Yueh-Tung Chen, Johann Sawatzky, and Juergen Gall. Two stream 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 2 +[9] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 3 +[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 7 +[12] David Hilbert. Dritter Band: Analysis·Grundlagen der Mathematik·Physik Verschiedenes: Nebst Einer Lebensgeschichte. Springer-Verlag, 2013. 2, 4 +[13] Yuenan Hou, Xiaoshui Huang, Shixiang Tang, Tong He, and Wanli Ouyang. Advances in 3d pre-training and downstream tasks: a survey. *Vicinagearth*, 1(1):6, 2024. 2 +[14] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 2, 4, 6, 7 +[15] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3351-3359, 2020. 2, 6, 7 +[16] Siqi Li, Changqing Zou, Yipeng Li, Xibin Zhao, and Yue Gao. Attention-based multi-modal fusion network for semantic scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11402-11409, 2020. 2 +[17] Xin Li, Botian Shi, Yuenan Hou, Xingjiao Wu, Tianlong Ma, Yikang Li, and Liang He. Homogeneous multi-modal feature fusion and interaction for 3d object detection. In European Conference on Computer Vision, pages 691-707. Springer, 2022. 2 +[18] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023. 2 +[19] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In AAAI, pages 1477-1485, 2023. 6 +[20] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 1 +[21] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 2, 3 +[22] Tianyang Lin, Yuxin Wang, Xiangyang Liu, and Xipeng Qiu. A survey of transformers. AI open, 3:111-132, 2022. 3 +[23] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3 +[24] Jiuming Liu, Ruiji Yu, Yian Wang, Yu Zheng, Tianchen Deng, Weicai Ye, and Hesheng Wang. Point mamba: A novel point cloud backbone based on state space model with octree-based ordering strategy. arXiv preprint arXiv:2403.06467, 2024. 3 +[25] Youquan Liu, Runnan Chen, Xin Li, Lingdong Kong, Yuchen Yang, Zhaoyang Xia, Yeqi Bai, Xinge Zhu, Yuexin Ma, Yikang Li, et al. Uniseg: A unified multi-modal lidar segmentation network and the openpcseg codebase. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21662-21673, 2023. 2 +[26] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: + +Visual state space model. arXiv preprint arXiv:2401.10166, 2024.3 +[27] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2, 3 +[28] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19936-19945, 2024. 4 +[29] Jianbiao Mei, Yu Yang, Mengmeng Wang, Tianxin Huang, Xuemeng Yang, and Yong Liu. Ssc-rs: Elevate lidar semantic scene completion with representation separation and bev fusion. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2023. 6, 2 +[30] Eric Nguyen, Karan Goel, Albert Gu, Gordon Downs, Preey Shah, Tri Dao, Stephen Baccus, and Christopher Ré. S4nd: Modeling images and videos as multidimensional signals with state spaces. Advances in neural information processing systems, 35:2846-2861, 2022. 3 +[31] Jingyi Pan, Zipeng Wang, and Lin Wang. Co-occ: Coupling explicit feature fusion with volume rendering regularization for multi-modal 3d semantic occupancy prediction. IEEE Robotics and Automation Letters, 2024. 2, 6, 7 +[32] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoqi Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 6, 2 +[33] Yancheng Pan, Biao Gao, Jilin Mei, Sibo Geng, Chengkun Li, and Huijing Zhao. Semanticposs: A point cloud dataset with large quantity of dynamic instances. In 2020 IEEE Intelligent Vehicles Symposium (IV), pages 687-693. IEEE, 2020. 2, 6 +[34] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. IEEE transactions on pattern analysis and machine intelligence, 44(10):7205-7218, 2021. 2 +[35] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020. 2, 6, 7 +[36] Luis Roldao, Raoul De Charette, and Anne Verroust-Blondet. 3d semantic scene completion: A survey. International Journal of Computer Vision, 130(8):1978-2005, 2022. 1, 2 +[37] Hao Shi, Song Wang, Jiaming Zhang, Xiaoting Yin, Zhongdao Wang, Zhijian Zhao, Guangming Wang, Jianke Zhu, Kailun Yang, and Kaiwei Wang. Occfiner: Offboard occupancy refinement with hybrid propagation. arXiv preprint arXiv:2403.08504, 2024. 7 +[38] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the + +IEEE conference on computer vision and pattern recognition, pages 1746-1754, 2017. 6 +[39] Pin Tang, Zhongdao Wang, Guoqing Wang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15035-15044, 2024. 6, 1 +[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 1, 2 +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[42] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 3 +[43] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6792-6802, 2023. 2 +[44] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 3 +[45] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17850-17859, 2023. 1, 2, 3, 4, 6, 7, 8 +[46] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 6, 2 +[47] Joey Wilson, Jingyu Song, Yuewei Fu, Arthur Zhang, Andrew Capodieci, Paramsothy Jayakumar, Kira Barton, and Maani Ghaffari. *Motionsc: Data set and network for real-time semantic mapping in dynamic environments*. IEEE Robotics and Automation Letters, 7(3):8439-8446, 2022. 6 +[48] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17642-17651, 2023. 2, 7 +[49] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 1, 2, 6, 7 +[50] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 3 + +[51] Shuo Zhang, Yupeng Zhai, Jilin Mei, and Yu Hu. Fusionocc: Multi-modal fusion for 3d occupancy prediction. In ACM Multimedia 2024, 2024. 1 +[52] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 1, 2, 6 +[53] Yanan Zhang, Jinqing Zhang, Zengran Wang, Junhao Xu, and Di Huang. Vision-based 3d occupancy prediction in autonomous driving: a review and outlook. arXiv preprint arXiv:2405.02595, 2024. 1 +[54] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 3 +[55] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 3 \ No newline at end of file diff --git a/occmambasemanticoccupancypredictionwithstatespacemodels/images.zip b/occmambasemanticoccupancypredictionwithstatespacemodels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cf1d62ac1e3da9a49ed7c9151593dd467ee5d11c --- /dev/null +++ b/occmambasemanticoccupancypredictionwithstatespacemodels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54f3aac4a35f2a97543d4be0176efb6466e6e6fade137d9163026377d2fea609 +size 721406 diff --git a/occmambasemanticoccupancypredictionwithstatespacemodels/layout.json b/occmambasemanticoccupancypredictionwithstatespacemodels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..18c6beef88d2acc06456c4f246ce5c1e7122cb46 --- /dev/null +++ b/occmambasemanticoccupancypredictionwithstatespacemodels/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:048a223dcdc933162d069b50956192b4f6f7ddf23780907296e7448f9da79dc1 +size 419537 diff --git a/octopusalleviatinghallucinationviadynamiccontrastivedecoding/ee076920-0b97-4ee1-ba36-19a9830c4fd0_content_list.json b/octopusalleviatinghallucinationviadynamiccontrastivedecoding/ee076920-0b97-4ee1-ba36-19a9830c4fd0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7db8712427b182c8febf4910ba11c16b3cae681e --- /dev/null +++ b/octopusalleviatinghallucinationviadynamiccontrastivedecoding/ee076920-0b97-4ee1-ba36-19a9830c4fd0_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:975b957b839036911a7fd6f4caca7f0699783b68630b8730120da89059793998 +size 83719 diff --git a/octopusalleviatinghallucinationviadynamiccontrastivedecoding/ee076920-0b97-4ee1-ba36-19a9830c4fd0_model.json b/octopusalleviatinghallucinationviadynamiccontrastivedecoding/ee076920-0b97-4ee1-ba36-19a9830c4fd0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..16d5ad4dbd168d3fbec6708d69978d6f23a84a1d --- /dev/null +++ b/octopusalleviatinghallucinationviadynamiccontrastivedecoding/ee076920-0b97-4ee1-ba36-19a9830c4fd0_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bff9843c0430e54fa12fa4ea2e5d3a906580845a46718b0916f4f0169fc35e1f +size 104381 diff --git a/octopusalleviatinghallucinationviadynamiccontrastivedecoding/ee076920-0b97-4ee1-ba36-19a9830c4fd0_origin.pdf b/octopusalleviatinghallucinationviadynamiccontrastivedecoding/ee076920-0b97-4ee1-ba36-19a9830c4fd0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c3a80647c3057e9eca6db85fc24477a45228e285 --- /dev/null +++ b/octopusalleviatinghallucinationviadynamiccontrastivedecoding/ee076920-0b97-4ee1-ba36-19a9830c4fd0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e3a6e247a678c4e9ca93b0cdcf547e66ff27547739013ea0774c57bad69ee16 +size 3162291 diff --git a/octopusalleviatinghallucinationviadynamiccontrastivedecoding/full.md b/octopusalleviatinghallucinationviadynamiccontrastivedecoding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bb50efe5678025e5186cd5ac51aabddf9c8086b5 --- /dev/null +++ b/octopusalleviatinghallucinationviadynamiccontrastivedecoding/full.md @@ -0,0 +1,328 @@ +![](images/9cf84f6f55cefd1c3da5a9e6a29132bba8e0ab11e9b5314b37268d45e5a62526.jpg) + +# Octopus: Alleviating Hallucination via Dynamic Contrastive Decoding + +Wei Suo $^{1,2,4*}$ , Lijun Zhang $^{1,2,4*}$ , Mengyang Sun $^{2,3,4}$ , Lin Yuanbo Wu $^{5}$ , Peng Wang $^{1,2,4\dagger}$ , Yanning Zhang $^{1,2,4}$ + +$^{1}$ School of Computer Science, $^{2}$ Ningbo Institute, $^{3}$ School of Cybersecurity, Northwestern Polytechnical University, China. + +$^{4}$ National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean. + +$^{5}$ Department of Computer Science, Swansea University, United Kingdom. + +{suowei1994,lijunzhang,sunmenmian}@mail.nwpu.edu.cn {peng.wang,ynzhang}@nwpu.edu.cn + +# Abstract + +Large Vision-Language Models (LVLMs) have obtained impressive performance in visual content understanding and multi-modal reasoning. Unfortunately, these large models suffer from serious hallucination problems and tend to generate fabricated responses. Recently, several Contrastive Decoding (CD) strategies have been proposed to alleviate hallucination by introducing disturbed inputs. Although great progress has been made, these CD strategies mostly apply a one-size-fits-all approach for all input conditions. In this paper, we revisit this process through extensive experiments. Related results show that hallucination causes are hybrid and each generative step faces a unique hallucination challenge. Leveraging these meaningful insights, we introduce a simple yet effective Octopus-like framework that enables the model to adaptively identify hallucination types and create a dynamic CD workflow. Our Octopus framework not only outperforms existing methods across four benchmarks but also demonstrates excellent deployability and expansibility. Code is available at https://github.com/LijunZhang01/Octopus. + +# 1. Introduction + +Large Vision-Language Models (LVLMs) have achieved significant success over the past few years [4, 7, 25, 26, 61]. They have facilitated various Vision-and-Language (VL) tasks [14, 22, 40, 41] by adapting to different input instructions. However, LVLMs are facing a grand challenge: they often fail to accurately capture the visual content and tend to generate fabricated responses (e.g., imaginary objects, incorrect attributes and inconsistent relationship), which is known as hallucination [13, 24]. The hallucination issue + +![](images/0806d3c6678f9954b612267e2e1dbf01cfc1d7d815a15690620d0c6d456b51df.jpg) + +![](images/ccf59fec9c42e69c75b9bf7ebf55f82f371a89eaf40d5766e2b0c014af139201.jpg) + +![](images/ed8f94f1bc1a33bec5636a2303378a8222c5f5f67f2910ba53db9072f19778dc.jpg) +Figure 1. Paradigm comparison of different hallucination alleviation methods. (a) Retraining method. Constructing high-quality data to retrain these LVLMs. (b) Contrastive Decoding. Comparing the output distributions from the original and distorted inputs. (c) Octopus. Our method focuses on dynamically selecting suitable strategies to reduce hallucinations caused by various factors. + +seriously affects user trust and confidence, especially in applications that require trustworthy outcomes such as medical reports [10, 52] and automatic driving [8]. + +To alleviate the hallucination issue, existing approaches can be roughly categorized into two research lines. As shown in Fig. 1 (a), the first paradigm relies on constructing high-quality instruction tuning data and re-training the models to suppress hallucinations [13, 24, 28, 44]. However, such a strategy requires a well-designed data construction process with complex verification and expensive costs. In addition, additional training is strictly prohibited for deployed models. + +In contrast, as shown in Fig. 1 (b), Contrastive Decod + +ing (CD) methods [11, 19, 29, 51] aim to contrast the logit scores at each generative step from the original inputs (i.e., image input $v$ and corresponding text $q$ ) with those derived from the modified inputs (i.e., $v^{*}$ and $q^{*}$ ). As a post-hoc corrective mechanism, such a methodology can effectively eliminate hallucination without complex training. In practice, these CD methods focus on designing a new pair of $v^{*}$ and $q^{*}$ to alleviate hallucination problems. For example, [19] utilizes Gaussian noise to mask the original image and overcome language priors. [51] employs enhanced visual input to mitigate attention bias. + +Despite their strong performance, these CD methods mostly apply the same disturbed manner for all samples and generative steps. This motivates us to question: 1) Is a single CD strategy suitable for different samples? 2) Do all generative steps (i.e., predicted tokens) experience the same type of hallucination? To answer the two questions, we conduct exploratory experiments to understand the causes of hallucinations at both sample and token levels. Specifically, we use a group of CD methods as off-the-shelf diagnostic tools to investigate the pattern of hallucinations. Our analysis reveals that each CD method is only effective on specific hallucinated samples, and using a single CD strategy would lead to sub-optimal results. Meanwhile, we thoroughly investigate the process of hallucination emergence at the token level. Through an enumeration method [5] and qualitative analysis, we find that fabricated responses are hybrid and each generative step faces a unique hallucination issue. + +The above results indicate that a one-size-fits-all approach struggles to correct different types of hallucinations effectively. Thus a natural idea is to combine multiple CD methods to reduce hallucinations from various sources. However, without well-defined labels, identifying the optimal strategy for different input samples is challenging. Additionally, token generation is sequentially dependent and involves a vast solution space, making it difficult to choose the best CD approach at each generative step. + +To tackle the above problems, as shown in Fig.1 (c), we introduce a simple yet effective framework, called Octopus. Different from previous works, our method focuses on guiding the model to dynamically organize the contrastive decoding workflow and selecting the appropriate CD strategy based on different inputs. In particular, we first build a transformer-based block and a learnable token to adaptively recognize the type of hallucination, similar to the Octopus's eyes. According to different decisions, each CD strategy is regarded as a "tentacle" to perform a specific contrastive operation. Finally, leveraging Direct Preference Optimization (DPO) [34] or reinforcement learning [37, 49], Octopus can be easily optimized. Benefiting from the above designs, the proposed method not only effectively reduces hallucinated content, but also scales well for deployments due to avoiding retraining weights of LVLMs. More importantly, as a + +general framework, subsequent CD methods can be seamlessly integrated without additional adjustments. In summary, we make the following contributions: + +1) Our work reveals that the mechanism of hallucination occurrence is a complex hybrid and different samples (or tokens) suffer from various forms of hallucination challenges. +2) We develop a new framework Octopus that can adaptively recognize the types of hallucination and build a dynamically contrastive decoding workflow to correct fabricated content. +3) Octopus achieves state-of-the-art performance across four benchmarks for both generative and discriminative tasks, while also demonstrating excellent deployability and expansibility. + +# 2. Related work + +# 2.1. Large Visual-Language Models + +Large Visual-Language Models (LVLMs) [4, 7, 25, 26, 61] usually consist of three key components: a visual encoder like CLIP [33], a Large Language Model (LLM) such as LLAMA [42] and a cross-modal alignment module that connects the visual encoder's output to the LLM. LVLMs have obtained impressive performance in visual content understanding and multi-modal reasoning such as image captioning [14, 61], referring expression comprehension [27, 56], human-object interaction [58], and visual question answering [9, 46]. + +# 2.2. Hallucination in LVLMs + +To alleviate the hallucinations in LVLMs, both data-driven retraining and Contrastive Decoding (CD) methods have been proposed. Data-driven methods aim to enhance data quality to reduce the hallucinations [13, 24, 28, 44]. For example, some works introduce negative data [24] and counterfactual data [54] to mitigate hallucination issues. [48] cleans the dataset to minimize noise and errors. [50, 55] annotates a high-quality hallucination dataset to suppress the occurrence of hallucinations by fine-tuning. In contrast, CD methods tackle hallucinations by comparing output distributions from original and distorted input without altering the model's weights. For instance, [19] alleviates hallucinations by counteracting language priors, while [11] tackles them through refined visual prompts. Different from the above methods, our work focuses on adaptively selecting the most suitable CD strategies to alleviate different hallucination issues. + +# 3. Preliminary + +# 3.1. Large Vision Language Models + +Given a Large Vision-Language Model (LVLM) with parameters $\theta$ , the model can effectively perform various multimodal tasks using a visual input $v$ and a textual instruction + +$q$ . Specifically, at each generative step $t$ , the auto-regressive LVLM conducts the following calculations: + +$$ +\ell_ {t} = \log p \left(y _ {t} \mid v, q, y _ {< t}; \theta\right), \tag {1} +$$ + +$$ +y _ {t} \sim \operatorname {S o f t m a x} \left(\ell_ {\mathrm {t}}\right), \tag {2} +$$ + +where $\ell_t$ is the logit score for the next token $y_t$ , while $y_{