diff --git a/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/b6444176-c113-4ffe-9cd2-8898c483f069_content_list.json b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/b6444176-c113-4ffe-9cd2-8898c483f069_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..442d9498e750732e8162bb2f1174a9dcef0e363b --- /dev/null +++ b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/b6444176-c113-4ffe-9cd2-8898c483f069_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8f85465a83404a2028103cd138fcffeab80740d6686eca347c7467e6792dad2 +size 83298 diff --git a/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/b6444176-c113-4ffe-9cd2-8898c483f069_model.json b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/b6444176-c113-4ffe-9cd2-8898c483f069_model.json new file mode 100644 index 0000000000000000000000000000000000000000..545718f5d81ab413261d4e28058d1964683ae7a5 --- /dev/null +++ b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/b6444176-c113-4ffe-9cd2-8898c483f069_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9831ccbe5c8d99d7df9b84318574f7213c184903e62c5c35071d8a3c95a2a3e8 +size 107994 diff --git a/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/b6444176-c113-4ffe-9cd2-8898c483f069_origin.pdf b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/b6444176-c113-4ffe-9cd2-8898c483f069_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a3c88961aa2b97cdefc4f7402fc601c615dc223a --- /dev/null +++ b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/b6444176-c113-4ffe-9cd2-8898c483f069_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:708b40e2d0129f35a46116fb1d6109f37e7f3b0a4932897afebb26cfa3190dbd +size 5218647 diff --git a/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/full.md b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8c17575d63f2e5739157f5d736ba5d4716bfc23a --- /dev/null +++ b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/full.md @@ -0,0 +1,310 @@ +# SegAgent: Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories + +Muzhi Zhu $^{1,2}$ Yuzhuo Tian $^{1}$ Hao Chen $^{1*}$ Chunluan Zhou $^{2}$ Qingpei Guo $^{2*}$ Yang Liu $^{1}$ Ming Yang $^{2}$ Chunhua Shen $^{1,2*}$ + +1 Zhejiang University, China 2 Ant Group + +# Abstract + +While MLLMs have demonstrated impressive image understanding capabilities, they still struggle with pixel-level comprehension, limiting their practical applications. Current evaluation tasks such as VQA and visual grounding remain too coarse to assess fine-grained pixel comprehension accurately. Although segmentation is foundational for pixel-level understanding, existing methods often require MLLMs to generate implicit tokens, decoded through external pixel decoders. This approach disrupts the MLLM's text output space, potentially compromising language capabilities and reducing flexibility and extensibility while failing to reflect the model's intrinsic pixel-level understanding. Thus, we introduce the Human-Like Mask Annotation Task (HLMAT), a new paradigm where MLLMs mimic human annotators using interactive segmentation tools. Modelling segmentation as a multi-step Markov Decision Process, HLMAT enables MLLMs to iteratively generate text-based click points, achieving high-quality masks without architectural changes or implicit tokens. Through this setup, we develop SegAgent, a model fine-tuned on human-like annotation trajectories, which achieves performance comparable to SoTA methods and supports additional tasks like mask refinement and annotation filtering. HLMAT provides a protocol for assessing fine-grained pixel understanding in MLLMs and introduces a vision-centric, multi-step decision-making task that facilitates the exploration of MLLMs' visual reasoning abilities. Our adaptations of policy improvement method StaR and PRM guided tree search further enhance model robustness in complex segmentation tasks, laying a foundation for future advancements in fine-grained visual perception and multi-step decision-making for MLLMs. Code can be found at https://github.com/aim-uofa/SegAgent. + +# 1. Introduction + +Recent advancements [33, 43] in large language models (LLMs) have brought them to near-human levels in many domains. However, multimodal LLMs (MLLMs) still face significant limitations in visual tasks [44, 45], especially in complex pixel-level understanding [56], which restricts their applicability in real-world scenarios. Despite efforts to develop MLLM vision capabilities through tasks like Visual Question Answering (VQA)[19, 20], and visual grounding [2, 46], these tasks fail to assess fine-grained pixel-level comprehension effectively. On the other hand, tasks like UI Operations [64, 66] and Robotic Control [40] require pixel-level precision but are often constrained to specialized environments with limited, heterogeneous data, making them difficult to scale to complex, open-world scenarios. + +Segmentation [11, 28, 29, 68, 70-72] is one of the most critical visual tasks for reflecting pixel-level understanding capabilities. Moreover, there is a wealth of segmentation data [15] available, which can sufficiently meet the needs of complex open-world scenarios. Establishing a method that enables MLLMs to effectively leverage this large-scale data is essential to further explore and enhance their pixel-level understanding capabilities. However, current approaches [16, 38, 49, 50, 67] integrating MLLMs with segmentation tasks often rely on implicit token outputs and require additional pixel decoders. While effective, these methods alter the original MLLM output space, potentially compromising semantic generalization by shifting away from language-based outputs. The use of implicit tokens also limits extensibility, complicating efforts to implement a unified textual framework across diverse tasks. Furthermore, reliance on an extra pixel decoder makes it challenging to accurately reflect the MLLM's intrinsic pixel understanding capabilities [4, 41]. There are also methods [6, 18, 26, 36] that attempt to perform segmentation using purely textual outputs, such as polygon vertex sequences or sampled points along the mask contour. However, this strategy limits mask precision, and generating lengthy coordinate sequences is particularly challenging. Even for a skilled human annotator, tracing + +the mask contour point by point is time-intensive and laborious [23]. Recent annotation practices [15] have introduced interactive segmentation tools [15, 27] that reduce workload by allowing iterative mask refinement through positive and negative clicks. It's natural to ask whether current MLLMs can perform segmentation by emulating the trajectories of human annotators using interactive segmentation tools. This approach would enable the model to complete the segmentation task more easily in a pure-text format. + +Therefore, we propose a new segmentation paradigm, enabling MLLMs to imitate human annotators using interactive segmentation tools by modeling the segmentation task as a multi-step Markov Decision Process. We refer to this as the Human-Like Mask Annotation Task (HLMAT). Compared to previous vision tasks, HLMAT imposes greater demands on the model's pixel-level understanding for both input and output. This paradigm allows MLLMs to iteratively generate text-based coordinates without relying on specialized architectures or implicit tokens, supporting fair comparisons across different MLLMs. We believe HLMAT offers a novel evaluation protocol to explore and advance MLLMs' fine-grained visual capabilities. + +We first develop an automated algorithm to transform existing datasets into annotation trajectories. Using the trajectories, we fine-tune MLLMs, resulting in our model, SegAgent. Our experiments show that, after fine-tuning, various MLLMs achieve performance comparable to state-of-the-art methods. Additionally, we demonstrate that SegAgent not only excels at segmentation tasks but also supports mask refinement and annotation filtering effectively. + +Furthermore, HLMAT, as a multi-step decision-making task, directly assesses pixel-level visual skills, unlike prior MLLM decision-making tasks that focus primarily on textual reasoning [63] or simple UI-based operations [1]. Our study lays the foundation for exploring MLLMs as vision-centered, multi-step decision-making agents, addressing the gap in pixel-level visual understanding in multi-step contexts. In this work, we first examine commonly used referring segmentation datasets like RefCOCO [57] and find that their annotation complexity and quality are insufficient for multi-step decision-making. To support deeper exploration, we introduce high-quality reference segmentation datasets. + +Next, we investigate whether certain enhancement methods for text-based LLM decision-making can be adapted to our proposed HLMAT task. We select StaR [62], a straightforward yet representative RL-style policy improvement method, and modify it for our purposes. Our findings reveal that incorporating the StaR algorithm, as opposed to solely relying on SFT data, significantly improves the model's performance on the test dataset. Additionally, inspired by frameworks such as Tree-of-Thoughts [55], we explore inference-stage techniques. By combining process reward modeling (PRM) [22, 30] with tree search, + +we demonstrate that for complex scene segmentation tasks, implementing tree search effectively mitigates errors arising from inaccurate predictions, substantially enhancing the model's robustness in challenging segmentation scenarios. This work lays a strong foundation for advancing vision-centered, multi-step decision-making agents in future research. + +Our main contributions are as follows: + +- We propose Human-Like Mask Annotation Task (HLMAT), a new segmentation framework where MLLMs imitate human annotators using interactive tools, modeled as a multi-step Markov Decision Process. HLMAT raises pixel understanding demands and only requires text-based coordinate outputs, allowing fair comparison across MLLMs and serving as a new protocol for fine-grained visual capability assessment. +- Using an automated algorithm to generate human-like annotation trajectories from existing datasets, we fine-tune MLLMs to create SegAgent. Experiments show SegAgent achieves competitive performance, supporting segmentation, mask refinement, and annotation filtering. +- HLMAT lays a foundation for multi-step, vision-centered decision-making research. We introduce the HRES dataset, suitable for complex reasoning, and demonstrate that text-based decision improvement methods like StaR+ (training) and tree search with PRM (inference) effectively enhance segmentation performance. + +# 2. Related Work + +# 2.1. Pixel Understanding in MLLMs + +Many studies have explored pixel understanding in MLLMs by combining these models with segmentation tasks. However, most approaches [17, 38, 39, 61, 65, 67] require MLLMs to learn an implicit token and involve additional fine-tuning with a separate pixel decoder. Among them, LISA [17] was the first to integrate MLLMs with SAM [15], while subsequent works [38, 67] introduced additional tasks and larger training datasets, so we did not include them in our experimental comparisons. Other methods attempt to let MLLMs perform segmentation through text-based outputs, such as using polygon vertex sequences [36] or randomly sampled points [9] where the model predicts if a point lies within the mask. However, this approach inherently limits mask precision. Our work, on the other hand, achieves segmentation by imitating the trajectory of human annotators using interactive segmentation tools. This approach enables the model to perform highly detailed segmentation tasks through simple text-based outputs. + +# 2.2. MLLMs as Decision-Making Agents + +In the text domain, extensive research has explored how to adopt LLMs as decision-making agents and enhance + +their reasoning and decision-making capabilities. Some works [48, 53, 54] improve LLM reasoning through prompt engineering during inference, while others [62, 69] investigate policy improvement methods during training. However, research on decision-making agents in the visual domain remains limited. A few studies integrate MLLMs [7, 32] into training processes, but they train simple MLP or transformer layers to align action spaces without using text-based actions. + +As RL4VLM [63] emphasized, prior work primarily examined MLLM capabilities in non-interactive tasks. RL4VLM represents the first systematic approach to fine-tuning an entire VLM as a decision-making agent using reinforcement learning, directly interacting with the environment through open-ended text. Our work shares this advantage, offering a novel perspective on evaluating MLLM decision-making in interactive tasks. However, unlike RL4VLM, which focuses on poker-like tasks that are primarily text-based, our proposed task directly involves fine-grained visual perception, laying the groundwork for further exploration of vision-centered, multi-step decision-making agents. + +# 2.3. Interactive Segmentation Tools + +The objective of interactive segmentation [8, 10, 13, 27, 42] is to segment regions of interest based on user-provided inputs, such as clicks, boxes, and scribbles. RITM [42] leverages COCO [3] and LVIS [12] as training data, achieving substantial segmentation quality. SimpleClick [27] was the first to explore the application of plain Vision Transformers (ViT) in interactive segmentation. InterFormer [13] decouples feature encoding and prompt fusion, effectively mitigating the high-latency issue observed in SimpleClick. SAM [15] adopts a similar design, trained on the large-scale SA1B dataset, and achieves notable generalization capabilities, supporting box inputs. However, existing interactive segmentation models, while capable of generating accurate masks, require human input. In contrast, our work aims to teach MLLMs to perform segmentation by imitating human annotators using interactive segmentation tools. + +# 3. Preliminaries and Task Formulation + +Given an image $I$ and a text prompt $P$ , a professional human annotator is able to accurately draw the mask $M$ corresponding to the object indicated by the text prompt $P$ . However, drawing the mask $M$ directly is time-intensive. Modern annotators commonly utilize an interactive segmentation tool, iteratively providing a series of positive and negative click points based on the current image $I$ and mask $M$ , making it easier to create a high-quality mask. This entire process can be modeled as a Markov Decision Process (MDP) consisting of $(S, A, T, R, \gamma)$ , where: + +- State $s_t \in S$ : Represents the current status of the mask + +$M_{t}$ in the segmentation process, along with the history of all previous actions $a_{0:t - 1}$ , i.e., $s_t = (M_t,a_{0:t - 1})$ . The initial state $s_0$ corresponds to an empty mask, while the goal state $s_{\mathrm{goal}}$ corresponds to the optimal target mask $M_{\mathrm{target}}$ . + +- Action $a_{t} \in \mathcal{A}$ : Here, each action $a_{t}$ refers to an annotator's click operation, consisting of two components: an attribute $\alpha_{t}$ and a coordinate $\mathbf{c}_{t}$ . The attribute $\alpha_{t} \in \{+1, -1\}$ indicates whether the point is a positive or negative click, while the coordinate $\mathbf{c}_{t} \in [0,1]^{2}$ represents the relative position of the point within the image, scaled to the [0, 1] range for both $x$ and $y$ axes. + +- Transition Function $T: \mathcal{S} \times \mathcal{A} \rightarrow \mathcal{S}$ : Represents the environment's response to the current state-action pair $(s_t, a_t)$ , mapping it to the next state $s_{t+1}$ . In this task, the environment's response is determined by the input image $I$ and the interactive segmentation network $F_{iter}$ . Specifically, the next mask $M_{t+1}$ is generated by $F_{iter}(I, s_t, a_t)$ , with $s_{t+1} = (M_{t+1}, a_{0:t})$ . + +- Reward Function $R: \mathcal{S} \to \mathbb{R}$ : The reward function evaluates the accuracy of the current mask $M_t$ relative to the target mask $M_{\text{target}}$ using the Intersection over Union (IoU) metric. Specifically, $R(s_t) = \mathrm{IoU}(M_t, M_{\text{target}})$ , where a higher IoU score indicates a better match between $M_t$ and $M_{\text{target}}$ . + +Since we are focused solely on the quality of the final mask, we omit the discount factor $\gamma$ in our setup. Given that the action space can be fully represented by text, and the state space can be fully represented by images, a competent MLLM should be able to perform this task as effectively as a human annotator. Our objective is to train a policy $\pi_{\theta}(a_t|s_t,I,P)$ for the MLLM that, given an image $I$ and a text prompt $P$ , outputs an action $a_{t}$ at each time step $t$ . The goal is to maximize the expected reward $R(s_{T})$ within a limited number of steps $T$ . We refer to this as the Human-Like Mask Annotation Task (HLMAT). In practice, to enable the MLLM to support more flexible use cases—such as mask refinement—we remove the dependency on $a_{0:t-1}$ in $s_{t}$ , allowing the model to determine the next action based solely on the current mask $M_{t}$ . + +# 4. SegAgent + +As mentioned in Sec. 3, our goal is to train an MLLM to learn a policy $\pi_{\theta}(a_t|s_t,I,P)$ that imitates the annotator's trajectory. In practice, to enable the MLLM to support more flexible use cases—such as mask refinement—we remove the dependency on $a_{0:t-1}$ in $s_t$ , allowing the model to determine the next action based solely on the current mask $M_t$ , so the policy becomes $\pi_{\theta}(a_t|M_t,I,P)$ . + +As illustrated in Fig. 1, the current mask $M_{t}$ and the original image $I$ are processed as visual inputs to the MLLM. Following common visualization techniques used by human annotators, we overlay the current segmentation result $M_{t}$ + +![](images/24c161d13ca27450c1c9d4d8ef95c8d4bfaa5852ebe07a81458301e3f6294238.jpg) +Figure 1. The overall framework of SegAgent. The image below shows a complete set of trajectories. We visualize current action $a_{t}$ and the resulting mask $M_{t + 1}$ in one image. + +as a semi-transparent green mask on the original image $I$ to create a new composite image $I_{t}$ . This image $I_{t}$ serves as the final input for the MLLM, allowing the model to simultaneously perceive information about $M_{t}$ and $I$ in a single frame. + +The text input $P$ to the MLLM includes two key components: (1) a description of the task and action space, and (2) a description of the target object to be segmented. This prompt design provides the model with context about the overall objective and the segmentation target, ensuring it can interpret both the visual input and the segmentation goal. Detailed prompt design can be found in the appendix. + +After processing and analyzing the inputs, the MLLM outputs the next action $a_{t+1}$ in text format, such as "Positive point: (175,483)." The action is then converted into a format compatible with the interactive segmentation model via a predefined post-processing function. Subsequently, all previous actions $a_{0:t}$ , the current mask $M_t$ , and the original image $I$ are fed into the interactive segmentation model $F_{iter}$ , which generates the updated mask $M_{t+1}$ . This iterative process continues until the segmentation achieves a satisfactory result. + +The remainder of this section is organized as follows: First, we describe our approach for collecting human-like annotation trajectories (Sec. 4.1). Next, we analyze the two essential capabilities that a robust SegAgent should possess (Sec. 4.2). Inspired by traditional RL-style policy improvement methods, we then adapt the StaR method to our task (Sec. 4.3). Finally, we combine the process reward model with tree search to enhance the model's robustness in complex scenarios (Sec. 4.4). + +# 4.1. Human Annotator Trajectories Generation + +Suppose we have an annotated dataset $D_{seg} = \{(I, M_{target}, P)\}$ , where each sample includes an image $I$ , a target mask $M_{target}$ , and a text prompt $P$ . Ideally, recording all state-action pairs $(s_t, a_t)$ in real-time during the human annotation process would yield the de + +![](images/c6d0af4aa8060dd1a4aecf3e0afa4b57ea3eabb72441a46a70ef894d0d16d99e.jpg) +Figure 2. An example of generated trajectory. We visualize current action $a_{t}$ and the resulting mask $M_{t + 1}$ in one image. Due to the noise from GT Mask, the action for Iteration 3,4 is meaningless + +sired trajectory $[s_0, a_0, s_1, a_1, \dots, s_T, a_T]$ . However, existing datasets lack such trajectories, and re-hiring human annotators to annotate the data while recording trajectories would be very costly. Thus, we consider whether it is possible to derive the trajectory dataset $D_{traj} = \{(I, M_{target}, P, [s_0, a_0, s_1, a_1, \dots, s_T, a_T])\}$ from the existing dataset $D_{seg} = \{(I, M_{target}, P)\}$ using a rule-based method. Thanks to researchers in interactive segmentation, an iterative click simulation strategy has been developed to automate the evaluation of interactive segmentation models. This strategy can be understood as a function $F_{sim}$ , which takes as input the current mask $M_t$ and the ground truth mask $M_{target}$ , and outputs the next action $a_{t+1}$ , i.e., $a_{t+1} = F_{sim}(M_t, M_{target})$ . Specifically, this function computes the false positive and false negative regions between the current mask $M_t$ and the ground truth mask $M_{target}$ , placing the next click action at the center of the error region based on the size and position of these regions. With the help of function $F_{sim}(\ref{eq:2})$ , we can simulate trajectories $[s_0, a_0, s_1, a_1, \dots, s_T, a_T]$ based on $M_{target}$ . However, the annotation quality of $M_{target}$ may vary, potentially generating suboptimal trajectories. This issue is especially pronounced toward the end of a trajectory (see Fig. 2), where noise in $M_{target}$ can cause erroneous actions that misguide the MLLM during training. To address this, we introduce the following steps during generation: + +- Limit Maximum Trajectory Length: We set a maximum trajectory length $T$ . For lower-quality datasets, we reduce $T$ to avoid excessive noisy actions at the trajectory's end. +- Terminate on Satisfactory Reward: We calculate the reward $R(s_{t})$ for the current state $s_{t}$ , and if $R(s_{t})$ reaches a threshold $\tau_{stop}$ , we terminate the trajectory generation. +- Discard Low-Impact Actions: We evaluate the reward difference before and after executing an action $a_{t}$ . If the reward difference $R(s_{t + 1}) - R(s_{t}) < \tau_{diff}$ , we discard action $a_{t}$ . This is because abnormal actions can cause the interactive segmentation model to collapse, significantly reducing the quality of the resulting mask $M_{t + 1}$ . + +To accommodate segmentation datasets of varying quality, we can adjust these three parameters to reduce noise in the generated trajectories. Following this process, we can generate the annotation trajectory dataset $D_{traj}$ from the existing segmentation dataset $D_{seg}$ and use these data to fine-tune the MLLM through instruction tuning, resulting in our SegAgent. + +# 4.2. What Makes a Good SegAgent? + +Now that we have obtained the annotation trajectory dataset $D_{traj}$ , in theory, we can fine-tune any MLLM to become a SegAgent. Here, we identify that HLMAT primarily relates to two main competencies of an MLLM: + +- Coarse-Grained Localization Ability Based on Text: This capability is primarily demonstrated when the MLLM executes the first or initial actions, where it must use the text prompt and global image context to locate the approximate position of the object. This is especially critical for complex or ambiguous prompts, such as "the second black motorcycle on the right." If the model fails in this initial localization, the final IoU score will likely be very low. This skill is closely related to tasks commonly known as Visual Grounding or Referring Expression Comprehension (REC), where previous work [2, 46] has shown that many different architectures of MLLMs can achieve good performance when trained on sufficient data. The only difference in our setup is that while other tasks typically output bounding box coordinates, our task supports point-based coordinates. Our experiments further reveal that the initial action can be represented as either a box or a point. However, this localization ability itself is not the primary focus of our work, as it has already been widely studied and optimized. + +- Fine-Grained Pixel Understanding and Mask Refinement: This capability is demonstrated when the MLLM performs subsequent actions, where it must accurately adjust the mask boundaries based on the current mask and local image information to gradually improve mask quality. This process requires the model to have an exceptionally detailed perception and understanding of both the input image and mask, as well as the ability to make precise localization decisions during output. Few studies have focused on this fine-grained pixel understanding of MLLMs, making this a key focus of our research. + +To better investigate the second capability, a complex and high-quality segmentation dataset is essential. Here, complexity refers to the requirement that the dataset should necessitate relatively long annotation trajectories to reach the predefined threshold $\tau_{stop}$ . High quality refers to minimal noise in annotations, ensuring accuracy at the trajectory endpoints. Our analysis of the commonly used Refcoco dataset revealed that it lacks the necessary complexity and annotation quality (see Sec. 5.2). Consequently, we select two alternative datasets that better meet these criteria: DIS5K [37] and ThinObject5K [21]. Our further exploration will primarily focus on these datasets. + +# 4.3. Policy Improment + +Training the MLLM with the annotation trajectory dataset $D_{traj}$ generated in Sec. 4.1 equips the model with foundational capabilities for the HLMAT task. However, because + +Algorithm 1 SegAgent Policy Improvement with StaR+ +1: Input: SegAgent model trained generated trajectory dataset $D_{\mathrm{traj}} = \{(I^i,M_i^i,\mathcal{T}^i)\}_{i = 1}^m$ +2: $S_0\gets \mathrm{SegAgent}$ // Initialize SegAgent model +3: $D_0\gets D_{\mathrm{traj}}$ // Initialize trajectory dataset +4: for $n = 1$ to $N$ do +5: $\hat{\mathcal{T}}^i\gets S_{n - 1}(I^i,M_i^i,\mathcal{T}^i)$ for all $i\in [1,m]$ // Generate trajectories +6: $D_{n}\gets \{(I^{i},M_{\mathrm{target}}^{i},P^{i},\mathrm{Refine}(\hat{\mathcal{T}}^{i}))\}_{i = 1}^{m}$ //Filter and correct trajectories based on reward function and $F_{sim}$ +7: $S_{n}\gets \mathrm{train}(S_{0},D_{0}\cup D_{n})$ // Fine-tune on combined dataset +8: end for + +these generated trajectories are nearly optimal and lack diversity, the model's generalization and robustness remain constrained. It is worth noting that our framework can directly interact with the environment, allowing the MLLM to engage in online learning or offline reinforcement learning (RL) to continually refine its policy through interaction. In this paper, we choose to implement and explore a straightforward RL-style policy improvement method, specifically the StaR [62] algorithm. + +As shown in Alg. 1, we start with a trained SegAgent model $S_0$ and the annotation trajectory dataset $D_{traj}$ . In each iteration, we first use $S_{n-1}$ to perform rollouts on the training images, generating a set of new trajectories $\hat{T}^i$ . Next, we filter and refine these trajectories based on the reward function and $F_{\mathrm{sim}}$ , producing an updated trajectory dataset $D_n$ . Finally, we merge $D_n$ with $D_{traj}$ and fine-tune the model on this combined dataset to obtain an improved SegAgent model $S_n$ . Unlike the original StaR algorithm (see comparison in Appendix), which filters based on the overall correctness of the trajectory, our StaR+ method calculates the reward change for each individual action. We retain actions that increase the reward, while for actions that decrease the reward, we replace them with actions generated by $F_{\mathrm{sim}}$ as corrections. + +# 4.4. Process Reward Model and Tree Search + +Process-supervised Reward Models [22] (PRM) refer to the techniques where a reward model is trained to evaluate the reward at each step of a multi-step task. PRM can guide the training process by integrating with approaches like RLHF [34] or reject sampling [60]. Previous work [30] has also shown that PRM can be used as guidance during the inference stage to enhance the model's search and reasoning capabilities. + +In our work, since the reward at each step can be directly obtained from the environment during training, PRM is un + +Algorithm 2 PRM with Heuristic Greedy Search +1: Input: SegAgent model $S$ , image $I$ , text prompt $P$ +2: $M_0 \gets$ initialize_mask(I) // Initialize the mask +3: $R_0 \gets \mathrm{PRM}(S, I, M_0, P)$ // Predict the initial reward +4: for $t = 1$ to $T$ do +5: $[a^1, a^2, \dots, a^K] \gets \text{get\_candidate}(S, I, M_{t-1}, P)$ // Get $K$ candidate actions +6: for $i = 1$ to $K$ do +7: $M^i \gets F_{iter}(I, M_{t-1}, a^i)$ +8: $R^i \gets \mathrm{PRM}(S, I, M^i, P)$ +9: end for +10: best_index $\leftarrow \operatorname{argmax}(R^1, R^2, \dots, R^K)$ // Get the best action guided by PRM +11: $M_t \gets M^{\text{best\_index}}$ // Predict the next mask +12: $R_t \gets R^{\text{best\_index}}$ // Predict the next reward +13: end for +14: best_index $\leftarrow \operatorname{argmax}(R_0, R_1, \dots, R_T)$ // Get the best step +15: return $M_{\text{best\_index}}$ + +necessary for the training phase. Therefore, we primarily explore the application of PRM techniques during the inference stage. There are two main reasons motivating our use of PRM techniques: + +- Providing a Stop Signal: We use PRM to provide a stop signal for the task. An alternative approach would be to add a stop action to the action space, but this would increase the complexity of the action space, thereby increasing the difficulty of model training. This is especially challenging when $M_{\text{target}}$ contains noise, as it becomes difficult to determine the exact stopping point, making it hard for the model to generate an accurate stop action. By directly predicting the reward, PRM allows the model to decide when to stop without adding complexity to the action space. +- Enhancing Robustness in Complex Scenarios: In complex scenarios, since the model cannot backtrack or undo actions, executing an inaccurate or incorrect action at any step can adversely affect subsequent actions, leading to lower quality in the final result. With the help of PRM, we can combine the model with advanced search strategies, such as Breadth-First Search (BFS) and Monte Carlo Tree Search (MCTS), to improve the model's robustness in complex situations. + +We implement PRM by asking the MLLM perform an additional text prediction, where the MLLM learns to predict the mIoU score of the current state before generating the next action, for example, "Current mIoU: 0". Based on this, we design a simple heuristic greedy search strategy, as shown in Alg. 2. At each step, we first use Multino + +mial Sampling to generate $K$ candidate actions. For each action, we perform one interaction with the segmentation model to obtain a new mask and calculate the reward using PRM. Finally, we select the action with the highest reward as the next action. This process continues until the maximum number of steps $T$ is reached or until the reward converges. We then return the mask with the highest reward as the final output. A more detailed illustration can be found in the appendix. Note that, in the algorithm description, we omit the historical actions included in the input of $F_{iter}$ for simplicity. It is worth noting that more complex search strategies, such as BFS and MCTS, are also feasible. However, there exists a trade-off between computational complexity and performance. Our experiments demonstrate that the simplest and most efficient heuristic greedy search strategy already provides significant performance improvements. We leave the exploration of more complex search strategies for future work. + +# 5. Experiments + +# 5.1. Evaluation Protocol + +As mentioned earlier, HLMAT serves as a novel evaluation protocol for assessing and advancing MLLMs' fine-grained pixel understanding capabilities. It also provides a framework to explore MLLMs as vision-centered, multi-step decision-making agents. We select two representative MLLMs as initialization models for SegAgent: a linear projector-based model, LLaVA-v1.5-7B [24, 25], and a stronger Q-former-based model, Qwen-VL-7B. We fine-tune each MLLM on the generated trajectory dataset for 2 epochs, keeping the image encoder frozen while allowing the LLM and projector layers to update. We use DeepSpeed's Zero2 mode for multi-GPU training, with all training experiments conducted on 8 NVIDIA 80GB GPUs. + +Other training parameters follow the default settings provided by the official repositories; details can be found in the Appendix. We used two main interactive segmentation models for this task: SimpleClick [27] and SAM [15]. SimpleClick supports point input only, while SAM also supports box input. During trajectory generation, we created trajectories for both segmentation models. As discussed in Sec. 4.2, SegAgent's first action can be either a point or a box. Our preliminary experiments demonstrated that SAM achieves better performance when provided with an additional box input. Therefore, unless otherwise specified, we include box input when using SAM in subsequent experiments. + +We conduct experiments in two settings: the widely adopted Referring Expression Segmentation (RES) and our newly proposed High-quality Referring Expression Segmentation (HRES). + +For inference evaluation, we set a fixed maximum num + +Table 1. Comparison of methods on RES dataset. We indicate which models use SAM. "SClick" denotes the use of SimpleClick as the interactive segmentation model. + +
MethodVenuerefCOCOrefCOCO+refCOCOg
valtestAtestBvaltestAtestBval(U)test(U)
traditional methods
MAttNet [59]CVPR'1856.5162.3751.7046.6752.3940.0847.6448.61
LAVT [51]CVPR'2272.775.868.862.168.455.161.262.1
CRIS [47]CVPR'2270.573.266.165.368.153.759.960.4
PolyFormer-L [26]CVPR'2376.9478.4974.8372.1575.7166.7371.1571.17
X-Decoder [73]CVPR'23------64.6-
SEEM [74]NeurIPS'23------65.7-
LLM based methods
LISA(SAM) [16]CVPR'2474.979.172.365.170.858.167.970.6
PixelLM [39]CVPR'2473.076.568.266.371.758.369.370.5
PerceptionGPT [35]CVPR'2475.178.671.768.573.961.370.371.7
GSVA(SAM) [49]CVPR'2477.278.973.565.969.659.872.773.3
SAM4MLLM(Qwen) [9]ECCV'2477.180.972.571.576.864.774.575.2
Qwen box [2] + SAM-71.7975.2067.2965.3971.6259.1266.9368.94
Our methods
SegAgent-LLaVA+SAM79.2081.4475.7271.5376.6865.4474.8074.90
SegAgent-Qwen+SAM78.0180.3474.9870.8675.5265.7574.4974.62
SegAgent-LLaVA+SClickthis work77.8180.0374.1266.7371.1659.8970.4571.25
SegAgent-Qwen+SClick79.6981.3576.5772.4975.8066.8975.1175.20
+ +ber of steps $T$ and, following LISA [16], used cumulative IoU (cIoU) as the evaluation metric. For RES, due to its relatively simple masks, we set $T = 7$ and $K = 1$ in Alg. 2. For HRES, with more complex segmentation requirements, we set $T = 11$ and $K = 3$ . + +# 5.2. Datasets + +We first conduct a fair comparison between our method and other state-of-the-art (SOTA) methods on the Referring Expression Segmentation (RES) dataset to demonstrate the effectiveness of our approach. Subsequently, we extend our method to more complex scenarios by experimenting on the proposed High-quality Referring Expression Segmentation (HRES) dataset. We begin by introducing the basic information of these two datasets and analyzing their characteristics to highlight the necessity of using the HRES dataset. + +Referring Expression Segmentation (RES): The RES dataset is composed of three subsets: refCOCO [58], refCOCO+ [58], and refCOCOg [31]. All three datasets are based on images from the COCO dataset, with differences in the types of referring expressions provided. RefCOCO is the earliest dataset, featuring relatively simple descriptions. Compared to RefCOCO, RefCOCO+ prohibits annotators from using location-based descriptions, making the descriptions more focused on object appearance and attributes. RefCOCOg includes longer descriptions with more detailed information. However, the mask quality in refCOCO(+/g) is relatively low, with many masks containing significant noise. This makes it challenging to generate long and precise annotation trajectories. On the other hand, the complexity of the datasets does not meet the requirements for + +![](images/01b01f9f32ba8e4bd6bee0f69d3c519bbea38cd639967d63af561b422312a37d.jpg) +Figure 3. Comparison of dataset complexity. + +the multi-step iteration in HLMAT. In most cases, only 1-2 steps are sufficient to complete the task. + +To address these limitations, we introduce a new dataset, the High-quality Referring Expression Segmentation (HRES) dataset, designed to facilitate further exploration and better assess the performance of MLLMs under more complex conditions. + +High-quality Referring Expression Segmentation (HRES): Our proposed HRES dataset features higher annotation quality, with more detailed and complex masks, which require additional steps to complete. The data in HRES is derived from the HQSeg-44K dataset proposed by HQ-SAM [14]. We selected two subsets, DIS5K [37] and ThinObject5K [21], as they contain texts or category descriptions with minimal ambiguity. The DIS5K (Dichotomous Image Segmentation) subset consists of high-resolution images and high-quality binary segmentation masks, specifically designed to handle challenging segmentation tasks. It includes 5,470 images spanning 22 groups and 225 fine-grained object categories. The ThinObject5K + +subset focuses on objects with thin, elongated structures, such as insect legs and bicycle spokes, which are typically challenging for segmentation models. + +As shown in Fig. 3, we use the number of clicks (required to reach the target IoU) as a measure of dataset complexity. The HRES dataset is significantly more complex than the RES dataset, requiring more steps to achieve the target IoU. + +# 5.3. Main Results + +In this section, we present the main results of SegAgent on the HLMAT task, comparing it with other state-of-the-art (SOTA) methods on RES datasets in Tab. 1. We compare SegAgent with both traditional and LLM-based methods. Our results demonstrate that SegAgent achieves competitive IoU scores, particularly SegAgent-LLaVA+SAM and SegAgent-Qwen+SClick, which outperforms all other methods except for GLaMM. + +We also implement a simple baseline method, Qwen box + SAM, by leveraging Qwen-VL's grounding capability to obtain a bounding box, which is then passed to SAM to generate the mask. For fair comparison, our SegAgent-LLaVA(Qwen) + SAM also uses the same box as the initial action to evaluate SegAgent's ability to refine the mask. Results show that SegAgent-LLaVA(Qwen) + SAM outperforms the Qwen box + SAM baseline across all datasets, indicating that our approach effectively refines the mask. + +An interesting observation is that although Qwen is trained on a larger dataset and generally outperforms LLaVA on standard vision-language tasks, LLaVA surpasses Qwen in mask refinement capability. We attribute this to Qwen's reliance on the Q-former structure, which appears to struggle with pixel-level tasks. Recent studies [5, 52] have reported similar limitations, suggesting that the vanilla QFormer might lose spatial locality during semantic abstraction. We envision HLMAT as a future protocol for assessing MLLM's fine-grained pixel-level visual abilities, inspiring further exploration into improving the pixel-level understanding abilities of these models. + +Additionally, when using SimpleClick as the interactive segmentation model, we observe that SegAgent-Qwen significantly outperforms SegAgent-LLaVA. This aligns with the first capability discussed in Sec. 4.2, as it plays a more dominant role here. Qwen's pretraining on large-scale visual grounding tasks gives it an advantage in locating the target during the initial interactions, leading to better overall performance. + +# 5.4. Exploration on More Complex Scenarios + +As discussed in Sec. 5.2, we validate our proposed PolicyImprovement strategy, as well as the effectiveness of PRM and Tree Search, on a more complex and accurate dataset. Here, we take the model fine-tuned on generated trajectories + +![](images/b398af5bb3872cf07ece9f048a058bcff88cea84cf5c59628d9c30d7ca76b45d.jpg) +Figure 4. Comparison of different strategies on the HRES dataset with SFT (Supervised Fine-Tuning) and apply a fixed-step greedy decoding during testing as our baseline. + +Policy-Improvement Strategies: As mentioned in Alg. 1, we use the baseline model trained on generated trajectories, and then conduct policy improvement using $\mathrm{StaR + }$ to enhance the model's performance. Although we only use a single round of $\mathrm{StaR + }$ , we find that this strategy significantly improves the model's performance. The improvement is particularly pronounced on the ThinObject5K dataset, where our model achieves an IoU of 86.57 compared to the baseline IoU of 71.45. + +PRM and Tree Search: As shown in the figure, introducing PRM as a stopping signal improves model performance compared to the baseline, which relies on fixed-step greedy decoding. Specifically, with PRM alone (equivalent to setting $K = 1$ ), the model performance on DIS increases by 2.36 and on ThinObject by 4.09. Furthermore, when we set $K = 3$ and apply our proposed Tree Search algorithm, the model performance is further enhanced, achieving an additional increase of 7.43 on DIS and 10.59 on ThinObject compared to using only PRM. These results align with our hypothesis that Tree Search effectively prevents the model from getting trapped in local optima when making uncertain actions, thus boosting performance in complex scenarios. + +# 6. Conclusion + +In this work, we introduced the Human-Like Mask Annotation Task (HLMAT), a novel segmentation paradigm that models MLLMs as multi-step decision-making agents, imitating human annotation paths in interactive segmentation. Through HLMAT, we assess and enhance MLLMs' pixel-level understanding without requiring additional model structures. We developed SegAgent by fine-tuning MLLMs on human-like trajectories and demonstrated its competitive performance across various segmentation tasks, including mask refinement and annotation filtering. Additionally, we introduced the high-quality HRES dataset and showed that decision-enhancement techniques like $\mathrm{StaR + }$ and tree search with PRM can further boost performance. Our study paves the way for future exploration of vision-centered, multi-step decision-making with MLLMs. + +# Acknowledgement + +This work is in part supported by the National Key R&D Program of China(No. 2022ZD0160101) and the National Natural Science Foundation of China (No. 62206244). This work is also supported by Ant Group Research Intern Program. + +# References + +[1] Hao Bai, Yifei Zhou, Mert Cemri, Jiayi Pan, Alane Suhr, Sergey Levine, and Aviral Kumar. Digirl: Training in-the-wild device-control agents with autonomous reinforcement learning. arXiv preprint arXiv:2406.11896, 2024. +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. +[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In CVPR, 2018. +[4] Shengcao Cao, Liang-Yan Gui, and Yu-Xiong Wang. Emerging pixel grounding in large multimodal models without grounding supervision. arXiv preprint arXiv:2410.08209, 2024. +[5] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. +[6] Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J Fleet, and Geoffrey E Hinton. A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems, 35:31333-31346, 2022. +[7] William Chen, Oier Mees, Aviral Kumar, and Sergey Levine. Vision-language models provide promptable representations for reinforcement learning. arXiv preprint arXiv:2402.02651, 2024. +[8] Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. Focalclick: Towards practical interactive image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1300-1309, 2022. +[9] Yi-Chia Chen, Wei-Hua Li, Cheng Sun, Yu-Chiang Frank Wang, and Chu-Song Chen. Sam4mllm: Enhance multimodal large language model for referring expression segmentation. In European Conference on Computer Vision, pages 323-340. Springer, 2025. +[10] Henghui Ding, Scott Cohen, Brian Price, and Xudong Jiang. Phraseclick: toward achieving flexible interactive segmentation by phrase and click. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 417-435. Springer, 2020. +[11] Chengxiang Fan, Muzhi Zhu, Hao Chen, Yang Liu, Weijia Wu, Huaqi Zhang, and Chunhua Shen. Divergen: Improving instance segmentation by learning wider data distribution with more diverse generative data. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3986-3995, 2024. +[12] 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. +[13] You Huang, Hao Yang, Ke Sun, Shengchuan Zhang, Liujuan Cao, Guannan Jiang, and Rongrong Ji. Interformer: Real-time interactive image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22301-22311, 2023. +[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. +[15] 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. arXiv preprint arXiv:2304.02643, 2023. +[16] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. +[17] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. +[18] Mengcheng Lan, Chaofeng Chen, Yue Zhou, Jiaxing Xu, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Text4seg: Reimagining image segmentation as text generation, 2024. +[19] 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. +[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. +[21] Jun Hao Liew, Scott Cohen, Brian Price, Long Mai, and Ji-ashi Feng. Deep interactive thin object selection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 305–314, 2021. +[22] Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. arXiv preprint arXiv:2305.20050, 2023. +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. +[24] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. + +[25] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. +[26] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In CVPR, 2023. +[27] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22290-22300, 2023. +[28] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. +[29] Yang Liu, Chenchen Jing, Hengtao Li, Muzhi Zhu, Hao Chen, Xinlong Wang, and Chunhua Shen. A simple image segmentation framework via in-context examples. Advances in Neural Information Processing Systems, 37:25095-25119, 2025. +[30] Qianli Ma, Haotian Zhou, Tingkai Liu, Jianbo Yuan, Pengfei Liu, Yang You, and Hongxia Yang. Let's reward step by step: Step-level reward model as the navigators for reasoning. arXiv preprint arXiv:2310.10080, 2023. +[31] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. +[32] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pretraining via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. +[33] OpenAI. Gpt-4, 2023. +[34] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. +[35] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into llm. CVPR, 2024. +[36] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14076-14088, 2024. +[37] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In European Conference on Computer Vision, pages 38-56. Springer, 2022. +[38] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M + +Answer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. +[39] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26374-26383, 2024. +[40] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. +[41] Mennatullah Siam. Pixfoundation: Are we heading in the right direction with pixel-level vision foundation models? arXiv preprint arXiv:2502.04192, 2025. +[42] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. +[43] 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. +[44] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. +[45] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. arXiv preprint arXiv:2401.06209, 2024. +[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. +[47] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In CVPR, 2022. +[48] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. +[49] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. CVPR, 2024. +[50] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. arXiv preprint arXiv:2312.17240, 2023. + +[51] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In CVPR, 2022. +[52] Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985, 2024. +[53] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601, 2023. +[54] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, 2023. +[55] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. +[56] Kaining Ying, Fanqing Meng, Jin Wang, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, et al. Mmt-bench: A comprehensive multimodal benchmark for evaluating large vision-language models towards multitask agi. arXiv preprint arXiv:2404.16006, 2024. +[57] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. +[58] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. +[59] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In CVPR, 2018. +[60] Hongyi Yuan, Zheng Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. Rrhf: Rank responses to align language models with human feedback. Advances in Neural Information Processing Systems, 36, 2024. +[61] Yuqian Yuan, Wentong Li, Jian Liu, Dongqi Tang, Xinjie Luo, Chi Qin, Lei Zhang, and Jianke Zhu. Osprey: Pixel understanding with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28202-28211, 2024. +[62] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems, 35:15476-15488, 2022. +[63] Yuexiang Zhai, Hao Bai, Zipeng Lin, Jiayi Pan, Shengbang Tong, Yifei Zhou, Alane Suhr, Saining Xie, Yann LeCun, Yi Ma, et al. Fine-tuning large vision-language models as decision-making agents via reinforcement learning. arXiv preprint arXiv:2405.10292, 2024. +[64] Jiwen Zhang, Jihao Wu, Yihua Teng, Minghui Liao, Nuo Xu, Xiao Xiao, Zhongyu Wei, and Duyu Tang. Android in the + +zoo: Chain-of-action-thought for gui agents. arXiv preprint arXiv:2403.02713, 2024. +[65] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. +[66] Zhuosheng Zhang and Aston Zhang. You only look at screens: Multimodal chain-of-action agents. arXiv preprint arXiv:2309.11436, 2023. +[67] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. In European Conference on Computer Vision, pages 74-91, 2024. +[68] Canyu Zhao, Mingyu Liu, Huanyi Zheng, Muzhi Zhu, Zhiyue Zhao, Hao Chen, Tong He, and Chunhua Shen. Disception: A generalist diffusion model for visual perceptual tasks. arXiv preprint arXiv:2502.17157, 2025. +[69] Banghua Zhu, Hiteshi Sharma, Felipe Vieira Frujeri, Shi Dong, Chenguang Zhu, Michael I Jordan, and Jiantao Jiao. Fine-tuning language models with advantage-induced policy alignment. arXiv preprint arXiv:2306.02231, 2023. +[70] Muzhi Zhu, Hengtao Li, Hao Chen, Chengxiang Fan, Weian Mao, Chenchen Jing, Yifan Liu, and Chunhua Shen. Seg-prompt: Boosting open-world segmentation via category-level prompt learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 999-1008, 2023. +[71] Muzhi Zhu, Chengxiang Fan, Hao Chen, Yang Liu, Weian Mao, Xiaogang Xu, and Chunhua Shen. Generative active learning for long-tailed instance segmentation. arXiv preprint arXiv:2406.02435, 2024. +[72] Muzhi Zhu, Yang Liu, Zekai Luo, Chenchen Jing, Hao Chen, Guangkai Xu, Xinlong Wang, and Chunhua Shen. Unleashing the potential of the diffusion model in few-shot semantic segmentation. arXiv preprint arXiv:2410.02369, 2024. +[73] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In CVPR, 2023. +[74] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. NeurIPS, 2024. \ No newline at end of file diff --git a/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/images.zip b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3cfcf32af15aa851ccc00a763ec6a0142b5cb763 --- /dev/null +++ b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4504b95bfd7ff57e2a0848376dd4db737351534072797f69a3762fab981bef9 +size 218265 diff --git a/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/layout.json b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dd21803422f65ce9cabf13551d0e9133b2524cb9 --- /dev/null +++ b/CVPR/2025/SegAgent_ Exploring Pixel Understanding Capabilities in MLLMs by Imitating Human Annotator Trajectories/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9c64ad5d4f2bceffd936462e3008f3f45396fdd4c502e366da929f93b9b5b9e +size 471885 diff --git a/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/e16e969d-226c-408e-ac45-2cd707494957_content_list.json b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/e16e969d-226c-408e-ac45-2cd707494957_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..21327c0355f4f07dc279621cb32d332e7ccc51cd --- /dev/null +++ b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/e16e969d-226c-408e-ac45-2cd707494957_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0de454325bc346c47f8e274f2a6b7eac2f9065c52c70345737c6a1da449069cb +size 90388 diff --git a/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/e16e969d-226c-408e-ac45-2cd707494957_model.json b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/e16e969d-226c-408e-ac45-2cd707494957_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5d7166e336a22a4a06ea0466ba2e44ab65ccb421 --- /dev/null +++ b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/e16e969d-226c-408e-ac45-2cd707494957_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:068e445dc830d985ac0ab99cd7ced620442bc8d56cd8cc9815cd30dbc34f005e +size 115598 diff --git a/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/e16e969d-226c-408e-ac45-2cd707494957_origin.pdf b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/e16e969d-226c-408e-ac45-2cd707494957_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4bc1060095e29e4e58c14243df54badc93d57f76 --- /dev/null +++ b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/e16e969d-226c-408e-ac45-2cd707494957_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10fca929ddc2554bb7a1dc3309926d4f9c0aa187e0295eee9d9325e2a150bfde +size 2129221 diff --git a/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/full.md b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..58ed3f664adedca3b8a0e487850c39bd0d399ad1 --- /dev/null +++ b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/full.md @@ -0,0 +1,357 @@ +# SegEarth-OV: Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images + +Kaiyu Li $^{1}$ , Ruixun Liu $^{2}$ , Xiangyong Cao $^{2,4\dagger}$ , Xueru Bai $^{6}$ , Feng Zhou $^{6}$ , Deyu Meng $^{3,4,5}$ , Zhi Wang $^{1}$ $^{1}$ School of Software Engineering, Xi'an Jiaotong University, Xi'an, 710049, China + $^{2}$ School of Computer Science and Technology, Xi'an Jiaotong University, Xi'an, 710049, China + $^{3}$ School of Mathematics and Statistics, Xi'an Jiaotong University, Xi'an, 710049, China + $^{4}$ Ministry of Education Key Laboratory of Intelligent Networks and Network Security, Xi'an Jiaotong University, Xi'an, 710049, China + $^{5}$ Pengcheng Laboratory $^{6}$ Xidian University + +![](images/4c12ca0ce3e074ee06c160cf061cc53a4e212b25043051a242e4ba904ef576be.jpg) + +![](images/c77a0d8576e6add75545b3ca3562e0995a2f17372a4c43748c9e1cc827212b1a.jpg) +Figure 1. Visualization and performance of SegEarth-OV on open-vocabulary semantic segmentation of remote sensing images. We evaluate our SegEarth-OV on 17 remote sensing datasets (including semantic segmentation, building extraction, road extraction, and flood detection tasks), and our SegEarth-OV consistently generates high-quality segmentation masks. + +![](images/e51ddbbbc291eb43ee21a2584514a771b5990a78ac0eea8f82e577dab1b1ceee.jpg) + +![](images/d1fc99ced470f990d402b4ddc264696fc14422ce6fbe7b8931c2b5f33e681397.jpg) + +![](images/0d42cf80436a41f84bc4042d7c2ad7b123080c20f1b4ee1c148dd0a559b8e736.jpg) + +![](images/d63e08da74277d2dac19fe7e87f6ddcecec2d965dc0098dc304d885438318ae5.jpg) + +![](images/313e2abca4e659a814db0a53c9bb5e5628a511af14744d3aeca9c59421fde1fe.jpg) + +![](images/9dcfa7ebc4aaa75ce3f725bdfc31778a3fc032fa31b49e9bbb52f132fdab51f5.jpg) + +![](images/78adede973ff9d1db6ca437a2b1e9932a4017b41c1f31c8bb8a2fa5a57d12e5d.jpg) + +# Abstract + +Current remote sensing semantic segmentation methods are mostly built on the close-set assumption, meaning that the model can only recognize pre-defined categories that exist in the training set. However, in practical Earth observation, there are countless new categories, and manual annotation is impractical. To address this challenge, we first attempt to introduce training-free1 open-vocabulary semantic segmentation (OVSS) into the remote sensing context. However, due to the sensitivity of remote sensing images to low-resolution features, distorted target shapes and ill-fitting boundaries are exhibited in the prediction mask. To tackle these issues, we propose a simple and universal upsampler, i.e. SimFeatUp, to restore lost spatial information of deep features. Specifically, SimFeatUp only needs to learn from a few unlabeled images, and can upsample arbitrary remote sensing image features. Furthermore, based on the observation of the abnormal response + +of patch tokens to the [CLS] token in CLIP, we propose to execute a simple subtraction operation to alleviate the global bias in patch tokens. Extensive experiments are conducted on 17 remote sensing datasets of 4 tasks, including semantic segmentation, building extraction, road detection, and flood detection. Our method achieves an average of $5.8\%$ , $8.2\%$ , $4.0\%$ , and $15.3\%$ improvement over state-of-the-art methods on the 4 tasks. Code is available at https://likyoo.github.io/SegEarth-0V. + +# 1. Introduction + +Remote sensing images have changed the way humans observe and understand the Earth. It enables us to monitor land cover/use types, respond effectively to natural disasters (e.g., fires, earthquakes, floods), gain insight into food and water resources, etc. Among the 17 Sustainable Development Goals (SDGs) $^{2}$ issued by the United Nations, remote sensing images can provide important data support for several goals including "Zero Hunger", "Clean Water + +![](images/566b62fbaa0f0f3901a9c958d2403f3748168d1ae40f4e60b0cffcb82dd6271b.jpg) +(a) MaskCLIP + +![](images/f9f58440daea50bc8b21815b0ea005c9aa46bf36c749d0423ca2f56af40e6c12.jpg) +(b) ClearCLIP +Figure 2. Limitations of state-of-the-art OVSS methods in remote sensing images. The two predictions on the left present distorted target shapes and ill-fitting boundaries (best viewed digitally with zoom, especially for the object edges). + +![](images/73428195efd77642028eb1a5d9d8ef0c1c7e411d9e05d13efee626e50b8005ba.jpg) +(c) SegEarth-OV + +![](images/9c49065108bb666ec57d7eb8fe6890c8e6164a3de62094108d072620293e47d8.jpg) +(d) GT + +and Sanitation", "Industry, Innovation and Infrastructure", "Climate Action", "Life on Land", etc. [60]. Notably, remote sensing data can be considered as a distinct modality in computer vision. It involves more diverse spatial resolutions (from centimeters to kilometers), temporal dimensions (from hours to decades), and object perspectives (overhead and oriented) than natural images. Therefore, solutions designed for other data modalities (e.g., natural images) may be sub-optimal for remote sensing data [53]. + +Recently, raw remote sensing images have been available from various sources (e.g., Landsat, Sentinel), but obtaining large-scale annotated categories is still a challenge due to countless potential categories and expensive manual costs. In addition, on the broad surface of the earth, "stuff" (e.g., grassland, forest, etc.) occupies much more area than "things" (e.g. buildings, ships, etc.) [74]. Therefore, for remote sensing images, pixel-level perception, i.e., segmentation, is applied more frequently than instance-level perception, and the demand for pixel-level annotation exacerbates the difficulty of obtaining large-scale labels. OpenStreetMap [21] is a popular solution that aims to create a freely usable, editable, and shareable map of the world. However, the completeness of the annotations in OpenStreetMap is affected by regional income levels, resulting in limited data availability [22]. The rise of vision language model (VLM) brings us new inspirations with its capabilities of open-vocabulary semantic segmentation (OVSS). However, through some exploratory experiments, we find that the solution designed for natural images is sub-optimal on remote sensing images. A notable phenomenon is the presence of distorted target shapes and ill-fitting boundaries in the prediction mask, as shown in Fig. 2 (a)(b). + +Empirically, these issues can be largely attributed to the excessively low resolution of the features [19, 76]. In the current CLIP-based OVSS paradigm, the feature maps from CLIP [51] are downsampled to $1/16th$ of the original image (ViT-B/16). Hence, in this paper, we propose a simple and general feature upsampler, i.e. SimFeatUp, which is trained with the goal of reconstructing content-invariant high-resolution (HR) features on a few unlabeled images and can upsample arbitrary remote sensing image features after training. Owing to this property of SimFeatUp, it can be used as a universal external unit for training-free + +OVSS framework. Further, CLIP is trained at the image level, it uses the [CLS] token as a representation of the entire image, and attaches global properties to the local token [48, 52, 62]. However, this global property biases local features against patch-level inference in OVSS. To effectively reduce global bias, we propose a simple subtraction operation of local patch features and global features. Extensive quantitative and qualitative experiments demonstrate the superior segmentation quality of our method over other state-of-the-art methods. + +In summary, our contributions are as follows. + +- We propose SimFeatUp, a general feature upsampler for training-free OVSS, which robustly upsamples low-resolution (LR) features and maintains semantic consistency with image content. +- We propose an extremely simple and straightforward way to alleviate the global bias problem of CLIP, i.e., executing subtraction operations of local and global tokens. +- Our proposed SegEarth-OV model can achieve state-of-the-art performance on 17 remote sensing datasets of 4 tasks, including semantic segmentation, building extraction, road extraction, and flood detection. + +# 2. Related Work + +Vision-Language Model. Recently, foundation models, especially visual language models, have energized the field of computer vision. One phenomenal advance is contrastive language-vision pretraining, i.e., CLIP [51], which elegantly bridges the gap between images and natural language. By training with massive data in a multimodal embedding space, CLIP gains strong transfer capabilities, achieving leaps in zero-shot learning and making OV learning possible [67]. Subsequently, related research has gradually emerged, from the data [10, 58, 70, 73], training [16, 35, 73] or model [33, 34] side. However, CLIP focuses only on global [CLS] tokens, and even though patch-level tokens can be generated, they are inevitably contaminated by global bias [48, 52, 62], which is detrimental to dense prediction. In addition, several remote sensing VLMs emerge, they adapt general VLMs to remote sensing contexts [37, 49, 65, 75] or mine the characteristics of remote sensing data [23, 50]. + +Supervised semantic segmentation. Semantic segmentation aims to discriminate images at the pixel level. The prediction head (aka decoder), as an essential component of segmentation models, is able to upsample LR feature maps into HR predictions. Typical prediction heads contain upsampling operators (e.g., bilinear interpolation, JBU [29]) and HR encoder features (as guidance), e.g., UNet [55], UperNet [69], Semantic FPN [27], MaskFormer [9], etc. Some works [38, 41, 78] focus on dynamic, learnable upsampling operators that make this process content-aware. FeatUp [17] constructs a model-agnostic upsam + +pling scheme that uses multi-view consistency loss with deep analogies to NeRFs [45]. However, it only explores the condition with labels. Inspired by FeatUp and built on top of it, the SimFeatUp proposed in this work is able to significantly improve OVSS without any labels. + +Open-Vocabulary Semantic Segmentation. As VLMs have shown remarkable zero-shot inference in image classification [51], which naturally extends to semantic segmentation. They empower the segmentation pipeline to recognize seen and unseen categories, and users can segment almost any category in an image using prompt vocabulary [67, 79]. We divide current CLIP-based OVSS methods into two groups: training-required and training-free. The former allows models to be trained on some base classes in a supervised or weakly supervised manner. Typically, some works [18, 42, 48, 52] try to train a localization-aware CLIP which can naturally make dense predictions, while others [11, 15, 32, 39, 71, 72] select a subset of the CLIP's pre-trained parameters and/or introduce a limited number of trainable parameters into the frozen CLIP, i.e., fine-tuning the CLIP to adapt to dense prediction on base classes. Still, training-free OVSS methods emphasize tapping into CLIP's inherent localization capabilities with limited surgery of features or structures. MaskCLIP [77] pioneers the removal of query and key projections at the attention pooling layer of CLIP's image encoder. Following it, subsequent studies [3, 30, 36, 62] adequately explore self-self attention (i.e., $q$ - $q$ , $k$ - $k$ or $\nu$ - $\nu$ self-attention), and these modifications somewhat mitigate noisy activations and spatial invariant perception of CLIP. Another stream [2, 25, 56, 59] is the two-stage method, which first generates category-agnostic mask proposals and then classifies the masks. Besides, some other foundation models (e.g. SAM [28], Stable Diffusion [54]) can be introduced to enhance the localization ability of CLIP, and these explorations also make sense [2, 31, 64]. + +Different from previous methods, we focus on the inherent characteristics of remote sensing images rather than the general attributes of natural images. The only contemporaneous work is [6], but it is training-required, like [11, 72]. Our SimFeatUp component, although it needs to be trained on a few images-only data beforehand, this process is independent of the semantic segmentation process and the trained weights can be used for almost any remote sensing data (like the foundation model in other works [2, 31]), so our method can still be seen as a training-free method. + +# 3. Preliminaries + +# 3.1. CLIP + +In ViT-based CLIP, the image encoder consists of a series of Transformer blocks. Let $X = [x_{cls}, x_1, \dots, x_{h \times w}]^\top \in \mathbb{R}^{(hw + 1, d)}$ denotes the input of the last block, where $h$ and $w$ denote the height and width of the feature map, $d$ denotes + +![](images/08b9ab25b328affae114290a784a9d0bef74755a1bf443b56ee799da637cff98.jpg) +(a) + +![](images/3924625ea5d7e36cda908b93a0785d7869857ad6512c90d08a1de223ed52d189.jpg) +(b) +Figure 3. Illustration of the proposed method. (a) is the training process of SimFeatUp. CLIP is frozen and only SimFeatUp is useful in reasoning. (b) is the reasoning process of SegEarth-OV. The LR feature maps from CLIP are upsampled by SimFeatUp and then the [CLS] token is subtracted to alleviate global bias. For better presentation, the color renderings follow [17]. + +the dimension of tokens, and $x_{cls}$ is a learnable global token and the others are local tokens from different image patches. The forward process of the last Transformer block can be formulated as follows: + +$$ +\begin{array}{l} \boldsymbol {q} = \operatorname {E m b} _ {q} (X), \boldsymbol {k} = \operatorname {E m b} _ {k} (X), \boldsymbol {v} = \operatorname {E m b} _ {v} (X), \\ \boldsymbol {y} = X + \operatorname {S A} (\boldsymbol {q}, \boldsymbol {k}, \boldsymbol {v}), \tag {1} \\ \boldsymbol {z} = \boldsymbol {y} + \operatorname {F F N} (\ln (\boldsymbol {y})) \\ \end{array} +$$ + +where $q, k, v$ denote Query, Key, and Value, respectively. Emb consists of a layer normalization (LN) layer and a linear layer, and FFN denotes the feed-forward network. SA denotes a standard self-attention module, i.e., $\mathrm{SA}(\pmb{q},\pmb{k},\pmb{v}) = \mathrm{softmax}\left(\frac{\pmb{q}\cdot\pmb{k}^{\top}}{\sqrt{d}}\right)\cdot \pmb{v}$ . Then, a projection layer maps $\pmb{z}$ to a multi-modal embedding space: + +$$ +\mathcal {O} = \operatorname {P r o j} (z), \tag {2} +$$ + +where $\mathcal{O} = [o_{cls}, o_1, \dots, o_{h \times w}]^\top \in \mathbb{R}^{(hw + 1, c)}$ denotes the output of the image encoder, $c$ denotes token dimension after the projection layer, and $c < d$ . During CLIP training, $o_{cls}$ is used for image-level learning; while during OVSS inference, $\mathcal{O}[1:hw + 1]$ is used for patch-level prediction. + +# 3.2.FeatUp + +FeatUp [17] aims to train a model-agnostic upsampler. It executes an upsampling operation on LR features $\mathcal{O}[1: h w + 1]$ from a frozen backbone network via a learnable upsampler $\sigma_{\uparrow}$ , and then reconstructs the LR features using a learnable downsampler $\sigma_{\downarrow}$ . Its critical insights can + +![](images/dc5d39782045cc704a240631c6f9dd716638d0a08be7af95c6850e9a92ee6a8c.jpg) +Figure 4. Comparison of with and without image reconstruction loss (Eq. (4)). The LR prediction is obtained directly using the output of CLIP (without bilinear interpolation). Color: building, tree, cropland, grass. + +be briefly summarized by the following loss function: + +$$ +\mathcal {L} _ {r e c} = \left\| \mathcal {O} [ 1: h w + 1 ] - \sigma_ {\downarrow} \left(\sigma_ {\uparrow} \left(\mathcal {O} [ 1: h w + 1 ]\right)\right) \right\| _ {2} ^ {2}. \tag {3} +$$ + +FeatUp instantiates $\sigma_{\uparrow}$ as stacked parameterized JBU operators [29]. The upsampled HR feature element is estimated by weighting the neighboring elements of the LR feature. For weight generation, JBU considers two factors, the similarity and distance between neighboring elements and the center element in the guidance feature, corresponding to kernel $k_{range}$ and $k_{spatial}$ . For brevity, we omit the multi-view consistency constraint in [17]. + +# 4. Method + +In the following, we introduce SegEarth-OV by first describing SimFeatUp's training, design and explaining why it is suitable for OVSS. Second, we discuss the impact of global token on dense prediction and present our method for alleviating global bias. + +# 4.1. SimFeatUp + +FeatUp provides us with an excellent training paradigm for general upsamplers. However, it lacks some considerations for the training-free setting, resulting in sub-optimal for the OVSS task, especially in remote sensing contexts. + +Image content retention. As described in Sec. 3.2, the goal of FeatUp is to minimize the original LR features and the LR features after the up-down-sampling (i.e. $\sigma_{\downarrow}(\sigma_{\uparrow}(\mathcal{O}[1:h w + 1]))$ ). Since both $\sigma_{\uparrow}$ and $\sigma_{\downarrow}$ are learnable, with such a weak constraint, the up-down-sampling process becomes a black box, and there is no guarantee that the intermediate HR features are complete and consistent with the original image in content. A direct example is shown in Fig. 4, + +where a small building in the original image is present in the LR prediction but disappears in the HR prediction (top right). To solve this issue, we introduce an additional image reconstruction loss to constrain the HR features: + +$$ +\mathcal {L} _ {i m g} = \left\| I - \operatorname {C R N} \left(\sigma_ {\uparrow} \left(\mathcal {O} [ 1: h w + 1 ]\right)\right) \right\| _ {2} ^ {2}, \tag {4} +$$ + +where $I$ denotes the input image, CRN denotes a content retention net. CRN is a very lightweight network that receives HR features as input and reconstructs the original image. Specifically, CRN consists of two 2D convolutional layers with normalization and a $\mathrm{Tanh}$ activation layer, where the $\mathrm{Tanh}$ layer is designed to constrain the output to [-1, 1], cf. VAEs [26]. Finally, the loss for training SimFeatUp consists of $\mathcal{L}_{rec}$ and $\mathcal{L}_{img}$ with a weight $\gamma$ , i.e., + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} + \gamma \mathcal {L} _ {\text {i m g}}. \tag {5} +$$ + +Which feature to upsample? FeatUp takes the final output of CLIP, i.e., $\mathcal{O}[1:h w + 1]$ in Eq. (2), as input to the upsampler. This can work well in training-based settings, e.g., linear probe [1]. However, in training-free OVSS, as described in Sec. 2, vanilla self-attention leads to inferior performance. Therefore, the current OVSS method modulates it to self-self attention, and this law also works in remote sensing images. Under this premise, the SA in Eq. (1) would be replaced by other modules, and direct upsampling of $\mathcal{O}[1:h w + 1]$ would lead to the mismatch between training and inference. Motivated by this, we propose to upsample CLIP features at an earlier layer. Specifically, we select the input of the last Transformer block of the CLIP's image encoder, i.e., $X[1:h w + 1]$ in Eq. (1). Furthermore, the high dimension of tokens in $X$ leads to a high-cost upsampler. Therefore, we retain the projection layer. Ultimately, the features $\mathcal{O}'$ which need upsampling are formulated as: + +$$ +\mathcal {O} ^ {\prime} = \operatorname {P r o j} (X [ 1: h w + 1 ]). \tag {6} +$$ + +Larger upsampling kernel. We follow the upsampling operator in FeatUp, i.e., the parameterized JBU. As mentioned in Sec. 3.2, the upsampling kernels $k_{range}$ and $k_{spatial}$ of the JBU are computed from the elements within a window in the guidance feature. The generation of $k_{range}$ and $k_{spatial}$ can be formulated as follows: + +$$ +k _ {\text {s p a t i a l}} (p, q) = \exp \left(\frac {- \| p - q \| _ {2} ^ {2}}{2 \tau_ {\text {s p a t i a l}} ^ {2}}\right), \tag {7} +$$ + +$$ +k _ {r a n g e} (p, q) = +$$ + +$$ +\operatorname {s o f t m a x} _ {(a, b) \in \Omega} \left(\frac {1}{\tau_ {\text {r a n g e}} ^ {2}} M L P (G [ i, j ]) \cdot M L P (G [ a, b ])\right), \tag {8} +$$ + +where $(p,q)$ denotes the position in the kernel. $\Omega$ denotes a window centered at $(i,j)$ in the guidance feature $G$ , which + +![](images/5740bc8de396ed7586134fee0b7ac33be5b4592f3d19b02649007a6eaaeabe42.jpg) +Figure 5. Comparison of before and after alleviating the global bias. (a) is the similarity map of patch tokens and cls tokens, some "non-building" regions also present high response, (b) is the original RGB image. Note that the right-hand histograms stretch the raw values for better presentation. + +is extracted from HR RGB image. $\tau_{\text{spatial}}$ and $\tau_{\text{range}}$ are learnable factors. In remote sensing images, unlike natural images, the size of the target presents a logarithmic scale spanning from the meter scale (e.g., trees, gardens) to the kilometer scale (e.g., forests, rangelands) [53]. Therefore, we set larger upsampling kernels to obtain a wider receptive field. Here, we expand the window size to $11 \times 11$ , compared to $7 \times 7$ in FeatUp. A possible concern is that a larger receptive field may introduce more irrelevant context, but with $k_{\text{spatial}}$ , more distant points consistently contribute lower weights, which makes it more reasonable to use larger upsampling kernels. + +Simplify. On the structural side, we simplify the components inFeatUp. InFeatUp, the parameterized JBU modules are stacked 4 times for $16 \times$ upsampling, and the parameters of each JBU module are independent. Although we fed HR features into the CRN to ensure the integrity of its content, the behavior of each JBU module is indeterminable. Therefore, in SimFeatUp, we change "JBUSTACK" to "JBU-One", i.e., only one parameterized JBU is used for upsampling. If $16 \times$ upsampling is required, then it only needs to repeat the execution 4 times. Further, "JBU-One" significantly reduces the number of trainable parameters in the upsampler and provides the possibility of upsampling arbitrary multiples. + +# 4.2. Alleviating global bias + +As described in Sec. 3.1, in the training phase of CLIP, the [CLS] token, which contains the global information of the + +whole image, is optimized with the text embedding in the multi-modal space via contrastive learning. However, in the inference phase of OVSS, [CLS] token is generally discarded and only patch tokens are used for similarity computation with the prompt vocabulary. This means that there is a gap between training and inference. Indeed, previous work [48, 52, 62] also demonstrates that: each local visual token in CLIP focuses on a wide range of positions, and the attention maps typically share similar patterns. This suggests that the global attribute is attached to the patch tokens in CLIP. This property is generally not a concern in classification task, but it significantly impairs performance in dense prediction. + +The visualization in Fig. 5 demonstrates the above elaboration. We extract the [CLS] token using CLIP for the RGB image in Fig. 5(b), and compute its similarity with the candidate text embeddings. The image is recognized as the building, which is reasonable because the building covers the maximum range in the image. Then, we calculate the similarity of the [CLS] token with patch tokens as shown in Fig. 5(a). The highly responsive regions are not only the regions with buildings, some roads and pavements are also activated, which indicates that the global bias contaminates the local patch tokens. Motivated by this observation, we propose to "subtract" some global bias from the patch token. This solution is very straightforward and simple, it can be formulated as follows. + +$$ +\hat {\mathcal {O}} = \mathcal {O} [ 1: h w + 1 ] - \lambda \mathcal {O} [ 0 ], \tag {9} +$$ + +where $\lambda$ denotes an intensity factor. $\mathcal{O}[0]$ is broadcast to the same dimension as $\mathcal{O}[1:h\omega +1]$ + +# 5. Experiments + +# 5.1. Dataset + +In remote sensing application contexts, not only multi-class semantic segmentation but also extraction of certain land cover types (e.g., buildings, roads, water bodies) is required, e.g., Google's Open Buildings project. Therefore, we select 17 typical datasets covering common semantic segmentation, building extraction, road extraction, and water body segmentation (flood detection) tasks. + +Semantic segmentation. We evaluate SegEarth-OV on 8 remote sensing semantic segmentation datasets including OpenEarthMap [68], LoveDA [63], iSAID [66], Potsdam, Vaihingen $^{4}$ , UAVid [43], UDD5 [8] and VDD [5]. Among them, the first 5 datasets consist of mainly satellite images and the last 3 consist of UAV images. They contain custom foreground classes and a background class. Detailed descriptions of these datasets can be found in Appendix 7.1. + +Table 1. Open-vocabulary semantic segmentation quantitative comparison on remote sensing datasets. Evaluation metric: mIoU. Best and second best performances are highlighted. + +
MethodsOpenEarthMapLoveDAiSAIDPotsdamVaihingenUAVidimgUDD5VDDAverage
CLIP [51]ICML'2112.012.47.514.510.310.99.514.211.4
MaskCLIP [77]ECCV'2225.127.814.531.724.728.632.432.927.2
SCLIP [62]arXiv'2329.330.416.136.628.431.438.737.931.1
GEM [3]CVPR'2433.931.617.736.524.733.441.239.532.3
ClearCLIP [30]ECCV'2431.032.418.240.927.336.241.839.333.4
SegEarth-OVOurs40.336.921.747.129.142.550.645.339.2
+ +Single-class extraction. We select 4 building extraction datasets (i.e., WHU $^{Aerial}$ [24], WHU $^{Sat.II}$ [24], Inria [44], and xBD [20]), 4 road extraction datasets (i.e., CHN6-CUG [80], DeepGlobe [14], Massachusetts [46], and SpaceNet [61]), and 1 flood detection dataset (i.e., WBS-SI $^5$ ) for the evaluation of single-class extraction. These datasets contain 1 foreground class (building, road or flood) and 1 background class. Detailed descriptions are in Appendix 7.2-7.4. + +Training dataset for SimFeatUp. SimFeatUp requires only image data for training, moreover, to avoid unfair comparisons, we use a public remote sensing image classification dataset, Million-AID [40], which is mainly collected from Google Earth. We randomly selected only 16k of these images to train SimFeatUp. + +# 5.2. Setup + +Implementation. Our implementations are based on MM-Segmentation [12] toolkit. If not specified, we use the original pretrained weights of CLIP (ViT-B/16) provided by OpenAI. For the text part, we use the OpenAI ImageNet template as input for the text encoder, e.g., "a photo of a {class name}". In addition, since the definition of certain classes may vary in some datasets, we use slight class rename tricks for all methods. For example, we rename "clutter" to "background" and "building" to {"building", "house", "roof"}, and the sub-class with the highest probability in {} will be the probability of that class. Detailed prompt class names for all datasets are listed in Appendix Tab. 7. For the image part, we resize input images with a long side of 448 and perform slide inference with a $224 \times 224$ window and 112 stride. For SimFeatUp training, we randomly crop $224 \times 224$ image patches on the original image. We use two 4090 GPUs to train 1 epoch with batch size set to 8. We retain the multi-view consistency constraint in FeatUp, and random flipping, translation and zoom are applied. For the hyper-parameters mentioned, the value of $\gamma$ is set to 0.1 and $\lambda$ is set to 0.3 for all datasets. + +Evaluation. We evaluate the semantic segmentation using the mean intersection over union (mIoU) metric. For single-class extraction, the IoU of the foreground class is used. + +Baseline. We take some lessons from natural image OVSS, which are also suitable for remote sensing scenes: we remove the FFN and residual connection of the last Transformer block, insights from [36] and [30]. In addition, the last self-attention is replaced by our modulated attention, i.e., the summation of $q - q$ , $k - k$ and $\nu - \nu$ as the weights of $\nu$ : + +$$ +\operatorname {M} - \operatorname {S A} (\boldsymbol {q}, \boldsymbol {k}, \boldsymbol {v}) = \sum_ {\boldsymbol {i} \in \{\boldsymbol {q}, \boldsymbol {k}, \boldsymbol {v} \}} \operatorname {s o f t m a x} \left(\frac {\boldsymbol {i} \cdot \boldsymbol {i} ^ {\top}}{\sqrt {d}}\right) \cdot \boldsymbol {v}. \tag {10} +$$ + +# 5.3. Comparison to State-of-the-art + +Since the proposed SegEarth-OV is a training-free method and there is no previous OVSS method designed for remote sensing images, we select 5 state-of-the-art training-free OVSS models of natural images for comparison, including vanilla CLIP [51], MaskCLIP [77], SCLIP [62], GEM [3] and ClearCLIP [30]. + +Semantic segmentation. As listed in Tab. 1, SegEarth-OV achieves the best performance on all 8 semantic segmentation datasets. SegEarth-OV achieves more than $40\%$ mIoU on 5 datasets and more than $50\%$ on the UDD5 dataset, which implies that the OVSS method is feasible in remote sensing scenarios. Compared to the previous method, SegEarth-OV achieves a performance gain of more than $5\%$ on 5 datasets and an average gain of $5.8\%$ on 8 datasets. On the iSAID dataset, the mIoU of SegEarth-OV is only $21.7\%$ , which is due to the fine-grained category delineation in this dataset, which covers 16 categories (see Appendix Tab. 7). + +Single-class extraction. In the building extraction task, the increase delivered by SegEarth-OV is more significant, as listed in Tab. 2. Considering that the "building" class occupies a small area (see Appendix Fig. 1), we evaluate the setup for larger scale images, i.e., resizing the long side of the input image to $896 \times 896$ . This setup significantly improves the IoU of Inria and xBD, which on the other hand supports our view that spatial detail preservation is essential for remote sensing OVSS. In the road extraction task, although SegEarth-OV achieves the best IoU, overall, the performance of all methods on the 4 datasets is unsatisfactory, with a best IoU of only $35.4\%$ . There may be two reasons for this phenomenon: (1) The special shape of the road makes it difficult to be extracted in a training-free OVSS manner; (2) The labels of some data are generated based on + +![](images/2bbed41f4b9a9c7ec01448a52363608613c2198740e86e064569b843b9ae8b22.jpg) +Figure 6. Qualitative comparison between different training-free OVSS methods on OpenEarthMap [68], UDD5 [8] and WHUAerial [24] datasets (best viewed digitally with zoom, especially for the edges of the object). + +Table 2. Open-vocabulary building / road / flood extraction quantitative comparison on remote sensing datasets. Evaluation metric: IoU of the foreground class, i.e. building, road or flood. Best and second best performances are highlighted. + +
MethodBuilding ExtractionRoad ExtractionFlood Detection WBS-SI
WHU AerialWHU Sat.IIInriaxBD preCHN6-CUGDeepGlobeMassachusettsSpaceNet
448 × 448:
CLIP [51]17.73.519.616.07.73.94.97.118.6
MaskCLIP [77]29.814.033.429.228.113.210.620.839.8
SCLIP [62]33.421.034.925.921.17.07.414.932.1
GEM [3]24.413.628.520.813.44.75.111.939.5
ClearCLIP [30]36.620.839.030.125.55.76.416.344.9
SegEarth-OV49.228.444.637.035.417.811.523.860.2
896 × 896:
SegEarth-OV49.9-48.943.132.820.117.229.157.9
+ +OpenStreetMap vector shapes with fixed widths attached, which are inherently imprecise. Again, the extraction of roads can generally benefit from larger size images. For the flood detection task, where "water" class features can be easily recognized, the IoU of SegEarth-OV is improved by $15.3\%$ over the previous best method, up to $60.2\%$ . Due to the small size of the original images in the WBS-SI dataset, resizing to a larger size does not result in a positive gain. + +Qualitative results. We present qualitative results for MaskCLIP, ClearCLIP, and SegEarth-OV in Fig. 6. Some observations are summarized as follows: (1) There are some incorrect category predictions in MaskCLIP, e.g., water on the road and bareland on the cropland. (2) ClearCLIP can generate correct category predictions, but lacks precise localization capability, with distorted target shapes and ill-fitting boundaries of the prediction mask. (3) SegEarth-OV is capable of generating more fine-grained masks that fit the target edges and maintain correct category discrimination. More visualizations can be found in Appendix Fig. 8-10. + +# 5.4. Ablation Study and Analysis + +Plug and play. Two key insights of this work, SimFeatUp and global bias alleviation, which can be attached to other OVSS methods as plug-and-play modules. As listed in Tab. 4, a revealing observation is that on both the Ope + +nEarthMap and WHU $^{Aerial}$ datasets, as the base capability of the model improves (from MaskCLIP to ClearCLIP), the increases delivered by our method also improve ( $\uparrow 3.3$ , $\uparrow 5.1$ , $\uparrow 8.1$ on OpenEarthMap, $\uparrow 5.6$ , $\uparrow 6.1$ , $\uparrow 14.5$ on WHU $^{Aerial}$ ). This suggests that our method has the potential to improve localization and discrimination for stronger models. + +Ablation study. To assess each of the proposed components, we perform a detailed ablation analysis, as listed in Tab. 5. FitUp (CLIP) denotes the original FitUp upsampler, which provides a $1.5\%$ improvement on OpenEarthMap but decreases the performance on $\mathrm{WHU}^{Sat.II}$ and WBS-SI (more comparisons between FitUp and the proposed method are shown in the bottom-right of Fig. 1). FitUp (MaskCLIP) denotes using $\mathbf{v}$ of self-attention as the upsampled feature, which somewhat mitigates the possible negative effects of FitUp (CLIP). In SimFitUp, the input feature $X$ of the last block is used to upsample, which presents a significant improvement in all 3 datasets. A substantial improvement is also delivered after replacing the training material for the upsampler from natural images to remote sensing images. "JBU-One" reduces the parameters by nearly $4\times$ while delivering a slight IoU gain (only $< 0.3\mathrm{M}$ parameters). The introduction of CRN with image reconstruction loss brings $1.7\%$ , $0.4\%$ , and $1.6\%$ improvement on 3 datasets, respectively. Note that the CRN only + +Table 3. Quantitative comparison of vanilla CLIP and remote sensing CLIPs (ViT-B/32). Evaluation metric: mIoU. + +
ModelsOpenEarthMapLoveDAiSAIDPotsdamVaihingenUAVidimgUDD5VDDAverage
CLIP [51]ICML'2125.727.216.240.025.131.639.739.130.6
RemoteCLIP [37]TGRS'2318.237.818.921.922.916.127.128.123.9
GeoRSCLIP [75]TGRS'2435.030.823.638.022.334.039.140.532.9
SkyCLIP [65]AAAI'2428.633.015.341.724.131.638.235.831.0
+ +Table 4. The proposed method is a plug-and-play module. “+ ours” indicates using SimFeatUp (Sec. 4.1) and alleviating the global bias (Sec. 4.2). + +
MethodsOpenEarthMapWHU AerialWBS-SI
MaskCLIP25.129.839.8
+ ours28.4↑3.335.4↑5.648.8↑9.0
SCLIP29.333.432.1
+ ours34.4↑5.139.5↑6.153.4↑21.3
ClearCLIP31.036.644.9
+ ours39.1↑8.151.1↑14.560.4↑15.5
+ +Table 5. Detailed ablation results for each component. "X"↑ indicates upsampling earlier stage features, i.e. Eq. (6). "+" RS Data" indicates using Million-AID [40] to train the upsampler, before using the images in COCO-Stuff [4]. + +
OpenEarthMapWHU Sat. IIWBS-SI
Baseline32.422.746.9
FeatUp (CLIP)33.920.239.6
FeatUp (MaskCLIP)33.825.245.9
“X”↑34.626.054.2
+ RS Data36.0 ↑1.426.256.4
+ JBU-One36.3 ↑0.326.057.1
+ Rec. Image37.6 ↑1.326.458.7
+ Alleviate Global Bias39.3 ↑1.727.959.5
+ Large Kernel40.3 ↑1.028.460.2
+ +participates in SimFeatUp's training and is discarded during inference. Global bias alleviation shows significant improvement for all 3 datasets, with an average $1.3\%$ improvement. Finally, expanding the upsampling kernel to $11 \times 11$ also exhibits consistent improvement across all datasets. + +Results on natural images. We evaluate SimFeatUp as an external unit on 3 natural image datasets: PASCAL Context59 [47], COCOStuff [4] and Cityscapes [13]. As listed in Tab. 6, after upsampling the visual features of MaskCLIP, SCLIP, and ClearCLIP using SimFeatUp, their mIoUs are improved by $5.7\%$ , $1.2\%$ , and $1.2\%$ , respectively. This reveals the potential of our method to inspire general vision. + +Remote sensing CLIPs for OVSS. We evaluate the performance of remote sensing CLIPs on OVSS, including RemoteCLIP [37], GeoRSCLIP [75], and SkyCLIP [65], which are trained on 0.8M, 5M, and 2.6M remote sensing data, respectively, without changing the model structure of + +Table 6. OVSS quantitative comparison on natural image datasets. The basic results are cited from [30]. + +
MethodsContext59 [47]Stuff [4]Cityscapes [13]Average
TCL [7]30.319.623.124.3
Reco [57]22.314.821.119.4
MaskCLIP26.416.412.618.5
+ SimFeatUp28.718.025.824.2 ↑5.7
SCLIP33.021.129.127.7
+ SimFeatUp34.122.030.528.9 ↑1.2
ClearCLIP35.923.930.029.9
+ SimFeatUp37.525.130.731.1 ↑1.2
+ +CLIP. Since these works do not provide the ViT-B/16, we uniformly use ViT-B/32. Hence, we repeat the JBU operation 5 times in SimFeatUp. For fair comparison, we train the respective SimFeatUp for each model. As listed in Tab. 3, RemoteCLIP performs suboptimally to vanilla CLIP, which indicates that a small amount of domain data diminishes the model's transfer capability. GeoRSCLIP achieves the best performance against SkyCLIP, which suggests that domain VLMs can benefit from more diverse domain-specific data. Moreover, the OVSS task effectively reflects the model's discrimination and localization capabilities, and can serve as an evaluation metric for remote sensing VLMs. + +# 6. Conclusion + +In this paper, we present SegEarth-OV, a training-free OVSS method for remote sensing images. The design of SegEarth-OV was motivated by the observation that OVSS methods currently used for natural images do not perform well on remote sensing images. The two key insights of SegEarth-OV, i.e., SimFeatUp and global bias alleviation, exhibit consistent improvements on 17 remote sensing datasets spanning the tasks of semantic segmentation, building extraction, road extraction, and flood detection, well beyond the previous state-of-the-art methods. More importantly, as the first exploration of training-free OVSS method in remote sensing scenarios, this work demonstrates that the OVSS solution is feasible in earth perception tasks even if the VLMs are pre-trained on natural images. We expect that this work will inspire more OVSS methods and more capable remote sensing VLMs, and open up new possibilities for the Earth vision community. + +Acknowledgment. This work is supported by the National Key R&D Program of China (2022YFA1004100), National Natural Science Foundation of China (No. 62272375, No. 62425113, No. 62192781), Major Key Project of PCL (PCL2024A06), and Tianyuan Fund for Mathematics of the National Natural Science Foundation of China (No. 12426105). + +# References + +[1] Guillaume Alain. Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644, 2016. 4 +[2] Luca Barsellotti, Roberto Amoroso, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Training-free open-vocabulary segmentation with offline diffusion-augmented prototype generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3689-3698, 2024. 3 +[3] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3828-3837, 2024. 3, 6, 7 +[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 8 +[5] Wenxiao Cai, Ke Jin, Jinyan Hou, Cong Guo, Letian Wu, and Wankou Yang. Vdd: Varied drone dataset for semantic segmentation. arXiv preprint arXiv:2305.13608, 2023. 5 +[6] Qinglong Cao, Yuntian Chen, Chao Ma, and Xiaokang Yang. Open-vocabulary remote sensing image semantic segmentation. arXiv preprint arXiv:2409.07683, 2024. 3 +[7] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11165-11174, 2023. 8 +[8] Yu Chen, Yao Wang, Peng Lu, Yisong Chen, and Guoping Wang. Large-scale structure from motion with semantic constraints of aerial images. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 347-359. Springer, 2018. 5, 7 +[9] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2 +[10] 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, pages 2818-2829, 2023. 2 +[11] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Catseg: Cost aggregation for open-vocabulary semantic seg- + +mentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4113-4123, 2024. 3 +[12] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.6 +[13] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 8 +[14] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. Deep globe 2018: A challenge to parse the earth through satellite images. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 172-181, 2018. 6 +[15] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 3 +[16] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 36, 2024. 2 +[17] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. Feuput: A model-agnostic framework for features at any resolution. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 4 +[18] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 3 +[19] Meng-Hao Guo, Cheng-Ze Lu, Qibin Hou, Zhengning Liu, Ming-Ming Cheng, and Shi-Min Hu. Segnext: Rethinking convolutional attention design for semantic segmentation. Advances in Neural Information Processing Systems, 35:1140-1156, 2022. 2 +[20] Ritwik Gupta, Richard Hosfelt, Sandra Sajeev, Nirav Patel, Bryce Goodman, Jigar Doshi, Eric Heim, Howie Choset, and Matthew Gaston. xbd: A dataset for assessing building damage from satellite imagery, 2019. 6 +[21] Mordechai Haklay and Patrick Weber. Openstreetmap: User-generated street maps. IEEE Pervasive computing, 7(4):12-18, 2008. 2 +[22] Benjamin Herfort, Sven Lautenbach, João Porto de Albuquerque, Jennings Anderson, and Alexander Zipf. A spatiotemporal analysis investigating completeness and inequalities of global urban building data in openstreetmap. Nature Communications, 14(1):3985, 2023. 2 +[23] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xi-uping Jia, et al. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 + +[24] Shunping Ji, Shiqing Wei, and Meng Lu. Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set. IEEE Transactions on geoscience and remote sensing, 57(1):574-586, 2018. 6, 7 +[25] Dahiyun Kang and Minsu Cho. In defense of lazy visual grounding for open-vocabulary semantic segmentation. In European Conference on Computer Vision and Pattern Recognition (ECCV), 2024. 3 +[26] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4 +[27] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6399-6408, 2019. 2 +[28] 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. 3 +[29] Johannes Kopf, Michael F Cohen, Dani Lischinski, and Matt Uytendaele. Joint bilateral upsampling. ACM Transactions on Graphics (ToG), 26(3):96-es, 2007. 2, 4 +[30] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. arXiv preprint arXiv:2407.12442, 2024. 3, 6, 7, 8 +[31] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In European Conference on Computer Vision, pages 70-88. Springer, 2025. 3 +[32] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022. 3 +[33] 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. 2 +[34] 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. 2 +[35] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23390-23400, 2023. 2 +[36] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 3, 6 +[37] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 8 + +[38] Wenze Liu, Hao Lu, Hongtao Fu, and Zhiguo Cao. Learning to upsample by learning to sample. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6027-6037, 2023. 2 +[39] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3491-3500, 2024. 3 +[40] Yang Long, Gui-Song Xia, Shengyang Li, Wen Yang, Michael Ying Yang, Xiao Xiang Zhu, Liangpei Zhang, and Deren Li. On creating benchmark dataset for aerial image interpretation: Reviews, guidances and million-aid. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:4205-4230, 2021. 6, 8 +[41] Hao Lu, Wenze Liu, Zixuan Ye, Hongtao Fu, Yuliang Liu, and Zhiguo Cao. Sapa: Similarity-aware point affiliation for feature upsampling. Advances in Neural Information Processing Systems, 35:20889-20901, 2022. 2 +[42] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning, pages 23033-23044. PMLR, 2023. 3 +[43] Ye Lyu, George Vosselman, Gui-Song Xia, Alper Yilmaz, and Michael Ying Yang. Uavid: A semantic segmentation dataset for uav imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 165:108 - 119, 2020. 5 +[44] Emmanuel Maggiori, Yuliya Tarabalka, Guillaume Charpiat, and Pierre Alliez. Can semantic labeling methods generalize to any city? the inria aerial image labeling benchmark. In 2017 IEEE International geoscience and remote sensing symposium (IGARSS), pages 3226-3229. IEEE, 2017. 6 +[45] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[46] Volodymyr Mnih. Machine Learning for Aerial Image Labeling. PhD thesis, University of Toronto, 2013. 6 +[47] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 8 +[48] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 2, 3, 5 +[49] Chao Pang, Jiang Wu, Jiayu Li, Yi Liu, Jiaxing Sun, Wei-jia Li, Xingxing Weng, Shuai Wang, Litong Feng, Gui-Song Xia, et al. H2srvlm: Towards helpful and honest remote sensing large vision language model. arXiv preprint arXiv:2403.20213, 2024. 2 + +[50] Li Pang, Datao Tang, Shuang Xu, Deyu Meng, and Xiangyong Cao. Hsigene: A foundation model for hyperspectral image generation. arXiv preprint arXiv:2409.12470, 2024.2 +[51] 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. 2, 3, 6, 7, 8 +[52] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in contrastive vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5571-5584, 2023. 2, 3, 5 +[53] Esther Rolf, Konstantin Klemmer, Caleb Robinson, and Hannah Kerner. Mission critical-satellite data is a distinct modality in machine learning. arXiv preprint arXiv:2402.01444, 2024. 2, 5 +[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3 +[55] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2 +[56] Tong Shao, Zhuotao Tian, Hang Zhao, and Jingyong Su. Explore the potential of clip for training-free open vocabulary semantic segmentation. In European Conference on Computer Vision. Springer, 2024. 3 +[57] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. Advances in Neural Information Processing Systems, 35:33754-33767, 2022. 8 +[58] Krishnakant Singh, Thanush Navaratnam, Jannik Holmer, Simone Schaub-Meyer, and Stefan Roth. Is synthetic data all we need? benchmarking the robustness of models trained with synthetic images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2505-2515, 2024. 2 +[59] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rn: Segment countless visual concepts without training endeavor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13171-13182, 2024. 3 +[60] Aysim Token, Marvin Eisenberger, Daniel Cremers, and Laura Leal-Taixe. Satsynth: Augmenting image-mask pairs through diffusion models for aerial semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27695-27705, 2024. 2 +[61] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. arXiv preprint arXiv:1807.01232, 2018. 6 +[62] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. arXiv preprint arXiv:2312.01597, 2023. 2, 3, 5, 6, 7 + +[63] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks. Curran Associates, Inc., 2021. 5 +[64] Jinglong Wang, Xiawei Li, Jing Zhang, Qingyuan Xu, Qin Zhou, Qian Yu, Lu Sheng, and Dong Xu. Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773, 2023. 3 +[65] Zhecheng Wang, Rajanie Prabha, Tianyuan Huang, Jiajun Wu, and Ram Rajagopal. Skyscript: A large and semantically diverse vision-language dataset for remote sensing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5805-5813, 2024. 2, 8 +[66] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 5 +[67] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, et al. Towards open vocabulary learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 3 +[68] Junshi Xia, Naoto Yokoya, Bruno Adriano, and Clifford Broni-Bediako. Openearthmap: A benchmark dataset for global high-resolution land cover mapping. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6254-6264, 2023. 5, 7 +[69] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018. 2 +[70] Hu Xu, Saining Xie, Xiaqing Ellen Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In The Twelfth International Conference on Learning Representations, 2024. 2 +[71] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. San: Side adapter network for open-vocabulary semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3 +[72] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945-2954, 2023. 3 +[73] Kaicheng Yang, Jiankang Deng, Xiang An, Jiawei Li, Ziyong Feng, Jia Guo, Jing Yang, and Tongliang Liu. Alip: Adaptive language-image pre-training with synthetic caption. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2922-2931, 2023. 2 +[74] D. Zanaga, R. Van De Kerchove, D. Daems, W. De Keersmaecker, C. Brockmann, G. Kirches, J. Wevers, O. Cartus, M. Santoro, S. Fritz, M. Lesiv, M. Herold, NE Tsendbazar, P. Xu, F. Ramoino, and O. Arino. Esa worldwide 10 m 2021 v200, 2022. 2 + +[75] Zilun Zhang, Tiancheng Zhao, Yulong Guo, and Jianwei Yin. Rs5m: A large scale vision-language dataset for remote sensing vision-language foundation model. arXiv preprint arXiv:2306.11300, 2023. 2, 8 +[76] Zhuo Zheng, Yanfei Zhong, Junjue Wang, Ailong Ma, and Liangpei Zhang. Farseg++: Foreground-aware relation network for geospatial object segmentation in high spatial resolution remote sensing imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[77] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3, 6, 7 +[78] Minghao Zhou, Hong Wang, Yefeng Zheng, and Deyu Meng. A refreshed similarity-based upsampler for direct high-ratio feature upsampling. arXiv preprint arXiv:2407.02283, 2024. 2 +[79] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[80] Qiqi Zhu, Yanan Zhang, Lizeng Wang, Yanfei Zhong, Qingfeng Guan, Xiaoyan Lu, Liangpei Zhang, and Deren Li. A global context-aware and batch-independent network for road extraction from vhr satellite imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 175:353-365, 2021. 6 \ No newline at end of file diff --git a/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/images.zip b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2b9131645efe32537c4477ce88b3a43a952efa03 --- /dev/null +++ b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c226704c02c96e60a84e112a566ef3b3f2fdce3b05fde32db60befa914d3b568 +size 637657 diff --git a/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/layout.json b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ecae9d1968a14411e25f5be72edd64757b25fc28 --- /dev/null +++ b/CVPR/2025/SegEarth-OV_ Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13fd3719ba28413aa9175e0e0a31d6673dc46c0552dedb8bca800aeb91d8d24e +size 475566 diff --git a/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/e4f2e2d4-e859-483d-9efe-a289f9871941_content_list.json b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/e4f2e2d4-e859-483d-9efe-a289f9871941_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44c0f5d12260fbcbbf879a9d2cd5dc728470dc63 --- /dev/null +++ b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/e4f2e2d4-e859-483d-9efe-a289f9871941_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb0629c9a302bbb8c4ae59a696746d231350367615c28b53fe8c3b0f914c48ca +size 83385 diff --git a/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/e4f2e2d4-e859-483d-9efe-a289f9871941_model.json b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/e4f2e2d4-e859-483d-9efe-a289f9871941_model.json new file mode 100644 index 0000000000000000000000000000000000000000..497b08bfd27212a990bd7050d1bad332bfb37c92 --- /dev/null +++ b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/e4f2e2d4-e859-483d-9efe-a289f9871941_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bc6eb8f45ef86ad2f96ab12725e41f3c71a6f3cfa994ea92bc86891de9eea83 +size 103906 diff --git a/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/e4f2e2d4-e859-483d-9efe-a289f9871941_origin.pdf b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/e4f2e2d4-e859-483d-9efe-a289f9871941_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..698b0abf139030f1106a61558f855ac00dfb6598 --- /dev/null +++ b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/e4f2e2d4-e859-483d-9efe-a289f9871941_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9aeb32138f338a8d18ee6d69086c003d5a555ed7c9a6ed53296afbbec527265 +size 6526179 diff --git a/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/full.md b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dbd8c029251811b749a316c47878943ce5e12aea --- /dev/null +++ b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/full.md @@ -0,0 +1,266 @@ +# SegMAN: Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation + +Yunxiang Fu* Meng Lou* Yizhou Yu† +School of Computing and Data Science, The University of Hong Kong +yunxiang@connect.hku.hk, loumeng@connect.hku.hk, yizhouy@acm.org + +# Abstract + +High-quality semantic segmentation relies on three key capabilities: global context modeling, local detail encoding, and multi-scale feature extraction. However, recent methods struggle to possess all these capabilities simultaneously. Hence, we aim to empower segmentation networks to simultaneously carry out efficient global context modeling, high-quality local detail encoding, and rich multi-scale feature representation for varying input resolutions. In this paper, we introduce SegMAN, a novel linear-time model comprising a hybrid feature encoder dubbed SegMAN Encoder, and a decoder based on state space models. Specifically, the SegMAN Encoder synergistically integrates sliding local attention with dynamic state space models, enabling highly efficient global context modeling while preserving fine-grained local details. Meanwhile, the MMSCopE module in our decoder enhances multi-scale context feature extraction and adaptively scales with the input resolution. Our SegMAN-B Encoder achieves $85.1\%$ ImageNet-1k accuracy $(+1.5\%)$ over VMamba-S with fewer parameters). When paired with our decoder, the full SegMAN-B model achieves $52.6\%$ mIoU on ADE20K $(+1.6\%)$ over SegNeXt-L with $15\%$ fewer GFLOPs), $83.8\%$ mIoU on Cityscapes $(+2.1\%)$ over SegFormer-B3 with half the GFLOPs), and $1.6\%$ higher mIoU than VWFormer-B3 on COCO-Stuff with lower GFLOPs. Our code is available at https://github.com/yunxiangfu2001/SegMAN. + +# 1. Introduction + +As a core computer vision task, semantic segmentation needs to assign a categorical label to every pixel within an image [35]. Accurate semantic segmentation demands three crucial capabilities. First, global context modeling establishes rich contextual dependencies regardless of spatial distance, enabling overall scene understanding [13]. Sec- + +![](images/51c3fad2e271e962b3996bc33fd3c9d4ef841a8f425698b5ebdeca43687023f0.jpg) + +![](images/04f67f0763cde4fe9f86c544c46a6caa8959873d929907ddb54fe9073a935a96.jpg) + +![](images/c0221811783acdd824af8726764f1033ed09282b198a59578ef62115774cb042.jpg) +Figure 1. SegMAN Encoder classification performance compared with representative vision backbones alongside semantic segmentation results of the full SegMAN model compared to prior state-of-the-art models. + +![](images/b3dc0d9af186a4795ab03b602ec1eb4a0649a8cf634de2155990532799098023.jpg) + +ond, local detail encoding endows fine-grained feature and boundary representations, crucial for differentiating semantic categories and localizing boundaries between adjacent regions with different semantic meanings [59]. Third, context modeling based on multiple intermediate scales promotes semantic representations across multiple scales, addressing intra-class scale variations while enhancing interclass discrimination [2, 57, 63]. + +Recent methods of semantic segmentation have been unable to simultaneously encapsulate all three of these capabilities effectively. For instance, VWFormer [57] introduces a varying window attention (VWA) mechanism that effectively captures multi-scale information through cross-attention between local windows and multiple windows with predefined scales. However, at higher input resolutions, the global context modeling capability of VWA diminishes because the predefined window sizes fail to + +![](images/f72c363bff6dcb884f6dd6a85788f7eb4b8a116f9d04c9d1b9c9091f7affb00b.jpg) +Figure 2. Qualitative analysis of receptive field patterns and segmentation performance for small-sized models (27M-29M parameters). Left: Visualization of effective receptive fields (ERF) on the Cityscapes validation set at $1024 \times 1024$ resolution, illustrating SegMAN's stronger global context modeling capacity in comparison to existing state-of-the-art models. Right: Segmentation maps highlighting SegMAN's superior capacity to encode fine-grained local details that are often missed by existing approaches. + +![](images/17a947d2086f3a7fa19b74bc534e8818df350732814b0a58d6191eafc65a7036.jpg) + +![](images/3c5f359fc11268f75effdcefecea150e22e04114f03cac8795bdf0a4c25d40bf.jpg) + +maintain full feature map coverage. In addition, larger windows significantly increase the computational cost as self-attention exhibits quadratic complexity. Likewise, EDAFormer [60] proposes an embedding-free spatial reduction attention (SRA) mechanism to implement global attention efficiently and an all-attention decoder for global and multi-scale context modeling. However, fine-grained details are lost due to its reliance on downsampled features for token-to-region attention. Moreover, the absence of dedicated local operators in the feature encoder limits fine-grained feature learning [56, 60]. + +To present a more intuitive understanding of the aforementioned issues, we visualize the effective receptive field (ERF) maps and segmentation maps of recent state-of-the-art models [56, 57, 60] with 27M-29M parameters. As depicted in Figure 2 (Left), the ERF of VWFormer [57] and EDAFormer [60] has limited feature map coverage when processing high-resolution Cityscapes images. Meanwhile, the segmentation maps in Figure 2 (Right) reveal that existing methods have a limited ability to recognize fine-grained local details, which we attribute to insufficient local feature modeling in these architectures. + +In this work, we aim to encapsulate omni-scale context modeling for a varying input resolution within a semantic segmentation network. As a result, we propose SegMAN, a novel segmentation network capable of carrying out efficient global context modeling, high-quality local detail encoding, and rich context representation at diverse scales + +simultaneously. Our SegMAN introduces two novel subnetworks: (1) a hybrid feature encoder that integrates Local Attention and State Space models (LASS) in the token mixer, and (2) a decoder with a Mamba-based Multi-Scale Context Extraction (MMSCopE) module. In the SegMAN Encoder, LASS leverages a two-dimensional state space model (SS2D) [28] for global context modeling and neighborhood attention [19] (Natten) for local detail encoding, both functioning in linear time. The dynamic global scanning in our method always covers the full feature map, enabling robust global context modeling across varying input resolutions and all encoder layers. Meanwhile, by performing local attention within a sliding window, Natten exhibits an outstanding capability in fine-grained local detail encoding. Our decoder complements the encoder with robust feature learning across multiple scales. The core MMSCopE module first aggregates semantic context from multiple regions of the encoder feature map. Then it uses SS2D to scan each regionally aggregated feature map to learn multiscale contexts at linear complexity. For higher efficiency, the feature maps at different scales are reshaped to the same resolution and concatenated so that scanning only needs to be performed on a single feature map. Note that our method differs from existing methods in two aspects. First, unlike multi-kernel convolution [2] and multi-scale window attention [57], our design adaptively scales with the input resolution. Second, by performing SS2D-based scanning over entire multi-scale feature maps, our method better preserves + +fine-grained details that are typically compromised by pooling operations [36, 63] or spatial reduction attention [60] in existing methods. + +As shown in Figure 1, our SegMAN Encoder shows superior performance compared to representative vision backbones on ImageNet-1k. When paired with our novel decoder, the full SegMAN model establishes new state-of-the-art segmentation performance on ADE20K [65], Cityscapes [8], and COCO-Stuff-164k [1]. To summarize, our contributions are threefold: + +- We introduce a novel encoder architecture featuring our LASS token mixer that synergistically combines local attention with state space models for efficient global context modeling and local detail encoding. +- We propose MMSCopE, a novel decoder module that operates on multi-scale feature maps that adaptively scale with the input resolution, surpassing previous approaches in both fine-grained detail preservation and omni-scale context learning. +- We demonstrate through comprehensive experiments that SegMAN, powered by LASS and MMSCopE, establishes new state-of-the-art performance while maintaining competitive computational efficiency across multiple challenging semantic segmentation benchmarks. + +# 2. Related work + +Segmentation encoders. Prior research in semantic segmentation architectures typically comprises two key components: an encoder for feature extraction from input images and a decoder for multi-scale context learning and pixel-wise classification. While conventional approaches often adapt popular classification networks such as ResNet [20], ConvNeXt [30], PVT [49], and SwinTransformer [29], along with advanced backbones [31-33, 41] as encoders, improvements in classification performance do not necessarily translate to enhanced segmentation capabilities [21], primarily due to the requirement for fine-grained spatial information in segmentation tasks. This observation has motivated the development of specialized segmentation-oriented encoders [18, 39, 45, 53, 56, 58, 60, 64]. In this paper, we introduce a novel segmentation encoder that integrates local attention with dynamic State Space Models (Mamba) [16]. In contrast to previous works, our design enables the simultaneous modeling of fine-grained details and global contexts while maintaining linear time complexity. + +Segmentation decoders. Multi-scale feature learning plays a fundamental role in semantic segmentation, leading to numerous innovations in decoder architectures. Seminal works established various approaches to multi-scale feature aggregation, including the adaptive pooling strategy of PSPNet [63], atrous convolutions in the DeepLab series [2-4], and feature pyramid network (FPN) based decoders + +such as UPerNet [54] and Semantic FPN [23]. Contemporary architectures have advanced both efficiency and performance through diverse approaches to multi-scale feature fusion, including MLP-based methods (Segformer [56]), Matrix decomposition modules (SegNeXt [15, 18]), Transformers (FeedFormer [43], VWFormer [57]), and convolutions (CGRSeg [36]). An approach relevant to ours is VWFormer [57], which employs cross-attention between local windows and multiple larger windows with predefined scales to capture multi-scale contexts. However, its reliance on fixed window scales limits its global context modeling capability: as the input resolution increases, the predefined windows fail to maintain complete feature map coverage. In contrast, our proposed decoder adaptively extracts context features at different scales while maintaining full feature map coverage regardless of the input resolution. + +Vision Mamba. Dynamic state space models represented by Mamba [16] have demonstrated promising performance in sequence modeling. Mamba combines structured state space models [17, 44] with selective scanning, enabling global context modeling with dynamic weights and linear time complexity. These favorable properties have led to promising results on vision tasks. For example, VMamba [28] and ViM [67] are Mamba-based vision backbone networks. Other works have explored Mamba for medical imaging [26, 40], 3D point clouds [27, 62], and image/video generation [12, 14, 46]. In this paper, we leverage Mamba to learn global contexts in the encoder and multiscale contexts in the decoder for segmentation. + +# 3. Method + +# 3.1. Overall Architecture + +As shown in Figure 3, our proposed SegMAN consists of a newly designed feature encoder and decoder. Specifically, we propose a novel hybrid feature encoder based on core mechanisms from both Transformer and dynamic State Space Models [16], and a Mamba-based Multi-Scale Context Extraction module (MMSCopE) in the decoder. The hierarchical SegMAN Encoder integrates Neighborhood Attention (Natten) [19] and the 2D-Selective-Scan Block (SS2D in VMamba) [28] within the token mixer termed Local Attention and State Space (LASS), enabling comprehensive feature learning at both global and local scales across all layers while maintaining linear computational complexity. Meanwhile, MMSCoPE dynamically and adaptively extracts multi-scale semantic information by processing feature maps at varying levels of granularity using SS2D [28], with the scales adaptively adjusted according to the input resolution. Collectively, the SegMAN Encoder injects robust global contexts and local details into a feature pyramid, where feature maps are progressively downsampled and transformed to produce omni-scale fea + +![](images/9814e5c48a327bbb74e528d1a4cf1ee1fe640fb3b411388efb26af3b9f9dd8d9.jpg) +Figure 3. Overall Architecture of SegMAN. (a) Hierarchical SegMAN Encoder. (b) LASS for modeling global contexts and local details with linear complexity. (c) The SegMAN Decoder. (d) The MMSCopE module for multi-scale contexts extraction. + +![](images/15ed16b7245d6b9a34af397fd2d8300c9044612461ce46d752f10e2cc58d5b0c.jpg) + +tures at different pyramid levels. This feature pyramid is then fed into MMSCopE, where features are further aggregated spatially as well as across scales, resulting in comprehensive multi-scale representations that can be used for dense prediction. + +# 3.2. Feature encoder + +Overview. As shown in Figure 3, our SegMAN Encoder is a standard four-stage network [29, 30], where each stage begins with a strided $3 \times 3$ convolution for spatial reduction, followed by a series of LASS Blocks. Each LASS Block includes pre-layer normalization, a novel LASS module, and a feedforward network (FFN) [10, 19, 29]. Our proposed LASS represents the first integration of local self-attention (Natten [19]) and state space models (SS2D [28]) for semantic segmentation, capturing local details and global contexts simultaneously with linear time complexity. Ablation studies (Table 5) confirm the necessity of both components. + +Local attention and state space module. Long-range dependency modeling is crucial for semantic segmentation as they enable comprehensive context learning and scene understanding, as demonstrated by the success of Transformer-based semantic segmentation models [34, 39, 56, 60, 64]. To efficiently learn global contexts across all network layers, we leverage the SS2D block in VMamba [28]. This variant of a dynamic state space model named Mamba [16] adopts four distinct and complementary scanning paths, enabling each token to integrate informa + +
SegMAN EncoderChannelsBlocksParams (M)FLOPs (G)
Tiny[32, 64, 144, 192][2, 2, 4, 2]3.50.65
Small[64, 144, 288, 512][2, 2, 10, 4]25.54.05
Base[80, 160, 364, 560][4, 4, 18, 4]45.09.94
Large[96, 192, 432, 640][4, 4, 28, 4]81.016.8
+ +Table 1. Configurations of the three SegMAN Encoder variants. FLOPs were measured at the $224 \times 224$ resolution. + +tion from all other tokens in four different directions. + +However, while Mamba [16] achieves linear time complexity by compressing the global context of each channel into a fixed-dimensional hidden state [16], this compression inherently leads to information loss, in particular, loss of fine-grained local spatial details, that are crucial in segmentation to accurately localize region boundaries. To this end, we utilize Neighborhood Attention (Natten) [19], where a sliding-window attention mechanism localizes every pixel's attention span to its immediate neighborhood. This approach retains translational equivalence, effectively captures local dependencies, and has linear time complexity. In practice, we serially stack Natten and SS2D, and a shortcut around SS2D is employed to merge local and global information. The merged output is then fed into a $1 \times 1$ convolution layer for further global-local feature fusion. + +Network architecture. Our SegMAN Encoder produces a feature pyramid. The spatial resolution of $F_{i}$ is $\frac{H}{2^{1 + i}} \times \frac{W}{2^{1 + i}}$ , where $H$ and $W$ are the height and width + +of the input image, respectively. We design three SegMAN Encoder variants with different model sizes, each with computational complexity comparable to some existing semantic segmentation encoders [18, 56, 60]. As the feature map resolution at Stage 4 is $(\frac{H}{32} \times \frac{W}{32})$ , global self-attention [48] becomes computationally feasible. Therefore we replace SS2D with this more powerful global context modeling mechanism. The configurations of these models are presented in Table 1. + +# 3.3. Feature Decoder + +Overview. As shown in Figure 3, our proposed decoder first aggregates features at various levels of abstraction (from low-level $F_{2}$ to high-level $F_{4}$ ) to obtain a feature map $F$ of size $\frac{H}{8} \times \frac{W}{8}$ , as in prior works [56, 57]. Specifically, each feature map $F_{i}$ is projected to a lower dimension using a $1 \times 1$ convolution followed by batch normalization and ReLU, collectively denoted as 'Conv' in Figure 3. Subsequently, $F_{3}$ and $F_{4}$ are upsampled using bilinear interpolation to match the spatial dimensions of $F_{2}$ . The upsampled features $F_{up3}$ and $F_{up4}$ are concatenated together with $F_{2}$ and passed through a Conv layer for channel dimension reduction, resulting in $F$ with size $\frac{H}{8} \times \frac{W}{8} \times C$ , where $C$ is the number of channels. The feature map $F$ is then processed by our proposed MMSCopE module to extract rich multi-scale context features, producing a new feature map $F'$ . We then add the feature map $F'$ to $F_{2}$ , $F_{up3}$ , and $F_{up4}$ to enrich their multi-scale contexts. The final prediction pathway concatenates $F'$ with these context-enhanced features and the global pooling result on $F_{up4}$ , passes the combined features through a two-layer MLP, and the result is upsampled to the input resolution. + +Mamba-based multi-scale context extraction. To effectively extract contexts at different scales, we propose a Mamba-based Multi-Scale Context Extraction (MMSCopE) module. As shown in Figure 3, we apply strided convolutions to downsample the feature map $F$ , generating $F_{s2}$ and $F_{s4}$ with a resolution equal to $1/2$ and $1/4$ of that of $F$ , respectively. The motivation behind this is to obtain multiple regionally aggregated contexts through two derived feature maps: $F_{s2}$ , which aggregates features from $3 \times 3$ neighborhoods using a convolution with stride 2, and $F_{s4}$ , which aggregates features from $5 \times 5$ neighborhoods with stride 4. Then, motivated by the observation that Mamba scans each channel independently, we use a single Mamba scan to simultaneously extract contexts from the three feature maps $F$ , $F_{s2}$ , and $F_{s4}$ . The main idea is to concatenate these feature maps along the channel dimension and perform a single Mamba scan, which is more efficient on the GPU than processing each feature map separately. However, the spatial dimensions of these feature maps are not consistent, which prevents direct concatenation. To resolve this, we employ lossless downsampling via the Pixel + +Unshuffle [42] operation to reduce the spatial dimensions of $F$ and $F_{s2}$ to match that of $F_{s4}$ , which is $\frac{H}{32} \times \frac{W}{32}$ . Specifically, this operation rearranges non-overlapping $4 \times 4$ patches from $F$ and $2 \times 2$ patches from $F_{s2}$ into their respective channel dimension, increasing the channel depth by a factor of 16 and 4 respectively, while preserving complete spatial information. The transformed feature maps retain information at their original scales. To reduce computational costs, we project each feature map to a fixed channel dimension $C$ through $1 \times 1$ convolutions. This step assigns equal importance to the context information from each scale. Then we concatenate the projected feature maps along the channel dimension, and pass the combined feature map to SS2D [28] to achieve multi-scale context extraction in a single scan. The resulting feature map is of size $\frac{H}{8} \times \frac{W}{8} \times 3C$ , where the three segments along the channel dimension correspond to earlier feature maps at three distinct scales ( $F$ , $F_{s2}$ , and $F_{s4}$ ). To facilitate context mixing across scales, we append a $1 \times 1$ convolution layer after the SS2D scan. Finally, we use bilinear interpolation to up-sample the mixed features to $1/8$ of the input resolution, and project the channel dimension down to $C$ using another $1 \times 1$ convolution. This yields the new feature map $F'$ with mixed multi-scale contexts. + +Multi-scale fusion. Instead of directly using the feature map $F'$ to predict pixel labels, we further exploit stage-specific representations from the encoder by adding $F'$ to $F_2$ , $F_{up3}$ , and $F_{up4}$ . This effectively injects multi-scale contexts into stage-specific feature maps, which inherits information from various levels of abstraction. Next, we concatenate the resulting context-enhanced feature maps from each stage, the multi-scale context feature $F'$ , and the global average pooling result on $F_{up4}$ . For pixel-wise label prediction, the diverse concatenated features are fused through a two-layer MLP, and further bilinearly interpolated to restore the original input resolution. + +# 4. Experiments + +# 4.1. Setting + +Datasets. Following previous works [18, 56], we firstly pre-train our SegMAN Encoder on the ImageNet-1K dataset [9]. Then, we evaluate the semantic segmentation performance on three commonly used datasets: ADE20K [65], a challenging scene parsing benchmark consisting of 20,210 images annotated with 150 semantic categories; Cityscapes [8], an urban driving dataset comprising 5,000 high-resolution images with 19 semantic categories; and COCO-Stuff-164K [1], which contains 164,000 images labeled across 172 semantic categories. + +Implementation details. We conducted ImageNet pretraining using the Timm library [52] with hyperparameters and data augmentation settings identical to those of Swin + +
MethodParams(M)ADE20KCityscapesCOCO-Stuff
GFLOPmIoUGFLOPmIoUGFLOPmIoU
Segformer-B0 [56]3.88.437.4125.576.28.435.6
SegNeXt-T [18]4.37.741.161.678.97.738.7
VWFormer-B0 [57]3.75.838.9112.477.25.836.2
EDAFormer-T [60]4.95.842.3151.778.75.840.3
CGRSeg-T† [36]9.44.842.565.578.34.840.4
SegMAN-T6.46.243.052.580.36.241.3
Swin UperNet-T [29]60.0236.044.4----
ViT-CoMer-S [53]61.4296.146.5----
OCRNet [61]70.5164.845.6----
Segformer-B2 [56]27.562.446.5717.181.062.444.6
MaskFormer [5]42.055.046.7----
Mask2Former [6]47.074.047.7----
SegNeXt-B [18]27.634.948.5279.182.634.945.8
FeedFormer-B2 [43]29.142.748.0522.781.5--
VWFormer-B2 [57]27.446.648.1415.181.746.645.2
EDAFormer-B [60]29.432.149.0605.981.632.145.9
CGRSeg-T† [36]35.716.547.3199.780.216.545.2
SegMAN-S29.425.351.3218.483.225.347.5
Segformer-B3 [56]47.379.049.4962.981.779.045.5
SegNeXt-L [18]48.970.051.0577.583.270.046.5
VWFormer-B3 [57]47.363.350.3637.182.463.346.8
SegMAN-B51.858.152.6479.083.858.148.4
Swin UperNet-B [29]121.0261.048.1----
ViT-CoMer-B [53]144.7455.448.8----
Segformer-B5 [56]84.7110.351.01149.882.4110.346.7
VWFormer-B5 [57]84.696.152.01139.682.896.148.0
SegMAN-L92.497.153.2796.084.297.148.8
+ +Table 2. Comparison with state-of-the-art semantic segmentation models on ADE20K, Cityscapes, and COCO-Stuff-164K. FLOPs are calculated at $512 \times 512$ (ADE20K, COCO-Stuff) and $2048 \times 1024$ (Cityscapes) resolutions. + +
ModelsParams(M)GFLOPAcc
MiT-B0 [56]3.80.6070.5
EFT-T [60]3.70.6072.3
MSCAN-T [18]4.20.8975.9
SegMAN-T Encoder3.50.6576.2
MiT-B2 [56]244.081.6
EFT-B [60]264.282.4
MSCAN-B [18]274.483.0
Swin-T [29]284.581.2
ConvNeXt-T [30]284.582.1
InternImage-T [51]305.083.5
ViM-S [67]26-81.6
VMamba-T [28]294.982.6
SparX-Mamba-T [32]275.283.5
SegMAN-S Encoder264.184.0
MiT-B3 [56]456.983.1
MSCAN-L [18]459.183.9
Swin-S [29]508.783.2
ConvNeXt-S [30]508.783.1
InternImage-S [51]508.084.2
VMamba-S [28]508.783.6
SparX-Mamba-S [32]479.384.2
SegMAN-B Encoder459.985.1
MiT-B5 [56]8211.883.8
Swin-B [29]8815.483.5
ConvNeXt-B [30]8915.483.8
InternImage-B [51]9716.784.9
ViM-B [67]98-83.2
VMamba-B [28]8915.483.9
SparX-Mamba-B [32]8415.984.5
SegMAN-L Encoder8116.885.5
+ +Table 3. Comparison with state-of-the-art vision backbones on ImageNet-1K. FLOPs are calculated at $224 \times 224$ resolution. + +Transformer [29]. Regarding semantic segmentation, we trained all models using the MMSegmentation library [7], combining the pre-trained SegMAN Encoder with a randomly initialized decoder. Following the Segformer training protocol [56], we employed standard data augmentation techniques, including random horizontal flipping and random scaling (ratio of 0.5 to 2.0). We implemented dataset-specific random cropping with dimensions of $512 \times 512$ for ADE20K and COCO-Stuff, and $1024 \times 1024$ for Cityscapes. We trained the models using the AdamW optimizer for 160K iterations with batch sizes of 16 for ADE20K and COCO-Stuff, and 8 for Cityscapes. The learning rate is set to 1e-6 with 1500 warmup iterations. We evaluated segmentation performance using mean intersection over union (mIoU) and computed FLOPs using the fvcore library [11]. All models were trained on 8 NVIDIA H800 GPUs. + +# 4.2. Comparison with state-of-the-art methods + +This section presents a comprehensive comparison of our proposed SegMAN with state-of-the-art segmentation methods on three widely used datasets. + +Baselines. We compare our method with other strong competitors including EDAFormer [60], VWFormer [57] and CGRSeg [36], Segformer [56], and SegNeXt [18]. + +Note that the Efficientformer-v2 [25] used in CGRSeg [36] significantly benefits from knowledge distillation from RegNetY-16GF [38] (83.6M parameters, 15.9 GFLOPs) and extended training duration of 450 epochs compared to the conventional 300 epochs, providing a stronger encoder network compared to other methods. To ensure a fair comparison, we retrained Efficientformer-v2 on ImageNet-1K using the same training configurations as Swin [29]. We report results for CGRSeg by fine-tuning this retrained encoder following official implementations. + +Results. Table 2 shows the superior performance of SegMAN across all three datasets: ADE20K, Cityscapes, and COCO-Stuff-164K. Among lightweight models, SegMAN-T achieves the best performance, with $43.0\%$ mIoU on ADE20K, $80.3\%$ mIoU on Cityscapes, and $41.3\%$ mIoU on COCO-Stuff, while maintaining comparable computational costs. Our SegMAN-S model achieves significant performance improvements compared to state-of-the-art methods. On the ADE20K benchmark, it surpasses the recently introduced EDAFormer-B [60] by $2.3\%$ mIoU, and even outperforms the larger SegNeXt-L [18] model, which has 48.9M parameters, by $0.3\%$ mIoU. The performance gains extend to the COCO-Stuff dataset, where SegMAN-S demonstrates improvements of $1.6\%$ mIoU over EDAFormer-B and $1.0\%$ + +
ModelParams (M)GFLOPsFPSmIoU
Segformer-B0 [56]3.8125.521.876.2
SegNext-T [18]4.361.625.878.9
VWFormer-B0 [57]3.7112.421.177.2
EDAFourner-T [60]4.9151.712.778.7
CGRSeg-T [36]9.365.527.678.3
SegMAN-T6.452.534.980.3
Segformer-B2 [56]27.5717.110.281.0
SegNext-B [18]27.6279.114.182.6
VWFormer-B2 [57]27.4415.19.8781.7
EDAFourner-B [60]29.4605.96.681.6
CGRSeg-B [36]35.6199.011.880.2
SegMAN-S29.4218.413.883.2
Segformer-B3 [56]46.3962.96.781.7
SegNext-L [18]48.9577.510.183.2
VWFormer-B3 [57]47.3637.16.382.4
SegMAN-B51.8479.07.283.8
Segformer-B5 [56]84.71460.43.581.7
VWFormer-B5 [57]84.61139.63.782.4
SegMAN-L92.4796.05.284.2
+ +mIoU over SegNeXt-L. Notably, SegMAN-S achieves these superior results while reducing computational complexity by over $20\%$ , requiring only 25.3 GFLOPs compared to EDAFormer-B. On Cityscapes, SegMAN-S achieves $83.2\%$ mIoU, surpassing EDAFormer-B by $1.6\%$ . Moreover, SegMAN-B surpasses previous best performing models by $1.6\%$ on both ADE20K and COCO-Stuff while incurring lower computational cost. On Cityscapes, SegMAN-B improves SOTA by $0.6\%$ mIoU using $17\%$ less GFLOPs. Similarly, our largest SegMAN-L model achieves $+1.2\%$ , $+1.0\%$ , and $+0.8\%$ gains over prior SOTA on ADE20K, Cityscapes, and COCO-Stuff, respectively. The consistent performance gains across different datasets and model scales validate the effectiveness of our proposed SegMAN, which can simultaneously capture global contexts, local details, and multi-scale information through its LASS-based encoder and MMSCopE-based decoder. + +Speed Analysis. We benchmark inference speed on Cityscapes [8] using an NVIDIA L40S GPU, averaging Frames Per Second (FPS) over 128 steps with batch size 2. As shown in Table 4, SegMAN-T achieves $+1.6\%$ mIoU at approximately 3 times the speed of EDAFormer-T, highlighting an excellent balance between performance and speed. + +# 4.3. Comparison with state-of-the-art backbones + +In Table 3, we evaluate the classification performance of our proposed SegMAN Encoder on ImageNet-1K [9], comparing against two categories of architectures: state-of-the-art segmentation encoders (MiT [56], MSCAN [56], EFT [60]) and representative vision backbones (Swin + +Table 4. Comparison of model complexity and inference speed on Cityscapes [8] (2048 × 1024 resolution). + +
Model VariantParams (M)GFLOPsFPSAccmIoU
SegMAN-S Encoder25.521.413984.051.3
Replace LASS
MaxViT [47]24.629.89683.5 (-0.5)47.2 (-4.1)
ACMix [37]24.919.310483.1 (-0.9)48.6 (-2.7)
PVT [49]29.522.016982.8 (-1.2)49.1 (-2.2)
BiFormer [66]25.230.59782.9 (-1.1)48.8 (-2.5)
LASS Design
Parallel Natten and SS2D25.521.412983.8 (-0.2)48.9 (-2.4)
w/o SS2D residual26.121.413283.7 (-0.3)50.2 (-1.1)
Stage 4 Attn → SS2D30.021.812183.7 (-0.3)50.6 (-0.7)
SS2D within LASS
Replace SS2D w SRA [49]26.022.314283.7 (-0.3)50.6 (-0.7)
Replace SS2D w L-Attn [22]27.621.815083.6 (-0.4)49.5 (-1.8)
Remove SS2D25.020.916283.6 (-0.4)47.4 (-3.9)
Natten within LASS
Replace Natten w Conv25.820.716083.1 (-0.9)49.5 (-1.8)
Replace Natten w Swin [29]30.021.614783.5 (-0.5)50.3 (-1.0)
Remove Natten [19]24.925.910083.6 (-0.4)49.8 (-1.5)
+ +Table 5. Ablation studies and comparison of token mixers for SegMAN encoder. FLOPs and FPS are calculated or measured at $512 \times 512$ resolution using the encoder. + +Transformer [29], ConvNeXt [30], InternImage [51] ViM [67], VMamba [28], SparX-Mamba [32]). SegMAN Encoder variants consistently achieve superior classification accuracy in comparison to architectures of similar computational complexity. Specifically, SegMAN-S Encoder significantly surpasses ConvNeXt-T and MSCAN-B by $1.9\%$ and $1.0\%$ in Top-1 accuracy, respectively. SegMAN-B Encoder demonstrates similar improvements over Swins and MSCAN-L by $1.9\%$ and $1.2\%$ , respectively. The promising performance across model variants demonstrates the benefits of integrating local attention with state space models to effectively capture both fine-grained details and global contextual information. A more detailed comparison can be found in the supplementary material. + +# 4.4. Ablation studies on SegMAN encoder + +We systematically evaluate SegMAN Encoder's architectural design on ImageNet-1K [9] and ADE20K [65]. Our LASS token mixer integrates Neighborhood Attention (Natten [19]) and 2D-Selective-Scan (SS2D), with a residual connection added to SS2D to fuse global and local contexts. In Stage 4 blocks, we replace SS2D with global attention for enhanced high-level feature modeling. + +To validate these architectural decisions, we conduct comprehensive ablation studies examining: (1) comparative performance against alternative token mixers, (2) design choices of LASS, and (3) the impact of individual components within the token mixer. In all experiments, including token mixer replacements, we maintain consistent model complexity by adjusting channel dimensions to achieve comparable parameter counts and computational costs. Table 5 presents our findings. + +Alternative token mixers. We evaluate the effectiveness of LASS by replacing it with recent token mixers + +
ModelParams (M)GFLOPsmIoU (%)
UPerNet [54]27.138.750.4
MLP Decoder [56]25.032.750.3
Ham Decoder [18]25.329.550.4
VWFormer Decoder [57]28.240.750.5
CGRHead [36]44.224.550.3
EDAFormer Decoder [60]28.526.250.2
SegMAN-S Decoder29.425.351.3
+ +Table 6. Comparison of complexity and accuracy of segmentation decoders. FLOPs are calculated at $512 \times 512$ resolution. + +(MaxViT [47], ACMix [37], PVT [49], and BiFormer [66]). Note that we implement MaxViT's Block-SA and Grid-SA modules in series to serve as a token mixer. As shown in Table 5, our LASS achieves consistent and substantial improvements across both classification and segmentation tasks while maintaining competitive computational costs. The performance gains stem from the complementary nature between Natten, which models fine-grained local features, and SS2D, which efficiently models global contexts. + +Alternative LASS mixer designs. We explore alternative methods to integrate SS2D and Natten together in LASS. Results indicate that parallel arrangement of Natten and SS2D and removing the residual connections for SS2D lead to substantial performance drops in both classification accuracy and segmentation mIoU. + +SS2D and Natten. We analyze the two key components of LASS through systematic ablations. When SS2D is replaced with alternative global attention mechanisms, performance consistently decreases: SRA [49] $(-0.3\%)$ accuracy, $-0.7\%$ mIoU) and Linear Attention [22] $(-0.4\%)$ accuracy, $-1.8\%$ mIoU). Likewise, replacing Natten with either convolutions $(-0.9\%)$ accuracy, $-1.8\%$ mIoU) or Shifted-Window Attention [29] $(-0.5\%)$ accuracy, $-1.0\%$ mIoU) leads to significant performance degradation, demonstrating the effectiveness of our component choices. + +# 4.5. Ablation studies on SegMAN decoder + +We investigate both the overall effectiveness of our proposed decoder and the performance of individual components of the MMSCopE module using the ADE20K dataset [65]. For a fair comparison, we ensure that different model variants have comparable computational complexity by adjusting the number of channels. Here, all ablations are conducted using the SegMAN-S decoder. + +Decoder comparisons and encoder generalization. Table 6 evaluates decoder efficacy under controlled conditions, where all models employ our SegMAN-S Encoder. We benchmark against decoders from representative methods [18, 36, 56, 57, 60] with matched model sizes and computational budgets—UPerNet's channel dimensions were adjusted to ensure equivalent GFLOPs. Our decoder achieves superior mIoU. + +Notably, integrating our SegMAN Encoder with third + +
Model VariantParams (M)GFLOPsFPSmIoU
SegMAN-S29.425.312851.3
Scan Strategy
Independent scans28.725.965 (-63)51.5 (+0.2)
Multi-scale feature
w/o 16×1628.924.8130 (+2)51.0 (-0.3)
w/o 32×3228.524.5132 (+4)50.8 (-0.5)
w/o 64×6428.324.3131 (+3)50.6 (-0.7)
64×64 only26.223.6137 (+9)50.9 (-0.4)
Compression
w/o compression30.326.798 (-30)51.0 (-0.3)
Channels 1C:2C:4C31.225.3124 (-4)50.8 (-0.5)
Architecture Components
w/o FFN28.424.8129 (+1)51.2 (-0.1)
+ +Table 7. Ablation studies on the MMSCopE module in SegMAN decoder. FLOPs and FPS are measured at $512 \times 512$ resolution. + +party decoders yields significant performance gains: $+2.1\%$ for VWFormer, $+1.9\%$ for SegNeXt, and $+1.2\%$ for EDAFormer over their original encoder-decoder configurations (Table 2). This highlights our encoder's generalization and plug-and-play capability. + +Ablation studies on the MMSCopE module. Table 7 presents the impact of various components within the MMSCopE module: (1) Switching to independently scanning each feature map instead of our proposed simultaneous scanning strategy significantly reduces FPS from 128 to 65; (2) The inclusion of multi-scale features yields consistent performance gains. Specifically, removing the feature map at $16 \times 16$ , $32 \times 32$ , or $64 \times 64$ resolution inside the MMSCopE module results in sub-optimal mIoU, with the removal of the $64 \times 64$ feature map resulting in the largest performance drop of $0.7\%$ . Additionally, using the $64 \times 64$ feature map alone without pixel unshuffling and compression leads to a $0.4\%$ decline. (3) Channel compression before SS2D proves beneficial, as its removal reduces mIoU by $0.3\%$ . Alternative channel scaling configuration (1C, 2C, 4C) for the three feature scales yields suboptimal results; (4) The FFN following SS2D provides a marginal benefit of $0.1\%$ . + +# 5. Conclusion + +This work presents a new segmentation network with two novel components: SegMAN Encoder, a hierarchical encoder that synergistically integrates state space models with local attention, and MMSCopE, an efficient decoder module that leverages state space models for multi-scale context extraction. SegMAN Encoder enables simultaneous modeling of fine-grained spatial details and global contextual information while MMSCopE adaptively scales with the input resolution while processing multi-scale features. We conduct comprehensive evaluations across multiple benchmarks, which demonstrates that our approach achieves state-of-the-art performance in semantic segmentation. + +# Acknowledgement + +This work was supported by Hong Kong Research Grants Council under the NSFC/RGC Collaborative Research Scheme (Grant CRS_HKU703/24). + +# References + +[1] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 3, 5 +[2] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 1, 2, 3 +[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. ArXiv, abs/1706.05587, 2017. +[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 3 +[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in neural information processing systems, 34:17864-17875, 2021. 6 +[6] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 6 +[7] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.6 +[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 3, 5, 7 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 7 +[10] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4 +[11] FAIR. fvcore. https://github.com/facebookresearch/fvcore, 2021.6 +[12] Zhengcong Fei, Mingyuan Fan, Changqian Yu, and Junshi Huang. Scalable diffusion models with state space backbone. arXiv preprint arXiv:2402.05608, 2024. 3 + +[13] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019. 1 +[14] Yunxiang Fu, Chaoqi Chen, and Yizhou Yu. Lamamba-diff: Linear-time high-fidelity diffusion models based on local attention and mamba. arXiv preprint arXiv:2408.02615, 2024. 3 +[15] Zhengyang Geng, Meng-Hao Guo, Hongxu Chen, Xia Li, Ke Wei, and Zhouchen Lin. Is attention better than matrix decomposition? arXiv preprint arXiv:2109.04553, 2021. 3 +[16] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3, 4 +[17] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3 +[18] Meng-Hao Guo, Cheng-Ze Lu, Qibin Hou, Zhengning Liu, Ming-Ming Cheng, and Shi-Min Hu. Segnext: Rethinking convolutional attention design for semantic segmentation. Advances in Neural Information Processing Systems, 35:1140-1156, 2022. 3, 5, 6, 7, 8, 2 +[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6185-6194, 2023. 2, 3, 4, 7 +[20] 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 +[21] Tong He, Zhi Zhang, Hang Zhang, Zhongyue Zhang, Junyuan Xie, and Mu Li. Bag of tricks for image classification with convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 558-567, 2019. 3 +[22] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 7, 8 +[23] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollar. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019. 3 +[24] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3041-3050, 2023. 2 +[25] Yanyu Li, Ju Hu, Yang Wen, Georgios Evangelidis, Kamyar Salahi, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Rethinking vision transformers for mobilenet size and speed. In Proceedings of the IEEE international conference on computer vision, 2023. 6 + +[26] Jiarun Liu, Hao Yang, Hong-Yu Zhou, Yan Xi, Lequan Yu, Yizhou Yu, Yong Liang, Guangming Shi, Shaoting Zhang, Hairong Zheng, et al. Swin-unamba: Mamba-based unet with imagenet-based pretraining. arXiv preprint arXiv:2402.03302, 2024. 3 +[27] 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 +[28] 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. 2, 3, 4, 5, 6, 7 +[29] 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. 3, 4, 6, 7, 8, 2 +[30] 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. 3, 4, 6, 7, 2 +[31] Meng Lou and Yizhou Yu. Overlock: An overview-firstlook-closely- next convnet with context-mixing dynamic kernels. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3 +[32] Meng Lou, Yunxiang Fu, and Yizhou Yu. Sparx: A sparse cross-layer connection mechanism for hierarchical vision mamba and transformer networks. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 6, 7, 2 +[33] Meng Lou, Shu Zhang, Hong-Yu Zhou, Chuan Wu, Sibei Yang, and Yizhou Yu. Transxnet: Learning both global and local dynamics with a dual dynamic token mixer for visual recognition. IEEE Transactions on Neural Networks and Learning Systems, 2025. 3 +[34] Chenyang Lu, Daan de Geus, and Gijs Dubbelman. Content-aware token sharing for efficient semantic segmentation with vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23631-23640, 2023. 4 +[35] Shervin Minaee, Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3523-3542, 2021. 1 +[36] Zhenliang Ni, Xinghao Chen, Yingjie Zhai, Yehui Tang, and Yunhe Wang. Context-guided spatial feature reconstruction for efficient semantic segmentation. arXiv preprint arXiv:2405.06228, 2024. 3, 6, 7, 8 +[37] Xuran Pan, Chunjiang Ge, Rui Lu, Shiji Song, Guanfu Chen, Zeyi Huang, and Gao Huang. On the integration of self-attention and convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815–825, 2022. 7, 8 +[38] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design + +spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 6 +[39] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 3, 4 +[40] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 3 +[41] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17773-17783, 2024. 3 +[42] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 5, 1 +[43] Jae-hun Shim, Hyunwoo Yu, Kyeongbo Kong, and SukJu Kang. Feedformer: Revisiting transformer decoder for efficient semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2263-2271, 2023. 3, 6 +[44] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 3 +[45] Quan Tang, Bowen Zhang, Jiajun Liu, Fagui Liu, and Yifan Liu. Dynamic token pruning in plain vision transformers for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 777-786, 2023. 3 +[46] Yao Teng, Yue Wu, Han Shi, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Dim: Diffusion mamba for efficient high-resolution image synthesis. arXiv preprint arXiv:2405.14224, 2024. 3 +[47] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In European conference on computer vision, pages 459-479. Springer, 2022. 7, 8, 1, 2 +[48] 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. 5 +[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 3, 7, 8 +[50] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 1, 2 +[51] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, + +Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14408-14419, 2023. 6, 7, 2 +[52] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.5 +[53] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5493-5502, 2024. 3, 6 +[54] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 3, 8 +[55] Yicheng Xiao, Lin Song, Jiangshan Wang, Siyu Song, Yixiao Ge, Xiu Li, Ying Shan, et al. Mambatree: Tree topology is all you need in state space model. Advances in Neural Information Processing Systems, 37:75329-75354, 2025. 1, 2 +[56] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 2, 3, 4, 5, 6, 7, 8 +[57] Haotian Yan, Ming Wu, and Chuang Zhang. Multi-scale representations by varying window attention for semantic segmentation. *ICLR*, 2024. 1, 2, 3, 5, 6, 7, 8 +[58] Chenglin Yang, Yilin Wang, Jianming Zhang, He Zhang, Zijun Wei, Zhe Lin, and Alan Yuille. Lite vision transformer with enhanced self-attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11998-12008, 2022. 3 +[59] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Learning a discriminative feature network for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1857-1866, 2018. 1 +[60] Hyunwoo Yu, Yubin Cho, Beoungwoo Kang, Seunghun Moon, Kyeongbo Kong, and Suk-Ju Kang. Embedding-free transformer with inference spatial reduction for efficient semantic segmentation. In European Conference on Computer Vision, pages 92-110. Springer, 2024. 2, 3, 4, 5, 6, 7, 8 +[61] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16, pages 173–190. Springer, 2020. 6 +[62] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024. 3 +[63] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In + +Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1, 3 +[64] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 3, 4 +[65] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 3, 5, 7, 8, 1 +[66] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. Bformer: Vision transformer with bi-level routing attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10323-10333, 2023. 7, 8 +[67] 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, 6, 7, 2 \ No newline at end of file diff --git a/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/images.zip b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3f7bc0f64194f70dc283950e946c862f05d4ada3 --- /dev/null +++ b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b22706ca031718045935c844a6df51734ca4e6bec1ed5b37cca0db4db3ff39d8 +size 649061 diff --git a/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/layout.json b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..47236a8e3055eb890275431a34a4028da22b7f6c --- /dev/null +++ b/CVPR/2025/SegMAN_ Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e98cd0d85a2766a666a2ac00a28bc73c6ccd96085a47296a6a7f202ba3cfe4ae +size 415397 diff --git a/CVPR/2025/Segment Any Motion in Videos/25801bc1-cb6e-4673-83cb-3317052d8a3e_content_list.json b/CVPR/2025/Segment Any Motion in Videos/25801bc1-cb6e-4673-83cb-3317052d8a3e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..80fa74a7b133a328c595de14aedde8f8e2202e41 --- /dev/null +++ b/CVPR/2025/Segment Any Motion in Videos/25801bc1-cb6e-4673-83cb-3317052d8a3e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2a678e14c08c5a984cd30e9e8c3561e174f2d3a44dba5f727db46cbe42c4758 +size 75774 diff --git a/CVPR/2025/Segment Any Motion in Videos/25801bc1-cb6e-4673-83cb-3317052d8a3e_model.json b/CVPR/2025/Segment Any Motion in Videos/25801bc1-cb6e-4673-83cb-3317052d8a3e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ddc6d964d3a89ac2e38bcf54b0f768b97baeb7c7 --- /dev/null +++ b/CVPR/2025/Segment Any Motion in Videos/25801bc1-cb6e-4673-83cb-3317052d8a3e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b5fc414dade56dc00a371bb47b593bbc62dfaa59bb6288258ee62aa0c108561 +size 99362 diff --git a/CVPR/2025/Segment Any Motion in Videos/25801bc1-cb6e-4673-83cb-3317052d8a3e_origin.pdf b/CVPR/2025/Segment Any Motion in Videos/25801bc1-cb6e-4673-83cb-3317052d8a3e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ffdbd9006dfc45e1d4cbb28f80da2e01569f3334 --- /dev/null +++ b/CVPR/2025/Segment Any Motion in Videos/25801bc1-cb6e-4673-83cb-3317052d8a3e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0b136555813bdc4350ca4efd75b711be680b3d2bd821b145bf2975ed2cf53fb +size 3279366 diff --git a/CVPR/2025/Segment Any Motion in Videos/full.md b/CVPR/2025/Segment Any Motion in Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..146d3129e49f272ceed74cf0325caaddcbfbd69f --- /dev/null +++ b/CVPR/2025/Segment Any Motion in Videos/full.md @@ -0,0 +1,287 @@ +# Segment Any Motion in Videos + +Nan Huang $^{1,2}$ Wenzhao Zheng $^{1}$ Chenfeng Xu $^{1}$ + +Kurt Keutzer1 Shanghang Zhang2 Angjoo Kanazawa1 Qianqian Wang1 + +$^{1}$ UC Berkeley $^{2}$ Peking University + +# Abstract + +Moving object segmentation is a crucial task for achieving a high-level understanding of visual scenes and has numerous downstream applications. Humans can effortlessly segment moving objects in videos. Previous work has largely relied on optical flow to provide motion cues; however, this approach often results in imperfect predictions due to challenges such as partial motion, complex deformations, motion blur and background distractions. We propose a novel approach for moving object segmentation that combines long-range trajectory motion cues with DINO-based semantic features and leverages SAM2 for pixel-level mask densification through an iterative prompting strategy. Our model employs Spatio-Temporal Trajectory Attention and Motion-Semantic Decoupled Embedding to prioritize motion while integrating semantic support. Extensive testing on diverse datasets demonstrates state-of-the-art performance, excelling in challenging scenarios and fine-grained segmentation of multiple objects. Our code is available at https://motion-seg.github.io/. + +# 1. Introduction + +Segmenting moving objects in videos is crucial for a range of applications, including action recognition, autonomous driving [10, 22, 65], and 4D reconstruction [58]. Many prior works address this problem under terms such as Video Object Segmentation (VOS) or motion segmentation. In this paper, we define our task as moving object segmentation (MOS) – segmenting objects that exhibit observable motion within the video. This definition differs from Video Object Segmentation, which includes objects that have the potential to move even if they remain static in the video, and from motion segmentation, which may also capture background motion, such as flowing water. This task is challenging as it implicitly requires distinguishing between camera motion and object motion, robustly tracking objects despite deformations, occlusions, rapid or transient movement, and segmenting them out with precise, clean masks. + +Recently, promptable visual segmentation has made sig- + +![](images/07a9fb0c4a3f2bb43614fb8a96b9368c7d223a5dab597264f772a8b7094c63c2.jpg) +Images Inputs + +![](images/7f4b5e5d77431c72fc717d29d86ea4d6bfef2736025714a0e333d8ee118ce27b.jpg) +Long-range Tracks Inputs + +![](images/17999daec2d298cfd1cf83cfbfafe2ab68f057c34adf14c8b242985204e42ca2.jpg) +Filtered Dynamic Tracks + +![](images/fb1d6e5661c408ab80fe66d3b08236f93ef6e30d8fc79449d03efc06992b9b6f.jpg) +Dynamic Masks +Figure 1. Our method is capable of handling challenging scenarios, including articulated structures, shadow reflections, dynamic background motion, and drastic camera movements, while producing per object level fine-grained moving object masks. + +nificant progress. Taking points, masks, or bounding boxes as prompts, SAM2 [51] segments and tracks the associated objects in videos effectively. However, SAM2 cannot natively handle MOS, as it has no mechanism to detect which objects are moving. + +We propose an innovative combination of long-range tracks with SAM2 for moving object segmentation to exploit the capabilities of SAM2. First, point tracking captures valuable long-range pixel motion information which is robust to deformation and occlusion, as shown in Fig. 2. At the same time, we incorporate DINO feature [12, 45], to add semantic context as a complementary source of information to support motion-based segmentation. We depart from traditional MOS approaches by training a model on extensive datasets that effectively combines motion and semantic information at a high level. Given a set of long-range 2D tracks, our model is designed to identify those tracks that correspond to moving objects. Once these dynamic tracks are identified, we apply a sparse-to-dense mask densification strategy, which uses an Iterative Prompting method in conjunction with SAM2 [51] to transform the sparse, point-level mask into a pixel-level segmentation. Since the primary objective is moving object segmentation, we em + +phasize motion cues while using semantic information as secondary support. To effectively balance these two types of information, we propose two specialized modules. (1) Spatio-Temporal Trajectory Attention. Given the long-term nature of input tracks, our model incorporates spatial attention to capture relationships between different trajectories and temporal attention to monitor changes within individual trajectories over time. (2) Motion-Semantic Decoupled Embedding. We implement special attention mechanisms to prioritize motion patterns and process semantic features in supplementary pathways. + +We trained our model on extensive datasets, including both synthetic [19, 28] and real-world data [36]. Due to the self-supervised nature of DINO features [45], our model demonstrates strong generalization capabilities, even when primarily trained on synthetic data. We evaluated our approach on benchmarks [34, 43, 47, 48] that were not part of the training data, and the results show that our method significantly outperforms baseline models in diverse tasks. + +While previous MOS methods leverage optical flow [6, 8, 9] to capture motion information, either by identifying different motion groups [6, 46, 53, 59] or by using learning-based models [8, 9, 18, 40, 49] to derive pixel masks from optical flow. However, optical flow is limited to short-range motion and can lose track over extended durations. Other methods [3, 7, 14, 42] rely on point trajectories as motion cues, but traditionally utilize spectral clustering on affinity matrices which struggle with complex motions. Though some methods also attempt to take advantage of appearance cues [24, 61] to help understand motion better, they typically handle different modalities in diverse separate stages, limiting the effective integration of their complementary information. Addressing these limitations, our unified framework achieves threefold integration: long-range trajectory, DINO feature, and SAM2. This design explains the model's exceptional capability in handling challenging cases like articulated motion and reflective surfaces as shown in the Fig. 1, and the superior performance in fine-grained segmentation of multiple objects. + +In summary, we make the following contributions: + +- We introduce an innovative combination of long-range tracks with SAM2, which enables efficient mask densification and tracking across frames. +- To obtain motion labels for trajectories, we propose a method that differs from traditional affinity matrix-based approaches and introduce the Motion-Semantic Decoupled Embedding, which enables a more effective integration of motion and semantic information, enhancing track-level segmentation by balancing these cues. +- Extensive results on multiple benchmarks demonstrate the effectiveness of our method, particularly in fine-grained moving object segmentation. + +![](images/ab45f340d5aba1c22638bfe22ecbb3971cec194d68e95e81ce73d22508750aaa.jpg) +Figure 2. The effectiveness of long-range tracks. Over longer periods of time, if a moving object experiences factors such as occlusion or changes in lighting, it can negatively affect the tracking performance of optical-flow-based methods for that object. + +# 2. Related Work + +Flow-based Moving Object Segmentation. Traditionally, optical flow based methods [6, 46, 53, 59] segment moving objects by grouping motion cues to create a moving object mask. These methods typically employ iterative optimization or statistical inference techniques to estimate motion models and identify motion regions simultaneously. Recently, numerous deep learning-based approaches [8, 9, 18, 37, 40, 49, 62] have used CNN encoders or transformer to extract motion cues from optical flow, followed by decoders to produce the final segmentation. The main distinctions among these methods lie in model architecture; for instance, methods that encode semantic information often utilize multiple CNN encoders to process different data modalities separately. In general, optical-flow-based methods struggle to distinguish independent object motion from apparent motion caused by depth differences. Furthermore, strong brightness changes also adversely affect these methods. Additionally, optical-flow-based methods are limited to short temporal sequences; they perform poorly if objects move slowly or are occluded. + +Trajectory-based Moving Object Segmentation. Trajectory-based methods can be typically classified into two categories: two-frame and multi-frame methods. Two-frame methods [3, 14, 25] generally estimate motion parameters by solving an iterative energy minimization problem, which are recently powered with various convolutional neural network (CNN) models [54, 74]. Multi-frame methods, in contrast, often utilize spectral clustering based on affinity matrices. These matrices are derived through techniques such as geometric model fitting [2, 23, 32, 63], subspace fitting [17, 50, 55, 57], or pairwise motion affinities that integrate motion and appearance information [7, 24, 30, 42]. Recent work has + +![](images/793c8605cb0f9ea599925f309d49bb867c5c5df877a34c5ad62678ba59658385.jpg) + +![](images/0da4e64087b6fcecac7b640e67dd85abdd8f2fe3b1e7db97b2a135266d5da7f9.jpg) +Figure 3. Overview of Our Pipeline. We take 2D tracks and depth maps generated by off-the-shelf models [15, 66] as input, which are then processed by a motion encoder to capture motion patterns, producing featured tracks. Next, we use tracks decoder that integrates DINO feature [45] to decode the featured tracks by decoupling motion and semantic information and ultimately obtain the dynamic trajectories(a). Finally, using SAM2 [51], we group dynamic tracks belonging to the same object and generate fine-grained moving object masks(b). + +focused on the search for more effective motion models. For instance, [1] uses the trifocal tensor to analyze point trajectories, arguing that it provides more reliable matches over three images than fundamental matrices can over two. However, the trifocal tensor also poses challenges: it is difficult to optimize and prone to failure when the three camera positions are nearly collinear[44]. Other studies [27, 64] have proposed geometric model fusion techniques to combine different models. Some recent work has explored integrating multiple motion cues [21, 41]. For example, [24] investigates combining point trajectories and optical flow, using well-crafted geometric motion models to fuse the two affinity matrices through co-regularized multi-view spectral clustering. However, these approaches still face inherent issues due to their reliance on affinity matrices. They tend to capture only local similarities, leading to poor global consistency, resulting in inconsistent segmentation. Furthermore, affinity matrices is difficult to capture dynamic changes in motion features like speed and direction over time. In contrast, we address the challenge of capturing motion similarities across diverse motion types. + +Unsupervised Video Object Segmentation. Unsupervised Video Object Segmentation (VOS) aims to automatically identify and track salient objects in raw video footage, while semi-supervised VOS relies on first-frame ground truth annotations to segment objects in subsequent frames [47, 48]. In this work, we focus on Unsupervised VOS, referred to here simply as "VOS". Recently, many approaches [68, 71] have combined motion and appearance information. For instance, MATNet [73] introduces a motion-attentive transition model for unsupervised VOS, leveraging + +ing motion cues to guide segmentation with a primary focus on appearance. RTNet [52] presents a method based on reciprocal transformations, using the consistency of object appearance and motion between consecutive frames to achieve segmentation. FSNet [26] employs a full-duplex strategy with a dual-path network to jointly model both appearance and motion. Overall, VOS generally targets salient objects in videos, regardless of whether the object is moving. Although many VOS methods incorporate motion information, it is often not their primary focus. + +# 3. Method + +Our objective is, given a video, to identify moving objects and generate pixel-level dynamic masks. Fig. 3 provides an overview of our pipeline. The central insight is that long-range tracks not only capture motion patterns that facilitate video understanding but also offer long-range prompts essential for promptable visual segmentation. Thus, we use long-range point tracks as motion cues, serving as the primary input in Sec. 3.1, where we apply spatial-temporal attention to capture context-aware feature. In Sec. 3.2, we further incorporate and decouple the use of semantic information with motion cues to decode features, helping the model predict the final motion labels. After identifying dynamic tracks, we leverage these long-range tracks to prompt SAM2 [51] iteratively, as described in Sec. 3.3. + +# 3.1. Motion Pattern Encoding + +Point trajectories carry valuable information for understanding motion, and related MOS methods can be typi + +cally classified into two categories: two-frame and multiframe methods. However, as discussed in Sec. 2, two-frame methods [3, 14, 25] often suffer from significant temporal inconsistencies and exhibit degraded performance when input flows are noisy. Multi-frame methods, in contrast, often utilize spectral clustering based on affinity matrices. Nevertheless, they remain highly sensitive to noise and struggle to handle global, dynamic, and complex motion patterns effectively. + +To address these limitations, and inspired by ParticleSFM [72], we propose a method that leverages long-range point tracks [15], processed through a specialized trajectory processing model, to predict per-trajectory motion labels. As illustrated in Fig. 3, our proposed network adopts an encoder-decoder architecture. The encoder directly processes long-range trajectory data and applies a Spatio-Temporal Trajectory Attention mechanism across trajectories. This mechanism integrates both spatial and temporal cues, capturing both local and global information across time and space, in order to embed the motion pattern of each trajectory. + +Given that the accuracy and quality of long-range trajectories significantly impact model performance, we utilize BootsTAP [15] to generate the tracks, which provides a confidence score for each track at each time step, enabling us to mask out low-confidence points. Furthermore, due to the movement of dynamic objects and camera motion, the visibility of long-range tracks can vary over time, as they may be occluded or move out of the frame. This variability in visibility and confidence makes each trajectory data highly irregular, motivating our use of a transformer model, inspired by sequence modeling approaches in natural language processing [56, 72], to handle the data effectively. + +Our input data comprises long-range trajectories, with each trajectory consisting of normalized pixel coordinates $(u_{i},v_{i})$ , visibility $\rho_{i}$ and confidence scores $c_{i}$ , where $i\in$ (0, time). Masks $\mathcal{M}_i$ is applied to indicate points where the pixel coordinates are either invisible or low-confidence. Additionally, we integrate monocular depth maps $d_{i}$ estimated by Depth-Anything [66], which, despite some noise, provide valuable insights into the underlying 3D scene structure, enhancing understanding of spatial layout and occlusions. To further enrich the input data and strengthen temporal motion cues, we compute frame-to-frame differences in both trajectory coordinates $(\Delta u_{i},\Delta v_{i})$ and depth $\Delta d_{i}$ for adjacent frames. + +Since adjacent sampling points in coordinates can lead to oversmoothing of spatially close features, we draw inspiration from NeRF [39] to address this issue. Specifically, we apply frequency transformations for positional encoding to better capture fine-grained spatial details. + +The final augmented trajectories pass through two MLPs to generate intermediate features, which are then fed into + +the transformer encoder. Given the long-range nature of the input data, we propose a Spatio-Temporal Trajectory Attention for our encoder $\mathcal{E}$ , interleaves attention layers that operate alternately across track and temporal dimensions [4, 29]. This design allows the model to capture both the temporal dynamics within each trajectory and the spatial relationships across different trajectories. Finally, to obtain a feature representation for each entire trajectory rather than individual points, we perform max-pooling along the temporal dimension, following [72]. This process yields a single feature vector for each trajectory, naturally forming a high-dimensional featured track that implicitly captures the unique motion pattern of each trajectory. + +# 3.2. Per-trajectory Motion Prediction + +Though we encoded motion pattern in Sec. 3.1, it is still challenging to distinguish moving objects based solely on motion cues, because learning to differentiate between object motion and camera motion from highly abstracted trajectories is difficult for the model. Providing the model with texture, appearance, and semantic information can simplify this task by helping it understand which objects are likely to move or be moved. Some approaches directly apply semantic segmentation models [5, 20, 67, 70] where potentially moving pixels are identified based on semantic labels. While these methods can be effective in specific scenarios, they are intrinsically limited for general moving object segmentation, as they depend entirely on predefined semantic classes. Recently, many MOS [61, 69] and VOS [11, 33, 35] methods combine appearance information and motion cues, but they do so in two separate stages, often using RGB images to refine masks. However, relying on raw RGB data may fail to capture high-level information, and applying the two modalities in separate stages limits the effective integration of their complementary information. + +To address these limitations, we incorporate DINO features predicted by DINO v2 [45], a self-supervised model, which helps generalize the inclusion of appearance information. However, we observed that simply introducing DINO features as input makes the model overly reliant on semantics as shown in Fig. 8 and discussed in Sec. 4.5, reducing its ability to differentiate between moving and static objects within the same semantic category. To overcome this issue, we propose a Motion-Semantic Decoupled Embedding, enabling the transformer decoder $\mathcal{D}$ to prioritize motion information while still considering semantic cues. + +We obtain the final embedded featured tracks $\mathcal{P}$ through the process described in Sec. 3.1: + +$$ +\mathcal {P} = \mathcal {E} ((\gamma (u), \gamma (v), \gamma (\Delta u), \gamma (\Delta v), d, \Delta d, \rho , c), \mathcal {M}). \tag {1} +$$ + +We then design a transformer-based decoder, where the encoder layer performs attention only on the embedded featured tracks, which contain motion information exclusively. + +After computing the attention-weighted feature, we concatenate the DINO feature and pass this concatenated feature through a feed-forward layer. In the decoder layer, self-attention is still applied only to the motion features; however, multi-head attention is used to attend to a memory that includes semantic information. Finally, we apply a sigmoid activation function to produce the final output, yielding the predicted label for each trajectory. + +We then compute the loss between these predicted labels and per-track ground truth labels using a weighted binary cross-entropy loss [72]. We assign ground truth labels to each trajectory by checking if the sampled point coordinates lie within the ground truth dynamic masks. If a point falls inside the mask, it is labeled as dynamic. + +# 3.3. SAM2 Iterative Prompting + +As depicted in Fig. 3, after obtaining the predicted label of each trajectory and filter dynamic trajectories, we use these trajectories as point prompts for SAM2[51] with an iterative, two-stage prompting strategy. The first stage focuses on grouping trajectories belonging to the same object and storing the trajectories of each distinct object in memory. In the second stage, this memory is used as a prompt for SAM2 [51] to generate dynamic masks. + +The motivation behind this approach is twofold. First, it is necessary because SAM2 requires object IDs as input. However, if we assign the same object ID to all dynamic objects (e.g., assigning 1 to represent all dynamic objects), SAM2 would struggle to simultaneously segment multiple objects that share the same ID. Second, this method offers the benefit of achieving finer-grained segmentation. + +In the first stage, we select the time frame with the maximum number of visible points and locate the densest point among all visible points in that frame. This point serves as the initial prompt for SAM2 [51], which then generates an initial mask for that frame. After generating this mask, we apply dilation to expand its boundaries, excluding all points within the expanded mask area to remove edge points and assume that these points belong to the same object. We then proceed to the next frame with the highest number of visible points and repeat this process until the remaining visible points across all frames are too few to process. The trajectories identified as belonging to the same object are stored in memory, with unique object IDs assigned to each. We only save the points within the undilated mask for each object. + +In the second stage, we use this memory to refine prompt selection by locating the densest point within the stored trajectories and the two points furthest from this point. Leveraging the long-range nature of trajectories, we prompt SAM2 at regular intervals to prevent it from losing track of the object over extended distances. Since SAM2 may generate partial object masks (e.g., parts of a person's clothing), we perform post-processing on all masks to merge + +those that overlap internally or appear within the same mask boundaries. This results in a complete mask for each distinct object. + +# 4. Experiments + +# 4.1. Implementation Details + +Training Dataset. We train our model using three datasets: Kubric[19], Dynamic Replica[28], and HOI4D [36], sampling them at a ratio of $35\%$ , $35\%$ , and $30\%$ respectively. Kubric [19] is a synthetic dataset composed of sequences of 24 frames showing 3D rigid objects falling under gravity and bouncing. We generate dynamic masks for each sequence based on the motion labels of individual objects. Dynamic Replica [28] is another synthetic dataset, created for 3D reconstruction, that includes long-term tracking annotations and object masks, featuring articulated models of humans and animals. We calculate dynamic masks by analyzing the 3D tracks to determine whether each object is in motion, providing accurate motion segmentation for this dataset. HOI4D [36] is a real-world, egocentric dataset that contains common objects involved in human-object interactions. This dataset provides official motion segmentation masks, making it ideal for real-world training of our model. + +Data Sampling. During training, we randomly sample a variable number of tracking points, enhancing the model's robustness to different track counts. For the Dynamic Replica dataset [28], which contains 300 frames, we speed up training by sampling $1/4$ of the frames at regular intervals randomly. This approach preserves the large camera motion characteristics of the dataset. We find that including the Dynamic Replica dataset is essential for helping the model understand camera motion effectively. + +# 4.2. Benchmark and metrics + +We evaluate our model using several established datasets for moving object video segmentation. DAVIS17-Moving[13] is a subset of the DAVIS2017 dataset[48], designed specifically for moving object detection and segmentation. In DAVIS17-Moving, all moving instances within each video sequence are labeled, while static objects are excluded. Following the same criteria, we created DAVIS16-Moving as a subset of the DAVIS2016 dataset [47]. Additionally, we report performance on other popular video object segmentation benchmarks, including DAVIS2016[47], SegTrackv2[34], and FBMS-59 [43]. + +For evaluation, we benchmark our moving object video segmentation performance using region similarity (J) and contour similarity (F) metrics, as outlined in [35, 38, 61]. + +# 4.3. Moving Object Segmentation + +We selected methods that specifically target moving object segmentation as baselines [35, 38, 60, 61, 69]. For + +![](images/669c46a5af54aa15521a07f6ab3ca6668d088c00bc66eb7ea926e617422f1044.jpg) +Figure 4. Qualitative comparison on DAVIS17-moving benchmarks. For each sequence we show moving object mask results. Our method successfully handles water reflections (left), camouflage appearances (middle), and drastic camera motion (right). + +![](images/83a3f772fb61d90f31d3879da5450b9c9ba684ef9b2460f28b6e266ff8fc4eac.jpg) +Figure 5. Qualitative comparison on FBMS-59 benchmarks. The masks produced by us are geometrically more complete and detailed. + +OCLR [60], we report results for two versions: OCLR-flow, which uses only flow input, and a second version OCLR-TTA that incorporates test-time adaptation on top of OCLR-flow. For RCF [35], the first stage, RCF-stage 1, focuses on motion information, while the second stage, RCF-All, further optimizes the results from the first stage. We report results for both stages. For all baselines, we apply a fully connected conditional random field (CRF) [31] to refine the masks and achieve the best possible results. + +Notably, for multi-object scenarios, we follow the common practice [16, 60, 61, 69] of grouping all foreground objects together for evaluation purposes, which we refer to as MOS. Although our approach is capable of generating highly accurate, fine-grained per-object masks, as detailed in Sec. 4.4, we term this second evaluation method + +as fine-grained MOS. Table 1 compares the performance of our model with several baseline methods on the MOS task. Our method achieves state-of-the-art F-scores across all datasets, and our region similarity (J) scores are either the best or second-best across multiple datasets, further validating the effectiveness of our approach. Fig. 4 shows our visual results on the DAVIS16-Moving dataset, where our method accurately identifies object boundaries without incorrectly labeling moving backgrounds. Moreover, our masks exhibit strong geometric structure, particularly in challenging scenarios with significant camera motion. Fig. 5 and Fig. 6 present qualitative results on the FBMS-59 and SegTrack v2 benchmarks, respectively. Our method performs exceptionally well in maintaining mask geometry, and even in cases where the RGB images are blurred or of + +![](images/c722510269c7038ca8fb740419f559bc1f024475532c9650e2cd62ceda4c85e3.jpg) +Figure 6. Qualitative comparison on SegTrack v2 benchmarks. Our method succeeds even under motion blur conditions. + +Table 1. Quantitative comparison on MOS task which grouping all foreground objects together for evaluation. + +
MethodsModel SettingsDAVIS2016-MovingSegTrackv2FBMS-59DAVIS2016
MotionAppearanceJ&F↑J↑F↑J↑J↑F↑J&F↑J↑F↑
CIS [69]Optical FlowRGB66.267.664.862.063.6-68.670.366.8
EM [38]Optical FlowX75.276.274.355.557.956.070.069.370.7
RCF-Stage1 [35]Optical FlowX77.378.676.076.769.9-78.580.276.9
RCF-All [35]Optical FlowDINO79.681.078.379.672.4-80.782.179.2
OCLR-flow [60]Optical FlowX70.070.070.067.665.564.971.272.070.4
OCLR-TTA [60]Optical FlowRGB78.580.276.972.369.968.378.880.876.8
ABR [61]Optical FlowDINO72.070.273.776.681.979.672.571.873.2
OursTrajectoryDINO89.589.289.776.378.382.890.990.691.0
+ +![](images/1cf10bdac16f86d73ecdf37c11ec154568dfacfe1c6f67287da32842ae073bc7.jpg) +Figure 7. Qualitative comparison on Fine-grained MOS task which will produce per-object level masks. + +low quality, our reliance on long-range trajectories enables accurate identification of moving objects. + +# 4.4. Fine-grained Moving Object Segmentation + +Building on the initial MOS task, this task not only identifies moving objects but also classifies them within their motion context to generate fine-grained, per-object masks. We evaluate our approach for multi-moving object segmentation specifically on the DAVIS2017-Moving dataset. For a fair comparison, we only include baselines that claim the + +Table 2. Quantitative comparison on DAVIS17-Moving dataset for MOS and Fine-grained MOS tasks. + +
MethodsMOSFine-grained MOS
J↑F↑J&F↑J↑F↑
OCLR-flow [60]69.970.044.442.146.8
OCLR-TTA [60]76.075.349.148.449.9
ABR [61]74.675.251.150.951.2
Ours90.089.080.577.483.6
+ +ability to perform this task. Table 2 shows that our method significantly outperforms the baselines, demonstrating its superior capability in producing accurate per-object masks. Additionally, Fig. 7 illustrates that, first, our method accurately identifies each object, effectively distinguishing different objects with similar motion patterns. Second, it ensures the completeness of each object mask, handling challenging cases such as articulated human structures and occluded objects while maintaining mask integrity. + +# 4.5. Ablation Study + +We investigate the effectiveness of our method and its various components on the DAVIS17-Moving and DAVIS16-Moving datasets. The former is used for fine-grained MOS, while the latter focuses on MOS. All models are trained for a full number of epochs. + +We conducted several experiments to assess the importance of each component. The w/o DINO configuration ex + +![](images/2bcd9e4108526de0d12303cd4b8e9aed5ff3981ff0e0931c1362eb703ac4fc3f.jpg) +Figure 8. Visual comparison for the ablation study on two critical and challenging cases. The top sequence shows scenarios involving drastic camera motion and complex motion patterns, while the bottom sequence with both static and dynamic objects of the same category. The experimental setup is detailed in Sec. 4.5. + +Table 3. Quantitative comparison for the ablation study on the DAVIS17-Moving and DAVIS16-Moving benchmarks, which evaluate fine-grained MOS and MOS tasks, respectively. The experimental setup is detailed in Sec. 4.5. + +
MethodsDAVIS17-MovingDAVIS16-Moving
J&F↑J↑F↑J&F↑J↑F↑
w/o Depth69.265.672.882.578.686.4
w/o Tracks19.614.524.720.99.831.9
w/o DINO65.062.167.975.571.479.5
w/o MOE72.068.775.481.881.082.7
w/o MSDE63.059.366.778.277.379.1
w/o PE66.464.768.282.081.582.5
w/o ST-ATT65.561.969.178.374.382.4
Ours-full80.577.483.689.189.089.2
+ +cludes DINO features entirely during training, while w/o MOE (Motion-only Encoding) concatenates DINO features with motion cues before the motion encoder, allowing both the encoder and decoder layers to incorporate DINO information throughout. w/o MSDE (Motion-Semitic Decoupled Embedding) excludes DINO features from the motion encoder but concatenates them with the embedded featured tracks from the encoder output, introducing semantic information through self-attention in the tracks decoder. We also test configurations w/o depth and w/o tracks, removing specific inputs to observe their impact on performance. Additionally, w/o PE (Positional Embedding) omits NeRF-like positional embedding in the motion encoder, and w/o ST-ATT (Spatial-temporal Attention) replaces spatial-temporal attention with conventional attention. + +Table 3 presents the quantitative results. We find that excluding depth as input or positional encoding impacts performance less than other components, but it still falls significantly short of the best results. When tracks are removed and only DINO features and depth maps are used, performance drops drastically, indicating that the model struggles to learn effectively without trajectory-based information. We further analyze the key components in two chal + +lenging scenarios presented below. + +Drastic Camera Motion. We observed that in highly challenging scenes, such as those with drastic camera movement or rapid object motion, relying solely on motion information is insufficient. As shown in the upper part of Fig. 8, the colored points represent dynamic points predicted by the model, while the hollow points indicate invisible points at that moment. In this example, without DINO feature information, the model incorrectly classifies the stationary road surface as dynamic, despite the fact that the road lacks the ability to move. This information can be effectively supplemented by incorporating DINO features. Additionally, we found that adding spatial-temporal attention within the motion encoder is particularly beneficial in these difficult scenarios, as it provides the model with richer motion information to capture the long-range motion patterns of tracks, as illustrated in Fig. 8. + +Distinguishing Moving and Static Objects of the Same Category. Results show that excluding DINO features entirely results in a performance drop, and the manner in which these features are integrated significantly affects the model's output. Simply incorporating DINO as an input during the motion encoding stage causes the model to rely heavily on semantic information, often leading it to assume that objects of the same type share the same motion state. In contrast, our Motion-Semantic Decoupled Embedding architecture effectively reduces this over-reliance on semantics, allowing the model to differentiate between moving and static objects within the same category, as illustrated in the lower part of Fig. 8. + +# 5. Conclusion + +In this work, we present a novel approach that leverages long-range tracks which departs from traditional affinity matrix-based methods. Trained on extensive datasets, our model accurately identifies dynamic tracks, which, when combined with SAM2, produce precise moving object masks. Our carefully designed model architecture is tailored to handle long-range motion information while effectively balancing motion and appearance cues. Experiments show that our method achieves state-of-the-art results across multiple benchmarks, with particularly strong performance in per-object-level segmentation. + +# 6. Limitation + +Our current model utilizes off-the-shelf long-range tracking estimators, whose accuracy can greatly influence overall performance, as shown in Tab. 3. However, we believe that our novel approach to moving object segmentation can, in turn, enhance the performance and understanding of these estimators, creating a mutually beneficial improvement. + +# References + +[1] Federica Arrigoni, Luca Magri, and Tomas Pajdla. On the Usage of the Trifocal Tensor in Motion Segmentation, page 514-530. 2020. 3 +[2] Federica Arrigoni, Luca Magri, and Tomas Pajdla. On the Usage of the Trifocal Tensor in Motion Segmentation, page 514-530. 2020. 2 +[3] Daniel Barath and Jiri Matas. Progressive-x: Efficient, anytime, multi-model fitting algorithm. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition, 2019. 2, 4 +[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? Cornell University - arXiv, Cornell University - arXiv, 2021. 4 +[5] Berta Bescos, Jose M. Facil, Javier Civera, and Jose Neira. Dynaslam: Tracking, mapping and inpainting in dynamic scenes. IEEE Robotics and Automation Letters, page 4076-4083, 2018. 4 +[6] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. Cornell University - arXiv, Cornell University - arXiv, 2016. 2 +[7] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In European conference on computer vision, pages 282-295. Springer, 2010. 2 +[8] M. Bösch. Deep learning for robust motion segmentation with non-static cameras. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition, 2021. 2 +[9] Zhe Cao, Abhishek Kar, Christian Hane, and Jitendra Malik. Learning independent object motion from unlabelled stereoscopic videos. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[10] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv:2311.18561, 2023. 1 +[11] Suhwan Cho, Minhyeok Lee, Seunghoon Lee, Dogyoon Lee, Heeseung Choi, Ig-Jae Kim, and Sangyoun Lee. Dual prototype attention for unsupervised video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19238-19247, 2024. 4 +[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 1 +[13] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 5 +[14] Andrew Delong, Anton Osokin, Hossam N. Isack, and Yuri Boykov. Fast approximate energy minimization with label costs. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2010. 2, 4 +[15] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, + +Ross Goroshin, João Carreira, and Andrew Zisserman. BootsTAP: Bootstrapped training for tracking any point. arXiv, 2024. 3, 4 +[16] Suyog Dutt Jain, Bo Xiong, and Kristen Grauman. Fusion-seg: Learning to combine motion and appearance for fully automatic segmentation of generic objects in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3664-3673, 2017. 6 +[17] E. Elhamifar and R. Vidal. Sparse subspace clustering: Algorithm, theory, and applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 2765-2781, 2013. 2 +[18] Muhammad Faisal, Ijaz Akhter, Mohsen Ali, and Richard Hartley. Epo-net: Exploiting geometric constraints on dense trajectories for motion saliency. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition, 2019. 2 +[19] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzehrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliausacchi. Kubric: a scalable dataset generator. 2022. 2, 5 +[20] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 386-397, 2020. 4 +[21] Christian Homeyer and RobertBosch Gmbh. On moving object segmentation from monocular video with transformers. 3 +[22] Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. S3gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323, 2024. 1 +[23] Yuxiang Huang and John Zelek. Motion segmentation from a moving monocular camera. 2023. 2 +[24] Yuxiang Huang, Yuhao Chen, and John Zelek. Zero-shot monocular motion segmentation in the wild by combining deep learning with geometric motion model fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2733-2743, 2024. 2, 3 +[25] Hossam Isack and Yuri Boykov. Energy-based geometric multi-model fitting. International Journal of Computer Vision, page 123-147, 2012. 2, 4 +[26] Ge-Peng Ji, Keren Fu, Zhe Wu, Deng-Ping Fan, Jianbing Shen, and Ling Shao. Full-duplex strategy for video object segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4922-4933, 2021. 3 +[27] Yangbangyan Jiang, Qianqian Xu, Ke Ma, Zhiyong Yang, Xiaochun Cao, and Qingming Huang. What to select: Pursuing consistent motion segmentation from multiple geometric + +models. Proceedings of the AAAI Conference on Artificial Intelligence, page 1708-1716, 2022. 3 +[28] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. CVPR, 2023. 2, 5 +[29] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In Proc. ECCV, 2024. 4 +[30] Laurynas Karazija, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Learning segmentation from point trajectories, 2024. 2 +[31] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Neural Information Processing Systems, Neural Information Processing Systems, 2011. 6 +[32] Taotao Lai, Hanzi Wang, Yan Yan, Tat-Jun Chin, and Wan-Lei Zhao. Motion segmentation via a sparsity constraint. IEEE Transactions on Intelligent Transportation Systems, page 973–983, 2017. 2 +[33] Minhyeok Lee, Suhwan Cho, Dogyoon Lee, Chaewon Park, Jungho Lee, and Sangyoun Lee. Guided slot attention for unsupervised video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3807-3816, 2024. 4 +[34] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M. Rehg. Video segmentation by tracking many figure-ground segments. In 2013 IEEE International Conference on Computer Vision, pages 2192-2199, 2013. 2, 5 +[35] Long Lian, Zhirong Wu, and Stella X. Yu. Bootstrapping objectness from videos by relaxed common fate and visual grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14582-14591, 2023. 4, 5, 6, 7 +[36] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21013-21022, 2022. 2, 5 +[37] Etienne Meunier and Patrick Bouthemy. Unsupervised motion segmentation in one go: Smooth long-term model over a video, 2024. 2 +[38] Etienne Meunier, Anais Badoual, and Patrick Bouthemy. Em-driven unsupervised learning for efficient motion segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4462-4473, 2023. 5, 7 +[39] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 4 +[40] Eslam Mohamed, Mahmoud Ewaisha, Mennatullah Siam, Hazem Rashed, Senthil Yogamani, Waleed Hamdy, Mohamed El-Dakdouky, and Ahmad El-Sallab. Monocular instance motion segmentation for autonomous driving: Kitti instancemotseg dataset and multi-task baseline. In 2021 IEEE Intelligent Vehicles Symposium (IV), 2021. 2 + +[41] Michal Neoral and Jan Sochman. Monocular arbitrary moving object discovery and segmentation. 3 +[42] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1187-1200, 2014. 2 +[43] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(6): 1187-1200, 2014. 2, 5 +[44] H. Opower. Multiple view geometry in computer vision. Optics and Lasers in Engineering, page 85-86, 2002. 3 +[45] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 1, 2, 3, 4 +[46] Anestis Papazoglou and Vittorio Ferrari. Fast object segmentation in unconstrained video. In 2013 IEEE International Conference on Computer Vision, 2013. 2 +[47] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition, 2016. 2, 3, 5 +[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 2, 3, 5 +[49] Mohamed Ramzy, Hazem Rashed, AhmadEl Sallab, and Senthil Yogamani. Rst-modnet: Real-time spatio-temporal moving object detection for autonomous driving. Cornell University - arXiv, Cornell University - arXiv, 2019. 2 +[50] S Rao, R Tron, R Vidal, and Yi Ma. Motion segmentation in the presence of outlying, incomplete, or corrupted trajectories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(10):1832-1845, 2010. 2 +[51] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 3, 5 +[52] Sucheng Ren, Wenxi Liu, Yongtuo Liu, Haoxin Chen, Guoqiang Han, and Shengfeng He. Reciprocal transformations for unsupervised video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15455-15464, 2021. 3 +[53] Hicham Sekkati and Amar Mitiche. A variational method for the recovery of dense 3d structure from motion. Robotics and Autonomous Systems, 55(7):597-607, 2007. 2 +[54] Pavel Tokmakov, Cordelia Schmid, and Karteek Alahari. Learning to segment moving objects. International Journal of Human Machine Interaction, 2017, 38(1-2), 1-16. + +nal of Computer Vision,International Journal of Computer Vision,2017.2 +[55] Roberto Tron and Rene Vidal. A benchmark for the comparison of 3-d motion segmentation algorithms. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, page 1-8, 2007. 2 +[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, AidanN. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Neural Information Processing Systems, Neural Information Processing Systems, 2017. 4 +[57] René Vidal. Subspace clustering. IEEE Signal Processing Magazine, IEEE Signal Processing Magazine, 2011. 2 +[58] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video, 2024. 1 +[59] Andreas Wedel, Annemarie Meißner, Clemens Rabe, Uwe Franke, and Daniel Cremers. Detection and Segmentation of Independently Moving Objects from Dense Scene Flow, page 14-27. 2009. 2 +[60] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. In NeurIPS, 2022. 5, 6, 7 +[61] Junyu Xie, Weidi Xie, and Andrew Zisserman. Appearance-based refinement for object-centric motion segmentation. In ECCV, 2024. 2, 4, 5, 6, 7 +[62] Junyu Xie, Charig Yang, Weidi Xie, and Andrew Zisserman. Moving object segmentation: All you need is sam (and flow), 2024. 2 +[63] Xun Xu, Loong-Fah Cheong, and Zhuwen Li. Motion segmentation by exploiting complementary geometric models. Cornell University - arXiv, Cornell University - arXiv, 2018. 2 +[64] Xun Xu, Loong Fah Cheong, and Zhuwen Li. Motion segmentation by exploiting complementary geometric models. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2859-2867, 2018. 3 +[65] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, and Yue Wang. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 1 +[66] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 3, 4 +[67] Shichao Yang and Sebastian Scherer. Cubeslam: Monocular 3d object slam. IEEE Transactions on Robotics, page 925-938, 2019. 4 +[68] Shu Yang, Lu Zhang, Jinqing Qi, Huchuan Lu, Shuo Wang, and Xiaoxing Zhang. Learning motion-appearance coattention for zero-shot video object segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1564–1573, 2021. 3 +[69] Yanchao Yang, Antonio Loquercio, Davide Scaramuzza, and Stefano Soatto. Unsupervised moving object detection via contextual information separation. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 879-888, 2019. 4, 5, 6, 7 +[70] Chao Yu, Zuxin Liu, Xin-Jun Liu, Fugui Xie, Yi Yang, Qi Wei, and Qiao Fei. Ds-slam: A semantic visual slam towards dynamic environments. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018. 4 +[71] Kaihua Zhang, Zicheng Zhao, Dong Liu, Qingshan Liu, and Bo Liu. Deep transport network for unsupervised video object segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8781-8790, 2021. 3 +[72] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In European conference on computer vision (ECCV), 2022. 4, 5 +[73] Tianfei Zhou, Shunzhou Wang, Yi Zhou, Yazhou Yao, Jianwu Li, and Ling Shao. Motion-attentive transition for zero-shot video object segmentation. In Proceedings of the AAAI conference on artificial intelligence, pages 13066–13073, 2020. 3 +[74] Tianfei Zhou, Shunzhou Wang, Yi Zhou, Yazhou Yao, Jianwu Li, and Ling Shao. Motion-attentive transition for zero-shot video object segmentation. In Proceedings of the AAAI conference on artificial intelligence, pages 13066–13073, 2020. 2 \ No newline at end of file diff --git a/CVPR/2025/Segment Any Motion in Videos/images.zip b/CVPR/2025/Segment Any Motion in Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a332c2619c3552b68079a5e7d52c2f58a316f6eb --- /dev/null +++ b/CVPR/2025/Segment Any Motion in Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:905afe34b31c2cb6b5ffaf21bdc0ea453e6d4fe898360e75cbbfd39c9d99a3fa +size 758678 diff --git a/CVPR/2025/Segment Any Motion in Videos/layout.json b/CVPR/2025/Segment Any Motion in Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..22ab6d004b67ac9d70e426291389f32524a1ddba --- /dev/null +++ b/CVPR/2025/Segment Any Motion in Videos/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:845a6cfeaac1c0c6960b193eb03908a9490c8db8cde16f20cba318d6043e3ccc +size 334475 diff --git a/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/1afb4c97-64af-47c3-b3be-9332e1082a6d_content_list.json b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/1afb4c97-64af-47c3-b3be-9332e1082a6d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6224f4ae882082dcb06038cb9996013a305bf544 --- /dev/null +++ b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/1afb4c97-64af-47c3-b3be-9332e1082a6d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5676bec1d9125df42b70e6d015604a25fb4f349ba95db0749c224281256674c5 +size 93455 diff --git a/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/1afb4c97-64af-47c3-b3be-9332e1082a6d_model.json b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/1afb4c97-64af-47c3-b3be-9332e1082a6d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9ae7ad036b0dca697b5b2375441af79dd86291ec --- /dev/null +++ b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/1afb4c97-64af-47c3-b3be-9332e1082a6d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0d951553cbb0996477c764b2daaa72852754985218dc2407488e406856ea57f +size 115375 diff --git a/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/1afb4c97-64af-47c3-b3be-9332e1082a6d_origin.pdf b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/1afb4c97-64af-47c3-b3be-9332e1082a6d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e689fe84f8e0c94d3567cc5cfd0caa36edd18318 --- /dev/null +++ b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/1afb4c97-64af-47c3-b3be-9332e1082a6d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:669379a3f5343274ebf3d55fddec41caff10078fd18d0c33a2febd95ae596543 +size 1762035 diff --git a/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/full.md b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fac131207e0cc409533ac745506f4783247451d2 --- /dev/null +++ b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/full.md @@ -0,0 +1,400 @@ +# Segment Any-Quality Images with Generative Latent Space Enhancement + +Guangqian Guo $^{1,2}$ Yong Guo $^{2}$ Xuehui Yu $^{3}$ Wenbo Li $^{4}$ Yaoxing Wang $^{1}$ Shan Gao $^{1}$ $^{1}$ Northwestern Polytechnical University $^{2}$ Huawei $^{3}$ Tencent $^{4}$ Huawei Noah's Ark Lab + $\{\mathrm{guogq}21, \mathrm{wangyx}24\} @$ mail.nwpu.edu.cn $\{\mathrm{guoyongcs, fenglinglwb}\} @$ gmail.com +xuehuiyu@tencent.com gaoshan@nwpu.edu.cn + +# Abstract + +Despite their success, Segment Anything Models (SAMs) experience significant performance drops on severely degraded, low-quality images, limiting their effectiveness in real-world scenarios. To address this, we propose GleSAM, which utilizes Generative Latent space Enhancement to boost robustness on low-quality images, thus enabling generalization across various image qualities. Specifically, we adapt the concept of latent diffusion to SAM-based segmentation frameworks and perform the generative diffusion process in the latent space of SAM to reconstruct high-quality representation, thereby improving segmentation. Additionally, we introduce two techniques to improve compatibility between the pre-trained diffusion model and the segmentation framework. Our method can be applied to pre-trained SAM and SAM2 with only minimal additional learnable parameters, allowing for efficient optimization. We also construct the LQSeg dataset with a greater diversity of degradation types and levels for training and evaluating the model. Extensive experiments demonstrate that GleSAM significantly improves segmentation robustness on complex degradations while maintaining generalization to clear images. Furthermore, GleSAM also performs well on unseen degradations, underscoring the versatility of our approach and dataset. + +# 1. Introduction + +Accurate object detection and segmentation [3, 12, 15, 16, 18, 26, 51, 55] in diverse scenarios is a fundamental task for various high-level visual applications, such as robotics and autonomous driving. The recently developed Segment Anything Models (SAMs), including SAM [27] and SAM2 [44], serving as a foundational model, have gained significant influence within the community [4, 5, 10, 16, 36, 37] due to their outstanding zero-shot segmentation abilities. + +Images w/ various quality + +![](images/49f2c4db074d74d67e34459a85c177f98ffb9c08334230d23f821ac0bd067b01.jpg) +Clear + +![](images/461cea2f23efb6afb3326b3c8c9d62cf0b204b666015c8ab26a80b6927a6f933.jpg) + +![](images/e9ed58cab824c0e18a60c394a32fda2ce29ca15d7ccf4993db7e900bcfc5e95e.jpg) ++ Gaussian N. + +![](images/7e3040ff0174a0f7e975f3b7a859b5c7927eaa64c4d85f4c9a717ab9e01a3836.jpg) ++Re-sam.N. ++ Noise degree +Figure 1. The comparison of qualitative results on low-quality images with varying degradation levels from an unseen dataset. To generate images with different degradation levels, we progressively added Gaussian Noise, Re-sampling Noise, and more severe Gaussian noise to an image. Results indicate that the baseline SAM [27] shows limited robustness to degradation. Although RobustSAM [8] retains some resilience against simpler degradations, it struggles with more complex and unfamiliar degradations. In contrast, our method consistently demonstrates strong robustness across images of varying quality. + +Qualitative Comparison + +![](images/481518b376a084d4bb4abec12805a743984c20be8370bc25d0aee95030f3f499.jpg) + +![](images/d8e7fdb1725d0301d1f483cebac8e4392d8f4dbf7ec711cbe3918604439820fc.jpg) + +![](images/b1dcd6f329a53a9f575d3e4f9ebe4bd81371c1b788b1f744085f01caf98b8b91.jpg) + +![](images/5ea1ae54c609584cdabd1937716b2cbaae64f10948b611e01d3ae1a76ef2ab6e.jpg) +SAM + +![](images/02514ad182f5a777b20db3784277a65f9e6a6e1d0bbf10a5a930b08064a630f6.jpg) + +![](images/756bd690003e290240abcc807ff9bd37ab3261aa84daead529eaa111e0e52a5e.jpg) + +![](images/6a6af336cd5d389e2ef6ce89830339dc7283cfec0ee907f606e8e10035c49f35.jpg) + +![](images/a88fad9c423a596276dbac1ad85145c9460a24e164dcdd4986f655fb327c31a9.jpg) +RobustSAM + +![](images/6dd471993be710accc10c6fc1edd701fd625ad7fc422ba8e4d83ac2ff03e46a7.jpg) + +![](images/9ce9986e329d3f91904a88060855b5395511fa0d3e97a73ffe61f8dc41957b3e.jpg) + +![](images/06631936e528d876672503e883e2731b7eda4d666e017949ff93375a30798601.jpg) + +![](images/509235a06cc95ee61b7aa01bb076118dac9539c7aaa3077070815df0ed31572d.jpg) +GleSAM (Ours) + +Despite their success, SAMs perform poorly on common low-quality images, such as those degraded by noise, blur, and compression artifacts [22, 42, 47, 56], which are often encountered in real-world scenarios [19, 34, 35]. Previous methods [8, 11] have employed distillation-based consistent learning to enhance degradation-robust features. + +Nonetheless, they still face challenges in handling severely degraded low-quality images, as illustrated in Fig. 1. As degradations become more complex (e.g. combining various types of degradation or increasing the level of degradation), the existing SAMs [8, 27] struggle to accurately segment edges and complete target areas, leading to incorrect segmentation. We analyze that it is caused by the limited feature representation for degraded images. The visualizations in Fig. 2 reveal that SAM's latent features from severely degraded images contain excessive noise, compromising the original representations and subsequently impacting the predictions of the decoder. Furthermore, the large gap between low-quality and high-quality features complicates consistency learning [38] in previous works [8], hindering performance improvement. Thus, achieving high-quality latent feature representations and robust segmentation across varying image quality, especially for degraded images, remains challenging. + +The recently developed generative Diffusion Models (DM) [20, 50], especially the large-scale pre-trained Latent Diffusion Models (LDM) [45] have demonstrated powerful content generation capabilities. Having been trained on internet-scale data [46], LDM that proceed diffusion and denoising in latent space, possess powerful representation prior, which can be well explored to enhance the latent representation of segmentation models. This inspires us to take full advantage of the generative ability of pre-trained diffusion models and incorporate it into the latent space of SAMs to enhance low-quality features, thus promoting accurate segmentation in low-quality images. + +To this end, we propose GleSAM, which reconstructs high-quality features (Fig. 2 (d)) in SAM's latent space through generative diffusion, enabling accurate segmentation across any-quality images. Starting with low-quality features, high-quality representations are generated through single-step denoising. To integrate LDM generative knowledge, we incorporate a pre-trained U-Net from LDM with learnable LoRA layers [21] to align with segmentation-specific features. Furthermore, to improve compatibility between the pre-trained diffusion model and the segmentation framework, we introduce two effective techniques: Feature Distribution Alignment (FDA) and Channel Replicate and Expansion (CRE). These techniques bridge feature distribution and structural alignment gaps between models. Built upon SAM and SAM2, GleSAM leverages the generalization of pre-trained segmentation and diffusion models, with a few learnable parameters added, and can be efficiently trained within 30 hours on four GPUs. + +In terms of data, we constructed LQSeg based on existing datasets [9, 17, 31, 32, 49] to train and assess segmentation models on low-quality images. LQSeg incorporates a greater diversity of degradation types than previous methods [8], combining basic degradation models (e.g., + +![](images/2b9b1ca303cad7de372bff2d460425e9e50f0438deea06818f5d986e08316abc.jpg) + +![](images/b062ae780362b86fa98a8d5a9f209228110ce58506097f6d86da7e3de8b70343.jpg) +(a) LQ-Img + +![](images/fbf39ec8b1eb98d7da7db4dd0b9347e27387082ec9e54b9f6c2c1f684380fe18.jpg) + +![](images/09d9094775c7cd7c74313bda9670654047ef50614d319cd5e15d548937f0bbf4.jpg) +(b) LQ-Feat + +![](images/acca9c60fa52591e1da4e4befd04f3dec0649e835d1b5004d892319dce191ed6.jpg) + +![](images/5ff0fd44433a47b5afe7eb73377bd482b0304525e283909d73232dbae51a4299.jpg) +(c) HQ-Feat +Figure 2. The visualization of latent features: (a) low-quality (LQ) images, (b) the SAM's latent features extracted from LQ images, which contain excessive noise and compromise the original representations, (c) the high-quality (HQ) features of the corresponding clear images, which are more salient than LQ ones, and (d) enhanced representation by our GleSAM. + +![](images/a48059ccce8f174be1092efa234cc6a18201cd4515b8ed8e79aab810434df659.jpg) + +![](images/a3b764976eb83a4aff2ee0d30c53f4045441732af9e434828fa16e4dc7539d53.jpg) +(d) Ours-Feat + +noise and blur) to simulate complex and real-world noise [54, 63]. We also introduce three degradation levels for a more comprehensive evaluation. We hope LQSeg will inspire the development of more robust segmentation models and contribute to future research. Overall, our contributions are summarized as: + +- We propose GleSAM, a SAM-based framework incorporating generative latent space enhancement, to generalize across images of any quality. GleSAM exhibits significantly improved robustness, particularly for low-quality images with varying degradation levels. +- Two effective techniques: FDA and CRE, are introduced to bridge feature distribution and structural gaps between the pre-trained latent diffusion model and SAM. +- We also construct the LQSeg dataset which includes a wide range of degradation types and levels, to effectively train and evaluate the model. +- Extensive experiments show that our method performs excellently on low-quality images with varying degrees of degradation while maintaining generalization to clear images. Additionally, our method achieves strong performance on unseen degradations, highlighting the adaptability of both our framework and dataset. + +# 2. Related Work + +# 2.1. Segmentation on Low-Quality Images + +Executing robust segmentation across various scenarios is a critical issue. Numerous studies [8, 24, 43, 56] have highlighted significant performance degradation in conventional segmentation models and foundational SAMs when confronted with low-quality images with degradation. Many related studies [11, 14, 24, 29, 43] have been proposed to enhance the robustness of segmentation models against low- + +quality data. These methods primarily consider a single type of degradation. Recently, RobustSAM [8] is introduced to enhance the robustness of the SAM against multiple image degradations through anti-degradation feature learning. However, its performance also struggles when dealing with complex degradations. The real-world image noise is often too complex to be modeled by a single degradation [19, 34, 35]. Therefore, robustly segmenting images of any quality remains challenging. + +# 2.2. Diffusion Models for Perception Tasks + +Recently, diffusion models [20, 39, 45, 50, 62] have garnered significant attention in research, due to their powerful generation capabilities. Numerous studies [1, 2, 6, 23, 33, 40, 53, 60, 64] explore how to extend their applications to a broader range of tasks, such as detection, segmentation, and image reconstruction, etc. For diffusion-based perception tasks, one category of methods [1, 7, 58, 59] reformulate the perception tasks as progressive denoising from random noise. such as DiffusionDet [6] and DiffusionInst [6]. Another route employs the denoising autoencoder pre-trained on the text-to-image generation as a backbone for downstream perception tasks [13, 23, 28, 40, 64]. For example, VPD [64] passes the image through a pre-trained diffusion model and extracts intermediate features for task prediction. Diverging from these existing works, we preserve the original segmentation structure and fine-tune a generative diffusion to enhance the segmentation model's latent representations for accurate segmentation of any quality images. + +# 2.3. Segment Anything Model and Variants + +Segment Anything Models (SAMs) [27, 44] have gained significant influence within the community due to their outstanding zero-shot segmentation capabilities. SAM [27] can interactively segment any object in an image using visual prompts such as points and bounding boxes. Most recently, the updated SAM2 [44] has been released, showing improved segmentation accuracy and inference efficiency. Their generalization abilities have led to breakthroughs and new paradigms in various downstream tasks [25, 30, 36, 48, 52, 57, 61, 65]. Although SAM is powerful, its performance decreases when facing complex scenarios, such as degraded images [22, 42, 56] and adverse weather conditions [47], which significantly hinders the real-world applications of SAM. Enhancing SAM's capability in such challenging scenarios is a worthwhile research topic. + +# 3. Generative Latent Space Enhancement for Any-Quality Image Segmentation + +In the following, we explore how to improve SAM's robustness for low-quality images through generative latent space enhancement, thus enabling it to generalize across varying image qualities. The overall framework of the proposed + +GleSAM is shown in Fig. 3. To begin, in Sec. 3.1, we propose incorporating diffusion models' generative capabilities into SAM's latent space to effectively and efficiently enhance low-quality feature representations. Next, to improve the compatibility of feature distribution and architecture between the pre-trained diffusion model and SAM, we introduce two techniques: Feature Distribution Alignment and Channel Replicate and Expansion, which are detailed in Sec. 3.2 and Sec. 3.3, respectively. Finally, the overall training method is outlined in Sec. 3.4. + +# 3.1. Latent Denoising Diffusion in Segmentation + +Recall that diffusion models [20, 45, 50] are a class of probabilistic generative models that progressively add noise to the latent space, and then they learn to reverse this process by predicting and removing the noise. Formally, in LDMs, the forward noise process iteratively adds Gaussian noise with variance $\beta_{t} \in (0,1)$ to the variable $z$ . The sample at each time point is defined as: + +$$ +z _ {t} = \sqrt {\overline {{\alpha}} _ {t}} z + \sqrt {1 - \overline {{\alpha}} _ {t}} \epsilon , \qquad (1) +$$ + +where $\alpha_{t} = 1 - \beta_{t},\overline{\alpha}_{t} = \prod_{s = 1}^{t}\alpha_{s}$ , and $\epsilon \in \mathcal{N}(0,\mathrm{I})$ . While the inverse diffusion process is modeled by applying a neural network $\epsilon_{\theta}(z_t,t)$ to predict the noise $\hat{\epsilon}$ and recover the original input $\hat{z}$ . LDMs model the above process in a latent space using a pre-trained Variational AutoEncoder (VAE) and then up-sample the latent output to the original resolution using the VAE decoder, enabling more efficient computations in the training and inference phases. + +A similar idea motivates us to introduce the generative latent space denoising process into the SAMs' framework to reconstruct low-quality segmentation features. Let's denote $\mathcal{E}_{\theta}$ and $\mathcal{D}_{\theta}$ the segmentation encoder and decoder of SAMs, respectively. As shown in Fig. 3, the input LQ image is first compressed by $\mathcal{E}_{\theta}$ and generates LQ feature $z_{L}$ . We consider $z_{L}$ to be a noisy version of $z_{H}$ , containing sufficient information to reconstruct a high-quality feature. Instead of the complex multi-step denoising from random noise, we start directly from $z_{L}$ and forward with a single denoising step. Specifically, based on Eq. 1, the clean latent variable $z$ can be directly predicted from the model's predicted noise $\hat{\epsilon}$ , as: + +$$ +\hat {z} = \frac {z _ {t} - \sqrt {1 - \overline {{\alpha}} _ {t}} \hat {\epsilon}}{\sqrt {\overline {{\alpha}} _ {t}}}, \tag {2} +$$ + +where $\hat{\epsilon}$ is the prediction of the network $\epsilon_{\theta}$ with given $z_{t}$ and t: $\hat{\epsilon} = \epsilon_{\theta}(z_t;t)$ . We re-parameterize the above generative denoising process to adapt low-quality latent space enhancement in segmentation, as: + +$$ +\hat {z} _ {H} = \operatorname {G L E} \left(z _ {L}\right) = \frac {z _ {L} - \sqrt {1 - \bar {\alpha} _ {T}} \epsilon_ {\theta} \left(z _ {L} ; T\right)}{\sqrt {\bar {\alpha} _ {T}}}, \tag {3} +$$ + +where we consider low-quality feature $x_{L}$ as the noised feature and perform one-step denoising at the $T$ -th diffusion + +![](images/6693a538551966b2a531ee42df6c36d781d37e091add082ad8662a033e631f67.jpg) +Figure 3. Given an input image, GleSAM performs accurate segmentation through image encoding, generative latent space enhancement, and mask decoding. During training, HQ-LQ image pairs are fed into the frozen image encoder to extract the corresponding HQ and LQ latent features. We then reconstruct high-quality representations in the SAM's latent space by efficiently fine-tuning a generative denoising U-Net with LoRA. Subsequently, the decoder is fine-tuned with segmentation loss to align the enhanced latent representations. Built upon SAMs, GleSAM inherits prompt-based segmentation and performs well on images of any quality. + +timestep. The denoised output $\hat{z}_H$ is expected to be closer to the features extracted from clear images $z_{H}$ . This single-step process significantly reduces computational overhead, making it more efficient when applied to segmentation models. After that, with $\hat{z}_H$ as input, the mask decoder can predict more precise masks, as: $m_p = \mathcal{D}_{\theta}(\hat{z}_H)$ + +# 3.2. Feature Distribution Alignment + +We employ the pre-trained U-Net in LDM as the denoising backbone. However, a significant challenge arises due to the substantial difference between the latent spaces in the original LDM (encoded by VAE) and segmentation models, leading to several technical issues for our application. + +Firstly, there is a distribution gap between the two spaces and directly feeding segmentation features into U-Net may prevent it from fully exerting its denoising capabilities, as shown in the right part of Fig. 3. To address this gap, we introduce a Feature Distribution Alignment (FDA) technique. Specifically, we add an adaptation weight $\gamma$ to scale the segmentation features, adjusting their variance to align more closely with the VAE's latent space. This adjustment ensures that the features are compatible with U-Net's optimal input space, improving the robustness and accuracy of the semantic interpretation and enhancing the denoising capability. The LQ feature denoising process in Eq. 3 can be updated as: + +$$ +\hat {z} _ {H} = \mathrm {G L E} (z _ {L}) = \frac {\gamma z _ {L} - \sqrt {1 - \overline {{\alpha}} _ {T}} \epsilon_ {\theta} (\gamma z _ {L} ; T)}{\gamma \sqrt {\overline {{\alpha}} _ {T}}}, \quad (4) +$$ + +where we divide by $\gamma$ to restore its original distribution. We experimentally verified in Sec. 5.4 that this simple operation effectively improves U-Net's denoising performance when applied to segmentation features. + +# 3.3. Channel Expansion for Head-tail Layers + +Another technical issue arises from the channel mismatch of the head and tail layers between the pre-trained U-Net and the segmentation features. The U-Net in LDMs is designed for 4-channel input and output ( $h \times w \times 4$ ), which does not match the dimension of SAM's latent space ( $h \times w \times 256$ ). We explore various methods to solve this problem (in Sec. 5.4) and empirically find that fine-tuning new head and tail layers or an encoder-decoder for segmentation features is ineffective. This is likely due to difficulties in aligning with the pre-trained model's parameters while preserving its generalization ability. To address this, we propose a Channel Replication and Expansion (CRE) method that replicates and concatenates the pre-trained weights of head and tail layers to match the required channel dimension. During training, the parameters of the head and tail layers remain frozen, while learnable LoRA layers are added to adapt to segmentation features. This approach effectively preserves the pre-trained generalization capacity while minimizing the number of learnable parameters. + +# 3.4. Training Method + +We employ a two-step fine-tuning process. In the first step, we fine-tune the denoising U-Net to reconstruct high + +quality features. In the second step, we fine-tune the decoder with the restored features to further align the feature space for more accurate segmentation. + +U-Net finetuning. To adapt the pre-trained U-Net to the segmentation framework while preserving its inherent generalization ability, we employ the LoRA [21] scheme to fine-tune all the attention layers in the U-Net. During this step, we freeze the pre-trained image encoder and U-Net layers and only fine-tune the added LoRA layers. The estimated feature is compared with the corresponding HQ feature $z_{H}$ by a reconstruction loss, as: + +$$ +\mathcal {L} _ {\mathrm {R e c}} = \mathcal {L} _ {\mathrm {M S E}} (\mathrm {G L E} \left(z _ {L}\right), z _ {H}). \tag {5} +$$ + +This step significantly enhances performance without fine-tuning SAM's parameters. + +Decoder finetuning. Next, we use the reconstructed high-quality features to fine-tune the mask decoder for more precise segmentation. Our experiments demonstrate that finetuning either the entire decoder or only the output tokens with these features further improves segmentation accuracy while maintaining generalization on clear images. Focal Loss and Dice Loss are employed as segmentation loss functions, as: + +$$ +\mathcal {L} _ {\text {S e g}} = \mathcal {L} _ {\text {D i c e}} \left(m _ {p}, m _ {g}\right) + \mathcal {L} _ {\text {F o c a l}} \left(m _ {p}, m _ {g}\right), \tag {6} +$$ + +where $m_p$ and $m_g$ indicate predicted and ground-truth masks respectively. + +# 4. Low-Quality Image Segmentation Dataset + +We construct a comprehensive low-quality image segmentation dataset dubbed $LQSeg$ that encompasses more complex and multi-level degradations, rather than relying on a single type of degradation for each image. The dataset is composed of images from several existing datasets with our synthesized degradations. In this section, we first model a multi-level degradation process of low-quality images and then detail the dataset composition. + +# 4.1. Multi-level Degradation Modeling + +To model a more practical and complex degradation process, inspired by the previous work in image reconstruction [54, 63], we utilize a mixed degradation method. Specifically, the degraded process is modeled as the random combination of the four common degradation models, including Blur, Random Resize, Noise, and JPEG Compression. Each degradation model encompasses various types, such as Gaussian and Poisson noise for Noise, ensuring the diversity of the degradation process. + +To enrich the granularity of degradation, we employ multi-level degradation by adjusting the downsampling rates. We employed three different resize rates, i.e., [1, 2, 4], which correspond to three degradation levels from slight + +to severe: LQ-1, LQ-2, and LQ-3. More implementation details are illustrated in Supplementary Material. + +# 4.2. Dataset Composition + +Based on the above multi-level degradation model, we construct LQSeg to train our model and evaluate the segmentation performance on different levels of low-quality images. The images in LQSeg are sourced from several well-known existing datasets in the community with our synthesized degradation. In detail, for the training set, we utilize the entire training sets of LVIS [17], ThinObject-5K [31], and MSRA10K [9] as the source data and procedurally synthesize corresponding low-quality images. The evaluation set is sourced from four subsets, i.e., 1) the whole test sets of ThinObject-5K and LVIS (seen sets), and 2) ECSSD [49] and COCO-val [32] (unseen sets). For each source image, We systematically generate three levels of degraded images to thoroughly assess the model's robustness. + +# 5. Experiment + +We conduct extensive experiments to verify our method across images of varying quality. All proposed techniques can be applied to SAM and SAM2, referred to as GleSAM and GleSAM2. In practice, our models perform well on low-quality (Tab. 1, 2) and clear images (Tab. 5) and they generalize effectively to unseen degradations (Tab. 3). + +# 5.1. Experimental Setup + +Implement Details. Our model is trained using the AdamW optimizer with the learning rate of $1 \times 10^{-4}$ and batch size of 4. The pre-trained U-Net in Stable Diffusion (SD) 2.1-base [45] is adopted as the denoising backbone. We use 3 clicks as SAM's prompts by default. Our approach can be efficiently trained on $4 \times$ A100 GPUs within approximately 30 hours, during which we fine-tune the U-Net for 100K iterations and the decoder for only 20K iterations. + +Evaluation Metrics. We employ three metrics to assess our model's performance, including Intersection over Union (IoU), Dice Coefficient (Dice), and Pixel Accuracy (PA). + +# 5.2. Performance Comparisons + +In this experiment, we evaluate the performance on the test set of our LQSeg, including seen-set (Tab. 1) and unseen-set (Tab. 2) evaluations. We compare our method with a set of comparison baselines to quantify the performance gains. For SAM-based comparisons, besides the original SAM [27], we also compare with the RobustSAM [8], which has improved robustness on the degraded dataset. Additionally, we compare with two-stage methods, i.e., reconstructing images first with image reconstruction (IR) networks and passing the restored clear images to the SAM and SAM2. We use two state-of-the-art IR networks for comparison: PromptIR [41], and diffusion-based DiffBIR [33]. + +
MethodThinObject-5KLVIS
LQ-3LQ-2LQ-1LQ-3LQ-2LQ-1
IoUDiceIoUDiceIoUDiceIoUDiceIoUDiceIoUDice
SAM0.62850.72860.70540.79390.75270.83430.40410.50050.48860.58380.53250.6282
RobustSAM0.70150.79650.76480.84630.79220.86580.45170.56700.49620.60790.52620.6356
PromptIR-SAM0.63800.73740.71460.80180.74340.82480.40200.49780.47050.56770.52220.6187
DiffBIR-SAM0.70550.79170.77720.85310.79270.86520.53160.63070.58120.69130.60210.7090
GleSAM (Ours)0.75940.84130.80330.87380.82770.89200.55350.67560.59160.70660.61310.7244
SAM20.71740.80000.76360.83730.78390.85360.51180.61740.56340.66330.60240.7004
PromptIR-SAM20.71190.79450.77530.84770.78010.84870.50170.60840.55290.65460.58750.6865
DiffBIR-SAM20.73480.81170.78320.85050.79740.86040.56510.66640.60040.70320.62780.7380
GleSAM2 (Ours)0.79470.86530.83000.88960.85270.90720.57380.68870.60820.71610.63610.7402
+ +Table 1. Performance comparison on the test set of Thinobject-5K [31] and LVIS [17] datasets (seen datasets) with different levels of degradation. From LQ-1 to LQ-3, the degree of degradation increases progressively. We report IoU and Dice for comparison. Our GleSAM and GleSAM2 consistently outperform other competitors, especially on the most challenging LQ-3 version. The words with boldface indicate the best results and those underlined indicate the second-best results. + +
MethodECSSDCOCO
LQ-3LQ-2LQ-1LQ-3LQ-2LQ-1
IoUDiceIoUDiceIoUDiceIoUDiceIoUDiceIoUDice
SAM0.52190.63830.61560.72020.67780.77350.40480.50140.48480.58100.53920.6350
RobustSAM0.65350.76590.72710.82400.77580.86060.45410.57110.49560.60870.53040.6401
PromptIR-SAM0.52770.64420.61090.71880.67520.77260.40060.49690.47840.57530.52570.6219
DiffBIR-SAM0.66750.76520.74080.82440.77370.85010.52600.62630.57420.69180.61120.7080
GleSAM (Ours)0.73320.82820.79440.87190.82570.89310.54950.67050.59000.70550.61580.7257
SAM20.65340.75400.72920.81570.76250.84170.51760.62300.56810.66920.60180.7000
PromptIR-SAM20.64900.75260.71750.80820.75190.83510.50880.61530.55790.66000.59050.6914
DiffBIR-SAM20.71580.80260.78530.85880.80700.87450.56800.67110.61090.71900.62450.7348
GleSAM2 (Ours)0.73350.82580.79760.87260.81750.88530.57240.68670.61650.72460.63370.7371
+ +Table 2. Zero-shot performance comparison on the ECSSD [49] and COCO [32] datasets (unseen datasets) with different levels of degradation. These results indicate that GleSAMs possesses significant robustness in zero-shot segmentation across different levels of degradations. + +Performance Comparison on Seen Datasets. In Tab. 1, we evaluate the performance of our GleSAM and GleSAM2 on two seen datasets: ThinObject-5K [31] and LVIS [17]. Each dataset contains three levels of degradation. Our method demonstrates superior performance across all degradation levels, effectively handling low-quality images. As the degradation level increases (from LQ-1 to LQ-3), GleSAMs show increasingly significant performance improvements compared to the baseline, highlighting its robustness against challenging degradations. + +Performance Comparison on Un-seen Datasets. In Tab. 2, we evaluate the zero-shot segmentation performance of GleSAMs on two unseen datasets: ECSSD [49] and COCO [32], all of which are synthesized with different levels of degradations. GleSAM consistently outperforms other methods, particularly on the most challenging LQ-3 version, underscoring its strong zero-shot generalization capabilities and potential for real-world applications. To further assess segmentation quality, we plot the density distribution maps of image quality and segmentation IoU in Fig. 4. Compared to the baselines, our method achieves overall + +improvement and more stable performance on low-quality images, especially for those of inferior quality. + +Validation with Other Degradations. To validate the model's generalization on other unseen degradations, we evaluated GleSAM and GleSAM2 on RobustSeg-style degradations [8], with the results presented in Tab. 3. These degradations were not used during training. Our method consistently outperforms SAM and SAM2 and even surpasses RobustSAM which is specifically trained on RobustSeg. This demonstrates the strong generalization of our method across diverse degradations. + +# 5.3. Ablation Study + +We conduct ablation experiments in Tab. 4, 5 to further understand the impact of each component of our framework. + +Effect of Each Component. In Tab. 4, we validated the effectiveness of each proposed module in GleSAM. "Gle" indicates our framework with generative latent space enhancement. The results show that each introduced method significantly improves the performance. We also make a qualitative visualization in Fig. 5, which shows that the + +![](images/0c996d3d21f3661365ab448bc827b0ec838ee9a368c3c8e19c9097f74ba29799.jpg) +Figure 4. Density distribution maps about IoU and image quality across different methods, including SAM, GleSAM, SAM2, and GleSAM2. The image quality is calculated using the Laplacian operator in OpenCV. The red dashed box highlights the area where our method demonstrates improved segmentation performance compared to SAM, particularly in lower-quality images. + +![](images/465ce12e70e8a50e0668cb3421a74c91c0224b0ced1a28ba5fb880433f7f62cb.jpg) + +![](images/2c547d0664123531e0f3eeefde53684a458885be159343ef2de948aedb418a47.jpg) + +![](images/74f8d94249e386f2186780ac9c9141a202b06049b107495177876069ff2b98c3.jpg) + +
MethodECSSDCOCO
IoUDiceIoUDice
SAM [27]0.78330.85730.65120.7457
RobustSAM [8]0.84270.90440.65430.7518
GleSAM (Ours)0.85680.91190.66780.7700
SAM20.82110.88170.68170.7753
GleSAM2 (Ours)0.86110.91330.68690.7836
+ +Table 3. Zero-shot performance comparison on RobustSeg-style [8] degradations. Performance is tested on the unseen ECSSD and COCO datasets. Note that our methods are not trained on such degradations. The superior performance of our method demonstrates robustness against various forms of degradation. + +
MethodECSSDCOCO
IoUDiceIoUDice
Baseline0.60540.71070.47630.5725
+ Gle & CRE0.65670.76570.49580.5963
+ Gle & CRE & FDA0.71040.80200.51660.6174
+ +clearest latent feature is obtained when combining all modules. Additionally, in Tab. 5, without fine-tuning the decoder, our method improves the results on LQ images by about 7 points, while also preserving the robustness of SAM on clear conditions, indicating the robustness of our method for both degraded and clear images. + +Effect of Fine-tuning SAM. We explore two common configurations to fine-tune SAM: fine-tuning the entire SAM's decoder and the output mask token. Our experiments reveal that directly fine-tuning SAM's decoder or output mask token on degraded images leads to a significant drop in zero-shot performance on clear images, with the IoU decreasing by nearly 20 points. In contrast, our method further improves performance on both low-quality and clear images by fine-tuning the entire decoder or mask token, enabling the segmentation of images with any quality. + +Table 4. Ablation study of each component in the proposed method, evaluated on the unseen ECSSD and COCO datasets. Each additional component positively affects the performance, demonstrating the effectiveness of the proposed methods. + +
MethodLQClearAverage
IoUDiceIoUDiceIoUDice
w/o Fine-tuning SAM:
SAM0.54070.64160.78300.85540.66190.7485
Ours0.61350.70970.78460.85670.69910.7832
+ +
Fine-tuning SAM:
SAM-FT-T0.63050.73740.58470.70290.60760.7202
SAM-FT-D0.63270.73850.60710.72420.61990.7314
Ours-FT-T0.67590.77510.80610.87470.74100.8249
Ours-FT-D0.68480.78250.80220.87040.74350.8265
+ +Table 5. Effect of Fine-tuning SAM. The performance is evaluated on the unseen ECSSD and COCO datasets. "FT-T" and "FT-D" indicate fine-tuning the SAM's mask token and decoder respectively. "LQ" indicates the mean performance on three levels of degraded data, and "Clear" indicates the results on the original clear images. + +![](images/f9c72fd87fb2dba082fce58ccef92fa4832cfb341ecd0eb56148ade9d01e6aba.jpg) +Figure 5. Qualitative visualization of the enhanced latent features. The clearest feature is obtained when combining all modules. + +# 5.4. Detailed Analysis + +Analysis of channel expansion methods. To address the issue of channel mismatch between SAM and the pretrained U-Net, we explored various strategies including (a) using two simple convolutional layers to reduce and expand the channels of segmentation features as needed, (b) fine-tuning new head and tail layers from scratch, and (c) our CRE method. Our results (in Tab. 6 and Fig. 7) show that strategies (a) and (b) are ineffective, likely because the new layers couldn't leverage the pre-trained knowledge. In contrast, our method effectively resolves this issue by replicating pre-trained parameters and fine-tuning with LoRA. + +Analysis of hyperparameter $\gamma$ . We use an adaption weight + +![](images/c5dcb088b2ae42449b2edaaaa54ffa86ac76c4bff0c288406cab5864c2ce65a1.jpg) +(a) ECSSD +Figure 6. Ablation study of adaption weight $\gamma$ . + +![](images/632d40407e9f22afaf7185829e5b57ebaa1f8edfafdd61894c580f085849e0f6.jpg) +(b) COCO + +
MethodIoUDicePA
(a) Additional encoder and decoder0.45440.58420.6106
(b) New head and tail layers0.60140.70770.7782
(c) Replicate and Expansion (Ours)0.65670.76570.8400
+ +![](images/4922927e210e8414e17ecf1e0236673027c312a734367e9dcf7f4a0b4a2ce36f.jpg) +Figure 7. Qualitative visualization of the enhanced latent features generated by different channel expansion methods. Our proposed CRE method (c) produces more salient features. + +![](images/a3a6d01c799d9ac63d2a4d2096798ae342ba7d5ef090064ad311df72455583a4.jpg) + +![](images/1a7d02d70d6c49e5c04377da352773dc6dd6603ceeb9fe11241e687c1a818191.jpg) + +![](images/cbaa8fb398c1ccb6e323099f35b1b2a0b0b00d0e2a2429efd6b58976d8c9c132.jpg) + +$\gamma$ to align the distribution of the latent space between LDM and SAM. To determine the optimal value of $\gamma$ , we empirically test five different values on the ECSSD and COCO datasets. The results, shown in Fig. 6, suggest that $\gamma = 5$ is the most effective setting, providing strong generalization across all models and datasets. Therefore, we adopt $\gamma = 5$ as the default value in our experiments. + +Analysis of LoRA ranks. We incorporate learnable LoRA layers to fine-tune the pre-trained denoising U-Net. Here, we evaluate the impact of different LoRA ranks on segmentation performance. The results are shown in Tab. 8. We tested the results and the corresponding number of learnable parameters for setting a rank to 4, 8, and 16. We found that setting the rank to 8 can obtain good results while maintaining an acceptable number of parameters. + +Analysis of degradation levels. To investigate the necessity of training on all three degradation levels together, we train GleSAM on LQ datasets for each degradation level (LQ-1, LQ-2, LQ-3) individually. We evaluate the models' performance on ECSSD and COCO datasets using three degraded levels and RobustSeg-style degradations. The results in Tab. 7 show that training on all three degrada + +Table 6. Analysis of the proposed CRE. It significantly outperforms alternative approaches, achieving higher scores. + +
Training DataOursLQ-RSAVG
LQ-1LQ-2LQ-3
LQ-10.69810.66630.60760.75870.6827
LQ-20.71680.68430.63180.76210.6988
LQ-30.71850.69100.64030.76190.7030
LQ-1,2,30.72080.69220.64140.76230.7042
+ +Table 7. Ablation study for degradation levels during training. LQ-RS indicates the RobustSeg-style [8] degradation. AVG indicates the average performance. We report the IoU for comparison. The results on the unseen ECSSD and COCO datasets reveal that multi-level degradation contributes to more robust performance. + +
RankLQLearnable +Params
IoUDicePA
40.77600.85970.943416.25M
80.78440.86440.945632.49M
160.76970.85430.941864.99M
+ +Table 8. Analysis of LoRA ranks in U-Net. + +
MethodLearnable ParamsNum. GPUsTraining Time (h)Inference Speed (s)
SAM1250 MB256N/A0.32
GleSAM47 MB4300.38 (+0.06)
+ +Table 9. Computational requirements of GleSAM vs SAM. + +tion levels together contributes to more robust performance across varying levels of image degradation. + +Analysis of computational requirements. In Tab. 9, we report detailed training and inference comparisons between our GleSAM and SAM. Although GleSAM demonstrates significantly improved robustness, it introduces only marginal learnable parameters and incurs a slight trade-off in inference speed. The additional parameters can be efficiently optimized in 30 hours on four A100 GPUs. + +# 6. Conclusion + +We present GleSAM, a solution to enhance Segment Anything Models (SAM and SAM2) for robust segmentation across images of any quality, particularly those with severe degradation. We incorporate the generative ability of pretrained diffusion models into the latent space of SAMs to enhance low-quality features, thus promoting more robust segmentation. Our approach is further supported by the LQSeg dataset, which includes diverse degradation types and levels, allowing for more comprehensive model training and evaluation. Extensive experiments demonstrate that GleSAM achieves superior performance on degraded images while maintaining generalization to clear ones, and it performs well on unseen degradations, highlighting its robustness and versatility. This work extends SAM capabilities, offering a practical approach for degraded scenarios. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China (NSFC) under Grant 62372382. + +# References + +[1] Tomer Amit, Tal Shaharbany, Eliya Nachmani, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021. 3 +[2] Dmitry Baranchuk, Andrey Voynov, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. In International Conference on Learning Representations, 2022. 3 +[3] Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. Just a hint: Point-supervised camouflaged object detection. In European Conference on Computer Vision, pages 332-348, 2024. 1 +[4] Huafeng Chen, Pengxu Wei, Guangqian Guo, and Shan Gao. Sam-cod: Sam-guided unified framework for weakly-supervised camouflaged object detection. In European Conference on Computer Vision, pages 315-331, 2024. 1 +[5] Keyan Chen, Chenyang Liu, Hao Chen, Haotian Zhang, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsprompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1 +[6] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusionnet: Diffusion model for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 19830-19843, 2023. 3 +[7] Ting Chen, Lala Li, Saurabh Saxena, Geoffrey Hinton, and David J Fleet. A generalist framework for panoptic segmentation of images and videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 909-919, 2023. 3 +[8] Wei-Ting Chen, Yu-Jiet Vong, Sy-Yen Kuo, Sizhou Ma, and Jian Wang. Robustsam: Segment anything robustly on degraded images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4081-4091, 2024. 1, 2, 3, 5, 6, 7, 8 +[9] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence, 37(3):569-582, 2014. 2, 5 +[10] Lei Ding, Kun Zhu, Daifeng Peng, Hao Tang, Kuiwu Yang, and Lorenzo Bruzzone. Adapting segment anything model for change detection in vhr remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1 +[11] Kazuki Endo, Masayuki Tanaka, and Masatoshi Okutomi. Semantic segmentation of degraded images using layer-wise feature adjustor. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3205-3213, 2023. 1, 2 +[12] Shan Gao, Guangqian Guo, Hanqiao Huang, and CL Philip Chen. Go deep or broad? exploit hybrid network architecture for weakly supervised object classification and localization. + +IEEE Transactions on Neural Networks and Learning Systems, 2023. 1 +[13] Rui Gong, Martin Danelljan, Han Sun, Julio Delgado Mangas, and Luc Van Gool. Prompting diffusion representations for cross-domain semantic segmentation. arXiv preprint arXiv:2307.02138, 2023. 3 +[14] Dazhou Guo, Yanting Pei, Kang Zheng, Hongkai Yu, Yuhang Lu, and Song Wang. Degraded image semantic segmentation with dense-gram networks. IEEE Transactions on Image Processing, 29:782-795, 2019. 2 +[15] Guangqian Guo, Pengfei Chen, Xuehui Yu, Zhenjun Han, Qixiang Ye, and Shan Gao. Save the tiny, save the all: hierarchical activation network for tiny object detection. IEEE transactions on circuits and systems for video technology, 34 (1):221-234, 2023. 1 +[16] Guangqian Guo, Dian Shao, Chenguang Zhu, Sha Meng, Xuan Wang, and Shan Gao. P2p: Transforming from point supervision to explicit visual prompt for object detection and segmentation. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2024. 1 +[17] 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. 2, 5, 6 +[18] 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. 1 +[19] Glenn E Healey and Raghava Kondepudy. Radiometric ccd camera calibration and noise estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 16(3):267-276, 1994. 1, 3 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, pages 6840-6851, 2020. 2, 3 +[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. 2, 5 +[22] Yihao Huang, Yue Cao, Tianlin Li, Felix Juefei-Xu, Di Lin, Ivor W Tsang, Yang Liu, and Qing Guo. On the robustness of segment anything. arXiv preprint arXiv:2305.16220, 2023. 1, 3 +[23] Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, and Ping Luo. Ddp: Diffusion model for dense visual prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21741-21752, 2023. 3 +[24] Christoph Kamann and Carsten Rother. Benchmarking the robustness of semantic segmentation models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8828-8838, 2020. 2 +[25] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 3 +[26] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceed- + +ings of the IEEE conference on computer vision and pattern recognition, pages 9404-9413, 2019. 1 +[27] 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. arXiv preprint arXiv:2304.02643, 2023. 1, 2, 3, 5, 7 +[28] Neehar Kondapaneni, Markus Marks, Manuel Knott, Rogério Guimaraes, and Pietro Perona. Text-image alignment for diffusion-based perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13883-13893, 2024. 3 +[29] Sohyun Lee, Taeyoung Son, and Suha Kwak. Fifo: Learning fog-invariant features for foggy scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18911-18921, 2022. 2 +[30] Bo Li, Haoke Xiao, and Lv Tang. Asam: Boosting segment anything model with adversarial tuning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3699-3710, 2024. 3 +[31] Jun Hao Liew, Scott Cohen, Brian Price, Long Mai, and Ji-ashi Feng. Deep interactive thin object selection. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 305–314, 2021. 2, 5, 6 +[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2, 5, 6 +[33] Xinqi Lin, Jingwen He, Zihan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diffir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 3, 5 +[34] Ce Liu, Richard Szeliski, Sing Bing Kang, C Lawrence Zitnick, and William T Freeman. Automatic estimation and removal of noise from a single image. IEEE transactions on pattern analysis and machine intelligence, 30(2):299-314, 2007. 1, 3 +[35] Jan Lukas, Jessica Fridrich, and Miroslav Goljan. Digital camera identification from sensor pattern noise. IEEE Transactions on Information Forensics and Security, 1(2): 205-214, 2006. 1, 3 +[36] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 3 +[37] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. 1 +[38] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In Proceedings of the AAAI conference on artificial intelligence, pages 5191-5198, 2020. 2 +[39] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distilla + +tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[40] Koutilya Pnvr, Bharat Singh, Pallabi Ghosh, Behjat Siddique, and David Jacobs. Ld-znet: A latent diffusion approach for text-based image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4157-4168, 2023. 3 +[41] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2024. 5 +[42] Yu Qiao, Chaoning Zhang, Taegoo Kang, Donghun Kim, Chenshuang Zhang, and Choong Seon Hong. Robustness of sam: Segment anything under corruptions and beyond. arXiv preprint arXiv:2306.07713, 2023. 1, 3 +[43] AN Rajagopalan et al. Improving robustness of semantic segmentation to motion-blur using class-centric augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10470-10479, 2023. 2 +[44] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 3 +[45] 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. 2, 3, 5 +[46] 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. 2 +[47] Xinru Shan and Chaoning Zhang. Robustness of segment anything model (sam) for autonomous driving in adverse weather conditions. arXiv preprint arXiv:2306.13290, 2023. 1, 3 +[48] Chuyun Shen, Wenhao Li, Yuhang Shi, and Xiangfeng Wang. Interactive 3d medical image segmentation with sam 2. arXiv preprint arXiv:2408.02635, 2024. 3 +[49] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence, 38(4): 717-729, 2015. 2, 5, 6 +[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 3 +[51] Chaowei Wang, Guangqian Guo, Chang Liu, Dian Shao, and Shan Gao. Effective rotate: Learning rotation-robust prototype for aerial object detection. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1 +[52] Di Wang, Jing Zhang, Bo Du, Minqiang Xu, Lin Liu, Dacheng Tao, and Liangpei Zhang. Samrs: Scaling-up remote sensing segmentation dataset with segment anything + +model. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3 +[53] Mengyu Wang, Henghui Ding, Jun Hao Liew, Jiajun Liu, Yao Zhao, and Yunchao Wei. Segrefiner: Towards model-agnostic segmentation refinement with discrete diffusion process. Advances in Neural Information Processing Systems, 36:79761-79780, 2023. 3 +[54] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 2, 5 +[55] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8741-8750, 2021. 1 +[56] Yuqing Wang, Yun Zhao, and Linda Petzold. An empirical study on the robustness of the segment anything model (sam). Pattern Recognition, page 110685, 2024. 1, 2, 3 +[57] Zhaoyang Wei, Pengfei Chen, Xuehui Yu, Guorong Li, Jianbin Jiao, and Zhenjun Han. Semantic-aware sam for point-prompted instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3585-3594, 2024. 3 +[58] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, Yehui Yang, Haoyi Xiong, Huiying Liu, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. In Medical Imaging with Deep Learning, pages 1623-1639. PMLR, 2024. 3 +[59] Junde Wu, Wei Ji, Huazhu Fu, Min Xu, Yueming Jin, and Yanwu Xu. Medsegdiff-v2: Diffusion-based medical image segmentation with transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6030-6038, 2024. 3 +[60] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 3 +[61] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In European Conference on Computer Vision, pages 162-179. Springer, 2025. 3 +[62] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[63] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 2, 5 +[64] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 3 + +[65] Jiayuan Zhu, Yunli Qi, and Junde Wu. Medical sam 2: Segment medical images as video via segment anything model 2. arXiv preprint arXiv:2408.00874, 2024. 3 \ No newline at end of file diff --git a/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/images.zip b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..90a687296986a74cf55f34e7d33864a3336a197d --- /dev/null +++ b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f38d5c0c3c47e1c053c67fbd17dfd445ac3099a61fadb57acbdc9119f1543ad +size 816221 diff --git a/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/layout.json b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..191b8bcdc8a81761bab4c0fcf59b3c3a12d60446 --- /dev/null +++ b/CVPR/2025/Segment Any-Quality Images with Generative Latent Space Enhancement/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c27059ef944cac7d9bffd8253285a7ba29c9e148bfbe1872e7c8224597b7c7fc +size 448042 diff --git a/CVPR/2025/Segment Anything, Even Occluded/91f2ed4a-96ac-4e81-9860-50746edbae3a_content_list.json b/CVPR/2025/Segment Anything, Even Occluded/91f2ed4a-96ac-4e81-9860-50746edbae3a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..58b1121bd789d84da445c2e2cf1e7d78d8330d8a --- /dev/null +++ b/CVPR/2025/Segment Anything, Even Occluded/91f2ed4a-96ac-4e81-9860-50746edbae3a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20cfe4de3070c6b72a3fc34aed189c5aa160d02cd2e63e553d903a279124e9a9 +size 69133 diff --git a/CVPR/2025/Segment Anything, Even Occluded/91f2ed4a-96ac-4e81-9860-50746edbae3a_model.json b/CVPR/2025/Segment Anything, Even Occluded/91f2ed4a-96ac-4e81-9860-50746edbae3a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9339efb3bdabd40ba38df27162f6dd002b942a96 --- /dev/null +++ b/CVPR/2025/Segment Anything, Even Occluded/91f2ed4a-96ac-4e81-9860-50746edbae3a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c122d4219b30068bd9b4dc623e62ea9b41ec209b58a411edeaad390ad179f72 +size 84054 diff --git a/CVPR/2025/Segment Anything, Even Occluded/91f2ed4a-96ac-4e81-9860-50746edbae3a_origin.pdf b/CVPR/2025/Segment Anything, Even Occluded/91f2ed4a-96ac-4e81-9860-50746edbae3a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8b23268bc8f64c06e8764dc07296145d7eebdab5 --- /dev/null +++ b/CVPR/2025/Segment Anything, Even Occluded/91f2ed4a-96ac-4e81-9860-50746edbae3a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74cd838b8a736ff276f52e4c53c1de0122f0dc71e4b75f026599e06604dfe07b +size 8872122 diff --git a/CVPR/2025/Segment Anything, Even Occluded/full.md b/CVPR/2025/Segment Anything, Even Occluded/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e96d5cb635192296fb6134e57ccb39ec4d1b1650 --- /dev/null +++ b/CVPR/2025/Segment Anything, Even Occluded/full.md @@ -0,0 +1,270 @@ +# Segment Anything, Even Occluded + +Wei-En Tai $^{1*}$ Yu-Lin Shih $^{1*}$ Cheng Sun $^{2}$ + +$^{1}$ National Tsing Hua University $^{2}$ NVIDIA + +Yu-Chiang Frank Wang $^{2,3}$ Hwann-Tzong Chen $^{1,4}$ + +$^{3}$ National Taiwan University $^{4}$ Aeolus Robotics + +# Abstract + +Amodal instance segmentation, which aims to detect and segment both visible and invisible parts of objects in images, plays a crucial role in various applications, including autonomous driving, robotic manipulation, and scene understanding. While existing methods require training both front-end detectors and mask decoders jointly, this approach lacks flexibility and fails to leverage the strengths of pre-existing modal detectors. To address this limitation, we propose SAMEO, a novel framework that adapts the Segment Anything Model (SAM) as a versatile mask decoder capable of interfacing with various front-end detectors to enable mask prediction even for partially occluded objects. Acknowledging the constraints of limited amodal segmentation datasets, we introduce Amodal-LVIS, a large-scale synthetic dataset comprising 300K images derived from the modal LVIS and LLVIS datasets. This dataset significantly expands the training data available for amodal segmentation research. Our experimental results demonstrate that our approach, when trained on the newly extended dataset, including Amodal-LVIS, achieves remarkable zero-shot performance on both COCOA-cls and D2SA benchmarks, highlighting its potential for generalization to unseen scenarios. + +# 1. Introduction + +Human visual perception extends beyond what is directly visible in a scene. We can naturally imagine and understand the complete shape of partially occluded objects through a combination of object recognition and prior knowledge about object categories. Even when correctly classifying objects is difficult, we can often infer the complete shape of partially visible objects by analyzing visible parts and reasoning about common occlusion patterns [9, 20, 24]. Amodal instance segmentation seeks to replicate this remarkable human capability by detecting and localizing objects in images and predicting their complete shapes, including both visible and occluded portions (Figure 1). + +![](images/28277ee719419dd090a46267b1b9114237aeb47cf2765c19693e60cd623b1098.jpg) +Figure 1. Amodal segmentation examples: The top row shows the original images. The middle row displays EfficientSAM predicted modal masks that only cover the visible parts of objects. The bottom row illustrates amodal masks that reveal the complete object shapes predicted by our method, SAMEO—a Segment Anything Model Even under Occlusion. + +![](images/0c942627b7d3b020fc2a00fc815876c0af364419260f3f03d048e8cb4ae1a983.jpg) + +![](images/12bbd276a74802f1d3957a7f6b961f28aa93c7dc1d4b08ce15199756a666f740.jpg) + +![](images/1210fba3f247425aca246803e5b46fe029ccc69c102389cb11aab2d4f2af74ba.jpg) + +An effective approach to addressing amodal instance segmentation is to divide the task into two main components: object detection and mask segmentation. In recent years, significant advances have been made in object detection, with state-of-the-art models such as RTMDet [16] and ConvNeXt-V2 [28] achieving impressive performance. However, current amodal segmentation approaches often require training both the detector and mask decoder jointly, which prevents them from fully utilizing these powerful pre-trained modal detectors. This limitation motivated us to develop a more flexible framework that can leverage existing modal detectors while still maintaining strong amodal segmentation capabilities. The emergence of foundation models for visual understanding has opened up new possibilities in segmentation tasks. Among these, the Segment Anything Model (SAM) [11] and its efficient variant, EfficientSAM [29], have demonstrated remarkable capabilities in prompt-based modal segmentation. We leverage EfficientSAM's architecture, which features a lightweight encoder for faster inference, and adapt it for amodal segmen + +tation through specialized training. Our approach enables the model to process both amodal and modal prompts for generating amodal mask predictions while maintaining potential zero-shot capabilities. Besides the improvements in algorithms and architectures, datasets are also crucial for learning-based methods, yet current amodal segmentation datasets encounter several challenges: + +- Limited Scale: Existing datasets contain relatively few images, hindering the development of robust models. +- Annotation Quality: Several datasets relying on automatic generation methods can lead to inconsistent and sometimes incorrect instance annotations when not properly validated. +- Irrelevant Objects: A significant portion of annotated objects, such as walls and floors, contribute little to meaningful scene understanding. + +To address these limitations, we present Amodal-LVIS, a new large-scale dataset derived from LVIS [6] and LVVIS [26]. Our dataset contains 300K carefully curated images, where each image contains one instance annotation. These annotations form paired examples between synthetic occluded instances and their original unoccluded versions. Furthermore, we have processed and refined existing datasets to create a comprehensive training collection comprising approximately 1M images and 2M instance annotations. + +Experimental results show that our method with the EfficientSAM architecture, when trained on our combined dataset, achieves remarkable zero-shot performance that surpasses previous supervised amodal segmentation methods. These results validate our approach of leveraging an efficient existing architecture with high-quality, large-scale training data for amodal segmentation tasks. + +Our main contributions can be summarized as follows: + +1. Flexible Amodal Framework: The proposed method, SAMEO, adapts EfficientSAM for amodal instance segmentation that works with both modal and amodal detector prompts through specialized training. +2. Large-scale Dataset: A new Amodal-LVIS dataset containing 300K images, forming paired examples between synthetic occluded instances and their original unoccluded versions. +3. Dataset Collection: A comprehensive training collection of 1M images and 2M instances created by combining and refining existing amodal datasets with AmodalLVIS. +4. Zero-shot Performance: State-of-the-art zero-shot results on both COCOA-cls and D2SA benchmarks, surpassing previous supervised methods. + +![](images/902242121c4bf82d48b1bc857a55b5634093eb84423b43f75e8bf1eefed73dd6.jpg) +Figure 2. Overview of our amodal segmentation pipeline. Given an input image, existing object detectors first generate either modal boxes (showing visible regions) or amodal boxes (showing complete object extent). Our SAMEO then processes these detections to produce amodal masks that recover the full shape of objects, including occluded parts. + +# 2. Related Work + +# 2.1. Instance Segmentation + +Instance segmentation is a fundamental computer vision task that simultaneously addresses object detection and segmentation, aiming to both locate objects in a scene and generate precise mask predictions for each detected instance. Initially focusing on visible parts of objects (modal instance segmentation), this field continues to evolve with the emergence of deep learning architectures. State-of-the-art methods demonstrate improvements through transformer-based feature extraction [27], modernized convolutions [21, 28], and optimized speeds [16]. Further detection models built upon the DETR architecture [1] have achieved additional advances through specialized query selection mechanisms and training schemes [33, 35]. + +Building upon modal instance segmentation, amodal instance segmentation extends the task to predict complete object shapes, including occluded regions. This extension is first formalized by Li and Malik [12], leading to various architectural innovations [5, 30, 31]. Notable approaches include ORCNN [4], ASN [22], which enhance Mask R-CNN [7] with occlusion reasoning capabilities, and BC-Net [10] with its bilateral layers for handling object overlaps. Currently, AISFormer [25] represents the state-of-the-art in amodal instance segmentation by introducing transformers to effectively model long-range dependencies. + +# 2.2. Segment Anything Model + +The Segment Anything Model (SAM) [11] represents a significant advancement in foundational computer vision models, capable of segmenting any visual object based on various prompts, including points or boxes. Trained on a dataset + +of 11M images, SAM demonstrates outstanding zero-shot generalization capabilities across diverse object categories and domains. + +The original SAM model, despite its strong performance, faces practical limitations due to high computational demands, including significant memory requirements and slow inference speed. EfficientSAM [29] addresses these challenges by using a Masked Autoencoder (MAE) [8] pre-training method to learn the feature embeddings from SAM's original ViT-H encoder, resulting in faster inference speed and reduced model size while maintaining comparable segmentation performance. + +# 2.3. Amodal Datasets + +Several datasets have been introduced for amodal segmentation. COCOA [34] is the first amodal dataset, providing semantic-level amodal annotations for COCO images. D2SA/COCOA-cls [4] extends this with instance-level annotations. DYCE [3] offers synthetic indoor scenes with accurate ground truth. The KINS dataset [22] focuses on traffic scenarios with 14K images of vehicles and pedestrians. More recently, MUVA [13] introduces a multi-view shopping scenario dataset, while MP3D-Amodal [32] provides real-world indoor scenes from Matterport3D. WALT [23] uniquely utilizes time-lapse imagery to obtain amodal ground truth, and KITTI-360-APS [18] extends KITTI-360 [14] with amodal panoptic annotations. Furthermore, datasets from related amodal completion works, such as pix2gestalt [19], have contributed to the development of the field. + +# 3. Our Approach + +# 3.1. Enabling Amodal Mask Prediction + +Preliminaries: EfficientSAM. Segment Anything Model (SAM) [11] is a foundation model for image segmentation that can generate high-quality object masks based on any prompt. The original SAM architecture consists of three main components: i) an image encoder that transforms the input image into image embeddings, ii) a lightweight transformer-based prompt encoder that converts prompts (points, boxes) into unified embeddings, and iii) a mask decoder that utilizes a transformer architecture with two cross-attention layers to process both image and prompt embeddings for generating the final segmentation mask. In our approach, we mainly use EfficientSAM [29], a compact adaptation of the original SAM model. EfficientSAM replaces SAM's image encoder with a lightweight ViT variant [2] while maintaining the original prompt encoder and mask decoder. + +Model Architecture. We propose SAMEO for amodal instance segmentation, retaining a lightweight image en + +coder $\mathcal{E}$ as in the original architecture of EfficientSAM, a transformer-based prompt encoder $\mathcal{P}$ , and a mask decoder $\mathcal{D}$ with dual cross-attention layers. Given an input image $I$ and a bounding box prompt $B$ , the proposed SAMEO pipeline predicts the amodal mask $\hat{M}$ and the estimated IoU $\hat{\rho}$ as follows: + +$$ +\hat {M}, \hat {\rho} = \mathcal {D} (\mathcal {E} (I), \mathcal {P} (B)). \tag {1} +$$ + +Training Strategy. During training, we exclusively fine-tune EfficientSAM's mask decoder while keeping the original weights of the image encoder and prompt encoder unchanged. The model receives two inputs: an image and a bounding box prompt derived from ground-truth annotations. The box prompts are randomly selected from modal and amodal ground-truth boxes with equal probability. The training objective combines Dice loss [17], Focal loss [15], and L1 loss for IoU estimation: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {D i c e}} + \mathcal {L} _ {\text {F o c a l}} + \lambda \mathcal {L} _ {\text {I o U}}, \tag {2} +$$ + +where + +$$ +\mathcal {L} _ {\mathrm {D i c e}} = 1 - \frac {2 | \hat {M} \cap M _ {\mathrm {g t}} |}{| \hat {M} | + | M _ {\mathrm {g t}} |}, \tag {3} +$$ + +$$ +\mathcal {L} _ {\text {F o c a l}} = - (1 - p _ {t}) ^ {\gamma} \log (p _ {t}), \tag {4} +$$ + +$$ +\mathcal {L} _ {\mathrm {I o U}} = \left| \hat {\rho} - \mathrm {I o U} (\hat {M}, M _ {\mathrm {g t}}) \right|. \tag {5} +$$ + +Here, $M_{\mathrm{gt}}$ represents the ground truth amodal mask, $p_t$ is the predicted probability for the target class, $\gamma$ is the focusing parameter set to 2 in our experiments, and $\lambda$ is empirically set to 0.05. + +Inference Pipeline. For inference, SAMEO can be flexibly integrated with various object detectors, including both amodal detectors (e.g., AISFormer [25]) and conventional modal detectors (e.g., RTMDet [16]). The detection outputs serve as box prompts for our model, which then generates corresponding amodal masks (Figure 2). This modular design allows our model to enhance existing detection systems with amodal segmentation capabilities while achieving state-of-the-art performance. + +# 3.2. Amodal Dataset Collection + +Limitations of Existing Amodal Datasets. Existing amodal datasets have inherent limitations in both manual and synthetic annotation mechanisms. Human-annotated datasets, while closely representing real-world scenarios, are costly to produce and prone to errors in occluded region estimation. Synthetic datasets, though efficient to generate, lack reliable verification mechanisms for object completeness and may not accurately reflect natural occlusion patterns (Figure 3). + +
DatasetType# Instances# ImagesPOI (%)Average ROR (%)
COCOA (no stuff) [34]Real32,9265,07354.930.2
COCOA-cls [4]Real10,5623,4994921.8
KINS [22]Real188,08514,99354.342.2
KITTI-360-APS [18]Real89,93812,4965033.9
D2SA [4]Synthetic28,7205,60053.623.7
MUVA [13]Synthetic198,57326,40676.732.1
WALT* [23]Synthetic485,36940,0003236.5
DYCE* [3]Synthetic66,4535,22977.829
MP3D-amodal* [32]Synthetic2,9682,54910041.1
pix2gestalt [19]Synthetic849,667849,66710035.9
Amodal-LVISSynthetic399,398301,4935034.5
Total-2,352,6591,267,424--
+ +Table 1. Amodal dataset collection. Datasets marked with * are generated or refined by ourselves. POI is calculated as the percentage of occluded instances (where modal mask is not equal to amodal mask) over all instances. Meanwhile, average ROR is the average ratio of occluded regions, considering only instances with occlusions. + +![](images/4c6c2a5846ebe11e65a6e1bd7846a01dc7874f8dee28b2af76acfc7f71619e9a.jpg) +(a) DYCE + +![](images/3187e777fb267d2b48737a06faefe51dd93890fa43d7ae511e42a8af6aa09198.jpg) +(b) MP3D-amodal + +![](images/f909521fb192f060ca747c75f4d91d3ee49dd56c38515fa1e3e9424c440fac5f.jpg) +(c) pix2gestalt +Figure 3. Examples of limitations in existing amodal datasets: (a) DYCE and (b) MP3D-amodal show meaningless architectural elements rendered from 3D meshes that dominate the image space, while (c) pix2gestalt contains potentially incomplete amodal masks due to restrictive generation criteria. + +Dataset Collection and Quality Control. To leverage the advantages of both mechanisms, we have collected and filtered datasets of both annotation types. Our cleaning process addresses specific issues in each dataset to ensure data quality while maintaining realistic occlusion representations (Table 1). + +For synthetic datasets DYCE and MP3D-amodal, generated using 3D furniture meshes, we identify and address two main quality issues: meaningless architectural elements (walls, floors, ceilings) occupying the majority of image space and objects with minimal visible areas. We implement filters to remove cases where visible parts are less than $10\%$ of the whole object, objects occupying more than $90\%$ of the image area, and architectural element annotations. + +The WALT dataset leverages road surveillance timelapse footage for synthetic data generation. It obtains bounding boxes for cars and people using a pre-trained detector and then identifies complete objects by analyzing these bounding box intersections. These discovered com + +![](images/9d69c4799297f7d0656c35531df66230454a14801633a8ff6a7047fc03c0a440.jpg) +Original Image +Figure 4. Amodal-LVIS dataset generation process. From left to right: original image with unoccluded objects, a selected occluder object, and the synthesized image with occlusion. Our dataset includes both the original and the synthesized image for each instance to prevent occlusion bias during training. + +![](images/bbbbe17969c3135948245e1eaa177f4ae0a8987e7a413352d77e4747bb1c8d8f.jpg) +Occluder + +![](images/e95b0ccbbe66bbd6ce01f9d26828cf35fc1df94dbd527b60c6397895c132c4e5.jpg) +Synthesized Image + +plete objects are then composited back into the same scenes to generate synthetic training data. However, their layer-by-layer placement can create unrealistic occlusions. We address this by implementing an occlusion threshold filter to ensure natural occlusion patterns. + +For other datasets with class annotation, such as CO-COA, the availability of semantic labels enabled straightforward quality control. We filter out "stuff" class annotations across these datasets to focus on meaningful objects that align with amodal instance segmentation goals. + +# 3.3.Amodal-LVIS + +We propose a synthetic dataset for amodal mask segmentation through precise object occlusion generation, incorporating complete object collection, synthetic occlusion generation, and a dual annotation mechanism to prevent model bias. Combined with existing datasets, our collection totals 1M images and 2M instance annotations. + +Complete Object Collection. To obtain complete objects for synthetic occlusion, we utilize SAMEO, which is pre-trained on the previously mentioned amodal datasets, to generate pseudo labels for instances within LVIS and LVVIS datasets. Our model predicts amodal masks for each instance, which are then compared with the ground-truth visible mask annotations. This comparison helps us identify complete, unoccluded objects. + +Synthetic Occlusion Generation. The occlusion generation process involves pairing randomly selected complete objects from our collected pool. To ensure that the occlusions look realistic, we normalize the paired objects to similar sizes while maintaining their aspect ratios. Object positioning and occlusion rates are controlled using bounding box annotations, which allow for precise management of how objects occlude one another. + +Dual Annotation Mechanism. Our experiments in Section 4 show that training solely on occluded masks leads to model confusion, resulting in an over-prediction of occluded objects even when the prompts are intended to target foreground instances. To resolve this issue, we include both occluded and original unoccluded versions of instances in our dataset (Figure 4). This dual annotation mechanism prevents occlusion bias while providing comprehensive training examples for both states. + +# 4. Experiments + +# 4.1. Settings + +Implementation Details. Our model is trained on NVIDIA Tesla V100/A100 GPUs for 1,440/2,340/22,500 iterations on COCOA-cls/D2SA/MUVA respectively. For zero-shot SAMEO, we increase the batch size to 32 and train for 40,000 iterations. We use the Adam optimizer with a learning rate of $1 \times 10^{-4}$ without any learning rate scheduler. For each instance during training, we randomly select either the amodal or modal ground truth bounding box as the prompt with equal probability. + +Datasets and Baselines. For training, we use our dataset collection and the proposed Amodal-LVIS dataset. As for evaluation, COCOA-cls, D2SA and MUVA datasets are utilized. We primarily compare our approach against AIS-Former, the current state-of-the-art model in amodal instance segmentation. Unlike conventional instance segmentation models that incorporate both object detection and mask decoding components, SAMEO functions solely as a mask decoder. This allows our model to flexibly integrate with existing amodal and modal instance segmentation models, using their object box predictions as prompts to + +generate refined amodal masks. For a comprehensive evaluation, we compare our results against the original mask predictions from these front-end models. + +To evaluate our zero-shot performance, we extend the comparisons beyond AISFormer to include both modal and amodal front-end models equipped with the original EfficientSAM as their mask decoder. This comparison demonstrates our successful adaptation of EfficientSAM for amodal mask segmentation while maintaining zero-shot capabilities. + +Evaluation Metrics. We evaluate our method using two standard metrics: mean Average Precision (AP) and mean Average Recall (AR). Since our model is class-agnostic, we compute both metrics without considering class labels. For a fair comparison, we reproduce baseline methods and evaluate them using the same class-agnostic AP and AR. For methods that use SAMEO as mask decoder, we refine the confidence score $\hat{\rho}_{\mathrm{front}}$ of front-end models using the estimated IoU $\hat{\rho}_{\mathrm{ours}}$ predicted by SAMEO when calculating these metrics. The refined confidence score $\hat{\rho}_{\mathrm{ref}}$ of these cases is computed as follows: + +$$ +\hat {\rho} _ {\text {r e f}} = \hat {\rho} _ {\text {f r o n t}} \times \hat {\rho} _ {\text {o u r s}}. \tag {6} +$$ + +# 4.2. Results + +Quantitative Results. We evaluate our proposed SAMEO on three widely-used datasets: COCOA-cls, D2SA, and MUVA. For each dataset, we train our model on their respective training sets and evaluate on the corresponding test sets (Table 2). To demonstrate SAMEO's effectiveness and versatility, we attach it to various pretrained front-end models, where modal front-ends are trained with modal annotations and amodal front-ends are trained with amodal annotations of these datasets. Our experimental results show that SAMEO significantly outperforms the current state-of-the-art method, AISFormer, achieving higher AP and AR across all datasets. Notably, our model exhibits robust performance in mask refinement independent of the front-end model's original mask type. Regardless of whether the front-end models produce modal or amodal mask predictions, SAMEO successfully refines them to achieve comparable high performance, demonstrating its strong capability to utilize both types of prompts. + +Qualitative Results. For qualitative evaluation, we compare our model's predictions against AISFormer on both COCOA-cls and MUVA datasets (Figure 5). Our method exhibits superior performance across various challenging scenarios, including complex still life arrangements with multiple overlapping objects (e.g., bottles and containers), + +
ModelCOCOAc-clsD2SAMUVA
AP\(AP_{50}\)\(AP_{75}\)ARAP\(AP_{50}\)\(AP_{75}\)ARAP\(AP_{50}\)\(AP_{75}\)AR
AISFormer [25]40.670.542.555.266.389.972.876.169.490.375.680.7
RTMDet* [16]49.871.254.769.659.781.363.478.246.068.246.457.8
ConvNeXt-V2* [28]46.767.552.070.966.189.169.875.347.969.046.661.7
ViTDet* [27]47.469.452.568.863.086.465.875.445.366.944.760.7
AISFormer+SAMEO54.374.059.769.379.892.784.884.276.290.880.985.2
RTMDet*+SAMEO55.375.260.874.372.785.877.584.275.889.279.983.1
ConvNeXt-V2*+SAMEO54.173.159.374.080.894.085.187.179.293.182.681.3
ViTDet*+SAMEO54.173.359.272.378.692.382.684.774.189.078.282.3
+ +Table 2. Quantitative comparison on various datasets. Models marked with * are modal instance segmentation methods that detect objects and segment their visible masks without handling occlusions, with performance metrics calculated using their modal mask predictions. SAMEO takes bounding box predictions from the front-end models as prompts to generate amodal instance masks. Evaluation metrics include Average Precision (AP) at different IoU thresholds and Average Recall (AR). Bold numbers indicate the best performance. + +
ModelCOCOA-clsD2SA
AP\(AP_{50}\)\(AP_{75}\)ARAP\(AP_{50}\)\(AP_{75}\)AR
AISFormer40.670.542.555.266.389.972.876.1
AISFormer+EfficientSAM†47.670.051.764.269.689.272.377.7
RTMDet*+EfficientSAM†48.771.153.265.963.082.764.874.3
RetinaNet* [21]+EfficientSAM†44.867.148.668.460.677.562.979.3
DINO* [33]+SAMEO†50.270.455.773.769.886.272.675.9
RetinaNet*+SAMEO†51.072.256.672.662.578.064.880.7
AISFormer+SAMEO†52.873.457.967.974.190.278.379.9
ViTDet*+SAMEO†53.273.058.671.272.189.074.980.7
ConvNeXt-V2*+SAMEO†53.473.059.073.074.391.678.083.1
CO-DETR* [35]+SAMEO†54.074.860.273.575.091.078.582.2
RTMDet*+SAMEO†54.475.060.273.468.484.571.077.7
+ +Table 3. Zero-shot performance on COCOA-cls and D2SA datasets. The results show SAMEO not only significantly outperforms AIS-Former but also successfully adapts EfficientSAM's modal segmentation capability to amodal segmentation, demonstrating consistent performance improvements when paired with various front-end detectors. $\dagger$ indicates zero-shot evaluation without training on the test dataset. * denotes modal object detectors that provide modal bounding boxes as prompts. Bold numbers indicate the best performance. + +scenes with intricate occlusions (e.g., people behind barriers), and diverse object categories and poses. Results show that our model generates significantly more precise amodal masks with sharper boundaries while providing more reasonable predictions for occluded parts. The qualitative comparison clearly demonstrates our method's improvements over the baseline method in both mask quality and occlusion reasoning capabilities, validating the effectiveness of SAMEO for real-world amodal segmentation tasks. + +# 4.3. Zero-shot Performance + +To evaluate SAMEO's zero-shot generalization capability, we train our model on our dataset collection and the proposed Amodal-LVIS dataset, excluding COCOA-cls and D2SA. During training, for each batch, a dataset is sampled with probability proportional to the logarithm of its + +size divided by the sum of log sizes across all datasets. We then test these two held-out datasets to demonstrate zero-shot performance (Table 3). For comparison, we include AISFormer and RTMDet (both trained on target datasets) combined with the original EfficientSAM, showing that our model successfully adapts EfficientSAM for amodal segmentation while preserving its zero-shot capability. Furthermore, we experiment with various pre-trained modal front-end detectors to demonstrate SAMEO's robust zero-shot performance regardless of the front-end choice. + +The results demonstrate SAMEO's superior performance, achieving up to 13.8 AP improvement over AIS-Former on COCOA-cls with RTMDet and 8.7 AP over on D2SA with CO-DETR, reaching state-of-the-art results and validating its strength as a robust zero-shot amodal segmentation solution. + +![](images/46211eb7441e155202e6bb7c76cc8c7515b9b83004767426817c0203ad466d36.jpg) +Figure 5. Qualitative comparison of amodal mask predictions. For each row: SAMEO's amodal prediction (top) with AISFormer box prompts, and AISFormer's prediction (bottom). Our method demonstrates superior mask quality, exhibiting more precise boundary delineation and robust handling of complex occlusion scenarios. Original images used for evaluation are available in supplementary materials. + +
IoU PredictionCOCOA-cls
APAP50AP75
×52.473.257.8
✓54.374.059.7
+ +Table 4. Ablation study of IoU prediction refinement on COCOAc-cls dataset, using AISFormer as front-end. $\times$ and $\checkmark$ indicate without and with IoU prediction refinement,respectively. + +
Box PromptCOCOA-cls
AP\(AP_{50}\)\(AP_{75}\)AR
amodal53.072.958.071.1
modal53.773.359.371.2
random54.273.559.571.6
+ +Table 5. Comparison of different prompt types during training. We evaluate three variants on the COCOA-cls dataset using both modal and amodal front-ends, reporting the averaged performance over both types. The results show that training with random box prompts has optimal performance across diverse front-end models. + +# 4.4. Ablation Study + +Effect of IoU Prediction. To validate the effectiveness of the IoU prediction branch in our model, we have conducted experiments comparing the performance metrics before and after confidence score refinement using SAMEO's predicted IoU. Using AISFormer as our front-end model and evaluating this setting on the COCOA-cls dataset, the experimental results demonstrate that SAMEO's precise IoU prediction significantly contributes to improving the ranking of segmentation results (Table 4). Specifically, we observe that incorporating the predicted IoU for confidence score refinement leads to notable improvements in AP metric, confirming that the IoU prediction branch plays a crucial role in enhancing the overall performance of our model. + +Impact of Training Prompt Types. We investigate the optimal prompt-type strategy for training SAMEO to achieve balanced performance across both modal and amodal front-end prompts. We train three variants of SAMEO using ground truth amodal boxes, modal boxes, and a random mixture of both with equal probability. For evaluation, we integrate each trained model variant with both amodal and modal front-end detectors and evaluate their performance separately. The final performance metric is calculated by averaging the AP and AR scores across both front-end scenarios (Table 5). Our findings reveal that training with random boxes with equal probability yields the most balanced performance when handling various front-end prompt types, demonstrating the model's ability to generalize across different input scenarios. + +![](images/11febd1ffacac072ccf72c7cc4aa508fdc1d01b9609d4fa2ef1f439624b9b4bf.jpg) +Figure 6. Visualization of over-prediction. Given a detector box clearly intended for the foreground dog, SAMEO(occ), trained solely on occluded instances, mistakenly predicts the mask of the background towel. After rebalancing our dataset to include both occluded and non-occluded instances, SAMEO Mix) produces results that closely match the EfficientSAM baseline. + +Dataset Composition Analysis. To understand the importance of diverse instance types in training data, we have conducted an experiment training SAMEO exclusively on datasets containing only occluded instances (e.g., pix2gestalt). Visualization results reveal a significant limitation: the model exhibits over-prediction of background instances, even when the input box prompt clearly indicates a foreground object (Figure 6). This observation motivates our design choice for the Amodal-LVIS dataset, which maintains an equal distribution of occluded and non-occluded annotations. This balanced composition prevents bias and ensures robustness across various scenarios. + +# 5. Conclusion + +We present a flexible approach to amodal instance segmentation by adapting foundation segmentation models to handle both visible and occluded portions of objects. Our framework successfully leverages pre-trained modal detectors while maintaining strong amodal segmentation capabilities. The introduction of Amodal-LVIS, containing 300K carefully curated images, along with our comprehensive collection of 1M images and 2M instance annotations, addresses critical limitations in existing datasets and provides the necessary scale for robust model development. + +Our extensive experiments demonstrate that SAMEO consistently outperforms state-of-the-art methods on COCOA-cls, D2SA, and MUVA datasets. Most notably, when trained on our dataset collection, including AmodalLVIS, SAMEO achieves strong zero-shot performance on unseen datasets. The model's robust generalization abilities persist across various front-end detectors, validating our approach of adapting foundation models for amodal segmentation without compromising performance. We further address the limitations and possible future work of SAMEO in the appendix. + +# Acknowledgements + +This work was supported in part by NSTC grants 113-2221-E-001-010-MY3 and 112-2221-E-A49-100-MY3 of Taiwan, as well as funding from NVIDIA Taiwan AI R&D Center. We appreciate the National Center for High-performance Computing for providing computational resources and facilities. + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 213-229. Springer, 2020. 2 +[2] 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 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3 +[3] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6144-6153. Computer Vision Foundation / IEEE Computer Society, 2018. 3, 4 +[4] Patrick Follmann, Rebecca König, Philipp Hartinger, Michael Klostermann, and Tobias Böttger. Learning to see the invisible: End-to-end trainable amodal instance segmentation. In IEEE Winter Conference on Applications of Computer Vision, WACV 2019, Waikoloa Village, HI, USA, January 7-11, 2019, pages 1328-1336. IEEE, 2019. 2, 3, 4 +[5] Jianxiong Gao, Xuelin Qian, Yikai Wang, Tianjun Xiao, Tong He, Zheng Zhang, and Yanwei Fu. Coarse-to-fine amodal segmentation with shape prior. In ICCV, pages 1262-1271. IEEE, 2023. 2 +[6] Agrim Gupta, Piotr Dollar, and Ross B. Girshick. LVIS: A dataset for large vocabulary instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 5356-5364. Computer Vision Foundation / IEEE, 2019. 2 +[7] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 2980-2988. IEEE Computer Society, 2017. 2 +[8] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 3 + +[9] Gaetano Kanizsa, Paolo Legrenzi, and Paolo Bozzi. Organization in vision: Essays on gestalt perception. 1979. 1 +[10] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4019-4028. Computer Vision Foundation / IEEE, 2021. 2 +[11] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3992-4003. IEEE, 2023. 1, 2, 3 +[12] Ke Li and Jitendra Malik. Amodal instance segmentation. In ECCV, pages 677-693. Springer, 2016. 2 +[13] Zhixuan Li, Weining Ye, Juan Terven, Zachary Bennett, Ying Zheng, Tingting Jiang, and Tiejun Huang. MUVA: A new large-scale benchmark for multi-view amodal instance segmentation in the shopping scenario. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 23447-23456. IEEE, 2023. 3, 4 +[14] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3292–3310, 2023. 3 +[15] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 2999-3007. IEEE Computer Society, 2017. 3 +[16] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. CoRR, abs/2212.07784, 2022. 1, 2, 3, 6 +[17] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In Fourth International Conference on 3D Vision, 3DV 2016, Stanford, CA, USA, October 25-28, 2016, pages 565-571. IEEE Computer Society, 2016. 3 +[18] Rohit Mohan and Abhinav Valada. Amodal panoptic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 20991-21000. IEEE, 2022. 3, 4 +[19] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: Amodal segmentation by synthesizing wholes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 3931-3940. IEEE, 2024. 3, 4 +[20] Stephen E. Palmer. Vision Science: Photons to Phenomenology. MIT Press, 1999. 1 +[21] Roberto Del Prete, Maria Daniela Graziano, and Alfredo Renga. Retinanet: A deep learning architecture to achieve + +a robust wake detector in SAR images. In 6th IEEE International Forum on Research and Technology for Society and Industry, RTSI 2021, Naples, Italy, September 6-9, 2021, pages 171-176. IEEE, 2021. 2, 6 +[22] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with KINS dataset. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 3014-3023. Computer Vision Foundation / IEEE, 2019. 2, 3, 4 +[23] N. Dinesh Reddy, Robert Tamburo, and Srinivasa G. Narasimhan. WALT: watch and learn 2d amodal representation from time-lapse imagery. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 9346-9356. IEEE, 2022. 3, 4 +[24] Thomas F Shipley and Philip J Kellman. From fragments to objects: Segmentation and grouping in vision. Elsevier, 2001. 1 +[25] Minh Q. Tran, Khoa Vo, Kashu Yamazaki, Arthur A. F. Fernandes, Michael Kidd, and Ngan Le. Aisformer: Amodal instance segmentation with transformer. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022, page 712. BMVA Press, 2022. 2, 3, 6 +[26] Haochen Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Cilin Yan, Weidi Xie, Shuai Wang, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 4034-4043. IEEE, 2023. 2 +[27] Liya Wang and Alex Tien. Aerial image object detection with vision transformer detector (vitdet). In IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2023, Pasadena, CA, USA, July 16-21, 2023, pages 6450-6453. IEEE, 2023. 2, 6 +[28] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext V2: co-designing and scaling convnets with masked autoencoders. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16133-16142. IEEE, 2023. 1, 2, 6 +[29] Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest N. Iandola, Raghuraman Krishnamoorthi, and Vikas Chandra. Efficientsam: Leveraged masked image pretraining for efficient segment anything. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 16111-16121. IEEE, 2024. 1, 3 +[30] Jianwei Yang, Zhile Ren, Mingze Xu, Xinlei Chen, David J. Crandall, Devi Parikh, and Dhruv Batra. Embodied amodal recognition: Learning to move to perceive objects. In ICCV, pages 2040-2050. IEEE, 2019. 2 +[31] Jian Yao, Yuxin Hong, Chiyu Wang, Tianjun Xiao, Tong He, Francesco Locatello, David P. Wipf, Yanwei Fu, and Zheng Zhang. Self-supervised amodal video object segmentation. In NeurIPS, 2022. 2 + +[32] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. Amodal ground truth and completion in the wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 28003-28013. IEEE, 2024. 3, 4 +[33] 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. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2, 6 +[34] Yan Zhu, Yuandong Tian, Dimitris N. Metaxas, and Piotr Dólár. Semantic amodal segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 3001-3009. IEEE Computer Society, 2017. 3, 4 +[35] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 6725-6735. IEEE, 2023. 2, 6 \ No newline at end of file diff --git a/CVPR/2025/Segment Anything, Even Occluded/images.zip b/CVPR/2025/Segment Anything, Even Occluded/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a4417cf3b294425021be60b612f8007de0156a22 --- /dev/null +++ b/CVPR/2025/Segment Anything, Even Occluded/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9927617525c81324c66375666c130a64ea5069877727904a81927dfb166c9cc3 +size 766686 diff --git a/CVPR/2025/Segment Anything, Even Occluded/layout.json b/CVPR/2025/Segment Anything, Even Occluded/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7f5831e0c1ae6f7ed256789b7587fb4da0e273e3 --- /dev/null +++ b/CVPR/2025/Segment Anything, Even Occluded/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ea37ecd3a8424f252ad7bf37ab98e53fd4a218b7a01fd5bb65539323cfedada +size 289762 diff --git a/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/f4675c76-12fd-46eb-8463-021215e9a2e0_content_list.json b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/f4675c76-12fd-46eb-8463-021215e9a2e0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..85bf3b673708f7ac694c1a140b7c440f40934b59 --- /dev/null +++ b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/f4675c76-12fd-46eb-8463-021215e9a2e0_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:594d6eacd513d5b4d0ca13be552f48862b1cfeced6dc60a74176b593c5d544ff +size 77488 diff --git a/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/f4675c76-12fd-46eb-8463-021215e9a2e0_model.json b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/f4675c76-12fd-46eb-8463-021215e9a2e0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b2f940cc07037cc3a902459ee015d57894de399c --- /dev/null +++ b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/f4675c76-12fd-46eb-8463-021215e9a2e0_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64406ba81baa2e819f544709d0874a59882c3b99a5701b8fa29b2c0c6510ad8d +size 93753 diff --git a/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/f4675c76-12fd-46eb-8463-021215e9a2e0_origin.pdf b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/f4675c76-12fd-46eb-8463-021215e9a2e0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ee834344a9d178a4b2b40e868bcf33da511d931d --- /dev/null +++ b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/f4675c76-12fd-46eb-8463-021215e9a2e0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32fd05c58afb11586fae4a18e6047afed424474f5c215b19e2e1e8ad210f2b99 +size 855889 diff --git a/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/full.md b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6ba03a20384ef2d4527e7778346ecd6fde6f3dfb --- /dev/null +++ b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/full.md @@ -0,0 +1,310 @@ +# Segment This Thing: Foveated Tokenization for Efficient Point-Prompted Segmentation + +Tanner Schmidt Meta Reality Labs + +tanner.schmidt@meta.com + +Richard Newcombe Meta Reality Labs + +newcombe@meta.com + +# Abstract + +This paper presents Segment This Thing (STT), a new efficient image segmentation model designed to produce a single segment given a single point prompt. Instead of following prior work and increasing efficiency by decreasing model size, we gain efficiency by foveating input images. Given an image and a point prompt, we extract a crop centered on the prompt and apply a novel variable-resolution patch tokenization in which patches are downsampled at a rate that increases with increased distance from the prompt. This approach yields far fewer image tokens than uniform patch tokenization. As a result we can drastically reduce the computational cost of segmentation without reducing model size. Furthermore, the foveation focuses the model on the region of interest, a potentially useful inductive bias. We show that our Segment This Thing model is more efficient than prior work while remaining competitive on segmentation benchmarks. It can easily run at interactive frame rates on consumer hardware and is thus a promising tool for augmented reality or robotics applications. + +# 1. Introduction + +Since its introduction in 2023, the Segment Anything Model (SAM) [20] has brought new focus to the image segmentation task and enabled numerous downstream applications with its flexible design and impressive performance. However, SAM's capabilities come at a cost: the model leverages a heavyweight image encoder (ViT-H in the standard model). + +In this paper we focus on segmentation for streaming use cases, for example robotics or augmented/virtual reality applications. With many frames arriving every second, it is likely that only one or a few entities need to be segmented in each image. Latency is critical, as frames may be separated in time by only tens of milliseconds. Depending on the relative locations of and communication channels between sensors and compute, bandwidth may also be a concern. + +![](images/ee0cb32d204e8d3f63080275484bc9d91aec4fcd384f522534ede351ed3b6d76.jpg) +Patch Tokenization +Figure 1. A visualization of the foveated tokenization employed by Segment This Thing. Pixels are grouped into small patches in the center and progressively larger patches towards the edges. + +![](images/2e368b77190a9a7191d1bc98a8eca84a90a4269fb0745580a930ef26930b6783.jpg) +Foveated Tokenization + +We introduce a new efficient segmentation model called "Segment This Thing" (STT) that was designed for streaming video applications. There have been many other follow-up works aimed at making SAM more efficient, but they have all done so by reducing the model size. We take a different approach, and instead apply a novel biologically-inspired tokenization scheme which results in a drastic reduction in token count and a concomitant increase in efficiency of the Transformer-based image encoder. This scheme, which we refer to as "foveated tokenization," is modeled after the variability of the visual acuity of human eye, which is highest in the fovea and decreases with increased eccentricity [3]. Instead of dividing the image into evenly sized patches on a uniform grid, we resample patches with variable resolution on an irregular grid centered on a single point prompt (see Figure 1). The further a patch is from the center, the lower its resolution. + +In the analogy with the human eye, it is as if the model directs its visual attention towards the prompt point and estimates the extent of the thing it is "looking at" (hence the name of our model). If the object is small or far away such that it subtends a small solid angle, it will be contained within the highest resolution patches and thus will not be affected by the downsampling (so long as there is sufficient + +context from the surrounding regions). If the object is large, it will extend into the periphery of the image and some information will be lost. However, we argue that for larger segments, high resolution is not as important. Our experiments show that our resampling of the image results in only moderate decreases in segmentation accuracy across a wide range of segment sizes. Importantly, there is no need to know the size of an object before attempting to segment it. + +Our foveated tokenization reduces the token count so significantly, we can actually use larger models relative to other efficient SAM variants while still requiring fewer FLOPs and thus lower latency, due to the quadratic scaling of self-attention with respect to the token count. This gives Segment This Thing more modeling flexibility and suggests that this tokenization and image encoding strategy could be applied gainfully towards other tasks besides segmentation. + +In addition to reducing latency, our foveated tokenization has the benefit of reducing the bandwidth required to get pixels from the sensor to the model. The tokenization results in a significant compression of the image, and is orthogonal to other compression schemes that could be applied to the patches (e.g. JPEG compression) for further bandwidth reduction. Furthermore, the resampling is computationally cheap and would be relatively easy to implement efficiently in hardware. Altogether, this makes the Segment This Thing model a promising option for power-constrained edge devices that need to stream data off-device to be segmented elsewhere. + +With foveated tokenization, each prompt leads to a unique foveation and thus requires running both the image encoder and the mask decoder, whereas models like SAM can amortize the cost of the image encoder across multiple prompts. However, our model is so much faster than the baselines that it would require a number of prompts on the same image before the cost is equivalent. Furthermore, the need to segment many objects is more common in static image analysis and less relevant to streaming video use cases. + +We benchmark Segment This Thing on a number of publicly-available segmentation datasets to quantify the performance degradation that comes with the increased efficiency. We also show qualitative experiments in which users wear glasses with front-facing cameras and eyetracking capabilities, applying STT to estimate what objects the user may be looking by using gaze-derived prompts. + +# 2. Related Work + +Our work is built on Segment Anything [20], which aims to segment any entity in an image given a prompt in the form of a bounding box or one or more 2D points. Our model aims to address the high computational cost of running the Segment Anything Model (SAM). + +Efficient Prompted Segmentation A number of other papers have also introduced techniques to reduce the cost of + +prompted segmentation. FastSAM [49] uses an efficient CNN-based object detection and segmentation backbone to produce segmentation candidates and then uses prompts to select from and / or merge these candidates into an output mask. SqueezeSAM [39] employs a UNet architecture with a transformer at the lowest-resolution level. Their early prompt fusion approach (providing prompts as additional channels in the input to the UNet rather than injecting them in a second stage) is similar to our centering of foveated inputs on the prompt, allowing the network to focus on relevant information from the start. EfficientViT-SAM [48] substitutes the ViT image encoder used by SAM for an EfficientViT [25] encoder. EfficientSAM [40], EdgeSAM [51], and MobileSAM [46] reduce the size of SAM by applying a knowledge distillation approach to transfer knowledge from the SAM image encoder into a lighter-weight encoder, then fine-tuning on the prompted segmentation task. For a recent review of efficient SAM methods, see Sun et al. [36]. + +Prompted Video Segmentation One of the aims of our work is to reduce the cost of prompted segmentation such that it is reasonable to run on every frame of a streaming video. Recently, Ravi et al. introduced SAM-2 [33], an extension of SAM to prompted video object segmentation (VOS). However, with SAM-2 each prompt recovers a segmentation of only one object across the video, with multiple segmentations requiring multiple prompts. Our model can segment a different object each frame depending on where the foveation center is placed. We leave the extension of foveated tokenization to video input to future work. + +Resolution reduction In the deep learning era of computer vision it has always been standard practice to rapidly reduce resolution in early network layers. For example, AlexNet [23] begins with a stride-4 convolution followed by a stride-2 pooling, reducing an input image of size $224 \times 224$ to a feature map of size $26 \times 26$ after just these two layers. This did not change as hardward advanced; ResNets [15] begin with a convolutional layer and a pooling layer each with a stride of two pixels, resulting in a quartering of the image resolution after progressing through a small fraction of the network depth. Our approach also reduces input resolution at the beginning of (in fact, before) the network, but with a rate that increases with increased distance from the prompt. + +Compression Other work reduces data sizes by learning from compressed representations. For example, Park and Johnson [31] train a Transformer-based image classifier directly on DCT co-efficients as used in JPEG encoding of images. Horton et al. take this idea even further and train a classifier directly on compressed byte encodings of images [17]. Other methods such as VQ-VAEs [38] or the more recent Lookup-Free Quantization [44] allow learning a compression of input data which can then be leveraged to solve downstream tasks more efficiently. However, depending on the specific architecture these techniques can incur + +significant costs to compute the compressed representations and are therefore not as widely applicable to deployment on power-constrained edge devices. Our foveated tokenization can be seen as a simple compression of the input image, but the compression technique is extremely cheap and largely orthogonal to these other approaches (i.e. one could apply further compression to the foveated patches). + +Efficient Tokenization Other prior works similarly addressed the issue of large token counts in transformer architectures. Yu et al. introduced TiTok [45], which generates a concise set of 32 or 64 tokens to represent an image. Yu et al. also make note of the ability to increase model size when token counts are low, but focused on the downstream task of image generation – the TiTok tokenizer itself is a Vision Transformer and is thus not particularly efficient. A large number of different token pruning or merging approaches have been proposed [2, 5, 13, 21, 27, 28, 32, 34, 35, 41, 42]. These methods typically assume a full input image, starting with a full token set and progressively reducing the token count from there. They therefore do not offer the same benefits resulting from input size reduction as our approach. + +Foveated Vision There have been some prior works applying foveation models to deep learning. Jonnalagadda et al. introduced FoveaTer [18], which uses variably-sized pooling windows applied to the output of a convolutional network to produce foveated feature maps. These foveated features are then processed by a Transformer, which like our image encoder benefits from the reduction in token count. However, the convolutional backbone of FoveaTer still requires full resolution images, which places higher constraints on sensors and bandwidth compared to our approach, in which foveation is applied as a first step (and can therefore be implemented on the sensor itself or in nearby accelerators). Peripheral Vision Transformer [29] also takes inspiration from biological vision systems, but focuses on the inductive bias aspect of foveation by modifying the position encoding of a Transformer-based image classifier. They do not achieve efficiency gains relative to their baseline. PeLK [4] applies the foveation concept to convolutional neural networks, introducing adaptive weight sharing that allocates more weights to the center of a large convolutional kernel and fewer towards the edges. This work actually moves in the opposite direction of ours by decreasing parameter counts; we instead decrease input size and increase parameter counts. Furthermore, the above works are evaluated on older datasets with small images such as ImageNet [10], ADE20K [50], or MS-COCO [24]. With our focus on efficiency, we evaluate Segment This Thing on more modern datasets with much larger images, showing the ability to segment objects ranging in size from a handful of pixels to millions of pixels with very low latency. + +Konrad et al. also explore foveated imaging with GazeGPT [22]. GazeGPT also investigates gaze-based + +![](images/45bbe3a97cffbe23bf55995fba925ac9391ef779219350aa75212d7661599b92.jpg) +Figure 2. (left) The foveation pattern used by our trained models. (right, top) A patch from the center maintains its original resolution. (right, bottom) A patch from the outer ring gets downsampled by a factor of 8. Best viewed digitally. + +![](images/65b9e25aef1f79d9a94e383093d97f8c2f282c8851d1107b8152c8c33527a171.jpg) + +prompting, in their case of a multimodal large language model. This model was pre-trained on uniform resolution images, which constrains the GazeGPT foveation model (it uses three overlapping images of the same size but varying receptive field). + +# 3. The Segment This Thing Model + +In this section we introduce the Segment This Thing model. First, we describe the foveation model used to map full resolution images to foveated token sets. Then we describe the architecture that maps these tokens to segmentation maps. + +# 3.1. Foveated Tokenization + +Each instance of the Segment This Thing model is trained with a fixed foveation pattern which divides an image region into a set of patches. An example of such a pattern is depicted in Figure 2. In the center of the pattern is a dense grid of patches corresponding to the center of the fovea in the analogy with human vision. Surrounding this block is a series of concentric rings of larger patches. The patches in each ring are larger than patches in the preceding ring and have a size that is an integer multiple of the central patch size. The patch sizes and the width of each ring are selected such that they can be nested without gaps or overlap1. + +Given an image and a point prompt, the first step is to extract a crop centered on the prompt and sized according to the pattern. If the crop region extends beyond the image boundary, we pad the image. Next, each patch in the foveation pattern is downsampled to the central patch size. This can be implemented efficiently using an integral image, which is equivalent to applying a box filter and subsampling with integer strides. This simplistic filter results in lower image quality than more sophisticated methods, but is hardware-friendly such that it could be deployed on power-constrained edge devices. The downsampling process is visualized in Figure 2. We follow SAM and use $16 \times 16$ patches. + +![](images/8adedbb155f2feba90dff87c140c6a96e0dd3c5d5b9f1ec1013f2a555617dbc6.jpg) +Figure 3. The foveated tokenization process. It begins by cropping a square region around the point prompt with a fixed size determined by the foveation pattern. This pattern divides the crop region into a set of variably-sized patches, each of which is resampled such that they are all the same size. The set is then concatenated to form a token tensor. + +After downsampling, the patches are all the same size, as is the output of the standard patch tokenization used e.g. in Vision Transformers [11]. We can thus flatten each patch into a vector and stack them to form a token matrix, hence the name foveated tokenization. The token set is then passed to the Segment This Thing model for processing. The full tokenization pipeline is depicted in Figure 3. + +# 3.1.1. Visualization + +It is often useful to visualize the foveated input to Segment This Thing. However, after foveated tokenization the image is represented as a $N \times T \times T$ tensor, where $N$ is the token count and $T$ is the patch size. To visualize such a tensor, we reverse the tokenization process: patches are upsampled back to their original size (typically using nearest neighbor interpolation) at which point they can be recomposed into an image. Note that while this process does not re-introduce any new information and thus provides a faithful view of what the network sees, it does introduce some distortion — the central regions take up much more space in the network input than in the visualizations thereof. + +![](images/6d2174e5978d08e7464d3fbf6658e6865a50287cb00e2e8b1778556d549f75bf.jpg) +Figure 4. Segment This Thing has a precision advantage for extremely small targets due to the additional decovolutional layers in the mask decoder. Left: an input image with prompt indicated in red. Right: a visualization of the segmentation output of SAM and STT. The crop corresponds to the region indicated by a rectangle at left. STT is able to output a smaller segment than SAM due to its higher resolution output in the central tokens. Note that this is not the most confident output mask for either model. + +![](images/f8255b2228b189b2d514b2b5a8ee5a0f8049eb2e5f226ab8404a357bc4b7c3c3.jpg) + +# 3.2. Segment This Thing + +Next we describe the architecture of the Segment This Thing model. It is largely based on the three-part SAM model. However, the prompt encoder has been omitted because the prompt is implicit in the centering of the crop. + +# 3.2.1. Image Encoder + +Due to the foveation, the input tokens to STT correspond to patches which are not arrayed on a 2D grid and thus the Vision Transformer model used by SAM (with local 2D windowed attention) is not applicable. However, the complexity of windowed attention was used in SAM to control compute cost, and our drastic cut in the token count removes that imperative. The image encoder for STT is thus a standard Transformer model. We simply project the patches (after downsampling to a common patch size) with a single linear layer, add learned position encodings, and pass them through the transformer to obtain a (foveated) feature map. We also include a register token to allow the model to collect and broadcast information, following Darcet et al. [8]. As the foveation pattern can extend beyond image boundaries, we also add a token mask to prevent the participation of out-of-bounds tokens in self-attention. + +# 3.2.2. MaskDecoder + +The mask decoder is also modeled on its counterpart in SAM, again with some simplifications. We do not support iterative segmentation – the model is always run just once per prompt. In SAM, the model outputs $N$ (typically $N = 3$ ) masks when prompted with a single point (which leads to scale ambiguity) and one mask when prompted with multiple points or a bounding box. Segment This Thing only supports single point prompts, so the decoder always outputs $N$ masks. The query token set is thus a set of $N + 1$ tokens, one for the intersection over union (IoU) prediction and one for each predicted mask. The context token set is the output of the encoder (including the register token). Out-of-bounds context tokens are masked during cross attention with the same token mask used in the encoder. The decoder otherwise operates in the same way as its counterpart in SAM, applying a two-way cross-attention trans + +![](images/7bca565e8a13f4b60789c5b4e8ecb3a4aeeb31cf849235bc23ff1c965e5cb344.jpg) +Figure 5. Despite the high downsampling rate in the periphery, the Segment This Thing model is capable of segmenting objects with a wide range of shapes and sizes, from those covering most of the receptive field (left) to those covering just a handful of pixels (right). All images are from SA-1B. They have been passed through foveated tokenization and are visualized as described in Section 3.1.1. + +![](images/db1b87c170f011896a51651f7ede96708577147f5fc7af751eb11b9265eace27.jpg) + +![](images/8574738d2bd69964550783bd4802b3bbe5d15720d1f6c10eaecbd77c3c8b0ef9.jpg) + +![](images/6caaae880352483b3d9f4292e99a6908ff23941349e9bad048c092d52d04f2e7.jpg) + +![](images/06fea92bf2485411e64cb7328102dc10d61afadbf48104c1d6d0eb08dfc218d8.jpg) + +![](images/6ead748982b5701df0f739ef3290670dd561f5beb67c7157776f8356dea35095.jpg) + +former, followed by learned upsampling of the image tokens, and finally a dot product between the mask tokens (transformed by an MLP) and the image token maps to produce the segmentation map. + +In SAM, the segmentation maps are estimated at a quarter of the resolution of the input image by applying two $2 \times 2$ deconvolutions with stride 2 to the token feature map. In other words, a $4 \times 4$ segmentation label patch is estimated for each $16 \times 16$ token. If full resolution maps are required, they are computed using bi-linear interpolation. This design decision was likely made to reduce memory costs during training. + +Another benefit of our lower token count is that we can afford to produce a full $16 \times 16$ segmentation map per patch by stacking four deconvolutions, each of which double the resolution. Thus the labels for the central patches are estimated at full resolution, with the output resolution dropping off towards the periphery at the same rate as the input resolution. If a full resolution segmentation map is required, it can be computed by interpolating each patch such that it matches its receptive field and then restructuring the patches to form an image. Note that this means that the output resolution of STT is actually higher than SAM in the center + +![](images/b940be34dd3e79b4feb35a666b9c5ecdd42c34cbad22912c198f87ef4fb45ba1.jpg) +Figure 6. A visualization of a training example for MAE pretraining of the image encoder. Left: the foveated input image (and reconstruction target), visualized as described in 3.1.1. Center: a visualization of the token set passed to the encoder. Right: a visualization of the reconstruction produced by the network. Note that the output resolution of the decoder matches the target image pixel-for-pixel, i.e. it is also foveated. + +![](images/4c06a4630fc7c90b2516ef0df7995112108a40830553cf654776af948a8fd18c.jpg) + +![](images/e129d6e4430a27e15d6463d7cc29f20ab3b6804407ec5438fe4e7539a1a26afb.jpg) + +of the segmentation map, at parity in the surrounding areas, and only at a lower resolution in the periphery. We have observed at least some cases where this allows STT to segment extremely small entities that SAM misses (see Figure 4; we suspect that this capability is limited by insufficiently granular labels in SA-1B, caused in part by the use of SAM itself to generate the masks). STT is also able to segment large objects, including segments that fill most or all of the foveation pattern, albeit at lower resolution. See Figure 5 for examples of segments estimated by STT across a wide range of scales. + +# 4. Experiments + +In this section we describe how we train the Segment This Thing model, and present evaluation results. + +# 4.1. Training + +We trained our STT models using a foveation pattern with 172 tokens, a total size of 1280 pixels, and a stride of 8 pixels in the outer ring. See the supplementary materials for details. In Table 1 we compare the input sizes for SAM and our STT model based on this pattern. + +Following SAM, we trained our Segment This Thing model exclusively on the SA-1B dataset [20]. However, because of the novel tokenization we cannot rely on an existing pre-trained image encoder as do Kirillov et al. Thus the first step in training STT is masked autoencoder (MAE) [16] pre-training of the image encoder. Other than the novel tokenization, we follow the standard MAE pre-training procedure. We center foveations at points randomly selected from SA-1B images that are more than a threshold distance + +
ModelTokensPixelsReceptive Field
SAM [20]40961.049M10242
STT (Ours)1720.044M12802
+ +Table 1. A comparison of the input size for SAM and STT. + +![](images/96fb24b0d4de269f54f2d43b103c968c9da56d01fe3fe0c88ea721888ced5736.jpg) +Figure 7. Results of evaluating our models and baselines on publicly available segmentation datasets. Models of the same family are grouped by hue, with larger models in darker shades. Results vary widely by dataset, but the trend is consistent and shows that STT outperforms MobileSAM, which runs with similar latency, and is competitive with more expensive models. See Table 5 in the supplementary for a full tabular listing of the results. + +from the image boundary. The reconstruction loss is applied to the in-bounds and downsampled tokens — we do not ask the network to predict the full-resolution source image. An MAE pre-training example is visualized in Figure 6. + +After pre-training, the decoder is discarded and the encoder weights are used to initialize the STT encoder. We then use SA-1B to train the mask decoder and fine-tune the encoder weights for the segmentation task. For each image, we select a set of segments uniformly at random, and for each segment select a foveation center uniformly within the segment. We use the same combination of Focal, Dice, and IoU prediction losses as SAM. However, we made two minor changes to handle the variable resolution of the STT output masks. + +First, each pixel in the STT output may represent a classification of many pixels in the binary segmentation target. Rather than upsampling the masks to the full image resolution as described in section 3.2.2, we map the ground truth to the foveated output space of STT. This is done by converting the masks to floating point and applying the same downsampling approach as used for input pixels described in section 3.1. The result is a real-valued foveated segmentation map. Each value in the map has a variably-sized receptive field and represents the probability that a pixel selected uniformly at random from within that receptive field has a positive label in the full-resolution ground truth segmentation map. After downsampling there is a 1:1 correspondence between pixels in target and estimated masks, and we can apply the Dice and Focal losses, both of which generalize to continuous target values. + +The second change is to the targets for IoU prediction. We found that thresholding the target and estimated masks and computing IoU of the binarized masks led to unstable loss values when significant portions of either mask are near a value of 0.5. We thus use an expected IoU of the real-valued segmentation maps. Given an estimated probability $p_i$ that pixel $i$ is part of the segment and probability $q_i$ that a pixel selected uniformly at random from its receptive field + +
MethodLatency (ms)GFLOPs
SAM [20]H572.76533.7
L347.63244.5
B153.91027.0
EfficientSAM [40]S78.6489.4
Ti39.8201.2
MobileSAM [46]20.7124.4
STT (Ours)H26.2223.2
L13.7108.0
B7.330.9
+ +Table 2. Comparing the efficiency of evaluated segmentation models. Latency is measured as the time to process a single image with a single prompt on an nVidia RTX 3080 GPU, and excludes preprocessing time for all models. See supplementary materials for details on calculating GFLOPs. Note that latency is highly dependent on hardware (and, to a lesser extent, software) and that there are competing ways to compute FLOPs; we were careful to run all models on the same hardware and used the same method of computing FLOPs, and can see that FLOPs and latency are roughly linearly correlated as expected. + +has a positive label, the expected IoU is given by: + +$$ +\frac {p _ {i} q _ {i}}{1 - \left(1 - p _ {i}\right) \left(1 - p _ {q}\right)} \tag {1} +$$ + +See the supplementary materials for further training details, including hyperparameter settings. + +# 4.1.1. Foveation Center Selection + +The randomized selection of the foveation center and the resulting variability in the resampling of the image plays much the same role as data augmentation. However, rather than altering the data contained in the training set, randomized foveation simply provides many different views of the same data. While we revisit the same images from SA-1B repeatedly throughout training, the model almost certainly + +![](images/d655495d499e58f2cc2f01c3d52cb500f1fe2ecca2ac0c297bf4287911d5db57.jpg) +SAM + +![](images/f2099d81f723cce36def48dfc3c82b0bc92530634f08d19746e01ae19ff94ba7.jpg) + +![](images/7ba540929b7dcc1a28721294964531a05fb5d17c16ba6968761ce14fa9ec04a9.jpg) +Figure 8. A side-by-side comparison of our Segment This Thing model on a set of images from the SA-1B dataset [20]. We also visualize the foveated input in the rightmost column as described in Section 3.1.1. This figure is best viewed digitally — zoom in to see the relative coarsity of STT's output for larger segments. For example, the ear of the goat in the third row is missed by STT. + +![](images/e09aaffcd5ad0b733e641cc390636a71e98e511e4611148c373e24c6f126e4e2.jpg) +EfficientSAM + +![](images/3ab3a37b5df79f874e578acb5034b6f43822e816e16195c9d9a659e1d8d9e111.jpg) + +![](images/9163d1b75af74771c7efe1b11cbc550f8aa41532f2a2857d8aecdfad7729408d.jpg) + +![](images/b9a83a8f104649247cb3e7bde8b0d27426429ed4fe6ec7291d0511e848c6f685.jpg) +MobileSAM + +![](images/a28f82cea6c48e1ba8b880b594771d1054e27360d26b839624904852cfe52c17.jpg) + +![](images/3f00ccaecdb0a5ca766927c23a2af4909cf845701f706c706eb2555d5bed49a1.jpg) + +![](images/c6e5ddead0eff0fd63d2885b7418d1a7799cd9a42f524e34cf44ce68e98b0c34.jpg) +STT (Ours) + +![](images/a537152f62eeb8559a694238de99d42852beb5144f7e5e9b657963d41b3a0403.jpg) + +![](images/adc01f5ccb2a36c10a50e4d0c8ce3c450d77eff8beb5dd5066fe6d4f0c6e8d02.jpg) + +![](images/427638cdc194b784b5702f0b9222f38dfa113c123fae11c2570322cf473ee21e.jpg) +Foveated Input + +![](images/eb2fa8619e490b630e2ed49ba8288388bd5dda2e9f594a08d1bfa6fd8b36c40e.jpg) + +![](images/8d73f3425d15e8fb45b099914a84e58c713cd7ab19af6ed6fd795005780d6bca.jpg) + +never receives exactly the same foveated input twice. We train without drop path as used by Kirillov et al., but have not observed issues with overfitting. + +In some cases, one may wish to prompt a segmentation model with some uncertainty as to whether the prompt is within the intended segment. In training STT, we can optionally add a small Gaussian-distributed offset vector between the point selected from within a target segment and the center point used for prompting the model. In this setup, the model is sometimes prompted with points somewhat outside the target segment. The model then learns to produce not only segments that include the prompt point but also other nearby segments. The same strategy employed by SAM to handle size ambiguity (i.e. allowing the model to produce multiple masks, and backpropagating only gradients of the loss associated with the best mask) also naturally handles the ambiguity introduced by prompt noise. If the model is prompted with an image that contains plausible segments both at and near the center, it can simply produce both segments as output. + +# 4.2. Evaluation + +We evaluate the Segment This Thing model on the "Zero-Shot Single Point Valid Mask Evaluation" task defined by Kirillov et al. [20] in which the model is prompted with a single point in each segment that is furthest from the segment boundary. We measure accuracy using the mean intersection over union (mIoU) across all segments in each evaluation dataset. + +We compare against SAM [20] and the faster EfficientSAM [40] and MobileSAM [46] as baselines, all of + +which involve resizing the image to $1024 \times 1024$ pixels, encoding a point prompt, and decoding a mask at $256 \times 256$ pixel resolution. To evaluate STT, we do cropping and resampling as usual, but map each output mask back to the input image pixel space before evaluating the IoU. During training, both the height and width of training images are larger than the pattern size, such that crops centered near a corner may need padding on up to two sides. Some of the evaluation images are significantly smaller, however. To avoid a domain gap, we upsample images as necessary before cropping such that the crop needs padding on at most two sides. + +We analyze the efficiency of our model and the baselines in Table 2. Note that STT requires fewer FLOPs and thus achieves lower latency than other efficient variants of SAM while retaining model capacity. + +We use a set of nine datasets to evaluate segmentation accuracy, and plot the results for multiple sizes of each model in Figure 7. We adopt the standard transformer size denotations used by Vision Transformers [11]. Our STT-L model significantly outperforms STT-B, an indication that maintaining model capacity is important to the performance of our model. However, we find diminishing returns with STT-H. Our models all significantly outperform MobileSAM, which has similar latency to STT-H. STT-L is competitive with EfficientSAM-Ti in terms of accuracy, but despite the larger parameter count is still nearly $3\mathbf{x}$ faster. This is again due to our reduced token count. + +A qualitative side-by-side comparison of the three models is shown in Figure 8. Note that some of the performance gap between STT and the more expensive models is + +![](images/633084403543624f17a93d688645529d696a075fe891636b266fa3f547f11cf8.jpg) +Figure 9. A visualization of prompting Segment This Thing with gaze. In a sequence from the Nymeria dataset [26], a user does a puzzle while wearing Aria glasses [12] equipped with eye-tracking capabilities and a forward-facing camera. The wearer's gaze is inferred and projected into the camera frame, and the resulting point is used to prompt Segment This Thing. A model trained with noise on the foveation center is capable of segmenting objects that the user is looking at even when the gaze is not entirely accurate. Left: the wear is likely looking at a puzzle piece, but the estimate of their gaze projects onto a nearby point on the table (shown in red). Center: in this zoomed-in view, we can see that the most confident segment in STT's estimation is the puzzle piece. Right: a few frames later, the wearer reaches for this exact piece, lending evidence that this indeed had been the object of the wearer's attention. + +![](images/4ff7f5e11333083e3fa7e6208c42172413a0a0463f73f8f299c2eab0b1f1da4b.jpg) + +![](images/4c96f6a14008129f941e27a2a0d335d8b49e2f1f0b3455408855632ed1a303c6.jpg) + +due to the relatively coarse segment boundaries caused by foveation (see the supplementary material for further analysis of the performance of STT as a function of segment size). However, we expect that this will not be critical for many downstream tasks. If a precise boundary is important, one can always prompt the model again near the boundary once its approximate contours are known. + +# 4.3. Gaze-based Prompting + +We also show qualitative results in which we prompt the Segment This Thing model with gaze data in Figure 9 and in the supplementary video. Specifically, we take sequences from the Nymeria dataset [26] in which users wear Aria glasses [12] while performing a variety of activities. The glasses have forward-facing cameras capturing the wearer's egocentric perspective as well as cameras aimed at the their eyes enabling gaze tracking. The estimated gaze vectors can then be re-projected into the frame of reference of the forward-facing camera image and used as a prompt for STT. However, there is some noise in the estimation of the wearer's gaze and therefore some uncertainty in the prompt. Instead of prompting the model with multiple points drawn from the distribution over the true gaze point, which would incur a cost that scales linearly in the number of prompts, we use a model trained with noise as described in Section 4.1.1. With this model, using a single prompt placed at our best estimate of where the wearer is looking still often results in a segmentation of the object they are actually looking at. Furthermore, if there is only one obvious segment near the gaze vector, as might be the case when the wearer gazes at an object against a featureless background, this segment will in many cases be the highest scoring segment according to the model's IoU predictions. The example shown + +in Figure 9 is such a case. + +# 5. Conclusion + +In this paper we introduced the Segment This Thing model, a new image segmentation model that builds on the success of SAM by increasing efficiency. Instead of simply reducing the size of the Transformer-based image encoder, we applied foveated tokenization to reduce the token count by roughly 24x compared to SAM. This reduces inference latency by roughly 42x compared to SAM-H or 5.6x relative to EfficientSAM. Foveated tokenization has the additional benefit of reducing the number of pixels required as input to the network, easing up on bandwidth requirements for transmitting image data between sensors and devices; our model requires only about 44K pixels as input (roughly equivalent to a $210 \times 210$ image), but retains the ability to segment small objects at full resolution. + +Despite the increased efficiency, Segment This Thing remains competitive on mIoU metrics on multiple public datasets. By maintaining a relatively large parameter count, our model retains the ability to do more complex computation over the reduced token set compared to smaller models that do simpler computations over more tokens. This helps our models cope with the information lost due to foveation. Furthermore, foveated tokenization concentrates the computational cost where it matters most: near the prompt. + +Altogether, the Segment This Thing model is a promising choice for streaming video use cases. While we have only evaluated foveated tokenization in the context of image segmentation, we believe the approach could be equally applicable to a variety of image and video processing tasks. We leave this exploration to future work. + +# References + +[1] Dina Bashkirova, Mohamed Abdelfattah, Ziliang Zhu, James Akl, Fadi Alladkani, Ping Hu, Vitaly Ablavsky, Berk Calli, Sarah Adel Bargal, and Kate Saenko. Zerowaste dataset: Towards deformable object segmentation in cluttered scenes. In CVPR, pages 21147-21157, 2022. 6 +[2] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022.3 +[3] Marisa Carrasco, Denise L Evert, Irene Chang, and Svetlana M Katz. The eccentricity effect: Target eccentricity affects performance on conjunction searches. Perception & psychophysics, 57:1241-1261, 1995. 1 +[4] Honghao Chen, Xiangxiang Chu, Yongjian Ren, Xin Zhao, and Kaiqi Huang. Pelk: Parameter-efficient large kernel convnets with peripheral convolution. In CVPR, pages 5557-5567, 2024. 3 +[5] Mengzhao Chen, Wenqi Shao, Peng Xu, Mingbao Lin, Kaipeng Zhang, Fei Chao, Rongrong Ji, Yu Qiao, and Ping Luo. Diffrate: Differentiable compression rate for efficient vision transformers. In ICCV, pages 17164-17174, 2023. 3 +[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 6 +[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1-23, 2022. 6 +[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 4 +[9] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In NeurIPS, 2022. 6 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 3 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 4, 7 +[12] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talattof, Arnie Yuan, Bilal Souti, Brighid Meredith, et al. Project aria: A new tool for egocentric multi-modal ai research. arXiv preprint arXiv:2308.13561, 2023. 8 +[13] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and + +Jürgen Gall. Adaptive token sampling for efficient vision transformers. In ECCV, pages 396-414. Springer, 2022. 3 +[14] Jean-Michel Fortin, Olivier Gamache, Vincent Grondin, François Pomerleau, and Philippe Giguère. Instance segmentation for autonomous log grasping in forestry operations. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6064-6071. IEEE, 2022. 6 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In ICCV, pages 770-778, 2016. 2 +[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 5 +[17] Maxwell Horton, Sachin Mehta, Ali Farhadi, and Mohammad Rastegari. Bytes are all you need: Transformers operating directly on file bytes. arXiv preprint arXiv:2306.00238, 2023. 2 +[18] Aditya Jonnalagadda, William Yang Wang, BS Manjunath, and Miguel P Eckstein. Foveater: Foveated transformer for image classification. arXiv preprint arXiv:2105.14173, 2021. 3 +[19] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 3 +[20] 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 CVPR, pages 4015-4026, 2023. 1, 2, 5, 6, 7 +[21] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In ECCV, pages 620-640. Springer, 2022. 3 +[22] Robert Konrad, Nitish Padmanaban, J Gabriel Buckmaster, Kevin C Boyle, and Gordon Wetzstein. Gazegpt: Augmenting human capabilities using gaze-contingent contextual ai for smart eyewear. arXiv preprint arXiv:2401.17217, 2024. 3 +[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 2 +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 3 +[25] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In CVPR, pages 14420-14430, 2023. 2 +[26] Lingni Ma, Yuting Ye, Fangzhou Hong, Vladimir Guzov, Yifeng Jiang, Rowan Postyeni, Luis Pesqueira, Alexander Gamino, Vijay Baiyya, Hyo Jin Kim, et al. Nymeria: A massive collection of multimodal egocentric daily motion in the wild. arXiv preprint arXiv:2406.09905, 2024. 8 + +[27] Dmitrii Marin, Jen-Hao Rick Chang, Anurag Ranjan, Anish Prabhu, Mohammad Rastegari, and Oncel Tuzel. Token pooling in vision transformers. arXiv preprint arXiv:2110.03860, 2021. 3 +[28] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, pages 12309-12318, 2022. 3 +[29] Juhong Min, Yucheng Zhao, Chong Luo, and Minsu Cho. Peripheral vision transformer. NeurIPS, 35:32097-32111, 2022. 3 +[30] Massimo Minervini, Andreas Fischbach, Hanno Scharr, and Sotirios A Tsaftaris. Finely-grained annotated datasets for image-based plant phenotyping. Pattern recognition letters, 81:80-89, 2016. 6 +[31] Jeongsoo Park and Justin Johnson. Rgb no more: Minimally-decodedJPEG vision transformers. In CVPR, pages 22334-22346, 2023. 2 +[32] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. NeurIPS, 34:13937-13949, 2021. 3 +[33] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2 +[34] Cedric Renggli, André Susano Pinto, Neil Houlsby, Basil Mustafa, Joan Puigcerver, and Carlos Riquelme. Learning to merge tokens in vision transformers. arXiv preprint arXiv:2202.12015, 2022. 3 +[35] Zhuoran Song, Yihong Xu, Zhezhi He, Li Jiang, Naifeng Jing, and Xiaoyao Liang. Cp-vit: Cascade vision transformer pruning via progressive sparsity prediction. arXiv preprint arXiv:2203.04570, 2022. 3 +[36] Xiaorui Sun, Jun Liu, Heng Tao Shen, Xiaofeng Zhu, and Ping Hu. On efficient variants of segment anything model: A survey. arXiv preprint arXiv:2410.04960, 2024. 2 +[37] Cameron Trotter, Georgia Atkinson, Matt Sharpe, Kirsten Richardson, A Stephen McGough, Nick Wright, Ben Burville, and Per Berggren. Ndd20: A large-scale few-shot dolphin dataset for coarse and fine-grained categorisation. arXiv preprint arXiv:2005.13359, 2020. 6 +[38] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 2 +[39] Balakrishnan Varadarajan, Bilge Soran, Forrest Iandola, Xiaoyu Xiang, Yunyang Xiong, Lemeng Wu, Chenchen Zhu, Raghuraman Krishnamoorthi, and Vikas Chandra. Squeezezam: User friendly mobile interactive segmentation. arXiv preprint arXiv:2312.06736, 2023. 2 +[40] Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, et al. Efficientsam: Leveraged masked image pretraining for efficient segment anything. In CVPR, pages 16111-16121, 2024. 2, 6, 7 +[41] Wilson Yan, Matei Zaharia, Volodymyr Mnih, Pieter Abbeel, Aleksandra Faust, and Hao Liu. Elastictok: Adap + +tive tokenization for image and video. arXiv preprint arXiv:2410.08368, 2024. 3 +[42] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In CVPR, pages 10809-10818, 2022. 3 +[43] Senthil Yogamani, Ciarán Hughes, Jonathan Horgan, Ganesh Sistu, Padraig Varley, Derek O'Dea, Michal Uricár, Stefan Milz, Martin Simon, Karl Amende, et al. Woodscape: A multi-task, multi-camera fisheye dataset for autonomous driving. In ICCV, pages 9308-9318, 2019. 6 +[44] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2 +[45] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arXiv preprint arXiv:2406.07550, 2024. 3 +[46] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 2, 6, 7 +[47] Lingzhi Zhang, Shenghao Zhou, Simon Stent, and Jianbo Shi. Fine-grained egocentric hand-object segmentation: Dataset, model, and applications. In ECCV, pages 127-145. Springer, 2022. 6 +[48] Zhuoyang Zhang, Han Cai, and Song Han. Efficientvit-sam: Accelerated segment anything model without performance loss. In CVPR, pages 7859-7863, 2024. 2 +[49] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 2 +[50] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. IJCV, 127: 302-321, 2019. 3, 6 +[51] Chong Zhou, Xiangtai Li, Chen Change Loy, and Bo Dai. Edgesam: Prompt-in-the-loop distillation for on-device deployment of sam. arXiv preprint arXiv:2312.06660, 2023. 2 \ No newline at end of file diff --git a/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/images.zip b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9ea9c643fa996cc0e6449cd096a6500247e9e6cf --- /dev/null +++ b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5620394b20f9d3d6221a6698ade067185d2ec17a26d97990d925d4d4e7728ce6 +size 459507 diff --git a/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/layout.json b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4bfa9a9d23f9c79c1a70b30ebebc5877e774c5a6 --- /dev/null +++ b/CVPR/2025/Segment This Thing_ Foveated Tokenization for Efficient Point-Prompted Segmentation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7cb3971e1b35e0f3fa7aa21657a79be37f395bd24c6a57c437095b5759b15a82 +size 338249 diff --git a/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/9fcc46a3-6daf-45d8-9a08-aa8a9a72949e_content_list.json b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/9fcc46a3-6daf-45d8-9a08-aa8a9a72949e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cdb652509ad6acfc5bd125b7ad6a727222ad559e --- /dev/null +++ b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/9fcc46a3-6daf-45d8-9a08-aa8a9a72949e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8602e5ef81037a2dac49e73125c7a2b52c1b7ade3380bc57f22e126c043dc02 +size 93303 diff --git a/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/9fcc46a3-6daf-45d8-9a08-aa8a9a72949e_model.json b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/9fcc46a3-6daf-45d8-9a08-aa8a9a72949e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e55aa0ac5bd4c33e367aab45d2d7b6f71978a821 --- /dev/null +++ b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/9fcc46a3-6daf-45d8-9a08-aa8a9a72949e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00cdb4a0272ab8b47c39d179db6134d046637d71a46194901d685090b46349c0 +size 114027 diff --git a/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/9fcc46a3-6daf-45d8-9a08-aa8a9a72949e_origin.pdf b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/9fcc46a3-6daf-45d8-9a08-aa8a9a72949e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a9a1400fc7d6672efc8851189bc5b3130d312ee9 --- /dev/null +++ b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/9fcc46a3-6daf-45d8-9a08-aa8a9a72949e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:357d90b031adf06ae989b9d0dfd178704ede2e35ce0e131d8f6aa5c4a9baeb2a +size 7722343 diff --git a/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/full.md b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..84c928869c5121c59cc7cdcbd83d5d0e2e7a91f7 --- /dev/null +++ b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/full.md @@ -0,0 +1,371 @@ +# Segmenting Maxillofacial Structures in CBCT Volumes + +Federico Bolelli1, *, Kevin Marchesini1, *, Niels van Nistelrooij2, 3, Luca Lumetti1, Vittorio Pipoli1, Elis Ficarra1, Shankeeth Vinayahalingam2, Costantino Grana1 + +1University of Modena and Reggio Emilia, Italy +2Radboud University Medical Center, the Netherlands +3Charité - Universitätsmedizin Berlin, Germany + +# Abstract + +Cone-beam computed tomography (CBCT) is a standard imaging modality in orofacial and dental practices, providing essential 3D volumetric imaging of anatomical structures, including jawbones, teeth, sinuses, and neurovascular canals. Accurately segmenting these structures is fundamental to numerous clinical applications, such as surgical planning and implant placement. However, manual segmentation of CBCT scans is time-intensive and requires expert input, creating a demand for automated solutions through deep learning. Effective development of such algorithms relies on access to large, well-annotated datasets, yet current datasets are often privately stored or limited in scope and considered structures, especially concerning 3D annotations. This paper proposes ToothFairy2, a comprehensive, publicly accessible CBCT dataset with voxel-level 3D annotations of 42 distinct classes corresponding to maxillofacial structures. We validate the dataset by benchmarking state-of-the-art neural network models, including convolutional, transformer-based, and hybrid Mamba-based architectures, to evaluate segmentation performance across complex anatomical regions. Our work also explores adaptations to the nnU-Net framework to optimize multiclass segmentation for maxillofacial anatomy. The proposed dataset provides a fundamental resource for advancing maxillofacial segmentation and supports future research in automated 3D image analysis in digital dentistry. + +# 1. Introduction + +Medical imaging is currently employed in the clinical practice of dental treatments. Different modalities are available to assist in diagnosis, treatment planning, and surgery: + +![](images/c6497b378783967f7652036f8dd6199e6639712c99dd347eec98ed8f4a680ced.jpg) +Figure 1. Fully annotated sample from the proposed dataset. + +2D panoramic X-rays, 3D Intra-Oral Scans (IOS), and 3D Cone-Beam Computed Tomography (CBCT) [29]. Among the available modalities, CBCT is the only one providing comprehensive 3D volumetric information on all the anatomical structures in the orofacial area, including jawbones, complete teeth, sinuses, and neurovascular canals, becoming a standard modality for maxillofacial image analysis. Segmenting individual structures from CBCT images to reconstruct a precise 3D model is essential in digital dentistry for various clinical applications, including surgical planning and implant placement [16, 53]. + +Manual segmentation of CBCT scans is a time-consuming and labor-intensive process that requires specialized expertise. For this reason, automated algorithms using deep learning techniques offer a promising solution to alleviate this burden, enhance efficiency, and improve the consistency of segmentation results [41]. However, developing robust deep learning models heavily depends on the availability of large, annotated datasets. Unfortunately, existing datasets often focus on a limited number of anatomical structures — i.e., teeth or alveolar canal — and usually lack 3D annotations. The few available methods trained on + +Table 1. Datasets used in literature to segment maxillofacial anatomical structures. Our analysis is limited to 3D imaging modalities, i.e., CBCTs and IOS. $\clubsuit$ means that only a portion of the data is available to the research community, i.e., 148 over 4531. Among these, only 97 are effectively usable for the training of automatic algorithms. + +
Anatomical Structure(s)Image ModalityAuthorsYearCountry# Train & Validation# TestLabel TypePublic
TrainTest
Inferior Alveolar Canal (IAC)CBCTJaskari et al. [28]2020Finland5091282D128 2D, 15 3DX
Lahoud et al. [33]2022Belgium205303D3DX
Usman et al. [48]2022South Korea5105003D3DX
Cipriano et al. [11]2022Italy33215332 2D, 76 3D15 2D, 15 3D✓
Chun et al. [10]2023South Korea32183D3DX
Bolelli et al. [6]2023Italy, Netherlands44350443 2D, 153 3D3D✓
Teeth Crown, Teeth RootCBCTCui et al. [16]2022China45314073D3D◇
Cui et al. [13]2022China22~3D~✓
Dou et al. [17]2022China3553D3DX
Tae Jun Jang et al. [26]2022China66766 3D, 31 2D7 3D, 4 2DX
Teeth CrownIOSBen-Hamadou et al. [3]2022Tunisia12006003D3D✓
Shankeeth et al. [49]2023Netherlands14003503D3DX
IACs, Teeth, OthersCBCTOurs2024Italy, Netherlands480503D3D✓
+ +3D datasets leverage very small-sized training and testing sets ( $< 50$ CBCTs) [10, 13, 17], most of the time unavailable to the research community, limiting their generalizability to scans acquired with other imaging protocols and their applicability to diverse patient populations. + +In this paper, we introduce a novel and unique dataset with 3D annotations of 42 distinct anatomical classes on maxillofacial CBCTs. The classes include jawbones, left and right Inferior Alveolar Canals (IACs), left and right maxillary sinus, pharynx, upper and lower teeth, including wisdom teeth, bridges, crowns, and dental implants. The proposed comprehensive dataset provides a rich resource for developing and benchmarking segmentation algorithms that require 3D anatomical annotations, providing a strong basis for the future development and evaluation of automated methods for maxillofacial image analysis (Fig. 1). + +To validate the dataset, we further explore the performance of various state-of-the-art neural networks used in the literature for medical image segmentation: classical Convolutional Neural Networks (CNNs) like nnU-Net and its variants [24], transformer-based models such as TransU-Net [7] and UNETR++ [47], and hybrid Mamba-based architectures like UMamba [38], Swin-UMamba [34], and VMamba [35]. + +Specifically, we explore how to adapt and optimize the nnU-Net framework [24], a state-of-the-art, self-configuring segmentation method, to enhance the segmentation of all 42 anatomical structures in our dataset, by adjusting hyperparameters and augmentation strategies. + +By leveraging our comprehensive dataset, we aim to advance the development of segmentation algorithms that can handle the complexities of 3D maxillofacial imaging. Our work not only contributes a valuable dataset to the research community but also provides insights into the optimization of neural network architectures for improved segmentation performance in complex anatomical regions. + +In summary, our contributions are: + +- The largest publicly available CBCT dataset with detailed 3D annotations covering 42 distinct maxillofacial anatomical structures, addressing a critical gap in existing research resources; +- A comprehensive comparative evaluation of state-of-the-art neural network architectures, encompassing convolutional networks, transformers, and hybrid models, to establish performance benchmarks using our dataset; +- Adaptation and optimization of the nnU-Net framework to target multi-class segmentation in complex maxillofacial anatomy, highlighting the impact of customized augmentation and hyperparameters tuning. + +# 2. Related Works + +# 2.1. Existing datasets + +In maxillofacial analysis, existing research predominantly emphasizes developing AI applications that demonstrate high accuracy on carefully selected datasets; however, limited consideration has been given to examining the datasets employed in the training and evaluation phases of AI model development [45]. Current research has primarily used academic databases, which may not fully address AI-specific data needs, including the FAIR (Findability, Accessibility, Interoperability, and Reusability) principles [52]. Only a few dental studies make datasets available to the research community [6, 12, 13, 16], making it essential to seek alternative data sources to mitigate these limitations and to prioritize the creation of high-quality, medically validated, and AI-ready datasets. All of these datasets are confined to specific anatomical structures, particularly the IAC [12] or the teeth complex [16], and others provide binary segmentation classes only, without considering tooth-level labels [16]. + +Tab. 1 presents a comprehensive comparison of datasets used in state-of-the-art research. + +# 2.2. Teeth segmentation + +Multiple approaches to segmenting teeth in 3D data have been proposed, each introducing different challenges in clinical and computational applications [8]. + +Binary segmentation. When employing a simple binary semantic segmentation, the primary goal is to separate teeth from surrounding tissues without any distinction between individual teeth. While it is useful for basic assessments, its clinical utility is limited, especially when dealing with adjacent or closely spaced teeth. To improve accuracy, recent methods have utilized CNN-based architectures with custom modules, hybrid loss functions [21], a combination of 2D and 3D networks [20], and post-processing techniques, such as posterior probability maps and dense conditional random fields [42]. + +Instance segmentation. By contrast, when performing instance segmentation, teeth are segmented as separate instances, though they have not yet been assigned specific labels. Instance segmentation methods usually leverage a two-stage approach consisting of detection followed by segmentation. The detection stage locates individual tooth instances through various techniques, including bounding box detection [14, 18], heatmap prediction [54], or offset regression [15-17]. Once identified, localized regions around each tooth instance can be extracted to perform binary segmentation and identify precise boundaries. + +Multi-class segmentation. With this approach, each tooth must be identified and assigned a specific label [32, 46] according to the FDI (Fédération Dentaire Internationale, also known as World Dental Federation) notation [25]. This is the most challenging of the three approaches, especially in cases with a complex anatomy, such as missing and impacted teeth, and will likely continue to benefit from more robust datasets and enhanced model architectures, motivating once again our proposed dataset. + +Current literature can be split into single- and multi-stage methods. While the former leverages a single framework to produce the final classification, multi-stage methods incorporate additional steps. The more commonly employed strategy consists of downsampling the original CBCT or splitting it into multiple regions (e.g., the four dental arch quadrants) to generate a coarse multi-class segmentation, later refining individual tooth segments by isolating volumes of interest and enhancing local details [46, 50]. + +# 2.3. IAC segmentation + +Since the advent of CBCT technology in the early 2000s, substantial effort has been dedicated by the scientific community to developing automated systems for segmenting the inferior alveolar canal from 3D scans [44]. With the rise of deep learning in medical diagnosis, learning-based networks [23, 27, 28] have outclassed traditional computer vision techniques [1, 4, 30, 31, 40, 51]. + +Early work by Jaskari et al. [28] applied the U-Net [43] architecture on a coarsely annotated dataset, achieving promising results compared to traditional methods, though limited by the lack of dense annotations. + +Lahoud [33] advanced this line of research by training a 3D U-Net with both interpolated control points and finer voxel-level annotations, though, again, access to data and code was restricted. Significant progress in the segmentation of the IAC came with the release of the first public CBCT dataset containing 2D and 3D annotations of the mandibular canal [11, 12]. Contextually, authors introduced PosPadUNet3D, a modified 3D U-Net that segments IAC in stages by expanding sparse 2D labels into 3D annotations, thus generating additional synthetic labels and using such generated data in combination with expert-labeled 3D voxel data to train the final U-Net based model. + +Usman et al. [48] proposed a two-stage U-Net approach to address class imbalance between the mandibular canal and background, using CNNs to isolate regions of interest (ROIs) before segmentation, while Zhao et al. [55] introduced a Frenet frame-based method to better capture mandibular topology and maintain canal structure during segmentation. More recently, Lv et al. [37] developed a transformer-based model with adaptive image processing and a "deep label fusion" technique to enhance label consistency across sparse public data, building on Cipriano's label expansion methodology. + +Finally, Lumetti et al. [36] tackled issues with patch-based learning by using a memory-augmented transformer encoder to improve spatial coherence in the U-Net bottleneck, enriching patch context and improving segmentation. + +IAC segmentation remains a challenging area with room for further advancements, largely dependent on access to comprehensive, high-quality 3D datasets. + +# 3. The ToothFairy2 Dataset + +The proposed dataset, ToothFairy2, represents the most extensive publicly available and fully annotated collection of dental CBCT scans. It comprises 530 3D volumes, of which 480 are entirely accessible for training purposes1 while the remaining 50 are reserved exclusively for evaluation. Although these 50 scans are not directly accessible for model training, researchers can evaluate their models on this test set via the grand-challenge platform,2 which automatically runs the submitted models and provides performance results. In this way, we ensure that researchers will not incur the risk of mixing training and testing data and that the reported results will not suffer from the random choice of the testing data subset selection. + +The training data (480 scans) were acquired by the Af + +![](images/c62f5df6e6670faef9858424eb30e7517b66151085efbd63dc75c6b21bf4fcaf.jpg) + +![](images/892b667ffd7904ce0080cdfa4a51863bdfce1264f010d79122b4802e1bc14dff.jpg) + +![](images/3d26511f772bf29bbba51f13a64f9e029685cb6a18769b09fb6e83f24fcb7c6c.jpg) + +![](images/6f09b7989ac5bed90862b732c669f72ef195dcd00fe729a4b9e3c127304b153a.jpg) + +![](images/27b27a259850fd96d2ed16232ffd86a61a29346343d3cc9e9e285e5972a0a2fe.jpg) + +![](images/1dde50e6d9231b74de8c8c4ccf1cac8c07d9cfed147a61ba702fc12d85b11305.jpg) + +![](images/d7063cc7e1721da1c808223d527d3c33cfcb2a21d92d629cbe0dcd490b0a2ed0.jpg) + +![](images/cc316baf243064c0b7ccad9330240050630a61a240f4e3b4ce4705919cfe9fc4.jpg) + +![](images/2b7f174e7dd67da0bb41bb4c9382c3928fa431e66cfc2c94e4d9daad4cab7a1d.jpg) + +![](images/96cef07db3696de64009751a4df59b0c8c304a4b2f41c2881e3d87cde3298eb1.jpg) + +![](images/410d861eb8a55c327e74b5f2c0c409f9ae75347c9a7a393845edaa4aea11fd13.jpg) + +![](images/1135e09da3c6521ac8d8948247a84be4b16160f73ae170521f843980fbb6e87f.jpg) + +![](images/5066c58cea10fc2e254ea477cd63300c603f4262d11a499795131e2df9ce855a.jpg) +Figure 2. Dataset samples, one patient per line. For each view, i.e., left, frontal, and right, both raw CBCT data and labels are provided. + +![](images/a8db2fd95d352d7a92b694a33353522ecee9478e548b15175aff376a5dc0d032.jpg) + +![](images/062b1c653c63a93480af8d5d8b6e3bde95df98693100483bf60bf331960a09df.jpg) + +![](images/64ea54a318a1650de52c08ab35ad06b351ea57c0d9757367ea59fce3947584dd.jpg) + +![](images/330c772e3095e469b1837dd7afe7df754f05e614f1a3dc62313235446d45a030.jpg) + +![](images/94c5d9027e226a36bebf5329d1a606328d5393ab226a016da7373a88f39f1661.jpg) + +fidea Center, a pan-European healthcare group that specializes in advanced diagnostics, laboratory analyses, rehabilitation, and cancer diagnosis and treatment. The scans were obtained through Cone Beam Computed Tomography (CBCT) via a NewTom/NTVGiMK4, operating at $3\mathrm{mA}$ , 110 kV, and offering $0.3\mathrm{mm}$ isotropic voxels (Fig. 2). The test data (50 scans) were provided by the Department of Oral and Maxillofacial Surgery at Radboud University Medical Centre, Nijmegen, obtained using a standard CBCT scanning protocol with the i-CAT 3D Imaging System. The scans were performed with a field of view measuring $16\mathrm{cm}$ in diameter and $22\mathrm{cm}$ in height. The scanning process involved two scans of 20 seconds each, resulting in an isotropic voxel size of $0.4\mathrm{mm}$ . These were later rescaled to match the voxel size of the training data before annotation. + +Every patient was anonymized and we only kept a few personal details—namely gender, age, and year of the scan. Specifically, $58.30\%$ of the patients are female (58.54% in the training set, 56.00% in the test set), all the scans were performed between 2019 and 2024, and volumes belong to patients with ages in range (10-100] with the highest frequencies in ranges (20-30] and (60-70] for the training set and (50-70] for the test set. + +# 3.1. Annotation protocol and tools + +The inclusion of densely labeled 3D data is necessary in order to achieve the full potential of CNNs [22]. 3D annotations refer to the annotation process carried out by medical professionals directly at the voxel level on CBCT scans. This annotation approach involves detailed markings applied to the individual voxels, providing a more precise depiction of the structures involved. + +As often happens in medical image annotation, accurate delineation of individual teeth and maxillofacial structures is a complex and time-consuming process, which requires + +expert eyes to be achieved. To reduce the burden of the annotators while ensuring high-quality labels, we employed a semi-automated approach that was fully supervised by clinicians. The annotation process was performed by 7 different maxillofacial experts. To prevent models trained on our dataset from being rewarded for learning annotator-specific biases and to allow for the understanding of their generalizability, 5 of the annotators focused on the training set only, while the remaining 2 handled the test cases only. + +Since labeling voxels directly on 3D views is ambiguous and prone to errors, our annotations have been performed on 2D slices. Moreover, to reduce the jagged effect that usually occurs when annotating 3D objects on 2D, the annotation is performed with multiple iterations, starting from the axial (or transversal) view and later refined by analyzing the scan from the sagittal and coronal (or frontal) planes. + +For each patient, a single annotation is provided, including the following classes: upper and lower jawbone, left and right inferior alveolar canal, maxillary sinuses, pharynx, and upper and lower teeth, including wisdom teeth, bridges, crowns, and dental implants. The annotation was produced by a single expert only but validated by a second one to reduce possible mistakes. + +In our work, we leverage five different base models built on the nnU-Net framework to provide an initial annotation that is later adjusted and refined by expert clinicians. Each of the models has been trained on a specific (group of) classes, including the jawbone (upper and lower), inferior alveolar canal (left and right branches), maxillary sinus (left and right), pharynx, and teeth. Regarding inferior alveolar canal branches and teeth, we leverage two datasets available in the literature, [12] and [16], respectively. Even when annotating the first volumes of our dataset, clinicians were provided with an initial segmentation of teeth and canals and were required to modify, update, or remove wrongly an + +![](images/437384607e6b776c109cee80ccd8efce4f2a2067ab8432ab9af2d0fbe477c63a.jpg) +(a) Dense Point Set + +![](images/374d69b7cae879f6c0f6d72f09c34df766f216429edb12533e798bf661059a6b.jpg) +(b) Voxelized Teeth + +![](images/d99b8de6de5372c3266eea0789923ad6e96984cdf720dc17f4d01e69e048f336.jpg) +Figure 3. Annotations post-processing. The annotations produced as described in Sec. 3 are dense and jagged (a). For this reason, we compute a concave $\alpha$ -shape, which, after voxelization, produces the volumes in (b). Such an approach is applied to all of the anatomical structures involved in our dataset but the jawbones, whose annotations contain multiple holes. +(a) Point Set +Figure 4. $\alpha$ -shape construction process for the point set of (a). In the Delaunay triangulation (b), triangles with circumradius $\leq -\frac{1}{\alpha}$ form a simplicial subcomplex known as $\alpha$ -complex (c), whose border is the $\alpha$ -shape (d). + +![](images/48ccff90583382753def5c2a9840ba91413e7ccd3a414a8bedc086ff34140422.jpg) +(b) Triangulation + +![](images/4ddb4d222bd2c3a6c9be1b2c0e5e213c77fb70df9cac67a8bef255ebd66a48e3.jpg) +(c) $\alpha$ -complex + +![](images/54083ad3159b4228138f2d246cc196dc5b99fff28d164a87329fd9b71fa959a9.jpg) +(d) $\alpha$ -shape + +notated voxels. Following an iterative approach, as soon as 20 more annotated volumes were available, the base models were retrained from scratch by including the new data. Following such an approach, clinicians are provided with better annotation proposals at each iteration, reducing the annotation effort and time at the increase of annotated volumes. Since jawbone, maxillary sinus, and pharynx classes were not available in any public dataset, these classes have been annotated from scratch in the first 20 volumes without leveraging any automatic proposal. + +# 3.2. Annotation refinement + +Even though the annotation process is performed on multiple planes, one after the other, artifacts due to 3D annotations performed on 2D views remain a challenge, albeit reduced (Fig. 3a). In order to obtain a smooth polygonal mesh out of expert-produced annotations, we compute a concave $\alpha$ -shape. The $\alpha$ -shape [19] is a generalization of the convex hull aimed at representing the intuitive concept of the shape of a point set. The only parameter of the algorithms is $\alpha \in \mathbb{R}$ , which regulates the "crudeness" of the result. Let's define a generalized disk of radius $\frac{1}{\alpha}$ as $D_{\alpha}$ : + +$$ +D _ {\alpha} \left\{ \begin{array}{l l} \text {T h e c o m p l e m e n t o f a d i s c o f r a d i u s} - \frac {1}{\alpha}, & \text {i f} \alpha < 0 \\ \text {A h a l f p l a n e ,} & \text {i f} \alpha = 0 \\ \text {A d i s c o f r a d i u s} \frac {1}{\alpha}, & \text {i f} \alpha > 0 \end{array} \right. \tag {1} +$$ + +Given a point set $S$ and a value for $\alpha$ , the $\alpha$ -shape is con- + +structed in this way: an edge is put between two points $p_i$ and $p_j$ whenever there exists a $D_{\alpha}$ with $p_i$ and $p_j$ lying on its boundary, and which contains the entire $S$ . When $\alpha = 0$ , this procedure constructs the convex hull; instead, cruder or finer shapes can be respectively obtained using positive or negative $\alpha$ values. Because of the geometrical nature of the alveolar nerve, we are specifically interested in concave $\alpha$ -shapes, achievable when $\alpha < 0$ . + +The most common method for computing a concave $\alpha$ -shape consists of taking the border of a simplicial subcomplex extracted from the Delaunay triangulation, containing only triangles with circumradius $\leq -\frac{1}{\alpha}$ . An example of the process is depicted in Fig. 4. + +The above concepts can be extended to the three-dimensional case by substituting disks and triangles with spheres and tetrahedra, respectively. After creating an $\alpha$ -shape polygonal mesh starting from dense annotations, it is converted into a binary raster volume by means of voxelization: the final result is given in Fig. 3b. + +# 3.3. Accessibility + +The resulting annotated volumes have been packed following the nnU-Net dataset format, which comprises three different components: raw images, corresponding segmentation maps, and a dataset. .json file specifying the metadata. The class IDs are an extension of the FDI notation and include a total of 42 classes. Final 3D volumes and labels are provided in the .mha format. + +Ethics approval. The training data received ethical committee approval from Comitato Etico dell'Area Vasta Emilia Nord (Approval Number 1374/2020/OSS/ESTMO SIRER ID 1275 - NAICBCT-D) and can be downloaded under the CC BY-SA license after user registration. + +# 4. Experiments + +In order to validate the effectiveness of the proposed dataset, experiments have been performed considering two different datasets: the one proposed with this paper and the publicly released subset of the dataset proposed by Cui et al. [15], which is the only dataset publicly available in the literature for multi-class tooth instance segmentation. Many of the dental-related scientific papers promising the release of the dataset upon request [2, 9] have ignored our emails. + +# 4.1. Evaluation metrics + +In our experiments, two widely accepted metrics for the segmentation task [39] have been employed, namely Dice Similarity Coefficient (DSC in $\%$ ) and the 95th percentile Hausdorff Distance (HD95 in mm). + +The DSC has the same meaning as the IoU (Intersection over Union), but DSC is better suited when the region of interest is significantly smaller than the background. In such + +Table 2. Results on the test set of the proposed dataset. Classes are grouped by the main anatomical structure to which they belong. Default nnU-Net plan is employed. The best results are in bold, while the second best are underlined. + +
ModelAverageL/R IACL/R SinusTeethJawbonesPharynxOthers
DSCHD95DSCHD95DSCHD95DSCHD95DSCHD95DSCHD95DSCHD95
CNNsnnU-Net [24]70.9217.8671.3429.1164.8128.3973.1718.3290.3112.5395.6619.2329.5020.95
nnU-Net ResEnc [24]74.1614.4873.0127.8165.7129.9976.4814.2691.7712.5395.2617.7537.1016.32
Transf.TransU-Net [7]70.3220.1781.9611.9959.6959.7672.4615.0490.3349.0287.8942.8727.6822.65
nnFormer [56]76.795.4572.2810.0675.118.2279.372.9491.5020.0990.8524.5318.9513.67
UNETR++ [47]71.4317.2368.8715.1074.5115.4873.7017.9192.385.4491.6018.9626.2119.70
MambaUMamba [38]85.055.2885.2616.2377.024.3586.582.2390.0522.1792.1825.2543.8913.97
VMamba [35]73.135.1760.629.9473.209.8975.993.3990.757.1788.2323.9514.8614.63
Swin-UMamba [34]79.642.9472.642.0287.752.3880.712.5994.294.2593.052.7925.3014.93
+ +a context, more weight is given to the correctly identified region, making the DSC more robust and informative than IoU. The relationship between DSC and IoU is expressed by the following formula: + +$$ +D S C (P, G T) = \frac {2 \times | P \cap G T |}{| P | + | G T |} = \frac {2 \times I o U}{1 + I o U} \tag {2} +$$ + +where $P$ is the model prediction and $GT$ is the ground truth. + +On the other hand, the HD95 computes the maximum distance between two sets of points, considering the 95th percentile of these distances. In general, the 95th percentile of the distances between boundary points in A and B is defined as follows: + +$$ +\mathrm {d} _ {9 5} (A, B) = \mathrm {x} _ {a \in A} ^ {9 5} \left\{\min _ {b \in B} d (a, b) \right\} \tag {3} +$$ + +where $x_{a \in A}^{95}$ denotes the 95th percentile of the elements in the set enclosed within the brackets. Given the set formed by the pixels in the predicted mask $(P)$ and the set of pixels belonging to the ground truth $(GT)$ , the Hausdorff distance is determined as the maximum value of the two distances between $P$ and $GT$ and $GT$ and $P$ at the 95th percentile: + +$$ +\mathrm {H D 9 5} (\mathrm {P}, \mathrm {G T}) = \max \left\{d _ {9 5} (P, G T), d _ {9 5} (G T, P) \right\} \tag {4} +$$ + +By using the 95th percentile, this metric provides a robust evaluation that is less sensitive to outliers or extreme differences between the sets of points. + +# 4.2. Compared algorithms + +Experimental analysis has been conducted on recently proposed general-purpose state-of-the-art algorithms for segmenting medical 3D volumes, considering all of the top-notch architecture (i.e., CNNs, Transformers, and Mamba-based hybrid solutions). For what concerns CNNs, we include the original nnU-Net [24] configuration making use of the U-Net architecture (nnU-Net), and its variations leveraging residual connections in the encoder (nnU-Net ResEnc). The considered Tranformer-based architectures are TransU-Net [7], UNETR++ [47], and the recently published nnFormer [56]. Finally, we include UMamba [38], VMamba [35], and Swin-UMamba [34] as the representatives of solutions based on state-space models. + +# 4.3. Experimental setting + +In all of the experiments carried out, we adopted a standardized scheme for hyperparameters configuration. More specifically, we leverage the planning provided by the nnU-Net framework. For intrinsically 3D algorithms, i.e., nnU-Net, TransU-Net, nnFormer, UNETR++, and UMamba, the nnU-Net planning applied to the training set of the proposed dataset suggests a patch size of $80 \times 160 \times 160$ , and a batch size of 2. The only exception is the nnU-Net ResEnc, which adopts a patch size of $122 \times 224 \times 256$ . On the other hand, models that produce 3D segmentations working slice-by-slice, i.e., VMamba and Swin-UMamba, have a suggested patch size of $384 \times 384$ and a batch size of 19. + +All the images in the training set of the proposed dataset have the same spacing (0.3 mm isotropic), so no resampling is required. The volumes from the secondary dataset employed in our experiments, Cui et al. [16], have a mixed voxel spacing instead, so they have been resampled to match $0.3\mathrm{mm} \times 0.3\mathrm{mm} \times 0.3\mathrm{mm}$ . + +Models are trained from scratch without any pre-training data. The nnU-Net three-fold cross-validation schema has always been employed for model selection. All the models have been trained for a total of 1000 epochs on a 48GB A40 Nvidia GPU using CUDA 11.8 and PyTorch 2.1.2. + +# 4.4. Results + +To fully explore the capabilities of state-of-the-art models on the proposed comprehensive dataset, Tab. 2 is provided. For space constraints, results are grouped by anatomical structures by averaging single-class results. The "others" column includes implants, artificial crowns, and bridges. Surprisingly, 2D models, i.e., VMamba and Swin-UMamba, outperform intrinsically 3D ones in many scenarios. Although performance is satisfying, 2D models lack the ability to fully capture and learn the spatial relationships between the 3D structures in the maxillofacial region due to a loss of contextual information and a limited capacity to learn effective representations of complex structures. This is also confirmed by the qualitative results discussed in Sec. 4.7, where it is evident that predictions are more noisy and seem fragmented. Anyway, their local ability to dis + +Table 3. Results on the test set of the proposed dataset when using variations of the nnU-Net ResEnc. + +
ModelDSCHD95
Default74.1614.48
w/o l/r mirroring80.7912.37
w/o l/r mirroring, increased depth82.1111.86
w/o l/r mirroring, increased depth, post-processing84.998.57
+ +criminate among classes ensures high-level performance. + +Overall, Mamba-based architectures are the most effective in this context. Among all, UMamba provides the best performance. Blending convolutions to model precise spatial information and state-space models to learn long-range voxel-level interactions ensures high accuracy. Mamba provides a global context alongside voxel-wise precision, the former missing in traditional convolutional layers due to limited receptive fields and the latter absent in Transformers due to computational complexity. Moreover, it requires less training data to converge, making it more suitable in the medical field where access to data is limited. + +Considering the different classes available in the dataset, jawbones and pharynx achieved the highest DSC scores, independently from the considered model. This could be explained by the consistent anatomical shape those structures display throughout the entire dataset. Performance drops when it comes to segmenting classes subject to more variability, like teeth; their position and orientation change considerably from patient to patient, there might be missing teeth, or they can be covered by artificial crowns, or completely replaced by implants or bridges. It is not uncommon for the models to confuse artificial with natural teeth. Independently from the model employed, the worst performances are obtained when segmenting "others" classes, which represent all of the artificial structures that restore or replace natural teeth. One reasonable explanation is that these classes are underrepresented in the proposed dataset. + +# 4.5. Improving default nnU-Net configuration + +To augment the training dataset and improve performance, the nnU-Net framework applies mirroring along multiple spatial dimensions. Such an augmentation usually improves model performance, providing additional useful training data without compromising the ability to differentiate between left and right organs. In the maxillofacial context, however, there is pronounced symmetry between structures: teeth, nerves, and maxillary sinuses are all symmetrical w.r.t. to the sagittal plane splitting the face. In such a scenario, ensuring precise orientation is fundamental. The application of the left/right mirroring augmentation has exactly the opposite effect: instead of increasing model performance and generalization capabilities, it leads to a drop in the model's ability to establish a reliable left/right distinction, downgrading overall performance (Tab. 3). Additionally, given the complexity of the proposed dataset, we + +Table 4. Results on the test set of the proposed dataset when using our plan w/o post-processing. + +
ModelDefault PlanOur Plan
DSCHD95DSCHD95
CNNsnnU-Net [24]70.9217.8678.2413.36
nnU-Net ResEnc [24]74.1614.4882.1111.86
Transf.TransUNet [7]70.3220.1775.127.17
nnFormer [56]76.795.4579.125.15
UNETR++ [47]71.4317.2383.573.04
MambaUMamba [38]85.055.2885.394.70
VMamba [35]73.135.1774.214.37
Swin-UMamba [34]79.642.9481.952.91
+ +enriched the network topology by introducing an additional layer. Such an enhancement ensures an increased receptive field in the bottleneck of the network, improving contextual information and the model's ability to identify long-range relationships between structures. Finally, considering the nature of the structures involved in the dataset, all of the voxels belonging to the same class should be in a close relationship and connected to each other. For this reason, we introduce a post-processing technique that computes connected components [5] and filters out predictions that are below a given threshold or the smallest prediction if multiple objects for the same class are identified by the model. Thresholds are computed per class using statistics on the training dataset, i.e., minimum connected component volume. Such a procedure ensures us an additional gain of more than 2 DSC points on the proposed dataset. + +Finally, in order to confirm the effectiveness of the improved nnU-Net planning on the proposed dataset, i.e., disabling $1 / \mathbf{r}$ mirroring and the deeper network topology, we apply the same strategy to all of the considered models. Results are reported in Tab. 4. + +# 4.6. Comparison with existing datasets + +Tab. 5 provides a comparison of state-of-the-art models, one from each category, trained on either the proposed data or a subset of it and Cui et al. [16]. To perform a fair comparison, we selected 82 random scans from both datasets to be used for training. Since [16] includes only tooth annotations, we set all non-teeth classes to background. Selected models are evaluated on our complete test set, the full Cui dataset, and its 15 volumes not used for training. + +A column-wise comparison of the results suggests that the volume of data has only a marginal impact on overall tooth segmentation performance but is critical for demographic generalizability. Additionally, models trained on our dataset generalize more effectively to the Cui dataset, while those trained exclusively on Cui experience a dice score reduction of up to 10 points when evaluated on our dataset compared to their own test set. Notably, UMamba demonstrates consistently strong performance across both datasets, highlighting its robust generalization capabilities. + +![](images/045e0734483b549b5f7392b2b4aaa5c21dd439bfd9c032bf7ddf3f65731ffb99.jpg) +(a) Ground-Truth + +![](images/3862bf74f08fc7e82fc722f2d86e53ef8a93b80759e8118ef6012bb7eb6b34f6.jpg) +(b) nnU-Net ResEnc + +![](images/395b9e71c268f0bb7ec736d799e3cd7b96a7cb3ec5a2eeeda961622e81a3f878.jpg) +(c) nnFormer +Figure 5. Qualitative results obtained with the proposed plan w/o post-processing, one patient per line. Red squares identify relevant errors. + +![](images/d03e3f9ff58e7f80876a654df3f95883029dca087e7d190d1789eccb73d1494a.jpg) +(d) UNETR++ + +![](images/f97b8b7e869583981ee87cb5264e8f0c9ae0264e2dd5dcef93dcc6d0c90fc4ae.jpg) +(e) UMamba + +![](images/5c3c739ce76bf7010a486f51f98885faa4330199cfe6c3fd3130e23080e2583a.jpg) +(f) VMamba + +![](images/3453eaa049d28da8dd11b1be0789002126466613f2d0eca2dca42c0fb5763bd5.jpg) +(g) Swin-UMamba + +Table 5. Dice comparison between our and Cui [16] datasets considering only teeth classes and converting others to background. Our plan w/o post-processing is employed. + +
Training DatasetModelTesting Dataset
Our (test)Cui (all)Cui (15 scans)
Our (all the 480 training scans)nnU-Net [24]91.1281.8682.13
nnFormer [56]85.1679.1980.25
UMamba [38]90.9787.6087.90
Our (82 scans from training)nnU-Net [24]90.0479.6781.12
nnFormer [56]82.7274.8577.42
UMamba [38]90.1886.2387.71
Cui (82 scans over 97 available)nnU-Net [24]78.0588.52
nnFormer [56]70.63~74.35
UMamba [38]79.9589.74
+ +# 4.7. Qualitative evaluation + +Fig. 5 depicts a qualitative comparison of the selected state-of-the-art segmentation models trained on the proposed dataset when using the previously introduced personalized nnU-Net planning. The lack of contextualized 3D receptive fields of the 2D VMamba is evident, especially on elongated objects like the inferior alveolar canal (first line). Transformer-based architectures, UNETR++ and nnFormer, produce more noisy labels. While some can be easily removed by employing post-processing techniques, the majority will remain. Interestingly, all of the 3D models mistakenly label the plastic chin support used to acquire the CBCT with a mix of artificial crown (yellow label, second line) and pharynx (gray label, second line). This has a negative impact on the HD95 of both "other" and pharynx classes. The local discriminative capabilities of 2D models make them stronger, avoiding this misidentification and explaining the lower HD95 values. + +All of the models except UNETR++ correctly recognize the first upper premolar implant (red label within the upper jawbone, second line). However, none of them label the artificial crown of the same tooth as such. + +# 5. Conclusions and Limitations + +In this paper, we presented the largest publicly available annotated CBCT dataset with 3D segmentations for 42 dis + +tinct maxillofacial anatomical structures, offering a valuable resource for developing and benchmarking automated segmentation models in digital dentistry and addressing a significant gap in the literature. Our dataset supports detailed multi-class segmentation across complex maxillofacial anatomy, enabling improved model training, evaluation, and generalization for clinical applications such as surgical planning and implant placement. Extensive benchmark evaluations of state-of-the-art neural network architectures on our dataset highlight the potential of existing models in enhancing the accuracy and efficiency of maxillofacial image analysis, also demonstrating the significance of the proposed dataset. Moreover, we demonstrated different techniques that can be adopted w.r.t. the standard nnU-Net planning to improve the overall performance of state-of-the-art models when trained on complex CBCT datasets such as the one proposed in this paper. + +Limitations and future work. Despite its significant contributions, our dataset has limitations that should be further addressed in the future. Although training and testing data come from different machines and acquisition centers, limiting potential biases in the evaluation, training data was acquired from a single center, potentially affecting the generalizability of models trained on this dataset to other clinical environments or patient demographics. Furthermore, while the dataset includes a wide range of classes, some—implants, crowns, and bridges—are underrepresented, affecting the robustness of model performance. Additionally, some components encountered in maxillofacial imaging, such as dental braces and bone plates, are missing due to their unavailability in the current patient cohort. Future work should focus on expanding this dataset to include multi-center data, ensuring a more representative sample of anatomical and demographic diversity. Including more cases with underrepresented and missing classes would further enhance the dataset's utility, supporting more comprehensive models. We believe that this dataset provides a strong foundation for advancing automated maxillofacial segmentation, and we encourage the research community to build upon it to address the outlined limitations. + +# Acknowledgments + +This work was supported by the University of Modena and Reggio Emilia and Fondazione di Modena, through the FAR 2024 and FARD-2024 funds (Fondo di Ateneo per la Ricerca 2024) and by the Italian Ministry of Research under the complementary actions to the NRRP "Fit4MedRob - Fit for Medical Robotics" Grant (#PNC0000007). + +# References + +[1] Fatemeh Abdolali, Reza Aghaeizadeh Zoroofi, Maryam Abdolali, Futoshi Yokota, Yoshito Otake, and Yoshinobu Sato. Automatic segmentation of mandibular canal in cone beam CT images using conditional statistical shape model and fast marching. International Journal of Computer Assisted Radiology and Surgery, 12:581-593, 2017. 3 +[2] Khalid Ayidh Alqahtani, Reinhilde Jacobs, Andreas Smolders, Adriaan Van Gerven, Holger Willems, Sohaib Shujaat, and Eman Shaheen. Deep convolutional neural network-based automated segmentation and classification of teeth with orthodontic brackets on cone-beam computed tomographic images: a validation study. European Journal of Orthodontics, 45(2):169-174, 2023. 5 +[3] Achraf Ben-Hamadou, Ouussama Smaoui, Ahmed Rekik, Sergi Pujades, Edmond Boyer, Hoyeon Lim, Minchang Kim, Minkyung Lee, Minyoung Chung, Yeong-Gil Shin, et al. 3DTeethSeg'22: 3D Teeth Scan Segmentation and Labeling Challenge. arXiv preprint arXiv:2305.18277, 2023. 2 +[4] Jonathan Blacher, Scott Van DaHuvel, Vijay Parashar, and John C Mitchell. Variation in Location of the Mandibular Foramen/Inferior Alveolar Nerve Complex Given Anatomic Landmarks Using Cone-beam Computed Tomographic Scans. Journal of Endodontics, 42(3):393-396, 2016. 3 +[5] Federico Bolelli, Stefano Allegretti, Luca Lumetti, and Costantino Grana. A State-of-the-Art Review with Code about Connected Components Labeling on GPUs. IEEE Transactions on Parallel and Distributed Systems, 2024. 7 +[6] Federico Bolelli, Luca Lumetti, Shankeeth Vinayahalingam, Mattia Di Bartolomeo, Arrigo Pellacani, Kevin Marchesini, Niels Van Nistelrooij, Pieter Van Lierop, Tong Xi, Yusheng Liu, et al. Segmenting the Inferior Alveolar Canal in CBCTs Volumes: the ToothFairy Challenge. IEEE Transactions on Medical Imaging, 2024. 2 +[7] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 2, 6, 7 +[8] Xiaokang Chen, Nan Ma, Tongkai Xu, and Cheng Xu. Deep learning-based tooth segmentation methods in medical imaging: A review. Proceedings of the Institution of Mechanical Engineers, Part H: Journal of Engineering in Medicine, 238 (2):115-131, 2024. 3 +[9] Zeyu Chen, Senyang Chen, and Fengjun Hu. CTA-UNet: CNN-transformer architecture UNet for dental CBCT images segmentation. Physics in Medicine & Biology, 68(17): 175042, 2023. 5 + +[10] So-Young Chun, Yun-Hui Kang, Su Yang, Se-Ryong Kang, Sang-Jeong Lee, Jun-Min Kim, Jo-Eun Kim, Kyung-Hoe Huh, Sam-Sun Lee, Min-Suk Heo, et al. Automatic classification of 3D positional relationship between mandibular third molar and inferior alveolar canal using a distance-aware network. BMC Oral Health, 23(1):794, 2023. 2 +[11] Marco Cipriano, Stefano Allegretti, Federico Bolelli, Mattia Di Bartolomeo, Federico Pollastri, Arrigo Pellacani, Paolo Minafra, Alexandre Anesi, and Costantino Grana. Deep Segmentation of the Mandibular Canal: a New 3D Annotated Dataset of CBCT Volumes. IEEE Access, 10:11500-11510, 2022. 2, 3 +[12] Marco Cipriano, Stefano Allegretti, Federico Bolelli, Federico Pollastri, and Costantino Grana. Improving Segmentation of the Inferior Alveolar Nerve through Deep Label Propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21137-21146. IEEE, 2022. 2, 3, 4 +[13] Weiwei Cui, Yaqi Wang, Qianni Zhang, Huiyu Zhou, Dan Song, Xingyong Zuo, Gangyong Jia, and Liaoyuan Zeng. CTooth: A Fully Annotated 3D Dataset and Benchmark for Tooth Volume Segmentation on Cone Beam Computed Tomography Images. In International Conference on Intelligent Robotics and Applications, pages 191-200. Springer, 2022. 2 +[14] Zhiming Cui, Changjian Li, and Wenping Wang. ToothNet: Automatic Tooth Instance Segmentation and Identification From Cone Beam CT Images. In Computer Vision and Pattern Recognition, pages 6368-6377, 2019. 3 +[15] Zhiming Cui, Bojun Zhang, Chunfeng Lian, Changjian Li, Lei Yang, Wenping Wang, and Min Zhu. Hierarchical Morphology-Guided Tooth Instance Segmentation from CBCT Images. In Information Processing in Medical Imaging, pages 150-162, 2021. 3, 5 +[16] Zhiming Cui, Yu Fang, Lanzhuju Mei, Bojun Zhang, Bo Yu, Jiameng Liu, Caiwen Jiang, Yuhang Sun, Lei Ma, Jiawei Huang, et al. A fully automatic AI system for tooth and alveolar bone segmentation from cone-beam CT images. Nature Communications, 13(1):2096, 2022. 1, 2, 4, 6, 7, 8 +[17] Wenhan Dou, Shanshan Gao, Deqian Mao, Honghao Dai, Chenhao Zhang, and Yuanfeng Zhou. Tooth instance segmentation based on capturing dependencies and receptive field adjustment in cone beam computed tomography. Computer Animation and Virtual Worlds, 33(5):e2100, 2022. 2, 3 +[18] Wei Duan, Yufei Chen, Qi Zhang, Xiang Lin, and Xiaoyu Yang. Refined tooth and pulp segmentation using U-Net in CBCT image. Dentomaxillofacial Radiology, 50(6): 20200251, 2021. 3 +[19] Herbert Edelsbrunner, David Kirkpatrick, and Raimund Seidel. On the shape of a set of points in the plane. IEEE Transactions on Information Theory, 29(4):551-559, 1983. 5 +[20] Kang Hsu, Da-Yo Yuh, Sc Lin, Pin-Sian Lyu, Guan-Xin Pan, Yi-Chun Zhuang, Chia-Ching Chang, Hsu-Hsia Peng, Tung-Yang Lee, Cheng-Hsuan Juan, Cheng-En Juan, Yi-Jui Liu, and Chun Jung Juan. Improving performance of deep learning models using 3.5D U-Net via majority voting for tooth + +segmentation on cone beam computed tomography. Scientific Reports, 12(1):19809, 2022. 3 +[21] Fengjun Hu, Zeyu Chen, and Fan Wu. A novel difficult-to-segment samples focusing network for oral CBCT image segmentation. Scientific Reports, 14:5068, 2024. 3 +[22] Kuofeng Hung, Andy Wai Kan Yeung, Ray Tanaka, and Michael M Bornstein. Current Applications, Opportunities, and Limitations of AI for 3D Imaging in Dental Research and Practice. International Journal of Environmental Research and Public Health, 17(12):4424, 2020. 4 +[23] Jae-Joon Hwang, Yun-Hoa Jung, Bong-Hae Cho, and MinSuk Heo. An overview of deep learning in the field of dentistry. Imaging Science in Dentistry, 49(1):1-7, 2019. 3 +[24] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 2, 6, 7, 8 +[25] ISO. Dentistry - designation system for teeth and areas of the oral cavity. https://www.iso.org/standard/68292.html, 2016. Accessed: 2024-11-11. 3 +[26] Tae Jun Jang, Kang Cheol Kim, Hyun Cheol Cho, and Jin Keun Seo. A Fully Automated Method for 3D Individual Tooth Identification and Segmentation in dental CBCT. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6562-6568, 2021. 2 +[27] Jorma Järnstedt, Jaakko Sahlsten, Joel Jaskari, Kimmo Kaski, Helena Mehtonen, Ari Hietanen, Osku Sundqvist, Vesa Varjonen, Vesa Mattila, Sangsom Prapayasatok, et al. Reproducibility analysis of automated deep learning based localisation of mandibular canals on a temporal CBCT dataset. Scientific Reports, 13(1):14159, 2023. 3 +[28] Joel Jaskari, Jaakko Sahlsten, Jorma Järnstedt, Helena Mehtonen, Kalle Karhu, Osku Sundqvist, Ari Hietanen, Vesa Varjonen, Vesa Mattila, and Kimmo Kaski. Deep Learning Method for Mandibular Canal Segmentation in Dental Cone Beam Computed Tomography Volumes. Scientific Reports, 10(1):5842, 2020. 2, 3 +[29] Touko Kaasalainen, Marja Ekholm, Teemu Siiskonen, and Mika Kortesniemi. Dental cone beam CT: An updated review. Physica Medica, 88:193-217, 2021. 1 +[30] Dagmar Kainmueller, Hans Lamecker, Heiko Seim, Max Zinser, and Stefan Zachow. Automatic Extraction of Mandibular Nerve and Bone from Cone-Beam CT Data. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2009: 12th International Conference, London, UK, September 20-24, 2009, Proceedings, Part II 12, pages 76-83. Springer, 2009. 3 +[31] Dirk-Jan Kroon. Segmentation of the Mandibular Canal in Cone-Beam CT Data. PhD thesis, University of Twente, Netherlands, 2011. 10.3990/1.9789036532808. 3 +[32] Pierre Lahoud, Mostafa EzEldeen, Thomas Beznik, Holger Willems, André Leite, Adriaan Van Gerven, and Reinhilde Jacobs. Artificial Intelligence for Fast and Accurate 3-Dimensional Tooth Segmentation on Cone-beam Computed Tomography. Journal of Endodontics, 47(5):827-835, 2021. 3 +[33] Pierre Lahoud, Siebe Diels, Liselot Nicales, Stijn Van Aelst, Holger Willems, Adriaan Van Gerven, Marc Quirynen, and + +Reinhilde Jacobs. Development and validation of a novel artificial intelligence driven tool for accurate mandibular canal segmentation on CBCT. Journal of dentistry, 116:103891, 2022. 2, 3 +[34] Jiarun Liu, Hao Yang, Hong-Yu Zhou, Yan Xi, Lequan Yu, Cheng Li, Yong Liang, Guangming Shi, Yizhou Yu, Shaoting Zhang, et al. Swin-UMamba: Mamba-based UNet with ImageNet-based pretraining. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 615–625. Springer, 2024. 2, 6, 7 +[35] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual State Space Model. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 6, 7 +[36] Luca Lumetti, Vittorio Pipoli, Federico Bolelli, Elisa Ficarra, and Costantino Grana. Enhancing Patch-Based Learning for the Segmentation of the Mandibular Canal. IEEE Access, pages 1-12, 2024. 3 +[37] Jinxuan Lv, Lang Zhang, Jiajie Xu, Wang Li, Gen Li, and Hengyu Zhou. Automatic segmentation of mandibular canal using transformer based neural networks. Frontiers in Bioengineering and Biotechnology, 11, 2023. 3 +[38] Jun Ma, Feifei Li, and Bo Wang. U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmentation. arXiv preprint arXiv:2401.04722, 2024. 2, 6, 7, 8 +[39] Lena Maier-Hein, Annika Reinke, Patrick Godau, Minu D Tizabi, Florian Buettner, Evangelia Christodoulou, Ben Glocker, Fabian Isensee, Jens Kleesiek, Michal Kozubek, et al. Metrics reloaded: Recommendations for image analysis validation. Nature Methods, pages 1-18, 2024. 5 +[40] Behnam Moris, Luc J. M. Claesen, Yi Sun, and Constantinus Politis. Automated tracking of the mandibular canal in CBCT images using matching and multiple hypotheses methods. 2012 Fourth International Conference on Communications and Electronics (ICCE), pages 327-332, 2012. 3 +[41] Archie Morrison, Marco Chiarot, and Stuart Kirby. Mental Nerve Function After Inferior Alveolar Nerve Transposition for Placement of Dental Implants. Journal-Canadian Dental Association, 68(1):46-50, 2002. 1 +[42] Yunbo Rao, Yilin Wang, Fanman Meng, Jiansu Pu, Jihong Sun, and Qifei Wang. A Symmetric Fully Convolutional Residual Network With DCRF for Accurate Tooth Segmentation. IEEE Access, 8:92028-92038, 2020. 3 +[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional Networks for Biomedical Image Segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3 +[44] A. Schramm, M. Rucker, N. Sakkas, R. Schön, J. Düker, and N.-C. Gellrich. The use of cone beam CT in craniomaxillofacial surgery. International Congress Series, 1281: 1200-1204, 2005. 3 +[45] Namrata Sengupta, Sachin C Sarode, Gargi S Sarode, and Urmi Ghone. Scarcity of publicly available oral cancer image datasets for machine learning research. Oral Oncology, 126:105737, 2022. 2 + +[46] Eman Shaheen, André Leite, Khalid Ayid Alqahtani, Andreas Smolders, Adriaan Van Gerven, Holger Willems, and Reinhilde Jacobs. A novel deep learning system for multiclass tooth segmentation and classification on cone beam computed tomography. a validation study. Journal of Dentistry, 115:103865, 2021. 3 +[47] Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. UNETR++: Delving Into Efficient and Accurate 3D Medical Image Segmentation. IEEE Transactions on Medical Imaging, 43(9):3377-3390, 2024. 2, 6, 7 +[48] Muhammad Usman, Azka Rehman, Amal Muhammad Saleem, Rabeea Jawaid, Shi-Sub Byon, Sung-Hyun Kim, Byoung-Dai Lee, Min-Suk Heo, and Yeong-Gil Shin. Dual-Stage Deeply Supervised Attention-Based Convolutional Neural Networks for Mandibular Canal Segmentation in CBCT Scans. Sensors, 22(24):9877, 2022. 2, 3 +[49] Shankeeth Vinayahalingam, Steven Kempers, Julian Schoep, Tzu-Ming Harry Hsu, David Anssari Moin, Bram van Ginneken, Tabea Flugge, Marcel Hanisch, and Tong Xi. Intraoral scan segmentation using deep learning. BMC Oral Health, 23(1):643, 2023. 2 +[50] Yiwei Wang, Wenjun Xia, Zhennan Yan, Liang Zhao, Xiaohe Bian, Chang Liu, Zhengnan Qi, Shaoting Zhang, and Zisheng Tang. Root canal treatment planning by automatic tooth and root canal segmentation in dental CBCT with deep multi-task feature learning. Medical Image Analysis, 85: 102750, 2023. 3 + +[51] Xueqiong Wei and Yuanjun Wang. Inferior alveolar canal segmentation based on cone-beam computed tomography. Medical Physics, 48(11):7074-7088, 2021. 3 +[52] Mark D Wilkinson, Michel Dumontier, IJsbrand Jan Aalbersberg, Gabrielle Appleton, Myles Axton, Arie Baak, Niklas Blomberg, Jan-Willem Boiten, Luiz Bonino da Silva Santos, Philip E Bourne, et al. The FAIR Guiding Principles for scientific data management and stewardship. Scientific data, 3(1):1-9, 2016. 2 +[53] Philip Worthington. Injury of the Inferior Alveolar Nerve during Implant Placement: a Literature Review. International Journal of Oral & Maxillofacial Implants, 19(5), 2004. 1 +[54] Xiyi Wu, Huai Chen, Yijie Huang, Huayan Guo, Tiantian Qiu, and Lisheng Wang. Center-Sensitive and Boundary-Aware Tooth Instance Segmentation and Classification from Cone-Beam CT. In International Symposium on Biomedical Imaging, pages 939–942, 2020. 3 +[55] Huanmiao Zhao, Junhua Chen, Zhaoqiang Yun, Qianjin Feng, Liming Zhong, and Wei Yang. Whole mandibular canal segmentation using transformed dental CBCT volume in Frenet frame. Heliyon, 9(7), 2023. 3 +[56] Hong-Yu Zhou, Jiansen Guo, Yinghao Zhang, Xiaoguang Han, Lequan Yu, Liansheng Wang, and Yizhou Yu. nnFormer: Volumetric Medical Image Segmentation via a 3D Transformer. IEEE Transactions on Image Processing, 2023. 6, 7, 8 \ No newline at end of file diff --git a/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/images.zip b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44fc0d0e6580b80fcc3b60a5e8a73691183dd7d9 --- /dev/null +++ b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13efa9c9bfb687d1dd4a393c258829781a3b1cc79a61f3cc95a5b1f72997cbc6 +size 559704 diff --git a/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/layout.json b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8a6ba6234f9cef6e64f603e7cf7f298dc7ee201b --- /dev/null +++ b/CVPR/2025/Segmenting Maxillofacial Structures in CBCT Volumes/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44f292e3537440b1f4c118c74522afe9dedc5a21cf2fb3bebb30d4710e0e6551 +size 422326 diff --git a/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/83e0a95c-7140-478a-a40d-1c8962e615cc_content_list.json b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/83e0a95c-7140-478a-a40d-1c8962e615cc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8cf8db6b03f3f006f364a4ced4e4d090bdc6bef5 --- /dev/null +++ b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/83e0a95c-7140-478a-a40d-1c8962e615cc_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04a1570d3dc1e596bf14ec5437a74f5aa27a53c13a794fce193a89201a12acfe +size 79902 diff --git a/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/83e0a95c-7140-478a-a40d-1c8962e615cc_model.json b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/83e0a95c-7140-478a-a40d-1c8962e615cc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aa4a9823759471a64ed890f8b2042589852732cd --- /dev/null +++ b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/83e0a95c-7140-478a-a40d-1c8962e615cc_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:082b7caede981fdf29ca37dd2fc7463048bf0508102aa096cd3f6e6e3e4df70d +size 101875 diff --git a/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/83e0a95c-7140-478a-a40d-1c8962e615cc_origin.pdf b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/83e0a95c-7140-478a-a40d-1c8962e615cc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b7172e647f5efe123cfc2d8fbb4263f391e8c2fa --- /dev/null +++ b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/83e0a95c-7140-478a-a40d-1c8962e615cc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f61edcecb70fbbc7a6265985b138df6aa65c14ce754f9598151c728c0e6adb25 +size 6103549 diff --git a/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/full.md b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6bddb4ffa7bb9f42e9982cab0a8ed5fd3f4e0ddc --- /dev/null +++ b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/full.md @@ -0,0 +1,371 @@ +# Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects + +Weimin Qiu Jieke Wang Meng Tang University of California Merced + +{wqiu5, jwang450, mtang4}@ucmerced.edu + +![](images/4d10c9824029f812d281c23d95f083b9c589b7a077a76af00ec4e5ee4f45097b.jpg) + +![](images/0c47bb4eaab500027209659b38ec606135bfabde03c0370d2a0744ddb9e4cd8c.jpg) +a rose and a carnation + +![](images/84f502694b48071d8632310b0ffd7c06e5ca308a0f3c08e3315c139731c36632.jpg) + +![](images/9023e3d6b1a0511b2e09b2cdbb3468682e1a61bb8628314a4ffd2bb125f9f2d3.jpg) + +![](images/eab78098becc94e6ffff88d9ec29e681a0d909367a74cfa7504a41877effac6e.jpg) +SD1.4 [41] +Figure 1. Our Self-Cross guidance addresses subject mixing in particular for similar subjects. Our training-free method can boost the performance of any Unet-based or transformer-based diffusion models such as Stable Diffusion 1, 2, and 3 (shown in Fig. 5 and 6.) + +![](images/603503a705b6217ae6768ee60e0c75715a4b22cfc4045ea6035807539912941a.jpg) +a beagle and a collie +INITNO [17] + +![](images/74aa5cf8beeb6934a4a0db0451724ff7ab2940eb3b85673f2bafdd73924f9405.jpg) +CONFORM [31] + +![](images/0595861c136cf5b6302a028ba08aceb8e65659dd81a3287baeb1728603b254e9.jpg) +Self-Cross (Ours) + +![](images/37c3f5d914fa17b8920b23504fcc67f67be5a91c7112020c06475f8ac142d761.jpg) + +![](images/7bfdbc136821a398ca67831f02f4ba410c6762cb8875b87bc7319a115dfe3873.jpg) +a leopard and a tiger + +![](images/53e880c49fc6b9a80fba4c6d1479f35735cbb532dada7482a2789db9208e1259.jpg) + +![](images/e088652ccdc29ba13eabc95a63c94c340cda3c7004dc3298391dfc2b75644fd2.jpg) + +![](images/7c170631c49f155bffad3c05c1c9c0082472e42c2b7924932395b260acb72d19.jpg) +SD1.4 [41] + +![](images/546c1e8e8d185c54b92e2696c076e4ac693959e1f60144938f1dbe1a2fd531f3.jpg) +a eagle and an owl +INITNO [17] + +![](images/55c2f36649b05ebdee0b708e7b4acc0cd554e7a70b72c91bf73bf367e1441360.jpg) +CONFORM [31] + +![](images/13570dffc092142d9fe0a2dc086b2141d3097150794e757f71bf47f1d6c81831.jpg) +Self-Cross (Ours) + +# Abstract + +Diffusion models achieved unprecedented fidelity and diversity for synthesizing image, video, 3D assets, etc. However, subject mixing is an unresolved issue for diffusion-based image synthesis, particularly for synthesizing multiple similar-looking subjects. We propose Self-Cross Diffusion Guidance to penalize the overlap between cross-attention maps and the aggregated self-attention map. Compared to previous methods based on self-attention or cross-attention alone, our guidance is more effective in eliminating subject mixing. What's more, our guidance addresses subject mixing for all relevant patches beyond the most discriminant one, e.g., the beak of a bird. For each subject, we aggregate self-attention maps of patches with higher cross-attention values. Thus, the aggregated self-attention map forms a region that the whole subject attends to. Our training-free method boosts the performance of both Unet-based and Transformer-based diffusion models such as the Stable Diffusion series. We also release a similar subjects dataset (SSD), a challenging benchmark, and utilize GPT-4o for automatic and reliable evaluation. Extensive qualitative and quantitative results demonstrate the effectiveness of our self-cross diffusion guidance. + +# 1. Introduction + +Diffusion-based generative models have made significant progress in recent years in synthesizing high-quality images [41], videos [5], 3D assets [37], etc. With a simple text prompt, diffusion-based generative models such as Stable Diffusion [41] can create highly photorealistic or artistic images of various styles and subjects. Such progress revolutionized many applications including the content creation. However, text-to-image diffusion models still have many issues such as subject neglect, subject mixing, and attribute binding. We are particularly interested in solving subject mixing in this work. Many mitigations [8, 17, 31] were proposed to enhance the faithfulness of text-to-image synthesis but the issue of subject mixing remains, see failure cases of previous methods [17, 31] in Fig. 1. This is more prominent when synthesizing multiple similar-looking subjects, e.g., a photo of a leopard and a tiger in Fig. 1. + +Since self-attention maps reflect the similarity of patches and cross-attention maps reflect the subjects. Intuitively, patches from a subject should not show large similarity to patches from other subjects at least for some timesteps and layers to avoid copying patches from other subjects. Empirically, Fig. 2 shows a failure case for Stable Diffusion [41] with subject mixing and an obvious overlap between cross + +![](images/8ed6ba9bda46081c342ae6bc23a94f4eff8852641af45fe326db95e4d2284a2e.jpg) +Figure 2. Results of Stable Diffusion [41] and our method with Self-Cross guidance for the same prompt "a bear and an elephant". Images are generated from the same random seed. "cross" means cross-attention map and "self" means the aggregated self-attention map. The overlap between self-attention and cross-attention leads to subject mixing, while Self-Cross guidance reduces overlapping. + +attention maps w.r.t. a subject token and self-attention maps w.r.t. patches of another subject. Our insight is that a subject should not attend to other subjects by its self-attention maps. As shown in Fig. 2, while overlapping between self-attention and cross-attention maps results in a failure case, our Self-Cross diffusion guidance penalizes such overlaps yielding synthetic images without subject mixing. + +Our method is different from previous methods based on cross-attention or self-attention maps alone [3, 17, 31]. We are the first to explore regularization between self-attention and cross-attention maps. What's more, we formulate Self-Cross diffusion guidance for self-attention maps of multiple image patches beyond the most discriminant one. It is well-known that neural networks for perceptron such as image classification tend to focus on the discriminant region of an image, e.g., the beak of a bird. However, apart from the most discriminant patch, other patches with more details are also particularly important for synthesizing multiple similar-looking subjects. We first adaptively identify all image patches corresponding to a subject and then avoid their attendance to other subjects by self-attention maps. As a training-free method, it outperforms other single-patch-based approaches [8, 17] qualitatively and quantitatively. + +Previous datasets with prompts for image synthesis are not challenging in terms of subject mixing. What's worse, the commonly used CLIP score doesn't correlate well with + +human judgment. Therefore, to facilitate image synthesis of similar objects, we release a similar-subject dataset (SSD) consisting of text prompts with similar subjects. To enable benchmarking at scale, we leverage state-of-the-art vision language models such as GPT-4o to evaluate synthetic images of different methods by visual question answering [22]. + +The main contributions of this paper are as follows. + +- We propose Self-Cross diffusion guidance for text-to-image synthesis of similar-looking subjects, which effectively addresses subject mixing as shown in our qualitative and quantitative results. +- Our guidance is training-free and can improve the performance of pre-trained models such as Stable Diffusion. +- We propose similar-subject dataset (SSD), a new benchmark for image synthesis of similar subjects. Our benchmark includes prompts for two or three similar subjects and an effective metric using VLM [22]. +- As a side effect evidenced by improved existence and recognizability scores, our method reduced subject neglect. + +# 2. Related Work + +Text-to-Image Diffusion Models Given text prompts, text-to-image synthesis aims to generate visually coherent images. Early approaches utilized GANs [16, 25, 43, 44, 48, 54, 59] and autoregressive models [7, 10, 11, 14, 27, 38, 55, 56]. Recently, with ground breaking advancements in diffusion models [9, 20, 21, 30, 33, 46, 47], the focus of text-to-image synthesis has shifted toward diffusion models [4, 13, 36, 39, 41, 43, 58]. Although diffusion models can generate photorealistic images, ensuring faithful adherence to the provided text prompt remains a significant challenge. To tackle this issue, methods like ReCap [45], DALLE3 [4], and SD3 [13] leverage improved image-caption pairs during training or incorporate multiple language encoders to capture more expressive language representations. However, these methods require training models from scratch, which entails substantial computational costs and makes them inapplicable to popular models like Stable Diffusion [41] and Imagen [43]. Furthermore, these models only partially address issues such as subject neglect, subject missing, and attribute binding, leaving room for improvement towards prompt-faithful synthesis. + +Guidance for Consistent Text-to-Image Generations Training-free inference-time optimization is an active research area to improve the consistency of the pre-trained text-to-image models. These methods typically extract internal representations of the denoising networks and correct the denoising trajectory to improve the alignment to the given prompt. One way to modulate the denoising trajectory is to replace internal features, such as attention modules in PnP-Diffusion [50], FreeControl [32], Prompt-to + +Prompt [19], DenseDiffusion [26], and MasaCtrl [6]. While effective for image editing and style transfer, these methods fail to address critical issues like subject mixing and attribute misalignment. Another line of work optimizes latents by minimizing guidance loss, in a way similar to classifier guidance [9] and classifier-gree guidance [20]. To address the problems of subject neglect, subject mixing, and attribute binding, many approaches have been proposed. Ge et al. [15], Self-Guidance [12], and DisenDiff [57] design loss functions for image editing and controllable generation. Attention Refocusing [35], BoxDiff [53], TokenCompose [52], and Ge et al. [15] use pre-defined layouts, either from external models or users, as inference-time supervision. However, these methods rely on prior knowledge, which is sometimes unreachable in real world. Therefore, other knowledge-free methods have been developed. Following Attend&Excite [8] and A-STAR [2] steps, CONFORM [31] takes contrastive loss for subject separation and attribute-binding. Linguistic Binding [40] introduces a variant of Kullback-Leibler divergence as the loss for improved consistency. INITNO [17] leverages both cross-attention maps and self-attention maps to refine the initial noise. While they use cross-attention and self-attention separately, we emphasize that the interaction between self-attention and cross-attention is key to eliminating subject mixing and improving faithfulness. Moreover, these methods typically consider the most discriminant patch, which is insufficient for removing subject mixing. + +# 3. Preliminaries + +Diffusion Model Latent diffusion model [41] operates in a latent space instead of the pixel space, which largely reduces the computational complexity of image generation. An encoder and a decoder are trained to encode images and decode lower-dimensional latents respectively. Furthermore, cross-attention between prompts and image patches allows controllable image generation with various prompts, such as layouts, semantics, and texts. + +In the latent space, the forward process gradually adds Gaussian noises on the latent code $z_0$ over time until it completely deteriorates to Gaussian noise $z_T$ . While in the reverse denoising process, a denoising network [42] $\epsilon_{\theta}$ denoises the latent code $z_t$ iteratively until time step zero $z_0$ . The training objective is formally defined as: + +$$ +\mathcal {L} = \mathbb {E} _ {z _ {t}, \epsilon \sim N (0, I), c (y), t} | | \epsilon - \epsilon_ {\theta} (z _ {t}, c (y), t) | | ^ {2} \tag {1} +$$ + +where $c(y)$ represents the condition embedding, for example CLIP embedding for text prompt $y$ . + +In cross-attention modules, condition embeddings $c$ are projected to query $Q$ and values $V$ . Accordingly, intermediate representations from UNet are projected to keys $K$ . + +Therefore, the cross-attention maps can be described as: + +$$ +A ^ {c} = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) \tag {2} +$$ + +where $\sqrt{d}$ is a scaling factor [51]. Each text token has a cross-attention map with shape $R^{h / P\times w / P}$ for patch size $P\times P$ . The cross-attention maps reflect the attendance of text tokens to patches. On the other hand, the self-attention map for each patch indicates the relationship between different patches. We denote the cross attention map of a text token $k$ as $A_{k}^{c}\in R^{h / P\times w / P}$ and the self attention map of a patch $(x,y)$ as $A_{x,y}^{s}\in R^{h / P\times w / P}$ . + +Attention Based Guidance As discussed in Sec. 2, one methodology for consistent text-to-image generation is attention-based guidance [3, 8, 17, 31, 35] using self attention or cross attention. Here we briefly introduce the most relevant methods to our Self-Cross diffusion guidance. + +To avoid subject neglect, Attend-and-Excite [8] finds the patch with maximum cross attention for each token, and penalizes if the maximum cross attention is small. In other words, it encourages the appearance of specified subjects. Spefically, the cross-attention response score is defined as, + +$$ +S _ {\text {c r o s s - a t t n}} = \max _ {k \in K} S _ {\text {c r o s s - a t t n}, k} \tag {3} +$$ + +where $K$ indicates the set of subjects' tokens and + +$$ +S _ {\text {c r o s s - a t t n}, k} = 1 - \max \left(A _ {k} ^ {c}\right). \tag {4} +$$ + +Unlike Attend-and-Excite[8], our Self-Cross diffusion guidance loss is between self-attention and cross-attention. Besides, it's formulated for all relevant subject patches beyond the most discriminant one. The most similar guidance to ours is the self-attention conflict score introduced in INITNO [17]. However, INITNO [17] is still limited to the most discriminant patch, and our method outperformed INITNO [17] by a large margin shown in our experiments. While separation loss [3] penalizes overlap between cross-attention maps, it isn't a training-free method. Our proposed Self-Cross guidance is novel and complementary to existing works on attention-based guidance. + +Optimization of Guidance Loss The standard method for optimizing a guidance loss is through gradient descent. To better minimize a guidance loss, Iterative Latent Refinement [8] runs multiple steps of gradient descent until the guidance loss is bellow a threshold. To find a better initial noise, Initial noise optimization [17] (INITNO) optimizes over the mean and variance of initial noise until the initial guidance loss is satisfactory. We adopted the two techniques above for our Self-Cross diffusion guidance. + +# 4. Our Method + +We describe Self-Cross diffusion guidance given a pair of similar subjects, namely "a bear and an elephant" with "bear" at index $i = 2$ and "elephant" at index at $j = 5$ . Our method can be easily extended to multiple pairs. + +Aggregation of Self-Attention Maps Fig. 3 describes the aggregation of self-attention maps. Given the cross-attention map of "bear", we select patches with high responses and visualize their self-attention maps. The diversity of self-attention for different patches means the self-attention map for the most discriminant patch alone can't cover all the regions the subject attends to. Thus, the key to our Self-Cross diffusion guidance is aggregating self-attention maps. If limited to the most discriminant patch (with the highest cross-attention value), other foreground patches may lead to subject mixing, as shown in Fig. 7. + +Firstly, we apply simple and efficient Otsu's method [34] on the cross-attention maps, which automatically returns the threshold and masked patches with relatively higher cross-attention values. Secondly, we aggregate all the self-attention maps of masked patches to better represent the whole region that a subject attends to. For subject at $i$ with cross-attention map $A_{i}^{c}$ , aggregated self-attention map is: + +$$ +A _ {i} ^ {s} = \frac {\sum_ {x _ {m} , y _ {n}} \left(A _ {i} ^ {c} \left[ x _ {m} , y _ {n} \right] \times A _ {x _ {m} , y _ {n}} ^ {s}\right)}{\sum_ {x _ {m} , y _ {n}} A _ {i} ^ {c} \left[ x _ {m} , y _ {n} \right]} \tag {5} +$$ + +Where $[x_m, y_n]$ are the coordinates of the selected patches, $A_i^c[x_m, y_n]$ is the cross-attention value for token at $i$ of the patch at $[x_m, y_n]$ , and $A_i^s[x_m, y_n]$ is the self-attention map for the patch at $[x_m, y_n]$ . + +In summary, we select patches corresponding to a subject and then take a weighted sum of the self-attention + +![](images/18263b272f2ddcbd434e7fb9bf3548de66194efb716c582c7702cd708c149b0e.jpg) +Figure 3. Self-Cross diffusion guidance between the cross-attention of "elephant" and the self-attention of "bear" + +maps for these patches, where the weights are the patches' cross attention values. Likewise, we obtain aggregated self-attention maps $A_{i}^{s}$ for all subjects at given indexes. + +Self-Cross Diffusion Guidance The aggregated self-attention maps highlight regions that the subject attends to. Our core assumption is that a subject should not attend to other subjects in the image at some time-steps and layers through its self-attention maps. Hence, we propose to penalize the overlap between the aggregated self-attention map of one subject and cross-attention maps of other similar subjects. This has a different effect compared to a loss between cross-attention maps, as aggregated self-attention can be different from cross-attention, shown in Fig. 3. + +The example prompt in Fig. 3 has two pairs of attention maps, aggregated self-attention of "bear" & cross-attention of "elephant" and aggregated self-attention of "elephant" & cross-attention of "bear". The Self-Cross diffusion guidance between the subject at $i$ and the subject at $j$ is defined as their overlap $g(i,j)$ , + +$$ +\begin{array}{l} g (i, j) = \sum_ {x, y} \min \left(A _ {i} ^ {s} [ x, y ], A _ {j} ^ {c} [ x, y ]\right) \\ + \sum_ {x, y} ^ {\prime \prime} \min \left(A _ {i} ^ {c} [ x, y ], A _ {j} ^ {s} [ x, y ]\right) \tag {6} \\ \end{array} +$$ + +where $A_{i}^{s}$ is the self-attention map of subject at $i$ and $A_{j}^{c}$ is the cross-attention map of subject at $j$ . + +Now, let's consider more similar subjects. If there are $N$ similar subjects, then mathematically we would have $C_N^2$ pairs of subjects. In this case, our Self-Cross diffusion guidance is the average of the $C_N^2$ ones, as in Eq.7, + +$$ +S _ {\text {s e l f - c r o s s}} = \sum_ {i, j \in \Omega} ^ {i \neq j} \frac {g (i , j)}{C _ {N} ^ {2}}. \tag {7} +$$ + +where $\Omega$ is the set containing all subject indexes. Similar to INITNO [17], we also include the cross-attention response score in our total loss. + +$$ +\mathcal {L} _ {\text {t o t a l}} = S _ {\text {s e l f - c r o s s}} + \lambda \cdot S _ {\text {c r o s s - a t t n}}. \tag {8} +$$ + +Spatial Relationship Among Subjects We define Our novel self-cross diffusion guidance with much thought to penalize appearance overlap while not hindering generation capabilities, as shown quantitatively in Tab. 5, Tab. 2 and qualitatively in Fig. 6. To achieve this, + +(a) We define the guidance only for early timestamps in the reverse process. Because attention maps from early timestamps are known to be semantically meaningful. +(b) It is known that attention maps of intermediate layers in diffusion models are highly correlated to semantics. So we enforce our guidance for these selected layers. + +(c) Other tokens such as verbs and adjectives in a prompt can render subject relationships for image synthesis. + +These tokens are not involved in our guidance. + +See more implementation details in the appendix A. + +Overall Pipeline Alg. 1 shows the overall pipeline, which involves initial noise optimization for our total loss 8. We apply Self-Cross guidance to the first half of the reverse process, which is more relevant for the semantic structure of synthesized images. Similar to previous work [8, 17], iterative refinement is conducted to ensure losses are below specified thresholds. + +Algorithm 1: T2I with Self-Cross Guidance +Input: A text prompt $P$ and indices $\Omega$ of subjects +A pre-trained T2I diffusion model such as $SD(\cdot)$ +Max iterations: $\tau_{MaxAlterStep}$ , $\tau_{MaxIter}$ +Thresholds: $\tau_{cross-attn}$ , $\tau_{self-cross}$ +A set of iterations for refinement $\{t_1, t_2, \dots, t_k\}$ . +Output: Generated image $z^0$ . + +Noise Initialization Step: Noise pool $\mathcal{P}\gets \{\}$ t $\leftarrow T$ do INITNO (loss $\mathcal{L}_{total} = S_{\mathrm{self - cross}} + \lambda S_{\mathrm{cross - attn}})$ return noise pool $\mathcal{P}$ $z^t\gets \arg \min_{z^t\in \mathcal{P}}\mathcal{L}_{total}$ + +Reverse Process: +while t ≥ τMaxAlterStep do + ▷ Compute attention and losses +Scross-attn, Sself-cross, ← SD( $z^t$ , P, t) +if t ∈ {t1, t2, ..., tk} and +(Scross-attn > τcross-attn or +Sself-cross > τself-cross) then + ▷Iterative Refinement Steps +i = 0 +while (Scross-attn > τcross-attn or +Sself-cross > τself-cross) and i < τMaxIter +do + $z^t \gets SGD(\mathcal{L}_{total})$ + i = i + 1 +end +else + $z^t \gets SGD(\mathcal{L}_{total})$ +end + $\begin{array}{rl} & z^t \gets SGD(z^t, P, t) \\ & t = t - 1 \end{array}$ +end +▷ Continue the remaining steps without guidance +while t > 0 do + -, -, $z^t \gets SGD(z^t, P, t)$ + t = t - 1 +end +return $z^0$ + +# 5. Experiments + +Datasets and Baselines Similar to previous work [8, 17, 31] on consistent text-to-image generation, we first report results on three datasets including animal-animal, animal-object, and object-object prompts based on Stable Diffusion models. After visualization of experimental results, We found some prompts are visually distinct and not challenging enough. Hence, we introduce our new dataset Similar Subjects Dataset (SSD) containing 31 prompts with two subjects(SSD-2) and 21 prompts with three subjects(SSD-3). The former ones are designed for primitive baselines, e.g., Stable Diffusion 1, and the latter ones are intended for stronger baselines, e.g., Stable Diffusion 3. In the new dataset, subjects usually share similar structures with distinguishable details, e.g., different textures for leopards and tigers shown in Fig. 1. Qualitative and quantitative results show more subject mixing for each method with the new dataset than the original ones. Additionally, We also verify the spatial reasoning capacity of our method on 2D-spatial and 3D-spatial from T2ICompBench [23] [24]. These subsets emphasize spatial relationships between two subjects. + +We compare our method to the original Stable Diffusion [41] as well as other training-free methods including Initial Noise Optimization (INITNO) [17] and CONFORM [31]. Separate-and-Enhance [3] is relevant to our work regarding subject mixing, but it necessitates fine-tuning, making it incomparable to training-free approaches. + +# 5.1. Qualitative results + +Self-Cross v.s. baselines We provide the qualitative comparisons in Figure 4, Figure 5, and Figure 6 using SD1.4, SD2.1, and SD3-medium respectively. For each prompt, we used the same list of random seeds for all methods. Self-Cross diffusion guidance successfully addressed the issue of subject mixing in most cases. For example, in Figure 4, given the prompt "a bird and a rabbit", Stable diffusion [41], INITNO [17], and CONFORM [31] generated birds with rabbits'ears, while our method generated faithful images without subject mixing. Our method synthesized better images for extremely similar subjects too, e.g., "a hummingbird and a kingfisher". Hummingbirds are recognized for their iridescent plumage and long, slender beaks, whereas kingfishers typically feature bold, vibrant colors like blue and have shorter, stout beaks. Appendices E and G show more qualitative results, failure cases, and discussion. + +Self-attention and Cross-attention maps Fig. 3 presents a cross-attention map and multiple self-attention maps for top-responsive image patches, along with an aggregated + +![](images/49bd9d0bccefcd99ec69deef104d781a861bc812771c21007dcd39f1cdb9f532.jpg) +Figure 4. Qualitative comparisons of Self-Cross (ours) to SD1.4 [41], INITNO [17], CONFORM [31]. For each prompt in the left column, we sample four seeds and show the results of different methods. + +self-attention map highlighting the region the entire subject("bear") attend to. Note that the aggregated self-attention map is different from the original cross-attention map, which motivated Self-Cross diffusion guidance rather than using cross-attention maps alone. The next section explores different self-attention aggregation schemes and their impact on image synthesis. + +Self-Cross guidance with the increasing number of patches To see the effect of self-attention aggregation, we compare our Self-Cross diffusion guidance using different patches. Specifically, given a cross-attention map, we select + +(a) the patch with the maximum cross-attention value. +(b) the top 16 patches w.r.t. cross-attention value. +(c) masked patches which typically have more than 16 patches. The details of masking are in Sec. 4. + +Fig. 7 shows results with three example prompts, which clearly demonstrate less subject missing with more patches. If only the patch with the maximum cross-attention is used in Self-Cross guidance, the subjects are still mixing. The mixing is less severe but not eliminated with 16 patches, for example mixing in terms of hair, teeth, and feet. The best is when we use all masked patches in the otsu mask and define an aggregated self-attention map for our guidance, in which case the suture mixing is eliminated. + +# 5.2. Quantitative results + +Quantitative results include TIFA-GPT4o scores and text-text similarities. Our TIFA-GPT4o is a variant of the recently proposed TIFA metric [22], which is much more cor- + +
MethodsSSD-3
Question TypesExtRecw/o Mt-t sim
SD3 Medium33.5430.3170.0873.82
Self-Cross(Ours)57.9253.0877.1574.96
+ +Table 1. Quantitative results using SD3 for prompts with three subjects. + +
%Animal-AnimalAnimal ObjObj-ObjSSD-2
SD1.476.579.276.470.0
INITNO82.284.082.372.0
CONFORM81.984.682.070.7
Self-Cross(Ours)84.384.782.573.6
+ +Table 2. Average text-text similarities (↑) on different methods + +
%Animal-AnimalSSD-22D-Spatial3D-Spatial
SD2.181.6873.0178.7177.73
INITNO83.1573.20--
Self-Cross(Ours)84.0273.5380.480.0
CONFORM85.2173.87--
CONFORM+Ours85.8874.91--
+ +Table 3. Average text-text similarities (↑) on different methods + +related with human judgment compared to the widely used CLIP score. Our main quantitative results are TIFA-GPT40 scores, while CLIP scores should be interpreted with caution (see appendix D for details). + +TIFA-GPT4o Scores TIFA [22] aims to assess the faithfulness of the generated image to the input prompt. It uses a vision-language model to perform Visual Question Answering (VQA) on the generated image, guided by specific questions about the image's contents. In our implementation, we leverage a more advanced vision-language model, + +
MethodsAnimal-AnimalAnimal-ObjectObject-ObjectSSD-2
Question TypesExtRecw/o MExtRecw/o MExtRecw/o MExtRecw/o M
SD1.4 [41]39.5129.7072.2467.8453.7290.3634.1531.8994.2230.7728.0977.47
INITNO [17]89.3977.0982.2698.3778.0095.9796.2090.4295.1761.3455.5379.70
CONFORM [31]89.6378.0084.2298.3775.0997.1678.5873.1297.4867.5459.9079.50
Self-Cross(Ours)94.5587.7992.9499.6075.1798.3098.9593.1998.6577.6770.9286.45
+ +Table 4. TIFA-GPT4o Scores (↑) on four benchmarks: Animal-Animal, Animal-Object, Object-Object, and our proposed Similar Subjects Dataset. Inspired by TIFA [22], we employ GPT4o [1] as the VQA model to evaluate three aspects: Existence of both subjects (Ext), Recognizability of both subjects (Rec), and Absence of Mixing of subjects (w/o M). GPT4o is prompted with several True/False questions, and we report the percentage of True responses as the scores. The list of question prompts is given in the supplementary materials. Best results are highlighted in bold and second best results are shown with underline. + +
MethodsAnimal-AnimalSSD-22D-Spatial3D-Spatial
Question TypesExtRecw/o MExtRecw/o MExtRecRelw/o MExtRecRelw/o M
SD2.1 [41]61.6338.2367.5152.8436.98\( 84.93^2 \)76.1673.3732.4690.2169.166.1747.7290.33
INITINO [17]80.9647.1863.8068.7046.1073.85--------
Self-Cross(Ours)89.5355.0378.7977.3545.3677.0887.3783.9035.4891.2487.0782.6057.9691.27
CONFORM [31]96.8870.1992.4983.7153.9089.76--------
CONFORM+Ours97.8369.0094.6681.3955.9893.20--------
+ +Table 5. Quantitative benchmarks with SD2.1 TIFA-GPT4o Scores (↑) on two challenging benchmarks: Animal-Animal and our proposed Similar Subjects Dataset. Best results are highlighted in bold and second best results are shown with underline. vanilla SD2.1 usually generates only one subject of the prompt so its 'w/o M' score is high. + +![](images/aad8e74d64f55117b69ce053dac98a3fbb1aca924e829fe2cf6d98ec961e9933.jpg) +Figure 5. Quantitative comparisons between original SD2.1 [41] and our method. + +![](images/ed85a03726bb571e823ebb50ad4be8246064298cb0695dbeb75eddd41ca7ed16.jpg) +Figure 6. Quantitative comparisons between SD3-medium [41] and our method. + +![](images/c960c7a7da9a042125571fc41f68206916f6c27c429ad148b535bae2bc3dc48c.jpg) +Figure 7. Our Self-Cross guidance works the best with masked patches, which verifies our assumption that all patches of a subject not just the most discriminant one need to be considered for eliminating subject mixing. + +![](images/3406ef9b5e337ba40622699edecca8b57fa57695c388c357a5ae3a9917e79e4c.jpg) + +![](images/7b1ab45811eec9064bf6dc474ec9534f1cd8be5b36aaec0aa38f87c1c490797b.jpg) + +GPT4o [1], to provide more precise feedback. Given a text prompt in the form of "a class A and a class B", we developed questions to (1) verify the existence of both subjects; (2) confirm the recognizability of both subjects, ensuring no artifacts or distortions are present; and (3) check that the image does not exhibit a mixture of the two subjects, i.e., subject mixing. These three aspects capture both local and global information about the image, serving as comprehensive and reliable metrics including Ext, Rec, w/o M, and Rel scores respectively. W/o score is most relevant to our study of subject mixing, while Ext and Rec scores are effective in assessing the issue of subject neglect. Rel measures the spatial relationship of subjects. The full set of question prompts we designed are in the appendix C. + +As shown in Table 4, when built on Stable Diffusion 1.4, our method is overwhelmingly better than baselines including SD [41], INITNO [17], and CONFORM [31] on Animal-Animal and Similar Subjects datasets(SSD-1), especially for reducing subject mixing (w/o M score). For animal-animal prompts, we achieved a w/o M score of $92.94\%$ , which is $8.7\%$ better than the second-best score $(84.22\%)$ CONFORM). For similar subjects dataset(SSD-1), our w/o M score of $86.45\%$ is $6.7\%$ better than the second-best score $(70.70\%)$ INITNO). Besides, for less challenging datasets with distinct subjects including Animal-Object and Object-Object, we still achieved better results. + +Interestingly, although our method focuses on addressing subjectmixing, it can also reduce subjectneglect and improve the recognizability/fidelity of generated subjects to some extent. However, visualization of the attention maps shows the inefficiency of attend&excite in Stable Diffusion 2.1. As we rely on attend&excite to encourage the existence of different subjects, our performance is influ + +enced in Stable Diffusion 2.1. So, for Stable Diffusion 2.1, we apply CONFORM for the first few steps to ensure the existence of subjects and apply our method in later steps(indicated as CONFORM+ours), as shown in Tab. 5. To further prove the effectiveness of our method on DiT, we also implement our method on Stable Diffusion 3 medium as shown quantitatively in Tab. 1 and qualitatively in Fig. 6. + +Text-text similarities Text-text similarities, also known as BLIP scores, are the similarity between captions generated by a vision-language foundation model [28] and the original prompts used to synthesize the images. This metric captures subjects and attributes from the original prompt, then measures the coherency and consistency of the generated content with the textual descriptions. As shown in Table 2, Table 3, and Table 1, compared with other methods, our approach, including the combined one, reaches SOTA performances on all benchmarks. + +# 6. Conclusion and Future Work + +Subject mixing remains a persistent issue for diffusion-based image synthesis, particularly for similar-looking subjects. We propose Self-Cross diffusion guidance to boost the performance of any diffusion-based image synthesis of similar subjects. Our method is motivated by the overlap between self-attention maps and cross-attention maps for mixed subjects, which is penalized by the proposed self-cross diffusion guidance loss during inference. Further more, we aggregate self-attention maps for multiple patches to a single attention map. In other words, our formulation involves all relevant patches of a subject beyond the most discriminant one. We are the first to reduce overlap between cross-attention and aggregated elf-attention maps, while previous methods are limited to the self-attention map from one patch or rely on cross-attention maps alone for guidance. We utilize standard gradient-based optimization and initial noise optimization [17] for minimizing our guidance loss during inference. We also released SSD, a new dataset of similar subjects for image synthesis, and leveraged the latest vision large language model (GPT-4o) for automatic and reliable evaluation of different methods. Qualitative and quantitative results show significant improvement over previous approaches. Our Self-Cross guidance greatly reduced subject mixing while also reducing the issue of object neglect as a side effect. + +In the future, we will extend our approach to video generation of similar subjects that face challenges of subject mixing. We also anticipate the issue of subject mixing to be less prominent with newer backbone models but not disappear. We will keep exploring variants of Sef-Cross diffusion guidance for finer-grained subjects synthesis and addressing other issues such as attribute binding. + +Acknowledgement This research was supported by the Department of Defense under funding award W911NF-24-1-0295 and by Google Cloud research credits program. + +# 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. 7, 8 +[2] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2283-2293, 2023. 3 +[3] Zhipeng Bao, Yijun Li, Krishna Kumar Singh, Yu-Xiong Wang, and Martial Hebert. Separate-and-enhance: Compositional finetuning for text2image diffusion models. In SIGGRAPH, 2024. 2, 3, 5 +[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 2 +[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1 +[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560-22570, 2023. 3 +[7] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, pages 4055-4075. PMLR, 2023. 2 +[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 1, 2, 3, 5, 4 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2, 3 +[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 2 +[11] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via + +hierarchical transformers. Advances in Neural Information Processing Systems, 35:16890-16902, 2022. 2 +[12] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3 +[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 1 +[14] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 2 +[15] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7545-7556, 2023. 3 +[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[17] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024, 1, 2, 3, 4, 5, 6, 7, 8 +[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. 2022. 1 +[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2023. 3 +[20] Jonathan Ho. Classifier-free diffusion guidance. ArXiv, abs/2207.12598, 2022. 2, 3, 1 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 2 +[22] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 2, 6, 7, 4 +[23] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv: 2307.06350, 2023. 5 +[24] Kaiyi Huang, Chengqi Duan, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2I-CompBench++: An Enhanced and Comprehensive Benchmark for Compositional Text-to-Image Generation. IEEE Transactions on Pattern Analysis Machine Intelligence, (01):1-17, 5555. 5 + +[25] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10124-10134, 2023. 2 +[26] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7701-7711, 2023. 3 +[27] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11523-11532, 2022. 2 +[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 8 +[29] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 1 +[30] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2021. 2 +[31] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9005-9014, 2024. 1, 2, 3, 5, 6, 7, 8, 4 +[32] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 2 +[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Proceedings of the 38th International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2 +[34] Nobuyuki Otsu. A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 1979. 4 +[35] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7942, 2024. 3, 4 +[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2 +[37] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. *ICLR*, 2023. 1 + +[38] 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 +[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. In *arxiv*, 2022. 2 +[40] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2024. 3 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8, 4 +[42] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3 +[43] 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 +[44] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International conference on machine learning, pages 30105-30118. PMLR, 2023. 2 +[45] Eyal Segalis, Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. A picture is worth a thousand words: Principled recaptioning improves image generation, 2023. 2 +[46] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, Lille, France, 2015. PMLR. 2 +[47] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2 +[48] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 2 +[49] Yifan Zhou Xingang Pan Tianyi Wei, Dongdong Chen. Enhancing mmdit-based text-to-image models for similar subject generation. https://arxiv.org/abs/2411.18301, 2024. 1 +[50] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, 2023. 2 +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia + +Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3 +[52] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Text-to-image diffusion with token-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8553-8564, 2024. 3 +[53] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 3 +[54] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2 +[55] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 2 +[56] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 2 +[57] Yanbing Zhang, Mengping Yang, Qin Zhou, and Zhe Wang. Attention calibration for disentangled text-to-image personalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4764-4774, 2024. 3 +[58] Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Ji-dong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, and Jie Tang. Cogview3: Finer and faster text-to-image generation via relay diffusion. arXiv preprint arXiv:2403.05121, 2024. 2 +[59] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 2 \ No newline at end of file diff --git a/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/images.zip b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..db77c52fb41af7548e60cc8c5754a13959b2d941 --- /dev/null +++ b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc0c40d939b598eaf838ad343959407272e30e826a5add631e7739ed23cf2985 +size 854043 diff --git a/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/layout.json b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3d6138e9343d8820d7135579d30475a52701a583 --- /dev/null +++ b/CVPR/2025/Self-Cross Diffusion Guidance for Text-to-Image Synthesis of Similar Subjects/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97da376a2876eefb1916d5b344a08e89fa5545dd2d38412bcdd81c07ea908439 +size 409363 diff --git a/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/62355ea3-56f2-42bd-a178-b9a4adbdf932_content_list.json b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/62355ea3-56f2-42bd-a178-b9a4adbdf932_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ff0a7365070cb32c999bb5b9936b099c93e142ab --- /dev/null +++ b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/62355ea3-56f2-42bd-a178-b9a4adbdf932_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3e62094c6a83ac88a604476ff94fa72322b0c190fbeaa32c771f70bcce413ee +size 72885 diff --git a/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/62355ea3-56f2-42bd-a178-b9a4adbdf932_model.json b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/62355ea3-56f2-42bd-a178-b9a4adbdf932_model.json new file mode 100644 index 0000000000000000000000000000000000000000..99eb8184c7d34f61516d78108d52e581f66992f4 --- /dev/null +++ b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/62355ea3-56f2-42bd-a178-b9a4adbdf932_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eae8208d560fac52c9558c41846879358b4a7bfcdcedb4a8a9976735dff6ea7c +size 92184 diff --git a/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/62355ea3-56f2-42bd-a178-b9a4adbdf932_origin.pdf b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/62355ea3-56f2-42bd-a178-b9a4adbdf932_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f163ddfff229ca602b08b3869b2e49e77fed8a25 --- /dev/null +++ b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/62355ea3-56f2-42bd-a178-b9a4adbdf932_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1a39d94b6bcdb39cca9dda86bd12faecf082f06172fde5de377072924d76f24 +size 1071887 diff --git a/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/full.md b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cd08a9687e3a4c682481371ddfa746c44fac352e --- /dev/null +++ b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/full.md @@ -0,0 +1,265 @@ +# Self-Evolving Visual Concept Library using Vision-Language Critics + +Atharva Sehgal $^{1*}$ Patrick Yuan $^{2}$ Ziniu Hu $^{3}$ Yisong Yue $^{3}$ Jennifer J. Sun $^{2}$ Swarat Chaudhuri $^{1}$ + +1University of Texas at Austin 2Cornell University 3California Institute of Technology + +# Abstract + +We study the problem of building a visual concept library for visual recognition. Building effective visual concept libraries is challenging, as manual definition is labor-intensive, while relying solely on LLMs for concept generation can result in concepts that lack discriminative power or fail to account for the complex interactions between them. Our approach, EsCHER, takes a library learning perspective to iteratively discover and improve visual concepts. EsCHER uses a vision-language model (VLM) as a critic to iteratively refine the concept library, including accounting for interactions between concepts and how they affect downstream classifiers. By leveraging the in-context learning abilities of LLMs and the history of performance using various concepts, EsCHER dynamically improves its concept generation strategy based on the VLM critic's feedback. Finally, EsCHER does not require any human annotations, and is thus an automated plug-and-play framework. We empirically demonstrate the ability of EsCHER to learn a concept library for zero-shot, few-shot, and fine-tuning visual classification tasks. This work represents, to our knowledge, the first application of concept library learning to real-world visual tasks. + +# 1. Introduction + +How do humans recognize different visual categories? Consider the example in Figure 1: while easily recognizable as a pastry, distinguishing between a "donut" and a "beignet" requires understanding visual concepts such circular shape with a hole for the donut, puffy texture, and the presence of powdered sugar specifically on the beignet. These visual concepts, including shape, texture, and the presence or absence of specific features, enable us to make distinctions between objects. Concept-bottleneck visual recognition [19, 30, 31] aims to leverage these discriminative visual concepts, to enable vision systems to more accurately recognize a wider range of classes. Here, we study general + +![](images/9367a63b01b253bbca70c09300acf9ad30512c01df2a3d55a56d50b6776ea2d3.jpg) +Figure 1. An overview of EsCHER. Prior work: concept-bottleneck visual recognition aims to leverage discriminative visual concepts to enable more accurate object classification. Ours: EsCHER is an approach for iteratively evolving a visual concept library using feedback from a VLM critic, to discover more effective visual concepts. + +approaches for improving concept-bottleneck visual recognition systems by evolving visual concept libraries to find more effective concepts. + +Existing concept-bottleneck visual recognition systems typically leverage a Large Language Model (LLM) to generate a set of potential visual concepts relevant to the task, then use a Vision-Language Model (VLM) to make predictions from these concepts. This process potentially improves both interpretability and accuracy for classification [19, 30, 31]. However, existing methods face limitations: manually defined concepts are labor-intensive, and LLM-generated concepts can be inaccurate or fail to account for interactions between them. We need more effective methods to construct and refine visual concepts. One promising approach to improve visual concept learning is to leverage library learning [6, 9, 10], which focuses on building a reusable collection of components. While library learning has shown success in domains that are naturally + +symbolically decomposable (e.g., equation learning [9]), it is not well-explored for visual concept learning. Our key insight is that library learning complements visual concept learning by providing a structured and evolving repository of concepts that is more effective for visual recognition. + +To achieve this, we introduce EsCHER, a novel self-evolving framework to automatically discover and refine a library of visual concepts. EsCHER employs an iterative algorithm where a VLM acts as a critic, providing feedback on the effectiveness of concepts generated by an LLM. Specifically, the VLM evaluates the similarity between each image and its associated concepts compared to other images. This evaluation, captured in a contrastive score, serves as a feedback signal to guide the LLM in refining its generated concepts. Furthermore, EsCHER provides the history of concepts and feedbacks to the LLM, enabling the LLM to effectively learn from its past performance and improving its proposals over time. Through this iterative process, EsCHER produces a set of concepts that are both accurate and highly informative for the VLM, enabling it to make more effective predictions. + +Our approach offers several key advantages. First, it is broadly applicable and complements a range of existing concept-bottleneck visual recognition frameworks, including those designed for zero-shot, few-shot, and fine-tuned settings. This adaptability ensures that as LLMs, VLMs, and visual concept learning frameworks continue to evolve, EsCHER remains relevant and applicable to emerging techniques. Second, EsCHER requires no human annotations or labeled datasets, making it a plug-and-play solution for various visual recognition tasks. Finally, the iterative approach of EsCHER leverages the in-context learning capabilities of LLMs, allowing them to learn from their concept history and generate increasingly effective concept concepts. This iterative refinement process ensures that the concept library continuously adapts and improves, leading to more accurate and discriminative visual representations. + +To summarize, our contributions are: + +- We present EsCHER, a novel VLM- and LLM-based framework for self-evolving visual concept libraries. Our method does not require human-labeled data and can improve the quality of the learned concepts via an open-ended learning loop. +- We develop an iterative concept refinement algorithm, leveraging the both the ability of VLMs to act as a critic and the ability of LLMs to incorporate history, to improve visual concepts based on past performance. +- We demonstrate that EsCHER is complementary to a range of different state-of-the-art baselines, and our learned concept library improves performance across zero-shot, few-shot, and fine-tuned image classification settings. + +# 2. Related Work + +Vision-language models. VLMs have emerged as powerful tools for a wide array of visual tasks, from visual question answering to zero-shot image classification [2, 18, 23]. These models are typically trained on large datasets of image-text pairs, enabling them to reason about the relationship between these modalities. Some VLMs [11, 23, 32] are trained using contrastive learning objectives to learn embedding spaces aligned between image and text representations. These models have shown promising results for zero-shot and few-shot classification. + +While VLMs have shown remarkable capabilities, they have limitations that motivate exploring alternative approaches, such as concept-bottleneck models. One limitation is their difficulty in perceiving and reasoning about fine-grained visual concepts. For example, recent work has shown that VLMs can struggle to distinguish between subtle visual differences [26]. As such, despite recent advances, these models can still often fail with accurate visual recognition. The standard zero-shot method does not provide any intermediate understanding or explanation of the model's reasoning process [19]. These challenges highlight the need for more interpretable and controllable approaches to visual recognition, such as those offered by concept-bottleneck methods. + +Concept-bottleneck Visual Recognition Models. Our work builds upon recent work on concept-bottleneck models [19, 22, 24, 30, 31, 33], which first identifies relevant concepts using a vision language model (VLM), and then uses those concepts to make a prediction. Compared to directly querying a VLM, this approach has advantages including interpretability, (if the concepts are interpretable) as well as stronger accuracy (if the concepts are useful in capturing the classification task). Similar ideas predate the rise of VLMs [12, 17], where a so-called concept bottleneck is built into neural network architectures, and concepts are learned via end-to-end training. Recent algorithms for learning concept-bottleneck models with VLMs generally fall within two categories: + +Non-parametric Algorithms: These methods focus on improving the visual concepts by employing non-parametric optimization techniques. One increasingly common approach is to use zero-shot queries to an LLM to select a list of concepts for each class that is useful for classification [19]. The aggregated score for a class is the mean over the scores of the selected concepts. [5] follows the LLM induction paradigm to initialize class concepts but, borrowing ideas from genetic mutation, repeatedly queries a finetuned LLM to generate new concepts and new concept selections for each class - using binary classification loss to rate each mutation. This mutation process focuses + +![](images/589b0c3a8a8a908f0c63a53fff6cf3245685adf832920e0bb9600012872a69d9.jpg) +Concept-Bottleneck Visual Recognition Learning $w_{y}$ in isolation + +![](images/7afb9756bdf757fab89f22cabee8cd4c86d0c155e4405e522cd53d8bb7de181b.jpg) +ESCHER: Self-Evolving Visual Concept Library Simultaneously Learning $w_{y}$ and $\mathcal{C}$ +Figure 2. (Left) Existing work on concept-bottleneck visual recognition, where a VLM scores a set of concepts to perform classification. The classification is based on the class with the maximum concept scores. (Right) EsCHER. (1) EsCHER follows previous work [19] in instantiating a set of concepts for each class using an LLM. (2) It initializes a concept-bottleneck model and collect the predictions for a classification dataset $\mathcal{D} = \{\left(\mathbf{x}_i,y_i\right)\}_{i=1}^N$ (labels optional). (3) A concept similarity heuristic identifies frequently confused classes. (4) A history bank then stores relevant information to guide (5) the LLM sampling procedure for improved concepts that disambiguate these classes. The new concepts are integrated into the next iteration. + +on each class in isolation and must be repeated for each class, which proves to be impractical for datasets with more than 20 classes. EsCHER, instead, reasons jointly about all classes and focuses on only those classes that are underperforming. As EsCHER is agnostic to the choice of VLM and the number of samples needed for training, it is technically possible to integrate the llm-mutate framework within EsCHER to maximize classification performance. + +Parametric Algorithms: These methods focus on improving the performance of a visual concept bottleneck classifier by training parametric adapters on top of the scores output by a frozen VLM model. These models may also subsample concepts from the concept library. These architectures generally consist of a 'concept bottleneck' [30, 31], with additional learning required such as a linear probing adapter or additional finetuning [24]. EsCHER is agnostic to the choice of concept tuning method, and focuses on using such methods to guide the concept discovery loop. One can 'plug' EsCHER into any of these other works and observe a performance improvement while retaining the key characteristics of the type of architecture used. + +Library Learning. Library learning is an emerging direction of program synthesis that aims to automate the construction of reusable components (libraries) for program generation. This is often framed as a hierarchical Bayesian + +optimization problem, where the goal is to simultaneously learn the library of components and the optimal way to combine them to solve a given task [6, 7, 16, 29]. While library learning has shown promise in domains like equation learning [9], its application to visual concept learning presents unique challenges. + +Unlike domains that are naturally symbolically decomposable, visual concepts often exhibit complex and subtle relationships that are difficult to capture with traditional library learning techniques. Moreover, the space of potential visual concepts is vast and diverse, making it challenging to design effective search strategies. Our work addresses these challenges by introducing a novel library learning approach designed for visual concept discovery. By leveraging a VLM as a critic and incorporating class resolution history into the LLM, our method can effectively explore the space of visual concepts and construct a dynamic library that adapts to the specific needs of the visual recognition task. + +# 3. Problem Formulation + +Concept-Based Visual Recognition Our work is rooted in the emerging area of concept-bottleneck visual recognition [19, 30, 31] (Figure 2). Given an image $\mathbf{x}$ , a set of concepts $\mathcal{C}$ , and a label $y$ , the basic setup is to use a vi + +sion language model (VLM) to score the likelihood of each concept $c \in \mathcal{C}$ for image $\mathbf{x}$ , denoted as $\text{score}_{\text{VLM}}(\mathbf{x}, c)$ . We also write the vector of scores over all concepts as $\text{score}_{\text{VLM}}(\mathbf{x}, \mathcal{C})$ . Afterward, the aggregate score of a label $y$ for image $\mathbf{x}$ is a weighted sum over the concept scores: $f(\mathbf{x}, y) = w_y^T \text{score}_{\text{VLM}}(\mathbf{x}, \mathcal{C})$ , where $w_y$ can be either a learned or fixed parameter vector. Finally, classification over a fixed label set $\mathcal{Y}$ is performed by choosing the class that maximizes image-concept similarity. + +$$ +y ^ {\star} = \arg \max _ {y \in \mathcal {Y}} w _ {y} ^ {\top} \operatorname {s c o r e} _ {\mathrm {V L M}} (\mathbf {x}, \mathcal {C}) \tag {1} +$$ + +This setup encompasses the bulk of recent work in concept-bottleneck models for image classification [19, 30, 31] and helps considerably in fine-grained and out-of-distribution classification scenarios. For instance, while a class $y =$ "SpaceX Starship" may not have been seen during CLIP training, we can construct a reasonably accurate and interpretable rocket classifier by aggregating the likelihoods over the feature set: $\{f(\mathbf{x}|c_y)\mid c_y\in [\text{Stainless Steel Rocket},\text{Grid Fins},\text{Space X logo}]\}$ . Selecting this feature set can be automated using a foundation model with access to an external database [25]. + +Typically, a concept-bottleneck model is developed by predefining a fixed set of concepts $\mathcal{C}$ , and then finetuning the weight matrix $w_{y}$ for each class. The optimization objective is then chosen to maximize the performance of the model and identify an interpretable set of concepts. A Bayesian formulation is presented in Eq 3. + +In contrast, scientists – when confronted with a new domain – rarely rely on a static set of concepts. The first reaction of a scientist is to learn more about the domain and expand their conceptual knowledge base. This newly gained knowledge is subsequently used to structurally discriminate between classes. For instance, even a trained ecologist might struggle to differentiate a Northern Curly-tailed Lizard from a Florida Scrub Lizard due to lack of prior knowledge, while a herpetologist can rely on their knowledge of lizard physiology to identify the correct characteristic feature difference2 + +Visual Recognition with Latent Concept Libraries. We model this evolving set of concepts as textual descriptions drawn from a latent concept library. We frame the relationship between the latent concept library and the classification model as a Hierarchical Bayesian model consisting of (i) a prior $p(\mathcal{C})$ representing the natural distribution over concept libraries; (ii) a model $p_{\mathcal{C}}(w_{\mathcal{Y}}|\mathcal{C})$ that quantifies the likelihood of assigning open-vocabulary classes for a given concept library; and (iii) an evaluation function $p(\mathcal{D}|\mathcal{C}) \coloneqq$ + +$\operatorname{argmax}_{y \in \mathcal{Y}} w_y^T \operatorname{score}_{\mathrm{VLM}}(\mathbf{x}, \mathcal{C})$ which grounds the performance of a concept library using an image classification dataset $(\mathcal{D})$ and a VLM-based recognition engine. We assume that the distributions $p(\mathcal{C})$ and $p(w_{\mathcal{Y}}|\mathcal{C})$ can be approximated using LLMs. That is, we can prompt an LLM to generate interesting concepts, and we can prompt an LLM to generate and discover new concepts that adhere to an open-vocabulary category. We also assume that the VLM-based visual recognition engine is well calibrated for confidence estimation. We now pose the problem of visual recognition with latent concept libraries as one of simultaneously inducing an optimal set of concepts and an optimal concept-bottleneck visual reasoning model: + +$$ +\begin{array}{l} w _ {\mathcal {Y}} ^ {\star}, \mathcal {C} ^ {\star} = \arg \max _ {w _ {\mathcal {Y}}, \mathcal {C}} p (w _ {\mathcal {Y}}, \mathcal {C} | \mathcal {D}) \tag {2} \\ = \arg \max _ {w _ {\mathcal {Y}}, \mathcal {C}} \underbrace {p (\mathcal {D} | w _ {\mathcal {Y}})} _ {\text {C B M t r a i n i n g}} \cdot \underbrace {p (w _ {\mathcal {Y}} | \mathcal {C})} _ {\text {B y L L M}} \cdot \underbrace {p (\mathcal {C})} _ {\text {B y L L M}} \\ \end{array} +$$ + +# 4. Method + +ESCHER performs a two stage evolution over the natural-language concepts and the weight matrix assignment. The two stages follow an alternating maximization strategy, as illustrated in Figure 2: (1) Concept Bottleneck Optimization: We fix a set of concepts and learn a concept-bottleneck model that maximizes the fitness to the dataset (Fig. 2, Left). (2) History-sensitive concept evolution: We leverage the best model to identify classes that appear to be confused and sample new concepts to resolve the confusion (Fig. 2, Right). + +The rest of this section first describes the classical concept-bottleneck model maximization strategies. Then, we discuss common heuristics for identifying confused classes. Finally, we show how the classes can be disambiguated by sampling new concepts conditioned on feedback derived from previous evolutions. + +Concept Bottleneck Optimization. Concept-bottleneck models [13] generate their predictions by learning to linearly combine the intermediate predictions over a fixed set of interpretable concepts. This yields a high-performing yet interpretable classifier that can be used for downstream classification tasks. We focus on two paradigms within this field for optimizing the adapter weight matrix. + +Zero shot maximization. In this setting, the adapter weights are instantiated by an LLM. Intuitively, the adapter will take the form of a block diagonal matrix, where each block represents the concepts selected by the LLM for a particular class and each element in the block is assigned the uniform weight $1 / |c_y|$ , where $c_{y}$ is the set of concepts the LLM generates for a particular label $y$ . As no labels are needed, this paradigm generates extremely flexible classifiers. However, the efficacy of the concepts is deeply tied to the backbone VLMs ability to score fine-grained concepts. + +Algorithm 1 Pseudocode for EsCHER. EsCHER takes as input a set of open vocabulary categories $\mathcal{V}$ , a dataset of images $\mathcal{D}$ (labels optional), a pretrained vision-language model $\theta_{\mathrm{VLM}}$ , an optional adapter $w$ (We drop the $\cdot_{\mathcal{V}}$ subscript for readability), and three hyperparameters: the number of iterations $T$ , the decay rate $\gamma$ for repeated categories, and the number of pairs to evolve $K$ . EsCHER outputs two artifacts: the adapter parameters after $T$ iterations, $w_{T}$ , and the corresponding evolved library of interpretable concepts $\mathcal{C}_T$ . + +1: function EsCHER(y, D = {(xi, yi})i=0, θvLM, T, γ, K) +2: $\mathcal{C}_0\gets$ INITCONCEPTS(y) Initialize concepts via zero-shot LLM queries +3: $H_0^{(i,j)} \gets \text{INITHISTORY}(\mathcal{Y}, T)$ Track concepts and feedback for each class pair per iteration +4: for $t$ in range $(T)$ do +5: $w_{t}^{\star}\gets \mathbf{fit}(w_{t},\mathcal{C}_{t},\mathcal{D},\theta_{\mathrm{VLM}})$ +6: $\hat{\mathbf{y}}\gets \mathsf{evaluate}(\mathcal{D},\theta_{\mathrm{VLM}},w_t^{\star})$ +7: $\{r_{ij}\}_{i,j = 1}^{|\mathcal{V}|^2}\gets \text{CALCULATE SIMILARITY} (\hat{\mathbf{y}})$ +8: $H_{t}^{(i,j)} \gets \text{UPDATEHISTORY}(\{r_{ij}\})$ ▷ Store the similarity of $y_{i}$ and $y_{j}$ for iteration $t$ +9: $s_{ij}\gets \mathbf{COMPUTESAMPLEPROB}(r_{ij},H_{[1:t]}^{(i,j)},\gamma)$ +10: $\{r_{ij}\}_{1}^{K}\gets$ subsample $(\{r_{ij}\} ,s_{ij})$ +11: for each $(i,j)\in \{r_{ij}\}_{1}^{K}$ do +12: $\hat{c}^{(i)},\hat{c}^{(j)}\gets \mathsf{CONCEPTEVOL}(y_i,y_j,\mathcal{C}_t^{(i)},\mathcal{C}_t^{(j)},H_{[1:t]}^{(i,j)})$ +13: $\mathcal{C}_{t + 1}^{(i)}\gets \mathcal{C}_t^{(i)}\cup \{\hat{c}^{(i)}\}$ +14: $\mathcal{C}_{t + 1}^{(j)}\gets \mathcal{C}_t^{(j)}\cup \{\hat{c}^{(j)}\}$ +15: $H_{t+1}^{(i,j)} \gets \text{UPDATEHISTORY}(\mathcal{C}_{t+1}^{(i)}, \mathcal{C}_{t+1}^{(j)})$ ▷ Store the updated concepts for iteration $t + 1$ +16: $w_{t + 1}\gets w_t^\star$ +17: $C_{t + 1}\gets C_t$ +18: return $w_{T}, \mathcal{C}_{T}$ + +Few shot / Fine-tuned maximization. Under this paradigm, the adapter weights are instantiated as a learnable linear layer of shape $\mathbb{R}^{|C| \times |\mathcal{Y}|}$ . As the set of concepts can grow very large, most approaches subsample the set of concepts as well. The linear layer is trained with cross-entropy loss, often with various regularizers [30]. In the few-shot setting, the number of images per class is fixed while in the fine-tuned setting, no restriction is placed on the number of images. This adapter training approach also overcomes the inherent weakness of the zero shot setting, as concepts that do not add to the performance of the model can be down-weighted or ignored by the linear adapter. + +Heuristics for disambiguation. After maximizing the model for the given set of concepts, the CALCULATESIMILARITY function identifies classes that are confused with each other. This function takes as input a matrix of scores for the dataset of images $\hat{\mathbf{y}}\in \mathbb{R}^{N\times |\mathcal{Y}|}$ . Each value in $\hat{\mathbf{y}}$ indicates the image-class similarity. We leverage this ma + +trix of scores to identify classes that are confused with each other across the images in the dataset. $\{r_{ij}\}_{i,j = 1}^{|\mathcal{Y}|^2}$ denotes the list of all possible confusion scores. EsCHER's modular framework allows us to admit any heuristics to model this confusion score. We list several heuristics to measure this disambiguation signal that perform well empirically. Additional studies are presented in $\S 8.4$ , Figure 4, and Figure 5. + +- Top-k Confusion Top-k confusion is computed by sorting each row in descending order of score and keeping the top $k$ class predictions. That is: $r_{ij} = \text{confusion\_mat}(\text{sort}(\hat{\mathbf{y}})[:,k])$ . Consult Figure 4 for more insight into how this heuristic works. +- Correlation: We compute Pearson's correlation between each class $r_{ij} = \operatorname{corrcoef}(\hat{\mathbf{y}}^{\top})$ . Pearson's correlation measures linear correlation between two classes ( $y_i$ and $y_j$ ) with -1 denoting maximal negative correlation and 1 denoting maximal positive correlation. This is visualized in Figure 5. The hyperparameter $k$ differs from the hyperparameter $K$ which tracks the number of classes to disambiguate. +- Agglomerative Clustering: Identifying classes with similar response signals can be viewed as an unsupervised clustering problem over the columns of $\hat{\mathbf{y}}$ . The number of clusters is kept as a tunable hyperparameter, and choose the values of $r_{ij}$ by greedily bottom-up traversing the dendrogram. +- Confusion Matrix: EsCHER assumes no access to labels during evolution. However, if labels are provided, we can construct a confusion matrix to identify classes that are confused for each other. + +History-sensitive concept evolution. Each value in $\{r_{ij}\}_{1}^{K}$ represents two classes that the model is unable to discriminate between. This suggests that the model lacks the correct discriminative concepts in the concept space $\mathcal{C}$ to conceptually separate $y_{i}$ and $y_{j}$ . Now we use the CONCEPTEVOLUTION function, which uses a zero-shot prompt to extract a list of natural language concepts for both classes $(\hat{c}_i,\hat{c}_j)$ that add additional concepts useful for disambiguating $y_{i}$ and $y_{j}$ to the concept library. + +This task requires good reasoning skills, so we provide the model with a scratchpad to enhance the model's reasoning abilities [21]. More details are presented in § 8.5. + +This process is repeated for each pair of confused concepts. However, the CALCULATESIMILARITY function generates $|\mathcal{Y}|^2$ pairs in each iteration, and processing all such pairs is extremely inefficient for real-world datasets. To increase feedback generation efficiency, we only sample the top $K$ pairs from a random distribution weighted by the confusion coefficient score and a penalty for repeat confusions which increases exponentially, controlled by a decay parameter $\gamma$ . This is implemented as + +COMPUTESAMPLEPROB $(r_{ij},H_{[1:t]}^{(i,j)},\gamma)$ where the size of $H_{[1:t]}^{(i,j)}$ indicates the number of times the $(i,j)^{\mathrm{th}}$ classes have been confused for each other. The $\gamma$ parameter helps guard against classes that are often misclassified due to failures in the backbone ViT model. + +Additionally, each pairwise disambiguation can cause collisions with other classes in later iterations, which makes it likely that some class pair will need multiple rounds of feedback. If we generate feedback in isolation, the model is likely to regurgitate previously proposed features. To ensure that each new round of feedback generates novel and interesting concepts, we borrow ideas from the program synthesis literature [1] and append past 'execution traces' to the model query. This history-conditioned prompt is presented in Figure 6. We maintain a history of past evolutions for each pair $(y_{i},y_{j})$ along with the similarity score derived from the VLM critic's score of the modified concepts $c_{i}\cup \hat{c}_{i}$ and $c_{j}\cup \hat{c}_{j}$ . Concretely, INITIALIZE HISTORY instantiates this datastructure, and UPDATE HISTORY updates the relevant fields in each iteration (the new concepts added to class $i$ and $j$ in iteration $t$ and the class confusion score in iteration $t + 1$ after incorporating feedback). These functions are explored in more detail in $\S 8.10$ . + +# 5. Experiments + +ESCHER is a meta-algorithm that aims to enhance the performance of concept-bottlenecked models (CBMs) by learning a library of concepts using an alternating maximization loop. Such CBMs operate in diverse data regimes, with some models requiring no human labels [19] and others requiring a fully annotated classification dataset ([30]). Our experiments focus on studying whether EsCHER can improve the performance of such preexisting algorithms that are characteristic of the data regime they operate in, with no additional modifications ( $\S 5.1$ , $\S 5.2$ , and $\S 5.3$ ) + +In addition, §5.4 presents an ablation of EsCHER's library learning component, and §5.5 explores EsCHER's performance under various LLM and VLM backbones. More studies are presented in the Appendix (§ 8). + +Datasets. We demonstrate the effectiveness of EsCHER on seven fine-grained and general purpose classification datasets that generally fall within three categories: fine-grained classification datasets with scientific applications (NABirds [27] and CUB [28]), fine-grained classification datasets with general purpose applications (Food-101 [3], Stanford Cars [14], Flowers-102 [20]), and general purpose categorization datasets (CIFAR100 [15]). + +Evaluation. We extend three concept-bottleneck visual classification models with publicly available and reproducible codebases at the time of writing: LM4CV (fine + +tuned adapter), LaBO (few shot adapter), and Classify by descriptions (zero-shot adapter) [19, 30, 31]. We follow previous work in using Top-1 accuracy to evaluate the performance of the CBM model on the test set of the classification dataset. + +Methodology. All experiments begin with an initial set of concepts generated with a backbone LLM (gpt-3.5-turbo-0125 [4]). A CBM conditioned on these concepts produces logits for the images in the validation set, and a heuristic identifies classes that are confused for each other. The specific heuristic and hyperparameters vary depending on the task and underlying algorithm, and grid-searching for such hyperparameters proves to be practical and effective. We use a generic history-conditioned prompt presented in Fig. 6. We use a ViT-L/14 CLIP model [23] as the vision backbone for all experiments unless otherwise clarified. + +# 5.1. Comparison against fine-tuned baselines + +
LM4CVLM4CV+ESCHER
CIFAR-10084.4889.63
CUB-200-201163.2683.17
Food10194.7794.90
NABirds76.5878.21
Oxford Flowers94.8096.86
Oxford IIIT Pets92.5092.86
Stanford Cars86.8493.76
+ +Table 1. Top-1 accuracy of LM4CV [30] and LM4CV evolved with EsCHER on multiple fine-grained classification problems. EsCHER improves upon LM4CV's performance in all datasets while utilizing no extra human annotations. + +Setup. For each dataset, we generate an initial set of concepts using queries to gpt-3.5-turbo and a generic visual concept learning prompt. We use LM4CV's suggested hyperparameters wherever possible. We run 60 iterations of Escher. Similarity is computed using the Top-k confusion metric $(k = 3)$ , only the top 50 pairwise confusions are evolved, and decay rate $\gamma$ is set to $1/30$ (i.e.: after 30 repeated evolution calls, the value drops to half of the original). LM4CV and LM4CV+Escher are trained for the same number of steps per iteration with the same batch size and learning rate. + +Observations. Our observations are presented in Table 1. We draw three observations from this experiment. First, LM4CV + EsCHER achieves a higher Top-1 accuracy than vanilla LM4CV on all datasets, suggesting that learning a + +concept library is an effective axis of improvement. Second, LM4CV+ESCHER significantly improves the performance of LM4CV on datasets where the initial accuracy is low, and finally, LM4CV+ESCHER did not require any human provided information to achieve this result. + +# 5.2. Comparison against few-shot baselines + +
Dataset8 shot16 shot
LaBOLaBO+ESCHERLaBOLaBO+ESCHER
CIFAR10074.2373.6277.6777.23
CUB72.7873.3778.7578.79
Food10187.0286.1088.4988.50
Oxford Flowers95.6695.3797.6997.80
Stanford Cars75.0775.9981.4882.56
+ +Setup. Our setup for LaBO follows the same setup as that of LM4CV. We evaluate all datasets for 8 shot and 16 shot. This setup necessitates a balanced set of images for each class. We drop NABirds from this evaluation, as some NABirds classes contain as low as 3 images per class. We keep the same hyperparameters as our LM4CV experiments but do not use a decay rate as the number of repeat classifications is generally unproblematic. + +Observations. We observe mixed results in few-shot evaluation, with similar performance compared to the baseline on all datasets except for Stanford Cars, which showed modest improvement, and CIFAR100, which showed very modest deterioration (Table 2). In general, LaBo+ESCHER performs considerably better in the 16-shot setting compared to the 8-shot setting. We hypothesize these mixed results are induced by poor calibration of LaBO's CBM's and as a result of model overfitting on the few available labels for each class. + +# 5.3. Comparison against zero-shot baselines + +Setup. Our zero shot setup compares against Classify by Descriptions. We sample all concepts from gpt-3.5-turbo. We use the same hyperparameters as LM4CV experiments, except for setting the decay rate to 50. We continue to use the ViT-L-14 backbone mode. + +Observations. Our observations are presented in Table 3. Overall, we find that CbD's performance improves consistently for all datasets when we evolve the concepts with Es-CHER. The relative improvement in performance is less + +Table 2. Top-1 accuracy for LaBO and LaBO evolved with Es-CHER on multiple fine-grained classification datasets in a few-shot learning setting. LaBO benefits from EsCHER's library guidance in the 16 shot setting more than in the 8 shot setting. We observe mixed results on EsCHER's efficacy in the few shot domain. + +
DatasetCLIPCbDCbD+ESCHER
CIFAR-10073.3076.2077.80
CUB-200-201164.8362.0063.33
Food10192.5193.1193.58
NABirds53.5353.6154.30
Oxford Flowers74.5179.4181.37
Stanford Cars74.5375.6577.14
+ +than that in LM4CV as the backbone (CLIP ViT-L-14) model's output scores are less calibrated than those produced by the finetuned adapter, which leads to noisier iterations. + +# 5.4. Library Learning Ablation + +Table 3. Top-1 accuracy of CLIP (ViT-L/14) [23], Classify by Descriptions (CbD) [19], and CbD evolved with EsCHER on multiple fine-grained classification datasets in a zero-shot learning setting. $\mathrm{CbD + EsCHER}$ improves upon CbD's performance in all datasets. + +
DatasetLM4CVLM4CV + Many ConceptsLM4CV + EsCHER
CIFAR-10084.4886.9189.63
CUB-200-201163.2666.0983.17
Food10194.7794.7794.90
NABirds76.5876.2878.21
Oxford Flowers94.8094.5196.86
Oxford IIIT Pets92.5092.0292.86
Stanford Cars86.8486.8493.76
+ +Table 4. Top-1 accuracy of an ablation of EsCHER's library learning component. For LM4CV, we replace the concepts learned with library learning with an equal number of concepts sampled from an LLM. We find that concepts evolved with EsCHER still outperform naively sampling more concepts - suggesting that feedback from a VLM critic is essential for LM4CV+EsCHER's performance. + +ESCHER focuses on using zero-shot disambiguation queries to maximize the performance of a concept library. To do this, EsCHER makes asymmetrically more calls to a zero-shot LLM model than the baseline (which only samples concepts once). To verify that EsCHER's performance is not simply a result of more concepts or more number of LLM calls, we rerun LM4CV with three times more concepts per class than the initial set of concepts (for a total of 3810 concepts). These concepts are sampled with the same LLM backbone we use in EsCHER. This represents an ablation on the Library learning component used in EsCHER. + +Results are highlighted in Table 4. We find that LM4CV's performance does not significantly improve even given the same number of LLM sampled concepts as LM4CV+ESCHER. This suggests that (1) the library learning component is essential for concept evolution and (2) sampling concepts without taking into account feedback + +![](images/66d33c0895d06970895719bcb8981464814608008a8144e1d3f3b42c09cd8c9a.jpg) +This is a Male Ring-necked pheasant. + +![](images/7a11bb90e40aa862f8f4f417d83551f5023af636dd969ce4294288dd94eb1c4c.jpg) +With no iterations, the baseline confuses this for an Female Ring-necked pheasant because: +While the true class has lower aggregate activation because: +After iteration with EsCHER, the baseline correctly predicts this to be a Male Ring-necked pheasant because: + +![](images/cbc06c837b31318dd84e06afc2f05552593685898ac31880dbdf35ba64d359ce.jpg) +Figure 3. A qualitative example of evolving concepts with $\mathrm{CbD + EsCHER}$ in NABirds. Initially, the model is confused between two similar categories with almost the same mean CLIP activation indicating that the concepts provide a coarse categorization signal, but miss subtle nuances. After training with EsCHER, the feedback mechanism identifies new characteristic features (e.g. metallic green head and neck) enabling the correct classification. Additional examples are provided in $\S 8.6$ . + +from the VLM results underperforms integrating a VLM 'critic' into the library learning loop. + +We visualize qualitative results of concepts for each dataset and corresponding activated images in the supplementary material. + +# 5.5. Backbone VLM/LLM Ablation + +
ViT-B/16CLIPLM4CVLM4CV +ESCHERCbDCbD +ESCHER
CIFAR-10063.2081.3581.7267.7069.90
CUB-200-201157.0670.9077.1756.1656.16
Food10187.2492.0092.2088.5189.19
NABirds44.5666.6968.7145.2545.25
Stanford Cars61.8680.8781.8265.9666.34
+ +Table 5. Top-1 accuracy for evolving EsCHER with a weaker LLM (Llama-3.3-70B-4bit) and visual critic (ViT-B/16). EsCHER consistently improves the performance of LM4CV and CbD across datasets. + +ESCHER is a meta-algorithm that enhances the performance of existing CBMs. The performance of such models is inherently bottlenecked by the quality of their CLIP-based visual backbone for capturing concept-image relationships and the quality of the GPT-based language backbone for querying relevant concepts. In this experiment, we investigate whether EsCHER can accelerate CBMs using other VLM/LLM backbones. Concretely, we instantiate LM4CV and CbD with a new backbone LLM (4bit quantized Llama-3.3-70B-Instruct [8]) and with new backbone VLMs (ViT-B/16 and ViT-B/32). These backbones are slightly weaker than the base models used in other experiments. As a result, we expect an overall reduction in performance and a weaker learning signal from the visual critic. We maintain the same training and evaluation setup as used in other experiments. + +Observations. Results for these experiments are presented in Table 5 (for ViT-B/16) and in the Appendix Table 10 (for ViT-B/32). Overall, we observe that CBMs tend to perform better when paired with stronger backbones. Additionally, in every case, refining CBMs iteratively using EsCHER leads to better performance than relying on a fixed set of concepts. + +# 6. Conclusion + +We present EsCHER, a framework for evolving visual concept libraries for visual recognition. EsCHER iteratively updates the library using a VLM as a critic to guide an LLM to generate more effective concepts. This process also enables the LLM to incorporate past histories of the feedback from the VLM. Notably, EsCHER does not require any additional annotations, and is compatible with a range of concept-bottleneck visual recognition systems. We demonstrate our results on concept-bottleneck models in zero-shot, few-shot, as well as fine-tuned settings. + +One direction of future work is to further improve the performance of evolving visual libraries on few-shot settings, by incorporating few-shot learning into the library evolution process and allowing the algorithm to leverage limited labeled data. Additionally, we aim to study the application of our approach to more complex visual reasoning tasks, where the learned concept libraries could provide a foundation for higher-level reasoning. By demonstrating the potential of combining library learning and concept-bottleneck visual recognition, we aim to encourage further research at this intersection, towards the development of more robust, interpretable, and intelligent visual recognition systems. + +Acknowledgments This research was partially supported by the NSF Expeditions award #CCF-1918651, a DARPA award #HR00112320018, an ARO award + +# References + +[1] Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and Charles Sutton. Program synthesis with large language models, 2021. 6 +[2] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. 2 +[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 6 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 6 +[5] Mia Chiquier, Utkarsh Mall, and Carl Vondrick. Evolving interpretable visual classifiers with large language models. arXiv preprint arXiv:2404.09941, 2024. 2 +[6] Kevin Ellis, Catherine Wong, Maxwell Nye, Mathias Sable-Meyer, Lucas Morales, Luke Hewitt, Luc Cary, Armando Solar-Lezama, and Joshua B Tenenbaum. Dreamcoder: Bootstrapping inductive program synthesis with wake-sleep library learning. In Proceedings of the 42nd acm sigplan international conference on programming language design and implementation, pages 835-850, 2021. 1, 3 +[7] Gabriel Grand, Lionel Wong, Maddy Bowers, Theo X. Olausson, Muxin Liu, Joshua B. Tenenbaum, and Jacob Andreas. LILO: Learning interpretable libraries by compressing and documenting code. In The Twelfth International Conference on Learning Representations, 2024. 3 +[8] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 8 +[9] Arya Grayeli, Atharva Sehgal, Omar Costilla-Reyes, Miles Cranmer, and Swarat Chaudhuri. Symbolic regression with a learned concept library. In Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3 +[10] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A Ross, Cordelia Schmid, and Alireza Fathi. Scenecraft: An llm agent for synthesizing 3d scenes as blender code. In *Forty-first International Conference on Machine Learning*, 2024. 1 +[11] 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 +[12] Eunji Kim, Siwon Kim, Minji Seo, and Sungroh Yoon. Xprotonet: diagnosis in chest radiography with global and local + +explanations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15719-15728, 2021. 2 +[13] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models, 2020. 4 +[14] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6 +[15] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 6 +[16] Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332-1338, 2015. 3 +[17] Oscar Li, Hao Liu, Chaofan Chen, and Cynthia Rudin. Deep learning for case-based reasoning through prototypes: A neural network that explains its predictions. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2 +[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[19] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 2, 3, 4, 6, 7 +[20] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6 +[21] Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, Charles Sutton, and Augustus Odena. Show your work: Scratchpads for intermediate computation with language models, 2021. 5 +[22] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15645-15655, 2022. 2 +[23] 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. 2, 6, 7 +[24] Oindrila Saha, Grant Van Horn, and Subhransu Maji. Improved zero-shot classification by adapting vlms with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17542-17552, 2024. 2, 3 + +[25] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools, 2023. 4 +[26] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 2 +[27] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 595-604, 2015. 6 +[28] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset, 2023. 6 +[29] Catherine Wong, Kevin Ellis, Joshua B. Tenenbaum, and Jacob Andreas. Leveraging language to learn program abstractions and search heuristics. In International Conference on Machine Learning, 2021. 3 +[30] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian McAuley. Learning concise and descriptive attributes for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3090-3100, 2023. 1, 2, 3, 4, 5, 6 +[31] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification, 2023. 1, 2, 3, 4, 6 +[32] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2 +[33] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Hongsheng Li, Yu Qiao, and Peng Gao. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. arXiv preprint arXiv:2303.02151, 2023. 2 \ No newline at end of file diff --git a/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/images.zip b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..095f63f678e3e588368640491ed4871298a5bd43 --- /dev/null +++ b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb60da7894fa9405e28867856320a2654cc8731696a1f5d6d02143561e7cd46a +size 355376 diff --git a/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/layout.json b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..90b41f74240ae05cfb78f700d91daf5987c3cca2 --- /dev/null +++ b/CVPR/2025/Self-Evolving Visual Concept Library using Vision-Language Critics/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e683ed25042a8511c320b3697ae8970ad3d4a2dc153aed9619e2b3dc9622e123 +size 382326 diff --git a/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/8dc52086-492d-48e3-b21b-9cfb917c190e_content_list.json b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/8dc52086-492d-48e3-b21b-9cfb917c190e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..86ccead6fdbd5657a1b4c6b02b45576082a6fbcc --- /dev/null +++ b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/8dc52086-492d-48e3-b21b-9cfb917c190e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f3893bb38c6bb393ab45b85395e3bcd56274fa026f267c8c97f165be5a6d80f +size 90575 diff --git a/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/8dc52086-492d-48e3-b21b-9cfb917c190e_model.json b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/8dc52086-492d-48e3-b21b-9cfb917c190e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..019859e8131122c3b64c4d8cdb41e54d9afa6357 --- /dev/null +++ b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/8dc52086-492d-48e3-b21b-9cfb917c190e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70b829a46280ecf3088c881ef657d99e37b1a9c5ef31455c79207c99a7de4519 +size 118473 diff --git a/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/8dc52086-492d-48e3-b21b-9cfb917c190e_origin.pdf b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/8dc52086-492d-48e3-b21b-9cfb917c190e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a115df6ff35d22c2f124bd494940bf90fb8ac55 --- /dev/null +++ b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/8dc52086-492d-48e3-b21b-9cfb917c190e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c0774e4b429341e1321c46b05b39dc1796859d73a9d57d00abb364fcb2f8165 +size 2311248 diff --git a/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/full.md b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..261faf3622f68e2914de4d93a1b2e306f562018b --- /dev/null +++ b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/full.md @@ -0,0 +1,328 @@ +# Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning + +Huiyi Wang $^{1,2}$ , Haodong Lu $^{1}$ , Lina Yao $^{2,1}$ , Dong Gong $^{1*}$ + +1University of New South Wales, 2CSIRO's Data61 + +{huiyi.wang, haodong.lu, dong.gong}@unsw.edu.au; lina.yao@data61.csiro.au + +# Abstract + +Continual learning (CL) aims to continually accumulate knowledge from a non-stationary data stream without catastrophic forgetting of learned knowledge, requiring a balance between stability and adaptability. Relying on the generalizable representation in pre-trained models (PTMs), PTM-based CL methods perform effective continual adaptation on downstream tasks by adding learnable adapters or prompts upon the frozen PTMs. However, many existing PTM-based CL methods use restricted adaptation on a fixed set of these modules to avoid forgetting, suffering from limited CL ability. Periodically adding task-specific modules results in linear model growth rate and impaired knowledge reuse. We propose Self-Expansion of pre-trained models with Modularized Adaptation (SEMA), a novel approach to enhance the control of stability-plasticity balance in PTM-based CL. SEMA automatically decides to reuse or add adapter modules on demand in CL, depending on whether significant distribution shift that cannot be handled is detected at different representation levels. We design modular adapter consisting of a functional adapter and a representation descriptor. The representation descriptors are trained as a distribution shift indicator and used to trigger self-expansion signals. For better composing the adapters, an expandable weighting router is learned jointly for mixture of adapter outputs. SEMA enables better knowledge reuse and sub-linear expansion rate. Extensive experiments demonstrate the effectiveness of the proposed self-expansion method, achieving state-of-the-art performance compared to PTM-based CL methods without memory rehearsal. Code is available at https://github.com/huiyiwang01/SEMA-CL. + +# 1. Introduction + +With the development of deep neural networks, deep learning models have achieved significant success in various fields, + +such as computer vision [15, 24]. However, real-world scenarios often present learning tasks in a dynamic data stream with non-stationary distributions [50]. Considering the need for efficient model updating and restricted budgets on storage and computation [35], it is not guaranteed to store all the historical data and repeatedly re-train the model. Continual learning (CL) is investigated to learn incrementally and accumulate knowledge efficiently from the non-stationary data stream without catastrophic forgetting [46, 54] of previously learned knowledge [14, 59, 65, 71]. It requires CL approaches to achieve a balance between knowledge expansion (i.e., plasticity) and knowledge retention (i.e., stability) [22, 55, 71]. Many CL approaches have been studied to tackle the challenge relying on different strategies, such as experience replay (ER) [7, 8, 77], regularization on parameters or representations [6, 39, 77], and architectures with modularization or isolation [55, 66, 70, 75, 78]. + +Given the progress in the pre-trained models (PTMs) with reliable representation, recent works explore the potential of using PTMs, such as Vision Transformer (ViT) [15], as the starting point of CL, unlike the "training-from-scratch" paradigm. PTM-based CL approaches [73, 74] usually keep the PTMs frozen to enable stable representation and alleviate forgetting. The PTMs are continually adapted to downstream tasks through parameter-efficient fine-tuning with newly expanded parameters as prompts and/or adapters [13, 51, 68, 73, 74, 83, 90, 91]. On the other hand, some methods enable continual fine-tuning of PTMs on real-world downstream tasks arriving in a streaming manner. Many PTM-based CL approaches mainly add and learn a fixed set/pool of prompts [33, 93] or adapters [9] shared by all downstream tasks in the stream [51, 73, 74, 90]. To alleviate forgetting caused by the interference on the newly added parameters, they restrict the parameter updating only on the first task seen in stream [51, 90] or use various regularization on the shared parameters [73, 74]. Their continual adaptation potentials are limited by the fixed and static size of prompt and adapter parameters. Some recent methods expand the PTMs with task-specific parameters to produce input-conditioned prompts [68] or ensemble of adapters [92]. + +![](images/f8418ffb85c472868a74d95c2dadd96751a90057f6ba3ba07f55142d304abefc.jpg) +Figure 1. An example of the self-expansion process. (a) The PTM (i.e., ViT) with $L$ transformer layers at the initial point of CL. (b) The first session adaptation – at Task 1, a modular adapter and a (dummy) router is added and trained in each transformer layer. (c) The modular adapters and routers added in the previous step (Task 1) are frozen to alleviate forgetting. When Task 2 arrives, only the representation descriptor in the $L$ -th layer detects feature distribution shift (with novel patterns) and generates expansion signal. A new module is added and trained in the $L$ -th layer, with the router expanded and updated. (d) At Task 3, new adapter is added at $L - 1$ -th layer after the expansion signal is firstly generated. In this demo example, the expansion is triggered and produced again in the $L$ -th layer, following the expansion in the $L - 1$ -th layer. If a task does not trigger expansion signal in any layer (implying no significantly different pattern), expansion would not happen, and existing adapters would be reused. More discussions are in Appendix A.1. + +The task-specifically added modules can help reduce the interference but cause a linearly-scaled model size (w.r.t. number of tasks) and restrained knowledge sharing and reuse. + +Considering that the PTM and the newly added parameters in expansion can provide a stable representation and a knowledge extension mechanism for CL, respectively, we focus on how to further enhance the control of the stability-plasticity balance during continual expansion. Although task-specific expansion of PTMs [68, 92] directly reduces the cross-task conflicts, it causes undesired linear scaling of model size and may impair knowledge transfer/reuse [55, 65, 70]. To address these issues, we propose SEMA, a CL approach with Self-Expansion of pre-trained models with Modularized Adaptation. It automatically expands PTMs with modularized adapters on demand and continually learns them to accommodate the distribution shifts without overwriting previously learned knowledge. Unlike existing methods that expand PTMs with a pre-defined fixed-size pool [51, 74, 83, 90] or task-specific components [68, 73, 92], we design modularized adapters to enable SEMA automatically decide when and where (i.e., which layer) to expand the PTM (i.e., a pre-trained ViT) on demand for tackling new requirements with sufficient and flexible plasticity, as shown in Fig. 1. The model continually learns how to compose the learned adapters. With the enhanced knowledge transfer and reuse, SEMA can thus perform better by only expanding the parameter size sub-linearly. + +We introduce modular/modularized adapters that can be + +identified and reused to solve new tasks, selectively adding and learning a subset of new adapters for unseen knowledge. Specifically, we design the modular adapter as a pair of a functional adapter and a representation descriptor (RD). The functional adapters produce specific feature representations to adapt to the different requirements of different tasks. The RDs are jointly trained to capture the feature distribution relevant to the coupled adapter at corresponding layers, serving as indicators of distribution shift at the representation level of intermediate layers. SEMA can use the representation descriptors for self-expansion – a new modular adapter is added at a specific layer if and only if all the representation descriptors indicate the input feature as a unseen pattern; otherwise, the existing frozen adapters are reused, resulting in sub-linear expansion. They can be implemented as a model with density estimation or novelty detection ability, such as autoencoder (AE) [27] or variational autoencoder (VAE) [38]. The module expansion at each layer can happen flexibly to supplement existing representation space, leading to sufficient plasticity. The on-demand expansion strategy strengthens the knowledge transfer and reuse, compared to the task-specific expansion [68, 92]. For example, cat images and dog images have more shared features than food images; the SEMA model trained only on cat images tends to expand more new adapters when training on food images than on dog images. To effectively compose the adapters, we design an expandable weighting router to produce layerwise weighted mixture of the adapters in a form of mixture + +of experts (MoE), which are expanded and learned in the self-expansion process. Despite the RDs may be used for adapter assignment by hard selection, the learned soft mixture router can perform more effectively (Appendix C.3). We summarize our contributions as follows: + +- We propose a novel continual learning approach via self-expansion of PTMs with modularized adapters, i.e. SEMA. In CL, it automatically determines the expansion necessity and location for new adapters, adding them at specific layers to accommodate new patterns in samples. The model enhances the control of stability-plasticity trade-off through adapter reuse and flexible expansion performed only on demand. SEMA enables sub-linear expansion and operates without the need for rehearsal. +- To achieve SEMA, we introduce modular adapters comprising a functional adapter and a representation descriptor. The representation descriptor maintains the distribution of pertinent input features, serving as a local novel pattern detector for expansion during training. The expandable weighting router is maintained simultaneously for composing the adapters via weighted mixture. +- Extensive experiments are conducted to validate the effectiveness and analyze the behavior of the proposed method, which demonstrates the model's ability on alleviating forgetting and knowledge transfer as well as the plausibility of the automated process. + +# 2. Related Work + +Continual Learning (CL). The mainstream taxonomy classifies continual learning methods into three categories: replay-based methods, regularization-based methods and architecture-based methods [14, 71]. Replay-based methods aim to alleviate catastrophic forgetting by retaining a memory buffer to store the information from old tasks for future replay [6, 8, 48, 59]. With simple intuition and effectiveness in preventing forgetting, these methods are limited by the size of the memory buffer and may also raise privacy concerns. An alternative approach is to implicitly maintain a generative model for producing pseudo-samples with similar distribution to old classes [11, 37, 60, 61, 67]. Regularization-based methods penalize significant changes to important parameters for seen tasks [2, 4, 39, 53, 84, 85], or consolidate the knowledge learnt from previous tasks with knowledge distillation [28, 41, 46, 88]. Instead of using all available parameters for all tasks, architecture-based methods allocate a subset of parameters dedicated to each task, which can be performed with task masking [36, 49, 66, 75] or dynamic architecture [3, 31, 43, 44, 55, 70, 78-81]. These methods tend to achieve optimal performance with less forgetting as isolating parameters and growing capacity for novel tasks reduce task interference during training, however, they are mostly restricted to simple applications due to the complex model design. + +Parameter-Efficient Fine-Tuning (PEFT). Parameter-efficient fine-tuning methods train a small set of additional parameters rather than the entire pre-trained model, which reduces the demands placed upon computational resources. Prompt tuning modifies input tokens/Prefixes via learnable prompts [33, 45]. LoRA [30] injects low-rank matrices to approximate weight updates and avoids additional inference latency via re-parameterization, which has been further utilized as experts with mixture modeling in recent works [16, 21, 72, 76]. Adapters introduced by [29], along with its variants [9, 34], insert lightweight learnable modules into the transformer. To enhance the efficacy of adapter learning, [23] investigates different insertion forms, and [12, 57, 63] explores the potential of adapter compositions. + +PTM-based CL. Recent works adopt PTMs, such as ViT and CLIP, as the backbone in the CL system to exploit its robust representational ability and enable further adaptation on downstream tasks [32, 62, 87, 89]. PTM can serve as a feature extractor for prototypes, which can be used for classification with distance measurement [51, 52, 56, 90]. PEFT techniques are also widely used to adapt PTMs in CL, including adaptation and prompting. L2P [74] and DualPrompt [73] apply a pool of prompts in CL through visual prompt tuning [33]. The prompt learning process is further improved by [68] with an attention mechanism and input-conditioned weights. ConvPrompt [62] adds parameter per task using linguistic knowledge from a large language model. Similar to prompt tuning in CL, some works also explore the use of a fixed set of adapters [13, 17, 82] or task-oriented expansion [47, 92] for better transfer of ViT to downstream CL tasks. [19] builds a unified framework incorporating both prompt and adapter-based methods. [10] adds experts in the pre-training of large language models (LLMs). + +# 3. Methodology + +# 3.1. Problem Definition + +Continual learning constructs a scenario where the model is required to learn from sequentially arriving tasks [14]. Consider a sequence of $T$ tasks $(\mathcal{D}^1,\mathcal{D}^2,\dots,\mathcal{D}^T)$ with distribution shift, where $\mathcal{D}^t = \{(x_i^t,y_i^t)\}_{i = 1}^{n_t}$ is the dataset containing $n_t$ data samples for the $t$ -th task. Only the training samples from $\mathcal{D}^t$ are accessible while seeing the $t$ -th task [74], if without additional ER process [8]. In a typical class-incremental learning (CIL) scenario [14], the classes in different tasks are non-overlapping, specifically, with the label space of the $t$ -th task denoted by $Y_{t}$ , $Y_{t}\cap Y_{t^{\prime}} = \emptyset$ for $t\neq t'$ . Let $F_{\theta}:X\to Y$ (with $X$ and $Y$ denoting the domain of input and label) be a model parameterized with $\theta$ . The goal of CL is to learn one model $F_{\theta}$ that can minimize the objective on each task $t$ in the stream: $\mathbb{E}_{(x,y)\in D^t}\mathcal{L}_{\mathrm{CE}}(F_\theta (x),y)$ , where $\mathcal{L}_{\mathrm{CE}}(\cdot ,\cdot)$ denotes the cross entropy loss in CIL. + +![](images/73dcfb26dfb50929c2cdeaae432f31b42e20d5b96a896a3519a07035899ae4d0.jpg) +Figure 2. Overview of the model architecture. (a) shows the structure of expandable adapter modules with adapters, RDs and router. (b) shows the scenario where expansion is triggered by representations with distribution different to previous tasks, estimated by RD. RDs are trained to align with the feature distribution of the corresponding task via only $\mathcal{L}_{\mathrm{RD}}$ , unaffected by gradients from the classification loss. (c) shows the scenario where incoming distribution can be handled by previously added modules, resulting in no expansion and adapter reuse. + +# 3.2. Overview + +We propose a PTM-based CL approach (i.e., SEMA) with a self-expansion mechanism to automatically add modularized adapters at arbitrary layers of the PTM (i.e., a pre-trained ViT with frozen parameters) on demand for handling automatically detected novel patterns in CL task stream, as shown in Fig. 1 and 2. The proposed method simultaneously learns a weighted mixture router for composing the adapters for different inputs. The design enhances the balance of knowledge transfer/reuse and plasticity for handling novelty, with only sub-linear expansion rate [5, 55]. + +To achieve the modularized design of SEMA, we introduce the modular adapters containing a pair of functional adapter $f_{\phi}(\cdot)$ and representation descriptor $g_{\varphi}(\cdot)$ , as defined in Sec. 3.3. Each added functional adapter works as a branch of a specific layer of the pre-trained transformer; and the representation descriptor indicates the feature distribution that can be handled by the paired $f_{\phi}(\cdot)$ . In CL, when new tasks arrive, $g_{\varphi}(\cdot)$ 's of the already-added adapters are used to detect novel feature patterns layer-by-layer. Only when novel pattern (i.e., representation-level distribution shifts) are detected, new adapters, i.e., pairs of $(f_{\phi}(\cdot), g_{\varphi}(\cdot))$ , are added and trained. After trained sufficiently, the adapters are kept frozen to alleviate forgetting and can be reused in future tasks. The details of the self-expansion strategy are in Sec. 3.6. At each layer of the PTM, an expandable weighting router is continually maintained and updated for composing the adapters via weighted mixture, as introduced in Sec. 3.4. When no adapters are added, the existing frozen adapters are retrieved and reused. + +# 3.3. Representation-Aware Modular Adapter + +The modular adapter $(f_{\phi}(\cdot), g_{\varphi}(\cdot))$ is designed as a pair of functional adapter $f_{\phi}(\cdot)$ and a representation descriptor + +$g_{\varphi}(\cdot)$ , which enables the module to be aware of the distribution of the local representation. One or more adapters can be added at arbitrary blocks/layers of the transformer. + +Functional adapter. In a (pre-trained) ViT, there are $L$ layers of transformer blocks, where each of them mainly contains a multi-head self-attention (MHSA) module and a multi-layer perceptron (MLP) module [15], as shown in Fig. 2. We keep all the parameters in the ViT frozen and perform adaptation through the learnable parameters in the continually added adapters. As a commonly used solution [9, 90], the functional adapter with learnable parameters is added as a side branch of the MLP in any layer of the ViT. + +Let $\mathbf{x}^l\in \mathbb{R}^d$ denote the feature input of the MLP at $l$ -th layer/block of ViT. In the proposed method, there can be different numbers (i.e., $K^{l}$ ) of adapters added at each layer through the self-expansion process. The $k$ -th functional adapter at $l$ -th layer is denoted as $f_{\phi_k^l}(\cdot)$ . Each $f_{\phi_k^l}(\cdot)$ takes $\mathbf{x}^l$ as input to bridge the representation gap between the pretrained model and the downstream tasks. By default, we implement $f_{\phi_k^l}(\cdot)$ as a lightweight adapter [9] containing a down-projection layer with parameters $\mathbf{W}_{\mathrm{down},k}^l\in \mathbb{R}^{d\times r}$ , an up-projection layer with parameters $\mathbf{W}_{\mathrm{up},k}^l\in \mathbb{R}^{r\times d}$ , and a non-linear ReLU activation [1] in between. By taking $\mathbf{x}^l$ as input, the output of each functional adapter is formulated as + +$$ +f _ {\phi_ {k} ^ {l}} \left(\mathbf {x} ^ {l}\right) = \operatorname {R e L U} \left(\mathbf {x} ^ {l} \cdot \mathbf {W} _ {\text {d o w n}, k} ^ {l}\right) \cdot \mathbf {W} _ {\text {u p}, k} ^ {l}, \tag {1} +$$ + +where $\phi_k^l\equiv \{\mathbf{W}_{\mathrm{up},k}^l,\mathbf{W}_{\mathrm{down},k}^l\}$ and $\mathbf{x}^l$ is treated as row vector for notation simplicity. If there is only one adapter at the $l$ -th layer (i.e., $K^{l} = 1$ ), the output representation of the MLP is adjusted as $\mathbf{x}_{\mathrm{out}}^l = \mathrm{MLP}(\mathbf{x}^l) + f_{\phi_k^l}(\mathbf{x}^l)$ . SEMA can continually expand the model with more than one adapters if needed. The number of adapters at each layer is automatically determined on demand, with a rate that is sub-linear w.r.t. number of tasks. Although similar adapter formulation + +have been used to handle CL, they only perform adaptation on the first task using only one adapter [51, 90] or periodically expand the PTM using task-specific adapters linearly [92]. In addition to Eq. 1, the functional adapters can also be implemented as other forms, such as LoRA [30], as discussed in Sec. 4.3. + +Representation descriptor. The representation descriptor (RD) $g_{\varphi_k^l}(\cdot)$ is paired with the functional descriptor $f_{\phi_k^l}(\cdot)$ to capture the characteristics of the local representation. It is designed and trained to indicate what kind of input representation can be handled by the corresponding functional adapter at each specific layer. Representation descriptors can be implemented as any model with density estimation or novelty detection ability. For simplicity, we implement them as AE [27], containing an encoder and a decoder. When a new pair of modular adapter is added at layer $l$ , the RD $g_{\varphi_k^l}(\cdot)$ is trained by minimizing the reconstruction loss on all the features fed to $f_{\phi_k^l}(\cdot)$ , i.e., $\mathcal{X}_k^l$ : + +$$ +\mathcal {L} _ {\mathrm {R D}, k} ^ {l} (x) = \sum_ {\mathbf {x} \in \mathcal {X} _ {k} ^ {l}} | | \mathbf {x} - g _ {\varphi_ {k} ^ {l}} (\mathbf {x}) | | _ {2} ^ {2}. \tag {2} +$$ + +In our expansion strategy (in Sec. 3.6), when a new task $t$ arrives, at each $l$ -th layer, if all existing RDs detect significantly novel distributions (based on the $z$ -score of reconstruction errors), the expansion signal is triggered. $f_{\phi_k^l}(\cdot)$ and $g_{\varphi_k^l}(\cdot)$ are trained on this task $t$ and then kept frozen in the future. $\mathcal{X}_k^l$ represents the input feature $\mathbf{x}^l$ of all the samples in this new expansion-triggering task $t$ . + +# 3.4. Expandable Weighting Router for Mixture Usage ofAdapters + +By definition, the representation descriptor can be used to compose the adapters, as in similar modular networks. However, it heavily relies on the statistics of similar inputs in a batch [55] and can be unreliable for individual inputs. We thus directly maintain and learn an expandable weighting router for a weighted mixture of the functional adapters. + +For any $l$ -th layer with $K^l$ adapters, the routing function is defined as $h_{\psi^l}(\cdot):\mathbb{R}^d\to \mathbb{R}^{K^l}$ . Similar to [16], we implement $h_{\psi^l}(\cdot)$ as a linear mapping function followed by a softmax operation $\mathbf{w}^l = h_{\psi^l}(\mathbf{x}^l)\equiv \mathrm{softmax}(\mathbf{x}^l\cdot \mathbf{W}_{\mathrm{mix}}^l)$ , where $\mathbf{W}_{\mathrm{mix}}^l\in \mathbb{R}^{d\times K^l}$ is the parameter of $\psi^l$ . As shown in Fig. 2, the weights $\mathbf{w}^l\in \mathbb{R}^{K^l}$ can produce the mixture of the added functional adapters to produce the output representation of the MLP in the transformer: + +$$ +\mathbf {x} _ {\text {o u t}} ^ {l} = \operatorname {M L P} \left(\mathbf {x} ^ {l}\right) + \sum_ {k = 1} ^ {K ^ {l}} w _ {k} ^ {l} \cdot f _ {\phi_ {k} ^ {l}} \left(\mathbf {x} ^ {l}\right). \tag {3} +$$ + +When new adapter is added at any layer $l$ , the router $h_{\psi^l}(\cdot)$ i.e., $\mathbf{W}_{\mathrm{mix}}^{l}$ , is expanded for producing weights with one more dimension. The expanded router is trained together with the added adapters. While expanding the router, the + +parameters corresponding to the existing adapters remain frozen and only the newly added ones (i.e., a newly added column in $\mathbf{W}_{\mathrm{mix}}^l$ ) are trained. This approach, similar to the common practice for training classification heads in CL [47, 68], controls and restricts forgetting in the expandable router (shown in Fig. 5), though it cannot fully eliminate it. + +# 3.5. Continual Learning Objective of SEMA + +In SEMA, the model $F_{\theta}(\cdot)$ for solving the tasks consists of learnable parameters from the functional adapters and router with learnable parameters, i.e., $\{\phi_k^l\}$ and $\{\psi^l\}$ . The learnable parameters are dynamically added and learned. The representation descriptors are learned jointly for maintaining a state of the local representation. The overall objective in SEMA optimizes all these parameters: + +$$ +\begin{array}{l} \min _ {\left\{\phi_ {k} ^ {l} \right\}, \left\{\psi^ {l} \right\}, \left\{\varphi_ {k} ^ {l} \right\}} \sum_ {t = 1} ^ {T} \mathbb {E} _ {(x, y) \in D ^ {t}} \left[ \mathcal {L} _ {\mathrm {C E}} \left(F _ {\left\{\phi_ {k} ^ {l} \right\}, \left\{\psi^ {l} \right\}} (x), y\right) \right. \\ \left. + \sum_ {l = 1} ^ {L} \sum_ {k = 1} ^ {K ^ {l}} \mathcal {L} _ {\mathrm {R D}, k} ^ {l} \left(x; \varphi_ {k} ^ {l}\right) \right]. \tag {4} \\ \end{array} +$$ + +Learning of modular adapters is executed only when new modules are added. The learned modules are kept frozen to prevent forgetting. Optimization of RDs can be parallel to other parameters. If no module is added in a specific task due to no significant pattern being identified by RDs, the existing modules can be reused without training. + +# 3.6. Self-Expansion Strategy + +The RDs provide the capacity to decide when and where to expand the model. We designed a more specific strategy to achieve the reliable self-expansion in the CL task stream. + +Task-oriented expansion. The expansion may occur at any time as new samples are seen during training. To incorporate the task identification prior knowledge in CL, especially CIL, we improve parameter efficiency and expansion stability with task-oriented expansion. We restrict the addition to at most one adapter per layer for each task. When a new task $t$ arrives, the method scans all samples in the first epoch to decide whether to expand the model. If the expansion signal is triggered, only one adapter is added and then trained for the whole task; otherwise, the task $t$ data can reuse learned modules and the learning process moves to the next task. + +$z$ -score based expansion signal. When scanning through the new task data, an expansion signal at layer $l$ is triggered when significantly new patterns are identified. It reflects that a $\mathbf{x}^l$ is out of the scope of all RDs, i.e., reconstruction error is high with each $g_{\varphi_k^l}(\mathbf{x})$ [20], as illustrated in Fig. 4. However, it is impractical to directly use reconstruction error due to the perturbation and heterogeneous characteristics of each task and adapter. We thus compute and maintain the running statistics $\mu_k^l$ and standard deviation $\sigma_k^l$ of reconstruction error on all relevant inputs used in training. Given any $x^l$ in the + +
MethodCIFAR-1005-Task IN-R10-Task IN-R20-Task IN-RImageNet-AVTAB
AANAANAANAANAANAAN
FT Adapter47.8830.953.9141.2345.3130.9338.5124.2229.7817.6459.9843.50
L2P84.7777.8777.4073.5966.9762.7270.6762.9047.1638.4881.1980.83
DualPrompt86.6080.4376.3972.2972.8366.7562.3361.9759.5450.2382.8979.79
CODA-P91.5586.1181.6376.9881.1175.2575.0070.0247.2935.0279.8881.58
SimpleCIL82.3176.2165.8361.3167.0961.3567.5961.3560.0549.2485.2983.61
ADAM90.5585.6279.9174.2579.1173.1575.8469.1060.1549.2485.2983.61
InfLoRA90.5185.0578.5872.5881.3975.3278.8772.6059.7146.2188.9087.63
SEMA91.3786.9884.7579.7883.5678.0081.7574.5364.5353.3291.2689.64
+ +Table 1. Comparison with ViT-based CL methods in CIL. All models adopt ViT-B/16-IN1K as the backbone. + +![](images/228f1470dfa8745bfc2d10baad03fa66b5a2438e1ef78f01e8262b0063d92a0e.jpg) +Figure 3. Incremental performance of different methods on class-incremental learning benchmarks. + +![](images/38e57c5abc753fc667ac9a46264fbfbd33616f6889a4bc5eade2d6b7b1c0029f.jpg) + +![](images/0f1d2d92fe27f31449697582de12fd0d2931db7bf669971d5832781a39064add.jpg) + +![](images/c9094e685d5715ab62a14e9bb1d2f4604a8bfddcc43e2cb4d68bf2dd1bc692da.jpg) + +scanning process for a future task, the $z$ -score corresponding to each existing RD can be calculated as $z_k^l = (r_k^l - \mu_k^l) / \sigma_k^l$ with $r_k^l$ as reconstruction error. If all $z_k^l$ 's for $k = 1, \dots, K^l$ are larger than a threshold, the expansion signal is triggered. Considering that the $z$ -score has normalized out perturbation and scale, the process can be very robust to the threshold setting, as shown in Sec. 4.3. + +Multi-layer expansion. We facilitate self-expansion across multiple layers through distinct decision processes. Upon encountering a new task, self-expansion operations are executed sequentially from shallow layers to deeper layers. As new adapters are introduced at shallow levels, training ensures representations are aligned accordingly. Subsequently, the model determines whether to continue expanding into subsequent layers. The adaptable multi-layer expansion facilitates the accommodation of various distribution shifts and enables flexible inter-class knowledge sharing [18, 42]. + +# 4. Experiments + +# 4.1. Setting and Implementation Details + +Datasets. Experiments are conducted on common datasets used for pre-trained ViT-based CIL: CIFAR-100 [40], ImageNet-R (IN-R) [25], ImageNet-A [26] and VTAB [86]. Baselines. We validate our method by comparing with PTM-based rehearsal-free CL approaches using similar backbone (e.g., ViT) and methodology, including fully fine-tuning of the adapter, L2P [74], DualPrompt [73], CODA-P [68], SimpleCIL [90], ADAM with Adapter [90] and InfLoRA [47]. + +Training details. We use the commonly used ViT-B/16 model [15] weights pre-trained on ImageNet-1K [64] as the PTM weights. We also conducted experiments with other + +pre-trained weights and left discussions in Appendix C.1. The batch size is set to 32. SGD is used as the optimizer with the initial learning rate set to 0.005 and 0.01 for adapters and RDs, respectively, decaying with cosine annealing. The hidden dimension of adapter is 16. In experiments, by default, we enable self-expansion in the last three transformer layers for simplicity without losing generality. + +# 4.2. Experimental Results + +We validate the proposed method by comparing with previous related state-of-the-art methods and reporting the average accuracy of all tasks $\mathcal{A}_N$ [7] and average incremental accuracy $\bar{\mathcal{A}}$ [59] metrics in Tab. 1. It shows that our method performs better than other related methods in terms of the average accuracy at the last step $\mathcal{A}_N$ , which reflects the final goal of CL. Fig. 3 shows the variation in accuracy during the continual learning process. It shows the consistently superior performance of SEMA in the process. Although most previous approaches exhibit strong performance on CIFAR-100, the proposed methods show more improvements on datasets containing adversarial samples similar to those found in ImageNet, due to its better stability-plasticity balance. + +# 4.3. Ablation Studies and Analyses + +Ablation studies on module expansion and adapter composing. We conduct ablation studies to demonstrate the effectiveness of the self-expansion process and investigate the influence of different adapter composing strategies, with the results reported in Tab. 2. We first conduct an experiment by removing the self-expansion process and only keeping the first-session adaptation (No Exp.), which is similar to ADAM [90] with slight difference on implementation. The + +results show that the self-expansion can work reliably to continually improve the adaptation results. + +
MethodImageNet-AVTAB
AANAAN
SEMA64.5353.3291.2689.64
No Exp.61.2049.9086.2183.66
Avg. W.56.8844.3190.8489.14
Rand. W.62.9549.7788.8785.17
Top-1 Sel.62.0050.5690.8388.61
Rand. Sel.61.7050.3690.8288.51
Top-1 Sel. Inf.61.9650.3690.9588.84
+ +Table 2. Ablation studies on adapter expansion and composing. + +To demonstrate the benefits of the weighted mixture routing, we investigate several variants of SEMA with different adapter composing strategies. Firstly, we study two variants with a soft mixture of adapters relying average weighting (Avg. W.) and random weighting (Rand. W.), respectively. Tab. 2 shows that the expandable weighting router learns an effective weighting function. We further study the variants that perform routing by selecting only a single adapter indicated by the highest value from the learned weighting router (Top-1 Sel.) or through random drawing (Rand. Sel.). Additionally, we evaluate SEMA trained with mixture routing, using an inference strategy that selects only the adapter with the highest weight (Top-1 Sel. Inf.). The results show that the weighted soft mixture of the learned adapters works more effectively by encouraging the better usage of the learned adapters. More experiments about adapter composing using representation descriptor are in Appendix C.3. + +Analysis on dynamic expansion process. To demonstrate how the representation descriptors are learned and how they work for self-expansion in CL, we visualize the reconstruction error of each AE-based RD corresponding to each sample seen during training, i.e., their representation features at specific layer, in Fig. 4. For more intuitive visualization and simplified experiment, in this analysis, we restrict the automatic self-expansion only to the last layer of transformer. The analysis is conducted on VTAB dataset. In this case shown in Fig. 4, the reconstruction error of each RD decreases and converges after training on the corresponding task, after the RD is added for handling this task. When a new task arrives, the reconstruction errors for the existing RDs are calculated and used to detect novelty. The expansion signal is generated when significantly high reconstruction errors (scaled as $z$ -scores) are detected from all the previous RDs (in Task 2 and 3). In Task 4 and 5, all samples can be well covered by at least one previous RD, which implies no significant distribution shift is detected and results in no expansion. Note that the $z$ -score (i.e., a normalized version of reconstruction error) is used for expansion in SEMA. + +Analysis on adapter usage. Fig. 5 demonstrates the average adapter usage of each task from VTAB. This analysis is produced by restricting self-expansion to the last layer, as in + +![](images/202035dc2f98c78b8cc728de929d3d971305ea1392e9431264af61cf21f108db.jpg) +Figure 4. Reconstruction error during training to show the dynamic expansion process. Expansion occurs for Tasks 1, 2, and 3, while no expansion is triggered for Tasks 4 and 5 due to no detected distribution shift. + +![](images/7f3a7f8bb72a7e45d1e12a7216fc41cc68d2860a909c13d5e2658017ea08aca0.jpg) +Figure 5. Visualization of adapter usage on VTAB. Adapters 1, 2, and 3 are added and trained on Tasks 1, 2, and 3, respectively. Tasks 4 and 5 primarily reuse Adapters 1 and 3 due to similar feature distributions with Tasks 1 and 3. + +Fig. 4. Self-expansion is automatically produced for Task 1, 2 and 3. For tasks that triggered expansion, the adapters used are primarily those they were trained with, as shown in the figure. Task 4 and 5 share a similar selection pattern with the tasks they are similar with (Task 1 and 3 respectively), showing that added adapters are effectively reused for new tasks. More details are in Appendix C.3. + +Study of expansion threshold. We investigate the impact of the expansion threshold on accuracy and the number of added adapters using ImageNet-A and VTAB. Firstly, the results in Fig. 6 show that the proposed method is not sensitive to the setting of the threshold, benefiting from the $z$ -score-based expansion signal. Fig. 6b and 6d show how the threshold influences the number of added adapters (at each layer), displaying trends consistent with those in Fig. 6a and 6c. Fig. 6a and 6b show that a smaller expansion threshold leads to more frequent expansion, which could boost the performance at some level through more parameters. A threshold that is too large (e.g., values over 1.5) minimizes the chance for expansion, which may lead to insufficient adaptation. In SEMA, a proper expansion threshold within a wide range can lead to a balance between the performance gain and the parameter size. + +Analysis of multi-layer expansion. In Fig. 7, we explore the effects on accuracy by implementing expansion across varying numbers of layers, ranging from the last 2 layers (#11-#12) to the last 4 layers (#9-#12). Intuitively, allowing expansion in deeper layers enables better adaptation + +![](images/ac243a3db8668f26d803c8137c67b9846bd0744bc52082695e75bc1ab2710565.jpg) +(a) Accuracy + +![](images/c76c2ed81ccb56009a84657129ae005692fc008afd4783ef7f599d93a2b2b66e.jpg) +(b) Num. of adapters + +![](images/2344d5f8287dcacfbd673e3d7a56207a99b800187c53194590163ab5a3a6a70d.jpg) +(c) Accuracy + +![](images/0933e2ca463dd1a1b01340b00189388ef83f008669d6a4ef692af90ab57f2e36.jpg) +(d) Num. of adapters + +![](images/0e86cb75d56ad8b1b1111e540b6b8ce144de86e00622f975f39fb024584c33b3.jpg) +(a) Accuracy + +![](images/4f1a2533936a77bcd0ff2023af08b09e523f43d176687eb29252169624f84f27.jpg) +Figure 6. Analysis of the impact of expansion threshold with (a)(b) ImageNet-A and (c)(d) VTAB. (a) and (c) show that SEMA can produce good accuracy stably with slight variation w.r.t. varying expansion threshold. (b) and (d) report how the number of added adapters (on the specific Transformer layers #10, #11, #12) changes with the varying threshold values, corresponding to (a) and (c), respectively. The proposed method is insensitive to the threshold. Adding more adapters may lead to higher accuracy, a proper threshold can achieve a balance between performance and model size. +(b) Num. of adapters + +![](images/5b614a1d0d7ec020f4a9c5825ddcd99b5df4c51df0caaa047dc888ce8d584774.jpg) +(c) Accuracy + +![](images/4f9e68a3f2a13d81d4389eff2c06d168a9e0b7dddb11d714404129fad9b692df.jpg) +(d) Num. of adapters +Figure 7. Analysis of the effect of multi-layer expansion, with (a)(b) ImageNet-A and (c)(d) VTAB. By enabling automatic self-expansion on multiple transformer layers, SEMA can achieve better performance than restricting that on a single layer. + +to different tasks. However, as shown in Fig. 7b and Fig. 7d, permitting expansion in early transformer layers also increases the overall number of added adapters, without a significant boost in performance as earlier layers tend to behave similarly despite distribution shifts. Also, enforcing addition of too many adapters may cause difficulty in training, especially in early transformer layers. + +
MethodImageNet-AVTAB
AANAAN
Adapter[9]64.5353.3291.2689.64
LoRA[30]63.5052.6791.8588.53
Convpass[34]63.4851.7490.6888.62
+ +Table 3. Different adapter variants. + +![](images/3b396879ccc3474ef29de8ff3b324947ec52b9006c8b3fb59c60c7150ac1cd6e.jpg) +Figure 8. Analysis on added parameters (in Millions) during model deployment on ImageNet-A. + +Ablation studies on adapter variants. Apart from Adapter [9], we extend our evaluation to other variants, namely LoRA [30] and Convpass [34]. As shown in Tab. 3, our proposed approach is robust to the choice of adapter methods, showing the broad applicability and effectiveness of our dynamic expansion strategy across different adapter methods. + +Sub-linear growth of parameters. In Fig. 8, instead of expanding w.r.t. number of tasks, SEMA adds parameters at a sub-linear rate, showing the efficiency of the self-expansion mechanism. Further analysis is provided in Appendix C.2. + +# 5. Conclusion + +In this paper, we propose a novel self-expandable modularized adaptation approach for continual learning. SEMA learns to reuse and add modules in an automated manner without memory rehearsal. We incorporate an efficient expansion strategy with detection for feature distribution shifts in different layers of transformer-based models, successfully mitigating the forgetting problem of jointly using the fixed set of parameters. Experimental results demonstrate the outstanding performance of SEMA to datasets with different levels of distribution shifts. + +Limitations and future work. We perform the task-oriented expansion at most once per layer for each task considering the CIL characteristics and parameter efficiency. The design can be more flexible to enable fully online dynamic expansion, which could open possibility in better adaptation for data with intra-task diversity and enable online CL. Moreover, the expansion of SEMA is based on the distribution shift detection ability from RDs, which could be further enhanced by elevating the optimization of RDs and expansion protocol to a meta level with a closed loop. + +# References + +[1] Abien Fred Agarap. Deep learning using rectified linear units (relu). arxiv 2018. arXiv preprint arXiv:1803.08375, 1803.4 +[2] Hongjoon Ahn, Sungmin Cha, Donggyu Lee, and Taesup Moon. Uncertainty-based continual learning with adaptive regularization. Advances in neural information processing systems, 32, 2019. 3 +[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3366-3375, 2017. 3 +[4] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139-154, 2018. 3 +[5] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 39-48, 2016. 4 +[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 1, 3 +[7] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with A-GEM. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 1, 6 +[8] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486, 2019. 1,3 +[9] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 1, 3, 4, 8 +[10] Wuyang Chen, Yanqi Zhou, Nan Du, Yanping Huang, James Laudon, Zhifeng Chen, and Claire Cui. Lifelong language pretraining with distribution-specialized experts. In International Conference on Machine Learning, pages 5383-5395. PMLR, 2023. 3 +[11] WU Chenshen, L Herranz, LIU Xialei, et al. Memory replay gans: Learning to generate images from new categories without forgetting [c]. In The 32nd International Conference on Neural Information Processing Systems, Montréal, Canada, pages 5966-5976, 2018. 3 +[12] Alexandra Chronopoulou, Matthew E. Peters, Alexander Fraser, and Jesse Dodge. Adaptersoup: Weight averaging to improve generalization of pretrained language models. In Findings of the Association for Computational Linguistics: EACL 2023, Dubrovnik, Croatia, May 2-6, 2023, pages 2009-2018. Association for Computational Linguistics, 2023. 3 + +[13] Yawen Cui, Zitong Yu, Rizhao Cai, Xun Wang, Alex C Kot, and Li Liu. Generalized few-shot continual learning with contrastive mixture of adapters. arXiv preprint arXiv:2302.05936, 2023. 1, 3 +[14] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern analysis and machine intelligence, 44(7):3366-3385, 2021. 1, 3 +[15] 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 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 1, 4, 6 +[16] Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Jun Zhao, Wei Shen, Yuhao Zhou, Zhiheng Xi, Xiao Wang, Xiaoran Fan, et al. Loramoe: Revolutionizing mixture of experts for maintaining world knowledge in language model alignment. arXiv preprint arXiv:2312.09979, 2023. 3, 5 +[17] Beyza Ermis, Giovanni Zappella, Martin Wistuba, Aditya Rawal, and Cedric Archambeau. Memory efficient continual learning with transformers. Advances in Neural Information Processing Systems, 35:10629-10642, 2022. 3 +[18] Chongyang Gao, Kezhen Chen, Jinmeng Rao, Baochen Sun, Ruibo Liu, Daiyi Peng, Yawen Zhang, Xiaoyuan Guo, Jie Yang, and VS Subrahmanian. Higher layers need more lora experts, 2024. 6 +[19] Qiankun Gao, Chen Zhao, Yifan Sun, Teng Xi, Gang Zhang, Bernard Ghanem, and Jian Zhang. A unified continual learning framework with general parameter-efficient tuning. arXiv preprint arXiv:2303.10070, 2023. 3 +[20] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1705-1714, 2019. 5 +[21] Yunhao Gou, Zhili Liu, Kai Chen, Lanqing Hong, Hang Xu, Aoxue Li, Dit-Yan Yeung, James T Kwok, and Yu Zhang. Mixture of cluster-conditional lora experts for vision-language instruction tuning. arXiv preprint arXiv:2312.12379, 2023.3 +[22] Raia Hadsell, Dushyant Rao, Andrei A Rusu, and Razvan Pascanu. Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences, 24(12):1028-1040, 2020. 1 +[23] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 3 +[24] 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. 1 + +[25] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6 +[26] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6 +[27] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313 (5786):504-507, 2006. 2, 5 +[28] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Lifelong learning via progressive distillation and retrospection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 437-452, 2018. 3 +[29] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 3 +[30] 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, 5, 8 +[31] Ching-Yi Hung, Cheng-Hao Tu, Cheng-En Wu, Chien-Hung Chen, Yi-Ming Chan, and Chu-Song Chen. Compacting, picking and growing for unforgetting continual learning. Advances in Neural Information Processing Systems, 32, 2019. 3 +[32] Saurav Jha, Dong Gong, and Lina Yao. CLAP4CLIP: Continual learning with probabilistic finetuning for vision-language models. In Thirty-eighth Conference on Neural Information Processing Systems, 2024. 3 +[33] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 1, 3 +[34] Shibo Jie and Zhi-Hong Deng. Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039, 2022. 3, 8 +[35] Daniel Justus, John Brennan, Stephen Bonner, and Andrew Stephen McGough. Predicting the computational cost of deep learning models. In 2018 IEEE international conference on big data (Big Data), pages 3873-3882. IEEE, 2018. 1 +[36] Zixuan Ke, Bing Liu, Nianzu Ma, Hu Xu, and Lei Shu. Achieving forgetting prevention and knowledge transfer in continual learning. Advances in Neural Information Processing Systems, 34:22443-22456, 2021. 3 +[37] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. arXiv preprint arXiv:1711.10563, 2017. 3 +[38] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 2 +[39] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan. + +John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13): 3521-3526, 2017. 1, 3 +[40] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009. 6 +[41] Kibok Lee, Kimin Lee, Jinwoo Shin, and Honglak Lee. Overcoming catastrophic forgetting with unlabeled data in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 312–321, 2019. 3 +[42] Yoonho Lee, Annie S. Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, and Chelsea Finn. Surgical finetuning improves adaptation to distribution shifts. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 6 +[43] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pages 3925-3934. PMLR, 2019. 3 +[44] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In International Conference on Machine Learning, pages 3925-3934. PMLR, 2019. 3 +[45] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3 +[46] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1, 3 +[47] Yan-Shuo Liang and Wu-Jun Li. Inflora: Interference-free low-rank adaptation for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23638-23647, 2024. 3, 5, 6 +[48] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 3 +[49] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In Proceedings of the European conference on computer vision (ECCV), pages 67-82, 2018. 3 +[50] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, pages 109-165. Elsevier, 1989. 1 +[51] Mark McDonnell, Dong Gong, Amin Parvaneh, Ehsan Abbasnejad, and Anton van den Hengel. RanPAC: Random projections and pre-trained models for continual learning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 5, 6 +[52] Lu Mi, Hao Wang, Yonglong Tian, Hao He, and Nir N Shavit. Training-free uncertainty estimation for dense regression: Sensitivity as a surrogate. In Proceedings of the AAAI Con- + +ference on Artificial Intelligence, pages 10042-10050, 2022. 3 +[53] Cuong V Nguyen, Yingzhen Li, Thang D Bui, and Richard E Turner. Variational continual learning. arXiv preprint arXiv:1710.10628, 2017. 3 +[54] Cuong V Nguyen, Alessandro Achille, Michael Lam, Tal Hassner, Vijay Mahadevan, and Stefano Soatto. Toward understanding catastrophic forgetting in continual learning. arXiv preprint arXiv:1908.01091, 2019. 1 +[55] Oleksiy Ostapenko, Pau Rodriguez, Massimo Caccia, and Laurent Charlin. Continual learning via local module composition. Advances in Neural Information Processing Systems, 34:30298-30312, 2021. 1, 2, 3, 4, 5 +[56] Francesco Pelosin. Simpler is better: off-the-shelf continual learning through pretrained backbones. arXiv preprint arXiv:2205.01586, 2022. 3 +[57] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, EACL 2021, Online, April 19 - 23, 2021, pages 487-503. Association for Computational Linguistics, 2021. 3 +[58] 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. 7 +[59] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1, 3, 6 +[60] Matthew Riemer, Tim Klinger, Djallel Bouneffouf, and Michele Franceschini. Scalable recollections for continual lifelong learning. In Proceedings of the AAAI conference on artificial intelligence, pages 1352-1359, 2019. 3 +[61] Mohammad Rostami, Soheil Kolouri, and Praveen K Pilly. Complementary learning for overcoming catastrophic forgetting using experience replay. arXiv preprint arXiv:1903.04566, 2019. 3 +[62] Anurag Roy, Riddhiman Moulick, Vinay K Verma, Saptarshi Ghosh, and Abir Das. Convolutional prompting meets language models for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23616-23626, 2024. 3 +[63] Andreas Rücklé, Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, and Iryna Gurevych. Adaptetrop: On the efficiency of adapters in transformers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pages 7930-7946. Association for Computational Linguistics, 2021. 3 +[64] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, + +Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6 +[65] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International conference on machine learning, pages 4528-4537. PMLR, 2018. 1, 2 +[66] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International conference on machine learning, pages 4548-4557. PMLR, 2018. 1, 3 +[67] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 3 +[68] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 1, 2, 3, 5, 6 +[69] Hai-Long Sun, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Pilot: A pre-trained model-based continual learning toolbox. arXiv preprint arXiv:2309.07117, 2023. 1 +[70] Tom Veniat, Ludovic Denoyer, and Marc'Aurelio Ranzato. Efficient continual learning with modular networks and taskdriven priors. arXiv preprint arXiv:2012.12631, 2020. 1, 2, 3 +[71] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 3 +[72] Yaqing Wang, Sahaj Agarwal, Subhabrata Mukherjee, Xiaodong Liu, Jing Gao, Ahmed Hassan Awadallah, and Jianfeng Gao. AdaMix: Mixture-of-adaptations for parameter-efficient model tuning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5744-5760, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 3 +[73] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 1, 2, 3, 6 +[74] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 1, 2, 3, 6 +[75] Mitchell Wortsman, Vivek Ramanujan, Rosanne Liu, Aniruddha Kembhavi, Mohammad Rastegari, Jason Yosinski, and Ali Farhadi. Supermasks in superposition. Advances in Neural Information Processing Systems, 33:15173-15184, 2020. 1, 3 + +[76] Xun Wu, Shaohan Huang, and Furu Wei. MoLE: Mixture of IoRA experts. In The Twelfth International Conference on Learning Representations, 2024. 3 +[77] Qingsen Yan, Dong Gong, Yuhang Liu, Anton van den Hengel, and Javen Qinfeng Shi. Learning bayesian sparse networks with full experience replay for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 109-118, 2022. 1 +[78] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3014-3023, 2021. 1, 3 +[79] Fei Ye and Adrian G Bors. Task-free continual learning via online discrepancy distance learning. Advances in Neural Information Processing Systems, 35:23675-23688, 2022. +[80] Fei Ye and Adrian G Bors. Self-evolved dynamic expansion model for task-free continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22102-22112, 2023. +[81] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 3 +[82] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219-23230, 2024. 3 +[83] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In CVPR, 2024. 1, 2 +[84] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987-3995. PMLR, 2017. 3 +[85] Chen Zeno, Itay Golan, Elad Hoffer, and Daniel Soudry. Task agnostic continual learning using online variational bayes. arXiv preprint arXiv:1803.10123, 2018. 3 +[86] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 6 +[87] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19148-19158, 2023. 3 +[88] Junting Zhang, Jie Zhang, Shalini Ghosh, Dawei Li, Serafettin Tasci, Larry Heck, Heming Zhang, and C-C Jay Kuo. Class-incremental learning via deep model consolidation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1131-1140, 2020. 3 +[89] Zangwei Zheng, Mingyuan Ma, Kai Wang, Ziheng Qin, Xi-angyu Yue, and Yang You. Preventing zero-shot transfer degradation in continual learning of vision-language models. + +In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19125-19136, 2023. 3 +[90] Da-Wei Zhou, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pre-trained models: Generalizability and adaptivity are all you need, 2023. 1, 2, 3, 4, 5, 6 +[91] Da-Wei Zhou, Yuanhan Zhang, Jingyi Ning, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Learning without forgetting for vision-language models, 2023. 1 +[92] Da-Wei Zhou, Hai-Long Sun, Han-Jia Ye, and De-Chuan Zhan. Expandable subspace ensemble for pre-trained model-based class-incremental learning. In CVPR, 2024. 1, 2, 3, 5, 6 +[93] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1 \ No newline at end of file diff --git a/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/images.zip b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f7973c036f2bf292fe31bd2241bf3b039c4257f0 --- /dev/null +++ b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfb60cfd6456bb72603caeacfc9ca05e3189c9e2c9290066fb9533e9fcf1cce2 +size 545161 diff --git a/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/layout.json b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0953c49007f770b0e0e8d5e15f69bb8dea913f87 --- /dev/null +++ b/CVPR/2025/Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a20a3d151b7eefc1d03cd44583f3cdbcf1c460e55801f14faf1bedeb9ecc4fe2 +size 466860 diff --git a/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/1ec491d6-a0a0-457b-bfa7-480f4910c1c2_content_list.json b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/1ec491d6-a0a0-457b-bfa7-480f4910c1c2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..380ff609d4da9aa0b91a51be249b5b36c254ae1e --- /dev/null +++ b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/1ec491d6-a0a0-457b-bfa7-480f4910c1c2_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:420b146154577e3ea26157abfe39095103366e0c7f99e216c055be3bebea8d00 +size 83247 diff --git a/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/1ec491d6-a0a0-457b-bfa7-480f4910c1c2_model.json b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/1ec491d6-a0a0-457b-bfa7-480f4910c1c2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ce160f10b8c0e52f6284825ac5e83de40928c8bd --- /dev/null +++ b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/1ec491d6-a0a0-457b-bfa7-480f4910c1c2_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:502ae76ce45fce75b3d802ce7e5229d9fc0825e2e644e7f42a418c3305838a48 +size 100863 diff --git a/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/1ec491d6-a0a0-457b-bfa7-480f4910c1c2_origin.pdf b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/1ec491d6-a0a0-457b-bfa7-480f4910c1c2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c2d4e4383a0dc112abd67d70ea1e3efb19f9551b --- /dev/null +++ b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/1ec491d6-a0a0-457b-bfa7-480f4910c1c2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1766a88ea6566608dca446d1008bd59f53e59707a580488051241e5da18678b1 +size 9912652 diff --git a/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/full.md b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..73640b9b9cadeb7a434328f1a7b9ddf4e79474cc --- /dev/null +++ b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/full.md @@ -0,0 +1,398 @@ +# Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model + +Jian Zhu, He Wang, Yang Xu\* Zebin Wu\* Zhihui Wei +School of Computer Science and Engineering, +Nanjing University of Science and Technology, Nanjing, China +{123106010719, he_wang, xuyangth90, wuzb, gswei}@njust.edu.cn + +# Abstract + +Hyperspectral and multispectral image (HSI-MSI) fusion involves combining a low-resolution hyperspectral image (LR-HSI) with a high-resolution multispectral image (HR-MSI) to generate a high-resolution hyperspectral image (HR-HSI). Most deep learning-based methods for HSI-MSI fusion rely on large amounts of hyperspectral data for supervised training, which is often scarce in practical applications. In this paper, we propose a self-learning Adaptive Residual Guided Subspace Diffusion Model (ARGS-Diff), which only utilizes the observed images without any extra training data. Specifically, as the LR-HSI contains spectral information and the HR-MSI contains spatial information, we design two lightweight spectral and spatial diffusion models to separately learn the spectral and spatial distributions from them. Then, we use these two models to reconstruct HR-HSI from two low-dimensional components, i.e., the spectral basis and the reduced coefficient, during the reverse diffusion process. Furthermore, we introduce an Adaptive Residual Guided Module (ARGM), which refines the two components through a residual guided function at each sampling step, thereby stabilizing the sampling process. Extensive experimental results demonstrate that ARGS-Diff outperforms existing state-of-the-art methods in terms of both performance and computational efficiency in the field of HSI-MSI fusion. + +# 1. Introduction + +Hyperspectral images (HSI) are specialized images that contain extensive continuous spectral band information. Compared to natural images, HSIs provide more detailed spectral data, reflecting the intrinsic properties of various objects. This distinctive characteristic of HSIs makes them widely applicable across various domains, such as agricultural monitoring [3, 9], environmental protection [11], ur + +ban planning [8], and mineral exploration [13, 26]. However, due to hardware limitations, existing hyperspectral sensors often fail to deliver sufficient spatial resolution. To achieve higher spatial resolution, a common approach is HSI-MSI fusion, which involves fusing low-resolution hyperspectral images (LR-HSI) with high-resolution multispectral images (HR-MSI) to generate high-resolution hyperspectral images (HR-HSI) [41]. In the past decades, HSI-MSI fusion has garnered significant attention, and numerous related methods have emerged. + +Existing HSI-MSI fusion methods can be roughly categorized into two types, i.e., model-based methods [15, 25, 27, 39] and deep learning (DL)-based methods [24, 31, 35, 38]. Model-based approaches enhance image resolution by integrating spatial and spectral information from low-resolution hyperspectral images and high-resolution multispectral images. However, they rely on manually designed feature extractors and often assume independence between spatial and spectral information, which poses limitations in handling high-dimensional data and impedes the effective capture of complex features. In particular, when faced with the high dimensionality and nonlinearity of hyperspectral images, these methods struggle to deliver optimal performance and frequently require precise prior knowledge, thereby restricting their applicability. + +In recent years, deep learning (DL) has made significant advancements in HSI-MSI fusion. These methods often rely on deep convolutional neural networks (DCNNs) [31, 37, 40] or Transformer networks [5, 14, 30, 32] to extract intricate spatial and spectral features from the images. DCNNs are particularly effective at capturing local patterns and hierarchical features, making them well-suited for spatial information extraction. On the other hand, Transformer networks excel at modeling long-range dependencies, allowing them to effectively capture global spectral information. While these methods outperform traditional model-based approaches in terms of feature extraction, they often rely heavily on large volumes of paired HSI-MSI data, which can be difficult to acquire in practical ap + +plications. Recently, diffusion models (DMs) [4, 7, 17, 29], a novel generative models, have shown tremendous potential in image generation. With a more stable training process and superior generation quality, diffusion models have been widely applied in various image tasks, including image super-resolution [23, 33, 36]. In the realm of hyperspectral image fusion [6, 22, 34], PLRDiff [22] introduced a low-rank diffusion model tailored for hyperspectral image pan-sharpening, which leverages a pretrained deep diffusion model to provide image priors through Bayesian principles. In [20], a spatial-spectral-bilateral cycle diffusion framework (S2CycleDiff) was proposed to progressively generate high-resolution hyperspectral images by learning a customized conditional cycle-diffusion network. However, despite these advancements, achieving high-quality results generally requires thousands of iterations, resulting in slow inference speeds and substantial computational costs. Some approaches [16, 18, 28] have developed accelerated sampling strategies to reduce the number of iterations to a few dozen. Nevertheless, each inference step still demands substantial memory and floating-point computations, particularly in HSI-MSI fusion. Moreover, most edge devices (e.g., drones) are limited by storage and computational resources, making it challenging to deploy these demanding methods in practical remote sensing applications. Consequently, it is crucial to design a model that achieves low resource consumption and fast inference speeds while maintaining generation quality. + +In this work, we propose a novel self-learning approach, the Adaptive Residual Guided Subspace Diffusion Model (ARGS-Diff). Based on the fact that HSI can be effectively reconstructed from the product of two low-dimensional components: the spectral basis and the reduced coefficient, we design and train two lightweight spectral and spatial networks and then use them to independently reconstruct the spectral basis and the reduced coefficient, respectively. The training process relies solely on the observed LR-HSI and HR-MSI to fully exploit the rich spectral and spatial information inherent in each of them. Furthermore, we introduce the Adaptive Residual Guided Module (ARGM) to ensure more stable convergence during the sampling process. In our method, both components need to be updated simultaneously at each step, and improper alignment between them may lead to instability or collapse. The ARGM alleviates this issue by applying a residual-guided function at each sampling step. It calculates the residual between the reconstructed two components and the target, then adjusts them to reduce the discrepancy, effectively guiding them toward better alignment with the true values. By incorporating this module, our approach demonstrates enhanced robustness to noise and achieves more stable and reliable reconstruction performance. + +In summary, our contributions are as follows: + +- We propose a self-learning Adaptive Residual Guided Subspace Diffusion Model (ARGS-Diff), which effectively learns spectral and spatial distributions from the observed LR-HSI and HR-MSI using specially designed lightweight spectral and spatial networks, and reconstructs the HR-HSI from the estimated spectral basis and reduced coefficient. +- We introduce the Adaptive Residual Guided Module (ARGM), which refines the spectral and spatial components through a residual-guided function, effectively enhancing the stability and convergence of the sampling process. +- Extensive experiments conducted on four datasets demonstrate the superior performance of the proposed method over existing state-of-the-art approaches. + +# 2. Preliminaries + +# 2.1. Diffusion Models + +The diffusion model reconstructs data by progressively adding noise during a forward process and learning the reverse denoising pathway to recover the original data [7, 28]. Typically, a diffusion process is composed of a forward process with $T$ steps of noise addition, followed by a reverse process over the same $T$ steps. Given a noise schedule $\{\bar{\alpha}_t\}_{t=1}^T$ and clean input data $\mathbf{x}_0$ , the noise addition in each step is defined by: + +$$ +\mathbf {x} _ {t} = \sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0} + (\sqrt {1 - \bar {\alpha} _ {t}}) \epsilon , \quad \epsilon \sim \mathcal {N} (0, I). \tag {1} +$$ + +The model is trained to predict the noise $\epsilon$ by minimizing the error between the estimated and true noise: + +$$ +\epsilon_ {\theta} \left(\mathbf {x} _ {t}, t\right) \approx \epsilon = \frac {\mathbf {x} _ {t} - \sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0}}{\sqrt {1 - \bar {\alpha} _ {t}}}. \tag {2} +$$ + +In the reverse process, starting from random noise, the model iteratively refines the estimate over $T$ steps. At each step $t$ , the model estimates the clean input $\hat{\mathbf{x}}_0$ as: + +$$ +\hat {\mathbf {x}} _ {0} = \frac {\mathbf {x} _ {t} - (1 - \bar {\alpha} _ {t}) \epsilon_ {\theta} (\mathbf {x} _ {t} , t)}{\sqrt {\bar {\alpha} _ {t}}}. \tag {3} +$$ + +Subsequently, $\mathbf{x}_{t - 1}$ is computed by: + +$$ +\mathbf {x} _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} \hat {\mathbf {x}} _ {0} + (\sqrt {1 - \bar {\alpha} _ {t - 1}}) \epsilon_ {\theta}. \tag {4} +$$ + +The final reconstructed output $\mathbf{x}_0$ is obtained after all reverse steps are completed. + +# 2.2. HSI-MSI Fusion + +In HSI-MSI fusion problem, the input data consists of a LR-HSI $\mathcal{X} \in \mathbb{R}^{h \times w \times C}$ and a HR-MSI $\mathcal{Y} \in \mathbb{R}^{H \times W \times c}$ , which are fused to produce a HR-HSI $\mathcal{Z} \in \mathbb{R}^{H \times W \times C}$ . Typically, + +![](images/a9a364fee09f531dcbefdc9b908f606bd548f000a6bd198eb704f3b975388438.jpg) +Figure 1. The overall framework of the proposed ARGs-Diff. + +the relationship between the LR-HSI, HR-MSI, and HR-HSI is represented through degradation models as follows: + +$$ +\mathcal {X} = \mathcal {H} (\mathcal {Z}) + \mathcal {N} _ {\mathcal {X}}, \tag {5} +$$ + +$$ +\mathcal {Y} = \mathcal {Z} \times_ {3} \mathbf {R} + \mathcal {N} _ {\mathcal {Y}}, \tag {6} +$$ + +where $\mathcal{H}(\cdot)$ represents downsampling in the spatial domain, $\mathbf{R}$ is the spectral response function (SRF) matrix, $\times_{3}$ denotes mode-3 multiplication, $\mathcal{N}_{\mathcal{X}}$ and $\mathcal{N}_{\mathcal{Y}}$ are the measurements noise, which is assumed to be the Gaussian noise. + +A common approach for HSI-MSI fusion is to decompose the HSI into two low-dimensional components, process each separately, and then combine them to synthesize a clean HR-HSI. The decomposition is expressed as: + +$$ +\mathcal {Z} = \mathcal {A} \times_ {3} \mathbf {E}, \tag {7} +$$ + +where $\mathcal{A} \in \mathbb{R}^{H \times W \times d}$ represents the reduced coefficient, $\mathbf{E} \in \mathbb{R}^{d \times C}$ represents the spectral basis, $d$ is the number of subspace dimensions, and $d \ll C$ . + +Compared to the high dimensional of HSI, these two low-dimensional components have reduced data complexity, making them easier to process. In this paper, we follow this approach by training a spectral and spatial network using the LR-HSI and HR-MSI. During the sampling process, these networks separately reconstruct the corresponding spectral basis and reduced coefficient. The final fused HR-HSI is obtained from the product of these two components. + +# 3. Proposed Method + +# 3.1. Overview + +As described in Eq. (7), the HSI can be decomposed into two low-dimensional components, i.e., the spectral basis and the reduced coefficient. We train two separate networks to reconstruct these components, which are then combined to form the final HR-HSI. Specifically, We design a spectral network and a spatial network, and train them using the LR-HSI and HR-MSI based on the forward process of the diffusion model to separately extract the spectral and spatial distribution information. The detailed network architecture and training procedure are described in Section 3.2. + +During the sampling process, we leverage the posterior sampling theory [2] to iteratively reconstruct the spectral basis and reduced coefficient using the well-trained networks, conditioning on the observed LR-HSI and HR-MSI. The detailed sampling process is outlined in Section 3.3. Additionally, since we need to update two components simultaneously at each sampling step, which may lead to instability if they are not properly aligned, we introduce the Adaptive Residual Guided Module (ARGM) to iteratively refine the spectral and spatial components at each step, ensuring the stability of the sampling process, as described in Section 3.4. After $T$ sampling steps, the HR-HSI is obtained by taking the product of the reconstructed spectral basis and reduced coefficient. + +# 3.2. Self-learning Subspace Networks + +Recent advancements in diffusion models have introduced many techniques, such as attention mechanisms, to enhance performance. However, these innovations often come with + +a significant increase in network size, resulting in higher resource consumption and longer image generation times. To address this challenge, we design two lightweight networks specifically tailored for hyperspectral image fusion: a spectral network and a spatial network. This design significantly reduces sampling time without compromising performance in our method. + +The spectral network is a fully connected network consisting of five layers, where the input and output feature dimensions match the number of spectral bands, $C$ , in the LR-HSI. The spatial network adopts a "UNet-like" architecture [21], comprising nine convolutional layers: four down-sampling layers, one intermediate layer, and four upsampling layers. Each convolutional layer includes two residual blocks, and the input and output channels are set to the number of subspace dimensions, $d$ . + +To optimize network training, we restructure the training samples. For the spectral network, LR-HSI, which contains abundant spectral information, is leveraged to represent the spectral distribution of the HR-HSI. Specifically, we randomly select $d$ pixel spectra from the LR-HSI and concatenate them into a sample of size $(d,C)$ for the spectral network's input. For the spatial network, HR-MSI, which retains rich spatial information, is used for training. We extract a patch from the HR-MSI, randomly select one band from $c$ spectral bands within the patch, and repeat this process $d$ times to form a sample of size $(H_{\mathrm{patch}},W_{\mathrm{patch}},d)$ . This approach not only facilitates effective training but also introduces diversity into the input samples, enabling the network to better capture the spatial and spectral features of the images. + +# 3.3. Subspace Reverse Diffusion Process + +Recently, several methods [1, 2] have been proposed to generate images using various conditional inputs. In general, given a condition $\mathbf{y}$ , these methods model the reverse sampling process as $p_{\theta}(\mathbf{x}_{t - 1}|\mathbf{x}_t,\mathbf{y})$ incorporating a gradient term into the noise estimation function $\epsilon_{\theta}$ . This is formulated as: + +$$ +\hat {\epsilon} _ {\theta} \left(\mathbf {x} _ {t}, t\right) = \epsilon_ {\theta} \left(\mathbf {x} _ {t}, t\right) - \rho \nabla_ {\mathbf {x} _ {t}} \mathcal {L} (\hat {\mathbf {x}} _ {0}, y), \tag {8} +$$ + +where $\rho$ denotes the step size, and $\mathcal{L}$ is a guidance function that measures the distance between the predicted image $\hat{\mathbf{x}}_0$ and the target image. In our case, we use the LR-HSI $\mathcal{X}$ and HR-MSI $\mathcal{Y}$ as conditions, employing spatial and spectral networks to estimate the spectral basis and the reduced coefficient. Specifically, Eq. (8) can be rewritten as: + +$$ +\hat {s} _ {\theta} \left(\mathcal {A} _ {t}, t\right) = s _ {\theta} \left(\mathcal {A} _ {t}, t\right) - \rho_ {1} \nabla_ {\mathcal {A} _ {t}} \mathcal {L} \left(\hat {\mathcal {A}} _ {0}, \hat {\mathbf {E}} _ {0}, \mathcal {X}, \mathcal {Y}\right), \tag {9} +$$ + +$$ +\hat {c} _ {\zeta} \left(\mathbf {E} _ {t}, t\right) = c _ {\zeta} \left(\mathbf {E} _ {t}, t\right) - \rho_ {2} \nabla_ {\mathbf {E} _ {t}} \mathcal {L} \left(\hat {\mathcal {A}} _ {0}, \hat {\mathbf {E}} _ {0}, \mathcal {X}, \mathcal {Y}\right), +$$ + +where $s_{\theta}(\cdot)$ and $c_{\zeta}(\cdot)$ represent the spatial and spectral networks, respectively, $\rho_{1}$ and $\rho_{2}$ are the step size for updating + +$s_{\theta}(\mathcal{A}_t,t)$ and $c_{\zeta}(\mathbf{E}_t,t)$ . The guidance function is defined as: + +$$ +\begin{array}{l} \mathcal {L} \left(\hat {\mathcal {A}} _ {0}, \hat {\mathbf {E}} _ {0}, \mathcal {X}, \mathcal {Y}\right) = \| \mathcal {H} \left(\hat {\mathcal {A}} _ {0} \times_ {3} \hat {\mathbf {E}} _ {0}\right) - \mathcal {X} \| _ {2} ^ {2} \tag {10} \\ + \lambda_ {1} \| \hat {\mathcal {A}} _ {0} \times_ {3} \hat {\mathbf {E}} _ {0} \times_ {3} \mathbf {R} - \mathcal {Y} \| _ {2} ^ {2}, \\ \end{array} +$$ + +where $\lambda_{1}$ is weight to balance the two observations. $\hat{\mathbf{A}}_0$ and $\hat{\mathbf{E}}_0$ can be estimated from Eq. (3) as: + +$$ +\begin{array}{l} \hat {\mathcal {A}} _ {0} = \frac {\mathcal {A} _ {t} - (\sqrt {1 - \bar {\alpha} _ {t}}) s _ {\theta} (\mathcal {A} _ {t} , t)}{\sqrt {\bar {\alpha} _ {t}}}, \\ \hat {\mathbf {E}} _ {0} = \frac {\mathbf {E} _ {t} - (\sqrt {1 - \bar {\alpha} _ {t}}) c _ {\zeta} (\mathbf {E} _ {t} , t)}{\sqrt {\bar {\alpha} _ {t}}}. \\ \end{array} +$$ + +Next, $\mathcal{A}_{t - 1}$ and $\mathbf{E}_{t - 1}$ are sampled from Eq. (1) as: + +$$ +\mathcal {A} _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} \hat {\mathcal {A}} _ {0} + (\sqrt {1 - \bar {\alpha} _ {t - 1}}) \hat {s} _ {\theta} (\mathcal {A} _ {t}, t), \tag {12} +$$ + +$$ +\mathbf {E} _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} \hat {\mathbf {E}} _ {0} + (\sqrt {1 - \bar {\alpha} _ {t - 1}}) \hat {c} _ {\zeta} (\mathbf {E} _ {t}, t). +$$ + +After $\mathcal{A}_{t - 1}$ and $\mathbf{E}_{t - 1}$ are sampled from Eq. (12) at current step, they are further refined using the proposed ARGM, as detailed in Section 3.4. Upon completion of the sampling process, the reconstructed HR-HSI $\mathcal{Z}_0$ is derived from $\mathcal{A}_0\times_3\mathbf{E}_0$ . The proposed method is summarized in Algorithm 1. + +Remark: To accelerate convergence, we integrate the Adam optimizer [10] into Eq. (9) with the following updates: + +$$ +\mathbf {m} _ {t - 1} ^ {(\mathcal {A})} = \beta_ {1} \mathbf {m} _ {t} ^ {(\mathcal {A})} + (1 - \beta_ {1}) \nabla_ {\mathcal {A} _ {t}} \mathcal {L} (\hat {\mathcal {A}} _ {0}, \hat {\mathbf {E}} _ {0}, \mathcal {X}, \mathcal {Y}), +$$ + +$$ +\mathbf {v} _ {t - 1} ^ {(\mathcal {A})} = \beta_ {2} \mathbf {v} _ {t} ^ {(\mathcal {A})} + (1 - \beta_ {2}) \nabla_ {\mathcal {A} _ {t}} \mathcal {L} (\hat {\mathcal {A}} _ {0}, \hat {\mathbf {E}} _ {0}, \mathcal {X}, \mathcal {Y}) ^ {2}, \tag {13} +$$ + +$$ +\mathbf {m} _ {t - 1} ^ {(\mathbf {E})} = \beta_ {1} \mathbf {m} _ {t} ^ {(\mathbf {E})} + (1 - \beta_ {1}) \nabla_ {\mathbf {E} _ {t}} \mathcal {L} (\hat {\mathcal {A}} _ {0}, \hat {\mathbf {E}} _ {0}, \mathcal {X}, \mathcal {Y}), +$$ + +$$ +\mathbf {v} _ {t - 1} ^ {(\mathbf {E})} = \beta_ {2} \mathbf {v} _ {t} ^ {(\mathbf {E})} + (1 - \beta_ {2}) \nabla_ {\mathbf {E} _ {t}} \mathcal {L} (\hat {\mathcal {A}} _ {0}, \hat {\mathbf {E}} _ {0}, \mathcal {X}, \mathcal {Y}) ^ {2}, +$$ + +where $\mathbf{m}_{t - 1}$ and $\mathbf{v}_{t - 1}$ represent the first and second moment estimates in the Adam optimizer, with superscripts $(\mathcal{A})$ and $(\mathbf{E})$ denoting the corresponding spatial and spectral matrices. The parameters $\beta_{1}$ and $\beta_{2}$ are the decay rates for these moment estimates. + +We then rectify the first and second moment estimates as: + +$$ +\hat {\mathbf {m}} _ {t - 1} ^ {(A)} = \frac {\mathbf {m} _ {t - 1} ^ {(A)}}{1 - \beta_ {1} ^ {T - t}}, \hat {\mathbf {v}} _ {t - 1} ^ {(A)} = \frac {\mathbf {v} _ {t - 1} ^ {(A)}}{1 - \beta_ {2} ^ {T - t}}, \tag {14} +$$ + +$$ +\hat {\mathbf {m}} _ {t - 1} ^ {(\mathbf {E})} = \frac {\mathbf {m} _ {t - 1} ^ {(\mathbf {E})}}{1 - \beta_ {1} ^ {T - t}}, \hat {\mathbf {v}} _ {t - 1} ^ {(\mathbf {E})} = \frac {\mathbf {v} _ {t - 1} ^ {(\mathbf {E})}}{1 - \beta_ {2} ^ {T - t}}. +$$ + +Finally, we rewrite Eq. (9) as: + +$$ +\hat {s} _ {\theta} \left(\mathcal {A} _ {t}, t\right) = s _ {\theta} \left(\mathcal {A} _ {t}, t\right) - \rho_ {1} \frac {\hat {\mathbf {m}} _ {t - 1} ^ {(\mathcal {A})}}{\sqrt {\hat {\mathbf {v}} _ {t - 1} ^ {(\mathcal {A})}} + \epsilon}, \tag {15} +$$ + +$$ +\hat {c} _ {\zeta} (\mathbf {E} _ {t}, t) = c _ {\zeta} (\mathbf {E} _ {t}, t) - \rho_ {2} \frac {\hat {\mathbf {m}} _ {t - 1} ^ {(\mathbf {E})}}{\sqrt {\hat {\mathbf {v}} _ {t - 1} ^ {(\mathbf {E})}} + \epsilon}, +$$ + +where $\epsilon$ is a small constant introduced to avoid division by zero. + +# 3.4. ARGM + +Unlike existing diffusion posterior sampling methods [2, 22], which update only a single intermediate sample at each sampling step, our approach requires simultaneous updates of both the spectral basis and the reduced coefficient. This introduces additional instability into the sampling process, as both components must be adjusted to align with each other in each iteration. To alleviate this issue, we introduce the Adaptive Residual Guided Module (ARGM), which facilitates better alignment between the spectral and spatial components, resulting in more stable convergence. + +Specifically, after obtaining $\mathcal{A}_{t-1}$ and $\mathbf{E}_{t-1}$ at the $t$ -th sampling step, we compute a residual loss that quantifies the discrepancy between the reconstructed two components and the target to refine them. The residual loss can be calculated by: + +$$ +\begin{array}{r l} \mathcal {L} (\mathcal {A} _ {t - 1}, \mathbf {E} _ {t - 1}, \mathcal {X}, \mathcal {Y}) = & \| \mathcal {H} (\mathcal {A} _ {t - 1} \times_ {3} \mathbf {E} _ {t - 1}) - \mathcal {X} \| _ {2} ^ {2} \\ & + \lambda_ {2} \| \mathcal {A} _ {t - 1} \times_ {3} \mathbf {E} _ {t - 1} \times_ {3} \mathbf {R} - \mathcal {Y} \| _ {2} ^ {2}, \end{array} \tag {16} +$$ + +where $\lambda_{2}$ is a weight balancing the two objectives. The first term ensures that the predicted LR-HSI $\mathcal{H}(\mathcal{A}_{t - 1}\times_3\mathbf{E}_{t - 1})$ is consistent with the observed image $\mathcal{X}$ , while the second term penalizes the discrepancy between the predicted HR-MSI $(\mathcal{A}_{t - 1}\times_3\mathbf{E}_{t - 1}\times_3\mathbf{R})$ and the target $\mathcal{V}$ . This ensures that the spectral and spatial components are aligned both individually and in combination. + +Then, the spatial and spectral components are updated by: + +$$ +\mathcal {A} _ {t - 1} := \mathcal {A} _ {t - 1} - \frac {\rho_ {1}}{r} \nabla_ {\mathcal {A} _ {t - 1}} \mathcal {L} \left(\mathcal {A} _ {t - 1}, \mathbf {E} _ {t - 1}, \mathcal {X}, \mathcal {Y}\right), \tag {17} +$$ + +$$ +\mathbf {E} _ {t - 1} := \mathbf {E} _ {t - 1} - \frac {\rho_ {2}}{r} \nabla_ {\mathbf {E} _ {t - 1}} \mathcal {L} (\mathcal {A} _ {t - 1}, \mathbf {E} _ {t - 1}, \mathcal {X}, \mathcal {Y}), +$$ + +where $r$ is a ratio that controls the relative step sizes for updating the spatial component and the spectral component, respectively. + +In this way, the ARGM computes the residual between the reconstructed two components and the target at each step, and then uses the residual-guided function to adjust both components. On the one hand, this brings the components closer to the target direction, reducing the instability that arises from updating both components simultaneously. On the other hand, it compensates for the prediction errors introduced by the model at each sampling step, thus making the model more robust and less sensitive to noise and discrepancies. + +By iteratively refining the spatial and spectral components using residual-guided function, the ARGM helps stabilize the convergence of the diffusion sampling process, + +leading to improved performance and robustness in the HSI-MSI fusion task. + +Algorithm 1: Reverse Diffusion Process of ARGs-Diff +Input: LR-HSI $\mathcal{X}$ ,HR-MSI $\mathcal{V}$ $\mathcal{A}_T$ and $\mathbf{E}_T$ sampled from $\mathcal{N}(0,I)$ , spatial network $s_\theta$ spectral network $c_{\zeta}$ , balance weight $\lambda_1,\lambda_2$ step size $\rho_{1},\rho_{2}$ ratio $r$ Adam params $\mathbf{m}_T^{(\mathcal{A})}$ $\mathbf{v}_T^{(\mathcal{A})}$ $\mathbf{m}_T^{(\mathbf{E})}$ $\mathbf{v}_T^{(\mathbf{E})}$ $\beta_{1},\beta_{2},\epsilon$ +Output: HR-HSI $\mathcal{Z}_0$ Initialize $\mathbf{m}_T^{(\mathcal{A})} = \mathbf{v}_T^{(\mathcal{A})} = \mathbf{m}_T^{(\mathbf{E})} = \mathbf{v}_T^{(\mathbf{E})} = 0$ +for $t = T,T - 1,\dots 1$ do step 1: estimate $\hat{\mathcal{A}}_0,\hat{\mathbf{E}}_0$ by Eq. (11) step 2: calculate $\mathcal{L}(\hat{\mathcal{A}}_0,\hat{\mathbf{E}}_0,\mathcal{X},\mathcal{Y})$ by Eq.(10) step 3: calculate $\hat{\mathbf{m}}_{t - 1}^{(\mathcal{A})},\hat{\mathbf{v}}_{t - 1}^{(\mathcal{A})},\hat{\mathbf{m}}_{t - 1}^{(\mathbf{E})},\hat{\mathbf{v}}_{t - 1}^{(\mathbf{E})}$ by Eq. (13), Eq. (14) step 4: estimate $\hat{s}_{\theta}(\mathcal{A}_t,t),\hat{c}_{\zeta}(\mathbf{E}_t,t)$ by Eq. (15) step 5: sample $\mathcal{A}_{t - 1},\mathbf{E}_{t - 1}$ by Eq. (12) step 6: calculate $\mathcal{L}(\mathcal{A}_{t - 1},\mathbf{E}_{t - 1},\mathcal{X},\mathcal{Y})$ by Eq. (16) step 7: update $\mathcal{A}_{t - 1},\mathbf{E}_{t - 1}$ by Eq. (17) +end +return $\mathcal{Z}_0 = \mathcal{A}_0\times_3\mathbf{E}_0$ + +# 4. Experiment + +# 4.1. Datasets and Evaluation Metrics + +We conducted a series of simulated experiments using three public datasets to assess the performance of the proposed method for HSI fusion: Pavia University, Chikusei, and KSC. The original dimensions of these datasets are $610 \times 340 \times 103$ , $2517 \times 2335 \times 128$ , and $512 \times 614 \times 176$ . We cropped the central regions of each to obtain HSIs with sizes $256 \times 256 \times 103$ , $256 \times 256 \times 128$ , and $256 \times 256 \times 176$ . The LR-HSI was generated by downsampling the HR-HSI using bicubic interpolation with a scale factor of 4, and the HR-MSI was obtained by performing mode-3 tensor multiplication of the HR-HSI with a simulated spectral response function. Additionally, to evaluate the method's real-world applicability, we conducted experiments on the DFC2018 Houston dataset, which provides real hyperspectral data. The following four commonly used evaluation metrics were employed to assess model performance: peak signal-to-noise ratio (PSNR), spectral angle mapper (SAM), relative dimensionless global error in synthesis (ERGAS), and structural similarity (SSIM). + +# 4.2. Implementation Details + +All experiments were carried out on an NVIDIA GeForce RTX 4090 GPU. To test the robustness of the model to + +noise, we added noise to all LR-HSI and HR-MSI inputs with a Signal-to-Noise Ratio (SNR) of 35. The diffusion sampling step $T$ was set to 500, using an exponential noise schedule [19]. The balance weights $\lambda_{1}$ and $\lambda_{2}$ were both set to 1, while the step sizes $\rho_{1}$ and $\rho_{2}$ were set to 0.05, and the ratio $r$ was set to 10. The Adam optimizer was used with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ , and $\epsilon = 1 \times 10^{-8}$ . The number of subspace dimensions $d$ was set to 8. + +# 4.3. Experimental Results + +# 4.3.1. Results on Simulated Data + +We compare the proposed method with six well-established approaches, including traditional methods such as CNMF [39] and HySure [27], deep learning-based methods like CuCaNet [38] and MIAE [12], and diffusion model-based approaches such as PLRDiff [22] and $\mathrm{S}^2$ CycleDiff [20]. The quantitative results are summarized in Tables 1, 2, and 3. From these results, it is clear that the proposed method consistently outperforms all other competing approaches across all evaluation metrics on three different datasets. In terms of PSNR, the proposed method achieves the highest values, surpassing the second-best method, MIAE, by 1.31 dB, 1.37 dB, and 1.27 dB on the respective datasets. This demonstrates the superior ability of ARGs-Diff to reconstruct high-fidelity hyperspectral images, preserving both spectral and spatial details with greater accuracy. When examining the computational efficiency, traditional methods such as CNMF and HySure offer faster inference speeds, but they result in poorer fusion performance. Deep learning methods like CuCaNet and MIAE show improved accuracy, but at the cost of longer processing times. Diffusion model-based methods like PLRDiff and $\mathrm{S}^2$ CycleDiff also demand significant inference time, while the proposed method achieves high-quality fusion results with a reduced runtime. + +Representative visual results are illustrated in Figure 2, showcasing the superior image quality of proposed method. To further highlight the performance, we include error maps that demonstrate the proposed method's ability to minimize bias across all three datasets, resulting in high-fidelity HR-HSI reconstructions. Overall, the proposed approach excels in both quantitative and qualitative evaluations, establishing its superiority in HSI-MSI fusion. + +# 4.3.2. Results on Real-Data + +Figure 3 presents the reconstructed pseudo-color images from the DFC2018 Houston dataset, generated using various methods. The reconstructed HSIs from the comparison methods exhibit varying degrees of deficiencies in color consistency, naturalness, and saturation control. In contrast, ARGs-Diff effectively preserves the authenticity and vibrancy of colors while avoiding pseudo-color artifacts and color distortions. The resulting spectral images exhibit high + +Table 1. The quantitative results on the Pavia dataset. The best and second-best values are highlighted. + +
MethodsPSNR↑SAM↓EGARS↓SSIM↑Time (s)
CNMF38.823.131.950.9747
HySure37.853.192.320.9705
CuCaNet38.862.951.980.970401
MIAE41.022.731.640.97520
PLRDiff40.113.401.940.97079
S2CycleDiff38.773.212.130.970297
ARGS-Diff42.332.641.490.97712
+ +Table 2. The quantitative results on the Chikusei dataset. The best and second-best values are highlighted. + +
MethodsPSNR↑SAM↓EGARS↓SSIM↑Time (s)
CNMF38.142.172.590.96711
HySure38.712.252.650.9675
CuCaNet39.012.092.550.968433
MIAE40.531.952.510.96924
PLRDiff38.442.402.790.96379
S2CycleDiff39.732.082.530.968318
ARGS-Diff41.901.772.100.97012
+ +Table 3. The quantitative results on the KSC dataset. The best and second-best values are highlighted. + +
MethodsPSNR↑SAM↓EGARS↓SSIM↑Time (s)
CNMF39.883.471.860.97715
HySure40.113.171.800.9786
CuCaNet40.393.051.720.977474
MIAE42.362.581.420.97932
PLRDiff40.073.631.830.97779
S2CycleDiff40.193.151.790.978349
ARGS-Diff43.632.541.250.98013
+ +quality, closely resembling real-world scenes. This demonstrates ARGs-Diff's superior ability to preserve fine details and maintain overall color balance, ensuring the consistency of both spatial and spectral information in the reconstructed hyperspectral images. These results highlight the effectiveness of the proposed method, particularly in real-world applications, where accurate color reproduction and detail preservation are critical. + +# 4.4. Ablation Study + +In this section, we assess the impact of the Adaptive Residual Guided Module (ARGM) on the performance of the proposed method. The results, summarized in Table 4, clearly demonstrate that incorporating ARGM leads to a significant improvement in the quality of the reconstructed images. Specifically, the inclusion of ARGM increased the PSNR by $0.47\mathrm{dB}$ , $0.57\mathrm{dB}$ , and $0.61\mathrm{dB}$ across the three datasets, while the SAM improved by $0.11$ , $0.12$ , and $0.16$ , underscoring its positive contribution to the fusion performance. Importantly, the computational overhead introduced + +![](images/e1c63aa863eabf19ea7ae1d3b6f26b53a8c51d9409c003bb18dbc468c5dce4c3.jpg) +Figure 2. Visual results and reconstruction error maps obtained by different methods. The first, third, and fifth rows are the pseudo-color images of Pavia, Chikusei, and KSC datasets, and the second, fourth, and sixth rows correspond to their error maps. + +![](images/a03a89081497a0cf03b756cd2b3efdb336ba6fea2bfdc9c783f73c45f315d09a.jpg) +LR-HSI + +![](images/3fb8b44557ba6a8c7d08ea06a4264734fc23c41d1ef3f7d9593091d924435449.jpg) +HR-MSI + +![](images/7d6e330dcf8ede40b3c4ddd2deefec1964ee0fa10613f4c60c04cd4692184755.jpg) +CNMF + +![](images/c17c0f8701bb767e9a9f80843303773667bc7bad7204eaac21cf2e20a189ce60.jpg) +HySure + +![](images/d09dfa6ee25539f5160f71bd24fe248536e452cb685aadeea18edc4bc0a51999.jpg) +CuCaNet + +![](images/580b96cba516ecf8b07b7d60d82767d06d7e1ad1ff8dc2440f51762dcb7dae03.jpg) +MIAE + +![](images/ebc0459f8aeb451d9fe2939e37d45f080953a3eacc94bc81bc5d9fffea5131ff.jpg) +PLRDiff + +![](images/008abd72d185fe1ac17aa8c3f62d1045344c1ff05ea521d4b340b1f209315c5f.jpg) +ARGS-Diff + +by ARGM is minimal, adding only 1-2 seconds to the overall runtime, which has a negligible impact on the total processing time. + +Additionally, Figure 4 presents the reconstruction results of the reduced coefficient $\mathcal{A}$ at different sampling steps $t = \{500, 400, 300, 200, 100, 0\}$ on the Pavia dataset. It is evident that, with the incorporation of ARGM, the reconstructed reduced coefficient becomes more informative and detailed. This visual improvement further confirms the significant enhancement brought by ARGM, which facilitates more accurate and robust fusion, contributing to better alignment of the spectral and spatial components. + +![](images/87f20e5f3d69194e90321d5fcb83419b6c2d629c5512e6b1d6d58e5ecae72b4d.jpg) +Figure 3. Visual results obtained by different methods on the Houston dataset. + +![](images/f2f468a724bbeb63cb66c4c3a2147e79f7264495be20ea8430b57a54f4848d5c.jpg) +Figure 4. Intermediate results of the reduced coefficient $\mathcal{A}$ . The first and second rows represent 'w/o ARGM' and 'w/ ARGM'. + +![](images/9f02cf2f38577374d1f6c2942929411133fb7a9f6576637d89172c341225888a.jpg) + +![](images/c3582c37a2b6974a3fa6d408df22de1d5eb40f778291e0ffc0f046360600b59c.jpg) + +![](images/cee978299960d964c3ef85eecfbecc7e7abdb2c615c78e1e8e69bc5b7ed1701a.jpg) + +![](images/4b0cfd3ce430ee47a662bb1bca436f77c155d990e9dfdd226ffeec238bd12cd7.jpg) + +Table 4. The quantitative ablation results of ARGM. + +
DatasetsMethodsPSNRSAMTime (s)
Paviaw/o ARGM41.862.7511
w/ ARGM42.332.6412
Chikuseiw/o ARGM41.331.8911
w/ ARGM41.901.7712
KSCw/o ARGM43.022.7011
w/ ARGM43.632.5413
+ +# 4.5. Analysis of Hyperparameters + +$d$ and $T$ . In this part, we examine the parameter sensitivity related to the number of subspace dimensions and the number of sampling steps. Using the Pavia dataset as a case study, the results are shown in Figure 5. As $d$ increases from 1 to 8, performance improves significantly; however, beyond this point, performance gradually declines. This drop can be attributed to the increasing number of parameters that need to be updated when $d$ becomes too large, which complicates the convergence process. Consequently, we set $d$ to 8. For the number of sampling steps $T$ , we observe substantial improvement as $T$ increases from 100 to 500. However, beyond 500 steps, the performance reaches a plateau and may even slightly decrease, suggesting that the model has reached saturation and further increases may lead to overfitting. Therefore, we set $T$ to 500. + +![](images/c2e24d7ed9d029954c07bf9212572a13e30fbc935437d5f86291b9eafaf9c0a6.jpg) +Figure 5. Sensitivity analysis of the parameters $d$ and $T$ . + +$\rho_{1}$ and $\rho_{2}$ . In Eq. (9), we introduce two hyperparameters, $\rho_{1}$ and $\rho_{2}$ , which control the update step sizes for the predicted components. To determine an appropriate combination of values, we conduct a hyperparameter search over $\rho_{1}, \rho_{2} \in \{0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07\}$ using the Pavia dataset. The PSNR results are shown in Table 5, from which we observe that the best performance is achieved when $\rho_{1} = \rho_{2} = 0.05$ . Therefore, for the Pavia dataset, we set $\rho_{1}$ and $\rho_{2}$ to 0.05. + +# 4.6. Inference Time and Model Scale Analysis + +In this section, we compare the proposed method with other diffusion model approaches in terms of parameter count, memory consumption, and runtime. Table 6 summarizes the results for the PLRDiff, S $^2$ CycleDiff, and the proposed + +Table 5. Sensitivity analysis of the parameters ${\rho }_{1}$ and ${\rho }_{2}$ on Pavia dataset. The best result is highlighted. + +
ρ2ρ10.010.020.030.040.050.060.07
0.0133.7038.9634.0936.8834.6735.0731.84
0.0240.7540.6940.8241.1541.4141.5040.73
0.0341.3541.5541.7141.5741.5941.9141.80
0.0441.7541.8941.7741.9341.9042.1041.69
0.0541.7242.0042.1342.2542.3342.2741.67
0.0641.6641.9742.1342.1542.2342.1141.95
0.0741.6741.8041.9742.1942.1441.7841.94
+ +ARGS-Diff, using the Pavia dataset as an example. The proposed method has a parameter count of just 21.85M, which is approximately 1/20th and 1/30th of the parameter counts for PLRDiff and $\mathrm{S}^2\mathrm{CycleDiff}$ , respectively. It also consumes only 2.11GB of memory, significantly less than the other two methods. Furthermore, the proposed method completes sampling in just 12 seconds, significantly outperforming PLRDiff (79 seconds) and $\mathrm{S}^2\mathrm{CycleDiff}$ (297 seconds) in terms of processing time. Overall, the proposed method not only excels in fusion accuracy but also offers substantial advantages in terms of model size, memory usage, and runtime, making it highly suitable for real-world applications. + +Table 6. The model size and computational consumption of the proposed method and other diffusion model-based methods. + +
MethodsPSNRParams (M)Memory (GB)Time (s)
PLRDiff40.11391.056.2179
S2CycleDiff38.77606.4423.34297
ARGS-Diff42.3321.852.1112
+ +# 5. Conclusion + +In this paper, we propose a self-learning Adaptive Residual Guided Subspace Diffusion Model (ARGS-Diff) for HSI-MSI fusion. We carefully design two lightweight spectral and spatial networks, training them using the observed LR-HSI and HR-MSI to extract the rich spectral and spatial information inherent in each. These well-trained networks are then employed to reconstruct two low-dimensional components: the spectral basis and the reduced coefficient. The HR-HSI is obtained by taking the product of these two components. Additionally, we introduce the Adaptive Residual Guided Module (ARGM), which further enhances the sampling process by stabilizing the updates of the spectral and spatial components. Extensive experimental evaluations demonstrate the superiority of the proposed method, achieving remarkable performance in the field of HSI-MSI fusion. + +# References + +[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 4 +[2] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. 3, 4, 5 +[3] Laura M Dale, André Thewis, Christelle Boudry, Ioan Rotar, Pierre Dardenne, Vincent Baeten, and Juan A Fernandez Pierna. Hyperspectral imaging applications in agriculture and agro-food product quality and safety control: A review. Applied Spectroscopy Reviews, 48(2):142-159, 2013. 1 +[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[5] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1 +[6] Wen-jin Guo, Weiying Xie, Kai Jiang, Yunsong Li, Jie Lei, and Leyuan Fang. Toward stable, interpretable, and lightweight hyperspectral super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22272-22281, 2023. 2 +[7] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[8] Xin Huang, Dawei Wen, Jiayi Li, and Rongjun Qin. Multi-level monitoring of subtle urban changes for the megacities of china using high-resolution multi-view satellite imagery. Remote Sensing of Environment, 196:56-75, 2017. 1 +[9] Sami Khanal, Kushal Kc, John P Fulton, Scott Shearer, and Erdal Ozkan. Remote sensing in agriculture—accomplishments, limitations, and opportunities. Remote Sensing, 12(22):3783, 2020. 1 +[10] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 4 +[11] Jun Li, Yanqiu Pei, Shaohua Zhao, Rulin Xiao, Xiao Sang, and Chengye Zhang. A review of remote sensing for environmental monitoring in china. Remote Sensing, 12(7):1130, 2020. 1 +[12] Jianjun Liu, Zebin Wu, Liang Xiao, and Xiao-Jun Wu. Model inspired autoencoder for unsupervised hyperspectral image super-resolution. IEEE Transactions on Geoscience and Remote Sensing, 60:1-12, 2022. 6 +[13] Lei Liu, Jun Zhou, Dong Jiang, Dafang Zhuang, Lamin R Mansaray, and Bing Zhang. Targeting mineral resources with remote sensing and field data in the xiemisitai area, west junggar, xinjiang, china. Remote sensing, 5(7):3156-3171, 2013. 1 +[14] 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. 1 +[15] Laetitia Loncan, Luis B De Almeida, José M Bioucas-Dias, Xavier Briottet, Jocelyn Chanussot, Nicolas Dobigeon, Sophie Fabre, Wenzhi Liao, Giorgio A Licciardi, Miguel Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geoscience and Remote Sensing Magazine, 3(3):27-46, 2015. 1 +[16] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan LI, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems, pages 5775-5787, 2022. 2 +[17] Yuchun Miao, Lefei Zhang, Liangpei Zhang, and Dacheng Tao. Dds2m: Self-supervised denoising diffusion spatiospectral model for hyperspectral image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12086-12096, 2023. 2 +[18] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2 +[19] Li Pang, Xiangyu Rui, Long Cui, Hongzhong Wang, Deyu Meng, and Xiangyong Cao. Hir-diff: Unsupervised hyperspectral image restoration via improved diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3005-3014, 2024. 6 +[20] Jiahui Qu, Jie He, Wenqian Dong, and Jingyu Zhao. S2cyclediff: Spatial-spectral-bilateral cycle-diffusion framework for hyperspectral image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4623-4631, 2024. 2, 6 +[21] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4 +[22] Xiangyu Rui, Xiangyong Cao, Li Pang, Zeyu Zhu, Zongsheng Yue, and Deyu Meng. Unsupervised hyperspectral pansharpening via low-rank diffusion model. Information Fusion, 107:102325, 2024. 2, 5, 6 +[23] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2 +[24] Giuseppe Scarpa, Sergio Vitale, and Davide Cozzolino. Target-adaptive cnn-based pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 56(9):5443-5457, 2018. 1 +[25] Vijay P Shah, Nicolas H Younan, and Roger L King. An efficient pan-sharpening method via a combined adaptive pca approach and contourlets. IEEE Transactions on Geoscience and Remote Sensing, 46(5):1323-1335, 2008. 1 +[26] Hojat Shirmard, Ehsan Farahbakhsh, R Dietmar Müller, and Rohitash Chandra. A review of machine learning in pro + +cessing remote sensing data for mineral exploration. Remote Sensing of Environment, 268:112750, 2022. 1 +[27] Miguel Simoes, José Bioucas-Dias, Luis B Almeida, and Jocelyn Chanussot. A convex formulation for hyperspectral image superresolution via subspace-based regularization. IEEE Transactions on Geoscience and Remote Sensing, 53 (6):3373-3388, 2014. 1, 6 +[28] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2 +[29] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 1 +[31] Wu Wang, Weihong Zeng, Yue Huang, Xinghao Ding, and John Paisley. Deep blind hyperspectral image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4150-4159, 2019. 1 +[32] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 1 +[33] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, 2023. 2 +[34] Chanyue Wu, Dong Wang, Yunpeng Bai, Hanyu Mao, Ying Li, and Qiang Shen. Hsr-diff: Hyperspectral image superresolution via conditional diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2023. 2 +[35] Qi Xie, Minghao Zhou, Qian Zhao, Zongben Xu, and Deyu Meng. Mhf-net: An interpretable deep network for multispectral and hyperspectral image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1457-1473, 2020. 1 +[36] Qiao Xue, Qingqing Ye, Haibo Hu, Youwen Zhu, and Jian Wang. Ddrm: A continual frequency estimation mechanism with local differential privacy. IEEE Transactions on Knowledge and Data Engineering, 35(7):6784-6797, 2022. 2 +[37] Jingxiang Yang, Yong-Qiang Zhao, and Jonathan Cheung-Wai Chan. Hyperspectral and multispectral image fusion via deep two-branches convolutional neural network. Remote Sensing, 10(5):800, 2018. 1 +[38] Jing Yao, Danfeng Hong, Jocelyn Chanussot, Deyu Meng, Xiaoxiang Zhu, and Zongben Xu. Cross-attention in coupled unmixing nets for unsupervised hyperspectral superresolution. In Computer Vision - ECCV 2020, pages 208-224, 2020. 1, 6 +[39] Naoto Yokoya, Takehisa Yairi, and Akira Iwasaki. Coupled nonnegative matrix factorization unmixing for hyperspectral + +and multispectral data fusion. IEEE Transactions on Geoscience and Remote Sensing, 50(2):528-537, 2011. 1, 6 +[40] Xueting Zhang, Wei Huang, Qi Wang, and Xuelong Li. Ssrnet: Spatial-spectral reconstruction network for hyperspectral and multispectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 59(7):5953-5965, 2020. 1 +[41] Man Zhou, Jie Huang, Keyu Yan, Danfeng Hong, Xiuping Jia, Jocelyn Chanussot, and Chongyi Li. A general spatial-frequency learning framework for multimodal image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 \ No newline at end of file diff --git a/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/images.zip b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6357b943c132752697179bd51c174daf1e306f4e --- /dev/null +++ b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48cbea0c3329c512e9576f6142ca152cab04ccf1c22bfd8d67e02d8f9471049b +size 719892 diff --git a/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/layout.json b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3156b7571bb7e85b5f9a35624ca64adff7530215 --- /dev/null +++ b/CVPR/2025/Self-Learning Hyperspectral and Multispectral Image Fusion via Adaptive Residual Guided Subspace Diffusion Model/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdca3815fef446258f580cc52798892f5b25c27ed1dc114f5859529cdb7a800d +size 459554 diff --git a/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/04de2c6b-9723-4b48-a712-82d3adfbf21e_content_list.json b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/04de2c6b-9723-4b48-a712-82d3adfbf21e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dd1810c6dd9f7bf5283cfc5f5fbe6454e49ec604 --- /dev/null +++ b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/04de2c6b-9723-4b48-a712-82d3adfbf21e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6e9e6727603e0f9ce095e97344b8244084025020a8f6543758c761860616832 +size 83187 diff --git a/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/04de2c6b-9723-4b48-a712-82d3adfbf21e_model.json b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/04de2c6b-9723-4b48-a712-82d3adfbf21e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b66b7d8ad75039c4da21aa8265efcf77e3a48302 --- /dev/null +++ b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/04de2c6b-9723-4b48-a712-82d3adfbf21e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2b3892d45ebd890ef75a3c151f3743c7686ef79ed820e25f4426eac2517674c +size 104672 diff --git a/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/04de2c6b-9723-4b48-a712-82d3adfbf21e_origin.pdf b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/04de2c6b-9723-4b48-a712-82d3adfbf21e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..05dd306bde2a97ec88eb7f19fe9f161c76afe17d --- /dev/null +++ b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/04de2c6b-9723-4b48-a712-82d3adfbf21e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb455823fface6b9c04d36febb13462593e2cbfb4049583dbe78a92b33a88c07 +size 2634935 diff --git a/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/full.md b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e85cf7024c4cecbd467e2e00751b0c599436207d --- /dev/null +++ b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/full.md @@ -0,0 +1,310 @@ +# Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency + +Alan Baade +The University of Texas at Austin +abaade@utexas.edu + +Changan Chen +Stanford University +cchangan@stanford.edu + +# Abstract + +Learning self-supervised visual correspondence is a long-studied task fundamental to visual understanding and human perception. However, existing correspondence methods largely focus on small image transformations, such as object tracking in high-framerate videos or learning pixel-to-pixel mappings between images with high view overlap. This severely limits their application in dynamic multi-view settings such as robot imitation learning. In this work, we introduce Predictive Cycle Consistency for learning object correspondence between extremely disjoint views of a scene without paired segmentation data. Our technique bootstraps object correspondence pseudolabels from raw image segmentations using conditional grayscale colorization and a cycle-consistency refinement prior. We then train deep ViTs on these pseudolabels, which we use to generate higher-quality pseudolabels and iteratively train better correspondence models. We demonstrate the performance of our method under both extreme in-the-wild camera view changes and across large temporal gaps in video. Our approach beats all prior supervised and prior SoTA self-supervised correspondence models on the EgoExo4D correspondence benchmark (+6.7 IoU Exo Query) and the prior SoTA self-supervised methods SiamMAE and DINO V1&V2 on the DAVIS-2017 and LVOS datasets across large frame gaps. Code at https://github.com/AlanBaade/PredictiveCycleConsistency + +# 1. Introduction + +At the core of learning is the discovery of recurring patterns. This is the purpose of visual correspondence: given multiple inputs to a scene—whether different camera angles, video frames, or other perspectives—how can we determine which objects are the same and which are different? + +In recent years, the introduction of large-scale datasets and powerful model architectures has led to strong results in learning visual correspondence without the need for costly labeled object pairings. However, existing self-supervised object correspondence methods [18, 29, 56, 58] + +![](images/541c841f091dfedb7586f21020529a53edb33ab1b1c9094cbbc566619122b993.jpg) +Figure 1. We extract cycle-consistent correspondence at the object level to guide self-supervision across extreme view changes + +have overwhelmingly focused on domains with continuous or small transformations, such as Video Object Segmentation (VOS) in continuous videos, or dense pixel-top-pixel methods which necessarily assume that object surfaces overlap between views. This limits the application of self-supervised correspondence to discontinuous inputs, such as the first-and-third person EgoExo4D dataset [16], where objects are small, frequently interacted with, and camera pose or depth changes between views are unavailable. Learning correspondence in these challenging, real-world domains is crucial for tasks such as robot imitation learning [43, 46], object reidentification [13, 48], and scene understanding [16]. For example, in robot imitation learning, visual correspondence allows a robot to map observed actions from a third-person perspective to its own first-person actions. + +The fundamental challenge of learning visual correspondence across discontinuities is that a system must take into account an object and the surrounding environment holistically to answer the question of where that mug is instead of where $a$ mug is. A particularly powerful approach for encouraging a holistic understanding in difficult data sources is predictive learning: training a model to de-corrupt a corrupted input. Predictive learning has shown strong results in a wide range of domains, such as natural language processing [9, 40], image classification [2, 22], and generative modeling [23]. Recently, predictive approaches have seen competitive results in object correspondence and representation learning through tasks such as image-conditioned grayscale coloration [56] and conditional masked autoencoders [18, 38]. + +![](images/8d028ba859c9413d48b16ce5be93c40ab86da839a127591b8659c6bd4303823e.jpg) + +![](images/f4a0ec32d18a51a55a63a9779a83c082f0e6ca68217094ae6d6aadc5d701c558.jpg) + +![](images/00a12ae831b091f63d58c53792e1fbaaafdfd27795868cc0519edf034264ca94.jpg) + +![](images/df3656a997d9ada79d535bd0c5118d440651855166ed7f922525bb85da8910ed.jpg) +Figure 2. Self-supervised object correspondence on EgoExo4D between Ego views (bottom row) and Exo views (top row) without any paired data. Objects with no match are not colored. + +![](images/82283991036de7573bf31ae550a504e0bae7d7d70800248f579f29ce532edc12.jpg) + +![](images/7b48dbeda2417fc786f0a79f2eb8512e063325377caf0f5ae9d9c1854168f33e.jpg) + +However, existing predictive methods for correspondence struggle to generalize effectively across challenging scenarios (Table 1). State-of-the-art approaches typically rely on K-Nearest-Neighbor matching of emergent learned representations, identifying corresponding regions by comparing the most similar model embeddings across views. Although the pretraining objectives of these methods promote spatial awareness within a scene, the resulting representations often entangle semantic object information with the spatial cues necessary for correspondence. This entanglement reduces robustness against semantically similar distractor objects, limiting generalization and performance in more demanding scenarios. + +An effective approach for encouraging a spatially-consistent mapping between views is cycle consistency, which leverages the fact that correspondence is generally invertible across time and view changes [49, 58, 66]. Objects that can be tracked forward in a video can also be tracked backward; camera view changes can be undone. Existing applications of cycle consistency bootstrap long-term correspondence relations via local techniques like optical flow or video palindromes. However, these techniques rely on continuous input data such as high-framerate videos and require hand-crafted biases such as local neighborhoods to restrict a large search space. As a result, cycle consistency approaches have fallen out of favor compared to more modern approaches in challenging scenarios. + +In this work, we propose Predictive Cycle Consistency (PCC) for the task of self-supervised object correspondence under extreme viewpoint changes in space (camera angle changes) and time (gaps between video frames). Our approach extends cycle-consistency to operate at the object-mask level rather than focusing on features or image patches, allowing for robust correspondence in challenging scenarios. We do this by building on existing predictive approaches [18] and [56] to train a directional correspondence + +model on the simple, asymmetric task of grayscale colorization of a target image conditioned on a colorful source view. Using this, we extract object pairs that correspond to each other when running the correspondence model in both directions—source to target and target to source—to generate high-quality paired pseudolabels. We then train deep ViTs on these pseudolabels, which we use to generate higher-quality pseudolabels and iteratively train better correspondence models. All together, our approach combines the power of deep transformers trained with a predictive pretext task for scene understanding with a method to extract refined object boundaries. + +Our approach learns object correspondences across a wide range of challenging scenarios, such as matching highly occluded objects (apron and hands, Fig. 2 left) and distinguishing between semantically similar objects (pieces of paper from a COVID test, Fig. 2 middle). We validate our method on several object correspondence benchmarks, achieving superior performance over all previous labeled approaches and state-of-the-art unlabeled correspondence methods on the EgoExo4D Object correspondence benchmark [16]. Additionally, our approach sets a new state of the art on widely used video tracking datasets DAVIS-2017 [39] and LVOS [24] under high viewpoint-change conditions. Our contributions are as follows: + +1. We introduce Predictive Cycle Consistency, a technique that combines the powerful representation learning of predictive approaches with the refinement of cycle-consistency for self-supervised correspondence. +2. We propose a pseudo-labeling method that incorporates cycle-consistency on top of existing correspondence models to iteratively refine self-supervised object correspondence outputs. +3. We obtain state-of-the-art results on a suite of correspondence tasks from EgoExo4D, DAVIS, and LVOS. + +# 2. Related Work + +Dense Visual Correspondence. The visual correspondence task is fundamental to human visual perception and has a long established history in computer vision. While our work focuses on object-mask level correspondence, early correspondence techniques focused on dense pixel-to-pixel correspondence using classical techniques such as optical flow [3, 25, 28, 31, 36, 44, 50, 52, 55, 61]. In recent years, dense correspondence learning has advanced with deep learning methods that address challenging conditions, including significant viewpoint changes. However, existing dense methods for correspondence either require large amounts of costly labeled data to train, use camera pose [57], depth information [11, 42], or extract annotations through simulated data which limits the diversity of available environments [53, 54]. Furthermore, implicit to dense correspondence is a restriction on allowed view changes: + +different views must contain overlapping object surfaces. As a result, our work focuses on the more semantically grounded task of object level correspondence given segmentation masks. + +Self-Supervised Object Correspondence Human vision excels at establishing visual correspondences across space and time without direct supervision, even in the face of occlusions, distractor objects, and object transformations. Inspired by this capability, much work has focused on self-supervised learning of object correspondence through video and scene data. + +Contrastive approaches [21, 65] for object tracking learn correspondence by creating ground-truth annotations through applying strong image transformations. This powerful inductive bias has been extraordinarily effective in tracking, but these approaches are criticized for the large amount of engineering and hand-crafting necessary to create robust augmentations. Cycle-consistency [49, 58, 66] emerged as an early technique for learning deep self-supervised object correspondence by leveraging the invertibility of object correspondence through techniques such as training on video palindromes. However, cycle-consistency as the sole training objective has struggled on its own to maintain stability over long videos or through occlusions, causing it to be replaced by more powerful representation learning techniques [27, 65]. Tracking-by-Matching approaches are a recent introduction to learning zero-shot video instance segmentation. These approaches build on top of supervised image segmentation models [60] to create object-level representations. These representations either transfer from labeled data across domains [7] or are learned through maintaining invariance under hand-crafted augmentations [34, 35]. + +Visual Representation Learning for Correspondence Current state-of-the-art approaches for both supervised and unsupervised correspondence learning rely on strong base models to construct clean visual representations. In the past few years, these base models have moved away from ResNets [20] towards larger models such as Vision Transformers (ViTs) [10] trained using self-supervised representation learning strategies on large datasets. For example, the DINO model family [4, 8, 37] emphasizes noncontrastive invariance under augmentation. Other methods use exponentially moving average models [1, 17] or siamese networks [5]. In cases involving multiple views, largescale contrastive models effectively align embedding spaces across domains, achieving strong results in view-invariant learning and language-image matching [14, 26, 41, 62]. + +Masked Autoencoding [22] trains high-quality visual representations by randomly masking patches of an image at a high ratio and reconstructing the missing areas. A recent advancement in this approach for correspondence learning is SiamMAE [18], which adapts the infilling task + +![](images/8e174816df81872f9aaa3570e40eca535d4d854ac1300e700115f60160620f66.jpg) +Figure 3. High-Level Approach. We 1) Enumerate all objects in each view, 2) Create pseudolabels that are cycle-consistent according to a previous correspondence model iteration, and 3) Use these pseudolabels to train a new correspondence model. + +by asymmetrically masking a future frame and predicting it based on an unmasked past frame. This approach, also used in cross-view masked completion models [63, 64], has achieved state-of-the-art performance in generating robust representations for object correspondence. In this work, we extend existing masked autoencoder architectures to enhance correspondence across extreme viewpoint changes. + +# 3. Approach + +We tackle the task of discovering object correspondence between two images $\mathcal{I}_1 \in \mathbb{R}^{3 \times H_1 \times W_1}$ and $\mathcal{I}_2 \in \mathbb{R}^{3 \times H_2 \times W_2}$ of the same spatiotemporal scene. Specifically, we want to create a model taking in a binary segmentation mask $\mathcal{O}_1$ of an object in image $\mathcal{I}_1$ and outputs $\mathcal{O}_2$ , the corresponding object segmentation mask in $\mathcal{I}_2$ , or indicates that no correspondence exists. + +At a high level, our correspondence pipeline (Fig. 3) operates by 1) Enumerating all objects in each image with existing segmentation models, 2) Using our proposed PCC technique to discover a subset of corresponding objects in each view, 3) Generating correspondence pseudolabels with these mined correspondences, and 4) Iteratively training new models and generating new pseudo labels to learn high-quality self-supervised visual correspondence. + +# 3.1. Image Segmentation + +We break down finding corresponding objects between two images into two subtasks: 1) Enumerate all objects in each image (Object detection and Image Segmentation) and 2) Discover which of these objects correspond to one another (Correspondence). This cascaded approach to correspondence has recently obtained state-of-the-art results in video object segmentation with low data or textual object descriptions [7, 35]. There is a great amount of prior work in object detection and image segmentation, with popular approaches being SAM [33] in the supervised setting and UnSAM [60] + +and CutLER [59] in the self-supervised setting. We follow prior object-centric correspondence work [7, 34, 35] and use SAM to segment an image into distinct object regions, although we suspect self-supervised approaches could be applied for a completely unsupervised pipeline. The rest of this paper focuses on the task of (2), learning object correspondence without labels given object segmentations. + +# 3.2. Grayscale Colorization + +Grayscale colorization [56] was an early yet remarkably effective method for self-supervised object tracking in continuous videos. However, as computational resources grew and advanced architectures like Vision Transformers (ViTs) became prevalent, grayscale colorization fell out of favor, replaced by more complex and general pretext tasks such as masked autoencoding [12, 18, 22, 30, 38, 51]. To address the challenge of learning correspondence under extreme discontinuities, we reinvestigate the grayscale colorization approach and find that it reemerges as strikingly successful technique. + +Our architecture for grayscale colorization is depicted in Figure 4 (a). Conditioned on our source image $\mathcal{I}_1$ we optimize to reconstruct $\mathcal{I}_2$ given its grayscale counterpart $\mathcal{I}_2^g$ . We adapt our architecture from prior correspondence work [18, 64] and use a two-stage ViT [10] that 1) passes $\mathcal{I}_1$ and $\mathcal{I}_2$ separately through an encoder module $\mathbf{E}$ to obtain $\mathbf{E}(\mathcal{I}_1), \mathbf{E}(\mathcal{I}_2^g)$ , and 2) passes the encoded grayscale image $\mathbf{E}(\mathcal{I}_2^g)$ into a decoder $\mathbf{D}$ that merges the colorful image conditioning using cross attention: $\mathbf{D}(\mathbf{E}(\mathcal{I}_1), \mathbf{E}(\mathcal{I}_2^g))$ . For optimization, we use a simple mean-squared-error loss in the RGB color space. + +# 3.3. Visual Correspondence from Colorization + +We now propose an approach to extract object correspondence from the inputs and outputs of conditional colorization models based on a simple observation: if we augment an object's color in the source view, we would expect the output of the colorization model to change the object's color in the target view as well. + +Our approach is demonstrated in Figure 5. Given a binary object segmentation map $\mathcal{O}_1$ in $\mathcal{I}_1$ we seek to obtain the corresponding segmentation mask $\mathcal{O}_2$ in view $\mathcal{I}_2$ . We perform two forward passes of our grayscale colorization model, $\mathbf{F}$ . First, we perform grayscale cross-view infilling of $\mathcal{I}_2^g$ given a colorized $\mathcal{I}_1$ from before: $y = \mathbf{F}(\mathcal{I}_2^g,\mathcal{I}_1)$ . For our second model pass, we augment $\mathcal{I}_1$ at the location of the segmentation mask $\mathcal{O}_1$ by adding a constant color offset to each channel. We denote this augmented image as $\mathcal{I}_1'$ , and the RGB augmentation as a vector $c\in \mathbb{R}^3$ . We then colorize $\mathcal{I}_2^g$ conditioned on this augmented image $y^\prime = \mathbf{F}(\mathcal{I}_2^g,\mathcal{I}_1')$ . To identify where objects correspond, we then take the average absolute difference between $y$ and $y^\prime$ over each color channel and normalize to sum to one, outputting an object + +segmentation heatmap $\mathcal{H}\in \mathbb{R}_+^{H_2\times W_2}$ + +$$ +\mathcal {H} _ {i j} = \frac {\sum_ {c = 1} ^ {3} | \mathbf {F} (\mathcal {I} _ {2} ^ {g} , \mathcal {I} _ {1}) _ {c i j} - \mathbf {F} (\mathcal {I} _ {2} ^ {g} , \mathcal {I} _ {1} ^ {\prime}) _ {c i j} |}{\sum_ {c , k , l} | \mathbf {F} (\mathcal {I} _ {2} ^ {g} , \mathcal {I} _ {1}) _ {c k l} - \mathbf {F} (\mathcal {I} _ {2} ^ {g} , \mathcal {I} _ {1} ^ {\prime}) _ {c k l} |} \quad (1) +$$ + +Importantly, this approach is entirely blackbox with respect to the grayscale colorization model. As a result, the spatial awareness of objects captured during grayscale model pretraining naturally emerges during extraction. This is not necessarily true for masked autoencoder approaches like SiamMAE, where internal model representations are used to extract correspondences that may not sufficiently encode cross-view information to disentangle semantically similar distractors. + +# 3.4. Generating Cycle-Consistent Correspondence + +In this section, our goal is to create a set of pseudolabels containing corresponding objects between two images $\mathcal{I}_1$ and $\mathcal{I}_2$ given their respective object segmentations $\{\mathcal{O}_{1,i}\}_{i = 1}^{N_1}$ and $\{\mathcal{O}_{2,j}\}_{j = 1}^{N_2}$ output by SAM, where $\mathcal{I}_1,\mathcal{I}_2$ have $N_{1},N_{2}$ segmented objects respectively. Following Sec. 3.3, we generate an augmented $\mathcal{I}_{1,i}^{\prime}$ for each segmentation mask $\mathcal{O}_{1,i}$ and calculate the output correspondence heatmap in $\mathcal{I}_2$ , which we write as $\mathcal{H}_i^{1\rightarrow 2}$ . Here, " $1\rightarrow 2"$ represents predicting $\mathcal{I}_2$ conditioned on $\mathcal{I}_1$ . To reduce variance, we use the same augmentation vector $c$ for each augmentation. By caching the unaugmented pass, calculating all $\mathcal{H}_i^{1\rightarrow 2}$ takes $N_{1} + 1$ forward passes. + +We then define the similarity between $\mathcal{H}_i^{1\to 2}$ and $\mathcal{O}_{2,j}$ to be the amount of weight $\mathcal{H}_i^{1\to 2}$ places on the segmentation mask of $\mathcal{O}_{2,j}$ , equal to the Frobenius inner product: + +$$ +\operatorname {S i m} \left(\mathcal {H} _ {i} ^ {1 \rightarrow 2}, \mathcal {O} _ {2, j}\right) = \left\langle \mathcal {H} _ {i} ^ {1 \rightarrow 2}, \mathcal {O} _ {2, j} \right\rangle_ {F} \tag {2} +$$ + +Then, we pair every object $\mathcal{O}_{2,j}$ with the most similar heatmap $\mathcal{H}_i^{1\to 2}$ , which we write as $\mathbf{P}_j^{1\to 2}$ : + +$$ +\mathbf {P} _ {j} ^ {1 \rightarrow 2} = \underset {i} {\arg \max } \operatorname {S i m} \left(\mathcal {H} _ {i} ^ {1 \rightarrow 2}, \mathcal {O} _ {2, j}\right) \tag {3} +$$ + +There is an important difference between taking the argmax over the source object augmentation heatmaps and taking the argmax over the objects in the target image. Outputs of grayscale colorization can be correlated with many factors such as object semantics (the sky is probably blue but a balloon can be any color) or lighting conditions, which may mean that some objects in the target view are more susceptible to having their predicted colors change with respect to augmented inputs than others. As a result, the object with the greatest color change in $I_{2}$ is often the same for many $H_{i}^{1\rightarrow 2}$ . By taking the most similar $H_{i}^{1\rightarrow 2}$ for each $\mathcal{O}_{2,j}$ , we normalize over the easiness of each object to be changed as the result of an augmentation in $I_{1}$ during grayscale colorization. + +![](images/d136995f6c886ad52578e7e1b897728f9349ff278af2c8c103d3d35fba920d24.jpg) +(a) + +![](images/759a13044394520f4dd1c18ecaa23764eb0bcb1e4b81d9fd804f452bf296082f.jpg) + +![](images/19c10e2e7a5b81c328866e7fb14da2fd4f7da64be5d391a92727afb8c4fe1fe4.jpg) +(b) +Figure 4. Model Architecture and Cycle Consistency. (a) Our architecture for grayscale color completion, using a shared Encoder for each view and a Decoder with Cross and Self Attention. (b) Our architecture for correspondence, which inserts a projection to represent the mask. (c) Our cycle-consistency pipeline. We run correspondence using the model from (a) or (b) on all objects in each view, find the best matching objects, and extract cycle-consistent matches. + +![](images/dd2f254322f788ccafecb4f62cba340c325908897284d31a5120a84356d8c5af.jpg) + +![](images/0dbe1d495441672d1a3f2d14fa793b28993f569ca7f41a06c3eaf87a6a10eb7d.jpg) +Figure 5. A simple method to extract object correspondence from a deep conditional grayscale colorization model. + +We repeat the above process in the reverse direction to extract $\mathcal{H}_j^{2\rightarrow 1}$ and $\mathbf{P}_i^{2\rightarrow 1}$ . Finally, we extract cycle-consistent pseudolabels at the object level by taking all pairs $(i,j)$ where $P_{j}^{1\rightarrow 2} = i$ and $P_{i}^{2\rightarrow 1} = j$ . This cycle consistency enforces that augmenting $O_{i}$ in $I_{1}$ leads to changes at $O_{j}$ in $I_{2}$ and augmenting $O_{j}$ in $I_{2}$ leads to changes at $O_{i}$ in $I_{1}$ during grayscale infilling, which strongly implies correspondence. We then use these extracted pseudolabels on each image to generate a dataset of labeled correspondence data with paired images and segmentation masks. + +# 3.5. Correspondence Model + +Our end goal is to have a correspondence model $\mathbf{G}$ than can take in a source image $\mathcal{I}_1$ , source object segmentation mask $\mathcal{O}_1$ , and target image $\mathcal{I}_2$ and output the corresponding object segmentation mask for the target image. To train such a model on our extracted pseudolabels or ground truth labeled data, we slightly modify our grayscale colorization + +architecture, shown in Fig 4 (b). We patchify the segmentation mask $\mathcal{O}_1$ , apply a linear projection, and add the output to $\mathbf{E}(\mathcal{I}_1)$ . The decoder then takes in the encoded target view $\mathbf{E}(\mathcal{I}_2)$ and uses cross attention to query the source view and source segmentation mask. As done in [16] we use a Dice Loss [47] and a BCE Loss to predict the target mask. + +# 3.6. Iterative Pseudolabeling and Training + +We repeat our pseudolabeling pipeline by substituting grayscale colorization with a trained correspondence model. Assuming that a correspondence model $\mathbf{G}$ outputs a binary mask, we replace the similarity score (Equation 2) with intersection over union (IoU) and make no other changes. We can then use these new pseudolabels to train new correspondence models, which in turn can generate new pseudolabels. In our experiments, we find that running three rounds of pseudolabeling (one grayscale, two iterative) saturates quality. + +# 4. Experiments + +For comprehensive evaluation, we train and evaluate our PCC pipeline from start to finish in two settings 1) high camera pose changes and 2) large temporal gaps in videos. We describe our training and evaluation approaches for each setting in Sec. 4.1 and present results in Sec. 4.2 and 4.3. + +# 4.1. Experimental Setup + +Datasets To train and evaluate the quality of PCC across extreme camera viewpoint changes, we use the EgoExo4D correspondence dataset [16]. EgoExo4D [16] builds on Ego4D [15] and captures over 1000 hours of complex human actions such as cooking and basketball from several + +time-synchronized camera views. The camera views consist of one egocentric (ego) view captured by person performing the action and one or more exocentric (exo) camera views that capture the holistic action environment from the third person. We train and evaluate on the object correspondence benchmark in EgoExo4D, which is unique across correspondence datasets in that it focuses on extremely challenging settings with many small objects, high occlusion, and extreme camera angle differences in a dynamic environment. Correspondence is evaluated in two directions, either with taking a query mask in the ego view and segmenting the corresponding object in a paired exo view (ego query), or the inverse task from the exo to the ego view (exo query). + +To run PCC on video object segmentation across large temporal gaps, we train on the commonly used Kinetics-400 dataset [32], consisting of large-scale action videos taken from the web. We evaluate temporal correspondence on DAVIS-2017 [39], which captures densely segmented short videos and LVOS [24], which measures video object segmentation across long dynamic videos with an average duration of over a minute. + +Backbone We implement our correspondence models using a ViT-B with a patch size of 16 to match the compute and parameter count of current state-of-the-art self-supervised correspondence work [18, 38]. We initialize our model from CroCo v2 [63, 64], which pretrains with the task of image conditioned cross-view autoencoding. + +Grayscale Model We train our grayscale colorization models separately for EgoExo4D and Kinetics-400. For EgoExo4D, for each ego view in the dataset, we 1) select a random corresponding exo camera, 2) choose a random synchronized frame from each view, and 3) choose with $50\%$ probability whether to grayscale the ego or exo view. For Kinetics-400, we follow SiamMAE and select two random frames between 4 and 48 frames apart and colorize the grayscale future frame conditioned on the past frame. We train each grayscale colorization model for 60k updates with a batch size of 256, which corresponds to 200 epochs for EgoExo4D and 50 epochs for Kinetics-400. Additional hyperparameters are in Appendix 6.1. + +Predictive Cycle Consistency We then run Predictive Cycle Consistency to extract paired pseudolabels. We use the ViT-H version of SAM to run image segmentation on video frames to extract $\{\mathcal{O}_{1,i}\}_{i = 1}^{N_1}$ and $\{\mathcal{O}_{2,j}\}_{j = 1}^{N_2}$ . For EgoExo4D, we segment every labeled (ego,exo) frame pair in the correspondence benchmark, which generally amounts to one frame per second. For Kinetics-400, we segment two random frames per video selected 2 seconds apart, and we ablate the length between frames in our experiments. + +Correspondence Model For a fair comparison, we match the compute of the EgoExo4D supervised baselines and train our correspondence models on 5M supervised examples, divided into 10k updates with a batch size of 512. + +We train for the same duration on Kinetics-400, and additional hyperparameters are provided in Appendix 6.1. For a given input example, we select a random cycle-consistent pseudolabel as a training target, and with a $50\%$ chance select whether to train with an ego or exo query. On $25\%$ of inputs, we replace the correspondence objective with learning correspondence existence accuracy, where we select a uniformly random SAM image segment, label it positive if it is also a cycle-consistent pseudolabel and negative otherwise. + +Baselines To the best of our knowledge, this is the first work to tackle self-supervised correspondence across the uniquely extreme viewpoint changes on EgoExo4D. As a result, we are rigorous about reimplementing the prior SoTA open source correspondence models. As demonstrated in the SiamMAE paper [18], the current best approaches for correspondence without labeled data are deep ViT models. In particular, SiamMAE makes two key observations 1) a small patch size (ViT-X/8) results in significantly stronger correspondence results (+9.5 on DAVIS-17 for ViT-S/8 vs SiamMAE ViT-S/16) and 2) despite deep work in prior correspondence approaches, DINO v1 [4], which focuses on learning robust and general visual representations at scale, outperforms all approaches except SiamMAE. As a result, for prior work we compare against 1) SiamMAE ViT-S/8, 2) SoTA DINO v1 and v2 models, and 3) CroCo v2 (to compare versus our baseline model architecture). For fairness, we continually pretrain SiamMAE and CroCO v2 on EgoExo4D. Additional implementation details are in Appendix 6. Furthermore, although our model outputs segmentations without using SAM at inference, we do use image segmentation for pseudolabel generation. To account for this, we implement a setting where we use SAM ViT-H to select the object with the highest IoU. For the DAVIS-17 and LVOS baselines, we similarly compare against the aforementioned models, however find that SAM did not help because it cut off objects (e.g. selecting a tire instead of a bike), which we discuss more in Appendix 6. + +# 4.2. Results Across Space + +We report quantitative results for PCC across extreme viewpoint changes on the EgoExo4D correspondence benchmark [16] in Table 1. Following the EgoExo4D baseline, we measure 1) IoU (denoted as $\mathcal{I}_m$ in VOS datasets), 2) Location Score, representing the mean-squared-error distance between the predicted and ground truth centroid, 3) Contour Accuracy (CA, denoted as $\mathcal{F}_m$ in VOS datasets), and 4) object presence balanced accuracy (Bal. Acc.). We use the official code for EgoExo4D for evaluation, and run all baselines at $480\mathrm{p}$ , as done in both [16] and [18]. + +Our results demonstrate that PCC outperforms all prior labeled and all prior SoTA self-supervised approaches across all metrics. We observe that the DINO family + +Table 1. Results on of the EgoExo4D [16] correspondence benchmark (v1, test set). $\diamond$ : Continual Pretraining on EgoExo4D $\clubsuit$ : Models with access to multiple frames per view. Bold is best, underlined is second best. Top labeled, middle and bottom self-supervised. + +
MethodBackboneEgo QueryExo Query
Bal. Acc. ↑IoU ↑Loc. Score ↓CA. ↑Bal. Acc. ↑IoU ↑Loc. Score ↓CA. ↑
XSegTx w/o finetuning [16]SegSwap [45] + ViT-B∅0.600.1160.017∅1.620.1970.027
XSegTx [16]SegSwap [45] + ViT-B62.6313.880.1540.23974.621.80.1330.265
XMem* [6]ResNet-50 [20] + Memory42.3313.070.3120.18256.9610.20.2490.125
XView-XMem* [16]XMem + ViT-B53.2822.140.1760.32559.3623.560.1860.308
XView-XMem (+ XSegTx)* [16]XMem + SegSwa + ViT-B54.6122.50.1390.34752.2819.390.2080.255
Ours SupervisedViT-B/1674.738.410.0370.60388.4543.700.0490.555
Ours Supervised + PCCViT-B/1676.939.010.0330.60087.2347.060.0540.590
SiamMAE* [18]ViT-S/8∅12.240.1700.185∅14.180.1590.198
CrocoV2* [63]ViT-B/16∅7.140.2000.138∅9.560.1640.136
DINO [4]ViT-B/8∅12.550.1530.178∅15.940.1370.246
DINOv2+Registers [37] [8]ViT-B/14∅20.260.1250.299∅24.60.1690.307
SiamMAE* +SAMViT-S/8+ViT-H∅17.970.1800.254∅24.050.1430.315
DINOv2+Registers+SAMViT-B/14+ViT-H∅28.920.1530.365∅34.780.1230.433
Grayscale Coloration + SAM*ViT-B/16+ViT-H∅20.820.1100.311∅19.500.1090.276
PCC Iter 1ViT-B/16∅26.410.0850.396∅34.350.0900.436
PCC Iter 2ViT-B/1665.2229.980.0830.44666.4040.410.0790.502
PCC Iter 3ViT-B/1660.6629.890.0940.43267.9041.450.0710.508
+ +of models, which learns robustness against strong image augmentations, outperforms the predictive SiamMAE and CroCo V2 approaches. Although applying SAM at inference time strongly improves self-supervised approaches in IoU and Contour Accuracy, it harms spatial accuracy as measured by Location Score, where PCC dramatically outperforms all prior work. We find that our self-supervised method performs relatively better in the Exo Query setting, where input objects are small, but output objects are generally larger. We find that iteratively training PCC strongly improves results, although performance largely saturates after the second iteration. + +We further apply our correspondence model architecture ("Ours" Sec. 3.5) on the labeled training data split of EgoExo4D, achieving state-of-the-art results compared to prior supervised approaches. When combining the supervised data from EgoExo4D with PCC Iteration 3 pseudolabels, we observe improvements over the supervised baseline, particularly in the Exo Query setting. This suggests that PCC captures complementary information to the EgoExo4D labels and can improve object correspondence even in high-data labeled regimes. + +We show qualitative results of PCC on EgoExo4D in 6. As shown in rows 3 and 4, our model can locate corresponding objects even under significant occlusions. However, it sometimes struggles to produce fine-grained masks. Examples of pseudolabels on EgoExo4D are visible in 2. + +# 4.3. Results Across Time + +We present quantitative results for PCC across large temporal gaps between video frames on the DAVIS-17 dataset [39] in Table 2. We follow the evaluation methodology in [11] and evaluate correspondences on all video frame pairs + +![](images/c25c617158a5c04858d1e5b4f0defb7b3bfcc0f3f278c1c469a9afc6354708c8.jpg) +Figure 6. Qualitative results of PCC on the EgoExo4D correspondence benchmark. + +in DAVIS with a temporal gap of 20 frames. Evaluation is restricted to only cover objects present in both views, and report $\mathcal{J} \& \mathcal{F}_m$ , $\mathcal{I}_m$ , and $\mathcal{F}_m$ . We report quantitative results on LVOS in Figure 7. Enabled by the large lengths of LVOS videos, we report results across a range of temporal gaps between 5 frames (the LVOS annotation rate) and 400 frames. We again restrict evaluation to objects present in both views and report $\mathcal{J} \& \mathcal{F}_m$ . + +On DAVIS, we find that PCC significantly outperforms the prior SOTA approaches for VOS. Similar to the EgoExo setting, PCC shows strong improvements after the first it- + +Table 2. SoTA Comparison of Video Object Correspondence Methods on DAVIS-17 [39] Val with a temporal gap of 20 frames + +
MethodJ&Fm↑Jm↑Fm
SiamMAE [18]60.758.462.9
CrocoV2 + Cont. Pretrain [63]40.037.442.5
Probalistic Warp Consistency [54]42.942.642.7
DINO ViTs/8 [4]64.561.667.5
DINO ViTb/8 [4]66.463.769.2
DINOv2 + Reg ViTb/14 [8]62.159.664.8
PCC Iter 164.461.367.5
PCC Iter 269.767.072.4
PCC Iter 370.267.872.7
+ +![](images/3dc65095e8d27dfc573b17db0d0ac29c1b568d95b5bdbdba113026ce5f20a213.jpg) +Figure 7. Results on LVOS-V1 Val at different temporal gaps. + +eration, but performance saturates after the second. On LVOS, although SiamMAE achieves SoTA for temporal gaps of 5 and 10 frames, PCC shows comparatively better performance as the frame gap increases, achieving SoTA results for all other temporal gaps. Interestingly, SiamMAE sees a sharper decline as the frame gap widens compared to DINO. We suspect this suggests that at shorter frame gaps, precise object segmentation is most critical, whereas at larger frame gaps, robustness to object semantics becomes more important. + +![](images/8133fd45171055140f7d671c5cedd6e4b3b58403210506ae2bd6b955483c60b6.jpg) +Figure 8. Qualitative results on correspondence across time. Representational approaches degrade quickly with larger frame gaps. + +We demonstrate qualitative results for object recognition across temporal gaps in Fig. 8. We observe that representational approaches such as DINO struggle to maintain object consistency (bottom row), handle semantically similar distractors with moderate movement (third row), and manage heavy object occlusions (second row). In contrast, PCC maintains high-quality correspondence throughout. + +# 4.4. Ablations + +Table 3. Ablation: Model configuration Supervised Training, EgoExo4D Validation Set. Grayscale continued training harms supervised training. A deeper decoder improves performance. + +
PretrainingDec. Params.Ego/Exo Query
IoU ↑Loc. Score ↓CA. ↑
MAE50M36.9/40.20.043/0.0780.57/0.51
MAE100M37.7/43.70.042/0.0600.60/0.56
Grayscale100M31.0/34.30.048/0.0890.49/0.43
+ +Table 4. Ablation: Frame Gap Davis 2017 Validation Set. We ablate the framerate gap on Kinetics-400 during pseudolabel extraction on Iteration 3. + +
Pseudolabel Temporal Gap2 FramesJ&Fm10 Frames30 Frames
2 seconds83.975.865.4
4 seconds83.576.266.3
6 seconds83.776.267.7
+ +We ablate our model configuration on supervised EgoExo4D correspondence in Table 3. Our results show that pretraining with a grayscale objective on EgoExo4D, rather than a MAE objective, significantly degrades performance. This suggests that the gains from PCC are not due to better representation learning, but rather arise from the full cycle-consistent pipeline. + +We investigate the impact of the temporal gap between images for pseudolabeling PCC on Kinetics-400 in Table 4. We find that training with a longer gap improves longer-term correspondence. + +# 5. Conclusion + +In this work, we address the task of learning self-supervised visual correspondence across significant camera view shifts and large temporal gaps in video. Our approach, Predictive Cycle Consistency, achieves state-of-the-art performance on the EgoExo4D correspondence benchmark and in low-frame-rate scenarios on the DAVIS-17 and LVOS datasets. Furthermore, we show that applying cycle consistency to correspondence predictions at the object level is an strikingly effective strategy for generating clean and consistent correspondence pseudolabels. Overall, our work extends the capabilities of self-supervised learning to establish correspondence in highly disjoint and challenging domains, paving the way for improved understanding of dynamic scenes in the natural world. + +Acknowledgments: This project was supported by the Vision Lab at UT Austin. We would like to thank Kristen Grauman for her valuable discussion and feedback on the paper. + +# References + +[1] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. 3 +[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1 +[3] Steven S. Beauchemin and John L. Barron. The computation of optical flow. ACM computing surveys (CSUR), 27(3):433-466, 1995. 2 +[4] 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. 3, 6, 7, 8 +[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3 +[6] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 7 +[7] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In ICCV, 2023. 3, 4, 2 +[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 3, 7, 8 +[9] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1 +[10] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 4 +[11] Mohamed El Banani, Ignacio Rocco, David Novotny, Andrea Vedaldi, Natalia Neverova, Justin Johnson, and Ben Graham. Self-supervised correspondence estimation via multiview registration. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1216-1225, 2023. 2, 7 +[12] Alexandre Eymäel, Renaud Vandeghen, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Efficient image pre-training with siamese cropped masked autoencoders. arXiv:2403.17823, 2024. 4 +[13] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In Proceedings of the IEEE/CVF conference on + +computer vision and pattern recognition, pages 5374-5383, 2019. 1 +[14] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 3 +[15] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino 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. 5 +[16] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. 1, 2, 5, 6, 7 +[17] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3 +[18] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 1, 2, 3, 4, 6, 7, 8 +[19] Bumsub Ham, Minsu Cho, Cordelia Schmid, and Jean Ponce. Proposal flow: Semantic correspondences from object proposals. IEEE transactions on pattern analysis and machine intelligence, 40(7):1711-1725, 2017. 2 +[20] 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 +[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3 +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 3, 4 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[24] Lingyi Hong, Wenchao Chen, Zhongying Liu, Wei Zhang, Pinxue Guo, Zhaoyu Chen, and Wenqiang Zhang. Lvos: A benchmark for long-term video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13480-13492, 2023. 2, 6 + +[25] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185-203, 1981. 2 +[26] Alexandros Iosifidis, Anastasios Tefas, and Ioannis Pitas. View-invariant action recognition based on artificial neural networks. IEEE transactions on neural networks and learning systems, 23(3):412-424, 2012. 3 +[27] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 3 +[28] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European conference on computer vision (ECCV), pages 690-706, 2018. 2 +[29] Zhenyu Jiang, Hanwen Jiang, and Yuke Zhu. Doduo: Dense visual correspondence from unsupervised semantic-aware flow. In arXiv preprint arXiv:2309.15110, 2023. 1 +[30] Zhouqiang Jiang, Bowen Wang, Tong Xiang, Zhaofeng Niu, Hong Tang, Guangshun Li, and Liangzhi Li. Concatenated masked autoencoders as spatial-temporal learner. arXiv preprint arXiv:2311.00961, 2023. 4, 1 +[31] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 557-572. Springer, 2020. 2 +[32] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6, 1 +[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3 +[34] Liulei Li, Wenguan Wang, Tianfei Zhou, Jianwu Li, and Yi Yang. Unified mask embedding and correspondence learning for self-supervised video segmentation. In CVPR, 2023. 3, 4, 2 +[35] Siyuan Li, Lei Ke, Martin Danelljan, Luigi Piccinelli, Mattia Segu, Luc Van Gool, and Fisher Yu. Matching anything by segmenting anything. CVPR, 2024. 3, 4, 2 +[36] Pengpeng Liu, Michael Lyu, Irwin King, and Jia Xu. Selfflow: Self-supervised learning of optical flow. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4571-4580, 2019. 2 +[37] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 7, 1 + +[38] Gensheng Pei, Tao Chen, Xiruo Jiang, Huafeng Liu, Zeren Sun, and Yazhou Yao. Videomac: Video masked autoencoders meet convnets. In CVPR, 2024. 1, 4, 6 +[39] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 2, 6, 7, 8, 1 +[40] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 1 +[41] 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 +[42] Tanner Schmidt, Richard Newcombe, and Dieter Fox. Self-supervised visual descriptor learning for dense correspondence. IEEE Robotics and Automation Letters, 2(2):420-427, 2017. 2 +[43] Jinghuan Shang and Michael S Ryoo. Self-supervised disentangled representation learning for third-person imitation learning. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 214-221. IEEE, 2021. 1 +[44] Xi Shen, François Darmon, Alexei A. Efros, and Mathieu Aubry. Ransac-flow: Generic two-stage image alignment. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV, page 618–637, Berlin, Heidelberg, 2020. Springer-Verlag. 2 +[45] Xi Shen, Alexei A Efros, Armand Joulin, and Mathieu Aubry. Learning co-segmentation by segment swapping for retrieval and discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5082-5092, 2022. 7 +[46] Bradley C Stadie, Pieter Abbeel, and Ilya Sutskever. Third-person imitation learning. arXiv preprint arXiv:1703.01703, 2017. 1 +[47] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MIC-CAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240–248. Springer, 2017. 5 +[48] Hao Tang, Kevin J Liang, Kristen Grauman, Matt Feiszli, and Weiyao Wang. Egotracks: A long-term egocentric visual object tracking dataset. Advances in Neural Information Processing Systems, 36, 2024. 1 +[49] Yansong Tang, Zhenyu Jiang, Zhenda Xie, Yue Cao, Zheng Zhang, Philip HS Torr, and Han Hu. Breaking shortcut: Exploring fully convolutional cycle-consistency for video correspondence learning. arXiv preprint arXiv:2105.05838, 2021. 2, 3 +[50] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV + +2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2 +[51] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In Advances in Neural Information Processing Systems, 2022. 4 +[52] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6258-6268, 2020. 2 +[53] Prune Truong, Martin Danelljan, Fisher Yu, and Luc Van Gool. Warp consistency for unsupervised learning of dense correspondences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10346-10356, 2021. 2 +[54] Prune Truong, Martin Danelljan, Fisher Yu, and Luc Van Gool. Probabilistic warp consistency for weakly-supervised semantic correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8708-8718, 2022. 2, 8 +[55] Prune Truong, Martin Danelljan, Radu Timofte, and Luc Van Gool. Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10247-10266, 2023. 2 +[56] Carl Vondrick, Abhinav Shrivastava, Alireza Fathi, Sergio Guadarrama, and Kevin Murphy. Tracking emerges by colorizing videos. In Proceedings of the European conference on computer vision (ECCV), pages 391-408, 2018. 1, 2, 4 +[57] Qianqian Wang, Xiaowei Zhou, Bharath Hariharan, and Noah Snively. Learning feature descriptors using camera pose supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 757-774. Springer, 2020. 2 +[58] Xiaolong Wang, Allan Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. In CVPR, 2019. 1, 2, 3 +[59] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 4 +[60] XuDong Wang, Jingfeng Yang, and Trevor Darrell. Segment anything without supervision. NeurIPS 2024, 2024. 3 +[61] Yang Wang, Yi Yang, Zhenheng Yang, Liang Zhao, Peng Wang, and Wei Xu. Occlusion aware unsupervised learning of optical flow. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4884-4893, 2018. 2 +[62] Daniel Weinland, Mustafa Özuysal, and Pascal Fua. Making action recognition robust to occlusions and viewpoint changes. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part III 11, pages 635-648. Springer, 2010. 3 +[63] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela + +Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. CroCo v2: Improved Cross-view Completion Pretraining for Stereo Matching and Optical Flow. In ICCV, 2023. 3, 6, 7, 8, 1 +[64] Weinzaepfel, Philippe and Leroy, Vincent and Lucas, Thomas and Brégier, Romain and Cabon, Yohann and Arora, Vaibhav and Antsfeld, Leonid and Chidlovskii, Boris and Csurka, Gabriela and Revaud Jérôme. CroCo: Self-Supervised Pre-training for 3D Vision Tasks by Cross-View Completion. In NeurIPS, 2022. 3, 4, 6 +[65] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10075-10085, 2021. 3 +[66] Tinghui Zhou, Philipp Krahenbuhl, Mathieu Aubry, Qixing Huang, and Alexei A Efros. Learning dense correspondence via 3d-guided cycle consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 117-126, 2016. 2, 3 \ No newline at end of file diff --git a/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/images.zip b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..394c0ded4b4d3973b9b854f593d59b9cc3c15f88 --- /dev/null +++ b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43106d821f46acb02a334917546690996f34cc1af75358d15732eca33a3f8c40 +size 610910 diff --git a/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/layout.json b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a5082ff6aac3ce436de520dc5f4cc75ff3ffcf7d --- /dev/null +++ b/CVPR/2025/Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a13bae3379db1251e6f49d9901e03459dc16cc70de4a6795278efd5de1cd702 +size 411612 diff --git a/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/3fdda516-00e3-4291-8037-b72ad987992c_content_list.json b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/3fdda516-00e3-4291-8037-b72ad987992c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..26eeacf1a6dd9c787263f9582e7429f89ea4deb6 --- /dev/null +++ b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/3fdda516-00e3-4291-8037-b72ad987992c_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fe5708f2e260f997e0dbb85ccf78da6728a74c2555b65c865c91efcde9e107b +size 79818 diff --git a/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/3fdda516-00e3-4291-8037-b72ad987992c_model.json b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/3fdda516-00e3-4291-8037-b72ad987992c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9e131136e82b0a71a6910514737e43088df41d34 --- /dev/null +++ b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/3fdda516-00e3-4291-8037-b72ad987992c_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8262d37ef557b68e139cf38369b9827d8e6239665905f797407ded72d747ad95 +size 96869 diff --git a/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/3fdda516-00e3-4291-8037-b72ad987992c_origin.pdf b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/3fdda516-00e3-4291-8037-b72ad987992c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..140f7546e4e7848eba4e36bb4ab5b6f27ad9803f --- /dev/null +++ b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/3fdda516-00e3-4291-8037-b72ad987992c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42b9867308a27d6cd164667ca3f45eabc5763bf1d632d58051aad66b988fb574 +size 8403541 diff --git a/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/full.md b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a78ffa3520fed28e319bac180c537cf1b0e19b19 --- /dev/null +++ b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/full.md @@ -0,0 +1,404 @@ +# Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration + +Aocheng Li1, James R. Zimmer-Dauphinee2, Rajesh Kalyanam1, Ian Lindsay1, Parker VanValkenburgh3, Steven Wernke2, Daniel Aliaga1, Purdue University, Vanderbilt University, Brown University + +li3922@purdue.edu, james.r.zimmer-dauphinee@vanderbilt.edu, rkalyana@purdue.edu +ilindsay@purdue.edu, parker.vanvalkenburgh@brown.edu, s.wernke@vanderbilt.edu, aliaga@purdue.edu + +![](images/e67ca49fbef3e6ef83defe6290c3ef318ec8d5bdc9c7e319b1bb3b88080ef837.jpg) +Figure 1. Restoration results of applying our method on a large structure consisting of 15 million points from Mawchu Llacta in Peru. The structure is completed and displayed over the original site for visualization. + +![](images/8225aeee0de8bcbdd3c31c7556f2c9031807a03578acad168a330ca26f185d23.jpg) + +# Abstract + +Point cloud completion helps restore partial incomplete point clouds suffering occlusions. Current self-supervised methods fail to give high fidelity completion for large objects with missing surfaces and unbalanced distribution of available points. In this paper, we present a novel method for restoring large-scale point clouds with limited and unbalanced ground-truth. Using rough boundary annotations for a region of interest, we project the original point clouds into a multiple-center-of-projection (MCOP) image, where fragments are projected to images of 5 channels (RGB, depth, and rotation). Completion of the original point cloud is reduced to inpainting the missing pixels in the MCOP images. Due to lack of complete structures and an unbalanced distribution of existing parts, we develop a self-supervised scheme which learns to infill the MCOP image with points resembling existing "complete" patches. Special losses are applied to further enhance the regularity and consistency of completed MCOP images, which is mapped back to 3D to form final restoration. Extensive experiments demonstrate the superiority of our method in completing $600+$ incom + +plete and unbalanced archaeological structures in Peru. + +# 1. Introduction + +Point clouds are important for a wide range of vision applications including autonomous driving, real-estate walk-throughs, pre/post-disaster understanding, and digital archaeology and historic preservation (e.g., [6, 13, 26, 35]). Effective algorithms for learning and understanding point clouds have been developed and validated on synthetic clean shapes [27, 39, 45]. However, for modeling large outdoor real world objects raw point clouds are sparse, noisy, and incomplete. The incompleteness stems from either occlusion (e.g., samples are behind a wall or tree) or missing surfaces (e.g., archaeological remains, post-disaster), which results in different probability distributions of point samples. Hence, point cloud completion for datasets that include significant missing surfaces/areas of large objects, and also lack ground-truth data, is a critical and necessary task. Such 3D reconstructions are vital to understanding the structuring of spaces within archaeological settlements and for modeling social dynamics in the past. + +To complete point clouds, prior works [4, 5, 16, 24, 37, 38] have attempted to fill-in the partial point clouds using supervised, unsupervised, or self-supervised learning methods. Supervised learning relies on the presence of groundtruth and labeled input-output pairs. Unsupervised learning omit labeled input-output pairs but still expect ground truth (e.g., examples of complete models spanning the distribution space). Self-supervised learning does not assume ground truth and generates supervision signals itself (e.g., additional self-masking [5, 24], consistency of multi-view partial inputs [16], or similarity of curvatures [38]) from fragmented data like KITTI [11] and MatterPort3D [3]. Self-supervision promises to be a good methodology to complete datasets of large point cloud with significant missing data and no ground truth (of the missing surfaces). + +In this work, we present a self-supervised point-cloud completion approach for datasets targeted to archaeological settings with a significant amount of missing surfaces due to weathering, erosion, physical stress and human activities [29], which would otherwise be extremely hard to restore physically to their original appearance. As opposed to prior methods (e.g.,[5, 16, 24, 38]), we are able to infer point clouds varying from $50\mathrm{k}$ to $2\mathrm{M}$ points, with up to $95\%$ missing surfaces, and to produce colors/textures. Our approach relaxes the typical uniform-probability assumption of missing points and instead extracts a probability from the partially provided data. This benefits point completion for more general, complex, and large-scale scenarios. + +Our method stores points clouds using multi-center-of-projection (MCOP) images and using patch-based self-supervised inpainting. Instead of processing the point cloud in full 3D, our approach uses an MCOP image to combine points observed from multiple viewpoints along a path. Such images are used to train a network using adversarial loss of sampled patches against complete patches to generate completed patches. The image-based representation scales well to point-cloud completion for millions of points (e.g., see Fig. 1 and Fig. 8). Further [35], we use Poisson reconstruction [17] on the completed point cloud to produce a 3D visualization. Our experiments show that our method not only out-performs previous self-supervised point cloud completion methods, but also yields finer geometric details than current self-supervised image inpainting methods for similar (archaeological) settings and produces high-quality completion of models with color data, even when far less than half of the point samples are available. + +Our main contributions include the following: + +(1) we provide a novel multiple-center-of-projection (MCOP) representation for point cloud data which unlocks the potential for large scale and high quality point cloud completion of extensive structures, +(2) we describe a patch-based point-cloud completion algorithm for a 3D scene encoded into a MCOP image repre + +![](images/078b1eb998506136e9d941ad7098aedf5fb36ed42914680ee38e90c55e6ccd3d.jpg) +(b) Pixel distribution for our archaeological sites from partial views. +Figure 2. Comparison of distribution of available points for incomplete point clouds suffering from occlusion/partial view and missing surfaces. The distributions are quite different and that of our outdoor sites is clearly vertically unbalanced. + +senting millions of points, + +(3) using (1) and (2) we construct a self-supervised point cloud completion scheme despite spatially unbalanced point distribution and lack of ground truth data (as in the case, at least, in outdoor archaeological settings), and +(4) we provide a large-scale 3D point-cloud dataset for improving future methods having to complete acquisitions with an extensive spatial range, millions of RGB points, and high-levels of incompleteness. + +# 2. Related Works + +# 2.1. Supervised/Unsupervised Completion + +Supervised point cloud completion generally requires complete and clean point clouds at training time. Further, it is assumed provided ground-truth data spans the distribution of target objects. Voxel-based methods [8, 40] benefit from 3D convolution but are challenged with processing speed and resolution. Point-based (e.g., [20, 27, 41, 43]) can adapt better but neither method works well when ground-truth data are not abundantly provided. Unsupervised point cloud + +completion does not rely on paired point clouds. Chen et al. [4] proposed an early unpaired point cloud completion method using auto-encoders. Other methods (e.g., [2, 22, 37, 44]) have made further improvements. However, this approach still relies on abundant complete point clouds during training and typically fails for uncommon, extensive and/or highly incomplete objects (e.g., [9, 19]). + +# 2.2. Self-supervised Completion + +For some real world scenarios, such as urban environments and archaeological sites, complete samples are costly or unavailable. Self-supervised methods are specifically designed to complete partial point clouds under such settings. Mittal et al. [24] proposed the first learning-based self-supervised method, PointPnCNet, which randomly masked out small regions during training in order to predict both the removed and the originally missing points. Hong et al. [16] developed a closed-loop point cloud completion system named ACL-SPC based on self consistency. Cui et al. [5] partitioned the partial point clouds into three disjoint groups, which served to improve performance. Wu et al. [38] introduced a Pattern Retrieval Network to further find similarities and subsequently densify the completion. + +These methods pave the way for point cloud completion without ground truth. Nevertheless, they still exhibit several limitations: 1) the size of the point cloud is limited (i.e., $10\mathrm{k}$ or less); 2) the methods generally assume missing data is caused by a limited field of view, which aggregates to a uniform distribution of all the available points. For object completion in outdoor and archaeological settings, the missing data is usually caused by weathering, erosion, physical stress, and activities. This exhibits a heavily unbalanced distribution, which misguides prior methods to overcompletion/undercompletion (see Section 4); 3) the methods usually lack support for color, which is highly important for downstream visualization and analysis applications. Our approach tackles all three of these prior limitations. + +# 2.3. Computational Archaeology + +Computing is essential to archaeology and to cultural heritage[33]. The Digital Michelangelo project [18] paved the way for large-scale archaeological object digitization. Later works addressed completion often by texture/image inpainting [30], finding and exploiting symmetries [19, 23], and/or a priori object databases [12, 14, 15, 34]. + +Despite the general success on reassembly of small scale archaeological objects, current methods heavily rely on prior knowledge like different types of symmetries and scanned complete shapes. This does not hold true for large-scale building structures, which also come with almost no available ground truth. However, this is the scenario our work addresses. + +# 3. Method + +For completing a target object, we first project its point cloud $P_{in}$ into 2D using our MCOP representation (Section 3.1). Then, completing the original point cloud is reduced to inpainting the MCOP image $I$ with an inpainter $C$ (Section 3.2), after which the completed shape $P_{out}$ is produced by reprojection $R$ of the inflated MCOP image $P_{out} = R(C(I))$ . Finally, Poisson reconstruction [17] is applied to form a water-tight mesh. An illustration of the whole pipeline is given in Fig. 3. + +# 3.1. Multiple Center of Projection Images + +To exploit convolutional deep networks ability to process millions of pixels, we construct a single MCOP image $I \in R^{W \times H \times 5}$ to store the raw point cloud data as seen from multiple viewpoints. The object can be represented as an MCOP image in different ways based on the trajectory and pose of the virtual cameras. For a faithful reconstruction of the original point cloud, we argue that by 1) moving the camera along a trajectory around/through the scene and 2) setting the camera to look perpendicular to the trajectory well preserves the scene's fine details. Hence in our work, a virtual camera flies around a target scene/object with its viewing orientation set perpendicular to the flying direction. The pixel-wide columns from each captured image (i.e., "slits") are concatenated to form the MCOP. Note that the MCOP projection is not necessarily conformal [7]. + +The MCOP image has 5-channels so that each column/slit samples the color and distance to both "side" and "top" of the target object. Unlike the original MCOP representation [28] which captures the side and top into different columns, we include a (viewing) rotation angle so that both parts are continuously placed into a single slit (Fig. 4d). + +The rotation channel also enables shape manipulation. As shown in supplemental, during generation the user can specify different rotations to essentially change the wall height or overall shape, enabling restoring and completing the archaeological structures based on optional simple user guidance. + +Formally, we represent the total process as + +$$ +I, M = M C O P \left(P _ {i n}, \mathbf {C}\right) \tag {1} +$$ + +where $\mathbf{C} = \{C_j|j\in [1,H]\}$ are the cameras used for each column of pixels, $I\in R^{W\times H\times 5}$ is the MCOP image, and $M\in R^{W\times H}$ is a binary mask defined by + +$$ +M _ {i j} = \mathbb {1} \left[ I _ {i j 4} < \infty \right] \tag {2} +$$ + +which is zero for pixels which do not intersect any point by camera $C_j$ at position $(i,j)$ . + +# 3.2. Self-supervised MCOP Completion + +Using MCOP image $I$ , the completion for an object fragment is achieved through infilling the missing pixels. We + +![](images/ee4b0b6697041c3a3c58bb976fba99880c809a967cd9182e66e20a1d6b3b25a4.jpg) +Figure 3. General pipeline of our MCOP based point cloud completion method. The input point cloud is first projected onto a 2D image with 5 channels (RGBD + rotation) using our MCOP representation. During training, a random sampling function $f_{\theta}$ is applied on the MCOP image to extract a group of local windows, which are trained against mostly complete local patches extracted from the dataset of all structures in an adversarial way. For inference, the MCOP image and an outline of the desired shape (e.g., wall height) is passed to our inpainting network adapted from [36]. The final completed point cloud is obtained by remapping the inpainted MCOP image back into 3D. + +tackle the scenario of high-levels of incompleteness (e.g., + +![](images/6650b0c0be5fa51e41aa812f790459e184a56e66735591b422716ebbb44477e5.jpg) + +![](images/5df52e25e2b157b20682c5692a507c72493eb389fa2f0fabeedf3418b1fef495.jpg) + +![](images/dc190762276b637300f6cce9c769e06d41bc5e0facb16fa73dd086fec12aa837.jpg) +(f) Ours: Rotation channel +Figure 4. MCOP [28] representation. For (a), we sample along a circular path while MCOP camera captures scan lines perpendicular to the path (b). Instead of the traditional MCOP image (c), we capture a 5 channel image (d, e, f). + +![](images/1ee85e7eec7361953cbe88da26058cd8305b642186dfeefd8c484f5a8173e634.jpg) +Figure 5. Histogram of random point cloud samples from the Mawchu Llacta dataset, with high completion ratios at the top to low-completion ratios at the bottom. + +over $90\%$ of the walls are missing over $90\%$ of pixels) and lack of ground truth. This results in the sampled point distribution exhibiting a non-uniform distribution (e.g., see Fig. 2b and Fig. 5 - most points are missing in the top). These problematic features still manifest in the MCOP images, for which we resort to a self-supervision mechanism which + +finds "signals" in subsets of the MCOP images to use in self-supervision during training. + +Image generation and completion in self- and unsupervised approaches have been addressed in the literature [1, 21, 25, 42]. In these approaches, given partial observation of MCOP images $\mathbf{D} = \{I_1, I_2, \dots, I_n\}$ and corresponding mask dataset $\mathbf{M} = \{M_1, M_2, \dots, M_n\}$ with $I_i, M_i = MCOP(P_i, \mathbf{C}), i = 1, \dots, n$ , an observation function $f_{\mathbf{M}}: R^{W \times H \times 5} \to R^{W \times H \times 5}$ + +$$ +f _ {\mathbf {M}} (x) = M _ {k} \otimes x \tag {3} +$$ + +with $M_{k}$ drawn at random from $M$ would be applied on the completer output $O$ to generate a masked sample of the completion with the masked pixels in $x$ filled with a given constant value (i.e., 0 in our case). In [25], the completer $C'$ is then trained adversarily to complete images from $\mathbf{D}$ to $\mathbf{D}' = \{I_1', I_2', \ldots, I_n'\}$ through + +$$ +I _ {i} ^ {\prime} = C \left(I _ {i}\right) \tag {4} +$$ + +by making the masked observation of $\mathbf{D}_{\mathbf{M}}^{\prime} = \{f_{\mathbf{M}}(I_1^{\prime}), f_{\mathbf{M}}(I_2^{\prime}), \ldots, f_{\mathbf{M}}(I_n^{\prime})\}$ to be indistinguishable from $\mathbf{D}$ . + +However, directly applying the above would fail under our targeted scenario where the distribution of available pixels in the image is heavily unbalanced (Fig. 2b). Because few fragments have the upper part of the wall, the completer $C$ will learn to escape from adversarial penalty by simply filling the upper part with a default value, as shown in Fig. 10b. It will not be heavily penalized under the masked observation of $D_M'$ . + +Instead, to provide effective supervision on both pixels from bottom and top, we apply the following patch-based observation function $f_{w}$ + +$$ +f _ {w} (I, x, y) = I [ x: x + w, y: y + w ] \tag {5} +$$ + +where $x$ and $y$ are the coordinates for the upper left corner of the window and $w$ is the fixed size of the patches. Despite the sparseness and unbalanced data in the MCOP images, there are smaller windowed samples of uniform completeness. By carefully selecting the observation window size $w$ , we extract a new observation dataset $D_w$ composed of all such patches under observation $f_w$ . One choice would be $f_{64}$ with windows of no less than $95\%$ completeness, whose average distribution will be (nearly) uniform, and enable the self-supervised strategy in [25, 42]. + +We empirically search for the optimal patch size $w$ and construct the patch ground truth dataset $\mathbf{D}_w$ based on the data distribution. For each MCOP image $I$ , we use a high resolution image inpainting network derived from LaMa [36] to obtain the output $O$ through $O = C(I)$ . To encourage the completion to be genuine, a random patch $o$ observed by + +$$ +o = f _ {w} \left(O, x _ {i}, y _ {i}\right) \tag {6} +$$ + +with $x_{i}$ and $y_{i}$ sampled from the image coordinates uniformly should be in the distribution modeled by $\mathbf{D}_w$ . The observation patches are fed into the discriminator $D$ adapted from [31] to train adversarially against random patches drawn from $D_w$ to minimize the adversarial loss + +$$ +L _ {a d v} = \mathbb {E} _ {I ^ {\prime} \sim D ^ {\prime}} [ (1 - \log (D (f _ {w} (C (I ^ {\prime}))) ] \tag {7} +$$ + +while discriminator $D$ is optimized by + +$$ +L _ {D} = \mathbb {E} _ {I \sim D _ {w}} [ D (I) ] + \mathbb {E} _ {I ^ {\prime} \sim D ^ {\prime}} [ (1 - \log (D (f _ {w} (C (I ^ {\prime}))) ] \tag {8} +$$ + +Regularity Terms. To further improve the quality of the generated MCOP images, additional loss terms are applied. To drive completion to reproduce the input MCOP image properly, consistency loss is applied on the known pixels as + +$$ +L _ {c o n s} = | | (O - I) \otimes M | | _ {2} = | | (C (I) - I) \otimes M | | _ {2} \tag {9} +$$ + +which encourages the completion to reproduce existing MCOP regions. To correctly reproject $O$ to a closed 3D object, each column of pixels in the rotation channel of $C$ should sum up to the angle camera rotated, giving + +$$ +L _ {r e g} = \left(\sum_ {j = 1} ^ {h} O _ {i j} - \pi\right) ^ {2} \tag {10} +$$ + +Appearance Consistency Terms. Due to the relative limited field of view of observation patches compared to the original image size, additional effort has to be taken to ensure the consistency of patterns and textures at a larger receptive field. Inspired by texture synthesis works [10], we use the pretrained VGG19 network to extract deep features on adjacent patch pairs randomly extracted from completion $O$ and minimize their correlations of features in order to yield a similar appearance. The term is defined as + +$$ +L _ {s i m} = \sum_ {i j} | | G \left(f _ {w} \left(O, x _ {1}, y _ {1}\right)\right) - G \left(f _ {w} \left(O, x _ {2}, y _ {2}\right)\right) | | _ {2} \tag {11} +$$ + +where $G$ computes the Gram matrix of deep features. The optimization goal is defined as + +$$ +L = L _ {a d v} + \lambda_ {\text {c o n s}} L _ {\text {c o n s}} + \lambda_ {\text {r e g}} L _ {\text {r e g}} + \lambda_ {\text {s i m}} L _ {\text {s i m}} \tag {12} +$$ + +where $\lambda_{cons},\lambda_{reg}$ and $\lambda_{sim}$ are weights controlling the importance of different terms. + +# 3.3. Implementation Details + +Based on the anticipated maximum height of structures, we generally divide the objects to complete into resolutions of $R^{256 \times W \times 5}$ and $R^{384 \times W \times 5}$ , which maximizes the pixel utility in the MCOP image for completion of 4 and 6 meter maximum height structures. Two different types of rotation + +channels are also preconstructed by making the rotation at the expected height and fed jointly with the original RGBD channels from the MCOP image. Noticing the influence of $f_{w}$ on different resolutions, we apply $\mathbf{D}_{64}$ and $\mathbf{D}_{96}$ separately on the above two distributions. To mitigate for efficient training with abundant self-supervision signals, we precomputed the necessary patch statistics and store the desired patches as an efficient way of accessing $\mathbf{D}_{w}$ . + +To train the high resolution image inpainter $C$ , we use the proposed self-supervised scheme on MCOP images with a batch size of 4 for 1 million iterations. We use the Adam optimizer with an initial learning rate of 0.0003 and decreasing by $2\%$ every 600 iterations. The model is trained on 4 NVIDIA V100 GPUs with NV-LINK for 3 days. + +With the model properly trained, we feed the MCOP images of the target objects into $C$ to get the completed MCOP representation $O$ , which is back projected into a 3D point cloud. To mitigate for minor issues and facilitate downstream processing/visualization, we apply Poisson reconstruction on the output to obtain a watertight mesh with colors as texture and normals computed per pixel using finite differences with the 4 neighboring pixels in MCOP. + +# 4. Experiments + +# 4.1. Dataset Construction + +We conduct qualitative and quantitative experiments on hundreds of point cloud datasets of buildings from three real-world archaeological sites, namely Mawchu Llacta, Huamanmarca, and Kuelap in Peru which suffer from heavy erosion and are missing a significant portion. Each site is reconstructed from multi-view high resolution UAV images as in [32], consisting of billions of points in total. Point clouds for individual walls/structures are extracted following coarse annotations from archaeologists in our team. To accommodate for efficient learning while not missing out important texture and geometric details, we sample the original point clouds with the proposed improved MCOP at a resolution of 2 centimeters, at which the tallest and biggest structure (as shown in Fig. 1) are represented within an MCOP image of $384 \times 9353$ . Since the original point clouds exhibit visible shadows due to lighting or self-occlusion in the input stage, we manually annotate out the shadowed regions in the MCOP images and "deshadow" them by either histogram matching to unshadowed regions or masking out the corresponding regions. Some important statistics for the datasets are listed in supplemental. We name the datasets Mawchu Raw, Huaman Raw and Kuelap Raw. + +# 4.2. Quantitative Comparisons + +Due to lack of ground truth, we randomly remove sets of points from Mawchu Raw to synthesize a more incomplete dataset Mawchu Incomplete. The removal is implemented + +by applying random masks from $D_{128}$ with completeness ratio in between $20\%$ to $80\%$ at random positions in the original MCOP images. As shown in Fig. 7, the masked out regions could serve as the ground truth for quantitative study when we train on Mawchu Incomplete only and evaluate against the masked out regions in Mawchu Raw. To fully evaluate the reconstruction quality both in inpainting and point cloud form, we measure the overall RGBD differences of the masked regions, patch/distribution similarity, and Chamfer distance of reconstructed point cloud. + +We compare our method on the aforementioned dataset against several state-of-the-art methods. These methods include self-supervised point-cloud approaches (ACL-SPC [16] and P2C [5]), self-supervised image inpainting methods (UAIR [25] and SSII [42]) on the MCOP images (i.e., MCOP+UAIR and MCOP+SSII). We also compare against MCOP+SSII(Patch) which uses image-patch self-supervision. ACL-SPC and P2C do not produce colors so we omit those metrics for them. + +
Feature TypeTextureGeometry
MAE↓ACL-SPCN/A0.172
P2CN/A0.195
MCOP+SSII0.2060.143
MCOP+UAIR0.2590.149
MCOP+SSII(Patch)0.2090.145
Ours0.1890.136
SSIM↑ACL-SPCN/A0.085
P2CN/A0.048
MCOP+SSII0.2360.291
MCOP+UAIR0.1730.261
MCOP+SSII(Patch)0.2130.278
Ours0.2460.353
PSNR↑ACL-SPCN/A15.06
P2CN/A13.52
MCOP+SSII11.7917.18
MCOP+UAIR12.3716.54
MCOP+SSII(Patch)11.6517.10
Ours14.2817.87
FID↓ACL-SPCN/A211.69
P2CN/A331.10
MCOP+SSII203.85184.84
MCOP+UAIR192.45158.79
MCOP+SSII(Patch)179.05169.06
Ours90.81128.41
+ +Table 1. Quantitative comparison of completion quality using image-level metrics. Our approach has the best performance (shown in bold). + +We first report metrics in image inpainting as shown in Table 1. For each completed point cloud, we project it into 2D using our proposed MCOP representation and perform image patch evaluation with the commonly used metrics of + +![](images/8ff3295a38a570ca96100d50851e04ff3555f38152d247010b6e385a746a60e9.jpg) + +![](images/71d185d83394d61fdffc907cee135af30c5d70c4155ce48908968996cb1280a6.jpg) + +![](images/72205fb1405e25b4cb6d363f22af10de9fa259eb1baf8755350d71c26854be7c.jpg) + +![](images/70765084f6149697363c5fd1372ecbb01b750d6b8520709870fce3c3a1ad0cb8.jpg) + +![](images/f30e426e7146d155eb4de293e85473f51df0099a874ac976aaa5a6b30f614863.jpg) + +![](images/d70f30a4d867b2b50dd792e825849ce129fe9a666f6ca85a4e790e494d12cf56.jpg) + +![](images/04d25eb23abdf7f929d1459b3df5e8aa2f06deea215dbb48ea36a82df815a0af.jpg) + +![](images/66c261da6c57a8dd0eacd7548eb89211cea0d2cb76a8b784858f025bb3e48659.jpg) + +![](images/a71b6b37c7a8610424cad6561e397c19725f3d7edb1c0750397751709558e6e5.jpg) + +![](images/784dcac526dcb940cca9ed4c94b4fa7f29018deae0c6e876cc7a4b0216c73e59.jpg) + +![](images/11a7c0fb250fd8a1294b9825a239e1647582cf65759a10fbbf5e900a801b0b78.jpg) + +![](images/7926fcba8692bb789fc83e2160bbb5087b2ca3fc5d93d60f753a3d4bb990076b.jpg) + +![](images/29254c0e99a3ccafca876771df1b83af191fd6d85e274bf150a4966da8320198.jpg) + +![](images/c6502b14f366634012c4e1ca0d7da8f15fee7d6608fd7d8a881b81d600eee081.jpg) + +![](images/c6de5de1e19d0d595329b193a77c06360e9a751eafb744f8059b7cb9a20704a4.jpg) + +![](images/91c020d5cb18c818f9b061151b7e6e0a5c442023eaa094c45e00970ea30ba079.jpg) +(a) Inputs + +![](images/570b707ca672e4342b4660785ffeaa569e538297468bdfb594004edaa973e22e.jpg) + +![](images/35dce98091de2e6cecb34bd2e845c0f3e74729e0eba21def8487b72f08f30563.jpg) +(b) P2C + +![](images/1dbae6d9842d2735608fcb9355363e43ebacb49a8675964da6bf44d86a623e86.jpg) + +![](images/50184f019bac630db78bd95ee2929cc5d18cb0175b33d2a90961e7f94c82474f.jpg) +(c) ACL-SPC + +![](images/5e1f7d6e4285c11a67c1d86a3930c76bb4515c354a56cdc6a2d8c7f4e4632fa9.jpg) + +![](images/78591d6deb5b7ae9fa9e37bf3225456f2050f5f0799db73a7ead95da090c71b0.jpg) +(d)UAIR +Figure 6. Point cloud completion results of different methods on Mawchu Raw. We overlay completion onto input. More visualization results can be found in the supplemental. + +![](images/1b93f2c992c7c35620555b8bc01318a70ecd60e087e1b2e7a817672d31ba8efb.jpg) + +![](images/a23017c70f01ceff3e0bf925035f129e557d3befe1078e2277a65713b6728103.jpg) +(e) SSII + +![](images/d41ba1f0ba5ee485e0c70e9637d07b9589fcb861bfc5caef6951cefeb16ef79f.jpg) + +![](images/f97948603be5f0db016a408cb786ad30559bed44d51d6392035e665121307049.jpg) +(f) SSII(Patch) + +![](images/92943d79a6425621305cd446f90aae69e82bfdb2465f6342d192b16b5e007039.jpg) + +![](images/e7db1b12b2c29016db8ad6832128d74700143f09f53705693cccbe5987e02805.jpg) +(g) Ours + +![](images/cf54ef8f51a3c575475e6d7f718ef9a1c86f95ac468cb43bd5f13422941b5fff.jpg) +(1) Original MCOP Image + +![](images/6c412dcb5b8fa518f919d6eacdaa819364450a7e568a59a5de84076d0e17fbda.jpg) +(2) MCOP Image Masked (Gray pixels serving as groundtruth) +Figure 7. Demonstration of the construction of Mawchu Incomplete used in evaluation. Using the original MCOP image, we randomly apply small masks from $D_{128}$ onto random locations producing an incomplete MCOP image. + +MAE, SSIM, PSNR and FID based on 500 pairs of randomly chosen windows covering the masked out regions. Our method achieves the best performance in both geometry and texture of predicted point cloud in all metrics. + +We further report the completion quality using the same point cloud metrics as in [16], where we compute the Chamfer distance of prediction $O$ and ground-truth part $G$ in the masked regions of Mawchu Incomplete as follows: + +$$ +C D (O, G) = \frac {1}{N _ {o}} \sum_ {p \in O} \min _ {q \in G} \| p - q \| _ {2} + \frac {1}{N _ {g}} \sum_ {q \in G} \min _ {p \in O} \| q - p \| _ {2} \tag {13} +$$ + +
MethodP↓C↓CD↓
ACL-SPC91.0735.00126.07
P2C11.3312.2423.57
MCOP+SSII5.204.589.79
MCOP+UAIR2.906.299.20
MCOP+SSII(Patch)4.274.919.18
Ours3.804.428.22
+ +Table 2. Quantitative comparison of completion quality using point cloud metrics: Chamfer distance (CD), precision (P), and coverage (C) as used in [16] and in centimeters. Our method is best or second best in all metrics. + +where $N_{o}$ and $N_{g}$ are the number of points in $O$ and $G$ . The first and second term are denoted separately as Precision $P$ and $C$ following [16]. We achieve best or second best in all metrics. + +# 4.3. Ablation Study + +MCOP representation. From Table. 1 and 2, methods using MCOP as their base representation generally achieve better performance by a large margin as compared to fully point clouds based ones. The behavior is also interpretable in Fig. 6, where P2C and ACL-SPC generate noisy and nonuniform points around the input, whereas MCOP efficiently regularizes the output point cloud to distribute along given boundary. + +Patch-level self supervision. We demonstrate using patch- + +![](images/4c91eaed72ec91849468e6d8f233f34a06a77baae2628e3d26f9af7aad49a26d.jpg) + +![](images/d51c39fe4abb4750d9041f25a2848cae0d30b5b7b406929156a51870d3de37a3.jpg) + +![](images/c9664776739a576d2d9e5256cfeb5020cc4d0f93116509e11ded1b45d6ada050.jpg) + +![](images/949c62a56ec2ca05a12955d6734f19971ddc0a50f3c48f4ae8af07ac591758e1.jpg) + +![](images/079708c925c4ceaa16f87355280fc5bd3f0ef80e98a79b35aa08c83893299d58.jpg) + +![](images/72ac3264be6fb15f4b16d2563b04c2492ba5ba518f14760fedabf6e812c5ff46.jpg) +Figure 10. Demonstration of the necessity of applying patch based self supervision for inpainting our MCOP images. + +![](images/26a7e47fc4d28212a23b4a8293d9dbd9c6d62db1124866eec1f485b693472724.jpg) +Figure 8. Example completion results from Mawchu Llacta and Huamanmarca. The roof in the fourth row is generated with additional synthesis for visualization purposes only. Completion results on Kuelap are included in the supplemental. + +![](images/38d063113a466950e55883979a6bf6fc27323fd8bf0f8c784120fbae6678033f.jpg) + +![](images/c3157bb366b442c8e270e2011a3203ae00d492658aa264b9d5ecc760525a2fbe.jpg) +(a) Original input + +![](images/d0b96e08bbdcb580e62f4521fd370252d4efad3a40355859d58564ed1b456d1d.jpg) +(b) Completion without similarity term + +![](images/4443edfdac8c36160b5719ba902a3caacad7c3adf82ba1ac307a1bc4d85d21ce.jpg) +(c) Completion with similarity term +Figure 9. Demonstration of need for texture similarity term. + +level self-supervision in Fig. 10. Given (a), directly applying UAIR on the completed patches would yield significantly smaller supervision at the higher pixels compared to the rest because of the unbalanced distribution of available pixels (b), for which the completer simply learns to escape adversarial penalty by infilling default pixel values at those positions. SSII also suffers from the unbalanced distribution and tends to produce blurry completions at pixels far from known regions (c). Our patch-based scheme successfully + +![](images/173f372df5978a8276b5c0ac4c9947fd6b869d5cf427ea48ee52c96e43fb800b.jpg) +(a) MCOP input + +![](images/f147bbc4d833942bb4f2b2e4ec97d7f0377e6ab5e0236fc5b8d9ea96f5dd1a4b.jpg) +(b) MCOP+UAIR output + +![](images/d0dd4cee01cef4246069b0ff02f77b125e9c5ca12425c215dfa19094640f97cc.jpg) +(c) MCOP+SSII output + +![](images/d1ffa27cbc3886bdd852810996452b3dea3383544787879c3a06a31fd4b3371c.jpg) +(d) MCOP+SSII(Patch) output + +overcomes the unbalanced distribution and helps produce a sharp completion everywhere (d). + +Texture similarity. As shown in Fig. 9, completion without texture similarity constraints produces inconsistent horizontal and vertical patterns at a receptive field larger than the observation window, leading to visual inconsistencies as highlighted in red circles. By incorporating the texture similarity term, the completer is able to maintain consistent completion in both directions at the whole image level. + +# 5. Conclusions and Future Work + +We propose the first self-supervised point cloud completion pipeline that handles large scale partial point cloud points suffering from missing surfaces, which is common for fragments in archaeological sites. By embedding the original structure in 2D using multiple-center-of-projection (MCOP) images, the approach can handle point clouds of up to millions of points. Patch-based self-supervision with texture similarity terms helps overcome the unbalanced distribution of available pixels resulting from missing surfaces. Results demonstrate our method performs best quantitatively and qualitatively. Nevertheless, our work relies on prior annotation for objects to collect high fidelity MCOP images, which could influence the performance and quality of point cloud completion. As future work, we would like to extend our method to 1) be fully automated to complete input fragments; 2) complete point clouds of general objects at high resolution, and 3) improve shadow elimination. + +Acknowledgement This research is partially funded by NSF Grant 2107096, 1835739 and 2411273. + +# References + +[1] Ashish Bora, Eric Price, and Alexandros G. Dimakis. AmbientGAN: Generative models from lossy measurements. In International Conference on Learning Representations, 2018. 5 +[2] Yingjie Cai, Kwan-Yee Lin, Chao Zhang, Qiang Wang, Xiaogang Wang, and Hongsheng Li. Learning a structured latent space for unsupervised point cloud completion, 2022. 3 +[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 2 +[4] Xuelin Chen, Baoquan Chen, and Niloy J Mitra. Unpaired point cloud completion on real scans using adversarial training. In Proceedings of the International Conference on Learning Representations (ICLR), 2020. 2, 3 +[5] Ruikai Cui, Shi Qiu, Saeed Anwar, Jiawei Liu, Chaoyue Xing, Jing Zhang, and Nick Barnes. P2c: Self-supervised point cloud completion from single partial clouds, 2023. 2, 3, 6 +[6] Yaodong Cui, Ren Chen, Wenbo Chu, Long Chen, Daxin Tian, Ying Li, and Dongpu Cao. Deep learning for image and point cloud fusion in autonomous driving: A review. IEEE Transactions on Intelligent Transportation Systems, 23(2): 722-739, 2022. 1 +[7] Sean Curry and A. Rod Gover. An introduction to conformal geometry and tractor calculus, with a view to applications in general relativity, 2015. 3 +[8] Angela Dai, Charles Ruizhongtai Qi, and Matthias Nießner. Shape completion using 3d-encoder-predictor cnns and shape synthesis, 2017. 2 +[9] Aref Enayati, Luca Palmieri, Sebastiano Vascon, Marcello Pelillo, and Sinem Aslan. Semantic motif segmentation of archaeological fresco fragments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1717-1725, 2023. 3 +[10] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks, 2015. 5 +[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 2 +[12] Robert Gregor, Ivan Sipiran, Georgios Papaioannou, Tobias Schreck, Anthousis Andreadis, and Pavlos Mavridis. Towards automated 3d reconstruction of defective cultural heritage objects. In Proceedings of the Eurographics Workshop on Graphics and Cultural Heritage, page 135-144, Goslar, DEU, 2014. Eurographics Association. 3 +[13] Lei Han, Tian Zheng, Yinheng Zhu, Lan Xu, and Lu Fang. Live semantic 3d perception for immersive augmented reality. IEEE Transactions on Visualization and Computer Graphics, 26(5):2012-2022, 2020. 1 + +[14] Gur Harary, Ayellet Tal, and Eitan Grinspun. Feature-preserving surface completion using four points. Computer Graphics Forum, 33(5):45-54, 2014. 3 +[15] Renato Hermoza and Ivan Sipiran. 3d reconstruction of incomplete archaeological objects using a generative adversarial network, 2018. 3 +[16] Sangmin Hong, Mohsen Yavartanoo, Reyhaneh Neshatavar, and Kyoung Mu Lee. Acl-spc: Adaptive closed-loop system for self-supervised point cloud completion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 6, 7 +[17] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2, 3 +[18] Marc Levoy, Kari Pulli, Brian Curless, Szymon Rusinkiewicz, David Koller, Lucas Pereira, Matt Ginzton, Sean Anderson, James Davis, Jeremy Ginsberg, Jonathan Shade, and Duane Fulk. The Digital Michelangelo Project: 3D Scanning of Large Statues. Association for Computing Machinery, New York, NY, USA, 1 edition, 2023. 3 +[19] Er Li, Xiaopeng Zhang, and Yanyun Chen. Symmetry based chinese ancient architecture reconstruction from incomplete point cloud. In 2014 5th International Conference on Digital Home, pages 157-161, 2014. 3 +[20] Shanshan Li, Pan Gao, Xiaoyang Tan, and Mingqiang Wei. Proxyformer: Proxy alignment assisted point cloud completion with missing part sensitive transformer, 2023. 2 +[21] Steven Cheng-Xian Li, Bo Jiang, and Benjamin Marlin. Learning from incomplete data with generative adversarial networks. In International Conference on Learning Representations, 2019. 5 +[22] Changfeng Ma, Yinuo Chen, Pengxiao Guo, Jie Guo, Chongjun Wang, and Yanwen Guo. Symmetric shape-preserving autoencoder for unsupervised real scene point cloud completion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13560-13569, 2023. 3 +[23] P. Mavridis, I. Sipiran, A. Andreadis, and G. Papaioannou. Object completion using k-sparse optimization. Computer Graphics Forum, 34(7):13-21, 2015. 3 +[24] Himangi Mittal, Brian Okorn, Arpit Jangid, and David Held. Self-supervised point cloud completion via inpainting. arXiv preprint arXiv:2111.10701, 2021. 2, 3 +[25] Arthur Pajot, Emmanuel de Bezenac, and Patrick Gallinari. Unsupervised adversarial image reconstruction. In International Conference on Learning Representations, 2019. 5, 6 +[26] Gabyong Park, Antonis Argyros, Juyoung Lee, and Woontack Woo. 3d hand tracking in the presence of excessive motion blur. IEEE Transactions on Visualization and Computer Graphics, 26(5):1891-1901, 2020. 1 +[27] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space, 2017. 1, 2 +[28] Paul Rademacher and Gary Bishop. Multiple-center-of-projection images. In Proceedings of the 25th Annual Conference on Computer Graphics and Interactive Techniques, + +page 199-206, New York, NY, USA, 1998. Association for Computing Machinery. 3, 4 +[29] Avi Resler, Reuven Yeshurun, Filipe Natalio, and Raja Giryes. A deep-learning model for predictive archaeology and archaeological community detection. *Humanities and Social Sciences Communications*, 8, 2021. 2 +[30] M. S. Sagiroglu and A. Erçil. A Texture Based Approach to Reconstruction of Archaeological Finds. In The 6th International Symposium on Virtual Reality, Archaeology and Cultural Heritage VAST. The Eurographics Association, 2005. 3 +[31] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5 +[32] Young-Ha Shin, Sang-Yeop Shin, Heidar Rastiveis, Yi-Ting Cheng, Tian Zhou, Jidong Liu, Chunxi Zhao, Günder Varinlioglu, Nicholas K. Rauh, Sorin Adam Matei, and Ayman Habib. Uav-based remote sensing for detection and visualization of partially-exposed underground structures in complex archaeological sites. Remote Sensing, 15(7), 2023. 6 +[33] Ivan Sipiran. The role of computing in the study of latin american cultural heritage. Commun. ACM, 67(8):58-62, 2024. 3 +[34] Ivan Sipiran, Alexis Mendoza, Alexander Apaza, and Cristian Lopez. Data-driven restoration of digital archaeological pottery with point cloud analysis. International Journal of Computer Vision, 130(9):2149-2165, 2022. 3 +[35] Andreas Stasinakis, Elisavet Chatzilari, Spiros Nikolopoulos, Ioannis Kompatsiaris, Dimitrios Karolidis, Annareta Touloumtzidou, Nikolaos Papas, and Dimitrios Tzetzis. A hybrid 3d object auto-completion approach with self-supervised data augmentation for fragments of archaeological objects. Journal of Cultural Heritage, 56:138-148, 2022. 1, 2 +[36] 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, 2021.4.5 +[37] Xin Wen, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Cycle4completion: Unpaired point cloud completion using cycle transformation with missing region coding. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13075-13084, 2021. 2, 3 +[38] Lintai Wu, Xianjing Cheng, Junhui Hou, Yong Xu, and Huanqiang Zeng. Self-supervised 3d point cloud completion via multi-view adversarial learning, 2024. 2, 3 +[39] 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, 2024. 1 +[40] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Jiageng Mao, Shengping Zhang, and Wenxiu Sun. Grnet: Gridding residual network for dense point cloud completion, 2020. 2 + +[41] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation, 2018. 2 +[42] Sriram Yenamandra, Ansh Khurana, Rohit Jena, and Suyash P. Awate. Learning image inpainting from incomplete images using self-supervision. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 10390-10397, 2021. 5, 6 +[43] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network, 2019. 2 +[44] Junzhe Zhang, Xinyi Chen, Zhongang Cai, Liang Pan, Haiyu Zhao, Shuai Yi, Chai Kiat Yeo, Bo Dai, and Chen Change Loy. Unsupervised 3d shape completion through gan inversion, 2021. 3 +[45] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer, 2021. 1 \ No newline at end of file diff --git a/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/images.zip b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..84cd860a8a9d4edc781a2a7282fb95cc40027b9f --- /dev/null +++ b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e81ba22cc536f23273c58aa9a1ed7db6fabb5d77434395026818dc7f8fd035b2 +size 894095 diff --git a/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/layout.json b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cb97027e32bbff543eae50e48a2ed26bc2804bb9 --- /dev/null +++ b/CVPR/2025/Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a33c08726b1c5ba9472f7edc8da2127359ea2e0a3af5bb42f1868a485f81373 +size 447235 diff --git a/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/48099590-89bf-4222-a359-563bb17451be_content_list.json b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/48099590-89bf-4222-a359-563bb17451be_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cb1303b8a8aef9877a60d92ece37c81fef3d89e8 --- /dev/null +++ b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/48099590-89bf-4222-a359-563bb17451be_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c4941ab987d5fd3e1f0b0b25a92357a191415cd6185f7c490caf20a99c78065 +size 79196 diff --git a/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/48099590-89bf-4222-a359-563bb17451be_model.json b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/48099590-89bf-4222-a359-563bb17451be_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ac9625934f6551d53579cc02f2406b65b528e7ec --- /dev/null +++ b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/48099590-89bf-4222-a359-563bb17451be_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9e3999442bf168acd8283a951cb6079072e062b4bdf0a4689792f6a8a6afe7a +size 96549 diff --git a/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/48099590-89bf-4222-a359-563bb17451be_origin.pdf b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/48099590-89bf-4222-a359-563bb17451be_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..792e555e1d5668c9f5c119ecb6949b6c86d6e90d --- /dev/null +++ b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/48099590-89bf-4222-a359-563bb17451be_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ac5291a049dc41b8143e95eca268adf7fe50a2bb62eea4427fa044047f7f7ab +size 6440818 diff --git a/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/full.md b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..963bc85ced507b89c5cb0260d8df125df355a8f3 --- /dev/null +++ b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/full.md @@ -0,0 +1,297 @@ +# Self-Supervised Learning for Color Spike Camera Reconstruction + +Yanchen Dong $^{1}$ , Ruiqin Xiong $^{1*}$ , Xiaopeng Fan $^{2}$ , Zhaofei Yu $^{3}$ , Yonghong Tian $^{1,4,5}$ , Tiejun Huang $^{1,3}$ $^{1}$ School of Computer Science, Peking University + $^{2}$ School of Computer Science and Technology, Harbin Institute of Technology + $^{3}$ Institute for Artificial Intelligence, Peking University + $^{4}$ School of Electronic and Computer Engineering, Peking University + $^{5}$ Peng Cheng Laboratory + +yanchendong@stu.pku.edu.cn, {rqxiong, yuzf12, yhtian, tjhuang}@pku.edu.cn, fxp@hit.edu.cn + +# Abstract + +Spike camera is a kind of neuromorphic camera with ultra-high temporal resolution, which can capture dynamic scenes by continuously firing spike signals. To capture color information, a color filter array (CFA) is employed on the sensor of the spike camera, resulting in Bayer-pattern spike streams. How to restore high-quality color images from the binary spike signals remains challenging. In this paper, we propose a motion-guided reconstruction method for spike cameras with CFA, utilizing color layout and estimated motion information. Specifically, we develop a joint motion estimation pipeline for the Bayer-pattern spike stream, exploiting the motion consistency of channels. We propose to estimate the missing pixels of each color channel according to temporally neighboring pixels of the corresponding color along the motion trajectory. As the spike signals are read out at discrete time points, there is quantization noise that impacts the image quality. Thus, we analyze the correlation of the noise in spatial and temporal domains and propose a self-supervised network utilizing a masked spike encoder to handle the noise. Experiments on real-world captured Bayer-pattern spike streams show that our method can restore color images with better visual quality, compared with state-of-the-art methods. The source codes are available at https://github.com/csycdong/SSL-CSC. + +# 1. Introduction + +With the development of machine vision, there has been an increasing demand for high-speed imaging, especially in some emerging applications, such as autonomous driving and unmanned aerial vehicles. How to capture dynamic scenes with high-speed motion and respond quickly becomes a key challenge. Conventional digital cameras usu + +ally require the scene to be still during the exposure time for photoelectric signal accumulation. The motion of captured objects will bring undesirable blur to snapshot images of the cameras, limiting their applications in high-speed imaging. + +Spike camera [7, 8, 10, 22, 35] is a type of neuromorphic camera mimicking the structure of the human visual system. Benefiting from its ultra-high temporal resolution (e.g., $40,000\mathrm{Hz}$ ), the spike camera shows great potential in recording high-speed scenes. Different from another kind of neuromorphic camera called event camera [25, 26], the spike camera captures absolute light intensity instead of light intensity changes. The first-generation spike camera is the gray-scale spike camera (GSC). To capture color information of dynamic scenes, the Bayer-pattern color filter array (CFA) is employed on the sensor, resulting in the color spike camera (CSC) [9, 11-14]. How to restore high-quality color images from the Bayer-pattern spike stream is still an open task, with the following challenges: + +- Missing pixels of each color channel. According to the Bayer-pattern color layout, $75\%$ (red and blue) or $50\%$ (green) pixels are missing in each color channel. +- High-speed motion. Spikes fired by a certain pixel do not always describe the same point of the object due to high-speed motion. Thus, it is likely to introduce blur. +- Quantization noise in spike signals. The spike signals are read out at discrete time points, bringing fluctuation and randomness to continuous photoelectric accumulation and resulting in quantization noise. + +Some spike camera reconstruction methods [12, 36, 37, 40, 42] tend to employ end-to-end convolutional neural networks (CNNs). As it is hard to collect spike streams with corresponding high-quality ground truth, they tried to synthesize massive training pairs from videos. However, there are gaps between synthetic data and real-world captured data, reducing the practical performance. + +In this paper, we develop a motion-guided reconstruc + +tion method for CSCs. To be specific, we propose to fill the missing pixel positions based on temporally neighboring pixels of the corresponding color, utilizing motion trajectory jointly estimated from color channels. We also fill the positions without available temporal pixels. As the discrete read-out of spike signals, there is quantization noise that impacts the image quality. Taking the preliminary reconstruction results as pseudo-labels, we propose a self-supervised network for CSC reconstruction. According to the analysis of quantization noise in temporal and spatial domains, we propose an encoder to extract features of each color channel and handle noise by our developed masked 3D partial convolution (MPConv) and temporal masked convolution (TMConv). In experiments, our self-supervised method achieved the best performance among methods without the supervision of ground truth. Although our method does not exceed supervised methods on synthetic data, it achieves better practical performance on real-world captured samples. The main contributions can be summarized as + +- Considering the color layout and motion, we propose a motion-guided scheme to search temporally neighboring pixels to estimate the missing pixels of each color channel. We also develop a self-supervised reconstruction network with the preliminary results as pseudo-labels. +- We analyze the correlation of quantization noise caused by discrete signal read-out in temporal and spatial domains and propose a masked spike encoder to extract features and suppress the noise. +- Experimental results on real-world captured Bayer-pattern spike streams show that our self-supervised reconstruction method achieves better visual quality than the state-of-the-art supervised methods. + +# 2. Related Work + +# 2.1. Reconstruction for Spike Cameras + +As the first attempt at spike camera reconstruction, Zhu et al. [45] utilized the number of spike signals within a temporal window (TFP) or temporally neighboring spike intervals (TFI) to infer the light intensity of dynamic scenes. However, the former strategy is likely to bring motion blur, and the latter often brings undesired noise. To overcome the issues, a Motion-Aligned Filter (MAF) [38, 41] was proposed to suppress both motion blur and noise. Differently, Zhu et al. [46, 47] employed a three-layer spiking neural model mimicking human vision. Zheng et al. [44] tried to handle the task with a Short-Term Plasticity (STP) mechanism. Inspired by deep learning, Zhao et al. [40] developed the first spike camera reconstruction network, Spk2ImgNet, bringing significant performance gains. Zhang et al. [37] utilized temporal-robust features of spike signals in time-frequency space with wavelet transforms, resulting in another gray-scale reconstruction network, WGSE. Consider + +ing the difficulty of collecting paired training data, Chen et al. [4] proposed a self-supervised joint learning framework for motion estimation and reconstruction. + +All the above methods are designed for the first-generation spike camera that can only generate gray-scale signals. Focusing on spike cameras with CFA, Dong et al. [9] presented 3D residual interpolation (3DRI) to first restore color images from the Bayer-pattern spike stream. Based on supervised learning, Dong et al. [13] developed a demosaicing network called CSpkNet to restore color visual signals from the Bayer-pattern spike stream. Starting from the observation model of spike cameras with CFA, an optimization-inspired network SJDD-Net [12] was also proposed for color image reconstruction of spike cameras. + +# 2.2. Reconstruction for Other Emerging Sensors + +The well-known event camera [1, 25, 26, 31] is another type of neuromorphic camera, designed to capture light intensity changes. As the first reconstruction attempt, Kim et al. [23] employed an Extended Kalman Filter to restore gradient images. Based on deep learning, E2VID [32] was proposed to reconstruct images in an end-to-end way, with great visual performance. To achieve lower computational complexity, Scheerlinck et al. [34] explored FireNet as a lightweight network. For color imaging, Scheerlinck et al. [33] proposed the first color event camera dataset and methods to restore color images from events. + +Spike cameras and event cameras are designed for high-speed applications. Differently, quanta image sensor (QIS) [2, 15-17, 19] is a sensor developed for applications in low-light imaging, which can detect individual photons through spatial and temporal oversampling. Choi et al. [6] presented the first deep neural network for QIS reconstruction. To achieve dynamic and low-light imaging with the QIS, Chi et al. [5] proposed a training strategy that transfers knowledge from motion and denoising teachers to a student network. Besides, Gnanasambandam et al. [18] designed a demosaicing method for color imaging of QIS. + +# 3. Preliminary + +# 3.1. Color Spike Camera + +Spike camera is a neuromorphic camera that accumulates photoelectric signals according to absolute light intensity and continuously outputs binary streams at ultra-high temporal resolution. The gray-scale spike camera (GSC) can only produce gray-scale visual signals. In contrast, the color spike camera (CSC) with a CFA on the sensor can capture color information of dynamic scenes. As shown in Fig. 1, the structure of CSCs mimics the human visual system. For each pixel, the photon receptor captures photons and transforms them into accumulations in the integrator, continuously monitored by the comparator. Whenever the accumu + +![](images/9d77914a9831a6dfff3d2f9f2f7ded341dfb2a97813a019324293b01b2cd4c98.jpg) +Figure 1. Working mechanism of the color spike camera. The symbol $\theta$ denotes the firing threshold. + +lated photoelectric signals reach a predetermined threshold $\theta$ , the pixel triggers a flag for spike firing. Then the pixel will be reset and continue to accumulate signals, restarting an accumulate-and-fire cycle [41]. For a certain pixel $(x,y)$ , the accumulated electric signal $\mathbf{E}_c$ of the integrator at arbitrary time point $t$ can be written as + +$$ +\mathbf {E} (t, x, y) = \int_ {0} ^ {t} \eta \cdot \mathcal {P} (\mathbf {I} (\tau , x, y)) d \tau \mod \theta , \tag {1} +$$ + +where $\eta$ denotes the photoelectric conversion rate, $\mathcal{P}(\cdot)$ denotes Poisson process of photon arrival, and $\mathbf{I}(\tau ,x,y)$ denotes the instantaneous intensity at time $\tau$ . According to the working mechanism, we develop a CSC simulator, with details in the Supplementary Material. + +# 3.2. Quantization Noise + +Ideally, the pixel is reset immediately after firing a spike signal. For a certain pixel $(x,y)$ , the number of spike signals fired before arbitrary time point $t$ can be written as + +$$ +\mathcal {N} (t, x, y) = \left\lfloor \frac {1}{\theta} \int_ {0} ^ {t} \mathcal {P} (\mathbf {I} (\tau , x, y)) d \tau \right\rfloor . \tag {2} +$$ + +In real hardware implementation, the check-and-reset process is controlled by a clock. Thus, the signals are read out at discrete time points. For the $n$ -th $(n \in \mathbb{N}^{+})$ time point, the read-out spike signal $\mathbf{S}_n(x,y)$ can be formulated by spike count difference of neighboring time points as + +$$ +\mathbf {S} _ {n} (x, y) = \left\{ \begin{array}{l l} 1, & \mathcal {N} (t _ {n}, x, y) - \mathcal {N} (t _ {n - 1}, x, y) > 0 \\ 0, & \mathcal {N} (t _ {n}, x, y) - \mathcal {N} (t _ {n - 1}, x, y) = 0 \end{array} , \right. \tag {3} +$$ + +where $t_n$ denotes the time of the $n$ -th discrete time point. Perfectly, we have $t_n - t_{n-1} \approx 0$ , while $t_n - t_{n-1} = 50\mu s$ according to the sensor frequency of CSCs. As a result, quantization noise is introduced into the output binary spike stream. For example, the difference between $\mathcal{N}(t_n,x,y)$ and $\mathcal{N}(t_{n-1},x,y)$ can be larger than 1, while the signal is read out like only one spike signal fired. + +# 3.3. Problem Statement + +As a neuromorphic camera, the CSC has an array of $H \times W$ pixels working independently on its sensor. After $N$ times of the check-and-reset process, it outputs a Bayer-pattern spike stream $\{\mathbf{S}_i\}_{i=1}^N$ . In this paper, we aim to restore a color image $\mathbf{T} \in [0,1]^{3 \times H \times W}$ with high visual quality from the $N \times H \times W$ binary data. To achieve it, we need to infer the light intensity of the missing pixels of each color channel and handle the quantization noise in spike signals. Denoting the Bayer-pattern color mask, which records the color layout of the three color channels, as $\mathbf{M} \in \{0,1\}^{3 \times H \times W}$ (0 means a missing pixel), the target of our method $\mathcal{F}$ can be formulated as $\mathcal{F}(\{\mathbf{S}_i\}_{i=1}^N, \mathbf{M}) \to \mathbf{T}$ . + +# 4. Motion-Guided Preliminary Reconstruction + +# 4.1. Overall Pipeline + +To restore a color image from the input Bayer-pattern spike stream $\{\mathbf{S}_i\}_{i = 1}^N$ , we develop a motion-guided reconstruction method as shown in Fig. 2. Considering the missing spatial pixels (75% or 50%) of each color channel, we propose to estimate the values of missing pixels by available temporally neighboring pixels. To be specific, we first perform pixel-wise multiplication between each spike frame and the color masks to split out the spike signals of each color channel. Then, we infer the coarse light intensity of color channels by the pixel-independent method TFI [45]. To fill the missing pixel positions, we try to estimate the values by pixels of other time points. However, due to the high-speed motion within the spike stream, it is likely to introduce blur when using temporal pixels. Thus, we propose a joint motion estimation method to obtain optical flows from the spike signals of the color channels. Along the motion trajectories, we collect the available temporal pixels to estimate the missing pixels, resulting in the color channels without missing pixels. Finally, the restored color image can be obtained by concatenating the three color channels. + +# 4.2. Joint Motion Estimation + +So far, some optical flow estimation methods [21, 39, 42] have been proposed to analyze motion from binary spike signals. However, these methods are designed for GSCs and unsuitable for CSCs that generate Bayer-pattern spike streams. Based on Spike2Flow [42] for grayscale spike signals, we propose a joint motion estimation strategy for Bayer-pattern spike streams. Since the input of Spike2Flow is intervals of temporally neighboring spike signals, we first transform the spike stream $\{\mathbf{S}_i^c\}_{i=1}^N$ of each color channel $c \in \{R, G, B\}$ into intervals as follows: + +$$ +\begin{array}{l} \Psi_ {n} ^ {c} (x, y) = \min \left\{k \mid \mathbf {S} _ {i} ^ {c} (x, y) = 1, k \geq N \right\} \\ - \max \left\{k \mid \mathbf {S} _ {i} ^ {c} (x, y) = 1, k < N \right\}, \tag {4} \\ \end{array} +$$ + +![](images/27874426c8f8873a140e8464482d0ef23fab282985f3ee6a16ee376a6b49d390.jpg) +Figure 2. Pipeline of the motion-guided reconstruction method for CSCs, with two parts: joint motion estimation and missing pixels filling. Spike2Flow [42] is a motion estimation method for GSCs. + +where $\Psi_{n}^{c}(x,y)$ denotes the shortest interval of two temporally neighboring spikes that covers the $n$ -th spike frame. Since there are missing pixels, we employ linear interpolation to estimate the values roughly, resulting in available input for the GSC optical flow estimation method. Then we can get the optical flows from the middle time point to the beginning and end time points of each color channel, $\mathbf{w}_{j,i}^{c}$ and $\mathbf{w}_{j,k}^{c}$ . As the missing pixels of each color impact the quality of optical flows a lot, we propose to exploit the motion consistency of color channels by jointly utilizing the optical flows, which can be formulated as + +$$ +\mathbf {w} _ {j, i}, \mathbf {w} _ {j, k} = \left[ \begin{array}{l} \mathbf {U} _ {j, i} \\ \mathbf {V} _ {j, i} \end{array} \right], \left[ \begin{array}{l} \mathbf {U} _ {j, k} \\ \mathbf {V} _ {j, k} \end{array} \right] = \frac {1}{3} \sum_ {c} \mathbf {w} _ {j, i} ^ {c}, \frac {1}{3} \sum_ {c} \mathbf {w} _ {j, k} ^ {c}, \tag {5} +$$ + +where $\mathbf{U}$ and $\mathbf{V}$ denotes motion components in the $x$ and $y$ directions, respectively. + +# 4.3. Missing Pixel Filling + +As there are spatially missing pixels in each color channel, we propose to collect available temporally neighboring pixels to estimate the values of missing pixel positions. The available pixels for missing positions refer to pixels along the motion trajectory and share the same color with the missing positions. As shown in Fig. 2, we collect the pixels based on the estimated optical flows. Specifically, for a certain non-missing position $(x_{t},y_{t})$ at the $t$ -th time point, the relation between it and its middle-time-point position $(x_{j},y_{j})$ along the motion trajectory can be formulated as + +$$ +\left(x _ {t}, y _ {t}\right) = \left(x _ {j} + \frac {t - j}{\beta - j} \cdot \mathbf {U} _ {j, \beta} \left(x _ {j}, y _ {j}\right), y _ {j} + \frac {t - j}{\beta - j} \cdot \mathbf {V} _ {j, \beta} \left(x _ {j}, y _ {j}\right)\right), \tag {6} +$$ + +where $\beta$ equals to $i$ (when $t < j$ ) or $k$ (when $t > j$ ), and $i, j, k$ denotes the beginning, middle and end time points. Thus, we can warp the non-missing pixels of other time points. The pixels of the middle time point can be estimated by summing the available temporal pixels. To normalize the result, we can divide the sum by the sum of the warped binary color masks, which can be written as + +$$ +\mathbf {I} ^ {c} = \frac {\sum_ {t = 1} ^ {n} \mathcal {W} \left(\mathbf {L} _ {i} ^ {c} , \frac {t - j}{\beta - j} \cdot \mathbf {w} _ {j , \beta}\right)}{\sum_ {t = 1} ^ {n} \mathcal {W} \left(\mathbf {M} ^ {c} , \frac {t - j}{\beta - j} \cdot \mathbf {w} _ {j , \beta}\right)}, c \in \{R, G, B \}, \tag {7} +$$ + +where $\mathbf{I}^c$ denotes the restored color channel, $\mathcal{W}(\cdot)$ denotes the wrapping operation, $\mathbf{L}_i^c$ denotes the preliminary light intensity by TFI, and $\mathbf{M}^c$ denotes the color mask. Considering the positions without available temporal pixels (i.e., $\sum_{t=1}^{n}\mathcal{W}(\mathbf{M}^c,\frac{t-j}{\beta-j}\cdot\mathbf{w}_{j,\beta}) = 0$ ), we can estimate the values via interpolation by non-missing spatial pixels. + +# 5. Self-Supervised Reconstruction Network + +# 5.1. Correlation Analysis of Quantization Noise + +![](images/1e0c9cfd6d1a808925551c2f64513f0d19d7f7d76f6d829cc977f8f1afcc1bed.jpg) +(a) + +![](images/5557a5094b886c71ab0b2aa96b57b15180ea8bf5746db582f63ec55708f9dc7a.jpg) + +![](images/8e6f8e0a9934f7ecbf0be03dc530b749e4b4aa2b1715fd6724e37c38d05a8b36.jpg) +(c) +Figure 3. (a) Hot map for the correlation of quantization noise in the spatial domain. (intensity coefficient $\alpha = 0.6$ ) (b) Hot map for the correlation of quantization noise in the temporal domain. (c) Correlation changes of 4 pixels in (a) over $\alpha$ . (d) Correlation changes of 2 pixels in (b) over $\alpha$ . + +![](images/de93e826009beea5c8599541ce5c7e170bda16356f83b776196b609709b952a9.jpg) +(b) +(d) + +Based on the neighboring pixels along the motion trajectory, we obtain the restored color image from the Bayer-pattern spike stream. However, the result suffers from undesired noise and artifacts. To restore a cleaner color image from the input spike stream, we try to statistically analyze the quantization noise via the simulator. Assuming the quantization noise does not exist, we can infer the nearly exact light intensity by the interval-based TFI [45] as $\delta/\Psi_{n}$ where $\delta$ denotes the maximum dynamic range. Thus, the + +![](images/2ae21568ed7c7982efcb249fbe07e31e852f837690ec04e6febc646bd2b1e8a3.jpg) +Figure 4. (a) Architecture of the self-supervised reconstruction network based on the motion-guided method. (b) The structure of the masked spike encoder (MSE). (c) The structure of the decoder. + +![](images/b5ed3d4cc0b134cb3ae0c15fb54874e924f1ac1afaa07e0ea7eb4952c486c776.jpg) + +![](images/ede2c14397ea208392e3575244bb5c67c2a32d43c05acecbf15390da0853f8ac.jpg) + +quantization noise can be estimated by the difference between the ground truth and the interval-based inference as + +$$ +\widetilde {\mathbf {n}} _ {q} ^ {c} = \left(\frac {\delta}{\boldsymbol {\Psi}} - \widehat {\mathbf {I}} ^ {c}\right) \odot \mathbf {M} ^ {c}, c \in \{R, G, B \}, \tag {8} +$$ + +where $\widetilde{\mathbf{n}}_q^c$ denotes the estimated quantization noise of color $c$ , $\widehat{\mathbf{I}}^c$ denotes the corresponding color channel of the ground truth image, and $\odot$ denotes element-wise multiplication. To explore the correlation of the quantization noise in both spatial and temporal domains, we employ the Pearson correlation coefficient, which can be calculated by + +$$ +\rho_ {s} ^ {c} (x, y) = \frac {\operatorname {c o v} \left(\widetilde {\mathbf {n}} _ {q} ^ {c} (n , \hat {x} , \hat {y}) , \widetilde {\mathbf {n}} _ {q} ^ {c} (n , x , y)\right)}{\sigma \left(\widetilde {\mathbf {n}} _ {q} ^ {c} (n , \hat {x} , \hat {y})\right) \sigma \left(\widetilde {\mathbf {n}} _ {q} ^ {c} (n , x , y))\right)}, \tag {9} +$$ + +$$ +\rho_ {t} ^ {c} (n) = \frac {\operatorname {c o v} \left(\widetilde {\mathbf {n}} _ {q} ^ {c} (\hat {n} , x , y) , \widetilde {\mathbf {n}} _ {q} ^ {c} (n , x , y)\right)}{\sigma \left(\widetilde {\mathbf {n}} _ {q} ^ {c} (\hat {n} , x , y)\right) \sigma \left(\widetilde {\mathbf {n}} _ {q} ^ {c} (n , x , y)\right))}, \tag {10} +$$ + +where $\rho_s^c (x,y)$ denotes the quantization noise correlation coefficient of a certain position $(x,y)$ in spatial domain, $\rho_t^c (n)$ denotes the coefficient of certain position $n$ in temporal domain, $\hat{n}$ , $\hat{x}$ and $\hat{y}$ denotes the central index of temporal, width and height dimensions. Based on synthetic spike streams and ground truth from REDS [29], we statistically obtain the hot maps of noise correlation. To explore the relation between noise correlation and brightness, we use multiple sets of light intensity coefficient $\alpha$ to linearly adjust the scene brightness. According to Fig. 3, the noise correlation in the spatial domain is weak and irregular, while that in the temporal domain shows a stronger correlation and regular tendency with the light intensity changes. + +# 5.2. Network Architecture + +Considering the correlation of quantization noise, we propose a self-supervised reconstruction network to restore a cleaner color image from $\{\mathbf{S}_i\}_{i=1}^N$ , with the results of the motion-guided reconstruction method as pseudo-labels. The architecture of the network is shown in Fig. 4, including a masked spike encoder (MSE) and a decoder. In particular, we develop masked 3D partial convolution (MPConv) to extract temporal features of each color channel from the binary spike signals. To handle the noise, we present temporal masked convolution (TMConv) with a temporal blind spot. Finally, we fuse the features of each color channel and restore the final image by residual blocks [20]. + +# 5.3. Masked Spike Encoder + +The Bayer-pattern spike stream contains binary signals with color and temporal information, which suffers from quantization noise as introduced in the Preliminary. To interpret the binary signals and suppress the noise of each color channel, we propose the masked spike encoder shown in Fig. 4, with $\{\mathbf{S}_i\}_{i=1}^N$ and the mask of a certain color as input. Inspired by partial convolution [27], we extend it to 3D and develop masked 3D partial convolution (MPConv), aiming to extract color and temporal information simultaneously. Given the color mask, we can perform 3D convolution on a temporal clip of spike signals of a certain color, splitting out color channels and extracting temporal features. + +According to the correlation analysis, the quantization noise can be roughly regarded as temporally correlated but spatially independent. Inspired by spatial blind spots [24] to handle correlated noise, we propose temporal masked convolution (TMConv) with a temporal blind spot in the 3D convolutional kernel to suppress the temporally correlated noise. As shown in Fig. 4 (b), part of the spatially central positions in the kernel along the temporal dimension are masked. To handle both the correlated and independent noise, we employ TMConv and the general 3D convolution in two branches and fuse the outputs as the final encoded features of a certain color channel. + +# 6. Experiments + +# 6.1. Implementation Details + +For training, the Bayer-pattern (RGGB) spike frames are randomly cropped into $128 \times 128$ patches. According to previous deep methods [3, 4, 37, 40] and discussion in [43], we set the number of input spike frames $N$ to 41. We employ the Adam optimizer with the $\ell_1$ loss function. The learning rate is $10^{-4}$ and will decay to 0.9 times after every 7500 iterations. Besides, we employ the PyTorch framework and an NVIDIA RTX 3090 GPU. + +# 6.2. Experimental Settings + +Datasets. With the CSC simulator, we generated $5 \times 240 = 1200$ samples from 240 dynamic scenes of REDS [29] as the training dataset. When generating the dataset, the + +
DatasetMethodTypeα = 0.6α = 0.8α = 1.0
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
RBSSWGSE [37]SL31.87dB0.89460.129929.89dB0.88430.128827.85dB0.87260.1332
WGSE*SL32.19dB0.85320.241130.01dB0.83170.254328.22dB0.81460.2658
SJDD-Net [12]SL35.47dB0.90430.119733.98dB0.89740.112632.68dB0.89290.1091
CSpkNet [13]SL35.62dB0.90880.107734.18dB0.90480.099732.75dB0.90030.0986
TFI [45]LF29.24dB0.72890.306325.95dB0.65960.353823.97dB0.62940.3731
TFP [45]LF27.86dB0.68590.389625.60dB0.64820.396023.78dB0.62070.3986
STP [44]LF25.50dB0.62960.335724.01dB0.59910.372223.32dB0.59780.3759
3DRI [9]LF29.59dB0.79000.212028.39dB0.80280.200226.95dB0.80290.1946
OursmgLF31.42dB0.80930.200329.65dB0.78500.197528.08dB0.76600.1990
SJRE [4]SS26.76dB0.52410.426225.77dB0.53410.416624.84dB0.54990.4050
SJRE*SS28.15dB0.60910.491626.90dB0.61080.457025.74dB0.61670.4292
SJRE + OursmgSS32.39dB0.84880.212130.73dB0.83770.208229.21dB0.82830.2089
SJRE* + OursmgSS30.40dB0.79020.307528.35dB0.76370.317026.64dB0.74290.3265
OursSS32.90dB0.85830.194931.34dB0.84860.186629.95dB0.84170.1829
DBSSWGSESL29.45dB0.90180.110027.22dB0.88720.120725.25dB0.87280.1316
WGSE*SL34.58dB0.90650.148532.43dB0.89230.157830.62dB0.88040.1663
SJDD-NetSL37.85dB0.93760.072336.10dB0.92640.074534.85dB0.92690.0709
CSpkNetSL37.66dB0.93910.071436.23dB0.93590.068134.64dB0.93110.0695
TFILF29.80dB0.72100.297826.11dB0.62870.377624.17dB0.59570.4101
TFPLF29.34dB0.75290.318827.25dB0.72770.320225.53dB0.71010.3207
STPLF25.77dB0.65850.321224.15dB0.59780.381423.49dB0.58740.3975
3DRILF30.65dB0.84630.164029.47dB0.85200.156328.07dB0.84780.1559
OursmgLF32.08dB0.82470.159530.19dB0.79320.168328.51dB0.76570.1836
SJRESS26.39dB0.50480.438225.34dB0.51540.431424.32dB0.52670.4237
SJRE*SS29.03dB0.63550.425828.07dB0.64400.384527.12dB0.65210.3540
SJRE + OursmgSS34.50dB0.89270.125632.96dB0.88560.124231.45dB0.87950.1263
SJRE* + OursmgSS32.52dB0.85440.194630.61dB0.83870.202028.96dB0.82620.2098
OursSS34.89dB0.89690.117933.37dB0.88790.114631.98dB0.88150.1126
+ +Table 1. Quantitative comparison on synthetic data with 3 light coefficients. SL, LF and SS denote the supervised learning, learning-free and self-supervised methods, respectively. Red and blue indicate the best performance in methods w and w./o. ground truth, respectively. + +
MethodMethods w Ground Truth (SL)Methods w./o. Ground Truth (LF + SS)
WGSEWGSE*SJDD-NetCSpkNetTFITFPSTP3DRIOursmgSJRESJRE*Ours
PSNR (dB) ↑28.4333.7738.3538.0327.1427.9724.8831.0432.5626.3928.8335.73
SSIM ↑0.93150.93660.96250.96410.70850.76530.66990.89730.85490.57010.69270.9323
LPIPS ↓0.07670.12930.03560.03340.31740.31020.32940.09800.10310.40500.36230.0671
FLOPs (G) ↓3828.13827.04727.43582.4-----3955.13960.23226.3
Params (M) ↓3.80673.80563.75296.6573-----1.19221.19201.0515
Running Time (s) ↓0.89130.89542.51382.61620.04870.085613.45758.1290.73853.27162.70631.5223
+ +Table 2. Quantitative and computational complexity comparison on GBSS with random light coefficient $\alpha \in \left\lbrack {{0.6},{1.0}}\right\rbrack$ . + +light intensity coefficient $\alpha$ , designed to control the brightness of dynamic scenes in the simulator, is randomly chosen from [0.6, 1.0] to improve the model's adaptability to various light intensities. To evaluate the performance, we generate three Bayer-pattern Spike Stream datasets, RBSS, DBSS and GBSS from REDS, DAVIS [30] and GoPro [28]. For RBSS and DBSS, we generate three versions with different brightness settings, $\alpha \in \{0.6, 0.8, 1.0\}$ , while the light intensity coefficient $\alpha$ of GBSS is randomly chosen from [0.6, 1.0]. The main motion sources of the datasets are different, aiming to evaluate the models more fully. We also verify the generalizability of the model trained on BSS [12]. + +Compared Methods. We have three types of compared methods: supervised learning (SL) methods, learning-free (LF) methods and self-supervised (SS) learning methods. For SL methods, we first choose the state-of-the-art (SOTA) CSC reconstruction networks SJDD-Net [12] and CSpkNet + +[13]. We also employ the GSC reconstruction network WGSE [37]. To adapt WGSE to the task, we use Bayer-pattern spike streams for end-to-end learning. Besides, we also adopt the strategy in [33] that estimates the intensities of each color channel separately and then combines them into a color image, denoted as WGSE*. These SL models are supervised by ground truth images. For LF methods, we choose the spike camera demosaicing method 3DRI. We also combine the pixel-independent GSC light inference methods TFI/TFP [45] and STP [44] with a linear interpolation-based demosaicing method. As there are no self-supervised (SS) methods for CSCs, we choose the SOTA self-supervised methods for GSCs, namely SJRE [4]. For color image reconstruction, we adopt the two strategies for WGSE to SJRE, resulting in SJRE and SJRE*. Besides, we replace the pseudo-label of SJRE and SJRE* with pseudo-labels by our motion-guided method ours $_{mg}$ . + +![](images/0c10e7378ced99bf3c2b420fd7d413b6872bc5ecbba617290b5c4a7230070554.jpg) +Figure 5. Visual comparison on the real-world captured Bayer-pattern spike streams from BSS [12]. The first sample records a fast-rotating water bottle (object motion). The second one is captured by a fast-shaking CSC (camera motion). Zoom in for better visual comparison. + +# 6.3. Comparison Results + +Quantitative Comparison. Table 1 presents the quantitative results on RBSS and DBSS. It can be found that the supervised learning methods perform better than most methods without supervised training. Supervised by ground truth images, the SL method CSpkNet achieves the best performance in most cases. Our SS method performs the best among the methods without ground truth $(\mathrm{LF} + \mathrm{SS})$ , achieving competitive performance with the supervised WGSE-based methods. $\mathrm{Ours}_{mg}$ achieves the best performance among all the LF methods. Trained with the pseudo-labels by $\mathrm{ours}_{mg}$ , our network (ours) obtains better results than $\mathrm{ours}_{mg}$ . Besides, WGSE* and SJRE*, which adopt the color-independent strategy [33], show better PSNR performance than WGSE and SJRE, respectively. With pseudolabels by $\mathrm{ours}_{mg}$ , the SS SJRE-based methods achieve significant performance gains. As shown in Table 2, results on the dataset with random $\alpha$ share a similar situation to that in Table 1. Besides, ours gains a minimal number of parameters and floating point of operations (FLOPs) among all the learning-based methods $(\mathrm{SL} + \mathrm{SS})$ . + +Visual Comparison. As shown in Fig. 5 and Fig. 6, we performed comparison experiments on real-world captured Bayer-pattern spike streams from BSS [12] to verify the practical performance. In contrast, the results of our self-supervised network achieve better details and less noise than other methods, including the SOTA supervised methods SJDD-Net and CSpkNet. Considering quantization noise, our network suppresses the noise compared with ours $_{mg}$ . Compared to WGSE and SJRE, WGSE* and SJRE* with another color-independent strategy obtain more precise color. With pseudo-labels by ours $_{mg}$ , the noise and artifacts from SJRE and SJRE* are reduced a lot, achieving better visual quality. Besides, the results of 3DRI enjoy good details but severe noise. The visual results of TFI and STP also tend to be noisy, while those of TFP look blurry. + +# 6.4. Ablation Study + +Motion-Guided Mehodt. Table 3 presents the ablation study on RBSS. First, we perform independent motion estimation for each color channel in Case (A). To verify the interpolation, we only extract pixels of each color with- + +![](images/122a9b854386c1393fd14617406b336fb1a6de2871476672d6c5156e856da38e.jpg) +Figure 6. Visual comparison on more real-world captured samples from BSS. BSS is a widely used real-world captured dataset for CSC reconstruction [11-13], involving both object and camera motion. Please enlarge the figure to better see the visual difference. + +out spatial interpolation for motion estimation in Case (B). Then, the temporal pixels are filtered directly without the motion-guided search in Case (C). We also replace the intensity inference method TFI by TFP with multiple windows in Cases (D), (E) and (F). After that, we remove the consideration of color layout and search pixels directly to verify its effectiveness in Case (G). Finally, we remove the no-temporal-pixel position filling operation in Case (H). + +
CaseSetting Descriptionα = 0.6α = 0.8α = 1.0Avg. ↑
(A)w./o. joint motion estimation31.2729.5027.9129.56
(B)w./o. spatial interpolation31.4429.5127.8129.59
(C)w./o. motion-guided search28.4825.9624.1726.20
(D)TFP (short) for inference30.2528.8427.4128.83
(E)TFP (medium) for inference30.0727.9926.2128.09
(F)TFP (long) for inference28.9826.7124.8726.85
(G)Pixel search w./o. color layout30.1627.8626.2328.08
(H)w./o. filling no-pixel positions31.3129.5427.9629.60
(I)Our motion-guided method31.4229.6528.0829.72
+ +Self-Supervised Network. We perform the study for the self-supervised network as shown in Table 4. To verify the performance of pseudo-labels by $\mathrm{ours}_{mg}$ , we replace it with TFI-based and TFP-based pseudo-labels in Cases (A), (B), (C) and (D). Then, we employ 2D/3D convolution as baselines in Case (E) and Case (F) to verify the performance of the MSE module. In Case (G), we remove the MPConv to show its effectiveness and the necessity of color extraction. To demonstrate the collaboration of convolution for spatial and temporal noise, we keep only the 3DConv or TMConv in Cases (H) and (I). We also employ the same type in the two branches to further verify it in Cases (J) and (K). + +Table 3. Ablation study for the motion-guided method. Deeper blocks denote higher PSNR(dB). + +
CaseSetting Descriptionα = 0.6α = 0.8α = 1.0Avg. ↑
(A)TFI pseudo-label30.6128.7627.0628.81
(B)TFP pseudo-label (short)28.0526.3424.6026.33
(C)TFP pseudo-label (medium)28.6526.8225.1426.87
(D)TFP pseudo-label (long)28.8726.5024.6026.66
(E)Replacing MSE with 2DConv30.1828.2226.4728.29
(F)Replacing MSE with 3DConv30.3128.3626.6628.44
(G)w./o. MPConv in MSE30.6028.7227.0728.80
(H)Only a single 3DConv in MSE32.8131.2529.8431.30
(I)Only a single TMConv in MSE32.8331.2729.8631.32
(J)Only 3DConv in two branches32.8731.2929.8631.34
(K)Only TMConv in two branches32.8431.2429.7831.29
(L)Our self-supervised network32.9031.3429.9531.40
+ +Table 4. Ablation study for the self-supervised network. + +# 7. Conclusion + +We propose a motion-guided reconstruction method for CSCs. To be specific, we develop a joint motion estimation pipeline for the Bayer-pattern spike stream. Then, we utilize the color layout and motion information to fill the missing pixel position of each color channel according to the temporally neighboring pixels of the corresponding color. To handle quantization noise, we analyze the correlation of the noise and develop a self-supervised network based on our masked spike encoder. Experiments on real-world captured data show that our method achieves better visual performance than the state-of-the-art supervised method. + +# Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China under Grants 22127807, 62072009, 62422601, and in part by the Beijing Nova Program (20230484362). + +# References + +[1] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \times 180$ 130db $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2 +[2] Claudio Bruschini, Samuel Burri, Scott Lindner, Arin C Ulku, Chao Zhang, I Michel Antolovic, Martin Wolf, and Edoardo Charbon. Monolithic spad arrays for high-performance, time-resolved single-photon imaging. In 2018 International Conference on Optical MEMS and Nanophotonics (OMN), pages 1-5. IEEE, 2018. 2 +[3] Shiyan Chen, Chaoteng Duan, Zhaofei Yu, Ruiqin Xiong, and Tiejun Huang. Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. IJCAI, 2022. 5 +[4] Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 350-358, 2023. 2, 5, 6 +[5] Yiheng Chi, Abhiram Gnanasambandam, Vladlen Koltun, and Stanley H Chan. Dynamic low-light imaging with quanta image sensors. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 122-138. Springer, 2020. 2 +[6] Joon Hee Choi, Omar A Elgendy, and Stanley H Chan. Image reconstruction for quanta image sensors using deep neural networks. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6543-6547. IEEE, 2018. 2 +[7] Siwei Dong, Tiejun Huang, and Yonghong Tian. Spike camera and its coding methods. In Data Compression Conference (DCC), page 437, 2017. 1 +[8] Siwei Dong, Lin Zhu, Daoyuan Xu, Yonghong Tian, and Tiejun Huang. An efficient coding method for spike camera using inter-spike intervals. In Data Compression Conference (DCC), page 568, 2019. 1 +[9] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In IEEE International Conference on Image Processing (ICIP), pages 1461-1465. IEEE, 2022. 1, 2, 6 +[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. High-speed scene reconstruction from low-light spike streams. In IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2022. 1 +[11] Yanchen Dong, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Superresolution reconstruction from bayer-pattern spike streams. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24871-24880, 2024. 1, 8 +[12] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Joint demosaicing and denoising for spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1582-1590, 2024. 1, 2, 6, 7 + +[13] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Learning a deep demosaicing network for spike camera with color filter array. IEEE Transactions on Image Processing, 33:3634-3647, 2024. 2, 6, 8 +[14] Yanchen Dong, Ruiqin Xiong, Xiaopeng Fan, Shuyuan Zhu, Jin Wang, and Tiejun Huang. Dynamic scene reconstruction for color spike camera via zero-shot learning. IEEE Transactions on Computational Imaging, 11:129-141, 2025. 1 +[15] Neale AW Dutton, Luca Parmesan, Andrew J Holmes, Lindsay A Grant, and Robert K Henderson. $320 \times 240$ oversampled digital single photon counting image sensor. In Symposium on VLSI Circuits Digest of Technical Papers, pages 1-2. IEEE, 2014. 2 +[16] Neale AW Dutton, Istvan Gyongy, Luca Parmesan, Salvatore Gnecchi, Neil Calder, Bruce R Rae, Sara Pellegrini, Lindsay A Grant, and Robert K Henderson. A spad-based qvga image sensor for single-photon counting and quanta imaging. IEEE Transactions on Electron Devices, 63(1):189-196, 2015. +[17] NA Dutton12, Luca Parmesan12, Salvatore Gnecchi12, Istvan Gyongy, Neil Calder, Bruce R Rae, Lindsay A Grant, and Robert K Henderson. Oversampled itof imaging techniques using spad-based quanta image sensors. In Proc. Int. Image Sensor Workshop, pages 170-173, 2015. 2 +[18] Abhiram Gnanasambandam, Omar Elgendy, Jiaju Ma, and Stanley H Chan. Megapixel photon-counting color imaging using quanta image sensor. Optics express, 27(12):17298-17310, 2019. 2 +[19] Istvan Gyongy, Neale Dutton, Parmesan Luca, and Robert Henderson. Bit-plane processing techniques for low-light, high speed imaging with a spad-based qis. In International Image Sensor Workshop, 2015. 2 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5 +[21] Liwen Hu, Rui Zhao, Ziluo Ding, Lei Ma, Boxin Shi, Ruiqin Xiong, and Tiejun Huang. Optical flow estimation for spiking camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17844-17853, 2022. 3 +[22] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, et al. $1000\times$ faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2023. 1 +[23] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew J Davison. Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ, 43:566-576, 2008. 2 +[24] Wooseok Lee, Sanghyun Son, and Kyoung Mu Lee. Apsbn: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17725-17734, 2022. 5 +[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ 120db $15\mu s$ latency asynchronous temporal con + +trast vision sensor. IEEE Journal of Solid-State Circuits, 43 (2):566-576, 2008. 1, 2 +[26] Martin Litzenberger, Christoph Posch, D Bauer, Ahmed Nabil Belbachir, P Schon, B Kohn, and H Garn. Embedded vision system for real-time object tracking using an asynchronous transient vision sensor. In 2006 IEEE 12th Digital Signal Processing Workshop & 4th IEEE Signal Processing Education Workshop, pages 173-178. IEEE, 2006. 1, 2 +[27] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European Conference on Computer Vision (ECCV), pages 85-100, 2018. 5 +[28] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3883-3891, 2017. 6 +[29] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1996-2005, 2019. 5 +[30] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 6 +[31] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A QVGA 143db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1): 259-275, 2010. 2 +[32] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 2 +[33] Cedric Scheerlinck, Henri Rebecq, Timo Stoffregen, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Ced: Color event camera dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 0–0, 2019. 2, 6, 7 +[34] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 156-163, 2020. 2 +[35] Yuanlin Wang, Ruiqin Xiong, Jing Zhao, and Tiejun Huang. Reconstruct dynamic scene for spike camera based on 3d space time similarity. In IEEE International Conference on Image Processing (ICIP), pages 1595-1601. IEEE, 2024. 1 +[36] Xijie Xiang, Lin Zhu, Jianing Li, Yixuan Wang, Tiejun Huang, and Yonghong Tian. Learning super-resolution reconstruction for high temporal resolution spike stream. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 1 + +[37] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 137–147, 2023. 1, 2, 5, 6 +[38] Jing Zhao, Ruiqin Xiong, and Tiejun Huang. High-speed motion scene reconstruction for spike camera via motion aligned filtering. In IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5, 2020. 2 +[39] Jing Zhao, Ruiqin Xiong, Rui Zhao, Jin Wang, Siwei Ma, and Tiejun Huang. Motion estimation for spike camera data sequence via spike interval analysis. In IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 371-374. IEEE, 2020. 3 +[40] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2ImgNet: Learning to reconstruct dynamic scene from continuous spike stream. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2021. 1, 2, 5 +[41] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE Transactions on Computational Imaging, 8:12-27, 2021. 2, 3 +[42] Rui Zhao, Ruiqin Xiong, Jing Zhao, Zhaofei Yu, Xiaopeng Fan, and Tiejun Huang. Learning optical flow from continuous spike streams. Advances in Neural Information Processing Systems, 35:7905-7920, 2022. 1, 3, 4 +[43] Rui Zhao, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Shuyuan Zhu, Lei Ma, and Tiejun Huang. Spike camera image reconstruction using deep spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 5 +[44] Yajing Zheng, Lingxiao Zheng, Zhaofei Yu, Boxin Shi, Yonghong Tian, and Tiejun Huang. High-speed image reconstruction through short-term plasticity for spiking cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6358-6367, 2021. 2, 6 +[45] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In IEEE International Conference on Multimedia and Expo (ICME), pages 1432-1437. IEEE, 2019. 2, 3, 4, 6 +[46] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Retina-like visual image reconstruction via spiking neural model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1438-1446, 2020. 2 +[47] 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. 2 \ No newline at end of file diff --git a/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/images.zip b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..179dd5af81629929cb405b8031990c9600fc9a12 --- /dev/null +++ b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ded37311b66d91e7c3654428e0b75eb162783bb65b87321652b3ef128ba0c2b +size 1237757 diff --git a/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/layout.json b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fcbc98ab290b5be2910bc32ecc15fa34dfda7e9a --- /dev/null +++ b/CVPR/2025/Self-Supervised Learning for Color Spike Camera Reconstruction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48d455ec05987dff081a52c0ff52acaee67a4f85cf9b32966ede0106f00da756 +size 389275 diff --git a/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/67b3acbf-55e8-459a-a4da-ba8ff6ef0c41_content_list.json b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/67b3acbf-55e8-459a-a4da-ba8ff6ef0c41_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4ef5200a2bf0f37024e1328e4429eaacc8d22340 --- /dev/null +++ b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/67b3acbf-55e8-459a-a4da-ba8ff6ef0c41_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c17f913ef2bc6c5d3832f65601c7ca2e99feb68c826f0329c56ec3c4f82ef7f2 +size 74485 diff --git a/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/67b3acbf-55e8-459a-a4da-ba8ff6ef0c41_model.json b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/67b3acbf-55e8-459a-a4da-ba8ff6ef0c41_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a18dd1dee38b2c9cec065fbb47e4ffae35d3b4fe --- /dev/null +++ b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/67b3acbf-55e8-459a-a4da-ba8ff6ef0c41_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:621b1b36186793813e1491565d7774c4899cf798226f34ac9916365fae532c1c +size 94616 diff --git a/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/67b3acbf-55e8-459a-a4da-ba8ff6ef0c41_origin.pdf b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/67b3acbf-55e8-459a-a4da-ba8ff6ef0c41_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..df691b57bbca1d79426097b04d484db2949a68e9 --- /dev/null +++ b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/67b3acbf-55e8-459a-a4da-ba8ff6ef0c41_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f994a3e203bb79bd417a4f46c94f3b598ae231ca7ed2561d630dac652311ca +size 7428509 diff --git a/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/full.md b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4bba5b5af1b8fa2d04acca38bfd84e15ef041459 --- /dev/null +++ b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/full.md @@ -0,0 +1,292 @@ +# Self-Supervised Spatial Correspondence Across Modalities + +Ayush Shrivastava Andrew Owens University of Michigan + +https://ayshrv.com/cmrw + +![](images/91f4b849deeb2b89b24427868d028905d05cd5e91069f2988f0ed91ce410e32c.jpg) +RGB + +RGB + +![](images/fc9ab78cca690fcb845d3f2b50f8ae4b66dce13f257699d2ee488ae63ba10443.jpg) +Depth + +![](images/bb57c5ce92a3818fd22a3e0364f14ced7239355e179951be3c3d1dc1c39b99ba.jpg) +RGB + +![](images/0919fa7921d0e2d701680e70c40ddadd341baaccbac5bccf2cf68ec05f1a1be4.jpg) +Metric Correspondence + +![](images/72a1cade89e777ea415415d6307cc8adbbc992428d14f28feb0d2a6466e92ccf.jpg) +Photo + +![](images/ce31d101e798a9c72d6f63e2ac8d318be1f9bd677ff4096f91ff84d4fb42aeff.jpg) +Sketch + +![](images/9721960c0cdf375614b12a86e0f2822a593b0c1c2e75f8884ebd745daabb420d.jpg) + +![](images/4fdb2c40925289e561ced46b8e67dddd3c9b2ecfeb216780da777307dc57e976.jpg) +Photorealistic style +(b) Semantic Correspondence +Figure 1. Finding spatial correspondences across modalities. We present a method for cross-modal matching, trained entirely through self-supervision using a simple formulation based on contrastive random walks [14]. (a) Given two images taken by different visual modalities and at different positions and times, we predict the pairs of image patches that physically correspond to the same points. (b) We also apply our method to semantic matching tasks, using a visual encoder initialized with pretrained DINOv2 [27] weights and fine-tuned during training. These include tasks such as photo-sketch alignment [26] and style-based matching between images of different styles generated with a text-to-image model [19]. + +![](images/157efb1a24bef9061f3c70ecea4a638c94c33b80ac362bea4088eaa920477d9a.jpg) +Anime style + +# Abstract + +We present a method for finding cross-modal space-time correspondences. Given two images from different visual modalities, such as an RGB image and a depth map, our model identifies which pairs of pixels correspond to the same physical points in the scene. To solve this problem, we extend the contrastive random walk framework to simultaneously learn cycle-consistent feature representations for both cross-modal and intra-modal matching. The resulting model is simple and has no explicit photo-consistency assumptions. It can be trained entirely using unlabeled data, without the need for any spatially aligned multimodal image pairs. We evaluate our method on both geometric and semantic correspondence tasks. For geometric matching, we consider challenging tasks such as RGB-to-depth and RGB-to-thermal matching (and vice versa); for semantic matching, we evaluate on photo-sketch and cross-style image alignment. Our method achieves strong performance across all benchmarks. + +# 1. Introduction + +Cameras take a multitude of different forms. While RGB, thermal, and depth cameras each record images, what is stored within each pixel differs drastically. Consequently, when a scene is captured using cameras that are based on different sensory modalities, it is difficult to determine which pixels within them correspond to the same physical points. These cross-modal pixel correspondences have a number of applications, such as cross-modal registration, 3D reconstruction, and multimodal data fusion [1, 6, 22, 43]. Recent work on multimodal learning has provided effective ways of learning cross-modal correspondences through self-supervision, yet these methods largely rely on having paired data from multiple sensors, which is not available for pixel-level correspondence. + +Traditional self-supervised methods for pixel correspondences often make strong assumptions about the visual appearance of pixels that should be matched together, such as by assuming that matching image patches should be photo- + +consistent [16, 17, 34, 47, 53], or that cross-modal translation methods can estimate one modality from another as part of the matching process [1]. However, these assumptions are frequently violated in multimodal data, making it challenging to use them as general-purpose multimodal matching methods. Depth images and RGB images, for example, may look alike to a human, but their intensity values record information that is so different—depth vs. lightness—that they cannot be matched using local information. And cross-modal prediction requires being able to solve a notoriously difficult problem—monocular depth estimation—without explicit paired data. + +We take inspiration from recent advances in self-supervised tracking [2, 14, 23, 44] that learn correspondences between video frames through cycle consistency. These methods assume that the content within a scene persists between frames, and that there should thus be a one-to-one correspondence between the pixels in them. Our approach extends the contrastive random walk [14] to the cross-modal pixel correspondence problem. We create a directed graph in which nodes correspond to image patches in each modality, and where edges connect patches across modalities. We train a network, based on the recently proposed global matching transformer [32, 45], to assign transition probabilities to pairs of patches for a random walk. A random walker uses these transition probabilities to step through the graph, moving from patches in one modality to another, and then back. To train the model, we maximize the walker's return probability, thus encouraging matches to be cycle-consistent. + +A key challenge in our setting is that, unlike many correspondence tasks, the patches we match are visually very different. We therefore incorporate intra-modal random walks between images of the same modality, such that the model must learn a representation that simultaneously allows it to match intra- and inter-modally. To get these image pairs, we apply data augmentation to input images, simulating the challenges of matching frames of a video. We also show that, in contrast to image-based matching [32], spatial smoothness priors are important for obtaining strong performance. Moreover, in contrast to other cross-modal correspondence methods [1, 17, 34], our method does not rely on hand-crafted measures of photoconsistency. Instead, it learns visual similarity through cycle-consistent representation learning, defined by the learned embedding space, thereby making it possible to apply it to different pairs of modalities without any modifications. + +We evaluate our model on both geometric and semantic cross-modal correspondence tasks. For geometric matching, we focus on two challenging tasks: RGB-to-depth and RGB-to-thermal matching (Figure 1a). In both cases, we match images captured at different times and positions, with different sensors. For semantic matching (Figure 1b), we + +use a visual encoder initialized with pretrained DINOv2 [27] weights and fine-tune it during training. We evaluate on photo-to-sketch alignment, where our method performs competitively with approaches specifically tailored to this task, and on a new cross-style image matching task, using images generated in different styles from the same diffusion model [19]. Our work makes several contributions: + +- We show that space-time cross-modal pixel correspondence can be learned from unlabeled data through cycle consistency. +- We extend the contrastive random walk framework for cross-modal pixel correspondence. +- Through experiments, we show that our model successfully finds correspondences in geometric tasks such as RGB-to-depth and RGB-to-thermal, significantly outperforming prior methods. In semantic tasks like photo-to-sketch matching, our method is competitive with other self-supervised approaches. +- We propose benchmarks for evaluating RGB-to-depth, RGB-to-thermal, and cross-style image matching. For cross-style matching, we use generative models to synthesize images with similar content but different styles. + +# 2. Related Work + +Cycle consistency for correspondence. Zhou et al. [51] proposed learning correspondences across 2D views of 3D objects, by matching the viewpoints to a 3D model, then projecting back to other images. In contrast, we focus on matching two images that differ in modality, space, and time, rather than just 3D viewpoint. Later work extended cycle consistency to space-time tracking. Wang et al. [44] trained a model to track forward and backward in time, penalizing deviation from the original position, using a formulation based on the spatial transformer [15]. Jabri et al.[14] framed this as a random walk on a space-time graph, enabling self-supervised tracking, followed by improvements such as shortcut removal[37] and smoothness priors [2] for fine-grained correspondence. Recent works further integrated this framework with global matching transformers [32, 45] to address the tracking-any-point problem [4]. + +While these works focus on temporal matching, we extend contrastive random walks to cross-modal matching, introducing both intra- and inter-modal random walks. Although contrastive random walks have been explored in multimodal settings like image-audio alignment [11] or audio-to-audio matching [3], these approaches operate at the patch or node level and do not aim for dense pixel-to-pixel correspondence, which is the focus of our work. + +Cross-modal pixel correspondence. Cross-modal correspondence has been explored through contrastive learning [10, 28, 40], but these approaches typically learn global or patch-level embeddings, not dense pixel-wise matches. + +![](images/5fe0aa7cb9870c4a955ae2b7e960bbbb45bc00f0222950859681334af5293c6c.jpg) +Figure 2. Model Architecture. We learn to find pixel-level correspondences between images that may differ in sensory modality, time, and scene position. Given images from two modalities (e.g., unpaired RGB and depth images from the same scene), we perform a contrastive random walk on a graph whose nodes come from patches within the two images using a global matching transformer architecture [32]. We simultaneously perform auxiliary intra-modal random walks within each modality's augmented crops of images to improve the model's ability to avoid local minima during optimization. Through this process, we learn to match in both directions (e.g., RGB-to-depth and depth-to-RGB). + +![](images/ba6564d5e21daa99a5bf35c484cbd9d4f94cfdaf9e251095248735eaac17271e.jpg) + +Arar et al.[1] proposed a method for RGB-infrared matching by jointly training a GAN for modality translation and a spatial transformation network. In contrast, our model does not require explicit cross-modal translation; it learns from cycle-consistency alone. Other methods such as[26] address photo-sketch alignment through a two-stage process involving cross-modal embeddings, whereas we learn fine-grained correspondence directly through self-supervision. Some recent work has explored matching across 2D-3D modalities (e.g., RGB and point clouds [22, 43]), but these approaches often rely on explicit geometric priors and operate on sparse data like LiDAR. Our method, by contrast, is designed for dense 2D image domains and does not rely on any handcrafted similarity metric. + +Self-supervised correspondence. Unsupervised learning of correspondence has a long history in optical flow [16, 17, 34, 47, 53], often relying on photometric loss as a training signal. Other approaches use cues like color similarity [20, 23, 42] or temporal coherence [7, 46] for self-supervised tracking. However, these techniques assume single-modality inputs and struggle to generalize to cross-modal settings. Simulated data has also been used for learning correspondence [4, 5, 8, 9, 35], but generating realistic, time-varying multimodal signals remains a significant challenge. In contrast, our method learns from real multimodal videos without requiring ground-truth annotations or simulation. + +# 3. Method + +Our goal is to learn pixel-wise correspondences between images from different modalities (e.g., RGB, thermal, depth, sketch, stylistically diverse images) solely through self-supervision. We build upon the contrastive random walk framework [14] and extend it to handle multiple modalities (Figure 2). The recently proposed Global Matching Random + +Walk (GMRW) model [32] allows for dense correspondences between RGB images by enforcing cycle consistency via contrastive random walks over videos [32, 45]. We leverage this framework to train on unlabeled videos from different modalities, learning cycle-consistent tracking across pairs such as RGB-depth, RGB-thermal, photo-sketch, and cross-style images. + +However, we find that directly optimizing the model for cross-modal cycle consistency leads to poor convergence: training plateaus at a suboptimal solution with high loss and non-semantic matches. We hypothesize that this difficulty arises due to the substantial domain gap between modalities. For instance, unlike space-time matching—where initial random features provide reasonable gradients—cross-modal inputs may align arbitrarily (e.g., bright regions in RGB could align with distant areas in depth due to similarly high intensity values). To mitigate this, we introduce auxiliary intra-modality random walk supervision, encouraging the model to learn representations effective for both intra- and cross-modality matching. + +# 3.1. Model architecture + +To address the challenge of cross-modal dense correspondence, we require an architecture capable of producing high-resolution matches while remaining modality-agnostic. We adopt the Global Matching Random Walk (GMRW) model [32], which performs all-pairs pixel matching [39, 45] and is trained using a cycle consistency objective via contrastive random walks. Crucially, it does not rely on modality-specific assumptions like photo-consistency. Each modality uses a dedicated visual encoder to produce input tokens, while a shared Transformer backbone performs the global matching across modalities. + +Image Features. For geometric tasks, we use a CNN-based encoder to extract visual features. For each image $I_{t}^{m}$ at time $t$ and modality $m \in \{\text{rgb}, \text{depth}, \text{thermal}\}$ , we + +compute $d$ -dimensional features $\phi(I_t^m) \in \mathbb{R}^{\frac{H}{c} \times \frac{W}{c} \times d}$ with a downsampling factor $c = 4$ . Each modality has its own encoder, and we append 2D positional encodings to the features. For depth inputs, the single-channel image is replicated across three channels to match the RGB architecture. + +For semantic tasks (e.g., photo-sketch, cross-style), we use DINOv2 [27] as a shared visual encoder across modalities. We initialize it with pretrained weights and finetune it, leveraging the rich semantic priors learned during DINOv2 pretraining. + +Cross-modal Matching Transformer. Given a pair of images $I_{t}^{m_{1}}$ and $I_{t + k}^{m_2}$ , we feed their extracted features into a shared Transformer consisting of six blocks of self-attention, cross-attention, and feed-forward layers. The final layer produces correlation features $F_{t}^{m_{1}}$ and $F_{t + k}^{m_2}$ . This module supports both intra-modality ( $m_{1} = m_{2}$ ) and cross-modality ( $m_{1} \neq m_{2}$ ) matching. + +Cross-modal Transition Matrix. We compute the transition matrix from $F_{t}^{m_{1}}$ and $F_{t + k}^{m_{2}}$ as $A_{t,t + k}^{m_1,m_2} = \mathrm{softmax}(F_t^{m_1}(F_{t + k}^{m_2})^\top /\tau)$ , which represents the likelihood of a pixel in modality $m_{1}$ at time $t$ corresponding to a pixel in modality $m_{2}$ at time $t + k$ . This matrix defines the transition probabilities for contrastive random walks (CRW) in both intra- and cross-modality matching. Following [32, 45], we normalize features using $\tau = \sqrt{d}$ instead of the L2 normalization used in earlier work [2, 14, 37]. + +The expected change in pixel position is computed as: + +$$ +\mathbf {f} _ {t, t + k} ^ {m _ {1}, m _ {2}} = \mathbb {E} _ {A _ {t, t + k}} \left[ A _ {t, t + k} ^ {m _ {1}, m _ {2}} D - D \right] \tag {1} +$$ + +where $\mathbf{f}_{t,t + k}$ is the predicted flow between modalities $m_{1}$ and $m_2$ , $D$ is the constant pixel coordinate grid and $A_{t,t + k}^{m_1,m_2}$ is the transition matrix from frame $t$ to $t + k$ . + +# 3.2. Learning cross-modal correspondences + +Cross-modal cycle-consistency. To learn cross-modal correspondences, we extend the cycle consistency objective from GMRW [32] to multi-modal data. Given a palindrome sequence $\{I_t^{m_1}, I_{t+k}^{m_2}, I_t^{m_1}\}$ , we treat it as a space-time graph and train our model to perform random walks across it. The model estimates two transition matrices: $A_{t,t+k}^{m_1,m_2}$ which maps pixels from modality $m_1$ at time $t$ to modality $m_2$ at time $t+k$ , and $A_{t+k,t}^{m_2,m_1}$ , which returns the walker back to modality $m_1$ at time $t$ . We chain these two transitions to obtain a round-trip path from a point in $I_t^{m_1}$ , through $I_{t+k}^{m_2}$ , and back. To enforce cycle consistency, we maximize the probability that the walker returns to its original position. This is done using the label-warping loss from [32]: + +$$ +\mathcal {L} _ {\text {c r o s s - c r w}} = \mathcal {L} _ {\mathrm {C E}} \left(A _ {t, t + k} ^ {m _ {1}, m _ {2}} A _ {t + k, t} ^ {m _ {2}, m _ {1}}, T _ {f} ^ {b} (I)\right), \tag {2} +$$ + +where $I$ is the identity matrix and $T_{f}^{b}(I)$ is the warped label target designed to prevent shortcut learning [32], and $\mathcal{L}_{\mathrm{CE}}$ is the cross-entropy loss. + +Intra-modal cycle-consistency. We further apply the same contrastive random walk objective within each modality to stabilize training. For a palindrome $\{I_{\mathrm{ori}}^{m_i}, I_{\mathrm{aug}}^{m_i}, I_{\mathrm{ori}}^{m_i}\}$ , where $I_{\mathrm{aug}}^{m_i}$ is an augmented crop of $I_{\mathrm{ori}}^{m_i}$ , the intra-modal cycle loss is: + +$$ +\mathcal {L} _ {\text {i n t r a - c r w}} = \sum_ {i = 1} ^ {2} \mathcal {L} _ {\mathrm {C E}} \left(A _ {\text {o r i , a u g}} ^ {m _ {i}} A _ {\text {a u g , o r i}} ^ {m _ {i}}, T _ {f} ^ {b} (I)\right) \tag {3} +$$ + +Smoothness loss. To encourage spatial coherence, we apply an edge-aware smoothness loss [17] — used in prior contrastive random walk work [2, 32]. It penalizes large second-order derivatives in flow, weighted by image gradients. We apply this loss only when the source image is from the RGB modality, where visual similarity is a reliable cue for perceptual grouping. The loss is defined as: + +$$ +\mathcal {L} _ {\text {s m o o t h}} = \mathbb {E} _ {p} \sum_ {d \in \{x, y \}} \exp (- \lambda_ {c} I _ {d} (p)) | \frac {\partial^ {2} \mathbf {f} _ {s , t} (p)}{\partial d ^ {2}} | \tag {4} +$$ + +where the $p$ is a pixel in the RGB image, $I_{d}(p)$ is the average gradient magnitude across color channels in direction $d$ . The coefficient $\lambda_{c}$ controls edge sensitivity. + +Overall loss. The final training objective combines all losses: + +$$ +\mathcal {L} _ {\text {c r o s s - c r w}} + \mathcal {L} _ {\text {i n t r a - c r w}} + \lambda_ {s} \mathcal {L} _ {\text {s m o o t h}} \tag {5} +$$ + +# 4. Experiments + +Our method estimates pixel-level correspondences between pairs of images across different modalities. We evaluate its performance on both geometric and semantic correspondence tasks. For geometric matching, we consider two settings: RGB-Depth and RGB-Thermal. For RGB-Depth, we use the NYU Depth V2 dataset [33] for both training and evaluation. For RGB-Thermal, we train and evaluate on two datasets: the indoor Thermal-IM dataset [38] and the outdoor KAIST dataset [13]. Notably, our method does not require spatially or temporally aligned image pairs during training. We sample frames at different time steps, allowing for scene changes and motion, while assuming partial scene overlap. For semantic correspondence, we consider two tasks. The first is photo-sketch alignment, using the PSC6K dataset [26], which includes annotated keypoints that match real-world photos with corresponding human-drawn sketches. The second task is cross-style image matching, where the goal is to match corresponding points across stylistic variants (e.g., photorealistic, anime, watercolor) of the same scene. We construct this benchmark using a text-to-image generation model conditioned on consistent prompts and varying style modifiers. For all benchmarks, evaluation sets are created by manually annotating keypoints across modalities or, where applicable, by propagating tracked points from the RGB domain using a point-tracking algorithm. + +Table 1. RGB-Depth and RGB-Thermal Matching. We compare our method with supervised and several unsupervised baselines on NYU-Depth V2, Thermal-IM and KAIST datasets. Our method performs best when computing cross-modality matching, significantly outperforming the baselines. D: Depth, T: Thermal, SSL: Self-Supervised Learning + +
MethodSSLNYU-Depth, < δxavg↑Thermal-IM, < δxavg↑KAIST, < δxavg↑
Intra-modalCross-modalIntra-modalCross-modalCross-modal
RGB → RGBD → DRGB → DD → RGBRGB → RGBT → TRGB → TT → RGBRGB → TT → RGB
RAFT [39]91.559.27.91.381.753.35.60.929.27.4
GMFlow [45]79.758.212.712.582.253.43.82.623.122.3
Arar et al. [1]✓--1.30.8--2.31.92.14.7
CycleGANRGB→D/T + GMFlow✓--8.57.8--7.97.18.38.2
CycleGAND/T→RGB + GMFlow✓--16.216.6--6.15.86.87.4
ARFlow [25]✓76.153.57.57.482.153.312.513.231.030.4
ARFlow (Retrained)✓--9.38.1--13.413.129.127.7
DIFT [36]✓40.618.53.34.368.250.317.518.27.18.8
SD-DINO [48]✓25.213.97.86.444.549.929.334.919.620.6
Ours✓80.162.333.534.381.653.141.847.935.234.1
+ +# 4.1. Geometric Correspondence + +RGB-Depth Training. We train our model on the NYU Depth V2 dataset [33], which contains approximately 400K unlabeled RGB-D frames. Training is performed in three stages: 1) intra-modality CRW applied to RGB-RGB and Depth-Depth pairs, 2) additionally cross-modality CRW applied to RGB-Depth and Depth-RGB pairs, 3) adding the smoothness loss to encourage spatial coherence. During training, we randomly sample two frames $(I_{t}^{m_{1}}, I_{t+k}^{m_{2}})$ from a scene, separated by a few random timesteps, drawn from both modalities. We then apply different random resized crops to forward $(I_{t}^{m_{1}}, I_{t+k}^{m_{2}})$ and backward images $(I_{t,\mathrm{aug}}^{m_{1}})$ [32] and train for cycle-consistency losses. + +RGB-Thermal Training. We use the Thermal-IM and KAIST datasets for RGB-Thermal training. Thermal-IM consists of 783 video sequences with RGB and thermal modalities, though they are spatially unaligned. KAIST contains 320 video sequences with spatially aligned RGB and thermal views. As with RGB-Depth training, we randomly sample two frames per scene from both modalities and follow the same three-stage training pipeline with intra- and cross-modality CRW losses. + +RGB-Depth Evaluation. Since RGB and depth frames in NYU are spatially aligned, we use $\mathrm{PIP}++$ [50], a point tracking method, to generate ground-truth correspondences. We create 10-frame video clips and track points across them, retaining those visible in all frames. These tracks are used to evaluate RGB-RGB, Depth-Depth, and RGB-Depth correspondence. Our final dataset includes 250 video clips with an average of 688 annotated tracks per clip. + +RGB-Thermal Evaluation. Thermal-IM videos are not spatially aligned, so we manually annotate 100 RGB-Thermal + +frame pairs across 5 timesteps, with 10 keypoints each, following the protocol from TAP-Vid [4], resulting in 1,000 evaluation points. For KAIST, where RGB and thermal frames are aligned, we use a strategy similar to NYU Depth, using CoTracker [18] to generate correspondences, as it performs better than other trackers on KAIST videos. + +Evaluation Metrics. We adopt the positional accuracy metric from TAP-Vid [4], denoted as $< \delta_{\mathrm{avg}}^x$ . This metric reports the fraction of visible keypoints that are predicted within a pixel distance threshold from ground truth, averaged over thresholds 1, 2, 4, 8, 16. + +# 4.2. Semantic Correspondence + +Photo-Sketch Matching. We use the PSC6K dataset [26] for both training and evaluation. The dataset contains 6,250 annotated photo-sketch pairs, with 8 keypoints per pair in the evaluation set, and over 130K photo-sketch pairs in the training set. Performance is measured using Percentage of Correct Keypoints (PCK) at thresholds $\alpha = 0.05$ and 0.1, denoted as PCK-5 and PCK-10. + +Cross-style Image Matching. We introduce a new benchmark to evaluate correspondence across different visual styles of the same scene. We generate images using the Flux text-to-image model [19], conditioned on a base prompt and a style modifier (added to the prompt). We consider 10 distinct styles: anime, dark, light, photorealistic, pixel art, watercolor, comic book, neon, pastel, and sci-fi (examples in Figure 6). Prompts are generated using BLIP-2[21] by captioning images from the ImageNet dataset [31], and are then used to generate stylistic variants via Flux. + +We generate 10K images per style, resulting in 100K training samples. For evaluation, we sample 50 unseen prompts and generate 10 images per prompt—one in each style— + +yielding 500 test images. We manually annotate corresponding keypoints across styles for each prompt. All possible pairs of styles are compared within each prompt, resulting in 2,250 total image pair comparisons in the test set. + +# 5. Results + +We train our model for RGB-Depth, RGB-Thermal, photosketch and cross-style matching tasks and present both quantitative and qualitative results alongside comparisons with baseline methods. + +# 5.1. Geometric correspondence + +Table 1 compares our method against a range of supervised and unsupervised baselines for RGB-Depth and RGB-Thermal matching, evaluated in the direct setting where query frames are matched directly to target frames without chaining [32]. Among the supervised baselines, we include RAFT [39] and GMFlow [45], state-of-the-art optical flow models trained on RGB images, applied to thermal and depth data by treating them as RGB inputs. For unsupervised baselines, we evaluate Arar et al.[1], a CycleGAN[52]-based cross-modal registration method, which struggles to generalize due to the difficulty of translating between modalities without paired supervision. To address this, we introduce a variant (CycleGAN+GMFlow) that first translates depth or thermal inputs to RGB using CycleGAN, followed by matching in the RGB domain using GMFlow. We also test the reverse direction—translating RGB to depth or thermal—before applying GMFlow. We further evaluate ARFlow [25], an unsupervised optical flow model trained with photometric losses, and include a retrained version adapted to our setting with modality-specific encoders. Finally, we compare against recent diffusion-based correspondence methods, including DIFT [36], which leverages Stable Diffusion features, and SD-DINO [48], which combines Stable Diffusion and DINoV2 features for robust geometric and semantic matching. Across both RGB-Depth and RGB-Thermal tasks, our method significantly outperforms all baselines. + +Qualitative results. We present geometric matching qualitative results showing the effectiveness of our method in cross-modal matching. As shown in Figures 3 and 4, our method consistently produces accurate and meaningful correspondences across RGB-Thermal and RGB-Depth pairs, significantly outperforming RAFT and GMFlow, which often yield noisy or incorrect matches. + +Model Ablations. Table 2 presents ablations on the NYU-Depth and Thermal-IM datasets. Training with only intramodal losses yields good performance within the same modality but fails on cross-modal tasks. In contrast, using only cross-modal losses leads to unstable training and poor convergence. Pretraining with intra-modal losses followed by joint intra- and cross-modal training significantly im + +![](images/69120d4f750d19b9110420740bd29d39e9f72333c17d0adb7b2ba4153b49c909.jpg) +Figure 3. RGB-Depth Matching. We show qualitative comparisons on the NYU Depth V2 dataset. RAFT and GMFlow struggle to establish accurate correspondences in the depth domain, while our method successfully matches keypoints across RGB and depth images. The points shown are randomly sampled from the dataset annotations. + +proves performance across both settings. Adding a smoothness loss further enhances cross-modal accuracy, highlighting its value as a regularization strategy. + +# 5.2. Semantic correspondence + +Photo-Sketch Matching. In Table 3, we compare our method on PSC6K with several baselines from [26], including DINOv2+NN [27], which performs nearest-neighbor search on DINO features, and SD-DINO [48]. When using a CNN encoder, our method performs poorly due to the lack of semantic priors. Replacing it with DINOv2 features and fine-tuning the encoder leads to a significant performance boost, highlighting the importance of DINOv2's semantic pretraining. Our final model, with DINOv2 and full training, performs best among our variants and is competitive with state-of-the-art methods. Qualitative results are shown in Figure 5. + +Cross-style Image Matching. We evaluate our model (using a DINOv2 encoder) alongside baselines including DINOv2+NN, DIFT, and SD-DINO in Table 4. We also compare with GeoAwareSC [49], a state-of-the-art supervised semantic correspondence method. Our model outperforms all baselines, including the supervised approach, albeit by a small margin. The performance gap between our method and DINOv2+NN demonstrates the effectiveness of our learned correspondence framework over simple feature similarity using pretrained DINO features. Qualitative results are presented in Figure 6. + +# 6. Discussion + +We have presented a modality-agnostic method for learning pixel-level correspondences between images taken using different visual sensors. Our method extends the contrastive + +Table 2. Model Ablations. We evaluate the impact of different loss functions and show that pretraining with intra-modal losses followed by fine-tuning with all losses achieves the best performance for our method. X denotes D (Depth) or T (Thermal), based on the dataset. + +
LossesNYU Depth, < δxavg↑Thermal-IM, < δxavg↑
Intra-modalCross-modalCross-modal
LRGB→RGB intra-crwLX→X intra-crwLRGB→X inter-crwLX→RGB inter-crwLSmoothRGB → RGBD → DRGB → DD → RGBRGB → TT → RGB
✓✓---78.849.22.52.24.96.2
--✓✓-18.54.45.64.56.28.3
✓✓✓✓-80.461.119.121.130.238.5
✓✓✓✓✓80.162.333.534.341.847.9
+ +![](images/3547e5c70733ff31f08c5f1dcce998991184c573d82ef4a538aaaff7df3f8fbd.jpg) +Figure 4. RGB-Thermal Matching. Qualitative comparisons on the Thermal-IM (left) and KAIST (right) datasets. Our method accurately tracks keypoints across RGB and thermal images, even in the presence of motion. In contrast, RAFT and GMFlow often fail to produce meaningful correspondences, frequently matching points to the same spatial location in the other modality. Here, we show all 10 annotated points from the dataset. + +Table 3. Photo-Sketch Matching. Qualitative comparison on the PSC6K dataset [26], demonstrating the competitive performance of our method against several baselines. + +
MethodTrained on PSC6KPCK-5PCK-10
CNNGeo [29]27.5957.71
CNNGeo [29]✓19.1942.57
DINOv2 + NN [27]11.4831.66
SD-DINO [48]33.1070.50
WeakAlign [29]35.6568.76
WeakAlign [29]✓43.5578.60
NC-Net [30]40.6063.50
DCCNet [12]42.4366.53
PMD [24]35.7771.24
WarpC-Net [41]48.7971.43
WarpC-Net [41]✓56.7879.70
PSCNet [26]✓57.9284.72
Ours (CNN + Stage 1, 2, 3)✓26.2260.89
Ours (DINO + Stage 2, 3)✓50.6680.70
Ours (DINO + Stage 1, 2, 3)✓53.6182.20
+ +Table 4. Cross-style Image Matching (Sec. 4.2). Quantitative comparisons on the cross-style image matching task. We compare our method against a few self-supervised and one supervised baseline, and show that it outperforms both. + +
MethodSSLPCK-5PCK-10
DINOv2 + NN [27]✓23.1356.52
DIFF [36]✓39.8157.97
SD-DINO [48]✓60.1480.72
GeoAwareSC [49]68.3079.67
Ours (DINO)✓69.2684.19
+ +random walk to perform cross-modal matching. A random walker transitions from patches in one modality to another, and then back, using transition probabilities that are produced by a global matching transformer [1, 32]. We avoid local minima in this process by including intra-modal random walks between frames of the same modality, as well as a spatial smoothness constraint. We evaluate our method on diverse tasks—including RGB-to-depth, RGB-to-thermal, photo-to-sketch, and cross-style image matching—and find + +![](images/b86009e5ea45b2df7adf464036f33edec68f67431489149713337fe8dafa5e01.jpg) +PSC6K Dataset +Figure 5. Photo-Sketch Matching. Qualitative results of our method on the PSC6K dataset [26]. For each image pair, all annotated points from the dataset are used as query keypoints. + +![](images/33524a0fdb6d57630da7f03e523186da2a567eaffd8e7e0d810b4af4399be426.jpg) +prompt: a dog standing next to a wicker +Figure 6. Cross-style Image Matching. We show qualitative results of our method on matching images across different styles. The images are generated using the Flux [19] model, with the prompts and styles used for generation shown above. + +that it significantly outperforms existing methods on geometric correspondence and is competitive with state-of-the-art methods on semantic matching. We see our work as opening new directions in multimodal matching, making it significantly easier to match images without hand-crafted measures of visual similarity. Our approach can be trained entirely using multimodal video data, which can be acquired at scale. We also see our work as experimentally demonstrating the generality of self-supervised matching methods, which can be applied in scenarios that would be difficult to simulate or acquire labeled data for. + +Limitations. We have demonstrated our model for four do + +mains, RGB-depth, RGB-thermal, photo-sketch, cross-style matching. While the model successfully handles these cases, it is possible that other modalities would present additional challenges, especially if they lacked other distinctive visual structures that may be used in matching, such as occluding contours (which are visible in all 3 current modalities, RGB, Thermal, Depth). In cross-style image matching, we observe occasional failures, such as confusion between left and right limbs in animals (see Figure 6, top-right). While our method places few assumptions on the underlying signal (e.g., no hand-crafted photo-consistency assumption), both of our datasets include RGB images due to their ubiquity. + +Acknowledgements. We thank Xuanchen Lu, Adam Harley, Qianqian Wang, Daniel Geng, Ziyang Chen, Jeongsoo Park, Yiming Dou and the reviewers for the valuable discussion and feedback. This work was supported by Toyota Research Institute and Cisco Systems. + +# References + +[1] Moab Arar, Yiftach Ginger, Dov Danon, Amit H Bermano, and Daniel Cohen-Or. Unsupervised multi-modal image registration via geometry preserving image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 1, 2, 3, 5, 6, 7 +[2] Zhangxing Bian, Allan Jabri, Alexei A Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6508-6519, 2022. 2, 4 +[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[4] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. arXiv preprint arXiv:2211.03726, 2022. 2, 3, 5 +[5] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637, 2023. 3 +[6] Yiming Dou, Fengyu Yang, Yi Liu, Antonio Loquercio, and Andrew Owens. Tactile-augmented radiance fields. Computer Vision and Pattern Recognition (CVPR), 2024. 1 +[7] Daniel Gordon, Kiana Ehsani, Dieter Fox, and Ali Farhadi. Watching the world go by: Representation learning from unlabeled videos. arXiv preprint arXiv:2003.07990, 2020. 3 +[8] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3749-3761, 2022. 3 +[9] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 3 +[10] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[11] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[12] Shuaiyi Huang, Qiuyue Wang, Songyang Zhang, Shipeng Yan, and Xuming He. Dynamic context correspondence + +network for semantic alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2010-2019, 2019. 7 +[13] Soonmin Hwang, Jaesik Park, Namil Kim, Yukyung Choi, and In So Kweon. Multispectral pedestrian detection: Benchmark dataset and baselines. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 4 +[14] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 1, 2, 3, 4 +[15] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2 +[16] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European conference on computer vision (ECCV), pages 690-706, 2018. 2, 3 +[17] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 557-572. Springer, 2020. 2, 3, 4 +[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In Proc. ECCV, 2024. 5 +[19] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.1,2,5,8 +[20] Zihang Lai, Erika Lu, and Weidi Xie. Mast: A memory-augmented self-supervised tracker. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6479-6488, 2020. 3 +[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: 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 +[22] Minhao Li, Zheng Qin, Zhirui Gao, Renjiao Yi, Chenyang Zhu, Yulan Guo, and Kai Xu. 2d3d-matr: 2d-3d matching transformer for detection-free registration between images and point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14128-14138, 2023. 1, 3 +[23] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019. 2, 3 +[24] Xin Li, Deng-Ping Fan, Fan Yang, Ao Luo, Hong Cheng, and Zicheng Liu. Probabilistic model distillation for semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7505-7514, 2021. 7 +[25] Liang Liu, Jiangning Zhang, Ruifei He, Yong Liu, Yabiao Wang, Ying Tai, Donghao Luo, Chengjie Wang, Jilin Li, and + +Feiyue Huang. Learning by analogy: Reliable supervision from transformations for unsupervised optical flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2020. 5, 6 +[26] Xuanchen Lu, Xiaolong Wang, and Judith E Fan. Learning dense correspondences between photos and sketches. In International Conference on Machine Learning, pages 22899-22916. PMLR, 2023. 1, 3, 4, 5, 6, 7, 8 +[27] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 4, 6, 7 +[28] 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. PMLR, 2021. 2 +[29] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. End-to-end weakly-supervised semantic alignment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6917-6925, 2018. 7 +[30] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovic, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in neural information processing systems, 31, 2018. 7 +[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5 +[32] Ayush Shrivastava and Andrew Owens. Self-supervised any-point tracking by contrastive random walks. In European Conference on Computer Vision (ECCV), 2024. 2, 3, 4, 5, 6, 7 +[33] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 4, 5 +[34] Austin Stone, Daniel Maurer, Alper Ayvaci, Anelia Angelova, and Rico Jonschkowski. Smurf: Self-teaching multi-frame unsupervised raft with full-image warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3887-3896, 2021. 2, 3 +[35] Deqing Sun, Daniel Vlasic, Charles Herrmann, Varun Jampani, Michael Krainin, Huiwen Chang, Ramin Zabih, William T Freeman, and Ce Liu. Autoflow: Learning a better training set for optical flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10093-10102, 2021. 3 +[36] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5, 6, 7 + +[37] Yansong Tang, Zhenyu Jiang, Zhenda Xie, Yue Cao, Zheng Zhang, Philip HS Torr, and Han Hu. Breaking shortcut: Exploring fully convolutional cycle-consistency for video correspondence learning. arXiv preprint arXiv:2105.05838, 2021. 2, 4 +[38] Zitian Tang, Wenjie Ye, Wei-Chiu Ma, and Hang Zhao. What happened 3 seconds ago? inferring the past with thermal imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17111-17120, 2023. 4 +[39] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3, 5, 6 +[40] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, 2020. 2 +[41] Prune Truong, Martin Danelljan, Fisher Yu, and Luc Van Gool. Warp consistency for unsupervised learning of dense correspondences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10346-10356, 2021. 7 +[42] Carl Vondrick, Abhinav Shrivastava, Alireza Fathi, Sergio Guadarrama, and Kevin Murphy. Tracking emerges by colorizing videos. In Proceedings of the European conference on computer vision (ECCV), pages 391-408, 2018. 3 +[43] Haiping Wang, Yuan Liu, Bing Wang, Yujing Sun, Zhen Dong, Wenping Wang, and Bisheng Yang. Freereg: Imageto-point cloud registration leveraging pretrained diffusion models and monocular depth estimators. arXiv preprint arXiv:2310.03420, 2023. 1, 3 +[44] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2566-2576, 2019. 2 +[45] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8121-8130, 2022. 2, 3, 4, 5, 6 +[46] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10075-10085, 2021. 3 +[47] Jason J. Yu, Adam W. Harley, and Konstantinos G. Derpanis. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In Computer Vision - ECCV 2016 Workshops, Part 3, 2016. 2, 3 +[48] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A Tale of Two Features: Stable Diffusion Complements DINO for Zero-Shot Semantic Correspondence. arXiv preprint arxiv:2305.15347, 2023. 5, 6, 7 +[49] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling + +left from right: Identifying geometry-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3076-3085, 2024. 6, 7 +[50] Yang Zheng, Adam W. Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J. Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In ICCV, 2023. 5 +[51] Tinghui Zhou, Philipp Krahenbuhl, Mathieu Aubry, Qixing Huang, and Alexei A Efros. Learning dense correspondence via 3d-guided cycle consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 117-126, 2016. 2 +[52] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Computer Vision (ICCV), 2017 IEEE International Conference on, 2017. 6 +[53] Yuliang Zou, Zelun Luo, and Jia-Bin Huang. Df-net: Unsupervised joint learning of depth and flow using cross-task consistency. In Proceedings of the European conference on computer vision (ECCV), pages 36-53, 2018. 2, 3 \ No newline at end of file diff --git a/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/images.zip b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..606e0445d8a37ffdae448501e1fc46ea0f40d6c1 --- /dev/null +++ b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e042717ee5eb4685eb2cad9442be489696d06ed46049c3be4d81cf35aa59277a +size 780495 diff --git a/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/layout.json b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9fccf44977439c33331a5cccc6096ff1be3691e6 --- /dev/null +++ b/CVPR/2025/Self-Supervised Spatial Correspondence Across Modalities/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce8ee61be3c92f3674c0f5244e4419ec7c0e9d2b902197d015aa240c69f1b44f +size 351293 diff --git a/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/17058328-37bf-4b7c-b282-396fbd60be04_content_list.json b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/17058328-37bf-4b7c-b282-396fbd60be04_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f8dd89180eb61cb80bce040fd9bfe6450031e019 --- /dev/null +++ b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/17058328-37bf-4b7c-b282-396fbd60be04_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b356b7a05cad2fb278e9940381cd2b436de9019a041e228272cfcd4601da565 +size 93551 diff --git a/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/17058328-37bf-4b7c-b282-396fbd60be04_model.json b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/17058328-37bf-4b7c-b282-396fbd60be04_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3eff3b9e44da121f466e6a00e77ad493e7c851f5 --- /dev/null +++ b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/17058328-37bf-4b7c-b282-396fbd60be04_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b26f7e448e086ad5f9c50926dc27d27334a4583d0b799030b451ddeb11126a0 +size 116407 diff --git a/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/17058328-37bf-4b7c-b282-396fbd60be04_origin.pdf b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/17058328-37bf-4b7c-b282-396fbd60be04_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..38cd45aa33c18f9eaa2f92ce59212a78cf7910ea --- /dev/null +++ b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/17058328-37bf-4b7c-b282-396fbd60be04_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69fbe7fed43c7c50c2486e26ecf09eeff8b54995b9c2ee509d7c0f06d32275d0 +size 3099150 diff --git a/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/full.md b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ef147955dd22df811c554b2adf632e6dc5ab08e6 --- /dev/null +++ b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/full.md @@ -0,0 +1,405 @@ +# Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution + +Shijun $\mathsf{Shi}^{1*}$ Jing Xu $^{2*}$ Lijing Lu $^{3\dagger}$ Zhihang Li $^{4\dagger}$ Kai Hu + +$^{1}$ Jiangnan University $^{2}$ University of Science and Technology of China $^{3}$ Peking University $^{4}$ Chinese Academy of Sciences + +ssj180123@gmail.com, xujing0@mail.ustc.edu.cn, lulijing1997@gmail.com, + +lizhihang.cas@gmail.com, hukai_wlw@jiangnan.edu.cn + +https://ssj9596.github.io/scst-project/ + +# Abstract + +Existing diffusion-based video super-resolution (VSR) methods are susceptible to introducing complex degradations and noticeable artifacts into high-resolution videos due to their inherent randomness. In this paper, we propose a noise-robust real-world VSR framework by incorporating self-supervised learning and Mamba into pre-trained latent diffusion models. To ensure content consistency across adjacent frames, we enhance the diffusion model with a global spatio-temporal attention mechanism using the Video State-Space block with a 3D Selective Scan module, which reinforces coherence at an affordable computational cost. To further reduce artifacts in generated details, we introduce a self-supervised ControlNet that leverages HR features as guidance and employs contrastive learning to extract degradation-insensitive features from LR videos. Finally, a three-stage training strategy based on a mixture of HR-LR videos is proposed to stabilize VSR training. The proposed Self-supervised ControlNet with Spatio-Temporal Continuous Mamba based VSR algorithm achieves superior perceptual quality than state-of-the-arts on real-world VSR benchmark datasets, validating the effectiveness of the proposed model design and training strategies. + +# 1. Introduction + +Video super-resolution (VSR) aims to restore high-resolution (HR) videos by leveraging complementary temporal information within low-resolution (LR) frames, which holds great value in practical usages, e.g., surveillance and high-definition display. Previous works mainly rely on the assumptions of simple and known image degradations (e.g., bicubic downsampling) or specific camera-related degradations, making it challenging to generalize the trained VSR + +![](images/1b62ef7fa6aab701da9cbf293fd1add77e6a41a11918d7672bda9e45e0dfa8e6.jpg) +Input +Frame +Figure 1. A side-by-side comparison of video super-resolution techniques StableSR [58], MGLD [69], and our SCST on two adjacent frames from the VideoLQ dataset. The zoomed-in regions, captured from the same local position, illustrate each method's performance. SCST stands out for maintaining temporal consistency while delivering crisp details in real-world scenarios. + +![](images/4a12f77c48bb8dd406a63dc21f48043b0af3637ae77b58e596541c9a709a648e.jpg) +MGLD + +![](images/dd6cdf58b40a18a5d446a0b00025735d79b82e257b7d7bcde1f722faaa6b61b2.jpg) + +![](images/4a8f7ed788e5ce7d02d43244d8c2848378c3f93a005f2360df967f1785ab9313.jpg) + +![](images/6e92828c8684a64ee73bbda2b08dc790019b613c69d835216fe818fa2b61f96c.jpg) + +![](images/092c1a233806fff1bb5d836211adf34e4b2b68eff6f65ff9188e6b94d75f70ff.jpg) +StableSR + +![](images/2ed5ac0045e320b9d0bc16e77effdb75c3e7a9f6b2ff598c4859df9d284a6aa3.jpg) +Frame 1 + +![](images/be77bb8d3fe707709510f7c8206ca329fabd64f55142349e63d6d43b31f7d80f.jpg) +SCST + +![](images/5e936c5604a1255b070e3a49c4e686e0836ff09f04fbb143e3a6db1984b273f3.jpg) + +![](images/84760ae743b7e185aaf14a81e8133bb5ceb3c3d901b6fcd292869ea9c9981889.jpg) + +![](images/a8a53bb6c1e3e336973e988f06fd44b0e239e67fd6336afd77744310c6a52280.jpg) +Frame T+1 + +models to real-world LR videos with unknown and more complex degradations. + +Recently, diffusion-based generative models [24] have achieved great success in tasks of general image/video generation [25, 51-53] and downstream tasks such as image editing [34], inpainting [45], colorization [6], etc, owing to its powerful ability in capturing diverse and complicated data distributions. There are also efforts to adapt these diffusion priors to image super-resolution [33, 59, 67]. [59, 67] rely on elaborately designed network architectures, such as ControlNet, to condition the diffusion models on low-resolution (LR) images for real-world image superresolution. Wang et al. [58] propose to inject LR images into the U-Net blocks in a LDM network using a SFT module. While showing promising results, directly applying pre-trained diffusion models for image super-resolution to degraded videos is challenging due to the inherent random + +ness in diffusion sampling, leading to temporal inconsistencies in generated videos. To address this problem, Zhou et al. [74] propose to employ 3D convolution and temporal attention in the network to ensure temporal consistency. Yang et al. [69] introduce to use optical flow to align latent features between adjacent frames, enhancing temporal coherence. However, the coupled problem of recovering the unknown and complex degradations in real-world scenarios and producing temporal consistent results in the same time causes great learning complexity. As shown in Figure 1, while MGLD [69] focuses on temporal consistency, it exhibits shortcomings in spatial recovery. Therefore, it is crucial for real-world VSR to extract clean features from complex degradation and hallucinate visual content while also modeling the spatial-temporal dependencies between different frames. + +To tackle these issues, we propose a Self-supervised ControlNet with Spatio-Temporal Continuous Mamba (SCST) for real-world VSR. SCST is a noise-robust temporal coherence diffusion model, aimed at reconstructing fine-grained textures from videos with unknown degradations. To our best knowledge, we are the first to introduce Spatial-Temporal Continuous Mamba (STCM) for global 3D attention in the VSR task. Specifically, we propose a 3D Selective Scan method tailored for traversing spatial-temporal domain, ensuring that every video patch acquires contextual knowledge through a compressed hidden state calculated along the respective scanning path with a linear computational complexity. To mitigate the impact of complex degradations in LR, we propose a ControlNet based on MoCo (MoCoCtrl) [22], which can distill degradation-insensitive features from LR towards the HR target. Finally, we design a multi-stage HR-LR hybrid training strategy to stabilize training. The self-supervised ControlNet is trained using contrastive learning, while the temporal module is optimized with de-noising loss. + +The main contributions of this paper are summarized as follows. + +- We propose a noise-robust temporal coherence diffusion model based on self-supervised learning and spatial-temporal continuous Mamba. The proposed self-supervised ControlNet distills degradation-insensitive features from LR videos while a global 3D attention based on Mamba is designed to model the spatial-temporal relationship of the video. +- To stabilize VSR training, we introduce a decoupled three-stage training strategy where HR and LR videos are mixed for training. Contrastive learning loss is incorporated to align LR features with HR, enabling the extraction of noise-free features. +- Our proposed SCST model achieves state-of-the-art performance on existing benchmarks, showing remarkable visual realism and temporal consistency. + +# 2. Related Work + +# 2.1. Video Super-Resolution + +The goal of VSR is to enhance a sequence of HR video frames from their degraded LR counterparts. Based on the paradigms, existing VSR algorithms [3, 7, 8, 27-29, 31, 42-44, 47, 60, 66, 70] could be roughly classified into two categories: temporal sliding-window based VSR and recurrent framework based VSR. Temporal sliding-window based VSR [31, 40, 60] utilize a fixed set of neighboring frames to super-resolve one or more target frames. However, the information accessible is constrained by the temporal window's size. Consequently, these methods can only exploit the temporal details of a restricted subset of input video frames. To exploit temporal information from more frames, recurrent framework based VSR [7, 8, 42, 43] utilizes multiple LR frames as input and employs recurrent neural networks to simultaneously produce their corresponding SR results. However, most existing approaches [3, 7, 8, 27-29, 31, 42, 43, 60, 66] assume a pre-defined degradation process [44, 47, 70]. In real-world scenes with more complicated degradations, these VSR methods may not perform well. Due to the lack of real-world paired data for training, Yang et al. [68] propose to collect LR-HR data pairs with iPhone cameras to better model realworld degradations. While the VSR model trained on such data can be effective to videos captured by similar mobile cameras, it is relatively labor-intensive and may not generalize well to videos collected by other devices. Recent studies have shifted towards employing diverse degradations for data augmentation during training, such as blur, downsampling, noise and video compression [10, 63]. However, maintaining temporal consistency while generating photorealistic textures remains a challenge. + +# 2.2. State Space Models + +Structured state space models (S4) [19], as a promising framework in handling long-distance sequences, has attracted widespread research interest. A variety of S4-inspired models, that capture long-range dependencies in sequential data, achieve competitive performance on various tasks [21, 26, 48, 54, 57]. The major reason behind this might be that S4's adherence to Linear Time Invariance (LTI), which guarantees consistent output for identical inputs regardless of their temporal application. Nevertheless, LTI systems come with some limitations, especially when it comes to handling dynamic changes over time. The constancy of the internal state transition matrix throughout the sequence constrains the model's adaptability to evolving content, thereby limiting its utility in contexts demanding content-driven reasoning. To address these constraints, Mamba [18] is recently introduced as a state-space model that dynamically adjusts its parameters in response to the in- + +![](images/3429c63789443fc8337ac3da30e4d22f60c191af15e9ebfd75b7b42fe380e8bf.jpg) +Figure 2. Overview of the proposed SCST framework for real-world VSR. SCST consists of several modules, including Spatial-Temporal Continuous Mamba (STCM) and Self-supervised ControlNet (MoCoCtrl). The STCM incorporates 3D-Mamba Block within its structure, which, with the addition of spatial-temporal continuous scan, ensures comprehensive 3D attention for both inter-frame and intra-frame modeling. The Self-supervised ControlNet adopts the MoCo architecture to employ contrastive learning between LR and HR features, aligning LR features to noise-free HR features, thus reducing the impact of degradation. + +put sequence. This adaptive strategy enables Mamba to engage in context-dependent reasoning, significantly enhancing its effectiveness across various domains [39, 41, 75]. However, the application of Mamba in video super resolution tasks remains unexplored. + +# 2.3. Self-Supervised learning + +Self-supervised learning (SSL), as an off-the-shelf representation techniques, has achieved excellent performance in various computer vision tasks [4, 5, 11, 16, 17, 22, 23, 64]. Recently, contrastive learning [1, 4, 5, 14, 15] has emerged as one of the most prominent self-supervised methods, making significant progress in exploring image representations based on instance discrimination tasks, where an instance's different views originating from the same instance are treated as positive examples for an anchor sample, while views from different instances serve as negative examples. The core idea is to promote proximity between positive examples and maximize the separation between negative examples within the latent space, thereby encouraging the model to capture meaningful relationships within the data. + +# 3. Methodology + +# 3.1. Overall Architecture + +Given a LR video sequence of $T$ frames $x^{l} \in R^{T \times H \times W \times 3}$ , the goal is to reconstruct the corresponding HR video sequence $x^{h} \in R^{T \times sH \times sW \times 3}$ , where $s$ is the scaling factor and $H, W$ are the height and width of input frames. We build our method on top of a pretrained SD model [52] to harness the powerful generative priors for real-world VSR. The key design principle of the SD model is to add progressively increasing Gaussian noise to the clean data sample $x$ + +according to a noise schedule $\{\beta_t\}_{t=1}^T$ . Given a noisy sample $x_t$ where $t$ is the diffusion step, a denoising UNet $D$ is trained to estimate the added noise. + +In our method, the training samples $x^{h}$ are drawn from a HR video dataset. We use ControlNet [71] as an Encoder $E$ for LR videos to extract multi-scale latent features, which are injected into the feature maps of the denoising UNet $D$ . In this way, $D$ is conditioned on LR videos to reconstruct the corresponding HR videos. Given the noise target $\epsilon_{t}$ , the objective function for training $D$ and $E$ is as follows: + +$$ +\underset {t, x ^ {h}} {\mathbb {E}} \left| \left| D \left(x _ {t} ^ {h}, t, E \left(x ^ {l}\right)\right) - \epsilon_ {t} \right| \right| ^ {2}, \tag {1} +$$ + +The stochastic nature of the diffusion denoising process leads to temporal instability in extended video sequences for VSR tasks. Two essential problems need to be addressed simultaneously: modeling the spatial-temporal dependencies between adjacent frames, and extracting clean and stable features for accurate reconstruction in the presence of unknown and complex video degradations. As depicted in Figure 2, our framework integrates the Spatial-Temporal Continuous Scan with the STCM into the Latent Diffusion Model (LDM) [52] to ensure spatial-temporal coherence both within and across frame segments. Additionally, we introduce an advanced self-supervised learning approach, the Self-supervised Controlnet (MoCoCtrl), aimed at utilizing HR and LR video pairs to effectively refine the model's ability to generate detailed and noise-robust representations for VSR. + +# 3.2. Mamba for 3D Attention + +We provide a brief overview of SSM and present how to introduce it for real-world VSR. + +# 3.2.1 Preliminaries: State Space Models + +Drawing from the Kalman filter [32], SSMs can be treated as linear time-invariant (LTI) systems that transform the input signal $x(t) \in \mathbb{R}$ to output signal $y(t) \in \mathbb{R}$ via the hidden state $\mathbf{h}(t) \in \mathbb{R}^N$ . In essence, continuous-time SSMs can be represented as linear ordinary differential equations (ODEs), where they encode and decode one-dimensional sequential inputs. + +$$ +\mathbf {h} ^ {\prime} (t) = \mathbf {A h} (t) + \mathbf {B} x (t), \tag {2} +$$ + +$$ +y (t) = \mathbf {C h} (t) + D x (t), +$$ + +where $\mathbf{A} \in \mathbb{R}^{N \times N}$ , $\mathbf{B} \in \mathbb{R}^{N \times 1}$ , $\mathbf{C} \in \mathbb{R}^{1 \times N}$ , and $D \in \mathbb{R}^1$ are the weighting parameters. + +Usually, natural language and two-dimensional vision inputs are discrete signals. Therefore, Mamba utilizes the zero-order hold (ZOH) rule for discretization. Consequently, the ODEs can be iteratively resolved: + +$$ +\bar {\mathbf {A}} = e ^ {\Delta \mathbf {A}}, +$$ + +$$ +\bar {\mathbf {B}} = \left(\boldsymbol {\Delta} \mathbf {A}\right) ^ {- 1} \left(e ^ {\boldsymbol {\Delta} \mathbf {A}} - \mathbf {I}\right) \cdot \boldsymbol {\Delta} \mathbf {B}, \tag {3} +$$ + +$$ +\mathbf {h} (t) = \bar {\mathbf {A}} \mathbf {h} (t - 1) + \bar {\mathbf {B}} x (t), +$$ + +Here, $\Delta$ represents a model parameter. Mamba [18] aims to enhance the adaptability of SSMs by transitioning the time-invariant parameters into time-varying ones. This adjustment entails substituting the fixed model weights $(\mathbf{B},\mathbf{C},\Delta)$ with dynamic weights [30] that are dependent on the input $x$ . This procedure, involving input-dependent parameters, is referred to as selective scanning. + +# 3.2.2 Spatial-Temporal Continuous Mamba + +Joint spatial-temporal modeling of videos plays a pivotal role in VSR. Approaches like Upscale-A-Video [74] incorporate 3D convolution layers and temporal attention, yet their receptive field is constrained. On the other hand, recent video generation methods [2, 37, 50] leverage full 3D attention for motion modeling, demonstrating superior performance against decoupled spatial and temporal attention, however at the expense of a much higher computation complexity. In this study, we introduce Spatial-Temporal Continuous Mamba (STCM), to strike a balance between efficiency and effectiveness. + +3D-Mamba Block. The 3D-Mamba Block, shown in Figure 2, is the core component of the STCM framework, specifically adapted for video-based tasks. It enhances spatiotemporal feature extraction by employing 3D depth-wise convolutions that capture both spatial and temporal dependencies. The block processes the input feature map using K types of scanning operations, generating sequences that are efficiently processed by a State Space Model (SSM) [18] to capture global context with linear complexity. After processing through the SSM, the K sequences are combined, + +![](images/40f55472596de103a780ea373b9d5af6ee439a1ce00900113fe3584235f3eb9e.jpg) +Figure 3. Diagram of Temporal-Spatial Continuous Scan Strategy with Global Consistency. Highlights a single continuous scan pathway across frames, emphasizing spatial-temporal alignment through intra-frame (orange) and inter-frame (cyan) continuity. + +resulting in an output feature map that maintains the original input dimensions. Details of the scanning operations will be discussed later. + +Spatial-Temporal Continuous Scan. As depicted in Figure 2, the 3D-Mamba Block utilizes the Spatial-Temporal Continuous Scan strategy to ensure a smooth, continuous information flow across both spatial and temporal dimensions. The scan involves three primary patterns, each featuring two distinct scanning trajectories: the original pattern and its flipped counterpart. This configuration results in a total of $K = 6$ scanning paths. Figure 3 visually highlights a continuous scan pathway across frames, represented by the orange and cyan areas. The orange areas show intra-frame continuity, where the scan follows a sequential, pixel-by-pixel approach across the horizontal and vertical dimensions, capturing spatial information in a dense, continuous manner. The cyan areas indicate inter-frame continuity, where the scan tracks the same spatial points across successive frames, preserving pixel positions over time. Unlike traditional 3D sweep scan [65], which flattens the input and resets both within and between frames, our method maintains a continuous scan path, ensuring spatial-temporal coherence. This approach is key to capturing consistent features in video data by maintaining pixel continuity both within and across frames, ensuring accurate temporal modeling. Our experiments in Section 4.3 validate its superiority over traditional methods. + +# 3.3. Momentum Contrastive ControlNet + +The proposed 3D-Mamba module empowers the model with the capability to capture global spatiotemporal correlations. However, we find that directly training the model with Eq. 1 conditioning on LR videos leads to unstable training and emergence of artifacts. The presence of unknown and complex degradation in the LR videos causes optimization difficulty. To stabilize the training progress, we propose to provide additional supervising signal for the ControlNet with self-supervised learning leveraging the ground-truth HR videos. + +![](images/0a923252b65a80ddcc1247050fe974fbeffd334816499a5e6af9e6d9188a8113.jpg) +Figure 4. Patch-Level Momentum Contrast. LR and HR images are separately processed by ControlNet to extract feature maps, followed by patch-level contrastive learning on these features. The ControlNet for LR is online updated, whereas the ControlNet for HR updates its weights using a momentum way. + +As shown in Figure 4, we devise a MoCo-like [22] training framework in the optimization of the ControlNet, named MoCoCtrl. We choose the MoCo [22] framework due to its effectiveness and memory-friendly nature. While MoCo is traditionally applied to classification tasks with globally pooled features, our method adapts it for super-resolution tasks by focusing on patch-level features to capture finer spatial details. In our MoCoCtrl framework, two encoders are used: a query encoder $E_{q}$ for LR frames and a momentum encoder $E_{k}$ for HR frames. The momentum encoder's weights are updated as an exponential moving average (EMA) of the query encoder's weights, ensuring stable representations of HR patches. During training, the positive sample pair $(x_{i}^{l}, x_{i}^{h})$ is mapped into feature maps through the query and key encoders as $q = E_{q}(x_{i}^{l})$ and $k_{+} = E_{k}(x_{i}^{h})$ . To capture spatial details, we use a projection head to generate $P \times P$ patch features for each encoded feature map. Negative samples are selected from a memory queue, which is defined as $Q = \{E_{k}(x_{j}^{h}), (E_{k}(x_{j}^{l}) \mid j \in \{0,1,\dots,K/2\}, j \neq i\}$ , where $K$ is the size of the memory queue. Both HR and LR samples are encoded in $Q$ to strengthen the contrastive learning signal and handle more diverse patch-level distinctions. + +For each encoded query $q$ , a patch-level contrastive loss is defined as: + +$$ +\mathcal {L} _ {q} = \frac {1}{P ^ {2}} \sum_ {p} - \log \frac {\exp \left(q ^ {p} \cdot k _ {+} ^ {p} / \tau\right)}{\exp \left(q ^ {p} \cdot k _ {+} ^ {p} / \tau\right) + \sum_ {Q} \exp \left(q ^ {p} \cdot Q / \tau\right)}, \tag {4} +$$ + +where $q^{p}$ and $k_{+}^{p}$ represent the $p^{th}$ patch features of the query and positive HR sample. This patch-level contrastive approach improves the model's ability to capture detailed spatial information, enhancing the performance of super-resolution tasks by precisely aligning LR and HR patches. + +# 3.4. Multi-stage HR-LR hybrid training strategy + +The proposed self-supervised ControlNet enables the model to find degradation-robust features, while the Mamba + +![](images/b2a3ec38502f97467097699e0368e10553e46ff298a4b31fdb56d7f3f5385b2b.jpg) +Figure 5. Multi-stage HR-LR hybrid training strategy. + +blocks model a comprehensive spatial-temporal correlation of the video sequences. We further propose a multi-stage HR-LR hybrid training strategy to facilitate the learning of the two modules. + +We initialize the network using pretrained weights from Stable Diffusion V2.1. The weights of the original 2D U-Net are kept fixed, only the newly introduced layers are trained. As shown in Figure 5, the training consists of three stages. In stage 1, the ControlNet is trained with a mixture of HR/LR videos, where the HR videos can be viewed as LR videos with minimal degradations. When the inputs to the ControlNet are HR videos, the model is essentially trained to reconstruct the inputs. Training with HR videos allows the ControlNet to extract the most accurate features for reconstructing the HR videos. The mixture ratio of HR/LR videos starts at 1 and gradually decreases to 0.3, thereby gradually adapting the ControlNet for real-world VSR. Additionally, a Reconstruction/SR label is introduced to enable the model to distinguish between the reconstruction and super-resolution tasks. In stage 2, the proposed MoCoCtrl is introduced. The training proceeds with a mixture of HR/LR videos, maintaining a fixed ratio at 1:1. The self-supervised learning more fully utilizes the reconstruction prior learned in Stage 1 to facilitate the training of real-world VSR. In stage 3, the proposed Spatial-Temporal Continuous Mamba is integrated into the Unet. During this stage, the ControlNet remains unchanged. Only LR videos are used for training at this stage. + +# 4. Experiments + +# 4.1. Experimental settings + +Training Datasets. We train our model using REDS [47] and YouHQ [74] datasets. Following Wang et al. [60], the REDS4 dataset1 within REDS is excluded from training the model. Additionally, we split YouHQ40 dataset for testing only following Zhou et al. [74], which contains 40 videos. Following the degradation pipeline of RealBasicVSR [9], we generate the LQ-HQ video pairs for training. + +Testing Datasets. We construct the test set with four synthetic testing datasets (i.e., REDS4, UDM10 [70], SPMCS + +Table 1. Quantitative comparisons of state-of-the-art VSR models on different VSR datasets. The best and second performances are highlighted in red and blue, respectively. + +
DatasetsMetricsBicubicRealESRGANStableSRDBVSRRealBasicVSRRealVformerUpscale-A-VideoMGLDSCST
REDS4PSNR ↑23.8122.6922.6422.3823.9424.2822.7322.5623.02
SSIM ↑0.63130.62010.62560.60150.65340.65130.59820.59430.6108
LPIPS ↓0.64850.29640.29920.49410.25450.25360.36390.26600.2518
DISTS ↓0.28580.14260.12770.25100.11960.13060.18400.11710.1094
UDM10PSNR ↑26.4025.6625.5424.8826.1027.1825.6525.8926.42
SSIM ↑0.77270.78170.76220.73430.76580.79480.74130.77130.7893
LPIPS ↓0.50390.27390.25670.46850.28120.25800.27990.25510.2156
DISTS ↓0.26320.15950.13430.24540.16190.15330.15440.13860.1328
SPMCSPSNR ↑23.2122.3822.2122.0123.1023.4321.7222.8722.17
SSIM ↑0.60820.60290.59320.56500.60490.62150.53270.60950.6098
LPIPS ↓0.63600.32380.30790.51600.31420.30300.37430.30410.2600
DISTS ↓0.30920.19240.17090.27250.18470.18840.22010.17690.1612
YouHQ40PSNR ↑24.4723.7623.4123.4123.2624.4423.4123.4824.31
SSIM ↑0.67870.67430.65550.64800.63060.67300.62520.63940.6759
LPIPS ↓0.54370.31260.30710.46990.37060.32020.33490.33090.2525
DISTS ↓0.24160.15370.13600.21370.17450.17400.16180.15400.1344
VideoLQCLIP-IQA ↑0.29490.36170.41600.24750.38810.34600.28180.34620.4859
MUSIQ ↑22.5649.8447.7731.2755.6152.0943.3450.9459.20
NIQE ↓8.0594.2034.4186.2783.6984.0574.87623.7273.566
DOVER ↑0.38820.71520.70290.52640.73670.71940.61990.73400.7443
+ +![](images/2a1debb748e8c0632db4aa0f37779305e03d3e4c7e658b22222e0db62151c576.jpg) + +![](images/002b8cc76e20d813be0853d23085d310aba2f12fb7b73005489789f0a6f2dce2.jpg) + +![](images/2fe8cddf98595967eb2163c0c2185c0ef341ae1d7802f42f92765884af70bea6.jpg) + +![](images/6dfe00e9d36119aeef6dc353002b449db6b1a42bc646c983c4a3f0ce04722c48.jpg) + +![](images/8c3662c308543fe05a102e6d1c01a459be74e028fc42acf680285cae4a5e8c59.jpg) + +![](images/d5fdd3accc652db36297424ae836f8a06887721afc44a02a9fa575acf097881f.jpg) + +![](images/133b91fb9d53bfaa49f8db0acd7d537348f88109290a179e20f5d3d666dd61f4.jpg) + +![](images/d09ee3e89a06378fe8b487d10ba2555314f2453dfc93048f1723859f7fc096bf.jpg) + +![](images/108fcc9f273515ea4ba7a6e07bcd3dec18f6b4389bb44eb5a6107970e668289c.jpg) + +![](images/20d70eadbb3d1083f685b45c4ff13973f2a9fb0624ef3c927b4ba37fb2cebb1a.jpg) + +![](images/f086028b033a8f29f066a44b1e564436ca08b8f27d6f401efc5a53a19c7b1ee8.jpg) + +![](images/ac9ebe9e6de0d77b61fe11f9db803f37bbb88fcbb822c41b9f38adfb58482496.jpg) + +![](images/87727a19168d767b087216614b08b2c8f8deae2a7b3998670d5389e3afe1f87e.jpg) + +![](images/9b6a4d2bf6c465484a2fb8eff47805fbf7b6ae98f173df8e72911062aac80d5a.jpg) + +![](images/d7ed1947895daa10ae53583fef88fdc832ea1a9b178dcd5ca35401e9ff44e874.jpg) +Input +Upscale-A-Video +Figure 6. Qualitative comparisons on synthetic low-quality videos. (Zoom-in for best view) + +![](images/2786586d03142ab11bfda0a6c0c42f74d401ab297169e63af99157c5f2067962.jpg) + +![](images/1cd61d055253b48c01a5f3703a5c1ba9633bcb8fb50edbf995592516f0169c3c.jpg) +StableSR + +![](images/198dfe12c26213f34bf8ebd62cb9e2c1a7159419403eeaf32dfbdff93c6c2110.jpg) +RealViformer + +![](images/0fad010878c530374d70ae03689c2640235e392c5dc635bcf19de991a9643624.jpg) +MGLD + +![](images/b639fdb6df1d0cb9a5d062dad4d216765b7ec74366df59671b788a5c5a4be967.jpg) +Ours + +![](images/c52403a164ebdf7bc2834d6195321a45d6b343c70377610954a95810d5ee20c1.jpg) +GT + +[55], and YouHQ40), which follow the same degradation pipeline in training to generate LQ videos. Additionally, we evaluate the models on a real-world dataset VideoLQ [9]. + +Implementation Details. The model is trained on 8 NVIDIA A100 GPUs with Adam [36] optimizer and a batch size of 32. The sequence length and video resolution are set to 8 and $512 \times 512$ respectively. To better leverage the prior knowledge of text-to-image diffusion models, we use Panda-70M model [12] to generate text prompts during training and inference. Training takes approximately 12 hours for stage 1, 30 hours for stage 2, and 30 hours for stage 3. During inference, owing to memory limitations, we segment LR videos into multiple sequences. The number of sampling steps is set to 20. + +Evaluation Metrics. In order to comprehensively evaluate real-world VSR methods, we utilize a range of metrics across both synthetic and real-world datasets. For + +synthetic datasets, we assess the video quality using four prevalent metrics in real-world VSR tasks: learned perceptual image patch similarity (LPIPS) [72], deep image structure and texture similarity (DISTS) [13], structural similarity index (SSIM), and the widely recognized peak signal-to-noiseratio (PSNR). When assessing the performance on the real-world dataset, we compute the no-reference image quality metrics: the natural image quality evaluator (NIQE) [46] and CLIP-IQA [56]. Additionally, we also include a deep learning based image-based metric MUSIQ [35] and a video quality assessment metric DOVER [62]. + +# 4.2. Experimental Results + +To comprehensively evaluate the performance of our SCST algorithm, we compare it with several state-of-the-art methods, including two real-world image super-resolution models (RealESRGAN [61], StableSR [58]), two real-world + +![](images/2ec800f345ac85023f6e2756d26b86f0653c72174b8991d895b2c72931d072d4.jpg) + +![](images/77360163a0e5bfb39da1cf1b3d0685605957366e3b46f07ada074349ebd15281.jpg) + +![](images/d36868688ffdef0f25729064c1158ebcee3f00a417bd64ea3f699218db71908e.jpg) + +![](images/86bbf72ae8c747cfbb196ea114f1a95321bb8df4b6edbab7fcf3efb595c08294.jpg) + +![](images/14d9749913c8c9e2a1ef4515df917d84d9ea7aa4016fa0312cbf59ab326048b9.jpg) +Figure 7. Qualitative comparisons on real-world test videos in VideoLQ. (Zoom-in for best view) + +![](images/6c3eca2a8a5a20767d1caa847182894fcb9ee898235c2ea73fc3c5bf029d984b.jpg) + +![](images/8b881e2c3d6b301904cacaaf96d844df5cf60d39ac131144e49d0ae272d4040d.jpg) + +![](images/7354010020008396c45dc29ad454f13635f700bb58cd5cf281c4eda1503202ab.jpg) + +![](images/94d5b4581585f30c6a65bf2394d4907b86223f9ae879d51f5e74eb8242073657.jpg) + +![](images/6b8b41013aa550b597b22e0144a9fdf7087765375d75cb1353d3aa1361920345.jpg) + +![](images/e8ae860334a12d340ca8c745f1dd6b565c8d212e135c17c1916b12e20e997b42.jpg) + +![](images/a808e9a9ca1e077f2ac6d1e288df96a38f493e817607acbd50960257a53910da.jpg) + +![](images/111dafb158823a4aa234ec0c8f1fea1248dfe6ba7684f81fd1a367410318f344.jpg) + +![](images/a470a2c26818a2efe111fea9a6c8c3600fecd312f3901668844fe526c9a11040.jpg) + +VSR models (DBVSR [49], RealBasicVSR [10], and recently proposed Upscale-A-Video [74], MGLD [69] and RealVformer [73]. + +Quantitative Comparison. Table 1 demonstrates the quantitative comparison on the synthetic datasets and real-world video benchmarks. From the Table 1, we can observe that our approach attains superior performance in terms of full-reference perceptual metrics LPIPS and DISTS across all synthetic test datasets. This suggests that our method can effectively restore high-quality, realistic details from sequences affected by intricate degradations. While methods such as DBVSR might exhibit improved performance concerning PSNR or SSIM on specific datasets, they often produce blurred outputs, as evidenced by the LPIPS and DISTS metrics. When considering the performance on the real-world VSR dataset VideoLQ, our approach achieves the best results in CLIP-IQA, MUSIQ, NIQE, and DOVER metrics, which indicate the robust capacity of our SCST to enhance real-world videos, producing authentic details and clean textures. + +Qualitative Comparison. To further demonstrate the effectiveness of our SCST, we conduct visual comparisons of these models on both synthetic datasets and real-world VideoLQ dataset, as shown in Figure 6 and Figure 7, respectively. For synthetic datasets, from the Figure 6, we can observe that SCST excels in reconstructing structures while generating cleaner details under complex degradations. The improvements are particularly evident in the enhanced clarity of distant mountain peaks and the textured details of brick walls, as well as the naturalistic rendering of a parrot's + +Table 2. Ablation study on the different comp YouHQ. Best marked in bold. + +
ModelsMoCoCtrlSTCMPSNR ↑SSIM ↑LPIPS ↓DISTS ↓
(a)21.220.63570.28240.1596
(b)✓23.630.65630.26710.1473
(c)✓23.180.65330.25810.1470
(d)✓✓24.310.67580.25250.1344
+ +plumage. For the real-world VSR, it is apparent that SCST surpasses other state-of-the-art algorithms in eliminating intricate spatially varying degradations while producing realistic details. Note that, SCST excels as the sole method capable of accurately delineating the intricate details of the eagle's eyes, showcasing its advanced resolution capabilities. Similarly, SCST provides a significantly clearer depiction of the vehicle's tires, accurately capturing the texture and contours with high fidelity. In contrast, other state-of-the-art methods result in blurred and less defined features. + +# 4.3. Ablation Study + +Baseline Design. To assess the impact of the two key components of our network, MoCoCtrl and STCM, we conducted a series of ablation experiments. We start by creating a baseline model removing all the major components of the network. Specifically, the baseline model excludes the second-stage contrastive training and directly trains the model using Eq. 1. + +Effectiveness of MoCoCtrl. As shown in Figure 8 (c) and (d), as well as Table 2, the MoCoCtrl module results in clearer super-resolution outputs and better performance metrics. We can observe that Model (d) outperforms + +![](images/bcdb0faf7e14e5f639bf5cd19e542926b48f907cb1435279f7cfa5cb65e85266.jpg) +Figure 8. Comparison of two adjacent frames in one video SR results synthesized by different components. + +Table 3. Comparison of different spatial-temporal modeling approaches on YouHQ and REDS4. "Mamba-S" employs a 3D Sweep Scan strategy. Best marked in blod. + +
ModelsPSNR↑ / LPIPS↓Warp Error↓
YouHQREDS4YouHQREDS4
w/o Temporal23.18 / 0.258122.73 / 0.26690.36033.619
Local Attention24.07 / 0.257022.60 / 0.26290.23112.889
Mamba-S24.12 / 0.253322.59 / 0.25900.24942.926
STCM24.31 / 0.252523.02 / 0.25180.22952.878
+ +Model (b) in terms of PSNR, SSIM, and LPIPS. Specifically, Model (d) shows better performance in PSNR (24.31 dB vs. 23.63 dB), higher SSIM (0.6758 vs. 0.6563), lower LPIPS (0.2525 vs. 0.2671) and DISTS (0.1344 vs. 0.1473). This comparison clearly demonstrates the improvement in video super-resolution quality in Model (d) due to the Mo-CoCtrl module, thus confirming its effectiveness. + +Effectiveness of STCM. In addition to the proposed Mo-CoCtrl, STCM further enhances the quality of our generated videos. Specifically, STCM extends its global receptive field to capture complex spatio-temporal relationships and employs continuous scanning to identify long-range dependencies within video sequences. STCM is designed to facilitate the model in better understanding video content, thereby enhancing temporal consistency while maintaining high-resolution output with increased fidelity. As illustrated in Figure 8, the temporal consistency is markedly inferior in the absence of the STCM module. Moreover, the integration of the STCM module into Model (c) is shown to yield substantial enhancements in performance metrics, with noteworthy increases of $1.13\mathrm{dB}$ in PSNR, 0.0225 in SSIM, 0.0056 in LPIPS, and 0.0126 in DISTS, as demonstrated in Table 2. + +Analysis on the Mamba design. To further elucidate the advantages of our STCM, we also conducted an analysis comparing various spatial-temporal modeling approaches in terms of restoration quality and temporal consistency. We replace the STCM component with local inter-frame attention (Local Attention) [20] and 3D Sweep Scan Mamba (Mamba-S) to compare different spatial-temporal modeling approaches. Table 3 shows STCM achieves the best performance among all methods in PSNR and LPIPS. Compared to Local Attention, which is limited by its spatial receptive field, STCM leverages multi-direction fusion to fully capture the 3D data pattern. As shown in Figure 9, Local At + +![](images/771cabfbceffc80d7f51c41fb3e7ce5f6076b3418df4dc1a0cf7c9ab80467081.jpg) +Figure 9. Different Spatial-temporal modeling approaches + +![](images/28bda01fc048d2e9f25cb2c4a2a83e08f5210ac3d7d668212c49801bb31e5b79.jpg) +Figure 10. Visual comparison on temporal profile with different spatial-temporal modeling approaches, with STCM exhibiting the best temporal consistency. (Zoom-in for best view) + +tention produces visible distortions and blurred edges, especially in complex regions, resulting in misaligned geometry. In contrast, STCM incorporates surrounding pixel information to reconstruct more rectilinear and well-aligned structures, significantly reducing artifacts in challenging areas. Beyond restoration quality, we use the Warping Error (WE) [38] to measure temporal consistency. Table 3 clearly shows that STCM outperforms all other methods on both datasets, achieving the lowest WE scores of 0.2295 and 2.878 on YouHQ and REDS4, respectively. It is worth mentioning that the excellent performance of STCM is driven by its spatial-temporal continuous scanning strategy. Unlike Mamba-S, which employs a 3D Sweep Scan that disrupts continuity by resetting between frames, our approach maintains an uninterrupted flow of information. This continuous scanning ensures spatial-temporal coherence and precise frame alignment. Figure 10 illustrates that STCM preserves steady structures across frames, while Mamba-S exhibits temporal fluctuations and misalignments. + +# 5. Conclusion + +In this paper, we propose a Self-supervised ControlNet with Spatio-Temporal Mamba algorithm dubbed SCST, designed for high-quality and temporally consistent real-world Video Super-Resolution (VSR). The distinctiveness of our proposed method lies in the idea of introducing a specific Self-supervised ControlNet as a degradation removal module, reducing the impact of complex degradations on the effectiveness of VSR. To further model spatio-temporal relationships for temporal consistency, we present an efficient 3D attention-based variant of the successful Mamba model. Finally, to ensure training stability, we propose a multi-stage HR-LR hybrid training strategy, decomposing real-world VSR into multiple subtasks, with each stage addressing a specific task. Our proposed SCST has achieved state-of-the-art results on existing VSR benchmarks. + +# References + +[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. 3 +[2] Fan Bao, Chengong Xiang, Gang Yue, Guande He, Hongzhou Zhu, Kaiwen Zheng, Min Zhao, Shilong Liu, Yaole Wang, and Jun Zhu. Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models. arXiv preprint arXiv:2405.04233, 2024. 4 +[3] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2 +[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 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 IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3 +[6] Hernan Carrillo, Michael Clément, Aurélie Bugeau, and Edgar Simo-Serra. Diffusart: Enhancing line art colorization with conditional diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3486-3490, 2023. 1 +[7] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 2 +[8] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 2 +[9] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5962-5971, 2022. 5, 6 +[10] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5962-5971, 2022. 2, 7 +[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3 +[12] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneca, Hsiang-wei Chao, Byung Eun Jeon, + +Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 6 + +[13] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 6 +[14] Chen Feng and Ioannis Patras. Adaptive soft contrastive learning. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 2721-2727. IEEE, 2022. 3 +[15] Chen Feng and Ioannis Patras. Maskcon: Masked contrastive learning for coarse-labelled dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19913-19922, 2023. 3 +[16] Chen Feng, Georgios Tzimiropoulos, and Ioannis Patras. Ssr: An efficient and robust framework for learning with unknown label noise. arXiv preprint arXiv:2111.11288, 2021. 3 +[17] Zheng Gao, Chen Feng, and Ioannis Patras. Self-supervised representation learning with cross-context learning between global and hypercolumn features. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1773-1783, 2024. 3 +[18] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 4 +[19] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2 +[20] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 8 +[21] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 2 +[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 3, 5 +[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3 +[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1 + +[26] Md Mohaiminul Islam and Gedas Bertasius. Long movie clip classification with state-space video models. In European Conference on Computer Vision, pages 87-104. Springer, 2022. 2 +[27] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 645-660. Springer, 2020. 2 +[28] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8008-8017, 2020. +[29] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. arXiv preprint arXiv:2008.05765, 2020. 2 +[30] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc Van Gool. Dynamic filter networks. In NIPS, pages 667-675, 2016. 4 +[31] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3224-3232, 2018. 2 +[32] Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. 1960. 4 +[33] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 1 +[34] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1 +[35] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6 +[36] D Kinga, Jimmy Ba Adam, et al. A method for stochastic optimization. In International conference on learning representations (ICLR), page 6. San Diego, California.;, 2015. 6 +[37] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 4 +[38] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, pages 170-185, 2018. 8 +[39] Shufan Li, Harkanwar Singh, and Aditya Grover. Mamba-nd: Selective state space modeling for multi-dimensional data. arXiv preprint arXiv:2402.05892, 2024. 3 +[40] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. Mucan: Multi-correspondence aggregation network for video super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X 16, pages 335–351. Springer, 2020. 2 + +[41] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaoqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 3 +[42] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 2 +[43] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 2024. 2 +[44] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. IEEE transactions on pattern analysis and machine intelligence, 36(2):346-360, 2013. 2 +[45] 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. 1 +[46] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6 +[47] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 2, 5 +[48] 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. 2 +[49] Jinshan Pan, Haoran Bai, Jiangxin Dong, Jiawei Zhang, and Jinhui Tang. Deep blind video super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4811-4820, 2021. 7 +[50] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 4 +[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[52] 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. 3 +[53] 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. 1 +[54] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 2 +[55] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In Proceedings of the IEEE international conference on computer vision, pages 4472-4480, 2017. 6 +[56] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 6 +[57] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6387-6397, 2023. 2 +[58] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1–21, 2024. 1, 6 +[59] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1–21, 2024. 1 +[60] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 2, 5 +[61] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 6 +[62] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20144-20154, 2023. 6 +[63] Liangbin Xie, Xintao Wang, Shuwei Shi, Jinjin Gu, Chao Dong, and Ying Shan. Mitigating artifacts in real-world video super-resolution models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2956-2964, 2023. 2 +[64] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 3 +[65] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 4 + +[66] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127:1106-1125, 2019. 2 +[67] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 1 +[68] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4781-4790, 2021. 2 +[69] Xi Yang, Chenhang He, Jianqi Ma, and Lei Zhang. Motion-guided latent diffusion for temporally consistent real-world video super-resolution. arXiv preprint arXiv:2312.00853, 2023. 1, 2, 7 +[70] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3106-3115, 2019. 2, 5 +[71] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. 3 +[72] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[73] Yuehan Zhang and Angela Yao. Realvformer: Investigating attention for real-world video super-resolution. arXiv preprint arXiv:2407.13987, 2024. 7 +[74] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-consistent diffusion model for real-world video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 2, 4, 5, 7 +[75] 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/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/images.zip b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a068e9801e8a04e6a6f304ac18341da5be947d0f --- /dev/null +++ b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01d62fa7bf19f5593b4f8389556e412029c61cfac0bf971ecd77387dd42601b8 +size 889187 diff --git a/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/layout.json b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..672ca9e31fd13f6658639f5b77aeecfc3f30c97f --- /dev/null +++ b/CVPR/2025/Self-supervised ControlNet with Spatio-Temporal Mamba for Real-world Video Super-resolution/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14a49350754cb1f3e528e4af59ab4b85876f5b4b8f4e14b1ab84666e6ba45943 +size 464198 diff --git a/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/bbd53a21-f039-4031-a580-5b3f999b5a9d_content_list.json b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/bbd53a21-f039-4031-a580-5b3f999b5a9d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1e3d00473b9a20c8bf3693bff1e4dfca246c5a17 --- /dev/null +++ b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/bbd53a21-f039-4031-a580-5b3f999b5a9d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d89b9f33766a84bed22e74770db2305bafea956eac4849c17eed76a829e5be6f +size 89779 diff --git a/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/bbd53a21-f039-4031-a580-5b3f999b5a9d_model.json b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/bbd53a21-f039-4031-a580-5b3f999b5a9d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5d022886ec3e1821e7f06089d97ee1c5414b1296 --- /dev/null +++ b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/bbd53a21-f039-4031-a580-5b3f999b5a9d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ece60ada972a0d6687c048a8ce0bf98df1a1ddb106556a77c60e2a5663de9668 +size 111516 diff --git a/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/bbd53a21-f039-4031-a580-5b3f999b5a9d_origin.pdf b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/bbd53a21-f039-4031-a580-5b3f999b5a9d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..edef720e0809079be54a007cb719be6b3178d087 --- /dev/null +++ b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/bbd53a21-f039-4031-a580-5b3f999b5a9d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d28b6fe7652957b12aa8b2ca91d2d804d857633b6ff4d8d4fb9a4ae29c71833f +size 3313082 diff --git a/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/full.md b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..17c98b6f43df943daf978d1b543f17d1ab818c2a --- /dev/null +++ b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/full.md @@ -0,0 +1,336 @@ +# SelfSplat: Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting + +Gyeongjin Kang $^{1*}$ Jisang Yoo $^{2*}$ Jihyeon Park $^{1}$ Seungtae Nam $^{3}$ Hyeonsoo Im $^{4}$ Sangheon Shin $^{4}$ Sangpil Kim $^{5}$ Eunbyung Park $^{3\dagger}$ + +Department of Electrical and Computer Engineering, Sungkyunkwan University1 +Department of Intelligent Robotics, Sungkyunkwan University2 +Department of Artificial Intelligence, Yonsei University3 +Hanhwa Systems4 Korea University5 + +# Abstract + +We propose SelfSplat, a novel 3D Gaussian Splitting model designed to perform pose-free and 3D prior-free generalizable 3D reconstruction from unposed multi-view images. These settings are inherently ill-posed due to the lack of ground-truth data, learned geometric information, and the need to achieve accurate 3D reconstruction without finetuning, making it difficult for conventional methods to achieve high-quality results. Our model addresses these challenges by effectively integrating explicit 3D representations with self-supervised depth and pose estimation techniques, resulting in reciprocal improvements in both pose accuracy and 3D reconstruction quality. Furthermore, we incorporate a matching-aware pose estimation network and a depth refinement module to enhance geometry consistency across views, ensuring more accurate and stable 3D reconstructions. To present the performance of our method, we evaluated it on large-scale real-world datasets, including RealEstate10K, ACID, and DL3DV. SelfSplat achieves superior results over previous state-of-the-art methods in both appearance and geometry quality, also demonstrates strong cross-dataset generalization capabilities. Extensive ablation studies and analysis also validate the effectiveness of our proposed methods. Code and pretrained models are available at https://gynjn.github.io/selfsplat/. + +# 1. Introduction + +The recent introduction of Neural Radiance Fields (NeRF) [36] and 3D Gaussian Splatting (3D-GS) [25] had marked a significant advancement in computer vision and graphics, particularly in 3D reconstruction and novel view synthesis. By training on images taken from various view- + +points, these methods can produce geometrically consistent photo-realistic images, providing beneficial for various applications, such as virtual reality [23, 55], robotics [15, 38], and semantic understanding [64, 65]. Despite their impressive capability in 3D scene representation, training NeRF and 3D-GS requires a large set of accurately posed images as well as iterative per-scene optimization procedures, which limits their applicability for broader use cases. + +To bypass the iterative optimization steps, various learning-based generalizable 3D reconstruction models [3, 13, 43, 53, 61] have been proposed. These models can predict 3D geometry and appearance from a few posed images in a single forward pass. Leveraging large-scale synthetic and real-world 3D datasets, they used pixel-aligned features to extract scene priors from input images and generate novel views through differentiable rendering methods such as volume rendering [35] or rasterization [27]. The generated images are then supervised with ground truth images captured from the same camera poses. While this approach enables 3D scene reconstruction without iterative optimization steps, a key limitation remains as follows: it relies on calibrated images (with accurate camera poses) for both training and inference, thereby constraining its use with less controlled, "in-the-wild" images or videos. + +Recent efforts have integrated camera pose estimation with 3D scene reconstruction, combining multiple tasks within a single framework. By relaxing the constraint of a posed multi-view setup, pose-free generalizable methods [6, 21, 29, 41] aim to learn reliable 3D geometry from uncalibrated images and generate accurate 3D representations in a single forward pass. While these approaches have demonstrated promising results, they still face significant challenges. For example, [41] relies on error-prone pretrained flow model for pose estimation, often leading to inaccuracies and performance degradation. [6, 29] achieve impressive results but require a per-scene fine-tuning stage, making them computationally expensive for real-world ap + +plications. Furthermore, both [41] and [6] inherit the limitation of NeRF-based approaches, demanding substantial computational costs due to the volumetric rendering. + +In this work, we present SelfSplat, a novel training framework for pose-free, generalizable 3D representations from monocular videos without pretrained 3D prior models or further scene-specific optimizations. We build upon the 3D-GS representation and leverage the pixel-aligned Gaussian estimation pipeline [3, 43], which has demonstrated fast and high-quality reconstruction results. By integrating 3D-GS representations with self-supervised depth and pose estimation techniques, the proposed method jointly predicts depth, camera poses, and 3D Gaussian attributes within a unified neural network architecture. + +3D-GS, as an explicit 3D representation, is highly sensitive to minor errors in 3D positioning. Even slight misplacements of Gaussians can disrupt multi-view consistency, significantly degrading rendering quality [3, 7]. This makes the simultaneous prediction of Gaussian attributes and camera poses especially challenging. The proposed approach, SelfSplat, mitigates this issue by leveraging the strengths of both self-supervised learning and 3D-GS. Exploiting the geometric consistency inherent in self-supervised learning techniques effectively guides the positioning of 3D Gaussians, leading to improved reconstruction accuracy in the absence of camera pose information. Also, harnessing 3D-GS representation and its superior view synthesis capabilities help enhance the accuracy of camera pose estimation, which would otherwise depend solely on 2D image features derived from CNNs [17, 42] or Transformers [9, 39]. + +While the proposed method is encouraging, simply combining self-supervised learning with explicit 3D geometric supervision has yielded suboptimal results, particularly in predicting accurate camera poses and generating multi-view consistent depth maps. This often results in misaligned 3D Gaussians and inferior 3D structure reconstructions. To address issues from pose estimation errors, we introduce a matching-aware pose network that incorporates additional cross-view knowledge to improve geometric accuracy. By leveraging contextual information from multiple views, this network improves pose accuracy and ensures more reliable estimates across views. Additionally, to support consistent depth estimation, crucial for accurate 3D scene geometry, we develop a depth refinement network. This module uses estimated poses as embedding features which contains spatial information from surrounding views, to achieve accurate and consistent 3D geometry representations. + +Once trained in a self-supervised manner, SelfSplat is equipped to perform several downstream tasks, including (1) pose, depth estimation, and (2) 3D reconstruction, including fast novel view synthesis. We demonstrate the efficacy of our method on RealEstate10k [67], ACID [33], and DL3DV [32] datasets providing higher appearance and ge + +ometry quality as well as better cross-dataset generalization performance. Extensive ablation studies and analyses also show the effectiveness of our proposed method. The main contributions can be summarized as follows: + +- We propose SelfSplat, a pose-free and 3D prior-free self-supervised learner from large-scale monocular videos. +- We propose to unify self-supervised learning with 3D-GS representation, harnessing the synergy of both frameworks to achieve robust 3D geometry estimation. +- To address pose estimation errors and inconsistent depth predictions, we introduce the matching-aware pose network and depth refinement module, which enhance geometry consistency across views, ensuring more accurate and stable 3D reconstructions. +- We have conducted comprehensive experiments and ablation studies on diverse datasets, and the proposed Self-Splat significantly outperforms the previous methods. + +# 2. Related work + +# 2.1. Pose-free Neural 3D Representations + +In the absence of camera pose information, recent efforts have aimed to jointly optimize camera poses and 3D scenes. Starting with optimization-based methods, BARF [30] and subsequent research [2, 16, 24] addressed this challenge by training poses along with implicit or explicit scene representations. Also, in a generalizable setting with NeRF representations, FlowCam [41] used the pretrained flow estimation model, RAFT [44], and found the rigid-body motion between 3D point clouds using the Procrustes algorithm [11]. DBARF [6] extended the previous optimization-based method [30] and utilized recurrent GRU [10] network for pose and depth estimation. Based on 3D-GS, several methods [22, 29, 58] have demonstrated pose-free generalizable approaches with explicit 3D representations. However, these methods face practical limitations due to their reliance on pretrained models [6, 29, 41, 58], the need for additional fine-tuning stages [6, 29], and the computationally intensive volume rendering process [6, 41], all of which hinder their scalability and efficiency in real-world applications. Also, CoPoNeRF [21] provides poses and radiance fields estimation at the inference stage, it still requires ground-truth pose supervision during training. In contrast, our method can reconstruct 3D scenes and synthesize novel views from unposed images, mitigating the preceding challenges, and offering a more scalable and efficient solution. + +# 2.2. Self-supervised Learning for 3D Vision + +Masked Autoencoder (MAE) [19, 46] is one of self-supervised representation learning framework on video datasets, leveraging their consistency in space and over time. The main objective of MAE is to reconstruct masked patch of pixels or latent features, thereby learn + +![](images/ffe475ab205900de3372014a30ef5037ed2c1d5c107ffde6a127477fdb878037.jpg) +Figure 1. Overview of SelfSplat. Given unposed multi-view images as input, we predict depth and Gaussian attributes from the images, as well as the relative camera poses between them. We unify a self-supervised depth estimation framework with explicit 3D representation achieving accurate scene reconstruction. + +ing spatiotemporal continuity without any 3D inductive bias. Recently, CroCo [51, 52], a cross-view completion method which extends previous single-view approaches, has demonstrated a pretraining objective well-suited for geometric downstream tasks, such as optical flow and stereo matching. Expanding on this, DUSt3R [49] and MAST3R [28] introduce a novel paradigm for dense 3D reconstruction from multi-view image collections. + +Another area of self-supervised learning for 3D vision is monocular depth estimation. In the absence of ground-truth depth and camera pose, previous methods have leveraged information from consecutive frames, using warped image reconstruction as a training signal. Starting with [66], which first introduced the method, and subsequent works [1, 8, 17] have developed upon this field. In this paper, we follow the framework of self-supervised depth estimation, but different from previous methods, we integrate 3D representation learning, which improves depth estimation and enables novel view synthesis with resulting scene representations. + +# 3. Preliminary + +# 3.1. Self-supervised Depth and Pose Estimation + +The self-supervised depth and pose estimation method is a geometric representation learning method from videos or unposed images, which does not require ground-truth depth and pose annotations [60, 66]. Typically, two separate networks are employed for each depth and pose estimation, though these networks may share common representations. Given a triplet of consecutive frames $I_{c_1}, I_t, I_{c_2} \in \mathbb{R}^{H \times W \times 3}$ , the pose network predicts the relative camera pose between two frames and the depth network produces the depth maps for each frame. While there exist many variants, a typical loss function, $\mathcal{L}_{\mathrm{proj}}$ , to train two networks is + +$$ +\mathcal {L} _ {\mathrm {p r o j}} = \mathrm {p e} \left(I _ {t}, I _ {c _ {1} \rightarrow t}\right) + \mathrm {p e} \left(I _ {t}, I _ {c _ {2} \rightarrow t}\right), \tag {1} +$$ + +$$ +\mathrm {p e} \left(I _ {a}, I _ {b}\right) = \frac {\omega}{2} \left(1 - \operatorname {S S I M} \left(I _ {a}, I _ {b}\right)\right) + (1 - \omega) \| I _ {a} - I _ {b} \| _ {1}, \tag {2} +$$ + +where $I_{c_1 \to t} \in \mathbb{R}^{H \times W \times 3}$ denotes the projected image from $I_{c_1}$ onto $I_t$ using the predicted pose and depth. $\mathrm{pe}(\cdot, \cdot)$ is a photometric reconstruction error, calculated using a combination of L1 and SSIM [50] losses, and $\omega$ is a hyperparameter that controls the weighting factor between them [17]. + +# 3.2. Feed-forward 3D Gaussian Splatting + +Feed-forward 3D Gaussian Splating methods infer 3D scene structure from input images through a single network evaluation, predicting Gaussian attributes based on pixel- or voxel-level tensors. Each Gaussian, defined as $g_{j} = (\mu_{j},\alpha_{j},\Sigma_{j},c_{j})$ , includes attributes such as a mean $\mu_{j}$ , a covariance $\Sigma_{j}$ , an opacity $\alpha_{j}$ , and spherical harmonics coefficients $c_{j}$ . In particular, our framework adopts a pixel-aligned approach, predicting per-pixel Gaussian primitives along with accurate depth estimations, achieving high-quality 3D reconstruction and fast novel view synthesis. Given multiple input views, the model generates pixel-aligned Gaussians for each image, and combine them to represent the full 3D scene [3, 43]. + +# 4. Methods + +# 4.1. Self-supervised Novel View Synthesis + +We begin with a triplet of unposed images, $I_{c_1}, I_t, I_{c_2} \in \mathbb{R}^{H \times W \times 3}$ , which are taken from different viewpoints. Building on the recent pixel-aligned 3D Gaussian Splitting methods, our goal is to predict dense per-pixel Gaussian parameters from input view images, + +$$ +\mathcal {G} _ {c _ {1}}, \mathcal {G} _ {c _ {2}} = f _ {\theta} \left(I _ {c _ {1}}, I _ {t}, I _ {c _ {2}}\right), \tag {3} +$$ + +where $f_{\theta}$ is a feed-forward network with learnable parameters $\theta$ , and $\mathcal{G}_{c_1} = \{(\mu_j,\alpha_j,\Sigma_j,c_j)\}_{j = 1}^{HW}$ is a generated Gaussians for the input image $I_{c_1}$ . Note that we only generate pixel-aligned Gaussians for two input views $I_{c_1}$ and $I_{c_2}$ while excluding the target view $I_{t}$ . This design encourages the network to generalize to novel views $I_{t}$ during training. In addition, we train a pose network $f_{\phi}$ + +![](images/f4fa5073b7e78bfa071ad9df98e349777e596fa78ce5c903df175ea73c91d2c9.jpg) +Figure 2. Matching-aware pose network (a) and depth refinement module (b). We leverage cross-view features from input images to achieve accurate camera pose estimation, and use these estimated poses to further refine the depth maps with spatial awareness. + +![](images/c36afda66fccc2120d9335b9fc27b6fcfb53070812d6322b7bf107a14708cfac.jpg) + +to estimate a relative transformation between two images, $T_{c_1 \rightarrow c_2} = f_\phi(I_{c_1}, I_{c_2})$ , where $T_{c_1 \rightarrow c_2} \in SE(3)$ consists of rotation, $R_{c_1 \rightarrow c_2} \in \mathbb{R}^{3 \times 3}$ , and translation, $t_{c_1 \rightarrow c_2} \in \mathbb{R}^{3 \times 1}$ , between two images, $I_{c_1}$ and $I_{c_2}$ . We utilize the estimated camera poses to transform the Gaussian positions in each frame's local coordinate system into an integrated global space. Then, we construct the 3D Gaussian representations for a scene by union of the generated Gaussians as follows, + +$$ +\mathcal {G} = \operatorname {T R} \left(\mathcal {G} _ {c _ {1}}, T _ {c _ {1} \rightarrow t}\right) \cup \operatorname {T R} \left(\mathcal {G} _ {c _ {2}}, T _ {c _ {2} \rightarrow t}\right), \tag {4} +$$ + +where $\mathbb{R}(\mathcal{G}_{c_1},T_{c_1\to t})$ transforms the generated Gaussian $\mathcal{G}_{c_1}$ into the $I_{t}$ 's coordinate system, and $\mathcal{G}$ is the final 3D Gaussians that are used to render images. The final loss function to jointly train both $f_{\theta}$ and $f_{\phi}$ is defined as follows, + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {1} \mathcal {L} _ {\text {p r o j}} + \lambda_ {2} \mathcal {L} _ {\text {r e n}}, \tag {5} +$$ + +$$ +\mathcal {L} _ {\text {r e n}} = \sum_ {I _ {k} \in \left\{I _ {c _ {1}}, I _ {c _ {2}}, I _ {t} \right\}} \gamma_ {1} \left(1 - \operatorname {S S I M} \left(I _ {k}, \hat {I} _ {k}\right)\right) + \gamma_ {2} \| I _ {k} - \hat {I} _ {k} \| _ {2}, \tag {6} +$$ + +where $\mathcal{L}_{proj}$ is the reprojection loss (Eq. (2)) and $\mathcal{L}_{ren}$ is the rendering loss that computes the error between input view images, $I_{k}$ , and the rendered images, $\hat{I}_k$ , from the constructed Gaussians $\mathcal{G}$ . Note that in $\mathcal{L}_{proj}$ , we use the rendered depth for $I_{t}$ to maintain a consistent scale with estimated depth maps from the context images, $I_{c_1}$ and $I_{c_2}$ . In accordance with the prior pose-free generalizable methods, we assume that the camera intrinsic parameters are given from the camera sensor metadata [6, 21, 29, 47]. + +# 4.2. Architecture + +As illustrated in Fig. 1, the proposed SelfSplat consists of four components: a multi-view and monocular encoder, a fusion and dense prediction block, a matching-aware pose estimation network, and a Gaussian decoder. + +Multi-view and monocular encoder. For multi-view feature extraction from input view images, we begin by processing each image independently through a weight-sharing CNN architecture, followed by a multi-view Transformer to exchange information across different views. Specifically, + +a ResNet-like architecture [18] is used to extract 4x downsampled features for each view. These features are then refined by a six-block Swin Transformer [34], which utilizes efficient local window self- and cross-attention mechanisms. The resulting cross-view-aware features are denoted as $F_{c_1}^{\mathrm{mv}}, F_{c_2}^{\mathrm{mv}} \in \mathbb{R}^{\frac{H}{4} \times \frac{W}{4} \times C^{\mathrm{mv}}}$ , where $C^{\mathrm{mv}}$ is the dimension. These features are subsequently processed to generate Gaussian attributes for rendering. As discussed in Sec. 4.1, since we do not generate Gaussian attributes for $I_t$ , $I_t$ is excluded from the feature extraction in this module. + +Despite substantial advancements in multi-view feature matching-based depth estimation methods, such as those leveraging epipolar sampling [3, 20] or plane-sweep techniques [4, 7, 57], these approaches continue to face challenges in handling occlusions, texture-less regions, and reflective surfaces. To address these limitations, we incorporate a monocular feature extractor, which has demonstrated robust performance across various downstream tasks [12, 45]. Specifically, we utilize a shared-weight Vision Transformer (ViT) model, CroCo [51, 52], as a monocular feature extractor. More specifically, input images are divided into non-overlapping patches with a patch size of 16 and processed by multi-head self-attention blocks and feed-forward networks in parallel. Then, we obtain robust monocular Transformer features $F_{c_1}^{\mathrm{mono}}$ , $F_{c_2}^{\mathrm{mono}} \in \mathbb{R}_{\frac{H}{16} \times \frac{W}{16} \times C^{\mathrm{mono}}}^{\frac{H}{16} \times \frac{W}{16} \times C^{\mathrm{mono}}}$ , where $C^{\mathrm{mono}}$ denotes the channel dimension. Similar to the multiview feature extraction, we do not extract the monocular feature from $I_t$ . It is important to note that, unlike previous methods [54, 62] which use a pretrained DepthAnything [56] model as a ViT backbone and thus incorporate 3D priors, we employ CroCov2 [52] weights, allowing us to maintain a fully self-supervised framework + +Feature fusion and dense prediction. To achieve consistent and fine-grained prediction of Gaussian primitives, we combine the multi- and single-view features, leveraging complementary information from both perspectives to enhance depth accuracy and robustness in complex scenes. We build our feature fusion block with Dense Prediction Transformer (DPT) [37] module. As the spatial resolutions between the two features are different, we first downsam + +ple the multi-view features by four to match with monocular ones. Then, CNN-based pyramidal architecture [31] is adopted to produce features at four different levels. Four intermediate outputs are pulled out from the encoder blocks for the monocular features. These are then simply concatenated at each level and used to produce dense predictions through a combination of reassemble and fusion blocks. + +Given the merged features, $F_{c_1}^{\mathrm{cat}}, F_{c_2}^{\mathrm{cat}}$ , we utilize two branches of dense prediction module, one for the depth of 3D Gaussians, $\mathrm{DPT}_{\mathrm{depth}}$ , and the other for the remaining Gaussian attributes $\mathrm{DPT}_g$ + +$$ +\tilde {D} _ {k} = \mathrm {D P T} _ {\text {d e p t h}} \left(F _ {k} ^ {\text {c a t}}\right), \tilde {\mathcal {G}} _ {k} = \mathrm {D P T} _ {g} \left(F _ {k} ^ {\text {c a t}}\right), k \in \left\{c _ {1}, c _ {2} \right\}, \tag {7} +$$ + +where $\tilde{\mathcal{G}}_k = \{(\delta x_j,\delta y_j,\alpha_j,\Sigma_j,c_j)\}_{j = 1}^{HW}$ is a set of the predicted Gaussians with all attributes except $z$ coordinates and $\tilde{D}_k$ is the predicted depth, further processed by the depth refinement module. $\delta x_{j}$ and $\delta y_{j}$ are the predicted offsets for each Gaussian, and the Gaussians for each input image in its coordinate system, $\mathcal{G}_{c_1},\mathcal{G}_{c_2}$ , can be obtained by adding the offsets to the pixel coordinates and unprojecting to 3D space using the refined depth $D_{c_1},D_{c_2}$ . Then, we construct the unified Gaussians by transforming them into a target coordinate system with the predicted poses (Eq. (4)). + +Matching-aware pose estimation. To enable high-quality rendering and reconstruction, it is crucial to predict accurate camera poses since it defines the transformation in 3D space. We begin by employing the CNN-based pose network from previous studies [17, 26] and introduce our matching-awareness module as a novel encoding strategy. As shown in Fig 2-(a), we use a 2D U-Net [40] with cross-attention blocks to extract multi-view aware features from unposed images. Unlike the dense prediction module, we also incorporate the target view $I_{t}$ as input and predict the relative camera poses for $(I_{c_1}, I_t)$ and $(I_{c_2}, I_t)$ . First, the matching network processes the triplet, + +$$ +F _ {c _ {1}} ^ {\mathrm {m a}}, F _ {t} ^ {\mathrm {m a}}, F _ {c _ {2}} ^ {\mathrm {m a}} = \text {M a t c h i n g N e t} \left(I _ {c _ {1}}, I _ {t}, I _ {c _ {2}}\right), \tag {8} +$$ + +where the matching aware features, $F_{k}^{\mathrm{ma}} \in \mathbb{R}^{H \times W \times 3}$ , have the same sizes as input images and inject these features into the pose network. + +$$ +T _ {c _ {1} \rightarrow t} = \operatorname {P o s e N e t} \left(\left[ F _ {c _ {1}} ^ {\mathrm {m a}}; I _ {c _ {1}}; E ^ {\mathrm {i n t}} \right], \left[ F _ {t} ^ {\mathrm {m a}}; I _ {t}; E ^ {\mathrm {i n t}} \right]\right), \tag {9} +$$ + +where $[\cdot ;\cdot ;\cdot ]$ concatenates input tensors along with the channel dimension, and $E^{\mathrm{int}}\in \mathbb{R}^{H\times W\times 3}$ is ray embedding of camera intrinsic matrix for scale-awareness. More specifically, $E_{x,y}^{\mathrm{int}} = K^{-1}p(x,y)\in \mathbb{R}^3$ , where $K\in \mathbb{R}^{3\times 3}$ is a camera intrinsic matrix and $p(x,y)\in \mathbb{R}^3$ is a homogeneous coordinate of a pixel coordinate $x,y$ . Note that the camera intrinsic parameters vary for different scenes but remain the same across different input views within the same scene. + +Pose-aware depth refinement. In this module, we refine the estimated depth map, $\tilde{D}_{c_1},\tilde{D}_{c_2}$ , derived from the + +dense prediction module, $\mathrm{DPT}_{\mathrm{depth}}$ , to improve the quality of rendering and reconstruction. The initial depth estimation, $\tilde{D}_{c_1}, \tilde{D}_{c_2}$ , yields inconsistent estimation between input views that negatively impact the overall accuracy of the reconstruction, e.g., incorrectly overlapping Gaussians. To resolve the limitation, we propose our refine module that leverages cross-view information with spatial awareness. While a few recent works have proposed depth refinement approaches [7, 62], our method uniquely differs by utilizing the predicted camera pose as additional information to resolve inconsistencies in the estimated depths across multiple input views. We employs a lightweight 2D U-Net, which takes current depth predictions, input images, and estimated poses as input and outputs residual depths for each view. The operation is defined as follows, + +$$ +\begin{array}{l} \Delta D _ {c _ {1}}, \Delta D _ {c _ {2}} = \operatorname {R e f i n e} \left(\left[ \tilde {D} _ {c _ {1}}; I _ {c _ {1}}; E ^ {\text {e x t}} \left(T _ {c _ {1} \rightarrow t}\right)\right], \right. \tag {10} \\ \left[ \tilde {D} _ {c _ {2}}; I _ {c _ {2}}; E ^ {\mathrm {e x t}} (T _ {c _ {2} \rightarrow t}) \right]), \\ \end{array} +$$ + +where $\Delta D_{k}$ is the residual for each view depth, and the final depth $D_{k} = \tilde{D}_{k} + \Delta D_{k}$ is obtained by adding the residual to the initial depth estimation for each view. Similar to our pose estimation module, there are cross-attention blocks in U-Net, and we utilize Plucker ray embedding to densely encode our estimated pose into a higher-dimensional representation space, e.g., $E^{\mathrm{ext}}(T_{c_1\rightarrow t})\in \mathbb{R}^{H\times W\times 6}$ (Fig. 2-(b)). + +
Pose-Freew/o 3D priorDepthFastw/o Finetune
VAE [26]✓✓X✓✓
DBARF [6]✓X²✓XX
FlowCAM [41]✓X✓X✓
CopoNeRF [21]X¹✓✓X✓
SelfSplat (Ours)✓✓✓✓✓
+ +Table 1. Baseline attributes compared to our proposed method. ${}^{1}$ CopoNeRF offers pose-free inference, but requires ground-truth pose supervision during training. ${}^{2}$ DBARF is trained from the pretrain generalizable NeRF, IBRNet [48] that has a 3D prior. + +# 5. Experiments + +# 5.1. Experiment Setup + +We train and evaluate our model on three large-scale datasets: RealEstate10K (RE10k) [67], ACID [33], and DL3DV [32], which include diverse indoor and outdoor real estate videos, aerial outdoor nature scenes, and diverse real-world videos, respectively. For RE10k, we use 67,477 training and 7,289 testing scenes; for ACID, 11,075 training and 1,972 testing scenes, consistent with previous works [3, 7]. Lastly, for DL3DV, we use subsets of the dataset amounting + +
RE10k
MethodAverageSmallMediumLarge
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
VAE [26]20.650.6430.32518.780.5850.41420.560.6460.31923.130.7010.241
DBARF [6]12.570.4940.47410.480.4970.52212.390.4870.47515.550.5130.415
FlowCAM [41]22.290.7110.31320.740.6790.37522.150.7090.31324.580.7540.241
CoPoNeRF [21]21.030.6930.25619.700.6700.34820.990.6950.28522.700.7150.217
Ours24.220.8130.18821.630.7490.25724.210.8200.18127.250.8640.132
+ +Table 2. Quantitative results of novel view synthesis on RE10k dataset. + +
MethodACID
AverageSmallMediumLarge
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
VAE [26]24.020.6660.28723.760.6790.30624.210.6640.29523.990.6580.263
DBARF [6]15.480.5720.39713.870.6180.42215.510.5840.39816.720.5210.378
FlowCAM [41]25.590.7210.29425.370.7300.31225.640.7140.30525.730.7230.267
CoPoNeRF [21]23.300.6680.27822.920.6920.30423.420.6670.28323.450.6490.251
Ours26.710.8010.19625.650.7760.23726.870.7970.19827.330.8260.159
+ +Table 3. Quantitative results of novel view synthesis on ACID dataset. + +![](images/0a16ba692f37846d10745ce867aa443fb892722aae4af9bec7088f076073270b.jpg) +Figure 3. Qualitative comparison of novel view synthesis on RE10k (top row) and ACID (bottom row) datasets. + +to 2,000 scenes (3K and 4K) for training and testing on 140 benchmark scenes, following PF3plat [22]. We assess our model's performance in reconstructing intermediate video frames between two context frames. + +Baselines. We compare our model against existing pose-free generalizable novel view synthesis methods, including VAE [26], DBARF [6], FlowCAM [41], and CoPoN-eRF [21], on two different tasks: novel view synthesis and relative camera pose estimation. We train all methods, including ours, using the same training curriculum, where the frame distance between context views increases gradually. We also provide an attribute overview in Tab. 1, showing the distinct features of our proposed method. + +Evaluation metrics. For novel view synthesis, we use standard metrics: PSNR, SSIM [50], and LPIPS [63]. Pose estimation is evaluated based on geodesic rotation and translation angular error, following [21]. For RE10k and ACID, we categorize test context pairs by image overlap ratios to evaluate performance across small(0.05-0.6), medium(0.6- + +0.8), and large $(0.8+)$ overlap, identified by a pretrained image matching method [14]. For DL3DV, overlap categories are defined by frame intervals between context images: 6 frames for large and 10 frames for small overlap. + +Implementation details. We employ the encoder part of pretrained CroCo [52] model as our monocular encoder, which is trained in a self-supervised manner, and utilized adapter [5] block designed to efficiently adapt pretrained ViT models to downstream tasks. For the Gaussian rasterizer, we implement it using gsplat [59], an open-source library for Gaussian Splatting [25], offering efficient computation and memory usage. We train RE10k and ACID with $256 \times 256$ resolution, and for DL3DV, we use $256 \times 448$ to accommodate the wider view in our experiments. + +# 5.2. Results + +Novel view synthesis. We report quantitative results in Tab. 2, 3 and qualitative results in Fig. 3 for RE10k and ACID datasets, while Tab. 6 and Fig. 4 for DL3DV dataset. + +
RE10k
MethodAverageSmallMediumLarge
RotationTranslationRotationTranslationRotationTranslationRotationTranslation
Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓
VAE [26]3.8592.818115.746113.5986.9716.936118.211116.2263.5953.037112.444109.9591.1270.525123.840132.181
DBARF [6]2.4711.47136.06924.3084.7713.00938.29225.8222.0431.44032.74920.8991.2220.91444.50035.788
FlowCAM [41]1.4381.16022.52415.9172.4831.74126.04419.5831.2641.13719.88914.3550.8010.74527.17919.328
CoPoNeRF [21]0.8390.5879.1755.5381.2810.83010.1736.0470.7840.5838.4315.1790.5110.32610.4876.232
Ours0.7500.3629.0954.4381.5230.70114.9547.2670.6030.3467.5933.9760.3440.2067.2813.799
+ +Table 4. Quantitative results of pose estimation on RE10k dataset. + +
MethodACID
AverageSmallMediumLarge
RotationTranslationRotationTranslationRotationTranslationRotationTranslation
Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓Med.°↓
VAE [26]1.6310.46186.18079.9872.7890.61781.05672.0601.5680.49785.37277.7610.8060.35591.187
DBARF [6]1.9750.86049.90695.3992.8991.06052.27536.8951.8790.75149.39636.7021.3730.74548.693
FlowCAM [41]3.8462.81992.78682.9764.7873.57484.59976.6594.2973.00191.26281.0192.5482.168101.074
CoPoNeRF [21]1.0580.35419.8889.7571.7310.50320.5729.4491.0250.36519.6179.4010.5750.26119.694
Ours0.9810.19916.3294.5351.7870.36221.6315.6810.9520.21916.0064.5180.3860.13412.592
+ +Table 5. Quantitative results of pose estimation on the ACID dataset. + +
ContextTarget ViewOursFlowCAM
+ +Figure 4. Qualitative comparison of novel view synthesis on DL3DV dataset. + +
DL3DV
Small (10 frame)Large (6 frame)
MethodPSNR ↑SSIM ↑RotationTranslationPSNR ↑SSIM ↑LPIPS ↓RotationTranslation
Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓Avg.°↓Med.°↓
FlowCAM [41]21.010.6080.4111.1381.01122.38516.43223.520.7100.3140.7050.62628.681
Ours21.910.7230.2790.9850.5739.6815.16424.820.8220.2000.6340.25612.057
+ +Table 6. Quantitative results of novel view synthesis and pose estimation on DL3DV dataset. + +Our method outperforms the baselines on all metrics, especially in terms of perceptual distance. These observations can be further confirmed by the rendering results that our method effectively captures fine details of 3D structure. + +Relative pose estimation. Tab. 4, 5, and 6 present the quantitative results for camera pose estimation between two images across the datasets. Our approach consistently achieves lower errors in both average and median deviations, highlighting its accuracy and robustness. The qualitative results in Fig. 5, visualizing epipolar lines from the estimated poses also demonstrates the effectiveness of our approach in capturing accurate geometric alignments. + +Cross-Dataset Generalization. To evaluate the generaliza- + +tion performance on out-of-distribution scenes, we train the models on RE10k (ACID) dataset and test them on ACID (RE10k) dataset without additional finetuning. As shown in Tab. 7, SelfSplat outperforms previous methods on unseen datasets, demonstrating robust generalization capabilities. + +# 5.3. Ablations and Analysis + +We provide quantitative and qualitative results on ablations studies in Tab. 8 and Fig. 6. All methods are trained for 50,000 iterations on RE10k dataset for a fair comparison. + +Importance of matching awareness in pose estimation. To measure the importance of adopting cross-view features in our pose network, we conduct a study ("No Matching + +
MethodRE10k → ACIDACID → RE10k
RotationTranslationRotationTranslation
PSNR ↑SSIM ↑LPIPS ↓Avg.°↓Med.°↓Avg.°↓Med.°↓PSNR ↑SSIM ↑LPIPS ↓Avg.°↓Med.°↓Avg.°↓Med.°↓
FlowCAM [41]25.310.7150.2971.6071.07432.31621.13521.130.6670.3294.4263.99574.87363.667
CoPoNeRF [21]23.560.6830.2872.3010.61919.6208.64218.890.6070.3644.1592.89320.43513.572
Ours26.600.7930.2061.1190.24918.6075.86421.650.7280.2421.6180.86717.99310.228
+ +![](images/19516655a7e2298f103da0c8a878a24e8c36535b823fff57ae3dce3bc86a4ca3.jpg) +Figure 5. Epipolar lines visualization. We draw the lines from reference to target frame using relative camera pose. + +awareness") by removing it from the pose network. Quantitatively, it leads to a drop in pose metrics: translation error increases by 1.5 degrees, which also negatively impacts rendering scores, decreasing PSNR by $0.4\mathrm{dB}$ . These results highlight that our encoding methods with multi-view awareness help capture relationships between frames, improving both pose estimation and novel view synthesis. + +Importance of depth refinement module. We conduct a study ("No Depth Refine") on our depth refinement module to validate its effectiveness. The results indicate a clear decline across all metrics: PSNR drops by 0.6 dB, and translation discrepancy increases by 1.1 degrees. Additionally, misalignment of overlapping Gaussians leads to degrading in visual quality, such as motion blur artifacts, demonstrating that our refinement scheme enhances the multi-view consistency of depth predictions. + +# How self-supervised depth estimation method and 3D-GS representation can make reciprocal improvement? + +We explore the benefits of combining self-supervised depth estimation with explicit 3D representation by comparing SelfSplat with two variants ("No Reprojection Loss", "No Rendering Loss"). Training without reprojection loss shows a significant performance decline across all metrics, particularly a $1.5\mathrm{dB}$ drop in PSNR, indicating challenges in accurately positioning Gaussians—a crucial factor for precise 3D reconstruction and novel view synthesis. In the "No Rendering Loss" variant, we replaced the rendered depth of the target view previously used in reprojection loss with an estimated depth map from the image using a dense prediction module. To validate the impact of incorporating 3D-GS, we also account for gradients of rotation, $R \in \mathbb{R}^{3 \times 3}$ , and translation, $t \in \mathbb{R}^{3 \times 1}$ , in camera poses. The rendering loss gradients with respect to translation and rotation are: + +$$ +\frac {\delta \mathcal {L} _ {\mathrm {r e n}}}{\delta t} = - \sum_ {j} \frac {\delta \mathcal {L} _ {\mathrm {r e n}}}{\delta \tilde {\mu} _ {j}}, \frac {\delta \mathcal {L} _ {\mathrm {r e n}}}{\delta R} = - \left[ \sum_ {j} \frac {\delta \mathcal {L} _ {\mathrm {r e n}}}{\delta \tilde {\mu} _ {j}} (\mu_ {j} - t) ^ {\top} \right] R, \tag {11} +$$ + +where $\tilde{\mu}_j$ is a splatted Gaussian in rendering viewspace. Excluding rendering loss results in degraded pose metrics, a common issue in self-supervised depth estimation methods with limited image overlap. Our framework effectively addresses this by combining explicit 3D-GS representation with rendering loss, improving depth and pose estimation. + +![](images/07b81e24e6e55bd81355dbaf87beeb42234869365c3a567548f5e46a17ce0d4e.jpg) +Figure 6. Ablation studies on our proposed component. + +Table 7. Cross-dataset generalization. We train the models on RE10k (ACID) dataset and directly evaluate on ACID (RE10k) dataset. + +
MethodPSNR ↑SSIM ↑LPIPS ↓Rot. Avg.° ↓Trans. Avg.° ↓
Ours22.650.7640.2221.03613.705
No Matching Network22.210.7350.2411.30815.171
No Depth Refine22.050.7440.2241.06414.8
No Reprojection Loss21.120.6720.2932.23628.584
No Rendering Loss---8.58164.436
+ +Table 8. Ablations. Our methods achieve better alignment of 3D Gaussians, with accurate pose and consistent depth estimations. + +# 6. Conclusion + +We present SelfSplat, a pose-free generalizable 3D Gaussian Splitting model that does not require pretrained 3D priors or an additional fine-tuning stage. Our method effectively integrates a 3D-GS representation with self-supervised depth estimation techniques to recover 3D geometry and appearance from unposed monocular videos. We conduct extensive experiments on diverse real-world datasets to demonstrate its effectiveness, showcasing its ability to produce photorealistic novel view synthesis and accurate camera pose estimation. We believe that SelfSplat represents a significant step forward in 3D representation learning, offering a robust solution for various applications. + +# Acknowledgements + +This work was supported by the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2024 (Research on neural watermark technology for copyright protection of generative AI 3D content, RS-2024-00348469; International Collaborative Research and Global Talent Development for the Development of Copyright Management and Protection Technologies for Generative AI, RS-2024-00345025), Korea Research Institute for defense Technology planning and advancement (KRIT) grant funded by the Korean Government, Defense Acquisition Program Administration (DAPA), under Grant KRIT-CT-23-021, Development of Multi-Modal Information Fusion and High-Realization Battlefield Data Generation Technology, in 2025, and the Ministry of Science and ICT (MSIT) of Korea, under the National Research Foundation (NRF) grant (RS-2024-00337548). + +# References + +[1] Jiawang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth and ego-motion learning from monocular video. Advances in neural information processing systems, 32, 2019. 3 +[2] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 2 +[3] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 1, 2, 3, 4, 5 +[4] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 4 +[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 6 +[6] Yu Chen and Gim Hee Lee. Dbarf: Deep bundle-adjusting generalizable neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24-34, 2023. 1, 2, 4, 5, 6, 7 +[7] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 2, 4, 5 + +[8] Boris Chidlovskii and Leonid Antsfeld. Self-supervised pretraining and finetuning for monocular depth and visual odometry. arXiv preprint arXiv:2406.11019, 2024. 3 +[9] Boris Chidlovskii and Leonid Antsfeld. Self-supervised pretraining and finetuning for monocular depth and visual odometry. arXiv preprint arXiv:2406.11019, 2024. 2 +[10] Kyunghyun Cho. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 2 +[11] Christopher Choy, Wei Dong, and Vladlen Koltun. Deep global registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2514-2523, 2020. 2 +[12] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4 +[13] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4970-4980, 2023. 1 +[14] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 6 +[15] Irving Fang, Kairui Shi, Xujin He, Siqi Tan, Yifan Wang, Hanwen Zhao, Hung-Jui Huang, Wenzhen Yuan, Chen Feng, and Jing Zhang. Fusionsense: Bridging common sense, vision, and touch for robust sparse-view reconstruction. arXiv preprint arXiv:2410.08282, 2024. 1 +[16] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 2 +[17] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 2, 3, 5 +[18] 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. 4 +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2 +[20] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shoou-I Yu. Epipolar transformers. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7779-7788, 2020. 4 +[21] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jiaolong Yang, Seungryong Kim, and Chong Luo. Unifying correspondence, pose and nerf for pose-free novel view synthesis + +from stereo pairs. arXiv preprint arXiv:2312.07246, 2023. 1, 2, 4, 5, 6, 7, 8 +[22] Sunghwan Hong et al. Pf3plat: Pose-free feed-forward 3d gaussian splatting. arXiv:2410.22128, 2024. 2, 6 +[23] Ying Jiang, Chang Yu, Tianyi Xie, Xuan Li, Yutao Feng, Huamin Wang, Minchen Li, Henry Lau, Feng Gao, Yin Yang, et al. Vr-gs: a physical dynamics-aware interactive gaussian splatting system in virtual reality. In ACM SIGGRAPH 2024 Conference Papers, pages 1-1, 2024. 1 +[24] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 2 +[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 6 +[26] Zihang Lai, Sifei Liu, Alexei A Efros, and Xiaolong Wang. Video autoencoder: self-supervised disentanglement of static 3d structure and motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9730-9740, 2021. 5, 6, 7 +[27] Christoph Lassner and Michael Zollhofer. Pulsar: Efficient sphere-based neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1440-1449, 2021. 1 +[28] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 3 +[29] Hao Li et al. Ggrt: Towards generalizable 3d gaussians without pose priors in real-time. arXiv:2403.10147, 2024. 1, 2, 4 +[30] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 2 +[31] Tsung-Yi Lin, Piotr Dollar, 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. 5 +[32] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2, 5 +[33] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 2, 5 +[34] 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. 4 +[35] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1995. 1 +[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1 +[37] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 4 +[38] Adam Rashid, Satvik Sharma, Chung Min Kim, Justin Kerr, Lawrence Yunliang Chen, Angjoo Kanazawa, and Ken Goldberg. Language embedded radiance fields for zero-shot task-oriented grasping. In 7th Annual Conference on Robot Learning, 2023. 1 +[39] Chris Rockwell, Justin Johnson, and David F Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 2 +[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 5 +[41] Cameron Smith, Yilun Du, Ayush Tewari, and Vincent Sitzmann. Flowcam: training generalizable 3d radiance fields without camera poses via pixel-aligned scene flow. arXiv preprint arXiv:2306.00180, 2023. 1, 2, 5, 6, 7, 8 +[42] Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[43] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 1, 2, 3 +[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 2 +[45] Hans Thisanke, Chamli Deshan, Kavindu Chamith, Sachith Seneviratne, Rajith Vidanaarachchi, and Damayanthi Herath. Semantic segmentation using vision transformers: A survey. Engineering Applications of Artificial Intelligence, 126:106669, 2023. 4 +[46] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35:10078-10093, 2022. 2 + +[47] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 4 +[48] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 5 +[49] Shuzhe Wang et al. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 3 +[50] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 3, 6 +[51] Philippe Weinzaepfel, Vincent Leroy, Thomas Lucas, Romain Brégier, Yohann Cabon, Vaibhav Arora, Leonid Antsfeld, Boris Chidlovskii, Gabriela Csurka, and Jérôme Revaud. Croco: Self-supervised pre-training for 3d vision tasks by cross-view completion. Advances in Neural Information Processing Systems, 35:3502-3516, 2022. 3, 4 +[52] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 3, 4, 6 +[53] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7467-7477, 2020. 1 +[54] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 4 +[55] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, et al. Vr-nerf: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia 2023 Conference Papers, pages 1–12, 2023. 1 +[56] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 4 +[57] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 4 +[58] Botao Ye, Sifei Liu, Haofei Xu, Li Xueting, Marc Pollefeys, Ming-Hsuan Yang, and Peng Songyou. No pose, no problem: + +Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2 +[59] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, et al. gsplat: An open-source library for gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 6 +[60] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1983-1992, 2018. 3 +[61] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 1 +[62] Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. arXiv preprint arXiv:2408.13770, 2024. 4, 5 +[63] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[64] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 1 +[65] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 1 +[66] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 3 +[67] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2, 5 \ No newline at end of file diff --git a/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/images.zip b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb03fee307391decd7705ed00123fc8c5fd2fe7b --- /dev/null +++ b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:acfaaa200201d01bbff11ae11f0fce2e63709c8e4f645765db650407343a9d33 +size 839586 diff --git a/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/layout.json b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..832b4737becb973068d59736e71c10e7e047a0b7 --- /dev/null +++ b/CVPR/2025/SelfSplat_ Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edefa8bdae3c1bd9d4d2d527464a65d5279850f2d285b95ade1ab36140be3ddc +size 417409 diff --git a/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/5cbc6a72-c5f6-4ba9-8961-eb87b27e63b6_content_list.json b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/5cbc6a72-c5f6-4ba9-8961-eb87b27e63b6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9729e51c293f86d47f512f9b2d6dc5b3dc3e4d77 --- /dev/null +++ b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/5cbc6a72-c5f6-4ba9-8961-eb87b27e63b6_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:821ac3fd1a9cfef27eb88ab66008cede1e1d8c8602ede3e172b999099f411945 +size 79323 diff --git a/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/5cbc6a72-c5f6-4ba9-8961-eb87b27e63b6_model.json b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/5cbc6a72-c5f6-4ba9-8961-eb87b27e63b6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9f4cfd5ccfe9a30aa8924e1a906663fa946b9530 --- /dev/null +++ b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/5cbc6a72-c5f6-4ba9-8961-eb87b27e63b6_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a44e09aa6de6ff72762dbbd466b0dd33317977dc1e7d8ad0a0eb33ffdde78e3 +size 96394 diff --git a/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/5cbc6a72-c5f6-4ba9-8961-eb87b27e63b6_origin.pdf b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/5cbc6a72-c5f6-4ba9-8961-eb87b27e63b6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9e6d6f94f5838b12280652d9e46cb14012c751a8 --- /dev/null +++ b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/5cbc6a72-c5f6-4ba9-8961-eb87b27e63b6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e12179ee02c189c7797f27c70fcf2181e7a15c29578dae923359c79c8d194797 +size 2377792 diff --git a/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/full.md b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0c8b1846db645cc71a2de94b8051074c96d3c1f8 --- /dev/null +++ b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/full.md @@ -0,0 +1,357 @@ +# SEMALIGN3D: Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations + +Krispin Wandel +Shanghai Jiao Tong University +krispin.wandel@sjtu.edu.cn + +Hesheng Wang* +Shanghai Jiao Tong University +wanghesheng@sjtu.edu.cn + +# Abstract + +Semantic correspondence made tremendous progress through the recent advancements of large vision models (LVM). While these LVMs have been shown to reliably capture local semantics, the same can currently not be said for capturing global geometric relationships between semantic object regions. This problem leads to unreliable performance for semantic correspondence between images with extreme view variation. In this work, we aim to leverage monocular depth estimates to capture these geometric relationships for more robust and data-efficient semantic correspondence. First, we introduce a simple but effective method to build 3D object-class representations from monocular depth estimates and LVM features using a sparsely annotated image correspondence dataset. Second, we formulate an alignment energy that can be minimized using gradient descent to obtain an alignment between the 3D object-class representation and the object-class instance in the input RGB-image. Our method achieves state-of-the-art matching accuracy in multiple categories on the challenging SPair-71k dataset, increasing the PCK@0.1 score by more than 10 points on three categories and overall by 3.3 points from $85.6\%$ to $88.9\%$ . Additional resources and code are available at dub.sh/semalign3d. + +# 1. Introduction + +Semantic correspondence aims at finding an alignment between two images based on the meaning or function of image regions rather than their exact spatial or visual similarity. This ability is crucial for robotics to learn policies with category-level generalization capabilities [4-6, 19, 29]. Beyond robotics, this task has also several applications in image editing [8, 23, 24, 31] and style transfer [7, 14]. + +Large Vision Models (VIT) such as DinoV2 [25] have pushed this field dramatically through the introduction of deep features that are able to capture the underlying seman + +![](images/3e70468a9212634884c6a8b52787764a31ca4d02b1b052f91a05dff7dc814a30.jpg) + +![](images/73b0c579fd052824b70f19932dae9aea3731558fab503c177e74e4009a29d9f6.jpg) +Figure 1. Our method aligns learned 3D object-class representations, displayed as the colored point cloud, with object instances to improve robustness of sparse semantic correspondence against extreme view variation and occlusion, which is still a major issue for existing methods like GeoAware [32]. + +tics of an object. However, as observed by [20, 32], these features still fall short when dealing with symmetric objects or extreme view variation. Recent work has proposed solutions such as feature fine-tuning or the introduction of spatial bias (see Sec. 2) but the same issues still persist in many instances, as depicted in Fig. 1. + +To derive a more robust method, we propose to learn category-level 3D object representations and then align these representations with the object instance shown in the image at inference time. Our hypothesis is that this process is significantly more data-efficient and robust than purely relying on features built at inference time. Towards this goal we make two contributions: + +1. We introduce a simple method to combine monocular depth estimates and deep features to build 3D object-class representations using a sparsely annotated dataset. (see Sec. 3.1). +2. We define an alignment loss function that can be minimized using gradient descent to obtain the alignment between the 3D representation and the object instance shown in the image. (see Sec. 3.2). + +Our method shows substantial performance gains over the previous state-of-the-art [32] in multiple categories on the SPAIR-71K [22] dataset, increasing the PCK@0.1 score by more than 10 points in three categories and the overall score by 3.3 points (from $85.6\%$ to $88.9\%$ ) (see Sec. 4 for details). + +# 2. Related Work + +The field of semantic correspondence focuses on finding meaningful correspondences between parts of objects in different images, often within an unsupervised or self-supervised framework. Several approaches have been developed to address this, categorized broadly into learning mappings, optimizing correlation matrices, and fine-tuning feature representations. + +Learning Mappings to Canonical Coordinates. Unsupervised methods in this category, like ASIC [8] and NeuCongeal [24], aim to learn mappings between deep features and canonical coordinates. ASIC minimizes contrastive and reconstruction losses, mapping features to canonical coordinates for both sparse and dense matching. NeuCongeal follows a similar approach but leverages deep feature-based reconstruction loss. Another notable approach, Viewpoint-Guided Spherical Maps [20], maps features to canonical spherical coordinates, integrating a weak geometric prior, though it faces challenges with complex or non-rigid objects. + +Optimizing the Correlation Matrix. Techniques such as SCOT [17] and PMNC [15] optimize correlation matrices to improve matching precision. SCOT frames semantic correspondence as an optimal transport problem, adding geometric consistency with regularized Hough Matching (RHM). However, RHM can be inaccurate for symmetrical objects. PMNC refines correlation matrices via learned patch scoring functions. CATs++ [3] combines multi-layer correlations using transformers, enhancing efficiency with convolutions. + +Fine-Tuning Deep Features. Recent works like SCorrSAN [11, 33] employ sparse annotations and generate dense pseudo labels to improve keypoint matching accuracy. Other studies utilize features extracted from + +models like DinoV2 [25, 31] or stable diffusion (SD) [16, 26], with methods like DHF [18] refining feature representations through cosine similarity loss. Fine-tuning DinoV2 and SD features across datasets with similar domains also demonstrates improved performance [32], e.g. by fine-tuning on bird datasets to enhance airplane features. + +# 3. Method + +# 3.1. Building 3D Object-Class Representations + +We aim to leverage the recent progress in monocular depth estimation [2, 12, 27] to build 3D object class representations from a small set of sparsely annotated RGB images. Specifically, we assume a set of $n$ images that display instances of an object class (e.g. passenger airplane), and each image $s$ is sparsely annotated with $L$ labeled keypoints $k_{i=1\dots L,\mathrm{im}}^s$ (e.g. nose, left/right wing end). For notational simplicity, we assume all images have the same number of keypoints. Our goal is to obtain 3D representations that capture the semantic regions and their geometric relationships for each object category. Importantly, these representations do not embody a specific object (e.g. Airbus 380), but a class of objects (e.g. passenger airplane). In the following, we propose a simple but effective procedure to construct these 3D object class representations. + +Keypoint World Coordinates. First, we compute world coordinates for the image keypoints using monocular depth estimates from DepthAnythingV2 [27]. Since camera intrinsics are unknown, we approximate them up to scale by estimating focal length parameters. Given a focal length estimate, we can back-project 2D keypoints to 3D world coordinates and construct scale-invariant geometric features: + +$$ +A _ {i j k l} ^ {s} = \angle \left(e _ {i j} ^ {s}, e _ {k l} ^ {s}\right) \tag {1} +$$ + +$$ +R _ {i j k l} ^ {s} = \frac {\left| e _ {i j} ^ {s} \right|}{\left| e _ {i j} ^ {s} \right| + \left| e _ {k l} ^ {s} \right|}, \tag {2} +$$ + +where $e_{ij}^{s} = k_{j,\mathrm{world}}^{s} - k_{i,\mathrm{world}}^{s}$ represents the edge between keypoints. We optimize focal length estimates by minimizing the variance of these features across $n$ train images: + +$$ +f _ {1 \dots n} ^ {*} = \min _ {f _ {1 \dots n}} \sum_ {i j k l} \operatorname {V a r} _ {s = 1 \dots n} \left(R _ {i j k l} ^ {s}\right) + \operatorname {V a r} _ {s = 1 \dots n} \left(A _ {i j k l} ^ {s}\right) \tag {3} +$$ + +Sparse 3D Representation. Next, we construct a sparse canonical point cloud $C_{\mathrm{sparse}} = \{k_{i,\mathrm{world}} \mid i = 1 \dots L\}$ by iteratively computing the most likely 3D location for each keypoint based on the Beta distributions $A_{ijkl}$ and $R_{ijkl}$ fitted to the geometric features on $n$ train images $A_{ijkl}^{s=1\dots n}$ and $R_{ijkl}^{s=1\dots n}$ , respectively. For semantics, we extract image-patch features using the pre-trained model from GeoAware + +![](images/355958e57361a6e456babc528a6545b685b5d1f97e62fd6a97cde97aac68b90f.jpg) +airplane + +![](images/a63299ab892f1e5dac7ab6a29cdf9013594f3b05dcbbda7c78464a5e10173730.jpg) +bicycle + +![](images/eec5ba108f12de4cf1c56beafd5b7acb1aec5a74ccecfd2eaead637d357866c5.jpg) +bird + +![](images/cea45fb88a8e0a0282dd06dbe6f1812f272ccca1883fb8d8e35e1da804efd10f.jpg) +boat + +![](images/8c8cca48e8d83ba7ff7c671c7d6134df14cbef26852efcfd6c05a9f25eb84d7c.jpg) +bottle + +![](images/fc8890d109eb6806ae4c7ceaf8fa35385a2b5288c7d1bdfead426bbe5db95322.jpg) +bus + +![](images/e6868e81a0c3194f0ae587debd2e6ecafb077db4de2108f7e982ba3a74cffa6e.jpg) +car + +![](images/d7b849554ed4d84188ebb9582ef4c29590d36086c9ccb5b9ee48e86645c0f5a8.jpg) +cat + +![](images/9b5b50275c7d2b1b0a84111907edf5687c0495dd63f62d9f4a4034307d811af1.jpg) +chair + +![](images/bdc555112d4f41d96579201bc1513227d915a4842376876b259a7bc4ebb3da8f.jpg) +COW + +![](images/d5074ad03f9b8103c9c995ea7ec80621790f419614a4515547fc98f7963c121f.jpg) +dog + +![](images/980df212801b898de18bbbfa48be43c01cad3a076e4f9844162573f34f1f5726.jpg) +horse + +![](images/1a512ebfa121d9049ba84757fcfce6630073ab7053193fe47bfc853030ae8bd8.jpg) +motorcycle + +![](images/83ed34e1ac68d80324eec0c495c337402cdb0ecc5d8b938224e80264d43753f1.jpg) +person + +![](images/0f051ea5d931350bcddb1ddba2862241e89a98c8e581d005c26b6c7fd957a7f1.jpg) +potted plant + +![](images/69899b62deda16dd630e4f79fc9c19bfe2c492753e596255254087a32b354265.jpg) +sheep + +![](images/e1c8a3fdf2021fcc388226db8263771852c0490a39f5c6e726638d46761f3d78.jpg) +train + +![](images/2d53c203fce05974f79a152f288c58647495caa6ab7e6fd633c4567a4474307f.jpg) +TV + +![](images/a41276a484700afe62e4fd810d68ac1d8637714bfc1598327779f3f9225ac78d.jpg) +Figure 2. Learned 3D model representations for all SPair-71k [22] categories. Colors represent the principal components of the point cloud features. +Figure 3. Construction of Geometric Features. Optimal focal lengths $f_{s}^{*}$ can be found by minimizing variance of scale-invariant geometric features $A_{ijkl}^{s=1\dots n}$ and $R_{ijkl}^{s=1\dots n}$ . We use the state-of-the-art model DepthAnythingV2 [27] for monocular depth estimation. Although the produced depth map (top right) looks visually appealing, it is still far from perfect as evident when back-projected to world coordinates. Nevertheless, in this work we demonstrate that we can still use these depth maps to build coherent 3D object-class representations. +Figure 4. Construction of Sparse Keypoint Point Cloud in Canonical Form. Built by iteratively computing the next most likely keypoint location based on the fitted Beta-distributions over $A_{ijkl}^{s=1\dots n}$ and $R_{ijkl}^{s=1\dots n}$ . + +[32]. Subsequently, we compute the semantic feature vector $q_{i}$ of $k_{i,\mathrm{world}}$ as the average of $q_{i}^{s = 1\dots n}$ , where $q_{i}^{s}$ denotes the image patch feature closest to $k_{i,\mathrm{img}}^{s}$ . In addition, we also compute the mean $a_{\mu_i}$ and the standard deviation $a_{\sigma_i}$ of the cosine similarities $a(q_i,q_i^{s = 1\dots n})$ between $q_{i}$ and $q_{i}^{s = 1\dots n}$ to obtain a keypoint presence certainty measure for $k_{i,\mathrm{world}}$ . + +![](images/a1b8cf39bc386248964c6ce03e8247bbd9b564dee3486bfb70e6c367f6201230.jpg) +Iterative Keypoint Sampling + +![](images/026602b86c04d6d6afac0fbf0df5a45f69f32139046f1faf39003711f875f37d.jpg) +Fitted Geometric Feature Distributions + +![](images/c451609a242dec661d85962e006c28ce8eee10440907cea6c976b5edba7bed2d.jpg) +Sparse Keypoint Point Cloud in Canonical Form + +Dense 3D Representation. Let $p_s = \pi_{f_s}^{-1}(d_s)$ denote the point cloud obtained by back-projecting the estimated depth image $d_s$ . First, we align $p_s$ with the sparse canonical point cloud $C_{\text{sparse}}$ via barycentric parametrization with respect to the labeled keypoints $k_{i,\text{world}}^s$ and merge all aligned point clouds $p_{\text{aligned}}^{s=1\dots n}$ into one large point cloud $p_{\text{merged}}$ , while discarding all points that are outside of the convex hull spawned by the image keypoint world coordinates. Following this, we obtain the dense point cloud $C_{\text{dense}}$ through k-mean clustering $p_{\text{merged}}$ , keeping only the clusters that are above a given density threshold. For semantics, let $N(c_i)$ of + +a point $c_i \in C_{\mathrm{dense}}$ be the index set of the $m$ nearest neighbors to cluster $c_i$ in $p_{\mathrm{merged}}$ . Furthermore, let $q_j^s$ denote the image-patch feature that is closest to the image coordinate of a point $j$ in $N(c_i)$ . Then, the semantic feature vector $q_{c_i}$ of $c_i \in C_{\mathrm{dense}}$ is the average over the neighbor features $q_{N(c_i)} = \{q_j^s | j \in \mathbf{N}(c_i)\}$ . As before, we also compute mean and variance of the cosine similarities between $q_{c_i}$ and $q_{N(c_i)}$ for presence certainty estimation. + +# 3.2. Aligning 3D Model Representations + +We now focus on the core of our method: utilizing the dense 3D object-class representations for semantic correspondence. Specifically, given an RGB image with an object instance, our goal is to determine the optimal 3D keypoint coordinates $k_{i}^{*} \in C_{\mathrm{sparse}}^{*}$ and focal length $f^{*}$ that yield the deformed dense point cloud $C_{\mathrm{dense}}^{*}$ aligning precisely with the instance when projected to the image plane. Recall that $C_{\mathrm{dense}}$ is parametrized by the sparse keypoint set $C_{\mathrm{sparse}}$ through barycentric coordinates, allowing for flexible deformation. To achieve this alignment, we formulate a simple yet effective loss function, $\mathcal{L}_{\mathrm{align}}(C_{\mathrm{sparse}},f)$ , minimized via gradient descent. This loss function consists of four key components: a reconstruction loss that maximizes image likelihood, a geometric consistency loss to enforce structural plausibility, an optional background mask penalization term, and a regularization loss for faster convergence during optimization. Each of these loss terms will be explained in detail below. + +Image Likelihood Maximization. We define a likelihood function over image pixel coordinates that is maximal when the 3D class representation is well-aligned with the object instance. Let $q_{\mathrm{img}}$ be the image embedding returned by the pre-trained GeoAware model [32]. Then, we can define the semantic and spatial likelihood at $(x,y)$ for point $i$ in $C_{\mathrm{dense}}$ as: + +$$ +p _ {\text {s e m}} ^ {x y i} = g \left(a _ {x y} ^ {\mathrm {u p}} \left(q _ {\mathrm {i m g}}, q _ {i}\right); a _ {\mu} ^ {i}, a _ {\sigma} ^ {i}\right) \tag {4} +$$ + +$$ +p _ {\text {s p a t i a l}} ^ {x y i} \left(C _ {\text {s p a r s e}}, f\right) = g \left(\left| (x, y) - \pi_ {f} \left(c _ {i}\right) \right|; \mu = 0, \sigma\right), \tag {5} +$$ + +where $g$ is the max-normalized normal probability density function, $a^{\mathrm{up}}$ is the cosine similarity up-scaled to image resolution, and $\sigma$ in Eq. (5) is a hyper-parameter. As we explain later, during optimization, it is crucial to start with a large $\sigma$ and decrease it over time. The combined reconstruction likelihood is: + +$$ +p _ {\text {r e c o n s t r u c t}} ^ {x y i} \left(C _ {\text {s p a r s e}}, f\right) = p _ {\text {s p a t i a l}} ^ {x y i} (f) \cdot p _ {\text {s e m}} ^ {x y i} \tag {6} +$$ + +Following this, the reconstruction loss is the negative total image likelihood defined as: + +$$ +P (\text {i m a g e} | C _ {\text {s p a r s e}}, f) = \mathbf {E} _ {x, y} \left[ \max _ {c _ {i} \in C _ {\text {d e n s e}}} p _ {\text {r e c o n s t r u c t}} ^ {x y i} \right] \tag {7} +$$ + +$$ +\mathcal {L} _ {\text {r e c o n s t r u c t}} ^ {\text {d e n s e}} = - P (\text {i m a g e} \mid C _ {\text {s p a r s e}}, f) \tag {8} +$$ + +Since we maximize the expected value, in practice we do not optimize over all pixel coordinates but only over a few samples. Lastly, we also maximize the image likelihood with respect to the keypoints $k_{i} \in C_{\mathrm{sparse}}$ to get a more precise position for the keypoints: + +$$ +\mathcal {L} _ {\text {r e c o n s t r u c t}} = w _ {\text {d e n s e}} \cdot \mathcal {L} _ {\text {r e c o n s t r u c t}} ^ {\text {d e n s e}} + w _ {\text {s p a r s e}} \cdot \mathcal {L} _ {\text {r e c o n s t r u c t}} ^ {\text {s p a r s e}}, \tag {9} +$$ + +where $w_{\text{dense}}$ and $w_{\text{sparse}}$ are hyper-parameters. In practice, we start the optimization procedure with high $w_{\text{dense}}$ to avoid local minima and increase $w_{\text{sparse}}$ over time for improved precision. + +Geometric Consistency. To enforce geometric consistency during alignment, we maximize the likelihood of keypoint angles and edge ratios according to the fitted Beta distributions: + +$$ +\mathcal {L} _ {\text {g e o m}} = - \mathbf {E} _ {i j k l} \left[ A _ {i j k l} \left(C _ {\text {s p a r s e}}\right) + B _ {i j k l} \left(C _ {\text {s p a r s e}}\right) \right] \tag {10} +$$ + +It is worth noting that $C_{\mathrm{sparse}}$ is initially already geometric consistent. This makes it much easier to optimize the loss function as we do not have to find a configuration that is geometric consistent but only keep the initial configuration geometric consistent. + +Background Mask Penalization. The reconstruction loss $\mathcal{L}_{\mathrm{reconstruct}}$ tries to maximize the spatial-semantic likelihood for every pixel in the image. However, it does not penalize wrong positives, that is, when the point cloud is projected to background image pixels that neither belong to the object segmentation mask $m_{\mathrm{seg}}$ nor a region $m_{\mathrm{occ}}$ that occludes the object. While we can obtain $m_{\mathrm{seg}}$ by combining image attention with Segment Anything [13], obtaining $m_{\mathrm{occ}}$ is very difficult. However, we observed that a background mask penalization loss $\mathcal{L}_{\mathrm{background}}$ that is only based on $m_{\mathrm{seg}}$ can still help to avoid local minima even when heavy occlusion is present. We define $\mathcal{L}_{\mathrm{background}}$ as follows: + +$$ +\mathcal {L} _ {\text {b a c k g r o u n d}} = \frac {1}{\left| C _ {\text {d e n s e}} \right|} \sum_ {c _ {i} \in C _ {\text {d e n s e}}} \mathbf {E} _ {x y \in m _ {\text {s e g}}} \left[ p _ {\text {s p a t i a l}} ^ {x y i} \right] \tag {11} +$$ + +Regularization. Since all features are scale invariant, we can constrain the motion of $C_{\mathrm{sparse}}$ to boost convergence rate. Concretely, we add a soft constraint to the mean of the z-coordinates of $C_{\mathrm{sparse}}$ : + +$$ +\mathcal {L} _ {\text {d e p t h}} = \left(d _ {\text {t a r g e t}} - \mu_ {\text {s p a r s e}, z}\right) ^ {2}, \tag {12} +$$ + +where $d_{\mathrm{target}}$ is a constant that can be set to an arbitrary value greater than zero. + +![](images/dd84be965d15b1b784276f1a8745fb370a1a230f5420ac79e20e5d92a2f3d1dc.jpg) +Figure 5. Overview of our method. After we constructed the object-class representation, we can minimize $\mathcal{L}_{\mathrm{align}}$ , which is based on similarity maps (here we only show 4) between the input RGB-image and the representation, to obtain the alignment. The green dots in the final image on the right represent the ground-truth keypoints. + +Final Loss Function. Putting it all together, the final alignment loss function is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {a l i g n}} = w _ {\text {r e c o n s t r u c t}} \cdot \mathcal {L} _ {\text {r e c o n s t r u c t}} + w _ {\text {g e o m}} \cdot \mathcal {L} _ {\text {g e o m}} \tag {13} \\ + w _ {\text {b a c k g r o u n d}} \cdot \mathcal {L} _ {\text {b a c k g r o u n d}} + w _ {\text {d e p t h}} \cdot \mathcal {L} _ {\text {d e p t h}} \\ \end{array} +$$ + +We will discuss the impact of each loss term and the choice of hyper-parameters $w_{i}$ in more detail in the next chapter. + +# 3.3. Semantic Correspondence + +After we found the optimal aligned point clouds $C_{\mathrm{sparse}, \{1,2\}}^*$ and focal lengths $f_{\{1,2\}}^*$ for images $s_1$ and $s_2$ , we can perform semantic correspondence between $s_1$ and $s_2$ simply by traversing the alignment process in reverse order. That is, given a query image point $p_1$ in image $s_1$ , we compute + +$$ +p _ {\text {r e c o n s t r u c t}} ^ {p _ {1}, i} \left(C _ {\text {s p a r s e}, 1} ^ {*}, f _ {1} ^ {*}\right) = p _ {\text {s p a t i a l}} ^ {p _ {1}, i} \cdot p _ {\text {s e m}} ^ {p _ {1}, i} \tag {14} +$$ + +similar to Eq. (6) and choose the point index $i^*$ for which $p_{\mathrm{reconstruct}}^{p_1,i}$ is maximal. Then, we can find its corresponding position in image $s_2$ by computing $p_{\mathrm{reconstruct}}^{xy,i^*}(C_{\mathrm{sparse},2}^*,f_2^*)$ for all image coordinates in $s_2$ and choosing the one with maximum likelihood. Another benefit of this approach is that we can adjust the variance of $p_{\mathrm{spatial}}$ and $p_{\mathrm{sem}}$ depending on how much we trust these terms. + +# 4. Experimental Evaluation + +# 4.1. Experimental Setup + +Datasets. Following previous work in semantic correspondence, we evaluate our method on the SPAIR-71K dataset [22]. SPAIR-71K contains around 71,000 image pairs, with each pair belonging to the same object category, such as animals, vehicles, or household items. Each pair includes detailed annotations like keypoints and bounding boxes. The dataset features significant variations in viewpoint, scale, occlusion, and variation among objects within each category, making it one of the most challenging + +datasets for semantic correspondence. For completeness, we also evaluate our method on a small subset of AP-10K [30] and PF-PASCAL [9, 10]. + +Evaluation Metrics. For a fair comparison with previous work we follow common practice and use the Percentage of Correct Keypoints (PCK) [28] as primary evaluation metric for sparse semantic correspondence. PCK at level $\alpha$ is defined as the percentage of predicted points that are no further away than $\alpha \cdot \max(w_{\mathrm{bbox}}, h_{\mathrm{bbox}})$ from the ground truth, where $w_{\mathrm{bbox}}$ and $h_{\mathrm{bbox}}$ are the height and width of the corresponding object-instance bounding boxes. While PCK is a popular metric for semantic correspondence, it primarily measures accuracy, not precision. Therefore, we also report results for $\mathrm{PCK}^{\dagger}$ [1] and KAP [20]. + +Optimization. We start the optimization procedure with a random initial guess for $C_{\mathrm{sparse}}$ and the focal length $f$ . To evaluate $\mathcal{L}_{\mathrm{reconstruct}}$ in Eq. (8), we sample a few image points based on the attention between the image and $Q_{\mathrm{dense}}$ . Although we observed that our method is good at finding the global minimum, there are initial configurations where the method can end up in a bad local minimum. To address this issue, we simply start with multiple initial guesses optimized in parallel. In our experiments, we start with $f \in [10.0, 5.0, 2.5, 1.25]$ and four $C_{\mathrm{sparse}}$ samples per focal length. After optimization, instead of simply choosing the solution with the smallest loss, we add two post-processing steps: First, for object-class representations that are almost flat, such as TV or bottle, we apply a normal check and discard solutions that face away from the viewer. Second, we use the depth estimate of the image and penalize solutions when the z-order of the corresponding $C_{\mathrm{dense}}^*$ is locally different from the depth estimate. + +
MethodAeroBikeBirdBoatBottleBusCarCatChairCowDogHorseMotorPersonPlantSheepTrainTVAll
UASIC [8]57.925.268.124.735.428.430.954.821.645.047.239.926.248.814.524.940.924.636.9
DINOv2+NN [25, 31]72.762.085.241.340.452.351.571.136.267.164.667.661.068.230.762.054.324.255.6
DIFT [26]63.554.580.834.546.252.748.377.739.076.054.961.353.346.057.857.171.163.457.7
SD+DINOv2 [31]73.064.186.440.752.955.053.878.645.577.364.769.763.369.258.467.666.253.564.0
SphericalMaps [20]74.864.587.145.652.777.871.482.447.782.067.373.967.660.049.969.878.559.167.3
SSCOT [17]34.920.763.821.143.527.321.363.120.042.942.531.129.835.027.724.448.440.835.6
PMNC [15]54.135.974.936.542.148.840.072.621.167.658.150.540.154.143.335.774.559.950.4
SCorrSAN [11]57.140.378.338.151.857.847.167.925.271.363.949.345.349.848.840.377.769.755.3
CATs++ [3]60.646.982.541.656.864.950.472.829.275.865.462.550.956.154.848.280.974.959.8
DHF [18]74.061.087.240.747.870.074.480.938.576.160.966.866.670.358.054.387.460.364.9
SD+DINOv2 (S) [31]81.266.991.661.457.485.383.190.854.588.575.180.271.977.960.768.992.465.874.6
GeoAware (AP-10k P.T.) [32]92.076.197.270.470.591.489.792.773.495.090.587.781.891.682.383.496.585.385.6
Ours95.680.494.066.582.292.692.591.488.396.191.387.484.292.387.785.395.996.188.9
+ +Table 1. Comparison of PCK@0.1 accuracy for all categories in the SPAIR-71K dataset. U=unsupervised, S=supervised. + +
MethodPCK@0.01PCK@0.05PCK@0.1
UDINOv2+NN [25, 31]6.338.453.9
DIFT [26]7.239.752.9
SD+DINOv2 [31]7.944.759.9
SphericalMaps [20]--69.8
SSCorrSAN [11]3.636.355.3
CATs++ [3]4.340.759.8
DHF [18]8.750.264.9
SD+DINOv2 (S) [31] [31]9.657.774.6
GeoAware (AP-10k P.T.) [32]22.075.385.6
Ours15.877.588.9
+ +Hyper-Parameters. For the specific hyper-parameters per category please refer to the supplementary material. However, there are a few general important considerations for choosing the hyper-parameters: + +First, it drastically improves performance when starting the optimization with a large $\sigma$ in $p_{\mathrm{spatial}}$ from Eq. (5) and gradually decreasing its value over time. Intuitively, this makes sense because our initial guess for $C_{\mathrm{sparse}}$ might be quite off which should be reflected in the spatial likelihood estimate. Another side benefit of this scheduling policy is that it is very robust against image pixel outliers with wrong semantics because, as we decrease $\sigma$ , the gradients with respect to these pixels will converge to zero. + +Another set of hyper-parameters includes the loss term weights in Eq. (13). Generally, we found that our method is not very sensitive with respect to these weights. However, one should choose $w_{\mathrm{geom}}$ within reasonable bounds in order to make $C_{\mathrm{sparse}}$ not too stiff. Furthermore, one should not start the optimization with a high value for $w_{\mathrm{depth}}$ , because this might push $C_{\mathrm{sparse}}$ outside the image depending on the level of object-instance occlusion. + +# 4.2. Results and Analysis + +Quantitative Analysis. Table 1 presents PCK@0.1 results for each SPAIR-71K category, showing that our approach outperforms the state of the art in 13 of 18 cate + +Table 2. PCK scores for different levels averaged over all categories in the SPAIR-71K dataset. $\mathbf{U} =$ unsupervised, $\mathbf{S} =$ supervised. + +
DatasetMethodPCK\( \mathbf{PCK}^{\dagger} \)KAP
\( \alpha_{1} \)\( \alpha_{2} \)\( \alpha_{3} \)\( \alpha_{1} \)\( \alpha_{2} \)\( \alpha_{3} \)\( \alpha_{1} \)\( \alpha_{2} \)\( \alpha_{3} \)
SPair-71kGeoAware85.675.322.080.773.222.775.366.348.9
Ours88.977.515.880.572.515.476.967.346.6
PF-PascalGeoAware89.684.872.083.279.769.486.581.770.1
Ours86.678.561.474.169.957.183.273.957.4
AP-10kGeoAware92.883.923.687.781.223.479.666.748.0
Ours89.168.17.382.364.97.277.458.840.4
+ +Table 3. Comparison of PCK, PCK†, and KAP across datasets. We evaluated AP-10K on a small subset {antelope, giraffe, elephant, hippo} and used features fine-tuned on SPAIR-71K for PF-Pascal. $\alpha_{1,2,3}$ are [0.15, 0.1, 0.05] on PF-PASCAL and [0.1, 0.05, 0.01] otherwise. + +gories, with comparable results in the remaining five. We achieve substantial improvements, including over 10 percentage points in Bottle, Chair, and TV categories, along with strong performance in other rigid object categories such as Airplane, Bike, and Plant. Our method demonstrates that even with simplistic representations, one can still achieve meaningful improvements. However, further enhancement of object-class representation learning is needed to improve performance on animals, where our method currently mainly benefits from feature averaging and certainty estimation. Overall, we see a $3.3\%$ increase (from $85.6\%$ to $88.9\%$ ), closing the gap to $100\%$ by $22.9\%$ . + +The overall PCK scores at different levels (0.01, 0.05, and 0.1) are displayed in Tab. 2. Our method shows a notable improvement of $2.2\%$ (from $75.3\%$ to $77.5\%$ ) over GeoAware [32] at the 0.05 level, though the gains are less pronounced than at the 0.1 level, with a slight drop in performance at the 0.01 level. This reduction may stem from the spatial bias in our approach, which could impact precision. We believe that refining model representations will help address this limitation and improve performance across all levels. + +Table 3 includes additional metrics and results on SPAIR-71K, AP-10K [30], and PF-PASCAL [9, 10] datasets, following GeoAware's evaluation protocol. Notably, the AP-10K test set does not include multi-instance + +cases. Our method does not perform as well here for several reasons. First, the spatial bias introduced through our method may negatively impact precision, as seen in results for $\mathrm{PCK}^{\dagger}$ and KAP. Second, as we discuss later, our representations are not yet effective in capturing exact joint limits and symmetries, leading to weaker performance on non-rigid objects, explaining the results on AP-10K. However, we anticipate that joint-angle aware representations will improve performance, particularly in challenging cases like crossed legs in animals. Lastly, PF-PASCAL favors traditional architectures due to pose similarities in source and target image, and test set images overlap with the training set. Our method excels in more realistic setups with high view variation. + +Qualitative Analysis. Figure 6 illustrates sparse matches predicted by DINO+SD [31], GeoAware [32], and our method for two image pairs $(s_{\mathrm{src}}^{1}, s_{\mathrm{tgt}}^{1})$ at the top and $(s_{\mathrm{src}}^{2}, s_{\mathrm{tgt}}^{2})$ at the bottom, each presenting different challenges. In the first image pair, the chair is rotated by almost 180 degrees between $s_{\mathrm{src}}^{1}$ and $s_{\mathrm{tgt}}^{1}$ , and in $s_{\mathrm{tgt}}^{2}$ of the second image pair, the airplane's front wheel is aligned with the left wheel. GeoAware [32] demonstrates that fine-tuning the backbone features of DINO+SD [31], which solely relies on combining unrefined DinoV2 [25] and Stable Diffusion features [26], can lead to significant better results. However, it still struggles with extreme view variations, likely requiring substantial more data to learn robust features. In contrast, our method seems data-efficient, relying only on locally semantic information, such as "wing-end", rather than "left wing end". Overall, qualitatively, our method shows a clear advantage in handling these challenging view changes. + +Representation Analysis. As depicted in Fig. 7, we qualitatively examine the behavior of our 3D object-class representations by distorting a keypoint of the airplane model while keeping other points fixed and observe the resulting deformation. When the airplane is stretched longitudinally, it scales as expected. However, when the tip of the left wing is moved laterally outwards, we expect the right wing to move in the opposite direction to maintain symmetry. Instead, the right wing stays in place. Symmetry would lead to distributions over ratios with very little variance, but due to the noisy nature of the depth estimates, the variance is naturally larger. Therefore, there may not be enough force to push the end of the right wing outward. This suggests that there is still much room for improvement in learning object-class representations. We believe it should be possible to obtain high-quality object-class representations even with very little data available, similar to techniques like NERF [21] for learning scene representations. + +
AblationBottleChairTV
DinoV2+SD backbone features60.268.764.3
Without l'sparse79.184.292.1
Without l'dense64.063.085.9
Without lffective80.583.594.8
Without l_depth81.485.194.3
Original82.288.396.1
+ +Table 4. PCK@0.1 scores of bottle, chair, and TV for various ablations. + +Failure Cases. Figure 8 depicts a failure case for the sheep category. The main cause of failure is misalignment between the object instance and the 3D representation. We expect that this problem can be solved by improving 3D object-class representations. Currently, if the alignment error is generally large for a specific category, we choose a large $\sigma$ for $p_{\mathrm{spatial}}$ in Eq. (14) to mitigate this issue. + +Ablation Studies In our ablation studies, summarized in Tab. 4, we investigate the influence of the backbone features and the loss terms. We present the ablation results for the bottle, chair, and TV category on which the ablation effects are most pronounced, as we obtained the most performance gains in these categories. Evidently, switching from the fine-tuned backbone features [32] to the unrefined DINoV2+SD [31] features significantly decreases performance. Notably, however, we still perform better than any other unsupervised method. It can also be seen that both $\mathcal{L}_{\mathrm{reconstruct}}^{\mathrm{sparse}}$ and $\mathcal{L}_{\mathrm{reconstruct}}^{\mathrm{dense}}$ are crucial for optimal performance. Generally, $\mathcal{L}_{\mathrm{reconstruct}}^{\mathrm{dense}}$ seems more important than $\mathcal{L}_{\mathrm{reconstruct}}^{\mathrm{sparse}}$ , underlying the significance of dense representations. Unsurprisingly, $\mathcal{L}_{\mathrm{background}}$ and $\mathcal{L}_{\mathrm{depth}}$ are not as critical as $\mathcal{L}_{\mathrm{reconstruct}}$ but can still be helpful to get the best performance. Lastly, we did not hold out $\mathcal{L}_{\mathrm{geom}}$ because this loss term is fundamental to our approach. + +# 4.3. Discussion + +Runtime. Since we have to solve an optimization problem each time we process an image, our method is relatively slow compared to previous work. We found that it can take up to 10-30 seconds on a NVIDIA-3090 GPU to process a single image depending on the number of initial guesses and the number of points in $C_{\mathrm{sparse}}$ and $C_{\mathrm{dense}}$ . That being said, there is likely a lot of potential for runtime optimization. However, in this work we did not focus on runtime because we are mainly interested in using our method for high-level planning for robotics, in which case we believe it is not prohibitive when the robot thinks for a few more seconds before executing a command. Lastly, we think it is exciting to explore trading off inference time for better results with less data. + +![](images/ccb3433a3ae1bcb8e760d791a56a1f0375d5c5883bfca488615925577fcdea3e.jpg) + +![](images/0110348646d38be42061456d69fb65580ddb01a6ff7c8ee64340362666ad0576.jpg) + +![](images/54eaef986ed0a11646fc53c53a5831caf1a7f3910094581cb221efe26ec993a8.jpg) + +![](images/6cb22caf055955e5a8687cef73d4c10544a3f73e93a2e03ce3de8fb6b0e33bc2.jpg) +DINO+SD + +![](images/332907780d406abdd14b1f66b7059598a44214e3c31b0ecaaf010b79e696cff6.jpg) +GeoAware + +![](images/a4a711042fbe91358412944e2a73d966cb3f7c1786a92cdb345b32ce3751798e.jpg) +Ours + +![](images/0fb28ce101b25b64646b108bfa9fe23806b046bfe1a8b07df218faa8b11a952e.jpg) +Longitudinal distortion + +![](images/b47cdbf2be14f90e011b7a4d05429b37e700a554a43e01a8276aba12b99016f3.jpg) +Lateral distortion + +Limitations and Future Work. Our method is mainly limited by the expressiveness of the 3D object-class representations. However, our hypothesis is that learning high-quality 3D object-class representations and then aligning them at inference time is significantly more data efficient and robust than purely relying on features built at inference time. This work showed that substantial performance gains can be achieved even with very simplistic 3D object-class representations, supporting our hypothesis. + +Another limitation of our method is that it currently does not support end-to-end learning as representation learning and alignment are two separated processes at the moment. However, it would be interesting to explore making the pipeline fully differentiable and adapting it to the unsupervised setting. + +![](images/63fed04ff814e1696e3eeeff0ec0d9a8acfbe10fbfb8b4b39fbaa0ae73967531.jpg) +Figure 6. Qualitative comparison for sparse correspondence between DINO+SD [31], GeoAware (AP-10K P.T.) [32], and our method. Green lines (o-o) denote correct matches and red lines depict wrong matches (x-x). +Figure 7. Behavior analysis of learned 3D object-class representation under external distortion. Black dots represent fixed points and the diamonds represent the target positions of fixed points that are being distorted. +Figure 8. Failure Example for Sheep. Round markers represent ground truth keypoints. Square markers represent the keypoints in $C_{\text{sparse}}^*$ projected onto the image plane. Keypoints are sometimes mislabeled (red) when the semantic information is too inaccurate for the spatial likelihood to correct. Similarly, the spatial likelihood can also be inaccurate due to improper alignment. + +# 5. Conclusion + +We introduce SEMALIGN3D, a novel approach for achieving robust semantic correspondence by aligning 3D object-class representations with RGB images. Leveraging monocular depth estimates and features from large vision models, we construct 3D representations that capture the semantics and geometric relations within an object class. Our method, demonstrated on the SPAIR-71K dataset, outperforms existing techniques by significantly enhancing matching accuracy, particularly on rigid object categories with extreme viewpoint variations. These results underscore the potential of our approach for data-efficient and robust semantic correspondence and we hope inspires future research on learning 3D object-class representations and alignment. + +# References + +[1] Mehmet Aygün and Oisin Mac Aodha. Demystifying unsupervised semantic correspondence estimation, 2022. 5 +[2] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second, 2024. 2 +[3] Seokju Cho, Sunghwan Hong, and Seungryong Kim. Cats++: Boosting cost aggregation with convolutions and transformers, 2022. 2, 6 +[4] Peter Florence, Lucas Manuelli, and Russ Tedrake. Self-supervised correspondence in visuomotor policy learning, 2019. 1 +[5] Peter R. Florence, Lucas Manuelli, and Russ Tedrake. Dense object nets: Learning dense visual object descriptors by and for robotic manipulation, 2018. +[6] Wei Gao and Russ Tedrake. kpm 2.0: Feedback control for category-level robotic manipulation, 2021. 1 +[7] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing, 2023. 1 +[8] Kamal Gupta, Varun Jampani, Carlos Esteves, Abhinav Shrivastava, Ameesh Makadia, Noah Snively, and Abhishek Kar. ASIC: Aligning sparse in-the-wild image collections, 2023. 1, 2, 6 +[9] Bumsub Ham, Minsu Cho, Cordelia Schmid, and Jean Ponce. Proposal flow, 2016. 5, 6 +[10] Bumsub Ham, Minsu Cho, Cordelia Schmid, and Jean Ponce. Proposal flow: Semantic correspondences from object proposals, 2017. 5, 6 +[11] Shuaiyi Huang, Luyu Yang, Bo He, Songyang Zhang, Xuming He, and Abhinav Shrivastava. Learning semantic correspondence with sparse annotations, 2022. 2, 6 +[12] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 4 +[14] Junsoo Lee, Eungyeup Kim, Yunsung Lee, Dongjun Kim, Jaehyuk Chang, and Jaegul Choo. Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence, 2020. 1 +[15] Jae Yong Lee, Joseph DeGol, Victor Fragoso, and Sudipta N. Sinha. Patchmatch-based neighborhood consensus for semantic correspondence. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13148-13158, 2021. 2, 6 +[16] Xinghui Li, Jingyi Lu, Kai Han, and Victor Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching, 2024. 2 +[17] Yanbin Liu, Linchao Zhu, Makoto Yamada, and Yi Yang. Semantic correspondence as an optimal transport problem. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4463-4472, 2020. 2, 6 +[18] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence, 2024. 2, 6 +[19] Lucas Manuelli, Wei Gao, Peter Florence, and Russ Tedrake. kpm: Keypoint affordances for category-level robotic manipulation, 2019. 1 +[20] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps, 2024. 1, 2, 5, 6 +[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 7 +[22] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence, 2019. 2, 3, 5, 1 +[23] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag-style manipulation on diffusion models, 2023. 1 +[24] Dolev Ofri-Amar, Michal Geyer, Yoni Kasten, and Tali Dekel. Neural congealing: Aligning images to a joint semantic atlas, 2023. 1, 2 +[25] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 1, 2, 6, 7 +[26] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion, 2023. 2, 6, 7 +[27] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data, 2024. 2, 3 +[28] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12):2878-2890, 2013. 5 +[29] Lin Yen-Chen, Pete Florence, Jonathan T. Barron, Tsung-Yi Lin, Alberto Rodriguez, and Phillip Isola. Nerf-supervision: Learning dense object descriptors from neural radiance fields, 2022. 1 +[30] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. Ap-10k: A benchmark for animal pose estimation in the wild, 2021. 5, 6 +[31] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence, 2023. 1, 2, 6, 7, 8, 3 + +[32] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence, 2024. 1, 2, 3, 4, 6, 7, 8 +[33] Tinghui Zhou, Philipp Krahenbihl, Mathieu Aubry, Qixing Huang, and Alexei A. Efros. Learning dense correspondence via 3d-guided cycle consistency, 2016. 2 + +Acknowledgment. This work was supported in part by the Natural Science Foundation of China under Grants 62225309, U24A20278, 62361166632, and U21A20480. \ No newline at end of file diff --git a/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/images.zip b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..176a4531b425e26dd718e39bee0e67fa0ce9829c --- /dev/null +++ b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd78b8ff68a29588d4141c4badd74b7800db6cc66ff1b5082f5ea068875dcb15 +size 712759 diff --git a/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/layout.json b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ebd9db1e485c92ba12a0c66d121a166d4bce1514 --- /dev/null +++ b/CVPR/2025/SemAlign3D_ Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62828919d64265f3725b2c6bb29e5c2a2d7ea67e587cb827aebe22246a122ffd +size 468133 diff --git a/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/9a2d022e-5c87-49e4-b542-4c957d3bcf16_content_list.json b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/9a2d022e-5c87-49e4-b542-4c957d3bcf16_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7a1d45db09478979ef1793a203adf30352a840ea --- /dev/null +++ b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/9a2d022e-5c87-49e4-b542-4c957d3bcf16_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e61a537088f21be8543210da891881ab29c8d755ae86ee56449247d71193102 +size 86587 diff --git a/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/9a2d022e-5c87-49e4-b542-4c957d3bcf16_model.json b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/9a2d022e-5c87-49e4-b542-4c957d3bcf16_model.json new file mode 100644 index 0000000000000000000000000000000000000000..516cf8b63afcfb72e294cd60d1f29f808fdc3d70 --- /dev/null +++ b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/9a2d022e-5c87-49e4-b542-4c957d3bcf16_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea0efd7ef113bf7997718686d418befbf4b51ede30fd7760f4791a594903cc47 +size 107979 diff --git a/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/9a2d022e-5c87-49e4-b542-4c957d3bcf16_origin.pdf b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/9a2d022e-5c87-49e4-b542-4c957d3bcf16_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..67639185e367a8451c0953a09e2cf91aebf3887a --- /dev/null +++ b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/9a2d022e-5c87-49e4-b542-4c957d3bcf16_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b95f3a149a4344474258f146021fc5e7d1f6e7f21659015c4622f80826df114 +size 7804686 diff --git a/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/full.md b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f75f975e8f7c226b1b8fca44ccc6be10dfcabe76 --- /dev/null +++ b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/full.md @@ -0,0 +1,373 @@ +# SemGeoMo: Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance + +Peishan Cong $^{1,2,*}$ , Ziyi Wang $^{1,*}$ , Yuexin Ma $^{1,\dagger}$ , Xiangyu Yue $^{2,\dagger}$ $^{1}$ ShanghaiTech University $^{2}$ The Chinese University of Hong Kong {congpsh, wangzy17, mayuxin}@shanghaitech.edu.cn + +# Abstract + +Generating reasonable and high-quality human interactive motions in a given dynamic environment is crucial for understanding, modeling, transferring, and applying human behaviors to both virtual and physical robots. In this paper, we introduce an effective method, SemGeoMo, for dynamic contextual human motion generation, which fully leverages the text-affordance-joint multi-level semantic and geometric guidance in the generation process, improving the semantic rationality and geometric correctness of generative motions. Our method achieves state-of-the-art performance on three datasets and demonstrates superior generalization capability for diverse interaction scenarios. The project page and code can be found at https://4dvlab.github.io/project_page/semgeomo/. + +# 1. Introduction + +Dynamic contextual human motion generation [19, 21, 22] aims to generate human interaction motions that are both commonsense and geometrically accurate, adapting seamlessly to real dynamic environments. Its core lies in constructing an interaction-oriented world model for humans, enabling reasonable adaptation to changes of interactive objects or people. This interaction-oriented world model can support a wide range of applications, including humanrobot interaction, closed-loop simulators, intelligent sports coaching, and immersive VR/AR gaming experiences. + +As the importance of interaction becomes increasingly recognized, some studies have evolved from text-driven human motion generation [4, 6, 16, 31, 32, 49] to text-driven joint generation of human-object or human-human interactions [2, 22, 23, 37, 42]. However, generating motions + +![](images/c979b0c41f743fa8ea4accb364302b78e2256abd73f32b6f3579f282fedf1646.jpg) +Figure 1. Given sequential point clouds of interactive targets, Sem-GeoMo generates realistic and high-quality human interactive motions along with corresponding textual descriptions. By leveraging both semantic and geometric guidance, our method ensures the semantic coherence and geometric accuracy of the generated results. jointly for both the human and the interactive target creates an excessively large search space, often leading to suboptimal generation results. Additionally, the lack of fine-grained control over the generated data hinders the creation of personalized interactive motions, limiting its applicability to real-world scenarios such as humanoid operation and human-robot interaction. + +Recently, contextual human motion generation has garnered increasing attention for its ability to produce controllable interactions based on specified scenario conditions. However, many works [3, 14, 35, 38, 40] only focus on generating coarse-grained human trajectories and motions in static environments with fixed furniture or layouts + +rather than the fine-grained details of interactions, limiting adaptability to dynamic conditions. A few recent studies [5, 22, 43] have begun to explore dynamic contextual human motion generation. However, these approaches have notable limitations: 1) they lack textual guidance, which undermines the semantic coherence of interactions and limits the generalization capability of the approach, and 2) they fail to incorporate fine-grained geometric representations, resulting in insufficient constraints on the geometric accuracy of generated interaction motions. + +In this work, we propose a novel dynamic contextual human motion generation method, named SemGeoMo, illustrated in Fig.1, which could generate reasonable and high-quality interactive motions by comprehensively integrating semantic information from textual descriptions with hierarchical geometric features extracted from interactive objects. The first challenge is how to construct the semantic guidance. Given that large language models (LLMs) possess general knowledge and can provide rich information on the attributes of interacting objects as well as guidance for the interaction process, we introduce an automated interaction text annotator. By leveraging careful prompt design and fine-tuning of the model [30, 33], our LLM Annotator eliminates the need for manual text labeling, offering strong support for the semantic coherence and generalization of generated human interaction behaviors. The second challenge is how to construct the geometric guidance. To ensure the geometric accuracy of interactive motions, such as avoiding geometric penetration and ensuring appropriate contact between human and object, we propose a two-stage framework that decouples contact geometry generation from interactive motion generation. In the first stage, SemGeo Hierarchical Guidance Generation, a diffusion model generates affordance-level and joint-level interaction cues guided by semantic information to capture both coarse and precise geometric positioning. In the second stage, SemGeo-guided Motion Generation, these cues are effectively utilized to guide the generation of detailed human motions, improving both semantic plausibility and geometric accuracy. It is worth noting that our model simultaneously generates human motions and language descriptions at varying levels of granularity, which not only improves the quality of the generated motions but also enhances the interpretability and comprehensibility of the interactions. + +Extensive experiments demonstrate that we achieve state-of-the-art performance on three human-object interaction datasets. Moreover, we demonstrate the generalization capability of our method in more challenging scenarios, including interactions with unseen objects, human-human interactions, and interactions with deformable objects. To summarize, our work makes the following contributions: + +- We propose a novel method that generates responsive human motions and corresponding textual descriptions + +based on observed dynamic interactive targets. + +- We introduce an automated text annotator that interprets reactions during interactions, reducing the burden of manual labeling and enhancing the generalization capability. +- Our method fully utilizes multi-level semantic and geometric guidance—including text, affordance, and joint-level cues—throughout the generation process, improving both the semantic rationality and geometric accuracy of interactive motion. +- Our method generates high-quality human motions and achieves state-of-the-art performance across three benchmarks and an unseen dataset. + +# 2. Related Work + +# 2.1. Text-guided Human Motion Generation + +Text-conditioned human motion generation [6, 7, 15, 31, 32, 34, 49, 50] have made significant progress with the rise of diffusion models [11, 12]. With the emergence of human-object interaction datasets [2, 22, 47, 51] and human-human interaction datasets [23, 42], some works have begun to explore text-driven interaction motion generation. Several studies [8, 27, 39, 44] dive into jointly generating a sequence of human and object poses based on textual conditions. HOI-Diff [27] emphasizes the significance of affordance information and Thor [39] refines object rotation during each inverse diffusion step for human-object interaction. Furthermore, CHOIS [21] further integrates the 2D waypoints with object geometry loss during sampling process. Other studies [23, 42] focus on human-human interaction, where two human motions are jointly generated based on language descriptions. However, the joint generation of human motions and interactive targets creates a large optimization search space during training, leading to lower quality in the generated motions. Moreover, conditioning solely on text lacks fine-grained control over the generation process, limiting its applicability in areas such as robotic operations, human-robot interaction, and AR/VR immersive experiences. + +# 2.2. Contextual Human Motion Generation + +Contextual human motion generation explores settings that are more applicable to real-world scenarios, where human motion is influenced by and interacts with the given environment. Several works [18, 20] focus on interactions with seating furniture [17] and others [3, 14, 35, 38] generate natural human motions in 3D indoor scenes [1, 9, 19, 36, 40, 45]. SceneDiff [14] utilizes point clouds as conditions to generate feasible interactions. AffordMotion [38] provides a two-stage framework that employs a scene affordance map as an intermediary. However, these works are limited to static environments, where interacted objects are often restricted to fixed furniture like beds and chairs, and + +![](images/4b76c7d2eaa677ea11a66bedc1ba9da8a6cc65d60499edfc5fbf34c6d0f7528a.jpg) +Figure 2. The pipeline of our two-stage framework. LLM Annotator provides the semantic guidance. SemGeo Hierarchical Guidance Generation takes textual information and sequential point cloud as condition and generate affordance-level and joint-level guidance. Then SemGeo-guided Motion Generation utilizes semantic and geometric information to generate responsive human motion. + +the category of motions is limited to sitting, lying, and walking. They focus on the trajectory and the goal state, rather than interacting with dynamic, ever-changing targets. + +Following works [5, 22, 43] introduce dynamic, interactive targets, such as movable objects or other people. ReGenNet [43] generates human reactions conditioned on given human motion in SMPL [25] representation and the action condition. However, SMPL representation requires additional processing on raw sensor data and is not suitable for all dynamic targets such as objects. A more flexible point cloud-based representation is used in [5] for interacting scenes or objects. However, it lacks a carefully designed feature modeling method for geometric and temporal information, resulting in suboptimal performance. OMOMO [22] presents a framework for generating human behaviors based on object motions, utilizing conditional diffusion to generate hand joint positions as extra guidance. While this work lacks the integration of textual information, which can provide important semantic guidance. By contrast, our method fully leverages textual reasoning and incorporates hierarchical semantic and geometric features to enhance contextual motion generation. + +# 3. Method + +Our goal is to generate responsive human motions conditioned solely on the interactive target, represented as a 4D sequential point cloud. To achieve this, we introduce the + +LLM Annotator to provide semantic features, which are then processed in the following two stages: SemGeo Hierarchical Guidance Generation and SemGeo-guided Motion Generation. In the first stage, multi-level geometric information is generated in a coarse-to-fine manner. These generated guidance signals are then used in the second stage to guide the motion generation process, enhancing both semantic plausibility and geometric accuracy. The overall architecture is shown in Fig. 2. + +# 3.1. Data Representation + +Motion Representations. We denote the human motion as $\mathbf{x}^h\in \mathbb{R}^{L\times D}$ , where $L$ represents the number of frames in the sequence and $D$ is the feature dimension, which includes the pelvis velocity, local joint positions, rotations and velocities of other joints in the pelvis space, as well as binary foot-ground contact labels, following the representation in MDM [32]. + +Point Cloud Representations. To extract geometric features, point clouds are well-suited for capturing the dynamic changes of objects, providing a unified input modality for representing deformable or non-rigid interactive targets without the need for additional preprocessing. The sequential point cloud is down-sampled to $N = 1024$ points per frame, denoted as $\mathbf{P} \in \mathbb{R}^{L \times N \times 3}$ . We then adopt Basis Point Set [28] (BPS) representation, following the approach in [22], to encode the object geometry. The process begins by uniform sampling the basis points from a unit ball with + +![](images/3f9b3473455845e9f5f0b7f05ebe592075d1a14b3e8eab5e7a4bbddfd5714237.jpg) +Sequential Point Clouds + +Coarse Text Description + +A person squats and lifts a monitor, and put down. + +![](images/198a8d3b0f34be4ec062eabedca0862ce98ad88759c3d5ef2a34a541136b8785.jpg) +Geometry feature + +![](images/27b703b93ae0ba6426c05a65cddb06237d65da6e41baaf81d43beb7478e5d2ff.jpg) +LoRA +Figure 3. LLM Annotator pipeline. It first takes a sequential point cloud to infer a coarse text description. Then uses joint positions, the coarse text, and geometric features to generate a fine-grained sentence with a designed prompt. + +nd Joint Positions + +Coarse Text + +Extract hand direction. + +- Infer detailed movements at each time step. +[start of rules] +1.Divide the total movement into three phases,... +2. Includes key actions,... +3. Infer the arm and legs movement [end of rules] + +Hand-object interaction direction left, right, top, bottom,... + +![](images/be5916b7d5f6b6ba6b48552309305a4c9abd709ddefcf47ab0ee24097acb038c.jpg) + +Fine-grained Text Description A person squats, touches the monitor with both hand contact the bottom of the monitor, then ... + +![](images/7390821950d99aadb0c9d4105725d129be5c280b40f413957aebb6bab455fa2e.jpg) + +Give me a fine-grained sentence. + +a radius of 1 meter, followed by calculating the minimum Euclidean distance from the basis points to their nearest neighbors. Finally, these distances are concatenated with the center position of the original point cloud. The resulting point cloud representation for each frame is denoted as $\mathbf{p} \in \mathbb{R}^{N \times 3 + 3}$ . An MLP is then applied to project the BPS representation into a lower-dimensional space, yielding the sequential geometry features $\mathbf{F}_{\mathbf{pc}} \in \mathbb{R}^{L \times 256}$ . + +Affordance Map Representations. The affordance map serves as an intermediate representation, providing crucial geometric information regarding which parts of the target are most likely to come into contact during the interaction. For each frame, we calculate the $\ell_2$ distance between each point and the skeleton joints, resulting in a per-frame distance map $d\in \mathbb{R}^{L\times N\times J}$ , where $J$ is the number of skeleton joints. We then transform this distance field into a normalized distance map, Affordance, denoted as $\mathbf{A}$ , which encodes the spatial relationship between the target and the interactive human. The affordance map is computed as: + +$$ +\mathbf {A f f o r d a n c e} (n, j) = \exp \left(- \frac {1}{2} \frac {\mathbf {d} (n , j)}{\sigma^ {2}}\right), \tag {1} +$$ + +where $\sigma$ is the normalizing factor. + +# 3.2. LLM Annotator + +Textual descriptions provide essential semantic information. Given the interactive targets, envisioning how to interact with them is crucial for generating realistic human motion. Previous work [31, 32, 49] on text-to-motion has also validated the effectiveness of feature mapping from textual descriptions to motion generation. + +To obtain textual guidance, we employ a Large Language Model (LLM) as an annotator in the initial stage to generate a coarse motion description. The whole pipeline is shown in Fig. 3. From the sequential point cloud, we extract the bounding box of the interactive target along with its movement trajectory. Utilizing a predefined list of actions and categories, we derive initial language-based guidance. Specifically, we utilize the pre-trained LLaMA model + +[33], enhanced with LoRA [13] finetuning on given text- interaction pairs from [22]. + +More detailed, context-aware cues can capture the dynamic changes in contact, guiding the model to generate more precise body movements. Leveraging the strong reasoning capabilities of LLMs, we design a coarse-to-fine automated language guidance annotation system. After generating initial textual guidance and predicting hand joint positions, we determine sequential contact information by calculating distances between predicted joint positions and interactive targets, inferring where the body should make contact with each part of the targets. The LLM then organizes this contact information into language descriptors (e.g., "the left-hand contacts the lower left of the box"). Simultaneously, we let the LLM divide the interaction into three steps, with the textual descriptions reflecting changes in contact at each step, with also deducing finer body movements of the arms and legs. + +Thus, the generated textual descriptions assist in the motion generation task by providing semantic information. On the other hand, our entire pipeline also enables reasoning capabilities, allowing for a more comprehensive understanding and generation of human motion. + +# 3.3. SemGeo Hierarchical Guidance Generation + +The joint position $\mathbf{J}_{\mathbf{h}} \in \mathcal{R}^{L \times J \times 3}$ provides precise spatial information, and the affordance map $\mathbf{A}$ offers coarse geometric clues. These two types of features are crucial for modeling interactions in a coarse-to-fine manner, thus we introduce a conditional diffusion model with dual-branch transformer to jointly generate contact information with capturing their mutual influence in the first stage. + +We utilize CLIP[29] as text encoder to obtain the text feature $F_{text}$ . At each step in the diffusion process, the model takes the text feature $F_{text}$ , point cloud feature $F_{pc}$ and noisy signals $x_{t}^{J}$ and $x_{t}^{A}$ as input and predict clean $x_{0}^{J}$ and $x_{0}^{A}$ . For JointTransformer, the inputs are concatenated together and then fed into the multi-head self + +attention blocks followed with position-wise feedforward layer. For AffordanceTransformer, inspired by [38], the affordance map is more closely related to the point cloud geometry. Therefore, we encode the point cloud feature $F_{pc}$ along with $x_{t}^{A}$ , which act as the key and value in attention module. The concatenation of the language feature and diffusion step embeddings serves as the query. After passing through the cross-attention mechanism and multiple self-attention blocks, the refined point features are obtained. To enhance the coarse to fine interactions and further refine the joint position, we introduce a mutual cross-attention mechanism. This takes the output from JointTransformer as query, output from AffordanceTransformer as key and value, which updates the hand joint position. + +The conditional diffusion model learns the reverse diffusion process to generate clean data from a Gaussian noise $x_{t}$ over $T$ consecutive denoising steps. Specifically, we use $c$ to represent the conditions, and the reverse diffusion process is modeled as: + +$$ +p _ {\theta} \left(x ^ {t - 1} \mid x ^ {t}, \boldsymbol {c}\right) := \mathcal {N} \left(x ^ {t - 1}; \boldsymbol {\mu} _ {\theta} \left(x ^ {t}, t, \boldsymbol {c}\right), \sigma_ {t} ^ {2} I\right). \tag {2} +$$ + +Finally, our model directly estimates the input signal. The training process optimizes from the reconstruction loss: + +$$ +\mathcal {L} = \mathbb {E} _ {\boldsymbol {x} ^ {0}, t} \left\| \hat {x} _ {\theta} \left(x ^ {t}, t, \boldsymbol {c}\right) - x ^ {0} \right\| _ {1}. \tag {3} +$$ + +# 3.4. SemGeo-guided Motion Generation + +In the second stage, a denoising network architecture is used to generate full-body motions based on the predicted joint positions $\mathbf{J}_h^{\prime}$ , affordance map $\mathbf{A}^{\prime}$ , and text descriptions. + +SemGeo Condition Module To effectively process and integrate these diverse features, we design SemGeo Condition Module to encode the combined features. After obtaining both coarse and fine-grained text descriptions, we use a text encoder to extract the semantic information. Since the language generated by the Fine-grained LLM Annotator contains phase-specific details and longer descriptions, we adopt LONGCLIP [46], which can better capture and represent fine-grained attributes without the length limitations inherent in CLIP. The extracted feature is $F_{\mathrm{text}}^{\prime}$ . + +To extract features from the affordance map, we first concatenate the point cloud feature $F_{pc}$ with the affordance map $\mathbf{A}'$ , and then pass the combined input through a 3-layer MLP. The temporal transformer is applied to extract latent features $F$ over time, which helps to capture both spatial and temporal dependencies. The operation is formalized as: + +$$ +F = \text {T e m p o r a l T r a n s f o r m e r} \left(M L P \left(F _ {p c} \oplus \mathbf {A} ^ {\prime}\right), \right. \tag {4} +$$ + +where $\oplus$ denotes concatenation. + +We apply mutual cross-attention to extract the mutual features between joint positions and the affordance map. The joint feature, after being mapped to a higher dimensional space with MLP, serves as the query. The latent feature from the previous step is used as key and value in the + +attention mechanism. This interaction between joint positions and the affordance map enables the model to capture the relationship between joint movements and object geometry, improving motion prediction: + +$$ +F _ {f u i s o n} = \text {C r o s s A t t e n t i o n} \left(M L P \left(\mathbf {J} _ {h} ^ {\prime}\right) _ {q}, F _ {k}, F _ {v}\right). \tag {5} +$$ + +The final condition $c$ is the concatenation result of $F_{text}$ , $F_{text}^{\prime}$ , $F_{fuison}$ . + +This comprehensive feature representation ensures that the model has a rich, multi-dimensional understanding, capturing both coarse-to-fine semantic information from the text and spatial-temporal geometric information from the joint positions and affordances. + +Motion ControlNet Inspired by [37, 41], we introduce Motion ControlNet to generate high-fidelity motions conditioned on $c$ . With MDM frozen during training, each transformer encoder layer in ControlNet [48] is linked to its MDM counterpart via a zero-initialized linear layer. + +Loss Guidance To refine our generated interactions, we employ joint guidance and foot guidance during sampling with classifier guidance. The joint-based guidance loss aligns the generated global joint positions $\mathbf{J}_{\mathrm{pred}}$ from the second stage with the target control joint positions $\mathbf{J}_{\mathrm{h}}^{\prime}$ from the first stage, ensuring consistency in the generated motion across both stages. Constraints are applied to the joints that are in contact with objects when the distance between the joint and the nearest point on the object is below a predefined threshold $\tau$ . Specifically, the mask is defined as $\mathrm{Mask} = \mathrm{Dis}(\mathbf{J}_{\mathrm{h}}, V) \leq \tau$ . The joint guidance function is then defined as: + +$$ +L _ {\text {j o i n t}} = \frac {1}{J} \sum_ {i = 1} ^ {L} \left| \mathbf {J} _ {\text {p r e d} _ {i}} - \mathbf {J} _ {\mathrm {h} _ {i}} ^ {\prime} \right| _ {2} \cdot \operatorname {M a s k} _ {i}. \tag {6} +$$ + +The foot-stability loss $L_{\mathrm{foot}}$ is designed to ensure that the foot stays near the ground and penalize sudden changes in velocity to eliminate foot sliding: + +$$ +L _ {\mathrm {f o o t}} = \frac {1}{L} \sum_ {i = 1} ^ {L} \left((y _ {i} - h _ {g}) ^ {2} + \alpha \mathbf {M} _ {c} (v _ {i} ^ {2}) + \beta \mathbf {M} _ {c} (a _ {i} ^ {2})\right), +$$ + +where $y_{i} = \min (h_{l,i},h_{r,i})$ is the height of the lower foot for each frame and $h_g$ is the empirical values indicating contact with the ground. $v_{i} = \| p_{i + 1} - p_{i}\|$ is the foot velocity at frame $i$ , calculated from the foot position $p$ and $a_{i} = \| v_{i + 1} - v_{i}\|$ is the foot acceleration. $\mathbf{M}_c$ represents the mask for contact with the floor. $\alpha$ and $\beta$ are hyperparameters. Inspired by [37], we employ L-BFGS for several iterations at each denoising step to update the posterior mean. + +# 4. Experiment + +# 4.1. Datasets + +FullBodyManipulation [22] takes a total duration of approximately 10 hours. It provides, paired object and human + +A person pull the floorlamp, and set it back down. + +![](images/acf92c265eec703603bc3cae77ff476dc3fa0a12807b72d26170193ccc4a08d6.jpg) + +![](images/02f00bb1fa1f1b64a58bab9b6702b9d3b8ee1e1e94a115640a9e317321ba43a8.jpg) + +![](images/a5573e479818aa24548b6eeb405b52b9de30098f8c0b7dfa29fb735f2767e1c1.jpg) + +![](images/5e3440364eb296ffdb49e987a7eb95f08e89f3cd4fd448d4bd84ed9421656c1d.jpg) + +A person lift the listable above your head, walk, and put the listable down. + +![](images/e68dc0d087954594ce5c827a624c506f94ef9d25cec99177ef592fe0f3fa349d.jpg) + +![](images/6c193d0402887c53f0318558b1da6eca1ad4b381e2151b07b4803a63868316d2.jpg) + +![](images/d2611fc60712a736304f43feeed4631e89f93767ae6d99cb5a72a5b02f9afcaa.jpg) + +![](images/0d451ef2bc107dc8a8f44a6f2a24c1329902bf779bd2474edb36bff7f6aa75fd.jpg) + +A person lift the woodchair, move the woodchair, and then place the woodchair on the floor. + +![](images/b9808cbf05834d156b60cfaecb13083cc9c95b95f3b3c3e8d709f8fb1e499ba6.jpg) +OMOMO +Figure 4. Qualitative results on the FullBodyManipulation dataset. We circle areas of low contact performance in pink and instances of contorted motion in green. + +![](images/1fa66e00516e57e67550df07f32ae5b1628705cf67bd21c80f643111a7902de7.jpg) +CHOIS + +![](images/24a446576ed4cf0c75cd41a0957526ca80b09e8c8779d13f6bb4aad6fc087bec.jpg) +MDM-PC + +![](images/8b2a60f676bb71cdec5a60fd137dc6e943e6329bb3f95e3894cd9dfe033d3355.jpg) +SemGeoMo + +Table 1. Human motion generation result on FullBodyManipulation. + +
HandJPE↓MPJPE ↓Cprec ↑Crec ↑Cacc ↑c%↑F1 ↑FID ↓R-score ↑Diversity ↑FS ↓
w/o textSceneDiff [14]95.3819.840.640.190.450.180.271.640.599.860.38
OMOMO [22]33.1818.060.770.710.740.610.751.980.388.990.50
w GT textMDM-PC [32]51.3518.250.710.410.620.330.490.650.579.520.51
CHOIS [21]31.6817.120.760.580.610.520.592.270.496.040.47
AffordMotion [38]98.6625.340.450.140.310.130.164.710.458.150.43
SemGeoMo27.8416.620.840.740.850.660.771.170.6610.150.57
w Gen textSemGeoMo30.3517.980.820.740.820.660.741.050.649.780.47
+ +motion, including interactions of 17 subjects with 15 different objects with text descriptions. We follow the official train/test split for evaluation. + +BEHAVE [2] consists of the interactions of 8 subjects with 20 different objects. We follow the official train/test split provided by BEHAVE. [27] provides the text annotations while we observe that these annotations merely combine action labels and categories of motion, offering phase-specific details. We provide a revised version of the annotations with phased interactions that more clearly describe the motions. + +$\mathbf{IMHD}^2$ [51] and HoDome [47] are challenging 3D human-object interaction datasets for motion capture, covering interactions between 10 objects and 10 subjects, and 23 diverse objects with 10 subjects, respectively. We annotate the text descriptions using our LLM annotation pipeline to facilitate further text-driven human-object interaction studies and enable comparisons with other text-guided methods. + +# 4.2. Evaluation Metrics + +We mainly follow the metrics in OMOMO [22] and MDM [32]. HandJPE and MPJPE represent mean hand joint position errors, and mean per-joint position errors, computed using the Euclidean distance between the pre + +dicted and ground truth in centimeters (cm). For measuring the interaction quality, we employ contact metrics including precision $(\mathbf{C}_{prec})$ , recall $(\mathbf{C}_{recall})$ , accuracy $(\mathbf{C}_{acc})$ and F1 score following [22]. The contact percentage ( $c\%$ ) is the proportion of frames where contact is detected. FID measures the distance of the generated motion distribution to the ground truth distribution in latent space. R-score measures the text and motion matching accuracy and Diversity measures the generation diversity. FS represents foot sliding metric and is computed following [10]. + +# 4.3. Results + +Baselines OMOMO [22] is the only work which is aligned with our setting. In addition, we adapt several related works, including SceneDiff [14], MDM-PC [32], AffordMotion [38], and CHOIS [21], to fit our problem setting. SceneDiff [14] utilizes a diffusion model conditioned on static scenes, so we modified the conditional model to accommodate sequential point cloud inputs. We modified the text-to-motion generation work MDM as MDM-PC, incorporating our sequential point cloud representation. CHOIS and AffordMotion are conditioned on both textual descriptions and scenarios. CHOIS requires a sequence of object + +Table 2. Human motion generation result on Behave. + +
HandJPE↓MPJPE ↓Cprec↑Crec↑Cacc↑c%↑F1↑FID↓R-score ↑Diversity ↑FS↓
w/o textSceneDiff [14]51.5818.250.730.380.400.320.471.690.135.320.33
OMOMO [22]45.3521.560.710.600.610.600.621.940.145.110.42
w Gen textMDM-PC [32]35.4118.610.730.480.510.470.571.520.105.450.32
CHOIS [21]36.7518.170.720.410.430.410.512.260.135.020.46
AffordMotion [38]55.6519.160.720.230.280.250.321.920.134.380.51
SemGeoMo27.9116.220.840.670.670.660.741.470.155.640.52
+ +Table 3. Human motion generation result on ${\mathrm{{IMHD}}}^{2}$ . + +
HandJPE↓MPJPE ↓Cprec ↑Crec ↑Cacc ↑c%↑F1 ↑FID ↓R-score ↑Diversity ↑FS ↓
w/o textSceneDiff [14]82.0125.080.450.180.210.160.221.890.155.220.57
OMOMO [22]39.4023.360.580.390.430.410.422.090.164.560.55
w Gen textMDM-PC [32]63.5723.810.480.240.280.300.241.730.145.230.59
CHOIS [21]44.9225.090.560.310.350.310.322.670.114.450.53
AffordMotion [38]75.3224.370.550.160.220.160.212.880.124.480.42
SemGeoMo35.4320.850.720.490.510.490.521.640.145.350.49
+ +Table 4. Human motion generation result on HoDome. + +
HandJPE↓MPJPE ↓Cprec↑Crec↑Cacc↑c%↑F1↑FID ↓R-score ↑Diversity ↑FS ↓
w/o textSceneDiff [14]107.5029.530.210.110.160.130.144.880.084.720.60
OMOMO [22]86.1227.070.420.230.310.220.255.470.104.930.31
w Gen textMDM-PC [32]95.9126.290.250.130.180.140.153.640.134.890.67
CHOIS [21]76.7424.170.550.120.260.110.186.120.114.140.28
AffordMotion [38]94.2429.310.490.110.150.100.135.290.115.140.45
SemGeoMo44.2224.280.780.470.470.450.544.290.135.220.35
+ +states in 2D waypoints, which we modify this part into 3D states. AffordMotion predicts the affordance map and subsequently generates motion in scenes, we adopted this approach for our interactive target setting. Note that for the aforementioned works, since FullBodyManipulation provide the textual description, we provided ground-truth (GT) text as input, and compared the variant of our model under the same conditions with ground-truth text as well. + +Results on the FullBodyManipulation Dataset. The results on the FullBodyManipulation dataset are illustrated in Tab. 1, and we provide a visualization comparison in Fig. 4. SceneDiff, MDM-PC, and CHOIS directly use the original point cloud as a condition in a single stage, the dynamic nature of the point cloud makes it challenging to infer low-level contact information, resulting in a significant drop in contact metrics. Even though MDM incorporates text guidance and improves performance on FID, the contact accuracy remains poor. CHOIS applies a human-object loss to refine contact, but such improvement in performance is limited. OMOMO predicts joint positions as the first-stage output, but without textual guidance, it performs poorly on FID and R-score, struggling to capture realistic interaction motions. Distortions and abnormal movements may occur, as shown in Fig. 4 circled in green. AffordMotion uses the affordance map for geometric guidance, while this approach lacks sufficient detail for fine-grained human-object interactions. In contrast, we leverage both semantic and geometric information by generating textual descriptions and predicting coarse-to-fine contact information, resulting in su + +perior performance. Compared to the variant, which utilizes ground-truth text information as a condition, our model achieves comparable performance, demonstrating the efficiency and accuracy of the language description generated by our method. + +Results on the Behave and $\mathbf{IMHD}^2$ Dataset. We further conduct experiments on the BEHAVE and $\mathrm{IMHD}^2$ datasets to validate our model's performance in Tab. 2 and Tab. 3. Notably, the $\mathrm{IMHD}^2$ dataset includes more challenging interactions, such as sports activities, which make precise contact and accurate motion generation more difficult. However, our proposed method still outperforms other approaches in these scenarios. + +Results on the HoDome Dataset. To test our model's ability to generate new scenarios, we further conduct the experiment on totally unseen objects with directly sampling on HoDome [47]. The result is illustrated in Tab. 4, our proposed method outperforms other methods. The visualization results are illustrated in Fig. 5. + +# 4.4. Ablation Studies + +Ablation studies on the impact of semantic and geometric information. To verify the effectiveness of our multilevel guidance, we conduct ablation studies to assess the impact of each component in Tab. 5. The multi-level geometric guidance and semantic information enhance human motion generation, reflected in the improvement of the contact and FID&R-score indicators, respectively. We also compare our designed SemGeo conditional module with cross + +Table 5. Ablation studies on the impact of semantic and geometric information with our model design. + +
HandJPE↓MPJPE ↓Cprec↑Crec↑Cacc↑c%↑F1↑FID ↓R-score ↑Diversity ↑FS ↓
w/o affordanc map29.8918.850.800.710.810.630.732.210.629.840.58
w/o joint31.2320.600.760.700.780.610.723.520.518.150.74
w/o attention29.1819.470.810.660.780.590.709.270.446.840.70
w/o text30.3624.440.780.620.720.540.641.780.419.400.61
w/o fine-grained text27.8416.620.840.740.850.660.771.170.6610.150.57
Full27.9717.010.840.750.860.670.771.030.6810.050.46
Full-GT6.4413.610.880.820.930.700.840.970.7010.480.58
+ +Table 6. Ablation study on guidance loss. + +
LJointLFSHandJPE↓MPJPE↓Cprec↑Crec↑Cacc↑c%↑F1↑FID↓R-score↑Diversity↑FS↓
✓27.8416.970.830.750.860.670.751.150.6510.130.61
✓85.2829.750.620.250.350.240.220.930.6610.050.31
✓✓27.8416.620.840.740.850.660.771.170.6610.150.57
+ +Behave +![](images/2d0ea9d628b3ebe17326dc9032e529dbd9c9a97f962137f70b9d97646824e88c.jpg) +A person lift the chair with both hand touches the top, then the hand keep the same position with rotating... + +IMHD +![](images/179b7f834634b6aa3345683cfae0ada63c94a8a5617ada79179c1e3c9526bc44.jpg) +A person standing, the arms lift the bat behind the head, the right hands gripping the bat on the bottom, then... + +HoDome +![](images/bd0dcb9880ceb0d0239fb7094e0ded9902cd9b3f6bfc528a11b2662b2cf7d648.jpg) +A person grad the keyboard with both hand contact on the left-bottom and right-top, then right hand ... + +![](images/3baa3eabbd1a131c52480130b93bf541e22769a52b3aacd825ae31cd4662546c.jpg) +Figure 5. Visualization on more datasets. + +![](images/a1e45323d173294c344d48af6817c6d5fb729cca331302d47cca1cffb7520cc7.jpg) +Figure 6. Extension on other scenarios. (a) is the interaction with humans and (b) is the interaction with objects of varying sizes. + +Table 7. Text generation result on FullBodyManipulation. + +
BLEU-4ROUGE -1ROUGE-2ROUGE-L
87.2190.9183.7989.39
+ +attention. The results show that cross-attention significantly improves performance by effectively sharing information between coarse-to-fine features. During implementation, we predict the hand joint and compare with generation process with ground truth position (denoted as Full-GT), it is worth noting that when the hand is fixed to the exact ground truth, this may reduce the FS score due to some foot sliding in the generated motion. Our results are comparable to the Full-GT setting, indicating that our joint and affordance predictions are sufficiently accurate to provide effective guidance for generating plausible results. + +Analysis on text annotation. To verify the correct- + +ness of our generated text, we calculate BLEU [26] and ROUGE [24] scores, compared with the textual information provided by FullBodyManipulation in Tab. 7. These metrics illustrate the accuracy of our reasoning results. As shown in Tab. 1, we achieve performance comparable to using ground truth text, further demonstrating the effectiveness of our textual descriptions. Additionally, the results on both coarse and fine-grained text highlight the effectiveness of incorporating multi-level semantic information. + +Ablation studies on loss guidance. We conduct experiments on the loss guidance described in Tab. 6. The joint loss plays a key role in improving contact performance, while the foot loss enhances the feasibility of the generated results. These improvements are reflected in the enhancement of FID and FS metrics. + +# 4.5. Extension + +We further evaluate our model on more challenging scenarios, including complex human-human interactions and shape-varying object manipulations, as Fig. 6 shows. The varying size of objects are simulated to assess how well our model generates responsive and adaptive interactions. The results demonstrate the feasibility of our approach, with generating realistic human responsive motions. + +# 5. Conclusion + +In this work, we introduce SemGeoMo, a novel method for generating responsive human motions and corresponding textual descriptions based on dynamic interactive targets. We design an automated text annotator to provide semantic information. By integrating text-affordance-joint semantic and geometric guidance, SemGeoMo ensures the semantic coherence of the generated text and the geometric precision of the corresponding motion. Our method achieves state-of-the-art performance on three benchmarks and demonstrates generalization abilities on an unseen dataset. + +# References + +[1] Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Jiajun Wu, Deepak Gopinath, Alexander William Clegg, and Karen Liu. Circle: Capture in rich contextual environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21211-21221, 2023. 2 +[2] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 1, 2, 6 +[3] Zhi Cen, Huaijin Pi, Sida Peng, Zehong Shen, Minghui Yang, Shuai Zhu, Hujun Bao, and Xiaowei Zhou. Generating human motion in 3d scenes from text descriptions. In Proceedings of the IEEE/CVF Conference on CVPR, pages 1855-1866, 2024. 1, 2 +[4] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF CVPR, pages 18000-18010, 2023. 1 +[5] Peishan Cong, Ziyi Wang, Zhiyang Dou, Yiming Ren, Wei Yin, Kai Cheng, Yujing Sun, Xiaoxiao Long, Xinge Zhu, and Yuexin Ma. Laserhuman: Language-guided scene-aware human motion generation in free environment. arXiv preprint arXiv:2403.13307, 2024. 2, 3 +[6] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9760-9770, 2023. 1, 2 +[7] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9760-9770, 2023. 2 +[8] Christian Diller and Angela Dai. Cg-hoi: Contact-guided 3d human-object interaction generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19888-19901, 2024. 2 +[9] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 2 +[10] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. Advances in Neural Information Processing Systems, 35:4244-4256, 2022. 6 +[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[13] 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. 4 +[14] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 1, 2, 6, 7 +[15] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2 +[16] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Y, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural IPS, 36, 2024. 1 +[17] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. Chairs: Towards full-body articulated human-object interaction. arXiv preprint arXiv:2212.10621, 3, 2022. 2 +[18] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Zhiyuan Zhang, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. Full-body articulated human-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9365-9376, 2023. 2 +[19] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. arXiv preprint arXiv:2403.08629, 2024. 1, 2 +[20] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. Nifty: Neural object interaction fields for guided human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 947-957, 2024. 2 +[21] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 1, 2, 6, 7 +[22] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 1, 2, 3, 4, 5, 6, 7 +[23] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 1, 2 +[24] C Lin. Recall-oriented understudy for gisting evaluation (rouge). Retrieved August, 20:2005, 2005. 8 +[25] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 3 +[26] K Papinesi. Bleu: A method for automatic evaluation of machine translation. In Proc. 40th Actual Meeting of the Association for Computational Linguistics (ACL), 2002, pages 311-318, 2002. 8 +[27] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. Hoi-diff: Text-driven synthe + +sis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2, 6 +[28] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4332-4341, 2019. 3 +[29] 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. 4 +[30] Konstantinos I Roumeliotis and Nikolaos D Tselikas. Chatgpt and open-ai models: A preliminary review. Future Internet, 15(6):192, 2023. 2 +[31] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 1, 2, 4 +[32] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 4, 6, 7 +[33] 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. 2, 4 +[34] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135, 2023. 2 +[35] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9401-9411, 2021. 1, 2 +[36] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems, 35:14959-14971, 2022. 2 +[37] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 1, 5 +[38] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can: Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF CVPR, pages 433-444, 2024. 1, 2, 5, 6, 7 +[39] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 2 +[40] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. arXiv preprint arXiv:2309.07918, 2023. 1, 2 + +[41] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. arXiv preprint arXiv:2310.08580, 2023. 5 +[42] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-x: Towards versatile human-human interaction analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22260-22271, 2024. 1, 2 +[43] Liang Xu, Yizhou Zhou, Yichao Yan, Xin Jin, Wenhan Zhu, Fengyun Rao, Xiaokang Yang, and Wenjun Zeng. Regennet: Towards human action-reaction synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1759-1769, 2024. 2, 3 +[44] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 2 +[45] Hongwei Yi, Justus Thies, Michael J Black, Xue Bin Peng, and Davis Rempe. Generating human interaction motions in scenes with text control. In ECCV, pages 246-263. Springer, 2025. 2 +[46] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. In ECCV, pages 310–325. Springer, 2025. 5 +[47] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Neuraldome: A neural modeling pipeline on multi-view human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8834-8845, 2023. 2, 6, 7 +[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 5 +[49] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 2, 4 +[50] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[51] Chengfeng Zhao, Juze Zhang, Jiashen Du, Ziwei Shan, Junye Wang, Jingyi Yu, Jingya Wang, and Lan Xu. I'm hoi: Inertia-aware monocular capture of 3d human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 729-741, 2024. 2, 6 \ No newline at end of file diff --git a/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/images.zip b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..010c87a8d6adcda848322b6b1f732c26ec2f9703 --- /dev/null +++ b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3fa7853e8b3fa8bd7df4e3b687d69fbed09ff0d52fc167ecb649fd8fb187d39 +size 678670 diff --git a/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/layout.json b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ba959808f5f37831f398ea429981d5274328a465 --- /dev/null +++ b/CVPR/2025/SemGeoMo_ Dynamic Contextual Human Motion Generation with Semantic and Geometric Guidance/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f2471e79358c8ffac115b043547452fc3608aee501e1e5d8e2c6fa42dc1a67c +size 422997 diff --git a/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/9ff37ee2-adc0-47f0-9f75-7ff2202be291_content_list.json b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/9ff37ee2-adc0-47f0-9f75-7ff2202be291_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0effb6caa02eaa409e3875a3914724e8db952983 --- /dev/null +++ b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/9ff37ee2-adc0-47f0-9f75-7ff2202be291_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beac39303e1f5f58b487c56a99818eb9d0d18200bff31371ff1e4552d26a4990 +size 96651 diff --git a/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/9ff37ee2-adc0-47f0-9f75-7ff2202be291_model.json b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/9ff37ee2-adc0-47f0-9f75-7ff2202be291_model.json new file mode 100644 index 0000000000000000000000000000000000000000..73eb23932443f40985a026cdfba2cb11c7e57a25 --- /dev/null +++ b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/9ff37ee2-adc0-47f0-9f75-7ff2202be291_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3af0105c3730f70f4a7cf4a085bf9eb3e69efd597cb63715e943f7c6637d0133 +size 125790 diff --git a/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/9ff37ee2-adc0-47f0-9f75-7ff2202be291_origin.pdf b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/9ff37ee2-adc0-47f0-9f75-7ff2202be291_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..abc0e58c3dc6a4150ec1c0f6c58326ab8d90cefc --- /dev/null +++ b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/9ff37ee2-adc0-47f0-9f75-7ff2202be291_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24a7ef9fab3c86cb0c763028e46a46d974bb64b752cb32e3d6498dccb29a5572 +size 2079150 diff --git a/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/full.md b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..33286e0f114e8454479b73d60edc96c404fc1efc --- /dev/null +++ b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/full.md @@ -0,0 +1,416 @@ +![](images/8d0d39042c89de66be8109f020492f9106ac92fd1a0ededd84ea919b49e29f34.jpg) + +# Semantic Library Adaptation: + +# LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation + +![](images/043cce7e254ae8c0621e68b06cc866700fd10efec1f6bb3b7b4cefbc6853d951.jpg) + +# Abstract + +Open-vocabulary semantic segmentation models associate vision and text to label pixels from an undefined set of classes using textual queries, providing versatile performance on novel datasets. However, large shifts between training and test domains degrade their performance, requiring fine-tuning for effective real-world applications. We introduce Semantic Library Adaptation (SemLA), a novel framework for training-free, test-time domain adaptation. SemLA leverages a library of LoRA-based adapters indexed with CLIP embeddings, dynamically merging the most relevant adapters based on proximity to the target domain in the embedding space. This approach constructs an ad-hoc model tailored to each specific input without additional training. Our method scales efficiently, enhances explainability by tracking adapter contributions, and inherently protects data privacy, making it ideal for sensitive applications. Comprehensive experiments on a 20-domain benchmark built over 10 standard datasets demonstrate SemLA's superior adaptability and performance across diverse settings, establishing a new standard in domain adaptation for open-vocabulary semantic segmentation. + +# 1. Introduction + +Semantic segmentation aims to classify images at the pixel level, assigning a label to each pixel in an image. Traditionally, models are trained to recognize a fixed set of classes, but recent advances have led to open-vocabulary (OV) semantic segmentation, where models identify categories from an undefined and unbounded set using textual queries. This task leverages the synergy between text and + +![](images/86e053269d7c7fdc47223294aa306194adce329a094d75a60a667a78db84e5c2.jpg) +Figure 1. Overview of SemLA. During test-time, SemLA uses CLIP as a domain navigator, to retrieve and fuse relevant adapters, to get a LoRA tailored to the target domain. + +visual embeddings [51], enabling flexible and dynamic labeling [10, 71]. However, like all models, OV semantic segmentation models are vulnerable to domain shifts – situations where the model encounters data distributions different from those it was trained on, leading to degraded performance. This issue is significant because it impacts both segmentation accuracy and open-vocabulary generalization capabilities, which are crucial in real-world applications. + +Data is plentiful but heterogeneous – varying in labels, annotation styles, and other factors – and often contains sensitive information, limiting its accessibility. Consequently, we find ourselves in a paradox where an abundance of data does not necessarily translate into robust models capable + +of handling domain shifts [67]. Although several domain adaptation methods have been proposed, they have not been applied to OV semantic segmentation settings. + +Classic domain adaptation approaches are limited: they usually focus on a single target domain, require access to source data during adaptation, involve slow and expensive processes, and often deteriorate performance on the source dataset [21, 23, 64, 68, 87]. Test-time and online adaptation address some issues but are usually too slow for real-time applications and lack explainability [3, 49]. Drawing inspiration from recent developments in Large Language Models (LLMs) and the use of Parameter-Efficient Fine-Tuning (PEFT) techniques enriched with metadata – such as adapters in Hugging Face and Civit-ai – we aim to transfer this idea to the field of OV semantic segmentation, by integrating external information beyond what was found in the static training data through retrieval mechanisms [84]. + +We propose a training-free test-time adaptation approach for OV semantic segmentation, Semantic Library Adaptation (SemLA), achieved through adapter retrieval and merging. Our approach begins with creating a set of adapters using Low-Rank Adaptation (LoRA) [26, 42]. However, simply having a collection of adapters is insufficient, as the model still can't discern which adapter to select. Just as a collection of books requires an indexing system to become a functional library, we introduce an indexing mechanism to guide retrieval. Specifically, we employ Contrastive Language-Image Pretraining (CLIP) [51] to represent each LoRA adapter as the centroid of the CLIP embeddings of its training data. At test time, we can then select the most relevant adapters based on the proximity of the target image embedding to these centroids. + +Real-world domains are, however, far too numerous to rely on single adapters for every possible scenario due to the inherent combinatorial complexity of the domain space. To address this, we aggregate the most relevant adapters based on their proximity to the target domain. Just as in a library where the exact book you seek may not be available, the necessary knowledge can still be gathered by consulting multiple related sources. By fusing several adapters based on their relevance to the target image, we effectively construct an ad-hoc model fine-tuned for each situation. Figure 1 presents an overview of SemLA at test-time. + +This is training-free, scales well to any library size, and accommodates vastly different domains and label sets, with adaptation happening almost instantaneously through CLIP inference and LoRA merging. Moreover, our approach enhances explainability by identifying which adapters are most useful for certain target domains, allowing us to understand the model's reasoning and capabilities. + +Our main contributions are summarized as follows: + +- Training-free test-time adaptation for OV semantic segmentation: We introduce the first method for adap + +tation in OV segmentation. It requires no training at test time, enabling adaptive responses to diverse input images. Our framework is simple, scalable with large adapter libraries, and backbone-agnostic. + +- Novel benchmark for OV domain adaptation: We provide comprehensive experiments with a 20-domain benchmark, built over 10 popular datasets, showcasing performances over vastly different datasets and superior performance compared to zero-shot and naive adapter merging [35] approaches. +- Explainability and LoRA contribution analysis: Our approach is inherently transparent and controllable even at test time, easily scaling to new targets. The adaptation phase occurs without accessing source data, thereby protecting data privacy. We present analyses showing how adapters synergize and how we can measure their contributions and influence over the adaptation process. + +# 2. Related Work + +Our work intersects the following research areas: + +Semantic Segmentation. Deep learning techniques have led to increasingly effective semantic segmentation models. Fully Convolutional Networks (FCN) [41] and SegNet [1] extended convolutional neural networks with upsampling layers (deconvolution) to produce pixel-wise predictions. Subsequent works improved both speed [46, 76] and accuracy [7-9]. Enhancements in accuracy involved enlarging the receptive field [7-9, 74, 83], designing refinement modules [20, 22, 86], introducing boundary cues [6, 14, 59], and exploiting attention mechanisms [19, 36, 65] and Vision Transformers [72, 73, 78]. + +With the introduction of vision-language models like CLIP [51], open-vocabulary (OV) semantic segmentation [10, 71] has emerged as a new approach to segmentation, where the set of classes can be arbitrarily defined at any time through textual queries. This flexibility, however, introduces additional sources of domain shifts, such as vocabulary misalignment across different domains. + +Domain Adaptation. Domain adaptation focuses on adapting a model pre-trained on a source domain to perform well on a new, unlabeled target domain [82]. Adaptation can be performed either offline or during deployment. In the offline case, usually known as unsupervised domain adaptation (UDA), early approaches used style transfer strategies [16, 23, 37, 68, 75, 87] or self-training to adapt the model, for example by carefully retrieving reliable pseudolabels for the target domain by exploiting the confidence of the model itself [43, 89, 90], class balancing [24, 88], or prototypes [5, 79, 80]. While the offline scenario usually focuses on domain shifts occurring only once, such as moving from synthetic [52] to real [12] images, the deployment approach aims to handle continuous adaptation to avoid catastrophic forgetting of the source domain. Some methods as + +sume availability of data from the training domain and involve replay buffers [2, 31, 32], style transfer [17, 18, 69], contrastive learning [57, 62], or adversarial learning [70]. Others consider a more constrained case, source-free or test-time adaptation, where no data from the source domain is available [56], and involve pseudo-source data generation [40], partial freezing of the original model [38], feature alignment [39], batch norm retraining through entropy minimization [63], or prototype adaptation [29]. Recently, online adaptation emerged to tackle multiple domain shifts occurring unpredictably during deployment [3, 49, 54, 58, 61]. Although these strategies are flexible for unseen domains, they introduce significant overhead at deployment. Finally, training-free adaptation [35, 55] has been proposed to address domain shifts by aggregating parameters of different single-domain experts. This approach relies on linear mode connectivity to uniformly merge different models. We demonstrate that a careful selection of a few meaningful experts allows for better results. + +Parameter-Efficient Fine-Tuning. The advent of large language models (LLMs) and vision-language models (VLMs), with billions of parameters trained on vast datasets, unveiled the need for new fine-tuning paradigms to adapt them to specific use cases. Low-Rank Adaptation (LoRA) [25] was proposed to fine-tune LLMs by optimizing new sets of low-rank weights to learn residuals over the original parameters, preserving the knowledge derived from large-scale datasets while reducing computational load during fine-tuning. Advanced strategies allow dynamically setting the rank of LoRA parameters [81]. Recently, the possibility of merging different LoRAs learned for different tasks has gained popularity [27], yielding various policies for combining multiple LoRAs [42]. + +Model Merging. Model merging has emerged as a strategy to fuse knowledge from multiple sources directly in the weight space, effectively representing an alternative to ensemble and federated learning [34]. In [67], merging parameters of zero-shot and fine-tuned models improves out-of-distribution performance. Similarly, [28, 30] use linear interpolation for multi-task learning. While these methods interpolate all weights of the models, our method merges LoRA adapters, which are orders of magnitude smaller in size. Inspired by Model Soups [66], AdapterSoup [11] averages adapters in parameter space, automatically retrieving and uniformly merging them for adaptation to downstream language tasks. LoraRetriever [84] introduced a similar pipeline but instead uses instruction fine-tuning to train a retriever for selecting the most relevant adapters. Compared to these methods, SemLA performs adapter retrieval and fusion simply relying on the powerful semantic indexing capabilities of CLIP [51], which effectively serves as a domain navigator. + +Positioning Our Work. Our framework falls into the + +category of domain adaptation methods, as it actively adapts to new, unseen domains by modifying the model weights based on the input. However, we diverge from classical UDA, as our source dataset isn't fixed but can be changed at test time. It is represented by the data used to train the model backbone and all the datasets employed in the training of the adapters composing our LoRA library. These effectively represent the points in the LoRA weight-space over which we interpolate to generate new models for every unseen target image. + +# 3. Method and Framework + +In this section, we introduce SemLA. It consists of two main stages: (1) Construction of the LoRA Adapters Library, and (2) Dynamic Test-Time Adaptation. We begin by detailing the backbone architecture and the integration of Low-Rank Adaptation (LoRA) into it, followed by the description of our library creation and the adaptation process. + +# 3.1. Backbone Architecture with LoRA Integration + +Our framework builds on the CAT-Seg architecture [10], a state-of-the-art model designed for OV semantic segmentation. It aligns visual and textual information by mapping images and textual queries into a shared semantic space using CLIP embeddings [51]. This allows the model to perform segmentation without being constrained to a predefined set of classes, enabling flexible and dynamic labeling. SemLA can be deployed with any model enjoying these properties – thus not only CAT-Seg, as we will discuss later. + +To facilitate efficient and adaptable domain-specific tuning, we integrate Low-Rank Adaptation (LoRA) [26] into CAT-Seg. LoRA introduces learnable low-rank matrices to each linear layer, allowing for parameter-efficient fin-tuning without altering the original weights. Specifically, for each linear layer with weights $\mathbf{W} \in \mathbb{R}^{d \times k}$ , we augment it with low-rank matrices $\mathbf{B} \in \mathbb{R}^{d \times r}$ and $\mathbf{A} \in \mathbb{R}^{r \times k}$ , where $r \ll \min(d, k)$ . The adapted weights become: + +$$ +\mathbf {W} ^ {\prime} = \mathbf {W} + \Delta \mathbf {W}, \quad \Delta \mathbf {W} = \mathbf {B A}. \tag {1} +$$ + +By training only the LoRA parameters $(\mathbf{B},\mathbf{A})$ for each domain, we create lightweight domain-specific adapters without the need to retrain the entire model. For consistency and to simplify the merging process, we use the same rank $r$ for all LoRA adapters. + +Moreover, we denote a LoRA adapter with $\Delta \mathcal{W}_i = (\mathcal{B}_i,\mathcal{A}_i)$ , where $\mathcal{B}_i = \{B_{1,i},\dots,B_{m,i}\}$ and $\mathcal{A}_i = \{A_{1,i},\dots,A_{m,i}\}$ denote all the LoRA parameters applied across all $m$ linear layers of the CAT-Seg model. Given a LoRA adapter $\Delta \mathcal{W}_i$ , the corresponding adapted CAT-Seg model is denoted by $\mathrm{CAT - Seg}(\Delta \mathcal{W}_i)$ , or $\mathrm{CAT - Seg}(\Delta \mathcal{W}_i,\mathbf{x}_t)$ if we wish to specify the dependence on the input image $\mathbf{x}_t$ . In the following, we explain the pro + +![](images/9ddf9361cf625224a196df128fb007e34395e57a8015f28d8680e290b9e63d44.jpg) +Figure 2. Construction and Expansion of the LoRA Adapter Library. Each LoRA adapter is created by fine-tuning on a specific dataset and subsequently added to the library. The library index for each adapter is represented by the CLIP centroid of its training data. + +cedure for a single linear layer; updates for all other linear layers are performed analogously. + +# 3.2. Construction of the LoRA Adapters Library + +The first stage involves building a library of domain-specific LoRA adapters, each associated with a representative embedding in the CLIP space. This library serves as the foundation for our dynamic adaptation at test time. This stage, depicted in Figure 2, is performed offline and can be repeated for each newly labeled domain that becomes available for training. + +# 3.2.1. Domain Embeddings from CLIP + +For each training dataset (domain) $\mathcal{D}_i$ , we compute a centroid embedding $\mathbf{c}_i$ to represent the domain in the CLIP embedding space as follows: + +(a) Embedding Computation: For each image $\mathbf{x}_j\in \mathcal{D}_i$ we obtain its CLIP embedding $\mathbf{e}_j$ .. + +$$ +\mathbf {e} _ {j} = \operatorname {C L I P} _ {\text {i m a g e}} \left(\mathbf {x} _ {j}\right). \tag {2} +$$ + +(b) Centroid Calculation: We calculate the centroid of the embeddings from the $N_{i}$ images in $\mathcal{D}_i$ : + +$$ +\mathbf {c} _ {i} = \frac {1}{N _ {i}} \sum_ {j = 1} ^ {N _ {i}} \mathbf {e} _ {j}, \tag {3} +$$ + +The centroid $\mathbf{c}_i$ captures the semantic essence of $\mathcal{D}_i$ , facilitating efficient similarity comparisons during adaptation. + +# 3.2.2. Training Domain-Specific LoRA Adapters + +For each domain $\mathcal{D}_i$ , we fine-tune the LoRA parameters $\Delta \mathcal{W}_i = (\mathcal{B}_i, \mathcal{A}_i)$ on CAT-Seg using $\mathcal{D}_i$ , while keeping the original model weights $\mathbf{W}$ frozen. This results in a domain-specific adapter $\mathrm{LoRA}_i = \Delta \mathcal{W}_i$ , which, when combined with the centroid $\mathbf{c}_i$ , forms a tuple $(\mathbf{c}_i, \Delta \mathcal{W}_i)$ representing domain $\mathcal{D}_i$ in our library. + +# 3.2.3. Assembly of the Adapters Library + +By collecting all domain-specific adapters and their centroids, we construct the LoRA Adapters Library: + +$$ +\mathcal {L} = \left\{\left(\mathbf {c} _ {i}, \Delta \mathcal {W} _ {i}\right) \mid i = 1, 2, \dots , M \right\}, \tag {4} +$$ + +where $M$ is the total number of training domains. This library enables the model to dynamically adapt to various domains by selecting and merging relevant adapters. + +# 3.2.4. Extending the Library + +Our approach is extremely scalable. In any point in the future, as soon as new annotated data become available for a new domain $\mathcal{D}_{*}$ , we can perform steps 3.2.1 and 3.2.2 to obtain $(\mathbf{c}_{*}, \Delta \mathcal{W}_{*})$ and append it to $\mathcal{L}$ to enrich it + +$$ +\mathcal {L} = \mathcal {L} + \left(\mathbf {c} _ {*}, \Delta \mathcal {W} _ {*}\right) \tag {5} +$$ + +# 3.3. Dynamic Test-Time Adaptation + +The second stage focuses on adapting the model for each input image at test time by selecting and merging the most relevant adapters from the library, guided by CLIP. + +# 3.3.1. Computing the Test Image Embedding + +Given a test image $\mathbf{x}_t$ , we compute its CLIP embedding $\mathbf{e}_t$ : + +$$ +\mathbf {e} _ {t} = \operatorname {C L I P} _ {\text {i m a g e}} \left(\mathbf {x} _ {t}\right). \tag {6} +$$ + +This embedding serves as a query to identify the most relevant adapters in the library. + +# 3.3.2. Selecting Relevant Adapters + +We measure the similarity between the test image embedding and each domain centroid, e.g. as Euclidean distance: + +$$ +d _ {i} = \left\| \mathbf {e} _ {t} - \mathbf {c} _ {i} \right\| _ {2}, \quad \forall (\Delta \mathcal {W} _ {i}, \mathbf {c} _ {i}) \in \mathscr {L}. \tag {7} +$$ + +We then select the top- $K$ adapters with the smallest distances, denoted by the index set $\mathcal{K} = \{i_1, i_2, \ldots, i_K\}$ . + +# 3.3.3. Computing Adapter Weights + +To quantify the relevance of each selected adapter, we apply a softmax function with temperature $\tau$ to the proximities, computed as distance reciprocals: + +$$ +w _ {i} = \frac {\exp \left(\frac {1}{d _ {i} \cdot \tau}\right)}{\sum_ {k \in \mathcal {K}} \exp \left(\frac {1}{d _ {k} \cdot \tau}\right)}, \quad \forall i \in \mathcal {K}. \tag {8} +$$ + +These weights $w_{i}$ determine the contribution of each adapter in the fusion process. + +# 3.3.4. Merging Adapters via Concatenation + +We specify here the merging procedure for a given linear layer adapter [42]. Since all adapters share the same rank $r$ , the merging is straightforward: + +$$ +\mathbf {A} _ {i} ^ {\prime} = w _ {i} \mathbf {A} _ {i}, \quad \forall i \in \mathcal {K}. \tag {9} +$$ + +$$ +\mathbf {A} _ {\text {f u s e d}} = \left[ \mathbf {A} _ {i _ {1}} ^ {\prime \top}, \mathbf {A} _ {i _ {2}} ^ {\prime \top}, \dots , \mathbf {A} _ {i _ {K}} ^ {\prime \top} \right] ^ {\top} \in \mathbb {R} ^ {r K \times d}. \tag {10} +$$ + +$$ +\mathbf {B} _ {\text {f u s e d}} = \left[ \mathbf {B} _ {i _ {1}}, \mathbf {B} _ {i _ {2}}, \dots , \mathbf {B} _ {i _ {K}} \right] \in \mathbb {R} ^ {k \times r K}. \tag {11} +$$ + +$$ +\Delta \mathbf {W} _ {\text {f u s e d}} = \mathbf {B} _ {\text {f u s e d}} \mathbf {A} _ {\text {f u s e d}}. \tag {12} +$$ + +This fused update $\Delta \mathbf{W}_{\mathrm{fused}}$ integrates knowledge from multiple domains, weighted by their relevance to the test image. + +# 3.3.5. Updating the Model and Prediction + +We update the model weights for each linear layer: + +$$ +\mathbf {W} ^ {\prime} = \mathbf {W} + \Delta \mathbf {W} _ {\text {f u s e d}}. \tag {13} +$$ + +The adapted model is then used to segment the test image: + +$$ +\hat {\mathbf {y}} _ {t} = \operatorname {C A T - S e g} \left(\Delta \mathcal {W} _ {\text {f u s e d}}, \mathbf {x} _ {t}\right). \tag {14} +$$ + +This process is executed for each test image individually, enabling real-time, training-free adaptation. + +# 4. Experiment Setup and Results + +To evaluate SemLA's performance in OV semantic segmentation under domain shifts, we design a benchmark with a diverse set of datasets covering substantial data and vocabulary variations. Unlike most of the adaptation literature, often using limited domains or fixed label spaces, our benchmark stress tests the OV capability of the network across varied scenarios, label sets, and weather conditions. + +Scalability is essential for practical deployment, so we constructed a large library of LoRA adapters to show that our method scales effectively even with several adapters. This setup mirrors real-world scenarios where models must handle numerous domains and conditions without compromising performance. + +# 4.1. Datasets + +Our benchmark consists of 20 diverse semantic segmentation domains, derived from 10 datasets, selected to represent a wide range of domains and to provide large data and vocabulary shifts. This includes: + +- Driving Datasets: Commonly used in adaptation scenarios, featuring complex urban scenes with varying conditions - Cityscapes (CS) [12], BDD100K (BDD) [77], Mapillary Vistas (MV) [47], and Indian Driving Dataset (IDD) [15]. +- Weather-Specific Datasets: Small but challenging datasets focusing on extreme weather conditions like fog, rain, and night-time scenarios – ACDC [53], MUSES [4]. + +- General-Purpose Datasets: Datasets with a vast number of classes, covering a broad spectrum of scenes and objects beyond driving scenarios - ADE20K [85] (ADE150), PC59 [44], NYU [45], and COCONutL [13]. + +This diverse collection ensures that our benchmark captures significant variations in both data distribution and label spaces, providing a rigorous evaluation for OV semantic segmentation. + +# 4.2. Evaluation Protocol + +To adhere to common conventions from the adaptation literature, we adopt a leave-one-out evaluation strategies + +1. For each dataset, we train an adapter on its training set. +2. When evaluating on a particular dataset, we remove its corresponding adapter from the library, to prevent the model from leveraging any direct knowledge of it. +3. We repeat the process for each dataset, ensuring that the model is always tested on unseen data, as it never accesses to domain-specific adapters. + +# 4.2.1. Implementation Details + +For our SemLA method, we select the top $K = 7$ closest LoRA adapters based on CLIP embedding proximity, using a softmax temperature $\tau = 0.01$ to weight the adapters during fusion. This configuration balances performance and computational efficiency. We refer to supplementary material for further details. + +# 4.3. Comparative Methods + +We compare the following methods: + +- Zero-Shot CAT-Seg [10]: The baseline model without any adaptation or LoRA integration. +- Uniform LoRA Merging: In this approach, all LoRA adapters (excluding the one corresponding to the test dataset) are uniformly averaged without considering their relevance to the target domain – which can be seen as an adapters-variant of [35] revised to our setting with supervised fine-tuning of the adapters. +- SemLA (Ours): Our proposed method, which performs adaptive LoRA merging based on domain proximity. +- Uniform (Late Fusion): Outputs from all LoRA adapters (excluding the one corresponding to the test dataset) are uniformly averaged at the softmax output level. +- SemLA (Late Fusion): Similar to SemLA, weights are computed based on domain proximity but applied to the softmax outputs. This method has higher computational complexity than SemLA (one forward pass per adapter). + +Furthermore, we report the performance achieved by each of the adapters trained on the specific dataset where we evaluate, although such adapters had the unfair advantage of having access to training data from the target domain – accordingly, we refer to them as Oracles. + +
MethodACDCMUSESCSBDDMVA150IDDPC59NYUCOCONutL*h-mean
rainsnowfognightclear (d)clear (n)rain (d)rain (n)fog (d)fog (n)snow (d)snow (n)
Zero-shot [10]46.5348.0447.0937.9344.4339.2938.9527.7853.7325.3543.5633.2947.1147.9525.6937.8335.3963.3349.38(68.26)39.39
Oracles70.9469.2269.9851.5569.3657.0954.2852.1175.8561.2666.2554.3567.4760.0649.5753.9964.3468.6861.90(70.44)61.05
Uniform (Late Fusion)64.5864.7669.6448.5160.5351.8850.6237.3079.0035.6862.3944.0461.8257.0830.3436.0437.4363.2447.39(66.40)49.26
SemLA (Late Fusion)66.0966.7771.0150.7962.6955.7053.2745.7168.8645.6565.2049.7162.8657.6029.6837.2839.6463.8750.64(65.84)52.09
Uniform [35]67.4066.3569.7149.9858.2855.7854.7045.0973.7545.1661.0249.0862.1858.1931.5137.2538.8363.0648.93(67.62)51.89
SemLA (ours)67.7168.9571.9251.7361.0960.0657.6047.3572.9752.3867.2855.9263.9157.3031.1238.1840.1664.7551.35(67.26)54.16
+0.31+2.60+2.21+1.75+2.81+4.28+2.90+2.26-0.78+7.22+6.26+6.84+1.73-0.89-0.39+0.93+1.33+1.69+2.42-0.36+2.27
+ +Table 1. Adaptation for OV semantic segmentation - CAT-Seg [10]. Performance comparison across our 20-domain benchmark, leave-one-out setting. On MUSES, (d) and (n) stand for day and night. () means excluded from h-mean. SemLA with $\tau = {0.05}$ ,and $K = 5$ . + +# 4.4. Main Results and Discussion + +Table 1 presents the mean Intersection over Union (mIoU) scores for each method across all datasets. Our method, SemLA, consistently outperforms both the zero-shot baseline and the uniform merging approach [35] (improvements over the latter are shown in the last row), reaching close to the performance of the single adapters - and, sometimes, even outperforming them (e.g., on ACDC fog and night). + +One particular dataset, *CoconutL* (labeled with *), largely overlaps with the pretraining data of CAT-Seg. Evaluating our performance on it - even in the leave-one-out protocol - would not fit the usual setting established for the adaptation task. Therefore, we exclude it from the final harmonic mean calculations. However, we provide its results in brackets for completeness, which show (unsurprisingly) an exceptional performance of the Zero-shot model. + +Performance Analysis. Our results demonstrate that SemLA significantly outperforms the zero-shot baseline across all datasets, highlighting the effectiveness of our domain adaptation strategy. Uniform merging [35], despite its simplicity, already shows notable improvements over the zero-shot model, primarily due to linear mode connectivity, showing that LoRA merging is effective when one has access to a LoRA library without any target awareness - e.g., in domain generalization settings - as already proven in Li et al. [35]. However, the introduction of careful adapters weighting by SemLA leads to substantial performance gains. This underscores that selecting relevant adapters based on domain proximity is more important than merely scaling up the number of adapters available in the library. Moreover, we observe that both Uniform and SemLA (Late Fusion) perform worse than their LoRA merging counterparts, despite their higher computational complexity. These likely underperform because operating solely at the output level and cannot adjust internal model representations, unlike parameter-level fusion. + +Adapter Contribution Analysis. To understand how the adapters contribute across datasets, we analyze the selection patterns shown in Figure 3, where a heatmap indicates the frequency and weight of each adapter's selection per test dataset. Adapter support is spread across the library, showing that the model benefits from multiple adapters rather than relying on a few only. On the one hand, we can appreciate high weights for strongly related domains - + +![](images/25b03680632e516f812c5a4464727d44edc6634fb0e5ecc355c0c83f2a8346fb.jpg) +Figure 3. Adapter contribution heatmap. Rows represent individual test datasets, and columns correspond to specific LoRA adapters. The color intensity of each cell indicates the frequency and weight of selection (with values below 0.1 omitted). The diagonal is empty due to the leave-one-out strategy. + +e.g., MUSES-snow-day heavily relies on ACDC-snow and vice-versa; in general, we can observe polarized selection for driving datasets (top left) or generic ones (bottom right). Nonetheless, even adapters from unrelated domains are often selected, suggesting the existence of shared useful features, though each category still predominantly influences related datasets. By examining columns (and aggregating them in a histogram), certain adapters emerge as "popular" choices, highlighting their stronger influence. + +# 4.5. Analysing CLIP as a Domain Navigator + +A key assumption of our framework is that CLIP embeddings provide a meaningful mapping of the domains and that proximity in the CLIP space correlates with segmentation performance. To verify this assumption, we examine i) how our method navigates complex domains by combining relevant adapters; and ii) whether proximity in the CLIP space is a good heuristic for predicting performance. + +
MethodACDCMUSESCSBDDMVA150IDDPC59NYUCOCONutL*h-mean
rainsnowfognightclear (d)clear (n)rain (d)rain (n)fog (d)fog (n)snow (d)snow (n)
Oracles70.9469.2269.9851.5569.3657.0954.2852.1175.8561.2666.2554.3567.4760.0649.5753.9964.3468.6861.90(70.44)61.05
Uniform67.7666.6969.9150.4458.6056.3055.7345.9673.3346.2961.3650.2762.7758.4232.8738.4839.8963.7750.12(68.13)52.78
SemLA τ = 0.00570.9769.4070.9451.6869.5157.9554.2552.1475.8861.2666.1651.2267.5260.2048.7049.4264.2768.2762.68(68.83)60.57
SemLA τ = 0.0568.8270.4373.3853.0667.1959.9359.3052.2172.4562.1069.3359.3867.4958.1838.0644.2049.4967.4557.93(68.96)58.79
+ +Table 2. Adaptation for OV semantic segmentation - CAT-Seg [10], all-inclusive setting. Performance comparison across our 20-domain benchmark, when all domains are available in the LoRA Library, hence to source. On MUSES, (d) and (n) stand for day and night. () means excluded from h-mean. SemLA with $K = 5$ . + +![](images/48c7e6a51994586b053f3d472351a659dff737aa8965b142a18fa1c3d8157b09.jpg) +Figure 4. Adapter weight distribution for MUSES-Fog-Night. The fused adapter combines knowledge from foggy and night-time conditions by weighting relevant adapters. Adapters with a weight lower than $5\%$ are not included. + +Domain Composition Analysis. We observe that CLIP effectively navigates the domain space, as shown in the adapter selection for MUSES-Fog-Night. This dataset, comprising foggy night images, lacks a direct match in our library. However, adapters trained on night and fog conditions are available. Examining LoRA contributions for MUSES-Fog-Night in Figure 4, we see that fog and night adapters dominate with around $40\%$ and $60\%$ contribution respectively, with higher weights for MUSES adapters, around $80\%$ , likely due to shared camera characteristics. This illustrates our ability to combine relevant knowledge even when an exact match is absent and showcases the easy interpretability of our approach. + +CLIP Distance - Performance Correlation. Having observed that CLIP embeddings effectively guide adapter selection, we proceed to the second question: Is image-LoRA proximity in the CLIP space a good heuristic for predicting segmentation performance? Figure 5 indicates that this is indeed the case. The negative slope of the regression line indicates an inverse relationship between CLIP distance and mIoU – i.e., smaller CLIP distances (higher proximity) generally correspond to better segmentation. This empirical observation confirms that CLIP embedding distance is a good heuristic for predicting adapter effectiveness. + +# 4.6. All-Inclusive Settings + +We evaluate the – unlikely – scenario where all the LoRA adapters corresponding to the benchmark datasets are in + +![](images/ca8a50f8cc37cb248abf42592a60f28213ccc0e056169ddc944e8f012a1834be.jpg) +Figure 5. CLIP-guidance effectiveness for LoRA selection on ACDC. Each point represents an image-adapter combination, with adapters separated by color. x-axis: distance from the corresponding image embedding to the adapter embedding. y-axis: improvement in mIoU when using the adapter relative to the zero-shot base network. The linear regression curve (dashed line) indicates that embedding similarity correlates with higher mIoU. We show the full adapter library, excluding those trained on ACDC. + +cluded in the library at test-time (i.e., in contrast with common adaptation settings). Table 2 reports SemLA tested with two $\tau$ and Uniform [35] settings. The results reveal a two-fold effect: when the system is directed to prioritize the target-domain LoRA adapter, the model's performance almost matches the Oracle. However, when the temperature $\tau$ is higher, hence LoRA fusion is promoted, the model - while slightly lower on average - overcomes the Oracles on 14 domains over 20, primarily in highly related datasets. This finding underscores that LoRA synergies can extend beyond individual adapters, achieving performance that, in some cases, surpasses even the Oracle. + +# 4.7. Results with a Different Backbone: SED + +Our method is inherently backbone-agnostic, as we only operate on LoRA retrieval and adaptation, relying on simple and widespread methods. To empirically prove this claim, we generate another library based on SED [71], another popular open-vocabulary semantic segmentation model. In Table 3, we collect the outcome of this new evaluation, carried out once again on our 20-domain benchmark. All our achievements with CAT-Seg are confirmed when switching to SED, with SemLA clearly outperforming both the zero-shot model and the uniform merging strategy [35], thus confirming the flexibility of our framework and the possibility of tailoring it around different OV semantic segmentation backbones. + +
MethodACDCMUSESCSBDDMVA150IDDPC59NYUCOCONuL*h-mean
rainsnowfognightclear (d)clear (n)rain (d)rain (n)fog (d)fog (n)snow (d)snow (n)
Zero-shot [10]42.6443.9746.8835.0638.0535.0135.1827.8748.0919.5342.6323.5341.9043.4125.0535.2735.2060.8747.20(66.36)35.56
Oracles61.7068.2569.3348.3962.0455.5553.6144.8269.4053.2065.6055.8568.4858.6147.2850.6663.1566.5560.60(68.01)58.05
Uniform [35]59.5961.0768.3347.2758.7953.8750.6537.1269.8235.5261.4644.0861.8554.7830.3734.9239.5960.8248.42(64.05)48.60
SemLA (ours)60.0167.0069.9649.7060.9854.7955.1538.8969.4638.2566.8048.8964.3554.4028.9936.9141.0862.2351.82(65.46)50.57
+0.42+5.93+1.63+2.43+2.19+0.92+4.50+1.77-0.38+2.70+5.34+4.81+2.50-0.38-1.38+1.99+1.59+1.69+3.40+1.41+1.97
+ +Table 3. Adaptation for OV semantic segmentation - SED [71]. Performance comparison across 20-domain benchmark, leave-one-out setting. On MUSES, (d) and (n) stand for day and night. () means excluded from h-mean. SemLA with $\tau = {0.05}$ ,and $K = 5$ . + +![](images/9a9fabd8eb7cc655d41cdf68f4c986dd3b8c8c57f322b03cbf47293c3e98e2aa.jpg) +Figure 6. Qualitative results on ACDC and NYU. From left to right: Input Image (RGB), ground-truth semantic masks (GT) and predictions by the Zero-Shot model [10], Uniform Merging [35], and SemLA. + +# 4.8. Qualitative Results + +Figure 6 presents qualitative comparisons between the predictions by different methods – zero-shot CAT-Seg, uniform merged model [35] and ours. SemLA produces more accurate and detailed segmentation maks, demonstrating its superiority over zero-shot and uniform merging models. + +# 4.9. Limitations and Future Work + +While our method demonstrates strong performance, several areas remain for future exploration. First, we do not explicitly address vocabulary alignment between source and target domains; incorporating vocabulary matching could enhance performance by ensuring compatibility with the target label space. Second, exploring automated domain discovery through CLIP clustering and unsupervised training of new LoRA adapters could lead to a self-augmenting library, enhancing adaptability without manual intervention. Scaling SemLA to thousands of adapters introduces challenges in recognizing and addressing library weaknesses and gaps; developing automated strategies to identify and + +mitigate them will be essential. Lastly, applying our approach to multi-tasks settings, e.g. panoptic segmentation or depth estimation [50], could validate its versatility. + +# 5. Conclusion + +We have presented SemLA, a framework for training-free test-time adaptation in open-vocabulary semantic segmentation. By constructing a library of LoRA adapters indexed with CLIP embeddings and dynamically merging them based on domain proximity, our method effectively mitigates domain shifts. Extensive experiments show that SemLA outperforms zero-shot baselines, uniform adapter merging, and sometimes even fine-tuned models across diverse datasets. We confirmed that CLIP embeddings reliably guide adapters selection and that proximity in the embedding space is a good heuristic for performance. With enhanced explainability, scalability, and data privacy, SemLA is suitable for practical applications. Future work will address vocabulary alignment, automated domain discovery, and extending the framework to other tasks. + +Acknowledgments. The authors thank Hedvig Kjellstrom for the helpful discussions and guidance, and acknowledge The European High Performance Computing Joint Undertaking (EuroHPC JU), EuroCC National Competence Center Sweden (ENCCS) and the CINECA award under the ISCRA initiative for the availability of high-performance computing resources and support. + +This study was funded by the European Union - Next Generation EU within the framework of the National Recovery and Resilience Plan NRRP - Mission 4 "Education and Research" - Component 2 - Investment 1.1 "National Research Program and Projects of Significant National Interest Fund (PRIN)" (Call D.D. MUR n. 104/2022) - PRIN2022 - Project reference: "RiverWatch: a citizen-science approach to river pollution monitoring" (ID: 2022MMBA8X, CUP: J53D23002260006). + +# References + +[1] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 39(12):2481-2495, 2017. 2 +[2] Andreea Bobu, Judy Hoffman, Eric Tzeng, and Trevor Darell. Adapting to continuously shifting domains. In ICLR 2018 Workshop Program Chairs, 2018. 00000. 3 +[3] Marc Botet Colomer, Pier Luigi Dovesi, Theodoros Panagiotakopoulos, Joao Frederico Carvalho, Linus Harenstam-Nielsen, Hossein Azizpour, Hedvig Kjellstrom, Daniel Cremers, and Matteo Poggi. To adapt or not to adapt? real-time adaptation for semantic segmentation. In IEEE International Conference on Computer Vision, 2023. ICCV. 2, 3, 6 +[4] Tim Brodermann, David Bruggemann, Christos Sakaridis, Kevin Ta, Odysseas Liagouris, Jason Corkill, and Luc Van Gool. Muses: The multi-sensor semantic perception dataset for driving under uncertainty. arXiv preprint arXiv:2401.12761, 2024. 5 +[5] Chaoqi Chen, Weiping Xie, Wenbing Huang, Yu Rong, Xinghao Ding, Yue Huang, Tingyang Xu, and Junzhou Huang. Progressive feature alignment for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 627-636, 2019. 2 +[6] Liang-Chieh Chen, Jonathan T Barron, George Papandreou, Kevin Murphy, and Alan L Yuille. Semantic image segmentation with task-specific edge detection using cnns and a discriminatively trained domain transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4545-4554, 2016. 2 +[7] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2 +[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4):834-848, 2018. + +[9] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2 +[10] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4113-4123, 2024. 1, 2, 3, 5, 6, 7, 8 +[11] Alexandra Chronopoulou, Matthew Peters, Alexander Fraser, and Jesse Dodge. AdapterSoup: Weight averaging to improve generalization of pretrained language models. In Findings of the Association for Computational Linguistics: EACL 2023, pages 2054-2063, Dubrovnik, Croatia, 2023. Association for Computational Linguistics. 3 +[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 5 +[13] Xueqing Deng, Qihang Yu, Peng Wang, Xiaohui Shen, and Liang-Chieh Chen. Coconut: Modernizing coco segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5 +[14] Henghui Ding, Xudong Jiang, Ai Qun Liu, Nadia Magnenat Thalmann, and Gang Wang. Boundary-aware feature propagation for scene segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6819-6829, 2019. 2 +[15] Shubham Dokania, AH Hafez, Anbumani Subramanian, Manmohan Chandraker, and CV Jawahar. Idd-3d: Indian driving dataset for 3d unstructured road scenes. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4482–4491, 2023. 5 +[16] A. Dundar, M. Y. Liu, Z. Yu, T. C. Wang, J. Zedlewski, and J. Kautz. Domain stylization: A fast covariance matching framework towards domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020. 2 +[17] Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, and Dengxin Dai. Normalization perturbation: A simple domain generalization method for real-world domain shifts. arXiv preprint arXiv:2211.04393, 2022.3 +[18] Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, and Dengxin Dai. Towards robust object detection invariant to real-world domain shifts. In The Eleventh International Conference on Learning Representations (ICLR 2023). OpenReview, 2023. 3 +[19] Jun Fu, Jing Liu, Hajie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019. 2 + +[20] Jun Fu, Jing Liu, Yuhang Wang, Yong Li, Yongjun Bao, Jinhui Tang, and Hanqing Lu. Adaptive context network for scene parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6748-6757, 2019. 2 +[21] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2 +[22] Junjun He, Zhongying Deng, Lei Zhou, Yali Wang, and Yu Qiao. Adaptive pyramid context network for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7519-7528, 2019. 2 +[23] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. CyCADA: Cycle-consistent adversarial domain adaptation. In Proceedings of the 35th International Conference on Machine Learning, pages 1989-1998, Stockholm, Sweden, 2018. PMLR. 2 +[24] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. arXiv preprint arXiv:2111.14887, 2021. 2 +[25] 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 +[26] 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. In International Conference on Learning Representations, 2022. 2, 3 +[27] Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. Lorahub: Efficient cross-task generalization via dynamic lora composition. arXiv preprint arXiv:2307.13269, 2023. 3 +[28] Gabriel Ilharco, Mitchell Wortsman, Samir Yitzhak Gadre, Shuran Song, Hannaneh Hajishirzi, Simon Kornblith, Ali Farhadi, and Ludwig Schmidt. Patching open-vocabulary models by interpolating weights. Advances in Neural Information Processing Systems, 35:29262-29277, 2022. 3 +[29] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In Advances in Neural Information Processing Systems, 2021. 3 +[30] Dong-Hwan Jang, Sangdoo Yun, and Dongyoon Han. Model stock: All we need is just a few fine-tuned models. In European Conference on Computer Vision, pages 207-223. Springer, 2025. 3 +[31] Yevhen Kuznietsov, Marc Proesmans, and Luc Van Gool. Towards unsupervised online domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 261-271, 2022. 3 +[32] Qicheng Lao, Xiang Jiang, Mohammad Havaei, and Yoshua Bengio. Continuous domain adaptation with vari- + +ational domain-agnostic feature replay. arXiv preprint arXiv:2003.04382, 2020. 3 +[33] 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. 3, 4 +[34] Weishi Li, Yong Peng, Miao Zhang, Liang Ding, Han Hu, and Li Shen. Deep model fusion: A survey, 2023. 3 +[35] Wenyi Li, Huan-ang Gao, Mingju Gao, Beiwen Tian, Rong Zhi, and Hao Zhao. Training-free model merging for multi-target domain adaptation. In European Conference on Computer Vision. Springer, 2024. 2, 3, 5, 6, 7, 8 +[36] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9167-9176, 2019. 2 +[37] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[38] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. CoRR, 2020. 3 +[39] Yuejiang Liu, Parth Kothari, Bastien Germain van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test-time training fail or thrive? In Advances in Neural Information Processing Systems, 2021. 3 +[40] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. 2021. 3 +[41] Jonathan Long, Evan Shelhamer, and Trevor Darrel. Fully convolutional networks for semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2 +[42] Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. Peft: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft, 2022.2,3,5 +[43] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. Lecture Notes in Computer Science, page 415-430, 2020. 2 +[44] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 5 +[45] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 5 +[46] Vladimir Nekrasov, Chunhua Shen, and Ian Reid. Lightweight refinenet for real-time semantic segmentation. In British Conference on Computer Vision (BMVC), 2018. 2 + +[47] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 5 +[48] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research. 3, 5, 6 +[49] Theodoros Panagiotakopoulos, Pier Luigi Dovesi, Linus Harenstam-Nielsen, and Matteo Poggi. Online domain adaptation for semantic segmentation in ever-changing conditions. In European Conference on Computer Vision (ECCV), 2022. 2, 3 +[50] Matteo Poggi and Fabio Tosi. Federated online adaptation for deep stereo. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 8 +[51] 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. 1, 2, 3, 5, 6 +[52] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In European conference on computer vision, pages 102-118. Springer, 2016. 2 +[53] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765-10775, 2021. 5 +[54] Mattia Segu, Bernt Schiele, and Fisher Yu. Darth: Holistic test-time adaptation for multiple object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9717-9727, 2023. 3 +[55] Mattia Segu, Alessio Tonioni, and Federico Tombari. Batch normalization embeddings for deep domain generalization. Pattern Recognition, 135:109115, 2023. 3 +[56] Serban Stan and Mohammad Rostami. Unsupervised model adaptation for continual semantic segmentation. In AAAI, 2021. 3 +[57] Peng Su, Shixiang Tang, Peng Gao, Di Qiu, Ni Zhao, and Xiaogang Wang. Gradient regularized contrastive learning for continual domain adaptation. 2020. 00000. 3 +[58] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. SHIFT: a synthetic driving dataset for continuous multi-task domain adaptation. In Computer Vision and Pattern Recognition, 2022. 3 +[59] Towaki Takikawa, David Acuna, Varun Jampani, and Sanja Fidler. Gated-scnn: Gated shape cnns for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5229-5238, 2019. 2 + +[60] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 3 +[61] Riccardo Volpi, Pau de Jorge, Diane Larlus, and Gabriela Csurka. On the road to online adaptation for semantic image segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[62] Vibashan VS, Poojan Oza, and Vishal M. Patel. Towards online domain adaptive object detection. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023. 3 +[63] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 3 +[64] Haoran Wang, Tong Shen, Wei Zhang, Lingyu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In The European Conference on Computer Vision (ECCV), 2020. 2 +[65] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 2 +[66] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Proceedings of the 39th International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 3 +[67] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7959-7971, 2022. 2, 3, 6 +[68] Zuxuan Wu, Xintong Han, Yen-Liang Lin, Mustafa Gokhan Uzunbas, Tom Goldstein, Ser Nam Lim, and Larry S. Davis. Dcan: Dual channel-wise alignment networks for unsupervised scene adaptation. Lecture Notes in Computer Science, page 535-552, 2018. 2 +[69] Zuxuan Wu, Xin Wang, Joseph Gonzalez, Tom Goldstein, and Larry Davis. ACE: Adapting to changing environments for semantic segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2121-2130. IEEE, 2019. 3 +[70] Markus Wulfmeier, Alex Bewley, and Ingmar Posner. Incremental adversarial domain adaptation for continually changing environments. 2018. 00000. 3 +[71] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 7, 8 +[72] Enze Xie, Wenjia Wang, Wenhai Wang, Peize Sun, Hang Xu, + +Ding Liang, and Ping Luo. Segmenting transparent objects in the wild with transformer. In *IJCAI*, 2021. 2 +[73] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2 +[74] Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3684-3692, 2018. 2 +[75] Yanchao Yang and Stefano Soatto. FDA: Fourier domain adaptation for semantic segmentation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4084-4094. IEEE, 2020. 2 +[76] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018. 2 +[77] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 5 +[78] Yuhui Yuan, Xiaokang Chen, Xin Chen, and Jingdong Wang. Segmentation transformer: Object-contextual representations for semantic segmentation. In European Conference on Computer Vision (ECCV), 2020. 2 +[79] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. 2021. 2 +[80] Qiming Zhang, Jing Zhang, Wei Liu, and Dacheng Tao. Category anchor-guided unsupervised domain adaptation for semantic segmentation. Advances in Neural Information Processing Systems, 2019. 2 +[81] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient fin-tuning. In The Eleventh International Conference on Learning Representations, 2023. 3 +[82] Youshan Zhang. A survey of unsupervised domain adaptation for visual recognition. arXiv:2112.06745, 2021. 2 +[83] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 2 +[84] Ziyu Zhao, Leilei Gan, Guoyin Wang, Wangchunshu Zhou, Hongxia Yang, Kun Kuang, and Fei Wu. LoraRetriever: Input-aware LoRA retrieval and composition for mixed tasks in the wild. In Findings of the Association for Computational Linguistics: ACL 2024, pages 4447-4462, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3 +[85] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through + +ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5 +[86] Yizhou Zhou, Xiaoyan Sun, Zheng-Jun Zha, and Wenjun Zeng. Context-reinforced semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4046-4055, 2019. 2 +[87] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2 +[88] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 2 +[89] Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jinsong Wang. Confidence regularized self-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5982-5991, 2019. 2 +[90] Yang Zou, Zhiding Yu, Xiaofeng Liu, B. V. K. Vijaya Kumar, and Jinsong Wang. Confidence regularized self-training. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2 \ No newline at end of file diff --git a/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/images.zip b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e0faa778c4270b855b4835322b19a01e36e5e5a2 --- /dev/null +++ b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cdb72f71db1f0c351f493e3d0c477ea3e9f41603b19a78b2ecb860a39e6f659 +size 607834 diff --git a/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/layout.json b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4a07c3d1871caccc7fa0be73b356baa28d7f5543 --- /dev/null +++ b/CVPR/2025/Semantic Library Adaptation_ LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fd46682dd51db45668fd52903e18b65d9ec6338497cea432b05ab4b083fbd15 +size 451621 diff --git a/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/465c8cd8-d1c4-4df0-a3b2-1dbb5fb3f3bb_content_list.json b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/465c8cd8-d1c4-4df0-a3b2-1dbb5fb3f3bb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1e1a35d0fc5fb220a1e4753a51cee56f4e599a47 --- /dev/null +++ b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/465c8cd8-d1c4-4df0-a3b2-1dbb5fb3f3bb_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e28e396d5afdc245f18ceb2c12ec4f84eaac8fe7800f16a4bdc7f5562ae7dc8 +size 98289 diff --git a/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/465c8cd8-d1c4-4df0-a3b2-1dbb5fb3f3bb_model.json b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/465c8cd8-d1c4-4df0-a3b2-1dbb5fb3f3bb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6194fec379dc747eb773d066918149188996f52d --- /dev/null +++ b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/465c8cd8-d1c4-4df0-a3b2-1dbb5fb3f3bb_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09ea2a02c194c8067c3b7a7f05e96c3410cd18390714cad5e0ca2b142147b2f4 +size 133542 diff --git a/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/465c8cd8-d1c4-4df0-a3b2-1dbb5fb3f3bb_origin.pdf b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/465c8cd8-d1c4-4df0-a3b2-1dbb5fb3f3bb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b0955184cf0383694b4eb6eb9fed4024d4224ab9 --- /dev/null +++ b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/465c8cd8-d1c4-4df0-a3b2-1dbb5fb3f3bb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7636f1e2144580c2deb25fde709043df33ba84c6e59e1e817966eba75810bb5a +size 5456016 diff --git a/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/full.md b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8f28469c59664a8dcc90aed68ee2afb50494252f --- /dev/null +++ b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/full.md @@ -0,0 +1,322 @@ +# Semantic and Expressive Variation in Image Captions Across Languages + +Andre Ye $^{1}$ , Sebastin Santy $^{1}$ , Jena D. Hwang $^{2}$ , Amy X. Zhang $^{1,2}$ , Ranjay Krishna $^{1,2}$ $^{1}$ University of Washington, $^{2}$ Allen Institute for AI Seattle, Washington + +andreye@uw.edu, ssanty@cs.washington.edu, axz@cs.uw.edu, ranjay@cs.washington.edu + +# Abstract + +Most vision-language models today are primarily trained on English image-text pairs, with non-English pairs often filtered out. Evidence from cross-cultural psychology suggests that this approach will bias models against perceptual modes exhibited by people who speak other (non-English) languages. We investigate semantic and expressive variation in image captions across different languages; we analyze both human-annotated datasets and model-produced captions. By analyzing captions across seven languages (English, French, German, Russian, Chinese, Japanese, Korean) in high-quality image captioning datasets (Crossmodal and Visual Genome), we find that multilingual caption sets tend to provide richer visual descriptions than monolingual (including English-only) ones; multilingual sets contain $46.0\%$ more objects, $66.1\%$ more relationships, and $66.8\%$ more attributes. We observe the same results with multilingual captions produced by LLaVA and the Google Vertex API: for example, compared to monolingual captions, they cover $21.9\%$ more objects, $18.8\%$ more relations, and $20.1\%$ more attributes. These suggest that, across a large number of samples, different languages bias people and models to focus on different visual concepts. Finally, we show that models trained on image-text data in one language perform distinctly better on that language's test set. Our work points towards the potential value of training vision models on multilingual data sources to widen the range/variation of descriptive information those models are exposed to. + +# 1. Introduction + +Typically, vision-language models are trained with large quantities of image-text pairs scraped from the web [33, 103, 116]. To obtain high quality image-text pairs, these web datasets are filtered with models like CLIP [51, 94] with the goal of ensuring that each text sample is sufficiently descriptive of or related to the corresponding image. However, this filtering tends to be biased towards En + +glish text samples, such that the majority of samples in popular image-text datasets are in English [37, 46, 103]. Many vision-language models are even trained only on English [59, 71, 94, 111], excluding other languages entirely. + +Prior work shows that models tend to fail at tasks where naturally occurring text is uncommon, such as negation [30]. Therefore, increasing the lexical diversity of naturally occurring text is important: it allows models to experience a broader range of linguistic patterns that might be underrepresented within a certain set of scraped text, while maintaining the scale that is only possible through training on naturally occurring data. Thus, biasing vision datasets towards English-only image-text pairs may risk excluding the equally relevant and possibly unique visual concepts produced in languages other than English. Indeed, a wide body of work from the cross-cultural psychology suggests that people from across the world, speaking different languages and living in different cultural contexts, tend to describe the visual world differently. Early work in psychology suggests that fundamental aspects of visual perception, such as perception of length [106], geometrical intuition [29, 89], and depth [52] vary across people from different geographic backgrounds. More recent work has shown that individuals from different cultural backgrounds exhibit differences in how they look at, understand, and talk about visual scenes [16, 62, 85]. For instance, when looking at the same image, Americans tend to describe the focal objects and its attributes, whereas Japanese tend to focus on the relations between objects [80]. + +Drawing on this work, we hypothesize the content distribution of visual descriptions varies by the structure and use of specific languages – and that these differences are detectable in computer vision datasets and model behaviors. This hypothesis is further supported by observed differences in spoken language: German's complex morphosyntactic system provides events with nuanced understanding of spatial relationships [92]; Russian verbs of motion require speakers to specify directionality [88]; English uses the generic verb of motion “to go” whereas a Russian speaker must specify whether the motion has a desti + +![](images/132f94ac2b63d2a6596866e8bc8aa0a366f58ca2ba37c0efbe56bd696270a729.jpg) +A Japanese speaker writing the caption +部屋には男性が座り、近くには荷物くださいますようお願いいたします。 +A man is sitting in a room, and there are several tables filled with luggage nearby. +An English speaker writing the caption Several suitcases leaned on large table, with several backpacks on top +Figure 1. People speaking different languages may caption images differently, noticing and emphasizing different aspects of the image. These examples are drawn from our user study. In this paper, we demonstrate that there are distributional differences between the concepts represented in different languages, in addition to the general variation in annotator subjectivity/noise. Illustrative example. + +nation/purpose, the mode of transportation, and other possibly visually salient variables. Similar factors associated with different languages might encourage the expression of different visual concepts in descriptions. If true, our hypothesis suggests that models trained on English-only data would be further improved if exposed to a more global view. + +In this paper, we investigate whether captions produced by people (in datasets) and by models tend to vary semantically and expressively across different languages. In our context, semantics refers to a caption's content (what a caption "says") as represented by scene graphs (objects, attributes, relations) [55, 63] and expression refers to how its content is communicated as measured by linguistic measures such as concreteness, authenticity, and tone [12]. We measure the variation of semantic and expression measurements between multilingual and monolingual sets of captions. If we systematically observe a greater variation in multilingual caption sets vs. monolingual sets (for example, if scene graphs created from multilingual caption sets are larger than monolingual ones), then this demonstrates that captions have different distributions of concepts in different languages. We analyze the human-produced captions from the high-quality Crossmodal dataset across seven diverse languages (English, French, German, Russian, Chinese, Japanese, Korean). The Crossmodal dataset keeps annotator instructions consistent across languages, ensuring fair comparison of multilingual and monolingual caption sets in our analysis by eliminating possible confounding sources of variation across languages. + +Our results demonstrate that multilingual caption sets demonstrate higher semantic and expression variation than monolingual caption sets. Importantly, while it is known that different people and models produce fairly widely varying captions in general [9, 50], our claim is that there are also broad distributional differences between captions produced in different languages, when evaluated over large numbers of samples. Multilingual scene graphs are larger overall (that is, they cover more content) than monolingual + +scene graphs, with an increase of $46.0\%$ objects, $66.1\%$ relationships, and $66.8\%$ attributes compared to scene graphs built from English-only captions. Multilingual caption sets also exhibit broader expressive variation, such as an increase of $53.4\%$ in range of tonality and $42.1\%$ in coverage of the embedding space. We extend this analysis to vision-language model behaviors (LLaVA and the Vertex API imagertext-001), and observe similar patterns in model-generated captions: multilingual scene graphs are larger than monolingual scene graphs overall by $21.9\%$ objects, $18.8\%$ relations, and $20.1\%$ attributes, and multilingual caption sets have $63.0\%$ wider range of tonality and nearly $92.4\%$ wider coverage of the embedding space. + +We further analyze how finetuning on linguistically diverse data affects a model's captioning capabilities. Models are often finetuned on captions produced by humans (in datasets) or by other models (distillation). To understand this, we fine-tune models on captions from one language and evaluate them on captions from another language. Models finetuned on language $X$ perform significantly better on the test set from language $X$ , suggesting that models internalize language-specific distributional characteristics. For example, a model finetuned on captions translated from Japanese attains a SPICE F-score of .27 on reference captions translated from Japanese, but only .23 from English. + +In sum, our primary contribution is providing evidence that captions produced by people and models vary by language. To do so, we give specific measurement methods on which multilingual distributions of captions have wider variation in content or information than monolingual distributions. Our work may help reframe the "curse of multilinguality" by emphasizing the diverse range of visual concepts in multilingual datasets. We are limited by our specific seven languages and mid-sized scale, but we believe our study provides a focused demonstration of the properties of multilingual data. Our work provides one possible explanation to findings in recent work [37] — that training on multilingual data at large scale improves vision represent- + +tations — by understanding how the information in multilingual caption distributions might differ from monolingual ones, and relatedly, how monolingual captions differ across languages. See §6 for related work and §7 for limitations. + +# 2. Measuring variation in datasets + +We want to understand how captions in computer vision datasets vary across languages. Variation across languages in datasets may have downstream effects in associated benchmarks and downstream models. We choose to perform a close analysis of the Crossmodal ("XM") dataset [115] because of its high quality and well-designed data annotation procedures. XM contains image descriptions in 36 languages over $3.6\mathrm{k}$ geographically diverse images. To eliminate potential bias in captioning instructions across annotators (e.g., varying instructions in different languages), all annotators are provided with the same annotation instruction and required to be reading-proficient in English (in addition to being native/fluent in the target language). To ensure consistency in caption style, annotators across languages are primed with the same base caption. These procedures ensure that the XM dataset serves as a prime exemplar to fairly study differences between captions produced in different languages for the same sets of images. + +Throughout the paper, we analyze captions produced in seven languages: English, French, German, Russian, Chinese, Japanese, and Korean. Together, these 7 languages encompass a wide range of typologically diverse linguistic families. Moreover, the speakers of these languages originate from a large variety of cultures and experiences. + +# 2.1. Translation for fair comparison + +We translate all captions into English with GPT-4 [86] (gpt-4-0613) to ensure fair comparison of caption content across languages by eliminating linguistic confounders. Our measurements of semantic ( $\S 2.2$ ) and expressive ( $\S 2.3$ ) variation use tools like parsers, embedding models, and tokenizers which are either language-specific (e.g., entirely separate parsers for different languages [17, 95]) or language-biased (e.g., multilingual embeddings encode language-specific information [18, 73, 87, 91]). When all captions are translated into English, we can analyze on the deeper differences in caption content and expression on "common ground", without these linguistic confounders. + +We chose GPT-4 because we can specify the manner and procedure for translation in the prompt to a degree not available for many other models; specifically, in our case it was more important for translations to be exact and preserve all factual details than to adhere to other translation values and metrics [15]. Several works have demonstrated GPT-4's competence and control on multilingual tasks [54, 70, 79, 99, 129], particularly in capturing nuances of expression and meaning specific to a language. + +We conduct a human evaluation to verify that the translations preserve salient conceptual information. We recruit 15 native speakers to rate 30 translations across each of the 6 non-English languages. Native speakers describe the translations as near perfect ( $\mu = 4.68$ on a 1-5 scale, where 5 is "perfect" and 4 is "most information preserved"). Importantly, participants considered $98.42\%$ of important information in a visual scene (objects, spatial relations, color, etc.) as faithfully preserved across translation. See Appendix 8.1 for more details. We also verify statistically that variation among translated captions is not explained by the original language of the captions in Section 4. Participants were allowed to give open-ended feedback in the user study, and we find that even the most extreme corrections were minor - e.g. replacing "on" with "above" and "memories" with "impressions" - and therefore do not noticeably alter measurements. Therefore, we believe the translations did not introduce significant analysis-biasing artifacts. + +Nonetheless, we repeat the representation variance experiment described in §2.3 and §3.4 with multilingual embeddings without translation and find that the same result holds as described (using English embeddings with translation) (see §9.2). This shows that our results still hold when using tools that can be theoretically applied across languages (even though they are language-biased). We describe all of our experiments in the main paper under the assumption of English translation for fair comparison. + +# 2.2. Semantic variation across languages + +We want to understand how captions across languages vary in their semantic content. A caption's semantic content is what it says about the objects in the image and how they relate to each other. We ask: do captions across languages mention the same objects and relations, or are there systematic differences? Scene graphs are a common representation for this kind of information [22, 53, 55]. + +Method: Formally, a scene graph $\mathcal{G}$ is defined as a list of tuples of the form (object, attribute) or (subject, predicate, object). Let $|\mathcal{G}|_{\mathrm{obj}}$ , $|\mathcal{G}|_{\mathrm{rel}}$ , and $|\mathcal{G}|_{\mathrm{attr}}$ be the unique number of objects, relations, and attributes in a scene graph $\mathcal{G}$ , respectively. These give measurements for the 'size' of $\mathcal{G}$ . Let $\mathsf{SG}(c)$ construct a scene graph from a caption $c$ , and $\mathcal{G}_1 \cup \mathcal{G}_2$ represent the "union" of scene graphs — a new scene graph with shared content (objects, relations, attributes) only counted once. For each image $i$ out of all images $I$ in the dataset, there is a corresponding set of captions $C_i$ , where each caption may be denoted by $c_{i,k}^l$ , where $l$ is a language and $k$ is the caption index for captions in that language. Let $\mathcal{L} = \{\text{en, fr, de, ru, zh, ja, ko}\}$ be the set of languages. We create a monolingual set of captions for language $l$ : $\text{mono}_i^l = \{c_{i,1}^l, \ldots, c_{i,K}^l\}$ and a multilingual set of captions for a set of languages $L$ : $\text{multi}_i^L = \{c_{i,1}^{L_1}, \ldots, c_{i,1}^{L_K}\}$ . Let + +Table 1. XM (dataset) evaluation results. Results compare monolingual (e.g., all-English) against multilingual (e.g., English, French, Chinese) caption sets against both semantic ("Sem.") and selected expressive ("Expr.") measurements. The multilingual "avg" column indicates the average of all multilingual language three-tuples, since not all can be displayed. Observe that multilingual measurements are almost always larger than monolingual ones. See all data aggregated at Table 10. + +
MetricMonolingualMultilingual
enfrderuzhjakoavgen-fr-zhfr-zh-rude-fr-ruavg
Sem.Objects2.592.923.163.032.993.412.712.983.713.923.934.35
Relations1.541.761.941.881.711.991.591.772.412.572.572.94
Attributes1.271.661.971.742.012.471.461.782.362.592.552.97
Expr.Concreteness1.802.242.112.032.262.252.142.122.082.132.102.17
Authenticity33.2140.0942.2232.3534.0233.8335.2835.8653.1254.4153.3553.21
Tone8.6212.169.7410.9010.599.189.0710.0413.7814.6915.4215.40
Embeddings.38.42.43.40.46.42.43.42.54.54.49.52
+ +$M(\mathcal{G}) \in \{|\mathcal{G}|_{\mathrm{obj}}, |\mathcal{G}|_{\mathrm{adj}}, |\mathcal{G}|_{\mathrm{attr}}\}$ be a metric for scene graph size. We want to compare $\mathbb{E}_{i \in I, l \in \mathcal{L}}[M(\bigcup \mathsf{SG}(\mathrm{mono}_i^l))]$ (mean size of scene graphs from monolingual captions) against $\mathbb{E}_{i \in I, L \subset \mathcal{L}; |L| = K}[M(\bigcup_{l \in L} \mathsf{SG}(\mathrm{multi}_i^L))]$ (mean size of scene graphs from multilingual captions). If the latter is significantly larger than the latter, then we conclude that, there tend to be distributional non-overlaps/differences across the concepts covered in captions across languages (apart from differences across captions due to general annotator subjectivity). To implement $\mathsf{SG}(c)$ which constructs a scene graph from a caption $c$ , we parse translated XM captions into scene graphs using FLAN-T5 [23] model finetuned on the FACTUAL-MR dataset. We find that LLM-based parsers were better able to resolve complex semantic relationships which often arose in the descriptions than conventional syntax-based parsers [105]. To implement the union of scene graphs, we concatenate their list representations and remove double counting of shared concepts (objects, relations, or attributes). We canonicalize the concepts using WordNet path similarity [81] and cosine similarity between concept embeddings [100] to merge concepts which have different text but refer to the same concept, erring on the side of merging. Multiple relationships between two entities are merged into a single edge. + +Results: We find that multilingual unioned scene graphs cover more objects, relations, and attributes than monolingual ones (Table 1). For instance, unioning the scene graph from an English caption with the scene graphs from a French and a Chinese caption increase the resulting scene graph by $46.0\%$ objects, $66.1\%$ relations, and $66.8\%$ attributes. Further computing the sizes of intersections between monolingual unioned scene graphs shows that any two languages share only $63.1\%$ of objects and $39.5\%$ of relations in common on average (Table 11). Since image captions are short and targeted to focus on salient parts of the image, these observed differences between languages likely indicate different tendencies in what counts as salient across caption sets written by humans who speak different lan + +guages. Qualitatively, we observe that many of the objects represented in multilingual scene graphs but not in English scene graphs tend to be background objects and objects at varying levels of object composition (e.g. keyboard and CTRL key). Moreover, multilingual scene graphs tend to include more information about perspective (e.g. "zooming", "fore/background"), color, size, and other details. See Figures 2 and 5 for detailed examples. Importantly, this demonstrates that there are distributional non-overlaps between the concepts represented in captions across languages. + +# 2.3. Expressive variation across languages + +We also want to understand how captions across languages vary by expression. Expression refers to how concepts are conveyed through language. + +Method: We consider five linguistic measures of expression: Concreteness is computed across noun objects and indicates how much a word refers to a perceptible entity [13]. For instance, "purple" has a high concreteness rating (4.04 on a 5-point scale), whereas "beautiful" has a low one (2.16). Text is analytic when it expresses logical and hierarchical thinking patterns, demonstrates clout when it exhibits social status and confidence, is authentic when it is spontaneous, and varies by tone. These last four measures are provided by the Linguistic Inquiry and Word Counts (LIWC) framework [112], and are computed using word dictionaries and aggregation algorithms informed by linguistic and psychological research [12]. Many previous works have demonstrated the use of these measures in the quantification of emotional and psychological expression in text from a variety of contexts [1, 56, 57, 102]. A sixth measure is a model embedding, which captures some semantic but also important expressive dimensions of text [5, 20, 25]. + +To measure expressive variation, we compare the expressive coverage of multilingual caption sets and monolingual caption sets, where the expressive coverage of a set of text $\mathcal{T}$ as measured by a measurement $M$ is $C_M(\mathcal{T}) = \max (M(\mathcal{T})) - \min (M(\mathcal{T}))$ , or the maximum width of $M$ spanned by elements of $\mathcal{T}$ . Note that because em + +beddings are not one-dimensional, expressive coverage is measured as the maximum pairwise cosine distance between two points in a set. Like previously, we compare $\mathbb{E}_{i\in I,l\in \mathcal{L}}[C_M(\mathrm{mono}_i^l)]$ and $\mathbb{E}_{i\in I,L\subset \mathcal{L}}[C_M(\mathrm{multi}_i^L)]$ . Following the interpretation methodology from §2.2, if the multilingual sets have more expressive coverage than monolingual sets, then this demonstrates some non-overlap in the distribution of captions' expressive dimensions (as measured by the aforementioned metrics). + +Results: Our results demonstrate variance in caption expression across languages. We find that across all metrics, expressive coverage is generally higher for multilingual rather than monolingual subsets (Table 1). For example, the range of tonality in caption sets widens by $53.4\%$ for multilingual sets. We also find that the mean width of caption sets in model representations is larger for multilingual sets than monolingual sets. For example, the average coverage in embedding space of English XM captions increases from 0.38 to 0.54 when switching to an English, French, and Chinese caption set. This means that training with captions from multiple languages as opposed to one language may expose the model to a measurably wider range of caption expressiveness. + +# 3. Measuring variation in model outputs + +In §2, we demonstrated semantic and expressive variations in the XM dataset. These results make sense in light of social sciences studies that show how linguistic factors associated with different languages encourage the production of specific kinds of information. However, we also want to study if multilingual vision models also exhibit similar differences when producing captions in different languages. We ask the question: "what kinds of information differences might there be across captions produced by models in different languages for the same images?" + +We study two models: LLaVA and the Google Vertex API. LLaVA [75] is a vision-language model capable of generating multilingual captions. Although LLaVA is trained with English data, it inherits multilingual capabilities from its large language model (LLM) component, LLaMA [119]. In total, we generate 54k captions for 3.6k Crossmodal images in 5 languages with 3 captions each, excluding Japanese and Korean for text quality issues. LLaVA allows us to study the significance of language factors across descriptions in an isolated fashion, because the vision representations are held constant across text prompts in different languages. See Appendix 8.2 for more details on probing multilingual behavior. The Google Vertex API is widely adopted to generate multilingual captions for images. The API outputs reflect established industry standards, having gone through extensive quality checks. + +# 3.1. Semantic variation across languages + +Method: We apply the same method as described in §2.2 to measure semantic variation, but substituting XM captions for LLaVA and Vertex-generated captions. + +Results: We find that multilingual unioned scene graphs cover more objects, relations, and attributes than monolingual ones (Table 2). For instance, unioning the scene graph from an English caption with the scene graphs from a French and a Chinese caption increase the resulting scene graph by $35.2\%$ objects, $28.0\%$ relations, and $45.6\%$ attributes for LLaVA captions; and by $13.2\%$ objects, $16.6\%$ relations, and $37.1\%$ attributes for Vertex API captions. The smaller increase for Vertex API captions is likely explained by the more standardized (lower-variance) captions produced by Vertex API compared to LLaVA. Figure 2 displays an example of such a scene graph. + +# 3.2. Can multilingual variation be explained by 'different language modes'? + +There may be a concern that multilingual captions generated in different languages by models are diverse because they are ' (built to) operate in different modes' for each language; for example, tokens from different languages might 'activate' language-specific representation axes [18], or APIs might route caption requests in different languages to model instances finetuned/trained on different data. Can all of the increase in concept coverage (as measured by larger scene graphs) for multilingual captions over monolingual captions be explained solely by models being built to work in 'different modes' for different languages? If the answer is 'yes', then the increase seems uninteresting. Although there are difficulties in measuring this, we make an argument that the answer is 'no'. To do so, we compare multilingual captions produced by one model (LLaVA or Vertex API) against English captions produced by multiple models (e.g., LLaVA, GIT[121], BLIP [71]) and find that their variation/coverage to be similar. One would expect the variation among multilingual captions produced by one model to be substantially smaller than monolingual captions produced by different models, since the difference between caption variation due to 'different language modes' inside one model should be much smaller than the caption variation due to totally different training procedures, model architectures, etc. across different models that happen to be in the same language [27, 126]. However, if the variation between the two is roughly the same, then we cannot explain the variation among multilingual captions produced by one model to the model operating in 'different language modes'. + +Method: Formally, let $c_{i,m}^{l}$ be a caption produced by model $m$ for image $i$ in language $l$ . Let $m$ either be LLaVA or Vertex. For a general variation metric $\mathcal{M}$ (e.g. size of unioned scene graphs or expressive coverage), we want to compare $\mathbb{E}_{i\in I,L\in \mathcal{L};|L| = 3}[\mathcal{M}(\{c_{i,m}^{l}:l\in L\})]$ (variation metric $\mathcal{M}$ + +Table 2. LLaVA and Vertex (model output) evaluation results. Results compare monolingual (e.g., all-English) against multilingual (e.g., English, French, Chinese) caption sets against both semantic ("Sem.") and selected expressive ("Expr.") measurements. The multilingual "avg" column indicates the average of all multilingual language three-tuples, since not all can be displayed. Observe that multilingual measurements are almost always larger than monolingual ones. See all data aggregated at Table 10. + +
ModelMetricMonolingualMultilingual
enfrderuzhjakoavgen-fr-zhfr-zh-rude-fr-ruavg
LLaVASem.Objects4.545.055.264.524.54--4.786.146.156.255.93
Relations3.794.214.423.673.66--3.954.854.764.974.54
Attributes2.753.473.502.763.25--3.154.003.994.073.86
Expr.Concreteness2.172.442.532.272.33--2.352.542.542.572.56
Authenticity35.9738.0134.3337.2044.63--38.0354.9454.6454.8253.15
Tone6.799.5316.6416.4811.56--12.2016.7420.7919.8016.56
Embeddings.22.29.28.29.36--.29.45.47.43.47
VertexSem.Objects3.653.513.603.863.463.133.183.484.134.244.134.17
Relations2.962.832.893.202.682.372.472.773.453.483.383.40
Attributes1.671.671.791.861.661.591.621.982.292.402.332.33
Expr.Concreteness1.641.641.671.661.511.501.561.601.751.741.731.81
Authenticity23.2122.8021.6823.0723.5125.0121.6722.9940.1636.9431.8538.06
Tone1.851.842.032.632.051.962.052.063.984.104.193.92
Embeddings.19.18.19.22.20.19.37.22.37.33.38.49
+ +across multilingual captions produced by the same model $m$ ) and $\mathbb{E}_{i\in I}[\mathcal{M}(\{c_{i,m}^{\mathrm{en}},c_{i,\mathrm{GIT}}^{\mathrm{en}},c_{i,\mathrm{BLIP}}^{\mathrm{en}}\})]$ (variation metric $\mathcal{M}$ across English captions produced by different models). If these two values are similar, then the variation in LLaVA and Vertex's multilingual output cannot be explained by models operating in 'different language modes'. + +Results: Our results suggest that demonstrated differences in scene graphs constructed from models' captions are not entirely explained by 'different language modes'. Interestingly, we find that the same-model across-language scene graphs are only slightly smaller than the same-language across-model scene graphs: $92.4\%$ objects, $98.4\%$ relations, and $97.7\%$ attributes. See Table 12 for more details. This supports the significance of caption content differences across languages. Furthermore, we find that caption sets generated by one model (Vertex) in different languages have large proportions of concepts not covered by English GIT and BLIP models: $47.8\%$ of objects, $71.4\%$ of relations, and $59.5\%$ of attributes. See Table 13 for more details. + +# 3.3. Grounding variation in scene graphs + +Another concern may be that captions generated by a model might contain concepts that are not 'objectively' grounded in the image. For example, one caption might mention a 'tree', another might mention 'tree trunk' and 'branches', possibly to describe specific attributes of each (e.g., the 'tree trunk is brown, and the branches are black') or describe a unique relation between them (e.g., 'the branches are all on the same side of the tree trunk'). Although these types of differences are important, since they reflect complementary information about the content of the image, if variation across caption content and expression can be ex + +plained solely by variation induced by these kinds of 'subjective' focuses, then one could argue it is not as significant or interesting. Therefore, we reproduce our scene graph experiment to measure content variation only on object sets documented in the Visual Genome object list [63] for that image. Since the Visual Genome object list is standard and shared across all models, it provides an 'objective' grounding for measuring the size of scene graphs by objects. + +Method: Formally, let $VG_{i}$ be the Visual Genome object set for an image $i$ . Recall from §3.1 that $\mathrm{mono}_{i}^{l}$ and $\mathrm{multi}_{i}^{L}$ are the monolingual and multilingual caption sets using language $l$ and language set $L$ , respectively, for image $i$ . Let $\mathbf{Obj}(C)$ take in a caption set and output the unique objects (as if constructing a unioned scene graph and then extracting only the objects). We want to compare $\mathbb{E}_{i\in I,l\in \mathcal{L}}[|\mathbf{Obj}(\mathrm{mono}_{i}^{l}\cap VG_{i})|]$ (# unique VG objects covered by monolingual caption set) and $\mathbb{E}_{i\in I,L\in \mathcal{L}}[|\mathbf{Obj}(\mathrm{multi}_{i}^{L}\cap VG_{i})|]$ (# unique VG objects covered by multilingual caption set). If the latter is significantly larger than the former, then this is a stronger confirmation that there tend to be distributional non-overlaps/differences across concepts covered in captions across languages. + +Results: The results demonstrate, like before, that multilingual scene graphs capture more Visual Genome objects/concepts than monolingual scene graphs (see Table 3). For example, unioning an English scene graph with a French and Chinese scene graph captures $23.9\%$ more Visual Genome-annotated objects than unioning with two other English scene graphs. Interestingly, we also find some evidence that multilingual scene graphs are able to identify objects which are unidentified even in Visual Genome's dense annotations (App. Table 16). + +Table 3. Ground truth coverage $\left( {\left| {\mathbb{C} \cap \mathbb{G}}\right| /\left| {\mathbb{G}}\right| }\right)$ increases when sampling multilingual captions. $\mathbb{C}$ refers to the caption concept set and $\mathbb{G}$ refers to the 'ground truth' Visual Genome concept set. $\left| {\mathbb{C} \cap \mathbb{G}}\right|$ (unnormized intersection size,number of objects shared) is provided for reference. 'avg' is the mean across all triplets of languages. + +
MonolingualMultilingual
3×en3×de3×fr3×ru3×zh3×ja3×koen,fr,zhfr,ru,zhde,fr,ruavg
Vertex|C∩G|/|G|9.6%9.7%9.9%10.3%9.8%9.9%9.3%11.7%11.7%11.2%11.5%
|C∩G|1.551.601.621.691.601.591.491.921.921.841.90
LLaVA|C∩G|/|G|9.5%10.9%11.6%9.3%9.1%--12.0%12.1%12.5%12.0%
|C∩G|2.863.303.552.832.75--3.713.703.863.69
+ +![](images/571193cf81308316515b5a5edd4408f9628e85251f521680b5e2096a5b691c49.jpg) +Figure 2. Semantic content evaluation. Captions of an image in different languages and their scene graphs, when unioned together produce more varied and complex scene graphs for multilingual distributions than monolingual ones. Captions from Vertex. + +# 3.4. Expressive variation across languages + +Method: We apply the same method as described in §2.3 to measure semantic variation, but substituting XM captions for LLaVA and Vertex-generated captions. + +Results: Our results demonstrate variance in models' caption expression across languages. We find that across all metrics, expressive coverage is generally higher for multilingual rather than monolingual subsets (Table 2). For example, the range of noun concreteness in caption sets widens by $8.9\%$ for multilingual sets on average for LLaVA, and by $13.1\%$ for Vertex. We also find that the mean width of caption sets in model representations is larger for multilingual sets than monolingual sets. For example, the average coverage in embedding space of English LLaVA captions increases from 0.22 to 0.45 when switching to an English, French, and Chinese caption set, and from 0.19 to 0.37 for Vertex. + +# 4. Effects of Fine-tuning Across Language + +In §2 and §3, we showed that dataset and model captions exhibit semantic and expressive variation across languages. What effect does this have on how we train and fine-tune captioning models? Do models internalize the specific dis + +tributional characteristics of captions from one language? + +Method: To evaluate this, we finetune a pretrained vision-language model, GIT [121] - chosen for its simple architecture and robust performance across benchmarks, on 8 constructed image caption sets. These finetuning datasets use the same set of 1.8k training images from the Crossmodal dataset ("XM") [114]. The first 7 training sets contain translations from a single language (e.g. 1.8k captions all from French, 1.8k captions all from Chinese). The eighth dataset is constructed with equal proportions of captions across all 7 languages. We inference the finetuned models on a holdout set of 1.8k images from XM and evaluate performance against each of the corresponding 8 sets of image caption labels. We use SPICE F1-score [4], which measures conceptual overlap over low-level syntax adherence, as our evaluation metric. Validating each of the 8 fine-tuned models on each of the 8 holdout caption sets yields 64 evaluations. We perform this entire process on captions from XM, LLaVA, and Vertex. + +We interpret our results using the following logic. Let $S(A; B)$ represent the score of a model fine-tuned on the data from language $B$ on the evaluation data from language $A$ (suppose bigger is better). If we observe superior model test performance on the test set corresponding to the lan + +Table 4. Model outputs evaluation results. SPICE F-scores when evaluating a model fine-tuned on the training set from the language on the left against the validation set from the language on the top. 'multi' refers to an even split across all languages. Red: best performance on a split, yellow highlights model fine-tuned on 'multi'. + +
Evaluated on
endefrruzhjakomulti
Fine-tuned onen0.2710.2250.2290.2190.2180.2290.2320.230
de0.2130.2450.2190.2170.2150.2100.2260.219
fr0.2480.2400.2590.2340.2360.2390.2530.246
ru0.2260.2340.2280.2540.2310.2360.2370.239
zh0.1990.2020.1990.2070.2470.2200.2240.216
ja0.2120.2120.2150.2120.2260.2660.2450.223
ko0.2180.2220.2240.2170.2420.2390.2710.235
multi0.2390.2330.2340.2330.2350.2430.2520.244
+ +guage it was trained on relative to other test sets (that is, $\forall_{A\in \mathcal{L},B\in \mathcal{L};A\neq B}\mathcal{S}(A;B)\ll \mathcal{S}(A;A))$ , the model has internalized distributional characteristics of each language. + +Results: We find that a model fine-tuned on language $X$ performs best on language $X$ (Table 4) for Vertex captions. We find the same results across models finetuned on data from LLaVA (Table 14), and from the human annotated XM dataset (Table 15), including when evaluating with default loss rather than SPICE F-score. Moreover, finetuning on a multilingual collection of captions yields a model which performs consistently well across all evaluation data compositions. In the case of training on human-generated multilingual captions (from the Crossmodal dataset), this model is the second-best model for 5/7 holdout-set languages out of 8 models, coming only behind a model finetuned on that language itself. Together, these suggest that models trained with multilingual data may represent diverse perceptual modes and learn more about the visual world. + +# 5. Discussion + +# 5.1. Implications + +Reframing the "curse of multilinguality". We contribute another framing to the ongoing debate related to the "curse of multilinguality." On one hand, arguments in favor of multilingual language and vision models are often made from an accessibility perspective of expanding information and technology access to individuals [47]. Meanwhile, empirical evidence suggests that increasing the number of languages a model handles degrades monolingual performance [26]. Multilingualism is, seen as a trade-off against model performance. Our work provides support for the possibility of a system benefit from including multilingual data. Using "organically" / human-produced multilingual data introduces substantive changes in the semantic concepts mentioned for the same images. Thus, incorporating the diverse content distribution may improve model ro + +bustness [35, 37, 58, 96]. $\circ$ Translating English content to other languages may be missing some information. Many efforts to build vision-language models for a non-English language $X$ train on translations from English to $X$ [19, 36, 41, 82, 90, 93]. Although these systems are usable in $X$ , our investigation suggests that they may miss out on the unique characteristics of an "organically $X$ content distribution (i.e., that native $X$ -speakers would provide). Indeed, other works which collect captions from $X$ -speakers and train models on them suggest quantitative and qualitative improvements over training from English-to- $X$ translated captions [38, 125]. Our work provides an explanation for such observations. $\circ$ Wariness of a "perceptual monoculture". The English-dominated language modeling landscape may not only introduce accessibility issues for non-English speakers but also possibly a "perceptual monoculture" [60]. In these ways, our contribution shares motivation and aims with a diverse array of previous works in how models and datasets encode and internalize particular limiting and possibly harmful knowledge, such as by gender [14], nationality [101], and language [2, 122]. + +# 5.2. Recommendations + +Multilingual modeling from organically multilingual data. Currently, many multilingual vision-language models are built by training models on captions translated from English into the desired language. When possible, multilingual vision-language applications should be built by training models collected from native speakers. Our results suggest that the advantage is not only syntactic (i.e. more "natural-sounding" than translations), but that annotations produced in the desired language have specific distributional semantic and expressive characteristics. + +Multilingual quality as visual descriptive diversity. Multilingual vision-language datasets should be explored as repositories of perceptually diverse annotations. At a basic level, simply training models on multilingual rather than monolingual vision datasets may expose them to a more diverse, rich set of concepts. At a more complex level, specifically designed structured learning paradigms such as contrastive learning may be able to exploit concept diversity across multilingual captions to yield richer image representations. + +# 5.3. Conclusion + +This paper demonstrates semantic and expressive in image captions across languages, revealing meaningful differences in how visual content is represented. Our results support that training models on multilingual captions exposes them to a wider range of semantic and expressive material. This increased range may result in more robust vision representations [37]. By embracing the richness that different languages bring, we can move toward more inclusive, nuanced models that better reflect the diversity of human perception. + +# Acknowledgements + +We thank Katharina Reinecke for the many insightful discussions that inspired this work; Quan Ze (Jim) Chen, Matthew Wallingford, Sarah Pratt, Thao Nguyen, Hyunwoo Kim, Wei-Chiu Ma, Chris Lin, Lisa Orii, Nino Migineishvili, Ashish Sharma, Jae Sung (James) Park and members of RAIVN lab for the multiple rounds of discussion, feedback and reviewing; Cheng-Yu Hsieh and Ashish Sharma for code/package assistance; Allen AI for running the surveys; and Chun-Liang Li at Google Cloud for running the Vertex API. We also thank participants of our study for carefully captioning the images. + +# References + +[1] Sanaz Aghazadeh, Kris Hoang, and Bradley Pomeroy. Using LIWC to Analyze Participants' Psychological Processing in Accounting JDM Research. AUDITING: A Journal of Practice & Theory, 41(3):1-20, 2022. 4 +[2] Badr AlKhamissi, Muhammad N. ElNokrashy, Mai AlKhamissi, and Mona Diab. Investigating cultural alignment of large language models. 2024. 8 +[3] Amith Ananthram, Elias Stengel-Eskin, Carl Vondrick, Mohit Bansal, and Kathleen McKeown. See it from my perspective: Diagnosing the western cultural bias of large vision-language models in image understanding, 2024. 1 +[4] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In Computer Vision - ECCV 2016, pages 382-398, Cham, 2016. Springer International Publishing. 7 +[5] Mikel Artetxe, Gorka Labaka, Inigo Lopez-Gazpio, and Eneko Agirre. Uncovering divergent linguistic information in word embeddings with lessons for intrinsic and extrinsic evaluation. In Proceedings of the 22nd Conference on Computational Natural Language Learning, pages 282-291, Brussels, Belgium, 2018. Association for Computational Linguistics. 4 +[6] Jesse Atuhurra, Iqra Ali, Tatsuya Hiraoka, Hidetaka Kamigaito, Tomoya Iwakura, and Taro Watanabe. Constructing multilingual visual-text datasets revealing visual multilingual ability of vision language models, 2024. 1 +[7] James Atwood, Yoni Halpern, Pallavi Baljekar, Eric Breck, D Sculley, Pavel Ostyakov, Sergey I Nikolenko, Igor Ivanov, Roman Solovyev, Weimin Wang, et al. The inclusive images competition. In *The NeurIPS'18 Competition: From Machine Learning to Intelligent Conversations*, pages 155–186. Springer, 2020. 1 +[8] Alexander C. Berg, Tamara L. Berg, Hal Daumé, Jesse Dodge, Amit Goyal, Xufeng Han, Alyssa C. Mensch, Margaret Mitchell, Aneesh Sood, Karl Stratos, and Kota Yamaguchi. Understanding and predicting importance in images. 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3562-3569, 2012. 1 +[9] Philipp Blandfort, Tushar Karayil, Damian Borth, and Andreas R. Dengel. Image captioning in the wild: How people caption images on flickr. Proceedings of the Workshop on + +Multimodal Understanding of Social, Affective and Subjective Attributes, 2017. 2 +[10] Lera Boroditsky. Linguistic relativity. Encyclopedia of cognitive science, 2006. 1 +[11] Lera Boroditsky, Lauren A Schmidt, and Webb Phillips. Sex, syntax, and semantics. Language in mind: Advances in the study of language and thought, 22:61-79, 2003. 1 +[12] R. L. Boyd, A. Ashokkumar, S. Seraj, and J. W. Pennebaker. The development and psychometric properties of liwc-22. Technical report, University of Texas at Austin, Austin, TX, 2022. 2, 4 +[13] Marc Brysbaert, Amy Beth Warriner, and Victor Kuperman. Concreteness ratings for 40 thousand generally known english word lemmas. Behavior Research Methods, 46:904-911, 2014. 4 +[14] Aylin Caliskan, Pimparkar Parth Ajay, Tessa Charlesworth, Robert Wolfe, and Mahzarin R. Banaji. Gender bias in word embeddings: A comprehensive analysis of frequency, syntax, and semantics. In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society, page 156-170, New York, NY, USA, 2022. Association for Computing Machinery. 8 +[15] Asli Celikyilmaz, Elizabeth Clark, and Jianfeng Gao. Evaluation of text generation: A survey, 2021. 3 +[16] Jiri Cenek and Cenek Sasinka. Cross-cultural differences in visual perception. The Journal of Education, Culture, and Society, 2015:187-206, 2020. 1 +[17] Pi-Chuan Chang, Huihsin Tseng, Dan Jurafsky, and Christopher D. Manning. Discriminative reordering with Chinese grammatical relations features. In Proceedings of the Third Workshop on Syntax and Structure in Statistical Translation (SSST-3) at NAACL HLT 2009, pages 51-59, Boulder, Colorado, 2009. Association for Computational Linguistics. 3 +[18] Tyler A. Chang, Zhuowen Tu, and Benjamin K. Bergen. The geometry of multilingual language model representations, 2022. 3, 5 +[19] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel M. Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. ArXiv, abs/2209.06794, 2022. 8, 1 +[20] Yanqing Chen, Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. The expressive power of word embeddings, 2013. 4 +[21] Zhongzhi Chen, Guang Liu, Bo-Wen Zhang, Fulong Ye, Qinghong Yang, and Ledell Wu. Altclip: Altering the language encoder in clip for extended language capabilities, 2022. 1 +[22] Woo Suk Choi, Yu-Jung Heo, Dharani Punithan, and Young-Tak Zhang. Scene graph parsing via Abstract + +Meaning Representation in pre-trained language models. In Proceedings of the 2nd Workshop on Deep Learning on Graphs for Natural Language Processing (DLG4NLP 2022), pages 30-35, Seattle, Washington, 2022. Association for Computational Linguistics. 3 +[23] Hyung Won Chung, Le Hou, S. Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Wei Yu, Vincent Zhao, Yanping Huang, Andrew M. Dai, Hongkun Yu, Slav Petrov, Ed Huai hsin Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models. ArXiv, abs/2210.11416, 2022. 4 +[24] Niv Cohen, Rinon Gal, Eli A. Meirom, Gal Chechik, and Yuval Atzmon. "this is my unicorn, fluffy": Personalizing frozen vision-language representations. *ArXiv*, abs/2204.01694, 2022. 1 +[25] Alexis Conneau, German Kruszewski, Guillaume Lample, Loïc Barrault, and Marco Baroni. What you can cram into a single vector: Probing sentence embeddings for linguistic properties, 2018. 4 +[26] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440–8451, Online, 2020. Association for Computational Linguistics. 8 +[27] Alexis Conneau, Shijie Wu, Haoran Li, Luke Zettlemoyer, and Veselin Stoyanov. Emerging cross-lingual structure in pretrained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6022–6034, Online, 2020. Association for Computational Linguistics. 5 +[28] William Croft and Esther J Wood. Construal operations in linguistics and artificial intelligence. Meaning and cognition: A multidisciplinary approach, pages 51-78, 2000. 1 +[29] John L. M. Dawson, Brian M. Young, and Peter P. C. Choi. Developmental influences on geometric illusion susceptibility among hong kong chinese children. Journal of Cross-Cultural Psychology, 4:49 - 74, 1973. 1 +[30] Radina Dobreva and Frank Keller. Investigating negation in pre-trained vision-and-language models. In BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, 2021. 1 +[31] Lior Elazary and Laurent Itti. Interesting objects are visually salient. Journal of vision, 8 3:3.1-15, 2008. 1 +[32] Jonathan Emami, Pierre Nugues, Ashraf Elnagar, and Imad Afyouni. Arabic image captioning using pre-training of deep bidirectional transformers. In Proceedings of the 15th International Conference on Natural Language Generation, pages 40-51, Stroudsburg, PA, USA, 2022. Association for Computational Linguistics. 1 +[33] Erkut Erdem, Menekse Kuyu, Semih Yagcioglu, Anette Frank, and Letitia Parcalabescu et al. Neural natural lan + +guage generation: A survey on multilinguality, multimodality, controllability and learning. J. Artif. Intell. Res., 73: 1131-1207, 2022. 1 +[34] Cristina España-Bonet and Alberto Barrón-Cedeno. The (undesired) attenuation of human biases by multilinguality. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2056-2077, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 1 +[35] Rohan Taori et al. Measuring robustness to natural distribution shifts in image classification, 2020. 8 +[36] Soravit Changpinyo et al. Towards multi-lingual visual question answering. ArXiv, abs/2209.05401, 2022. 8 +[37] Thao Nguyen et al. Multilingual diversity improves vision-language representations, 2024. 1, 2, 8 +[38] Yuya Yoshikawa et al. STAIR captions: Constructing a large-scale Japanese image caption dataset. In Proceedings of the 55th ACL, pages 417-421. Association for Computational Linguistics, 2017. 8 +[39] Xingyu Fu, Ben Zhou, Ishaan Preetam Chandratreya, Carl Vondrick, and Dan Roth. There's a time and place for reasoning beyond the image. ArXiv, abs/2203.00758, 2022. 1 +[40] Jiahui Gao, Yi Zhou, Philip L. H. Yu, Shafiq Joty, and Jiumiang Gu. Unison: Unpaired cross-lingual image captioning, 2022. 1 +[41] Gregor Geigle, Radu Timofte, and Goran Glavas. Babel-imagenet: Massively multilingual evaluation of vision-and-language representations. ArXiv, abs/2306.08658, 2023. 8, 1 +[42] Gregor Geigle, Abhay Jain, Radu Timofte, and Goran Glavas. mblip: Efficient bootstrapping of multilingual vision-llms, 2024. 1 +[43] Christopher S Hill. Perceptual experience. Oxford University Press, 2022. 1 +[44] Musashi Hinck, Carolin Holtermann, Matthew Lyle Olson, Florian Schneider, Sungduk Yu, Anahita Bhiwandiwalla, Anne Lauscher, Shao-Yen Tseng, and Vasudev Lal. Why do llava vision-language models reply to images in english? ArXiv, abs/2407.02333, 2024. 1 +[45] Julian Hitschler, Shigehiko Schamoni, and Stefan Riezler. Multimodal pivots for image caption translation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), page 2399-2409. Association for Computational Linguistics, 2016. 1 +[46] Rachel Hong, William Agnew, Tadayoshi Kohno, and Jamie Morgenstern. Who's in and who's out? a case study of multimodal clip-filtering in datacomp, 2024. 1 +[47] Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalization. In Proceedings of the 37th International Conference on Machine Learning. JMLR.org, 2020. 8 +[48] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A. Smith. Tifa: + +Accurate and interpretable text-to-image faithfulness evaluation with question answering. ArXiv, abs/2303.11897, 2023. 2 +[49] Edmund Husserl. Cartesian Meditations: An Introduction to Phenomenology. Routledge, London, UK, translated by dorion cairns edition, 2012. 1 +[50] Ben Hutchinson, Jason Baldridge, and Vinodkumar Prabhakaran. Underspecification in scene description-to-depiction tasks. arXiv preprint arXiv:2210.05815, 2022. 2 +[51] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 1 +[52] Gustav Jahoda and Harry McGurk. Pictorial depth perception: a developmental study. British journal of psychology, 65 1:141-9, 1974. 1 +[53] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatio-temporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236-10247, 2020. 3 +[54] Wenxiang Jiao, Wenxuan Wang, Jen tse Huang, Xing Wang, Shuming Shi, and Zhaopeng Tu. Is chatgpt a good translator? yes with gpt-4 as the engine, 2023. 3 +[55] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3668-3678, 2015. 2, 3 +[56] Doerte U Junghaenel, Joshua M Smyth, and Laura Santner. Linguistic dimensions of psychopathology: A quantitative analysis. Journal of Social and Clinical Psychology, 27(1), 2008. 4 +[57] Jeffrey H. Kahn, Renee M. Tobin, Audra E Massey, and Jemima Asabea Anderson. Measuring emotional expression with the linguistic inquiry and word count. The American journal of psychology, 120 2:263-86, 2007. 4 +[58] Sanjay Kariyappa and Moinuddin K. Qureshi. Improving adversarial robustness of ensembles with diversity training, 2019. 8 +[59] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision, 2021. 1 +[60] Jon M. Kleinberg and Manish Raghavan. Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences, 118, 2021. 8 +[61] Wolfgang Kohler. Gestalt psychology. Psychologische Forschung, 31(1):XVIII-XXX, 1967. 1 +[62] Minkyung Koo, Jong An Choi, and Incheol Choi. Analytic versus holistic cognition: Constructs and measurement. 2018. 1 +[63] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and + +vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32 - 73, 2016. 2, 6 +[64] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding, 2020. 1 +[65] Zhengfeng Lai, Haotian Zhang, Wentao Wu, Haoping Bai, Aleksei Timofeev, Xianzhi Du, Zhe Gan, Jiulong Shan, Chen-Nee Chuah, Yinfei Yang, et al. From scarcity to efficiency: Improving clip training via visual-enriched captions. arXiv preprint arXiv:2310.07699, 2023. 1 +[66] George Lakoff. Fire, Women, and Dangerous Things. University of Chicago Press, 1987. 1 +[67] George Lakoff and Mark Johnson. Metaphors we live by. University of Chicago press, 2008. 1 +[68] Ronald W Langacker. Universals of construal. In Annual Meeting of the Berkeley Linguistics Society, pages 447-463, 1993. 1 +[69] Ronald W Langacker. Construal. Cognitive linguistics: Foundations of language, pages 140-166, 2019. 1 +[70] Jaechan Lee, Alisa Liu, Orevaoghene Ahia, Hila Gonen, and Noah A. Smith. That was the last straw, we need more: Are translation systems sensitive to disambiguating context?, 2023. 3 +[71] 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. 1, 5 +[72] Xirong Li, Chaoxi Xu, Xiaoxu Wang, Weiyu Lan, Zhengxiong Jia, Gang Yang, and Jieping Xu. Coco-cn for crosslingual image tagging, captioning and retrieval, 2019. 1 +[73] Jindrich Libovický, Rudolf Rosa, and Alexander Fraser. On the language neutrality of pre-trained multilingual representations, 2020. 3 +[74] Fangyu Liu, Emanuele Bugliarello, Edoardo Maria Ponti, Siva Reddy, Nigel Collier, and Desmond Elliott. Visually grounded reasoning across languages and cultures. arXiv preprint arXiv:2109.13238, 2021. 1 +[75] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. ArXiv, abs/2304.08485, 2023. 5 +[76] Zhixuan Liu, You Won Shin, Beverley-Claire Okogwu, Youngsik Yun, Lia Coleman, Peter Schaldenbrand, Jihie Kim, and Jean Oh. Towards equitable representation in text-to-image synthesis models with the cross-cultural understanding benchmark (ccub) dataset. ArXiv, abs/2301.12073, 2023. 1 +[77] Zhixuan Liu, Peter Schaldenbrand, Beverley-Claire Okogwu, Wenxuan Peng, Youngsik Yun, Andrew Hundt, Jihie Kim, and Jean Oh. Scoft: Self-contrastive fine-tuning for equitable image generation. ArXiv, abs/2401.08053, 2024. 1 +[78] Ngan Luu-Thuy Nguyen, Nghia Hieu Nguyen, Duong T.D. Vo, Khanh Quoc Tran, and Kiet Van Nguyen. Evjvqa + +challenge: Multilingual visual question answering. Journal of Computer Science and Cybernetics, page 237-258, 2023. 1 +[79] Shushen Manakhimova, Eleftherios Avramidis, Vivien Macketanz, Ekaterina Lapshinova-Koltunski, Sergei Bagdasarov, and Sebastian Möller. Linguistically motivated evaluation of the 2023 state-of-the-art machine translation: Can ChatGPT outperform NMT? In Proceedings of the Eighth Conference on Machine Translation, pages 224-245, Singapore, 2023. Association for Computational Linguistics. 3 +[80] Takahiko Masuda and Richard E. Nisbett. Attending holistically versus analytically: comparing the context sensitivity of japanese and americans. Journal of personality and social psychology, 81 5:922-34, 2001. 1 +[81] George A. Miller. Wordnet: A lexical database for english. Commun. ACM, 38(11):39-41, 1995. 4 +[82] Santosh Kumar Mishra, Rijul Dhir, Sriparna Saha, and Pushpak Bhattacharyya. A Hindi image caption generation framework using deep learning. ACM Trans. Asian Low-Resour. Lang. Inf. Process., 20(2), 2021. 8 +[83] Youssef Mohamed, Mohamed Abdelfattah, Shyma Alhuwaider, Feifan Li, Xiangliang Zhang, Kenneth Ward Church, and Mohamed Elhoseiny. Artelingo: A million emotion annotations of wikiart with emphasis on diversity over language and culture. arXiv preprint arXiv:2211.10780, 2022. 1 +[84] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. arXiv preprint arXiv:2307.10350, 2023. 1 +[85] Richard E. Nisbett and Takahiko Masuda. Culture and point of view. Proceedings of the National Academy of Sciences of the United States of America, 100:11163 - 11170, 2003. 1 +[86] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 3, 1 +[87] Yulia Otmakhova, Karin Verspoor, and Joy Han Lau. Cross-linguistic comparison of linguistic feature encoding in BERT models for typologically different languages. In Proceedings of the 4th Workshop on Research in Computational Linguistic Typology and Multilingual NLP, pages 27–35, Seattle, Washington, 2022. Association for Computational Linguistics. 3 +[88] A. Pavlenko and M. Volynsky. Motion encoding in russian and english: moving beyond talmy's typology. The Modern Language Journal, 99:32-48, 2015. 1 +[89] Darhl M. Pedersen and Jay Wheeler. The Müller-lyer illusion among navajos. The Journal of social psychology, 121 1st Half:3-6, 1983. 1 +[90] Jonas Pfeiffer, Gregor Geigle, Aishwarya Kamath, Jan-Martin O. Steitz, Stefan Roth, Ivan Vulic, and Iryna Gurevych. xgqa: Cross-lingual visual question answering. In Findings, 2021. 8 +[91] Fred Philippy, Siwen Guo, and Shohreh Haddadan. Identifying the correlation between language distance and crosslingual transfer in a multilingual representation space, 2023. 3 + +[92] Jakob Prange and Nathan Schneider. Draw mir a sheep: A supersense-based analysis of german case and adposition semantics. KI-Künstliche Intelligenz, 35(3-4):291-306, 2021. 1 +[93] Chen Qiu, Dan Oneata, Emanuele Bugliarello, Stella Frank, and Desmond Elliott. Multilingual multimodal learning with machine translated text. In Conference on Empirical Methods in Natural Language Processing, 2022. 8 +[94] 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. 1 +[95] Anna Rafferty and Christopher D. Manning. Parsing three German treebanks: Lexicalized and unlexicalized baselines. In Proceedings of the Workshop on Parsing German, pages 40-46, Columbus, Ohio, 2008. Association for Computational Linguistics. 3 +[96] Vivek Ramanujan, Thao Nguyen, Sewoong Oh, Ludwig Schmidt, and Ali Farhadi. On the connection between pretraining data diversity and fine-tuning robustness, 2023. 8 +[97] Ankit Rathi. Deep learning approach for image captioning in Hindi language. In 2020 ICCECE., pages 1-8, 2020. 1 +[98] Alexander J. Ratner, Stephen H. Bach, Henry R. Ehrenberg, Jason Alan Fries, Sen Wu, and Christopher Ré. Snorkel: Rapid training data creation with weak supervision. Proceedings of the VLDB Endowment. International Conference on Very Large Data Bases, 11 3:269-282, 2017. 1 +[99] Vikas Raunak, Amr Sharaf, Yiren Wang, Hany Hassan Awadallah, and Arul Menezes. Leveraging gpt-4 for automatic translation post-editing, 2023. 3 +[100] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 2019. 4, 2 +[101] Sebastin Santy, Jenny T Liang, Ronan Le Bras, Katharina Reinecke, and Maarten Sap. Nlpositionality: Characterizing design biases of datasets and models. In Annual Meeting of the Association for Computational Linguistics, 2023. 8 +[102] Kristin L Schaefer, Jorge Rosales, and Jerrod A Henderson. Literature adventures with linguistic inquiry and word count. In 2023 ASEE Annual Conference & Exposition, 2023. 4 +[103] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirtysixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 1 +[104] Candice Schumann, Susanna Ricco, Utsav Prabhu, Vittorio Ferrari, and Caroline Pantofaru. A step toward more + +inclusive people annotations for fairness. Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, 2021. 1 +[105] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 4 +[106] Marshall H. Segall, Donald T. Campbell, and Melville J. Herskovits. The influence of culture on visual perception. 1967. 1 +[107] Merrielle Spain and Pietro Perona. Some objects are more equal than others: Measuring and predicting importance. In European Conference on Computer Vision, 2008. 1 +[108] Otilia Stretcu, Edward Vendrow, Kenji Hata, Krishnamurthy Viswanathan, Vittorio Ferrari, Sasan Tavakkol, Wenlei Zhou, Aditya Avinash, Enming Luo, Neil Gordon Alldrin, MohammadHossein Bateni, Gabriel Berger, Andrew Bunner, Chun-Ta Lu, Javier A Rey, Giulia DeSalvo, Ranjay Krishna, and Ariel Fuxman. Agile modeling: From concept to classifier in minutes. In International Conference on Computer Vision, 2023. 1 +[109] Leonard Talmy. Force dynamics in language and cognition. Cognitive Science, 12(1):49-100, 1988. 1 +[110] Leonard Talmy. Fictive motion in language and 'ception'. In Ten Lectures on Cognitive Semantics, pages 125-162. Brill, 2018. 1 +[111] Hao Tan and Mohit Bansal. LXMERT: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5100-5111, Hong Kong, China, 2019. Association for Computational Linguistics. 1 +[112] Yla R. Tausczik and James W. Pennebaker. The psychological meaning of words: Liwc and computerized text analysis methods. Journal of Language and Social Psychology, 29 (1):24-54, 2010. 4 +[113] Hamed Rezazadegan Tavakoli, Rakshith Shetty, Ali Borji, and Jorma T. Laaksonen. Paying attention to descriptions generated by image captioning models. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2506-2515, 2017. 1 +[114] Ashish V. Thapliyal and Radu Soricut. Cross-modal language generation using pivot stabilization for web-scale language coverage. ArXiv, abs/2005.00246, 2020. 7 +[115] Ashish V. Thapliyal, Jordi Pont-Tuset, Xi Chen, and Radu Soricut. Crossmodal-3600: A massively multilingual multimodal evaluation dataset. In Conference on Empirical Methods in Natural Language Processing, 2022. 3 +[116] Bart Thomee, David A. Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Commun. ACM, 59(2):64-73, 2016. 1 +[117] Michael Tomasello. Usage-based linguistics. Geeraerts, D, pages 439-458, 2006. 1 +[118] Imad Eddine Toubal, Aditya Avinash, Neil Gordon Alldrin, Jan Dlabal, Wenlei Zhou, Enming Luo, Otilia Stretcu, Hao + +Xiong, Chun-Ta Lu, Howard Zhou, Ranjay Krishna, Ariel Fuxman, and Tom Duerig. Modeling collaborator: Enabling subjective vision classification with minimal human effort via llm tool-use. In International Conference on Computer Vision and Pattern Recognition, 2024. 1 +[119] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. ArXiv, abs/2302.13971, 2023. 5 +[120] Sean Trott, Tiago Timponi Torrent, Nancy Chang, and Nathan Schneider. (Re)construing meaning in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5170-5184, Online, 2020. Association for Computational Linguistics. 1 +[121] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. ArXiv, abs/2205.14100, 2022. 5, 7 +[122] Chris Wendler, Veniamin Veselovsky, Giovanni Monea, and Robert West. Do llamas work in english? on the latent language of multilingual transformers. 2024. 8 +[123] Max Wertheimer. Experimentelle studien über das sehen von bewegung. Zeitschrift fur psychologie, 61:161-165, 1912. 1 +[124] Ludwig Wittgenstein. Philosophical Investigations. Wiley-Blackwell, New York, NY, USA, 1953. 1 +[125] Jiahong Wu, He Zheng, Bo Zhao, Yixin Li, Baoming Yan, Rui Liang, Wenjia Wang, Shipei Zhou, Guosen Lin, Yanwei Fu, Yizhou Wang, and Yonggang Wang. Large-scale datasets for going deeper in image understanding. In 2019 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2019. 8 +[126] Liming Xu, Quan Tang, Jiancheng Lv, Bochuan Zheng, Xianhua Zeng, and Weisheng Li. Deep image captioning: A review of methods, trends and future challenges. Neurocomputing, 546:126287, 2023. 5 +[127] Da Yin, Feng Gao, Govind Thattai, Michael F. Johnston, and Kai-Wei Chang. Givl: Improving geographical inclusivity of vision-language models with pre-training methods. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10951-10961, 2023. 1 +[128] Kiwon Yun, Yifan Peng, Dimitris Samaras, Gregory J. Zelinsky, and Tamara L. Berg. Studying relationships between human gaze, description, and computer vision. 2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 739-746, 2013. 1 +[129] Wenhao Zhu, Hongyi Liu, Qingxiu Dong, Jingjing Xu, Shujian Huang, Lingpeng Kong, Jiajun Chen, and Lei Li. Multilingual machine translation with large language models: Empirical results and analysis, 2024. 3 \ No newline at end of file diff --git a/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/images.zip b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4ddba6ad3c08922e5dda1ff3df413017ab243145 --- /dev/null +++ b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:180e967e57f11902838d409f8d26da3f442a3869564c55bc799f01c2257648dd +size 306388 diff --git a/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/layout.json b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3c32e63653eea2bfc2b087c6fa17832db71cde95 --- /dev/null +++ b/CVPR/2025/Semantic and Expressive Variations in Image Captions Across Languages/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddb6466b4dbd566c469bad4f1cda40838d66bec7c088347b09958aa3cd6e1df7 +size 504524 diff --git a/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/0f152076-9d57-482c-a9cc-897a0bc8eb28_content_list.json b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/0f152076-9d57-482c-a9cc-897a0bc8eb28_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a9a747f1a7238c3f3d09ea697c99e90fd528854c --- /dev/null +++ b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/0f152076-9d57-482c-a9cc-897a0bc8eb28_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98d26189d8ed364b75deb4464e75c480e5705469d84c4bed4aca76cb95750688 +size 78994 diff --git a/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/0f152076-9d57-482c-a9cc-897a0bc8eb28_model.json b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/0f152076-9d57-482c-a9cc-897a0bc8eb28_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8a5fd223bf0fb4964353d68709c8a380c17dfeb1 --- /dev/null +++ b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/0f152076-9d57-482c-a9cc-897a0bc8eb28_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18ebd796ec0e41726c74b59753a89e9694a531a17940846a699ff779b57ba578 +size 100683 diff --git a/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/0f152076-9d57-482c-a9cc-897a0bc8eb28_origin.pdf b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/0f152076-9d57-482c-a9cc-897a0bc8eb28_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..910d74f24f9150fed5c6fead7cb6301dd72c725d --- /dev/null +++ b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/0f152076-9d57-482c-a9cc-897a0bc8eb28_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cce9fc00d7b7e78748000d76828542bbac81e3fe3c7301c5f37dafd666457e1 +size 3592914 diff --git a/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/full.md b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c88c34a0898e22ba88e9728114d57047d20a5072 --- /dev/null +++ b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/full.md @@ -0,0 +1,364 @@ +# Semantic and Sequential Alignment for Referring Video Object Segmentation + +Feiyu Pan1 Hao Fang1 Fangkai Li1 +Yanyu Xu1 Yawei Li2 Luca Benini2,3 Xiankai Lu1* +1Shandong University ETH Zürich 3University of Bologna +https://github.com/tavarich/SSA + +# Abstract + +Referring video object segmentation (RVOS) seeks to segment the objects within a video referred by linguistic expressions. Existing RVOS solutions follow a "fuse then select" paradigm: establishing semantic correlation between visual and linguistic feature, and performing frame-level query interaction to select the instance mask per frame with instance segmentation module. This paradigm overlooks the challenge of semantic gap between the linguistic descriptor and the video object as well as the underlying clutters in the video. This paper proposes a novel Semantic and Sequential Alignment (SSA) paradigm to handle these challenges. We first insert a lightweight adapter after the vision language model (VLM) to perform the semantic alignment. Then, prior to selecting mask per frame, we exploit the trajectory-to-instance enhancement for each frame via sequential alignment. This paradigm leverages the visual-language alignment inherent in VLM during adaptation and tries to capture global information by assembling trajectories. This helps understand videos and the corresponding descriptors by mitigating the discrepancy with intricate activity semantics, particularly when facing occlusion or similar interference. SSA demonstrates competitive performance while maintaining fewer learnable parameters. + +# 1. Introduction + +Referring video object segmentation (RVOS) is an emerging multimodal video understanding task, which aims at segmenting the target from a video based on linguistic description [7, 10, 18, 47]. It has diverse applications in fields such as human-robot interaction [40, 41] and video editing [3], garnering attention from the academic community. + +Traditional RVOS approaches [4, 54] typically focus on simple yet static scenes with prominent object, where one static expression refers to only one target [10, 18, 47]. Such isolated objects can be easily captured by static attributes in + +![](images/4dbbd01b1b24262c5b524a098f5e5745ffe25c0e0d220c7dddec35081e910d68.jpg) +(a) Previous frameworks + +![](images/1b5a1a4cf7066099858b432a8bcf2ea659f91ed13664612b70e1fb6ca5f00576.jpg) +(b) Our framework +Figure 1. (a) Previous frameworks extracted features from independent extractors, while not effectively leverage the consistency of the object query sequence. (b) Our SSA framework yields sufficient visual-linguistic interaction through semantic alignment by Vision-Language Model and global video modeling via sequential alignment. + +a single frame. However, in real-world video scenarios, existing pipelines struggle to seamlessly interpret the motion and multi-targets expressions with complex and dynamic scenarios due to substantial movement, abrupt scene change and non-salient targets [7, 26]. This leads to the challenge of aligning both video-text and trajectory-instance relations. + +RVOS in real-world environment requires explicitly identifying the particular target trajectory referred by the motion guidance. This underscores the necessity for robust visual-text semantic understanding to establish the corre + +![](images/69060ed967c615b2c7d0d86f3fb290b3ff5c38b1c36d4d394accab4f577f039d.jpg) +Figure 2. Comparison of correlation between performance and trainable parameters on MeViS dataset [7]. Circle size represents the numbers of trainable parameters for each model. + +spondence between video content and motion description, as well as the capacity to comprehend the entire video that ensure consistent object sequence segmentation. Existing methods [4, 7, 12, 13, 53, 54] utilize independent backbones to learn visual-text relationships from scratch, which may cause semantic gap during training (Fig. 1 (a)). After the challenging relationship modeling, they segment the target for each frame individually [4, 53, 54] or introduce video query [14] initialized solely by text feature, then interacting with heavily processed frame contents to generate sequence output [7, 13], both of which are suboptimal due to overlooking the comprehensive video representation from object sequences. To enhance the solution in real-world scenarios, a paradigm shift is needed, where both multi-modality semantic features and sequential information are uniformly aligned to improve multi-modal understanding and consistent temporal perception. + +To alleviate the misalignment issues in real-world scenes overlooked by pervious methods, we propose a novel framework called SSA (Fig. 1 (b)), aimed at enhancing the semantic alignment across visual-text features and sequential alignment across frame contents. Specifically, we leverage a frozen Vision-Language Model (VLM) to provide semantically correlated visual-textual features. We regard RVOS as a complex concept that involves various fine-grained entities, such as motions, scenes, and instances parsing. So we further propose a lightweight adapter to refine the visual-text semantic alignment between pixel-level video content and referring description for sufficient interactions, as we hypothesize that, the textual representations of the relevant entities should be aligned with specific regions of the image. + +Furthermore, to improve the global understanding of the entire video context, we incorporate a sequential alignment module to enhance the trajectory-to-instance correspondence. By ensembling these global trajectories, we can obtain a video-level context about the whole video thereby + +facilitating the consistent mask generation. This synergistic integration of sequential alignment and semantic alignment further improves the overall performance. + +The contributions of this paper are as follows: + +- We propose the Semantic and Sequential Alignment (SSA) framework that leverages video-to-text correspondences as well as trajectory-to-instance correspondence to guide the adaptation from VLM to RVOS. +- We propose a new lightweight adapter to refine the visual-language alignment of VLM. It encourages SSA to extract the semantics of the most important image regions given the referring descriptor and alleviate training difficulties. +- We further conduct sequential alignment on the trajectory to fully exploit global context, improving the overall interpretation of the video. +- SSA achieves new state-of-the-art results, especially on the challenging MeViS dataset [7] by $2.2\% \mathcal{J} \& \mathcal{F}$ with fewer trainable parameters, and exhibits significant improvement compared to LVLM methods (Fig. 2). + +# 2. Related Work + +Referring Video Object Segmentation. Referring Video Object Segmentation (RVOS) requires segmenting the target within a video based on the corresponding language expression [10, 24]. Seo et al. [47] establish the first large-scale RVOS benchmark, driving progress in the filed of RVOS. Many prior approaches in referring video segmentation have primarily adapted VIS [9, 56] (e.g. VisTR [51]) as streamlined pipeline, such as MTTR [51] and ReferFormer [54]. To better modeling temporal context [32], subsequent explorations include OnlineRefer [53] that leverages query propagation to explore online RVOS, TempCD [49] that interacts with a global refer token with sequences of intra-frame queries. + +Recently, a new large-scale benchmark called MeViS [7] is introduced. Compared to the previous RVOS datasets [10, 18, 47], MeViS emphasizes the motion attributes of videos with a large variety of motion expressions, as well as flexible expressions that refer to multi-targets. Thus, most of the current methods [4, 12, 47, 49, 53, 54, 60] can not handle this complex setting well, which reflects real-world scenarios. To address this new issue, a "fuse then select" paradigm has been raised. Typical works include LMPM [7] and DsHmp [13]. These methods first employ an advanced instance segmentation model to segment the possible objects according to fusing the text features. Then, a decoder module works for selecting the referred object and performing heavily cross-frame association. Despite handling multiple instances segmentation successfully, these methods either yield suboptimal performance [7] or have complex network structures [13]. + +Although our approach also builds upon query-based segmentation architecture, it differs in key aspects. We + +![](images/803e2c46e7838944fcb2b7ee78062419de79b51fbe956b724d5c43da5efc1ae5.jpg) +Figure 3. The overview architecture of the proposed method. It primarily contains Semantic Feature Alignment Encoders (Sec. 3.2) to capture semantic aligned visual-linguistic features, Instance Query Generation Module for extracting object-centric representation, and Sequential Object Alignment Module (Sec. 3.3) to align instance queries across the timeline for global video understanding. + +revisit RVOS from a multi-modal perception perspective and decompose it into video-to-text correspondences and trajectory-to-instance correspondence. Correspondingly, we propose semantic alignment and sequential alignment modules. First, this perspective enhances the model's ability to associate textual queries with relevant video content. Second, the alignment paradigm enables precise and consistent sequence-instance modeling, leading to better results. + +Vision-Language Models. Vision-Language Models (VLM) learn from image and text simultaneously, demonstrating strong cross-modal reasoning and generalization capabilities [2, 21, 23, 42, 46, 61]. VLMs typically use contrastive learning [16, 37, 42], generative methods [45, 48, 57] and alignment objectives [22, 27] to achieve cross-modal pre-training. With the powerful pre-trained knowledge of VLMs, they are now widely applied to various visual analysis tasks [19, 52, 59]. In the video domain, early works [9, 11, 50] utilize CLIP [42] as knowledge prior to achieve VIS in open vocabulary scenarios. Meanwhile, recent works, such as VISA [55] and VideoLISA [1], explore large VLMs (LVLMs) to enable textual reasoning and referring capabilities for video segmentation. Nevertheless, typical RVOS models use standalone vision-text encoders, leading to training difficulties caused by feature misalignment. Additionally, the current LVLM-based solutions [1, 55] and SAM2-based attempts [15, 43, 44] can be costly and yields suboptimal results on the challenging RVOS benchmark. In contrast, inspired by the success of VLM's fine-tuning trend, we utilize frozen VLM (CLIP [42]) as powerful feature extractors and introduce a context-aware adapter to alleviate the training challenges of RVOS. + +# 3. Methods + +# 3.1. Revisiting "Fuse then Select" Paradigm + +Given a video clip with $T$ frames $\mathcal{I} = \{I_t\}_{t=1}^T$ and a referring expression $\mathcal{E} = \{e_n\}_{n=1}^N$ consisting of $N$ words, RVOS needs to generate the binary masks $\mathcal{M} = \{m_t\}_{t=1}^T$ , $m_t \in \mathbb{R}^{H \times W}$ of the targets guided by the expression $\mathcal{E}$ . Before elaborating on our method (§3.2), we first give a brief introduction to "fuse then select" paradigm in RVOS. + +Existing approaches [4, 12, 13, 49, 53, 54] often utilize pretrained vision encoder and language encoder to extract visual features $\mathcal{F}_v$ and textual feature $\mathcal{F}_s$ independently: + +$$ +\mathcal {F} _ {v} = \varPhi_ {i m a g e} (\mathcal {I}) \in \mathbb {R} ^ {T \times C _ {i} \times H _ {i} \times W _ {i}}, +$$ + +$$ +\mathcal {F} _ {s} = \Phi_ {\text {t e x t}} (\mathcal {E}) \in \mathbb {R} ^ {1 \times C _ {l}}, \tag {1} +$$ + +where $\varPhi_{image}$ and $\varPhi_{text}$ denote Swin Transformer [29] and RoBERTa [28], separately. $i\in [1,4]$ is the stage index of the image encoder, $C_l$ is the dimension of textual features. + +After that, "fuse then select" methodology [7, 13] employs a decoder to fuse the visual and textual features together, and generate language-guided object embeddings which correspond to the potential objects. Then, these potential objects are associated into the trajectory candidates. Finally, the textual features are reused to select the predicted object trajectories from the candidates. + +# 3.2. Semantic Feature Alignment + +In contrast to the prior RVOS methods, we propose a new framework named SSA that learns Semantic and Sequential Alignment for RVOS. Specifically, it fully exploits both the semantic correspondence between text and visual features, and the temporal alignment within video sequences. Building on this, our approach enables us to construct an elegant + +![](images/1019e75f08d7447941d427959e8235819b9a83077130d6fe99e2478a4f5d6ebd.jpg) +Figure 4. Illustration of the Semantic Alignment Refine Module. + +model that is both strong in performance and effective in training. The overview of our approach SSA is depicted in Fig. 3. + +To obtain more accurate grounding region according to the text embeddings, we first devise a semantic feature alignment in a progressive manner. + +Coarse Feature Alignment. To address the rich semantic diversity in real-world RVOS scenarios, we firstly leverage frozen Vision-Language Model (i.e., CLIP [42]) to obtain semantic correlated visual-linguistic representations jointly: + +$$ +\left(\mathcal {F} _ {v}, \mathcal {F} _ {s}, \mathcal {F} _ {w}\right) = \Phi_ {\text {c l i p}} (\mathcal {I}, \mathcal {E}), \tag {2} +$$ + +where $\varPhi_{clip}$ denotes the CLIP model and $\mathcal{F}_w\in \mathbb{R}^{N\times C_l}$ denotes textual features of each word. + +Compared to Eq. (1), Eq. (2) facilitates the model's comprehensive understanding of both the video and expression while reducing training cost. + +Semantic Alignment Refine Module. Considering CLIP is pre-tained with large-scale image-text pairs, it may not handle fine-grained visual perception tasks well, such as image segmentation [17]. Therefore, we further design a lightweight adapter (as shown in Fig. 4) for each stage of CLIP to enhance vision-language semantic alignment, promoting a refined grounding of relevant entities according to the referring sentences. + +Due to referring expression provide fine-grained representations of the sentence, and we argue that they can ground the corresponding regions in video frames. Thus the proposed adapter consists of two-granularity cross-attention among visual features and textual features: $\{\mathrm{visual~feature}$ $\mathcal{F}_v$ and sentence-level textual features $\mathcal{F}_s\}$ , $\{\mathrm{visual~feature}$ $\mathcal{F}_v$ and word-level textual features $\mathcal{F}_w\}$ . In addition, to preserve the general knowledge of CLIP, we also utilize convolution blocks to build residual path: + +$$ +\begin{array}{l} f _ {m w} = \operatorname {s o f t m a x} \left(\frac {\mathcal {F} _ {v} \cdot \mathcal {F} _ {w}}{\sqrt {C}}\right) \mathcal {F} _ {w}, \\ f _ {m s} = \operatorname {s o f t m a x} \left(\frac {\mathcal {F} _ {v} \cdot \mathcal {F} _ {s}}{\sqrt {C}}\right) \mathcal {F} _ {s}, \tag {3} \\ f _ {m r} = \varPhi_ {c o n v} (\mathcal {F} _ {v}) + \mathcal {F} _ {v}, \\ \end{array} +$$ + +where $f_{mw}$ , $f_{ms}$ and $f_{mr}$ mean visual outputs from three streams, respectively. $C$ denotes dimension of the joint embedding space, $\varPhi_{conv}$ indicates convolution operation. + +Through Eq. (3), the adapter generates refined visual-textual aligned features from multiple streams. To balance the semantics introduced by different feature streams, we employ an MLP serving as gating mechanism to learn dynamic weights for each stream. The final output at each stage is as follows: + +$$ +\mathcal {W} _ {\alpha} = \operatorname {s o f t m a x} (\operatorname {M L P} ([ g (f _ {m w}), g (f _ {m r}), g (f _ {m s}) ])), \tag {4} +$$ + +$$ +\mathcal {F} _ {v} = \mathcal {W} _ {\alpha 1} \cdot f _ {m w} + \mathcal {W} _ {\alpha 2} \cdot f _ {m r} + \mathcal {W} _ {\alpha 3} \cdot f _ {m s}, +$$ + +where $g(\cdot)$ denotes global average pooling, $\mathcal{W}_{\alpha} \in \mathbb{R}^{3 \times 1}$ is the weight of different streams, $[\cdot]$ is the concatenation operation. Through the proposed adapter, multi-granularity textual features progressively align within the visual feature, while retaining VLM's powerful knowledge. + +# 3.3. Sequential Object Alignment + +Instance Query Generation. Once fused features $\mathcal{F}_v$ is obtained from VLM, the existing RVOS solutions often employ object query to serve as implicit representation of the target in the video. Typically, they are either initialized randomly for each frame [4] or sorely generated via sentence-level textual feature (Eq. (1)) [7, 12, 53, 54]. + +Considering the variations in the position and shape of objects between different video frames over time, we propose an object query generating mechanism to ensure that queries have variability across frames via cross-modal feature interaction (as depicted in Fig. 3). + +In concrete, we firstly leverage $1 \times 1$ convolution block and linear layer to project the visual and textual features into the joint space, respectively. Subsequently, bidirectional cross-attention operation is adopted at bi-directional encoder to conduct interaction between visual and textual features frame by frame: + +$$ +\hat {\mathcal {F}} _ {v} ^ {t} = \operatorname {s o f t m a x} \left(\frac {\mathcal {F} _ {v} ^ {t} W _ {t 2 v} ^ {Q} \cdot \left(\mathcal {F} _ {s} W _ {t 2 v} ^ {K}\right) ^ {\top}}{\sqrt {C}}\right) \mathcal {F} _ {s} W _ {t 2 v} ^ {V}, \tag {5} +$$ + +$$ +\hat {\mathcal {F}} _ {s} ^ {t} = \operatorname {s o f t m a x} (\frac {\mathcal {F} _ {s} W _ {v 2 t} ^ {Q} \cdot (\mathcal {F} _ {v} ^ {t} W _ {v 2 t} ^ {K}) ^ {\intercal}}{\sqrt {C}}) \mathcal {F} _ {v} ^ {t} W _ {v 2 t} ^ {V}, +$$ + +where $W_{\{t2v,v2t\}}^{Q}, W_{\{t2v,v2t\}}^{K}, W_{\{t2v,v2t\}}^{V}$ are learnable projection weights, $t$ denotes frame index. $\hat{\mathcal{F}}_s^t \in \mathbb{R}^{1 \times C}$ is updated textual feature of $t$ -th frame. Additionally, $\mathcal{F}_v$ is utilized to generate mask feature $\mathcal{F}_{mask}$ , which is used to produce the final mask sequence. + +Instance queries $\{Q^t\}_{t = 1}^T,Q^t\in \mathbb{R}^{N\times C}$ generated by repeating $\hat{\mathcal{F}}_s$ for $N$ times, form as initialized sequences. Then, $\{Q^t\}_{t = 1}^T$ are fed into the frame query decoder to learn the potential target representation $\{\tilde{Q}^t\}_{t = 1}^T$ for each frame. + +Instance Trajectory Generation. Object queries $\{\tilde{Q}^t\}_{t = 1}^T$ are extracted independently for each frame, which may lead + +![](images/a3c66e7bd71928c62a061f55ef065b6f1c8ea370c3c11de8857642a24356d17b.jpg) +Figure 5. Illustration of Sequential Object Alignment Module. + +to misalignment for the same instance along the temporal dimension. To address this issue, we utilize Hungarian matching algorithm [20] to match $\tilde{Q}$ of adjacent frames along timeframes to model trajectories for each potential instance (as shown in Fig. 5). + +Upon establishing trajectories for $\{\tilde{Q}_i|i\in [1,N],\tilde{Q}_i\in$ $\mathbb{R}^{T\times C}\}$ , we leverage convolution block for short-term information and self-attention layer for global view for trajectory's temporal enhancement: + +$$ +\text {S h o r t T e r m} (\tilde {Q} _ {i}) = \Phi_ {C o n v 1 D} (\tilde {Q} _ {i}) \in \mathbb {R} ^ {T \times C}, +$$ + +$$ +\operatorname {L o n g T e r m} \left(\tilde {Q} _ {i}\right) = \Phi_ {\text {T e m p A t t n}} \left(\tilde {Q} _ {i}\right) \in \mathbb {R} ^ {T \times C}, +$$ + +where $\tilde{Q}_i$ is the $i$ -th instance of $N$ trajectories, $\varPhi_{Conv1D}$ and $\varPhi_{TempAttn}$ denote 1D convolution and temporal attention. + +Trajectory to Instance Correspondence. Once obtained the temporal enhanced instance trajectories $\tilde{Q}_{TE}$ , we can ensemble it as a global description of the whole video. The global representation of instance (video query) is then obtained by temporal weighted fusion: + +$$ +\mathcal {A} = \operatorname {s o f t m a x} \left(\Phi_ {F C} \left(\tilde {Q} _ {T E}\right)\right) \in \mathbb {R} ^ {T \times N \times 1}, +$$ + +$$ +Q _ {v} = \sum_ {t = 1} ^ {T} \mathcal {A} ^ {t} \cdot \tilde {Q} _ {T E} ^ {t} \in \mathbb {R} ^ {N \times C}, \tag {7} +$$ + +where $\varPhi_{FC}$ denotes linear projection layer with learnable weight in $\mathbb{R}^{1\times C}$ , $\mathcal{A}$ is the temporal weight in each frame through the softmax function. + +Compared to previous methods that utilize frame queries directly for sequence output [4, 12, 53, 54] or rely on solely text features as global video query [7, 13, 33], the proposed approach integrates global modeling from the entire video that sufficiently mines the global information. + +# 3.4. Video Query Decoding + +Similar to previous works [7, 13, 14], given the video queries $Q_{v}$ , frame queries $\tilde{Q}_{TE}$ and language features $\hat{\mathcal{F}}_s$ , video decoder works for generating binary classification scores $S_{cls}$ and mask embeddings $Q_{emb}$ , separately. + +Then $Q_{emb}$ is filtered through $S_{cls}$ by argmax operation for single object scenarios [38, 47] or threshold operation for multiple instance scenarios [7]. Then, it multiplies with mask feature $\mathcal{F}_{mask}$ to obtain the final mask sequence of the identified target(s). + +# 3.5. Model Implementation and Training + +Model Implementation. Specifically, we use ConvNeXt [30] based CLIP [42] as feature extractors (§ 3.2), as ConvNeXt naturally extracts multi-scale feature maps from image, making it suitable for segmentation tasks. We utilize modified mask2former [5] to implement instance query generation and transformer blocks for video query decoder in §3.3. + +Training. Following [7, 13, 14], we employ the $\mathcal{L}_f$ to calculate the loss between per-frame predictions and framewise annotations, along with $\mathcal{L}_v$ as video-level supervision. We implement the $\mathcal{L}_f$ and $\mathcal{L}_v$ by binary cross-entropy loss with dice loss. The overall training objective is given by: + +$$ +\mathcal {L} _ {\text {t r a i n}} = \mathcal {L} _ {v} + \mathcal {L} _ {f} + \lambda_ {\text {s i m}} \mathcal {L} _ {\text {s i m}}, \tag {8} +$$ + +where $\lambda_{sim}$ is the weight for the similarity loss $\mathcal{L}_{sim}$ [14]. + +# 4. Experiments + +# 4.1. Datasets and Evaluation Metrics + +Datasets. The proposed methods are evaluated on three popular RVOS datasets: Ref-Youtu-VOS [47], RefDAVIS17 [18] and MeViS [7]. Ref-Youtu-VOS [47] establishes the first large-scale dataset consisting of 3,978 videos with approximate 15k language descriptions and 202 video samples are used for online testing. RefDAVIS17 [18] contains 90 videos modified DAVIS17 [39] by providing additional textual annotations, which allocates 30 videos for validation. MeViS [7] is a new large-scale benchmark that highlights motion information analysis in video. It includes 2,006 video sequences with 443k fine-grained segmentation annotations, and 28,570 sentences describing 8,171 instances in complex video scenarios. The Val dataset of MeViS contains 140 video for online evaluation, while Valu includes 50 videos for local validation. + +We further validate on two additional datasets [10], and the details are provided in the supplementary materials. + +Evaluation Metrics. Following previous works [7, 54], jaccard index $\mathcal{I}$ (region similarity), mean boundary similarity $\mathcal{F}$ (contour accuracy), and their average $\mathcal{J} \& \mathcal{F}$ are employed as evaluation metrics. + +# 4.2. Experimental Settings + +MeViS. For MeViS dataset, following previous works [7, 13], without pre-training on the image-level referring + +Table 1. Comparison with state-of-the-art methods on MeViS [7] Val and Valu dataset. † denotes our model is trained on MeViS [7] and Ref-Youtu-VOS [7] jointly. See §4.3 for details. + +
MethodsReferenceValValu
J&F↑J↑F↑J&F↑J↑F↑
Compared with expert methods
URVOS [47]ECCV'202027.825.729.9---
LBDT [8]CVPR'202229.327.830.8---
MTTR [4]CVPR'202230.028.831.2---
ReferFormer [54]CVPR'202231.029.832.2---
VLT+TC [6]TPAMI'202235.533.637.3---
HTR [36]TCSVT'202442.739.945.5---
LMPM [7]ICCV'202337.234.240.240.236.543.9
DsHmp [13]CVPR'202446.443.049.855.351.060.4
Ours-48.644.053.256.951.762.2
Compared with LVLM methods
VideoLISA-3.8B [1]NeurIPS'202444.441.347.6---
VISA-7B [55]ECCV'202443.540.746.351.248.054.4
Ours †-48.944.353.457.852.363.1
+ +datasets RefCOCO/+/g [34, 58], we train the model on MeViS directly. The optimization is performed with AdamW [31] for 50k iterations, with a learning rate initialized to $5e^{-4}$ . + +Ref-Youtu-VOS and Ref-DAVIS17. For these two datasets, following [13], we first conduct image-level pretraining on the RefCOCO+/g datasets [34, 58]. After that, we fine-tune the model with videos from the Ref-Youtu-VOS. During the pre-training phase, the model is trained for 100k iterations. In the video-level training phase, we trained the model for 50k iterations. For Ref-DAVIS17, we directly apply the learned model from Ref-YouTube-VOS to Ref-DAVIS17 [18] without online fine-tuning. + +Joint Video Datasets Co-training. Considering current VLM methods [1, 55] tend to be trained on image and video datasets jointly, we adopt a similar strategy to explore the performance potential by training our model on MeViS and Ref-Youtu-VOS jointly during the video training phase. + +Table 2. Comparison with SAM2 models on MeViS [7] Val dataset. Results are adopted from [15]. See §4.3 for details. + +
MethodsJ&F↑J↑F↑
Grounded-SAM 2 [44]‡38.935.742.1
AL-Ref-SAM 2 [15]‡42.839.546.2
Ours48.644.053.2
+ +# 4.3. Main Results + +MeViS. In Tab. 1, we validate the proposed approach on the newly released motion expressions guided video segmentation dataset MeViS [7]. With comparable training costs, SSA outperforms other leading expert methods, surpassing the main counterpart DsHmp [13] by $2.2\%$ $\mathcal{J}\& \mathcal{F}$ on val set + +and $1.6\%$ on valu set $(2.5\%)$ on val and valu under our constraining strategy). + +Additionally, in comparison with methods leveraging large vision-language models, SSA exceeds the contemporary methods and establishes a new state-of-the-art performance in terms of mean $\mathcal{I}\&\mathcal{F}(48.9\%)$ val set and $\mathcal{I}\&\mathcal{F}(57.8\%)$ on valu set. Notably, our method shows a marked performance gain over the best method VideoLiSA [1] by $4.5\%$ $\mathcal{I}\&\mathcal{F}$ . These results demonstrate the importance of semantic alignment and sequential alignment. Furthermore, as shown in Fig. 2, our approach improves performance while reducing approximate 46M trainable parameters compared to current state-of-the-art DsHmp [13], alleviating the training difficulties for RVOS task. + +Moreover, the recently introduced Segment Anything Model 2 (SAM2) [43] has achieved breakthrough progress in VOS tasks, and some works have migrated it to the RVOS domain [15, 44]. For experimental completeness, we also conduct experimental comparisons with SAM2 based pipelines. Nevertheless, as shown in Tab. 2, our method behaves better than the existing SAM2 attempts in the complex MeViS scenarios with a distinct performance gap: (AL-Ref-SAM 2 [15]: 48.6 vs. 42.8, 44.0 vs. 39.5, 53.2 vs. 46.2) and (Grounded-SAM 2 [44]: 48.6 vs. 38.9, 44.0 vs. 35.7, 53.2 vs. 42.1). + +Ref-YouTube-VOS. The results on the Ref-YouTube-VOS dataset are presented in Tab. 3. Our method exceeds existing approaches on Ref-Youtu-VOS across all metrics by achieving $64.3\%$ $\mathcal{J}\& \mathcal{F}$ , which is $0.7\%$ higher than DsHmp. Ref-DAVIS17. On Ref-DAVIS17 (Tab. 3), our approach achieves $67.3\%$ $\mathcal{J}\& \mathcal{F}$ and consistently outperforming the top method (i.e., DsHmp) by more than $3.3\%$ in terms of $\mathcal{J}\& \mathcal{F}$ . These experiment results exhibit the generalization + +Table 3. Comparison with state-of-the-art models on Ref-Youtu-VOS [47] and Ref-DAVIS17 [18] datasets. See §4.3 for details. + +
MethodsReferenceBackboneRef-Youtube-VOSRef-DAVIS17
J&F↑J↑F↑J&F↑J↑F↑
LBDT [8]CVPR'2022ResNet-5049.448.250.654.3--
MTTR [4]CVPR'2022Video-Swin-T55.354.056.6---
ReferFormer [54]CVPR'2022Video-Swin-T59.458.060.959.656.562.7
OnlineRefer [53]ICCV'2023Video-Swin-B62.961.064.762.459.165.6
HTML [12]ICCV'2023Video-Swin-T61.259.563.0---
R2VOS [25]ICCV'2023Video-Swin-T61.359.663.1---
SgMg [35]ICCV'2023Video-Swin-T62.060.463.561.959.064.8
TempCD [49]ICCV'2023Video-Swin-T62.360.564.062.259.365.0
SOC [33]NeurIPS'2023Video-Swin-T62.461.163.763.560.266.7
LoSh [60]CVPR'2024Video-Swin-T63.762.065.462.960.165.7
DsHmp [13]CVPR'2024Video-Swin-T63.661.865.464.060.867.2
Ours-CLIP64.362.266.467.364.070.7
+ +capability of our method. + +# 4.4. Ablation Studies + +Table 4. Ablation study on main components of SSA. Semantic indicates Semantic Feature Alignment, Sequential denotes Sequential Object Alignment. See §4.4 for details. + +
ComponentsJ&F↑J↑F↑
SemanticSequential
XX42.437.447.4
✓X46.341.551.1
X✓45.140.849.4
✓✓48.6↑6.244.0↑6.653.2↑5.8
+ +In this section, we conduct ablation studies on the challenging dataset MeViS [7] to investigate our proposed approach. By removing the proposed core components, we build vanilla baseline that shares a similar structure with LMPM [7]. + +Table 5. Comparison with DsHmp [13] variant models on MeViS dataset. See §4.4 for details. + +
MethodsJ&F↑J↑F↑
DsHmp [13]46.443.049.8
+ CLIP Encoders (Eq. (2))47.242.951.5
+ Correspond (Eq. (7))47.643.451.8
Ours48.644.053.2
+ +Component Analysis. We first verify the effectiveness of main components of our framework. As shown in Tab. 4, semantic feature alignment mechanism provides an significant performance gain by $3.9\%$ of $\mathcal{J}\& \mathcal{F}$ compared to baseline who utilizes Swin [29] and RoBERTa [28] as encoders ( $2^{nd}$ row). This indicates the necessity of semantic alignment in RVOS modeling. + +Adding sequential object alignment to baseline improves the performance by $2.7\%$ of $\mathcal{I} \& \mathcal{F}(3^{rd}$ row). Finally, after + +integrating all the main components together, we observe a substantial improvement up to $6.2\%$ of $\mathcal{I}\& \mathcal{F}$ compared to the baseline (last row). + +Additionally, we extend the DsHmp [13] to validate the fairness. As shown in Tab. 5, when employing the same CLIP encoders, DsHmp slightly improves by $0.8\%$ , but still remains $1.4\%$ lower than SSA. Moreover, adding instance correspondence (Eq. 7) further improves the valina Dshmp by $1.2\%$ . These analysis demonstrate the effectiveness as well as suitability of the proposed components. + +Table 6. Comparisons of different Semantic Feature Alignment strategies. Frozen-CLIP denotes freezing the entire CLIP model, Fine-tuning indicates fine-tune the last stage of CLIP, Adapter refers to the proposed adapter. See §4.4 for details. + +
Frozen-CLIPFine-tuningAdapterTrainable-params↓J&F↑
XXX72.3M45.1
✓XX39.4M47.5
✓✓X100.6M47.9
✓X✓46.5M48.6
+ +Effectiveness of Adapter. Further ablation studies on Semantic Feature Alignment is shown in Tab. 6. Compared to the baseline, it can be seen $2.4\%$ $\mathcal{J}\& \mathcal{F}$ performance improvement is achieved by introducing frozen CLIP model, while reduce approximately $46\%$ trainable parameters. Compared with fine-tuning the last stage of ConvNext, the training cost is increased but bring a slight performance gain of $0.4\%$ . Finally, when adopting the proposed adapter, SSA further improves the $\mathcal{J}\& \mathcal{F}$ by $1.1\%$ with affordable extra parameters increase than fine-tuning. + +In-depth Analysis of Sequential Object Alignment. Finally, we study the components of Sequential Object Alignment. As shown in Tab. 7, by removing bi-directional attention in Eq. (5) ( $2^{nd}$ row) for instance query generation, + +![](images/85a12b8a7dc5dc3199ce52f19f94f29303ee78528cfc4aab9333a9d76636e7c6.jpg) +(a) "The feline is constantly moving in different directions." + +![](images/ddb53e3f8f63726857b0227b67b137f80da6464b6098b4866cee81531392c41f.jpg) + +![](images/739c0e71243ac3444a601c6c027794b7b6ed0736389a0988cc0c0f3a81a9843e.jpg) +(b) "Three people riding elephants, racing against each other." +Figure 6. Qualitative Results on Mevis dataset [7]. The first row illustrates the segmentation results of DsHmp [13] while the second row indicates the segmentation results of ours. + +Table 7. Ablation study on Sequential Object Alignment. Sequential indicates the overall module, the rest rows indicate the variants by removing specific components. See §4.4 for details. + +
ArchitectureJ&F↑J↑F↑
Sequential48.644.053.2
w/o Bi-directional47.8↓0.843.2↓0.852.3↓0.9
w/o Ensemble47.2↓1.442.6↓1.451.9↓1.3
w/o Trajectory/Ensemble46.1↓2.541.2↓2.851.0↓2.2
+ +the segmentation performance drops $0.8\%$ of $\mathcal{J} \& \mathcal{F}$ . Without trajectory ensemble ( $3^{rd}$ row) in Eq. (7) for instance correspondence, we observe a $1.4\%$ of $\mathcal{J} \& \mathcal{F}$ performance decrease. Furthermore, the omission of instance trajectory generation via Hungarian algorithm [20] and temporal enhancement in Eq. (6), as well as trajectory ensemble in Eq. (7) ( $4^{th}$ row), result in a performance drop of $2.5\%$ of $\mathcal{J} \& \mathcal{F}$ . The empirical evidence suggests that importance of each component. + +# 4.5. Qualitative Results + +As shown in Fig. 6, we present the qualitative comparison between our approach and DsHmp [13] on the more challenging dataset, MeViS [7]. Specifically, the target (cat) in Fig. 6 (a) suffers from the distraction of the similar object. Except for the referring target, DsHmp erroneously segments the stationary cat as well. In contrast, SSA can segment the target precisely. For multi-instance referring + +scenario (Fig. 6 (b)), SSA accurately segments three riders on the elephants during the race. These qualitative results show that the proposed semantic and sequential alignment help the model better understand text-described objects and ensure consistent temporal segmentation. + +# 5. Conclusion + +In this paper, we propose a novel Semantic and Sequential Alignment framework for RVOS in real-world scenarios. Specifically, we achieve multi-modality semantic alignment by utilizing pretrianed VLM with proposed adapter for better video-to-text correspondence. Besides, the trajectory-to-instance correspondence established in the sequential alignment module enhances the model's ability to achieve global instance awareness. Extensive experiments on the challenging RVOS benchmarks demonstrate the effectiveness and generalization capacity of our method. + +# Acknowledgement + +This work was supported in part by the Shandong Excellent Young Scientists Fund (ZR2024YQ006), Shandong Province Higher Education Institutions Youth Entrepreneurship and Technology Support Program (2023KJ027) and Taishan Scholars Program (Grant No. tsqn202312067). + +# References + +[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. NeruIPS, 2024. 3, 6 +[2] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. NeurIPS, 2022. 3 +[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 1 +[4] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7 +[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 5 +[6] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In ICCV, 2021. 6 +[7] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[8] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In CVPR, 2022. 6, 7 +[9] Hao Fang, Peng Wu, Yawei Li, Xinxin Zhang, and Xiankai Lu. Unified embedding alignment for open-vocabulary video instance segmentation. In ECCV, 2024. 2, 3 +[10] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In CVPR, 2018. 1, 2, 5 +[11] Pinxue Guo, Tony Huang, Peiyang He, Xuefeng Liu, Tianjun Xiao, Zhaoyu Chen, and Wenqiang Zhang. Openvis: Open-vocabulary video instance segmentation. arXiv preprint arXiv:2305.16835, 2023. 3 +[12] Mingfei Han, Yali Wang, Zhihui Li, Lina Yao, Xiaojun Chang, and Yu Qiao. Http: Hybrid temporal-scale multimodal learning framework for referring video object segmentation. In ICCV, 2023. 2, 3, 4, 5, 7 +[13] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In CVPR, 2024. 2, 3, 5, 6, 7, 8 +[14] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. NeurIPS, 2022. 2, 5 +[15] Shaofei Huang, Rui Ling, Hongyu Li, Tianrui Hui, Zongheng Tang, Xiaoming Wei, Jizhong Han, and Si Liu. Unleashing the temporal-spatial reasoning capacity of gpt for training-free audio and language referenced video object segmentation. In AAAI, 2025. 3, 6 + +[16] 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 ICML, 2021. 3 +[17] Siyu Jiao, Yunchao Wei, Yaowei Wang, Yao Zhao, and Humphrey Shi. Learning mask-aware clip representations for zero-shot segmentation. NeurIPS, 2023. 4 +[18] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In ACCV, 2018. 1, 2, 5, 6, 7 +[19] Seoyeon Kim, Minguk Kang, Dongwon Kim, Jaesik Park, and Suha Kwak. Extending clip's image-text alignment to referring image segmentation. In NAACL-HLT, 2024. 3 +[20] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 5, 8 +[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022, 3 +[22] 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 ICML, 2023. 3 +[23] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 3 +[24] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In CVPR, 2018. 2 +[25] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust referring video object segmentation with cyclic structural consensus. In ICCV, 2023. 7 +[26] Chang Liu, Xudong Jiang, and Henghui Ding. Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence, 2024. 1 +[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 3 +[28] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3,7 +[29] 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 ICCV, 2021. 3, 7 +[30] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 5 +[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[32] Xiankai Lu, Wenguan Wang, Chao Ma, Jianbing Shen, Ling Shao, and Fatih Porikli. See more, know more: Unsupervised video object segmentation with co-attention siamese networks. In CVPR, 2019. 2 + +[33] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. NeurIPS, 2023. 5, 7 +[34] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 6 +[35] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In ICCV, 2023. 7 +[36] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, Mubarak Shah, and Ajmal Mian. Temporally consistent referring video object segmentation with hybrid memory. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 6 +[37] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 3 +[38] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In ICCV, 2019. 5 +[39] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5 +[40] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In CVPR, 2020. 1 +[41] Zheyun Qin, Cheng Han, Qifan Wang, Xiushan Nie, Yilong Yin, and Lu Xiankai. Unified 3d segmenter as prototypical classifiers. NeurIPS, 2023. 1 +[42] 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 ICLR, 2021. 3, 4, 5 +[43] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3, 6 +[44] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 3, 6 +[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3 +[46] 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. NeurIPS, 2022. 3 + +[47] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, 2020. 1, 2, 5, 6, 7 +[48] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 3 +[49] Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In ICCV, 2023. 2, 3, 7 +[50] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In ICCV, 2023. 3 +[51] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2 +[52] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In CVPR, 2022. 3 +[53] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In ICCV, 2023. 2, 3, 4, 5, 7 +[54] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7 +[55] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. In ECCV, 2024. 3, 6 +[56] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2 +[57] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 3 +[58] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, 2016. 6 +[59] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. NeurIPS, 2023. 3 +[60] Linfeng Yuan, Miaojing Shi, Zijie Yue, and Qijun Chen. Losh: Long-short text joint prediction network for referring video object segmentation. In CVPR, 2024. 2, 7 +[61] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 3 \ No newline at end of file diff --git a/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/images.zip b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a26e3ad8a102fc49fe6d24bcc19e9197c3c92eb8 --- /dev/null +++ b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb1981a1dad25bfbd8e553061eb4339ae46df8668a977f737d56baa1355fe302 +size 789008 diff --git a/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/layout.json b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..414254330581239a771c9093f983ec30b99d4da7 --- /dev/null +++ b/CVPR/2025/Semantic and Sequential Alignment for Referring Video Object Segmentation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1fb06fd22ef0136e030c2954debd0d6cfca64d339252e439d0da1cbaa5158a6 +size 429157 diff --git a/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/433d7f52-bf59-406e-8aa1-aae038005c90_content_list.json b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/433d7f52-bf59-406e-8aa1-aae038005c90_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3d44bdfb0eb08724dc8dd3ac1409c4c2b4ad83b9 --- /dev/null +++ b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/433d7f52-bf59-406e-8aa1-aae038005c90_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16cdf227e0f0cac1bd9436cd589fa5a79df9148a9e80d34541c6367a0f43acea +size 87993 diff --git a/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/433d7f52-bf59-406e-8aa1-aae038005c90_model.json b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/433d7f52-bf59-406e-8aa1-aae038005c90_model.json new file mode 100644 index 0000000000000000000000000000000000000000..56bc33a333e24a725bddef4da8e213ad9e411c85 --- /dev/null +++ b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/433d7f52-bf59-406e-8aa1-aae038005c90_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0874ec59507bf05df7d6f32533b9053bb301d863e2997a0c1c755d58fa388726 +size 106419 diff --git a/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/433d7f52-bf59-406e-8aa1-aae038005c90_origin.pdf b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/433d7f52-bf59-406e-8aa1-aae038005c90_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd71b031148c7c484ff38af7276a80b213fdb8ed --- /dev/null +++ b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/433d7f52-bf59-406e-8aa1-aae038005c90_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e74b1dc7b8bf572d498e5905ff8fc0f491476c9c0d920ecb12f07535bca76921 +size 3332558 diff --git a/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/full.md b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..96a73a46cf8b833f6ff8a9787e4cc5f25f6f0174 --- /dev/null +++ b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/full.md @@ -0,0 +1,306 @@ +# Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition + +Anqi Zhu $^{1}$ , Jingmin Zhu $^{2}$ , James Bailey $^{1}$ , Mingming Gong $^{1}$ , Qiuhong Ke $^{2,*}$ $^{1}$ The University of Melbourne; $^{2}$ Monash University + +Parkville VIC 3052, Australia; Wellington Road, Clayton VIC 3800, Australia + +azzhl@student.unimelb.edu.au, jzhu0105@student.monash.edu, {baileyj, mingming.gong}@unimelb.edu.au, qiuhong.ke@monash.edu + +# Abstract + +Skeleton-based human action recognition is promising due to its privacy preservation, robustness to visual challenges, and computational efficiency. Especially, the practical necessity to recognize unseen actions has led to increased interest in zero-shot skeleton-based action recognition (ZSSAR). Existing ZSSAR approaches often rely on manually crafted action descriptions or visual assumptions to enhance knowledge transfer, which is limited in flexibility and prone to inaccuracies and noise. To overcome this, we introduce Semantic-guided Cross-Modal Prompt Learning (SCoPLe), a novel framework that replaces manual guidance with data-driven prompt learning for refinement and alignment of skeletal and textual features. Specifically, we introduce a dual-stream language prompting module that preserves the original semantic context from the pre-trained text encoder while still effectively tuning its output for ZSSAR task adaptation. We also introduce a joint-shaped prompting module that learns tuning for skeleton features and incorporate an adaptive visual representation sampler that leverages text semantics to strengthen the cross-modal prompting interactions during skeleton-to-text embedding projection. Experimental results on the NTU-RGB+D and PKU-MMD datasets demonstrate the state-of-the-art performance of our method in both ZSSAR and generalized ZSSAR scenarios. + +# 1. Introduction + +Human Action Recognition (HAR) is essential in various fields, including surveillance [7, 17, 32], healthcare [1, 46], human-computer interaction [2, 40], and smart environments [4, 24, 41, 45]. Within HAR, skeleton-based action recognition has garnered significant attention due to its unique capability to capture class-determinant body joint movements while ensuring privacy [13, 27, 38]. Moreover, skeleton data + +
ModelNTU-RGBD 60NTU-RGBD 120
55/5 split48/12 split110/10 split96/24 split
CLIP62.5833.1663.9248.15
CLIP-CoOp [49]75.88 (+13.3%)36.32 (+3.16%)64.8 (+0.88%)46.02 (-2.13%)
CLIP-MaPLe [18]65.59 (+3.01%)34.93 (+1.77%)66.45 (+2.47%)47.42 (-0.73%)
CLIP-DSLP (ours)70.21 (+7.63%)40.69 (+7.53%)68.88 (+4.96%)48.17 (+0.02%)
+ +Table 1. ZSSAR accuracy for adapting CLIP with different common prompting techniques or our DSLP (see Sec. 3.2) on $NTU-RGB + D$ 60 and $NTU-RGB + D$ 120. For previous approaches, we only preserve their prompting frameworks for the text branch because the shape of the skeleton features is not compatible with their original designs for visual prompting which are applied for image encoding. + +is robust to common visual challenges such as background clutter and lighting variations [6], and is computationally efficient, making it suitable for resource-constrained environments [28, 31, 47]. These advantages have led to increased research efforts on the topic in recent years [5, 21, 26, 44]. + +Most of the present skeleton-based action recognition approaches [8, 10, 35-37, 43] focus on fully supervised learning, where both training and testing involve the same fixed set of action classes. However, in practical scenarios, the vast range of human actions, combined with the high cost and time required to collect and annotate data for each possible category, demands models that can be effectively adaptable to new recognition tasks involving unseen classes. Zero-shot skeleton-based action recognition (ZS-SAR) is designed to tackle this challenge – enabling models to recognize novel actions without requiring explicit training examples for them. + +In addressing ZSSAR, the major challenge is achieving effective knowledge transfer, which requires constructing a shared semantic space to bridge a generalized visual representation distribution for seen and unseen actions. Prior research in ZSSAR [9, 14, 42, 51] has solved this by leveraging label embeddings from pre-trained language models (e.g., BERT) as generalizable intermediate references for distribution learning and developing various strategies to expand the scope of transferable knowledge. While effective, + +they often rely on careful assumptions and manually crafted techniques to guide the training. For example, some latest solutions involve semantic decomposition that makes simplified assumptions about action labels. [14] assumes that every action label follows a "verb+noun" syntactic structure. [51] and [9] require the manual generation of local movement descriptions for each label by GPT [3] for body-part-based learning. The success of these approaches is highly sensitive to the quality and consistency of the assumptions and descriptions, which might introduce noise or irrelevant information that does not directly aid in skeleton-text alignment. + +In this work, we shift away from manually crafted guidance and delve into ZSSAR solutions through purely data-driven prompt tuning for cross-modal knowledge refinement and alignment. In particular, our aim is to investigate both text and skeleton side prompt learning to automatically extract essential contextual signals from action labels and transmit them across the modalities, thereby reducing reliance on artificial assumptions and operations. By automatically capturing task-refined cross-modal semantics from the labels, the model can focus on tuned consistent contexts shared across similar actions, allowing it to establish generalizable visual patterns that define each potential action. This ultimately enhances knowledge transfer, enabling the model to recognize new actions without support from specific analyses or detailed descriptions. + +Antecedent works on prompt learning laid the methodology foundation for enhancing pre-trained multi-modal models (PMMs) such as CLIP [30] by concatenating learnable prompt embeddings with class text tokens or image patch embeddings before feeding them into the encoders [49], or by inserting learnable prompts at each encoder layer to fine-tune the attention distribution [18]. Ensuring learning interaction and synchronization between prompts in different data-mode encoders is another important technique to develop tuning efficiency [18, 48]. Showing success in image-based supervised task adaptation, these methods also remain valid in some zero-shot learning (ZSL) scenarios. Yet, according to the 11 dataset experiments in [18], the improvement is unstable and sometimes still fail to surpass the baseline of only using CLIP. In Tab. 1, we conducted a validation experiment by applying common prompting methods (e.g. CoOp [49] and MaPLe [18]) to CLIP for ZSSAR, and the results yield similar volatile performance gains. We argue that this instability may arise from the impact of learnable prompts on the generalizability of the original CLIP model to unseen actions. Specifically, a CLIP text encoder, built on a Transformer architecture with a self-attention mechanism, is pre-trained on extensive visual-text pairs to support broad generalization. However, prompts trained exclusively on seen classes can alter the attention outputs of text labels in ways that can significantly impact the model's inherent ability to generalize to the semantic context of new classes. + +In such cases, the model is overly reliant on these modified attention distributions and feature adjustments introduced by the learnable prompts, which can sometimes enhance performance, but may also compromise the fundamental strengths of CLIP. Additionally, these methods' original prompting strategy for the image branch does not directly accommodate skeleton features due to differences in data shape. The absence of prompting on the visual side will further exacerbate the tuning instability. + +To address this, we argue that both the attention output of the original and prompted encodings should be considered for optimal generalization. We propose a dual-stream language prompting (DSLP) module. It consists of a pretrained CLIP text encoder that retains the original encoding stream, along with a second stream that learns layer-level prompts, which tunes the prompted feedforward features while adaptively integrating the original attention distribution from the first stream. This enables the model to balance between the preservation of the original context structure and the tuning of the features. Furthermore, to facilitate the corresponding tuning for skeleton feature refinement, we propose a semantic-guided skeleton prompting module that uses joint-shaped prompts to optimize task-relevant contexts. The design is grounded in the unique data structure of skeleton joint features. The module then incorporates an adaptive visual representation sampler that synchronizes the prompted text semantics from the language branch as queries to gather class-relevant prompted visual clues for projecting skeleton-to-text embeddings. To this end, we establish all the core components of our proposed Semantic-guided Cross-Modal Prompt Learning (SCoPLe). In summary, the main contributions of this work include: + +- We propose a dual-stream language prompting module that enhances the semantic extraction capabilities of the pre-trained CLIP text encoder, particularly improving its generalization to refine textual features for novel classes. +- We introduce a semantic-guided skeleton prompting module. It learns joint-shaped prompts to refine skeleton features, utilizing an adaptive visual representation sampler to implement cross-modal prompted feature alignment. +- We evaluate SCoPLe in both ZSSAR and Generalized ZSSAR scenarios, demonstrating its ability to achieve state-of-the-art performance on the unified baselines of NTU-RGB+D 60, NTU-RGB+D 120, and PKU-MMD, without requiring extra manual semantic knowledge definition/enrichment or local movement segmentation. + +# 2. Related Works + +# 2.1. Skeleton-based Zero-shot Action Recognition + +Zero-shot skeleton-based action recognition (ZSSAR) addresses the challenge of recognizing skeleton-based actions without access to training samples for target classes. This + +paradigm represents a specialized knowledge transfer problem in which the source and target domains maintain non-overlapping label distributions while sharing the same input domain of uniformly shaped skeleton sequences. Generalized ZSSAR extends this challenge by incorporating discrimination from the seen class during testing. + +Early methodologies primarily focused on two types of strategies: visual-to-semantic projection and shared encoding space construction. In 2019, [42] initially proposed fundamental learning frameworks based on these two theories. They demonstrated that one can either utilize common-space metric learning under a Relation-Net architecture, which transforms both skeleton and text format features, or apply visual-to-semantic embedding projection using a DeViSE [12] model, followed by similarity-based semantic matching. ReViSE [39] aligned latent space outputs from individual data modality autoencoders by minimizing maximum mean discrepancy loss. Moreover, CADA-VAE [33] and SynSE-ZSL [14] employed variational autoencoders for generative latent space with cross/intra-modal reconstruction alignment, introducing part-of-speech decomposition to enhance semantic overlap and improve knowledge transfer. PURLS [51] and STAR [9] regard action recognition as the discrimination of combinations of overlapping sub-motions appearing in both seen and unseen classes. By utilizing GPTs to extract sub-motion semantics, they established a multi-scale cross-modal representation matching to enrich seen-to-unseen knowledge transfer. During semantic alignment, PURLS uses fixed preamble prompts for each sub-motion description, while STAR employs naive prompting concatenation before applying feature projection for each local part, where the main prediction improvements for both solutions arise from the introduction of labor-intensive local pattern learning. SMIE [50] focuses on maximizing mutual information for distributional semantic alignment while incorporating temporal rank loss for more comprehensive feature matching. SA-DVAE [22] tackles information imbalance by disentangling skeleton features into semantically related and irrelevant components. + +While preceding baselines incorporate sophisticated artificial designs to boost knowledge transfer in skeleton sequence features, our method fully harnesses the pre-trained CLIP knowledge through purely data-driven task adaptation, eliminating the need for additional manual transferrable knowledge definition, enrichment, or local movement segmentation. + +# 2.2. Prompt Learning + +Before the proliferation of prompting techniques, fine-tuning was pivotal for adapting pre-trained large models (PLMs) to specific tasks. Although it efficiently leveraged PLM knowledge, it required labor-intensive supervised training to adjust the model's backbone weights for targeting task + +beneficial features. Depending on the domain discrepancies between pre-training and task data, fine-tuning can vary its learning object as either the final layers or the entire model. However, the process eventually becomes cumbersome and costly due to the exponential increase in model size, with parameters ranging from millions to trillions [11, 20, 29]. + +To mitigate these challenges, researchers have explored lightweight plug-in modules that can freeze all pre-trained parameters, using small training tasks to find additional learnable embeddings concatenated to the standard inputs to optimize the output. Prompt learning, which originated from manual label prompting operations in CLIP [30], was first systematized by Lester et al. [19], who introduced learnable vectors to Transformer inputs. [49] refines this concept by enabling the training of independent prompt sub-embeddings that are attached to the front of the original label embeddings. However, static prompts lack dynamic and visual-level generalization across various classes [48], leading to the integration of visual cues via Meta-Net for generating conditional prompts based on instances. Concurrently, [23] emphasized that the inclusion of layer-level learnable prompts can further improve task adaptability in feature extraction during inference, and [18] correspondingly designed a prompting framework tailored for visual transformer branches. LoRA [16] advocates for inserting trainable low-rank matrices within the query projection of a Transformer attention module to facilitate nuanced parameter tuning for task adaptation. In 2022, [15] consolidated the theoretical derivation regarding adapters, prefix tuning (i.e. prompt learning) and low-rank matrix updates, pointing out that these solutions serve the same adaptation intuition but differ only in design choices. Based on their conclusive formula, they introduced Parallel Adapters for competitive accuracy through combined parallel learning and non-linear low-rank adjustments. + +Previous approaches have demonstrated the effectiveness of prompt learning in fully supervised training environments by experimenting with various prompt insertion positions within model architectures and diverse integration methodologies. However, their research focuses on RGB-based and textual inputs, lacking in-depth exploration of adapting skeleton-based features, and present deterioration risks on base-to-novel generalization. Our approach addresses these issues by developing skeleton-feature-friendly prompting and introducing attention normalization to preserve the authentic label contexts for better new class generalization. + +# 3. Methodology + +Problem Definition. Let $\mathcal{D}_{tr} = \{(x_{tr}^{sc},y_{tr}^{sc})\}$ represents the set of $N_{tr}$ training samples from the available seen classes $\mathcal{V}^{sc}$ . Each skeleton sequence $x_{tr}^{sc}\in \mathbb{R}^{L\times J\times M\times 3}$ records the 3-D locations of $J$ body joints for each actor across $L$ frames. Here, $M$ denotes the maximum number of actors per sequence, and $y_{tr}^{sc}\in \mathcal{V}^{sc}$ is the corresponding action + +![](images/4c731817dd2850b25478dcc1a90f45f8717dec7e20031dceec21843616237a1f.jpg) +(a) + +![](images/eccc028fccb6eaf3c5da3f160aa73a9bd6bec24f23fb4a96a83db10165e1af99.jpg) +(b) +Figure 1. (a) Overview of our proposed SCoPLe to solve (G)ZSSAR within a cross-modal semantic alignment framework. SCoPLe tunes both the skeleton and language branches by learnable prompts, while the original weights of the backbones remain frozen. On the linguistic side, SCoPLe enhances the text branch with a dual-stream language prompting (DSLP) module, effectively preserving the strengths of the original pre-trained CLIP knowledge to normalize textual prompt learning. On the visual side, SCoPLe introduces a semantic-guided skeleton prompting module, which includes joint-shaped prompts and an adaptive visual representation sampler to optimize the skeleton feature projection for synergistic semantic alignment with text-side tuning. (b) Illustration of the dual-stream language prompting module. + +label from the seen class label set. Similarly, let $\mathcal{D}_{te} = \{(x_{te}^{uc},y_{te}^{uc})\}$ denote the set of $N_{te}$ testing samples from the unseen classes $\mathcal{Y}^{uc}$ . In a standard ZSSAR setting, we have $\mathcal{Y}^{sc}\cap \mathcal{Y}^{uc} = \phi$ . By training solely on seen-class samples, we aim to develop a model capable of robustly projecting skeleton features from unseen classes, enabling accurate predictions $\hat{y}\in \mathcal{Y}^{uc}$ by mapping these features to their corresponding label embeddings. To extend the problem to GZSSAR, we require the model to predict $\hat{y}\in \mathcal{Y}^{sc}\cup \mathcal{Y}^{uc}$ , where it should accurately identify both seen and unseen classes. This presents extra challenges for the model, as it must not only generalize to unseen classes but also distinguish them from seen classes. + +# 3.1. Model Overview + +Fig. 1 (a) presents the overall architecture of the proposed Semantic-guided Cross-Modal Prompt Learning (SCoPLe) based on a cross-modal embedding alignment framework. Unlike previous approaches that rely on strong assumptions about label semantics or visual information structures to enhance learnable knowledge transfer, our method emphasizes automatic learning of textual and visual prompts to improve feature representation and cross-modal alignment. To achieve this, we propose two key components: 1) a dual-stream language prompting module and 2) a semantic-guided skeleton prompting module. Below we describe the details. + +# 3.2. Dual-stream Language Prompting + +We propose using learnable textual prompts on a pre-trained CLIP text encoder $\mathcal{T}$ to assist the extraction of key context knowledge signals from action labels, reducing the dependence on using manual descriptions for refining the linguistic semantic targets used for skeleton-side projection alignment. + +This allows the model to automatically adapt for learning consistent and task-specific features across actions, improving knowledge transfer and facilitating the recognition of new actions. However, due to the learning mechanism in the CLIP text encoder (based on a Transformer architecture), common prompt learning methods that concatenate learnable prompts with label text tokens [18] for integrated tuning can disrupt the entire self-attention distributions that interpret the semantic contexts of the original labels, consequently limiting and degrading the innate generalizability of CLIP, if the prompts are only trained with seen class samples. + +To address this, for each layer in $\mathcal{T}$ , we attach a dual-stream language prompting module. As shown in Fig. 1 (b), the module processes the layer input in two streams: a regular encoding stream to extract original label token features, which preserves the authentic semantics of the input text label $y$ , and a prompted encoding stream that enhances contextual information shared across actions for task adaptation. For the second stream, before conducting prompt-inclusive attention and feedforward calculations, we first decorate its input with a layer-level prompt and apply a learnable weighted aggregation to combine the attention output from the first stream. This enables the model to selectively preserve the original semantic context structure while effectively incorporating the prompted tuning to refine the token feature content. Formally, before sending $y$ to $\mathcal{T}$ we first follow the operation from [49] and concatenate a preamble prompt initialized as the phrase 'A skeleton video for the action' before the label text. Since each word will have its own token embedding, this allows us to replace some parts of the preamble phrase with learnable prompts without modifying the label texts. We swap the first two tokens in the preamble prompt with a learnable vector $P_0^{\mathcal{T}}$ of the + +same length to incorporate basic input channel tuning, and these positions are also used as the swapping slots for later layer-wise prompts. The input for $\mathcal{T}$ is now defined as $E_0$ . Then, as shown in Fig. 1 (b), for any arbitrary layer $\mathcal{T}_s$ where $s \in [1, S]$ , $S \leq Z$ and $Z$ denotes the number of layers in $\mathcal{T}$ , it is primitively composed of a multihead attention layer $\mathrm{Att}_s$ and a feedforward layer $\mathrm{FFN}_s$ with pre-trained parameter weights. In our learning process, it handles two streams of input: $E_{s-1}$ and $E_{s-1}'$ . $E_{s-1}$ is the output of the previous layer $\mathcal{T}_{s-1}$ that encodes $E_0$ without any additional layer-level tuning modification, while $E_{s-1}'$ is the second stream output from $\mathcal{T}_{s-1}$ that incorporates encoding tuning by the previous layer-wise prompt. For the original encoding stream, the output $E_s$ of $\mathcal{T}_s$ is calculated from $E_{s-1}$ as + +$$ +\begin{array}{l} \hat {E} _ {s} = \operatorname {A t t} _ {s} \left(\operatorname {L a y e r N o r m} \left(E _ {s - 1}\right)\right), (1) \\ \hat {E} _ {s} ^ {\text {r e s}} = E _ {s - 1} + \hat {E} _ {s}, (2) \\ E _ {s} = \mathcal {T} _ {s} (E _ {s - 1}) \\ = \hat {E} _ {s} ^ {r e s} + \operatorname {F F N} _ {s} (\text {L a y e r N o r m} (\hat {E} _ {s} ^ {r e s})), (3) \\ \end{array} +$$ + +where $\hat{E}_s$ is the output of $\mathrm{Att}_s$ and $\hat{E}_s^{res}$ is the result of adding residual connections. For the prompted encoding stream, $E_{s - 1}^{\prime}$ first replaces its decorated positions prompted by the previous layer with the current layer prompt $P_{s}^{\mathcal{T}}$ . We denote this output by $E_{s - 1}^{*}$ . Then, it is enhanced with the attention output of $\hat{E}_s$ to emphasize the information of the original label semantics. This is achieved through a weighted aggregation controlled by a learning layer parameter $\alpha_{s}$ that balances between the tuning effect and the original encoding for a best generalized adaptation. The resulting $E_{s - 1}^{**}$ then goes through the ordinary attention learning and feedforward network in $\mathcal{T}_s$ using the pre-trained parameters and produces $E_s^\prime$ as the prompted stream input for the next layer. Note that the location decorated with layer prompts is now encoded as $\hat{P}_s^T$ rather than the native value of $P_{s}^{\mathcal{T}}$ . The actual procedures are implemented as + +$$ +E _ {s - 1} ^ {*} = \left[ P _ {s} ^ {\mathcal {T}}, E _ {s - 1} ^ {\prime} \backslash \hat {P} _ {s - 1} ^ {\mathcal {T}} \right], \tag {4} +$$ + +$$ +E _ {s - 1} ^ {* *} = \left(1 - \alpha_ {s}\right) E _ {s - 1} ^ {*} + \alpha_ {s} \hat {E} _ {s}, \tag {5} +$$ + +$$ +\hat {E} _ {s} ^ {\prime} = \operatorname {A t t} _ {s} \left(\operatorname {L a y e r N o r m} \left(E _ {s - 1} ^ {* *}\right)\right), \tag {6} +$$ + +$$ +\hat {E} _ {s} ^ {\prime r e s} = E _ {s - 1} ^ {* *} + \hat {E} _ {s} ^ {\prime}, \tag {7} +$$ + +$$ +\begin{array}{l} E _ {s} ^ {\prime} = \mathcal {T} _ {s} (E _ {s - 1} ^ {\prime}) \\ = \hat {E} _ {s} ^ {\prime r e s} + \operatorname {F F N} _ {s} \left(\operatorname {L a y e r N o r m} \left(\hat {E} _ {s} ^ {\prime r e s}\right)\right), \tag {8} \\ \end{array} +$$ + +where $[\cdot, \cdot]$ refers to the concatenation operation and $\backslash$ refers to the removal operation. We implement layer-wise prompting on the first $S$ layers while keeping the remaining blocks in $\mathcal{T}$ unchanged as regular layers. This ensures that the subtle contextual tuning in each layer eventually accumulates and is comprehended in a non-linear manner at the end to influence the final output. The final output $F \in \mathbb{R}^{d_{\text{text}}}$ of $\mathcal{T}$ + +is computed from the subsequent encoding of the rest $Z - S$ layers according to $E_S'$ from the $S$ th prompted layer as + +$$ +\left(E _ {S}, E _ {S} ^ {\prime}\right) = \left(\mathcal {T} _ {s} \left(E _ {S - 1}\right), \mathcal {T} _ {s} \left(E _ {S - 1} ^ {\prime}\right)\right), \tag {9} +$$ + +$$ +F = \operatorname {T e x t P r o j} \left(E _ {S} ^ {\prime}\right). \tag {10} +$$ + +$d_{text}$ is the final text feature dimension and $\mathrm{TextProj}(\cdot)$ is the combination of the remaining $Z - S$ layers. $S$ is a pre-definable hyperparameter. + +# 3.3. Semantic-guided Skeleton Prompting + +While each skeleton joint feature captures spatial-temporal visual information, they lack direct semantic context unless learned in conjunction with the entire human body. The primary objective of visual prompting is to strengthen semantic tuning interactions between the skeleton joint features and the corresponding label embeddings. To achieve this, we propose a semantic-guided skeleton prompting module that learns joint-shaped prompts for the visual side based on the prompted textual output. This refines the skeleton features to enable a more accurate representation and understanding of the visual data in relation to textual cues. + +Joint-shaped Prompts. Previous work on learnable prompts [18, 48, 49] has established a basic paradigm for effectively integrating prompts by integrating them into the original data structure of the sample during the feature feedforward process. Therefore, we propose a joint-shaped prompt that matches the unique encoding format of skeleton features. Given a 3D skeleton sequence input $x \in \mathbb{R}^{L \times J \times M \times 3}$ , we first use a universal pre-trained Shift-GCN [10] backbone to extract its joint visual features $G = \mathcal{G}(x) \in \mathbb{R}^{J \times d_{skel}}$ , where $d_{skel}$ is the skeleton feature dimension. To simplify the calculations, we average the features across $M$ performers and the temporal dimensions. The visual prompt is then established in the format of $P^{\nu} \in \mathbb{R}^{L_{P^{\nu}} \times d_{skel}}$ , where $L_{P^{\nu}}$ is the prompt length and the feature shape aligns with the real joint feature. The prompt is initialized with random values $z \sim N(0,1)$ . The design intuition is to simulate $L_{P^{\nu}}$ fake joint features that introduce contextual information facilitating text semantic alignment. + +Adaptive Visual Representation Sampler. To extract textual-relevant alignment information from the skeleton features, we propose an Adaptive Visual Representation Sampler that leverages the textual branch output to perform cross-attention sampling on the original joint features and visual prompts, with weights determined by their semantic relevance to the targeted labels (see Fig. 2). This obtains a semantic-relevant visual feature for better skeleton-text alignment. Given the skeleton feature output $G$ , the joint-level prompt $P^{\mathcal{V}}$ and the textual embedding $F$ , we establish three attention matrices $\mathbf{W}_Q \in \mathbb{R}^{d_{text} \times h}$ , $\mathbf{W}_K \in \mathbb{R}^{d_{skel} \times h}$ , and $\mathbf{W}_V \in \mathbb{R}^{d_{skel} \times d_R}$ . $h$ refers to the dimension size of the projected features and $d_R$ is the dimension size of the output feature. We project $F$ to get the query vector $F_Q \in$ + +![](images/73c027d3dcdbed44a5abed0c6b403d620c366e296f2df21e2768bfd98c584849.jpg) +Figure 2. Illustration of the internal architecture of Adaptive Visual Representation Sampler. Each colored block represents either an text encoding output, a fake joint feature (i.e. skeleton-shaped prompt), or a true joint feature. + +$\mathbb{R}^{1 \times h}$ and find its semantic relevance with the projected key vector $G_{K} \in \mathbb{R}^{(J + L_{P^{\nu}}) \times h}$ of the concatenated vector of $P^{\nu}$ and $G$ . After obtaining the relevance score $A \in \mathbb{R}^{1 \times (J + L_{P^{\nu}})}$ through the external product between $F_{Q}$ and $G_{K}$ , we can apply a dot production between $A$ and the projected value vector $G_{V} \in \mathbb{R}^{(J + L_{P^{\nu}}) \times d_{R}}$ to produce the visual representation $R \in \mathbb{R}^{d_{R}}$ that adaptively highlights the most class-relevant prompted visual contexts according to the given label. + +$$ +F _ {Q} = F \mathbf {W} _ {Q}, \tag {11} +$$ + +$$ +G _ {K} = \left[ P ^ {\mathcal {V}}, G \right] \mathbf {W} _ {K}, G _ {V} = \left[ P ^ {\mathcal {V}}, G \right] \mathbf {W} _ {V}, \tag {12} +$$ + +$$ +A = \operatorname {S o f t m a x} \left(\left(F _ {Q} \times G _ {K} ^ {T}\right) / \sqrt {h}\right), \tag {13} +$$ + +$$ +R = A G _ {V}. \tag {14} +$$ + +As indicated in Fig. 1 (a), $R$ then goes through an MLP $\mathcal{V}$ to generate $V\in \mathbb{R}^{d_{text}}$ , which expresses the projected visual embedding of $x$ in the text encoding space. $\mathbf{W}_Q,\mathbf{W}_K,\mathbf{W}_V$ are randomly initialized and learn along with $\mathcal{V}$ . + +# 3.4. Learning Losses + +At this point, to achieve a complete learning optimization for generalization towards dual-modal alignment in ZSSAR, we have proposed respective prompting solutions for each branch that appropriately enhance the visual and textual representations of $V$ and $F$ from the pre-trained skeleton and CLIP backbones. During the training stage, we employ the standard contrastive learning loss [30] to minimize the embedding similarity between the projected visual and textual representations. This can be written as + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t r a i n}} (V, F) = - \frac {1}{2} \log \frac {\exp \left(\frac {V F}{\tau}\right)}{\sum_ {o} ^ {\mathcal {Y} _ {\text {b a t c h}}} \exp \left(\frac {V F _ {o}}{\tau}\right)} (15) \\ - \frac {1}{2} \log \frac {\exp \left(\frac {V F}{\tau}\right)}{\sum_ {w} \exp \left(\frac {V _ {w} F}{\tau}\right)}, (16) \\ \end{array} +$$ + +where $F_{o}$ refers to the class textual embedding of all the other $o$ (negative) samples in the same batch, and $V_{w}$ represents the + +projected visual embedding from all the other $w$ (negative) skeleton samples. The temperature parameter $\tau$ regulates the training gradient. + +During the inference stage, a skeleton sequence input $x_{te}$ goes through the frozen skeleton backbone to get $G_{x_{te}}$ . Using $G_{x_{te}}$ and the learned prompted text encoder $\mathcal{T}$ to obtain the textual embedding $F_{y}$ and the corresponding representation $V_{y}^{x_{te}}$ for each candidate $y \in \mathcal{Y}^{uc}$ based on cross-attention query, the classification is based on the search for the nearest class neighbors of $V_{y}^{x_{te}}$ . The following equations interpret the mentioned prediction process. + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t e s t}} \left(V _ {y} ^ {x _ {t e}}, F _ {y}\right) = - \frac {1}{2} \log \frac {\exp \left(\frac {V _ {y} ^ {x _ {t e}} F _ {y}}{\tau}\right)}{\sum_ {o} y _ {u c} ^ {y u c} \exp \left(\frac {V _ {o} ^ {x _ {t e}} F _ {o}}{\tau}\right)}, (17) \\ \hat {y} _ {t e} = \underset {y \in \mathcal {Y} ^ {u c}} {\arg \min } \mathcal {L} _ {t e s t} \left(V _ {y} ^ {x _ {t e}}, F _ {y}\right). (18) \\ \end{array} +$$ + +# 4. Results + +# 4.1. Datasets + +We follow [22] to evaluate the ZSSAR performance of our model on three datasets according to the metrics of a) general performance and b) random three-split prediction from [22]. The splits are based on cross-subjects, ensuring that the training and testing subjects do not overlap. In a test split formatted as 'A/B', A refers to the number of seen classes, while B refers to the number of unseen classes. + +NTU-RGB+D 60 [34] is a prominent benchmark dataset in skeleton-based HAR, offering a vast and diverse collection of data for 60 different common actions. It comprises 56,880 video samples from 40 subjects, captured from 80 distinct camera viewpoints. In our experiment, we used its skeleton sequences collected from 25 body joints per person (setting 2 performers in each sample). Following the zero-shot evaluation protocols in [14], we tested our model under a 55/5 split and a 48/12 split. + +NTU-RGB+D 120 [25] is an extension of NTU-60, which stands as one of the most extensive datasets for HAR at present. This dataset includes 114,480 video samples across 120 action classes, significantly expanding the diversity and complexity of actions compared to its predecessor. The actions are performed by 106 subjects and captured from 155 different camera viewpoints. Analogous to the previous dataset, we follow [14] and use the splits of $110 / 10$ and $96 / 24$ for the evaluation setups. + +PKU-MMD is a mid-scale dataset that encompasses 51 action categories and provides more than 20,000 action instances. It is notable for its inclusion of complex and interactive actions, which makes it suitable for evaluating algorithms in both individual and group activity recognition. We calculated the average accuracy from three random split settings in the format of 46/5. The selected splits are referenced from the same experiments in [22]. + +
ModelBackboneNTU-RGBD 60NTU-RGBD 120
ZSL55/5 splitHZSL48/12 splitHZSL110/10 splitHZSL96/24 split
SUSUSUSUH
ReViSE [39]BERT53.9174.2234.7347.3217.4962.3620.7731.1655.0448.6944.8446.6832.3849.6625.0633.31
JPoSE [42]BERT64.8264.4450.2956.4928.7560.4920.6230.7551.9347.6646.4047.0532.4438.6222.7928.67
CADA-VAE [33]BERT76.8469.3861.7965.3728.9651.3227.0335.4159.5347.1649.7848.4435.7741.1134.1437.31
SynSE [14]BERT75.8161.2756.9359.0233.3052.2127.8536.3362.6952.5157.6054.9438.7056.3932.2541.04
SMIE [50]BERT77.98///40.18///65.74///45.30///
STAR [9]CLIP81.4069.0069.9069.4045.1062.7037.0046.6063.3059.9052.7056.1044.3051.2036.9042.90
PURLS [51]CLIP79.23///40.99///71.95///52.01///
SA-DVAE [22]BERT82.3762.2870.8066.2741.3850.2036.9442.5668.7761.1059.7560.4246.1258.8235.7944.50
SCoPLe (ours)CLIP84.1069.6071.9470.7552.9654.4961.8357.9374.5363.5161.0862.2752.1753.3351.1852.23
+ +Table 2. ZSL and GZSL accuracy (%) on the NTU RGB+D datasets. Each baseline chooses either a pretrained Sentence-BERT or CLIP-ViT-B/32 as its text backbone. S is for seen classes, U is for unseen classes, H is for the harmonic mean of seen/unseen classes. + +
ModelNTU-RGBD 60NTU-RGBD 120PKU-MMD
55/5 split48/12 split46/5 split
ZSLSUHZSLSUHZSLSUH
ReViSE [39]60.9471.7552.0660.3444.9048.2934.6440.3459.3460.8942.1649.82
JPoSE [42]59.4466.2554.9260.0546.6949.4339.1443.6957.1760.2645.1851.64
CADA-VAE [33]61.8477.3558.1466.3845.1551.0941.2445.6460.7463.1735.8645.75
SynSE [14]64.1975.8460.7767.4747.2841.7345.3643.4753.8563.0940.6949.47
SMIE [50]65.08///46.40///60.83///
SA-DVAE [22]84.2078.1672.6075.2750.6758.0940.2347.5466.5458.4951.4054.72
SCoPLe (ours)83.7275.3280.1777.6753.3470.4744.2954.0871.4162.1749.6954.85
+ +Table 3. Average ZSL and GZSL accuracy $(\%)$ under the random split settings on the NTU-60, NTU120, and PKU-MMD datasets. + +# 4.2. Implementation Details + +To fairly highlight the superiority of our method, we use Shift-GCN as the skeleton backbone, consistent with the previous baseline (see Sec. 3.1 in our Supplementary Materials for ablation study with various backbones). Our text prompt is embedded within a pre-trained CLIP-ViT-B/32 text encoder. Adhering to the training protocols in [14], we only use samples from the seen classes to pre-train the skeleton backbone for each split. The visual features are then obtained from the $d_{skel}$ -dimensional penultimate layer of the backbone, preserving the standard feature format of $G \in \mathbb{R}^{J \times d_{skel}}$ where $d_{skel}$ is 256. The textual features $F$ are represented by the 512-dimensional encoding outputs. + +We implement the projection layer $\mathcal{V}$ as a two-layer MLP with SiLU activations and a Tanh output layer. The size $d_{proj}$ of each hidden layer in $\mathcal{V}$ is 512. We set the length of prefix prompts $P_0^T$ to 2 and the depth of layer prompt tuning $S$ to 3. The number of joint-shaped prompts $L_{P\nu}$ is 3. Following the operations in [18], we randomly initialize the layer-level prompts $P_s^T$ for each Transformer block of $s\in [0,S)$ according to a normal distribution $\sim N(0,1)$ . The Adaptive Visual Representation Sampler holds $\mathbf{W}_Q\in \mathbb{R}^{512\times 150}$ , $\mathbf{W}_K\in \mathbb{R}^{256\times 150}$ and $\mathbf{W}_V\in \mathbb{R}^{256\times 256}$ where $h = 150$ and $d_R = 256$ , using single-headed attention learning. $\alpha_{s}$ in each prompted layer in $\mathcal{T}$ is initialized to 0.5 to balance contributions from the original text encoding and the prompted tuning streams at the beginning. We conducted our experiments using PyTorch on a machine equipped with a single A100 GPU. Each training process takes 100 epochs and a learning rate of $1e - 4$ . The batch size is set to 256. We use an Adam optimizer with $\beta 1 = 0.9$ and $\beta 2 = 0.999$ . + +Adaptation to Generalized ZSSAR: We design an extra step to mitigate the classification confusion between seen + +and unseen classes. Drawing inspiration from OOD domains, we implement a gating mechanism based on the prediction entropy confidence. In detail, we first let SCoPLe make similarity predictions $p_y = \mathcal{L}_{test}(V_y^{x^{te}}, F_y)$ for a testing sample $x_{te}$ with only seen class candidates $y \in \mathcal{Y}^{sc}$ . Using this as an inference logit, we obtain the entropy $H_y$ by calculating $H_y = -p_y \log p_y$ for all seen class candidates. Given a pre-defined entropy threshold $\delta$ , we assume that the model will more confidently predict $H_y$ to be below $\delta$ if $x_{te}$ has a ground truth label $y$ from seen classes, since its feature should exhibit more representative characteristics when being predicted as a seen class. If all $H_y$ exceed $\delta$ , we then perform the standard zero-shot classification, in which the model classifies $x_{test}$ as one of the unseen classes in $\mathcal{Y}^{uc}$ . + +![](images/aa5d5961d1769f524aec6f3598a3114a25fb381378959ffc7fcd1a61e92dc9d4.jpg) +(a) + +![](images/916486bf9101d1b38ac83ec93dbedeb553723af0c6c4aefd94fe9bfa1d1ca686.jpg) +(b) +Figure 3. Ablation study on Prefix Prompt Length (a), Layer Prompt Depth (b), and Joint-shaped Prompt Number (c) in SCoPLe with the testing split of 12/48 in NTU-60. + +![](images/1da38b1cd6b7e7d6fab9987095e2ab82c232896d3187f470dbec2ef5d55592f2.jpg) +(c) + +# 4.3. Overall Performance + +As shown in Tab. 2, we compare our method with all previous ZSSAR solutions following the same testing protocols originally established by [14]. For zero-shot learning, SCoPLe outperforms other baselines and achieves state-of-the-art prediction accuracy in each split case. Note that, due to varying design requirements, the textual backbones used among methods can differ and may significantly impact performance improvement. Nevertheless, during the random 3-split testing experiment, all solutions are equipped with the same CLIP backbone, and Tab. 3 demonstrates that our method still has an advantage over the previous SOTA baseline of SA-DVAE [22] for the 48/12 split and the 46/5 split, with accuracy increases of $+2.67\%$ , and $+4.87\%$ . Similarly, for the more challenging task of Generalized ZSSAR in both the general and 3-split protocols, SCoPLe improves + +
ModelNTU-RGBD 60NTU-RGBD 120
ZSL55/5 splitHZSL48/12 splitHZSL110/10 splitHZSL96/24 split
SUSUSUSU
SA-DVAE [22] (last SOTA)82.3762.2870.8066.2741.3850.2036.9442.5668.7761.1059.7560.4246.1258.8235.7944.50
SCoPLe w/o all prompts62.5841.1680.6354.5033.1653.7858.3555.9763.9239.1362.3548.0848.1532.0264.3342.76
SCoPLe w/o text prompts63.3937.2182.8751.3632.4153.6757.4855.5164.6949.6961.0054.7746.1432.2264.2442.92
SCoPLe w/ text prompts (no dual)75.3257.8570.6163.5944.7362.7451.1456.3565.9358.6656.9057.7646.0254.1841.7447.15
SCoPLe w/o skel. prompts70.2167.9261.0364.2940.6955.3555.9455.6468.8867.8243.7553.1948.1751.8449.4650.62
SCoPLe w/skel. prompts (no sampler)70.7265.6066.0465.8240.7848.9061.9354.6567.9267.4745.1754.1140.4148.6652.9350.71
SCoPLe (full)84.1069.6071.9470.7552.9654.4961.8357.9374.5363.5161.0862.2752.1753.3351.1852.23
+ +Table 4. Ablation study on each proposed module with all testing splits for ZSL and GZSL on NTU-RGB+D datasets. + +the best hybrid accuracy of previous baselines with $+1.35\%$ , $+11.33\%$ , $+1.85\%$ , $+7.73\%$ in Tab. 2, and $+2.40\%$ , $+6.54\%$ , $+0.13\%$ in Tab. 3. In Sec. 2.4 of our Supplementary Materials, we include additional experiments to examine the limit when training data becomes significantly insufficient to produce a valid effect on ZSSAR. + +Furthermore, we made a training efficiency comparison between PURLS and SCoPLe using the 110/10 split on NTU 120. The latter simplifies the operations of preparing eight labor-intensive text description embedding (7 local&1 global) and uses the original single label embedding with cross-attention query, achieving better performance with a lower training time of 53 mins while PURLS took 72 mins. + +# 4.4. Ablation Study + +Tuning impacts of different prompting modules. Tab. 4 records our ablation experiment results on NTU datasets by removing each proposed prompting module or mechanism along the linguistic or visual branches. Specifically, we thoroughly examine the gradual ablation by testing the removal of all prompt learning (w/o all prompts), all layer-level text prompting (w/o text prompts), the attention normalization introduced by the dual-stream mechanism (w/ text prompts (no dual)), all skeleton prompting (w/o skeleton prompts), and the Adaptive Visual Representation Sampler (w/ skeleton prompts (no sampler)). Their results are horizontally compared with the performance of the full model (full). Based on the outcomes, we can arrive at three obvious conclusions. First, by comparing the ablation study among w/o all prompts, w/o text prompts, w/o skeleton prompts, and full, we can validate that each prompting brings orthogonal prediction improvements over directly using CLIP. Especially, using only skeleton prompting without any language guidance and textual prompting (i.e. w/o text prompts) results in unstable advancement. This again corroborates the fact that ZSL is highly dependent on the quality of language embedding to enhance the efficiency of knowledge transfer. In addition, comparing the study between w/ text prompts (no dual) and full, we find that adding dual-stream normalization not only improves prediction in ZSL protocols, but also better balances prediction accuracy between seen and unseen classes for GZSL protocols, thus improving overall performance in this scenario. Finally, comparing the study between w/ skeleton prompts (no sampler) and full, we + +reveal that the introduction of textual guidance through the adaptive visual representation sampler significantly improves the auxiliary effect of visual side prompting. + +Hyperparameter Selection. In Fig. 3, we conduct a hyperparameter grid search and illustrate the effect of changing the length of $P_0^T$ (left), the layer prompting depth of $S$ (middle), and $L_{P^{\nu}}$ (right). We find that optimal design choices occur when the length of $P_0^T$ is 2, $S$ is 3, and $L_{P^{\nu}}$ is 3. The performance for values larger than these begins to decrease, which can be interpreted as an overfitting situation where the model cannot effectively allocate new parameters to enhance prediction generalization. + +A key characteristic of SCoPLe is that each module is entirely data-driven and functions at full feature level. This theoretically enhances its versatility beyond skeleton-based recognition, especially for the text-based prompting brought by DSLP. In Sec. 4 of our Supplementary Materials, we have also conducted a preliminary exploration of image-based base-to-novel generalization tasks. + +# 5. Conclusion + +In conclusion, this paper presents SCoPLe, a novel semantic-guided cross-modal prompt learning framework. By introducing dual-stream language prompting learning and semantic-guided adaptive skeleton prompting learning modules, our approach shifts from manually crafted assumptions to a purely data-driven solution for adapting PLMs to skeleton-based zero-shot learning tasks. The experimental results demonstrate that SCoPLe achieves state-of-the-art performance on the NTU-RGB+D 60, NTU-RGB+D 120, and PKU-MMD datasets. A limitation of our method compared to manually refined solutions like PURLS [51] is the deficiency of interpretability, considering that SCoPLe trades it for improved generalization and efficiency through automatic learning on prompt tuning. In addition, future research can further explore the extension of the concept of our prompting normalization to other data modalities. + +# 6. Acknowledgement + +This research was partially supported by the Australian Government through the Australian Research Council's DECRA funding scheme (Grant No: DE250100030) and Discovery Projects scheme (Grant No: DP230101534). + +# References + +[1] Automatic video annotation of human health care action via clustering. J. Medical Imaging Health Informatics, 10(10): 2512-2521, 2020. +[2] 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 (CVPR), 2016. +[3] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. +[4] Cesar Cadena, Anthony Dick, and Ian Reid. Multi-modal auto-encoders as joint estimators for robotics scene understanding. 2016. +[5] Yi Cao, Chen Liu, Zilong Huang, Yongjian Sheng, and Yongjian Ju. Skeleton-based action recognition with temporal action graph and temporal adaptive graph convolution structure. 80(19):29139-29162, 2021. +[6] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. +[7] Ronald Juárez Chambi, Juan Carlos Gutierrez Cáceres, and César Armando Beltrán Castañón. Densenet 3d for violent action recognition in surveillance video sequences. In 41st International Conference of the Chilean Computer Science Society, SCCC 2022, Santiago, Chile, November 21-25, 2022, pages 1-8. IEEE, 2022. +[8] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13359-13368, 2021. +[9] Yang Chen, Jingcai Guo, Tian He, and Ling Wang. Fine-grained side information guided dual-prompts for zero-shot skeleton action recognition, 2024. +[10] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 180-189, 2020. +[11] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity, 2022. +[12] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc' Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2013. + +[13] Thomas Golda, Johanna Thiemich, Mickael Cormier, and Jürgen Beyerer. For the sake of privacy: Skeleton-based salient behavior recognition. In 2022 IEEE International Conference on Image Processing, ICIP 2022, Bordeaux, France, 16-19 October 2022, pages 3983-3987. IEEE, 2022. +[14] Pranay Gupta, Divyanshu Sharma, and Ravi Kiran Sarvadevabhatla. Syntactically guided generative embeddings for zero-shot skeleton action recognition. In 2021 IEEE International Conference on Image Processing (ICIP), pages 439-443, 2021. +[15] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning, 2022. +[16] 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, 2021. +[17] Mustaqem Khan, Jamil Ahmad, Abdulmotaleb El-Saddik, Wail Gueaieb, Giulia De Masi, and Fakhri Karray. Drone-hat: Hybrid attention transformer for complex action recognition in drone surveillance videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024 - Workshops, Seattle, WA, USA, June 17-18, 2024, pages 4713-4722. IEEE, 2024. +[18] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multimodal prompt learning, 2023. +[19] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. +[20] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online, 2020. Association for Computational Linguistics. +[21] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In CVPR, pages 3595–3603. Computer Vision Foundation / IEEE, 2019. +[22] Sheng-Wei Li, Zi-Xiang Wei, Wei-Jie Chen, Yi-Hsin Yu, Chih-Yuan Yang, and Jane Yung jen Hsu. Sa-dvae: Improving zero-shot skeleton-based action recognition by disentangled variational autoencoders, 2024. +[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online, 2021. Association for Computational Linguistics. +[24] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quaan Ngiam, Daiyi Peng, Junyang Shen, Bo Wu, Yifeng Lu, + +Denny Zhou, Quoc V. Le, Alan Yuille, and Mingxing Tan. Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection, 2022. +[25] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10): 2684-2701, 2019. +[26] Qiang Nie and Yunhui Liu. View transfer on human skeleton pose: Automatically disentangle the view-variant and view-invariant information for pose representation learning. IJCV, 129(1):1-22, 2021. +[27] Yoshihisa Nitta and Yuko Murayama. Privacy-aware remote monitoring system by skeleton recognition. In 52nd Hawaii International Conference on System Sciences, HICSS 2019, Grand Wailea, Maui, Hawaii, USA, January 8-11, 2019, pages 1-8. ScholarSpace, 2019. +[28] Quang-Tien Pham, Duc-Anh Nguyen, Tien-Thanh Nguyen, Thanh Nam Nguyen, Duy-Tung Nguyen, Dinh-Tan Pham, Thanh-Hai Tran, Thi-Lan Le, and Hai Vu. A study on skeleton-based action recognition and its application to physical exercise recognition. In The 11th International Symposium on Information and Communication Technology, SoICT 2022, Hanoi, Vietnam, December 1-3, 2022, pages 239-246. ACM, 2022. +[29] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. +[30] 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, 2021. +[31] M Rashmi and Ram Mohana Reddy Guddeti. Skeleton based human action recognition for smart city application using deep learning. In 2020 international conference on communication systems & networks (COMSNETS), pages 756-761. IEEE, 2020. +[32] Justin Sanchez, Christopher Neff, and Hamed Tabkhi. Real-world graph convolution networks (rw-gcns) for action recognition in smart video surveillance. CoRR, abs/2201.05739, 2022. +[33] E. Schonfeld, S. Ebrahimi, S. Sinha, T. Darrell, and Z. Akata. Generalized zero- and few-shot learning via aligned variational autoencoders. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8239–8247, Los Alamitos, CA, USA, 2019. IEEE Computer Society. +[34] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1010-1019, 2016. +[35] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In CVPR, pages 12026-12035. Computer Vision Foundation / IEEE, 2019. +[36] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. + +In CVPR, pages 7912-7921. Computer Vision Foundation / IEEE, 2019. +[37] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks. IEEE Transactions on Image Processing, 29:9532-9545, 2020. +[38] Zehua Sun, Qiuhong Ke, Hossein Rahmani, Mohammed Bennamoun, Gang Wang, and Jun Liu. Human action recognition from various data modalities: A review. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-20, 2022. +[39] Yao-Hung Hubert Tsai, Liang-Kang Huang, and Ruslan Salakhutdinov. Learning robust visual-semantic embeddings, 2017. +[40] Thang Vu, Kookhoi Kim, Tung M. Luu, Xuan Thanh Nguyen, and Chang D. Yoo. Softgroup for 3d instance segmentation on point clouds, 2022. +[41] Christian Wojek, Stefan Walk, Stefan Roth, and Bernt Schiele. Monocular 3d scene understanding with explicit occlusion reasoning. pages 1993-2000, 2011. +[42] Michael Wray, Gabriela Csurka, Diane Larlus, and Dima Damen. Fine-grained action retrieval through multiple parts-of-speech embeddings. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 450-459, 2019. +[43] Wentian Xin, Ruyi Liu, Yi Liu, Yu Chen, Wenxin Yu, and Qiguang Miao. Transformer for skeleton-based action recognition: A review of recent advances. Neurocomputing, 537: 164-186, 2023. +[44] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition, 2018. +[45] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11784-11793, 2021. +[46] Yan Zhang. Human action parsing in untrimmed videos and its applications for elderly people healthcare. PhD thesis, University of Ulm, Germany, 2020. +[47] Yunkai Zhang, Yinghong Tian, Pingyi Wu, and Dongfan Chen. Application of skeleton data and long short-term memory in action recognition of children with autism spectrum disorder. Sensors, 21(2):411, 2021. +[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models, 2022. +[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. +[50] Yujie Zhou, Wenwen Qiang, Anyi Rao, Ning Lin, Bing Su, and Jiaqi Wang. Zero-shot skeleton-based action recognition via mutual information estimation and maximization. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 5302-5310. ACM, 2023. +[51] Anqi Zhu, Qiuhong Ke, Mingming Gong, and James Bailey. Part-aware unified representation of language and skeleton for zero-shot action recognition. CoRR, abs/2406.13327, 2024. \ No newline at end of file diff --git a/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/images.zip b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3805e75dff0a569faee483574dd386d7d7ece83e --- /dev/null +++ b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81b9c442c8333ade8e53d0144541cbf25e4928f1cad2fa8a22884f859c0287a0 +size 340011 diff --git a/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/layout.json b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..223b3240e072d339518e6daf4268880b7e65160d --- /dev/null +++ b/CVPR/2025/Semantic-guided Cross-Modal Prompt Learning for Skeleton-based Zero-shot Action Recognition/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83ab4e93ced7647b91f0d2051ee205a8cfb8f844bb59aa25caafafa1d18b8e8e +size 453190 diff --git a/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/42ac13af-fc7a-4c50-8fb5-d598ec99a242_content_list.json b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/42ac13af-fc7a-4c50-8fb5-d598ec99a242_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c21621f7550795519f0e7345db34c0e97579c3b0 --- /dev/null +++ b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/42ac13af-fc7a-4c50-8fb5-d598ec99a242_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3b23c4732857dcc35c8de208aefd6cf2a3c4b99d29c5f54cfc6a47f64a7883a +size 77343 diff --git a/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/42ac13af-fc7a-4c50-8fb5-d598ec99a242_model.json b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/42ac13af-fc7a-4c50-8fb5-d598ec99a242_model.json new file mode 100644 index 0000000000000000000000000000000000000000..96932726ff7cf97fb78e9a1f1f5159b14abe083d --- /dev/null +++ b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/42ac13af-fc7a-4c50-8fb5-d598ec99a242_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02486bded34ac962031938dd7ab13cc4eabc484db0bedcf35c7f898ad895b1b6 +size 99048 diff --git a/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/42ac13af-fc7a-4c50-8fb5-d598ec99a242_origin.pdf b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/42ac13af-fc7a-4c50-8fb5-d598ec99a242_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a403f8e3fdae22861ddf70806b99287cc8a5e65 --- /dev/null +++ b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/42ac13af-fc7a-4c50-8fb5-d598ec99a242_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64a01c01b762cb685ca47f62e6b58c9930890fce973c80abf94f49c284fdfc9b +size 7091803 diff --git a/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/full.md b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d6424b1d231dd8800365dbf071df1d99e92dc143 --- /dev/null +++ b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/full.md @@ -0,0 +1,307 @@ +# SemanticDraw: Towards Real-Time Interactive Content Creation from Image Diffusion Models + +Jaerin Lee $^{1}$ Daniel Sungho Jung $^{2,3*}$ Kanggeon Lee $^{1}$ Kyoung Mu Lee $^{1,2,3}$ + +$^{1}$ ASRI, Department of ECE, $^{2}$ Interdisciplinary Program in Artificial Intelligence, $^{3}$ SNU-LG AI Research Center, Seoul National University, Korea + +{ironjr,dqj5182,dlrkdrjs97,kyoungmu}@snu.ac.kr + +inputs: drawing with semantic palette + +outputs: image stream + +![](images/ea289ce450aefa9c3b805f511cbac73161010cdecd5b9befb8a91c1041c57122.jpg) + +![](images/44f893b02e3161537f982a3ae724d64790978b32070fac9b8c228ae6666014e4.jpg) +Figure 1. Overview. Our SEMANTICDRAW is a sub-second (0.64 seconds) solution for region-based text-to-image generation. This streaming architecture enables an interactive application framework, dubbed semantic palette, where image is generated in near instant interactivity based on online user commands of hand-drawn semantic masks. + +![](images/0a8df98d2e610bba8325f09c9210d9d17df8507276c4568fe6d0042134c12271.jpg) +Prompts: Row 1: Background: “A sunset at a grassland”, Red: “A horse”. Row 2: Background: “A person is time standing in front of a white background”, Red: “A girl smiling at viewer”, Blue: “Doggy body part”, Green: “Beautiful flower garden”. Row 3: Blue: “Underwater coral reef”, Yellow: “Yellow fish”, Red: “A shark”. + +# Abstract + +We introduce SemanticDraw, a new paradigm of interactive content creation where high-quality images are generated in near real-time from given multiple hand-drawn regions, each encoding prescribed semantic meaning. In order to maximize the productivity of content creators and to fully realize their artistic imagination, it requires both quick interactive interfaces and fine-grained regional controls in their tools. Despite astonishing generation quality from recent diffusion models, we find that existing approaches for regional controllability are very slow (52 seconds for $512 \times 512$ image) while not compatible with acceleration methods such as LCM, blocking their huge potential in interactive + +content creation. From this observation, we build our solution for interactive content creation in two steps: (1) we establish compatibility between region-based controls and acceleration techniques for diffusion models, maintaining high fidelity of multi-prompt image generation with $\times 10$ reduced number of inference steps, (2) we increase the generation throughput with our new multi-prompt stream batch pipeline, enabling low-latency generation from multiple, region-based text prompts on a single RTX 2080 Ti GPU. Our proposed framework is generalizable to any existing diffusion models and acceleration schedulers, allowing subsecond (0.64 seconds) image content creation application upon well-established image diffusion models. The code is https://github.com/ironjr/semantic-draw. + +# 1. Introduction + +Recent massive advancements and widespread adoptions of generative AI [1, 41, 42, 44, 46, 58] are fundamentally transforming the landscape of content creation, demonstrating huge potential for improving efficiency of production processes and expanding the boundaries of creativity. Especially, diffusion models [44] are gaining significant attention in generative AI for image content creation because of their ability to produce realistic, high-resolution images. Nevertheless, in the perspective of content creators, a pure generative quality is not the only point of consideration [35]. Diffusion models for content creators should require efficient, interactive tools that can swiftly translate their artistic imaginations into refined outputs, supporting a more responsive and iterative creative process with fine-grained controllability under straightforward control panels as illustrated in Figure 1 and 7. These goals should all be satisfied simultaneously. + +The academic community had several attempts to address these criteria in isolated areas, but has yet to tackle them comprehensively. On one hand, there is a line of works dealing with acceleration of the inference speed [8, 25, 32, 33, 43, 51, 52] of diffusion models. Acceleration schedulers including DDIM [51], latent consistency models (LCM) [32, 33, 52], SDXL-Lightning [25], HyperSD [43], and Flash Diffusion [8] reduced the number of required inference steps from several thousand to a few tens and then down to 4. Focusing on the throughput directly, StreamDiffusion [21] reformed diffusion models into a pipelined architecture, enabling streamlined generation and real-time video styling. On the other hand, methods to enhance the controllability [4, 5, 57, 58] of the generative framework were also heavily sought. ControlNet [58] and IP-Adapter [57] enabled image-based conditioning of the pre-trained diffusion models. SpaText [4] and MultiDiffusion [5] achieved image generation from multiple region-based texts, allowing more fine-grained controls over the generation process from localized text prompts. Those two areas of research have developed largely independently. This suggests a straightforward approach for fast yet controllable generation: simply combining achievements from both, e.g., acceleration technique such as LCM [33] can serve a pair of a noise schedule sequence and fine-tuned model weights. + +However, directly combining multiple works together does not work as intended. Figure 2 illustrates an example where diffusion models fail when extended to complex real-world scenarios. Here, inspired from the famous yet complex artwork of Korean royal folding screen, Irworobongdo ("Painting of the Sun, Moon, and the Five Peaks")1, we generate an image of size $768 \times 1920$ from nine region- + +![](images/a15d1fef2567d6ee90b9eb1669b95a5df840df2270ffb47bb9a8f23ff29dcfcf.jpg) + +![](images/fb8badabeea295176bff2dcccc759d17f4ba621c70d896d5e30ef88cbdb60618.jpg) + +![](images/145dc0b7d233aa101e91bd9ffad5dce4786fd86a227c9c6fb00f3456802c9525.jpg) +Figure 2. Example of large-size region-based text-to-image synthesis inspired by Korean traditional art, Irworobongdo. Our SEMANTICDRAW can synthesize high-resolution images from multiple, locally assigned text prompts with $\times 52.5$ faster speed of convergence. The size of the image is $768 \times 1920$ and we use 9 text prompt-mask pairs including the background. The time is measured with a RTX 2080 Ti GPU. Note that time takes longer than regular sized images (e.g., $512 \times 512$ ) due to panoramic shape. + +![](images/7d3b71b00ccb39ab3afabaa6eeffba117743f08ec1e2d9910e9bf7eea2de8e2f.jpg) +Text prompt: Background: "Clear deep blue sky", Green: "Summer mountains", Red: "The Sun", Pale Blue: "The Moon", Light Orange: +"A giant waterfall", Purple: "A giant waterfall", Blue: "Clean deep blue lake", Orange: "A large tree", Light Green: "A large tree" + +ally assigned text prompts as defined by a user under Figure 2. At this scale, previous state-of-the-art (SOTA) region-based controlling pipeline [5] fails to match the designated mask regions and text prompts despite its extremely slow and, hence, cautious reverse diffusion process. Applying a famous acceleration method LCM [33] on the diffusion model [5] does not solve high-latency problem, producing noisy output in the second row in Figure 2. This proves that the problem of controllability and acceleration cannot be scaled to real-world scenarios when simply com + +bining the existing diffusion models and acceleration methods, due to their poor compatibility. + +Our goal is to build a real-time pipeline for image content creation, ready for interactive user applications. The system should be operated at least in near real-time, while maintaining stability of fine-grained regional controls. In the end, we propose SEMANTICDRAW which solves the problems from existing methods as shown in Figure 3. Elaborated in Section 3.2, we establish a stable pipeline for accelerated image synthesis with fine-grained controls, given through multiple, locally assigned text prompts. Building upon the rapid development from both acceleration schedulers [8, 25, 32, 33, 43] and network architectures [39, 44, 48] for diffusion models, we propose the first method to allow the acceleration schedulers to be compatible with region-based controllable diffusion models. We achieve up to $\times 50$ speed-up of the multi-prompt generation while maintaining or even surpassing the image fidelity of the original algorithm [5]. + +Even after resolving the compatibility problem between the acceleration and controllability modules, generation throughput remains to be a main obstacle to interactive application. To this end, as illustrated in Section 3.3, we restructure our multi-prompt reverse diffusion process into a pipelined architecture [21], which we call the multi-prompt stream batch architecture. By bundling multi-prompt latents at different timesteps as a batched sequence of requests for image generation, we can perform the multi-prompt text-to-image synthesis endlessly by repeating a single, batched reverse diffusion. The result is a sub-second interactive image generation framework, achieving 1.57 FPS in a single 2080 Ti GPU. This high, stable throughput from SEMANTICDRAW allows a novel type of application for image content creation, named semantic palette, in which we can draw semantic masks in real-time to create an endless stream of images as in Figure 1 and 7. Our model-agnostic and acceleration-agnostic design allows the framework to be suitable for any existing diffusion pipelines [39, 44, 48]. We highly recommend readers to try our technical demo application of semantic palette in our Supplementary Material for better understanding. + +# 2. Related Work + +Accelerating Inference from Diffusion Models. Diffusion model [15, 44, 51] is a branch of generative models that sample target data distributions, e.g., images, videos, sounds, etc., by iteratively reducing randomness from pure noise. Its earliest form [15, 50, 51] traded off inference efficiency against sample diversity and quality. These have required thousands of iterations to generate a single image, raising a need for acceleration of inference to gain practicality. Majority of works [29, 30, 51] achieved speed through reformulating the reverse diffusion process. + +DDIM [51] utilized a non-Markovian graphical model, and DPM-Solvers [29, 30] interpreted the generation process as Euler's method for solving ordinary differential equations, cutting the required number of inference steps from thousands down to 20. Later, Consistency Models [52] exploited identity map boundary condition, and Flow Matching [27] adopted optimal transport for efficient sampling. These became the foundations of the most recent accelerated schedulers, including latent consistency model (LCM) [32, 33], SDXL-Lightning [25], Hyer-SD [43] and Flash Diffusion [8], which are utilized by large-scale latent diffusion models [10, 39, 44] in form of low-rank adaptations (LoRA) [16], a weight modifier upon baseline diffusion models. Alternatively, StreamDiffusion [21] introduced a novel pipelined architecture for video-to-video transfer, video stylization, and streamed image generation from a latent consistency model [32]. Our multi-prompt stream batch pipeline for interactive semantic drawing extends this philosophy with multi-prompt-based generation. + +Controlling Generation from Diffusion Models. Enhanced controllability of diffusion models is another intensely investigated field of study. There are five major subgroups: (1) modifying from intermediate latent vectors, (2) modifying from inpainting masks, (3) attaching separate conditional branches, (4) connecting a subset of prompt tokens to positions in an image, and (5) enabling finer-grained generation from multiple, region-based prompts. The first group including ILVR [9], RePaint [31], and SDEdit [34] attempt to hijack the intermediate latent variables in the reverse process. SSI [23] accelerates this group of methods by utilizing locality of edit command. LazyDiffusion [38] take advantage of the transformer architecture to progressively edit and generate images within few seconds of latency, whereas our method builds upon arbitrary architecture and achieves sub-second generation time. The second major group utilizes the in-painting functionality [37] of diffusion models for editing [3, 17, 31, 37, 55]. After diffusion models have become massively publicized as image generation [2, 44] and editing [12, 19, 20, 28, 34, 36, 53, 56] tools, the demand for easier, modularized controls on behalf of professional creators has increased. In the third group, ControlNet [58] and IP-Adapter [57] introduce simple yet effective way to append image conditioning feature to existing pre-trained diffusion models. Our method applies orthogonally with the control methods in this group. Various text-conditioning [12, 19, 20, 28, 36] and image-conditioning [11, 53, 56] methods can also be placed in this group. The fourth group, including GLIGEN [24] and InstanceDiffusion [54] attach add-on modules to the diffusion model that focus on increasing the positional accuracy of a single prompt. Alternatively, we are mainly interested in a scenario where image diffusion models continuously cre + +![](images/8682d3563eb0183f9fd4420e30649f7a03a19a66e595254dec396511d9bb5b4b.jpg) +(a) Prompt + +![](images/ac844b385602c1af99ece098bf54cefdedd486af76334e1fc54bf3c15abd6571.jpg) +(b) MD 50steps 52s + +![](images/c301966f6c1fa3dc5f7dab13a7f5fef0a496fb711ffb8b7165292fa930d9b82e.jpg) +(c) MD+LCM 5steps 5s +Figure 3. Our SEMANTICDRAW enables fast region-based text-to-image generation by stable acceleration of MultiDiffusion [5]. PreAvg, Bstrap, and QMask stand for the latent pre-averaging, mask-centering bootstrapping, and quantized masks, our first three proposed strategies. Each method used in (d), (e), (f) contains the method used in the previous image. The images are single tiles of size $768 \times 512$ . + +![](images/b99a7091c8fd7355bb95cdd7dc703da84dbc2588febfb4b1a1ce2a8bf7c0ec11.jpg) +(d) +PreAvg 5steps 5s + +![](images/e5eb266b2ae084265bf4f78b042b33cb787981fe7ade86c35d479feaf9a0b943.jpg) +(e) +Bstrap 5steps 5s +(f) +QMask(Ours)5steps 5s + +![](images/f49b809c98fc796ff1b794f02147e0cc5df8a5e73a7b5a9a3db4a5f73406475c.jpg) + +ate new images from multiple, dynamically moving, regionally assigned text prompts. This is most related to the final group [4-6, 40] which focus on controlling the semantic composition of the generated images. + +Content Creation from Regional Text Prompts. The last group mentioned above provides a way to flexibly integrate multiple regionally assigned text prompts into a single image. SpaText [4] achieves generation from multiple spatially localized text prompts by utilizing CLIP-based spatio-temporal representation. Differential Diffusion [22] and Mixture of Diffusers [6] similarly operate on mask-based generation but differs in their approaches to overlapping regions and noise addition. MultiDiffusion [5] and more recent LRDiff [40] present simple yet effective way to generate from multiple different semantic masks: to iteratively decompose and recompose the latent images according to different regional prompts during reverse diffusion process. This simple formulation works not only with irregular-shaped regions, but also with irregular-sized canvases. However, as mentioned in Section 1 and depicted in Figure 2, this breakthrough has not been developed in aware of modern acceleration methods, reducing their practical attraction in this era of rapid diffusion models. Starting from the following section, we will establish the compatibility between these type of pipeline architecture with accelerated samplers. This opens a new type of semantic drawing application, SEMANTICDRAW, where users draw images interactively with brush-type tools that paints semantic meanings as shown in Section 5. + +# 3. Method + +# 3.1. Preliminary + +A latent diffusion model (LDM) [44] $\epsilon_{\theta}$ is an additive Gaussian noise estimator defined over a latent space. The model + +$\epsilon_{\theta}$ receives a combination of a noisy latent $\pmb{x}$ , a text prompt embedding $\pmb{y}$ , and a timestep $t \in [0, T]$ . It outputs an estimation of the noise $\epsilon$ that was mixed with the true latent $\pmb{x}_0$ . At inference, the diffusion model $\epsilon_{\theta}$ is consulted multiple times to estimate a latent $\hat{\pmb{x}}_0 \approx \pmb{x}_0$ , which correlates to the information described in the conditional input $\pmb{y}$ , starting from a pure noise $\pmb{x}_T \sim \mathcal{N}(0,1)^{HWD}$ . Each of the recursive calls to the reverse diffusion process can be expressed as a summation of a denoising term and a noise-adding term to the intermediate latent: + +$$ +\boldsymbol {x} _ {t _ {i - 1}} = \operatorname {S T E P} \left(\boldsymbol {x} _ {t _ {i}}, \boldsymbol {y}, i, \epsilon ; \epsilon_ {\theta}, \alpha , \boldsymbol {t}\right), \tag {1} +$$ + +where, we denote $i$ as the index of the current time step $t_i$ . Note that the newly added noise $\epsilon$ depends on the type of scheduler. + +Although this abstract form embraces almost every generation algorithm of diffusion models [15, 29, 51], it does not consider practical scenarios of our interest: (1) when the desired shape $(H^{\prime}\times W^{\prime})$ of the latent $\hat{\pmb{x}}_0^{\prime}$ is different from that of the training set $(H\times W)$ or (2) multiple text prompts $\pmb{y}_1,\dots ,\pmb{y}_p$ correlate to different regions of the generated images. MultiDiffusion [5] is one of the pioneers to deal with this problem. Their main idea is to aggregate (AGGRSTEP) multiple overlapping tiles of intermediate latents with simple averaging. That is, for every sampling step $t_i$ , perform: + +$$ +\begin{array}{l} \boldsymbol {x} _ {t _ {i - 1}} ^ {\prime} = \operatorname {A G G R S T E P} \left(\boldsymbol {x} _ {t _ {i}} ^ {\prime}, \boldsymbol {y}, i, \mathcal {W}; \text {S T E P}\right) (2) \\ = \frac {\sum_ {\boldsymbol {w} \in \mathcal {W}} \operatorname {S T E P} \left(\operatorname {c r o p} \left(\boldsymbol {w} \odot \boldsymbol {x} _ {t _ {i}} ^ {\prime}\right) , \boldsymbol {y} _ {\boldsymbol {w}} , i , \boldsymbol {\epsilon}\right)}{\sum_ {\boldsymbol {w} \in \mathcal {W}} \boldsymbol {w}}, (3) \\ \end{array} +$$ + +where $\odot$ is an element-wise multiplication, $\boldsymbol{w} \in \mathcal{W} \subset \{0,1\}^{H'W'}$ is a binary mask for each latent tile, $\boldsymbol{y}_{\boldsymbol{w}}$ is a conditional embedding corresponding to the tile $\boldsymbol{w}$ , and crop is a cropping operation to chop down large $\boldsymbol{x}_{t_i}'$ into tiles of same size as training image latents. + +![](images/13749d4f99a68326bfbb6c1706344f5caadee02d0641b40f8dc244f4634d3260.jpg) +(a) Bootstrapping strategy overview. + +![](images/1eee4e57638acc8ebdd3f02dd9c17e41b6701ca804cfc611527abb9acc5e6dcf.jpg) +(b) Multi-prompt stream batch architecture. +Figure 4. SEMANTICDRAW pipeline. Our acceleration technique for region-based multi-prompt generation consists of three strategies. Figure 4a summarizes the first two of three: (1) latent pre-averaging and (2) mask-centering bootstrapping. In Figure 4b, we devise multi-prompt stream batch pipeline that aggregates foreground and background latents from different time steps to maximize the throughput of generation, enabling near real-time content creation. Further, text embeddings are cached for interactive brush-like interface, elaborated in Section 5. Our method can be applied to arbitrary diffusion pipelines. We also provide the full algorithm in the Supplementary Material. + +# 3.2. Acceleration-Compatible Regional Controls + +Our objective is to build an accelerated solution to image generation from multiple regionally assigned text prompts. Unfortunately, simply replacing the Stable Diffusion (SD) model [10, 39, 44] with an acceleration module, such as Latent Consistency Model (LCM) [32] or SDXL-Lightning [25], etc., and updating the default DDIM sampler [51] with the corresponding accelerated sampler [18, 32] does not work in general. This incompatibility greatly limits potential applications of both acceleration [8, 25, 32, 43] and region-based control techniques [4, 5]. We discuss each of the causes and seek for faster and stronger alternatives. In summary, our stabilization trick consists of three strategies: (1) latent pre-averaging, (2) mask-centering bootstrapping, and (3) quantized masks. + +Step 1: Achieving Compatibility through Latent Pre-Averaging. The primary reason for the blurry image of the second row of Figure 2 is that the previous algorithm [5] is not aware of different types of underlying reverse diffusion step functions STEP. While the reverse diffusion algorithms can be categorized into two types: (1) additional noise at each step [29, 32], (2) no additional noise at each step [5, 51], the previous SOTA region-based controllable method [5] falls into the latter. Hence, applying the averaging aggregation of the method [5] cancels the prompt-wise added noises in STEP, which leads to overly smooth latents. We can avoid this problem with a simple workaround. First, we split the STEP function into a deterministic denoising part (DENOISE) and an optional noise addition: + +$$ +\begin{array}{l} \boldsymbol {x} _ {t _ {i - 1}} = \tilde {\boldsymbol {x}} _ {t _ {i - 1}} + \eta_ {t _ {i - 1}} \epsilon (4) \\ = \operatorname {D E N O I S E} \left(\boldsymbol {x} _ {t _ {i}}, \boldsymbol {y}, i; \epsilon_ {\theta}, \alpha , t\right) + \eta_ {t _ {i - 1}} \epsilon , (5) \\ \end{array} +$$ + +where $\eta_{t}$ is an algorithm-dependent parameter. The averaging of equation (3) is then applied to the output of the denoising part $\tilde{\pmb{x}}_{t_{i - 1}}$ , instead of the output of the full step $\pmb{x}_{t_{i - 1}}$ . Note that the noise is added after aggregation step. + +$$ +\boldsymbol {x} _ {t _ {i - 1}} ^ {\prime} = \operatorname {A G G R S T E P} \left(\boldsymbol {x} _ {t _ {i}} ^ {\prime}, \boldsymbol {y}, i, \mathcal {W}; \text {D E N O I S E}\right) + \eta_ {t _ {i - 1}} \epsilon . \tag {6} +$$ + +As it can be seen in Figure 3d, this change alleviates the compatibility issue with acceleration methods like LCM. + +Step 2: Mask-Centering Bootstrapping for Few-Step Generation. The second cause of the incompatibility lies in the bootstrapping stage of the previous method [5]. MultiDiffusion [5] introduced bootstrapping stages that replace the background latents with random colors in the first $40\%$ of total steps. This is performed to cut out the generated regions outside of object masks, which claims to enhance mask-fidelity. In original form, the perturbation introduced by the bootstrapping cancels out during long inference steps. However, as we decrease the number of timesteps in ten-fold from $n = 50$ steps to $n = 4$ or 5 steps, the number of bootstrapping stage is reduced down to $n = 2$ . Regrettably, this magnifies the effect of perturbation introduced by the random color latents in the bootstrapping phase, and results in leakage of mixed colors onto the final image as shown in Figure 3. Instead, we propose to use a mixture of white background and aggregation of contents co-generated from other regional prompts (blue in Figure 4a), which alleviates the problem and allows compatibility with the accelerated generation as seen in Figure 3e. + +Furthermore, we empirically found that first two steps of reverse diffusion process determine the overall structure of generated images when sampling with accelerated schedulers. Even after the first step, the network formulates the + +![](images/378ec14cef482e9919df41b8123cb6bcfaf309c1fa5ba6857037e108cc452776.jpg) +Figure 5. Quantized mask. As the last of three stabilizing techniques, binary masks are blurred and quantized by scheduler noise level to trade off between mask-fidelity and overall harmonization. See Supplementary Material for utilizing this trade-off. + +rough structure of the objects being created. The problem is that diffusion models exhibit a strong tendency to generate screen-centered objects than off-centered ones, following image datasets [49] they are trained for. After the first step, the object for every mask is generated at the center of the screen, not at the center of the mask. Off-centered objects are often masked out by the pre-averaging step (yellow in Figure 4a). The final results often neglect small, off-centered regional prompts, and the large objects are often unnaturally cut, lacking harmonization within the image. To prevent this, we propose mask centering strategy (pink in Figure 4a) to exploit the center-bias of the diffusion model. Especially, for the first two steps of generation, we shift the intermediate latents from each prompt to the center of the frame before being handled by the noise estimator $\epsilon_{\theta}$ . The result of Step 2 can be seen in Figure 3e. + +Step 3: Quantized Mask for Seamless Generation. Another problem from the reduced number of inference steps is that harmonization of the generated content becomes more difficult. As Figure 3e shows, all the objects appear to be salient and their abrupt boundaries are visible between regions. This is because the number of later sampling steps that contribute to the harmonization is now insufficient. In contrast, the baseline with long reverse diffusion steps 3b effectively smooth out the mask boundaries by consecutively adding noises and blurring them. To mitigate this issue, we develop an alternative way to seamlessly amalgamate generated regions: quantized masks, shown in Figure 5. Given a binary mask, we obtain a smoothened mask by applying Gaussian blur. Then, we quantize the real-numbered mask by the noise levels of the diffusion sampler. As Figure 4a illustrates, for each denoising step, we use a mask with corresponding noise level. Since the noise levels monotonically decrease throughout iterations, the coverage of a mask gradually increases along with each sampling step, gradually mixing the boundary regions. The final result can be seen from Figure 3f. This relaxation of semantic masks also provides intuitive interpretation of brushes, one of the most widely used tool in professional graphics editing + +software. We will revisit this interpretation in Section 5. + +# 3.3. Optimization for Throughput + +As mentioned in Section 1, achieving real-time response is important for practical end-user application. Inspired by StreamDiffusion [21], we reconstruct our region-based text-to-image synthesis framework into a pipelined architecture to maximize the throughput of image generation. + +Multi-Prompt Stream Batch Architecture. Figure 4b illustrates the architecture and the interfaces of our pipeline. Instead of the typical mini-batched use of diffusion model with synchronized timesteps, the noise estimator is fed with a new input image every timestep along with the last processed batch of images. In other words, each image in a mini-batch has different timestep. This architecture hides the latency caused by multi-step algorithm of reverse diffusion. Restructuring our stabilized framework in 4a takes several steps. The quantized masks, the background images, the noises, and the prompt embeddings differ along each timesteps and should be saved separately. Instead of a single image, we change the architecture to process a minibatch of images of different prompts and masks to the U-Net at every timestep, as depicted in Figure 4b. We call this the multi-prompt stream batch architecture. To further reduce the latency, we added asynchronous pre-calculation step applied only when a user command changes the configuration of the text prompts and masks. This allows interactive brush-like interfaces elaborated in Section 5. + +Optimizing Throughput. Additional increase of throughput can be achieved by using a compressed autoencoder such as Tiny AutoEncoder [7]. Detailed analysis on the effect of throughput optimization is in Table 6. + +# 4. Experiment + +We provide comprehensive evaluation of our SEMANTIC-DRAW using various types of acceleration modules and samplers. We compare our experiments based on the public checkpoints of Stable Diffusion 1.5 [44], SDXL [39], and SD3 [48]. However, we note that our method can be applied to any community-trained models using DreamBooth [45]. More results can be found in Section S2 of our Supplementary Materials. + +# 4.1. Quality of Generation + +Generation from Multiple Region-Based Prompts. We first demonstrate the stability and speed of our algorithm for image generation from multiple regionally assigned text prompts. The evaluation is based on COCO validation dataset [26], where we generate images from the image captions as background prompts and object masks with categories as foreground prompts. The public latent diffusion + +Table 1. Comparison of generation from region-based prompts between DDIM [51] (default) and LCM [32] sampler. + +
MethodSamplerFID ↓IS ↑CLIPfg ↑CLIPbg ↑Time (s) ↓
SD1.5 (512 × 512)
MultiDiffusion (Ref.)DDIM [51]70.93 ●16.24 ●24.09 ●27.55 ●14.1 ●
MultiDiffusion (MD)LCM [32]270.55 ●2.653 ●22.53 ●19.63 ●1.7 ●
SemanticDraw (Ours)LCM [32]93.93 ●14.12 ●24.14 ●24.00 ●1.3 ●
+ +Table 2. Comparison of generation from region-based prompts between DDIM [51] (default) and Hyper-SD [43] sampler. + +
MethodSamplerFID ↓IS ↑CLIPfg ↑CLIPbg ↑Time (s) ↓
SD1.5 (512 × 512)
MultiDiffusion (Ref.)DDIM [51]70.93 ●16.24 ●24.09 ●27.55 ●14.1 ●
MultiDiffusion (MD)Hyper-SD [43]168.34 ●10.12 ●20.08 ●15.90 ●1.7 ●
SemanticDraw (Ours)Hyper-SD [43]98.60 ●14.90 ●24.48 ●23.31 ●1.3 ●
+ +Table 3. Comparison of generation from region-based prompts between DDIM [51] (default) and Euler Discrete [18] sampler. + +
MethodSamplerFID ↓IS ↑CLIPtg ↑CLIPbg ↑Time (s) ↓
SDXL (1024 × 1024)
MultiDiffusion (Ref.)DDIM [51]73.77 ●16.31 ●24.16 ●28.11 ●50.6 ●
MultiDiffusion (MD)EulerDiscrete [18]572.95 ●1.328 ●21.02 ●17.36 ●4.3 ●
SemanticDraw (Ours)EulerDiscrete [18]84.27 ●15.04 ●24.19 ●24.22 ●3.6 ●
+ +Table 4. Comparison of generation from region-based prompts between Flow Match Euler Discrete [10] (default) and Flash Flow Match Euler Discrete [8] sampler. + +
MethodSamplerFID ↓IS ↑CLIPfg ↑CLIPbg ↑Time (s) ↓
SD3 (1024 × 1024)
MultiDiffusion (Ref.)FlowMatch [10]166.42 ●8.517 ●20.66 ●16.39 ●46.3 ●
MultiDiffusion (MD)FlashFlowMatch [8]209.36 ●5.347 ●19.83 ●14.48 ●4.0 ●
SemanticDraw (Ours)FlashFlowMatch [8]79.2 ●17.41 ●23.59 ●27.83 ●3.2 ●
+ +models [39, 44, 48] are trained for specific range of image sizes, and reportedly fail when given image sizes are small. Since COCO datasets consists of relatively small images compared to the default size the models were trained for, we rescale the object masks with nearest neighbor interpolation to the default size of each model. This is $512 \times 512$ for SD1.5 [44] and $1024 \times 1024$ for SDXL [39] and SD3 [48]. To compare the image fidelity, we use Fréchet Inception Distance (FID) [14] and Inception Score (IS) [47]. We also use CLIP scores [13] to compare the text prompt fidelity. We separate the foreground score $(\mathrm{CLIP}_{\mathrm{fg}})$ , which is obtained by taking the average CLIP score between each generated image and corresponding set of foreground object categories, from the background score $(\mathrm{CLIP}_{\mathrm{bg}})$ , which is a measured between images and their corresponding COCO captions. Tables 1 through 4 summarize the results. + +We implement MultiDiffusion [5] for SDXL [39] and SD3 [48] simply by changing the pipelines, accelerator LoRAs [16], and schedulers, from the official implementation. Even though schedulers with higher numbers of iterations generally produce better quality images [51], the tables show that our accelerated pipeline achieves comparable quality with more than $\times 10$ reduction of time. These results demonstrate that our method provides universal acceleration under different types of diffusion pipelines (SD1.5 [33] + +Background: "Plain wall", Yellow: "A desk", Red: "A flower vase", Blue: "A window" + +![](images/bcb2e90ef95beea96287759f3fd975e39d4c8a1c8968eca38d79494a2bf32ef4.jpg) + +![](images/946845a3e248beac00f099806ca0abd9f63cab212cc4ce2d803955d8904b61c4.jpg) +Background: "A photo of backyard", Yellow: "A yellow bird", Red: "A red bird" + +![](images/7aeb0cfed1362d1616095535f28b9d08284b12f5cc150cb939d775e48894bd91.jpg) +Background: "A floor", Yellow: "A box", Red: "A tiny head of a cat" + +![](images/158a23e7482ebff19b3b9c1072d3962fca14d630e666724af8015c6c5b95399c.jpg) +Background: "A photo", Yellow: "A smiling girl", Red: "A cool beret hat", Blue: "Sky at noon" +(a) Prompt +(b) MD, 50 steps (c) MD+LCM, 5 steps +(d) Ours, 5 steps +Figure 6. Region-based text-to-image synthesis results. Our stabilization methods accelerate MultiDiffusion [5] up to $\times 10$ while preserving quality. + +, SDXL [39], SD3 [10]), noise schedulers (DDIM [51], LCM [32], Euler Discrete [18], Flow Match Euler Discrete [10]), and acceleration methods (LCM [32], Lightning [25], Hyper-SD [43], Flash Diffusion [8]), without compromising the visual quality. Figure 6 shows a random subset of generation from the experiments in Table 1. Comparable visual quality from our method is consistent to the quantitative comparisons. + +Stabilized Acceleration of Region-Based Generation. Next, we evaluate the effectiveness of each stabilization step introduced in Section 3.2. Figure 3 and Table 5 summarize the result on region-based text-to-image generation from the same setup as Table 1. Applying each strategy consistently boosts both perceptual quality, measured by FID score [14], and text prompt-fidelity, measured by the two CLIP scores [13]. This reveals that our techniques help alleviating the incompatibility as intended. + +Throughput Maximization. Table 6 compares the effect of throughput optimization. We have already achieved $\times 9.7$ speed-up by establishing the compatibility with acceleration modules. This is further enhanced though our multi- + +Table 5. Ablation on the effectiveness of our stabilization techniques on the fidelity of region-based generation. + +
MethodFID ↓CLIPfg ↑CLIPbg ↑
No stabilization270.5522.5319.63
+ Latent pre-averaging80.6422.8026.95
+ Mask-centering bootstrapping79.5423.0626.72
+ Quantized masks (σ = 4)78.2123.0826.72
+ +Table 6. Ablations on throughput optimization techniques, measured with a single RTX 2080 Ti. Images of $512 \times 512$ are generated from three prompt-mask pairs. + +
MethodThroughput (FPS)Relative Speedup
Baseline [5]0.0189×1.0
+ Stable Acceleration0.183×9.7
+ Multi-Prompt Stream Batch1.38×73.0
+ Tiny AutoEncoder [7]1.57×83.1
+ +Table 7. User preference regarding quality of generation. + +
MethodOursMD [5]MD [5]+LCM [32]
Preference90.9%9.1%0.0%
+ +prompt stream batch architecture. With low-memory autoencoder [7] to trade quality off for speed, we could finally achieve 1.57 FPS (0.64 seconds per frame). This near real-time, sub-second generation speed is a necessary step towards practical applications of generative models. + +User Study. Finally, we conduct a user study on the methods of Table 1. Its result, summarized in Table 7, shows that our method greatly increases generation quality with multiple region-based text prompts. + +# 5. SemanticDraw + +Our real-time interface of SEMANTICDRAW opens up a new paradigm of user-interactive application for image generation. We discuss the key features of the application. + +Concept. Responsive region-based text-to-image synthesis enabled by our streaming pipeline allows users to edit their prompt masks similarly to drawing. Since the standard text encoding by large text encoders (e.g., CLIP) accounts for approximately $40\%$ of our sub-second runtime (1.57 FPS), caching and reusing these encodings when only mask modifications occur hides this latency and provides even faster feedback to users. This allows them to iteratively refine their commands according to the generated image. In scenarios where users change text prompts, standard text processing occurs while still maintaining the original sub-second runtime. This enables users to paint with text prompts just like they can paint a drawing with colored brushes, hence the name: SEMANTICDRAW. + +![](images/54125044579792e1a6aeea47266fda2693841f571c8afa8f3b219ab697b542ad.jpg) +(a) Semantic Draw. + +![](images/bd97870016538bfd65842a5e6ace3741fdc6383cf47fe87e734b49c8bc8df7ca.jpg) +(b) Streaming Semantic Draw. +Figure 7. Screenshot of the sample applications of SEMANTIC-DRAW. After registering prompts and optional background image, the users can create images in real-time by drawing with text prompts. We invite the readers to play with the application. + +Sample Application Design. We briefly describe our minimal demo application in Figure 7. The application consists of a front-end user interface and a back-end server that runs SEMANTICDRAW. Each user input is either a modification of the background image, the text prompts, the masks, and the tweakable options for the text prompts and the masks such as mix ratios and blur strengths. When commanding major changes requiring preprocessing stages, such as a change of prompts or the background, the backend pipeline is flushed and reinitialized with the newly given context. Otherwise, the pipeline is repeatedly called to obtain a stream of generated images. The user first selects the background image and creates a palette of semantic masks by entering a pair of positive and negative text prompts. The user can then draw masks corresponding to the created palette with a familiar brush tool, a shape tool, or a paint tool. The application automatically generates a stream of synthesized images according to user inputs. We gently invite the readers to play with our technical demo provided with the official code2. + +# 6. Conclusion + +We proposed SEMANTICDRAW, a new type image content creation where users interactively draw with a brush tool that paints semantic masks to endlessly and continuously create images. Enabling this application required high generation throughput and well-established compatibility between regional control pipelines and acceleration schedulers. We devised multi-prompt regional control pipeline that is both scheduler-agnostic and model-agnostic in order to maximize the compatibility. We further proposed multi-prompt stream batch architecture to build a near real-time, highly interactive image content creation system for professional usage. Our SEMANTICDRAW achieves up to $\times 50$ faster generation of large scale images than the baseline, bringing the latency of multi-prompt irregular-sized generation down to a practically meaningful bounds. + +# Acknowledgment + +This work was supported in part by the IITP grants [No.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No. 2021-0-02068, and No.2023-0-00156], and the NOTIE grant (No. RS2024-00432410) by the Korean government. + +# 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. 2 +[2] AUTOMATIC1111. Stable diffusion WebUI. https://github.com/AUTOMATIC1111/stable-diffusion-webui, 2022.3 +[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 3 +[4] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. SpaText: Spatio-textual representation for controllable image generation. In CVPR, 2023. 2, 4, 5 +[5] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. MultiDiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2, 3, 4, 5, 7, 8 +[6] Álvaro Barbero Jiménez. Mixture of Diffusers for scene composition and high resolution image generation, 2023. 4 +[7] Ollin Boer Bohan. Tiny autoencoder for stable diffusion. https://github.com/madebyollin/taesd, 2023.6,8 +[8] Clement Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin. Flash Diffusion: Accelerating any conditional diffusion model for few steps image generation. In AAAI, 2025. 2, 3, 5, 7 +[9] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. In ICCV, 2021. 3 +[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 3, 5, 7 +[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. *AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning*. In *ICLR*, 2023. 3 +[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2022. 3 +[13] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 7 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by + +a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 7 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4 +[16] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language models. In ICLR, 2021. 3, 7 +[17] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. BrushNet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In ECCV, 2024.3 +[18] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 5, 7 +[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, 2023. 3 +[20] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. DiffusionCLIP: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 3 +[21] Akio Kodaira, Chenfeng Xu, Toshiki Hazama, Takanori Yoshimoto, Kohei Ohno, Shogo Mitsuhori, Soichi Sugano, Hanying Cho, Zhijian Liu, and Kurt Keutzer. StreamDiffusion: A pipeline-level solution for real-time interactive generation. arXiv preprint arXiv:2312.12491, 2023. 2, 3, 6 +[22] Eran Levin and Ohad Fried. Differential diffusion: Giving each pixel its strength, 2023. 4 +[23] Muyang Li, Ji Lin, Chenlin Meng, Stefano Ermon, Song Han, and Jun-Yan Zhu. Efficient Spatially Sparse Inference for Conditional GANs and Diffusion Models. In NeurIPS, 2022. 3 +[24] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-Set Grounded Text-to-Image Generation. In CVPR, 2023. 3 +[25] Shanchuan Lin, Anran Wang, and Xiao Yang. SDXL-Lightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 2, 3, 5, 7 +[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6 +[27] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow Matching for Generative Modeling. In ICLR, 2023. 3 +[28] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! Image synthesis with semantic diffusion guidance. In WACV, 2023. 3 +[29] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 3, 4, 5 + +[30] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3 +[31] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. RePaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 3 +[32] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 3, 5, 7, 8 +[33] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. LCM-LoRA: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2, 3, 7 +[34] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 3 +[35] Niloy J. Mitra, Duygu Ceylan, Or Patashnik, Danny CohenOr, Paul Guerrero, Chun-Hao Huang, and Minhyuk Sung. Diffusion models for visual content creation. In SIGGRAPH Tutorial, 2024. 2 +[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 3 +[37] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3 +[38] Yotam Nitzan, Zongze Wu, Richard Zhang, Eli Shechtman, Daniel Cohen-Or, Taesung Park, and Michael Gharbi. Lazy Diffusion Transformer for Interactive Image Editing. In ECCV, 2024. 3 +[39] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2, 3, 5, 6, 7 +[40] Zipeng Qi, Guoxi Huang, Chenyang Liu, and Fei Ye. Layered Rendering Diffusion Model for Controllable Zero-Shot Image Synthesis. In ECCV, 2024. 4 +[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2 +[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[43] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, XING WANG, and Xuefeng Xiao. Hyper-SD: Trajectory segmented consistency model for efficient image synthesis. In NeurIPS, 2024. 2, 3, 5, 7 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image syn + +thesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5, 6, 7 +[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 6 +[46] 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. In NeurIPS, 2022. 2 +[47] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved Techniques for Training GANs. In NeurIPS, 2016. 7 +[48] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion Distillation, 2024. 3, 6, 7 +[49] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open dataset of CLIP-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 6 +[50] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep Unsupervised Learning using Nonequilibrium Thermodynamics. In ICML, 2015. 3 +[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 2, 3, 4, 5, 7 +[52] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3 +[53] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. In ICLR, 2022. 3 +[54] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. InstanceDiffusion: Instance-level Control for Image Generation. In CVPR, 2024. 3 +[55] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. SmartBrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 3 +[56] 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 CVPR, 2023. 3 +[57] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3 +[58] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3 \ No newline at end of file diff --git a/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/images.zip b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7b47485cfbe48e3fda99cad989a799999f02727c --- /dev/null +++ b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9080ee7d1b4aedd29800b442349d6bd0349ed94bc9e3bc10ec397f35be9cac08 +size 721798 diff --git a/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/layout.json b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9ae5ebae10a64c2c7652291375cc6eb96abe19e7 --- /dev/null +++ b/CVPR/2025/SemanticDraw_ Towards Real-Time Interactive Content Creation from Image Diffusion Models/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ecb4124d033adbae5210034d512b43ad326ae8d26adbe9e5710ca0c54afc938 +size 381581 diff --git a/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/cb0aa053-c5a4-4837-8b51-1c448c661a37_content_list.json b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/cb0aa053-c5a4-4837-8b51-1c448c661a37_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..435cd4bc3df20da1ed97fc91ba601dc4d62622b8 --- /dev/null +++ b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/cb0aa053-c5a4-4837-8b51-1c448c661a37_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c7402129a011be4771b2d88bac876be4a9142118995e539868cb53d542272e3 +size 91059 diff --git a/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/cb0aa053-c5a4-4837-8b51-1c448c661a37_model.json b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/cb0aa053-c5a4-4837-8b51-1c448c661a37_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3d9561437b73276eb7b1ef47d5a131dce16f0634 --- /dev/null +++ b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/cb0aa053-c5a4-4837-8b51-1c448c661a37_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7dbf823b5bdc2170ab804c4fcb588943e37b494091c62f098c352134c764e85 +size 116828 diff --git a/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/cb0aa053-c5a4-4837-8b51-1c448c661a37_origin.pdf b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/cb0aa053-c5a4-4837-8b51-1c448c661a37_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e3f6c60120ecbd03068d7e0982986e279577646e --- /dev/null +++ b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/cb0aa053-c5a4-4837-8b51-1c448c661a37_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e500d7e2a0e229977250b8a94c550d0b1f5f65df87eae22ac63c645d36713129 +size 6717829 diff --git a/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/full.md b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..105d177142296572f0792f07b6408d1aba74e1f1 --- /dev/null +++ b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/full.md @@ -0,0 +1,498 @@ +# Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining + +Shangquan Sun $^{1,2}$ + +Shu Wang + +Wenqi Ren $^{3,4,5}$ + +Jianhou Gan + +Juxiang Zhou + +Xiaochun Cao $^{3\dagger}$ + +$^{1}$ Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China + +$^{2}$ School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China + +$^{3}$ School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University + +$^{4}$ MoE Key Laboratory of Information Technology + +$^{5}$ Guangdong Provincial Key Laboratory of Information Security Technology + +$^{6}$ Key Laboratory of Educational Information for Nationalities, Yunnan Normal University + +$^{7}$ School of Mechanical Engineering and Automation, Fuzhou University + +shangquansun@gmail.com, {renwq3,caoxiaochun}@mail.sysu.edu.cn + +# Abstract + +Significant progress has been made in video restoration under rainy conditions over the past decade, largely propelled by advancements in deep learning. Nevertheless, existing methods that depend on paired data struggle to generalize effectively to real-world scenarios, primarily due to the disparity between synthetic and authentic rain effects. To address these limitations, we propose a dual-branch spatiotemporal state-space model to enhance rain streak removal in video sequences. Specifically, we design spatial and temporal state-space model layers to extract spatial features and incorporate temporal dependencies across frames, respectively. To improve multi-frame feature fusion, we derive a dynamic stacking filter, which adaptively approximates statistical filters for superior pixel-wise feature refinement. Moreover, we develop a median stacking loss to enable semi-supervised learning by generating pseudoclean patches based on the sparsity prior of rain. To further explore the capacity of deraining models in supporting other vision-based tasks in rainy environments, we introduce a novel real-world benchmark focused on object detection and tracking in rainy conditions. Our method is extensively evaluated across multiple benchmarks containing numerous synthetic and real-world rainy videos, consistently demonstrating its superiority in quantitative metrics, visual quality, efficiency, and its utility for downstream tasks. + +# 1. Introduction + +Rain streaks in captured videos or images obscure the background, diminishing image quality and potentially + +![](images/0db9a9310040a68f3d6bc180b1f6b1a357953c30f9fb59e1e29d82fd59907dda.jpg) + +![](images/db25c76dd08681ffea4cd8240bc7c217c92ce74cce48fa6596c0b1382eee1990.jpg) + +![](images/d47862319c60243bffbdcef7c91fd11562d58435bed6ba19e1eaedc383f9140c.jpg) +(a) Rain Streak Removal + +![](images/ab3080c093ee294a0a6780a7550b0dfb1f89bfaac1f0921982fd07bbaa7cf803.jpg) + +![](images/1edba14dbe04ad42908c7a63acbeefdadf68f29b44c3b41c8db63c5e49c3c75f.jpg) + +![](images/ba8c9d36445708a752cfe2e58f81603bc1f6fc606b18dd91e54f5fe93ccd01dd.jpg) + +![](images/bd5272a3decb4312fb7bae5e68cac582204042628a8ec8503e050adc02996787.jpg) + +![](images/cf393155ce08caae4a2323df719cd87a4cb73abaa6ffe2423ec6e3458e8ae4ac.jpg) + +![](images/4e44d2ef35aaa7a2d2ef854fc9be617123a4575dadf5f05ac250d24ba4d49ddb.jpg) +(b) Noise Reduction +Figure 1. The functionality of various stacking filters in video restoration. With aligned candidate frames, (a) shows that the min and median filters effectively remove falling rain streaks, while (b) and (c) illustrate how the mean and max filters aid in noise reduction and filling missing pixels, respectively. Our dynamic stacking filter can adaptively approximate these filters at the pixel level, facilitating versatile frame fusion. + +![](images/a6697d974c4ea13e5d2147ea4b8fc3c1a7b6fd1bf69856321d1407b4fc812be7.jpg) + +![](images/e10a366d209c86ab9f4bb53202944dca5f05d40ae7c3cb128cd2bb0e6c66c4fa.jpg) + +![](images/163bdbd0de056663de2b46e42a5d7f90eb52d39f6b9e723a8be087886c63d33b.jpg) +(c) Empty Filling + +![](images/782cc12e8873a2cb3b63e169dc82511a38422f0660fe5c448a3db8aa589cb306.jpg) + +![](images/3732cc419d122921609348ad8a40fa3249f8b68968e97444977680e6a33fea28.jpg) + +![](images/8446a169307e2766389f68c34f40e98329fe9492b87e0c3871a61b702af13f2d.jpg) +(d) Ours + +impairing downstream high-level vision tasks such as object detection [23, 43, 45] and tracking [20]. Given the prevalence of rain as an adverse weather condition, effective rain streak removal is essential. + +Since the advent of deep learning, numerous architectures have been developed to tackle this issue, including convolutional neural networks (CNNs) [7, 60, 66], Transformers [8, 46], and sequential models [54]. Despite these advancements, two key challenges persist: the disparity between synthetic and real-world rain streaks, and the practical application of deraining methods in real-world downstream tasks. Model-based methods can effectively remove rain streaks from real-world rainy frames by solving complex, prior-based objective functions [27, 28, 52]. However, the high computational demands and manually defined hyper-parameters of model-based methods limit their real + +world applicability and generalization. Conversely, data-driven approaches that rely on paired synthetic data struggle to bridge the gap between synthetic and real-world rain streaks, thus failing to achieve effective rain removal in real-world scenarios [54, 66]. Additionally, despite the existence of several synthetic benchmarks [26, 46], the potential of leveraging video deraining to enhance other vision tasks in real-world rainy conditions remains largely unexplored. + +To address these challenges, we introduce a semi-supervised dual-branch spatio-temporal state-space model for video deraining, named VDMamba. Specifically, we implement one-step multi-frame warping by incorporating degradation-free optical flow knowledge into our model. Our dual-branch architecture comprises a spatial feature extraction branch and a temporal feature fusion branch, empowered by two key components: the spatial state-space model layer (S3ML) and the temporal state-space model layer (TSML). To further improve multi-frame information fusion, we derive a stacking-based dynamic filter that selectively integrates pixels from neighboring frames. This filter adaptively approximates various statistical functions, including the min and median filters for rain removal, the mean filter for noise suppression, and the max filter for filling misaligned pixels, as illustrated in Fig. 1. Utilizing the sparsity prior of rain, we incorporate a median stacking loss to enable semi-supervision, thereby enhancing generalization for effective rain streak removal in real-world scenarios. As shown in Fig. 2, our method achieves state-of-the-art performance in video deraining with real-time inference speed and a compact model size. To further assess the effectiveness of video deraining in supporting rain-affected downstream tasks, we establish a benchmark for object detection and tracking in rainy videos. + +In summary, our contributions are as follows: + +- We introduce a spatio-temporal state-space model for video deraining, featuring two branches dedicated to spatial feature refinement and temporal information integration, respectively. +- To achieve improved real-world deraining, we develop a dynamic stacking filter for adaptive pixel aggregation, along with a semi-supervised stacking loss function based on the sparsity prior of rain. +- We also create a real-world rainy video benchmark to assess the effectiveness of deraining methods in facilitating downstream tasks of object detection and tracking. Extensive experiments show our method's superiority across synthetic, real-world, and downstream benchmarks. + +# 2. Related Works + +Single Image Deraining. The rainy image is typically modeled as an additive composition [29, 30], + +$$ +\mathbf {I} = \mathbf {B} + \mathbf {R}, \tag {1} +$$ + +![](images/eff26bddc37c114cfe34c1466dd94b0592d4aa9ee765c750e2adc6732050f2cf.jpg) +Figure 2. The comparison of performance, time and space complexities among video deraining methods. The red line denotes the boundary of real-time inference. + +where $\mathbf{I}$ , $\mathbf{B}$ , and $\mathbf{R} \in \mathbb{R}^{H \times W \times 3}$ represent the rainy image, clean background, and rain streak layer, respectively. Due to the occlusion introduced by $\mathbf{R}$ , recovering $\mathbf{B}$ becomes an ill-posed and inherently challenging task. Numerous approaches have been proposed over the years, primarily falling into two categories: model-based and data-driven methods. Model-based techniques aim to decompose the rain and background layers, employing methods equipped with different priors [6, 29, 30, 38] In contrast, data-driven methods have recently emerged and largely dominated the field, leveraging CNNs [53], Transformers [8], and state-space models [74] to remove rain from the high-frequency components of rainy images [12, 13]. Besides, various techniques like multi-scale approaches [25, 63], convolutional sparse coding [51], adversarial learning [50, 68], transfer learning [53], and self-attention mechanisms [8, 47, 67] have been incorporated. Although single-image deraining methods can be adapted for video deraining, they often fail to fully utilize temporal information. As a result, specialized deraining models tailored for video data are required to achieve optimal performance. + +Video Deraining We can similarly express the relationship for a rainy video: + +$$ +\mathbf {I} _ {n} = \mathbf {B} _ {n} + \mathbf {R} _ {n}, \tag {2} +$$ + +where $\mathbf{I}_n$ , $\mathbf{B}_n$ , and $\mathbf{n}_t$ denote the respective components of the $n$ -th frame in a rainy video. Garg and Nayar were pioneers in video deraining, developing a rain detector based on photometric appearance [14-16]. Drawing inspiration from their work, numerous video deraining methods have since emerged, exploring various sources of prior knowledge such as chromatic properties [70], rain shape [5], low rankness [10, 27, 28], frequency space properties [3], rain sparsity and dynamics [28, 34], and temporal correlations [64]. Techniques such as local phase congruency [42], Gaussian mixture models [52], tensor factorization [27, 31, 40], and statistical models [2] have been employed. However, methods relying on manually tuned + +hyper-parameters and heavy computation often struggle to efficiently capture the complex distribution of rain. + +Recent advancements in video deraining have been driven largely by deep learning-based methods. Various frameworks have been developed, including CNNs [7, 37] recurrent networks [36, 58, 61], self-learning approaches [60, 62], dynamic rain generators [66], a combination of convolutional and recurrent networks [69], multi-patch progressive learning [46], and state-space models [54]. However, existing data-driven methods continue to face challenges in bridging the gap between synthetic and real-world data. Furthermore, they often overlook the application of real-world rain streak removal in videos for downstream tasks, such as object detection and tracking. + +Vision Mamba. State-space models (SSMs) [18, 35, 44, 49, 55, 73], such as Mamba [17, 21], have recently demonstrated their efficiency in long-range modeling. While several works have applied Mamba to vision tasks [19, 33, 56, 71], attention-based and CNN models are still reported to outperform Mamba in local and short-sequence modeling [65]. To enhance global modeling capabilities, existing Mamba-based deraining methods [33, 56, 71] leverage frequency-domain transformations, applying computations in the Fourier space to improve performance. + +# 3. Method + +# 3.1. Video Deraining Pipeline + +Multi-Frame Optical Flow Estimation. Previous optical flow-aware video deraining methods [57, 60] require multiple pairwise optical flow estimations to warp and align several neighboring frames, leading to significant latency during flow prediction. Additionally, optical flow directly estimated from degraded frames may be unreliable [57]. Inspired by knowledge distillation [22, 48], we adopt a pipeline using degradation-free multi-frame optical flow transfer. Given a sequence of $N$ rainy frames $\mathbf{I}_1,\dots,\mathbf{I}_N$ and their corresponding ground-truths, $\mathbf{B}_1,\dots,\mathbf{B}_N$ , a pre-trained flow estimator, $\mathcal{F}$ , predicts the pairwise optical flow for the clean pairs, + +$\tilde{\mathbf{O}}_{n\rightarrow c} = \mathcal{F}\left(\mathbf{B}_n,\mathbf{B}_c\right)$ $n\in \{1,\dots,c - 1,c + 1,\dots,N\}$ (3) where $c$ denotes the index of the current center frame, and $\tilde{\mathbf{O}}_{n\rightarrow c}\in \mathbb{R}^{H\times W\times 2}$ represents the optical flow of the $n$ -th frame towards the center frame. We define our multi-frame flow estimator as $f$ , which estimates multi-frame flows by + +$$ +\mathbf {O} _ {1 \rightarrow c}, \dots , \mathbf {O} _ {N \rightarrow c} = f \left(\mathbf {I} _ {1}, \dots , \mathbf {I} _ {N}\right). \tag {4} +$$ + +The multi-frame flow estimator is trained by minimizing the optical flow transfer loss, which is formulated as: + +$$ +\mathcal {L} _ {\text {f l o w}} = \sum_ {n = 1, n \neq c} ^ {N} \| \mathbf {O} _ {n \rightarrow c} - \tilde {\mathbf {O}} _ {n \rightarrow c} \| _ {1}. \tag {5} +$$ + +![](images/b3b386a13a10d5bc1218b3f9a0557fa45b45859577b78c51088118c83a1cad91.jpg) + +![](images/946bd406df594eec175c2b5065dc7d0873247fbf0947fbb34025841b780993f3.jpg) + +![](images/400f50dd7faee39549f0cd086fdea9e7911852486ab6ca293826f8991444cb5a.jpg) + +![](images/853e01601c71007fbeb856ae9bc919aebe4c90cf872fc86f5bdc74a77f59770b.jpg) +(a) Rainy $\mathbf{I}_n$ +(d) Rainy $\mathbf{I}_{n + 1}$ +Figure 3. Given two adjacent rainy frames (3a & 3d), rain degradation leads to errors in optical flow estimation and subsequent frame alignment (3b & 3c). In contrast, our degradation-free multi-frame estimation pipeline facilitates more accurate warping. + +![](images/5eafc2995095b4507e778ef84573031be1ed798d134850b92a17ad0a5a244fbd.jpg) +(b) Flow by $\mathcal{F}$ +(e) Flow by ours + +![](images/57b9051fa7f32526bbf57531e4a86e490394d0127a87b30d5912b64433b05544.jpg) +(c) Warped by (b) +(f) Warped by (e) + +In this manner, a single model forward pass generates multiple optical flows simultaneously, effectively reducing runtime. Furthermore, this approach helps mitigate the degradation of flow inference caused by rain distortions, as shown in Fig. 3. With the trained multi-frame flow estimation, we can efficiently align neighboring frames for the deraining task. For both multi-frame flow estimation and subsequent video deraining, we use the same structure of our VDMamba, except that the projected output channel dimension is $2 \times (N - 1)$ for flow estimation and 3 for deraining. We also include the restored previous frames along with the rainy frames, though this detail is omitted for brevity. + +Overall Structure. The overall architecture of our VDMamba is shown in Fig. 4. It consists of two distinct branches, one for spatial feature extraction and the other for temporal feature fusion. Both branches follow a U-Net [41] architecture, comprising four levels of feature transformation through upsampling and downsampling. The encoders and decoders are connected through skip connections using residual learning. The network includes two core components: the spatial state-space model layer (S3ML) and the temporal state-space model layer (TSML). Given the spatial and temporal branch features, $\mathbf{F}_S^i\in \mathbb{R}^{H\times W\times C_S}$ and $\mathbf{F}_T^i\in \mathbb{R}^{H\times W\times C_T}$ at the $i$ -th level, with the height $H$ , the width $W$ , and the channels of $C_S$ and $C_T$ , we have + +$$ +\mathbf {F} _ {S} ^ {i} = \operatorname {S 3 M L} \left(\mathbf {F} _ {S} ^ {i - 1}\right), \mathbf {F} _ {T} ^ {i} = \operatorname {S 3 M L} \left(\mathbf {F} _ {T} ^ {i - 1}\right), +$$ + +$$ +\mathbf {F} _ {T} ^ {i + 1} = \mathrm {T S M L} (\mathbf {F} _ {S} ^ {i}, \mathbf {F} _ {T} ^ {i}). +$$ + +At the beginning and end of the branches, two convolutions with a kernel size of $3 \times 3$ are used for patch embedding and output projection, respectively. + +Spatial State-Space Model Layer is designed for spatial feature extraction. Given a spatial feature $\mathbf{F}_S^i$ , we can express S3ML as: + +$$ +\mathbf {F} _ {S, 1} ^ {i}, \mathbf {F} _ {S, 2} ^ {i} = \mathrm {C} _ {1} (\mathbf {F} _ {S} ^ {i}), +$$ + +$$ +\mathbf {F} _ {S, 1, 2} ^ {i} = \mathrm {C} _ {1} \big (\sigma \left(\mathrm {C} _ {5} ^ {\mathrm {D}} \left(\mathbf {F} _ {S, 1} ^ {i}\right)\right) \otimes \mathrm {S S M} \left(\mathrm {L} \left(\mathrm {C} _ {5} ^ {\mathrm {D}} \left(\mathbf {F} _ {S, 2} ^ {i}\right)\right)\right) \big), +$$ + +![](images/936239b569c55f7069a71c7db4ba3764caffafe5d1286c4383794989df762f07.jpg) +Figure 4. The architecture of our proposed VDMamba for video deraining, consisting of spatial state-space model layers (S3ML) for single-frame feature extraction and temporal state-space model layers (TSML) for multi-frame feature fusion. Due to the modular design of the spatial branch, the sub-model, VDMamba-S, which contains only S3ML, is also capable of performing single-image deraining. + +$$ +\mathbf {F} _ {S, 3} ^ {i} = \mathcal {I} \left(\mathrm {C} _ {1} \left(\sigma \left(\mathrm {C} _ {5} ^ {\mathrm {D}} \left(\mathrm {C} _ {1} \left(\mathcal {I} (\mathbf {F} _ {S} ^ {i})\right)\right)\right)\right)\right), +$$ + +$$ +\mathbf {F} _ {S} ^ {i + 1} = \mathrm {L N} \left(\mathbf {F} _ {S, 1, 2} ^ {i} \odot \mathbf {F} _ {S, 3} ^ {i}\right) + \mathbf {F} _ {S} ^ {i}, +$$ + +where $C_1$ , $C_5^{\mathrm{D}}$ , and $L$ denote point-wise convolution of a kernel size of $1 \times 1$ , depth-wise convolution with a kernel size of $5 \times 5$ , and a linear layer, respectively. LN represents layer normalization, while $\sigma$ is the SiLU [11] activation function. The operations $\odot$ and $\otimes$ refer to elementwise multiplication and matrix multiplication, respectively. $\mathcal{I}$ represents the interpolation process. The selective scan module (SSM) operates along the channel dimension to capture global dependencies, while we refine the local information by computing an additional $\mathbf{F}_{S,3}^{i}$ feature. + +Temporal State-space Model Layer is designed for temporal feature fusion. Given a spatio-temporal feature $\mathbf{F}_T^i$ , we can express TSML as: + +$$ +\mathbf {F} _ {T, 1} ^ {i}, \mathbf {F} _ {T, 2} ^ {i} = \mathrm {C} _ {1} (\mathbf {F} _ {T} ^ {i}), +$$ + +$$ +\mathbf {F} _ {T, 1, 2} ^ {i} = \mathrm {C} _ {1} \left(\sigma \left(\mathrm {C} _ {5} ^ {\mathrm {D}} \left(\mathbf {F} _ {T, 1} ^ {i}\right)\right) \otimes \operatorname {S S M} \left(\mathrm {L} \left(\mathrm {C} _ {5} ^ {\mathrm {D}} \left(\mathbf {F} _ {T, 2} ^ {i}\right)\right)\right)\right), \tag {7} +$$ + +$$ +\mathbf {F} _ {T, 3} ^ {i} = \sigma \left(\mathrm {C} _ {5} ^ {\mathrm {D}} \left(\mathrm {C} _ {1} \left(\mathrm {D S F} \left([ \mathbf {F} _ {S} ^ {i}; \mathbf {F} _ {T} ^ {i} ]\right)\right)\right)\right), +$$ + +$$ +\mathbf {F} _ {T} ^ {i + 1} = \operatorname {L N} \left(\mathbf {F} _ {T, 1, 2} ^ {i} \odot \mathbf {F} _ {T, 3} ^ {i}\right) + \mathbf {F} _ {T} ^ {i}, +$$ + +where $[\cdot ;\cdot ]$ denotes the concatenation operation. The SSM operation is performed along the temporal dimension to integrate multi-frame information. DSF refers to the proposed dynamic stacking filter, whose derivation is provided below. + +# 3.2. Dynamic Stacking Filter + +Given the feature candidates of aligned neighboring frames, conventional operations, such as a convolution layer with consistent parameters, struggle to selectively integrate pixel-adaptive features [1]. Inspired by the benefits of stack- + +ing techniques in video deraining, as illustrated in Fig. 1, we propose a dynamic stacking filter to adaptively select a statistics from the mean, median, and extreme values (maximum and minimum) to fuse pixels across the temporal dimension. However, among the statistics, only the mean is a differentiable value. To address the indifferentiability of the others, we define their differentiable approximations. + +Differentiable Approximation to Max and Min. Given a set of pixels $G = \{x_1, \dots, x_N\}$ , if we know the index of the minimum, i.e., + +$$ +\operatorname {a m i n} (G) _ {n} = \left\{ \begin{array}{c l} 0 & , \text {i f} x _ {n} \neq \min (G), \\ \frac {1}{| \{x | x = \min (G) \} |} & , \text {o t h e r w i s e}, \end{array} \right. \tag {8} +$$ + +then we can compute the minimum by + +$$ +\min (G) = \sum_ {n = 1} ^ {N} x _ {n} \operatorname {a m i n} (G) _ {n}. \tag {9} +$$ + +The index of the minimum in Eq. 8 can be approximated using a generalized softmax function, + +$$ +\operatorname {a m a x} (G; a) _ {n} \approx \frac {e ^ {a x _ {n}}}{\sum_ {j = 1} ^ {N} e ^ {a x _ {j}}}, \tag {10} +$$ + +$$ +\operatorname {a m i n} (G) = \operatorname {a m a x} (\{- x _ {1}, \dots , - x _ {N} \}), \tag {11} +$$ + +where $a$ is a scaling coefficient, analogous to the temperature used in distillation [22], and the subscript of $n$ denotes the $n$ -th element of function's output. The approximation converges to equality as $a \to \infty$ . Thus, a differentiable expression for the approximated minimum can be + +$$ +\min (G; a) \approx \frac {\sum_ {n = 1} ^ {N} x _ {n} e ^ {- a x _ {n}}}{\sum_ {j = 1} ^ {N} e ^ {- a x _ {j}}}. \tag {12} +$$ + +Similarly, the differentiable estimate of maximum is obtained by computing $\sum_{n=1}^{N} x_n \operatorname{amax}(G)_n$ . + +Differentiable approximation to median. We begin with the fact that the median is a solution to minimizing the mean absolute deviation, i.e., + +Theorem 1. Given a set of values $G = \{x_{n}\}_{n = 1}^{N}$ , its median is a solution to minimizing the mean absolute deviation, as expressed by the equation: + +$$ +\operatorname {m e d i a n} (G) \in \arg \min _ {x} \frac {1}{N} \sum_ {n = 1} ^ {N} | x - x _ {n} |, \tag {13} +$$ + +where $\mathrm{median}(\cdot)$ denotes the median of a set. + +The proof is provided in the Supplements. Similar to the definition of amin in Eq. 8, we can define the index of the median with the assistance of Eq. 13, as follows: + +$$ +\operatorname {a m e d i a n} (G) _ {n} = \operatorname {a m i n} \left(\left\{\frac {1}{N} \sum_ {j = 1} ^ {N} | x _ {n} - x _ {j} | \right\} _ {n = 1} ^ {N}\right). \tag {14} +$$ + +Thus, we can obtain the differentiable approximation to the median as follows: + +$$ +\begin{array}{l} \operatorname {m e d i a n} (G; a) = \sum_ {n = 1} ^ {N} x _ {n} \operatorname {a m e d i a n} (G) _ {n} \tag {15} \\ \approx \frac {\sum_ {n = 1} ^ {N} x _ {n} e ^ {a \frac {1}{N} \sum_ {j = 1} ^ {N} | x _ {n} - x _ {j} |}}{\sum_ {k = 1} ^ {N} e ^ {a \frac {1}{n} \sum_ {j = 1} ^ {N} | x _ {k} - x _ {j} |}}. \\ \end{array} +$$ + +Note that the absolute value function is differentiable except when the input equals zero. It can also be approximated by $|x| \approx \operatorname{amax}(\{x, -x\}; a)$ . This approximation converges to equality as $a \to \infty$ . + +Dynamic stacking filter. To design a adaptive stacking filter that can degrade to four options of mean, median, maximum and minimum values, we introduce a gate coefficient $b$ in Eq. 15, resulting in the following expression: + +$$ +\begin{array}{l} \operatorname {D S F} (G; a, b) = \frac {\sum_ {n = 1} ^ {N} x _ {n} e ^ {a \frac {1}{n} \sum_ {j = 1} ^ {n} | x _ {n} - b x _ {j} |}}{\sum_ {k = 1} ^ {n} e ^ {a \frac {1}{n} \sum_ {j = 1} ^ {n} | x _ {k} - b x _ {j} |}} \\ = \left\{\begin{array}{l}\operatorname {m e a n} (G), \text {i f} a \rightarrow 0, b \rightarrow \infty ,\\\operatorname {m e d i a n} (G), \text {i f} a \rightarrow \infty , b \rightarrow 1,\\\max (G), \text {i f} a \rightarrow \infty , b \rightarrow 0, x _ {n} \geq 0,\\\min (G), \text {i f} a \rightarrow - \infty , b \rightarrow 0, x _ {n} \geq 0.\end{array}\right. \tag {16} \\ \end{array} +$$ + +The cases where $x_{n} < 0$ are symmetric and thus omitted for brevity. In the case of matrix inputs $[\mathbf{F}_S^i, \mathbf{F}_T^i]$ , the filter DSF is applied to each pixel, with constant coefficients becoming learnable matrix parameters $\mathbf{a}, \mathbf{b} \in \mathbb{R}^{H \times W}$ . + +# 3.3. Loss functions for Semi-Supervision + +We train the model for multi-frame optical flow estimation using the optical flow transfer loss in Eq. 5. To train VDMamba for video deraining, we employ three loss terms in a semi-supervision manner. Given the ground truth, the first + +two loss terms represent reconstruction losses in the form of the $L_{1}$ norm: + +$$ +\mathcal {L} _ {r e c} = \left\| \mathbf {B} _ {c} - \tilde {\mathbf {B}} _ {c} \right\| _ {1}, \quad \mathcal {L} _ {s p a} = \left\| \overline {{\mathbf {B}}} _ {c} - \tilde {\mathbf {B}} _ {c} \right\| _ {1}, \tag {17} +$$ + +where $\tilde{\mathbf{B}}_c$ and $\overline{\mathbf{B}}_c$ denote the current frames restored by our dual-branch model and the spatial-branch-only model, respectively. The loss term $\mathcal{L}_{spa}$ is designed to enhance the reconstruction capacity of the spatial branch. In the case of unpaired real rainy data, we construct pseudo ground-truths using a masked median stacking technique designed to remove moving objects. Specifically, we first obtain aligned neighboring frames $\tilde{\mathbf{I}}_n$ by + +$$ +\tilde {\mathbf {I}} _ {n} = \mathcal {W} \left(\mathbf {I} _ {n}; \mathbf {O} _ {n \rightarrow c}\right), \text {f o r} n = 1, \dots , N, \tag {18} +$$ + +where $\mathcal{W}(\cdot ;\cdot)$ is the backward warping function. With the aligned frame, we compute the median of the consecutive aligned frames, denoted by + +$$ +\mathbf {M} _ {c} = \operatorname {m e d i a n} \left(\left\{\mathbf {I} _ {c}, \tilde {\mathbf {I}} _ {1}, \dots \tilde {\mathbf {I}} _ {N} \right\}\right). \tag {19} +$$ + +However, directly applying the temporal median to the aligned sequence may result in mismatches caused by occlusions and mistaken removal of moving backgrounds instead of rain streaks. To mitigate this, we first crop the frames into $P^2$ sub-patches, each with width $\frac{W}{P}$ and height $\frac{H}{P}$ . The corresponding sub-patch of the derained frame, the median frame, and the original central frame are denoted as $\hat{\mathbf{B}}_{c,p}$ , $\hat{\mathbf{M}}_{c,p}$ , and $\hat{\mathbf{I}}_{c,p}$ for $p = \{1,\dots,P^2\}$ . Due to the sparsity of rain streaks, the median sub-patch $\hat{M}_{t,i}$ is adopted as a pseudo-clean image when the percentage of unchanged pixels compared to $\hat{\mathbf{I}}_{c,p}$ exceeds a threshold $\theta$ . + +$$ +\begin{array}{l} \mathcal {M} \left(\hat {\mathbf {M}} _ {c, p}, \hat {\mathbf {I}} _ {c, p}\right) \\ = \left\{ \begin{array}{l} 1, \text {i f} \frac {P ^ {2}}{W H} \sum_ {j, k} \mathbb {1} _ {\left\{\left| \hat {\mathbf {M}} _ {c, p} [ j, k ] - \hat {\mathbf {I}} _ {c, p} [ j, k ] \right| \leq \delta \right\}} \geq \theta , \\ 0, \text {o t h e r w i s e}, \end{array} \right. \tag {20} \\ \end{array} +$$ + +where $\delta$ is a slack variable that controls the tolerance for the decision condition, $\mathcal{M}$ is a masking function to select rain-free and well-aligned patches, and $\mathbb{1}_{\{\text{condition}\}}$ is an indicator function that returns 1 if the specific condition is met and 0 otherwise. The stacking loss is subsequently computed only on the masked sub-patches: + +$$ +\mathcal {L} _ {s t a} = \sum_ {p = 1} ^ {P ^ {2}} \mathcal {M} \left(\hat {\mathbf {M}} _ {c, p}, \hat {\mathbf {I}} _ {c, p}\right) \left\| \hat {\mathbf {B}} _ {c, p} - \hat {\mathbf {M}} _ {c, p} \right\| _ {1}. \tag {21} +$$ + +To further improve the temporal consistency, we incorporate a temporal loss [60], + +$$ +\mathcal {L} _ {t e m} = \frac {1}{N - 1} \sum_ {n = 1, n \neq c} ^ {N} \left\| \mathcal {W} \left(\tilde {\mathbf {B}} _ {c}; \mathbf {O} _ {c \rightarrow n}\right) - \mathbf {I} _ {n} \right\| _ {1}, \tag {22} +$$ + +where $\mathbf{O}_{c\rightarrow n} = \mathcal{F}(\tilde{B}_c,I_n)$ represents the optical flow warping central frame inversely to the neighboring frames. The total loss for training deraining is calculated by sum + +Table 1. Quantitative results of various video deraining methods on RainSynLight25 [36], RainSynComplex25 [36] and NTURain [7]. + +
DatasetMetricsRainyFastDeRain [28]SpacCNN [7]J4RNet [59]DRSformer [8]ESTINet [69]MFGAN [61]RainMamba [54]Ours
RainSynLight25PSNR25.2329.4232.7832.9636.8436.1236.9936.7437.53
SSIM0.90990.86830.92390.94340.97390.96310.97600.97410.9812
RainSynComplex25PSNR16.8519.2521.2124.1331.6128.4832.7032.6532.89
SSIM0.67910.53850.58540.71630.92580.82420.93570.93610.9425
NTURainPSNR30.4130.3033.1132.1436.9337.4838.9237.8739.74
SSIM0.91080.92740.94750.94800.95910.97000.97640.97380.9791
+ +ming the individual loss terms, + +$$ +\mathcal {L} = \left\{ \begin{array}{l l} \mathcal {L} _ {r e c} + \mathcal {L} _ {s p a} + \lambda_ {2} \mathcal {L} _ {t e m} & , \text {i f p a i r e d d a t a}, \\ \lambda_ {1} \mathcal {L} _ {s t a} + \mathcal {L} _ {s p a} + \lambda_ {2} \mathcal {L} _ {t e m} & , \text {i f u n p a i r e d d a t a}, \end{array} \right. +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ regulate the balance between the strengths of the supervision terms. + +# 3.4. Real-World Rainy Task Benchmark + +In addition to enhancing visual quality, deraining models hold practical potential for restoring degraded videos, enabling more effective performance in downstream high-level vision tasks. To this end, we introduce the real-world rainy video object detection and tracking benchmark, referred to as RVDT. Specifically, we curated 57 long video clips, totaling 10,258 frames, selected from the internet, each exhibiting noticeable rain streak degradation. We have annotated 33,077 objects for detection and 408 unique objects for tracking. Detailed statistics and examples are provided in the Supplements. + +# 4. Experiment + +# 4.1. Experimental Settings + +Training Details. We train our model for 100 epochs and 50,000 iterations. The input data are cropped into patches of size $256 \times 256$ with a sequence length of 7 frames. The initial learning rate is set to 0.0001, which is reduced by half after 50 epochs. We utilize the optimizer Adam [32] with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ . The threshold $\theta$ and the slack number $\delta$ in Eq. 20 are set to $80\%$ and 0.1, respectively. The loss weights are $\lambda_{1} = 0.1$ and $\lambda_{2} = 0.1$ . The pretrained optical flow estimator is LiteFlowNet3 [24]. All experiments are conducted on Nvidia Tesla V100 32G GPU. + +**Benchmarks.** We evaluate our model on three widely used video deraining benchmarks: RainSynLight25 [36], RainSynComplex25 [36], and NTURain [7]. To assess the potential of improving downstream high-level tasks, we also utilize our proposed RVDT benchmark. + +Baselines. We compare our method with state-of-the-art deraining methods, including FastDeRain [28], DRSformer [8], SpacCNN [7], SLDNet [60], J4RNet [59], S2VD [66], ESTINet [69], MFGAN [61], and RainMamba [54]. Among them, FastDeRain [28] is a model-based method, and DRSformer [8] is a recent model for single image. The others are data-driven video-based methods. Metrics. In addition to the commonly used SSIM and + +Table 2. The object detection and tracking results on RVDT. $\mathrm{mAP}_1$ and $\mathrm{mAP}_2$ denote the results of YOLO-v3 [39] and MEGA [9], respectively. The tracking result is obtained by GTR [72]. + +
Downstream Task MetricsObject DetectionObject Tracking
\( \mathrm{mAP}_{1} \uparrow \)\( \mathrm{mAP}_{2} \uparrow \)IDF1↑MOTA↑MOTP↑
Rainy39.7344.6964.528.019.9
FastDeRain [28]40.4542.9066.529.820.9
SLDNet [60]41.5942.0666.730.720.4
DRSformer [8]39.1044.0167.028.820.0
S2VD [66]41.3743.7368.129.420.4
ESTINet [69]40.7143.0567.629.719.9
MFGAN [61]40.8343.9767.128.620.1
RainMamba [54]40.6744.2267.229.520.3
Ours42.5146.2768.331.821.2
+ +Table 3. The quantitative results of ablation analysis. + +
Dual-branch U-Netw/o S3MLw/o TSMLOurs
PSNR38.5439.0239.2439.74
SSIM0.97330.97540.97630.9791
Strategy 1Strategy 2Strategy 3Ours
PSNR39.6539.0338.5139.74
SSIM0.97580.97540.97340.9791
w/o DSFmedianmeanOurs
PSNR39.2139.0939.3139.74
SSIM0.97550.97490.97610.9791
+ +PSNR metrics, we propose evaluating performance improvements in various object detection and trakcing tasks before and after the deraining process on selected rainy videos. Specifically, we use mAP (mean average precision) for object detection, and IDF1, MOTA, and MOTP [4] for object tracking. We utilize two detection models: the image-based YOLO-v3 [39] and the video-based MEGA [9], along with GTR [72] as the tracking method. + +# 4.2. Video Deraining Result + +Synthetic Rain Removal. We present the quantitative comparisons on three synthetic datasets in Tab. 1. Our method achieves the highest PSNR and SSIM values across all the benchmarks. For a visual quality comparison, we provide examples in Fig. 5, illustrating that only our method successfully removes all rain streaks. + +Real-World Rain Removal. We then assess the effectiveness of the existing deraining methods for real-world video deraining. A group of visual comparisons are presented in Fig. 6. As demonstrated, only our method successfully removes all rain streaks, regardless of their scale. + +Complexity Analysis We compare our VDMamba with + +![](images/a55ce7778135140afea35b277acd999d64d1a8a62a5c6ca2392bc9740a76c5a7.jpg) +Figure 5. The qualitative comparison among the existing deraining methods on synthetic datasets RainSynLight25 [36], RainSynComplex25 [36] and NTURain [7]. Please zoom in for better view. + +![](images/d54499c928ae994aa454de9ff6a55fe91f6a56f39324ffe72534b4e6fd0e018e.jpg) +Figure 6. The qualitative comparison among the existing deraining methods on real-world rainy videos from NTURain [7] and RVDT. Please zoom in for better view. + +other deraining methods in terms of time and space complexities in Fig. 2. "Ours-Single" refers to the lightweight version of our VDMamba, which includes only the spatial branch. Our method achieves the best overall balance between deraining performance and the complexities. + +# 4.3. Video Deraining for Real-World Tasks + +Rainy Object Detection. We conduct experiments of real-world rainy object detection using our RVDT. The quantitative results are presented in Tab. 2, while visual comparisons are shown in the first two examples of Fig. 7. Only our method successfully removes the rain streaks, while simultaneously improving detection accuracy. + +Rainy Object Tracking. We also evaluate the tracking performance on the videos restored by the deraining methods. The results are presented in the right half of Tab. 2. Our method demonstrates the most significant improvement compared to direct tracking on the rainy videos. Additionally, a visual comparison is provided in the final example of Fig. 7. It is evident that only our method effectively removes all rain streaks and enables the tracking method to correctly identify both the object "umbrella" and the "person". + +# 4.4. Ablation Studies + +To validate the necessity of the components we introduced, we conduct four groups of ablation studies. + +![](images/09e888cd3a019ceb5daf59d82e71a8764d6c5a3200987ecac3255dec0cc30971.jpg) +Figure 7. Rainy-scene object detection and tracking results on RVDT after video deraining. Three results are obtained using YOLOv3 [39], MEGA [9], and GTR [72], respectively. Please zoom in for better view. + +![](images/b1b741443de67a99f0f883866b65f4ea126e814f9cc31a07cec67fc51ea3589c.jpg) +(a) Rainy + +![](images/19aaec00ad1a9a65d9d828dec98d102a8a0e9790d31bee02181f557f06638bfa.jpg) + +![](images/4f50e281a2a4ad89b4cd32f9974fe112396594fd9b828462e6f6ee4b12805e0c.jpg) + +![](images/d2a0f0480693bf7b3314f69c1a4e08c272313957552d613ecc5b8c9ae26d5a6d.jpg) +(d) $\theta = 80\%$ + +![](images/0a2f09ad06156936023c96eef5183852f49d72bea2e4e4c28dec68bd7488f3c0.jpg) +(b) $\theta = 40\%$ +(e) $\theta = 100\%$ + +![](images/85d7c3e84f5f03ccd0d4d50f7e79454cee463139ad83cfcf1840a3a852c33e43.jpg) +(c) $\theta = 60\%$ +(f) No stacking loss +Figure 8. The qualitative ablation of the stacking loss. + +Network Structure. We design three additional model variants: (1) replacing both S3ML and TSML with residual blocks to form a pure dual-branch U-Net [41], (2) replacing S3ML alone, and (3) replacing TSML alone. The results in the first row of Tab. 3 shows that both S3ML and TSML contribute to improving the deraining performance. + +Optical Flow Estimation. We design three optical flow estimation strategies. Strategy 1 directly uses the pre-trained model to estimate flows for multiple frame pairs. Strategy 2 and 3 involve conducting the transfer on rainy frames and clean frames, respectively. The results are presented in the second row of Tab. 3. Though Strategy 1 achieves a comparable result, its speed is significantly slower. + +Dynamic Stacking Filter. Apart from the proposed dynamic stacking filter, we evaluate three additional fusion strategies: removing the filter, using the median filter, and using the mean filter. The results are shown in the last row of Tab. 3. Our method demonstrates a performance improvement over the other strategies. + +Threshold of Stacking Loss. We experiment with different values for the threshold $\theta$ to generate the mask index for learning sub-patches in Eq. 20. The resulting deraining examples are shown in Fig. 8. We observe that the introduction of the stacking loss enhances the model's capability for real-world deraining. However, a too slack threshold, such as $\theta = 40\%$ , causes the network to learn from many mismatched sub-patches, resulting in blurry outputs. As a threshold range from $60\%$ to $90\%$ provides satisfactory results, we select $80\%$ as the default value. + +# 5. Conclusion + +In this work, we propose a dual-branch spatio-temporal state-space model for video restoration in rainy conditions, addressing the gap between synthetic and real-world rain effects. By integrating spatial and temporal feature extraction, dynamic stacking filters, and a median stacking loss, our method effectively enhances rain streak removal. Extensive evaluations on synthetic and real-world rainy videos demonstrate its superior performance in visual quality, efficiency, and our proposed downstream vision benchmarks. + +# Acknowledgment + +This work was supported in part by the National Key R&D Program of China (Grant No.2022ZD0119200), in part by National Natural Science Foundation of China (No. 62322216, 62025604, 62172409), and in part by Shenzhen Science and Technology Program (Grant No. JCYJ20220818102012025, KQTD20221101093559018, RCYX20221008092849068). + +# References + +[1] Neena Aloysius and M Geetha. A review on deep convolutional neural networks. In ICCSP, 2017. 4 +[2] Peter Barnum, Takeo Kanade, and Srinivasa G Narasimhan. Spatio-temporal frequency analysis for removing rain and snow from videos. In PACV, 2007. 2 +[3] Peter Barnum, Srinivasa Narasimhan, and Takeo Kanade. Analysis of rain and snow in frequency space. IJCV, 2010. 2 +[4] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. JIVP, 2008. 6 +[5] J. Bossu, N. Hautiere, and J.-P. Tarel. Rain or snow detection in image sequences through use of a histogram of orientation of streaks. *IJCV*, 2011. 2 +[6] Duan-Yu Chen, Chien-Cheng Chen, and Li-Wei Kang. Visual depth guided color image rain streaks removal using sparse coding. IEEE TCSVT, 2014. 2 +[7] Jie Chen, Cheen-Hau Tan, Junhui Hou, Lap-Pui Chau, and He Li. Robust video content alignment and compensation for rain removal in a cnn framework. In CVPR, 2018. 1, 3, 6, 7 +[8] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In CVPR, 2023. 1, 2, 6, 7, 8 +[9] Yihong Chen, Yue Cao, Han Hu, and Liwei Wang. Memory enhanced global-local aggregation for video object detection. In CVPR, 2020. 6, 8 +[10] Yi-Lei Chen and Chiou-Ting Hsu. A generalized low-rank appearance model for spatio-temporally correlated rain streaks. In ICCV, 2013. 2 +[11] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural Netw., 2018. 4 +[12] Xueyang Fu, Jiabin Huang, Xinghao Ding, Yinghao Liao, and John Paisley. Clearing the skies: A deep network architecture for single-image rain removal. IEEE TIP, 2017. 2 +[13] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In CVPR, 2017. 2 +[14] K. Garg and S. Nayar. Detection and removal of rain from videos. In CVPR, 2004. 2 +[15] Kshitiz Garg and Shree K. Nayar. When does a camera see rain? In ICCV, 2005. +[16] Kshitiz Garg and Shree K. Nayar. Vision and rain. IJCV, 2007. 2 + +[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3 +[18] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3 +[19] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024.3 +[20] Mahmoud Hassaballah, Mourad A Kenk, Khan Muhammad, and Shervin Minaee. Vehicle detection and tracking in adverse weather using a deep learning framework. IEEE TITS, 2020. 1 +[21] Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024. 3 +[22] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531, 2015. 3, 4 +[23] Mazin Hnewa and Hayden Radha. Object detection under rainy conditions for autonomous vehicles: A review of state-of-the-art and emerging techniques. IEEE SPM, 2020. 1 +[24] Tak-Wai Hui and Chen Change Loy. Liteflownet3: Resolving correspondence ambiguity for more accurate optical flow estimation. In ECCV, 2020. 6 +[25] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In CVPR, 2020. 2 +[26] Kui Jiang, Wenxuan Liu, Zheng Wang, Xian Zhong, Junjun Jiang, and Chia-Wen Lin. Dawn: Direction-aware attention wavelet network for image deraining. In ACM MM, 2023. 2 +[27] Tai-Xiang Jiang, Ting-Zhu Huang, Xi-Le Zhao, Liang-Jian Deng, and Yao Wang. A novel tensor-based video rain streaks removal approach via utilizing discriminatively intrinsic priors. In CVPR, 2017. 1, 2 +[28] Tai-Xiang Jiang, Ting-Zhu Huang, Xi-Le Zhao, Liang-Jian Deng, and Yao Wang. Fastderain: A novel video rain streak removal method using directional gradient priors. IEEE TIP, 2019. 1, 2, 6, 7, 8 +[29] Li-Wei Kang, Chia-Wen Lin, and Yu-Hsiang Fu. Automatic single-image-based rain streaks removal via image decomposition. IEEE TIP, 2012. 2 +[30] Jin-Hwan Kim, Chul Lee, Jae-Young Sim, and Chang-Su Kim. Single-image deraining using an adaptive nonlocal means filter. In ICIP, 2013. 2 +[31] Jin-Hwan Kim, Jae-Young Sim, and Chang-Su Kim. Video deraining and desnowing using temporal correlation and low-rank matrix completion. IEEE TIP, 2015. 2 +[32] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6 +[33] Dong Li, Yidi Liu, Xueyang Fu, Senyan Xu, and Zheng-Jun Zha. Fourier mamba: Fourier learning integration with state space models for image deraining. arXiv preprint arXiv:2405.19450, 2024. 3 + +[34] Minghan Li, Qi Xie, Qian Zhao, Wei Wei, Shuhang Gu, Jing Tao, and Deyu Meng. Video rain streak removal by multiscale convolutional sparse coding. In CVPR, 2018. 2 +[35] Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, et al. Jamba: A hybrid transformer-mamba language model. arXiv preprint arXiv:2403.19887, 2024. 3 +[36] Jiaying Liu, Wenhan Yang, Shuai Yang, and Zongming Guo. Erase or fill? deep joint recurrent rain removal and reconstruction in videos. In CVPR, 2018. 3, 6, 7 +[37] Jiaying Liu, Wenhan Yang, Shuai Yang, and Zongming Guo. D3r-net: Dynamic routing residue recurrent network for video rain removal. IEEE TIP, 2019. 3 +[38] Yu Luo, Yong Xu, and Hui Ji. Removing rain from a single image via discriminative sparse coding. In ICCV, 2015. 2 +[39] Joseph Redmon. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 6, 8 +[40] Weihong Ren, Jiandong Tian, Zhi Han, Antoni Chan, and Yandong Tang. Video desnowing and deraining based on matrix decomposition. In CVPR, 2017. 2 +[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3, 8 +[42] Varun Santhaseelan and Vijayan K. Asari. Utilizing local phase information to remove rain from video. IJCV, 2015. 2 +[43] Vishwanath A Sindagi, Poojan Oza, Rajeev Yasarla, and Vishal M Patel. Prior-based domain adaptive object detection for hazy and rainy conditions. In ECCV, 2020. 1 +[44] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 3 +[45] Shangquan Sun, Wenqi Ren, Tao Wang, and Xiaochun Cao. Rethinking image restoration for object detection. NeurIPS, 2022. 1 +[46] Shangquan Sun, Wenqi Ren, Jingzhi Li, Kaihao Zhang, Meiyu Liang, and Xiaochun Cao. Event-aware video deraining via multi-patch progressive learning. IEEE TIP, 2023. 1, 2, 3 +[47] Shangquan Sun, Wenqi Ren, Xinwei Gao, Rui Wang, and Xiaochun Cao. Restoring images in adverse weather conditions via histogram transformer. In ECCV, 2024. 2 +[48] Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, and Xiaochun Cao. Logit standardization in knowledge distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15731-15740, 2024. 3 +[49] Shangquan Sun, Wenqi Ren, Juxiang Zhou, Jianhou Gan, Rui Wang, and Xiaochun Cao. A hybrid transformer-mamba network for single image deraining. arXiv preprint arXiv:2409.00410, 2024. 3 +[50] Chaoyue Wang, Chang Xu, Chaohui Wang, and Dacheng Tao. Perceptual adversarial networks for image-to-image transformation. IEEE TIP, 2018. 2 +[51] Hong Wang, Qi Xie, Qian Zhao, and Deyu Meng. A model-driven deep neural network for single image rain removal. In CVPR, 2020. 2 + +[52] Wei Wei, Lixuan Yi, Qi Xie, Qian Zhao, Deyu Meng, and Zongben Xu. Should we encode rain streaks in video as deterministic or stochastic? In ICCV, 2017. 1, 2 +[53] Wei Wei, Deyu Meng, Qian Zhao, Zongben Xu, and Ying Wu. Semi-supervised transfer learning for image rain removal. In CVPR, 2019. 2 +[54] Hongtao Wu, Yijun Yang, Huihui Xu, Weiming Wang, Jinni Zhou, and Lei Zhu. Rainmamba: Enhanced locality learning with state space models for video deraining. In ACM MM, 2024. 1, 2, 3, 6, 7, 8 +[55] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 3 +[56] Shugo Yamashita and Masaaki Ikehara. Image deraining with frequency-enhanced state space model. arXiv preprint arXiv:2405.16470, 2024. 3 +[57] Wending Yan, Robby T. Tan, Wenhan Yang, and Dengxin Dai. Self-aligned video deraining with transmission-depth consistency. In CVPR, 2021. 3 +[58] Wenhan Yang, Jiaying Liu, and Jiashi Feng. Frameconsistent recurrent video deraining with dual-level flow. In CVPR, 2019. 3 +[59] Wenhan Yang, Robby T. Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE TPAMI, 2020. 6 +[60] Wenhan Yang, Robby T. Tan, Shiqi Wang, and Jiaying Liu. Self-learning video rain streak removal: When cyclic consistency meets temporal correspondence. In CVPR, 2020. 1, 3, 5, 6 +[61] Wenhan Yang, Robby T. Tan, Jiashi Feng, Shiqi Wang, Bin Cheng, and Jiaying Liu. Recurrent multi-frame deraining: Combining physics guidance and adversarial learning. IEEE TPAMI, 2022. 3, 6, 7, 8 +[62] Wenhan Yang, Robby T. Tan, Shiqi Wang, Alex C. Kot, and Jiaying Liu. Learning to remove rain in video with self-supervision. IEEE TPAMI, 2022. 3 +[63] Rajeev Yasarla and Vishal M. Patel. Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. In CVPR, 2019. 2 +[64] Shaodi You, Robby T. Tan, Rei Kawakami, Yasuhiro Mukaigawa, and Katsushi Ikeuchi. Adherent raindrop modeling, detection and removal in video. IEEE TPAMI, 2016. 2 +[65] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024.3 +[66] Zongsheng Yue, Jianwen Xie, Qian Zhao, and Deyu Meng. Semi-supervised video deraining with dynamical rain generator. In CVPR, 2021. 1, 2, 3, 6 +[67] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2 +[68] He Zhang, Vishwanath Sindagi, and Vishal M. Patel. Image de-raining using a conditional generative adversarial network. IEEE TCSVT, 2020. 2 + +[69] Kaihao Zhang, Dongxu Li, Wenhan Luo, Wenqi Ren, and Wei Liu. Enhanced spatio-temporal interaction learning for video deraining: A faster and better framework. IEEE TPAMI, 2022, 3, 6, 7, 8 +[70] Xiaopeng Zhang, Hao Li, Yingyi Qi, Wee Kheng Leow, and Teck Khim Ng. Rain removal in video by combining temporal and chromatic properties. In ICME, 2006. 2 +[71] Zou Zhen, Yu Hu, and Zhao Feng. Freqmamba: Viewing mamba from a frequency perspective for image deraining. arXiv preprint arXiv:2404.09476, 2024. 3 +[72] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In CVPR, 2022. 6, 8 +[73] 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 +[74] Zhen Zou, Hu Yu, Jie Huang, and Feng Zhao. Freqmamba: Viewing mamba from a frequency perspective for image deraining. In ACM MM, 2024. 2 \ No newline at end of file diff --git a/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/images.zip b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f77f8dcc628bcbbd701a2c2c082349e18696f9f3 --- /dev/null +++ b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b0247eb72041a8da57ae7057f8f295578633369ec926a61d361e6bae1053f42 +size 1170790 diff --git a/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/layout.json b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..64558b81ac5658b730d5e84c461043d82bfd22df --- /dev/null +++ b/CVPR/2025/Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49edf39fc30cd4b1be002f8681b8f6e5f300ad8396d3ae658068552e39150e5a +size 518223 diff --git a/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/bef13b7f-93f6-4bb3-a82f-a8785de2a581_content_list.json b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/bef13b7f-93f6-4bb3-a82f-a8785de2a581_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8129e007c313833ea909cbc5867abc4f954441c7 --- /dev/null +++ b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/bef13b7f-93f6-4bb3-a82f-a8785de2a581_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:703af2eaf9be03dbfe649d55a41ca974a3a7fefcdab414e2f27d8cd018dcea5c +size 110884 diff --git a/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/bef13b7f-93f6-4bb3-a82f-a8785de2a581_model.json b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/bef13b7f-93f6-4bb3-a82f-a8785de2a581_model.json new file mode 100644 index 0000000000000000000000000000000000000000..637cf58044478c8e29925d63aa38bc0c5f8df044 --- /dev/null +++ b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/bef13b7f-93f6-4bb3-a82f-a8785de2a581_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9a62f2d6ea884864c48695f52ad866164b624aec02965d36d280458c26b2844 +size 140101 diff --git a/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/bef13b7f-93f6-4bb3-a82f-a8785de2a581_origin.pdf b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/bef13b7f-93f6-4bb3-a82f-a8785de2a581_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7f728b4467f585633315e878a6433c8d66aecfa5 --- /dev/null +++ b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/bef13b7f-93f6-4bb3-a82f-a8785de2a581_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82fc75209bc4ffbd3be16aefaf76b0d6d0f32743d26457fcd565ade6ef5fb6b3 +size 3694876 diff --git a/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/full.md b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..645b42e54f0faf94291f6f6529d548776a7c2a93 --- /dev/null +++ b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/full.md @@ -0,0 +1,349 @@ +# SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation + +Hritam Basak*, Zhaozheng Yin + +Dept. of Computer Science, Stony Brook University, NY, USA + +$^{*}$ hbasak@cs.stonybrook.edu + +# Abstract + +Domain Adaptation (DA) and Semi-supervised Learning (SSL) converge in Semi-supervised Domain Adaptation (SSDA), where the objective is to transfer knowledge from a source domain to a target domain using a combination of limited labeled target samples and abundant unlabeled target data. Although intuitive, a simple amalgamation of DA and SSL is suboptimal in semantic segmentation due to two major reasons: (1) previous methods, while able to learn good segmentation boundaries, are prone to confuse classes with similar visual appearance due to limited supervision; and (2) skewed and imbalanced training data distribution preferring source representation learning whereas impeding from exploring limited information about tailed classes. Language guidance can serve as a pivotal semantic bridge, facilitating robust class discrimination and mitigating visual ambiguities by leveraging the rich semantic relationships encoded in pre-trained language models to enhance feature representations across domains. Therefore, we propose the first language-guided SSDA setting for semantic segmentation in this work. Specifically, we harness the semantic generalization capabilities inherent in vision-language models (VLMs) to establish a synergistic framework within the SSDA paradigm. To address the inherent class-imbalance challenges in long-tailed distributions, we introduce class-balanced segmentation loss formulations that effectively regularize the learning process. Through extensive experimentation across diverse domain adaptation scenarios, our approach demonstrates substantial performance improvements over contemporary state-of-the-art (SoTA) methodologies. + +# 1. Introduction + +The remarkable progress in deep learning has significantly enhanced the performance of visual understanding tasks, including image classification [7], object detection [103], and, more recently, semantic segmentation [61, 71]. These advancements have been particularly notable when a wealth of labeled training data is available. However, as noted in + +![](images/a68d12ce48676792a55962d9bd1777b44e7bef0258e7868736097a1cd09bf9e3.jpg) + +![](images/0c60000710665c5ade67b1175c484c79f091d813fbdea046d04950228169f7fa.jpg) + +![](images/278d14df107ae17af43bb51c821038b75e97601e4286855a939166ebb56c0876.jpg) +Figure 1. Major contributions of SemiDAVil: (1) We propose the first language-guided SSDA framework for semantic segmentation, (2) Utilizing spatial context via dense language guidance (DLG) improves segmentation performance, (3) Our proposed DyCE loss dynamically reweights imbalanced class distributions, resulting in precise segmentation of minority classes. + +![](images/4333d7219284cc8a23620a34f4128cb6197766765bba3acea90b2f3bdabb4426.jpg) + +[68], their performance degrades precipitously when confronted with annotation-scarce environments, especially in the context of semantic segmentation, where dense pixelwise annotations are essential. Furthermore, these sophisticated models exhibit substantial vulnerability when tasked with generalizing across domains characterized by significant distributional shifts [30, 36] - a challenge particularly evident in real-world applications where models trained on synthetic data must maintain robust performance in naturalistic settings, such as autonomous navigation systems [19, 66]. This inherent limitation in cross-domain generalization has catalyzed the emergence of two pivotal research paradigms: Domain Adaptation (DA) and Semi-supervised Learning (SSL). + +The confluence of DA and SSL has given rise to Semi-supervised Domain Adaptation (SSDA), a hybrid approach that strategically leverages three distinct data streams: com + +prehensively labeled source domain data, sparsely labeled target domain samples, and a wealth of unlabeled target domain instances [42, 65]. While SSDA holds intuitive appeal for real-world applications, existing methods encounter critical limitations when applied to semantic segmentation tasks. Specifically, (1) despite achieving accurate segmentation boundaries, current approaches [52, 92] often suffer from misclassification among visually similar classes, due to restricted supervision within the target domain; (2) the SSDA framework tends to over-prioritize source domain features, driven by abundant source labels, while generating error-prone pseudo-labels for target data, which hampers adaptation performance; (3) class-imbalance, a common issue in real-world datasets, exacerbates these challenges, limiting effective exploration and representation of minority (tail) classes in the target domain. + +To address the identified SSDA challenges, we augment the SSDA paradigm with vision-language (VL) guidance using VLMs (e.g., CLIP [60]) to enrich semantic representation, leveraging their large-scale image-caption pretraining. By incorporating VLM features into a global-local context exploration module, we mitigate misclassification among visually similar classes. To tackle the over-reliance on source features, we introduce a joint embedding space guided by language priors, enhancing instance separability and reducing domain bias, unlike traditional divergence-based alignment methods [41, 94]. Finally, to combat class imbalance, we design a tailored cross-entropy loss that dynamically reweighs minority classes, thereby facilitating more equitable exploration and representation of tail classes in the target domain. Specifically, our contributions can be summarized as: + +1. Language-Guided SSDA Framework: We pioneer the first language-guided SSDA framework for semantic segmentation by harnessing the rich semantic knowledge encoded in pre-trained Vision-Language Models (VLMs). Our novel attention-based fusion mechanism seamlessly integrates visual features with dense language embeddings, establishing a robust semantic bridge between source-target domains while providing enhanced contextual understanding. +2. Enhanced Feature Localization: Recognizing that VL pre-training primarily operates at the image level, we address the critical challenge of feature localization in semantic segmentation through targeted fine-tuning. To mitigate the risks of overfitting and semantic knowledge degradation inherent in limited-annotation scenarios, we develop a sophisticated consistency regularization framework that preserves the rich semantic representations acquired during pre-training. +3. Adaptive Class-Balanced Loss: To tackle class imbalance in a limited annotation scenario, we introduce a Dynamic Cross-Entropy (DyCE) loss formulation that + +dynamically calibrates the learning emphasis on tail classes. This innovative, plug-and-play loss mechanism demonstrates broad applicability across various class-imbalanced learning scenarios. + +4. State-of-the-Art Performance: Through detailed evaluation across diverse domain-adaptive and class-imbalanced segmentation benchmarks, our methodology demonstrates superior performance and robustness, consistently surpassing contemporary state-of-the-art approaches by significant margins. + +# 2. Related Works + +# 2.1. Semi-supervised Domain Adaptation + +Recent advances in Semi-Supervised Domain Adaptation (SSDA) for semantic segmentation have focused on utilizing limited labeled target data and abundant unlabeled data to bridge the domain gap at the pixel level [1, 2]. Early approaches like MME [65] and ASDA [58] used entropy minimization for feature alignment, but their classification-centric strategies struggled with fine-grained segmentation tasks, leading to suboptimal boundary delineation. To address this, SSL-based methods such as DECOTA [91] and SS-ADA [89] employed teacher-student frameworks with consistency constraints, generating pseudo-labels for unlabeled target data. However, these methods faced issues with noisy pseudo-labels, particularly for minority and boundary classes. More recent methods have explored novel directions: S-Depth [32] leverages self-supervised depth estimation as an auxiliary task to enhance feature learning, while DSTC [24] introduces a domain-specific teacher-student framework that dynamically adapts to target domain characteristics. IIDM [23] proposes an innovative inter-intradomain mixing strategy to address domain shift and limited supervision simultaneously. However, these methods often struggle with two critical limitations: class confusion due to limited supervision and skewed data distribution favoring source domain representations. + +# 2.2. Vision Language Model + +Vision-Language Models (VLMs), like CLIP and its extensions [34, 50, 60, 99], leverage large-scale image-text pre-training for semantic segmentation via a shared embedding space that aligns visual and textual features. Initial zero-shot methods, such as MaskCLIP [98] and GroupViT [87], struggled with boundary precision due to reliance on high-level features. Later, fine-tuned models like OpenSeg [25] and LSeg [40] improved segmentation accuracy using labeled data and text embeddings. Techniques such as ZegFormer [17] and OVSeg [46] utilize frozen CLIP features for mask proposal classification, while ZegCLIP [100] aligns dense visual-text embeddings in a streamlined manner. Recently, SemiVL [33] has shown that language cues + +can enhance semantic insights and mitigate class confusion; however, their application in domain adaptation remains under-explored. The recent LIDAPS model [53] and follow-up works [20, 37, 84] apply language guidance for domain bridging in panoptic segmentation but rely on manual thresholding for pseudo-mask filtering and a complex multi-stage training process. Despite their improvements, these methods still misclassify tail classes (e.g., fence, bike, wall), posing critical risks for applications like autonomous driving, where errors in identifying such objects can lead to severe consequences. + +# 2.3. Class-Imbalance Handling + +Class imbalance significantly hinders real-world semantic segmentation, as small object classes often appear less frequently and cover fewer pixels than dominant background classes, unlike balanced datasets like CIFAR-10/100, ImageNet, and Caltech-101/256 [15, 26, 38]. Data-level methods like oversampling/undersampling adjust sampling probabilities for minority classes but struggle in dense tasks due to uneven class distribution [64, 101]. Algorithmic strategies such as class-weighted losses address bias by penalizing rare classes more [47], but treating small object classes equally often leads to instability [69, 88]. + +In unsupervised domain adaptation (UDA) for segmentation, common strategies include data-level adjustments using source domain frequencies [29-31], and adaptive weighting based on target statistics [88], but these are computationally costly for dense predictions. Approaches that relax pseudo-label filtering for rare classes still inherit source biases, causing misclassifications [102]. Most UDA methods prioritize data-level sampling [56], overlooking the synergy of combining data and algorithmic approaches, which remains impractical and lacks generalizability for diverse tasks [68, 76, 82]. + +# 3. Proposed Method + +In our SSDA setting, we utilize image-label pair from the source domain $\mathcal{D}^{Sr}$ : $\{(x_i^{Sr},y_i^{Sr})\}_{i = 1}^{\mathbb{N}^{Sr}}$ , a limited set of labeled target samples $\{(x_{i}^{Tr_{\mathcal{L}}},y_{i}^{Tr_{\mathcal{L}}})\}_{i = 1}^{\mathbb{N}^{Tr_{\mathcal{L}}}}$ , and a large pool of unlabeled target data $\{(x_i^{Tr_{\mathcal{U}}})\}_{i = 1}^{\mathbb{N}^{Tr_{\mathcal{U}}}}$ , where $\mathbb{N}^{Tr_{\mathcal{U}}}\gg \mathbb{N}^{Tr_{\mathcal{L}}}$ . Our proposed SemiDAVI-L framework effectively tackles the challenges of SSDA by leveraging VL-pretrained encoders (subsection 3.1) for enriched semantic representation learning from $S_r\cup Tr_{\mathcal{L}}\cup Tr_{\mathcal{U}}$ , addressing the issue of misclassification among visually similar classes. We incorporate dense semantic guidance from language embeddings (subsection 3.2) to enhance instance separability and reduce domain bias. Consistency-regularized SSL (subsection 3.3) mitigates over-reliance on source features, while the class-balancing DyCE loss (subsection 3.4) combats class imbalance by reweighting tail classes. The + +![](images/154cbd8a81a2f06f6459f9b0a21f980966fd27f823573c54bafb8a5c3ccf778c.jpg) +Figure 2. Overview of SemiDAViL: We leverage Vision-Language (VL) Pre-training (top) to initialize the language encoder $\mathcal{E}_{\mathcal{L}}$ and vision encoders $\mathcal{E}_{\mathcal{V}}^{\{S, T\}}$ in a semi-supervised setting (bottom), where $S$ and $T$ denote the student and teacher branches, respectively. To bridge image-level VL features for dense pixel-level tasks, we utilize a captioning model $\mathcal{C}$ to generate text descriptions of images and a Dense Language Guidance (DLG) module. The framework is trained with a supervised loss $\mathcal{L}_{DyCE}$ for labeled data and a consistency loss $\mathcal{L}_{CT}$ for unlabeled data. + +overall architecture of SemiDAViL is outlined in Figure 2. + +# 3.1. Vision-Language Pre-training + +Previous regularization-based SSDA methods have shown effectiveness in semi-supervised semantic segmentation by enforcing stable predictions on unlabeled data. However, as discussed in section 2, they often struggle with distinguishing visually similar classes, especially when only a limited set of labeled target samples $\{(x_{i}^{\mathcal{T}_{\mathcal{C}}},y_{i}^{\mathcal{T}_{\mathcal{C}}})\}$ is available. The primary issue arises due to the lack of diverse semantic coverage, leading to errors in class discrimination. To address this, we leverage Vision-Language Models (VLMs) like CLIP [60], which are trained on large-scale image-text datasets, $\mathcal{D}_{\mathrm{clip}} = \{(x,t)\}$ , where $x$ and $t$ are images and their associated captions. CLIP consists of a vision encoder $\mathcal{E}_{\mathcal{V}}$ and a language encoder $\mathcal{E}_{\mathcal{L}}$ , optimized jointly using a contrastive loss: + +$$ +\mathcal {L} _ {\text {c o n t r a s t}} = - \sum_ {i = 1} ^ {N} \log \frac {\exp \left(\left\langle \mathcal {E} _ {\mathcal {V}} \left(x _ {i}\right) , \mathcal {E} _ {\mathcal {L}} \left(t _ {i}\right) \right\rangle / \tau\right)}{\sum_ {j = 1} ^ {N} \exp \left(\left\langle \mathcal {E} _ {\mathcal {V}} \left(x _ {i}\right) , \mathcal {E} _ {\mathcal {L}} \left(t _ {j}\right) \right\rangle / \tau\right)}, \tag {1} +$$ + +where $\langle \cdot, \cdot \rangle$ denotes the cosine similarity and $\tau$ is a temperature parameter. This objective aligns the visual and textual + +![](images/53b36cde9af505f550086bf6da5c978006f328e5a55422b1968aba80a394ee90.jpg) +Figure 3. Overall architecture of our proposed DLG module: it is based on dense similarity maps of the vision and text embeddings. More details are provided in subsection 3.2. + +embeddings into a shared semantic space, learning robust, class-agnostic representations that generalize across diverse classes. Unlike conventional ImageNet pre-training that relies on manually annotated labels, CLIP's training with web-crawled image-caption pairs allows it to capture richer semantics without restricting to a fixed set of categories. + +To mitigate the limited semantic knowledge in standard consistency training in our SSDA framework, we initialize our (student-teacher) segmentation encoders $\mathcal{E}_{\mathcal{V}}^{\{S,\mathcal{T}\}}$ with CLIP's pre-trained vision encoder $\mathcal{E}_{\mathcal{V}}$ , rather than using an ImageNet-trained backbone. This transfer of rich semantic priors enables enhanced feature extraction and better semantic differentiation (as found in Table 3 and well supported in [33]), particularly for visually ambiguous classes, leading to more robust segmentation performance. + +# 3.2. Dense Language Guidance (DLG) + +Most prior VLM methods employ a standard attention mechanism for multi-modal feature integration [16, 49, 96], i.e., features from two modalities (query and key) generate an attention matrix to aggregate vision features based on language-derived weights. However, this approach only utilizes the language feature to compute attention scores, without directly incorporating it into the fused output, effectively treating the result as a reorganized single-modal vision feature. Consequently, the output vision feature dominates the decoder, leading to a substantial loss of language information. Based on our empirical findings (provided in supplementary file, and well supported in [49]), we argue that while generic attention effectively processes value inputs, it fails to fully exploit query features for deep cross-modal interaction, resulting in insufficient fusion of vision and language modalities. + +To address this, we utilize Dense Language Guidance (DLG) that transforms both the vision and language features into key-query pairs and treats them equally, as shown in Figure 3. First, visual features $\mathcal{F}_{\mathcal{V}} \in \mathbb{R}^{h \times w \times c}$ with $h \times w$ dimension and $c$ channels for image $\mathcal{X}$ are extracted through $\mathcal{E}_{\mathcal{V}}^1$ with frozen weight $\phi_{\mathcal{V}}: \mathcal{F}_{\mathcal{V}} \gets \mathcal{E}_{\mathcal{V}}(\mathcal{X}; \phi_{\mathcal{V}}^*)$ . To further utilize language embeddings $\mathcal{F}_{\mathcal{L}} \in \mathbb{R}^{n_{\mathcal{L}} \times c}$ , we extract text description with $n_{\mathcal{L}}$ tokens for $\mathcal{X}$ using off-the-shelf captioning model $\mathcal{C}$ , followed by CLIP-initialized language encoder $\mathcal{E}_{\mathcal{L}}$ with frozen weights $\phi_{\mathcal{L}}: \mathcal{F}_{\mathcal{L}} \gets \mathcal{E}_{\mathcal{L}}(\mathcal{C}(\mathcal{X}); \phi_{\mathcal{L}}^*)$ . This is followed by projection of $\mathcal{F}_{\{\nu, \mathcal{L}\}}$ to key-value pairs using linear layers: $\mathcal{F}_{\{\nu, \mathcal{L}\}}^{K, V} \gets \text{Linear}(\mathcal{F}_{\{\nu, \mathcal{L}\}})$ . Next, multi-modal key values are used to generate an attention matrix $\mathcal{A} \in \mathbb{R}^{n_{\mathcal{L}} \times h \times w}$ : + +$$ +\mathcal {A} = \frac {1}{\sqrt {c}} \mathcal {F} _ {\mathcal {V}} ^ {K} \cdot \left(\mathcal {F} _ {\mathcal {L}} ^ {K}\right) ^ {T} \tag {2} +$$ + +Instead of applying attention to a single modality as in conventional methods, we normalize across both dimensions and compute cross-attention on vision and language features. Specifically, we employ a SoftMax activation followed by attention over $\mathcal{F}_{\mathcal{V}}^{V}$ and $\mathcal{F}_{\mathcal{L}}^{V}$ to generate language-attended vision features and vision-attended language features, respectively, ensuring balanced and comprehensive feature fusion: + +$$ +\begin{array}{l} \mathcal {F} _ {\mathcal {V}} ^ {A} = \text {S o f t M a x} [ \mathcal {A} ] \mathcal {F} _ {\mathcal {V}} ^ {V} \tag {3} \\ \mathcal {F} _ {\mathcal {L}} ^ {A} = \operatorname {S o f t M a x} [ \mathcal {A} ] \mathcal {F} _ {\mathcal {L}} ^ {V} \\ \end{array} +$$ + +Finally, these two attended feature maps are combined to generate a true multimodal feature representation $\mathcal{F}_{\mathcal{M}} \in \mathbb{R}^{n_{\mathcal{L}} \times h \times w}$ : $\mathcal{F}_{\mathcal{M}} = \mathcal{F}_{\mathcal{V}}^{\mathcal{A}} \cdot \left(\mathcal{F}_{\mathcal{L}}^{\mathcal{A}}\right)^T$ , which is the basis of our VL-guided SSDA pipeline. This is thereafter passed to the student-teacher decoders $\mathcal{D}_{\mathcal{V}}^{\{S, \mathcal{T}\}}$ for consistency training. + +# 3.3. Consistency Training (CT) + +To effectively utilize labeled and unlabeled data in our SSDA setting, we utilize a student-teacher network [70] for consistency training. Specifically, for unlabeled target data $\{(x_i^{T\mathcal{r}\mathcal{U}})\}_{i = 1}^{\mathbb{N}^{T\mathcal{r}\mathcal{U}}}\in \mathcal{D}^{T\mathcal{r}\mathcal{U}}$ , we obtain two multimodal features $\mathcal{F}_M^S,\mathcal{F}_M^T$ (from DLG), and pass them through two identical but differently initialized decoders $\{\mathcal{D}_{\nu}^{S},\mathcal{D}_{\nu}^{T}\}$ with trainable parameters $\{\theta_{\nu}^{S},\theta_{\nu}^{T}\}$ , respectively, and enforce their predictions to be consistent: + +$$ +L _ {C T} = \frac {1}{\mathbb {N} ^ {T r _ {\mathcal {U}}}} \sum_ {x _ {i} \in \mathcal {D} ^ {\mathcal {T} _ {\mathcal {U}}}} \sum_ {p = 1} ^ {h \times w} \underset {m a x \left(y _ {p} ^ {S}\right) \geq T h} {\mathbb {1}} C E \left(y _ {p} ^ {S}, y _ {p} ^ {\mathcal {T}}\right) \tag {4} +$$ + +where $y_{p}^{\mathcal{S}}$ and $y_{p}^{\mathcal{T}}$ are the $p^{\mathrm{th}}$ pixel prediction from student and teacher model: $y^{m}\gets \mathcal{D}_{\mathcal{V}}^{m}(\mathcal{F}_{\mathcal{M}}^{m};\theta_{\mathcal{V}}^{m});m = \{\mathcal{S},\mathcal{T}\}$ + +$Th$ is a threshold to exclude noisy pseudo-labels in $L_{CT}$ . To further utilize labeled target data $\{\mathcal{D}^{Sr}\cup \mathcal{D}^{Tr_{\mathcal{L}}}\}$ , we can employ a supervised CE loss between ground truth $y$ and student prediction $y^S$ : + +$$ +L _ {S} = \frac {1}{\mathbb {N} ^ {\mathcal {S} r} + \mathbb {N} ^ {T r _ {\mathcal {L}}}} \sum_ {x _ {i} \in \{\mathcal {D} ^ {\mathcal {S} r} \cup \mathcal {D} ^ {T r _ {\mathcal {L}}} \}} \sum_ {c = 0} ^ {N _ {C}} C E \left(y _ {c} ^ {\mathcal {S}}, y _ {c}\right), (5) +$$ + +where $N_{C}$ represents the number of classes. However, $L_{S}$ might incur suboptimal performance due to inherent class imbalance, as discussed in subsection 2.3 and evident in previous methods in Table 4, Table 5. We propose a dynamic CE loss to alleviate this shortcoming, as detailed in subsection 3.4. The student branch is updated using a combined consistency and DyCE loss, whereas the teacher model is updated using an exponential moving average (EMA) of the student parameters: + +$$ +\theta_ {\nu} ^ {\mathcal {T}} (t) \leftarrow \alpha \theta_ {\nu} ^ {\mathcal {T}} (t - 1) + (1 - \alpha) \theta_ {\nu} ^ {\mathcal {S}} (t) \qquad (6) +$$ + +where $t$ is step number, $\alpha$ is the momentum coefficient [28]. + +# 3.4. Class-balanced Dynamic CE Loss (DyCE) + +The Cross-Entropy (CE) loss measures the difference between predicted probabilities and ground truth labels by computing a negative log-likelihood for each class, averaged over all instances in the mini-batch: + +$$ +L _ {\mathrm {C E}} = - \frac {1}{S} \sum_ {i = 1} ^ {S} \sum_ {c = 0} ^ {N _ {C}} y _ {i, c} \log p _ {i, c}, \tag {7} +$$ + +where $y_{i,c}$ and $p_{i,c}$ are the GT and predicted probability for class $c$ , $S$ is the batch size. Taking the gradient of $L_{CE}$ for each instance, we have: + +$$ +\frac {\partial \mathcal {L} _ {\mathrm {C E}}}{\partial p _ {i , c}} = \left\{ \begin{array}{l l} - \frac {1}{S} \frac {1}{p _ {i , c}}, & \text {i f} y _ {i, c} = 1, \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {8} +$$ + +Hence, CE loss only updates the gradient for the target class per instance, using a uniform weight of $-\frac{1}{S}$ . This leads to two key problems in large imbalanced datasets: (1) equal weighting across classes overlooks class imbalance, treating frequent and rare classes the same; (2) the gradient magnitude becomes vanishingly small as $N$ scales to millions, causing ineffective updates (gradient vanishing). While recent studies have proposed reweighting schemes to address the class imbalance issue [59, 73], they fail to tackle the core problem of diminished gradient magnitudes (refer to supporting evidence in supplementary file), limiting the optimization efficiency in dense segmentation tasks. + +To address this, we propose a Dynamic CE (DyCE) loss that dynamically adjusts the weighting of gradients based on the class distribution within each mini-batch, addressing + +the persistent class imbalance issue that remains even after discarding simple instances. The key idea is to adaptively align the gradient contributions to the real-time class distribution at every training step. This is formalized as: + +$$ +L _ {\mathrm {D y C E}} = - \frac {1}{f _ {H} ^ {\omega}} \sum_ {c = 0} ^ {N _ {C}} \frac {1}{f _ {c} ^ {(1 - \omega)}} \sum_ {i = 1} ^ {S} \mathbb {1} _ {i \in H} y _ {i, c} \log p _ {i, c}, \tag {9} +$$ + +where $f_{c} = \sum_{i\in H}y_{i,c}$ is the total count of class $c$ in the mined subset $H$ which consists of $h\%$ hardest instances from the batch, $f_{H} = |H|$ is the count of instances in subset $H$ . The loss computation involves four key steps: (1) computing the standard CE loss for each sample; (2) creating a subset $H$ from the batch, similar to [85]; (3) assigning dynamic class weights $\frac{1}{f_c^{(1 - \omega)}}$ , inversely proportional to the mined class frequency; and (4) scaling the loss by a volume weight $-\frac{1}{f_H^{\omega}}$ , which adjusts for the batch size and mined subset size. The hyperparameter $\omega \in (0,1)$ acts as a weight-balancing factor, balancing the influence of instance-level and class-level weighting. The gradient of DyCE loss is: + +$$ +\frac {\partial L _ {D y C E}}{\partial p _ {i , c}} = \left\{ \begin{array}{l l} - \frac {1}{f _ {H} ^ {\omega}} \frac {1}{f _ {c} ^ {(1 - \omega)}} p _ {i, c} & \text {i f} y _ {i, c} = 1, \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {10} +$$ + +Here $\frac{1}{f_H^\omega} \frac{1}{f_c^{(1 - \omega)}} \geq \frac{1}{S}$ as compared to Equation 8, as $S \geq f_H \geq f_c$ and hence the vanishing gradient issue is resolved. + +# 4. Experimental Results + +# 4.1. Dataset Description + +We evaluate our proposed SSDA method on a segmentation task by adapting from two synthetic datasets, GTA5 [62] and SYNTHIA [63], to the real-world Cityscapes dataset [14]. The Cityscapes dataset consists of 2,975 training images and 500 validation images, all manually annotated with 19 classes. Since the test set annotations are not publicly available, we evaluate on the validation set, and tune the hyper-parameters on a small subset of the training set, following previous works [4, 32]. GTA5 provides 24,966 training images, and we consider the 19 classes that overlap with Cityscapes. The SYNTHIA dataset includes 9,400 fully labeled images, and we evaluate results based on the 16 classes it shares with Cityscapes. + +Furthermore, to validate our DyCE loss's effectiveness, we evaluate on an extremely imbalanced medical dataset, Synapse [39]. The Synapse dataset comprises 30 CT scans covering 13 different organs (i.e., foreground classes): spleen (Sp), right and left kidneys (RK/LK), gallbladder (Ga), esophagus (Es), liver (Li), stomach (St), aorta (Ao), inferior vena cava (IVC), portal and splenic veins (PSV), pancreas (Pa), and right and left adrenal glands (RAG/LAG). In this dataset, foreground voxels make up + +Table 1. Quantitative comparison of our proposed method with existing unsupervised domain adaptation (UDA), semi-supervised learning (SSL), and semi-supervised domain adaptation (SSDA) methods on $\mathbf{GTA5} \rightarrow$ Cityscapes benchmark. We report 19-class mIoU scores on the Cityscapes validation set across 0, 100, 200, 500, 100, and 2975 (100%) labeled target images. Our results are highlighted whereas the previous best and second-best results are marked in red and blue. + +
TypeMethodsLabeled Target Samples
010020050010002975
SupervisedDeepLabV2 [8]-43.048.354.858.366.1
UDAPRoDA [97]57.5-----
DaFormer [29]56.0-----
CONFETI [45]62.2-----
DIGA [67]62.7-----
SSLCowMix [22]-50.854.861.764.8-
ClassMix [57]-54.458.662.164.3-
CPS (S) [10]-55.059.563.065.7-
CPS (E) [10]-55.360.063.666.3-
DusPerb [92]-61.866.768.472.1-
SSDAASS [86]-54.256.060.264.569.1
DECOTA (S) [91]-60.761.864.266.369.2
DECOTA (E) [91]-61.362.364.767.069.9
DLDM [9]-61.260.564.366.669.8
SSDDA [13]-60.162.965.766.8-
DSTC (S) [24]-64.565.869.270.371.9
DSTC (E) [24]-65.266.470.070.972.6
S-Depth [32]-66.167.369.970.571.7
IIDM [23]-69.570.070.672.873.3
Ours w/o DyCE66.970.371.872.173.974.7
Ours w/ DyCE67.771.172.572.974.875.2
+ +only $4.37\%$ of the entire dataset, with $95.63\%$ background, and the right adrenal gland contributes a mere $0.14\%$ of foreground, whereas liver consists of $53.98\%$ foreground, underscoring the severe class imbalance. Following the setup of [76], we split the dataset into 20 scans for training, 4 for validation, and 6 for testing. + +# 4.2. Implementation Details + +Following SemiVL [33], we utilize ViT-B/16 vision encoder [18] and a Transformer text encoder [74], both initialized with CLIP pre-training [60] and generate dense embeddings following [98]. The initial learning rate is set to $10^{-4}$ , decaying exponentially with a factor of 0.9. We set the weight decay to $2 \times 10^{-4}$ and momentum to 0.9. Following [35], we use BLIP-2 [43] as our off-the-shelf captioning model $\mathcal{C}$ for all domains. $Th, \alpha$ is set to 0.95, 0.999, following [28, 33]. Following [23, 72], source images are resized to $760 \times 1280$ and target images to $512 \times 1024$ , followed by random cropping to $512 \times 512$ . SemiDAViL is trained for 40k iterations which takes $\sim 15$ hours on an NVIDIA RTX4090 GPU using Python environment. + +# 4.3. Findings and Comparison with SoTA + +Our proposed SemiDAViL framework demonstrates significant gains on both GTA5→Cityscapes and Synthia→Cityscapes benchmarks (Table 1, Table 2). + +Table 2. Quantitative comparison of our method with existing UDA, SSL, and SSDA methods on Synthia $\rightarrow$ Cityscapes benchmark. We report 16-class mIoU scores on the Cityscapes validation set and follow the same settings as in Table 1. DACS++ represents implementation of DACS [72] from UDA to SSDA. + +
TypeMethodsLabeled Target Samples
010020050010002975
SupervisedDeepLabV2 [8]-53.058.961.067.570.8
UDADaCS [72]54.8-----
PRoDA [97]62.0-----
DaFormer [29]58.8-----
DIGA [67]67.9-----
SSLCowMix [22]-61.3-
ClassMix [57]-61.467.672.373.1
DMT [21]-59.7-
DusPerb [92]-68.471.474.276.1-
SSDAASS [86]-62.164.869.873.077.1
ComplexMix [12]-70.675.6
DACS++†[72]-64.967.771.372.874.4
DLDM [9]-68.4
SSDDA [13]-70.671.872.674.0-
ALFSA [81]-68.973.577.579.079.9
SLA [95]-63.7
S-Depth [32]-72.473.575.476.377.1
IIDM [23]-74.276.477.078.879.2
Ours w/o DyCE69.574.976.877.779.279.6
Ours w/ DyCE70.276.977.278.679.780.5
+ +Table 3. Ablation experiments using three different SSDA settings on $\mathbf{GTA5}\rightarrow \mathbf{Cityscapes}$ and $\mathrm{Synthia}\rightarrow \mathrm{Cityscapes}$ to identify the contribution of individual components: Consistency Training (CT), Dynamic Cross-Entropy loss (DyCE), Vision-Language Pre-training (VLP), and DenseLanguage Guidance (DLG). + +
ComponentsGTA5 → CityscapesSynthia → Cityscapes
CTDyCEVLPDLG10020050010001002005001000
✓---54.558.262.364.660.265.371.572.0
✓✓--63.364.565.968.268.770.172.274.7
✓-✓-65.666.869.169.971.973.074.775.9
✓-✓✓70.371.672.173.974.976.877.779.2
✓✓✓✓71.172.572.974.876.977.278.679.7
+ +We compare against state-of-the-art UDA, SSL, and SSDA techniques, highlighting its robustness with varying levels of target annotations. + +In the GTA5→Cityscapes scenario, (A) using only 100 labeled target samples, SemiDAViL achieves $71.1\%$ mIoU, outperforming the previous best, IIDM [23], by $1.6\%$ . The advantage grows with 200 labeled samples, where we attain $72.5\%$ mIoU, showcasing our framework's strength in leveraging limited annotations through language-guided features; (B) in the fully unsupervised setting, our method achieves $67.7\%$ mIoU, a $5\%$ improvement over the previous best, DIGA [67], owing to our language-guided joint embedding, which provides more robust semantic alignment than divergence-based methods like DaFormer [29]; and (C) with 2975 labeled samples, our model reaches $75.2\%$ mIoU, surpassing the previous best, IIDM by $1.9\%$ , at + +Table 4. Class-wise performance evaluation of our proposed method (with and without the proposed class-balancing DyCE loss), and comparison with the existing class-balanced UDA and SSDA methods. We report 19-class and 16-class mIoU scores on the GTA5 → Cityscapes and Synthia → Cityscapes settings, respectively with 100 labeled target samples. The segmentation performance of tailed classes significantly improves by incorporating our DyCe loss in both settings. Our results are highlighted whereas the previous-best and second-best results are marked in red and blue. Please refer to supplementary file for detailed class distribution and improvement analysis. + +
TypeMethodsTarget +LabelsGTAL → Cityscapes
RoadSidewalkBuildingWallsFencePoleT-LightT-signVegTerrainSkyPersonRiderCarTruckBusTrainMotorBikemIoU
SupervisedUniMatch [92]297597.279.390.636.552.156.764.272.191.159.093.677.553.593.473.879.867.849.671.271.5
U2PL [79]97.682.190.739.353.458.064.274.090.960.293.777.049.893.764.877.947.951.273.670.5
UDACADA [90]091.346.084.534.429.732.635.836.484.543.283.060.032.283.235.046.70.033.742.249.2
IAST [54]93.857.885.139.526.726.243.134.784.932.988.062.629.087.339.249.623.234.739.651.5
DACS [72]89.939.787.930.739.538.546.452.888.044.088.867.235.884.545.750.20.027.334.052.1
Shallow [4]91.948.986.038.628.634.845.643.086.242.487.665.638.686.838.448.20.046.559.253.5
ProDA+distill [44]87.856.079.746.344.845.653.553.588.645.282.170.739.288.845.559.41.048.956.457.5
CPSL+distill [44]92.359.984.945.729.752.861.559.587.941.585.073.035.590.448.773.926.353.853.960.8
SSDAALFSA [81]10095.971.587.439.939.044.652.660.489.150.791.373.148.391.355.363.726.355.868.763.4
SS-ADA+UniMatch [89]96.475.089.243.745.153.358.268.890.755.493.875.849.791.654.667.443.647.269.466.8
SS-ADA+U2PL [89]96.575.589.747.147.755.360.668.190.655.392.177.452.592.567.167.841.249.970.868.3
Ours w/o DyCE97.380.289.550.249.258.362.369.390.657.793.278.653.992.968.467.947.956.570.970.3
Ours w/DyCE98.180.990.351.951.560.262.571.190.959.593.279.956.893.271.668.149.158.671.471.1
SINHIA → Cityscapes
SupervisedUniMatch [92]297597.582.191.252.453.060.766.375.392.3-94.179.957.594.4-82.1-57.974.575.7
U2PL [79]97.581.790.036.950.956.859.971.791.6-93.176.543.593.6-75.4-45.272.171.0
UDAFADA [77]084.540.183.14.80.034.320.127.284.8-84.053.522.685.4-43.7-26.827.845.2
IAST [54]81.941.583.317.74.632.330.928.883.4-85.065.530.886.5-38.2-33.152.749.8
DACS [72]80.625.181.921.52.637.222.724.083.7-90.867.638.382.9-38.9-28.547.648.3
Shallow [4]90.451.183.43.00.032.325.331.084.8-85.559.330.182.6-53.2-17.545.648.4
ProDA+distill [44]87.845.784.637.10.644.054.637.088.1-84.474.224.388.2-51.1-40.545.655.5
CPSL+distill [44]87.243.985.533.60.347.757.437.287.8-88.579.032.090.6-49.4-50.859.857.9
SSDASS-ADA+U2PL [89]10091.062.086.738.933.453.658.969.091.0-92.573.944.692.3-69.3-37.367.266.4
SS-ADA+UniMatch [89]97.179.490.249.849.856.958.272.291.6-93.478.153.392.8-69.1-48.472.172.0
Ours w/o DyCE98.578.991.652.752.862.363.974.391.5-94.479.756.993.1-76.6-55.574.974.9
Ours w/DyCE98.980.992.257.656.263.867.176.791.9-95.980.659.993.8-78.9-59.876.676.9
+ +attributed to the effective handling of class imbalance via the DyCE loss. + +For Synthia $\rightarrow$ Cityscapes, similar gains are observed against the previous best methods like IIDM [23] and ALFSA [81], demonstrating the strength of our dense language guidance and adaptive DyCE loss. Competing methods, such as DACs++ [72], attempt to address class imbalance via pseudo-label refinement, but their reliance on multi-stage training and hyperparameter tuning limits scalability. Unlike these, our end-to-end solution with DyCE loss adaptively re-weights based on class distribution, yielding balanced learning without the need for manual adjustments. + +Overall, our framework consistently outperforms across all benchmarks and supervision levels, demonstrating its robustness and scalability. Detailed class-wise analysis is provided in subsection 4.5. + +# 4.4. Ablation Experiments + +We conduct ablation experiments to analyze the effectiveness of each component in our proposed framework: Consistency Training (CT), Dynamic Cross-Entropy loss (DyCE), Vision-Language Pre-training (VLP), and Dense Language Guidance (DLG) in Table 3. (A) Starting with CT as our baseline, we observe moderate performance with mIoU scores of $54.5\%$ and $60.2\%$ on GTA5 and Synthia respectively with 100 labeled samples. (B) The addition of our DyCE loss significantly improves performance by addressing class imbalance issues, boosting the mIoU by $8.8\%$ (to $63.3\%$ ) and $8.5\%$ (to $68.7\%$ ) respectively. This, + +along with the findings in Table 4 and Table 5 demonstrates the effectiveness of the proposed dynamic loss function to address class imbalance. (C) When incorporating VLP without DyCE, we achieve better results than DyCE alone, with mIoU improvements of $11.1\%$ (to $65.6\%$ ) and $11.7\%$ (to $71.9\%$ ) over the baseline. This demonstrates the effectiveness of leveraging semantic knowledge from pretrained vision-language models. (D) The addition of DLG further enhances performance substantially, reaching $70.3\%$ and $74.9\%$ mIoU, as it enables fine-grained semantic understanding through dense language embeddings. (E) Finally, our full model combining all components achieves the best performance across all settings, with notable improvements of $16.6\%$ (to $71.1\%$ ) and $16.7\%$ (to $76.9\%$ ) over the baseline with 100 labeled samples. The consistent performance gains across different label ratios demonstrate the complementary nature of our proposed components in addressing the challenges of SSDA for semantic segmentation. + +# 4.5. Detailed Analysis on the DyCE Loss + +We provide a comprehensive evaluation of our proposed method (with and without the DyCE loss) against existing UDA and SSDA methods that use various solutions on the class-imbalance problem on the GTA5 $\rightarrow$ Cityscapes and Synthia $\rightarrow$ Cityscapes benchmarks, using 100 labeled target samples in Table 4. (A) Without DyCE, our model already achieves competitive mIoU scores, but the addition of DyCE enhances class separability by adaptively reweighting underrepresented classes based on their occur + +Table 5. Comparative performance improvement by incorporating our proposed DyCE loss to address class imbalance in $20\%$ labeled Synapse dataset in SSL setting. Please refer to supplementary file for detailed class-distribution and improvement analysis. + +
TypeMethodsAverageSpRKLKGaEsLiStAoIVCPSVPARAGLAG
DSC ↑ASD ↓
SupervisedV-Net [55]62.110.384.677.273.873.338.294.668.472.171.258.248.517.929.0
General SSLUA-MT [93]20.371.748.231.722.20.00.081.229.123.327.50.00.00.00.00.0
URPC [51]25.772.760.738.256.80.00.085.333.933.114.80.05.10.00.00.0
CPS [11]33.641.262.855.245.435.90.091.131.341.949.28.814.50.00.00.0
SS-Net [83]35.150.862.767.960.934.30.089.920.961.744.80.08.74.20.00.0
DST [5]34.537.757.757.246.443.70.089.033.943.346.99.021.00.00.00.0
DePL [78]36.336.062.861.048.254.80.090.236.042.548.210.717.00.00.00.0
Class-balanced SSLAdsh [27]35.339.655.159.645.852.20.089.432.847.653.08.914.40.00.00.0
CreST [80]38.322.962.164.753.843.88.185.927.254.447.714.413.018.74.64.6
SimiS [6]40.133.062.369.450.761.40.087.033.059.057.229.211.80.00.00.0
ACISSMIS [3]33.243.857.453.848.546.90.087.828.742.345.46.315.00.00.00.0
CLD [48]41.132.262.066.059.361.50.089.031.762.849.428.618.50.05.05.0
DHC [75]48.610.762.869.559.266.013.285.236.967.961.537.030.931.410.610.6
A&D [76]60.92.585.266.967.052.762.989.652.183.074.941.843.444.827.227.2
SSL+Our DyCE lossUA-MT[93]+DyCE48.1(+27.9)10.9(-60.8)55.6(+7.4)52.3(+20.6)53.4(+31.2)49.5(+49.5)51.5(+51.5)88.7(+7.5)46.0(+16.9)48.7(+25.4)47.3(+19.8)35.6(+35.6)33.4(+33.4)39.1(+39.1)24.4(+24.4)
URPC[51]+DyCE48.9(+23.2)10.1(-62.6)63.2(+2.5)59.9(+21.7)59.4(+2.6)52.5(+52.5)52.2(+52.2)89.3(+4.0)45.6(+11.7)50.6(+17.5)50.3(+35.5)30.1(+30.1)30.4(+25.3)26.7(+26.7)25.1(+25.1)
DePL[78]+DyCE51.8(+15.6)9.2(-26.1)65.9(+3.1)64.2(+3.2)65.1(+16.9)59.5(+4.7)58.2(+58.2)90.5(+0.3)41.7(+5.7)54.9(+12.4)55.6(+7.4)32.2(+21.5)35.8(+18.8)26.9(+26.9)23.3(+23.3)
DHCP[75]+DyCE57.9(+9.3)6.4(-4.3)77.5(+14.7)72.5(+3.0)64.3(+5.1)70.3(+4.3)53.2(+40.0)88.7(+3.5)44.8(+7.9)79.1(+11.2)67.8(+6.3)42.0(+5.0)38.2(+7.3)34.8(+3.4)20.1(+9.5)
A&D[76]+DyCE65.5(+4.6)5.8(-0.7)87.3(+2.1)78.5(+11.6)72.4(+5.4)65.1(+12.4)64.7(+1.8)90.9(+1.3)55.3(+3.2)84.1(+1.1)77.5(+2.6)47.3(+5.5)49.8(+6.4)48.4(+3.6)30.1(+2.9)
+ +![](images/daaeec046b38edf831598028dd936888d3022d975aff9b6ae6319ae41f1b3b4e.jpg) +Figure 4. Qualitative segmentation performance of SemiDAViL with and without DyCE loss on 100 labeled target data. + +rence, addressing inherent dataset imbalances. (B) The integration of DyCE loss shows significant improvements across a variety of challenging classes, particularly in tailed categories with high intra-class variation such as Fence, Wall, Terrain, Rider, and Pole. We provide a qualitative visualization of improvement in segmentation performance by incorporating DyCE loss in Figure 4. For GTA5→Cityscapes, the performance on tail classes like Wall (51.9%), Fence (51.5%), and Pole (60.2%) shows substantial improvement over previous state-of-the-art methods SS-ADA+U2PL [89] (47.1%, 47.7%, and 55.3% respectively) and SS-ADA+UniMatch [89] (43.7%, 45.1%, and 53.3% respectively). (C) We observe a significant improvement over UDA approaches like [4, 44, 72]. (D) The overall mIoU rises from 70.3% to 71.4% by integrating DyCE loss, showing consistent performance gains across all label splits. Similar improvements are also evident in Synthia→Cityscapes, demonstrating the robust superiority of our language-guided and class-balanced approach. + +To further validate the effectiveness of DyCE loss, we evaluate it on a more challenging scenario: Synapse medical dataset with severe class imbalance (95.63% background, 4.37% foreground: foreground classes vary from 53.98% to 0.14%). As summarized in Table 5, we highlight + +the significant impact of our DyCE loss as a plug-in enhancement across various SSL methods [51, 75, 76, 78, 93]. (A) For the lowest-performing general SSL method in Table 5, i.e., UA-MT [93], its mDSC leaps from $20.3\%$ to $48.1\%$ , with minority classes like Gallbladder improving from $0.0\%$ to $49.5\%$ , showcasing DyCE's impressive adaptive weighting. Similarly, URPC's mDSC increases from $25.7\%$ to $48.9\%$ , whereas the previous best general SSL-based DePL sees a boost from $36.3\%$ to $51.8\%$ , underscoring DyCE's capability to mitigate severe class imbalance by prioritizing underrepresented organs (e.g., Right Adrenal Gland from $0.0\%$ to $26.9\%$ ). (B) Recent SoTA of balanced SSL like DHC and A&D also benefit, with DHC's DSC rising from $48.6\%$ to $57.9\%$ , and A&D reaching $65.5\%$ (up from $60.9\%$ ). Notable gains include improvements in Gallbladder (DHC: $66.0\%$ to $70.3\%$ ) and Left Adrenal Gland (A&D: $27.2\%$ to $30.1\%$ ), demonstrating DyCE's effectiveness as a plug-in loss across class-imbalanced datasets. + +Further experimental findings, qualitative and quantitative analysis are provided in the supplementary material. + +# 5. Conclusion + +In this work, we introduced SemiDAViL, a novel SSDA framework that leverages vision-language guidance and a dynamic loss formulation to address key challenges in domain-adaptive semantic segmentation. Through comprehensive experiments on several SSDA and SSL benchmarks, our method demonstrated consistent improvements over state-of-the-art techniques, especially in low-label regimes and class-imbalanced scenarios. The integration of vision-language pre-training, dense language embeddings, and the proposed DyCE loss contributes to discriminative feature extraction, better handling of minority classes, and enhanced semantic understanding. Overall, SemiDAViL sets a new benchmark in SSDA, showcasing strong generalizability across diverse domain shifts and label constraints. + +# References + +[1] Hritam Basak and Zhaozheng Yin. Quest for clone: Test-time domain adaptation for medical image segmentation by searching the closest clone in latent space. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 555-566. Springer, 2024. 2 +[2] Hritam Basak and Zhaozheng Yin. Forget more to learn more: Domain-specific feature unlearning for semi-supervised and unsupervised domain adaptation. In European Conference on Computer Vision, pages 130-148. Springer, 2025. 2 +[3] Hritam Basak, Sagnik Ghosal, and Ram Sarkar. Addressing class imbalance in semi-supervised image segmentation: A study on cardiac mri. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 224-233. Springer, 2022. 8 +[4] Adriano Cardace, Pierluigi Zama Ramirez, Samuele Salti, and Luigi Di Stefano. Shallow features guide unsupervised domain adaptation for semantic segmentation at class boundaries. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1160-1170, 2022. 5, 7, 8 +[5] Baixu Chen, Junguang Jiang, Ximei Wang, Pengfei Wan, Jianmin Wang, and Mingsheng Long. Debiased self-training for semi-supervised learning. Advances in Neural Information Processing Systems, 35:32424-32437, 2022. 8 +[6] Hao Chen, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Marios Savvides, and Bhiksha Raj. An embarrassingly simple baseline for imbalanced semisupervised learning. arXiv preprint arXiv:2211.11086, 2022. 8 +[7] Leiyu Chen, Shaobo Li, Qiang Bai, Jing Yang, Sanlong Jiang, and Yanming Miao. Review of image classification algorithms based on convolutional neural networks. Remote Sensing, 13(22):4712, 2021. 1 +[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 6 +[9] Shuaijun Chen, Xu Jia, Jianzhong He, Yongjie Shi, and Jianzhuang Liu. Semi-supervised domain adaptation based on dual-level domain mixing for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11018-11027, 2021. 6 +[10] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2613-2622, 2021. 6 +[11] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2613-2622, 2021. 8 + +[12] Ying Chen, Xu Ouyang, Kaiyue Zhu, and Gady Agam. Semi-supervised domain adaptation for semantic segmentation. arXiv preprint arXiv:2110.10639, 2021. 6 +[13] Ying Chen, Xu Ouyang, Kaiyue Zhu, and Gady Agam. Semi-supervised dual-domain adaptation for semantic segmentation. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 230-237. IEEE, 2022. 6 +[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5 +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3 +[16] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 4 +[17] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 2 +[18] 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. ICLR, 2021. 6 +[19] Ozgür Erkent and Christian Laugier. Semantic segmentation with unsupervised domain adaptation under varying weather conditions for autonomous vehicles. IEEE Robotics and Automation Letters, 5(2):3580-3587, 2020. 1 +[20] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul de Charette. A simple recipe for language-guided domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23428-23437, 2024. 3 +[21] Zhengyang Feng, Qianyu Zhou, Qiqi Gu, Xin Tan, Guangliang Cheng, Xuequan Lu, Jianping Shi, and Lizhuang Ma. Dmt: Dynamic mutual training for semi-supervised learning. Pattern Recognition, 130:108777, 2022. 6 +[22] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. BMVC, 2020. 6 +[23] Weifu Fu, Qiang Nie, Jialin Li, Yuhuan Lin, Kai Wu, Jian Li, Yabiao Wang, Yong Liu, and Chengjie Wang. Iidm: Inter and intra-domain mixing for semi-supervised domain adaptation in semantic segmentation, 2024. 2, 6, 7 +[24] Yuan Gao, Zilei Wang, and Yixin Zhang. Delve into source and target collaboration in semi-supervised domain adaptation for semantic segmentation. In 2024 IEEE International + +Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2, 6 +[25] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2 +[26] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech 256, 2022. 3 +[27] Lan-Zhe Guo and Yu-Feng Li. Class-imbalanced semi-supervised learning with adaptive thresholding. In International conference on machine learning, pages 8082-8094. PMLR, 2022. 8 +[28] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 5, 6 +[29] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9924-9935, 2022. 3, 6 +[30] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Hrda: Context-aware high-resolution domain-adaptive semantic segmentation. In European conference on computer vision, pages 372-391. Springer, 2022. 1 +[31] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11721-11732, 2023. 3 +[32] Lukas Hoyer, Dengxin Dai, Qin Wang, Yuhua Chen, and Luc Van Gool. Improving semi-supervised and domain-adaptive semantic segmentation with self-supervised depth estimation. International Journal of Computer Vision, 131 (8):2070-2096, 2023. 2, 5, 6 +[33] Lukas Hoyer, David Joseph Tan, Muhammad Ferjad Naeem, Luc Van Gool, and Federico Tombari. Semivl: Semi-supervised semantic segmentation with vision-language guidance. In European Conference on Computer Vision, pages 257-275. Springer, 2025. 2, 4, 6 +[34] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 2 +[35] Tarun Kalluri, Bodhisattwa Prasad Majumder, and Manmohan Chandraker. Tell, don't show!: Language guidance eases transfer across domains in images and videos. arXiv preprint arXiv:2403.05535, 2024. 6 +[36] Guoliang Kang, Yunchao Wei, Yi Yang, Yueting Zhuang, and Alexander Hauptmann. Pixel-level cycle association: A new perspective for domain adaptive semantic segmentation. Advances in neural information processing systems, 33:3569-3580, 2020. 1 +[37] Young-Eun Kim, Yu-Won Lee, and Seong-Whan Lee. Lc-msm: Language-conditioned masked segmentation model + +for unsupervised domain adaptation. Pattern Recognition, 148:110201, 2024. 3 +[38] Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html, 5(4):1, 2010. 3 +[39] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, page 12, 2015. 5 +[40] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2023. 2 +[41] Jingjing Li, Erpeng Chen, Zhengming Ding, Lei Zhu, Ke Lu, and Heng Tao Shen. Maximum density divergence for domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 43(11):3918-3930, 2020. 2 +[42] Jichang Li, Guanbin Li, Yemin Shi, and Yizhou Yu. Cross-domain adaptive clustering for semi-supervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2505-2514, 2021. 2 +[43] 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. 6 +[44] Ruihuang Li, Shuai Li, Chenhang He, Yabin Zhang, Xu Jia, and Lei Zhang. Class-balanced pixel-level self-labeling for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11593-11603, 2022. 7, 8 +[45] Tianyu Li, Subhankar Roy, Huayi Zhou, Hongtao Lu, and Stéphane Lathuilière. Contrast, stylize and adapt: Unsupervised contrastive learning framework for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4869-4879, 2023. 6 +[46] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2 +[47] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 3 +[48] Yiqun Lin, Huifeng Yao, Zezhong Li, Guoyan Zheng, and Xiaomeng Li. Calibrating label distribution for class-imbalanced barely-supervised knee segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 109-118. Springer, 2022. 8 +[49] Chang Liu, Henghui Ding, Yulun Zhang, and Xudong Jiang. Multi-modal mutual attention and iterative interaction for referring image segmentation. IEEE Transactions on Image Processing, 32:3054-3065, 2023. 4 + +[50] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7086-7096, 2022. 2 +[51] Xiangde Luo, Guotai Wang, Wenjun Liao, Jieneng Chen, Tao Song, Yinan Chen, Shichuan Zhang, Dimitris N Metaxas, and Shaoting Zhang. Semi-supervised medical image segmentation via uncertainty rectified pyramid consistency. Medical Image Analysis, 80:102517, 2022. 8 +[52] Jie Ma, Chuan Wang, Yang Liu, Liang Lin, and Guanbin Li. Enhanced soft label for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1185-1195, 2023. 2 +[53] Elham Amin Mansour, Ozan Unal, Suman Saha, Benjamin Bejar, and Luc Van Gool. Language-guided instance-aware domain-adaptive panoptic segmentation. arXiv preprint arXiv:2404.03799, 2024. 3 +[54] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 415-430. Springer, 2020. 7 +[55] Fausto Miletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 8 +[56] Khan Muhammad, Tanveer Hussain, Hayat Ullah, Javier Del Ser, Mahdi Rezaei, Neeraj Kumar, Mohammad Hijji, Paolo Bellavista, and Victor Hugo C de Albuquerque. Vision-based semantic segmentation in scene understanding for autonomous driving: Recent achievements, challenges, and outlooks. IEEE Transactions on Intelligent Transportation Systems, 23(12):22694-22715, 2022. 3 +[57] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1369-1378, 2021. 6 +[58] Can Qin, Yizhou Wang, and Yun Fu. Robust semi-supervised domain adaptation against noisy labels. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 4409-4413, 2022. 2 +[59] Shoumeng Qiu, Xianhui Cheng, Hong Lu, Haiqiang Zhang, Ru Wan, Xiangyang Xue, and Jian Pu. Subclassed loss: Rethinking data imbalance from subclass perspective for semantic segmentation. IEEE Transactions on Intelligent Vehicles, 2023. 5 +[60] 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. 2, 3, 6 +[61] Wenqi Ren, Yang Tang, Qiyu Sun, Chaoqiang Zhao, and Qing-Long Han. Visual semantic segmentation based on + +few/zero-shot learning: An overview. IEEE/CAA Journal of Automatica Sinica, 2023. 1 +[62] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102-118. Springer, 2016. 5 +[63] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 5 +[64] Manisha Saini and Seba Susan. Tackling class imbalance in computer vision: a contemporary review. Artificial Intelligence Review, 56(Suppl 1):1279-1335, 2023. 3 +[65] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Darrell, and Kate Saenko. Semi-supervised domain adaptation via minimax entropy. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8050-8058, 2019. 2 +[66] Manuel Schwonberg, Joshua Niemeijer, Jan-Aike Termöhlen, Nico M Schmidt, Hanno Gottschalk, Tim Fingscheidt, et al. Survey on unsupervised domain adaptation for semantic segmentation for visual perception in automated driving. IEEE Access, 11:54296-54336, 2023. 1 +[67] Fengyi Shen, Akhil Gurram, Ziyuan Liu, He Wang, and Alois Knoll. Diga: Distil to generalize and then adapt for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15866-15877, 2023. 6 +[68] Wei Shen, Zelin Peng, Xuehui Wang, Huayu Wang, Ji-azhong Cen, Dongsheng Jiang, Lingxi Xie, Xiaokang Yang, and Qi Tian. A survey on label-efficient deep image segmentation: Bridging the gap between weak supervision and dense prediction. IEEE transactions on pattern analysis and machine intelligence, 45(8):9284–9305, 2023. 1, 3 +[69] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastian Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 3 +[70] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 4 +[71] Hans Thisanke, Chamli Deshan, Kavindu Chamith, Sachith Seneviratne, Rajith Vidanaarachchi, and Damayanthi Herath. Semantic segmentation using vision transformers: A survey. Engineering Applications of Artificial Intelligence, 126:106669, 2023. 1 + +[72] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1379-1389, 2021. 6, 7, 8 +[73] Thanh-Dat Truong, Ngan Le, Bhiksha Raj, Jackson Cothren, and Khoa Luu. Freedom: Fairness domain adaptation approach to semantic scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19988-19997, 2023. 5 +[74] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 6 +[75] Haonan Wang and Xiaomeng Li. Dhc: Dual-debiased heterogeneous co-training framework for class-imbalanced semi-supervised medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 582-591. Springer, 2023. 8 +[76] Haonan Wang and Xiaomeng Li. Towards generic semi-supervised framework for volumetric medical image segmentation. Advances in Neural Information Processing Systems, 36, 2024. 3, 6, 8 +[77] Haoran Wang, Tong Shen, Wei Zhang, Ling-Yu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In European conference on computer vision, pages 642-659. Springer, 2020. 7 +[78] Xudong Wang, Zhirong Wu, Long Lian, and Stella X Yu. Debiased learning from naturally imbalanced pseudolabels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14647-14657, 2022. 8 +[79] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo-labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4248-4257, 2022. 7 +[80] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10857-10866, 2021. 8 +[81] Lu Wen, Yuanyuan Xu, Zhenghao Feng, Jiliu Zhou, Luping Zhou, and Yan Wang. Semi-supervised domain adaptation for semantic segmentation via active learning with feature-and semantic-level alignments. IEEE Transactions on Intelligent Vehicles, 2024. 6, 7 +[82] Wenbo Qi1 Jiafei Wu and SC Chan. Gradient-aware for class-imbalanced semi-supervised medical image segmentation. ECCV, 98:1-86, 2024. 3 +[83] Yicheng Wu, Zhonghua Wu, Qianyi Wu, Zongyuan Ge, and Jianfei Cai. Exploring smoothness and class-separation for semi-supervised medical image segmentation. In International conference on medical image computing and computer-assisted intervention, pages 34-43. Springer, 2022. 8 + +[84] Yao Wu, Mingwei Xing, Yachao Zhang, Yuan Xie, and Yanyun Qu. Clip2uda: Making frozen clip reward unsupervised domain adaptation in 3d semantic segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8662-8671, 2024. 3 +[85] Zifeng Wu, Chunhua Shen, and Anton van den Hengel. Bridging category-level and instance-level semantic image segmentation. arXiv preprint arXiv:1605.06885, 2016. 5 +[86] Binhui Xie, Longhui Yuan, Shuang Li, Chi Harold Liu, and Xinjing Cheng. Towards fewer annotations: Active learning via region impurity and prediction uncertainty for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8068-8078, 2022. 6 +[87] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 2 +[88] Hongliang Yan, Zhetao Li, Qilong Wang, Peihua Li, Yong Xu, and Wangmeng Zuo. Weighted and class-specific maximum mean discrepancy for unsupervised domain adaptation. IEEE Transactions on Multimedia, 22(9):2420-2433, 2019. 3 +[89] Weihao Yan, Yeqiang Qian, Yueyuan Li, Tao Li, Chunxiang Wang, and Ming Yang. Ss-ada: A semi-supervised active domain adaptation framework for semantic segmentation. arXiv preprint arXiv:2407.12788, 2024. 2, 7, 8 +[90] Jinyu Yang, Weizhi An, Chaochao Yan, Peilin Zhao, and Junzhou Huang. Context-aware domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 514-524, 2021. 7 +[91] Luyu Yang, Yan Wang, Mingfei Gao, Abhinav Shrivastava, Kilian Q Weinberger, Wei-Lun Chao, and Ser-Nam Lim. Deep co-training with task decomposition for semi-supervised domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8906-8916, 2021. 2, 6 +[92] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7236-7246, 2023. 2, 6, 7 +[93] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Medical image computing and computer assisted intervention-MICCAI 2019: 22nd international conference, Shenzhen, China, October 13–17, 2019, proceedings, part II 22, pages 605–613. Springer, 2020. 8 +[94] Lei Yu, Wanqi Yang, Shengqi Huang, Lei Wang, and Ming Yang. High-level semantic feature matters few-shot unsupervised domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11025-11033, 2023. 2 + +[95] Yu-Chu Yu and Hsuan-Tien Lin. Semi-supervised domain adaptation with source label adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24100-24109, 2023. 6 +[96] Jing Zhang, Yingshuai Xie, Weichao Ding, and Zhe Wang. Cross on cross attention: Deep fusion transformer for image captioning. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):4257-4268, 2023. 4 +[97] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12414-12424, 2021. 6 +[98] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 2, 6 +[99] Jinghao Zhou, Li Dong, Zhe Gan, Lijuan Wang, and Furu Wei. Non-contrastive learning meets language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11028–11038, 2023. 2 +[100] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175-11185, 2023. 2 +[101] Zheng Zhou, Change Zheng, Xiaodong Liu, Ye Tian, Xiaoyi Chen, Xuexue Chen, and Zixun Dong. A dynamic effective class balanced approach for remote sensing imagery semantic segmentation of imbalanced data. Remote Sensing, 15(7):1768, 2023. 3 +[102] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 3 +[103] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 1 \ No newline at end of file diff --git a/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/images.zip b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0597d5831c786d1172d77ad8ed6c5c3c53810494 --- /dev/null +++ b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1087d1d2fe92faff18e5fe3350c06f11f0729ef59d0c1328d492ff744dace60e +size 686410 diff --git a/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/layout.json b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e192b2bcd2c6a26cfeb720f04dec5af850b4f6c0 --- /dev/null +++ b/CVPR/2025/SemiDAViL_ Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df89f02d000bd2f14b820dfac92f1edd2537f4b18a22eaa1f941c4f3df32f1d5 +size 520423 diff --git a/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/93ac07f9-5c5a-4084-8709-6f1dc8a9bd3d_content_list.json b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/93ac07f9-5c5a-4084-8709-6f1dc8a9bd3d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a60d0d40adbd2e51e1fdabcd1c170dbe8498d7f9 --- /dev/null +++ b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/93ac07f9-5c5a-4084-8709-6f1dc8a9bd3d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0719bc505834ba152670953a45e27adb84df6a6b3a771da85548a8a588993482 +size 95046 diff --git a/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/93ac07f9-5c5a-4084-8709-6f1dc8a9bd3d_model.json b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/93ac07f9-5c5a-4084-8709-6f1dc8a9bd3d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a002c0dd630b66a8a9ae48477b09493d5a253085 --- /dev/null +++ b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/93ac07f9-5c5a-4084-8709-6f1dc8a9bd3d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31d68e5324b1078e70161f387039354514aa60bdc761a3193f89e4baa0062f16 +size 117924 diff --git a/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/93ac07f9-5c5a-4084-8709-6f1dc8a9bd3d_origin.pdf b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/93ac07f9-5c5a-4084-8709-6f1dc8a9bd3d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..892f65abc7acb8fa683d272e62cb5f1eafb599de --- /dev/null +++ b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/93ac07f9-5c5a-4084-8709-6f1dc8a9bd3d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:faa1b5767e003350d25934d57d1736f351399518de7bfc56be8165d6485ffd92 +size 2604352 diff --git a/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/full.md b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b56c649153bfd6d5cddb1b012df7822d8fe58518 --- /dev/null +++ b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/full.md @@ -0,0 +1,403 @@ +# SemiETS: Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting + +Dongliang Luo* Hanshen Zhu* Ziyang Zhang Dingkang Liang Xudong Xie + +Yuliang Liu Xiang Bai + +Huazhong University of Science and Technology + +{ldl, zhs_china, zzyzz, dkliang, xdxie, ylliu, xbai}@hust.edu.cn + +# Abstract + +Most previous scene text spotting methods rely on high-quality manual annotations to achieve promising performance. To reduce their expensive costs, we study semi-supervised text spotting (SSTS) to exploit useful information from unlabeled images. However, directly applying existing semi-supervised methods of general scenes to SSTS will face new challenges: 1) inconsistent pseudo labels between detection and recognition tasks, and 2) sub-optimal supervisions caused by inconsistency between teacher/student. Thus, we propose a new Semi-supervised framework for End-to-end Text Spotting, namely SemiETS that leverages the complementarity of text detection and recognition. Specifically, it gradually generates reliable hierarchical pseudo labels for each task, thereby reducing noisy labels. Meanwhile, it extracts important information in locations and transcriptions from bidirectional flows to improve consistency. Extensive experiments on three datasets under various settings demonstrate the effectiveness of SemiETS on arbitrary-shaped text. For example, it outperforms previous state-of-the-art SSL methods by a large margin on end-to-end spotting $(+8.7\%, +5.6\%)$ , and $+2.6\%$ H-mean under $0.5\%$ , $1\%$ , and $2\%$ labeled data settings on TotalText, respectively). More importantly, it still improves upon a strongly supervised text spotter trained with plenty of labeled data by $2.0\%$ . Compelling domain adaptation ability shows practical potential. Moreover, our method demonstrates consistent improvement on different text spotters. Code will be available at https://github.com/DrLuo/SemiETS. + +# 1. Introduction + +Text spotting aims at concurrently localizing and recognizing texts from images. Most previous works [6, 7, 18, 23, 44, 46] highly rely on high-quality annotations for satisfac + +![](images/b41f35c8dc61e005bac67d2e1830706500b65a1a07fbc16f6dd8cefe9f990f41.jpg) + +![](images/eea44acfdedf07450d7e8a0ceac2711d9d340e0b84a4457cd7a0cfed3f8b2947.jpg) +(a) Semi-supervised text detection + +![](images/3832dbaf1377793475b3e73bffb68bebb4431bc3f2e7dc13af14c5bd597169f2.jpg) + +![](images/5702d10cd5b930b9862c1877aa21982e06fd7a7690e270a775c8b94baa42800e.jpg) +$\rightarrow$ + +![](images/6697d0c4515f7d13eb16d5c9f3565979b1290448c88796ae57bd29d4673438f6.jpg) +Unlabeled image + +![](images/8869ee2033919152050bfb1bc46c2bfd8fb9de3101305b2f683b3145e0381c5e.jpg) +(b) Semi-supervised text recognition +(c) Semi-supervised text spotting +Transcription +# +Coordinates +Pseudo labels +Figure 1. Comparison of different semi-supervised tasks related to text, including (a) semi-supervised text detection, (b) semi-supervised text recognition, and (c) semi-supervised text spotting. The red arrow indicates the supervision flow of pseudo labels. + +tory performance. However, annotating such data is expensive and laborious due to its complex formats. To alleviate the cost of labeling, exploring effective information from unlabeled text images is a valuable research direction. To achieve that, semi-supervised learning (SSL) is a natural and effective approach. + +Recently, SSL methods have made considerable progress in general scenarios such as semi-supervised object detection (SSOD) [15, 29, 40, 45]. The common practice is to use a teacher model to generate pseudo labels on unlabeled sets, which act as the ground truth (GT) for the student model [31, 34]. Thus, migrating them to optical character recognition (OCR) is natural and straightforward. Although semi-supervised methods have been applied to OCR tasks, most focus on one single task of text detection [16, 27, 35] or recognition [1, 12, 47]. The former is mostly the modification of SSOD frameworks to fit the irregular shape of texts [27] or improve label selection [36, 50], while not involving the recognition part, as in Fig. 1 (a). The latter, illustrated in Fig. 1 (b), usually performs pseudo-labeling [12, 24] or consistency-based regu + +larization [43, 47] on cropped text regions, assuming the detection is already accurate beforehand. Differently, in semi-supervised text spotting (SSTS), both the location and content of text need to be carefully considered, as in Fig. 1 (c), increasing the complexity. Not only the optimization goal is different, but also the relationship between detection and recognition is vital, hence bringing new challenges to SSL methods: 1) inconsistency between tasks that the pseudo labels of detection and recognition are not always accurate concurrently; 2) ambiguous and ineffective supervision signals caused by inconsistency between teacher and student. + +Therefore, we propose a new framework for Semi-supervised End-to-end Text Spotting named SemiETS following the teacher-student architecture with two key designs: a Progressive Sample Assignment (PSA) module to tackle the inconsistency between tasks, and a Mutual Mining Strategy (MMS) to boost the effective supervision signals and reduce ambiguity. Specifically, PSA selects reliable pseudo labels by comprehensively considering the joint constraints and then gradually assigns hierarchical labels to each task. The quality of pseudo labels and the rationality of label assignment is thereby improved. + +Furthermore, leveraging the complementarity between text detection and recognition, we excavate effective information using the proposed MMS to relieve the inconsistency problem from bidirectional flows. On the one hand, it estimates the reliability of transcription pseudo labels using the localization discrepancy between teacher and student. On the other hand, it propagates recognition information to the detection flow to softly amplify informative detection supervision signals to guide the student. + +We conduct extensive experiments on a variety of scene text datasets. Experimental results demonstrate consistent superior performances of SemiETS over the state-of-the-art (SOTA) semi-supervised frameworks on arbitrary-shaped texts under various settings. Particularly, the improvements are more pronounced when using smaller proportions of labeled data. For example, We significantly outperform the SOTA SSL methods [31, 34] by $8.7\%$ and $4.7\%$ without lexicon on Total-Text and CTW1500 using only $0.5\%$ labeled data, respectively. Furthermore, it still improves upon a strong supervised baseline already trained using extensive labeled data with $2.0\%$ H-mean in E2E (None) on Total-Text, which verifies the potential of our method to further boost the ability of text spotters utilizing unlabeled data. Moreover, we prove SemiETS is compatible with various text spotters and brings consistent improvements. + +The advantages of this paper can be summarized as: 1) We propose a semi-supervised framework, namely SemiETS, for end-to-end text spotting that can mine effective information from extensive unlabeled data. To our knowledge, this is the first effort exploring this task. 2) Aiming at the problem of inconsistency in semi-supervised + +text spotting, SemiETS consists of a Progressive Sample Assignment module to improve label reliability using hierarchy and a Mutual Mining Strategy to relieve ambiguity and amplify effective supervision signals. 3) SemiETS not only significantly outperforms existing SSL methods under various settings but can continue to boost performance upon a well-trained text spotter. Especially, it outperforms previous SSL SOTA by $8.7\%$ when only using $0.5\%$ labeled data. + +# 2. Related Work + +# 2.1. End-to-End Text Spotting + +Text spotting aiming at concurrently localizing and recognizing texts concurrently [11]. Most studies mainly address two challenges: representing arbitrary shapes texts [13, 14, 18, 21, 23, 26, 37, 39] and enhancing the intrinsic synergy of text detection and recognition [6, 7, 44, 46]. However, these methods require expensive labeling costs. To reduce label dependency, SPTS series [22, 25] simplify the spatial representation to a single point. Some [9, 32, 41] even remove detection labels and only use transcriptions. However, at least the recognition GT is given to these methods. Thus, they cannot utilize information from unlabeled data. We explore the semi-supervised paradigm, focusing on generating high-quality pseudo labels for text spotting. + +# 2.2. Semi-supervised Object Detection + +It is a common practice for SSOD to generate pseudo labels using a teacher model and expect the student detectors to make consistent predictions on augmented input images. STAC [31] first trains a teacher detector with labeled data. Subsequent studies simultaneously update the teacher by EMA inherited from Mean-Teacher [34]. Researchers are committed to improving the quality of pseudo labels for classification and regression in one-stage [15, 20, 40, 49], two-stage [10, 19, 33, 42] and DETR-based [29, 45] detectors, or extending to oriented objects [5]. However, SSOD methods for general scenarios hardly consider the characteristics of the text. Moreover, SSTS requires both text detection and recognition predictions, increasing its complexity. + +# 2.3. Semi-supervised Text Detection & Recognition + +Most semi-supervised methods for OCR are designed only for text detection or recognition, while frameworks for text spotting are rarely studied. + +Semi-supervised Text Detection. WeText [35] proposes a semi-supervised framework for character detection, later extended to curved texts by Qin et al. [27]. Subsequent methods refine pseudo-labeling by improving label quality. Some explore thresholding techniques [50] or apply clustering [36], while others use character-level information and context refinement to suppress false positives [16]. However, similar to SSOD methods, these approaches seldom + +involve the recognition aspect. + +Semi-supervised Text Recognition. Baek et al. [1] apply pseudo-labeling to improve scene text recognition. Due to the sequential property of the text, subsequent approaches aim to develop pseudo-labeling [12, 24] and regularization [4, 43, 47] strategies for sequence. To select reliable pseudo labels, Li et al. [12] set dynamic thresholds per character, while Seq-UPS [24] uses sequential uncertainty estimation. In regularization, some methods [4, 43] apply sequence-level consistency, while Others use character-level regularization and tackle misalignment of characters by sharing context information [47] or reinforcement learning [43]. However, the input images for them are already focused on text regions, which is not guaranteed in SSTS, where detected regions might deviate. + +Unlike the above works, in this paper, we explore semi-supervised text spotting, which reduces the annotation cost and boosts text spotters. In particular, we carefully consider the detection and recognition tasks concurrently. + +# 3. Preliminary + +# 3.1. Task Definition + +Given a labeled image set $D_{l} = \{x_{i}^{s},y_{i}^{s}\}_{i = 1}^{N_{l}}$ and an unlabeled image set $D_{u} = \{x_{i}^{u}\}_{i = 1}^{N_{u}}$ , where $x_{i}^{s} / x_{i}^{u}$ represents labeled/unlabeled images and $N_{l} / N_{u}$ is the number of labeled/unlabeled images, respectively. Each label $y_{i}^{s}$ contains ground truth information for text spotting, including the coordinates $p$ and transcriptions $q$ of all text instances. Semi-supervised text spotting (SSTS) aims to leverage both labeled and unlabeled data to train a strong text spotter. + +Following the pseudo-labeling paradigm inheriting from Mean-Teacher [34], our framework consists of a teacher model and a student model. The teacher generates pseudo labels from a weakly augmented image to guide the student, which takes a strongly augmented version as input. Additionally, the labeled dataset $D_{l}$ supervises the student in a standard manner. The teacher is updated simultaneously using Exponential Moving Average (EMA). + +# 3.2. Inconsistency Investigation + +In this part, we mainly summarize the inconsistent issues in semi-supervised text spotting from two aspects. + +Inconsistency between tasks. One challenge of SSTS is that the predictions contain two tasks, i.e., text detection and recognition. However, ensuring both pseudo labels are always reliable is not easy. For example, even though the location of the text region is precise, the recognition result may still be incorrect, as shown in Fig. 2 (a). Naturally, simply using every predicted text instance equally to supervise the student is unsuitable, as it may introduce noisy labels. Therefore, we propose a Progressive Sample Assignment module to distinguish reliable information. + +![](images/e4a647f88d7a643fbbde851450faa81860b33a6c30276ef43ea539b143383b9c.jpg) +(a) Inconsistency between tasks +(b) Inconsistency between teacher/student +Figure 2. Illustration of the inconsistency issues including inconsistency (a) between tasks, and (b) between teacher and student. + +![](images/7812363274ed54ee92f079542297fb39cc94e4bd26603455165b02751395fa7c.jpg) + +Inconsistency between teacher & student. The teacher's and student's predictions may differ in text position or content. Since the recognition results are pretty sensitive to the detected regions, a slight deviation would cause a change in recognition results, as shown in Fig. 2 (b). Intuitively, it is sub-optimal to directly force the student to learn the content of a misaligned region, which would lead to ambiguity. Therefore, we propose a Mutual Mining Strategy to improve the alignment of recognition supervision and boost proper supervision signals to guide the student. + +# 4. Methodology + +The overall framework of SemiETS is illustrated in Fig. 3 focusing on unlabeled data flow. It consists of two key components: A Progressive Sample Assignment (PSA) module improves the quality of pseudo labels to supervise the student; the Mutual Mining Strategy (MMS) further mines useful information from E2E labels to relieve inconsistency. We take DeepSolo [44], a recent DETR-based text spotter with a concise architecture, as an example. + +# 4.1. Progressive Sample Assignment + +Generating high-quality pseudo labels is essential for semi-supervised learning but is challenging in SSTS. Due to taskwise inconsistency, selecting pseudo labels solely from a single perspective is inadequate and rigid. Thus, we propose PSA to progressively distinguish useful localization and recognition labels, constructing hierarchical supervisions to suppress noisy information. + +As shown in Fig. 3, the PSA consists of two steps. Firstly, a joint constraint filter first removes predictions with the classification score $s^t$ lower than a threshold $\mathcal{T}_D$ or with void decoded text. Then, the student's predictions are matched to the remaining samples through the cost-based assignment. The cost function comprehensively integrates the detection and recognition, which is formulated as: + +$$ +\begin{array}{l} \mathcal {C} \left(\hat {Y} _ {i} ^ {s}, Y _ {j} ^ {t}\right) = \lambda_ {\mathrm {c l s}} \operatorname {F L} \left(s _ {i} ^ {s}, B \left(s _ {j} ^ {t}\right)\right) + \lambda_ {\text {t e x t}} \mathcal {L} _ {\text {t e x t}} \left(\hat {q} _ {i} ^ {s}, q _ {j} ^ {t}\right) \tag {1} \\ + \lambda_ {\mathrm {c o o r d}} \| p _ {i} ^ {s}, p _ {j} ^ {t} \| _ {1}, i \in N, j \in M, \\ \end{array} +$$ + +where $\hat{Y}_i^s$ is the predictions of student and $Y_j^t$ is the coarsely selected pseudo labels. $\mathrm{FL}(*)$ is derived from the focal loss. $B$ is binarization. $\mathcal{L}_{text}$ is the recognition loss between the + +![](images/6037af955b4413c8c34c9270d74e554f55c3b7ebf166ad6c11814cc33615250a.jpg) +Figure 3. Overview of the proposed framework where the labeled data flow is omitted. Given an unlabeled image, Progressive Sample Assignment selects reliable pseudo labels and splits them into Det-only and E2E labels. Then, the Mutual Mining Strategy explores effective information in E2E labels in a crossover strategy with Spatial-aware Consistency Integration and Content-aware Region Calibration. + +predicted characters $\hat{q}_i^s$ and the decoded transcription $q_j^t$ . $p_i$ refers to the points on the polygon. $M$ and $N$ denote the number of student's predictions and pseudo labels, respectively. $\lambda_{\mathrm{cls}}$ , $\lambda_{\mathrm{text}}$ , and $\lambda_{\mathrm{coord}}$ are set to 1.0, 1.0, and 0.5 by default, respectively. Hungarian algorithm is leveraged to establish optimal bipartite matching. To improve the training efficiency in DETR-based text spotters, we introduce the hybrid matching (HM) strategy [45] that adopts the one-to-many assignment in early iterations. Here, each matched pseudo label is valid to supervise the detection flow. + +In the second step, we further select reliable recognition labels with a recognition filter combining the following process: 1) filtering by threshold: the confidence score $c^t$ of the pseudo label is larger than a certain threshold $\mathcal{T}_R$ ; 2) confidence comparison (CC): $c^t$ is larger than the matched student's confidence score $c^s$ . CC is used to ensure the teacher has positive guidance for the student. + +As a result, pseudo labels are categorized into Det-only and E2E. The transcriptions of Det-only labels are not included in the optimization. Unlike treating every instance equally, our differentiated assignment maximizes the utilization of data while reducing misleading information, thereby improving the quality of supervision for both tasks. + +# 4.2. Mutual Mining Strategy + +Since text detection and recognition have a natural synergy, we further develop MMS to mine useful information by explicitly bridging the correlation between these two tasks. As illustrated in Fig. 3, it consists of a Spatial-aware Consistency Integration (SCI) and a Content-aware Region Calibration (CRC) to conduct mutual mining in bidirectional flows. Specifically, SCI integrates the detection accuracy + +![](images/f6585a8349c9a9381ac53fab36ded1f7339b811df0f0104d72104b75555898eb.jpg) +(a) Spatial-aware Consistency Integration (SCI) + +![](images/4c6fac9d9c256d7cd93c9e0227f457943c723b016a3bb48d6c0e5d8373859141.jpg) +(b) Content-aware Region Calibration (CRC) +Figure 4. The details of SCI and CRC in the proposed Mutual Mining Strategy. We distinguish text instances using green and blue in (b) and the wrongly recognized characters indicated in red. + +into the recognition flow. In turn, CRC facilitates the learning of detection incorporated with recognition information. + +# 4.2.1. Spatial-aware Consistency Integration + +Since text recognition results are sensitive to the detected text regions, the misaligned regions between teacher and student may cause ambiguous recognition supervision. Meanwhile, certain pseudo labels could be erroneous despite having high confidence scores. Therefore, we propose SCI to integrate detection information into the recognition flow, as shown in Fig. 4 (a). Firstly, softly adjusting the importance of the recognition supervision according to region alignment can relieve ambiguity, and enable the framework to better understand the geometric priors of recognition. Secondly, the discrepancy in detected regions between teacher and student can be regarded as the uncertainty of pseudo labels, thereby estimating their reliability. + +Considering that the scale and shape of text vary, and characters are usually distributed in order, we migrate Distance-IoU (DIoU) [48] to polygon to measure spatial consistency rather than vanilla IoU. For the $i$ -th matched pairs whose detected regions are represented by polygons $p_i^t$ and $p_i^s$ , the polygon DIoU is formulated as: + +$$ +\mathrm {D I o U} \left(p _ {i} ^ {t}, p _ {i} ^ {s}\right) = \mathrm {I o U} \left(p _ {i} ^ {t}, p _ {i} ^ {s}\right) - \frac {\frac {1}{K} \sum_ {k = 1} ^ {K} E \left(\widetilde {p} _ {(i , k)} ^ {t} , \widetilde {p} _ {(i , k)} ^ {s}\right)}{\max _ {1 \leq m , n \leq 2 K} E \left(\widehat {p} _ {(i , m)} ^ {t} , \widehat {p} _ {(i , n)} ^ {s}\right)}, \tag {2} +$$ + +where $\widetilde{p}_{(i,m)}$ and $\widehat{p}_{(i,n)}$ are the $m$ -th center point and the $n$ -th boundary point of the corresponding region. $K$ is the number of center points within a text instance and $E(*)$ means the Euclidean distance of two points. It integrates the deviations in center points, overlap between polygons, and text instance scale. Thus, we obtain the adaptive factor to adjust the recognition loss of the semi-supervised flow: + +$$ +\alpha_ {i} = 1 + \mathrm {D I o U} \left(p _ {i} ^ {t}, p _ {i} ^ {s}\right). \tag {3} +$$ + +Using the spatial-aware modulating factor, SCI suppresses ambiguous optimization, facilitating the learning process. + +# 4.2.2. Content-aware Region Calibration + +Conversely, the recognition results can help calibrate the predicted text regions. We assume that the discrepancy between text contents predicted by the teacher and student implicitly indicates the deviation of the detected text regions. Therefore, we propose CRC to amplify important supervision signals for detection from recognition results as shown in Fig. 4 (b). Such design lets the optimization focus on the imprecise regions that cause incorrect transcriptions in particular. Specifically, we leverage Levenshtein distance to measure the disparity between words, formulated as: + +$$ +D \left(q _ {i} ^ {t}, q _ {i} ^ {s}\right) = \frac {\text {L e v e n s h t e i n} \left(q _ {i} ^ {t} , q _ {i} ^ {s}\right)}{\max \left(\left| q _ {i} ^ {t} \right| , \left| q _ {i} ^ {s} \right|\right)}, \tag {4} +$$ + +where $q_{i}^{t}$ and $q_{i}^{s}$ are the $i$ -th pair of decoded words from teacher and student, respectively. The Levenshtein distance is normalized by the maximum length of words. Then, the corresponding factor $\beta_{i}$ for regression loss is derived as: + +$$ +\beta_ {i} = \left\{ \begin{array}{l l} 1 + \lambda D \left(q _ {i} ^ {t}, q _ {i} ^ {s}\right), & \text {i f} c _ {i} ^ {t} > \mathcal {T} _ {R} \text {a n d} c _ {i} ^ {t} > c _ {i} ^ {s}, \\ 1, & \text {o t h e r w i s e}, \end{array} \right. \tag {5} +$$ + +where $\lambda$ is a scale factor. CRC is only applied to E2E labels. The content-aware modulating factor highlights the inconsistent detection and explicitly enhances the task synergy. + +# 4.3. Training Objective + +The overall loss $\mathcal{L}$ of SemiETS is formulated as: + +$$ +\mathcal {L} = \omega_ {l} \mathcal {L} _ {l} + \omega_ {u} \mathcal {L} _ {u} ^ {c l s} + \omega_ {u} \left(\beta \cdot \mathcal {L} _ {u} ^ {r e g} + \alpha \cdot \mathcal {L} _ {u} ^ {r e c}\right), \tag {6} +$$ + +where $\mathcal{L}_l$ is the supervised loss inherit from the supervised baseline. $\mathcal{L}_u^{cls}$ , $\mathcal{L}_u^{reg}$ and $\mathcal{L}_u^{rec}$ are semi-supervised classification, regression and recognition losses in the unlabeled data flow, respectively. $\omega_{l}$ and $\omega_{u}$ are the weighting factors. + +# 5. Experiments + +# 5.1. Experimental Settings + +Datasets. We conduct experiments on three widely used scene text datasets, i.e., Total-Text [3], ICDAR 2015 [8] (IC15), and CTW1500 [17]. + +Partially Labeled Data. We randomly sample $0.5\%$ , $1\%$ , $2\%$ , $5\%$ , and $10\%$ images from the training set of each dataset as labeled data and set the remaining images as unlabeled data following the data split in SSOD [40, 45]. The smallest labeled subset only contains 5 samples. + +Fully Labeled Data. We set the training set of Total-Text or IC15 as labeled data and images from TextOCR [30] as additional unlabeled data. + +Domain Adaptation. We set the training set of Total-Text and IC15 as the labeled and unlabeled data, respectively. Models are evaluated on the test set of IC15. + +# 5.2. Implementation Details + +We first pre-trained the text spotter on Synth150K [18] to initialize the model's text perception ability. We use Deep-Solo [44] as our base text spotter. The labeled-to-unlabeled data ratio for Partially Labeled Data settings is 1:2 on Total-Text and CTW1500, and 1:1 on ICDAR 2015. AdamW is the optimizer. $\lambda$ is set to 20, and $\mathcal{T}_R$ is set to 0.7. $\omega_{s}$ and $\omega_{u}$ are set to 1 and 2, respectively. We use the data augmentation strategy from Semi-DETR [45], excluding flipping and polarization to avoid ambiguity in recognized texts. + +# 5.3. Main Results + +# 5.3.1. Partially Labeled Data + +We compare our method to several popular semi-supervised learning methods, including STAC [31], MeanTeacher [34], Soft Teacher [42], Unbiased Teacher v2 [20], and Semi-DETR [45]. We conduct the partially labeled data settings on three popular benchmarks for text spotting. In general, our method achieves the best results on all of them. In addition, its advantage is more pronounced using less proportions of labeled data such as $0.5\%$ . + +Results on Total-Text. As shown in Tab. 1, our framework archives state-of-the-art end-to-end recognition results on arbitrary-shaped texts under all proportions. Especially, when only using $0.5\%$ labeled data, it outperforms the previous SOTA by $8.7\%$ without lexicon. Moreover, the supervised baseline can be improved by $13.2\%$ . Surprisingly, the performance of most EMA-based methods [20, 34, 42] even degrades in low data proportions. We assume it is because models with limited capacity are likely to introduce + +Table 1. The end-to-end spotting results on curved text datasets Total-Text and CTW1500 under the Partially Labeled Data setting. None (Full) denotes the F1-measure without (with) using the lexicon that includes all words in the test set. Experiments are conducted on various data proportions. \* indicates our adaptation of SSOD methods that fit curved texts and the recognition loss from pseudo labels is added. + +
MethodsTotal-TextCTW1500
0.5%1%2%5%10%0.5%1%2%5%10%
NoneFullNoneFullNoneFullNoneFullNoneFullNoneFullNoneFullNoneFullNoneFullNoneFull
Supervised [44]58.871.161.274.963.476.866.978.869.980.930.159.440.666.248.171.253.074.856.976.9
STAC* [31]63.373.766.777.267.377.772.180.973.282.832.760.743.268.049.772.055.276.358.477.5
Mean-Teacher* [34]55.557.764.469.068.673.773.580.673.881.948.871.052.773.155.075.157.877.158.177.3
Soft Teacher* [42]58.460.961.263.868.273.972.478.973.481.341.759.852.571.253.673.458.576.258.977.1
Unbiased Teacher v2* [20]56.658.964.769.570.877.373.380.775.083.547.566.150.468.052.369.356.773.856.574.5
Semi-DETR* [45]59.463.067.273.068.575.571.579.474.282.233.149.738.253.241.355.647.763.657.275.3
SemiETS (Ours)72.078.572.880.673.482.275.483.176.484.853.575.656.276.358.677.660.079.060.979.6
+ +Table 2. The end-to-end spotting results on ICDAR 2015 under the Partially Labeled Data setting. 'S', 'W' and 'G' refer to using strong, weak and generic lexicons, respectively. Experiments are conducted on $0.5\%$ , $1\%$ , $2\%$ , $5\%$ and $10\%$ labeled data settings. + +
Methods0.5%1%2%5%10%
SWGSWGSWGSWGSWG
Supervised [44]71.266.159.969.965.659.872.067.461.677.372.567.379.375.069.2
STAC* [31]76.671.264.475.069.763.776.671.764.880.875.369.681.276.871.2
Mean-Teacher* [34]77.767.960.676.868.561.978.670.563.881.674.067.982.376.070.2
Soft Teacher* [42]76.567.260.476.268.261.877.769.162.581.473.567.181.474.368.4
Unbiased Teacher v2* [20]74.964.757.575.466.359.277.269.161.579.671.765.680.774.067.5
Semi-DETR* [45]79.971.665.380.072.666.281.273.967.582.775.569.383.277.171.1
SemiETS (Ours)81.974.567.882.575.368.882.676.270.084.077.471.284.177.971.8
+ +cumulative error without specified designs for SSTS. + +Results on ICDAR 2015. SemiETS also achieves the best results under all proportions on multi-oriented scene text as presented in Tab. 2, indicating its effectiveness in complex scenarios. Specifically, it surpasses the SOTA methods by $2.5\%$ , $2.6\%$ , and $2.5\%$ H-mean under $0.5\%$ , $1\%$ , and $2\%$ labeled data settings using generic lexicon, respectively. + +Results on CTW1500. The results in Tab. 1 also demonstrate the consistent superiority of SemiETS in spotting curved text at the text-line level. Even only using 5 labeled images, SemiETS achieves decent results, bringing $23.4\%$ performance gain to the supervised baseline in E2E (None). + +# 5.3.2. Fully Labeled Data + +Under this setting, we can explore the potential of the proposed semi-supervised framework even if it has been trained with extensive labeled data. As shown in Tab. 3, although the supervised baseline has already achieved a high performance of $79.7\%$ H-mean in E2E (None) on Total-Text, our SemiETS still takes $2\%$ improvements on it. In addition, our method performs better than other semi-supervised methods with the same baseline, especially regarding E2E results. Similarly, SemiETS also outperforms them on IC15, but the improvements are smaller due to a larger domain shift. It indicates the promising potential of SemiETS to utilize unlabeled data further. Notably, recent generalist VLLMs lack of text spotting ability, indicating the value of our work in the era of VLLMs. + +# 5.3.3. Generalization of our method + +Generalization to text spotters. We also conduct experiments on ABCNet [18], a representative RoI-based text spotter. As shown in Tab. 4, SemiETS also brings significant performance gain to the supervised baseline on different text spotting baselines. The proposed strategies achieve consistent improvements over the baseline SSL framework, demonstrating that SemiETS is compatible with both RoI-based and DETR-based text spotters. + +Domain Adaptation. We further study the domain adaptation ability using the training set of Total-Text and IC15 as the labeled and unlabeled data, respectively. From Tab. 5, the generalization ability of the supervised baseline is limited. Our SemiETS can effectively adapt the text spotter to new scenarios without extra labeling costs, facilitating detection and E2E results by $17.3\%$ and $14.4\%$ (generic lexicon), respectively. Moreover, it surpasses other SSL frameworks by a large margin, showing great practical value. + +# 5.4. Ablation Study + +In this section, we conduct extensive studies to validate our designs. We use DeepSolo [44] as the baseline text spotter. + +# 5.4.1. Effect of each component + +We conduct the experiments using various data proportions on Total-Text, demonstrating the effectiveness of the proposed designs shown in Tab. 6. Each component can provide improvements to the baseline in both detection and recognition. For example, under $2\%$ data proportion, PSA + +Table 3. Results under Fully Labeled Data setting on Total-Text and ICDAR 2015. + +
ParadigmsMethodsTotal-TextICDAR2015
Det-F1NoneFullDet-F1SWG
SupervisedABCNet v2 [21]87.070.478.188.182.778.573.0
GLASS [28]88.179.986.285.784.780.176.3
TESTR [46]86.973.383.990.085.279.473.6
SwinTextSpotter [6]88.074.384.1-83.977.370.5
TTS (poly) [9]-78.286.3-85.281.777.4
DeepSolo [44]87.379.787.090.086.881.976.9
ESTextSpotter [7]90.080.887.191.087.583.078.1
VLLMsInternVL2-8B [2]0.30.00.10.10.00.00.0
Qwen2VL2-VL-7B [38]1.80.61.416.3315.614.413.3
Semi-supervisedSupervised baseline87.379.787.090.086.881.976.9
Mean-Teacher* [34]86.9(-0.4)80.1(+0.4)87.1(+0.1)89.4(-0.6)86.7(-0.1)81.2(-0.7)76.8(-0.1)
Soft-Teacher* [42]86.1(-1.2)80.6(+0.9)86.7(-0.3)87.9(-2.1)86.2(-0.6)78.9(-3.0)74.0(-2.9)
Unbiased Teacher v2* [20]87.5(+0.2)80.3(+0.6)87.3(+0.3)88.5(-1.5)85.2(-1.6)79.7(-2.2)75.1(-1.8)
Semi-DETR* [45]84.0(-3.3)79.6(-0.1)85.5(-1.5)87.9(-2.1)85.4(-1.4)79.6(-2.3)75.0(-1.9)
SemiETS (Ours)87.5(+0.2)81.7(+2.0)87.6(+0.6)90.0(+0.0)87.0(+0.2)81.6(-0.3)77.0(+0.1)
+ +Table 4. The end-to-end spotting results on Total-Text under Partially Labeled Data setting with different text spotters. + +
SpottersSettings0.5%1%2%5%10%
NoneFullNoneFullNoneFullNoneFullNoneFull
ABCNet [18]Supervised40.960.143.763.647.666.050.470.154.772.9
SSL Baseline46.161.448.164.052.266.953.170.857.374.6
SemiETS54.271.956.074.457.774.859.276.561.878.7
DeepSolo [44]Supervised58.871.161.274.963.476.866.978.869.980.9
SSL Baseline55.557.764.469.068.673.773.580.673.881.9
SemiETS72.078.572.880.673.482.275.483.176.484.8
+ +Table 5. Results of domain adaptation. TT is short for Total-Text. + +
MethodsDlDuDetectionE2EWord Spotting
PRF1SWGSWG
Supervised--64.829.940.938.936.134.038.535.933.7
SupervisedTT-71.363.967.466.861.857.566.362.257.8
STAC*TTIC1565.062.763.862.556.551.062.256.751.0
Mean-Teacher*TTIC1576.189.582.382.271.765.281.772.565.7
Soft Teacher*TTIC1577.188.082.283.474.767.883.075.768.5
UT v2*TTIC1569.581.074.876.267.161.475.968.062.0
Semi-DETR*TTIC1574.586.279.980.072.867.579.674.068.5
SemiETSTTIC1581.788.084.784.977.871.984.578.872.7
+ +Table 6. The effects of each component. The experiments are conducted on the Partially Labeled Data setting using Total-Text. + +
Settings2%5%
DetectionE2EDetectionE2E
PSAMMSPRF1NoneFullPRF1NoneFull
96.258.272.568.673.795.071.481.573.580.6
✓94.171.681.372.480.493.676.684.374.383.2
✓95.266.478.271.378.594.472.982.373.781.5
✓✓93.574.282.773.482.294.277.985.275.183.7
+ +significantly improves the detection recall and F1 score by $13.4\%$ and $8.8\%$ , respectively. The E2E H-mean increases by $3.8\%$ $(6.7\%)$ with (without) lexicon, respectively. On this basis, MMS can further bring performance gain, achieving the best result on text detection and spotting. Consistent improvements are verified in $5\%$ data proportion. Unless specified, the following experiments are conducted using $2\%$ labeled data setting on Total-Text. + +Table 7. Ablation study on Progressive Sample Assignment. + +
SettingsDetectionE2E
HM\( \mathcal{T}_R \)CCPRF1NoneFull
96.258.272.568.673.7
✓95.761.474.870.175.8
✓95.863.476.371.877.3
✓✓95.066.177.972.178.1
✓93.871.681.170.679.7
✓✓93.570.780.572.279.8
✓✓93.871.781.372.080.3
✓✓✓94.171.681.372.480.4
+ +Table 8. Ablation study on Mutual Mining Strategy. + +
SettingsDetectionE2E
SCICRCPRF1NoneFull
94.171.681.372.480.4
✓93.173.682.273.381.9
✓93.473.782.473.381.9
✓✓93.574.282.773.482.2
+ +Table 9. Ablation study on the detailed designs in the MMS +(a) The measurements in SCI. + +
SettingsDetectionE2E
PRF1NoneFull
w/o SCI94.171.681.372.480.4
IoU93.872.281.672.881.0
D IoU93.173.682.273.381.9
+ +(b) The measurements in CRC. + +
SettingsDetectionE2E
PRF1NoneFull
w/o CRC94.171.681.372.480.4
DIoU93.372.281.472.980.8
Levenshtein93.473.782.473.381.9
+ +# 5.4.2. Analysis on PSA + +We remove MMS to study the details of PSA. As shown in Tab. 7, both $\mathcal{T}_r$ and CC can boost the performance individually, and using both simultaneously can further improve performance, indicating the hierarchical label assignment can reduce noisy recognition labels. Specifically, the detection F1 score and E2E H-mean increase by $5.4\%$ and $3.5\%$ , respectively. HM can further enhance model performance, especially in detection recall. Due to the complementary nature of the recognition and detection tasks, both tasks benefit concurrently. Integrating these all designs lead to optimal overall performance, particularly on E2E. + +![](images/b4a00949068dca34f7939d61c1b93d97d760ad0fbe226ce70bccf88cce3eb059.jpg) +Figure 5. Statistical analysis on the relationship of the accuracy of detected regions and text similarity. + +![](images/e0de95842024ac60b23af65386e0b11d3dacccd4f26f54eda4c78cf118e33784.jpg) + +![](images/2f0b811fbb50704b3fc6bf5c3ad4062c5f5398ce35fe1a417c680165eb68caa7.jpg) +(a) Ground truth +Figure 6. Some typical qualitative results on Total-Text. True positives are indicated in green. Instances in blue are detected correctly but recognized wrongly, while red are falsely detected. + +![](images/f16f089c7a0d0ddc333167c1b6319f06c64d9d41a247995e9d4cec0c6b441cc3.jpg) +(b) Supervised baseline + +![](images/45c9ff438a7a247342901ef1f176540a1a86f8fba9089b0a5c376f792e855858.jpg) +(c) Semi-DETR + +![](images/a9b9b7494664a6e727ce100153145eb350ed4f4ecd285b262b893073381acbbd.jpg) +(d) SemiETS +Figure 7. Visualizations of (a) usage of pseudo labels; (b) attention maps of queries in DeepSolo [44] and corresponding predictions. + +# 5.4.3. Analysis on Mutual Mining Strategy + +We first discuss two key designs in MMS. PSA is equipped to ensure the quality of pseudo labels. As shown in Tab. 8. Both SCI and CRC can boost detection and recognition results while combining them leads to the optimal. + +Analysis of SCI. The statistical analysis shown in Fig. 5 (a) reveals a strong positive correlation between the region deviation and the average text similarity. This proves that SCI can estimate the reliability of pseudo labels for recognition. Furthermore, we compare the spatial descriptions in Tab. 9a. Our polygon DIoU offers more benefits than vanilla IoU, indicating its advantage in representing text regions. + +Analysis of CRC. CRC aims to rectify the imprecise text regions with recognition results. The positive correlation between text similarity and the region alignment measured by IoU shown in Fig. 5 (b) supports our design through statistical analysis. Yet, another approach is directly using the spatial deviation of the predictions from the teacher and student to assess the quality of regions. We use the normalized DIoU to describe their deviation. Our approach performs better, demonstrating the superiority of mutual mining that effectively boosts task synergy. + +# 5.5. Visualization Analysis + +Qualitative results. We visualize typical qualitative results in Fig. 6. The first row shows a curved text case, while the second presents a dense text case with varied font sizes. Compared to the supervised baseline and SOTA SSL method, SemiETS achieves higher accuracy and recall with fewer false positives in both scenarios. We believe this is because SemiETS extracts richer information from unlabeled data through hierarchical label assignment and inter + +![](images/ca691ffac3edafa2dc3a03a88238259d57dcf4c6e6a739cfd7ec3062eabd9764.jpg) + +![](images/1d22e8fdefdc6cd2c64f692b0d972ad411d92caf694291b11b1edcbf7379a7d1.jpg) +Semi-DETR (CVPR 2023) +SemiETS + +![](images/7cd1ac7cffa85cbd6c14bfae9583d18ff5424e78dd3b539ab9cf008f7fcac3c2.jpg) + +![](images/32c5c5a17fc5b4b6a35d78ab68fcf87b74a5ffea4d452d47032e7564b336ec1e.jpg) +(a) Comparison of pseudo labels +Teacher +(b) Comparison of predictions and attention maps + +![](images/894a021cc167ce4632bb116545b35dd099e90e44d66f1639248b4420f2f49baa.jpg) + +![](images/227a7a32f57b0a1dfbc445ca1ed3d974914c6bd13f263906faa86672271a696e.jpg) +Student + +task complementarity, highlighting its effectiveness. + +Hierarchical v.s. unitary labels. We visualize the used pseudo labels in Fig. 7 (a). In Semi-DETR, the simple label assignment approach introduces noisy recognition supervisions (red), which misleads the recognition flow. In contrast, SemiETS progressively differentiates between Det-only labels (blue) and E2E labels (green), reducing noise and enabling a more reliable label assignment. + +Implicit correlation. Although the detection and recognition do not have an explicit sequential order in DETR-based spotters, visualization analysis in Fig. 7 (b) still supports this property. Most detected regions are highly consistent with their attentions that directly influence the recognition predictions, such as regions in red-dashed circles. + +# 5.6. Limitations + +Although our method achieves promising results, the usage of the characteristics of texts and the synergy of text detection and recognition is to be explored more deeply. At present, SemiETS mainly focuses on spotting Latin text. For our future work, we will further explore semi-supervised text spotting in multilingual scenarios. + +# 6. Conclusion + +The paper presents a straightforward yet effective framework named SemiETS for semi-supervised end-to-end text spotting. Focusing on the challenges of inconsistent pseudo labels, we customize the Progressive Sample Assignment module and Mutual Mining Strategy. The former enhances the quality of pseudo labels for text spotting. The latter introduces a crossover strategy to excavate information using the complementarity of text detection and recognition. Compared with the baseline, SemiETS obtains a remarkable improvement and outperforms existing SSOD approaches on several datasets, with extensive experiments demonstrating its generalization and scalability. + +# Acknowledgments + +This work was supported by NSFC (Grant No.62225603, 62206104). + +# References + +[1] Jeonghun Baek, Yusuke Matsui, and Kiyoharu Aizawa. What if we only use real datasets for scene text recognition? toward scene text recognition with fewer labels. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3113-3122, 2021. 1, 3 +[2] 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 IEEE Conf. Comput. Vis. Pattern Recog., pages 24185–24198, 2024. 7 +[3] Chee-Kheng Chng, Chee Seng Chan, and Cheng-Lin Liu. Total-text: toward orientation robustness in scene text detection. Int. J. Document Anal. Recognit., 23(1):31-52, 2020. 5 +[4] Yunze Gao, Yingying Chen, Jinqiao Wang, and Hanqing Lu. Semi-supervised scene text recognition. IEEE Trans. Image Process., 30:3005-3016, 2021. 3 +[5] Wei Hua, Dingkang Liang, Jingyu Li, Xiaolong Liu, Zhikang Zou, Xiaoqing Ye, and Xiang Bai. Sood: Towards semi-supervised oriented object detection. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15558-15567, 2023. 2 +[6] Mingxin Huang, Yuliang Liu, Zhenghao Peng, Chongyu Liu, Dahua Lin, Shenggao Zhu, Nicholas Jing Yuan, Kai Ding, and Lianwen Jin. Swintextspotter: Scene text spotting via better synergy between text detection and text recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4583-4593, 2022. 1, 2, 7 +[7] Mingxin Huang, Jiaxin Zhang, Dezhi Peng, Hao Lu, Can Huang, Yuliang Liu, Xiang Bai, and Lianwen Jin. Estextspotter: Towards better scene text spotting with explicit synergy in transformer. In Int. Conf. Comput. Vis., pages 19438-19448, 2023. 1, 2, 7 +[8] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In Int. Conf. Doc. Anal. Recognit., pages 1156-1160. IEEE, 2015. 5 +[9] Yair Kittenplon, Inbal Lavi, Sharon Fogel, Yarin Bar, R. Manmatha, and Pietro Perona. Towards weakly-supervised text spotting using a multi-task transformer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4594-4603, 2022. 2, 7 +[10] Gang Li, Xiang Li, Yujie Wang, Yichao Wu, Ding Liang, and Shanshan Zhang. Pseudo labeling and consistency training for semi-supervised object detection. In Eur. Conf. Comput. Vis., pages 457-472, 2022. 2 +[11] Hui Li, Peng Wang, and Chunhua Shen. Towards end-to-end text spotting with convolutional recurrent neural networks. In Int. Conf. Comput. Vis., pages 5238-5246, 2017. 2 +[12] Xiaoyu Li, Xiaoxue Chen, Zuming Huang, Lele Xie, Jingdong Chen, and Ming Yang. Fine-grained pseudo labels for scene text recognition. In ACM Int. Conf. Multimedia, pages 5786-5795, 2023. 1, 3 +[13] Minghui Liao, Guan Pang, Jing Huang, Tal Hassner, and Xiang Bai. Mask textspotter v3: Segmentation proposal net + +work for robust scene text spotting. In Eur. Conf. Comput. Vis., pages 706-722, 2020. 2 +[14] Minghui Liao, Pengyuan Lyu, Minghang He, Cong Yao, Wenhao Wu, and Xiang Bai. Mask textspotter: An end-to-end trainable neural network for spotting text with arbitrary shapes. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):532-548, 2021. 2 +[15] Chang Liu, Weiming Zhang, Xiangru Lin, Wei Zhang, Xiao Tan, Junyu Han, Xiaomao Li, Errui Ding, and Jingdong Wang. Ambiguity-resistant semi-supervised learning for dense object detection. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15579-15588, 2023. 1, 2 +[16] Juhua Liu, Qihuang Zhong, Yuan Yuan, Hai Su, and Bo Du. Semitext: Scene text detection with semi-supervised learning. Neurocomputing, 407:343-353, 2020. 1, 2 +[17] Yuliang Liu, Lianwen Jin, Shuaiqiao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse and longitudinal sequence connection. Pattern Recognition, 90: 337-345, 2019. 5 +[18] Yuliang Liu, Hao Chen, Chunhua Shen, Tong He, Lianwen Jin, and Liangwei Wang. Abcnet: Real-time scene text spotting with adaptiveBezier-curve network. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9806-9815, 2020. 1, 2, 5, 6, 7 +[19] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In Int. Conf. Learn. Represent., 2021. 2 +[20] Yen-Cheng Liu, Chih-Yao Ma, and Zsolt Kira. Unbiased teacher v2: Semi-supervised object detection for anchor-free and anchor-based detectors. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9809-9818, 2022. 2, 5, 6, 7 +[21] Yuliang Liu, Chunhua Shen, Lianwen Jin, Tong He, Peng Chen, Chongyu Liu, and Hao Chen. Abcnet v2: AdaptiveBezier-curve network for real-time end-to-end text spotting. IEEE Trans. Pattern Anal. Mach. Intell., 44(11):8048-8064, 2022. 2, 7 +[22] Yuliang Liu, Jiaxin Zhang, Dezhi Peng, Mingxin Huang, Xinyu Wang, Jingqun Tang, Can Huang, Dahua Lin, Chunhua Shen, Xiang Bai, and Lianwen Jin. SPTS v2: Single-point scene text spotting. IEEE Trans. Pattern Anal. Mach. Intell., 45(12):15665-15679, 2023. 2 +[23] Pengyuan Lyu, Minghui Liao, Cong Yao, Wenhao Wu, and Xiang Bai. Mask textspotter: An end-to-end trainable neural network for spotting text with arbitrary shapes. In Eur. Conf. Comput. Vis., pages 71-88, 2018. 1, 2 +[24] Gaurav Patel, Jan P. Allebach, and Qiang Qiu. Seq-ups: Sequential uncertainty-aware pseudo-label selection for semisupervised text recognition. In Winter Conf. Appl. Comput. Vis., pages 6169-6179, 2023. 1, 3 +[25] Dezhi Peng, Xinyu Wang, Yuliang Liu, Jiaxin Zhang, Mingxin Huang, Songxuan Lai, Jing Li, Shenggao Zhu, Dahua Lin, Chunhua Shen, Xiang Bai, and Lianwen Jin. SPTS: single-point text spotting. In ACM Int. Conf. Multimedia, pages 4272-4281, 2022. 2 +[26] Liang Qiao, Ying Chen, Zhanzhan Cheng, Yunlu Xu, Yi Niu, Shiliang Pu, and Fei Wu. MANGO: A mask attention guided + +one-stage scene text spotter. In AAAI, pages 2467-2476, 2021. 2 +[27] Xugong Qin, Yu Zhou, Dongbao Yang, and Weiping Wang. Curved text detection in natural scene images with semi- and weakly-supervised learning. In Int. Conf. Doc. Anal. Recognit., pages 559–564, 2019. 1, 2 +[28] Roi Ronen, Shahar Tsiper, Oron Anschel, Inbal Lavi, Amir Markovitz, and R Manmatha. Glass: Global to local attention for scene-text spotting. In Eur. Conf. Comput. Vis., pages 249-266, 2022. 7 +[29] Tahira Shehzadi, Khurram Azeem Hashmi, Didier Stricker, and Muhammad Zeshan Afzal. Sparse semi-detr: Sparse learnable queries for semi-supervised object detection. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5840-5850, 2024. 1, 2 +[30] Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. Textocr: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8802-8812, 2021. 5 +[31] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 1, 2, 5, 6 +[32] Yipeng Sun, Jiaming Liu, Wei Liu, Junyu Han, Errui Ding, and Jingtuo Liu. Chinese street view text: Large-scale Chinese text reading with partially supervised learning. In Int. Conf. Comput. Vis., pages 9085-9094, 2019. 2 +[33] Yihe Tang, Weifeng Chen, Yijun Luo, and Yuting Zhang. Humble teachers teach better students for semi-supervised object detection. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3132-3141, 2021. 2 +[34] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Adv. Neural Inform. Process. Syst., 30, 2017. 1, 2, 3, 5, 6, 7 +[35] Shangxuan Tian, Shijian Lu, and Chongshou Li. Wetext: Scene text detection under weak supervision. In Int. Conf. Comput. Vis., pages 1501-1509, 2017. 1, 2 +[36] Zichen Tian, Chuhui Xue, Jingyi Zhang, and Shijian Lu. Domain adaptive scene text detection via subcategorization. arXiv preprint arXiv:2212.00377, 2022. 1, 2 +[37] Hao Wang, Pu Lu, Hui Zhang, Mingkun Yang, Xiang Bai, Yongchao Xu, Mengchao He, Yongpan Wang, and Wenyu Liu. All you need is boundary: Toward arbitrary-shaped text spotting. In AAI, pages 12160-12167, 2020. 2 +[38] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 7 +[39] Wenhai Wang, Enze Xie, Xiang Li, Xuebo Liu, Ding Liang, Zhibo Yang, Tong Lu, and Chunhua Shen. PAN++: towards efficient and accurate end-to-end spotting of arbitrarily-shaped text. IEEE Trans. Pattern Anal. Mach. Intell., 44(9): 5349-5367, 2022. 2 +[40] Xinjiang Wang, Xingyi Yang, Shilong Zhang, Yijiang Li, Litong Feng, Shijie Fang, Chengqi Lyu, Kai Chen, and + +Wayne Zhang. Consistent-teacher: Towards reducing inconsistent pseudo-targets in semi-supervised object detection. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3240-3249, 2023. 1, 2, 5 +[41] Jingjing Wu, Zhengyao Fang, Pengyuan Lyu, Chengquan Zhang, Fanglin Chen, Guangming Lu, and Wenjie Pei. We-cromcl: Weakly supervised cross-modality contrastive learning for transcription-only supervised text spotting. In Eur. Conf. Comput. Vis., pages 289-306, 2025. 2 +[42] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In Int. Conf. Comput. Vis., pages 3060-3069, 2021. 2, 5, 6, 7 +[43] Mingkun Yang, Biao Yang, Minghui Liao, Yingying Zhu, and Xiang Bai. Sequential visual and semantic consistency for semi-supervised text recognition. Pattern Recognit. Lett., 178:174-180, 2024. 2, 3 +[44] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Tongliang Liu, Bo Du, and Dacheng Tao. Deepsolo: Let transformer decoder with explicit points solo for text spotting. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19348-19357, 2023. 1, 2, 3, 5, 6, 7, 8 +[45] Jiacheng Zhang, Xiangru Lin, Wei Zhang, Kuo Wang, Xiao Tan, Junyu Han, Errui Ding, Jingdong Wang, and Guanbin Li. Semi-detr: Semi-supervised object detection with detection transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 23809-23818, 2023. 1, 2, 4, 5, 6, 7 +[46] Xiang Zhang, Yongwen Su, Subarna Tripathi, and Zhuowen Tu. Text spotting transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9509-9518, 2022. 1, 2, 7 +[47] Caiyuan Zheng, Hui Li, Seon-Min Rhee, Seungju Han, Jae-Joon Han, and Peng Wang. Pushing the performance limit of scene text recognizer without human annotation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 14096-14105, 2022. 1, 2, 3 +[48] 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 AAAI, pages 12993-13000, 2020. 5 +[49] Hongyu Zhou, Zheng Ge, Songtao Liu, Weixin Mao, Zeming Li, Haiyan Yu, and Jian Sun. Dense teacher: Dense pseudolabels for semi-supervised object detection. In Eur. Conf. Comput. Vis., pages 35-50, 2022. 2 +[50] Yu Zhou, Hongtao Xie, Shancheng Fang, and Yongdong Zhang. Semi-supervised text detection with accurate pseudolabels. IEEE Sign. Process. Letters, 29:1272-1276, 2022. 1, 2 \ No newline at end of file diff --git a/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/images.zip b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e241e14f234f1cb179a1190b6e46501a2654b2ce --- /dev/null +++ b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28afda58ef8d15e6f511dbc42e0c204cf90724f4d5004b19d8351ec42e94b617 +size 733825 diff --git a/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/layout.json b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf80891b5844349f68d57eecef343de0b60cfcd7 --- /dev/null +++ b/CVPR/2025/SemiETS_ Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb265072fa88f3528d1ed73d5f9850cc152888e63665eb1dcccef0505c4e8af9 +size 480251 diff --git a/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/a68a581f-a077-45a9-9839-521859c5961e_content_list.json b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/a68a581f-a077-45a9-9839-521859c5961e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..809ffebae222598c647c4deaa3d2d5e49a9da997 --- /dev/null +++ b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/a68a581f-a077-45a9-9839-521859c5961e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97c81ade455028de50ddcb5308dea71d759f317057fed2626a90249de3bc5627 +size 91257 diff --git a/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/a68a581f-a077-45a9-9839-521859c5961e_model.json b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/a68a581f-a077-45a9-9839-521859c5961e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..799c96d6ef43aa4393aa6b960fc4352acaf913af --- /dev/null +++ b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/a68a581f-a077-45a9-9839-521859c5961e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82f6bb36efe3036339ea6e0ed3e39784812f96acfedf7467d730d7fe6b971b60 +size 111107 diff --git a/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/a68a581f-a077-45a9-9839-521859c5961e_origin.pdf b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/a68a581f-a077-45a9-9839-521859c5961e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e5abc40919105e4b1f2ed05cdff14cc10ab2ce2a --- /dev/null +++ b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/a68a581f-a077-45a9-9839-521859c5961e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56e185d2838128650f6b35791eb667eaa8fbbf1246fe5811dea428caef873ef4 +size 2502010 diff --git a/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/full.md b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4ceeb05652a8fa7b8f39d447c7e16c9facf66bfe --- /dev/null +++ b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/full.md @@ -0,0 +1,368 @@ +# Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation + +Tianran Chen $^{1}$ , Jiarui Chen $^{1}$ , Baoquan Zhang $^{1*}$ , Zhehao Yu $^{1}$ , Shidong Chen $^{1}$ , Rui Ye $^{2}$ , Xutao Li $^{1}$ , Yunming Ye $^{1}$ $^{1}$ Harbin Institute of Technology, Shenzhen $^{2}$ Guizhou Power Grid Co., Ltd. Intelligent Operation Center + +{23S151121, jiaruichen, 210110629, chenshidong}@stu.hit.edu.cn {baoquanzhang, lixutao, yeyunming}@hit.edu.cn, {yeruigzyj}@163.com + +# Abstract + +Parameter-Efficient Fine-Tuning (PEFT) is a fundamental research problem in computer vision, which aims to tune a few parameters for efficient storage and adaptation of pre-trained vision models. Recently, sensitivity-aware parameter efficient fine-tuning method (SPT) addresses this problem by identifying sensitive parameters and then leveraging its sparse characteristic to combine unstructured and structured tuning for PEFT. However, existing methods only focus on the sparse characteristic of sensitive parameters but overlook its distribution characteristic, which results in additional storage burden and limited performance improvement. In this paper, we find that the distribution of sensitive parameters is not chaotic, but concentrates on a small number of rows or columns in each parameter matrix. Inspired by this fact, we propose a Compact Dynamic-Rank Adaptation-based tuning method for Sensitivity-aware Parameter efficient fine-Tuning, called CDRA-SPT. Specifically, we first identify the sensitive parameters that require tuning for each downstream task. Then, we reorganize the sensitive parameters by following its row and column into a compact sub-parameter matrix. Finally, a dynamic-rank adaptation is designed and applied at sub-parameter matrix level for PEFT. Its advantage is that the dynamic-rank characteristic of sub-parameter matrix can be fully exploited for PEFT. Extensive experiments show that our method achieves superior performance over previous state-of-the-art methods. + +# 1. Introduction + +With the development of large vision models on representation and generation, the paradigm of pre-training and fine-tuning has attracted increasing interest on computer vision [5, 15, 22, 25, 41, 42]. A widely used fine-tuning technique is the full fine-tuning, which leverages the pre-trained weights as the initial parameters and then tunes all param + +![](images/1d67f891d448e78fb9a7612b07ae585371c81dc2a84cdb09ee7e5e22f5d7844a.jpg) +(a) Original Parameters. + +![](images/76cbbcb17aa30ce8cf64254693d697697294c676cca46ab7bffd6111f91129bc.jpg) +(b) Reorganized Parameters. +Figure 1. Distribution illustration of of sensitive parameter. In fact, the distribution of sensitive parameters is not chaotic on origin parameter but concentrates in a small number of rows or columns (a). After reorganizing its row and column, these sensitive parameters will be clustered into a compact sub-parameter matrix (b). Inspired by this fact, we propose to apply structured tuning at sub-parameter matrix levels instead of entire matrix. Its advantage is the distribution character of sensitive parameters concentrating in a small of row and column can be finely leveraged for PEFT. + +eters for adaptation of downstream tasks. However, as the number of model parameters increases, the full fine-tuning method needs to store many parameters and takes lots of computing power to tune pre-trained vision models for each downstream task. This results in these methods are difficult to apply to resource-limited scenarios. To address this problem, parameter-efficient fine-tuning (PEFT) is proposed and has received wide attention recently. PEFT, as a fundamental problem in computer vision, aims to achieve an efficient transfer from a pre-trained vision model to target downstream tasks by freezing most of the pre-trained parameters and only tuning a few model parameters [29, 31]. + +Existing studies [4, 6, 31, 43, 48] effectively address this problem by selecting a subset of parameters (called unstructured tuning) [4, 6, 31, 43, 48] or adding a few additional parameters (called structured tuning) [4, 6, 31, 43, 48] to perform adaptation of downstream tasks. Although these methods are all different, He et al. [14] recently found that these unstructured and structured tuning methods can combine and then present sensitivity-aware parameter efficient fine-tuning methods (called SPT [14]) for enhancing PEFT + +performance. Their main idea is to identify sensitive parameters and then tune these parameters in a manual combination manner between unstructured and structured tuning [14]. Such a design can fully exploit the sparse characteristic of sensitive parameters, such that the finetuning performance can obtain significant improvement. + +In this paper, we also focus on sensitivity-aware efficient fine-tuning, but argue the distribution characteristic of sensitive parameters is not fully explored and utilized in existing SPT [14]. This is because existing SPT [14] methods evaluate sensitivity at the level of the entire parameter matrix (see Figure 1a), which only focuses on the sparse characteristic of sensitive parameters, but overlooks its distribution characteristic at the row and column vector levels. This results in additional storage burden and limited performance improvement. In fact, as shown in Figure 1a, we find that the distribution of sensitive parameters is not chaotic on entire parameter matrix but concentrates in a small number of rows or columns in the parameter matrix. Based on this fact, our intuitive idea is can we further improve the efficiency of SPT based on such distribution characteristics? To this end, we propose a novel compact dynamic-rank adaptation-based tuning method for sensitivity-aware efficient fine-tuning, which reorganizes the sensitive row and column parameters as a compact sub-parameter matrix (see Figure 1b) and designs a compact dynamic-rank adaptation for SPT, called CDRA-SPT. Specifically, we first quickly identifies and organizes the sensitive row and column parameters as a compact sub-parameter matrix. After that, we design a dynamic rank allocation strategy for sub-parameter matrix adaptation. Its advantage is that the dynamic-rank characteristic of sub-parameter matrix can be fully exploited for enhancing fine-tuning performance on downstream tasks. Our main contributions can be summarized as follows: + +- We find a new distribution characteristic of sensitive parameters overlooked in previous existing studies, i.e., the distribution of sensitive parameters is actually not chaotic on the entire parameter matrix, but concentrates in a small number of rows or columns in the parameter matrix. +- Resorting such distribution characteristic, we design a novel compact dynamic-rank adaptation-based tuning method to replace the combination fine-tuning between unstructured and structured tuning. Its advantage is that the relationship and sparse distribution between parameters can be fully leveraged to enhance PEFT. +- We conduct various experiments on various datasets, which verify the effectiveness of our CDRA-SFT method. + +# 2. Related Works + +# 2.1. Parameter-Efficient Fine-Tuning + +Parameter-efficient fine-tuning [13, 16, 19, 44] (PEFT) is an important machine learning method, which aims to tune only a tiny portion of parameters to achieve adaptation of pre-trained models to downstream tasks. Depending on the difference of parameter types, existing PEFT methods can be divided into three types. 1) Addition-based PEFT Approaches. This type of approach attaches additional trainable parameters to the backbone and only tunes these parameters for adaptation of pre-trained models [16, 19]. For example, in [19], Jia et al. propose introducing learnable prompts as additional trainable parameters (called VPT) and then freeze the pre-trained weights to perform PEFT in a prompt learning manner. 2) Reparameterization-based PEFT Approaches. This line of methods aims to avoid extra computational costs by tuning parameters that are inherently in or can be reparameterized into the backbone during inference [9, 17, 26, 45]. For instance, in [17], Hu et al. propose a low-rank adaptation (called LoRA) for PEFT, which freezes the pre-trained model weights and injects trainable rank decomposition matrices into each layer of Transformer for achieving adaptation of pre-trained models. With the superiority of LoRA, various LoRA variants are proposed recently, such as DoRA [26], VeRA [23], AdaLoRA [45], LoRA+ [13], FourierFT [9], and HydraLoRA [30]. 3) Selection-based PEFT Approaches. This methods focus on how select a subset from pretrained parameters and tuning them in a masked finetuning manner for adaptation of pre-trained models [14, 37, 47]. In early works, existing studies mainly focus on parameter types to select parameter subset, e.g., selecting bias [37] or last k-layer [16, 19] parameters as parameter subset to tune. Recently, He et al. turn into the perspective of parameter sensitivity to select subset of parameters. For instance, in [14], Hu et al. propose to select a subset of parameters from the perspective of sensitivity and then perform PEFT in a combination manner between unstructured and structured tuning, called SPT. Unlike SPT, Zhang et al. [47] perform PEFT by fining sensitive parameters in a mask-tuning manner. + +In this paper, we mainly focus on selection-based PEFT approaches due to its superiority. However, different from these existing methods, 1) we find a new distribution characteristic of sensitive parameters overlooked in previous existing studies, i.e., the distribution of sensitive parameters is actually not chaotic on entire parameter matrix, but concentrates in a small number of rows or columns in the parameter matrix; and 2) we present a novel compact dynamic-rank adaptation for enhancing sensitivity-aware PEFT methods. + +# 2.2. Sub-Networks Learning + +Sub-network learning [2, 8] is an important model pruning techniques, which aims to identify an important subnetwork from a big model meanwhile remains superiority performance similar to the origin big model. In [8], Frankle et al. propose a lottery ticket hypothesis, which verifies that small sub-networks from big models can achieve the same level of accuracy as the original model. After that, with such a lottery ticket hypothesis theory, various subnetworks learning methods are proposed [12, 35]. For example, in [12], Guo et al. design a SpotTune method that employs a policy network to make routing decisions for sub-networks. Similarly, Xu et al. [35] develop a ChildTuning method, which iteratively updates a subset of parameters by masking certain gradients during the backpropagation process for sub-network learning. With the superiority of sub-network learning, some studies have explored sub-network learning to various domains, such as continual learning [11, 20, 24, 36], multi-task learning [1, 18, 27, 30] and few-shot learning [3, 36, 39, 40]. Recently, several PEFT techniques relying on sub-network learning have been proposed. In [14], a sensitivity-aware parameter efficient fine-tuning method (SPT) is developed based on weight sensitivity, which aims to only finetune a sub-set of sensitive parameters to achieve PEFT. Zhang et al. [47] design a gradient-based parameter selection method (GPS) to select sub-network and then perform PEFT by mask-tuning. + +In this paper, we also focus on exploring sub-networks learning to enhance PEFT. However, different from existing methods that performs PEFT by using mask-tuning manner, we present a novel compact dynamic-rank adaptation method to explore sub-networks learning for PEFT. Its advantage is the distribution character of sensitive subnetworks can be finely characterized for PEFT. + +# 3. Methodology + +# 3.1. Preliminaries on SPT + +SPT [14], as a pioneering work on sensitivity-aware parameter efficient fine-tuning, aims to explore where to introduce and how to allocate trainable parameters for a pretrained models in a sensitivity-aware manner. Formally, let $\{w_0,\dots ,w_i,\dots ,w_N\}$ denotes the pre-trained model weights where $N$ is the number of parameters, and $\mathcal{D}^{train}$ denotes the training set of a given downstream task. The SPT performs the PEFT of downstream by the following two steps: (1) evaluating the sensitivity $s_i$ of each parameter $w_{i}$ by calculating the square of its accumulation gradient $g_{i}$ around 400 training samples from $\mathcal{D}^{train}$ , i.e., $s_i = g_i^2$ , and then selecting the top- $\gamma$ most sensitive weight as the sensitive parameters; and (2) performing the PEFT of downstream tasks in an adaptive combination between unstructured and structured tuning, i.e., using unstructured tuning when the + +![](images/c238ba42c8a564e047b347245d9ba8f8fc2c657e00b7799ae28731f1c1634be0.jpg) +(a) Sparsity analysis at row +Figure 2. Sparsity analysis of sensitive parameter. + +![](images/802b590a1716cce722ce2fb3aed10ba44ceef3a531111a909ae7d4ff5b6e08e2.jpg) +(b) Sparsity analysis at column + +sensitive weight matrices that have a small number of sensitive parameters otherwise structured tuning is employed. Finally, the adapted parameter $w_{i}^{\prime}$ can be expressed as: + +$$ +w _ {i} ^ {\prime} = \left\{ \begin{array}{l l} w _ {i} - \eta g _ {w _ {i}} \odot m _ {i}, & i f \sum_ {i} m _ {i} < \alpha , \\ w _ {i} + w _ {d o w n} w _ {u p}, & \text {o t h e r w i s e .} \end{array} \right. \tag {1} +$$ + +where $\eta$ is a learning rate, $m_{i}$ denotes a binary mask marking sensitive parameters, $g_{w_i}$ is the gradient of parameter $w_{i}$ , $w_{down}w_{up}$ denotes the low-rank adaptation (LoRA) [17], and $\alpha$ is the number of parameters of $w_{down}$ and $w_{up}$ . + +# 3.2. Motivation Analysis on SPT + +Although the existing SPT has shown superior performance, it only focuses on the level of the entire parameter matrix to evaluate the sparse characteristic of sensitive parameters (see Eq. 1), which overlooks its distribution characteristic at the level of row and column vectors. This results in its additional storage burden and limited performance improvement. Next, we conduct a detailed analysis on this point. + +Case 1: Parameter Matrix $w_{i}$ Satisfying $\sum_{i} m_{i} < \alpha$ . As shown in Eq. 1, existing SPT employs a full fine-tuning with sensitivity mask $m_{i}$ to tune such parameters, which is because the sensitive parameters are very sparse. However, in this paper, we find that these sensitive parameters actually are not sparse. The reason it's considered sparse is that existing SPT only focuses on the level of entire parameter matrix to evaluate its sparsity, i.e., comparing the number of sensitive parameters and LoRA parameters of entire parameter matrix. However, these sensitive parameters will become dense when we evaluate its sparsity from the perspective of row and column. This is because the distribution of sensitive parameters is actually not chaotic on entire parameter matrix but concentrates in a small number of rows or columns. To illustrate this, in Figure 2, we randomly select a query parameter matrix $w_{i}$ from pre-trained ViT weights and then show its sparsity (i.e., the number of sensitive parameters divided by the total number of parameters) for each row (see Figure 2a), column (see Figure 2b), and entire matrix (see dotted lines of Figure 2). From results, we can see that the sensitive parameters at some row and column (around $43\%$ and $52\%$ ) are more dense than entire matrix. This inspires us an intuitive idea that can we reor + +![](images/85ac9c41527d71a63693b9b001a2e489fdfc04a015e0291dc5b79f9601827469.jpg) +(a) Existing SPT method. + +![](images/3afceec62df072e0c793e7e84a6f540d5976692a3d5026e2e5a21802a28005bd.jpg) +(b) Our method. +Figure 3. An example illustration of weight change. + +ganize the dense row and column parameters as a compact sub-parameter matrix and then tune the sub-parameter in a low-rank adaptation manner. Its advantage are: (1) using the low-rank adaptation to model the weight change of sensitive parameters can further reduce trainable parameters; and (2) the correlation between sensitive row and column vectors can be fully characterized for enhancing SPT. + +Case 2: Parameter Matrix $w_{i}$ Satisfying $\sum_{i} m_{i} \geq \alpha$ . As shown in Eq. 1, to preserve efficient storage and superior performance, existing SPT employs a structured tuning (e.g. LoRA [17]) to replace unstructured tuning for tuning these parameters. Although such method can ensure its accuracy and efficiency, it also introduces lots of redundant parameters to achieve adaptation of non-sensitive parameters. To illustrate this issue, given a parameter $w_{i}$ , we assume the matrix shown in Figure 3a is its sensitive parameter distribution. For tuning such parameters, LoRA employs two low-rank matrices $A$ and $B$ to model its change, i.e., $\triangle w_{i} = AB$ with same size as $w_{i}$ . As shown in Figure 3a, the weight change actually can be divided into four sub-matrices $(\triangle w_{i}^{00}, \triangle w_{i}^{01}, \triangle w_{i}^{10}, \triangle w_{i}^{11})$ . That is, + +$$ +\triangle w _ {i} = A B = \left[ \begin{array}{c c} \bigtriangleup w _ {i} ^ {0 0} & \bigtriangleup w _ {i} ^ {0 1} \\ \hline \bigtriangleup w _ {i} ^ {1 0} & \bigtriangleup w _ {i} ^ {1 1} \end{array} \right]. \tag {2} +$$ + +In fact, among them, only the low-rank sub-matrix $\triangle w_{i}^{10}$ accounts for modeling the weight change of sensitive row and column parameters, and others low-rank sub-matrixs (i.e., $\triangle w_{i}^{00}, \triangle w_{i}^{01}$ , and $\triangle w_{i}^{11}$ ) models either non-sensitive row parameters or non-sensitive column parameters. This means that the three low-rank sub-matrixs (i.e., $\triangle w_{i}^{00}, \triangle w_{i}^{01}$ , and $\triangle w_{i}^{11}$ ) actually are non-sensitive for downstream tasks. Thus, as shown in Figure 3b, an intuitive idea that whether can we only use the low-rank sub-matrix $\triangle w_{i}^{10}$ to tune the parameter $w_{i}$ ? The advantage of such design is that the trainable parameter will be significantly reduce when we remove $\triangle w_{i}^{00}, \triangle w_{i}^{01}$ , and $\triangle w_{i}^{11}$ such that the over-fitting issue can also be effectively alleviated. + +# 3.3. The Proposed CDRA-SPT + +Based on the above motivation analysis, we find that the above unstructured tuning (i.e., Case 1 described in Section 3.2) and structured tuning (i.e., Case 2 described in Section 3.2) used in SPT [14] can all be unified to a structured tuning applied at the compact sub-parameter matrix + +Algorithm 1 PyTorch-style pseudocode for CDRA-SPT. +class CDRA_SPT(nnModule): + def __init__(self, m_i, in_d, out_d, r): + super(CDMA_SPT, self).__init() + self.bar_d, self.bar_k, self.c_map, self.r_map = $\backslash$ self.reorganize(selfmasked_w, self.d, self.k) + self.d = in_d + self.k = out_d + self.lora_rank = lora_r + self.w = nn_PARAMETER(torch randn(self.d, self.k)) + self.bias = nn_PARAMETER(torch.randint(1, self.k)) + self.A = nn_PARAMETER(torch.randint(self.bar_d, lora_r)) + self.B = nn_PARAMETER(torch.randint(lora_r, self.bar_k)) + # binary mask matrix obtained in Step 1 + self-mask_w = m_i + def reorganize(self, mask_w, in_dim, out_dim): + sensitive_row = torch.sum(mask_w, dim=1) + _, r_map = torch.sort(sensitive_row, descending=True) + sensitive_col = torch.sum(mask_w[r_map], dim=0) + _, c_map = torch.sort(sensitive_col, descending=True) + # Calculate the dimension of compact matrix + row_dim = in_dim + col_dim = out_dim + for i in range(in_dim): + if torch.sum(mask_w[r_map[i], :]) == 0: + row_dim = i + break + for i in range(out_dim): + if torch.sum(mask_w[:, c_map[i]]) == 0: + col_dim = i + break + return col_dim, row_dim, c_map, r_map + def forward(self, x): + # Reorganize in Step2 + bar_w = self.w[self.r_map, :][:, self.c_map] + delta_w = self.A @ self.B + # Eq 6 in Step 3 + bar_w[:self.bar_d, :][:, self.bar_k] += delta_w + w = torch.zeros_like(self.w) + # Reverse the reorganization in Step 3 + w[self.r_map, :][:, self.c_map] = bar_w + return x @ w + self.bias + +level. Based on this fact, we propose a novel compact dynamic-rank adaptation-based tuning method for SPT, called CDRA-SPT. The main idea is that with the row/column distribution characteristic of sensitive parameters, we reorganize the sensitive parameters by following its row and column into a compact sub-parameter matrix, and then perform unified structured tuning at sub-parameter matrix level to enhance existing SPT performance. + +As shown in Algorithm 1, given a pre-trained model $\{w_0,\dots ,w_i,\dots ,w_N\}$ and a downstream task $\tau = \{\mathcal{D}_{tr},\mathcal{D}_{te}\}$ where $\mathcal{D}_{tr}$ and $\mathcal{D}_{te}$ denote its training and test set, our CDRA-SPT achieve PEFT by three steps: + +Step 1. we first evaluate the sensitivity $s_i$ for each parameter $w_i \in \{w_0, \dots, w_i, \dots, w_N\}$ . Here, the sensitivity evaluation manner is similar to SPT, i.e. calculating and regarding its accumulated square gradient from all training samples of downstream tasks as parameter sensitivity (i.e., + +![](images/71ce009521a7c112dc5cb8a98da7d0ce069082cadf4865bb0af4cf63b34323fa.jpg) +(a) Query weight matrix at layer 11. + +![](images/0813cf1c5f48ca463ad5daa9eaa667c818c00253b22b61f9785bfd57581ea5a5.jpg) +(b) FFN weight matrix at layer 11. +Figure 4. Illustration of compact sub-parameter matrices. Here, the red box represents the transformed compact sub-parameter matrices. + +![](images/3aed008d3e1d2fb8c81350f2321bdb4ee8767a77b982c8058a41325e344b7b62.jpg) +(c) FFN weight matrix at layer 10. + +![](images/77d7f724ad5798b93ef47df4d3b131df9fa76760815e6da6ff8ebb47cc2aca6f.jpg) +(d) Value weight matrix at layer 6. + +$s_i = g_i^2$ where $g_{i}$ is its gradient). That is, + +$$ +g _ {i} ^ {2} = \frac {\sum_ {x \in D ^ {t r}} \left(\nabla \mathcal {L} _ {x} \left(w _ {i}\right) ^ {2}\right)}{\| D ^ {t r} \|}, s _ {i} = g _ {i} ^ {2}, \tag {3} +$$ + +where $\|\cdot\|$ is the set size and $\nabla \mathcal{L}_x(w_i)$ denotes the gradients from training sample $x$ . Finally, we select the top- $\gamma$ sensitive weight as the sensitive parameters. In particular, we denote these sensitive parameters for each parameter matrix $w_i$ by using a binary mask $m_i$ where "1" denotes sensitive parameters and "0" denotes non-sensitive parameters. + +Step 2. To fully exploit the distribution character of sensitive parameters, we propose to reorganize each parameter matrix $w_{i}$ in a descending order of importance along its rows and columns. As a result, a new parameter $\bar{w}_{i}$ can be obtained where these sensitive parameters can be clustered into a compact sub-parameter matrix. That is, + +$$ +\bar {w} _ {i} = \operatorname {R e o r g a n i z e} \left(w _ {i}\right) = \left[ \begin{array}{c c} \bar {w} _ {i} ^ {0 0} & \bar {w} _ {i} ^ {0 1} \\ \hline \bar {w} _ {i} ^ {1 0} & \bar {w} _ {i} ^ {1 1} \end{array} \right], \tag {4} +$$ + +where $Reorganize()$ denotes parameter transformation and all sensitive parameters will be clustered to sub-parameter matrix $\bar{w}_i^{10}$ . Note that the parameter transformation $Reorganize()$ is only the reordering of rows and columns, which is inverse. Thus, the original parameter $w_i$ can be again obtained by using a reverse transformation. That is, + +$$ +w _ {i} = \text {R e o r g a n i z e} ^ {- 1} (\bar {w} _ {i}), \tag {5} +$$ + +where $Reorganize^{-1}()$ is the reverse transformation. + +Step 3. Considered that all sensitive parameters are distributed in sub-parameter matrix $\bar{w}_i^{10}$ , instead of performing PEFT on the entire parameter matrix, we propose to directly perform low-rank adaptation on the sub-parameter matrix $\bar{w}_i^{10}$ to model its weight change $\triangle \bar{w}_i$ . As a result, an adapted parameter matrix $\bar{w}_i^{\prime}$ can be obtained. That is, + +$$ +\bar {w} _ {i} ^ {\prime} = \bar {w} _ {i} + \triangle \bar {w} _ {i} = \left[ \begin{array}{c c} \bar {w} _ {i} ^ {0 0} & \bar {w} _ {i} ^ {0 1} \\ \hline \bar {w} _ {i} ^ {1 0} + A B & \bar {w} _ {i} ^ {1 1} \end{array} \right], \tag {6} +$$ + +where $A$ and $B$ denotes its learnable low-rank matrices. Finally, we employ the reverse transformation described in Eq. 5 to achieve the adapted parameters $w_{i}^{\prime}$ in original parameter space and then perform PEFT of downstream task. + +# 3.4. Dynamic-Rank Allocation Strategy + +Until now, we have introduced all framework details of our CDRA-SPT, a remaining problem is: how to allocate the rank for low-rank matrices $A \in \mathcal{R}^{d^i \times r}$ and $B \in \mathcal{R}^{r \times k^i}$ used to model the weight change of sub-parameter matrix $\bar{w}_i^{10}$ . A direct and simple allocation method is following existing low-rank adaptation methods [17] and setting fixed rank (e.g., $r = 8$ or $r = 16$ ) for modeling all sub-parameter matrix $\bar{w}_i^{10}$ . The assumption behind such approach is that the intrinsic dimension of all sub-parameter matrix $\bar{w}_i^{10}$ from pretrained models are all the same. However, a nature question is "Is this assumption reasonable?" + +To answer this question, in Figure 4, we randomly select four parameter matrices from various parameter types (i.e., query matrix, value matrix, and feedforward matrix from transformer layers) and visualize the sensitive parameters of reorganized compact parameter matrices (i.e., the parameter matrix $\bar{w}_i$ obtained by Eq. 4). The results are shown in Figure 4. From the results, we can see that although the size of parameter matrix $w_i \in \{w_0, \dots, w_i, \dots, w_N\}$ are all the same (i.e., $768 \times 768$ ), its size of compact sub-parameter matrices $\bar{w}_i^{10}$ is different. For example, the size of compact sub-parameter matrices $\bar{w}_i^{10}$ from 15-layer FNN weight matrix is $589 \times 284$ , but the ones from 10-layer FNN weight matrix is $318 \times 533$ . Besides, we also calculate the proportion of the sensitive parameters in the compact sub-parameter matrices $\bar{w}_i^{10}$ (see the red number of Figures. $5a \sim 5e$ ). From results, we can see that the proportion of the sensitive parameters are all different. The above results suggest that setting fixed rank (e.g., $r = 8$ or $r = 16$ ) for modeling all sub-parameter matrix change is difficult to fit its change of size and proportion of the sensitive parameters of sub-parameter matrices. + +Based on the above analysis, an intuitive idea is how allocate the matrix rank in an adaptive manner such that the dynamic characteristics of compact sub-parameter matrices $\bar{w}_i^{10}$ can be finely modeled. Based on this, we propose a novel dynamic-rank allocation strategy. Our key insight is that the number of sensitive parameters and the compact matrix's dimension may reflect the intrinsic dimension of compact sub-parameter matrix $\bar{w}_i^{10}$ , i.e., (1) the smaller the number of sensitive parameters, the smaller the intrinsic dimension and (2) the smaller the dimensions of a compact matrix, the smaller the intrinsic dimension. Based on the + +
Method\DatasetNaturalSpecializedStructuredVTAB
CIFAR-100Caltech101DTDFlowers102PetsSVHNSun397Mean Acc.Patch CamelyonEuroSATResis45RetinopathyMean Acc.Clev/countClev/distanceDMLabKITTI/distancedSprites/locdSprites/oriSmallNORB/aziSmallNORB/eleMean Acc.Mean Params. (%)
Full [19]68.987.764.397.286.987.438.875.979.795.784.273.983.456.358.641.765.557.546.725.729.147.665.57100.00
Addition-based methods
MLP-3 [19]63.884.762.397.484.732.549.267.877.088.070.256.172.847.832.832.358.112.921.215.224.830.653.211.50
VPT-Shallow [19]77.786.962.697.587.374.551.276.878.292.075.672.979.750.558.640.567.168.736.120.234.147.064.850.13
VPT-Deep [19]78.890.865.898.088.378.149.678.581.896.183.468.482.468.560.046.572.873.647.932.937.855.069.430.70
Adapter [16]69.290.168.098.889.982.854.379.084.094.981.975.584.180.965.348.678.374.848.529.941.658.571.440.39
NOAH [46]69.692.770.299.190.486.153.780.284.495.483.975.884.982.868.949.981.781.848.332.844.261.373.250.50
Reparameterization-based methods
Linear [19]63.485.064.397.086.336.651.069.178.587.568.674.077.234.330.633.255.412.520.09.619.226.953.000.05
PARTIAL-1 [19]66.885.962.597.385.537.650.669.478.689.872.573.378.541.534.333.961.031.332.816.322.434.256.528.38
Bias [37]72.887.059.297.585.359.951.473.378.791.672.969.878.361.555.632.455.966.640.015.725.144.162.050.16
LoRA [17]68.191.469.899.090.586.453.179.885.195.884.774.284.983.066.950.481.480.246.632.241.160.272.630.90
Selection-based methods
GPS [47]70.893.974.899.482.491.451.680.687.295.786.176.186.380.961.854.081.484.252.630.245.561.373.680.25
SPT-LoRA-8 [14]72.892.772.699.391.286.755.281.585.595.885.775.685.782.068.149.481.480.248.629.439.759.973.260.51
SPT-LoRA-16 [14]71.993.271.099.391.289.655.281.686.496.285.676.286.183.267.052.082.981.849.632.339.061.073.871.36
CDRA-SPT (Ours)73.893.372.799.491.689.955.882.486.896.286.276.086.383.068.951.582.180.951.733.043.261.874.530.49
+ +Table 1. Performance comparisons on VTAB-1k with ViT-B/16 models pre-trained on ImageNet-21K. Best results are highlighted in bold. + +above insight, we propose dynamically allocating compact matrix's rank $r^i$ according to the sensitive parameters and compact matrix's dimension (i.e., its row $\bar{d}^i$ and column $\bar{k}^i$ ). That is, finding a maximum rank $r$ as $r^i$ and satisfy: + +$$ +\bar {r ^ {i}} = \max (r), +$$ + +$$ +\mathbf {s . t .} \sum m _ {i} \geq \left(\bar {d} ^ {i} \times r + \bar {k} ^ {i} \times r\right) \times \frac {d ^ {i} \times k ^ {i}}{\bar {d} ^ {i} \times \bar {k} ^ {i}}, \tag {7} +$$ + +where $d^i$ and $k^i$ denote the dimension of the original matrix, and $m_i$ denotes the binary mask where its value '1' denotes the corresponding weights are sensitivity parameters. The goal of Eq. 7 is to ensure that the number of trainable parameters of our CDRA is always lower than the number of sensitive parameters, and the smaller the dimension of compact sub-parameter matrix $\bar{w}_i^{10}$ , the lower its number. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. Following SPT [14], we evaluate our method of CDRA-SPT and various baselines on a series of datasets, including image classification and semantic segmentation tasks. For image classification, we use: VTAB-1k: [38] Visual Task Adaptation Benchmark (VTAB), which consists of 19 visual classification tasks grouped into three types: natural image classification tasks, specialized image classification tasks, and structured image classification tasks. FGVC: Fine-Grained Visual Classification (FGVC) benchmark, which includes five downstream tasks: CUB-200-2011 [34], NABirds [33], Oxford Flowers [28], StanfordDogs [21], and StanfordCars [10]. + +For semantic segmentation, we evaluate on: ADE20K: [49]A large-scale dataset, with over 20,000 images annotated across 150 object and stuff categories, is commonly + +used for semantic segmentation tasks due to its diverse scenes and detailed annotations. VOC: [7] The PASCAL VOC dataset, consisting of 20 object categories with pixel-level annotations, is known for its well-defined evaluation protocols and balanced class distribution. + +Baselines and Evaluation Metrics. In order to demonstrate the effectiveness of our CDRA-SPT method, we compare our CDRA-SPT with four type parameter-efficient fine-tuning methods: (i) Full: As the most commonly used method, full fine-tuning updates all parameters of the whole model for adaptation to downstream tasks. (ii) Addition-based: The trainable parameters are added to the model in this method, but require extra computational cost during inference. Consisting of MLP-k, Adapter [16], VPT [19] and NOAA [46]. (iii) Reparameterization-based: The trainable parameters can be merged into the original parameter matrix after fine-tuning without computational cost during inference, including LINEAR, PARTIAL-k, BIAS [37], and LORA [17]. (iv)Selection-based: select some important parameters as a subset of parameters in the backbone for fine-tuning, including SPT [14] and GPS [47]. In particular, our CDRA-SPT method can be categorized into the type of selection-based PEFT approaches, thus existing SPT [14] and GPS [47] methods are our key competitors in this paper. + +Implementation Details. Following SPT [14], we employ AdamW optimizer with cosine learning rate decay to fine-tune the model for 100 epochs. And we use the standard data augmentation pipeline to process images in all datasets of the FGVC, VTAB-1k. We set the same number of training samples as the SPT to select sensitive parameters. All experiments are conducted on NVIDIA 4090 GPU. + +
DatasetCUB-2011NABridsOxfordFlowersStan.DogsStan.CarsMeanAcc.Params.(%)
Full [19]87.382.798.889.484.588.54100.00
Linear [19]85.375.997.986.251.379.320.21
Bias [37]88.484.298.891.279.488.400.33
Adapter [16]87.184.398.589.868.685.660.48
LoRA [17]85.679.898.987.672.084.780.90
VPT-Shallow [19]86.778.898.490.768.784.620.29
VPT-Deep [19]88.584.299.090.283.689.110.99
SPT-LoRA-8 [14]88.083.599.389.886.789.461.20
SPT-LoRA-16 [14]88.283.899.389.787.489.692.28
CDRA-SPT (Ours)88.683.499.389.888.289.860.86
+ +Table 2. Results of image classification on FGVC datasets. + +
DatasetADE20kVOC
MethodmDice (↑)mIoU (↑)Params.(%)mDice (↑)mIoU (↑)Params.(%)
Bias72.860.70.1274.362.40.12
LoRA73.861.91.2075.063.90.34
Adapter72.960.90.3476.765.90.34
SPT-LoRA-873.361.40.8376.565.20.87
SPT-LoRA-1673.361.41.6076.966.31.70
CDRA (ours)74.062.00.8277.466.40.70
+ +# 4.2. Discussion of Results + +Image Classification Tasks. We evaluate various baselines and our CDRA-SPT on the FGVC and VTAB benchmarks. The results are reported in Tables 1 and 2, respectively. From the results, it can be found that our CDRA-SPT outperforms all baseline methods on both FGVC and VTAB, which sufficiently proves the effectiveness of our method. Specifically, on VTAB dataset, we can see that 1) our CDRA-SPT exceeds existing addition-based methods by a large margin. This is reasonable since our CDRA-SPT explores weight sensitivity, which can finely characterize distribution of downstream task; 2) different from reparameterization-based methods, our CDRA-SPT focuses on selecting sensitive sub-parameters to perform PEFT of downstream tasks. This improvement results verify the superiority of our CDRA-SPT; and 3) our CDRA-SPT obtains around $1\% \sim 2\%$ accuracy improvement of the mean accuracy compared to selection-based methods. In particular, our CDRA-SPT method also exceeds our key baseline (i.e., SPT-LoRA-8 even SPT-LoRA-16), while it only use $0.49\%$ of trainable parameters and is below the parameter number of SPT-LoRA-8, which verifies our CDRA-SPT superiority. Besides, on FGVC, we also similar performance improvements, i.e., 1) our CDRA-SPT also achieves best classification over previous PEFT methods and 2) our CDRA-SPT exceeds our key competitors (i.e., SPT) by around $2\%$ . This further shows the superiority of our CDRA-SPT. + +Semantic Segmentation Tasks. In Table 3, we conduct semantic segmentation experiments on ADE20K[49] and VOC [7] datasets to show the universality of our CDRA-SPT. In particular, following [14], 1) we select DINOV2 [32] as our pre-trained models and report mDice and mIoU as our comparison metrics; and 2) we implement Bias [37], LoRA [17], Adapter [16], SPT-LoRA-8, and SPT-LoRA-16 as our baselines. From these results, we can see that 1) our CDRA-SPT achieves the best segmentation performance on all datasets. This further verifies the effectiveness; and 2) + +Table 3. Result of semantic segmentation on ADE20K and VOC. + +
MethodNaturalSpecializedStructuredMean AccParams. (%)
(i)Dynamic rank82.486.361.876.80.49
(ii)Fixed Rank-881.084.559.475.00.42
(iii)Fixed Rank-1681.085.259.275.10.76
(iv)Fixed Rank-3280.885.260.175.41.45
+ +Table 4. Analysis of compact dynamic-rank adaptation on VTAB. + +
Case1Case2NaturalSpecializedStructuredMean AccParams. (%)
(i)80.985.360.075.40.94
(ii)✓81.285.560.375.70.88
(iii)✓81.485.860.275.80.46
(iv)✓✓81.785.860.376.00.37
+ +Table 5. Ablation study of our CDRA on Cases 1 and 2 of Eq. 1. + +our CDRA-SPT only use around $0.4\%$ trainable parameters, which is below the number of our key baseline (i.e., SPT-LoRA-8, and SPT-LoRA-16). This means that our CDRA-SPT indeed alleviates the parameter redundancy issue of existing SPT while achieving better PEFT performance. + +# 4.3. Ablation Study + +Is our idea (i.e., applying dynamic-rank adaptation at compact sub-parameter matrix) effective? In Table 4, we conduct an ablation study on VTAB dataset to show the effectiveness of introducing dynamic rank adaptation at compact sub-parameter matrix. Specially, (i) we implement our CDRA-SPT as our baseline (i.e. applying dynamic-rank adaptation at compact sub-parameter matrix); Then, on (ii) $\sim$ (iv) we replace the dynamic-rank adaptation of our CDRA-SPT by using a fixed-rank (i.e., 8-rank adaptation, 16-rank adaptation, and 32-rank adaptation) adaptation used in existing SPT method, respectively. From the results, we can see that the performance of (i) outperforms (ii) $\sim$ (iv) by a large margin and the number of trainable parameters on (i) is lower than the ones of (iii) $\sim$ (iv), which is reasonable because the dynamic-rank characteristic of subparameter matrix can be fully exploited for PEFT. This suggests that our idea, i.e., applying dynamic-rank adaptation at compact sub-parameter matrix, is very effective for SPT. Is our CDRA (i.e., compact dynamic-rank adaptation) all effective for cases 1 and 2 of SPT? To answer this question, in Table 5, we conduct an ablation study on VTAB dataset. Specially, (i) we remove our CDRA at Cases 1 and 2 of Eq. 1 as our baseline (i.e. using Eq. 1 to perform PEFT); (ii) we replace the mask tuning of Eq. 1 by using our CDRA on (i), i.e., applying our CDRA at Case 1 of Eq. 1; (iii) we replace the low-rank adaptation of Eq. 1 by using our CDRA on (i), i.e., applying our CDRA at Case 2 of Eq. 1; and (iv) we apply our CDRA at both Case 1 and Case 2 of Eq. 1 on (i), which is exactly our CDRA-SPT method. From these results, we can see that 1) the performance of (ii) $\sim$ (iv) exceeds (i) by around $0.1\% \sim 1\%$ , which suggests that our CDRA is very effective, which can alleviate parameter redundancy meanwhile achieving performance improvement; and 2) the setting of (iv) achieves best PEFT performance and minimum number of trainable parameters, + +
MethodNaturalSpecializedStructuredMean AccParams. (%)
(i)Baseline (SPT)81.585.759.975.70.51
(ii)+ CDRA-SPT (×1.0)82.486.361.876.80.49
(iii)+ CDRA-SPT (×1.5)81.085.058.975.00.54
(iv)+ CDRA-SPT (×0.5)81.084.957.974.60.29
+ +Table 6. Analysis of sub-parameter matrix's size on VTAB. + +
MethodNaturalSpecializedStructuredMean AccParams. (%)
(i)Baseline (SPT)81.585.759.975.70.51
(ii)+ CDRA (\( \triangle w_i^{00} \))81.285.359.275.20.74
(iii)+ CDRA (\( \triangle w_i^{01} \))80.785.358.474.80.48
(iv)+ CDRA (\( \triangle w_i^{10} \))82.486.361.876.80.49
(v)+ CDRA (\( \triangle w_i^{11} \))81.284.658.974.90.73
+ +Table 7. Analysis of different compact sub-parameter matrices. + +
MethodInference Time (ms/img)Inference Memory (GB)Fine-tuning Memory (GB)
Full2.81.311.9
VPT-Deep3.81.913.2
LoRA2.81.38.2
SPT2.81.39.8
CDRA-SPT2.81.38.3
+ +Table 8. Analysis of computational cost on VTAB. + +which verifies the superiority of our CDRA-SPT method. + +How does the size of sub-parameters impact the performance of our method? In Table 6, we analyzed the impact of the size of sub-parameters on PEFT performance. Specifically, (i) we first implement existing SPT as our baseline; Then, (ii) we apply our CDRA on (i), which is exactly our CDRA-SPT; (iii) we expand the size of compact sub-parameter matrix on (ii) to 1.5 times; (iv) we reduce the size of compact sub-parameter matrix on (ii) to 0.5 times. From these results, we can see that 1) our CDRA-SPT achieves the best PEFT performance; and 2) the performance of our CDRA-SPT decreases around $1\% \sim 2\%$ , which may be because too large or too small sub-parameter matrix leads to overfitting or underfitting issues of downstream tasks. + +How does tuning different sub-parameters impact performance? To answer this question, we achieve our CDRA-SPT on different sub-parameter matrices (i.e., $\triangle w_{i}^{00}$ , $\triangle w_{i}^{01}$ , $\triangle w_{i}^{10}$ , and $\triangle w_{i}^{11}$ ) on the VTAB dataset. The results are reported in Table 7. From these results, it can be observed is that our CDRA-SPT applying at sub-parameter matrices $\triangle w_{i}^{10}$ achieve the best PEFT performance. This is reasonable since the sub-parameter matrices $\triangle w_{i}^{10}$ indeed is more sensitive than sub-parameter matrices $\triangle w_{i}^{00}$ , $\triangle w_{i}^{01}$ , and $\triangle w_{i}^{11}$ ), after reorganizing the parameter matrix in theory. Thus, fine-tuning $\triangle w_{i}^{10}$ is more effective. + +How much our CDRA-SPT take computational cost? In Table 8, we report the computational cost of our CDRA-SPT where we select fullfine-tuning, VPT-Deep, LORA, and SPT as our comparison baselines. From results, we observe that 1) the VPT-Deep method takes higher inference latency and GPU memory, which is because it introduces additional prompts as inputs; and 2) our CDRA-SPT achieves comparable inference latency and GPU memory with Full, LoRA, SPT methods. This is reasonable because the learned parameter (i.e., weight changes) by our CDRA-SPT can also be reparameterized and merged into the pre + +![](images/26e0feda93a0d60e93b9149420e782c870c0952ee44785bd5f3f00da3f150b83.jpg) +(a) CIFAR + +![](images/3114cce3f84f0547eeef871c3a9b1f6c45cb3bca9d5dbf8196b06d178b63e67b.jpg) + +![](images/f831048b7b66ff02f66f5d1d8e14c9e5f21c5ac212378692b5e8690eba974e03.jpg) +(c) Pets + +![](images/c9b17a5ec169bbf27381e825c8f8f258c1ae19ab49b1d26e1d07ba4bc61c87f0.jpg) +(d) Patch Camelyon + +![](images/8cb33980d75252a95984cf9433394a83d172682da68b974863f052d99e485312.jpg) +(b) DTD + +![](images/4feececcfb9ba433f9a77718c6c54a4690ff79bd5794fb03b15ab3f85ee8728f.jpg) +(f) Resisc45 + +![](images/18d686505a0583ef365d5b93396013273d696e4535f02f6df4a6be2e7639a316.jpg) +(g) Clevr/count + +![](images/6dd9fd44c73c4fa8dae038f31b69f0c33a14543bd02c49b9ed44bb2b72633524.jpg) +(e) Eurosat +(h) DSprites/loc + +![](images/6b77f64ac0414976b8379f66687c76ef5700815ad1e329768d9d183fb5759ff9.jpg) +(i) SmallNORB/azi +Figure 5. Distribution of the size number of compact subparameters at natural, specialized, and structured datasets. + +trained model. Thus, our CDRA-SPT does not consume additional cost to perform downstream tasks. Besides, it can be found that our CDRA-SPT takes less memory than baselines during fine-tuning, which verifies our efficiency. + +How does the row/column dimension of compact sub-parameter matrix distribute at different datasets? To answer this question, in Figure 5, we visualize the dimension distribution of compact sub-parameter matrix at all layers. Specifically, we randomly select three datasets from natural image classification tasks, specialized image classification tasks, and structured image classification task, respectively, and then count the number of rows (see red line) and columns (see blue line) of all compact sub-parameter matrices. From the results, we can see that most parameter matrices all are grouped into a very compact sub-parameter matrix with mean row dimension of $121\sim 221$ and mean column dimension of $384\sim 567$ , which verifies the rationality of our idea (i.e., compact dynamic-rank adaptation). + +# 5. Conclusions + +In this paper, we find a new distribution character of sensitive parameters, i.e., these sensitive parameters usually concentrate in a small number of rows or columns. Resorting to such characters, this paper proposes a novel compact dynamic-rank adaptation method for sensitivity-aware parameter efficient fine-tuning, called CDRA-SPT. Its advantage is that the dynamic-rank characteristic of subparameter matrix can be fully exploited for PEFT. Experimental results on various datasets show that our CDRA-SPT achieves superior performance over previous methods. + +# Acknowledgements + +This work was supported by the Guangdong Basic and Applied Basic Research Foundation under Grant No. + +2025A1515011674, the Shenzhen Peacock Program under Grant No. ZX20230597, NSFC under Grant No. 62272130 and Grant No. 62376072. + +# References + +[1] Ahmed Agiza, Marina Neseem, and Sherief Reda. Mtlora: Low-rank adaptation approach for efficient multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16196-16205, 2024. 3 +[2] Tianlong Chen, Jonathan Frankle, Shiyu Chang, Sijia Liu, Yang Zhang, Zhangyang Wang, and Michael Carbin. The lottery ticket hypothesis for pre-trained bert networks. ArXiv, abs/2007.12223, 2020. 3 +[3] Daiki Chijiwa, Shin'ya Yamaguchi, Atsutoshi Kumagai, and Yasutoshi Ida. Meta-ticket: Finding optimal subnetworks for few-shot learning within randomly initialized neural networks. Advances in Neural Information Processing Systems, 35:25264-25277, 2022. 3 +[4] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1 +[6] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1 +[7] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision, 111(1):98-136, 2015. 6, 7 +[8] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2018. 3 +[9] Ziqi Gao, Qichao Wang, Aochuan Chen, Zijing Liu, Bingzhe Wu, Liang Chen, and Jia Li. Parameter-efficient fine-tuning with discrete fourier transform. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 2 +[10] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Fine-grained car detection for visual census estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 6 +[11] Dipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twardowski, Joost van de Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28525-28534, 2024. 3 + +[12] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogerio Feris. Spottune: transfer learning through adaptive fine-tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4805-4814, 2019. 3 +[13] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. In *Forty-first International Conference on Machine Learning*. 2 +[14] Haoyu He, Jianfei Cai, Jing Zhang, Dacheng Tao, and Bohan Zhuang. Sensitivity-aware visual parameter-efficient finetuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11825-11835, 2023. 1, 2, 3, 4, 6, 7 +[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1 +[16] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 6, 7 +[17] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations. 2, 3, 4, 5, 6, 7 +[18] Wooseong Jeong and Kuk-Jin Yoon. Quantifying task priority for multi-task optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 363-372, 2024. 3 +[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2, 6, 7 +[20] Haeyong Kang, Rusty John Lloyd Mina, Sultan Rizky Hikmawan Madjid, Jaehong Yoon, Mark Hasegawa-Johnson, Sung Ju Hwang, and Chang D Yoo. Forget-free continual learning with winning subnetworks. In International Conference on Machine Learning, pages 10734-10750. PMLR, 2022. 3 +[21] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), 2011. 6 +[22] 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. 1 +[23] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki M Asano. Vera: Vector-based random matrix adaptation. In The Twelfth International Conference on Learning Representations. 2 +[24] Yan-Shuo Liang and Wu-Jun Li. Inflora: Interference-free low-rank adaptation for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23638-23647, 2024. 3 + +[25] Yan-Bo Lin, Yi-Lin Sung, Jie Lei, Mohit Bansal, and Gedas Bertasius. Vision transformers are parameter-efficient audiovisual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2299–2309, 2023. 1 +[26] Shih-yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. In *Forty-first International Conference on Machine Learning*. 2 +[27] Pierre Marza, Laetitia Matignon, Olivier Simonin, and Christian Wolf. Task-conditioned adaptation of visual features in multi-task policy learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17847-17856, 2024. 3 +[28] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6 +[29] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10775–10784, 2021. 1 +[30] Chunlin Tian, Zhan Shi, Zhijiang Guo, Li Li, and Chengzhong Xu. Hydralora: An asymmetric lora architecture for efficient fine-tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3 +[31] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1 +[32] Rob van Gastel. Finetuning dinov2 with lora for image segmentation. 2024. 7 +[33] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 595-604, 2015. 6 +[34] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6 +[35] Runxin Xu, Fuli Luo, Zhiyuan Zhang, Chuanqi Tan, Baobao Chang, Songfang Huang, and Fei Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. arXiv preprint arXiv:2109.05687, 2021. 3 +[36] Jaehong Yoon, Sultan Madjid, Sung Ju Hwang, Chang-Dong Yoo, et al. On the soft-subnetwork for few-shot class incremental learning. In International Conference on Learning Representations (ICLR) 2023. International Conference on Learning Representations, 2023. 3 +[37] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1–9, 2022. 2, 6, 7 + +[38] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 6 +[39] Baoquan Zhang, Xutao Li, Yunming Ye, Zhichao Huang, and Lisai Zhang. Prototype completion with primitive knowledge for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3754-3762, 2021. 3 +[40] Baoquan Zhang, Chuyao Luo, Demin Yu, Xutao Li, Huiwei Lin, Yunming Ye, and Bowen Zhang. Metadiff: Meta-learning with conditional diffusion for few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16687-16695, 2024. 3 +[41] Baoquan Zhang, Huaibin Wang, Chuyao Luo, Xutao Li, Guotao Liang, Yunming Ye, Xiaochen Qi, and Yao He. Codebook transfer with part-of-speech for vector-quantized image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7757–7766, 2024. 1 +[42] Baoquan Zhang, Bingqi Shan, Aoxue Li, Chuyao Luo, Yunming Ye, and Zhenguo Li. Zookt: Task-adaptive knowledge transfer of model zoo for few-shot learning. Pattern Recognition, 158:110960, 2025. 1 +[43] Jiahui Zhang, Fangneng Zhan, Christian Theobalt, and Shijian Lu. Regularized vector quantization for tokenized image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18467-18476, 2023. 1 +[44] Longteng Zhang, Lin Zhang, Shaohuai Shi, Xiaowen Chu, and Bo Li. Lora-fa: Memory-efficient low-rank adaptation for large language models fine-tuning. arXiv preprint arXiv:2308.03303, 2023. 2 +[45] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2 +[46] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 6 +[47] Zhi Zhang, Qizhe Zhang, Zijun Gao, Renrui Zhang, Ekaterina Shutova, Shiji Zhou, and Shanghai Zhang. Gradient-based parameter selection for efficient fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28566-28577, 2024. 2, 3, 6 +[48] Chuanxia Zheng and Andrea Vedaldi. Online clustered codebook. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22798-22807, 2023. 1 +[49] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6, 7 \ No newline at end of file diff --git a/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/images.zip b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a462ddf8e0645b557aaa36dee580f59d9d4ced02 --- /dev/null +++ b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96534424e530838d2f51bffb5d3e7b88fc6842d3d11fda72f500f9e44861fdf7 +size 464543 diff --git a/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/layout.json b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..716b26a1a5cc2329724a86180b9840b26ff253da --- /dev/null +++ b/CVPR/2025/Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b1c61472d45ebc7290416e4d255f2203cad808b1b343b459f6ee89b5a45bd34 +size 455251 diff --git a/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/b0b27d1a-186c-41cf-a28c-523d8082bf6d_content_list.json b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/b0b27d1a-186c-41cf-a28c-523d8082bf6d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..257342d89894671807845011d670e831c26d8851 --- /dev/null +++ b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/b0b27d1a-186c-41cf-a28c-523d8082bf6d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1a452f160e361ef981eceb17e1f9dcf36ffd795e95a2398e8a4af8e028cb60e +size 77395 diff --git a/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/b0b27d1a-186c-41cf-a28c-523d8082bf6d_model.json b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/b0b27d1a-186c-41cf-a28c-523d8082bf6d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d51ec37363ba13d269cf97e2e562f079330c86f8 --- /dev/null +++ b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/b0b27d1a-186c-41cf-a28c-523d8082bf6d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65cab5396230f1d09004456537de0cac04da65c554604bcf89de0d4451db9e35 +size 96601 diff --git a/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/b0b27d1a-186c-41cf-a28c-523d8082bf6d_origin.pdf b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/b0b27d1a-186c-41cf-a28c-523d8082bf6d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..50bc96b745eb100022bd6210862a0c9d5e4500f8 --- /dev/null +++ b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/b0b27d1a-186c-41cf-a28c-523d8082bf6d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ad85e621d1ba6b76d54853fe8650fef17390065bc8d30af935fd2bb02cfec1b +size 2925155 diff --git a/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/full.md b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..99ae03f89b1e7651b71c8e36c9bff1e503086b67 --- /dev/null +++ b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/full.md @@ -0,0 +1,342 @@ +# Separation of powers: On segregating knowledge from observation in LLM-enabled knowledge-based visual question answering + +Zhen Yang $^{1,2*}$ Zhuo Tao $^{1,2*}$ Qi Chen $^{3}$ Liang Li $^{1\dagger}$ Yuankai Qi $^{4\dagger}$ + +Anton van den Hengel3 Qingming Huang1,2 + +$^{1}$ Institute of Computing Technology, Chinese Academy of Sciences $^{3}$ University of Adelaide + +$^{2}$ University of Chinese Academy of Sciences $^{4}$ Macquarie University + +# Abstract + +Knowledge-based visual question answering (KBVQA) separates image interpretation and knowledge retrieval into separate processes, motivated in part by the fact that they are very different tasks. In this paper, we transform the KBVQA into linguistic question-answering tasks so that we can leverage the rich world knowledge and strong reasoning abilities of Large Language Models (LLMs). The captionthen-question approach to KBVQA has been effective, but relies on the captioning method to describe the detail required to answer every possible question. We propose instead a Question-Aware Captioner (QACap), which uses the question as guidance to extract correlated visual information from the image and generate a question-related caption. To train such a model, we utilize GPT-4 to build a corresponding high-quality question-aware caption dataset on top of existing KBVQA datasets. Extensive experiments demonstrate that our QACap model and dataset significantly improve KBVQA performance. Our method, QA-Cap, achieves $68.2\%$ accuracy on the OKVQA validation set, $73.4\%$ on the direct-answer part of the A-OKVQA validation set, and $74.8\%$ on the multiple-choice part, all setting new SOTA benchmarks. + +# 1. Introduction + +Visual Question Answering (VQA) requires a model to answer open-set questions about a given image, bridging the gap between visual perception and language understanding. Traditional VQA tasks [2, 9] focus solely on the content within the image, requiring basic image comprehension, such as object detection and identifying visual attributes. Knowledge-Based Visual Question Answering (KBVQA) [37] expands this task by incorporating external world knowledge into the reasoning process [24, 29, 38]. + +![](images/408be746c36539a9618e26f977fe9b4c22952edd19e623914bba33c539e2020c.jpg) + +![](images/27305a843e49a661a776d2faf63d2cbf9bb9ca63553791bca1c770bc560ec4e4.jpg) +Question +Q: "What toy is this?" +GT Answer: Stuffed animal + +![](images/6913a54464b9364fc5be1811d9fb2753741f04f7f7e10020d55b32d7402d333e.jpg) + +![](images/29d1c6d9cf4b5ef6296a974d4f8185b32eeed6d05b525fb772cb19dd81d53c8c.jpg) +BLIP: a boy is sitting in a pew. + +![](images/9c02751d714e48d0d0e363452e6a44b65c7a5949fb9cfd8a71639e436272437e.jpg) + +![](images/86de3e1a6a4aeb9744551cefc96f3721de94f800590fe407b8d7df6980812cce.jpg) +QACap: a boy holds a stuffed animal in his backpack. + +![](images/434b558949ef3aab1c70afba94987b480d118c753a32aabb58b8a7b470dcf890.jpg) +Q: "Can you guess to which part of body this dish is good?" +GT Answer: Eye + +![](images/87070f0cb3d7b289d00cb45a50e058f67f7235d0171d67ac435e48dc55a73cec.jpg) +BLIP: four people standing against a brick wall eating food. +Figure 1. Comparative illustration of predictions on the OKVQA dataset generated by BLIP and our proposed QACap. Using GradCAM, we visualize the focus areas of each model during the generation of question-aware captions. It shows our underlying motivation of QACap: by integrating the question into the captioning process, the model is guided to attend to the more relevant regions of the image, thereby producing more appropriate captions that align with the given question. + +![](images/67d9ca04ed7233f51337eb91a7a521584c3d9d3e3afbd8d6ea06deaa62259dd0.jpg) +QACap: a group of four people, including a man with carrots in his mouth, standing by a brick wall. + +For example, in Figure 2, if shown an image of a natural landscape in the wild and asked "What South American country usually has this climate?", a KBVQA system needs to recognize the landscape and retrieve knowledge about its location. Thus, the key challenges in KBVQA are: 1) accurately identifying information in the image relevant to the question, and 2) effectively integrating relevant background knowledge not directly present in the image for reasoning. + +To retrieve question-relevant background knowledge, prior methods such as [25, 40, 45] have leveraged external databases. However, these approaches often fall short due to the limited scale, domain-specific focus, and low retrieval efficiency of these knowledge bases. To address these chal + +![](images/67be693324df552c13dd7cbdcec445801b77268b81e78dc93687b6efd03e489f.jpg) +Figure 2. The pipeline of our method. The proposed QACap model takes in the image and the question and outputs a question-relevant image caption, serving as the context. The task instruction is designed to enable the LLM to fully understand its task and what to do. The off-the-shelf LLM is prompted with the task instruction, the context, and the question, from which it generates the answer. + +lenges multimodal large language models (MLLMs) have been introduced to align visual and text features within a shared semantic space. However, this multimodal training is highly data-intensive, computationally demanding, and often reduces language capabilities, limiting the LLM's effectiveness in integrating visual information with knowledge for multi-step reasoning. Therefore, to maintain language capabilities, recent works [11, 30] instead convert images into textual descriptions, which are then input into LLMs for prediction (reasoning). This method has two key advantages over MLLMs. First, captions provide a clear format, allowing researchers to easily diagnose errors, and then determine whether they stem from the caption generation or the LLM's reasoning. Second, using off-the-shelf LLMs increases the generality of the method, and allows it to exploit the rapid pace of development in the area. + +While many caption-driven LLM-based KBVQA methods [15, 42] have achieved promising performance, we show here that generating image captions directly relevant to the question improves performance, as general captions often lack the specific detail needed to enable accurate responses. PromptCap [15] has a similar goal, in that it uses question-answer (QA) pairs to prompt the image captions. The challenge with that model is that, due to its training, the PromptCap captioner often answers the VQA question directly, rather than providing the information required for the LLM to answer it. This does not achieve the desired separation of image analysis from knowledge source. We seek here a solution that focuses on providing captions that enable an LLM to answer general VQA questions, rather than have the LLM just confirming the answer generated by the captioner. + +To address the above issues, we propose a framework that contains a Question-Aware Captioner (QACap) model with an LLM (see the overall pipeline in Figure 2). Concretely, the QACap takes both the image and question as in + +puts to generate a caption that is specifically relevant to the question's context. The LLM perceives the question-aware image caption as the problem-solving context and predicts the answer jointly using the context and self-contained knowledge. In QACap, we design a Question-Aware Feature Distillation (QAFD) module that focuses on image areas relevant to the question while minimizing irrelevant regions, effectively incorporating question information into image feature extraction, as shown in Figure 1. Besides, we introduce a Question-Aware Regional Consistency Learning (QARCL) strategy to enhance alignment between interpreting the question and generating the caption, guiding question-aware caption generation. To support this, we create a high-quality question-aware caption dataset using GPT-4, refining the initial dataset to prevent answer leakage and enhance descriptiveness with natural expressions. Following principles of accuracy and helpfulness, the best caption is selected as the training instance from several candidates. Extensive experiments demonstrate that our model and high-quality dataset lead to substantial improvements in KBVQA tasks. Consequently, our method achieves state-of-the-art accuracy, with $68.2\%$ on the OKVQA dataset and $74.8\%$ on the A-OKVQA dataset. + +We summarize our contribution as follows: + +- We propose a Question-Aware Captioner (QACap), which integrates both image and question inputs to generate contextually relevant captions. It contains a novel Question-Aware Feature Distillation (QAFD) module that focuses on image regions relevant to the question, enhancing feature extraction based on question context. +- We design a Question-Aware Regional Consistency Learning (QARCL) approach, which improves alignment between question interpretation and caption generation, ensuring that captions remain closely tied to question-relevant image areas. +- We construct a new dataset, curated with GPT-4 assis + +tance, which provides descriptive and contextually appropriate captions. Experiments show our QACap supported by this dataset achieves SoTA performance on two main KBVQA benchmarks. + +# 2. Related Work + +# 2.1. Knowledge-based Visual Question Answering + +Knowledge-based VQA is a much more difficult task where answering questions requires additional world knowledge beyond the image content and certain reasoning ability. Early datasets like KB-VQA [37], FVQA [38] annotate questions by selecting a fact (a knowledge triplet such as "dog is mammal") from a fixed knowledge base. OK-VQA dataset [24] is the first large-scale dataset with questions that need to be answered using external knowledge instead of a provided fixed knowledge base, and A-OKVQA provides more questions with Multiple-Choice (MC) as well as Direct Answer evaluation settings. + +Previous studies [10, 25, 26, 40, 45] try to retrieve different knowledge from various external knowledge resources, e.g., ConceptNet [31], Wikipedia[36], Google Images[40], etc. These approaches jointly reason over the retrieved knowledge and the image-question pair to form the answer. However, PICa [42] finds this two-step method may be suboptimal due to the representation mismatches during the retrieving stage and the infer stage, resulting in the potentially limited performance. At the same time, with the rapid development of deep learning [4, 13, 22, 34, 43, 44], large language models have witnessed rapid advancement and demonstrated powerful capabilities in natural language processing. Researchers resort to LLMs to help accomplish knowledge-based VQA tasks. + +# 2.2. LLM for Knowledge-based VQA + +Trained on extensive corpora, LLMs exhibit not only excellent language comprehension abilities but also possess rich world knowledge and reasoning capabilities. Therefore, LLMs are considered excellent implicit knowledge engines for addressing knowledge-based VQA. The difficulty lies in that LLMs lack visual modality. Besides doing multimodal training, many methods explicitly convert images into caption captions to tackle this issue. + +This approach can directly make use of the LLMs. PICa [42] is the first method to adopt GPT-3 for solving the KBVQA task in a few-shot manner by just providing a few in-context VQA examples. [10] propose to use both implicit (i.e. GPT-3) and explicit (i.e. KBs) knowledge based on CLIP retrieval [28] which are combined by a novel fusion module called KAT (based on T5 or Bart). [21] propose to integrate local visual features and positional information (bounding box coordinates), retrieved external and implicit knowledge (using a GPT-3) into a transformer + +![](images/2ad804a5609e49b223eda21d1f06ef271e9ee12ce7f24f290e367dcece4e4ed1.jpg) +Q: How old do you have to be in canada to do this? A: 18 +OC: A couple, who in Canada must be at least 18 years of age, is illustrated holding a martini in their hands, indicating the legal drinking age for this action. +MC: A man and woman celebrate together, each holding a martini with olives, symbolizing a shared moment of enjoyment. + +![](images/04de1414ac49c36224fd4a7ffdf201f607988515deb9c53d4a79ca1ab0efdaf4.jpg) +Q: Is the scene indoors or outdoors? +A: outdoor +OC: A pedestal sink and a commode, repurposed as outdoor decor, sit on the street in front of a building, showcasing creativity in outdoor settings. +MC: A white sink and toilet, repurposed as planters, are situated on a sidewalk in front of a building. +Figure 3. Two examples during the caption refinement. "Q", "A", "OC" and "MC" denote the question, ground truth answer, original GPT-4 synthesized caption, and modified caption after regeneration. + +based question-answering model. [15] propose PromptCap, a novel task-aware captioning model that uses a natural language prompt to control the generation of the visual content that can be used in conjunction with GPT-3 in-context learning. Img2LLM [11] is a zero-shot VQA method that generates image-relevant exemplar prompts for the LLM. Their key insight is that synthetic question-answer pairs can be generated using image captioning and question-generation techniques as in-context exemplars from the provided image. Prophet [30] proposes to prompt GPT-3 with answer heuristics (answer candidates and answer-aware examples) encoded into the prompts to enable GPT-3 to comprehend the task better, thus enhancing its capacity. + +# 3. Method + +# 3.1. Overview + +The general pipeline of our method is illustrated in Figure 2, including a Question-Aware Captioner (QACap) and an LLM. Specifically, the QACap model first generates a question-aware caption (i.e., context) by taking in both the image and the question. The prompt provided to the LLM includes the task instruction, context, and question, after which the LLM generates an answer based on both image-relevant and background knowledge it has. Due to lacking such data (i.e., image-question-context triplet), we carefully construct a safe, bias-free, high-quality dataset. Moreover, to enhance question-awareness for the QACap model, we design a question-aware feature distilling module and a question-aware regional consistency learning strategy. We will provide more details in the following section. + +# 3.2. QACap Dataset + +# 3.2.1. Training Examples Generation and Refinement + +For QACap training, we annotate the image-question-context triplet on top of KBVQA datasets. To avoid costly human annotation, we use GPT-4 to create a high-quality + +question-aware caption dataset based on OKVQA [24], A-OKVQA[29], and FVQA[38], totaling 29,569 entries. Concretely, we generally follow the PromptCap [15] pipeline for dataset annotation. Images in these three KBVQA training splits are sourced from the MSCOCO dataset, allowing us to use the original 5 human-annotated COCO captions as initial scene descriptions. These captions, along with the question-answer pair, are provided to GPT-4, prompting it to fully interpret the scene and question context. GPT-4 then generates 3 candidate captions that align semantically with the originals and contain helpful information for QA. + +After the synthesis process, we identify two main issues. First, there is a risk of directly injecting the answer information into the caption, especially when it isn't visible in the image. For example, figure 3 shows two cases where this occurs. In the top case, the answer '18' is inferred from the scene of people drinking and background knowledge about Canada's legal drinking age. A robust question-aware caption should simply focus on the drinking event in the image without inferring the answer from the question, as this could introduce significant bias into the model. Moreover, GPT-4-generated captions often display issues like redundancy, verosity and unnatural expressions. We address these issues by expressing them in a GPT-4 prompt and requesting the model regenerate new captions. In practice, we find that both issues are related to particular forms of phrasing, such as 'known for' or 'demonstrating'. We collect a word list of this kind and filter out captions containing these improper expressions. This subset is refined by emphasizing the above problems in the GPT-4 prompt. Once revised in this manner the dataset shows marked improvements in caption quality. The prompts and more refinement cases can be seen in the supplementary materials. + +# 3.2.2. Training Examples Selection + +To select the best synthetic caption, we evaluate two aspects: helpfulness and truthness. Helpfulness means the target caption should aid the LLM to predict a correct answer, truthness means the target caption should accurately describe the image's content and should not include inconsistent objects or not related expressions. Figure 4 illustrates the general selection pipeline. First, we follow our VQA pipeline to get the LLM's answer with each candidate caption. The answer score represents the count of ground truth s appearing in the LLM response. Then, we calculate the clip score of the image and candidate captions using a CLIP model[28]. The higher clip score means the caption is semantically closer to the image. Among the candidates achieving the same high answer score, the caption with the highest clip score will be selected as the final training instance. + +![](images/4fc8c192fd613abebef2ed4828ff3737a35f27060d3146d8c4c876cc9c53c215.jpg) +Figure 4. Our evaluation pipeline to select the best caption as training instance from synthesized candidates. We mainly evaluate the caption from the aspective of helpfulness and truthness. + +# 3.3. Question-Aware Captioner + +In our pipeline, we need a lightweight visual module that can convert images into question-relevant captions. In light of this, we introduce Question-Aware Captioner (QACap), a question-aware image captioning model finetuned from BLIP [19]. QACap takes a question and an image as input and outputs a caption about the image specific to the question. The caption describes related visual content and provides relevant information for LLMs to infer the answer. Formally, given a training sample $D = \{Q, I, C\}$ in which $Q$ is the input question, $I$ is the image, and $C$ is the synthesized question-aware caption. The overall architecture is shown in Fig. 5, which contains three components: (a) an image text encoding module, (b) a question-aware feature distilling module, and (c) a caption generation module. + +# 3.3.1. Image Encoding + +Given an image $I$ , we use a pre-trained vision transformer module [8] to extract the feature representations, which are denoted as $F_{v}$ ( $F_{v} \in \mathbb{R}^{L_{v} \times d}$ ). Here, $L_{v}$ and $d$ indicate the number of patches and dimensions, respectively. + +# 3.3.2. Question-Aware Feature Distilling + +The goal of this module is to adjust the image features based on the input question, enhancing the features of image regions that are relevant to the question while diminishing those of irrelevant regions. Additionally, it should also incorporate the semantic features of the question. First, we use the BERT [7] tokenizer to tokenize the sentence and get the sentence embedding $F_{q} \in \mathbb{R}^{L_{q} \times d}$ . After that, we pass the question features through the multi-head self-attention (MHSA) [35] layer to fully integrate the contextual semantic information: + +$$ +\hat {F} _ {q} = \mathrm {M H S A} (F _ {q}, F _ {q}, F _ {q}). \tag {1} +$$ + +With the question feature $\hat{F}_q$ , we compute the question-relevant image feature on each image by the multi-head + +![](images/630b5510a7d042aa3a73acfd5532c00b55ea327a4a5f5955a46003941894bf61.jpg) +Figure 5. Architecture of the proposed QACap model. Orange squares indicate the same sentence token in QAFD. We first perform (a) Image Text Encoding on the input image and question followed by (b) Question-aware Feature Distilling (QAFD), which adjusts the image features based on the input question, enhancing the features of image regions that are relevant to the question while diminishing those of irrelevant regions. These are fed into the (c) Caption Generation module. During the training phase, we apply a Question-Aware Regional Consistency Learning (QARCL) strategy $\mathcal{L}_{\mathrm{KL}}$ which is a JS Divergence loss to align the cross-attention map from the last cross attention layer in question-aware feature distilling and caption generation. + +cross-attention (MHCA) [35] mechanism: + +$$ +\tilde {F} _ {q} = \operatorname {M H C A} \left(\hat {F} _ {q}, F _ {v}, F _ {v}\right). \tag {2} +$$ + +Then, we take cross-attention matrix $A \in \mathbb{R}^{L_q \times L_v}$ as the channel correlation matrix between $\hat{F}_q$ and $F_v$ . The differences in each column represent how much attention the word pays to the image patch. Thus, we consider using the maximum one in each column as the token-wise maximum similarity of the image patches with the complete query: + +$$ +S _ {q} = \max (A) = \left[ \max \left(s _ {. 1}\right), \max \left(s _ {. 2}\right), \dots , \max \left(s _ {. L _ {v}}\right) \right]. \tag {3} +$$ + +Beyond token-wise matching similarity, we also consider the overall semantic similarity between the question and the image. We start by calculating the sentence representation using a max-pooling function: + +$$ +\hat {F} _ {q} ^ {s} = \operatorname * {m a x - p o o l i n g} (\hat {F} _ {q}). \tag {4} +$$ + +Then, we compute the cosine similarity between the question and the image features to get sentence-wise semantic similarity: + +$$ +S _ {w} = \cos \left(F _ {v}, \hat {F} _ {q} ^ {s}\right). \tag {5} +$$ + +We capture the token-wise and sentence-wise similarities through average pooling of $S_{w}$ and $S_{q}$ : + +$$ +S _ {o} = \left(S _ {w} + S _ {q}\right) / 2, \tag {6} +$$ + +and enhance the image features with an element-wise product using $S_{o}$ : + +$$ +\hat {F} _ {v} = F _ {v} \odot S _ {o}. \tag {7} +$$ + +A weighted mean approach controls the original image feature information, strengthening the image features while minimizing noise from excessive cross-modal information. The process is as follows: + +$$ +\tilde {F} _ {v} = \alpha F _ {v} + (1 - \alpha) \hat {F} _ {v}. \tag {8} +$$ + +where $\alpha$ stands for the mean weighted factor. + +# 3.3.3. Caption Generation + +After obtaining $\bar{F}_v\in \mathbb{R}^{L_v\times D}$ , we employ a transformer decoder to generate a caption. Here, we adopt the prompt enhanced tuning (PET) inspired by [15] to generate question-aware captions. Specifically, we first rewrite the caption with a prompt that includes the question at the beginning, + +such as, "Please describe this image according to the following question: What toy is this?". We then extract embedding features for all $m$ words in the captions $C$ and apply masked self-attention to model relationships between these word features. Next, we concatenate the question features $\tilde{F}_q$ and question-aware image features $\tilde{F}_v$ to get question-aware token sequence $\tilde{F}_t$ . Cross-attention is then used to capture the interactions between word features and then the question-aware token sequence, allowing the model to focus on the relevant features during word generation. These selected features are passed through a feed-forward network to produce an enhanced caption representation, denoted as $\hat{C} \in \mathbb{R}^{m \times d}$ . Finally, the probability distribution of the words is computed by a single hidden layer: + +$$ +W = \operatorname {S o f t m a x} \left(\hat {C} W _ {c} + b _ {c}\right). \tag {9} +$$ + +where $W_{c} \in \mathbb{R}^{D \times u}$ and $b_{c} \in \mathbb{R}^{u}$ are the learnable parameters, and $u$ represents the vocabulary size. + +Question-aware Regional Consistency Learning To ensure that the image regions considered during caption generation align with those relevant to the question in question-aware feature distilling, we introduce a question-aware region consistency loss. We extract the cross-attention maps from the last cross-attention layer in both question-aware feature distilling and caption generation, denoted as $A_{q}$ and $A_{c}$ . A Jensen-Shannon (JS) divergence loss is then applied to these attention maps, ensuring that the attended regions during caption generation and question understanding are consistent: + +$$ +\begin{array}{l} \mathcal {L} _ {J S} = \frac {1}{2} \left(D _ {K L} \left(A _ {q} \| M\right) + D _ {K L} \left(A _ {c} \| M\right)\right), \tag {10} \\ \text {w h e r e}, M = \frac {1}{2} \left(A _ {q} + A _ {c}\right). \\ \end{array} +$$ + +Here, $D_{KL}$ is the Kullback-Leibler (KL) divergence and $M$ is the average of the two attention maps. This loss encourages better alignment between the image areas focused on by the model when interpreting the question and those used during caption generation, thereby enhancing the relevance of generated captions to the question. + +# 3.4. Joint Training + +We train our method in an end-to-end manner by maximizing the likelihood of the observed word sequence. Given the ground truth words $(w_{1}^{*},\dots,w_{m}^{*})$ , we aim to minimize the negative log-likelihood loss, which can be expressed as: + +$$ +\mathcal {L} _ {\text {c a p}} = - \sum_ {t = 1} ^ {m} \log p _ {\theta} \left(w _ {t} ^ {*} \mid w _ {< t} ^ {*}\right), \tag {11} +$$ + +where $p_{\theta}(w_t^* | w_{MethodsImage RepresentationLarge Language ModelICL shotsAccuracy(%)BLIP-2[20]FeatureFlanT5045.9Flamingo[1]FeatureChinchilla1657.8CogVLM[39]FeatureVicuna-7B058.9CogAgent[14]FeatureVicuna-7B061.2InstructBLIP[5]FeatureVicuna-7B062.1PaLI-X-VPD[16]FeaturePaLI-X066.8KCKR[12]FeatureT5-XL065.1PICa-Full[42]Caption+TagsGPT-31648.0Img2LLM [11]Caption+TagsOPT1645.6KAT[10]Caption+Tags+FeatureGPT-3+Wikidata054.4REVIVE[21]Caption+FeatureGPT-3+Wikidata1658.0PromptCap[15]CaptionGPT-33260.4TOA[41]Caption+Visual EvidenceChatGPT2460.6Prophet[30]Caption+Answer HeuristicsGPT-32061.1QACapCaptionClaude 3.5068.2 + +Table 2. Results comparison with existing systems on A-OKVQA validation set. There are two evaluations, namely Multiple Choice and Direct Answer. Both are measured by accuracy(%) + +
MethodDirect AnswerMultiple ChoiceAvg
ClipCap [29]44.018.131.0
Pythia [17]49.025.237.1
ViLBERT [23]49.130.640.0
LXMERT [32]51.430.741.0
KRISP [25]51.933.742.8
GPV-2 [18]60.348.654.5
PromptCap + GPT-3[15]56.373.264.8
QACap + Claude3.566.376.771.5
+ +knowledge base and zero in-context learning shots, achieves a state-of-the-art performance accuracy of $68.2\%$ , surpassing all existing methods on OKVQA. We list some representative successful cases of our method in the appendix. + +Results on A-OKVQA. To further demonstrate the effectiveness of our method, we conduct additional tests on the A-OKVQA validation datasets. A-OKVQA is a larger knowledge-based VQA dataset, with questions in the form of both Multiple Choice(MC) and Direct Answer(DA). Table 2 shows the comparative results on the challenging A-OKVQA dataset. QACap with Claude3.5 reaches $76.7\%$ on Multiple Choice setting and $66.2\%$ on Direct Answer setting, achieving a sate-of-the-art average accuracy of $71.5\%$ . Our performance on Direct Answer part surpasses PromptCap by a large margin, showing strong capability of handling open-ended question answering. + +# 4.3. Ablation Study + +# 4.3.1. Effect of Proposed Module + +We conduct ablation studies of proposed modules on OKVQA and A-OKVQA val dataset as shown in Table 3. (a) stands for the baseline is BLIP fine-tuned on the PromptCap dataset and our constructed QACap dataset. To prove + +the necessity of question guidance in image feature extraction, we compare our method with question-aware feature distilling (QAFD) to baseline. Comparison between (a) and (b) shows a $1.4\%$ improvement in terms of accuracy on OKVQA, which illustrates that enhancing the image features with question guidance can lead to question-aware captions, thus improving VQA performance. Besides, to illustrate the effect of question-aware regional consistency learning strategy (QARCL), we compare the baseline with/without QARCL. From the comparison between (d) and (e), we note that QARCL can bring $2.1\%$ performance gain, which indicates the importance of modeling the focus area in the image in question understanding and caption generation process. We also notice that from the comparison of (b, c), the performance declines. The results show that QARCL alone leads to a negative improvement, but when combined with PET, it results in an improvement, indicating that PET is necessary first to make the decoder understand the meaning of the question in caption generation. In addition, we illustrate the impact of prompt-enhanced tuning (PET) which can lead to a $3.2\%$ improvement from the comparison between (b) and (d). This result demonstrates that PET can make the decoder understand the question and better generate question-aware captions to boost VQA performance. + +# 4.3.2. Effect of Language Models + +One notable advantage of our method is its flexibility in leveraging the latest and most powerful LLMs. Consequently, we focus on selecting top-performing LLMs from both open-source and closed-source models. Among opensource options, we test the newest Qwen2.5 [33] and Intern2.5 [3] series. For the closed-source models, we use Claude3.5, DeepSeek-V2.5 [6], and GPT series [27] via API access. Tests are conducted on both OKVQA and A- + +Table 3. Ablation studies of proposed modules on OKVQA val and A-OKVQA val split. QAFD stands for question-aware feature distilling, QARCL stands for question-aware regional consistency learning strategy, and PET stands for prompt-enhanced tuning. + +
SettingBaselineQAFDQARCLPETOKVQAA-OKVQA
DAMC
(a)✓61.552.467.7
(b)✓✓62.953.865.6
(c)✓✓✓60.952.944.1
(d)✓✓✓66.160.854.7
(e)✓✓✓✓68.266.269.2
+ +![](images/182ecd82eefe9900e3d6e19e54f295f8ce8ff82daff4b97a476768de090872b9.jpg) +Figure 6. QACap with different LLMs' accuracy on OKVQA val and AOKVQA val. + +OKVQA datasets, using the same captions generated by QACap and the same prompt. The results, presented in Figure 6, show that large language model performance varies across different datasets. Typically, language models perform worse on A-OKVQA than on OKVQA, though the GPT series exhibits the opposite trend, likely benefiting from their broader knowledge base. While model capabilities are often associated with parameter size, Intern2.5, with only 20 billion parameters, achieves the highest score of $74.8\%$ on the A-OKVQA Multiple Choice task, surpassing all larger models. Claude3.5 achieves the highest $68.2\%$ accuracy on OKVQA and places 3rd and 2nd on A-OKVQA DA and MC, respectively. As a result, we select Claude3.5 as the primary LLM for our other experiments. + +# 4.3.3. Effect of Training Data + +We train our QACap on two datasets: the PromptCap dataset and our self-built QACap dataset. As shown in Table 4, our high-quality dataset consistently improves KBVQA performance. With additional fine-tuning on the QACap dataset, our QACap paired with Claude3.5 achieves a $2.9\%$ improvement on OKVQA, a $5\%$ increase on A-OKVQA DA, and a $0.2\%$ gain on A-OKVQA MC. To further validate our dataset's quality, we randomly sample an equivalent amount of data from the PromptCap dataset to match the size of the QACap dataset. We then fine-tune the QACap model on each dataset separately. The model fine-tuned solely on our dataset outperforms the model fine-tuned on PromptCap by $3.5\%$ , $4.8\%$ , and $1.5\%$ , respectively. + +Table 4. Impact of different training datasets. "Datasets" means datasets used for training. The results are tested on OKVQA and A-OKVQA validation sets. The used LLM is Claude3.5. + +
DatasetsOKVQAA-OKVQA
DAMC
PromptCap65.361.269.0
PromptCap + QACap68.2 (+2.9)66.2 (+5.0)69.2 (+0.2)
PromptCap (29k)64.056.768.2
QACap67.5 (+3.5)61.5 (+4.8)69.7 (+1.5)
+ +![](images/b895cf5233510882b35e769f589042b3d698846f788cad13f0d7d7616bfb28b2.jpg) +Figure 7. Ablation on the caption type and prompt type. + +![](images/2e6a936563a3c5d9c1eebd4b41fefd7425158e26332947c8dedb272e20feeea6.jpg) + +# 4.3.4. Ablation on Caption Type and Prompt Design + +As we define the core ability as focusing on question-relevant visual information, we explore the difference between general captions and question-aware captions. We use BLIP to generate general captions and QACap to generate question-aware captions. Figure 7's left part shows question-aware captions outperform general captions extensively in the Direct Answer setting on both datasets, though they provide only a minor advantage in the Multiple Choice setting. This is because, with the question and answer choices provided, the LLM can guess the correct answer without detailed context, thereby reducing the impact of an improved caption. As discussed in Section 4.1, the traditional VQA paradigm typically yields a single word or phrase as an answer. However, open-ended answers that include reasonable inferences are increasingly popular and desirable. We generate both types of answers by editing the LLM prompt (See the Appendix). The right part of Figure 7 demonstrates that open-ended answers achieve higher accuracy while also being more readable and explainable. + +# 5. Conclusion + +We propose QACap, a new question-aware image captioning model to capture relevant information from images based on questions. To improve performance, we construct a high-quality training dataset using GPT-4. The pipeline, QACap with GPT-4, works effectively on knowledge-based VQA tasks. Our system achieves SoTA accuracy of $68.2\%$ on OK-VQA, and an average of $71.5\%$ on A-OKVQA. Ablations show that the QACap dataset gives consistent gains over the PromptCap dataset. QACap is a lightweight, plug-and-play visual module that connects with LLMs. + +# 6. Acknowledgement + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 7 +[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 1, 6 +[3] Zheng Cai, Maosong Cao, Haojiong Chen, et al. Internl m2 technical report, 2024. 7 +[4] Gaoxiang Cong, Jiadong Pan, Liang Li, Yuankai Qi, Yuxin Peng, Anton van den Hengel, Jian Yang, and Qingming Huang. Emodubber: Towards high quality and emotion controllable movie dubbing. arXiv preprint arXiv:2412.08988, 2024.3 +[5] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 7 +[6] DeepSeek-AI. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model, 2024. 7 +[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4 +[8] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4 +[9] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1 +[10] Liangke Gui, Borui Wang, Qiuyuan Huang, Alex Hauptmann, Yonatan Bisk, and Jianfeng Gao. Kat: A knowledge augmented transformer for vision-and-language. arXiv preprint arXiv:2112.08614, 2021. 3, 7 +[11] Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From images to textual prompts: Zero-shot visual question answering with frozen large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10867-10877, 2023. 2, 3, 7 + +This work was supported by National Natural Science Foundation of China: 62322211, 62336008, “Pioneer” and “Leading Goose” R&D Program of Zhejiang Province (2024C01023). Qi Chen, Yuankai Qi and Anton van den Hengel are not supported by the aforementioned fundings. +[12] Dongze Hao, Jian Jia, Longteng Guo, Qunbo Wang, Te Yang, Yan Li, Yanhua Cheng, Bo Wang, Quan Chen, Han Li, and Jing Liu. Knowledge condensation and reasoning for knowledge-based vqa, 2024. 7 +[13] SUN Haoran, WANG Yang, LIU Haipeng, and QIAN Biao. Fine-grained cross-modal fusion based refinement for text-to-image synthesis. Chinese Journal of Electronics, 32(6): 1329-1340, 2023. 3 +[14] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. arXiv preprint arXiv:2312.08914, 2023. 7 +[15] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 2, 3, 4, 5, 6, 7 +[16] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. arXiv preprint arXiv:2312.03052, 2023. 7 +[17] Yu Jiang, Vivek Natarajan, Xinlei Chen, Marcus Rohrbach, Dhruv Batra, and Devi Parikh. Pythia v0.1: the winning entry to the vqa challenge 2018, 2018. 7 +[18] Anita Kamath, Christopher Clark, Tanmay Gupta, Eric Kolve, Derek Hoiem, and Aniruddha Kembhavi. Webly supervised concept expansion for general purpose vision models, 2022. 7 +[19] 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. 4 +[20] 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. 7 +[21] Yuanze Lin, Yujia Xie, Dongdong Chen, Yichong Xu, Chenguang Zhu, and Lu Yuan. Revive: Regional visual representation matters in knowledge-based visual question answering. Advances in Neural Information Processing Systems, 35:10560-10571, 2022. 3, 7 +[22] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Zechao Li, Qi Tian, and Qingming Huang. Entity-enhanced adaptive reconstruction network for weakly supervised referring expression grounding. IEEE PAMI, 45(3):3003-3018, 2023. 3 +[23] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks, 2019. 7 +[24] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 1, 3, 4 + +[25] Kenneth Marino, Xinlei Chen, Devi Parikh, Abhinav Gupta, and Marcus Rohrbach. Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14111-14121, 2021. 1, 3, 7 +[26] Medhini Narasimhan, Svetlana Lazebnik, and Alexander Schwing. Out of the box: Reasoning with graph convolution nets for factual visual question answering. Advances in neural information processing systems, 31, 2018. 3 +[27] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, et al. Gpt-4 technical report, 2024. 7 +[28] 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, 4 +[29] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European Conference on Computer Vision, pages 146-162. Springer, 2022. 1, 4, 7 +[30] Zhenwei Shao, Zhou Yu, Meng Wang, and Jun Yu. Prompting large language models with answer heuristics for knowledge-based visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14974-14983, 2023. 2, 3, 7 +[31] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In Proceedings of the AAAI conference on artificial intelligence, 2017. 3 +[32] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers, 2019. 7 +[33] Qwen Team. Qwen2.5: A party of foundation models, 2024. 7 +[34] Yunbin Tu, Liang Li, Li Su, Zheng-Jun Zha, and Qingming Huang. SMART: syntax-calibrated multi-aspect relation transformer for change captioning. IEEE PAMI, 46(7): 4926-4943, 2024. 3 +[35] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4, 5 +[36] Denny Vrandecic and Markus Krötzsch. Wikidata: a free collaborative knowledgebase. Communications of the ACM, 57(10):78-85, 2014. 3 +[37] Peng Wang, Qi Wu, Chunhua Shen, Anton van den Hengel, and Anthony Dick. Explicit knowledge-based reasoning for visual question answering. arXiv preprint arXiv:1511.02570, 2015. 1, 3 +[38] Peng Wang, Qi Wu, Chunhua Shen, Anthony Dick, and Anton Van Den Hengel. Fvqa: Fact-based visual question answering. IEEE transactions on pattern analysis and machine intelligence, 40(10):2413-2427, 2017. 1, 3, 4 +[39] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 7 + +[40] Jialin Wu, Jiasen Lu, Ashish Sabharwal, and Roozbeh Mottaghi. Multi-modal answer validation for knowledge-based vqa. In Proceedings of the AAAI conference on artificial intelligence, pages 2712-2721, 2022. 1, 3 +[41] Xiaoying Xing, Mingfu Liang, and Ying Wu. Toa: Task-oriented active vqa. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 7 +[42] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3081–3089, 2022. 2, 3, 7 +[43] Zhaoda Ye, Xiangteng He, and Yuxin Peng. Unsupervised cross-media hashing learning via knowledge graph. Chinese Journal of Electronics, 31(6):1081-1091, 2022. 3 +[44] Beichen Zhang, Liang Li, Shuhui Wang, Shaofei Cai, Zheng-Jun Zha, Qi Tian, and Qingming Huang. Inductive state-relabeling adversarial active learning with heuristic clique rescaling. IEEE PAMI, 46(12):9780-9796, 2024. 3 +[45] Zihao Zhu, Jing Yu, Yujing Wang, Yajing Sun, Yue Hu, and Qi Wu. Mucko: multi-layer cross-modal knowledge reasoning for fact-based visual question answering. arXiv preprint arXiv:2006.09073, 2020. 1, 3 \ No newline at end of file diff --git a/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/images.zip b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e74c86ae74fff89de92121fdcbb3723c01992a44 --- /dev/null +++ b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43b3079e72242b48070c0e1f6bebed41b8832ee743ab4ec80e6079097b94eaf1 +size 537947 diff --git a/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/layout.json b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ea2a452a982f4e5c227f6298a3177a87aaac197a --- /dev/null +++ b/CVPR/2025/Separation of Powers_ On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ecc686c5a79b6e03ec0d16f30b3f4799887a106ca2be118fd2ac7f4085a9cbd +size 377794 diff --git a/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/767e680b-2744-4105-a84b-e4f95953240e_content_list.json b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/767e680b-2744-4105-a84b-e4f95953240e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f19cb7a95f8739eac85e4f3f1326f637f990ebc2 --- /dev/null +++ b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/767e680b-2744-4105-a84b-e4f95953240e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a68244933e65dd5d22bdd84cbe914bca4d2f7d195c14c4d895a3a238a16641a +size 75210 diff --git a/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/767e680b-2744-4105-a84b-e4f95953240e_model.json b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/767e680b-2744-4105-a84b-e4f95953240e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..335f7dbce490907d8b1d4ade4ef79068a66125b6 --- /dev/null +++ b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/767e680b-2744-4105-a84b-e4f95953240e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1550693b5d28968a565e55eb7c42863b074166b7455b88327bda6495199a3897 +size 94764 diff --git a/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/767e680b-2744-4105-a84b-e4f95953240e_origin.pdf b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/767e680b-2744-4105-a84b-e4f95953240e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9ba1a5e109f5e36351987bb51c681ec79b6e3b75 --- /dev/null +++ b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/767e680b-2744-4105-a84b-e4f95953240e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcc01172c88fde268f54098ddc281171c3f0ba6c9c811435a5b352d13b5adbda +size 2670189 diff --git a/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/full.md b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..392d149a53dac2fbeb5b32c11d9a9795d9054251 --- /dev/null +++ b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/full.md @@ -0,0 +1,304 @@ +# Seq2Time: Sequential Knowledge Transfer for Video LLM Temporal Grounding + +Andong Deng $^{1,2,*}$ , Zhongpai Gao $^{2,\dagger}$ , Anwesa Choudhuri $^{2}$ , Benjamin Planche $^{2}$ , Meng Zheng $^{2}$ , Bin Wang $^{2,3}$ , Terrence Chen $^{2}$ , Chen Chen $^{1}$ , Ziyan Wu $^{2}$ + +1University of Central Florida, 2United Imaging Intelligence, 3Northwestern University + +# Abstract + +Temporal awareness is essential for video large language models (LLMs) to understand and reason about events within long videos, enabling applications like dense video captioning and temporal video grounding in a unified system. However, the scarcity of long videos with detailed captions and precise temporal annotations limits their temporal awareness. In this paper, we propose Seq2Time, a data-oriented training paradigm that leverages sequences of images and short video clips to enhance temporal awareness in long videos. By converting sequence positions into temporal annotations, we transform largescale image and clip captioning datasets into sequences that mimic the temporal structure of long videos, enabling self-supervised training with abundant time-sensitive data. To enable sequence-to-time knowledge transfer, we introduce a novel time representation that unifies positional information across image sequences, clip sequences, and long videos. Experiments demonstrate the effectiveness of our method, achieving a $27.6\%$ improvement in F1 score and $44.8\%$ in CIDEr on the YouCook2 benchmark and a $14.7\%$ increase in recall on the Charades-STA benchmark compared to the baseline. Project available at: https://seq2time.github.io/ + +# 1. Introduction + +Developing versatile video large language models (LLMs) capable of perceiving temporal dynamics in long-term videos is crucial for advancing toward artificial general intelligence (AGI) and has numerous impactful real-world applications, e.g., enabling intelligent surveillance systems to detect anomalous activities, assisting in medical diagnostics through the analysis of patient behavior, and facilitating ad + +vanced sports analytics by recognizing patterns and strategies over entire games. + +Building upon well-trained LLMs [2, 7, 8, 14, 39], visual instruction tuning [24]—which projects visual representations into the language embedding space—has significantly advanced the ability to perform various multimodal tasks, such as image captioning [18, 23, 42], video question answering [6, 19, 20, 26, 37], and 3D perception [11, 17, 31]. The success of multimodal LLMs (MLLMs) in general vision-language tasks has further inspired the development of time-sensitive MLLMs [10, 12, 13, 29, 30, 34] for long-term video understanding, such as dense video captioning [15] and temporal video grounding [1, 9]. + +Creating time-sensitive instruction tuning data from existing video datasets with timestamp annotations [1, 9, 15, 32, 40] is essential for temporal grounding. However, current time-related video datasets face significant limitations. First, they are limited in scale due to the labor-intensive process of timestamp annotation; datasets like TimeIT [30] and VTG-IT [10] contain only around $125\mathrm{K}$ and $120\mathrm{K}$ videos respectively, which is substantially fewer than general video instruction tuning datasets without detailed time annotation, e.g., VideoChat2 [20] includes over $800\mathrm{K}$ videos. However, the number of training data is crucial in time-sensitive MLLM training. As illustrated in Table 1, when randomly removing $12\%$ task-irrelevant training data for TimeChat [30], the average performance of dense video captioning on YouCook2 could drop $13.4\%$ ; while intentionally removing task-relevant training data (e.g., dense video captioning), the performance sharply degrades $65.5\%$ . Second, the captions in these datasets often lack richness and precision; the visual-language alignment is less accurate compared to high-quality datasets. For example, LLaVA-Instruct-150K [24] provides detailed image captions and specifies key object locations, greatly enhancing perception during instruction following. These limitations—the constrained dataset sizes and suboptimal caption quality—not only hinder the training of temporally aware video LLMs but also impede their performance in under + +![](images/a988c45072643e1c099ca307311542bbd77ea40db4c5d5644d76109ff09e3d27.jpg) +Figure 1. Illustration of our proposed Seq2Time. While existing approaches [10, 30] rely on timestamp-annotated long videos, Seq2Time leverages image and short video datasets. We introduce a novel time representation of unified relative position token that bridges different sequence types by encoding positions as 4-digit codes—each digit becomes a learnable embedding in the language space, enabling seamless knowledge transfer in the LLM embedding space. + +standing and reasoning over long-term videos. + +To address these challenges, we propose Seq2Time, a data-oriented training paradigm that leverages sequences of independent images and short video clips to enhance temporal grounding in long videos. Our intuition is that MLLMs do not inherently "perceive time" but recognize the correspondence between visual content and its position in a sequence. Therefore, we construct two types of data: i) image sequence data and ii) short video clip sequence data. For the image sequence data (IS), based on correlations between image captions and their indices in the sequence, we design three sophisticated pretext tasks: a) image index grounding, b) indexed image captioning and c) adjacent location reasoning, enforcing the MLLM to link the sequential location of a specific image to its textual content and vice versa. For the clip sequence data (CS), we first generate captions using LongVA [38] given the video and its action labels; then, we combine several short clips along with their sequential positions to form a long video. Similarly, the model is trained to learn the correspondence between the captions of the short clips and their sequential positions. Due to the large scale of image [16] and short video clip [4] datasets, we can easily sample a vast number of long sequences with high-quality caption annotations, significantly improving the quality and quantity of the instruction tuning training data. + +It is worth noting that the image sequence data and clip + +sequence data in Seq2Time offer distinct advantages. The three pretext tasks associated with image sequences are more challenging than the two downstream tasks, as they require locating one or few specific images within a long sequence. Additionally, the quality of image captions is generally higher than that of video data. On the other hand, the clip sequence data aligns more closely with real-world long video sequences, both in terms of data characteristics and training objectives. Overall, the two types of data are complementary in our Seq2Time. + +Importantly, to facilitate seamless knowledge transfer from image and clip sequence to temporal understanding in real real-world settings, we introduce a novel time representation called the unified relative position token. This representation unifies positional information across image sequences, clip sequences, and long videos within the LLM embedding space. By encoding the relative positions of images or clips as tokens, we enable the model to generalize the concept of position in a sequence to temporal locations in videos. This unification allows the video LLMs to interpret positional cues consistently across different modalities, effectively bridging the gap between artificial sequences and continuous video content. Our contributions are summarized as follows: + +- We propose Seq2Time, a novel data-oriented training paradigm that enhances temporal understanding in video + +LLMs through image and clip sequences, enabling effective temporal grounding in long videos without requiring extensive timestamp annotations. + +- We introduce three well-designed pretext tasks for image sequence data (image index grounding, indexed image captioning, and adjacent location reasoning) that effectively enhance MLLMs' ability to locate and describe temporal events in sequences. +- We develop a unified relative position token representation that bridges image indices and video timestamps in the LLM embedding space, facilitating efficient knowledge transfer between different sequence types and temporal understanding tasks. +- Experiments validate the effectiveness of Seq2Time, achieving significant improvements over the baseline model: $44.8\%$ gain in CIDEr and $27.6\%$ in F1 score on YouCook2, and $17.1\%$ improvement in R@1 (IoU=0.7) on Charades-STA. + +# 2. Related Work + +# 2.1. Visual Instruction Tuning Data + +The emergence of multimodal LLMs [6, 19, 24, 26, 42] has revolutionized vision-language understanding, largely due to advances in visual instruction tuning [24] and carefully curated instruction datasets [5, 17, 19, 24]. A significant milestone was established in [24], where they introduced a standardized protocol using GPT-4 [28] to transform COCO dataset [22] annotations into LLM-compatible instruction data. This approach was later extended to video understanding, with [19] leveraging WebVid-10M [3] to create comprehensive video-centric instruction tuning datasets. + +Several efforts have been made to create time-sensitive instruction-tuning datasets. LITA [13] developed specialized instruction data for temporal localization. TimeIT [30] focused on dense video captioning and grounding. However, these time-sensitive datasets, particularly TimeIT, face limitations in both scale and quality due to the resource-intensive nature of timestamp annotation. Our work addresses these challenges by introducing a novel approach that leverages images and short clips to create rich instruction tuning data, as illustrated in Figure 1, ultimately enhancing video LLMs' temporal understanding capabilities. + +# 2.2. Time-sensitive MLLMs + +Recent works have made significant strides in developing time-sensitive MLLMs through architectural innovations. TimeChat [30] introduces a dual Q-Former approach: one for timestamp embedding through cross-modal attention and another utilizing sliding windows for long video sequences. VTimeLLM [12] employs a three-stage pipeline that progressively learns visual-text alignment, event boundaries, and timestamp perception. Momen + +tor [29] emphasizes the significance of temporal representation and segment-level modeling, while HawkEye [34] utilizes coarse-grained clip descriptions as intermediate representations with recursive temporal grounding. VTGLLM [10] introduces absolute-time tokens to minimize quantization errors in time perception, alongside slot-based token compression to address context length limitations. Grounded-VideoLLM [33] combines relative time tokens with a two-stream visual encoding architecture for improved temporal understanding. While these approaches focus on architectural innovations, our work takes a fundamentally different approach by investigating the potential of sequence learning from images and short clips. We address the critical challenge of insufficient timestamp annotations in long video datasets by exploring, for the first time, how sequential knowledge can be effectively transferred to enhance temporal understanding in long videos. + +# 2.3. Time-aware Video Understanding + +Dense video captioning [15] and temporal video grounding [1, 9] represent tasks in video understanding that require precise temporal comprehension. Recent advances in dense video captioning have explored various innovative approaches: Zhu et al. [43] developed an end-to-end framework that integrates ASR signals with simultaneous timestamp and caption prediction, while Yang et al. [36] introduced a specialized token-based temporal framework trained on large-scale data. To enhance efficiency and modularity, Zhu et al. [41] proposed a clustering-based memory bank that can be readily integrated into existing architectures. While unified frameworks [21, 35] have emerged to handle multiple temporal understanding tasks simultaneously, these approaches are typically designed for specific video understanding tasks with dedicated architectures. In contrast, our work focuses on enhancing Video LLMs' temporal understanding capabilities while preserving their general-purpose reasoning abilities, enabling a more versatile and comprehensive solution for video understanding. + +# 3. Seq2Time + +Traditional approaches to training time-sensitive video LLMs have been constrained by the limited availability of temporal annotations in long video datasets. To address this challenge, we propose Seq2Time, a novel paradigm that leverages abundant image and short video data to create rich sequential training data with self-constructed temporal information. Our approach consists of three components: image sequence data (Section 3.1) that exploits index-caption correspondence in long sequences, clip sequence data (Section 3.2) that utilizes LongVA [38] to generate high-quality captions for short video clips, and a unified relative position token (Section 3.3) that bridges different types of sequential data in the LLM embedding space. + +![](images/514e192a2d67e3a9a529d21f885d12da846000c46bb3de8311e0d0ba359e5ee3.jpg) + +![](images/61166057a2f5f8fc3c6818164899484ee3845fec2d953ae6379c822942c6d8a7.jpg) +image8 + +Figure 2. Image sequence data in Seq2Time, featuring three complementary pretext tasks designed to leverage index-caption correspondence. IIG (Image Index Grounding) mimics temporal grounding through position prediction, IIC (Indexed Image Captioning) parallels dense video captioning, and ALR (Adjacent Location Reasoning) enhances sequential understanding through neighbor relationships. + +![](images/e12e0374df31f818b3303824110b20589a442bb00cf8f5535e4e2d1bd16c1a31.jpg) +image9 + + +![](images/5b0eefee095fd3bef48e59a743db7c76b48882947f7da0e1c187432fcab95ef1.jpg) +image10 + + + + +![](images/b5f8f2ba78633c604d3894e42a974e52cdc6dd695e81692ffc4168a50d1c4bc1.jpg) +image11 + + +![](images/d4961e85b9bf6826810d2cef4d4519d9f8c7159334e8f829fa7b61e33a67933c.jpg) +image69 + + +![](images/19f50c7f67a6cb1ce97bc3fa9be23e596334a8348eb7795ef3ac2800c2dab693.jpg) +image70 +caption70> + +![](images/c213dc9a3facb7fa37cffb61ffacfed8a14f2eb3904cfdb198fe94be5da63ab8.jpg) +image71 + + +![](images/ebb0dcc4a0d0cadd00bb216943e8a0a4d338c94a69bea65c8dd5ce8cbe38d5e6.jpg) +image72 + +96 images + +# Preview task 1: Image Index Grounding (IIG) + +# Single-target: + +Q: Which image matches the following description: ? Please output the image index. + +A: The image index is $<1><0><4><2>$ . + +# Multiple-target: + +Q: Which images match the following descriptions: , ? Please output the image indices in the same order as the captions. + +A: The image indices are $<1><1><4><6>, <7><5><0><0>$ . + +# Preview task 2: Indexed Image Captioning (IIC) + +# Single-target: + +Q: Please describe the image $<0><9><3><8>$ . + +A: The image $<0><9><3><8>$ describes + +# Multi-target: + +Q: Please describe the image $<0><9><3><8>$ and image $<7><3><9><6>$ . +A: The image $<0><9><3><8>$ describes and image $<7><3><9><6>$ describes . + +# Preview task3: Adjacent Location Reasoning (ALR) + +Q: What is the image right after the image describing $<$ caption69>? Please output the index and describe this image. +A: The image index is $< 7>< 2>< 9>< 2>$ . It describes + +
Training DataSODAcCIDErMeteorF1
Full1.02.91.112.7
w/o irr1.22.00.810.5
w/o dvc0.40.90.34.8
+ +Table 1. Impact of data loss on the performance of TimeChat on dense video captioning (YouCook2). We compare the full training data results and $12\%$ less training data. "irr" means task-irrelevant data, and "dvc" means dense video captioning data. The average performance drops $13.4\%$ and $65.5\%$ , respectively. + +# 3.1. Image Sequence Data + +We utilize three LLaVA-ReCap datasets [16] (COCO118K, BLIP558K, and CC3M) as our data source due to their rich textual content and consistent image-text alignment. To adapt the image data for downstream time-aware video understanding tasks, we design three pretext tasks: Image Index Grounding, Indexed Image Captioning, and Adjacent Location Reasoning, as shown in Figure 2, aimed at enhancing sequence localization abilities in Video LLMs. + +Image Index Grounding (IIG) trains the model to locate specific images within a sequence based on their descriptions. Given one or several captions, the model must identify the corresponding image index or indices, mimicking temporal grounding in videos. The input format is: Which image matches the description: $<\text{CAPTION}>$ ? Please output the image index, with output: The image index is $<\text{INDEX}>$ . + +Indexed Image Captioning (IIC) reverses this process by requiring the model to generate descriptions for specific indexed images, analogous to dense video captioning. A typical input is: Please describe the image with index , with output: The image with index describes . + +Adjacent Location Reasoning (ALR) challenges the model to understand sequential relationships by identifying and describing neighboring images. The task takes input: What is the image right before/after the image described as $$ ? Please provide the index and describe this image, and expects output: The image index is $$ . It describes $$ . By combining both index grounding and image description in a sequential context, this task strengthens the model's ability to perform both temporal localization and content understanding, which are essential for video understanding tasks like dense captioning and temporal grounding. + +In our implementation, we limit the maximum number of target images to 5 for both IIG and IIC tasks, with 10 different template variations for questions and answers (detailed templates in Supplementary). Each training sequence consists of 96 randomly sampled images, from which up to 5 are selected as targets for caption-based prompts. From our source datasets containing $\sim 3.7\mathrm{M}$ images, we randomly sample 300K training instances, evenly distributed across the three pretext tasks (100K each). This approach not only ensures rich textual descriptions to enhance text generation + +![](images/7dc3a47665d9129a09009f0ea79794c228ed03428ca5fe7897ae2ddd9dedf4a1.jpg) +Figure 3. Clip sequence data in Seq2Time. We use LongVA to generate captions for short video clips from Kinetics-700, then combine multiple clips from different action categories to simulate longer videos. Temporal annotations are derived from sequence positions of clips. The resulting sequences serve as training data for both temporal grounding and dense video captioning tasks. + +capabilities but also strengthens sequence localization abilities through our carefully designed pretext tasks. + +# 3.2. Clip Sequence Data + +To better simulate real long videos with multiple events, we utilize the Kinetics-700 [4] dataset to construct clip sequence data. We employ LongVA [38] to generate detailed descriptions for each clip, conditioning on the action labels to ensure high-quality captions. The captioned clips are then combined to create longer sequences, which serve as training data for both dense video captioning and temporal video grounding tasks. To create diverse temporal structures, we randomly sample between 2 to 10 clips per sequence, mimicking multi-event scenarios in long videos. While most Kinetics-700 clips are 10 seconds long, we intentionally avoid uniform temporal spacing by applying varying frame rates during sampling. This strategy, illustrated in Figure 3, ensures that our model learns to handle non-trivial temporal relationships and variable-length segments within sequences. + +# 3.3. Unified Relative Position Tokens + +Previous works have represented time in video LLMs using either free-form language (e.g., “2.05 seconds”) [30] or special learnable tokens (e.g., ) [10]. However, learning absolute time is often suboptimal for two reasons: 1) videos may have different frame rates, making absolute time less meaningful across datasets, and 2) the primary goal is to capture the correspondence between visual + +content and its relative sequential position. Furthermore, to enable effective sequence-to-time knowledge transfer, we need to bridge the semantic gap between image indices and video timestamps through a unified representation. + +We propose using relative position tokens that work seamlessly across both image sequences and video data (Figures 2 and fig:clipseq). Instead of using direct indices which could lead to complex token sequences (e.g., $<1><2>\ldots<99>\ldots$ ), we normalize each position (i.e., image or frame index) to a 4-digit decimal representation: $I_{\mathrm{norm}} = \text{round}\left(\frac{i}{L},4\right)$ , with $i$ the image (or frame) index and $L$ the sequence length. For example, in a 96-image sequence, the 7th image's normalized index would be 0.0729, encoded as $<0><7><2><9>$ (Figure 1). The 4-digit precision is carefully chosen to balance accuracy and complexity. For a typical 1-minute video at 30 fps sampled at 96 frames, this representation achieves an average temporal error of just $0.13\%$ . More specifically, each digit serves a different purpose: the first digit represents the coarse position, while subsequent digits provide increasingly fine-grained localization within that section. This hierarchical structure helps the model learn temporal relationships at different scales. + +Our approach introduces only 10 new tokens ( $<0>$ through $<9>$ ) to the LLM vocabulary, making it highly efficient for training while providing a unified representation that bridges image sequences and video timestamps in the LLM embedding space. During inference, absolute timestamps can be easily reconstructed from these relative positions using the video frame rate, enabling precise temporal localization without the need for learning video-specific + +timing patterns. + +# 4. Experiments + +# 4.1. Experiment Setup + +Training Data Our training dataset comprises three main components: 1) Seq2Time instruction data, consisting of 300K image sequence instances and 100K clip sequence instances; 2) 110K temporal understanding samples from TimeIT [30]; and 3) general video caption data, including 40K instances from Valley [25] and 93K from ShareGPT4Video [5]. For Seq2Time data, we employ balanced sampling across tasks: the image sequence tasks (IIG, IIC, and ALR) are randomly selected with equal probability; and similarly, the clip sequence tasks (dense video captioning and temporal video grounding) are sampled uniformly. + +Implementation Details We evaluate Seq2Time using TimeChat [30] as our baseline model. The training process consists of two phases: initial training for 1 epoch with the complete dataset, followed by 3 epochs of fine-tuning using only TimeIT and Valley (40K) data. Training parameters include a LoRA rank of 32, batch size of 8, and 96 sampled frames per video, following TimeChat's configuration. All experiments are conducted on 8 A100 GPUs (80GB each). For evaluation, we follow TimeChat's metrics: dense video captioning is assessed using $SODA_c$ , $CIDEr$ , and Meteor for text quality, and $F1$ score for temporal accuracy. Temporal video grounding is evaluated using $R@1$ with IoU thresholds of 0.5 and 0.7. Note that our reproduced baseline results are slightly lower than those reported in [30] due to partial Valley [25] data availability. + +# 4.2. Main Results + +We evaluate Seq2Time on two standard benchmarks for temporal understanding: YouCook2 for dense video captioning and Charades-STA for temporal video grounding (Table 2). Our approach demonstrates consistent improvements across all metrics compared to the TimeChat baseline, even without the unified relative position token (RPT). The improvements are particularly notable in YouCook2's CIDEr and F1 scores, and Charades-STA's R@1(IoU=0.5), validating that our sequence-based training effectively enhances temporal understanding capabilities. + +The addition of RPT further boosts performance, yielding significant improvements: $+30\%$ in SODA_c and $+27.6\%$ in F1 score on YouCook2, and $+17.1\%$ in R@1(IoU=0.7) on Charades-STA. Notably, while TimeChat relies on approximately 120K manually annotated temporal samples, Seq2Time achieves superior performance using largely self-supervised data, leveraging the natural sequential correspondence between images/clips + +and their captions. We provide detailed ablation studies of individual components in the following section. Additional results of VTG-LLM can be found in Supplementary. + +# 4.3. Dataset Effect + +To analyze how different data components influence temporal understanding in video LLMs, we conduct comprehensive ablation studies on TimeChat, as shown in Table 3. Removing clip sequence (CS) data leads to performance drops in both dense video captioning and temporal grounding tasks, demonstrating its importance in the training pipeline. + +Interestingly, incorporating additional video captions (MC) from ShareGPT4video [5] shows unexpected effects: while it doesn't significantly improve captioning quality, it substantially enhances temporal perception, yielding a $19.5\%$ improvement in YouCook2's F1 score and $7.3\%$ in Charades-STA's R@1(IoU=0.5). We hypothesize that exposure to more captions improves the model's ability to understand long-form video content, leading to better event boundary detection. The impact of image sequence data (IS) is particularly striking, especially in text generation quality. The CIDEr score shows a remarkable $48.3\%$ improvement over the baseline, providing strong evidence that temporal understanding can be effectively learned from static image sequences. Notably, while the addition of MC and CS improves temporal localization, it results in a slight degradation in text generation quality. This may be because image sequence data inherently provides stronger sequential visual-language alignment (i.e. per-image captions), whereas video caption data represents visual-language correspondence at the video level. Additionally, the limitations of LongVA [38] may also impact the quality of clip-level captions, further contributing to this effect. + +To further validate the effectiveness of the image sequence data, we tested our image sequence data on VideoLLaMA [6], a general video LLM without prior temporal training. As shown in Table 4, while the base model struggles with temporal events (F1 score of 0.2), adding image sequence (w/ IS) improves this to 3.3, showcasing initial temporal awareness. Notably, this improvement occurs even without our unified relative position tokens, using only free-form language for index representation. This confirms that sequential knowledge can be effectively transferred across different data modalities. + +In addition to dense video captioning performance with common metrics, we also evaluate the impact of different training datasets on the text richness of the generated captions in two aspects: generated caption length and lexical diversity. As shown in Table 5, adding new training data has a noticeable effect on the text richness of the captions. To assess text richness, we measure the average caption length $(L_{\mathrm{avg}})$ and the Type-Token Ratio (TTR) [27], which quantifies lexical diversity. As indicated in the results, incor + +
ModelsYouCook2 (DVC)Charades-STA (TVG)
SODA_c↑CIDEr↑Meteor↑F1↑R@1 (IoU=0.5) ↑R@1 (IoU=0.7) ↑
VTimeLLM[12] (CVPR 2024)\\\\27.511.4
Monmentor [29] (arXiv 2024)\\\\26.611.6
TimeChat [30]* (CVPR 2024)1.02.91.112.727.211.7
TimeChat + Seq2Time (Ours) w/o RPT1.23.71.415.729.312.8
TimeChat + Seq2Time (Ours)1.3+30.0%4.2+44.8%1.3+18.2%16.2+27.6%31.2+14.7%13.7+17.1%
+ +Table 2. Comparison on YouCook2 and Charades-STA datasets. * denotes our reproduction of TimeChat with partial training data availability, resulting in lower scores than reported [30]. RPT denotes our proposed unified relative position token for time representation. + +
ISMCCSYouCook2Charades-STA
S↑C↑M↑F1↑R@0.5↑R@0.7↑
✓✓✓1.34.21.316.231.213.7
✓✓X1.24.01.215.930.913.1
✓XX1.44.31.413.328.812.5
XXX1.02.91.112.727.211.7
+ +Table 3. Dataset ablation of Seq2Time on TimeChat. IS: image sequence data, MC: more video captions, CS: clip sequence data. Dark green and light green indicate the best and second-best results, respectively, and this color scheme applies to all tables unless stated otherwise. + +
Video LLaMAS ↑C ↑M ↑F1 ↑
w/o IS0.10.00.00.2
w/ IS0.10.10.23.3
+ +Table 4. Impact of image sequence data on Video-LLaMA. Training with only image sequence data enables temporal awareness as demonstrated by improved F1 scores $(0.2\rightarrow 3.3)$ . + +
MethodTimeChat+IS+IS+MC+Seq2Time
Lavg6.818.198.7411.53
TTR0.2680.4110.3840.396
+ +porating additional training data consistently enhances the richness and diversity of the generated captions, resulting in more descriptive and meaningful outputs. More details can be found in Supplementary. + +# 4.4. Image-based Pretext Tasks Ablation + +We analyze the contribution of each pretext task while controlling for training data volume by maintaining a fixed total amount of image sequence data. Table 6 shows our systematic evaluation where omitted tasks are compensated by increasing the sampling of remaining tasks (e.g. when ALR is + +Table 5. Impact of caption quality metrics across different training data configurations. $L_{\mathrm{avg}}$ : average caption length, TTR: Token Ratio (lexical diversity). Higher values indicate more detailed and diverse descriptions. + +
IICIIGALRYouCook2Charades-STA
S↑C↑M↑F1↑R@0.5↑R@0.7)↑
✓✓✓1.44.31.413.328.812.5
X✓✓1.23.41.212.927.911.8
✓X✓1.12.60.911.926.211.3
✓✓X1.02.71.014.427.211.4
XXX1.02.91.112.727.211.7
+ +Table 6. Effect of individual pretext tasks in image sequence training. The first row shows results with only image sequence data. IIC: Indexed Image Captioning, IIG: Image Index Grounding, ALR: Adjacent Location Reasoning. + +removed, IIC and IIG samples increase to 150K each). + +Our analysis reveals distinct roles for each task. Indexed Image Captioning (IIC) primarily enhances text generation quality, as evidenced by the significant CIDEr score drop when IIC is absent. Image Index Grounding (IIG) proves crucial for overall temporal understanding, with its removal causing performance degradation across both benchmarks, suggesting a stronger impact than IIC. Adjacent Location Reasoning (ALR) substantially influences text generation capabilities—its removal leads to a $37.2\%$ decrease in CIDEr and $40\%$ in SODA_c scores. + +Importantly, comparing the baseline TimeChat with variants using only IIC and IIG shows that performance improvements depend more on task diversity than data volume—adding 300K samples of just IIC and IIG data yields only a $13.4\%$ improvement. This underscores the importance of our complementary task design in developing robust temporal understanding capabilities. + +# 4.5. Dataset Scaling + +To investigate the scaling properties of Seq2Time, we conduct experiments with varying proportions (5% and 20%) of our image sequence and clip sequence data using TimeChat as the baseline model. As shown in Table 7, even with just 5% of our data, we observe notable improvements in dense video captioning performance, particularly in the CIDEr metric which measures text generation quality. The performance scales positively with data volume, showing consis + +![](images/853d3a9e5c0204272c74eecd318af3dda9d2c989edffadb3966f4d22114e62e3.jpg) +Figure 4. Qualitative examples of our Seq2Time on TimeChat. IS: image sequence data, CS: clip sequence data, MC: more video captions. Text highlighted in red shows repetitive patterns in outputs, while brown indicates incorrect predictions in timestamps or event descriptions. + +tent improvements across both dense video captioning and temporal video grounding tasks as we increase the proportion of Seq2Time data. Detailed scaling analysis and additional results are provided in the Supplementary. + +
IS/CSYouCook2Charades-STA
SCMF1R@0.5R@0.7
100%1.34.21.316.231.213.7
20%1.13.81.314.828.311.9
5%1.23.61.215.026.811.2
0%1.02.91.112.727.211.7
+ +Table 7. Impact of Seq2Time data scaling on TimeChat. Even with $5\%$ data, significant improvements are observed. Performance continues to scale positively with increased data volume. + +# 4.6. Qualitative Study + +To qualitatively demonstrate the effectiveness of Seq2Time, we analyze different model variants in Figure 4. The baseline TimeChat tends to generate repetitive patterns, e.g., repeatedly mentioning knead the dough throughout the video when this action only occurs between 120s and 200s. This phenomenon is consistent with TimeChat's low TTR in Table 5. This limitation likely stems from insufficient captioning data in TimeIT. However, our Seq2Time, leveraging rich captions from both image and clip sequence datasets, significantly improves temporal and descriptive accuracy. + +Specifically, the addition of image sequence data (IS) yields two key improvements. First, it eliminates repetitive captioning patterns, as indicated by the absence of repetitive patterns in the green patch. Second, it enhances event recognition—for example, the model accurately localizes the event water and salt are added into the bowl to 34.8-76.4 seconds, closely matching the ground truth of 48-79 seconds. This improvement aligns with our quantitative findings in Table 3, where image sequence data substantially improves the CIDEr score. In the full version (blue + +patch), the model demonstrates even more precise temporal and descriptive capabilities. For instance, it correctly identifies and localizes the event the mixture is then put in a bowl and water is poured over it to 95.2-123.2 seconds, matching the ground truth event add salt, plain yogurt, and water and mix to form a dough (80-133 seconds). While some fine-grained object recognition challenges still remain, temporal localization is accurate. + +Failure Case Analysis As illustrated in Figure 4, with only image sequence data, object recognition errors occur—for example, misidentifying a blender as hands during dough mixing. Adding more video caption data (+IS+MC) reduces but doesn't eliminate all repetitive patterns, and sometimes confuses similar objects (e.g., identifying dough as bread). Even in the full version, we notice object recognition errors (pan mistaken for oven) and hallucinated events (e.g., cut into different shapes). These issues primarily stem from visual recognition limitations and suggest that advanced visual encoders could further improve performance. + +# 5. Conclusion + +We propose Seq2Time to address the data scarcity challenge in training time-sensitive video LLMs. Our approach constructs training data from both image sequences and video clips in a self-supervised manner. Furthermore, we propose a unified relative position token, which bridges different sequence types by standardizing position representation in the LLM embedding space. Extensive experiments demonstrate the effectiveness and promising scalability of our approach, showing that sequential knowledge can be successfully transferred from sequence data to long video understanding. The framework could be also extended to other sequence types, such as image patches, where patch-level correlations could provide additional supervisory signals for temporal understanding. This suggests a broader potential for leveraging various forms of sequential data to enhance the temporal capabilities of video LLMs. + +# References + +[1] 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. 1, 3 +[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1 +[3] Max Bain, Arsha Nagrani, Güll Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 3 +[4] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 2, 5 +[5] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 3, 6 +[6] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in videolms. arXiv preprint arXiv:2406.07476, 2024. 1, 3, 6 +[7] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 1 +[8] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1 +[9] 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. 1, 3 +[10] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Xi Chen, and Bo Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 1, 2, 3, 5 +[11] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 1 +[12] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 1, 3, 7 + +[13] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. arXiv preprint arXiv:2403.19046, 2024. 1, 3 +[14] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1 +[15] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In International Conference on Computer Vision (ICCV), 2017. 1, 3 +[16] Bo Li, Hao Zhang, Kaichen Zhang, Dong Guo, Yuanhan Zhang, Renrui Zhang, Feng Li, Ziwei Liu, and Chunyuan Li. Llava-last: What else influences visual instruction tuning beyond data?, 2024. 2, 4 +[17] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 3 +[18] 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. 1 +[19] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 3 +[20] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1 +[21] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 3 +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3 +[23] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1 +[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 3 +[25] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023. 6 + +[26] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 1, 3 +[27] David Malvern, Brian Richards, Ngoni Chipere, and Pilar Durán. Lexical diversity and language development. Springer, 2004. 6 +[28] OpenAI. Gpt-4 technical report. https://openai.com/research/gpt-4, 2023. Accessed: 2024-11-07. 3 +[29] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 1, 3, 7 +[30] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1, 2, 3, 5, 6, 7 +[31] Jaehong Yoon Shoubin Yu and Mohit Bansal. Crema: Generalizable and efficient video-language reasoning via multimodal modular fusion. arxiv, 2024. 1 +[32] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 1 +[33] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videoollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3 +[34] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training videotext llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 1, 3 +[35] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. Unloc: A unified framework for video localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13623-13633, 2023. 3 +[36] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, 2023. 3 +[37] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In Advances in Neural Information Processing Systems, pages 76749-76771. Curran Associates, Inc., 2023. 1 +[38] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, + +Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2,3,5,6 +[39] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1 +[40] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1 +[41] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18243-18252, 2024. 3 +[42] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3 +[43] Wanrong Zhu, Bo Pang, Ashish V Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning as sequence generation. arXiv preprint arXiv:2204.08121, 2022. 3 \ No newline at end of file diff --git a/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/images.zip b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7faff2916fb64760770f96e68fd9cf6c06b0618f --- /dev/null +++ b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cc269c2f1100c05df9313766a0091c0a7adabb180bd0197955f47042acb81b0 +size 634722 diff --git a/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/layout.json b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5f99ed8e796e4916e732552d040ec977c82ce2f4 --- /dev/null +++ b/CVPR/2025/Seq2Time_ Sequential Knowledge Transfer for Video LLM Temporal Grounding/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10e793cd04ce2f12aac0a6a5365071268fb16afc79576e091463b7754cc58f80 +size 367659 diff --git a/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/a3234d97-d927-4684-915f-578f426ea470_content_list.json b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/a3234d97-d927-4684-915f-578f426ea470_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b1c978cc8bce5283e672317e1b880e710fc754b0 --- /dev/null +++ b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/a3234d97-d927-4684-915f-578f426ea470_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c956d925b1151b684a5ffe5bb1cf0c4e4f9d396bf238fd62d642b178b0474e6d +size 71047 diff --git a/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/a3234d97-d927-4684-915f-578f426ea470_model.json b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/a3234d97-d927-4684-915f-578f426ea470_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cef2a7338c245c17f21a2f0bbf772c6e483f6e34 --- /dev/null +++ b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/a3234d97-d927-4684-915f-578f426ea470_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d99f8b4a7119839852632252419da60943689fdebdb6e40d789ce3a2f8538827 +size 88947 diff --git a/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/a3234d97-d927-4684-915f-578f426ea470_origin.pdf b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/a3234d97-d927-4684-915f-578f426ea470_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2928b7421d03f7fe9544b54a4be89d69b60ee64d --- /dev/null +++ b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/a3234d97-d927-4684-915f-578f426ea470_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c42aa687e48edc40773a7020ef5b4ee10761f608070171b7eb6ccaf27ce9fe8c +size 1340283 diff --git a/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/full.md b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cc43db891b18b51dc52922974b3d2f9140ac68bf --- /dev/null +++ b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/full.md @@ -0,0 +1,266 @@ +# SeqAfford: Sequential 3D Affordance Reasoning via Multimodal Large Language Model + +Chunlin $\mathrm{Yu}^{1,\dagger}$ Hanqing Wang $^{1,\dagger}$ Ye Shi $^{1}$ Haoyang Luo $^{1}$ Sibei Yang $^{1}$ Jingyi Yu $^{1}$ Jingya Wang $^{1, *}$ $^{1}$ ShanghaiTech University, Shanghai, China + +{yuchl, wanghq2024, shiye, luohy2024, yangsb, yujingyi, wangjingya}@shanghaiitech.edu.cn + +![](images/7b939c120e60d62c81050ea739131125025e85df2ddbc354ef17072aae877698.jpg) +Figure 1. Sequential 3D affordance reasoning task with different types of interactions. We introduce SeqAfford, a Multi-Modal Language Model (MLLM) capable of serialized affordance inference implied in human instructions: 1) Single Affordance Reasoning; 2) Sequential Affordance Reasoning; 3) Sequential Affordance Reasoning with Multiple Objects + +![](images/49afbd0219d2714423179519617cfbf9feb942f4301bd02b50be057e5f2a17f1.jpg) + +![](images/be3b0736a5987f674d2dde63bc44358dbb37bf91aeaf1e248c0932c0f629183e.jpg) + +# Abstract + +3D affordance segmentation aims to link human instructions to touchable regions of 3D objects for embodied manipulations. Existing efforts typically adhere to single-object, single-affordance paradigms, where each affordance type or explicit instruction strictly corresponds to a specific affordance region and are unable to handle long-horizon tasks. Such a paradigm cannot actively reason about complex user intentions that often imply sequential affordances. In this paper, we introduce the Sequential 3D Affordance Reasoning task, which extends the traditional paradigm by reasoning from cumbersome user intentions and then decomposing them into a series of segmentation maps. Toward this, we + +construct the first instruction-based affordance segmentation benchmark that includes reasoning over both single and sequential affordances, comprising 180K instruction-point cloud pairs. Based on the benchmark, we propose our model, SeqAfford, to unlock the 3D multi-modal large language model with additional affordance segmentation abilities, which ensures reasoning with world knowledge and fine-grained affordance grounding in a cohesive framework. We further introduce a multi-granular language-point integration module to endow 3D dense prediction. Extensive experimental evaluations show that our model excels over well-established methods and exhibits open-world generalization with sequential reasoning abilities. Project page: https://seq-afford.github.io/. + +# 1. Introduction + +Affordance is a crucial lens through which humans and embodied agents interact with various objects of the world. When provided with human instructions, affordance aims to highlight the actionable possibilities of these objects, linking visual perception with manipulation. While 2D affordance [14, 35, 47] offers visual cues that suggest potential actions to embodied systems, 3D affordance [3, 6, 8, 24, 51] provides a more direct and intuitive guidance for executing tasks in the realistic 3D world, and thus solidify the foundation for downstream robot manipulation tasks [11, 30, 44, 45]. + +Previous work on 3D affordances has largely focused on the single-object, single-affordance paradigm, where affordance maps are grounded either in affordance categories [7, 29] or 2D demonstration images [8, 50]. Recently, language models have been employed to pair 3D objects with natural language questions [16], each designed to elicit a specific affordance. For example, the question "How can you go through the door?" can be interpreted by language models like BERT [12] or RoBERTa [20] as referring to the "openable" affordance component (e.g., the door handle). However, regardless of whether the grounding source is an affordance phrase, HOI image, or a natural question, each grounding source generally corresponds to a fixed affordance type. + +Current systems can't actively reason complex user intentions, break them down into actionable primitives, or formulate chains of affordances from each primitive. Real-world physical interactions require modeling complex human instructions involving multiple objects and sequential affordances. For example, to 'reheat food in a bowl using the microwave', agents must first grasp the bowl, open the microwave, then contain the bowl inside. Thus, multi-object sequential reasoning is crucial for next-generation affordance systems. + +Recently, Large-Language Models (LLMs) [2, 31, 55] have demonstrated exceptional sequential reasoning abilities, ingrained with internalized common-sense knowledge encoded from vast text data corpora. On top of that, the emergence of 3D Multimodal Large-Language Models (MLLMs) [34, 48] have further expanded their possibilities in understanding various object shapes in the 3D world. However, even the latest 3D MLLMs are not panaceas for reasoning about visual affordances from 3D objects, as they primarily focus on object-centered text generation tasks. This, therefore, highlights a pressing question: Can we devise a 3D multi-modal large language model to sequentially reason and segment multi-object affordances based on long-horizon human instructions? + +In this paper, we introduce a new task called Sequential 3D Affordance Reasoning, designed to narrow the gap with real-world demands. Towards this, we construct a large- + +scale sequential affordance reasoning benchmark, containing 180K instruction-point cloud pairs. To ensure the diversity of the instruction data, four distinct methods are used for generating instructions. These methods include utilizing immersive 2D HOI images and detailed, colorful mesh renderings to prompt GPT-4o to generate diverse instructions, drawing on its inherent world knowledge, as illustrated in Fig. 2. Supported by this benchmark, we introduce our model SeqAfford, which unlocks the current 3D MLLMs with sequential affordance segmentation abilities. To further bolster the 3D dense prediction and reasoning task, we introduce a multi-granular language-point integration module, where dense point features conditioned on segmentation tokens of the large language model are integrated with sparse point features for subsequent dense prediction tasks. This module not only effectively injects the reasoning results of large language models into the dense point features, but enhances the affordance segmentation task with multi-granular levels of representation. + +To summarize, our contributions are as follows: + +- We introduce the Sequential 3D Affordance Reasoning task, which involves the sequential reasoning and segmentation of affordances based on complex human instructions. This paradigm is crucial for the development of next-generation affordance systems. + +- We develop a large-scale sequential affordance reasoning benchmark with 180K instruction-point cloud pairs, serving as a comprehensive resource to advance research in affordance reasoning. + +- We leverage the internalized world knowledge of prealigned 3D MLLMs to achieve fine-grained affordance-level multi-target sequential reasoning and explanations in a cohesive framework. + +# 2. Related Work + +3D Affordance Segmentation. With the rapid advancement of embodied AI, research on 3D affordance has increasingly garnered significant attention from both academia and industry. 3D AffordanceNet [7], built on point cloud data from PartNet [26], was the first to construct a fine-grained 3D affordance dataset, establishing a benchmark for 3D affordance research. Building on this, some researchers [38, 50] proposed leveraging universal knowledge extracted from 2D human-object interaction (HOI) images to assist in 3D affordance segmentation. However, these methods rely solely on visual information to infer affordances, without considering that embodied agents in the real world communicate with humans through language. Consequently, such methods are limited in their applicability for direct deployment in embodied agents. Recently, Li et al. [16] introduced a language-based affordance segmentation task, promoting the integration of natural language and affordance understanding. + +![](images/7f77ae2db65424ad80585fc3e6a141cf6907c2e8deefe09b67facf6e82268c6c.jpg) +a) + +![](images/fd01af694813148f0f8a5c9afa20f89d40aede88b183dcbf5000332c23381b1b.jpg) +b) + +![](images/25e04a57d3cfb6a5f06b5720b1d7c70b446211d8d78836dab840287892c31676.jpg) +c) + +![](images/68f647bd28a40407d0eb3fc19f06aa6d6d015bd2282387ba0e14f272d69311ac.jpg) +d) +Figure 2. Preparing the instructions. To better utilize the world knowledge of GPT4, We prompt GPT-4o to generate diverse instructions based on 4 types of system prompts containing different modalities as input. Instructions are generated based on input prompts with modalities from a) purely textual affordance type, object name; b) the mesh-rendered image of the object; c) the mesh-rendered image and HOI images that reveal affordances of the object; d) the mesh-rendered image and textual description of the scenario. + +Following this, Chu and Zhang [4] utilized LLMs to locate objects in 2D images and subsequently retrieve corresponding objects from a 3D dataset. Although these approaches leverage the reasoning capabilities of LLMs, they lack the ability to perform joint visual and language alignment, failing to bridge the gap between 2D and 3D modalities. Furthermore, their research assigns each text statement to a single specific affordance, overlooking the complexity of scenarios that often require the coordination of multiple affordances. In response to these limitations, we propose the integration of 3D multimodal large language models (MLLMs) into affordance segmentation. This novel approach enables the model to simultaneously comprehend contextual semantics and point cloud data, thereby facilitating affordance reasoning across a wide range of complex scenarios. By incorporating both natural language and 3D data, our model offers a more comprehensive and versatile understanding of affordances, rendering it better suited for real-world applications in embodied AI. + +Multimodal Large Language Models. Large Language Models (LLMs) have achieved remarkable success in processing natural language, and researchers have been working to extend these models' reasoning capabilities into the visual domain. Early efforts [2, 18, 28, 55] have made significant strides by enabling LLMs to process both visual and textual information concurrently, thereby introducing an initial level of human-like multimodal reasoning. However, for many practical applications, such as visual segmentation, these models lack the necessary fine-grained perception required for detailed visual tasks. To address this issue, research + +efforts [21, 41, 52, 54] enable the localization of specific regions within images by encoding spatial coordinates as tokens, improving the models' ability to reason about precise areas within the visual data. + +Building on these 2D MLLM advancements, research has increasingly expanded into the 3D domain. Following the paradigm of LLaVA, PointLLM [48] replaces the vision encoder with a 3D point encoder, enabling the processing of 3D data within the latent space of LLMs and facilitating 3D object understanding. Similarly, ShapeLLM [34] is built upon the enhanced 3D encoder RECON++, which strengthens geometric understanding through multi-view image distillation. Other models, such as 3D-LLM [9], leverage 2D foundational models, like CLIP ViT [36], to process multi-view rendered images of 3D point clouds, thereby integrating the 3D world into LLMs. However, despite these advancements, the majority of existing MLLMs are primarily focused on scene-level and object-level understanding, lacking the ability to recognize and segment fine-grained affordances of 3D objects in diverse semantic contexts. Addressing this limitation, our study aims to endow MLLMs with affordance-aware perception, enabling them to interpret and act upon 3D objects more effectively in context-sensitive scenarios. + +# 3. Dataset + +Affordance segmentation involves understanding the operability of objects in various contexts, and the varying complexity of intentions adds significant challenges to the process. Simple instructions typically pertain to the direct usage + +
Method#Sequential#World Knowledge#Multi-object#Instruction-Point Pairs#Point Cloud
3D AffordanceNet [7]××××23k
O2O-Afford [27]××××1.7k
Partafford [46]××××25k
IAGNet [50]××××7k
LASO [16]×××19k8.4k
Ours✓✓✓180k18k
+ +Table 1. Comparison of Existing 3D Affordance Datasets with Ours. #Point Cloud and #Instruction-Point Cloud Pairs denote the number of point clouds and instruction-point cloud pairs, respectively. $\times$ indicates that the dataset does not possess this attribute. + +of an object, such as grasping a cup or opening a door. In contrast, complex instructions may involve multi-step actions or require contextual understanding, such as using a cup for a specific occasion or purpose. To address the challenge of affordance segmentation based on both simple and complex instructions, we constructed a dataset of instruction-point cloud pairs based on 3D AffordanceNet [7], encompassing both simple and complex types of intentions, which includes 162,386 instruction-point cloud pairs in the single affordance segmentation setting and 20,847 pairs in the sequential affordance segmentation setting, comprising a total of 18,371 point cloud instances across 23 object categories. + +# 3.1. Dataset Collection + +Point Cloud. Our point cloud data and affordance annotations are entirely sourced from 3D AffordanceNet [7]. In the simple instruction setting, we generated five instructions for each affordance of every point cloud instance. For the sequential affordance segmentation setting, we carefully selected point cloud categories that support this configuration and generated corresponding instructions for each affordance sequence combination. + +Instruction. To create instructions, we developed four methods for generating instructions using GPT-4 [1]. Unlike LASO [16] which generates the same texts for all point cloud instances of a specific affordance type for each category, we generate text for each point cloud by utilizing the point cloud instance from 3D AffordanceNet [7] to trace back to the Mesh-rendered images in the PartNet [26] dataset. Additionally, we collect HOI images corresponding to the affordance types from IAGNet [50] to alleviate GPT's hallucination issues, enabling better understanding. Fig. 2 illustrates our ways of generating instructions in detail. + +# 3.2. Statistics and Analysis + +Our dataset comprises 162,386 instruction-point cloud pairs in the single affordance segmentation setting and 20,847 pairs in the sequential affordance segmentation setting, making up a total of 18,371 point cloud instances across 23 object categories. + +Based on the complexity of the instructions, we have + +divided them into two settings. The first setting is based on instructions that can only infer a single specific affordance, which we define as the "Single Affordance Segmentation Task". The second setting is based on instructions that can infer a combination of multiple affordances in sequence, which we define as the "Sequential Affordance Segmentation Task". + +Inspired by the evaluation settings presented in LASO [16] and IAGNet [50], we propose two types of distinct dataset configurations including Seen and Unseen. Seen: This default setting maintains similar distributions of object classes and affordance types across both training and testing phases. Unseen: This configuration is specifically designed to evaluate the model's ability to generalize affordance knowledge. In this setup, certain affordance-object pairings are deliberately omitted from the training set but introduced during testing. For instance, while the model may learn to grasp objects like bags and mugs during training, it is expected to generalize the "grasp" affordance to earphones, a pairing absent from the training data. + +# 4. Method + +# 4.1. Architecture Overview + +The overall architecture of SeqAfford is presented in Fig. 3. Generally, SeqAfford mainly consists of three components: 1) a 3D vision encoder benefited from large-scale 3D representation learning, which provides solid foundations for dense prediction tasks; 2) a 3D Multi-modal Large Language Model (MLLM) $\mathcal{F}$ that exhibits affordance reasoning ability with the aid of internalized world knowledge; 3) a Multi-Granular Language-Point Integration module that considers the effective integration the point features and the segmentation tokens of MLLM, synergizing both reasoning and segmentation tasks from a multi-granular feature perspective. + +# 4.2. Network Architecture + +3D MLLM Backbone. Recently, a series of 3D multimodal language models have been proposed to deepen the understanding of open-world 3D objects, among which + +![](images/082dcaa94987e9e47f38599185c880716a6a8b80aa6b2d193e02434c2e7b0851.jpg) +Figure 3. Main Pipeline. Given the point clouds of the target objects and a piece of complex human instruction, SeqAfford first reasons from this instruction and decomposes it into several hidden tokens extracted from the last-layer embeddings, each representing an intermediate affordance segmentation result. Then, for each , the point features extracted by the 3D vision encoder dynamically interact with the token before being sent to the decoder for mask generation. The interaction is achieved through multi-granular language-point integration, synergizing both reasoning and affordance segmentation. We use LoRA for efficient fine-tuning. + +ShapeLLM is recently pretrained for understanding various embodied interactions. In light of this, we adopt ShapeLLM [34] as our backbone, denoted as $\mathcal{F}$ , where the point cloud encoder $\mathrm{ReCon}++$ is pre-trained via multi-view distillation based on ReCon [33], and the LLM is drawn from LLaMa [40]. Previous work on 3D affordance tasks typically employed 3D backbones [29, 50] or utilized separate point-language encoders [16], which may fall short in reasoning and open-world generalization abilities. Here, we leap ahead by directly utilizing a unified 3D MLLM instead of relying solely on pure LLMs or other visual structures as our backbones, for two main reasons: 1) it opens new possibilities for open-world 3D objects understanding, bolstering the generalization of unseen objects or affordances; 2) it internalizes affordance perception ability, compressing it into natural language form, thus preparing for the subsequent affordance reasoning task. + +Squential Affordance Reasoning. Despite the efficacy of 3D MLLMs in aligning 3D representations with natural language, they are primarily designed for object-oriented text generation tasks and lack the capability for 3D dense prediction tasks, particularly in fine-grained affordance segmentation. To encapsulate the segmentation ability into 3D MLLMs, a specific segmentation token $<\mathrm{SEG}>$ can be appended to the vocabulary set of the MLLM, inspired by [13]. + +Formally, when provided with point cloud $\mathbf{X}_{\mathrm{point}}$ and a text instruction $\mathbf{X}_{\mathrm{txt}}$ that demonstrates the user intentions on these potential objects, the 3D MLLM absorbs this multimodal information and generates a text response $\tilde{\mathbf{y}}_{\mathrm{txt}}$ . It can be formulated as + +$$ +\tilde {\mathbf {y}} _ {\mathrm {t x t}} = \mathcal {F} \left(\mathbf {X} _ {\text {p o i n t}}, \mathbf {X} _ {\mathrm {t x t}}\right), \tag {1} +$$ + +where the output $\tilde{\mathbf{y}}_{xt}$ would include several tokens, where a single indicates a segmentation result within the sequence. We then extract the last-layer embeddings $\{\mathbf{h}_{\mathrm{seg}}^{(i)}\}_{i = 0}^{S - 1}$ corresponding to the tokens, where $S$ is the number of predicted affordance sequences. Afterwards, an MLP projection layer to obtain $\{\mathbf{H}_{seg}^{(i)}\}_{i = 0}^{S - 1}$ as follows + +$$ +\mathbf {H} _ {\mathrm {s e g}} ^ {(i)} = \operatorname {P r o j} \left(\mathbf {h} _ {\mathrm {s e g}} ^ {(i)}\right). \tag {2} +$$ + +Multi-Granular Language-Point Integration. After obtaining several segmentation tokens that indicate a sequence of regions for reasoning where the given objects can be afforded, the remaining work entails integrating the abstracted reasoning results into 3D point clouds for dense affordance predictions. Therefore, the multi-granular language-point integration module mainly consists of two stages: 1) the multi-granular feature propagation process, which iteratively up-samples the point cloud features into dense features with multiple granularities considered; 2) the point-language integration stage, which distills the informative language features (i.e. the segmentation tokens) into the dense visual features (i.e. the 3D point features), and fuses the integrated dense features with global sparse features for final affordance segmentation. This serves as a crucial step towards reasoning and segmenting affordances in a cohesive framework. + +To enable multi-granular feature propagation, as shown in Fig. 4, we hierarchically up-sample (UpS. in Fig. 4) the intermediate features from the 3D encoder, and propagate features through farthest point sampling (FPS in Fig. 4) process to sequentially generate $\mathbf{f}_1$ , $\mathbf{f}_2$ , and the ultimate dense feature $\mathbf{f}_{\mathrm{dense}}$ . During the up-sampling process of intermediate features, various feature up-sample techniques + +![](images/6c708b221dcc8adf27e4cce480e8fb5926b0924107637ad206149cb951235bac.jpg) +Figure 4. Multi-Granular Language-Point Integration Module. We propose an interaction module between tokens from LLM and point features from the 3D vision encoder, to synergize both reasoning and segmentation in a cohesive framework. This module consists of the multi-granular feature propagation process, and the point-language integration stage. + +are adopted, inspired by PointNet++ [32] and DGCNN [42]. More details about the multi-granular feature propagation process are revealed in the supplementary details. + +During the point-language integration stage, the model takes the dense point cloud features $\mathbf{f}_{\mathrm{dense}}$ , the sparse point cloud features $\mathbf{f}_{\mathrm{sparse}}$ and the instruction-rich $\mathbf{H}_{\mathrm{seg}}^{(i)}$ as input. The $\mathbf{H}_{\mathrm{seg}}^{(i)}$ and dense features $\mathbf{f}_{\mathrm{dense}}$ are used as $\mathbf{Q}$ and $\mathbf{K}$ , $\mathbf{V}$ respectively to perform cross-attention, and FFN, and the results obtained are fused with the sparse features $\mathbf{f}_{\mathrm{sparse}}$ to get $\mathbf{A}_f^{(i)}$ . Finally, the decoder takes $\mathbf{A}_f^{(i)}$ as the input to get affordance mask $\tilde{\mathbf{y}}_{\mathrm{mask}}$ + +$$ +\mathbf {A} _ {f} ^ {(i)} = \mathcal {G} (\mathbf {f} _ {\text {d e n s e}}, \mathbf {f} _ {\text {s p a r s e}}, \mathbf {H} _ {\text {s e g}} ^ {(i)}), \quad \tilde {y} _ {\text {m a s k}} = \mathcal {D} (\mathbf {A} _ {f} ^ {(i)}), \tag {3} +$$ + +where $\mathcal{G}$ denotes point-language integration, and $\mathcal{D}$ denotes the decoder. + +# 4.3. Training objectives + +Our objective is to train an end-to-end MLLM capable of generating diverse texts while simultaneously predicting point-wise affordance masks. To this end, we employ autoregressive cross-entropy loss $\mathcal{L}_c$ for text generation, Dice loss $\mathcal{L}_d$ and Binary Cross-Entropy loss $\mathcal{L}_b$ for guiding the segmentation mask prediction. + +$$ +\begin{array}{l} \mathcal {L} = \lambda_ {c} \mathcal {L} _ {c} (\mathbf {y} _ {\mathrm {t x t}}, \tilde {\mathbf {y}} _ {\mathrm {t x t}}) + \lambda_ {b} \mathcal {L} _ {b} (\mathbf {y} _ {\text {m a s k}}, \tilde {\mathbf {y}} _ {\text {m a s k}}) + \tag {4} \\ \lambda_ {d} \mathcal {L} _ {d} (\mathbf {y} _ {\text {m a s k}}, \tilde {\mathbf {y}} _ {\text {m a s k}}), \\ \end{array} +$$ + +where the weights $\lambda_c, \lambda_b, \lambda_d$ are utilized to balance the different loss items. + +# 5. Experiment + +We conduct extensive experiments to evaluate the effectiveness of our proposed dataset, task, and method, including both Single and Sequential Affordance segmentation tasks. In Sec. 5.1, we assess the capability of our model to ground the Single Affordance with simple instruction. Sec. 5.2 studies a more challenging task where the model is requested to predict the sequential affordances. Various ablation experiments on our model are performed in Sec. 5.3. + +Implementation Details. We employ ShapeLLM [34] as our 3D MLLM module in this paper, with the ShapeLLM-7B checkpoint as the default setting and we freeze the 3D encoder during training. We adopt Uni3D as the 3D vision encoder to enhance the 3D dense prediction tasks. Unless otherwise stated, the projection layer is implemented as a multi-layer perceptron. We employ LoRA [10] for efficient fine-tuning and set the rank of LoRA to 8 by default. Additionally, we utilize AdamW [23] optimizer with the learning rate and weight decay set to 0.0002 and 0, respectively. We adopt a cosine learning rate scheduler, with the warm-up iteration ratio set to 0.03. All attentions in ShapeLLM [34] are replaced by flash-attention [5] during training. The training is done on one A100 GPU for 10 epochs for the main experiments and during training, we use all mentioned datasets in Sec. 3 for joint training by leveraging task-specific prompts. For evaluation on a specific dataset, we finetune the trained model on the corresponding dataset. + +Evaluation Metrics and Baseline. To provide a comprehensive and effective evaluation, we follow previous works + +
MethodmIoU↑AUC↑SIM↑MAE↓
SeenReferTrans [15]11.477.20.4490.135
ReLA [17]12.176.30.4800.130
3D-SPS [25]10.175.20.4130.141
IAGNet [50]14.281.70.5100.117
PointRefer [16]16.384.30.5680.108
Ours19.586.90.5940.098
UnseenReferTrans [15]9.167.40.4270.151
ReLA [17]9.368.20.4230.147
3D-SPS [25]7.166.90.3970.162
IAGNet [50]11.773.60.4380.143
PointRefer [16]12.476.10.5020.132
Ours13.882.40.5180.128
SquentialReferTrans* [15]10.874.50.4250.142
ReLA* [17]11.474.80.4630.136
3D-SPS* [25]9.973.10.4070.148
IAGNet* [50]13.578.20.4960.131
PointRefer* [16]14.380.70.5210.124
Ours14.684.20.5730.118
+ +Table 2. Main Results. The overall results of all comparative methods, the best results are in bold. Seen and unseen are two partitions of the Single Affordance segmentation dataset. AUC and mIoU are shown in percentage. * means that baseline methods use ground-truth sequential order as all existing methods cannot predict sequential affordances. + +and finally chose four evaluation metrics: Area Under the Curve (AUC) [22], Mean Intersection Over Union (mIOU) [37], SIMiliarity (SIM) [39] and Mean Absolute Error (MAE) [43]. To the best of our knowledge, LASO [16] is the most similar to our work. For a thorough comparison of our method, we conduct comparisons based on its setup in the Single Affordance segmentation task, comparisons with other baselines were also implemented following the approach mentioned in it. While in the sequential affordance segmentation task, we offer sequential information to these models enabling them to perform "sequential" reasoning. + +# 5.1. Results on Language-Guided Single Affordance Segmentation Task + +As detailed in Table. 2, our model demonstrates superior performance across all evaluation metrics compared to the baseline methods. Unlike conventional segmentation tasks, language-guided Single Affordance segmentation demands not just identification but the integration of perception and cognition, necessitating the model's reasoning capabilities and access to world knowledge. Existing approaches struggle with implicit queries due to their lack of integration of perception and cognition, which further underscores the task's inherent challenges. In contrast, our model lever + +ages MLLMs to bridge this gap, demonstrating superior performance by comprehending and interpreting the queries accurately. + +# 5.2. Results on Language-Guided Sequential Affordance Reasoning Task + +The sequential affordance reasoning task implies the ability to infer multiple affordances from a single text, which requires a more profound integration of cognitive and perceptual capabilities compared to the Single Affordance segmentation task. In our model, to make the Multimodal Large Language Model (MLLM) more comprehensible, we use the format $<\mathrm{SEG}>$ to represent the sequence of affordances and decode to obtain the affordance mask based on them. The previous models did not possess this capability. To compare them with our method, we used GPT to decompose the original instructions into new sequence instructions, then fed the decomposed instructions separately as inputs to these models, enabling them to perform "sequential" reasoning. The main results are shown in Table 2, and we speculate that the reason our model performs better is that, compared to LASO [16] which merely uses language models to encode the input text, we have introduced a multimodal 3D large language model. This model has a much stronger capability for integrating 3D data and text than a purely linguistic model, and it possesses a richer knowledge of the world, enabling it to handle multimodal sequential tasks more effectively. The sequential reasoning results are visualized in Fig. 5 (b), where our model can understand how a task is connected with sequential actionable affordances involving multi-objects. This ability is not only attributed to the challenging benchmark collected from diverse sources but also the powerful world knowledge internalized in 3D MLLMs. + +# 5.3. Ablation Study + +We conduct various ablation studies to assess the impact of different model implementations on our model SeqAfford's performance, including the multi-granular language-point integration module and the different choices of 3D vision encoder backbones. + +Multi-Granular Language-Point Integration Module. Introducing the Multi-Granular Language-Point Integration (MGLP) module results in a substantial improvement over the baseline, as shown in Table 3. This underscores our method's capability to minimize information loss and this indicates that our method enables deep integration of high-dimensional semantic features representing instructions with dense features from point clouds and semantically rich but sparse features from point clouds, thereby making affordance reasoning more effective. + +Choice of 3D Vision Encoder Backbone. We conducted ablation experiments to study the influence of the 3D Vision Encoder backbone, and we investigated the performance of + +![](images/156b1a8f4e8b5551093dfb211a3388ad6a12046b83bcbea0315760a57aabe501.jpg) +Figure 5. Qualitative results of our model. SeqAfford understands human instruction and accurately segments the target affordance. + +3D SeqAfford with some alternative backbones. As shown in Table. 4, Uni3D [53] performs better in this task due to its strong representation ability. Thus, we set it as our default 3D vision encoder backbone. + +
VariantsTaskmIoU↑AUC↑SIM↑MAE↓
w/o MGLPSingle12.183.40.5520.117
OursSingle19.586.90.5940.098
w/o MGLPSequential11.780.30.5180.129
OursSequential14.684.20.5730.118
+ +# 6. Conclusion + +In this paper, we introduce a novel task called the Sequential 3D Affordance Reasoning Task, designed to address complex user intentions that may involve sequential or long-horizon subtasks, transcending the conventional single-object, single-affordance paradigm. Our contributions include the construc + +Table 3. Ablation study on multi-granular language-point module. + +
3D Vision BackbonemIoU↑AUC↑SIM↑MAE↓
ULIP [49]17.984.80.5740.109
OpenShape [19]18.485.30.5820.103
Recon++ [34]19.186.40.5880.099
Uni3D [53]19.586.90.5940.098
+ +Table 4. Ablation study on the choices of 3D vision backbone. + +tion of a large-scale 3D affordance benchmark comprising 180K instruction-point pairs collected from diverse sources, and the development of an advanced 3D multimodal LLM that leverages world knowledge to interpret complex user intentions, producing sequential, actionable affordance maps with reasonable explanations. SeqAfford advances 3D affordance reasoning by integrating LLM capabilities, enhancing the model's ability to handle complex, sequential tasks in real-world contexts. Our future work will focus on scene-level fine-grained reasoning to enable more sophisticated and context-aware affordance segmentation for embodied intelligent agents. + +# 7. Acknowledgement + +This work was supported by the Shanghai Local College Capacity Building Program (23010503100), National Natural Science Foundation of China (No.62406195), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), HPC Platform of ShanghaiTech University and Shanghai Engineering Research Center of Intelligent Vision and Imaging. + +# 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. 4 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2, 3 +[3] Hengshuo Chu, Xiang Deng, Xiaoyang Chen, Yinchuan Li, Jianye Hao, and Liqiang Nie. 3d-affordancelm: Harnessing large language models for open-vocabulary affordance detection in 3d worlds. arXiv preprint arXiv:2502.20041, 2025. 2 +[4] Meng Chu and Xuan Zhang. Iris: Interactive responsive intelligent segmentation for 3d affordance analysis. arXiv preprint arXiv:2409.10078, 2024. 3 +[5] 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. 6 +[6] Alexandros Delitzas, Ayca Takmaz, Federico Tombari, Robert Sumner, Marc Pollefeys, and Francis Engelmann. Scenefun3d: fine-grained functionality and affordance understanding in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14531-14542, 2024. 2 +[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. 2, 4 +[8] Xianqiang Gao, Pingrui Zhang, Delin Qu, Dong Wang, Zhigang Wang, Yan Ding, Bin Zhao, and Xuelong Li. Learning 2d invariant affordance knowledge for 3d affordance grounding. arXiv preprint arXiv:2408.13024, 2024. 2 +[9] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3 +[10] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 + +[11] 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. In European Conference on Computer Vision, pages 222–239. Springer, 2024. 2 +[12] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 2 +[13] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 5 +[14] Gen Li, Deqing Sun, Laura Sevilla-Lara, and Varun Jampani. One-shot open affordance learning with foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3086-3096, 2024. 2 +[15] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in neural information processing systems, 34:19652-19664, 2021. 7 +[16] Yicong Li, Na Zhao, Junbin Xiao, Chun Feng, Xiang Wang, and Tat-seng Chua. Laso: Language-guided affordance segmentation on 3d object. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14251-14260, 2024. 2, 4, 5, 7 +[17] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23592-23601, 2023. 7 +[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3 +[19] 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. 8 +[20] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 364, 2019. 2 +[21] Zhaoyang Liu, Yinan He, Wenhai Wang, Weiyun Wang, Yi Wang, Shoufa Chen, Qinglong Zhang, Zeqiang Lai, Yang Yang, Qingyun Li, et al. Interrupt: Solving vision-centric tasks by interacting with chatgpt beyond language. arXiv preprint arXiv:2305.05662, 2023. 3 +[22] Jorge M Lobo, Alberto Jiménez-Valverde, and Raimundo Real. Auc: a misleading measure of the performance of predictive distribution models. Global ecology and Biogeography, 17(2):145–151, 2008. 7 +[23] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[24] Dongyue Lu, Lingdong Kong, Tianxin Huang, and Gim Hee Lee. Geal: Generalizable 3d affordance learning with cross-modal consistency. arXiv preprint arXiv:2412.09511, 2024. 2 +[25] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage + +3d visual grounding via referred point progressive selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16454-16463, 2022. 7 +[26] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909–918, 2019. 2, 4 +[27] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on robot learning, pages 1666-1677. PMLR, 2022. 4 +[28] Soroush Nasiriany, Sean Kirmani, Tianli Ding, Laura Smith, Yuke Zhu, Danny Driess, Dorsa Sadigh, and Ted Xiao. Rt-affordance: Affordances are versatile intermediate representations for robot manipulation. arXiv preprint arXiv:2411.02704, 2024.3 +[29] Toan Nguyen, Minh Nhat Vu, An Vuong, Dzung Nguyen, Thieu Vo, Ngan Le, and Anh Nguyen. Open-vocabulary affordance detection in 3d point clouds. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5692-5698. IEEE, 2023. 2, 5 +[30] Chuanruo Ning, Ruihai Wu, Haoran Lu, Kaichun Mo, and Hao Dong. Where2explore: Few-shot affordance learning for unseen novel categories of articulated objects. Advances in Neural Information Processing Systems, 36:4585-4596, 2023. 2 +[31] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2 +[32] 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. 6 +[33] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In International Conference on Machine Learning, pages 28223-28243. PMLR, 2023. 5 +[34] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. Shapellm: Universal 3d object understanding for embodied interaction. arXiv preprint arXiv:2402.17766, 2024. 2, 3, 5, 6, 8 +[35] Shengyi Qian, Weifeng Chen, Min Bai, Xiong Zhou, Zhuowen Tu, and Li Erran Li. Affordancelm: Grounding affordance from vision language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7587-7597, 2024. 2 +[36] 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 + +[37] Md Atiqur Rahman and Yang Wang. Optimizing intersection-over-union in deep neural networks for image segmentation. In International symposium on visual computing, pages 234–244. Springer, 2016. 7 +[38] Yawen Shao, Wei Zhai, Yuhang Yang, Hongchen Luo, Yang Cao, and Zheng-Jun Zha. Great: Geometry-intention collaborative inference for open-vocabulary 3d object affordance grounding. arXiv preprint arXiv:2411.19626, 2024. 2 +[39] Michael J Swain and Dana H Ballard. Color indexing. International journal of computer vision, 7(1):11-32, 1991. 7 +[40] 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. 5 +[41] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. Advances in Neural Information Processing Systems, 36, 2024. 3 +[42] 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. 6 +[43] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. *Climate research*, 30(1):79-82, 2005. 7 +[44] Ruihai Wu, Kai Cheng, Yan Zhao, Chuanruo Ning, Guanqi Zhan, and Hao Dong. Learning environment-aware affordance for 3d articulated object manipulation under occlusions. Advances in Neural Information Processing Systems, 36:60966-60983, 2023. 2 +[45] Shijie Wu, Yihang Zhu, Yunao Huang, Kaizhen Zhu, Jiayuan Gu, Jingyi Yu, Ye Shi, and Jingya Wang. Afforddp: Generalizable diffusion policy with transferable affordance. arXiv preprint arXiv:2412.03142, 2024. 2 +[46] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance discovery from 3d objects. arXiv preprint arXiv:2202.13519, 2022. 4 +[47] Peiran Xu and MU Yadong. Weakly-supervised affordance grounding guided by part-level semantic priors. In The Thirteenth International Conference on Learning Representations. 2 +[48] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. PointLLM: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 2, 3 +[49] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, 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. 8 +[50] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from + +2d interactions in images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10905-10915, 2023. 2, 4, 5, 7 +[51] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, and Zheng-Jun Zha. Lemon: Learning 3d human-object interaction relation from 2d images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16284-16295, 2024. 2 +[52] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 3 +[53] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3d: A unified baseline for multi-dataset 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9253-9262, 2023. 8 +[54] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 3 +[55] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3 \ No newline at end of file diff --git a/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/images.zip b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..43dd35971f461ddb72df0bb12f608578fafcc5ce --- /dev/null +++ b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d48272a78a437d5c8889bf31eaab9114e9e128e4323f8ad1fc5b6c1af6585454 +size 641458 diff --git a/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/layout.json b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..493c4496b310b0ab951700f1f1c4e10c736b4a11 --- /dev/null +++ b/CVPR/2025/SeqAfford_ Sequential 3D Affordance Reasoning via Multimodal Large Language Model/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eef071029dcb982d46f3a771eef788eeabeffbf1d1842db56474376904a112c7 +size 313645 diff --git a/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/1514ab4f-633b-4f04-8a2e-1c9625f283df_content_list.json b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/1514ab4f-633b-4f04-8a2e-1c9625f283df_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f41b02c781321703c99108964223cce5394febe4 --- /dev/null +++ b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/1514ab4f-633b-4f04-8a2e-1c9625f283df_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2bc2753b6c60be784932636e0b8b6c53d4a8df0d5490998ab9bae67a75d9794 +size 79216 diff --git a/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/1514ab4f-633b-4f04-8a2e-1c9625f283df_model.json b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/1514ab4f-633b-4f04-8a2e-1c9625f283df_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2d1e7457e7d744a634d335552c3491028e819e66 --- /dev/null +++ b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/1514ab4f-633b-4f04-8a2e-1c9625f283df_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91b8fd66a362f4222fc7a40acb1b1d46e293e0f057db9af9f0984a2cef56e008 +size 97352 diff --git a/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/1514ab4f-633b-4f04-8a2e-1c9625f283df_origin.pdf b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/1514ab4f-633b-4f04-8a2e-1c9625f283df_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d12bcef1c5367fb561220c352762c11f8c0abd98 --- /dev/null +++ b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/1514ab4f-633b-4f04-8a2e-1c9625f283df_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:234712820b16a84b097feffa2f9f8d9707cd363d2ba555400c8c1775093ea655 +size 2010523 diff --git a/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/full.md b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d548d332f88defdf6d2f1728fdb7c1481e3dc3ea --- /dev/null +++ b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/full.md @@ -0,0 +1,326 @@ +# SeqMvRL: A Sequential Fusion Framework for Multi-view Representation Learning + +Ren Wang $^{1,2}$ Haoliang Sun $^{1*}$ Yuxiu Lin $^{3}$ Chuanhui Zuo $^{1}$ Yongshun Gong $^{1}$ Yilong Yin $^{1}$ Wenjia Meng $^{1*}$ School of Software, Shandong University + +$^{2}$ Department of Computer Science and Technology, Tsinghua University + +$^{3}$ School of Computing and Artificial Intelligence, Shandong University of Finance and Economics {xxlifelover, zch1758440889} $@$ gmail.com, {haolsun, ysgong, ylyin, wjmeng} $@$ sdu.edu.cn, linyx0109@mail.sdufe.edu.cn + +# Abstract + +Multi-view representation learning integrates multiple observable views of an entity into a unified representation to facilitate downstream tasks. Current methods predominantly focus on distinguishing compatible components across views, followed by a single-step parallel fusion process. However, this parallel fusion is static in essence, overlooking potential conflicts among views and compromising representation ability. To address this issue, this paper proposes a novel Sequential fusion framework for Multiview Representation Learning, termed SeqMvRL. Specifically, we model multi-view fusion as a sequential decision-making problem and construct a pairwise integrator (PI) and a next-view selector (NVS), which represent the environment and agent in reinforcement learning, respectively. PI merges the current fused feature with the selected view, while NVS is introduced to determine which view to fuse subsequently. By adaptively selecting the next optimal view for fusion based on the current fusion state, SeqMvRL thereby effectively reduces conflicts and enhances unified representation quality. Additionally, an elaborate novel reward function encourages the model to prioritize views that enhance the discriminability of the fused features. Experimental results demonstrate that SeqMvRL outperforms parallel fusion schemes in classification and clustering tasks. + +# 1. Introduction + +Multi-view representation learning (MvRL) aims to unify diverse views into a comprehensive and robust entity representation [1, 2]. These views can include multi-modal sensor data (e.g., visual, auditory, and tactile inputs from a robot) or multi-feature data from distinct extraction tech- + +![](images/716fc7650e3abdfcadc4458415c419ba168e12eebe9c3018d8991336025eac7b.jpg) +(a) A multi-modal sentiment analysis example. + +![](images/cec227fb5d337de4f0adfd3c34e381dbfacb9bcbc289105f0df087b56b538f67.jpg) +(b) Parallel Fusion +(c) Sequential Fusion +Figure 1. Comparison of two fusion pipelines using the mean fusion strategy. (a) shows an example of sentiment analysis with conflicts between views. Conflicts hinder Parallel Fusion (b). Sequential Fusion (c) can explicitly reduce the conflict and achieve a more flexible and effective fusion. + +niques like Sobel, Gabor, and HOG, with each contributing to a complete representation of the underlying entity [3]. MvRL has achieved great success across numerous downstream tasks, from traditional classification and clustering tasks to emerging embodied perception tasks [4-7]. However, the inherent structural heterogeneity and potential redundancy among views pose significant challenges to effective fusion in MvRL. + +To tackle these challenges, existing methods strive to integrate as much consistency and complementary information as possible while minimizing redundancy in + +heterogeneous multi-view data [8]. Current integration strategies can be broadly categorized into parameter-free and parameter-based methods. Compared to traditional parameter-free methods, such as concatenation and weighted fusion, parameter-based methods are more effective in handling complex information components and have been extensively explored [9]. For example, mutual information-based methods [10, 11] enhance consistency by maximizing mutual information between views while preserving complementary information. Similarly, attention-based methods [12, 13] and adversarial learning-based methods [14] reduce redundancy by learning view weights and introducing adversarial loss functions, respectively. Despite these advancements, most existing methods adopt a parallel fusion framework that fuses all views simultaneously, lacking the flexibility needed to resolve potential conflicts among views. + +Sequential fusion is a promising strategy for reducing conflicts in multi-view data by detecting conflicts and redundancy at each step, preserving valuable information, and enabling the effective use and reuse of views, yet it remains underexplored [15]. Figure 1 (a) presents a sentiment analysis example, where the textual view clearly conveys positive information, while the other two views reflect conflicting information due to diverse sentiment expression. In this example, the parallel fusion in Figure 1 (b) fails to produce a decisive conclusion, while the sequential fusion in Figure 1 (c) utilizes its inherent flexibility to alleviate these conflicts and preserve more meaningful information. Despite the effectiveness of sequential fusion, the core challenge in sequential fusion is identifying the optimal view sequence without effective supervised information, as highlighted by preliminary experiments in Subsection 5.2, which emphasize the importance of a well-chosen sequence. + +This paper introduces SeqMvRL, a novel sequential fusion method designed to tackle this challenge. We model multi-view fusion as a sequential decision-making problem and address it using reinforcement learning. SeqMvRL consists of two main components: the Pairwise Integrator (PI) and the Next View Selector (NVS), corresponding to the environment and agent in reinforcement learning, respectively. During agent-environment interactions, the PI maintains a fused feature state, while the NVS selects views based on a novel reward function that emphasizes cluster compactness and separation. At each fusion step, the NVS takes an action based on the current state, selecting the next view to fuse by maximizing the cumulative reward, effectively discarding conflicting views with low rewards. The PI then integrates the selected view into the current feature, updating the state. The process terminates when the "End" action is chosen, producing a unified, discriminative representation optimized for downstream tasks. The main contributions of this work can be summarized as follows: + +- We propose SeqMvRL, a sequential fusion method for multi-view representation learning, enabling flexible and effective view fusion and high-quality representations. +- We build a reinforcement learning agent for the next-view selector and design a novel reward function emphasizing cluster compactness and separation, aiming to select views that enhance the discriminability of fused features. +- Our SeqMvRL surpasses existing state-of-the-art methods in clustering and classification tasks, especially when dealing with low-quality or numerous views. + +# 2. Related Works + +# 2.1. Multi-View Representation Learning + +MvRL is a classical topic that received sustained attention, with many studies systematically reviewing its advancements [4, 16, 17]. This paper focuses on "how to fuse" multiple views and roughly categorizes existing fusion strategies into parameter-free and parameter-based fusion. Parameter-free fusion refers to simple handcrafted strategies without trainable parameters, such as feature concatenation [18, 19] and weighted averaging [9, 20]. These methods are efficient for handling low-noise and simple inter-view relationships but may fail when dealing with more complex, heterogeneous, or inconsistent views. Parameter-based fusion leverages trainable parameters to capture complex inter-view relationships, excelling in multi-modal and heterogeneous environments. Examples include sample-level weighted fusion [9, 21], attention-based fusion [12, 13], and GNN-based fusion [22, 23]. Despite careful design, a fundamental yet easily overlooked problem lies in how these strategies are utilized. Most existing MvRL methods employ them in parallel, fusing all views at once, which overlooks potential conflicts and noise between views. In contrast, sequential fusion integrates views in sequence based on their importance, offering greater flexibility. However, exploration of sequential fusion remains limited, particularly at the sample level, which is one of the core contributions of this work. + +# 2.2. Reinforcement Learning + +Reinforcement learning (RL) is a learning paradigm that optimizes decision-making by refining strategies based on feedback signals[24-26]. This paradigm enables adaptive responses to complex, dynamic environments and improves performance in multi-step tasks. This adaptability has driven RL's integration with other learning paradigms [27-29]. In transfer learning, RL supports effective cross-domain knowledge transfer, especially for tasks with structural similarities to the source domain. For example, MetaRL [30, 31] enables models to quickly adapt to new tasks or environments. In multi-task learning, RL facilitates joint training across related tasks by maximizing shared informa + +![](images/47d5bceb5f079278a3fc63b538a333b43c303d276ffaa18120fadd9f6716bc22.jpg) +Figure 2. The overall framework of SeqMvRL consists of three modules. Heterogeneous multi-view data is first mapped to latent features with consistent dimensions by the Latent Feature Extractor (LEF). Then, the Next View Selector (NVS) and Pairwise Integrator (PI) alternately perform sequential fusion. The former selects the next view to be fused based on the current state, while the latter fuses them and updates the current state. Once the NVS selects the "End" action, the sequential fusion is complete, and the current state serves as the final unified representation for downstream tasks. + +tion [32-34]. In embodied learning, RL optimizes control strategies for robots and virtual agents through environment interactions [35, 36]. However, RL remains underutilized in the multi-view field. We believe that RL possesses unique advantages in learning sequential fusion for multiple views, potentially offering a new perspective for multi-view learning and inspiring future research. + +# 3. Problem Definition + +Given a multi-view dataset $\mathcal{D} = \{X_i\}_{n=1}^N$ , where $n$ and $N$ represent the index and number of samples, respectively. Each $X_i = \{x_i^v\}_{v=1}^V$ consists of $V$ views, with the superscript $v$ indicating the view index. Correspondingly, the matrix composed of the $v$ -th view across all samples is denoted by $X^v$ . For a sample $X_i$ , the goal of MvRL is to integrate information from its $V$ views to obtain a unified representation $U_i = F_\circ(X_i)$ , where $F_\circ$ denotes the fusion process employing strategy $\circ$ . This representation $U_i$ is expected to comprehensively capture the characteristics of $X_i$ , thereby enhancing performance in downstream tasks. + +Parallel fusion. Existing research follows a parallel fusion manner, where all views are fused at once. A typical parallel fusion pipeline can be refined as: + +$$ +U _ {i} = F _ {\circ} ^ {\operatorname {p a r}} \left(X _ {i}\right) \triangleq \circ \left(x _ {i} ^ {1}, x _ {i} ^ {2}, \dots , x _ {i} ^ {V}\right). \tag {1} +$$ + +As mentioned above, parallel fusion overlooks differences in information richness and relevance across views, lacking the flexibility to resolve inter-view conflicts, which ultimately reduces the quality of the fused representation. + +Sequential fusion. As an alternative, this work explores a novel sequential fusion framework. Here, we consider a simple pairwise sequential fusion for clarity. Given a fusion sequence $\pi_i = \{\pi_{i,1},\pi_{i,2},\dots,\pi_{i,T}\}$ for sample $X_{i}$ and applying the same fusion strategy $\circ$ , the sequential fu + +sion pipeline can be formalized as: + +$$ +U _ {i} = F _ {\circ} ^ {\operatorname {s e q}} \left(X _ {i}, \pi\right) \triangleq \left(\left(\left(x _ {i} ^ {\pi_ {1}} \circ x _ {i} ^ {\pi_ {2}}\right) \circ \dots\right) \circ x _ {i} ^ {\pi_ {V}}\right). \tag {2} +$$ + +The proposed sequential fusion enables controlled and flexible integration of information from each view, effectively preventing conflicts between views. It is also scalable and generalizable, compatible with the most carefully designed existing fusion strategies. + +# 4. Methodology + +In MvRL tasks, a key challenge for the sequential fusion framework is determining the view sequence. To tackle this, we propose a specific solution, SeqMvRL. As shown in Figure 2, the framework consists of three modules: Latent Feature Extractor, Pairwise Integrator, and Next-View Selector. The following subsections describe them in detail. + +# 4.1. Latent Feature Extractor + +The Latent Feature Extractor (LFE) transforms the original multi-view data into latent features with consistent dimensions, aiming to reduce disparities arising from differing data formats. Existing work [37] demonstrates that such dimensional alignment strategy facilitates more effective representation learning in subsequent stages and has become a widely adopted fundamental operation [38, 39]. In this work, we instantiate the LEF as a set of view-specific autoencoders, $F_{\mathrm{LEF}} = \{f_{\mathrm{enc}}^v, f_{\mathrm{dec}}^v\}_{v=1}^V$ , where each encoder $f_{\mathrm{enc}}^v$ with the parameter $\theta_{\mathrm{enc}}^v$ transforms the corresponding raw view data $X^v \in \mathbb{R}^{d^v \times n}$ into latent feature $Z^v \in \mathbb{R}^{d \times n}$ : + +$$ +Z ^ {v} = f _ {\mathrm {e n c}} ^ {v} \left(X ^ {v}; \theta_ {\mathrm {e n c}}\right). \tag {3} +$$ + +The latent features are expected to retain as much information as possible from the original data. Therefore, the decoder $f_{dec}^{v}$ with a symmetrical structure subsequently maps + +these features back to the original view space. By minimizing a reconstruction loss, the LEF learns to retain key information from the original views in the latent features. The reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ is typically defined as: + +$$ +\mathcal {L} _ {\mathrm {r e c}} = \sum_ {v = 1} ^ {V} \| X ^ {v} - f _ {\mathrm {d e c}} ^ {v} \left(Z ^ {v}; \theta_ {\mathrm {d e c} ^ {v}}\right) \| _ {2} ^ {2}, \tag {4} +$$ + +where $\theta_{\mathrm{dec}}$ and $\| \cdot \| _2$ denote the parameter of decoder $f_{dec}^{v}$ and $\ell_2$ norm. As an essential pre-processing step in multiview learning, this module's model design and optimization are not central to our contributions but provide a robust foundation for sequential fusion. + +# 4.2. Pairwise Integrator + +The pairwise integrator (PI) is essentially a specialized multi-view fusion model, designed to incrementally produce a fused representation based on a specified view sequence. It takes two inputs: the current fusion state and the features of the next view to be fused. Based on this analysis, existing two-view fusion models can serve as PI's specific implementations. Inspired by previous works [40, 41], our pairwise integrator consists of a two-layer neural network. For a sample $X_{i}$ and its corresponding fusion order $\pi_{i} = \{\pi_{i,1},\pi_{i,2},\dots,\pi_{i,t},\dots,\pi_{i,T}\}$ , the fusion process at step $t$ can be formalized as: + +$$ +U _ {i} ^ {(t)} = F _ {\mathrm {P I}} \left(U _ {i} ^ {(t - 1)} \oplus z _ {i} ^ {\pi_ {i, t}}; \theta_ {\mathrm {P I}}\right), \tag {5} +$$ + +where $\oplus$ denotes feature concatenation, $\theta_{\mathrm{PI}}$ is the parameter of PI, $U_{i}^{(t)}$ and $U_{i}^{(t - 1)}$ are the fusion state at step $t$ and $t - 1$ respectively. The $z_{i}^{\pi_{i,t}}$ is the latent feature of the $\pi_{i,t}$ -th view for the sample $X_{i}$ . + +Considering that multi-view representation learning typically lacks supervisory information, we adopted the Deep Divergence-based Clustering (DDC) loss [14], which has demonstrated state-of-the-art performance in both single-view and multi-view clustering [42, 43]. To this end, we added an additional clustering head to the PI, consisting of a fully connected layer with a softmax activation, to obtain the $k$ -dimensional vector of soft cluster assignments, $\alpha_{i}$ . The DDC loss consists of three components: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {D D C}} = \sum_ {i = 1} ^ {k - 1} \sum_ {j = i + 1} ^ {k} \frac {\left(\frac {k}{2}\right) ^ {- 1} \sum_ {a = 1} ^ {n} \sum_ {b = 1} ^ {n} \alpha_ {a i} \kappa_ {a b} \alpha_ {b j}}{\sqrt {\sum_ {a = 1} ^ {n} \sum_ {b = 1} ^ {n} \alpha_ {a i} \kappa_ {a b} \alpha_ {b i} \sum_ {a = 1} ^ {n} \sum_ {b = 1} ^ {n} \alpha_ {a j} \kappa_ {a b} \alpha_ {b j}}} \\ + \sum_ {i = 1} ^ {k - 1} \sum_ {j = i + 1} ^ {k} \frac {\left(\frac {k}{2}\right) ^ {- 1} \sum_ {a = 1} ^ {n} \sum_ {b = 1} ^ {n} m _ {a i} \kappa_ {a b} m _ {b j}}{\sqrt {\sum_ {a = 1} ^ {n} \sum_ {b = 1} ^ {n} m _ {a i} \kappa_ {a b} m _ {b i} \cdot \sum_ {a = 1} ^ {n} \sum_ {b = 1} ^ {n} m _ {a j} \kappa_ {a b} m _ {b j}}} \\ \end{array} +$$ + +$$ ++ \left(\frac {n}{2}\right) ^ {- 1} \sum_ {i = 1} ^ {n - 1} \sum_ {j = i + 1} ^ {n} \alpha_ {i} ^ {T} \alpha_ {j}, \tag {6} +$$ + +where $k$ is the number of clusters, $\kappa_{ij} = \exp (-||U_i^{(T)} - U_j^{(T)}||^2 /(2\sigma^2))$ , $m_{ij} = \exp (-||\alpha_i - e_j||^2)$ , $e_j$ is the $j$ -th vertex of the standard simplex in $\mathbb{R}^k$ , and $\sigma$ is a hyperparameter. The first term promotes cluster separability and compactness, while the latter two terms encourage sharpen cluster boundaries and distinct assignments, respectively. Note that the DDC loss can be replaced by other clustering losses, as it is not the main contribution of this paper. + +# 4.3. Next-View Selector + +Now, we can flexibly obtain a unified representation of samples through the PI based on a given fusion sequence. The final key question is how to select an appropriate sequence. The next-view selector (NVS) is designed for this purpose based on reinforcement learning. In this context, the PI and NVS can be regarded as the terms of environment and agent in reinforcement learning, respectively. The environment PI maintains a current fused state. In each fusion step, the agent NVS takes an action to select the next view to be fused. The PI then integrates it into the current fused features, updates the state, and provides reward feedback. Guided by the reward, the NVS learns the optimal selection strategy. We begin with several key terms in reinforcement learning, then introduce the learning process. + +State. The state refers to the information observable by the NVS, including the current fused features and the historical action vector. The current fused features serve as the basis for the NVS to decide its next action, while the historical action vector helps determine if a view has been reused. In Subsection 5.4, we experimentally compare three scenarios: allowing view reuse, disallowing reuse, and conditional reuse with a cooldown window. + +Action. In the sequential fusion framework, each action represents the latent features of a specific view, with an additional action representing "END". Thus, for data with $V$ views, the action space has $(V + 1)$ dimensions, encompassing all possible actions. At each step, NVS outputs a $(V + 1)$ -dimensional vector indicating the value of each action. At step $t$ , the output can be expressed as: + +$$ +a = F _ {\mathrm {N V S}} \left(U _ {i} ^ {(t)}; \theta_ {\mathrm {N V S}}\right), \tag {7} +$$ + +where $U_{i}^{(t)}$ denotes the fusion state at step $t$ , $a \in \mathbb{R}^{V + 1}$ represents action values, and $F_{\mathrm{NVS}}$ is the agent model with the parameter $\theta_{\mathrm{NVS}}$ . Like PI, the agent model is structured as a two-layer MLP. Based on this, the NVS selects an action, i.e., the corresponding view. Once the "END" action is chosen, the sequential fusion process terminates, and the current fusion state becomes the final unified representation. + +Reward. The reward is essential for guiding the agent to select the optimal view at each fusion step. Due to the lack of supervision, "optimal" is defined here as achieving smaller intra-cluster distances and larger inter-cluster distances. Thus, we initialize with features in a default order and compute each cluster's center feature using KMeans. At each step, the changes in clusters before and after fusion are evaluated by calculating the intra-cluster distances $d_{\mathrm{intra}} = ||U_i - c_k||$ and inter-cluster distances $d_{\mathrm{inter}} = \sum_{j \neq k} ||U_i - c_j||$ , where $c_k$ is the corresponding cluster center. We further use the cluster assignments $\alpha_i$ derived from the PI as pseudo-labels to maintain assignment consistency. The final reward $r$ is calculated by: + +$$ +r = (1 - \gamma) \left(\frac {\alpha}{1 + d _ {\text {i n t r a}}} + \beta \cdot d _ {\text {i n t e r}}\right) + \gamma \cdot \left(2 \cdot \delta_ {\hat {y} _ {i}, y _ {i}} - 1\right), \tag {8} +$$ + +where $\alpha$ , $\beta$ , and $\gamma$ are balancing parameters. The $\delta_{\hat{y}_i, y_i}$ is the Kronecker delta, defined as $\delta_{\hat{y}_i, y_i} = 1$ if $\hat{y}_i = y_i$ and 0 otherwise, ensuring a reward of $+1$ when the prediction is correct and $-1$ when incorrect. This reward guides the NVS to select the view that best preserves cluster structure at each step. Appendix A provides an intuitive explanation. Agent. Guided by the reward feedback from the environment, the agent learns the optimal policy to maximize cumulative rewards in the given task. Existing reinforcement learning algorithms can serve as specific implementations for the agent. In this work, we choose Deep Q-Network (DQN) [44], which effectively handles discrete action spaces, making it well-suited for next-view selection [45]. Moreover, the experience replay and target network strategies in DQN enhance training stability, which is crucial for MvRL tasks. + +We denote $Q^{*}(s,a)$ as the optimal action-value function, which represents the maximum expected cumulative reward the agent can achieve by taking action $a$ in state $s$ and following the optimal policy thereafter. DQN uses a deep neural network to estimate Q-values, which can be updated iteratively using the Bellman equation: + +$$ +Q ^ {*} (s, a) = \mathbb {E} _ {(s, a, r, s ^ {\prime})} \left[ \left(r + \mu \cdot \max _ {a ^ {\prime}} Q \left(s ^ {\prime}, a ^ {\prime}\right) - Q (s, a)\right) ^ {2} \right], \tag {9} +$$ + +where $r$ is the immediate reward, $\mu$ is the discount factor, set to $\mu = 0.98$ in this paper, and $s'$ is the next state after action $a$ . This iterative process enables the agent to learn a policy that optimally selects the next view at each step, maximizing the long-term quality of the sequential fusion. + +# 4.4. Training Process + +Our framework adopts a three-stage training strategy: + +Stage 1: Pre-training the LFE. By minimizing reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ , we ensure that essential information from each view is preserved in the latent space. + +Algorithm 1 SeqMvRL +1: Input: Multi-view dataset $\mathcal{D} = \{X_i\}_{i=1}^N$ , where $X_i = \{x_i^v\}_{v=1}^V$ , LFE $F_{\mathrm{LFE}}(\cdot; \theta_{\mathrm{enc}}, \theta_{\mathrm{dec}})$ , PI $F_{\mathrm{PI}}(\cdot; \theta_{\mathrm{PI}})$ , and NVS $F_{\mathrm{NVS}}(\cdot; \theta_{\mathrm{NVS}})$ and target NVS. +2: Output: Fused representations $\{U_i\}_{i=1}^N$ . +3: Stage 1: Pre-training the LEF +4: while not converged do +5: Update $\theta_{\mathrm{enc}}, \theta_{\mathrm{dec}}$ by minimizing loss $\mathcal{L}_{\mathrm{LFE}}$ . +6: end while +7: Stage 2: Training PI +8: while not converged do +9: Sample a batch of data $\mathcal{B} \subset \mathcal{D}$ . +10: Randomly select a sequence length $T$ . +11: Construct a view sequence of length $T$ by randomly selecting from available views. +12: Update $\theta_{\mathrm{PI}}$ by minimizing loss $\mathcal{L}_{\mathrm{PI}}$ . +13: end while +14: Stage 3: Reinforcement Learning for the NVS +15: Obtain cluster centers $c$ and assignments $\alpha$ from PI. +16: while not converged do +17: Sample a data, initialize state $s$ and action $a$ +18: while $a \neq$ "End" do +19: Compute action value, reward $r$ and new state $s$ +20: Store $[s, a, r, s', \text{done}]$ in sequence pool $\mathcal{P}$ . +21: end while +22: Sample a batch of sequences from $\mathcal{P}$ . +23: Update parameters $\theta_{\mathrm{NVS}}$ by optimizing the DQN objective, and periodically update the target network. +24: end while + +Stage 2: Training the PI. After obtaining latent features for each view via the pre-trained autoencoder, we train the PI module by minimizing multi-view clustering loss $\mathcal{L}_{\mathrm{DDC}}$ . During training, the PI is exposed to randomly generated fusion sequences of varying lengths to enhance the robustness and generalization capabilities. + +Stage 3: Reinforcement Learning for the NVS. Based on the trained PI, we initialize cluster centers and assignments. By optimizing the DQN objective with established experience replay and target network strategies, the optimal fusion sequence is learned for each sample. + +This three-stage training strategy allows each module in the framework to be flexibly replaced with more advanced models. Algorithm 1 outlines the complete training process. + +# 5. Experiments + +We conduct extensive experiments on five datasets. Qualitative preliminary experiments verify the feasibility of sequential fusion. Comparative experiments on classification and clustering tasks show that SeqMvRL outperforms the state-of-the-art methods. Additional analyses validate its + +
MethodsFlower17COIL-20Caltech-7Caltech-20
ACCcluNMIARIACCcluNMIARIACCcluNMIARIACCcluNMIARI
EAMC (CVPR-20)30.1540.6611.4158.0070.5145.2837.7150.3028.7429.6153.8919.95
SiMVC (CVPR-21)30.8040.8112.4260.3372.6146.1545.7954.6638.6129.5349.0023.48
CoMVC (CVPR-21)31.5541.5812.2968.8381.3057.4039.8948.3425.3028.6052.3226.68
DCP (TPAMI-22)39.3450.3931.7470.6785.4768.9648.1554.6238.0839.1349.9730.78
DSMVC (CVPR-22)34.7944.5516.5777.0785.6270.5251.5250.2434.5648.2466.4437.36
MFLVC (CVPR-22)31.9935.7616.3966.1874.5458.1245.3248.1230.6539.3555.7234.27
PDMF (AAAI-23)44.1252.6525.3469.3380.2960.6847.1446.3733.6536.8556.1024.90
RCML (AAAI-24)38.6046.3538.6477.0080.5971.6049.6148.9534.8545.8665.6134.06
CLOVEN (TCSVT-24)30.5135.8016.1775.3780.1269.4647.7150.8228.1947.6164.8935.36
SeqMvRL (Our)46.2355.1941.0279.6887.4774.6258.2556.2444.6447.8567.7237.30
Δ SOTA↑2.11↑2.54↑2.38↑2.61↑1.85↑3.02↑6.73↑1.58↑6.03↓0.39↑1.28↓0.06
+ +Table 1. Clustering results (%) of all methods on four datasets. Bold and underline denote the best and the second-best results, respectively. + +flexibility and effectiveness. + +# 5.1. Experimental Setup + +Datasets. To assess the effectiveness of our SeqMvRL, we conduct experiments on four commonly used multi-view datasets varying in number of views and categories. Flower17 contains 1360 images belonging to 17 different types of flowers, and each image corresponds to seven views [46]. COIL-20 includes 1440 gray images from 20 classes. Three distinct features, i.e., intensity, LBP, and Gabor, are extracted as three views. Caltech-7 and Caltech-20 are two subsets of Caltech-101, which contains 8677 object images from 101 classes. For Caltech-7, 1474 images with only 7 classes are chosen. Caltech-20 is composed of 2386 images with a total of 20 classes. Similar to [47], we extract six kinds of visual features (Gabor, WM, CENTRIST, HOG, GIST, and LBP) for both datasets. Each dataset is scaled to [0, 1] and divided into training set (60%), validation set (20%), and test set (20%). + +Baselines. Nine state-of-the-art MvRL methods are used for comparison with SeqMvRL as follows: end-to-end adversarial-attention networks (EAMC) [14]; contrastive multi-view clustering (SiMVC and CoMVC) [43]; dual contrastive prediction (DCP) [48]; deep safe multi-view clustering (DSMVC) [42]; multi-level feature learning for contrastive multi-view clustering (MFLVC) [41]; progressive deep multi-view fusion (PDMF) [49]; reliable conflictive multi-view learning (RCML) [50]; and clustering-guided contrastive fusion (CLOVEN) [51]. + +Evaluation Metrics. We evaluate the performance of SeqMvRL and baselines for clustering and classification tasks. Three standard metrics, including clustering accuracy $(\mathrm{ACC}_{clu})$ , normalized mutual information (NMI), and adjusted rand index (ARI), are employed to evaluate the clustering performance. For the multi-view classification task, accuracy $(\mathrm{ACC}_{cls})$ , precision (Prec), and F-score are + +used in experiments. A higher value indicates better performance on downstream tasks for all metrics. Further details on these metrics can be referred to [51]. + +Implementation Details. All experiments are implemented with PyTorch on a single NVIDIA RTX 4090 GPU, and average results of all evaluation metrics across 10 independent runs are reported to mitigate randomness. For a fair comparison, we fix the unified representation dimension to 256 across all methods. Regarding clustering-specific methods, we only utilize latent representations learned from them. Following previous works [40], the results on clustering and classification tasks are obtained by post-processing K-means and SVM classifiers, respectively. For our SeqMvRL, we train LFE, PI, and NVS modules for 500, 2000, and 200 epochs, respectively, employing the Adam optimizer with a learning rate of $10^{-4}$ . Both the fusion network in PI and the agent network in NVS are MLP networks with a hidden layer dimension of 768. The balance weight $\gamma$ in the reward is set to 0.5, while $\alpha$ and $\beta$ are set to 1 for simplicity. It is important to note that all hyper-parameters of our method are selected based on the validation set. + +# 5.2. Qualitative Preliminary Experiments + +Before comprehensive experiments, we conducted a preliminary clustering experiment on the Caltech-7 to compare SeqMvRL with conventional parallel and permutated sequential fusion. Parallel fusion was implemented by modifying the PI's input and training form, while permutated sequential fusion tested all view orders exhaustively. Figure 3 shows representative results and reveals three key observations: 1) The best result of permutated sequential fusion outperforms parallel fusion, owing to its flexibility; 2) The view order is crucial for sequential fusion. On the one hand, the default order ([0,1,2,3,4,5]) is suboptimal; on the other hand, an inappropriate order can lead to performance degradation; 3) Our SeqMvRL model learns sample-level view + +
MethodsFlower17COIL-20Caltech-7Caltech-20
ACCclsPrecF-scoreACCclsPrecF-scoreACCclsPrecF-scoreACCclsPrecF-score
EAMC (CVPR-20)36.4036.6332.5967.6765.9864.7387.8846.9848.0975.1647.1844.10
SiMVC (CVPR-21)34.6332.5031.7768.0070.5265.7887.2142.1442.0061.6246.1047.36
CoMVC (CVPR-21)35.9636.8032.8976.7777.3774.6188.3250.1149.0472.9148.0541.85
DCP (TPAMI-22)37.4239.6433.5280.6780.5680.4490.7756.2248.3178.2661.4960.77
DSMVC (CVPR-22)34.7238.0731.8277.6775.3874.8081.4823.0725.3380.3360.3753.11
MFLVC (CVPR-22)37.5741.4634.7882.6481.6481.5884.6041.7332.4383.4966.0061.03
PDMF (AAAI-23)40.7841.3836.0689.3387.3886.9391.2960.0763.1384.4167.7358.74
RCML (AAAI-24)39.7338.4529.7574.0371.5869.3280.6737.5532.2179.6257.2848.42
CLOVEN (TCSVT-24)37.6339.9330.0475.9174.8473.4580.7227.0523.1383.8159.7856.27
SeqMvRL (Our)41.9143.2635.7390.3386.3384.5193.2764.0665.0386.1368.3760.48
Δ SOTA↑1.13↑1.80↓0.33↑1.00↓1.05↓2.42↑1.98↑3.99↑1.90↑1.72↑0.64↓0.55
+ +Table 2. Classification results of all methods on four datasets. Bold and underline denote the best and the second-best results, respectively. + +![](images/eb2f98e0521246e65a4baa5ffe5f8d10d80d2184b5051c977a88b66fecf53a8a.jpg) +Figure 3. Comparison of our SeqMvRL with the parallel and permuted sequential fusion baselines. The last eight bars show a subset of view orders obtained through exhaustive search. + +sequences, offering greater flexibility than permutated sequential fusion and yielding the best results. These observations validate the feasibility of sequential fusion and emphasize the importance of view order. + +# 5.3. Quantitative Comparison Experiments + +Clustering Tasks. Table 1 summarizes the clustering results of all methods across four datasets, leading to the following observations: 1) Our proposed SeqMvRL consistently outperforms other methods on all datasets, highlighting its effectiveness in clustering tasks. For example, on the Flower17 dataset, SeqMvRL achieves improvements of $2.11\%$ in $\mathrm{ACC}_{\mathrm{clu}}$ , $2.54\%$ in NMI, and $2.38\%$ in ARI compared to the best state-of-the-art method. 2) DSMVC delivers the second-best performance by mitigating conflicts among views through adaptive weighting. In contrast, SeqMvRL's sequential pipeline identifies conflicts at each fusion step and explicitly discards conflicting views, further enhancing its robustness. + +![](images/e44d74d9b8547e00d87358f39888698cac4d237bedfa863153824a20e8acced1.jpg) +Figure 4. Results of different combinations of balance parameters in the reward function on the COIL-20 dataset. + +Classification Tasks. Table 2 presents the classification results, where SeqMvRL surpassing all compared methods by an average accuracy improvement of $1.5\%$ over the state-of-the-art. Notably, while the representations learned by DSMVC are well-suited for clustering tasks, they underperform in classification. In contrast, the SeqMvRL's representations exhibit versatility across different tasks. + +# 5.4. Additional Analyses + +Ablation Study. We respectively test the representations obtained from the pre-trained LFE, the trained PI, and the complete SeqMvRL with NVs on COIL20 and Caltech-7, with results presented in Table 3. The objective of LFE is to retain as much information from each view as possible, including conflicting parts, which significantly impacts performance. PI addresses this issue by minimizing clustering loss but lacks the ability to determine the optimal sequence. + +Reward Function. The reward function is crucial for sequential selection and significantly impacts performance. This section focuses on tuning balance parameters within the reward function. Figure 4 illustrates the effects of different parameter combinations on the Caltech-7 dataset. Specifically, certain parameter combinations notably enhance clustering accuracy, highlighting the importance of + +
ModulesCOIL-20Caltech-7
LFEPINVSACCcluNMIARIACCclsPrecF-scoreACCcluNMIARIACCclsPrecF-score
✓54.3354.0357.0584.6782.8681.5941.0849.3729.8784.5144.3734.60
✓✓75.6783.4171.3188.3384.4782.3055.5655.0342.6390.9556.0060.93
✓✓✓79.6887.4774.6290.3386.3384.5158.2556.2444.6493.2764.0665.03
+ +Table 3. Ablation study results of key modules. + +
MethodsOxford-IIIT PetCAT
ACCcluNMIARIACCclsPrec.F-scoreACCcluNMIARIACCclsPrec.F-score
DSMVC86.8688.0180.8298.9999.0698.8327.8624.0017.3538.0632.2832.52
SeqMvRL (Our)87.8288.2382.2999.1999.2499.2030.6628.6719.0841.1135.2535.27
Δ SOTA↑0.96↑0.22↑1.47↑0.20↑0.18↑0.37↑2.80↑4.67↑1.73↑3.05↑2.97↑2.75
+ +![](images/10f916af6a862d1ca15747b1b840eeade859b9648ea2e8c68a472a056b3e11bf.jpg) +Figure 5. Visualization of sequential view selection process (View $4 \rightarrow$ View $2 \rightarrow$ View 1). + +tuning for optimal results. Overall, the hyperparameter settings demonstrate robustness within a reasonable range, and we set the default value to 1 in our experiments. + +Multi-modal Dataset. To validate the effectiveness of the SeqMvRL in multi-modal data scenarios, experiments were conducted on two multi-modal datasets: Oxford-IIIT Pet and CAT. The results presented in Table 4 demonstrate the superiority of our method in integrating multiple modalities. More details are provided in Appendix B. + +View Reuse. Another key attribute of the sequence is whether view reuse is permitted. Intuitively, as fused features are updated, the components within each view that can be effectively fused may change, suggesting that reusing views could add flexibility. We compared different view reuse constraints on the Caltech-7 dataset. The results are shown in Table 5, where 'w/ Reuse' and 'w/o Reuse' indicate whether reuse is allowed or not, respectively, and 'w/o Reuse- $t$ represents a cooldown mechanism allowing reuse only after step $t$ . Experimental results support this analysis. However, the benefit of allowing reuse depends on view quality; for instance, 'w/ Reuse' achieved the best results on COIL-20, indicating this should be set as a hyperparameter. + +Visualization of The View Selection Process. Figure 5 shows an example of view selection process, where view 4, view 2, and view 1 are sequentially selected step by step. At each step, all view choices are evaluated based on inter + +Table 4. Results on two multi-modal datasets. + +
ConstraintsACCcluNMIARIACCclsPrecF-score
w/ Reuse58.2556.2444.6493.2764.0665.03
w/o Reuse-158.0553.9742.3491.7963.1263.14
w/o Reuse-258.0553.8243.2991.7963.1863.42
w/o Reuse-358.0954.4243.9892.8164.1364.88
w/o Reuse-456.4651.7240.6390.7961.7761.83
w/o Reuse-556.7851.2641.0390.2561.8261.51
w/o Reuse56.8050.0940.1790.6361.7661.88
+ +Table 5. Comparison of performance for different reuse strategies on the Caltech-7 dataset. + +cluster distance, intra-cluster distance, and assignment accuracy. Our reward value $r$ integrates these factors, ensuring NVS selects views that maximize both cluster separation and assignment accuracy. + +# 6. Conclusion + +This work investigates a sequential fusion framework for multi-view representation learning, highlighting the importance of learning sample-specific fusion sequences and introducing the novel SeqMvRL method. By leveraging the proposed Latent Feature Extractor, Pairwise Integrator, and Next-View Selector, SeqMvRL dynamically determines optimal fusion sequences for each sample from heterogeneous multi-view data. Furthermore, a novel reward function is designed to emphasize cluster compactness and separation, guiding the model to select the most beneficial view for cluster partitioning in subsequent fusion based on the current fusion state. Experimental results demonstrate that the flexibility of sequential fusion leads to substantial performance improvements in both clustering and classification tasks. This research also offers new perspectives on multiview fusion strategies, including the role of fusion models in sequence selection and more flexible approaches beyond traditional pairwise fusion constraints. + +# 7. Acknowledgment + +The work is supported by Natural Science Foundation of China (No. 62206158), Natural Science Foundation of Shandong Province (No. ZR2024MF101, ZR2022QF097), Young Expert of Taishan Scholars (No. tsqn202312026), and Shandong University Young Scholar Future Plan. THE AUTHOR GRATEFULLY ACKNOWLEDGES THE SUPPORT OF K.C.WONG EDUCATION FOUNDATION. + +# References + +[1] Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu, and Qinghua Hu. Deep partial multi-view learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2402-2415, 2020. 1 +[2] Siwei Wang, Xinwang Liu, En Zhu, Chang Tang, Jiyuan Liu, Jingtao Hu, Jingyuan Xia, and Jianping Yin. Multi-view clustering via late fusion alignment maximization. In IJCAI, pages 3778-3784, 2019. 1 +[3] Xiaoqiang Yan, Shizhe Hu, Yiqiao Mao, Yangdong Ye, and Hui Yu. Deep multi-view learning methods: A review. Neurocomputing, 448:106-129, 2021. 1 +[4] Xuefei Li, Baodi Liu, Kai Zhang, Honglong Chen, Weijia Cao, Weifeng Liu, and Dapeng Tao. Multi-view learning for hyperspectral image classification: An overview. Neurocomputing, 500:499-517, 2022. 1, 2 +[5] Lihua Zhou, Guowang Du, Kevin Lu, Lizheng Wang, and Jingwei Du. A Survey and an Empirical Evaluation of Multi-View Clustering Approaches. ACM Computing Surveys, 56(7):1-38, 2024. +[6] Rundong He, Zhongyi Han, Xiushan Nie, Yilong Yin, and Xiaojun Chang. Visual out-of-distribution detection in open-set noisy environments. International Journal of Computer Vision, 132(11):5453-5470, 2024. +[7] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multi-View Transformer for 3D Visual Grounding. In CVPR, pages 15524–15533, 2022. 1 +[8] Yuxiu Lin, Hui Liu, Ren Wang, Gongguan Chen, and Caiming Zhang. Efficient and Effective Multi-View Subspace Clustering for Large-scale Data. arXiv preprint, arXiv:2310.09718, 2023. 2 +[9] Liu Yang, Chenyang Shen, Qinghua Hu, Liping Jing, and Yingbo Li. Adaptive Sample-Level Graph Combination for Partial Multiview Clustering. IEEE Transactions on Image Processing, 29:2780-2794, 2020. 2 +[10] Karthik Sridharan and Sham M. Kakade. An Information Theoretic Framework for Multi-View Learning. In $COLT$ , number 114, pages 403–414, 2008. 2 +[11] Chenhang Cui, Yazhou Ren, Jingyu Pu, Jiawei Li, Xiaorong Pu, Tianyi Wu, Yutao Shi, and Lifang He. A novel approach for effective multi-view clustering with information-theoretic perspective. In NeurIPS, volume 36, 2024. 2 +[12] Meng Qu, Jian Tang, Jingbo Shang, Xiang Ren, Ming Zhang, and Jiawei Han. An Attention-based Collaboration Framework for Multi-View Network Representation Learning. In CIKM, pages 1767–1776, 2017. 2 + +[13] Zhiyong Xu, Weibin Chen, Ying Zou, Zihan Fang, and Shiping Wang. Attention-based stackable graph convolutional network for multi-view learning. *Neural Networks*, 180:106648, 2024. 2 +[14] Runwu Zhou and Yi-Dong Shen. End-to-End Adversarial-Attention Network for Multi-Modal Clustering. In CVPR, pages 14607-14616, 2020. 2, 4, 6 +[15] Sami Labbaki and Peter Minary. Orthogonal sequential fusion in multimodal learning. 2024. 2 +[16] Uno Fang, Man Li, Jianxin Li, Longxiang Gao, Tao Jia, and Yanchun Zhang. A Comprehensive Survey on Multi-View Clustering. IEEE Transactions on Knowledge and Data Engineering, 35(12):12350-12368, 2023. 2 +[17] Yingming Li, Ming Yang, and Zhongfei Zhang. A Survey of Multi-View Representation Learning. IEEE Transactions on Knowledge and Data Engineering, 31(10):1863-1883, 2019. 2 +[18] Zhuowei Cai, Limin Wang, Xiaojiang Peng, and Yu Qiao. Multi-View Super Vector for Action Recognition. In CVPR, pages 596–603, 2014. 2 +[19] Yongshan Zhang, Jia Wu, Zhihua Cai, and Philip S. Yu. Multi-View Multi-Label Learning With Sparse Feature Selection for Image Annotation. IEEE Transactions on Multimedia, 22(11):2844-2857, 2020. 2 +[20] Bin Zhang, Qianyao Qiang, Fei Wang, and Feiping Nie. Flexible Multi-View Unsupervised Graph Embedding. IEEE Transactions on Image Processing, 30:4143–4156, 2021. 2 +[21] Suyuan Liu, Junpu Zhang, Yi Wen, Xihong Yang, Siwei Wang, Yi Zhang, En Zhu, Chang Tang, Long Zhao, and Xinwang Liu. Sample-Level Cross-View Similarity Learning for Incomplete Multi-View Clustering. In AAAI, pages 14017-14025, 2024. 2 +[22] ErLin Pan and Zhao Kang. Multi-view Contrastive Graph Clustering. In NeurIPS, volume 34, pages 2148-2159, 2021. 2 +[23] Hao Wang, Yan Yang, and Bing Liu. GMC: Graph-Based Multi-View Clustering. IEEE Transactions on Knowledge and Data Engineering, 32(6):1116-1129, 2020. 2 +[24] Abhijit Gosavi. Reinforcement Learning: A Tutorial Survey and Recent Advances. INFORMS Journal on Computing, 2008. 2 +[25] L. P. Kaelbling, M. L. Littman, and A. W. Moore. Reinforcement Learning: A Survey. Journal of Artificial Intelligence Research, 4:237-285, 1996. +[26] Pawel Ladosz, Lilian Weng, Minwoo Kim, and Hyondong Oh. Exploration in deep reinforcement learning: A survey. Information Fusion, 85:1-22, 2022. 2 +[27] Zhuangdi Zhu, Kaixiang Lin, Anil K. Jain, and Jiayu Zhou. Transfer learning in deep reinforcement learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13344-13362, 2023. 2 +[28] Qingpeng Cai, Can Cui, Yiyuan Xiong, Wei Wang, Zhongle Xie, and Meihui Zhang. A survey on deep reinforcement learning for data processing and analytics. IEEE Transactions on Knowledge and Data Engineering, 35(5):4446-4465, 2023. + +[29] Ke Yu, Chao Dong, Liang Lin, and Chen Change Loy. Crafting a Toolchain for Image Restoration by Deep Reinforcement Learning. In CVPR, pages 2443-2452, 2018. 2 +[30] Abhishek Gupta, Russell Mendonca, YuXuan Liu, Pieter Abbeel, and Sergey Levine. Meta-Reinforcement Learning of Structured Exploration Strategies. In NeurIPS, volume 31, 2018. 2 +[31] Mandi Zhao, Pieter Abbeel, and Stephen James. On the Effectiveness of Fine-tuning Versus Meta-reinforcement Learning. In NeurIPS, volume 35, pages 26519-26531, 2022. 2 +[32] Siming Lan, Rui Zhang, Qi Yi, Jiaming Guo, Shaohui Peng, Yunkai Gao, Fan Wu, Ruizhi Chen, Zidong Du, Xing Hu, Xishan Zhang, Ling Li, and Yunji Chen. Contrastive modules with temporal attention for multi-task reinforcement learning. In NeurIPS, 2023. 3 +[33] Zhihai Wang, Jie Wang, Dongsheng Zuo, Ji Yunjie, Xilin Xia, Yuzhe Ma, Jianye Hao, Mingxuan Yuan, Yongdong Zhang, and Feng Wu. A Hierarchical Adaptive Multi-Task Reinforcement Learning Framework for Multiplier Circuit Design. In ICML, 2024. +[34] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Meta-world: A benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning. In CoRL, pages 1094–1100, 2020. 3 +[35] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 518(7540):529-533, 2015. 3 +[36] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Haoran Tan, Chencheng Jiang, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Octopus: Embodied Vision-Language Programmer from Environmental Feedback. In ECCV, pages 20-38, 2024. 3 +[37] Changqing Zhang, Huazhu Fu, Qinghua Hu, Xiaochun Cao, Yuan Xie, Dacheng Tao, and Dong Xu. Generalized Latent Multi-View Subspace Clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(1):86–99, 2020. 3 +[38] Ren Wang, Haoliang Sun, Xiushan Nie, Yuxiu Lin, Xiaoming Xi, and Yilong Yin. Multi-View Representation Learning via View-Aware Modulation. In ACM MM, pages 3876-3886, 2023. 3 + +[39] Changqing Zhang, Yeqing Liu, and Huazhu Fu. AE2-Nets: Autoencoder in Autoencoder Networks. In CVPR, pages 2577-2585, 2019. 3 +[40] Ren Wang, Haoliang Sun, Yuling Ma, Xiaoming Xi, and Yi-long Yin. MetaViewer: Towards a Unified Multi-View Representation. In CVPR, pages 11590–11599, 2023. 4, 6 +[41] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level Feature Learning for Contrastive Multi-view Clustering. In CVPR, pages 16030-16039, 2022. 4, 6 +[42] Huayi Tang and Yong Liu. Deep Safe Multi-View Clustering: Reducing the Risk of Clustering Performance Degradation Caused by View Increase. In CVPR, pages 202–211, 2022. 4, 6 +[43] Daniel J. Trosten, Sigurd Løkse, Robert Jenssen, and Michael Kampffmeyer. Reconsidering representation alignment for multi-view clustering. In CVPR, pages 1255–1265, 2021. 4, 6 +[44] Volodymyr Mnih. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. 5 +[45] Jianqing Fan, Zhaoran Wang, Yuchen Xie, and Zhuoran Yang. A theoretical analysis of deep q-learning. In Learning for dynamics and control, pages 486-489. PMLR, 2020. 5 +[46] Maria-Elena Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In CVPR, pages 1447-1454, 2006. 6 +[47] Shixin Yao, Guoxian Yu, Jun Wang, Carlotta Domeniconi, and Xiangliang Zhang. Multi-view multiple clustering. In IJCAI, pages 4121-4127, 2019. 6 +[48] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4447-4461, 2023. 6 +[49] Cai Xu, Wei Zhao, Jinglong Zhao, Ziyu Guan, Yaming Yang, Long Chen, and Xiangyu Song. Progressive deep multiview comprehensive representation learning. In AAAI, pages 10557-10565, 2023. 6 +[50] Cai Xu, Jiajun Si, Ziyu Guan, Wei Zhao, Yue Wu, and Xiyue Gao. Reliable conflictive multi-view learning. In AAAI, pages 16129-16137, 2024. 6 +[51] Guanzhou Ke, Guoqing Chao, Xiaoli Wang, Chenyang Xu, Yongqi Zhu, and Yang Yu. A clustering-guided contrastive fusion for multi-view representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2056-2069, 2024. 6 \ No newline at end of file diff --git a/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/images.zip b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f8f4109c099ba4f9cbf5a39693b472d36beb02c3 --- /dev/null +++ b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03fa4fe20f6c7f407683be259d95ce7d3d1bbbd5b525405a8e818aec1e230141 +size 629824 diff --git a/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/layout.json b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..679dd4905800d011d67fe2cc0339c4f934d614f8 --- /dev/null +++ b/CVPR/2025/SeqMvRL_ A Sequential Fusion Framework for Multi-view Representation Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e7d077c1cdd037a984f8e3878a49f6533c47cdc3fbc766f5796962a799b5af8 +size 413355 diff --git a/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/4db92e6e-13aa-4914-acf1-41ac9de88e3c_content_list.json b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/4db92e6e-13aa-4914-acf1-41ac9de88e3c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ea264370b7279bcb85b51f3f33d5f9e269457e0 --- /dev/null +++ b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/4db92e6e-13aa-4914-acf1-41ac9de88e3c_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a18e90bf1b68f0bfd49b1d54396f38e6f45033481656543fb3275c28305bfbf8 +size 68299 diff --git a/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/4db92e6e-13aa-4914-acf1-41ac9de88e3c_model.json b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/4db92e6e-13aa-4914-acf1-41ac9de88e3c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4e46cec1cbce44f6b61b3c5c1d550a11d7c29269 --- /dev/null +++ b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/4db92e6e-13aa-4914-acf1-41ac9de88e3c_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2aa271e409a97dc466f7cdacf080c06ab5e9e127a2fca98edded64cc29c94ce4 +size 82263 diff --git a/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/4db92e6e-13aa-4914-acf1-41ac9de88e3c_origin.pdf b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/4db92e6e-13aa-4914-acf1-41ac9de88e3c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..088c5ff1e9c72efd441f76eb943083c23635ff50 --- /dev/null +++ b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/4db92e6e-13aa-4914-acf1-41ac9de88e3c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4ebbd19f05a77acb4f684ed5f263134194a340c342125a61ab21aa6371a89ce +size 9766270 diff --git a/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/full.md b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f74298150277d6f8564f9822da60972e9a857164 --- /dev/null +++ b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/full.md @@ -0,0 +1,294 @@ +# SerialGen: Personalized Image Generation by First Standardization Then Personalization + +Cong Xie* + +Han Zou* + +Ruiqi Yu + +Yan Zhang† + +Zhenpeng Zhan† + +Global Business Unit, Baidu Inc. + +https://serialgen.github.io + +![](images/b28fcfe8d1e987336108f9a48ed538a9b22bed42c748fd43b74affd0543a893e.jpg) + +![](images/894d7c60eedad1a83d8b061986b53c252223a520983e07ebf0b8b5a4a8bdfd85.jpg) +Figure 1. Serial images generated by SerialGen. Our method can produce personalized images that faithfully recover the reference image's whole-body appearance while accurately responding to a wide range of text prompts. More results are in the supplementary material. + +![](images/0de1557f6c531bbcdf2b5244306ccafe5f92d0129af5511c303a58e46e243628.jpg) + +![](images/e7f0e0350424eb41a0b3d901de6dcd82cd26672100ee21f9916325469a0c119c.jpg) + +![](images/e35333fb709a3d04b6439a2df32b6512aa32483be22a56db17c7377bfc35305c.jpg) + +![](images/141a216b12bb6460eeccc24c257d3b92cf0ad69e4f603dcac4698bc726efc06b.jpg) + +![](images/68576baa67c33d00fd700982e7dfd06b3c2baf37898c507f0f1c61b34b464bee.jpg) + +![](images/8c55cf2deec434587bf82e3fbb70afee5420a92c83e9d94362d6d6cbe781ff46.jpg) + +![](images/84b172c8540187f6af865f720b5bb74c9a405595c11f68727de9bca960206c01.jpg) + +![](images/39e3e99c1ca2a0d6fea20db6f1cb59c8ec071c015ad4f67a06183c8ed312928d.jpg) + +![](images/dc8b192a2cb7a2b1b245afae1e64f9099caf65cb04a1ea889b7e7007d364d5a7.jpg) + +![](images/436c05e209719f339f20d2884d8606d6b2efb87856424258fec4a92f1c940626.jpg) + +# Abstract + +In this work, we are interested in achieving both high text controllability and whole-body appearance consistency in the generation of personalized human characters. We propose a novel framework, named SerialGen, which is a serial generation method consisting of two stages: first, a standardization stage that standardizes reference images, and then a personalized generation stage based on the standardized reference. Furthermore, we introduce two modules aimed at enhancing the standardization process. Our experimental results validate the proposed framework's abl + +ity to produce personalized images that faithfully recover the reference image's whole-body appearance while accurately responding to a wide range of text prompts. Through thorough analysis, we highlight the critical contribution of the proposed serial generation method and standardization model, evidencing enhancements in appearance consistency between reference and output images and across serial outputs generated from diverse text prompts. The term "Serial" in this work carries a double meaning: it refers to the two-stage method and also underlines our ability to generate serial images with consistent appearance throughout. + +# 1. Introduction + +Recently, text-to-image generation models based on diffusion methods [19, 21-23, 25] have experienced rapid development. These models demonstrate a remarkable capability to control the generated content based on the text prompt. Concurrently, the personalized text-to-image generation tasks [1, 3, 7, 9, 10, 18, 24, 27, 30, 33, 36] have garnered widespread attention due to their broad range of application scenarios. Given a reference image of a subject, the objective is to output additional images of the subject based on text prompts with high text controllability and appearance consistency. Text controllability demands a close match between the generated images and the text prompt. Appearance consistency ensures consistency between the generated images and the input subject. One category of work requires fine-tuning [7, 17, 24] the model for each new input subject. The practical application of this type of method is limited due to the high cost of fine-tuning. + +In this work, we focus on tuning-free methods for personalized generation of human characters [30, 33, 36], ensuring text controllability and whole-body appearance, including facial features, hairstyles, and clothing. To exempt from fine-tuning for every new subject, another type of approach typically involves a training phase on a substantial amount of subject data. The trained model is expected to generalize to new subjects, thereby avoiding the need for fine-tuning during the inference stage. Despite being more practical, this approach faces significant challenges. The training pairs of input (reference) and output (target) images utilized in existing methods typically derive from the same source image, generated by simple data augmentation operations such as cropping and flipping. Such training data poses significant challenges to model training due to the highly consistent content between the input and output images. + +In this study, we empirically observe that models trained with such data do not effectively achieve the goal of personalized generation tasks: they either compromise text controllability to maintain high appearance consistency or sacrifice appearance consistency to enhance text controllability. If the model is sufficiently powerful, it can easily replicate the reference image in the output to achieve minimal loss, thus achieving high appearance consistency but displaying inadequate responsiveness to text prompts such as an action different from the reference. Conversely, if the model is constrained to extract only compressed information from the reference, it often fails to capture all the appearance-related visual features and discards irrelevant features. This leads to a model that better aligns with text prompts, albeit at the cost of compromised appearance consistency. + +To alleviate the above-mentioned problems, we propose to train personalized models using paired images of (stan + +dardized reference image, target image), where both images share the same appearance but differ in non-appearance elements (NAE), such as background, pose, expression, and viewpoint. The standardized reference image will possess a standardized NAE, whereas the target image will feature non-standard NAEs. In addition to addressing the replication issue, another motivation for using standardized reference is that they simplify the generation task. It is easier for a feature extractor to process standardized input than to handle inputs with complex and varied characteristics. By standardizing NAE of the input images, the model can focus more on capturing the appearance features, leading to better personalized character generation. + +To achieve this goal, we introduce SerialGen, which is a serial generation method consisting of two stages: first, a standardization stage that standardizes reference images, and then a personalized generation stage based on the standardized reference. The advantages of this framework are threefold: + +- Employing different images as reference-target pairs can mitigate the issue of input-output replication, even when utilizing a powerful model. + +- Utilization of standardized reference enhances appearance consistency between reference image and output image. + +- Standardization enhances appearance consistency across serial images generated from different text prompts, which is highly beneficial in practical applications, such as comic story generation. For instance, if a reference image displays only a character's head, without standardization, the body's appearance could vary in different generation processes. By standardizing, the body is pregenerated in the standardized reference image, ensuring consistent appearance across various text prompts. + +Our contributions can be summarized as follows: 1) We introduce a novel first-standardization-then-personalization framework enabling to generation of images with high text controllability and maintaining superior whole-body appearance consistency. 2) Our comprehensive analysis underlines the pivotal role of the serial generation method and the standardization model, highlighting several key benefits. 3) We propose two innovative modules designed to augment the standardization model. The effectiveness of these modules is validated through experimental results. + +# 2. Related Works + +Personalized text-to-image generation based on diffusion models has experienced rapid development. These models can be categorized into two primary types: methods that require subject-specific fine-tuning and those that don't. Models necessitating fine-tuning, such as those proposed by [7, 8, 17, 24], demand additional training for each new input subject, which may limit their practicality. Conversely, + +models that forego fine-tuning offer enhanced flexibility, eliminating the need for further training. The bulk of research in this domain has concentrated on preserving facial identity in generated images by conditioning on facial images, as noted in works by [9-12, 18, 27, 33]. The facial features are extracted by pre-trained face recognition models [4, 5] and then integrated into diffusion models. Recently, efforts [30, 33, 36] have been directed towards generating images that maintain the full body appearance of characters, encompassing facial identity, clothing, hairstyle, and other attributes. + +Many recent methods train with unpaired images, leading to significant training difficulties and causing models to replicate these images. Efforts to collect paired human face data [12, 18] have involved generating and filtering images through facial recognition for similarity. However, this struggles with maintaining consistency in whole-body appearance, including clothing and hairstyles, due to difficulties in varying poses, backgrounds, and perspectives. Additionally, there's no model that effectively filters data based on whole-body appearance. + +Recent developments have made significant strides in generating multiple images or videos from a single reference image of a character using diffusion models [2, 14-16, 26, 28, 31, 34, 37]. These models are adept at capturing all details from the reference image, maintaining both the whole-body appearance of character and the background scene unchanged across a sequence of generated images. + +Furthermore, StoryDiffusion [35] proposes a consistent self-attention mechanism to preserve character consistency across a sequence of generated images. However, maintaining appearance consistency between the reference and target images is not the primary focus of their work. + +# 3. Methods + +# 3.1. Preliminaries + +Personalized Text-to-Image Generation Models The widely adopted framework for personalized text-to-image generation typically comprises two main components: a diffusion model and a reference encoder. The reference encoder is tasked with extracting visual features from the reference image, which are then integrated into the diffusion model. Latent diffusion models [19, 23] are frequently utilized within this context. A training image $\mathbf{x}$ is first converted into a latent representation through the VAE [6] encoder $\mathbf{z} = \varepsilon (\mathbf{x})$ . Subsequently, a noise $\epsilon$ is imposed to $\mathbf{z}$ at timestep $t$ , resulting in a noised latent $\mathbf{z}_t$ . A denoising UNet $\epsilon_{\theta}$ is employed to estimate the imposed noise. The training loss is defined as follows: + +$$ +\mathcal {L} = \mathbb {E} _ {\mathbf {x}, c, \epsilon , t} \| \epsilon - \epsilon_ {\theta} (\mathbf {z} _ {t}, t, c, \phi_ {\eta} (\mathbf {x})) \| _ {2} ^ {2} \tag {1} +$$ + +where $c$ refers to the text embeddings extracted from text prompts corresponding to $\mathbf{x}$ and $\phi_{\eta}$ is the reference encoder. + +Human Image Animation Models Our proposed framework is also closely related to a diffusion-based model for human image animation [15]. This model employs a dual-stream structure with a Denoising UNet and a ReferenceNet, which share the same backbone architecture. The Denoising UNet processes a guided pose alongside a noised image to predict the noise, whereas the ReferenceNet is responsible for encoding the reference image and integrating this information into the Denoising UNet. Training an animation model requires paired images of reference-target: $(\mathbf{x}^r,\mathbf{x}^o)$ . The training loss is defined as follows: + +$$ +\mathcal {L} = \mathbb {E} _ {(\mathbf {x} ^ {r}, \mathbf {x} ^ {o}), c, \epsilon , t} \| \epsilon - \epsilon_ {\theta} (\mathbf {z} _ {t} ^ {o}, t, c, \phi_ {\delta} (\mathbf {x} ^ {r})) \| _ {2} ^ {2} \tag {2} +$$ + +where $\mathbf{z}_t^o$ is $\mathbf{z}^o = \varepsilon (\mathbf{x}^o)$ diffused with noise $\epsilon$ at timestep $t$ , $c$ is the guide poses, $\phi_{\delta}$ denotes the ReferenceNet. The integration of ReferenceNet features is detailed as follows: For a given self-attention module within the ReferenceNet, the input feature map $\mathbf{f}_r\in \mathbb{R}^{\mathrm{h}\times \mathrm{w}\times c}$ is concatenated to the input feature map $\mathbf{f}_d\in \mathbb{R}^{\mathrm{h}\times \mathrm{w}\times c}$ of the identical self-attention module in the Denoising UNet. This concatenation occurs along the spatial dimension to produce $\mathbf{f}_c\in \mathbb{R}^{2\mathrm{h}\times \mathrm{w}\times c}$ . Subsequently, this concatenated feature map $\mathbf{f}_c$ is input into the self-attention module of the Denoising UNet. For further details, refer to the original paper. + +# 3.2. Overall Framework + +The overall framework of the proposed SerialGen is illustrated in Figure 2. SerialGen comprises two stages. In the first stage, a standardization model is introduced and trained on synthetic data. In the second stage, this standardization model is utilized to convert a set of real training images to standardized images. This process results in pairs of (standardized reference image, target image), which are then used to train the personalized text-to-image model. The inference pipeline is also divided into two steps. Upon receiving a reference image, it is first standardized, followed by the personalized text-to-image generation. + +# 3.3. Stage I: Standardization + +Given any reference image, Stage I transforms it into a new image with standardized NAE, while preserving the referenced appearance unchanged. To accomplish this, we introduce a standardization model. + +Standardization Model Inspired by the tasks of human image animation, which also involves altering human poses while maintaining the appearance invariant, we use a state-of-the-art human image animation framework [15] without temporal attention module as our baseline model. To + +![](images/16c70ed3407574ae5af1080c3429d5bea018b75b3ffa16bc44cdb0e6d97f53e2.jpg) +Stage I: Standardization + +![](images/8e55c6e10a310a4931eb728847f31aa09652f7de130db0584754cb8acfdf8124.jpg) +Stage II: Personalization + +![](images/b0b6ba9dc83dd8e8ee67e4c19dcb22c6b60568cbd41f9f9d6acb6ea4fcb977ee.jpg) +Inference Pipeline +Figure 2. Overview of the proposed SerialGen with two stages: (1) Standardization - training a standardization model on synthetic data, and (2) Personalization - using the standardization model to create (standardized reference, target) pairs for personalized text-to-image model training. During inference, once a reference image is standardized, serial images can be generated based on different text prompts. + +enhance the standardization of the background and pose, we propose two modules designed to improve the baseline model: 1) a Foreground-Background Distinction Module (FBDM), which is intended to explicitly integrate the background mask of the reference image into the model, and 2) a Reference Pose Injection Module (RPIM), aimed at explicitly incorporating the pose information from the reference image into the model, thereby enabling the model to accurately locate body parts. An overview of the standardization model is given in Figure 3. + +Foreground-Background Distinction Module Within each self-attention module of the ReferenceNet, we introduce two learnable class tokens: foreground token $\mathbf{v}_f \in \mathbb{R}^{1 \times 1 \times c}$ and background token $\mathbf{v}_b \in \mathbb{R}^{1 \times 1 \times c}$ . Given a background mask $\mathbf{m}_b \in \mathbb{R}^{h \times w}$ , $\mathbf{v}_f$ and $\mathbf{v}_b$ are separately added to the foreground and background regions of input feature map $\mathbf{f}_r \in \mathbb{R}^{h \times w \times c}$ in accordance with the mask: + +$$ +\tilde {\mathbf {f}} _ {r} = \left(\mathbf {f} _ {r} + \mathbf {v} _ {b}\right) \circ \mathbf {m} ^ {\prime} + \left(\mathbf {f} _ {r} + \mathbf {v} _ {f}\right) \circ \left(1 - \mathbf {m} ^ {\prime}\right) \tag {3} +$$ + +where $\mathbf{m}' \in \mathbb{R}^{h \times w}$ is resized version of $\mathbf{m}$ to fit the resolution of $\mathbf{f}_r$ . This method enables the explicit incorporation of mask information into the self-attention mechanism. + +Reference Pose Injection Module This module integrates pose information into the model. Specifically, the pose of the reference image is extracted utilizing DW-Pose [32]. Following this, a pose feature map is derived from the pose image via a light convolutional network composed of four layers. Within each self-attention module of the ReferenceNet, an additional convolutional layer is utilized to adjust the channel number of the pose feature to match the self-attention module. More details are given in the supplementary material. Furthermore, the resolution of the pose feature is interpolated to match that of the self-attention module. This processed pose feature $\mathbf{f}_p$ is subsequently added to $\tilde{\mathbf{f}}_r$ : + +$$ +\tilde {\mathbf {f}} _ {r} = \tilde {\mathbf {f}} _ {r} + \mathbf {f} _ {p} \tag {4} +$$ + +$\tilde{\mathbf{f}}_r$ is then concatenated to the feature $\mathbf{f}_d$ of Denoising U-Net for further processing. + +Synthetic Data To train the standardization model, paired images consisting of (non-standardized, standardized) are essential. However, the collection of such real-world data is expensive. In response to this challenge, we create a synthetic dataset by rendering animatable 3D character models. As illustrated in Figure 2, given a 3D character model, + +![](images/e0e64a1f77fc6d78edb1713dc04be75a610664c854539bdd7c1fd5d8ca535025.jpg) +Figure 3. Illustration of the standardization model. The pose and mask of the reference image are input into ReferenceNet to enhance the effect. + +images of the character in various poses, backgrounds, expressions, and viewpoints are rendered to represent the non-standardized images. Concurrently, utilizing the same character model, the standardized image is rendered adhering to a standard pose, against a white background, with a neutral expression, and the face scaled to a fixed position. This rendering pipeline was automated to generate a substantial volume of paired images for training purposes. + +# 3.4. Stage II: Personalization + +At this stage, pairs of images, consisting of (standardized reference image, target image), can be constructed utilizing a standardization model. As illustrated in Figure 2, a real image is standardized using a standardization model. The standardized output acts as the standardized reference image and the original real image serves as the target image. By applying standardization to a substantial corpus of real images, we can generate numerous paired images, which + +are crucial for training a personalized model. + +One might question the domain gap problem that arises when the standardization model, trained on synthetic data, is applied to real data for inference. Our experimental results demonstrate that while the appearance of the output image remains unchanged, its style is indeed biased towards the 3D style used during training. This is exemplified in the supplementary material. However, this bias does not impede the training or inference of personalized models. The reason is that the biased 3D style can be considered an integral part of the standardization. This bias can be effectively mitigated after training with real target images. + +Given paired data, a personalized model can be trained using the loss function described in Equation (1), with modifications applied to the input of the reference encoder: + +$$ +\mathcal {L} = \mathbb {E} _ {\mathbf {x}, c, \epsilon , t} \| \epsilon - \epsilon_ {\theta} (\mathbf {z} _ {t}, t, c, \phi_ {\eta} (\varphi (\mathbf {x})) \| _ {2} ^ {2} \tag {5} +$$ + +where $\varphi$ is the standardization model frozen during the training of the second stage. + +# 4. Experiments + +# 4.1. Implementation Details + +The Standardization Stage For the creation of synthetic data (non-standardized, standardized), we collect 3D character models sourced from publicly accessible websites. We filtered out characters that were not animatable, exhibited duplicate shapes, lacked skeletal structures, or presented texture defects. Following this filtration process, a total of 2,924 3D character models were retained. These models were subsequently paired with 18 distinct motion models, encapsulating 3,817 unique motion poses. The ensuing combination resulted in the generation of 10,513,090 data pairs, each rendered with a transparent background. During the training of the standardization model, non-standardized images were dynamically merged with 9,186 background images. The training was conducted utilizing 8 NVIDIA A100 GPUs, with a batch size of 4 per GPU. The training was conducted at a resolution of $512 \times 768$ for a duration of 50,000 steps. During the inference phase, reference images are padded to $512 \times 768$ and generated utilizing the DDIM [26] sampler, which employs 20 denoising steps. + +The Personalization Stage We adopt the existing model architecture, IP-Adapter [33], as the implementation for our second-stage personalization model. Specifically, we utilize Stable Diffusion XL (SDXL) [19] as our diffusion model, OpenCLIP ViT-H/14 [20] as the reference encoder, and a multilayer perceptron (MLP) layer for the adapter module, which generates 257 token features for cross-attention. The training dataset comprises approximately 300,000 character images, all of which have been standardized by trained standardization model, resulting in around 300,000 image pairs. + +During the training process, the adapter modules undergo training, while the rest part is kept frozen. The training procedure is conducted at a resolution of $512 \times 768$ , spanning 200,000 steps on 8 NVIDIA A100 GPUs, with a batch size of 8 per GPU. During the inference phase, we employ the DDIM sampler with 20 denoising steps. + +# 4.2. Quantitative and Qualitative Results + +We compare our method with three peer personalized models, all of which are tuning-free and capable of utilizing a full-body reference image. These models include StoryMaker [36], IP-Adapter [33], and FastComposer [30]. Following previous practices [33], we utilize the CLIP image similarity score (CLIP-I) to assess appearance consistency and the CLIP text similarity score (CLIP-T) to evaluate text controllability. To more precisely measure the similarity between two character images, we first remove the background regions of the images before computing the CLIP-I scores. Additionally, we employ face similarity (Face Sim.), a metric quantifying the cosine similarity between identity embeddings extracted using ArcFace [4]. To evaluate performance, we compiled a test dataset consisting of 40 characters, which includes 20 anime characters and 20 real-life individuals. We generated 20 distinct text prompts using ChatGPT-4, encompassing a diverse range of actions, backgrounds, viewpoints, and expressions. For each combination of prompt and character, we produced four images with randomness introduced. Unless specified otherwise, all subsequent experiments were conducted on this test dataset. + +As presented in Table 1, our method achieves the highest scores for both CLIP-T and CLIP-I. These results suggest that our proposed method surpasses others in terms of appearance consistency and text controllability. For the Face Sim. metric, our method ranks second, following StoryMaker. Notably, StoryMaker employs an additional face encoder to enhance facial consistency. Qualitative evidence supporting our findings is illustrated in Figure 4, where it is evident that the characters generated by our method exhibit superior performance compared to those produced by alternative approaches, particularly in the context of anime reference images. More results and analysis are given in the supplementary material. This underlines our method's superiority over competing methodologies. + +In addition to automatic evaluation, we also conducted a user study to further assess these methods. During each round of testing, participants were asked to select their preferred image according to a specific metric from among four methods. For more detailed information, please refer to the supplementary material. The results demonstrate that our methods received the majority of preferences across all metrics, thus surpassing other methods. This finding is consistent with the outcomes of the automatic evaluations. + +
MethodCLIP-I ↑CLIP-T↑Face Sim.↑
IP-Adapter [33]81.3120.220.39
FastComposer [30]73.9821.430.32
StoryMaker [36]84.2420.450.54
Ours85.4921.760.53
+ +Table 1. Comparison to other methods. + +
MethodCLIP-I ↑CLIP-T↑Face Sim.↑
unpaired-one-stage89.6216.080.52
paired-one-stage80.0021.990.41
two-stage (ours)85.4921.760.53
+ +Table 2. Comparison to one-stage methods. + +# 4.3. Analysis and Ablation Study + +# 4.3.1. Comparison to One-stage + +We compare our two-stage personalization model against two one-stage models, each employing distinct training strategies: unpaired and paired. All three compared models share the same model architecture and are initially given the same training dataset, $\mathcal{D}$ , comprising 300,000 images. The unpaired one-stage model is trained directly on $\mathcal{D}$ using the loss function described in Equation (1). Our model first performs standardization on $\mathcal{D}$ to get 300,000 pairs of (standardized reference image, target image) and then trains on it. The paired one-stage model is trained on paired data. However, these pairs are generated by applying our two-stage model to dataset $\mathcal{D}$ , resulting in 300,000 pairs of images. The key distinction here is that the reference image of paired one-stage is not standardized and only one stage inference is required. + +The results are presented in Table 2 and Figure 5. It is observed that the unpaired one-stage model exhibits a replication issue, where major contents of the reference image are duplicated in the output. This results in a high CLIP-I score of 89.62 and a Face Sim. score of 0.52 but a significantly low CLIP-T score of 16.08, indicating that the text prompts are not accurately represented in the output images. As illustrated in Figure 5(b), the background and action do not closely align with the prompt. This highlights the limitations associated with unpaired training. Further experimental analysis is provided in the supplementary materials, where we demonstrate that unpaired training may also suffer from low appearance consistency while exhibiting high text controllability. On the other hand, the remaining two methods, which utilize paired data, exhibit higher scores of 21.99 and 21.76 on the CLIP-T metric. The visualization outcomes further reveal that these methods can accurately respond to the text prompts. This demonstrates the super + +![](images/8f79772bd346af27a0c8711029ee60c00b2eab9e752842240fd9cb6fbddb59e4.jpg) +Figure 4. Comparison with other methods. Our method is capable of generating images with high text controllability and appearance consistency. + +# riority of paired training over unpaired training. + +Reducing our model from a two-stage to a paired one-stage configuration results in a significant decrease of 5.49 in the CLIP-I score. As evident from Figure 5(c) and (d), the lower body clothing of the character does not match the reference image in the paired one-stage model, whereas ours maintains consistency. This highlights that utilizing standardized references enhances the consistency between the reference image and the output. + +We demonstrate that standardization improves appearance consistency across serial images generated + +from different text prompts. This form of appearance consistency among multiple outputs is quantified by averaging all pairwise CLIP-I (AP-CLIP-I) scores. Our model achieves an AP-CLIP-I score of 83.17. In contrast, the score decreases to 77.74 for the paired one-stage model without standardization. The qualitative results presented in Figure 6 further underline the importance of standardization. As illustrated, when the reference image displays only a character's head, the absence of standardization can result in variations in the body's appearance across different generative processes. Standardization addresses this issue by + +![](images/8f8ffbbe0f0905499ee0489a061de7075cf4bf76b40499a4344960f53882300a.jpg) +(a) + +![](images/a804d269b1861bbe5b4b5bbab6e3f26d3e3e00ff4b416691409a29e5e2d8d294.jpg) +(b) + +![](images/521447eefbc329016dea720de0c458de9285b2052fa01b29ef995a41a3822859.jpg) +(c) + +![](images/92037e440886ac4e050119e9dc5b65f993b706ed5cb9630aa685d5f3733194a5.jpg) +(d) + +![](images/ed5773793d13f909d3c9ddfeb9adf599249b02b2f33a069058573bc4dc4da2de.jpg) +Figure 5. Visual comparison of different training strategies. Prompt: 1girl, cooking the meal, in the park. (a) Reference Image; (b) Unpaired one-stage; (c) Paired one-stage; (d) Ours. +Figure 6. Consistency between two prompts. The inconsistent part is highlighted by rectangle. + +
PSNR↑SSIM↑
baseline12.600.5820
baseline+FBDM12.790.5898
baseline+RPIM+FBDM12.840.5916
+ +pre-generating the body in the standardized reference image, thereby ensuring a consistent appearance across various text prompts. + +# 4.3.2. The Architecture of Standardization Model + +In this part, we verify the effectiveness of the proposed FBDM and RPIM modules. We first conduct experiments on our synthetic data, which is divided into a training set (80%) and a testing set (20%) based on the character ID. Table 3 presents the PSNR [13] and SSIM [29] for predictions made by the standardization model, both with and without the integration of the proposed modules. The results illustrate the superiority of the proposed modules. Qualitative results, depicted in Figure 7, demonstrate that the incorpo + +![](images/201007f41ea1d43b5fcb5db84499ea03ef0d0db48c52d95c178e955836f9ccfd.jpg) +Figure 7. Standardization results with and without proposed FBDM&RPIM. Leftmost is the input. + +Table 3. Ablation study of the standardization modules. + +
ModuleCLIP-I↑CLIP-T↑AP-CLIP-I↑
without84.9121.7182.86
with85.1621.8583.15
+ +Table 4. Two-stage generation results with and without proposed FBDM&RPIM in the standardization model. + +ration of the proposed modules enhances the standardization outcomes, yielding improved consistency. + +Next, we demonstrate the impact of the proposed modules on the final results of the two-stage generation. As shown in Table 4, the inclusion of these modules improves all evaluated metrics, emphasizing their importance. + +Finally, we discuss the rationality of overall architecture choice made for our standardization model. Given that the standardization task fundamentally pertains to human image animation, we compare our architecture against leading human image animation models using the benchmark dataset TikTok [28]. No additional training data was utilized to ensure a fair comparison. Finally, our method achieves an FVD of 149.95 and FID-VID of 14.75, outperforming state-of-the-art methods and justifying our architectural choices. More quantitative results are provided in the supplementary material. + +# 5. Conclusion + +Our work focuses on balancing text controllability with appearance consistency. By developing SerialGen which employs a serial generation method, we have empirically shown that it is possible to achieve high appearance consistency and adequate responsiveness to diverse text prompts. This work not only contributes valuable insights to the community but also showcases the practical implications of our approach in applications requiring high fidelity and personalized output, such as in comic story generation. + +# References + +[1] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H. Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. In SIGGRAPH Asia 2023 Conference Papers, pages 1–10, 2023. 2 +[2] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In Forty-first International Conference on Machine Learning, 2023. 3, 5 +[3] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. Advances in Neural Information Processing Systems, 36:30286–30305, 2023. 2 +[4] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. 3, 6 +[5] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In Proceedings of the IEEE Conference on European Conference on Computer Vision, 2020. 3 +[6] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3 +[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[8] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2 +[9] Rinon Gal, Or Lichter, Elad Richardson, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Lcmlookahead for encoder-based text-to-image personalization. In European Conference on Computer Vision, pages 322-340. Springer, 2024. 2, 3, 6 +[10] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. arXiv preprint arXiv:2404.16022, 2024. 2 +[11] Yue Han, Junwei Zhu, Keke He, Xu Chen, Yanhao Ge, Wei Li, Xiangtai Li, Jiangning Zhang, Chengjie Wang, and Yong Liu. Face-adapter for pre-trained diffusion models with fine-grained id and attribute control. In European Conference on Computer Vision, pages 20–36. Springer, 2025. +[12] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine + +yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024. 3 +[13] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. sSIM. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 8 +[14] 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. In International Conference on Learning Representations, 2022. 3 +[15] Li Hu. Animate anyone: Consistent and controllable imaged-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3, 5 +[16] Jeongho Kim, Min-Jung Kim, Junsoo Lee, and Jaegul Choo. Tcan: Animating human images with temporally consistent pose guidance using diffusion models. arXiv preprint arXiv:2407.09012, 2024. 3, 5 +[17] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2 +[18] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 2, 3 +[19] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 5 +[20] 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. 5 +[21] 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 +[22] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. +[23] 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. 2, 3 +[24] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2 + +[25] 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 +[26] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 3, 5 +[27] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2, 3 +[28] Tan Wang, Linjie Li, Kevin Lin, Yuanhao Zhai, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for realistic human dance generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9326-9336, 2024. 3, 8, 5 +[29] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 8 +[30] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. International Journal of Computer Vision, pages 1-20, 2024. 2, 3, 6 +[31] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 3, 5 +[32] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 4 +[33] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3, 5, 6, 4 +[34] Yong Zhong, Min Zhao, Zebin You, Xiaofeng Yu, Changwang Zhang, and Chongxuan Li. Posecrafter: One-shot personalized video synthesis following flexible pose control. In European Conference on Computer Vision, pages 243-260. Springer, 2025. 3 +[35] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. arXiv preprint arXiv:2405.01434, 2024. 3 +[36] Zhengguang Zhou, Jing Li, Huaxia Li, Nemo Chen, and Xu Tang. Storymaker: Towards holistic consistent characters in text-to-image generation. arXiv preprint arXiv:2409.12576, 2024. 2, 3, 6 +[37] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: + +Controllable and consistent human image animation with 3d parametric guidance. In European Conference on Computer Vision (ECCV), 2024. 3, 5 \ No newline at end of file diff --git a/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/images.zip b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c718b46bf438f6c4242d388dbd38ccd5391d9cd7 --- /dev/null +++ b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48c6eefd8e0ca9ced4c2113473309dcf9429ffb28ab2692690918b6b9fae00f4 +size 874235 diff --git a/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/layout.json b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8bbcd69c7f88141b3376cbf3708bb73babe0acca --- /dev/null +++ b/CVPR/2025/SerialGen_ Personalized Image Generation by First Standardization Then Personalization/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0669e6b26923d8f65fd579730a3d5136d0b52512f95e7d6dd8e9c16ad02cd57 +size 316533 diff --git a/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/a7b4ece0-fc55-49e5-bbf2-188de74cecca_content_list.json b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/a7b4ece0-fc55-49e5-bbf2-188de74cecca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6a8a278d2148eaaeacb41d20e4913e1e0d129b19 --- /dev/null +++ b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/a7b4ece0-fc55-49e5-bbf2-188de74cecca_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ff37680fc20ac8df40c7afd69e372b0ff0cfb5951667e973692d1f58314c979 +size 80556 diff --git a/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/a7b4ece0-fc55-49e5-bbf2-188de74cecca_model.json b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/a7b4ece0-fc55-49e5-bbf2-188de74cecca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..54fea803a6a6a0af8f166d2b5c4c9882abbf74d9 --- /dev/null +++ b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/a7b4ece0-fc55-49e5-bbf2-188de74cecca_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1da1874ee6b9f896d26566b7c166788091e2cd05b9ea221ff8c512ab4e52a82 +size 96081 diff --git a/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/a7b4ece0-fc55-49e5-bbf2-188de74cecca_origin.pdf b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/a7b4ece0-fc55-49e5-bbf2-188de74cecca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..49a17d9a0cd728708c5fa4914317ba0f8b0fb629 --- /dev/null +++ b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/a7b4ece0-fc55-49e5-bbf2-188de74cecca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b65a41a6a0b7ec8bffa79ac405b0f9f9365c3d78a7a592169810770ae0ea7c6c +size 742328 diff --git a/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/full.md b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c98e3255d1bee6f38a96caac2739f080868b0860 --- /dev/null +++ b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/full.md @@ -0,0 +1,259 @@ +# SeriesBench: A Benchmark for Narrative-Driven Drama Series Understanding + +Chenkai Zhang $^{1,3*}$ Yiming Lei $^{1,3*}$ Zeming Liu $^{2\dagger}$ Haitao Leng $^{4}$ ShaoGuo Liu $^{4}$ Tingting Gao $^{4}$ Qingjie Liu $^{1,3\dagger}$ Yunhong Wang $^{1}$ + +$^{1}$ State Key Laboratory of Virtual Reality Technology and Systems, Beihang University $^{2}$ School of Computer Science and Engineering, Beihang University, Beijing, China $^{3}$ Hangzhou Innovation Institute, Beihang University, Hangzhou China $^{4}$ Kuaishou Technology + +$^{*}$ Co-first authors: {zhangchenkai, ymlei}@buaa.edu.cn +†Corresponding authors: {zmliu, qingjie.liu}@buaa.edu.cn + +# Abstract + +With the rapid development of Multi-modal Large Language Models (MLLMs), an increasing number of benchmarks have been established to evaluate the video understanding capabilities of these models. However, these benchmarks focus on standalone videos and only assess "visual elements" like human actions and object states. In reality, contemporary videos often encompass complex and continuous narratives, typically presented as a series. To address this challenge, we propose SeriesBench, a benchmark consisting of 105 carefully curated narrative-driven series, covering 28 specialized tasks that require deep narrative understanding to solve. Specifically, we first select a diverse set of drama series spanning various genres. Then, we introduce a novel long-span narrative annotation method, combined with a full-information transformation approach to convert manual annotations into diverse task formats. To further enhance the model's capacity for detailed analysis of plot structures and character relationships within series, we propose a novel narrative reasoning framework, PC-DCoT. Extensive results on SeriesBench indicate that existing MLLMs still face significant challenges in understanding narrative-driven series, while PC-DCoT enables these MLLMs to achieve performance improvements. Overall, our SeriesBench and PC-DCoT highlight the critical necessity of advancing model capabilities for understanding narrative-driven series, guiding future MLLMs development. SeriesBench is publicly available at https://github.com/zackhxn/SeriesBench-CVPR2025. + +# 1. Introduction + +In recent years, multimodal large language models (MLLMs) have made remarkable strides in the field of visual understanding [4, 9, 16, 31, 32, 37, 38, 47]. Building + +0.20 The director agrees to install air conditioning for Yingyan. + +![](images/218b222cdf6e4affc1f99a59052f0e4fcb4eb5fac4bcde07da17f8f5b8a7580b.jpg) +Figure 1. An example from SeriesBench. The task involves multiple events and characters spanning a long time period in a video. + +1:12 The director forbids Tuojiangkai from using the air conditioning. 2:24 The director allows Meiyang to turn on the air conditioning. + +Why does the director repeatedly enter the classroom? + +A. Announces the removal of Meiyang's air conditioning rights. +B. Announces defeat due to the plan executed by Simayi. +C. Announces that Yingyan's air conditioning is not working. + +upon these advancements, a range of benchmarks [7, 20, 23, 35, 42, 44] have been established to evaluate the visual-language capabilities of these models, primarily through question-answering (QA) tasks based on static images. Although these benchmarks effectively assess static visual comprehension, real-world visual understanding must account for complex and dynamic changes over time. As a result, an increasing emphasis has emerged on evaluating MLLMs' ability to comprehend dynamic video content and perform temporal reasoning. Several studies [5, 8, 18, 21, 25, 26] have introduced video-based benchmarks to assess MLLMs' temporal analysis capabilities through questions like "Is the man walking towards the right-side exit?" + +However, despite the growing attention to temporal issues in video analysis, current evaluations primarily focus on standalone videos, often neglect the narrative structure and character development across multi-video series. Furthermore, these evaluations often focus solely on visual ele + +ments, overlooking the broader multimodal nature of modern videos. In reality, most modern videos are not merely sequences of static images but rather complex multimodal compositions [10, 27], integrating script, audio, and special effects through editing techniques, as found in contemporary films, television series, and drama series. This oversight in current evaluations significantly limits model performance, as MLLMs frequently struggle to understand the intricate narrative structures and character interactions essential for comprehensive video understanding, especially in series-based content. Such limitations could impair the application of MLLMs in fields like series recommendation, interactive media, and autonomous video summarization, potentially leading to biased conclusions or skew public understanding of events, particularly in public crises or media representations. + +To this end, we propose SeriesBench, the first-ever comprehensive benchmark designed to evaluate MLLMs' understanding of narrative-driven series. As exemplified in Fig. 1, we meticulously curate 105 series, comprising a total of 1,072 videos, which span a wide variety of themes, including daily life, anime, time-travel, historical drama, fantasy, and other narrative-driven genres. Alongside the video collection, we also compiled subtitles for each video, as well as detailed information on the themes and character backgrounds specific to each series. To enable a more comprehensive evaluation of large models that aligns with the diverse modalities found in modern videos [10, 27], SeriesBench is organized into five primary task dimensions corresponding to the core compositional elements of modern videos: [Visuals], [Script], [Audio], [Augmentation], and [Comprehension], collectively covering 28 fine-grained sub-tasks. To ensure the quality of SeriesBench, we engage over 30 professional annotators and propose a novel long-span narrative annotation method, which requires annotators to label narrative events and characters across extended temporal spans. After the annotation process, we apply a full-information transformation technique to convert the annotated data into a diverse array of question types, including multiple-choice, judgment-based, and open-ended formats. + +Inspired by human binge-watching behavior, we propose Plot & Character Dual Chain of Thought (PC-DCoT) to enhance MLLMs' understanding of narrative-driven series by constructing Plot Event Chains and Character Temporal Chains. We conduct extensive experiments and analyses on 10 leading Video-MLLMs and apply the PC-DCoT framework to 4 top-performing models. Results show that: 1) Video-MLLMs struggle significantly with understanding narrative structure across series; 2) A notable performance gap persists between open-source and commercial MLLMs; 3) Even MLLMs with limited audio processing capabilities show partial comprehension of audio elements. Our evalu + +ations demonstrate that utilizing the PC-DCoT framework can improve MLLMs' performance in narrative-driven series by over $10\%$ . Despite these advancements, current MLLMs still fall short of human-level understanding of narrative-driven series, highlighting the significant potential for further research to advance MLLM series understanding. We summarize our contributions as follows: + +(1) Novel and Challenging Tasks: We introduce a new set of challenging tasks focused on understanding video narratives within a series, involving the analysis of events and character relationships. +(2) New Video Understanding Benchmark: We are the first to establish a video understanding benchmark specifically designed for narrative-driven series, addressing a critical gap in current evaluation benchmarks. +(3) Innovative Method and Comprehensive Evaluation: We propose the PC-DCoT framework that improves MLLMs' capacity for understanding narrative-driven series by reconstructing inherent structure. Extensive experiments on state-of-the-art baselines and our method provide valuable insights for further research in narrative comprehension of series-based videos. + +# 2. Related work + +Benchmark. Traditional multimodal visual benchmarks [1, 7, 42, 44] primarily focused on static image understanding tasks to evaluate Image LLMs. As research shifted towards video understanding, early explorations concentrated primarily on action recognition and event detection [13, 41], lacking the capability to comprehensively evaluate MLLMs. Most current video benchmarks, including Video-Bench [11], SEEDBench [14], and VLM-Eval [19], addressing more complex tasks such as video-specific understanding, knowledge-based QA, and decision-making. For long temporal sequences, benchmarks like EgoSchema [28] and Video-MME [8] emphasize fine-grained temporal reasoning, while TempCompass [25], MVBench [18], and AutoEval-Video [3] assess models with tasks demanding temporal understanding. Additionally, VideoVista [21] and MLVU [48] focus on fine-grained video understanding, with MLVU testing videos from 3 minutes to 2 hours. + +Video-MLLM. The rapid development of large language models (LLMs) has accelerated multimodal research, enabling models [17, 22] to combine visual and textual information for a variety of tasks. With the extension of context lengths, MLLMs have been applied to video comprehension. Video-MLLMs, such as VideoChat [17] and Video-LLaVA [22], typically include a visual encoder, a projector for modal alignment, and a LLM for reasoning, significantly enhancing multimodal comprehension in complex video scenarios. Modern Video-MLLMs, such as InternVL2 [36], Qwen2-VL [38], and MiniCPM-V-2.6 [43], leverage more powerful LLMs and dynamic visual encoders + +
DimensionSub-dimensionFine-grainedExample
VisualsFiguresActionsWhat was Zhao Dezhu's action when facing the room check? +(A) Hiding behind the door. (B) Acting on the floor. (C) Jumping out of the window.
InteractionsWas Allen asked by Xuefeng to help him take a photo? (A) True. (B) False.
ScenesScene TransitionsWhat scenes appear in the video? GT: Outside the office, parking lot.
Spatiotemporal ShiftsWhere did they go if the princess's dress changed from red to white? +(A) Demon Realm. (B) Human Realm. (C) Magic Realm. (D) Fairy Realm.
ObjectsPresenceIs the building behind the fairy's owner called Dacheng Hall? (A) True. (B) False.
InteractionWhat item did Manager Min hand to Xiaoqian to help take a photo? +(A) Tablet. (B) Folder. (C) Camera. (D) Phone.
ScriptBackgroundWorld-BuildingWhat type of life is described in the video? +(A) Workplace life. (B) School life. (C) Martial arts life. (D) Rural life.
Time and LocationAccording to the video, these students slept for 8 hours tonight. (A) True. (B) False.
PlotPlot DevelopmentWhat was Yingyan's wish after passing the exam, and what did the director decide? +GT: Yingyan wanted air conditioning or an eye cure and the Director agreed.
Foreshadowing and PayoffWhat happened after the dorm manager praised Sima Yi's well-folded quilt? +GT: Dorm manager was scared by spider.
Twists and ConflictsWhat event marks the turning point of the plot in the story? +(A) Zhao Dezhu's appearance. (B) Director's lecture. (C) Dorm manager's intrusion.
Climaxes and Build-upsThe chairman forgave Manager Min after expressing disappointment (A) True. (B) False.
Suspense and ContinuityTujiang Kai shot at Yingyan, but Yingyan killed him instead. (A) True. (B) False.
Emotional DynamicsWhat unexpected funny moments appear in the video? +GT: Everyone thought Yingyan was out of breath, but he could still blow music.
CharactersReferenceWho is the person wearing a pink T-shirt and glasses? +(A) Xiaojin. (B) Manager Huang. (C) Brother Hong. (D) David.
MotivationsWhy did Director establish the rules for air conditioning management? +GT: Because he did not want students to use the air conditioning easily.
AudioDialogueDialogue AttributionThe chairman said "My family has a daughter". (A) True. (B) False.
Pronoun ReferencesWho does "she" refer to in the line "She stole the company's confidential files"? +(A) Zhiyang. (B) Yue Xin. (C) An Xin. (D) General Manager.
Tone and EmotionWhat emotions did Teacher Du display towards the top student? +(A) Relief and anger. (B) Worry and anger. (C) Anger and frustration. (D) Relief and worry
MusicAtmosphereWhat is the main purpose of the background music in the video? +(A) To create tension. (B) To evoke anxiety. (C) To provide calmness. (D) To convey intensity.
Sound EffectsImpactHow do sound effects in the video express the characters' inner feelings? +GT: They convey emotions such as anxiety and sadness through tense music and sighs.
AugmentationSubtitlesRecognitionDoes the video depict the life stories of "Shanghai boys"? (T) True. (F) False.
LabelsPurposeThe prohibition label indicates no electronic devices are allowed. (T) True. (F) False.
VFXEffectivenessWhich character does Zhao Dezhu portray through the use of special effects in the play? +(A) Iron Man. (B) Spider-Man. (C) Hulk. (D) Thor.
ComprehensionEngagementFuture PredictionsHow does the relationship between Zhiyang and Yuexin develop in the subsequent plot? +(A) They let go of their grievances. (B) They go their separate ways. (C) They argue again.
Current InterpretationIs the little fairy struck twice, by the princess and the master? (T) True. (F) False.
EmpathyCharacter ResonanceWhich character shows the greatest passion for studying in the video? +GT: Sima Yi demonstrates the greatest passion for learning, showing diligence and curiosity.
+ +Table 1. Task Diversity in SeriesBench. The table presents different hierarchical tasks, including examples. + +to improve real-time interactions and performance. VideoLLaMA2 [4] uses spatial-temporal convolutions and an audio projector, expanding the integration of audio with other modalities. However, despite these advances, most video-MLLMs still rely on sparse frame extraction within standalone videos. + +# 3. SeriesBench + +To construct SeriesBench, we first define task dimensions based on the core elements of modern videos, as outlined in Tab. 1. Then, we present our meticulously designed annotation method. Finally, we compare SeriesBench with other benchmarks, as shown in Tab. 2. + +# 3.1. Task Dimension Definitions + +Existing benchmarks for evaluating MLLMs are largely confined to purely visual elements and often rely on stan- + +dalone video clips, limiting their capacity to capture complex and cohesive narratives. For instance, these benchmarks typically assess continuous image frames that depict human actions or object states, such as asking, "What object is this person passing?" However, they rarely delve into deeper narrative questions like "Why is this person passing this object?" and fail to explore broader narrative reasoning across connected videos. To address this challenge, SeriesBench offers a more comprehensive evaluation, with a strong emphasis on narrative understanding by utilizing videos from drama series. To align with the diverse elements inherent in modern videos, which include frames, scripts, audio, and post-production components, we have meticulously designed five major task categories for SeriesBench: [Visuels], analyzing frames and figure appearances critical to the narrative; [Script], focusing on the design and development of plot and character dynamics; + +
BenchmarkSeriesDimensionSub TokensAnno.VideoMulti Tasks
VSSCAUAGCO#ClipsLen.
MSRVTT-QA [41]X✓XXXXX&2,99015.2X
MSVD-QA [41]X✓XXXXX&5049.8X
TGIF-QA [12]X✓XXXXX& &9,5753.0X
ActivityNet-QA [45]X✓XXXXX&800111.4X
STAR [39]X✓XXXXX&91411.9X
NExT-QA [40]X✓XXXXX&1,00039.5X
MVBench [18]X✓XXXXX&3,64116.0X
Video-Bench [30]X✓XXXXX& &5,91756.0X
EgoSchema [29]X✓XXXXX& &5,063180.0X
AutoEval-Video [3]X✓XXXXX&32714.6X
TempCompass [24]X✓XXXXX& &41011.4X
Video-MME-S [8]X✓XXXX198.6&30080.7X
VideoVista [21]X✓XXXXX&3402131.0X
TVBench [5]X✓XXXXX&2525446.1X
MMBench-Video [6]X✓XXXXX&609165.0X
SeriesBench (Ours)✓✓✓✓✓✓250.2&107279.2✓
+ +Table 2. Comparison between SeriesBench and other existing benchmarks. Compared to other benchmarks, SeriesBench extends video understanding to narrative-driven series, while other benchmarks are limited to the visual comprehension of standalone videos. In the table, Series indicates whether the dataset contains episodic or series-based videos. VS represents visuals, SC represents script, AU represents audio, AG represents augmentation, and CO represents comprehension. Sub Tokens refers to the average number of subtitle tokens per video. denotes manual annotation, and represents automatic annotation. #Clips indicates the number of video clips, Len. is the average video length. Multi Tasks reflects whether the benchmark includes a variety of task types. + +[Audio], interpreting dialogue, sound effects, and music in context; [Augmentation], examining special effects and labels affecting narrative; and [Comprehension], integrating all elements to assess overall narrative comprehension. All tasks are aimed at enhancing the understanding of narrative-driven series, with examples provided in Tab. 1. + +# 3.2. Task Annotation Design + +The construction of SeriesBench is a meticulous process designed to capture the complexity of a narrative-driven series. We introduce a novel long-span narrative annotation method, requiring annotators to label events and characters across extended temporal spans. To further support task generation, we apply a full-information transformation method to convert annotations into diverse questions, enhancing the benchmark's quality and challenge. + +# 3.2.1 Long-span Narrative Annotation + +After collecting the raw series, we established detailed annotation guidelines and trained professional annotators to carry out the video annotations. We also implemented a trial annotation phase, through which we selected the 32 annotators with the highest annotation quality. Firstly, annotators were required to fully understand the video narrative before annotating specific key segments, including the events occurring and the actions of key characters. Then, they were tasked with summarizing these annotations in a concise declarative statement sentence that connects all the annotation content. As a result, this statement is related to + +long-term events and characters within the video and often involves narrative reasoning. After this, annotators needed to select a task dimension most relevant to this statement. Finally, all annotated content underwent a rigorous quality control process conducted by dedicated reviewers. Each annotation was meticulously examined for accuracy, coherence, and consistency. If any annotation failed to meet the established quality standards, it was sent back for correction and refinement. Ultimately, a random sampling of 500 annotations showed that $96\%$ of the annotations met the required standards, demonstrating a high level of quality. This annotation workflow ensures that each statement covers information spanning a long duration in the video and incorporates multi-hop narrative reasoning content, thereby guaranteeing both the quality of the annotations and the complexity required for subsequent question generation. + +# 3.2.2 Full-Information Transformation + +Based on the annotated data, we proposed a full-information transformation method to generate tasks. Due to each video annotation containing manual declarative statements, subtitles, themes, and character introductions gathered during the video collection phase, we can utilize the video's full information to transform the manual annotations effectively. Leveraging this data, we use GPT-4o to generate questions of various task types, including true-or-false questions, multiple-choice questions, and open-ended questions. This method ensures that annotated declarative sentences are seamlessly transformed into question stems and correct answers, with related video information crafted + +![](images/0b19ebafb4d1be246d2f31190525437844f5e82806f1add812fffe0b185c826c.jpg) +Figure 2. Illustration of PC-DCoT. The process begins by extracting relevant events and characters from the input question and the original video. Subsequently, a Plot Event Chain and a Character Temporal Chain are constructed independently. Finally, these chains are merged, enabling a reasoning process that generates an answer to the posed question based on the integrated dual-chain framework. + +![](images/358b941dd17540f559825580d8423f40c7a8c802cd7d63edd26e0c752c81b443.jpg) + +![](images/3198da010497b1abd34512d367647252580b8b31d1f1fdd046e750e4e6f18ecd.jpg) + +as distractors. Finally, A total of 29,196 tasks were generated. Specific examples can be found in Tab. 1. + +# 3.3. Comparison with Previous Benchmarks + +Previous benchmarks primarily focus on standalone videos and evaluate only the visual elements within them. In comparison, as shown in Tab. 2, videos in SeriesBench are organized as interconnected series collections, allowing for a more comprehensive evaluation of narrative-driven content. To achieve this, we crawled the top 11 trending series genres from Kuaishou1, and through data cleaning, we finalized a selection of 105 series with the highest engagement ratios. Furthermore, SeriesBench expands task dimensions beyond visual elements to include narrative, audio, augmentation, and comprehension, promoting a more comprehensive understanding of the videos. As these videos encompass complete narratives, they contain a higher volume of subtitles. To further enhance the quality, we enlisted over 30 professional annotators to meticulously label the videos, supported by a rigorous quality review mechanism. In addition to various new features, SeriesBench comprises 1,072 narrative-driven videos, and diverse task types, enabling comprehensive evaluation of video understanding. + +# 4. Plot & Character Dual Chain of Thought + +Current MLLMs [4, 32, 36, 38] are capable of describing character actions and object states in a video but struggle to comprehend narrative-related events, character identities, and inter-character relationships. However, humans naturally focus on the distinct identities and roles of characters and the complex events unfolding between them when watching narrative-driven videos, particularly in drama se + +ries. Drawing inspiration from this human approach, we propose the PC-DCoT framework. This framework begins by extracting key events and characters that need to be tracked based on the given video and questions, then constructs separate analytical chains for events and characters, which are subsequently integrated into a dual-chain structure to facilitate comprehensive answer reasoning. Fig. 2 illustrates the overall framework of PC-DCoT. + +# 4.1. Event and Character Extraction. + +For narrative-driven videos, the plot and characters are critical components that drive the storyline. To effectively analyze this complex content, the method begins by leveraging the MLLM to receive raw video frame input, along with relevant questions. Then, the MLLM extracts and identifies significant events pertinent to the questions and key characters that need to be tracked. This comprehensive process establishes primary targets for tracking, thereby laying the groundwork for a deeper and more comprehensive understanding of the narrative. + +# 4.2. Plot and Character Chains Construction. + +In narrative-driven videos, character appearances are often discontinuous, while events typically progress through interconnected and cohesive segments. To align input videos with task-relevant events and characters, we trained a video clip model [34] to identify and isolate key scenes. By taking the raw video input along with extracted events relevant to the questions, the model searches for video frames that correspond to each identified event. It then aggregates these frames with adjacent frames to form cohesive sets of independent event sequences. Subsequently, the MLLM is employed to generate detailed descriptions of each continuous event, constructing the Plot Event Chain. Simultaneously, + +the model utilizes key character portraits to identify and retrieve all corresponding character frames from the video. By recognizing and tracking the appearances of these characters, even when their presence is disjointed and spread across different segments, the MLLM generates detailed behavioral descriptions for each individual. This approach ensures consistent tracking of specific characters throughout the entire video, resulting in the construction of multiple Character Temporal Chains that capture their behaviors, interactions, and roles over the video timeline. + +# 4.3. Dual-Chain Synthesis and Reasoning. + +Since the Plot Event Chain and Character Temporal Chain both contain specific video frames, each event or character is associated with a precise time interval. This precise temporal annotation allows for direct and accurate alignment of events and character appearances based on their temporal axes, making it possible to determine exactly which characters are involved within the timeframe of each event. Subsequently, the MLLM synthesizes the interactions between events and characters, merging their occurrences and behaviors within the aligned timeframes. This integration yields a more cohesive and detailed narrative representation, capturing the nuances of character dynamics, interactions, and event progression with enhanced clarity. + +# 5. Experiments + +# 5.1. Settings + +We partitioned SeriesBench into training, validation, and test sets using an 8:1:1 ratio through stratified random sampling from the original dataset, ensuring no overlap between different sets. For multiple-choice and judgment tasks, we evaluated the correctness of discrete predictions using accuracy. For open-ended tasks, we measured lexical overlap and semantic relevance using BLEU-2 [33] and ME-TEOR [2], and BERTScore F1 [46] for semantic similarity. + +We evaluate a comprehensive set of recent Video-MLLMs following their official inference and frame extraction configurations similar to Video-MME-S [8] and Video-Vista [21]: (a) Open sourced Video-MLLMs including InternVL2 [36], LLaVA-OneVision [15], Qwen2-VL [38], LLaVA-Video [47], MiniCPM-V 2.6 [43], Aria [16], and VITA [9]. (b)Open sourced Video-Audio-MLLMs, i.e., VideoLLaMA2.1-AV [4]. (c) Proprietary commercial models such as GPT-4o-mini [31] and GPT-4o [32]. + +# 5.2. Main Results on SeriesBench + +We include random choice and frequent guessing as naive baselines, alongside human performance as the upper bound. The overall evaluation results averaged across subtasks, are presented in Tab. 3. We further apply PC-DCoT + +to the models that performed best, specifically InternVL2, Qwen2-VL, MiniCPM-V 2.6, and GPT-4. + +Strong Video-LLMs on existing benchmarks struggle on SeriesBench. While state-of-the-art(SOTA) MLLMs (e.g., GPT-4o and Qwen2-VL) have achieved around $80\%$ accuracy on benchmarks(e.g., Video-MME-S and VideoVista), their performance on SeriesBench remains limited, particularly for tasks that require fine-grained visual analysis and deep comprehension, which demands enhanced content understanding and reasoning. We attribute the models' inefficiencies to disrupted visual continuity caused by discrete frame sampling and an overemphasis on image-level captioning rather than on narrative storyline in episodes. + +Performance gap between open-source and commercial MLLMs remains evident. We observe that GPT-4o achieves the highest performance on SeriesBench, particularly in [Script] tasks involving background, plot, and character comprehension, outperforming the runner-up Qwen2-VL by $5.3\%$ . However, without subtitles, open-source video models like InternVL2, Qwen2-VL, and MiniCPM-V 2.6 can surpass GPT-4o in [Visual] and [Comprehensive] tasks (see Appendix). This indicates that open-source multimodal models with optimized visual processing modules are capable of approaching the visual task performance of commercial models, although a significant gap remains in text comprehension and reading capabilities relative to GPT-4o. + +MLLMs with LIMITED audio processing capabilities can comprehend audio components. While most video MLLMs lack audio input support, our research reveals that audio provides limited benefits in video-audio MLLMs. For instance, VideoLLaMA2.1-AV showed decreased performance when audio is included. In contrast, textual inputs like subtitles help models better capture semantic and contextual cues, significantly improving performance(see Appendix). Furthermore, inefficiencies in complex multimodal reasoning tasks highlight the challenges of aligning audio with other modalities and underscore the critical role of high-density textual information in MLLMs. + +PC-DCoT enhances the performance of both open-source and commercial MLLMs. PC-DCoT uses plot-event and character-temporal chains to extract fine-grained video segments, enhancing the model's ability to reconstruct complex storylines sequentially. Evaluation results show that PC-DCoT consistently achieves SOTA performance across all tasks on the narrative-driven SeriesBench, highlighting its adaptability and scalability across various MLLM architectures. However, a notable gap remains between models and human comprehension of intricate narrative causal reasoning and multi-character plot analysis. These findings indicate substantial potential for improving models in capturing detailed event dependencies and multi-character interactions within narrative contexts. + +
ModelSizeFramesVSSCAUAGCOOverallBL-2METF1BERT
Heuristic baselines
Random Choice--39.338.235.536.538.837.7---
Frequent Guess--43.646.043.141.150.644.4---
Human--98.294.494.697.292.695.8---
Open-source Video MLLMs
InternVL2[36]7B3252.056.857.778.657.759.28.4321.2268.40
LLaVA-OneVision[15]7B3251.554.056.270.659.656.97.3020.2067.82
LLaVA-Video[47]7B6454.956.857.471.251.958.37.2719.0667.95
Qwen2-VL[38]7B6455.757.558.675.359.660.311.4127.9770.71
MiniCPM-V 2.6[43]8B6453.357.357.776.655.159.18.5730.0668.66
Aria[16]8×3.5B12852.456.651.875.650.056.810.0026.0870.35
VITA[9]8×7B3246.946.053.368.948.151.89.3227.0868.93
Open-source Video-Audio MLLMs
VideoLLaMA2.1-AV[4] +audio7B3247.452.253.366.649.453.17.6123.3467.53
45.352.251.366.246.251.77.2422.7167.09
Commercial MLLMs
GPT-4o-mini[31]N/A5046.742.747.970.944.949.89.9929.3068.76
GPT-4o[32]N/A5055.862.860.679.959.662.89.6125.1068.94
MLLMs with PC-DCoT
InternVL2†7B-76.571.467.381.166.773.3 (↑14.1)8.2620.7167.85
Qwen2-VL†7B-75.273.569.780.967.473.9 (↑13.6)8.9323.4168.08
MiniCPM-V 2.6†8B-76.368.266.582.560.072.0 (↑12.9)14.1633.6472.98
GPT-4o†N/A-78.676.173.882.161.776.2 (↑13.4)12.6128.2972.75
+ +# 5.3. Analysis + +To comprehensively examine the factors influencing the series understanding of Video-MLLMs in narrative-driven tasks, we design a set of research questions (RQs) for empirical analysis, including: (1) investigating the consistency of MLLMs in understanding multi-episode series (RQ1); (2) analyzing the impact of different input modalities on series understanding (RQ2); and (3) conducting an ablation study of Dual-Chains in PC-DCoT (RQ3). + +RQ1: Can MLLMs consistently understand multi-episode series? We investigate the performance of different models on SeriesBench across varying episode contexts and report results for a subset of tasks related to multi-episode series, including [Plot], [Engagement], and [Empathy]. These tasks require the comprehension of interleaved modalities and the ability to extract relevant information from multiple videos. The increased token count in multi-episode inputs also demands models capable of processing longer context. Tab. 4 summarizes the performance, with detailed results available in the Appendix. + +Experimental results show improved performance when earlier episodes are used as input. We attribute this improvement to the richer contextual information in earlier episodes, which enhances comprehension of plot-related tasks. In contrast, later episodes, which are not available in the current context, may introduce irrelevant details and hin + +Table 3. Performance of MLLMs on SeriesBench. Size means the LLM size. Judgement and multichoice metrics Accuracy and open-ended metrics BLEU-2(BL-2), METEOR (MET), and BERTScore F1 (F1BERT) are reported in percentage (%). “-” indicates that results are not feasible with open-ended metrics in heuristic baselines. †: MLLMs using the PC-DCoT framework during inference. The best, second-best, and third-best results are marked purple, orange, and gray, respectively. + +
ModelEpisodesPlotEng.Emp.Total
InternVL2-63.167.556.864.0
Prev260.969.254.163.1
Prev164.968.054.165.2
Next161.266.967.663.5
Next261.267.556.862.9
Qwen2-VL-61.570.456.864.0
Prev260.970.470.364.6
Prev162.570.464.965.2
Next158.268.054.161.0
Next260.369.256.862.9
MiniCPM-V 2.6-64.666.945.964.0
Prev257.564.554.159.5
Prev160.365.140.560.5
Next156.655.051.455.7
Next256.651.548.654.4
+ +Table 4. Performance of Top-Performing MLLMs on Multi-Episode Series Tasks. Plot, Engagement (Eng.), and Empathy (Emp.), and Total scores. $\text{Prev}_i$ and $\text{Next}_i$ represent the inclusion of $i$ episodes before and after the target episode, respectively. Evaluation for GPT-40 is not feasible due to API limitations. + +der comprehension. Notably, Qwen2-VL is the only model that shows improved performance with earlier episodes, demonstrating its ability to perform cross-context reasoning, while InternVL2 and MiniCPM-V 2.6 experience declines under the same conditions. This discrepancy highlights the importance of series-based narrative tasks over individual video understanding tasks for evaluating MLLMs' reasoning capabilities. + +
ModelInputsVSSCAUAGCOOverall
InternVL2Q only42.543.245.743.543.643.6
Q, F50.2 (+7.7)51.3 (+8.1)52.6 (+6.9)73.6 (+30.1)55.1 (+11.5)55.2 (+11.6)
Q, S51.1 (+8.6)55.0 (+11.8)57.9 (+12.2)74.9 (+31.4)55.1 (+11.5)57.7 (+14.1)
Q, F, S52.0 (+9.5)56.8 (+13.6)57.7 (+12.0)78.6 (+35.1)57.7 (+14.1)59.2 (+15.6)
Q, F, S, TC55.3 (+12.8)61.4 (+18.2)59.9 (+14.2)76.9 (+33.4)60.3 (+16.7)61.7 (+18.1)
Qwen2-VLQ only44.737.442.840.535.341.1
Q, F52.9 (+8.2)55.0 (+17.6)56.4 (+13.6)72.2 (+31.7)57.7 (+22.4)57.7 (+16.6)
Q, S52.2 (+7.5)54.0 (+16.6)59.1 (+16.3)75.9 (+35.4)53.8 (+18.5)58.1 (+17.0)
Q, F, S55.7 (+11.0)57.5 (+20.1)58.6 (+15.8)75.3 (+34.8)59.6 (+24.3)60.3 (+19.2)
Q, F, S, TC53.5 (+8.8)61.9 (+24.5)58.9 (+16.1)74.2 (+33.7)63.5 (+28.2)60.9 (+19.8)
MiniCPM-V 2.6Q only46.540.947.039.540.443.6
Q, F48.4 (+1.9)53.1 (+12.2)54.7 (+7.7)73.2 (+33.7)56.4 (+16.0)55.6 (+12.0)
Q, S48.7 (+2.2)53.4 (+12.5)55.5 (+8.5)72.9 (+33.4)51.3 (+10.9)55.4 (+11.8)
Q, F, S53.3 (+6.8)57.3 (+16.4)57.7 (+10.7)76.6 (+37.1)55.1 (+14.7)59.1 (+15.5)
Q, F, S, TC50.4 (+3.9)55.9 (+15.0)55.7 (+8.7)74.9 (+35.4)56.4 (+16.0)57.3 (+13.7)
GPT-4oQ only46.444.346.038.141.744.1
Q, F52.4 (+6.0)56.4 (+12.1)56.0 (+10.0)72.2 (+34.1)47.4 (+5.7)56.9 (+12.8)
Q, S55.3 (+8.9)61.2 (+16.9)61.1 (+15.1)75.3 (+37.2)57.1 (+15.4)61.3 (+17.2)
Q, F, S55.8 (+9.4)62.8 (+18.5)60.6 (+14.6)79.9 (+41.8)59.6 (+17.9)62.8 (+18.7)
Q, F, S, TC57.3 (+10.9)65.1 (+20.8)65.0 (+19.0)76.9 (+38.8)55.1 (+13.4)63.8 (+19.7)
+ +RQ2: How do different modalities affect series understanding? Most benchmarks use only video frames and questions as inputs, often neglecting other crucial information. Similar to Video-MME [8] and VideoVista [21], we generated subtitle information and obtained thematic and character descriptions during video collection, ensuring synchronization with visual content. We explore how additional descriptive elements influence series understanding by varying input modalities. Experimental results are presented in Tab. 5. We observed: + +(1) Video frames or subtitles significantly improve accuracy and consistency. Supplementing text-only questions with frames (Q, F) or subtitles (Q, S) leads to performance gains across all models, showing the validity of these elements. (2) Textual components outweigh visual information. Combining text questions with subtitles (Q, S) yields higher accuracy than with frames (Q, F), which is likely due to the narrative nature of videos, where subtitles function as a SCREENPLAY that aids storyline understanding. (3) MLLMs process video information similarly to human cognition. As humans benefit from subtitles in films, models improve comprehension of thematic and character information. Integrating thematic and character data (Q, F, S, TC) with frames and subtitles (Q, F, S) significantly boosts performance, achieving optimal results across all modalities. + +RQ3: How do dual-chains impact PC-DCoT? We conduct a comprehensive analysis of the Plot-Event Chain and Character-Temporal Chain to assess their effectiveness in PC-DCoT during inference. Ablation results are presented in Tab. 6. Both the dual-chains are essential for the reasoning process. Comprehensive series understanding requires precise alignment of events and characters, omitting + +either chain decreases performance, highlighting the importance of joint events and characters for effective reasoning. + +Table 5. Performance of 4 Top-Performing MLLMs with Different Input Modalities. (a) Question text (Q) only, (b) Q with Frames (F), (c) Q with Subtitles (S), (d) Q with F and S, (e) Q with F, S, and Thematic-Character information (TC). + +
ModelVSSCAUAGCOOverall
InternVL2†76.571.467.381.166.773.3
(w/o) Cha-Temp71.166.364.175.462.468.5
(w/o) Plot-Event69.367.462.380.465.268.8
Qwen2-VL†75.273.569.780.967.473.9
(w/o) Cha-Temp64.466.062.079.661.066.4
(w/o) Plot-Event61.767.658.679.364.565.4
MiniCPM-V 2.6†76.368.266.582.560.072.0
(w/o) Cha-Temp60.158.756.778.954.261.6
(w/o) Plot-Event57.658.157.480.656.561.3
GPT-4o†78.676.173.882.161.776.2
(w/o) Cha-Temp70.166.768.176.958.369.0
(w/o) Plot-Event71.176.566.386.555.872.3
+ +Table 6. Ablation study of dual-chains on SeriesBench. †: MLLMs using the PC-DCoT framework during inference. + +# 6. Conclusion + +To address the challenge of narrative understanding across multi-video series, this paper introduces SeriesBench, a novel benchmark designed to comprehensively evaluate MLLMs' understanding of narrative structures and character relationships across series. We further propose PC-DCoT, an adaptable and robust framework that enables MLLMs to achieve over a $10\%$ improvement on Series-Bench. Experiments across a wide range of video-MLLMs reveal significant challenges in comprehending narrative structures and character relationships within series, emphasizing the necessity for continued advancements in this area. We hope that our SeriesBench and PC-DCoT will inspire future research focused on enhancing the capacity of MLLMs to understand narrative-driven drama series. + +# Acknowledgements + +This work is supported by "Pioneer" and "Leading Goose" R&D Program of Zhejiang (No. 2024C01020), the National Natural Science Foundation of China (No. 62406015), research funding from Kuaishou Technology, the Emergency Management Research and Development Project of Zhejiang Province (No. 2024YJ018). + +# References + +[1] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. Touchstone: Evaluating vision-language models by language models, 2023. 2 +[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6 +[3] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. arXiv preprint arXiv:2311.14906, 2023. 2, 4 +[4] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in videolms. arXiv preprint arXiv:2406.07476, 2024. 1, 3, 5, 6, 7 +[5] Daniel Cores, Michael Dorkenwald, Manuel Mucientes, Cees G. M. Snoek, and Yuki M. Asano. Tvbench: Redesigning video-language evaluation, 2024. 1, 4 +[6] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 4 +[7] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1, 2 +[8] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 1, 2, 4, 6, 8 +[9] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, YifanZhang, Shaoqi Dong, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, Ran He and Rongrong Ji, Yunsheng Wu, Caifeng Shan, and Xing Sun. Vita: Towards open-source interactive omni multimodal llm, 2024. 1, 6, 7 +[10] Megan Fulwiler and Kim Middleton. After digital storytelling: Video composing in the new media age. Computers and Composition, 29(1):39-50, 2012. 2 +[11] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, + +Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 + +[12] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. TGIF-QA: toward spatio-temporal reasoning in visual question answering. In CVPR, 2017. 4 + +[13] Yunseok Jang, Yale Song, Chris Dongjoo Kim, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Video Question Answering with Spatio-Temporal Reasoning. IJCV, 2019. 2 + +[14] Bohao Li, Yuying Ge, Yi Chen, Yixiao Ge, Ruimao Zhang, and Ying Shan. Seed-bench-2-plus: Benchmarking multimodal large language models with text-rich visual comprehension. arXiv preprint arXiv:2404.16790, 2024. 2 +[15] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 6, 7 +[16] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model, 2024. 1, 6, 7 +[17] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding, 2024. 2 +[18] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 2, 4 +[19] Shuai Lin Li, Yuang Zhang, Yucheng Zhao, Qiuyue Wang, Fan Jia, Yingfei Liu, and Tiancai Wang. Vlm-eval: A general evaluation on video large language models. arXiv preprint arXiv:2311.11865, 2023. 2 +[20] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji rong Wen. Evaluating object hallucination in large vision-language models. ArXiv, abs/2305.10355, 2023. 1 +[21] Yunxin Li, Xinyu Chen, Baotian Hu, Longyue Wang, Haoyuan Shi, and Min Zhang. Videovista: A versatile benchmark for video understanding and reasoning, 2024. 1, 2, 4, 6, 8 +[22] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection, 2024. 2 +[23] Yuanzhan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? ArXiv, abs/2307.06281, 2023. 1 +[24] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Temp-compass: Do video llms really understand videos? ArXiv preprint, 2024. 4 + +[25] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv: 2403.00476, 2024. 1, 2 +[26] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Chang Wen Chen, and Ying Shan. E.t. bench: Towards open-ended event-level video-language understanding. In Neural Information Processing Systems (NeurIPS), 2024. 1 +[27] Monica Mak. Visual postproduction in participatory videomaking processes. Handbook of participatory video, pages 194-207, 2012. 2 +[28] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 2 +[29] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. In NeurIPS, 2024. 4 +[30] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. ArXiv preprint, 2023. 4 +[31] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence, 2024. Accessed: 2024-10-31. 1, 6, 7 +[32] OpenAI. Gpt-4o system card, 2024. Accessed: 2024-10-31. 1, 5, 6, 7 +[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6 +[34] 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, 2021. 5 +[35] Xiaoming Shi, Zeming Liu, Yiming Lei, Chenkai Zhang, Haitao Leng, Chuan Wang, Qingjie Liu, Wanxiang Che, Shaoguo Liu, Size Li, et al. Kwaichat: A large-scale video-driven multilingual mixed-type dialogue corpus. arXiv preprint arXiv:2503.06899, 2025. 1 +[36] InternVL Team. Internvl2: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024. Accessed: 2024-10-31. 2, 5, 6, 7 +[37] Ke Wang, Jiahui Zhu, Minjie Ren, Zeming Liu, Shiwei Li, Zongye Zhang, Chenkai Zhang, Xiaoyu Wu, Qiqi Zhan, Qingjie Liu, et al. A survey on data synthesis and augmentation for large language models. arXiv preprint arXiv:2410.12896, 2024. 1 +[38] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2, 5, 6, 7 + +[39] Bo Wu and Shoubin Yu. Star: A benchmark for situated reasoning in real-world videos. In NeurIPS, 2024. 4 +[40] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9777–9786, 2021. 4 +[41] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In ACM Multimedia, 2017. 2, 4 +[42] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Meng Lei, Fanqing Meng, Siyuan Huang, Yu Qiao, and Ping Luo. Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265, 2023. 1, 2 +[43] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 2, 6, 7 +[44] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities, 2023. 1, 2 +[45] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, 2019. 4 +[46] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations, 2020. 6 +[47] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data, 2024. 1, 6, 7 +[48] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2 \ No newline at end of file diff --git a/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/images.zip b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d07fe9f139353c94e3ce2a88b49ce65a6665afbd --- /dev/null +++ b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e859268febe3159c61fee04d3c2646329e4ed7ca29fd026683eb8d8ce9a3f646 +size 958113 diff --git a/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/layout.json b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3f52673e434b22de872bf5d998e7189221d7f542 --- /dev/null +++ b/CVPR/2025/SeriesBench_ A Benchmark for Narrative-Driven Drama Series Understanding/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3a091077be93f32c7a73eb24834f0b9bf732138170545ee006083971a47931a +size 280404 diff --git a/CVPR/2025/Seurat_ From Moving Points to Depth/7b500a33-6c02-4d6d-8fe4-184de0c5a7f2_content_list.json b/CVPR/2025/Seurat_ From Moving Points to Depth/7b500a33-6c02-4d6d-8fe4-184de0c5a7f2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..da2a4dca043704f3a33c5cb94cb22c6b298d2000 --- /dev/null +++ b/CVPR/2025/Seurat_ From Moving Points to Depth/7b500a33-6c02-4d6d-8fe4-184de0c5a7f2_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d456318a087fbb60f4bb4f6f9bfa9c6d6cf36d046c850143b572829cc4e3128 +size 87303 diff --git a/CVPR/2025/Seurat_ From Moving Points to Depth/7b500a33-6c02-4d6d-8fe4-184de0c5a7f2_model.json b/CVPR/2025/Seurat_ From Moving Points to Depth/7b500a33-6c02-4d6d-8fe4-184de0c5a7f2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..79176b54ccf9b14c834fe43a0578a7456d9bde31 --- /dev/null +++ b/CVPR/2025/Seurat_ From Moving Points to Depth/7b500a33-6c02-4d6d-8fe4-184de0c5a7f2_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:961ac59419490d84e101295d51af030229dce6d74161e8e0ed66f73aaaaaec74 +size 107794 diff --git a/CVPR/2025/Seurat_ From Moving Points to Depth/7b500a33-6c02-4d6d-8fe4-184de0c5a7f2_origin.pdf b/CVPR/2025/Seurat_ From Moving Points to Depth/7b500a33-6c02-4d6d-8fe4-184de0c5a7f2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..41f454b4e2244b8c5b8c5b21dab60861e183f0db --- /dev/null +++ b/CVPR/2025/Seurat_ From Moving Points to Depth/7b500a33-6c02-4d6d-8fe4-184de0c5a7f2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c8e284005bb01452cb4d2117375b13d003f85f1825172bd16a61cb4445dbfcb +size 6176000 diff --git a/CVPR/2025/Seurat_ From Moving Points to Depth/full.md b/CVPR/2025/Seurat_ From Moving Points to Depth/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ceeb54853d2cfdb647bebdbe0d18df3af908d3b4 --- /dev/null +++ b/CVPR/2025/Seurat_ From Moving Points to Depth/full.md @@ -0,0 +1,361 @@ +# Seurat: From Moving Points to Depth + +Seokju Cho + +Jiahui Huang + +Seunrgyong Kim + +Joon-Young Lee + +1KAISTAI + +2Adobe Research + +![](images/dc4ace6e1acc1009a918fd73b70b891122987f245d0e9063ab3bab0cf5416063.jpg) +Figure 1. Seurat predicts precise and smooth depth changes for dynamic objects over time by only looking at the 2D point trajectories, which encode depth cues in their motion patterns. The figure illustrates 2D point tracks lifted into 3D space with our depth predictions on videos from the DAVIS dataset [41]. + +![](images/dafd2e169b787fa3580e83d0aa4501b15d02036bbb1549d160a090e03ba7114d.jpg) +3D Trajectories + +![](images/cb0b4a4f189c2044fa067b77fa6da3faac234eb96b9567ae13266e86ed663891.jpg) + +![](images/87b7de6cd33db9e468b96b505918631a93d54afd57b9cc024e7790e2c7215b9c.jpg) + +![](images/9af71e22a18139733387a2fec11bc9a37e54a9cba2cf4c33bf4c3a654f87e30e.jpg) + +# Abstract + +Accurate depth estimation from monocular videos remains challenging due to ambiguities inherent in single-view geometry, as crucial depth cues like stereopsis are absent. However, humans often perceive relative depth intuitively by observing variations in the size and spacing of objects as they move. Inspired by this, we propose a novel method that infers relative depth by examining the spatial relationships and temporal evolution of a set of tracked 2D trajectories. Specifically, we use off-the-shelf point tracking models to capture 2D trajectories. Then, our approach employs spatial and temporal transformers to process these trajectories and directly infer depth changes over time. Evaluated on the TAPVid-3D benchmark, our method demonstrates robust zero-shot performance, generalizing effectively from synthetic to real-world datasets. Results indicate that our approach achieves temporally smooth, high-accuracy depth predictions across diverse domains. + +# 1. Introduction + +Understanding the 3D structure of a scene is essential for numerous applications, including image and video generation [56], robotics [12], autonomous driving [40], and 3D reconstruction [32]. However, obtaining accurate depth in + +formation from monocular images is challenging due to inherent ambiguities [17], difficulties in scale and shift determination [42], and the considerable diversity of real-world scenes [42, 55]. In monocular video sequences, additional complexities arise from dynamic objects exhibiting intricate movements [36] and lengthy sequences complicating the maintenance of temporal coherence in depth estimation [21]. + +A classical method for obtaining precise depth information is Structured Light 3D Scanning [13, 14, 33], which involves projecting known patterns onto surfaces. The deformation of these patterns upon interacting with surfaces encodes valuable data about the 3D structure of the objects. This effectively transforms the spatial variations in the scene into measurable distortions in the projected pattern, allowing for accurate depth reconstruction. + +Drawing inspiration from Structured Light methods [13, 14, 33], we propose that similar principles can be applied to monocular video sequences, assuming local object rigidity. Just as the deformation of a projected pattern reveals depth information, the patterns formed by the trajectories of tracked points [9-11, 23] in video sequences can reveal the 3D structure of a scene. These trajectories inherently capture spatial relationships and motion patterns relative to the camera, providing robust cues for depth estimation. For example, points moving away from the camera create denser + +![](images/71ad7afec2e0c2435eae13f158e666467cfca5668c7766f8901e66225130edaf.jpg) +(a) From the points alone, we can tell if the car is moving away or approaching. + +![](images/fdc072e0035eabd1a23b4a144788d56af6a8ad1281c9123c43e8c2a3e1653013.jpg) +(b) Object recedes, points converge toward center. +Figure 2. Motivation of our work. (a) By only looking at the tracked points, we can easily perceive that the object (here, a car) is moving away. (b) As a 3D object (here, a sphere) moves away from the camera, the pattern of its projected 2D points on the image plane changes, providing depth cues. In the initial frame (left), points are spaced farther apart on the image plane. As the object recedes (right), these 2D points converge toward the center, indicating increasing depth. This change in the density of projected points allows for inference of relative depth changes from motion in monocular video. + +trajectory patterns, as illustrated in Figure 2. Analyzing the relative motion of these points allows us to discern whether objects or surfaces approach or recede from the camera, facilitating accurate temporal depth estimation. + +Specifically, we begin by extracting trajectories using off-the-shelf point-tracking models [9, 24]. To exploit depth information encoded in these trajectories, we employ both spatial and temporal transformers [1, 24], which respectively model spatial relationships and ensure temporal smoothness. Furthermore, we explicitly decouple the supporting trajectories from query trajectories to avoid potential biases during depth estimation. + +We observe that processing an entire video sequence simultaneously results in highly complex and unstable depth predictions, particularly for objects exhibiting rapid movements relative to the camera. To mitigate this issue, we predict temporal depth changes within sliding windows, assuming that depth variations are more consistent and manageable in shorter segments. Additionally, we introduce a specially designed window-wise log-ratio depth loss to achieve accurate depth supervision. We found this approach critical for reliably learning relative depth. + +Our work introduces a novel trajectory-based framework for depth estimation from monocular videos, capitalizing on the temporal evolution of point trajectories without relying on stereo [7] or multi-view setups [35], additional sensors [49], strong feature backbones [4, 28, 37], or extensive annotated datasets [42]. Our approach infers depth change over time in a strictly zero-shot manner, trained solely on a synthetic dataset [18] without any pre-trained feature backbone. Despite this simplicity, our model demonstrates ro + +bust generalization capabilities and performs effectively on real-world datasets. + +We evaluate our proposed method on the TAPVid-3D benchmark [27], highlighting its effectiveness and robustness in diverse scenarios, including driving scenes [15, 16], egocentric viewpoints [38], and videos containing deformations [22]. Qualitative analyses further demonstrate that our method consistently produces temporally smooth and highly accurate depth predictions. + +# 2. Related Work + +Point tracking in 2D. Track Any Point (TAP) [9, 11, 19, 23, 26, 27], or point tracking, aims to track any given query point throughout a video sequence along with its visibility status. PIPs [19] iteratively refines trajectories using an MLP-Mixer [50] architecture. TAP-Net [10] constructs a global cost volume followed by convolutions and applies a soft-argmax operation for point tracking. TAPIR [11] initializes trajectories with TAP-Net and refines them using PIPs' iterative refinement method. CoTracker [23] tracks multiple points simultaneously and models the interactions between them using a Transformer [52] architecture. TAPTR [29, 30] introduces the perspective of object detection to the point tracking. LocoTrack [9] enhances the cost volume with the neighboring points of query points, constructing 4D local correlation volume. Another approach involves test-time training [46, 51, 53], which can benefit from optimization regularization. In our work, we leverage these off-the-shelf point tracking models and extract depth information from the trajectories they produce. + +Point tracking in 3D. Recently, SpatialTracker [54] + +demonstrated that point tracking can be improved by operating in 3D space. This is accomplished by lifting tracked points into 3D using a monocular depth estimation model [3], followed by iterative refinement within a triplane representation [6], which enhances the accuracy of the resulting 2D point trajectories. While SpatialTracker leverages 3D structural information to enable more robust tracking, our approach instead focuses on uncovering the 3D geometry inherently encoded within the trajectories themselves. TrackTo4D [25] performs Structure-from-Motion [44] from trajectory data, representing dynamic parts with a low-dimensional basis [5]. Unlike TrackTo4D, we do not introduce extra constraints or assumption for dynamic parts. + +Monocular depth estimation. With advancements in deep learning, MDE has achieved rapid progress by learning depth from large and diverse datasets. MegaDepth [31] demonstrated that training on extensive, varied datasets leads to better generalization and improved robustness to domain shifts. MiDaS [42] built on this by blending datasets from multiple sources. Recent work has further expanded data sources by incorporating self-training on unlabeled data [55] or using foundational models as a backbone, such as Stable Diffusion [43] or DINOv2 [37]. Although these approaches show strong results on per-frame depth estimation, they often show flickering when applied to video. + +# 3. Method + +# 3.1. Motivation and Overview + +Estimating depth from monocular videos is challenging due to limited depth cues and inherent ambiguities in monocular images [17]. Traditional methods primarily focus on spatial relative depth [3, 4, 42, 55], which concerns the depth differences between points within the same frame. In contrast, we infer depth changes over time, exploiting the rich temporal information that monocular videos inherently contains. + +We observe that point trajectories over time encapsulate valuable temporal depth information. Specifically, the motion patterns of tracked points can reveal whether objects or surfaces are moving towards or away from the observer, as illustrated in Figure 2. Our approach leverages the temporal evolution of 2D point trajectories to predict depth changes, avoiding reliance on stereo [7, 49], multiview setups [20, 28], additional sensors [49], or large-scale dataset [42]. + +Formally, given a monocular video with $T$ frames, we employ an off-the-shelf point tracking model [9, 24] to extract $N$ trajectories along with their occlusion statuses: + +- Trajectories: $\mathcal{T} = \{\mathbf{p}_i\}_{i=1}^N$ , where each trajectory $\mathbf{p}_i = \{p_{i,t}\}_{t=1}^T$ consists of the 2D positions of point $i$ across frames. +- Occlusion Statuses: $\mathcal{V} = \{\mathbf{v}_i\}_{i=1}^N$ , where $\mathbf{v}_i =$ + +$\{v_{i,t}\}_{t = 1}^{T}$ indicates the visibility of point $i$ at each frame. + +We develop a model that takes the extracted trajectories $\mathcal{T}$ and occlusion statuses $\nu$ as input to predict temporal changes in depth. Specifically, our model estimates the depth ratio along each trajectory $i$ : $r_{i,t} = d_{i,t} / d_{i,t_0}$ , relative to a reference frame $t_0$ , where $d_{i,t}$ represents the depth of point $i$ at frame $t$ . + +We begin by detailing the theoretical basis behind extracting depth cues from trajectories in Sec. 3.2. Subsequently, we introduce our Transformer-based model architecture for depth ratio prediction in Sec. 3.3. Next, we describe the sliding window training and inference in Sec. 3.4 and outline the inference strategy in Sec. 3.5, which combines predicted depth ratios $\hat{r}$ with a metric monocular depth estimation model to produce the final metric depth estimates $\hat{d}_{i,t}$ . + +# 3.2. Theoretical Analysis + +Point trajectories over time encapsulate valuable temporal relative depth information. Specifically, the motion patterns of tracked points can reveal depth changes based on the density variation of points projected onto the image plane. Under the assumption of a pinhole camera model, consider a small surface patch with area $A_{\mathrm{surface}}$ at depth $d_t$ and orientation $\theta_t$ in frame $t$ , where $\theta_t$ is the angle between the surface normal and the camera's optical axis. The projected area of this patch onto the image plane is given by: + +$$ +A _ {t} ^ {\text {i m a g e}} = \left(\frac {f}{d _ {t}}\right) ^ {2} A ^ {\text {s u r f a c e}} \cos \theta_ {t}, \tag {1} +$$ + +where $f$ is the focal length of the camera. The density of projected points $\rho_t^{\mathrm{image}}$ is inversely proportional to $A_t^{\mathrm{image}}$ : + +$$ +\rho_ {t} ^ {\text {i m a g e}} \propto \frac {(d _ {t}) ^ {2}}{f ^ {2} A _ {\text {s u r f a c e}} \cos \theta_ {t}}. \tag {2} +$$ + +Assuming local rigidity in a small area $A^{\mathrm{surface}}$ , the ratio of densities at times $t$ and $t_0$ is: + +$$ +\frac {\rho_ {t _ {0}} ^ {\text {i m a g e}}}{\rho_ {t} ^ {\text {i m a g e}}} = \left(\frac {d _ {t _ {0}}}{d _ {t}}\right) ^ {2} \left(\frac {\cos \theta_ {t}}{\cos \theta_ {t _ {0}}}\right) ^ {2}. \tag {3} +$$ + +Taking the square root yields the depth ratio $r_t$ , which is the depth variation over time with respect to the anchor timestep $t_0$ : + +$$ +r _ {t} = \frac {d _ {t}}{d _ {t _ {0}}} = \sqrt {\frac {\rho_ {t} ^ {\text {i m a g e}}}{\rho_ {t _ {0}} ^ {\text {i m a g e}}}} \frac {\cos \theta_ {t}}{\cos \theta_ {t _ {0}}}. \tag {4} +$$ + +It is important to note that the theoretical derivation above assumes the object is locally rigid and requires precise knowledge of its rotational orientation over time $\cos \theta_{t} / \cos \theta_{t_{0}}$ . Calculating depth changes deterministically + +![](images/270822ed71dc74c8f22406b7fe87a27422e5387520109a74792cacab05199d78.jpg) +Figure 3. Overall architecture. We first use off-the-shelf point tracker [9, 24] to extract 2D trajectories of query points and a dense supporting grid, then, these trajectories are processed with a temporal and a spatial transformer in two separate branches. Motion information encoded by the supporting branch is injected into the query branch via cross-attention. Finally, two regression heads output ratio depths of both supporting and query trajectories. + +using these formulas demands accurate measurements of rotation and local rigidity, which can be impractical in real-world scenarios, as validated in Table 4. + +To overcome these limitations, we employ a transformer-based framework that implicitly captures the complexities of object motion and rotation through learned representations. This approach allows us to estimate ratio depth without relying solely on the explicit calculations provided by the equations, making the prediction more robust to variations in object properties and motion dynamics. + +# 3.3. Processing Trajectories with Transformer + +For effective depth estimation, it's crucial that the model captures the comprehensive motion patterns of the entire scene. Relying solely on user-defined or dataset-provided query points may not suffice, as they might not adequately represent the scene's overall motion due to their uneven distribution, often biased to the salient objects [2, 10, 57]. This limitation can hinder the model's ability to infer depth accurately, especially in regions lacking sufficient trajectory. + +To address this challenge, we introduce supporting trajectories derived from a grid of uniformly sampled points across the image. These grid-shaped trajectories offer a uniform and dense representation of the scene's motion, effectively capturing both local and global movements. By incorporating these supporting trajectories, the model gains a holistic understanding of the scene's dynamics, which is essential for accurate depth estimation. + +Additionally, to prevent the biased distribution of query points from influencing the supporting trajectories, we decouple the model into two branches. The supporting trajectory branch processes the supporting grid trajectories to capture global motion information, while the query trajectory branch processes the query trajectories. The motion information encoded by the supporting branch is then injected into the query branch using cross-attention [52]. This design ensures that the depth predictions for query points are informed by the overall scene motion without being biased by the distribution of the query points. The overall architecture is illustrated in Figure 3. + +# 3.3.1 Supporting Trajectory Branch + +The supporting trajectory branch processes a uniform grid of trajectories $\mathcal{T}_s$ , which we refer to as supporting trajectories. These are obtained by tracking a predefined grid of points using the point tracking model [9, 24]. The supporting trajectories capture the overall motion dynamics of the scene and provide contextual information that aids in accurate depth estimation. + +Let $\mathcal{T}_s = \{\mathbf{p}_i^{\mathrm{supp}}\}_{i=1}^{i=N_s}$ denote the supporting trajectories, where $N_s$ is the number of supporting points. Each supporting trajectory $\mathbf{p}_i^{\mathrm{supp}}$ is associated with an occlusion status $\mathbf{v}_i^{\mathrm{supp}} \in \mathcal{V}_s$ . The encoder processes these trajectories using a Transformer architecture [52] with alternating temporal and spatial attention layers [1, 8, 24]. + +Formally, the supporting trajectories are embedded and processed through $L$ layers: + +$$ +\mathbf {h} _ {s} ^ {0} = \operatorname {E m b e d d i n g} \left(\mathcal {T} _ {s}, \mathcal {V} _ {s}\right) +$$ + +$$ +\mathbf {h} _ {s} ^ {l} = \operatorname {T r a n s f o r m e r L a y e r} _ {s} ^ {l} \left(\mathbf {h} _ {s} ^ {l - 1}\right), \quad l = 1, \dots , L, \tag {5} +$$ + +where $\mathbf{h}_s^0$ is the initial embedding of the supporting trajectories from trajectory position and occlusion status. Each + +Transformer layer TransformerLayer $_s^l$ consists of temporal attention followed by spatial attention, following recent works in point tracking [8, 24]. The intermediate features $\mathbf{h}_s^l$ captures the encoded motion information from the supporting trajectories, serving as a rich representation of the scene's dynamics. + +# 3.3.2 Query Trajectory Branch + +The query trajectory branch processes trajectories $\mathcal{T}_q$ and their visibility status $\nu_{q}$ , obtained by tracking user-defined query points. The query trajectory branch predicts the ratio depth by incorporating motion information from the supporting trajectories. To achieve this, we employ a Transformer architecture with cross-attention that attend to the motion information encoded from supporting trajectories. This allows the query decoder to leverage the global motion context while focusing on the specific query points. + +The decoding process is defined as: + +$$ +\begin{array}{l} \mathbf {h} _ {q} ^ {0} = \operatorname {E m b e d d i n g} \left(\mathcal {T} _ {q}, \mathcal {V} _ {q}\right) \\ \mathbf {h} _ {q} ^ {l} = \operatorname {T r a n s f o r m e r L a y e r} _ {q} ^ {l} \left(\mathbf {h} _ {q} ^ {l - 1}, \mathbf {h} _ {s} ^ {l - 1}\right), \quad l = 1, \dots , L, \tag {6} \\ \end{array} +$$ + +where $\mathbf{h}_q^0$ is the initial embedding of the query trajectories derived from their positions and occlusion status. Each Transformer layer $\mathrm{TransformerLayer}_q^l$ includes a cross-attention that aggregates information from the supporting trajectories with $\mathbf{h}_s^{l - 1}$ . The output $\mathbf{h}_q^{(L)}$ provides a refined representation for depth ratio prediction. Note that for each query point, trajectory is independently processed to prevent the predictions from being influenced by the distribution of query points. + +Both the supporting encoder and the query decoder have ratio depth prediction heads attached to their final layers, which output the estimated relative depth changes for the trajectories. Additionally, inspired by [8, 9, 19, 24, 48], we iteratively refine the predicted depths by feeding the current predictions back into the model in subsequent iterations. + +# 3.4. Sliding Window Prediction + +Processing entire video sequences at once may lead to complex and divergent depth results due to the increased complexity of long-range motion patterns and the potential for supporting trajectories to move out of the frame. To address this, we employ a sliding window approach during both training and inference. + +The sliding window approach involves segmenting the video sequence into shorter temporal windows of length $W$ , with an overlap to ensure continuity, setting the window stride as $S$ . In our model, the query trajectories persist across windows, while the supporting trajectories are re-initialized for each window, making the supporting trajectories more likely to remain within the frame. + +Training with window-wise log ratio depth loss. To align our training objective with Eq. 4, we focus on predicting the log depth ratio for each frame with respect to the starting frame of the current window. By predicting the log depth ratios, the model becomes invariant to the absolute scale of depth, focusing instead on the relative changes. The log depth ratio for point $i$ at time $t$ within window $w$ is defined as: + +$$ +\ell_ {i, t} ^ {w} = \log \left(r _ {i, t} ^ {w}\right) = \log \left(\frac {d _ {i , t} ^ {w}}{d _ {i , 0} ^ {w}}\right), \quad t \in [ 0, W - 1 ], \tag {7} +$$ + +where $W$ is the window size, and $d_{i,t}^{w}$ is the ground-truth depth of point $i$ at time $t$ within $w$ -th window. + +Our model predicts $\hat{\ell}_{i,t}^{w}$ , the predicted log depth ratios for each point $i$ at time $t$ within window $w$ . We train the model using an $L_{1}$ loss between the predicted and ground-truth log depth ratios $\ell_{i,t}$ : + +$$ +\mathcal {L} = \sum_ {i} \sum_ {w} \sum_ {t} \left| \hat {\ell} _ {i, t} ^ {w} - \ell_ {i, t} ^ {w} \right|. \tag {8} +$$ + +This loss encourages the model to accurately predict the depth changes within each window, focusing on the temporal relative depth. We apply the loss to the depth predictions of both the query trajectories and the supporting trajectories. + +Inference with sliding window. Within each sliding window, the model predicts log depth ratios, denoted as $\hat{\ell}_{i,t}^{w}$ , relative to the first frame of that window. To obtain depth predictions for the entire sequence, these log depth ratios are accumulated, and then exponentiated to convert them back to linear-scale depth ratios. Importantly, during inference, the supporting point depth ratios are discarded after each window; only the query point depth ratios are accumulated, as the supporting points serve solely to improve depth estimation within the current window. Specifically, the accumulated depth ratio $\hat{r}_{i,t}$ for point $i$ at time $t$ is calculated as: + +$$ +\hat {r} _ {i, t} = \exp \left(\hat {\ell} _ {i, t _ {k}} ^ {k} + \sum_ {w = 0} ^ {k - 1} \hat {\ell} _ {i, S} ^ {w}\right), \quad t _ {k} = t - S k, \tag {9} +$$ + +where $\hat{\ell}_{i,t}^{w}$ represents the predicted log depth ratio for $t$ -th frame in window $w$ , $S$ denotes the stride of the sliding window, $k$ is the index of the last window that contains time $t$ , and $t_k$ represents the time index of $t$ within window $k$ . + +# 3.5. Incorporating Depth Ratio and Metric Depth + +Although our predicted depth ratios accurately capture changes over time, the accumulated ratio $\hat{r}_{i,t}$ for the entire video reflects only the depth change relative to the initial frame. Therefore, in the final step, we correct each accumulated depth ratio using an off-the-shelf metric depth estimator [3, 4]. + +
Depth EstimatorPoint TrackerAria [38]DriveTrack [2]PStudio [22]Average
3D-AJ ↑APD ↑TC ↓3D-AJ ↑APD ↑TC ↓3D-AJ ↑APD ↑TC ↓3D-AJ ↑APD ↑TC ↓
Oracle depth*CoTracker [24]55.970.3-53.271.7-46.965.0-52.069.0-
TAPIR-3D [27]-8.514.9-10.217.0-7.213.1-8.615.0-
Per-frame Depth Estimator
ZoeDepth [3]Oracle tracker*19.128.60.0510.817.41.2715.524.30.0515.123.40.46
CoTracker [24]16.525.50.069.516.21.3411.820.40.0512.620.70.48
LocoTrack [9]15.725.40.0610.016.31.3311.820.10.0412.520.60.48
DepthPro [4]Oracle tracker*13.421.30.146.712.03.728.915.30.169.716.21.34
CoTracker [24]11.318.70.165.410.53.876.712.70.167.814.01.40
LocoTrack [9]10.818.60.155.610.34.186.712.50.177.713.81.50
Video Depth Estimator
DepthCrafter [21]Oracle tracker*17.527.10.029.816.40.7513.822.60.0313.722.00.26
CoTracker [24]15.124.30.048.415.20.9811.119.50.0311.519.70.35
LocoTrack [9]13.923.50.038.815.10.9511.019.30.0311.219.30.34
ChronoDepth [45]Oracle tracker*12.318.70.227.112.73.943.56.50.337.612.61.50
CoTracker [24]11.017.82.376.111.77.433.05.90.376.711.83.39
LocoTrack [9]10.517.60.236.411.61.493.05.90.386.611.70.70
Seurat (Ours)CoTracker [24]25.136.90.0111.620.40.1517.328.50.0118.028.60.05
LocoTrack [9]24.134.60.0812.720.40.5517.427.60.0718.127.40.23
+ +Table 1. Quantitative results on TAPVid-3D [27] minimal split with per-trajectory depth scaling. Oracle tracker* rows use ground-truth 2D trajectories, while Oracle depth* row uses ground-truth depth estimation to determine the upper bound. + +Concretely, we perform piecewise scale matching for each visible subsequence $S_{i,t}$ . Here, $S_{i,t}$ is the set of consecutive frames that includes time $t$ , during which a particular point $i$ remains visible. We compute a scale factor $s_{i,t}$ by matching the median of $\hat{r}_{i,t}$ to the median of the MDE model's depth estimates over all $t' \in S_{i,t}$ : + +$$ +s _ {i, t} = \frac {\operatorname {m e d i a n} _ {t ^ {\prime} \in \mathcal {S} _ {i , t}} \left(d _ {\mathrm {M D E}} \left(p _ {i , t ^ {\prime}}\right)\right)}{\operatorname {m e d i a n} _ {t ^ {\prime} \in \mathcal {S} _ {i , t}} \left(\hat {r} _ {i , t ^ {\prime}}\right)}, \tag {10} +$$ + +where $d_{\mathrm{MDE}}(p_{i,t'})$ is the MDE-predicted depth at point $p_{i,t'}$ , obtained by bilinear interpolation from the depth map. We then apply this scale factor to our predicted ratio to obtain the final depth estimate: + +$$ +\hat {d} _ {i, t} = s _ {i, t} \cdot \hat {r} _ {i, t}. \tag {11} +$$ + +By matching medians in each subsequence, we align our model's temporal depth changes with the MDE model's metric depth estimates. This integration leverages the strengths of both approaches: our model offers temporally coherent depth changes, while the MDE model ensures reliable spatial depth relationships within each frame. The resulting estimates are therefore both temporally stable and spatially precise. + +# 4. Experiment + +# 4.1. Evaluation Protocol and Baselines + +Evaluation protocol. We assess trajectory depth accuracy using the TAPVid-3D benchmark [27], which encom + +passes both outdoor and indoor scenarios through egocentric videos, driving scenes, and studio setups. For details on the datasets included in the TAPVid-3D benchmark, please refer to the supplementary materials. + +We measure the accuracy of the predicted position using APD, which represents the percentage of points within a threshold $\delta$ . Unlike 2D point tracking, which defines threshold values in pixels, we use a depth-adaptive threshold [27]. We report the average score across $\delta = 1, 2, 4, 8, 16$ . 3D-AJ reflects combined accuracy for both position and occlusion. In this study, we prioritize position accuracy over occlusion, as we rely on the occlusion accuracy of an off-the-shelf 2D point tracker. We also measure the temporal coherence (TC) [8, 53] of the predicted 3D tracks, which is the $L2$ distance between the ground-truth acceleration of the trajectory and that of the predicted trajectory. + +Baselines. For comparison, we construct simple baselines by pairing a 2D point tracking model with a monocular depth estimation (MDE) model. By unprojecting the 2D tracks generated by the tracking module using depth estimates from the MDE model, we obtain a 3D trajectory over the video sequence. For 2D point tracking, we use recent state-of-the-art models for their accuracy, specifically LocoTrack [9] and CoTracker [24]. We carefully select MDE models, ensuring that their training sets do not overlap with our evaluation benchmark dataset. Some state-of-the-art MDE models, such as UniDepth [39], use the Waymo dataset [47] for training. To avoid overlap, we use ZoeDepth [3] and DepthPro [4] as our baseline models. + +
Depth EstimatorPoint TrackerAria [38]DriveTrack [2]PStudio [22]Average
3D-AJ ↑APD ↑TC ↓3D-AJ ↑APD ↑TC ↓3D-AJ ↑APD ↑TC ↓3D-AJ ↑APD ↑TC ↓
Oracle depth*CoTracker [24]55.970.3-53.271.7-46.965.0-52.069.0-
DepthCrafter [21]Oracle tracker*9.515.10.0169.615.20.48413.421.10.01510.817.10.172
CoTracker [24]7.412.20.0417.212.30.5949.316.40.0158.013.60.217
LocoTrack [9]6.712.20.0307.011.50.5848.214.50.0177.312.70.210
ChronoDepth [45]Oracle tracker*12.318.70.0338.013.30.91311.818.90.02010.717.00.322
CoTracker [24]10.115.90.0456.711.90.8958.615.40.0208.514.40.320
LocoTrack [9]9.415.70.0446.210.70.8888.314.70.0228.013.70.318
Seurat (Ours) + ZoeDepth [3]CoTracker [24]11.318.00.0127.512.90.24411.419.20.01210.116.70.089
LocoTrack [9]10.716.70.0867.612.30.56710.416.80.0699.615.30.241
Seurat (Ours) + DepthPro [4]CoTracker [24]15.122.50.0128.413.90.21912.520.50.01312.019.00.081
LocoTrack [9]14.521.40.0868.713.50.55112.018.80.07011.717.90.236
+ +Table 2. Quantitative results of affine-invariant depth on TAPVid-3D [27] minival split. Compared to video depth estimators, our model shows superior performance. Oracle tracker* rows use ground-truth 2D trajectories, while Oracle depth* row uses ground-truth depth estimation to determine the upper bound. + +
MethodAverage# of LayersAverage
3D-AJAPD3D-AJAPD
(I)Seurat (Ours)18.028.6115.525.1
(II)(I) - Two-branch design13.723.12 (Ours)18.028.6
(III)(I) - Sliding window8.816.1317.828.3
(IV)(I) - Window-wise loss14.924.5417.828.1
+ +Table 3. Ablation studies. Left: Ablation of main components. We conduct gradually exclude core components from our full model. Right: Ablation on the number of layers $L$ . + +
Depth Ratio EstimatorPoint TrackerAria 3D-AJ \( \uparrow \)APD \( \uparrow \)DriveTrack 3D-AJ \( \uparrow \)APD \( \uparrow \)PStudio 3D-AJ \( \uparrow \)APD \( \uparrow \)Average 3D-AJ \( \uparrow \)APD \( \uparrow \)
Hand-crafted baselineCoTracker8.615.25.310.44.28.16.011.2
Seurat (Ours)CoTracker25.136.911.620.417.328.518.028.6
+ +# 4.2. Implementation Details + +We train our model using the AdamW optimizer [34] with a learning rate of $5 \times 10^{-4}$ and a weight decay of $1 \times 10^{-5}$ . We linearly decayed the learning rate during training, with warm-up step of 1,000. We conduct training for 100,000 steps on eight NVIDIA RTX 3090 GPUs, using a batch size of 1 per GPU. We generate 90,000 training samples from the Kubric [18] MOVi-F dataset generator. We use the number of Transformer layers [52] as $L = 2$ , where each layer has a hidden dimension of 384 and uses 8 attention heads. The supporting grid size is set to $24 \times 24$ . For LocoTrack [9], we use the base model with a resolution of $256 \times 256$ . For CoTracker [24], we use the global grid of $6 \times 6$ . We set the temporal window size $W = 8$ . + +# 4.3. Main Results + +Quantitative results. We conduct a quantitative evaluation on the TAPVid-3D benchmark in Table 1 and Table 2. For comparison, we combine existing point tracking methods [9, 24] with monocular depth estimators [3, 4] and video depth estimators [21, 45] by unprojecting the tracked 2D trajectories into 3D space using the predicted depth maps. + +Table 4. Comparison with handcrafted baseline. Handcrafted implementation of Eq. 4 exhibits lower performance. + +
Depth Ratio EstimatorPoint TrackerAriaDriveTrackPStudioAverage
3D-AJ↑APD↑3D-AJ↑APD↑3D-AJ↑APD↑3D-AJ↑APD↑
Seurat + Texture patchCoTracker21.031.810.318.616.127.115.825.8
Seurat (Ours)CoTracker25.136.911.620.417.328.518.028.6
+ +Table 5. Texture patch ablation. Additional texture information reduces performance. + +
MethodAriaDriveTrackPStudio
3D-AJ ↑APD ↑3D-AJ ↑APD ↑3D-AJ ↑APD ↑
ZoeDepth + CoTracker9.815.87.212.310.217.9
ZoeDepth + CoTracker + 1 iter. of Gaussian smoothing4.88.66.311.17.714.1
ZoeDepth + CoTracker + 3 iters. of Gaussian smoothing4.48.06.010.77.413.6
Seurat (Ours)14.621.96.911.812.720.7
+ +Table 6. Comparison to simple Gaussian smoothing. Simple Gaussian smoothing of per-frame depth estimates proves insufficient for achieving precise video depth. + +Despite the simplicity of this approach, their strong tracking and depth performance provide a robust baseline. + +Table 1 presents the tracking precision evaluated on a per-trajectory basis, with each trajectory scaled according to the depth of its query point. By individually scaling each trajectory, we can assess how accurately our method captures depth changes over time for each point, aligning with the focus of our approach. Our method significantly outperforms other baselines in position precision (APD), especially in the Aria dataset. Notably, unlike other depth prediction models [3, 4] trained on 14-21 large-scale annotated real datasets, our method trained on single synthetic dataset demonstrates more robust generalizability in depth prediction. Additionally, it excels in temporal coherence, particularly on the DriveTrack dataset, achieving more than $40\times$ better coherence, underscoring the stability of our approach. + +Table 2 presents the tracking precision obtained when predicted depths are adjusted to ground truth depths using a single scale and shift value per video, determined via least squares. This evaluation is similar to the common practice in video depth estimation studies [21, 45]. This metric reflects the accuracy of depth predictions across an en + +![](images/5667b09ed13eb156ef342956acb9f9565e282761fd29d60c1a31e3b293ab09f8.jpg) +Figure 4. Qualitative comparisons to baselines. We visualize 3D trajectories using the TAPVid-3D [27] benchmark. Compared to baselines that use combinations such as CoTracker with ZoeDepth, our model achieves superior depth smoothness and accuracy. + +tire video sequence. Our method achieves strong results on both the Aria and PStudio datasets, particularly when combined with DepthPro [4]. Additionally, we show that our model provides even better temporal coherence compared to existing video depth estimation methods. + +Qualitative results. In Figure 4, we compare our method with baseline models. Our model demonstrates exceptionally smooth depth predictions with high accuracy relative to the ground truth. + +# 4.4. Ablation and Analysis + +Component ablation. We conduct an ablation study, presented on the left of Table 3. In (II), we remove the two-branch design (query and supporting trajectory branches) of our architecture, instead processing both supporting and query trajectories together using a Transformer encoder. In (III), we eliminate the sliding window approach, processing the entire video sequence at once. In (IV), rather than using ratio depth with respect to the first frame of the sliding window, we use ratio depth with respect to the query point. + +The observed performance degradation when omitting each design choice demonstrates the effectiveness of our approach. Specifically, in (II), joint processing of supporting and query points negatively impacts performance. In (III), processing the full sequence at once significantly reduces performance, suggesting that handling the whole sequence may complicate the learning process due to complex motion patterns. Finally, (IV) shows that predicting ratio depth within a sliding window is beneficial for the model. + +Ablation on the number of layers $L$ . On the left of Table 3, we show the score with varying number of Transformer layers. We found that more than 2 layers does not guarantee the performance boost. + +Analysis on hand-crafted deterministic baseline. In Table 4, we validate the handcrafted baseline that implements Eq. 4. To measure how the local spatial density of semidensely sampled trajectories evolves over time, we use a kernel-based approach that compares each point's neighborhood structure across frames. Specifically, for each trajectory point, we fix its $k$ -nearest neighbors based on their positions in the first frame and then compute a kernel den + +sity estimate in each frame using Gaussian kernels applied to the Euclidean distances between the point and its fixed neighbors in that frame. This provides a smooth, robust estimate of local density that reflects how tightly clustered a point remains with respect to its original neighborhood over time. The density at each frame is then normalized by the corresponding density in the first frame, yielding a relative density change per point across the temporal window. + +The results imply that the hand-crafted baseline performs significantly worse than ours, which we believe is due to its sensitivity to prediction noise and its inability to estimate surface normals, demonstrating the impracticality of estimating trajectory density using a handcrafted method. + +Analysis of the impact of texture as input. To investigate whether texture information can help the model infer depth changes, we conduct an analysis by adding an RGB input path to our model, as shown in Table 5. Specifically, we extract local RGB patches around the trajectory, flatten them, and concatenate them with the Transformer input. The results show that incorporating texture information actually degrades performance. We suspect this is because training on texture data may lead to overfitting to the synthetic dataset. + +Is Seurat simply smoothing the per-frame depth jitter? Table 6 addresses whether Seurat is just trajectory smoothing. Simple Gaussian smoothing of the lifted trajectory, using per-frame depth estimates, yielded worse results than Seurat, indicating that Seurat enforces temporal geometric consistency beyond simple smoothing. + +# 5. Conclusion + +We introduce a novel trajectory-based framework for depth estimation that leverages the temporal evolution of point trajectories in monocular videos. Using spatial and temporal transformers within a sliding window approach, our method processes trajectory data to accurately capture depth changes with high temporal smoothness. This approach offers a fresh perspective on 3D understanding from monocular videos—simple, effective, and straightforward to implement. We hope this method will inspire further exploration and innovation in related fields. + +Acknowledgements. This research was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2019-II190075, RS-2024-00509279, RS-2025-II212068, RS-2023-00227592) and the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism (RS-2024-00345025, RS-2024-00333068), and National Research Foundation of Korea (RS-2024-00346597). + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 2, 4 +[2] Arjun Balasingam, Joseph Chandler, Chenning Li, Zhoutong Zhang, and Hari Balakrishnan. Drivetrack: A benchmark for long-range point tracking in real-world videos, 2023. 4, 6, 7 +[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 3, 5, 6, 7 +[4] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 2, 3, 5, 6, 7, 8 +[5] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 3 +[6] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 3 +[7] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5418, 2018. 2, 3 +[8] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19268-19277, 2024. 4, 5, 6 +[9] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 1, 2, 3, 4, 5, 6, 7 +[10] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew + +Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2, 4 +[11] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 1, 2 +[12] Xingshuai Dong, Matthew A Garratt, Sreenatha G Anavatti, and Hussein A Abbass. Towards real-time monocular depth estimation for robotics: A survey. IEEE Transactions on Intelligent Transportation Systems, 23(10):16940-16961, 2022. 1 +[13] Philipp Fechteler and Peter Eisert. Adaptive colour classification for structured light systems. IET Computer Vision, 3 (2):49-59, 2009. 1 +[14] Philipp Fechteler, Peter Eisert, and Jurgen Rurainsky. Fast and high resolution 3d face scanning. In 2007 IEEE International Conference on Image Processing, pages III-81. IEEE, 2007. 1 +[15] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 2 +[16] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 2 +[17] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 1, 3 +[18] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3749-3761, 2022. 2, 7 +[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2, 5 +[20] Richard Hartley. Multiple view geometry in computer vision. Cambridge university press, 2003. 3 +[21] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 1, 6, 7 +[22] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In Proceedings of the IEEE international conference on computer vision, pages 3334-3342, 2015. 2, 6, 7 +[23] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co + +tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 1, 2 +[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635v2, 2023. 2, 3, 4, 5, 6, 7 +[25] Yoni Kasten, Wuyue Lu, and Haggai Maron. Fast encoder-based 3d from casual videos via point track processing. arXiv preprint arXiv:2404.07097, 2024. 3 +[26] Inès Hyeonsu Kim, Seokju Cho, Jiahui Huang, Jung Yi, Joon-Young Lee, and Seungryong Kim. Exploring temporally-aware features for point tracking. arXiv preprint arXiv:2501.12218, 2025. 2 +[27] Skanda Koppula, Ignacio Rocco, Yi Yang, Joe Heyward, João Carreira, Andrew Zisserman, Gabriel Brostow, and Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d. arXiv preprint arXiv:2407.05921, 2024. 2, 6, 7, 8 +[28] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3 +[29] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Feng Li, Tianhe Ren, Bohan Li, and Lei Zhang. Taptrv2: Attention-based position update improves tracking any point. arXiv preprint arXiv:2407.16291, 2024. 2 +[30] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. In European Conference on Computer Vision, pages 57-75. Springer, 2025. 2 +[31] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 3 +[32] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 1 +[33] Kai Liu, Yongchang Wang, Daniel L Lau, Qi Hao, and Laurence G Hassebrook. Dual-frequency pattern scheme for high-speed 3-d shape measurement. Optics express, 18(5): 5229-5244, 2010. 1 +[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7 +[35] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis, 2023. 2 +[36] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1 +[37] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3 +[38] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new + +benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 2, 6, 7 +[39] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation, 2024. 6 +[40] Marc Pollefeys, David Nister, J-M Frahm, Amir Akbarzadeh, Philippos Mordohai, Brian Clipp, Chris Engels, David Gallup, S-J Kim, Paul Merrell, et al. Detailed real-time urban 3d reconstruction from video. International Journal of Computer Vision, 78:143-167, 2008. 1 +[41] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 1 +[42] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 1, 2, 3 +[43] 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. 3 +[44] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3 +[45] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv preprint arXiv:2406.01493, 2024. 6, 7 +[46] Yunzhou Song, Jiahui Lei, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Track everything everywhere fast and robustly. In European Conference on Computer Vision, pages 343-359. Springer, 2025. 2 +[47] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6 +[48] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 5 +[49] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2, 3 +[50] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, + +Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in neural information processing systems, 34:24261-24272, 2021. 2 +[51] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. In European Conference on Computer Vision, pages 367-385. Springer, 2025. 2 +[52] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 4, 7 +[53] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 2, 6 +[54] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 2 +[55] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 3 +[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1 +[57] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 4 \ No newline at end of file diff --git a/CVPR/2025/Seurat_ From Moving Points to Depth/images.zip b/CVPR/2025/Seurat_ From Moving Points to Depth/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1ce8070c506c05c09ec4fb3ad6774516c13d5a91 --- /dev/null +++ b/CVPR/2025/Seurat_ From Moving Points to Depth/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c4a0136e6fd6984fd088fcbe11e6b26d6222bf91b3637ba02de87e39ff19434 +size 686173 diff --git a/CVPR/2025/Seurat_ From Moving Points to Depth/layout.json b/CVPR/2025/Seurat_ From Moving Points to Depth/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e7f5350ca5684212402b0460593a6081d173fc13 --- /dev/null +++ b/CVPR/2025/Seurat_ From Moving Points to Depth/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c89119785cd53c05fa893bce47618b74589ecc6db547aaa4b2635011d4f0d831 +size 423790 diff --git a/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/7a80aa8f-8f76-44f7-8c3e-53c972bd2f84_content_list.json b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/7a80aa8f-8f76-44f7-8c3e-53c972bd2f84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9448c9ec49e61ba72050c4cea92c1959121ac97f --- /dev/null +++ b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/7a80aa8f-8f76-44f7-8c3e-53c972bd2f84_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6a7f88ace635aec37d872b37f74f8c5ef42a40d13ee915d51e430ce6e02ab58 +size 84584 diff --git a/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/7a80aa8f-8f76-44f7-8c3e-53c972bd2f84_model.json b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/7a80aa8f-8f76-44f7-8c3e-53c972bd2f84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dd23691931a5c28d0fb1b813e5b28f155dc34d68 --- /dev/null +++ b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/7a80aa8f-8f76-44f7-8c3e-53c972bd2f84_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:031a530dca232116369cd60633a9cf5fb0b1487c37a07c9113140c86b7a18c99 +size 101183 diff --git a/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/7a80aa8f-8f76-44f7-8c3e-53c972bd2f84_origin.pdf b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/7a80aa8f-8f76-44f7-8c3e-53c972bd2f84_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..64c017aa51bfa60907fc80be91457f36c2c5fb52 --- /dev/null +++ b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/7a80aa8f-8f76-44f7-8c3e-53c972bd2f84_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:394f84e9fea68f2c0878651a1310cc63a40bf50723fc0557191be6e14b42afe8 +size 5155024 diff --git a/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/full.md b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8c2782939fd24e71ad6bfe6992bbab4b3d14c5e6 --- /dev/null +++ b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/full.md @@ -0,0 +1,348 @@ +# SfM-Free 3D Gaussian Splitting via Hierarchical Training + +Bo Ji Angela Yao National University of Singapore {jibo,ayao}@comp.nus.edu.sg + +![](images/242cd5f6914cb3cb9a5c43add6b16b6e08b61dc9afc1079a7f45d0df093028ad.jpg) + +![](images/9d962976a5f498d8da3a892865eb7c4105842ea66e0731c0ff068c2129704539.jpg) + +![](images/c759d55c231940fc2c9fa54f945a2cd1afaf894bce0769df0fcf82afa09f4017.jpg) + +![](images/931e11da062446e1e3f6f5ce1b4f092a73cc6f2db42bfb596f4d331b0ccdae3a.jpg) +Figure 1. Novel view synthesis results (left) alongside the projected centers of 3D Gaussians (right). Each blue dot represents a projected 3D Gaussian center. Our proposal offers two key advantages: 1) Our 3D Gaussians are well-distributed across the scene, whereas CF-3DGS [11] has a notable absence of 3D Gaussians on the image's left side (e.g., in the red region); 2) Our learned 3D Gaussians are of high quality. While CF-3DGS places numerous 3D Gaussians in the green region, the rendering quality there is notably inferior to ours. + +![](images/2b6ab661b422c372e0494aa35656ffbc4a667225b23f4464ee2490acc9a0dbc9.jpg) + +![](images/0fdcd8291a8c8c856e4a873154c79ab1ed5b93f0f19a03582e95d656ecc33527.jpg) + +![](images/b46d0e22062635e113b9422d872dabb2146d2a3b11c1bcab643f4b7da8f4400b.jpg) + +![](images/587f01683e0054460f7c0487028a1da0e823489246b10e97e8cf89c187e85abe.jpg) + +![](images/176aad3a220cf35c56ea14671bd36b7ad05efd15d0843331f5dec65c4d6fd251.jpg) + +![](images/76ada8540d897d53cd438df61bfdafdad7c7d33a2d4c2db7e1d7d9c3688382b0.jpg) + +![](images/1148167c1d09ae4cbdb62f429d1d35b2c5336ba7bbd97c46cfe5ec433ca3acd7.jpg) + +![](images/1a6e8e3e03cc42b6b16574e6ca0b891bda4c17d7e55ea46689db8ef2ecc560ae.jpg) + +# Abstract + +Standard 3D Gaussian Splatting (3DGS) relies on known or pre-computed camera poses and a sparse point cloud, obtained from structure-from-motion (SfM) preprocessing, to initialize and grow 3D Gaussians. We propose a novel SfM-Free 3DGS (HT-3DGS) method for video input, eliminating the need for known camera poses and SfM preprocessing. Our approach introduces a hierarchical training strategy that trains and merges multiple 3D Gaussian representations - each optimized for specific scene regions - into a single, unified 3DGS model representing the entire scene. To compensate for large camera motions, we leverage video frame interpolation models. Additionally, we incorporate multi-source supervision to reduce overfitting and enhance representation. Experimental results reveal that our approach significantly surpasses state-of-the-art SfM-free novel view synthesis methods. On the Tanks and Tem + +plies dataset, we improve PSNR by an average of 2.25dB, with a maximum gain of 3.72dB in the best scene. On the CO3D-V2 dataset, we achieve an average PSNR boost of 1.74dB, with a top gain of 3.90dB. The code is available at https://github.com/jibo27/3DGS_Hierarchical_Training. + +# 1. Introduction + +3D Gaussian Splatting (3DGS) [15] represents a 3D scene from multi-view images based on camera intrinsic and extrinsic parameters along with an initial point cloud. Obtaining camera poses and the initial point cloud requires preprocessing, which is often performed using a structure-from-motion (SfM) algorithm [29]. However, SfM can be time-consuming and may struggle with repetitive patterns, textureless regions, or feature extraction errors. Additionally, SfM lacks differentiability, which can limit its applicability in future research [4]. As such, a class of new methods for novel view synthesis is trying to eliminate the need for SfM + +preprocessing [9, 11, 14, 20]. + +Removing SfM preprocessing introduces two obvious questions for 3DGS. First, how can the camera poses of the input images be estimated? Second, how can 3D gaussians be initialized and grown within the scene? Inspired by CF-3DGS [11], we address the challenge of constructing an SfM-free 3DGS from video sequences. Assuming video input with small camera movement, we address the issue of camera pose estimation by predicting the relative poses between temporally adjacent frames. By sequentially stacking these relative poses, we obtain the overall camera poses. + +To improve camera pose estimation, a key innovation in our work is leveraging a video frame interpolation (VFI) model to generate additional frames. We using an off-the-shelf deep model [18] to double the input video length by interpolating between frames. Although these interpolated frames are not rendered from an underlying 3D model and may lack perfect geometric consistency, they provide sufficient quality (see Fig. 3) to bridge relative poses between frames, which is particularly beneficial for sequences with larger camera movements. They also provide additional supervision, covering viewpoints not present in the original training frames. Incorporating these interpolated frames into 3DGS training yields a $0.35\mathrm{dB}$ performance boost on the Tanks and Temples dataset [17]. + +To address the second question of initializing and growing the 3D Gaussians, a straightforward way would be to use a point cloud derived from the depth map of the first frame; however, this often leads to sparse Gaussian coverage for regions not visible in the first frame. The standard adaptive density control [15] – which adjusts 3D Gaussians by splitting, cloning, and pruning – struggles in these sparsely covered regions. In these areas, the Gaussians may have very small gradients, making it challenging to activate densification processes [5, 39]. + +To this end, we propose a novel hierarchical training strategy that merges multiple base 3DGS models, each optimized for specific parts of the scene, into a unified model representing the entire scene. Intuitively, the adaptive density control encounters difficulties in regions with sparse 3D Gaussians; however, with our strategy, these regions are populated with 3D Gaussians merged from other 3DGS models. Interestingly, this merging strategy can be viewed as a densification process: we discard unimportant 3D Gaussians and densify the representation by merging essential Gaussians from different base 3DGS models. Fig 1 illustrates the improved 3D Gaussian coverage achieved by our approach compared to a naive strategy without hierarchical training. This strategy boosts PSNR by $1.19 - 1.58$ dB on the Tanks and Temples dataset[17]. + +Furthermore, we enhance the representation quality through multi-source supervision, leveraging both base 3DGS models and interpolated frames from VFI. + +Our approach achieves a significant PSNR improvement of 2.25 dB on the Tanks and Temples [17] and 1.74 dB on the CO3D-V2 [28] over state-of-the-art SfM-free novel view synthesis methods. Even without known camera intrinsics, our method surpasses the state-of-the-art methods by 0.89 dB in PSNR. Our contributions are as follows: + +- We improve pose estimation by leveraging video frame interpolation to smooth camera motion. +- We introduce a hierarchical training strategy to address initialization and density control challenges without SfM preprocessing. Interestingly, this approach can be interpreted as a densification step. +- We employ multi-source supervision, reusing base 3DGS models and VFI-interpolated frames to reduce overfitting. +- Together, these innovations yield a 3DGS approach that requires no SfM preprocessing, significantly outperforming existing SfM-free novel view synthesis methods. + +# 2. Related Works + +Novel view synthesis is the task of predicting realistic images from an unobserved viewpoint. NeRFs [24] are an implicit 3D representation that encode scenes within an MLP. They are remarkable at rendering images from novel views, but despite the various proposed improvements [1, 2, 10, 12], they are still relatively slow and may require up to several minutes to render a scene. In contrast, 3D Gaussian splats [15], as an explicit representation, are able to achieve high-quality renderings at much faster speeds. Subsequent works build upon 3D Gaussian splatt-ting, dealing with revising the density [5, 39], compressing the representation [19, 25, 26] and anti-aliasing [35, 37]. Others extend 3D Gaussian splatting to dynamic [8, 30, 33] and large, city-scale scenes [16, 22]. + +SfM-free novel view synthesis for both NeRFs and 3D Gaussian splatting is a class of works that try to do away with known or estimated camera pose from SfM. Examples include i-NeRF [36], which estimates camera poses by aligning keypoints using a pre-trained NeRF. Followups like NeRFmm [32], SiNeRF [34], BARF [21] and GARF [7] learn both the NeRF model and camera pose embeddings simultaneously [32], addressing the gradient inconsistency [7, 21], leveraging pre-trained networks for monocular depth estimation or optical flow, incorporating prior geometric knowledge or correspondence information [4, 6, 23]. + +For 3D Gaussian splatting, CF-3DGS [11] and GGRt [20], InstantSplat [9], COGS [14] was developed to support SfM-free optimization. CF-3DGS [11] performs an affine transformation on the positions of the 3D Gaussians to predict relative poses, progressively expanding the representations from the first frame to the last frame. However, its performance is limited by the accuracy of estimated cam + +era poses. It also suffers when there is insufficient initialization of 3D Gaussians, and has challenges in density control. GGRt [20] jointly learns two modules for iterative pose optimization and a generalizable 3DGS. On the other hand, InstantSplat [9] and COGS [14] are designed primarily for scenarios with sparse image views. In this paper, we focus on the video input with small camera movement, similar to CF-3DGS [11]. We address two main challenges associated with applying 3DGS in SfM-free tasks: improving camera pose estimation and enhancing the initialization and learning of 3D Gaussians. + +# 3. Approach + +# 3.1. Overview + +Consider a video sequence $\mathcal{I} = \{I_i \mid i = 1, \dots, N\}$ captured with small camera movements. We aim to reconstruct a 3D Gaussian splatting representation (3DGS) $S$ from $\mathcal{I}$ and camera intrinsics $K$ . We first estimate a series of camera poses $\mathcal{P} = \{P_i \mid i = 1, \dots, N\}$ (Section 3.2). Then, we partition the video into overlapping segments $\{C_j\}$ . For each segment $C_j$ , a base 3DGS model $S_{C_j}$ is trained. These models are then iteratively merged from adjacent segments to form a unified representation (Section 3.3). After each merge, we retrain the merged 3DGS model using original training frames, pseudo-view frames from the base models, and interpolated frames from VFI on the combined segments (Section 3.4). This merging and retraining process continues until we obtain the final 3DGS model $S$ , representing the entire sequence $\mathcal{I}$ . Fig. 2 provides an overview of the pipeline. + +# 3.2. Camera pose estimation + +We estimate the camera poses $\mathcal{P}$ by stacking relative camera poses between temporally adjacent pairs of frames. The camera pose of the first frame is set to have no rotation or translation, i.e., $P_{1} = [\mathbb{I}|\mathbf{0}]$ , serving as the reference frame. The estimated poses for all subsequent frames are with respect to this first frame. For each frame pair $(I_i, I_{i+1})$ , we estimate the relative pose $T_{i\rightarrow i+1}$ across all $N-1$ pairs. The camera pose for the $i$ -th frame is the matrix multiplication of the previous relative camera poses: + +$$ +P _ {i} := T _ {1 \rightarrow i} = T _ {i - 1 \rightarrow i} \odot \dots \odot T _ {2 \rightarrow 3} \odot T _ {1 \rightarrow 2}. \tag {1} +$$ + +While stacking relative poses can accumulate error, directly estimating each frame's pose with respect to the first frame is more challenging due to the larger camera displacement. + +Relative pose estimation. As identified in [11], the relative poses between two frames can be approximated by estimating an affine transformation, denoted as $A$ , which is applied to 3D Gaussians from the first frame. After the transformation, the rendered image with respect to the second frame + +should align with the second frame. Specifically, the 2D projection $\mu_{\mathrm{2D}}$ of a 3D Gaussian with position $\mu$ under the pose $P$ is given by $\mu_{\mathrm{2D}} = K(P\mu) / (P\mu)_z$ . This can be approximated by applying an affine transformation $A$ to $\mu$ , followed by a projection using the identity camera pose $[\mathbb{I}|0]$ , yielding $\mu_{\mathrm{2D}} = K(\mathbb{I}\mu') / (\mathbb{I}\mu')_z$ , where $\mu' = A\mu$ . As a result, the relative pose can be estimated from $A$ . + +In practice, to estimate the relative pose from $I_{i}$ to $I_{i + 1}$ , we first construct a single-image 3DGS model $\mathcal{S}_i$ optimized exclusively on $I_{i}$ . We then apply the affine transformation $A$ to each Gaussian in $\mathcal{S}_i$ and render the image $\hat{I}_{i + 1}$ using the camera pose $[\mathbb{I}|\mathbf{0}]$ . The reconstructed $\hat{I}_{i + 1}$ is expected to match $I_{i + 1}$ . We optimize $A$ by minimizing the photometric loss between the rendered image $\hat{I}_{i + 1}$ and the target image $I_{i + 1}$ . During optimization, the attributes of 3D Gaussians in $\mathcal{S}_i$ are fixed, and only $A$ is optimized. The optimized transformation matrix $A$ corresponds to the estimated relative pose $T_{i\rightarrow i + 1}$ . + +Relative pose estimation with video frame interpolation. When the camera movement between adjacent frames is small, the relative pose estimation described above performs well because of sufficient frame overlap. However, with larger camera motions, performance decreases, and the fixed single-image 3DGS model from the previous frame may fail to render a high-quality image $\hat{I}_{i+1}$ . This is because larger camera motions introduce more unseen content that the previous frame may not cover, leading to optimization objectives impacted by artifacts and resulting in poorly estimated poses. For example, the rendered $\hat{I}_{i+1}$ exhibits artifacts in such regions, as shown in Fig. 3d. + +A key insight of our work is compensating for large camera motions with a well-trained video frame interpolation (VFI) model [18]. Let $I_{i + 0.5}$ represent an interpolated frame between $I_{i}$ and $I_{i + 1}$ , where the decimal 0.5 indicates the interpolated result. We estimate the relative poses between $I_{i}$ and $I_{i + 0.5}$ , and between $I_{i + 0.5}$ and $I_{i + 1}$ , separately. The overall relative pose is then given by $T_{i\rightarrow i + 1} = T_{i\rightarrow i + 0.5}\odot T_{i + 0.5\rightarrow i + 1}$ . By reducing the relative camera motion in each step, fewer artifacts are introduced (see Figs. 3e and 3f). + +# 3.3. Hierarchical Training + +After estimating the camera poses $\mathcal{P}$ , the next step is to initialize and grow 3D Gaussians from the input video. A straightforward approach is to initialize the 3D Gaussians using point clouds lifted from the depth map [3, 27] of the first frame, then gradually grow the 3D Gaussians by processing frames sequentially. This initialization, based on a single frame, is incomplete in covering the scene. Moreover, the standard adaptive density control [15], which relies on accumulated gradients to split and clone Gaussians or prunes them based on low opacity, struggles + +![](images/88f9fe82498afa0875548ac8213cbb57ec7c6315e245a1ac5b13a282634bf366.jpg) +Figure 2. Overview of our proposal. We partition the video into multiple segments, train a base 3DGS model on each segment individually, and then iteratively merge these base models into a single, unified 3DGS model representing the entire scene. + +![](images/2be6b0c914798c720e4e6bb8e99d76d0b7bb3d98b2ae9e9a0d4f1e3a181d5e94.jpg) +(a) $I_{i}$ + +![](images/5b9a5b28b7eaa9b5988a0a4994892de08177c80aa110c33e38dce8964f58173b.jpg) +(b) $I_{i + 0.5}$ + +![](images/3b18237f91812a65d3fb8b17d44c00d0d1e7171f0da7a5c2b1a34c94af056dda.jpg) +(c) $I_{i + 1}$ + +![](images/cbe42b23b291d1c33a00b3df8d978be543bb94253d6b000b8330e246b0bfd560.jpg) +(d) $\hat{I}_{i + 1}\mathrm{w / o}$ VFI + +![](images/c58ccd7d6bc4ff15ad809bd8eaa06d81d120c3d63b2fe06851e6ad2adf7fc73d.jpg) +(e) $\hat{I}_{i + 0.5}\mathrm{w / VFI}$ + +![](images/e09f6325dac85532fb80a22f451cda62199b338662fd1c8296a912c3d6bdb712.jpg) +(f) $\hat{I}_{i + 1}\mathrm{w} / \mathrm{VFI}$ +Figure 3. Effect of VFI on relative pose estimation between $I_{i}$ (3a) and $I_{i+1}$ (3c). Fig 3b shows the interpolated frame. In Fig 3d, artifacts are noticeable in regions affected by camera movement, which VFI helps reduce. Fig 3e and 3f show fewer artifacts in the rendered interpolated and original frames. + +in regions with sparse Gaussians distributions. In such areas, Gaussians may have very small gradients, and thus fail to activate the densification process. Consequently, as shown in Fig. 1, areas in the scene not covered in the first frame exhibit a noticeable lack of Gaussians. To overcome this limitation, we propose the hierarchical training strategy. + +Video partitioning. We first partition the $N$ frames into overlapping segments, with the $j$ -th segment denoted as $C_j$ . Ideally, each segment features similar scene content so that the 3DGS model is trained without encountering many unseen regions. To achieve this, we reuse the estimated relative camera poses. Segments $C_j$ and $C_{j+1}$ are separated by locating adjacent image pairs with the largest camera movement. Empirically, we observe that evenly partitioning the video into segments yields similar results. + +Training of base 3DGS models. After partitioning the video into segments, we train the 3DGS model $S_{C_j}$ for each segment $C_j$ . The 3D Gaussians are initialized using a point cloud predicted from the first frame of $C_j$ by depth + +estimators [3, 27]. The Gaussians are then optimized from the first to the last frame in each segment, growing as needed to capture the scene details. + +Merging of base 3DGS models. To represent the entire scene, we merge these base 3DGS models. Consider the merging of two base 3DGS models $S_{C_j}$ and $S_{C_{j+1}}$ . Since each model is optimized with its 3D Gaussian positions aligned to the first frame of its respective segment, it is necessary to align the two models before merging. The alignment is performed by transforming the 3D Gaussians in $S_{C_{j+1}}$ to match those in $S_{C_j}$ , based on the relative pose $T_{C_{j+1,1} \rightarrow C_{j,1}}$ , where $C_{j,1}$ and $C_{j+1,1}$ denote the first frames of their respective segment. + +For merging two 3DGS models, one potential idea is to identify correspondences between 3D Gaussians in the two models and interpolate matched pairs. However, this idea presents several challenges in our task. First, establishing accurate correspondences between Gaussians is non-trivial. For instance, one should not interpolate the Gaussians in the non-overlapping regions in the scene. Yet distinguishing between overlapping and non-overlapping regions can be difficult. Secondly, even with correspondences identified, variations in properties like opacity, scale, covariance, and color among matched Gaussians could complicate the interpolation process [16]. + +Therefore, instead of interpolation, we take a simple yet effective merging strategy: first, prune unimportant 3D Gaussians in each model, then concatenate the remaining Gaussians. Specifically, we first assign an importance score to each 3D Gaussian. Our importance score is inspired by a 3DGS compression strategy [26], which defines the importance of a given parameter $p$ for a 3D Gaussian as: + +$$ +f (p) = \frac {1}{\sum_ {i = 1} ^ {N} H _ {i} W _ {i}} \sum_ {i = 1} ^ {N} \left| \frac {\partial \left(\sum_ {x , y} \hat {I} _ {i} (x , y)\right)}{\partial p} \right|. \tag {2} +$$ + +Above, $\sum_{x,y}\hat{I}_i(x,y)$ represents the sum of RGB values in the rendered image $\hat{I}_i$ , and $H_{i}$ and $W_{i}$ are the height and + +width of image $\hat{I}_i$ . Parameter $p$ is a general notation representing variables such as color $c$ , opacity $\alpha$ , or covariance. The importance score evaluates the sensitivity of the rendering quality with respect to changes to parameter $p$ of a 3D Gaussian. If a minor change in the parameter significantly affects the rendered image, that parameter is considered important. We compute the importance score for each 3D Gaussian in base 3DGS models using Eq 2 and keep the top $\gamma$ percent from base 3DGS models. Finally, we take the union of the selected 3D Gaussians to form the merged representation. + +It is feasible to take the union since 3DGS is an explicit representation. While this approach may introduce redundant Gaussians, we empirically observe that adaptive density control [15] is more effective at pruning redundant Gaussians than at generating new or high-quality Gaussians in sparse regions. If we anchor the model to $S_{C_j}$ , our merging strategy can be viewed as a densification step: it removes less important Gaussians and densifies the representation by adding importance Gaussians from $S_{C_{j + 1}}$ . + +Hierarchical training. After defining the partition and merging strategy, we describe the hierarchical training pipeline that merges multiple base 3DGS models, each optimized for individual segments, into a unified model. First, we define a hierarchical level $L$ and partition the input into $2^{L}$ overlapping segments, resulting in $2^{L}$ base models. We iteratively merge adjacent pairs, reducing the number of models by half in each step, until only one unified model remains. For example, with $L = 2$ , we create four base models: $S_{C_1}$ , $S_{C_2}$ , $S_{C_3}$ , and $S_{C_4}$ . The first merge yields $S_{C_1 \cup C_2}$ and $S_{C_3 \cup C_4}$ , and the final merge gives $S_{1:N} := S_{C_1 \cup C_2 \cup C_3 \cup C_4}$ . Hierarchical training requires all input frames in advance. We also explore an online approach in which we sequentially merge $S_{C_1 \cup C_2}$ with $S_{C_3}$ , then with $S_{C_4}$ , yielding $S_{1:N}$ . We refer to this variant as progressive training. Both strategies improve the baseline by at least 1.32dB (Table 5), with slightly better results for hierarchical training. + +# 3.4. Multi-source supervision + +After merging $S_{C_j}$ and $S_{C_{j+1}}$ and pruning unimportant 3D Gaussians, the newly merged representation needs to be retrained. Simply retraining on the set of images from $C_j \cup C_{j+1}$ leads to overfitting on those specific images. To address this, we propose to augment training with two additional sets of images: (1) from the base 3DGS models $S_{C_j}$ and $S_{C_{j+1}}$ before merging, and (2) from the interpolated frames (see Section 3.2). + +Supervision from base 3DGS models. As the base 3DGS models are better-optimized for their respective segments, novel pseudo-views rendered from these models can serve + +as a source of training images for the merged 3DGS. To generate novel views, we first sample a virtual camera pose $P_{i + \tau}$ between two poses, $P_{i}$ and $P_{i + 1}$ , using the formula: + +$$ +P _ {i + \tau} = P _ {i} \exp \left(\tau \log \left(P _ {i} ^ {- 1} \cdot P _ {i + 1}\right)\right), \tag {3} +$$ + +where $\tau \in (0,1)$ and $P_{i}, P_{i+1} \in \mathrm{SE}(3)$ represent camera poses in the $\mathrm{SE}(3)$ space. This smooth interpolation enables the creation of pseudo-views, which are then rendered as additional supervision for the merged 3DGS model. + +Supervision from video frame interpolation. The interpolated images from VFI are of sufficiently high quality at viewpoints not covered by the training frames, making them suitable for supervision. For instance, Fig. 3b shows an example of an interpolated frame. To supervise the merged 3DGS with the interpolated frames, we need to estimate their corresponding camera pose. For an interpolated frame $I_{i+0.5}$ between frames $I_i$ and $I_{i+1}$ , the camera pose $P_{i+0.5}$ is computed as $P_{i+0.5} = T_{i\rightarrow i+0.5}\odot \dots \odot T_{1.5\rightarrow 2}\odot T_{1\rightarrow 1.5}$ . Each relative pose was previously calculated in Section 3.2, so no additional computational overhead is required. + +Loss function. We optimize the 3DGS model using the photometric loss between the rendered image and the training image or the pseudo image [15]: + +$$ +\mathcal {L} = (1 - \lambda) \mathcal {L} _ {1} + \lambda \mathcal {L} _ {\mathrm {D} - \mathrm {S S I M}}, \tag {4} +$$ + +where $\mathcal{L}_1$ is the L1 loss, and $\mathcal{L}_{\mathrm{D - SSIM}}$ is the D-SSIM term. + +# 4. Experiment + +# 4.1. Experimental setting + +Datasets. We conduct experiments on the Tanks and Temples dataset [17] and the CO3D-V2 dataset [28]. For Tanks and Temples, following [4, 11], we sequentially divide the frames into groups of 8, using one of the frames as the test frame and the remaining frames for training. For the Family scene on Tanks and Temples, we alternate images, using every other image as a test image. The CO3D-V2 dataset is more challenging due to its larger camera motions. We adopt the same sampling strategy, using every eighth frame for testing and the rest for training. + +Metrics. We use PSNR, SSIM [31], and LPIPS [38] to evaluate the effectiveness of novel view synthesis. For camera pose estimation, we report the Absolute Trajectory Error (ATE) and Relative Pose Error (RPE), similar to [4, 11]. ATE measures the difference of the camera positions. RPE measures the relative pose errors, containing relative rotation error $(\mathrm{RPE}_r)$ and relative translation error $(\mathrm{RPE}_t)$ . Additionally, we report the memory size required to store the optimized parameters. + +Implementation details. The hierarchical training level is set to $L = 2$ as it reaches the saturation point (see Table 4). + +
ScenesBARF [21]SC-NeRF [13]Nope-NeRF [4]CF-3DGS [11]Ours
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Church23.170.620.5221.960.600.5325.170.730.3930.230.930.1131.340.940.08
Barn25.280.640.4823.260.620.5126.350.690.4431.230.900.1034.950.970.05
Museum23.580.610.5524.940.690.4526.770.760.3529.910.910.1131.590.950.08
Family23.040.610.5622.600.630.5126.010.740.4131.270.940.0734.710.970.05
Horse24.090.720.4125.230.760.3727.640.840.2633.940.960.0535.820.980.03
Ballroom20.660.500.6022.640.610.4825.330.720.3832.470.960.0734.120.970.04
Francis25.850.690.5726.460.730.4929.480.800.3832.720.910.1434.090.930.13
Ignatius21.780.470.6023.000.550.5323.960.610.4728.430.900.0931.640.950.06
Mean23.420.610.5423.760.650.4826.340.740.3931.280.930.0933.530.960.07
+ +Table 1. Novel view synthesis results on Tanks and Temples [17]. We achieve the best performance among all competitors. + +![](images/2c638c1771b29a3a43dc84c104afd04b26cc6a5f860b5a2f00f500c67e26b13b.jpg) + +![](images/b78cd7c3fa4ca2b2d33b525e3d5bdac895161c2a0014bd6678e89977f4296e97.jpg) +Nope-NeRF + +![](images/3fc9c63eefc8f09e8b53f855d2e7dc494c119e0665543df124a8d39fc6f41cbb.jpg) +CF-3DGS + +![](images/0c8cc552ee98a5aefcf642520673018f8c1813f5eb8bca11ae26eadc97aa9a61.jpg) +Ours + +![](images/258cbac882e1b14c8e27f161d6f98e58fb2b17f17dc833cd4dd8a961f8257c1e.jpg) +Figure 4. Qualitative novel view synthesis results on Tanks and Temples [17]. Our proposal achieves superior rendering quality. + +![](images/78d2c605f3f7c6c550f5283efbb532d99ca063cbd15f061d408b760ee3c898d0.jpg) +Nope-NeRF + +![](images/2f8f6a16348018a81005fea7f8af2b31111a9a347ae45c7baa642313e6ff4bc3.jpg) +CF-3DGS + +![](images/439e49de60f7f87ec53aa31045ccb7f6ddb294d4b1f459f80635c310030da636.jpg) +Ours + +Each segment's base 3DGS model is trained from start to end frame with 300 iterations per frame. Multi-source supervision also employs 300 iterations per frame. During the merging, we select the top $50\%$ of 3D Gaussians from each model based on importance score. Multi-source supervision involves two steps: first, using pseudo-view images from base models and original frames; second, using interpolated frames and original frames, with a $50\%$ probability of selecting pseudo-view or interpolated frame for each step. Gaussians are grown and pruned every 100 and 2000 iterations on the Tanks and Temples and CO3D-V2 datasets, respectively. All experiments are conducted on a single RTX A5000 GPU. + +# 4.2. Evaluation on the Tanks and Temples Dataset + +Quantitative comparison. We perform the comparison with the state-of-the-art novel view synthesis methods without SfM preprocessing, including BARF [21], SC-NeRF [13], Nope-NeRF [4] and CF-3DGS [11] on Tanks and Temples [17]. As shown in Table 1, our method achieves superior performance compared to all of them. Specifically, compared to CF-3DGS, we improve the av + +erage PSNR by 2.25 dB, SSIM by 0.03, and reduce LPIPS by 0.02. The most significant improvement is observed in the Barn scene, with a PSNR increase of 3.72 dB, SSIM improvement of 0.07, and LPIPS reduction of 0.05. + +Qualitative comparison. Fig. 4 shows that we achieve finer detail and higher fidelity, especially in highly detailed regions where CF-3DGS struggles to grow 3D Gaussians, highlighting the advantages of our training strategy. + +# 4.3. Evaluation on the CO3D-V2 Dataset + +Quantitative comparison. Due to the challenges posed by this dataset, we limit our comparison to the most advanced methods, namely Nope-NeRF [4] and CF-3DGS [11]. Table 2 shows that our method outperforms these approaches, with an average PSNR boost of $1.74\mathrm{dB}$ . The most significant improvement is seen in the scene 34_1403_4393 (teddybear), where we increase the PSNR by over $3.90\mathrm{dB}$ , improve SSIM by 0.07, and reduce LPIPS by 0.06. + +Qualitative comparison. Fig. 5 demonstrates that even with challenging input videos, our method maintains high performance compared to CF-3DGS, which exhibits blur and unrealistic red artifacts in the 34_1403_4393 (teddy + +
Method34_1403_4393106_12648_23157110_13051_23361245_26182_52130415_57112_110099
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nope-NeRF [4]28.620.800.3520.410.460.5826.860.730.4725.050.800.4924.780.640.55
CF-3DGS [11]27.750.860.2022.140.640.3429.690.890.2927.240.850.3026.210.730.32
Ours32.520.930.1423.430.730.2829.950.870.1928.590.870.2727.230.780.30
+ +Table 2. Novel view synthesis results on CO3D-V2 [28]. We achieve superior performance over all competitors, with the largest improvement on 34_1403_4393, where our method increases PSNR by 3.90 dB compared to Nope-NeRF. + +![](images/4ec299d9790dbe2e70ba2736b984ac45fb70d83fcba889ed40d7aed5ed52613a.jpg) + +![](images/d31b55d4275dee63b3c4be8fa071f24e4a3129d68d5a7c061f6fd1573a3be6a5.jpg) +CF-3DGS + +![](images/ade1964c2106c0d1e7e2fa9dda0e7c73496626101924ee77ee2ee536b5cb075b.jpg) +Ours + +![](images/36a2cf63b185673b2a0e1b546f3011253714cafd605e4d770c735d43dc9ece17.jpg) +Figure 5. Qualitative novel view synthesis results on CO3D-V2 [28]. Our proposal achieves superior rendering quality. + +![](images/99f889062307025708b038df03e116a8cc02a3c365ae90a829749751712f01bc.jpg) +CF-3DGS + +![](images/a199bbec112a4ee7b7d64599652edfd60c9ddf4b1bb6faa141fade1cc437b779.jpg) +Ours + +![](images/fdb02104fff763e847d2068be59339900d6d12e0c51d4772fdd84b4beed0c760.jpg) + +![](images/da5ab157e415747addf898faa949b3667c57471b6ed20d1e4fb1b41573aafb37.jpg) +CF-3DGS + +![](images/6f7b7d7a62b87d1929d62cbe2152bd4a3d3bd41949993089525038394e691534.jpg) +Ours + +![](images/b893ea09dfe8fc214752db13fff8381fb18d3a845352ac3aecf77fcf3fefb374.jpg) + +![](images/ccc5f1a31716ef60004b85dbebd7a4913fa8bb3510d5d2e56aee21e951b8c084.jpg) +CF-3DGS + +![](images/147ff07c9db274e34527346955ba436e6bd52f666171047e6653c25292729caf.jpg) +Ours + +bear) scene due to suboptimal 3D Gaussian learning. + +Camera pose estimation. We conduct camera pose estimation comparisons only on CO3D-V2, as it provides ground-truth camera poses. As shown in Table 3, our method achieves comparable or superior performance to all competitors, reducing $\mathrm{RPE}_t$ and $\mathrm{RPE}_r$ by up to 0.464 and 0.078, respectively. Our approach's improvements on ATE are less consistent. We hypothesize that this is due to errors in the interpolated images generated by the VFI model, to which camera position estimation is particularly sensitive. + +# 4.4. Ablation study + +We conduct an ablation study on the Tanks and Temples dataset [17]. Table 5 reports the average PSNR, SSIM, and LPIPS across all scenes. The baseline generally follows CF-3DGS [11], with adjustments to certain hyperparameters to align with our strategy. + +Progressive vs. hierarchical training. Table 5 shows that hierarchical training (HT) outperforms progressive training (PT) by $0.2\mathrm{dB}$ (Variant 6 vs. 7), with both improving PSNR over the baseline (Variant 3) by more than $1.32\mathrm{dB}$ . This supports our claim that merging 3D Gaussians from different base models enhances results. HT performs better + +by balancing training across early and late frames, whereas PT, which allocates more iterations to early frames, tends to overfit them. However, PT is suitable for online tasks, unlike HT. Moreover, to test if the performance gains result from additional training iterations, we retrain the baseline model on the entire input (referred to as global training, GT). GT shows minimal improvement, indicating that retraining alone does not improve the performance. + +Hierarchical training level. We evaluate the effectiveness of hierarchical training at various levels $L$ , with results in Table 4. At $L = 0$ , the 3DGS model is trained by treating the entire input as a single segment. Our strategy notably boosts PSNR by 1.19-1.58 dB and SSIM by 0.02, while reducing LPIPS by 0.01-0.02. Performance increases with higher levels, saturating at $L = 2$ , where the video is divided into four segments, sufficient for 3D Gaussian learning on the about 150-frame Tanks and Temples dataset. Our proposal also lowers memory storage by 0.22-0.26 GB. While it may seem counterintuitive, hierarchical training reduces memory storage for two reasons: (1) pruning unimportant Gaussians before merging, maintaining a stable count; (2) optimized segment-wise Gaussians are more representative, unlike the baseline model, which redundantly clones and + +
Method34_1403_4393106_12648_23157110_13051_23361245_26182_52130415_57112_110099
RPEt ↓RPEr ↓ATE ↓RPEt ↓RPEr ↓ATE ↓RPEt ↓RPEr ↓ATE ↓RPEt ↓RPEr ↓ATE ↓RPEt ↓RPEr ↓ATE ↓
Nope-NeRF [4]0.5911.3130.0530.3871.3120.0490.4001.9660.0460.5871.8670.0380.3261.9190.054
CF-3DGS [11]0.5050.2110.0090.0940.3600.0080.1400.4010.0210.2390.4720.0170.1100.4240.014
Ours0.0410.1700.0090.0450.2820.0140.0930.3310.0200.0640.4380.0170.0490.3510.024
+ +Table 3. Camera pose estimation results on CO3D-V2 [28]. We achieve significant improvements in $\mathrm{RPE}_t$ and $\mathrm{RPE}_r$ , with a slight decrease in performance on ATE. We hypothesize that ATE is more sensitive to errors in the interpolated frames from VFI. + +
ScenesL=0L=1L=2L=3
PSNR ↑SSIM ↑LPIPS ↓Mem ↓PSNR ↑SSIM ↑LPIPS ↓Mem ↓PSNR ↑SSIM ↑LPIPS ↓Mem ↓PSNR ↑SSIM ↑LPIPS ↓Mem ↓
Church30.440.930.091.0631.400.940.080.8431.670.950.080.8531.610.940.080.83
Barn30.090.880.111.4432.060.920.081.4332.270.920.081.4132.200.920.081.38
Museum30.240.910.101.3530.950.940.081.1731.750.940.071.1031.970.950.071.09
Family33.120.960.051.4433.710.960.051.1834.200.970.051.2134.150.970.051.18
Horse34.080.960.051.0735.440.980.040.9635.440.980.040.9035.540.980.030.92
Ballroom32.820.960.051.3933.170.970.051.1433.410.970.041.1433.670.970.041.13
Francis32.840.920.140.8133.640.920.130.6833.660.920.130.7533.620.920.130.64
Ignatius28.370.910.092.0931.150.940.061.5031.780.940.061.4331.900.950.061.40
Mean31.500.930.091.3332.690.950.081.1133.020.950.071.1033.080.950.071.07
+ +Table 4. Ablation study of the hierarchical training level. Memory storage is measured in gigabytes. + +
IdVariantPSNR ↑SSIM ↑LPIPS ↓
(1)CF-3DGS31.280.930.09
(2)CF-3DGS + VFI31.450.930.09
(3)Baseline31.500.930.09
(4)+ Global Training (GT)31.520.930.08
(5)+ GT + VFI31.950.930.08
(6)+ Progressive Training32.820.950.07
(7)+ Hierarchical Training (HT)33.020.950.07
(8)+ HT + VFI33.370.950.07
(9)+ HT + VFI + Base (Ours)33.530.960.07
(10)Ours w/o intrinsic32.170.940.09
(11)NeRF + SfM26.610.750.38
(12)3DGS + SfM30.200.920.10
+ +Table 5. Ablation study on Tanks and Temples [17]. Incorporating all components yields the best performance. + +splits Gaussians in regions with sparse Gaussians. + +Effectiveness of video frame interpolation (VFI). We evaluate the impact of VFI in Table 5. VFI smooths camera motion and provides additional supervision, resulting in an average PSNR gain of $0.17\mathrm{dB}$ (Variant 1 vs. 2), $0.43\mathrm{dB}$ (Variant 4 vs.5) $0.35\mathrm{dB}$ (Variant 7 vs.8). VFI improves various 3DGS models, including CF-3DGS, beyond our hierarchical training. Moreover, hierarchical training contributes a PSNR gain exceeding $1.52\mathrm{dB}$ , showing that the core improvement stems from hierarchical training rather than augmented data from VFI. + +Effectiveness of supervision from base 3DGS models. Table 5 shows that the supervision from base 3DGS models enhances performance, with an average PSNR increase of $0.16\mathrm{~dB}$ and an SSIM increase of 0.01 (Variant 8 vs. 9). Pseudo-views generated by base 3DGS models mitigate overfitting and provide additional supervision for views not covered by the training images. + +Unknown camera intrinsics. We experiment with heuris- + +tics instead of known camera intrinsics by setting the FoV to $70^{\circ}$ . As shown in Table 5, PSNR dropped from 33.53dB to 32.17dB (Variant 9 vs. 10). Inaccurate intrinsics hinder pose estimation and may introduce scale ambiguity. Nonetheless, our method, even without known intrinsics, outperforms CF-3DGS by 0.89dB (Variant 1 vs. 10). + +Comparison to 3DGS and NeRF with SfM poses. Variants 11 and 12 demonstrate that incorporating SfM results in lower performance for both NeRF and 3DGS compared to our method, with reductions of $6.92\mathrm{~dB}$ and $3.33\mathrm{~dB}$ , respectively. This performance gap arises from COLMAP's challenges in accurately estimating poses in low-texture environments, such as those found in Tanks & Temples. + +# 5. Conclusion + +We propose a hierarchical training strategy for 3D Gaussian splatting without known camera poses or SfM preprocessing, merging segment-specific base 3DGS models for enhanced representation. We further incorporate video frame interpolation to smooth camera motion and mitigate overfitting by reusing interpolated images and base models. This approach outperforms state-of-the-art SfM-free novel view synthesis methods, enabling broader generalization across datasets without SfM preprocessing. + +Limitations. Our approach requires longer training and can face challenges with large camera motion or low-quality inputs. While training time increases, rendering is faster due to fewer 3D Gaussians. In practice, training time can be reduced by lowering iterations or removing VFI, which is less necessary with small camera motion or abundant input frames. Large motion or poor inputs may cause alignment errors in 3DGS model merging. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 2 +[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 3, 4 +[4] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 1, 2, 5, 6, 7, 8 +[5] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. arXiv preprint arXiv:2404.06109, 2024. 2 +[6] Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs. arXiv preprint arXiv:2306.05410, 2023. 2 +[7] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Garf: Gaussian activated radiance fields for high fidelity reconstruction and pose estimation. arXiv eprints, 2022. 2 +[8] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. arXiv preprint arXiv:2402.03307, 2024. 2 +[9] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds, 2024. 2, 3 +[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2 +[11] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 1, 2, 3, 5, 6, 7, 8 +[12] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19774–19783, 2023. 2 + +[13] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In ICCV, 2021. 6 +[14] Kaiwen Jiang, Yang Fu, Mukund Varma T, Yash Belhe, Xiaolong Wang, Hao Su, and Ravi Ramamoorthi. A construct-optimize approach to sparse view synthesis without camera pose. SIGGRAPH, 2024. 2, 3 +[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 5 +[16] Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time rendering of very large datasets. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. 2, 4 +[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 2017. 2, 5, 6, 7, 8 +[18] Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, and Jie Yang. Ifrnet: Intermediate feature refine network for efficient frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1969-1978, 2022. 2, 3 +[19] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719-21728, 2024. 2 +[20] Hao Li, Yuanyuan Gao, Chenming Wu, Dingwen Zhang, Yalun Dai, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Junwei Han. Ggrt: Towards pose-free generalizable 3d gaussian splatting in real-time. CoRR, 2024. 2, 3 +[21] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 2, 6 +[22] Yang Liu, He Guan, Chuanchen Luo, Lue Fan, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. arXiv preprint arXiv:2404.01133, 2024. 2 +[23] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16539-16548, 2023. 2 +[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2 +[25] Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3d scene representation via self-organizing gaussian grids. arXiv preprint arXiv:2312.13299, 2023. 2 +[26] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3d gaussian splatting for accelerated + +novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10349-10358, 2024. 2, 4 +[27] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 3, 4 +[28] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 2, 5, 7, 8 +[29] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1 +[30] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Pérez-Pellitero. Swings: Sliding windows for dynamic 3d gaussian splatting. 2 +[31] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 2004. 5 +[32] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2 +[33] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2 +[34] Yitong Xia, Hao Tang, Radu Timofte, and Luc Van Gool. Sinerf: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. 2022. 2 +[35] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. 2 +[36] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. In IROS, 2021. 2 +[37] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2 +[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[39] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 2 \ No newline at end of file diff --git a/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/images.zip b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..01e455d4667f43268221f14081fbf4bce8155254 --- /dev/null +++ b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41d8e5fe9da98dedfd2ffabb68b810bc0640cabd8639a8dc330f2ef76264b0cd +size 774455 diff --git a/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/layout.json b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fbda5793df87741e45197d2e59cdd5dbfbee4a92 --- /dev/null +++ b/CVPR/2025/SfM-Free 3D Gaussian Splatting via Hierarchical Training/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bf9f1937e0118af641808b5a1d4fc8162202e6d3b7f3912828dc28bea0683a5 +size 476670 diff --git a/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/86916c80-f47e-47cb-9eef-47dac43d7bb0_content_list.json b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/86916c80-f47e-47cb-9eef-47dac43d7bb0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cd4d41fd2d530e9ba9f50a3df0e8cef70edd3fa0 --- /dev/null +++ b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/86916c80-f47e-47cb-9eef-47dac43d7bb0_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e03c535f59b690d62c348b0f98a35c231635a8ac72b939d6398ef9c3bdd4d06 +size 98504 diff --git a/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/86916c80-f47e-47cb-9eef-47dac43d7bb0_model.json b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/86916c80-f47e-47cb-9eef-47dac43d7bb0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7fff29831282e41bdbfea664666104bab26fea67 --- /dev/null +++ b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/86916c80-f47e-47cb-9eef-47dac43d7bb0_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f110b57ba411bdc87893c32d9dc2fa8f23bdf799d1813ae9257dc6cd3dfd51f +size 125730 diff --git a/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/86916c80-f47e-47cb-9eef-47dac43d7bb0_origin.pdf b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/86916c80-f47e-47cb-9eef-47dac43d7bb0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5bf732648851c0854e7ca84214f828a498e216a7 --- /dev/null +++ b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/86916c80-f47e-47cb-9eef-47dac43d7bb0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50f9bb99d397cc704088bde532e3cd187d06781c647cb58ea02f4dcdb99c7be7 +size 2195532 diff --git a/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/full.md b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5b3fdbda658e6ea7f00bc4fe5a5596a7b6bb4f86 --- /dev/null +++ b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/full.md @@ -0,0 +1,420 @@ +# Shading Meets Motion: Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion + +Guoyu Lu +Intelligent Vision and Sensing Lab +Binghamton University +guoyulu62@gmail.com + +# Abstract + +Scene reconstruction has a wide range of applications in computer vision and robotics. To build practical constraints and feature correspondences, rich textures and distinguished gradient variations are particularly required in classic and learning-based SfM. When building low-texture regions with repeated patterns, especially mostly-white indoor rooms, there is a significant drop in performance. In this work, we propose Shading-SfM-Net, a novel framework for simultaneously learning a shape-from-shading network based on the inverse rendering constraint and a structure-from-motion framework based on warped keypoint, room layout, and geometric consistency, to improve structure-from-motion and surface reconstruction for low-texture indoor scenes. Shading-SfM-Net tightly incorporates the surface shape consistency and 3D geometric registration loss in order to dig into their mutual information and further overcome the instability on flat regions. We evaluate the proposed framework on texture-less indoor scenes (NYUv2 and ScanNet), and show that by simultaneously learning shading, motion and shape, our pipeline is able to achieve state-of-the-art performance with superior generalization capability for unseen texture-less datasets. + +# 1. Introduction + +For robotics in the indoor environment, the ability to understand its surrounding scenes is of utmost importance for the facilitation of the tasks, such as scene understanding and navigation [4] [51] [52] [90] [14] [3] [32]. Over the years, scene reconstruction has become a key area of research in 3D computer vision and robot vision, leading to considerable improvement. For common outdoor environments, direct geometry-based SfM approaches are usually reliable and accurate under favorable conditions, such as sufficient lighting illumination and rich texture information to establish adequate and accurate correspondences [15]. For an indoor environment, minimum changes in room layout are required to establish stable correspondences on textured re + +gions. Recent deep learning-based algorithms on monocular video sequences [89] [93] [77] [41] [81] [2] [10] [92] [85] [65] [18] [47] [49] have achieved notable improvement on outdoor driving datasets such as KITTI, Cityscapes, etc, to facilitate autonomous driving tasks [30] [50] [46]. However, improving prediction accuracy on low-texture regions is still a challenge. For example, a majority of the currently available methods, e.g., [89] [77], usually flop on textureless indoor datasets such as on NYUv2 or ScanNet. With those regions (e.g., walls and floors) frequently containing a large proportion of few textures or repeated patterns, it always fails to apply photometric constraints continuously. In those regions, the estimated inverse depths are generally infinite or random. However, the photometric error will remain close to zero. If depth estimation is inaccurate, it will cause a dramatic performance decrease in the camera motion estimation, resulting in an ineffective training process of the framework. + +Existing 3D reconstruction methods mainly rely on textures to build correspondences. To improve the results of state-of-the-art self-supervised approaches in low-textured indoor scenes such as walls, floors, roofs and curtains, we propose a novel framework to simultaneously learn shapefrom-shading (SfS) and structure-from-motion (SfM). As SfM and SfS reconstruct 3D scenes from different perspectives and principles, SfM-Net performs well for textured objects and scenes since prominent structures facilitate pixel matching [34]. However, it suffers from an absence of fine surface details and tends to be over-smooth. SfS-Net can help produce a detailed reconstruction. However, the decomposition of the observation to intrinsic images is ambiguous and thus hard to be directly applied in real applications. To tackle this issue, we utilize the shading information to optimize the scene shape estimated from the SfM network with surface shape and scene registration consistencies. The surface shape consistency benefits from the albedo characteristic to constrain depth continuity on the curved surfaces of the albedo's segments, and scene registration consistency minimizes the 3D distance between the + +![](images/2e7f4a198b48b353d67eb27ca7970ebbc3b4fe95c91cf2e0903ea5165dc129c1.jpg) +Figure 1. Overview of the proposed framework. We propose to train the shading network and structure-from-motion network simultaneously to overcome the potential failure on low-texture scenes. Blue arrows and lines represent the data flow and predictions; Green lines mark the loss constraints for guiding the learning process. + +transformed point cloud of the previous frame and the point cloud of the current frame in the SfM network, as well as the distance between the group of points from the two networks at the same time. The consistencies can effectively avoid spatial and spectral confusion due to limited textures and occurrence of repeated patterns in input color images. + +The main contributions are summarized as follows: i) we propose Shading-SfM-Net. To our best knowledge, this is the first framework that leverages the benefits from both SfS and SfM networks simultaneously, to aid scene reconstruction in texture-less indoor scenes; ii) the designed network introduces an adaptive keypoint and layout-based consistency and a curved surface consistency between Albedo and depths, which firmly constrain the learning process on low and non-textured regions; iii) the 3D point clouds from SfM at different time are further optimized by registering with the point cloud from shape-from-shading network, to build a 3D geometric registration loss, which exploits their mutual information and overcomes the spatial confusion of SfM systems in low or no texture regions of challenging indoor images; iv) the unsupervised manner in which the entire framework is designed mitigates the training requirement while enhancing the robust generalization capability. + +# 2. Related Work + +Supervised depth estimation 2.5D prediction from a single image has been intensively studied in the past. Many of them recently were designing different Convolutional Neural Networks (CNN) to estimate dense depth by regress- + +ing annotated data. Eigen et al. [17] proposed a multiscale network for performing coarse depth estimation, and refining the initial estimation. Liu et al. [44] propagated single-view image depth estimation based on a deep CNN together with conditional random field (CRF). Fu et al. [21] suggested a deep ordinal regression network and spacing-increasing discretization to address this issue. Recent works mostly focus on combining depth estimation with other relevant tasks to enhance performance. For instance, as observed in surface normal prediction [60] [76] [84] [61], semantic segmentation [83] [84] and contour detection [61]. As supervised methods for structure-from-motion, Wang et al. [71] proposed a recurrent convolutional neural network to learn monocular SfM from videos. Ummenhofer et al. [67] and Teed et al. [66] proposed to learn depth and motion together in an end-to-end way by treating structure from motion as a supervised problem. Despite having delivered a progressive performance, these methods rely heavily on huge amounts of labeled data for training. This makes them cost-effective, pertaining to increased human labor and hardware costs. Supervised methods are also sensitive to data distribution, making them hard to be applied in unseen scenes during real applications. + +Unsupervised depth estimation Unsupervised estimation of monocular depth started from training stereo pairs and has been explored to avoid large-scale annotated data [13]. After the depth is estimated from the network, the photometric error for image reconstruction is minimized [22] + +[23] [55] [59] [24] [53] [27] [70]. However, these methods are limited by binocular settings. Other works propose to jointly learn depth and ego motion from only monocular sequences [12] [68] [39] [82] [5] [48]. Zhou et al. [89] first introduced SfMLearner to estimate depth map from a depth CNN and ego-motion from pose CNN that used unstructured video sequences. Following works utilize generative adversarial networks to improve it [58] [1] [2] [20]. UnDeepVO [41] and Depth-VOFeat [81] applied stereo video sequences during the training to solve the scale ambiguity, while SC-SfMLearner [10] dealt with the scale inconsistency by learning a global depth consistency. Recent works replaced standard convolutional encoder-decoder structures with 3D packing and unpacking block [25] and self-attention context module [31] to improve the depth estimation performance. Another types of works put more efforts on joint learning with other tasks such as optical flow [93] [77], key-frame detection [64] and uncertainty [74]. However, almost all these works focus on outdoor driving scenes such as KITTI, Cityscapes, etc. The reason is that, these datasets have diverse objects and rich features, learning could always be easier than low-texture challenging scenes like indoor scenes. Moving Indoor proposed by Zhou et al. [88] was the first work that targeted this problem by introducing rigid flow as a supervisory signal. Yu et al. [79] encodes the piece-wise planar prior into their unsupervised depth estimation network for depth estimation. Bian et al. [11] proposed a weak image rectification strategy when processing data, to increase the effectiveness of depth estimation. However, the performance of such works still heavily relies on the learning of accurate feature matching, and shading information is ignored. + +Intrinsic image decomposition Intrinsic image decomposition splits an image into a shading map and an albedo map. Previous works assessed the local image deviations both in shading maps and albedo using several different constraints and assumptions to reduce unknown variables [63] [72] [6] [86]. Given the success of deep neural networks, recent works could produce high-quality decomposition [19] [9] [36]. Fan et al. [19], Bi et al. [9] and Lettry et al. [36] all apply supervised constraints on synthetic dataset. Very few works attempted self-supervised learning on this topic, e.g., [42] [45]. Li et al. [42] explored single image decomposition under varying illumination. Liu et al. [45] suggested learning the albedo and shading maps latent feature from unlabeled and uncorrelated data. Inverse rendering estimates one or more of illumination, albedo and shape simultaneously over varying appearance. Recent works mostly performed inverse rendering constraints on faces [75] and other objects rather than scenes. Janner et al. [29] predicted reflectance, shape, and lighting conditions given a single image and apply a learning model to recompose the original image with strong supervised sig + +nals. SfSNet [62] was proposed to decompose human face images under diverse illuminations, and learn from a mixture of labeled synthetic data and unlabeled real images together during training. Yu et al. [78] proposed an inverse rendering based framework InverseRenderNet for uncontrolled outdoor images by using multi-view stereo as prior for training. Different from the aforementioned works, we jointly leverage the benefits from both inverse rendering of shading and structure-from-motion framework, to prevent uncertainty and wrong prediction in non-textured or low-textured regions without any priors as supervision. + +# 3. Shading Constrained SfM Framework + +Our model is an end-to-end deep learning framework that takes monocular image sequences as inputs and simultaneously outputs dense depths, camera motion, albedo and shading maps. As illustrated in Fig. 1, the entire network mainly consists of two pathways: the shape-from-shading pathway based on inverse rendering relationship and structure-from-motion pathway based on warped keypoints, layout and geometric depth consistency. The shape-from-shading pathway takes each single frame as input to estimate albedo and normal maps directly. The estimated albedo is then formulated with the depth maps from the SfM pathway to build the cross-surface consistency, and the shading map and light source are also inferred from the estimated albedo and normal maps. The SfM pathway takes image sequences as input, to output depth maps and egomotion together. The estimated depth maps are then converted to 3D maps to build the scene registration consistency. The two pathways are learned and promoted with each other to produce a more accurate and denser 3D map on challenging texture-less scenes by a large margin. + +# 3.1. Shading Navigated Learning Scheme + +In a single image, shape-from-shading relies more on shading information to generate surface shapes, which is widely applied in astronomy, e.g., planet surface reconstruction. Accurate feature matching for indoor scenes and unstructured outdoor scenes, such as forests, mountains and grassland, remains to be a challenge to accomplish. Therefore, we introduce a deep learning-based approach for inverse rendering from a single image to support shape reconstruction on texture-less regions, which is shown in Fig. 1. Considering a scene as a Lambertian surface, the designated network takes an image as the input of the encoder, and learns the representation to generate the estimated albedo map $A$ and surface normal map $N$ . The shading map $S$ is obtained from albedo $A$ and input $I$ . The light source $L$ can also be inferred via the illumination model from $N$ and $S$ by computing $(I / A)f^{-1}(N)$ . $f^{-1}$ is the pseudo inverse. Our SfS-Net can be denoted as a mapping function $\psi(\cdot)$ : + +$$ +S f S - N e t: \psi (I) \rightarrow \{A, N, (S), (L) \} \tag {1} +$$ + +With the estimated Albedo $A$ , the estimated surface normal $N$ , the inferred shading $S$ and light source $L$ , the reconstructed image $I_{rec}$ can be rendered by $I_{rec} = A \cdot S = A \cdot Lf(N)$ , where $I_{rec}$ is stacked linearly on RGB pixel values to a matrix of $R^3$ . $A \in R^{3 \times l}$ and $N \in R^3$ are the stacked Albedo map and normal map formed in the same way. $f$ represents the simplified rendering equation under the 2nd order basis: $[1, n_x, n_y, n_z, 3n_z^2 - 1, n_x \times n_y, n_x \times n_z, n_y \times n_z, n_x^2 - n_y^2]^T$ , where $\cdot$ denotes the Hadamard (elementwise) multiplication operator. + +The self-supervision reconstruction loss $L_{shading\_rec}$ thus is designed to promote the consistency between the reconstructed image $I_{rec}$ and the raw image $I_{raw}$ in CIELAB color space, which is demonstrated to better process perceptual uniformity and outperform RGB in image reconstruction [28] [7]: + +$$ +L _ {\text {s h a d i n g} - \text {r e c}} = \| L * A * B * \left(I _ {\text {r a w}}\right) - L * A * B * \left(I _ {\text {r e c}}\right) \| _ {1} \tag {2} +$$ + +In the CIELAB color space, the dimensions L, A, and B quantify objective color and facilitate the computation of color differences. The L dimension scales lightness from black to white, ranging from zero to 100. Conversely, A and B dimensions denote chromaticity and do not adhere to a fixed numerical range. + +# 3.2. Adaptive Brightness and Scale-invariant SfM + +In order to learn dense depth and relative pose jointly on unlabeled monocular video sequences, we propose to use keypoint and brightness based consistencies as supervisions to guide the learning. Given a pixel point $p = (x,y)$ and the corresponding estimated depth value $d(p)$ , we build the inverse projection mapping to reconstruct the pixel coordinated point to 3D space as: + +$$ +\pi^ {- 1} (p, D (p)) = \left(\left(x - c _ {x}\right) D / f _ {x}, \left(y - c _ {y}\right) D / f _ {y}, D, 1\right) \tag {3} +$$ + +where $f_{x}, f_{y}, c_{x}, c_{y}$ represents focal length and principal point in $x$ and $y$ coordinates, respectively. The projection mapping can be obtained as $\pi(\cdot)$ . + +Provided two adjacent frames $I_{t}$ and $I_{t + 1}$ , the depth network predicts the corresponding pixel-level depth maps $(D_{t}, D_{t + 1})$ with shared weights. We can use the relative pose vector $T_{t\rightarrow t + 1}$ output from the pose estimation network to infer the warping transform operation as: + +$$ +\omega (p, T _ {t \rightarrow t + 1}) = \pi \left(T _ {t \rightarrow t + 1} \pi^ {- 1} (p, D (p))\right) \tag {4} +$$ + +The goal of the unsupervised loss is to synthesize $I_{t}$ with the warping transform operation $\omega$ by generating the synthesized image $I_{t - 1}(\omega (p,T_{t\rightarrow t + 1}))$ . Different from [89] [24], we compute dense depths on both source and reference images with a shared weight instead of on source images only. The initial photometric spectral consistency loss on appearance across temporal frames is formulated as: + +$$ +L _ {\text {s p e c t r a l}} = \frac {1}{N} \sum_ {i j} \left| I _ {t + 1} \left(\omega \left(p (i, j), T _ {t \rightarrow t + 1}\right)\right) - I _ {t} \left(p (i, j)\right)\right| \tag {5} +$$ + +Most existing works apply photometric consistency on all image pixels $I(p)$ . We deployed a different strategy and paid attention only to the extracted BRISK [38] keypoints $\mathbf{I}(\mathbf{Kp})$ to prevent the negative impact of the ever-challenging texture-less or non-textured regions in the learning process. That is because most scenes like floors and walls usually have similar appearances that decrease matching performance. Moreover, to further mitigate the influence of dramatic image intensity variation between two consecutive images, we use two trainable parameters $m$ and $b$ based on the brightness constancy assumption, to model the adjacent frames' brightness. After updating, the image can be expressed as: $I' = m * I + b$ . As a result, the refined spectral consistency loss in appearance is formulated as: + +$$ +\begin{array}{l} L _ {\text {s p e c t r a l}} = \frac {1}{N} \sum_ {i j} \left| I _ {t + 1} ^ {\prime} \left(\omega \left(\mathbf {K p} (i, j), T _ {t \rightarrow t + 1}\right)\right)\right. \tag {6} \\ - I _ {t} ^ {\prime} (\mathbf {K p} (i, j)) | \\ \end{array} +$$ + +The geometric constraint further ensures the 3D structure to be consistent between the estimated depth at $t$ time $D_{t}$ and synthesized depth $D_{t}^{\prime}$ via warping function from adjacent depth map $D_{t + 1}$ across the entire sequence. The synthesized depth map $D_{t}^{\prime}$ can be expressed as: + +$$ +D _ {t} ^ {\prime} = D _ {t + 1} \left(\omega \left(p, T _ {t \rightarrow t + 1}\right)\right) \tag {7} +$$ + +To penalize the depth inconsistency, we normalize the depths with their mean value to prevent monocular scale ambiguity across the sequences. Therefore, the geometric depth consistency loss $L_{geo}$ turns into: + +$$ +\begin{array}{l} L _ {g e o} = \frac {1}{N} \sum_ {i j} \left| D _ {t} ^ {\prime} \left(\omega \left(\mathbf {K p} (i, j), T _ {t \rightarrow t + 1}\right)\right)\right. \tag {8} \\ - D _ {t} (\mathbf {K p} (i, j)) | \\ \end{array} +$$ + +where $N$ stands for the number of valid points in the estimated depth map. As depth discontinuities usually occur in the gradients of images, we adopt an edge-preserving smoothness term to prevent inaccurate predictions and retain the sharp local pixels, which is described as: + +$$ +L _ {\text {s m o o t h}} = \left\| \frac {| \bigtriangledown^ {2} (D \otimes G) |}{\| \bigtriangledown I \|} \right\| _ {1} \times \frac {1}{\| D \|} \tag {9} +$$ + +where $\bigtriangledown$ and $\bigtriangledown^2$ refer to the gradient and Laplacian operators respectively. $G$ represents a $5\times 5$ Gaussian kernel. $D$ and $I$ correspond to the predicted depth map and the input image, respectively. + +# 3.3. Layout Consistency Constraint + +To enhance geometric consistency in partial indoor 3D reconstruction, we introduce a layout consistency constraint + +that aligns the structural layout lines detected from the input image with those extracted from the back-projected image rendered from the reconstructed 3D scene. This constraint encourages the reconstructed geometry and camera pose to faithfully preserve dominant layout structures such as wall-floor and wall-wall boundaries, even in partially visible scenes. Layout samples are shown in Fig. 2. We represent room structure using a set of straight 2D line segments: + +$$ +\mathcal {L} = \left\{\ell_ {i} = \left(\left(x _ {1} ^ {i}, y _ {1} ^ {i}\right), \left(x _ {2} ^ {i}, y _ {2} ^ {i}\right), a _ {i}, b _ {i}\right) \right\} _ {i = 1} ^ {M} \tag {10} +$$ + +Each line $\ell_{i}$ is defined by its two endpoints in image coordinates and its slope-intercept parameters, given by: + +$$ +a _ {i} = \frac {y _ {2} ^ {i} - y _ {1} ^ {i}}{x _ {2} ^ {i} - x _ {1} ^ {i} + \epsilon}, \quad b _ {i} = y _ {1} ^ {i} - a _ {i} x _ {1} ^ {i} \tag {11} +$$ + +where $\epsilon$ is a small constant to avoid division by zero. To extract layout lines, we employ a general-purpose layout-aligned line detector such as LayoutNet [91] and RoomNet [35]. These methods detect semantically meaningful line segments aligned with structural elements in perspective indoor scenes. We apply the detector to both the input image $I$ and the back-projected image $\hat{I}$ generated from the reconstructed 3D scene using estimated camera poses. + +Given two sets of detected lines, line segments on the input images $\mathcal{L}_{\mathrm{img}} = \{\ell_i^{\mathrm{img}}\}$ and line segments on projected images $\mathcal{L}_{\mathrm{proj}} = \{\ell_j^{\mathrm{proj}}\}$ , we compute correspondences by minimizing a geometric and parametric similarity metric: + +$$ +\operatorname {m a t c h} (i) = \arg \min _ {j} \mathcal {D} \left(\ell_ {i} ^ {\mathrm {i m g}}, \ell_ {j} ^ {\mathrm {p r o j}}\right) \tag {12} +$$ + +$$ +\mathcal {D} \left(\ell_ {i}, \ell_ {j}\right) = w _ {p} \cdot d _ {\text {p o i n t}} \left(\ell_ {i}, \ell_ {j}\right) + w _ {s} \cdot \left| a _ {i} - a _ {j} \right| + w _ {o} \cdot \left| b _ {i} - b _ {j} \right| \tag {13} +$$ + +where $d_{\mathrm{point}}$ is the average distance between line endpoints, and $a_i, b_i$ are the slope and intercept. We enforce mutual best matching to form the final correspondence set $\mathcal{C}$ : + +$$ +\mathcal {C} = \left\{\left(\ell_ {i} ^ {\text {i m g}}, \ell_ {j} ^ {\text {p r o j}}\right) \mid \operatorname {m a t c h} (i) = j \text {a n d} \operatorname {m a t c h} ^ {- 1} (j) = i \right\} \tag {14} +$$ + +We define the layout consistency loss over matched line pairs as: + +$$ +\begin{array}{l} L _ {\text {l a y o u t}} = \frac {1}{| \mathcal {C} |} \sum_ {(\ell^ {\operatorname {i m g}}, \ell^ {\operatorname {p r o j}}) \in \mathcal {C}} \left(\left\| \left(x _ {1}, y _ {1}\right) ^ {\operatorname {i m g}} - \left(x _ {1}, y _ {1}\right) ^ {\operatorname {p r o j}} \right\| _ {1} \right. \\ + \left\| \left(x _ {2}, y _ {2}\right) ^ {\mathrm {i m g}} - \left(x _ {2}, y _ {2}\right) ^ {\mathrm {p r o j}} \right\| _ {1} \\ + \lambda_ {s} \left| a ^ {\mathrm {i m g}} - a ^ {\mathrm {p r o j}} \right| + \left. \lambda_ {b} \left| b ^ {\mathrm {i m g}} - b ^ {\mathrm {p r o j}} \right|\right) \tag {15} \\ \end{array} +$$ + +This loss penalizes both endpoint misalignment and orientation discrepancies. It is differentiable and can be used to backpropagate errors to the reconstruction network and pose estimator. + +With the introduced keypoint driven spectral consistency, the learnable brightness constancy, the geometric depth consistency, and layout prediction consistency, the + +![](images/a8b9c0b653966e0afa88a46c7bdd5188db751425a9a98bbddbafc72838f50cef.jpg) +Figure 2. Layout samples for indoor environments. + +![](images/aab861f53d88b0dc978268bb124bfd8947ddb3769736280a38bbe2c07a6aba1b.jpg) + +![](images/f4c088690946b41fd7d341e81b72055c42d862b5a62696a80ba0dc1cc6cf42d2.jpg) + +comprehensive constraint for SfM network $L_{SfM}$ can be written as: + +$$ +L _ {S f M} = \lambda_ {1} L _ {\text {s p e c t r a l}} + \lambda_ {2} L _ {\text {g e o}} + \lambda_ {3} L _ {\text {s m o o t h}} + \lambda_ {4} L _ {\text {l a y o u t}} \tag {16} +$$ + +# 3.4. Joint Framework Optimization + +SfM and SfS understand the 3D world from different perspectives. However, there is rich mutual information for the two algorithms to be exploited and optimized together. Constraints from inverse rendering relationships can further be utilized to optimize surface details in the SfM network, especially on those regions without many textures. Therefore, we propose two consistency terms between them to learn the mutual information and add more effective restrictions on low-textured or non-textured regions for better scene understanding and reconstruction. + +Surface shape consistency: Depth and normal vectors from the shading network and SfM should be consistent. Assuming the camera is in a 3D coordinate system and the depth at each pixel $p(x,y)$ is $D(p)$ , the 3D point coordinate $P$ corresponding to pixel $p$ can be represented as $P(p) = \left[\frac{x - c_x}{f} D(p),\frac{y - c_y}{f} D(p),D(p)\right]$ , where $D(p)$ is the depth value at image pixel $p, f$ is the focal length, and $c_{x},c_{y}$ are the principal points. Then the tangent vectors to the surface are formulated as: + +$$ +\frac {\partial P}{\partial x} = \left[ \begin{array}{c} \frac {- 1}{f} (x - c _ {x}) \frac {\partial D}{\partial x} - \frac {1}{f} D \\ \frac {- 1}{f} (y - c _ {y}) \frac {\partial D}{\partial x} - \frac {1}{f} D \\ \frac {\partial D}{\partial x} \end{array} \right] \tag {17} +$$ + +$$ +\frac {\partial P}{\partial y} = \left[ \begin{array}{c} \frac {- 1}{f} (x - c _ {x}) \frac {\partial D}{\partial y} - \frac {1}{f} D \\ \frac {- 1}{f} (y - c _ {y}) \frac {\partial D}{\partial y} - \frac {1}{f} D \\ \frac {\partial D}{\partial y} \end{array} \right] \tag {18} +$$ + +With the two crossing directions, we can compute normal vector $N$ at $p$ through cross product: + +$$ +N (p) = \frac {\partial P}{\partial x} \times \frac {\partial P}{\partial y} = \left[ \begin{array}{c} - f \frac {\partial D (p)}{\partial x} \\ - f \frac {\partial D (p)}{\partial y} \\ \left(x - c _ {x}\right) \frac {\partial D (p)}{\partial x} + \left(y - c _ {y}\right) \frac {\partial D (p)}{\partial y} + D (p) \end{array} \right] \tag {19} +$$ + +We associate the estimated depth from the SfM network with the estimated surface normal from the SfS network. Given the strong correlation between depth and normal, we introduce a surface normal cross consistency loss. This loss function optimizes the depth estimation from the SfM network by enhancing the normal maps $\tilde{\mathbf{N}}$ derived from the SfM depth maps $\tilde{\mathbf{D}}$ through the SfS network. + +$$ +L _ {c r o s s - n o r m a l} = \frac {1}{N} \sum_ {i} \left(1 - \frac {\langle \tilde {d} _ {i} , \tilde {n} _ {i} \rangle}{\| \tilde {d} _ {i} \| \cdot \| \tilde {n} _ {i} \|}\right) \tag {20} +$$ + +The computed vector $\tilde{n}_i = (\tilde{n}_x^i,\tilde{n}_y^i)^T$ represents the x and y direction components of normal $(\tilde{n}_x^i,\tilde{n}_y^i,\tilde{n}_z^i)^T$ which is estimated from SfS normal estimation. $\tilde{d}_i = (\partial_x\tilde{D}_i,\partial_y\tilde{D}_i)$ represents the x and y normal direction calculated by the gradient of the estimated depth map from the SfM network. The inner product $< \tilde{d}_i,\tilde{n}_i>$ calculates the similarity of the x and y direction normal vector calculated from SfM and SfS, which is further normalized by their magnitude $(||\tilde{d}_i||\cdot ||\tilde{n}_i||)$ to be in the range of [0, 1]. With that, if the normal from SfS and SfM are close, $\frac{< \tilde{d}_i,\tilde{n}_i > }{||\tilde{d}_i||\cdot||\tilde{n}_i||}$ will be close to 1, and the loss $1 - \frac{< \tilde{d}_i,\tilde{n}_i > }{||\tilde{d}_i||\cdot||\tilde{n}_i||})$ will be close to 0. Otherwise, the loss will be close to 1. This correlation constrains the estimated input depth to conform to the requirements specified in Eq. 2, which is crucial for the accurate reconstruction of shapes in low-textured areas. + +To further facilitate a continuous depth distribution on the same curved surface regions, we incorporate a surface shape consistency into our entire framework. The motivation behind the idea is that most low-textured or nontextured regions (e.g., floor, wall) usually fall on the same surface and share the same materials, thus contributing to continuously changing depths on those surface regions. Specially, we extract superpixel segments on the albedo images from the SfS-Net, and then constrain the corresponding 3D points on the same segments to meet the surface parameters by converting the estimated depth maps into point cloud with the introduced inverse projection function $\pi^{-1}(\cdot)$ . Different from [79], we propose to extract segments on the estimated albedo map from SfS-Net rather than the normal color images. Meanwhile, we extend the piece-wise planar assumption for indoor scenes in [43] to a more general curved surface condition, based on the fact that most outdoor roads and indoor floors are more suitable to be approximated as curved surfaces. Given the depth value at each image pixel $p(x,y)$ is $D(p)$ , the corresponding 3D points are $P = \pi^{-1}(p,D(p))$ . The curved surface can be fitted as: + +$$ +M ^ {T} P = [ 1, 1, \dots , 1 ] ^ {T} \quad a n d \quad P \left(p _ {i}\right) = D \left(p _ {i}\right) K ^ {- 1} p _ {i} \tag {21} +$$ + +where $M \in R^3$ is the parameters of each curved surface, and it can be induced with the least square fitting as: + +$$ +M = N P ^ {T} [ 1, 1,.. 1 ] ^ {T} \tag {22} +$$ + +where normal $N$ can be represented by $N = (P^T P)^{-1}$ . We therefore compute the distance between the depths fitted in the curved surface and the estimated depths on the superpixel segments in the estimated albedo map and minimize the distance error continuously by the surface shape consistency term as: + +$$ +L _ {\text {s u r f a c e - s h a p e}} = \sum_ {m} \sum_ {i} | D (p _ {i}) - \left(M ^ {T} K ^ {- 1} p _ {i}\right) ^ {- 1} | \tag {23} +$$ + +Scene registration consistency: Instead of applying constraints on pixel intensity only, we establish a global scene registration consistency between the two networks for better shape completion, as illustrated in Fig. 1. If using $P_{SfS}^{t}$ to represent the inverse projected 3D points from the inverse render network at time $t$ , making $P_{SfM}^{t}$ and $P_{SfM}^{t+1}$ to stand for the 3D points from the SfM network at time $t$ and $t+1$ respectively, we establish a mapping from $P_{SfS}^{t}$ to $P_{SfM}^{t+1}$ and from $P_{SfS}^{t}$ to $P_{SfM}^{t}$ simultaneously and then minimize the 3D distances between them. If the estimated motion vector is $T$ , the 3D registration loss is to minimize the Chamfer distance between the transformed point set and the raw point set as: + +$$ +\begin{array}{l} L _ {3 D \_ r e g} = \sum_ {P _ {S f M} ^ {t + 1}} P _ {S f S} ^ {t} \| T \cdot P _ {S f M} ^ {t + 1} - P _ {S f S} ^ {t} \| _ {2} ^ {2} + \tag {24} \\ \sum_ {P _ {S f S} ^ {t}} P _ {S f M} ^ {t + 1} | | T ^ {- 1} \cdot P _ {S f S} ^ {t} - P _ {S f M} ^ {t + 1} | | _ {2} ^ {2} \\ \end{array} +$$ + +The point clouds from the SfS-Net and SfM-Net on the same frame are forced to be consistent in the two recovered point clouds as: + +$$ +\begin{array}{l} L _ {3 D \_ c o n s i s} = \sum_ {P _ {S f M} ^ {t}} P _ {S f S} ^ {t} \| P _ {S f M} ^ {t} - P _ {S f S} ^ {t} \| _ {2} ^ {2} + \tag {25} \\ \sum_ {P _ {S f S} ^ {t}} P _ {S f M} ^ {t} | | P _ {S f S} ^ {t} - P _ {S f M} ^ {t} | | _ {2} ^ {2} \\ \end{array} +$$ + +Hence, our full scene registration consistency loss includes both the computed 3D registration loss between the raw and transformed point clouds, and the 3D consistency term acting on the two projected point clouds from the two networks. The full loss function can be used to back-propagate the error and update the weights for both shape and camera motion network on both network optimizations, which can be expressed as: + +$$ +L _ {\text {s c e n e - r e g}} = L _ {3 D \_ r e g} + L _ {3 D \_ c o n s i s} \tag {26} +$$ + +# 4. Experiments + +We first introduce the public low-texture datasets and our configuration, including indoor datasets NYU V2 and ScanNet, and an outdoor dataset FinnForest though our main target is indoor scenes. Then we provide both quantitative and qualitative results on public datasets with the corresponding analysis. Finally, the ablation experiments are designed to show the effectiveness of each part in the entire framework. + +# 4.1. Datasets and Configurations + +NYUv2 Depth dataset consists of 582 indoor video sequences captured with a Microsoft Kinect camera and we select 283 scenes as the training split and the official testing split from another 25 scenes for evaluation, following + +
MethodsTypeErrorAcc
AbsRelRMSLog10δ1δ2δ3
Wang et al. [69]S0.2200.824-0.6050.8900.970
Eigen et al. [16]S0.1580.641-0.7690.9500.988
Liu et al. [44]S0.2130.7590.0870.6500.9060.976
Li et al. [40]S0.1430.6350.0630.7880.9580.991
Xu et al. [73]S0.1430.613-0.7890.9460.984
DORN [21]S0.1150.5090.0510.8280.9650.992
SharpNet [61]S0.1390.5020.0470.8360.9660.993
Adabins [8]S0.1030.3640.0440.9030.9840.997
P3Depth [57]S0.1040.3560.0430.8980.9810.996
NeWCRFs [80]S0.0950.3340.0410.9220.9920.998
Trap Attention [56]S0.0900.3290.0390.9270.9910.998
GeoNet [77]U\\\\\\
SC-SfMLearner [10]U0.2390.8320.0980.6380.8970.964
Monodepth2 [24]U0.1760.6390.0740.7340.9370.983
MovingIndoor [88]U0.2080.7120.0860.6740.9000.968
TrianFlow [87]U0.2010.7080.0850.6870.9030.968
PackNet-SfM [25]U0.1670.6080.0710.7490.9430.986
Johnston et al. [31]U0.1640.6020.0700.7600.9430.987
P2Net [79](3-frame)U0.1590.5990.0680.7720.9420.984
HR-Depth [54]U0.1610.5960.0690.7620.9440.983
Lite-Mono [82]U0.1580.5880.0680.7660.9450.984
RA-Depth [27]U0.1520.5760.0630.7750.9490.987
Ours fullU0.1290.5200.0510.8280.9670.992
+ +Table 1. Quantitative depth comparisons between our proposed Shading-SfM scheme and state-of-the-art learning-based depth estimation methods [69] [16] [44] [40] [73] [21] [61] [8] [57] [80] [56] [77] [10] [24] [88] [87] [25] [31] [79] [54] [82] [27] on NYUv2 Depth dataset. 'S' and 'U' indicate supervised and unsupervised training types, and $\backslash$ indicates training collapsed. + +
MethodsTypeErrorAcc
AbsRelRMSLog10δ1δ2δ3
DORN [21]S0.1400.2900.0540.7980.9600.991
SharpNet [61]S0.1670.3580.0520.8010.9610.991
Monodepth2 [24]U0.1960.4610.0720.6920.9350.982
MovingIndoor [88]U0.2380.570-0.631--
P2Net [79]U0.1810.4400.0680.7170.9370.983
HR-Depth [54]U0.1740.4420.0630.7250.9420.983
RA-Depth [27]U0.1690.4230.0610.7370.9460.985
Lite-Mono [82]U0.1720.4340.0620.7290.9440.984
Ours fullU0.1510.3520.0580.7980.9560.990
+ +Table 2. Quantitative depth comparison with [21] [61] [24] [88] [79] [54] [27] [82] on ScanNet dataset. + +the same setting with [88]. We sampled training images every ten frames to reduce redundant repetitions and applied a fixed 3-frame length in the SfM network, following similar settings as [88] [79]. Besides NYUv2, we also evaluated the proposed framework on a large ScanNet indoor dataset with 1513 video sequences for about 2.5 million images in total. For outdoor scenes, we adopt sequences in the winter condition of FinnForest dataset, which contains challenging texture-less pathways mostly covered with snow. + +We apply the encoder-decoder structure (ResNet-18 [26] pre-trained on ImageNet) in the SfM-Net and SfS-Net, which further integrate the constraints for SfM network and multiple introduced constraints designated for low/nontextured regions, as shown in the Fig. 1. The training hyper-parameters for $L_{shading\_rec}$ , $L_{spectral}$ , $L_{layout}$ , $L_{geo}$ , $L_{smooth}$ , $L_{surface\_shape}$ and $L_{scene\_reg}$ are 1.0, 1.0, 0.5, 0.5, 0.1, 0.2 and 0.2, respectively. All the indoor images are resized into $640 \times 480$ , and the outdoor images are resized into $512 \times 256$ for training. The Adam [33] optimizer is + +![](images/f781679fccdf76b57239addecba688c386a9c5148a7a6b9b5930dc29a3a67c8f.jpg) +Figure 3. Visual comparisons of 3D reconstruction on NYUv2 dataset. From left to right: input (1st column); result from our method (2nd column), [82] (3rd column), [54] (4th column), and [27] (5th column). Each reconstructed 3D scene is shown from two perspectives for comparison. + +![](images/a1c9832b9e8acdeb4acd439d5fe7053bba529945a7abe23fd09642da4886640d.jpg) +Figure 4. Visual comparisons of 3D reconstruction on ScanNet dataset. From left to right: input (1st column); result from our method (2nd column), [82] (3rd column), [54] (4th column), and [27] (5th column). Each reconstructed 3D scene is shown from two perspectives for comparison. + +![](images/0e5b4c57cd467f1597b1b4153ce0e573ec37d8686c9ee3bef66ec783155acfed.jpg) +Figure 5. With input (1st row), our depth estimation results (2nd row) on NYUv2 (the first three columns) and ScanNet (the last one column) datasets in comparison with RA-Depth [27] (3rd row) and HR-Depth [54] (4th row). + +applied with $\beta_{1} = 0.9$ , $\beta_{2} = 0.99$ and a weight decaying rate as $5 \times 10^{-4}$ . The entire network is trained for a total of 50 epochs. The initial 20 epochs have a learning rate of 1e-4 and then gradually reduce to half for the rest epochs. + +# 4.2. Evaluation on Indoor NYUv2 and ScanNet + +The evaluation results of single-view depth estimation on NYUv2 and ScanNet datasets are provided in Table 1 and + +![](images/1d798e036ebdc8af951877fdfb56cbbe0ab1049ab58512bac0ca09490adacc81.jpg) +Figure 6. With input image (1st row), visual comparison of the estimated albedo and shading of our results (2-3rd row) on indoor NYUv2 (the first three columns) and outdoor Finnforest (the last one column) datasets, with [37] (4-5th row) and [78] (6-7th row). + +![](images/999866a14e536844ee3cd439bd8dd7e1b61623062db78d8236813b4f18258672.jpg) + +![](images/94436390882802ff011ed1118318f5e406c4a8ef66525e5ded191d0dc0b1dc71.jpg) + +![](images/c53a65c5b394303a8bc448f0fa10c28da0120e79ea6e5218b9f1c0689e7e1835.jpg) +Figure 7. Using network trained on NYUv2 data only, our model generalizes well to other unseen indoor (ScanNet) images without further training to generate estimated Albedo, shading and normal. + +Table 2, respectively. We compare our method with both supervised learning methods and unsupervised approaches in error and accuracy metrics. From the comparison, it can be observed that our method achieves state-of-the-art accuracies and the lowest errors on both datasets when compared with other unsupervised methods, and obtains comparable performance with most recent supervised methods. In addition, the proposed framework performs better on accuracy under three thresholds which reflect the overall accuracy of the prediction results. + +As depicted in Fig. 3 and Fig. 4, we evaluate our method qualitatively on NYUv2 and Scannet through 3D reconstruction models, as 3D models are easier to observe the difference (e.g., distortion, stretching) in real 3D space. It can be observed that the proposed method is able to prevent stretching and distortion effects that appeared in the generated point clouds from other compared methods, especially on the uniform and low-textured surfaces. It can be contributed to the integration of the SfS-Net and the proposed joint consistencies for better reconstruction. The depth estimation results are also provided in Fig. 5. + +Visual comparisons of the estimated albedo and shading on indoor NYU and outdoor Finnforest datasets are shown in Fig. 6. Our unsupervised inverse rendering-based network Shading-SfM-Net, in general, produces high-quality decompositions. Among all of the three compared approaches, our shading network achieves the best visual results. For instance, our model better recovers the albedo of + +![](images/47d671c031f753c7c23a2553009694c0e20d2f4744dac82840c4c55880d1c98a.jpg) +Input + +![](images/604bfbddc804f83780a6831619c0dd8ec4f5adeedd4527c76687b3aee17de282.jpg) +albedo + +![](images/ddaf67d449400b4b8877db1bed46eca4222fa30cfcfc12ce83cd6a4c41680c22.jpg) +shading + +![](images/7db02b25d3c5dd28031bd33812991bd8b9fbbec28f526b5e4fab32429e2a61f1.jpg) +normal + +
w/ SfS-Netw/ SSCw/ SRCw/ KPCw/ LPCAbsRelδ1
-----0.1780.733
✓----0.1610.758
✓✓---0.1550.770
✓✓✓--0.1490.781
✓✓✓✓-0.1380.805
✓✓✓✓✓0.1290.828
+ +Table 3. Ablation study of depth estimation on different combinations of network components on NYUv2 dataset. + +the wall and the shading of the bed. Though our method targets indoor scenes, it can still display stable and accurate albedo and shading maps when testing on challenging outdoor snow scenes, which provides possibilities to extend the proposed method to large-scale texture-less outdoor scenes. + +# 4.3. Generalization and Discussion + +Figure 7 provides a cross-dataset generalization from our trained SfS-Net on the indoor NYUv2, when applying directly on the ScanNet dataset. It can be noticed that the trained network can perform well to generate albedo, shading and normal on data that has never been learned before. + +To verify the effectiveness of our proposed mechanisms, a series of ablation studies on NYUv2 are conducted. We train different versions of network using different network components as shown in Table 3. The full pipeline is compared with the settings of SfM-Net by gradually incorporating SfS-Net, surface shape consistency (SSC), scene registration consistency (SRC), keypoint-based consistency (KPC), and layout prediction consistency (LPC). It can be observed that the designed constraints significantly enhance the initial depth estimation performance, and the full pipeline achieve the best accuracy and the lowest error. + +# 5. Conclusion + +This work tackles with the challenging unsupervised shape-from-shading and structure-from-motion tasks in non-texture and low-texture indoor scenes. We propose Shading-SfM-Net to simultaneously learn the two tasks, with the designated keypoint, room layout and brightness based consistencies for SfM-Net, the self-supervised reconstruction for SfS-Net, and the joint consistencies for bridging the network branches. Extensive experiments on diverse low-texture indoor scenes and ablation studies validate the effectiveness of the proposed approach. + +# Acknowledgement + +This work is supported by NSF Awards NO. 2340882, 2334624, 2334246, and 2334690. + +# References + +[1] Filippo Aleotti, Fabio Tosi, Matteo Poggi, and Stefano Mattoccia. Generative adversarial networks for unsupervised monocular depth prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 0–0, 2018. 3 +[2] Yasin Almalioglu, Muhamad Risqi U Saputra, Pedro PB de Gusmao, Andrew Markham, and Niki Trigoni. Ganvo: Unsupervised deep monocular visual odometry and depth estimation with generative adversarial networks. In 2019 International Conference on Robotics and Automation (ICRA), pages 5474-5480. IEEE, 2019. 1, 3 +[3] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3674-3683, 2018. 1 +[4] Mohamed M Atia, Shifei Liu, Heba Nematallah, Tashfeen B Karamat, and Aboelmagd Noureldin. Integrated indoor navigation system for ground vehicles with automatic 3-d alignment and position initialization. IEEE Transactions on Vehicular Technology, 64(4):1279-1292, 2015. 1 +[5] Antyanta Bangunharcana, Ahmed Magd, and Kyung-Soo Kim. Dualrefine: Self-supervised depth and pose estimation through iterative epipolar sampling and refinement toward equilibrium. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 726-738, 2023. 3 +[6] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 3 +[7] Amanda Berg, Jorgen Ahlberg, and Michael Felsberg. Generating visible spectrum images from thermal infrared. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 1143-1152, 2018. 4 +[8] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 7 +[9] Sai Bi, Xiaoguang Han, and Yizhou Yu. An 11 image transform for edge-preserving smoothing and scene-level intrinsic decomposition. ACM Transactions on Graphics (TOG), 34(4):1-12, 2015. 3 +[10] Jiawang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth and ego-motion learning from monocular video. In Advances in neural information processing systems, pages 35-45, 2019. 1, 3, 7 + +[11] Jia-Wang Bian, Huangying Zhan, Naiyan Wang, Tat-Jun Chin, Chunhua Shen, and Ian Reid. Unsupervised depth learning in challenging indoor video: Weak rectification to rescue. arXiv preprint arXiv:2006.02708, 2020. 3 +[12] Behzad Bozortabar, Mohammad Saeed Rad, Dwarikanath Mahapatra, and Jean-Philippe Thiran. Syndemo: Synergistic deep feature alignment for joint learning of depth and egomotion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210–4219, 2019. 3 +[13] Jaehoon Cho, Dongbo Min, Youngjung Kim, and Kwanghoon Sohn. Deep monocular depth estimation leveraging a large-scale outdoor stereo dataset. Expert Systems with Applications, 178:114877, 2021. 2 +[14] Jiang Dong, Marius Noreikis, Yu Xiao, and Antti Ylä-Jäaski. Vinav: A vision-based indoor navigation system for smartphones. IEEE Transactions on Mobile Computing, 18(6):1461-1475, 2018. 1 +[15] Wei Dong, Christopher Choy, Charles Loop, Or Litany, Yuke Zhu, and Anima Anandkumar. Fast monocular scene reconstruction with global-sparse local-dense grids. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4263-4272, 2023. 1 +[16] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In Proceedings of the IEEE international conference on computer vision, pages 2650-2658, 2015. 7 +[17] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NIPS, 2014. 2 +[18] Fahim Faisal, Md Abdus Salam, Mufti Bin Habib, Md Shariful Islam, and Mirza Muntasir Nishat. Depth estimation from video using computer vision and machine learning with hyperparameter optimization. In 2022 4th International Conference on Smart Sensors and Application (ICSSA), pages 39-44. IEEE, 2022. 1 +[19] Qingnan Fan, Jiaolong Yang, Gang Hua, Baoquan Chen, and David Wipf. Revisiting deep intrinsic image decompositions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8944-8952, 2018. 3 +[20] Tuo Feng and Dongbing Gu. Sganvo: Unsupervised deep visual odometry and depth estimation with stacked generative adversarial networks. IEEE Robotics and Automation Letters, 4(4):4431-4437, 2019. 3 +[21] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2002-2011, 2018. 2, 7 +[22] Ravi Garg, Vijay Kumar Bg, Gustavo Carneiro, and Ian Reid. Unsupervised cnn for single view depth estimation: Geometry to the rescue. In European conference on computer vision, pages 740-756. Springer, 2016. 2 +[23] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, pages 270-279, 2017. 3 +[24] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular + +depth estimation. In Proceedings of the IEEE international conference on computer vision, pages 3828-3838, 2019. 3, 4, 7 +[25] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2485-2494, 2020. 3, 7 +[26] 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. 7 +[27] Mu He, Le Hui, Yikai Bian, Jian Ren, Jin Xie, and Jian Yang. Ra-depth: Resolution adaptive self-supervised monocular depth estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 565-581. Springer, 2022. 3, 7 +[28] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Let there be color! joint end-to-end learning of global and local image priors for automatic image colorization with simultaneous classification. ACM Transactions on Graphics (ToG), 35(4):1-11, 2016. 4 +[29] Michael Janner, Jiajun Wu, Tejas D Kulkarni, Ilker Yildirim, and Josh Tenenbaum. Self-supervised intrinsic image decomposition. In Advances in Neural Information Processing Systems, pages 5936-5946, 2017. 3 +[30] Jianhao Jiao, Ruoyu Geng, Yuanhang Li, Ren Xin, Bowen Yang, Jin Wu, Lujia Wang, Ming Liu, Rui Fan, and Dimitrios Kanoulas. Real-time metric-semantic mapping for autonomous navigation in outdoor environments. IEEE Transactions on Automation Science and Engineering, 2024. 1 +[31] Adrian Johnston and Gustavo Carneiro. Self-supervised monocular trained depth estimation using self-attention and discrete disparity volume. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4756–4765, 2020. 3, 7 +[32] Peter Karkus, Shaojun Cai, and David Hsu. Differentiable slam-net: Learning particle slam for visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2815-2825, 2021. 1 +[33] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7 +[34] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 1 +[35] Chen-Yu Lee, Vijay Badrinarayanan, Tomasz Malisiewicz, and Andrew Rabinovich. Roomnet: End-to-end room layout estimation. In Proceedings of the IEEE international conference on computer vision, pages 4865-4874, 2017. 5 +[36] Louis Lettry, Kenneth Vanhoey, and Luc Van Gool. Darn: a deep adversarial residual network for intrinsic image decomposition. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1359-1367. IEEE, 2018. 3 + +[37] Louis Lettry, Kenneth Vanhoey, and Luc Van Gool. Unsupervised deep single-image intrinsic decomposition using illumination-varying image sequences. In Computer Graphics Forum, volume 37, pages 409-419. Wiley Online Library, 2018. 8 +[38] Stefan Leutenegger, Margarita Chli, and Roland Y Siegwart. Brisk: Binary robust invariant scalable keypoints. In 2011 International conference on computer vision, pages 2548-2555. IEEE, 2011. 4 +[39] Hanhan Li, Ariel Gordon, Hang Zhao, Vincent Casser, and Anelia Angelova. Unsupervised monocular depth learning in dynamic scenes. In Conference on Robot Learning, pages 1908-1917. PMLR, 2021. 3 +[40] Jun Li, Reinhard Klein, and Angela Yao. A two-streamed network for estimating fine-scaled depth maps from single rgb images. In Proceedings of the IEEE International Conference on Computer Vision, pages 3372-3380, 2017. 7 +[41] Ruihao Li, Sen Wang, Zhiqiang Long, and Dongbing Gu. Undeepvo: Monocular visual odometry through unsupervised deep learning. In 2018 IEEE international conference on robotics and automation (ICRA), pages 7286-7291. IEEE, 2018. 1, 3 +[42] Zhengqi Li and Noah Snavely. Learning intrinsic image decomposition from watching the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9039-9048, 2018. 3 +[43] Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz. Planercnn: 3d plane detection and reconstruction from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4450-4459, 2019. 6 +[44] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian D Reid. Learning depth from single monocular images using deep convolutional neural fields. TPAMI, 38(10):2024-2039, 2016. 2, 7 +[45] Yunfei Liu, Yu Li, Shaodi You, and Feng Lu. Unsupervised learning for intrinsic image decomposition from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3248-3257, 2020. 3 +[46] Guoyu Lu. Image-based localization for self-driving vehicles based on online network adjustment in a dynamic scope. In IEEE International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2022. 1 +[47] Guoyu Lu. Bird-view 3d reconstruction for crops with repeated textures. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4263-4270, 2023. 1 +[48] Guoyu Lu. Deep unsupervised visual odometry via bundle adjusted pose graph optimization. In IEEE International Conference on Robotics and Automation (ICRA), pages 6131-6137, 2023. 3 +[49] Guoyu Lu. Slam based on camera-2d lidar fusion. In IEEE International Conference on Robotics and Automation (ICRA), pages 16818-16825, 2024. 1 +[50] Guoyu Lu, Xue-Iuan Wong, and James McBride. From mapping to localization: A complete framework to visually estimate position and attitude for autonomous vehicles. In IEEE + +International Conference on Image Processing (ICIP), pages 3103-3107, 2019. 1 +[51] Guoyu Lu, Yan Yan, Li Ren, Jingkuan Song, Nicu Sebe, and Chandra Kambhametu. Localize me anywhere, anytime: a multi-task point-retrieval approach. In Proceedings of the IEEE International Conference on Computer Vision, pages 2434-2442, 2015. 1 +[52] Guoyu Lu, Yan Yan, Nicu Sebe, and Chandra Kambhametu. Indoor localization via multi-view images and videos. Computer Vision and Image Understanding, 161:145-160, 2017. 1 +[53] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (TOG), 39(4):71-1, 2020. 3 +[54] Xiaoyang Lyu, Liang Liu, Mengmeng Wang, Xin Kong, Lina Liu, Yong Liu, Xinxin Chen, and Yi Yuan. Hr-depth: High resolution self-supervised monocular depth estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2294-2301, 2021. 7 +[55] Ishit Mehta, Parikshit Sakurikar, and PJ Narayanan. Structured adversarial training for unsupervised monocular depth estimation. In 2018 International Conference on 3D Vision (3DV), pages 314-323. IEEE, 2018. 3 +[56] Chao Ning and Hongping Gan. Trap attention: Monocular depth estimation with manual traps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5033-5043, 2023. 7 +[57] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610-1621, 2022. 7 +[58] Andrea Pilzer, Dan Xu, Mihai Puscas, Elisa Ricci, and Nicu Sebe. Unsupervised adversarial depth estimation using cycled generative networks. In 2018 International Conference on 3D Vision (3DV), pages 587-595. IEEE, 2018. 3 +[59] Matteo Poggi, Fabio Tosi, and Stefano Mattoccia. Learning monocular depth estimation with unsupervised trinocular assumptions. In 2018 International Conference on 3D Vision (3DV), pages 324-333. IEEE, 2018. 3 +[60] Xiaojuan Qi, Renjie Liao, Zhengzhe Liu, Raquel Urtasun, and Jiaya Jia. Geonet: Geometric neural network for joint depth and surface normal estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 283-291, 2018. 2 +[61] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 0–0, 2019. 2, 7 +[62] Soumyadip Sengupta, Angjoo Kanazawa, Carlos D Castillo, and David W Jacobs. Sfsnet: Learning shape, reflectance and illuminance of facesin the wild'. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6296-6305, 2018. 3 +[63] Li Shen, Ping Tan, and Stephen Lin. Intrinsic image decomposition with non-local texture cues. In 2008 IEEE Con- + +ference on Computer Vision and Pattern Recognition, pages 1-7. IEEE, 2008. 3 +[64] Lu Sheng, Dan Xu, Wanli Ouyang, and Xiaogang Wang. Unsupervised collaborative learning of keyframe detection and visual odometry towards monocular deep slam. In Proceedings of the IEEE International Conference on Computer Vision, pages 4302-4311, 2019. 3 +[65] Aditya Singh, Raghav Narula, Hatem A Rashwan, Mohamed Abdel-Nasser, Domenec Puig, and GC Nandi. Efficient deep learning-based semantic mapping approach using monocular vision for resource-limited mobile robots. Neural Computing and Applications, 34(18):15617-15631, 2022. 1 +[66] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. In 8th International Conference on Learning Representations, 2020. 2 +[67] Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig, Nikolaus Mayer, Eddy Ilg, Alexey Dosovitskiy, and Thomas Brox. Demon: Depth and motion network for learning monocular stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5038-5047, 2017. 2 +[68] Anjie Wang, Zhijun Fang, Yongbin Gao, Songchao Tan, Shanshe Wang, Siwei Ma, and Jenq-Neng Hwang. Adversarial learning for joint optimization of depth and egomotion. IEEE Transactions on Image Processing, 29:4130-4142, 2020. 3 +[69] Peng Wang, Xiaohui Shen, Zhe Lin, Scott Cohen, Brian Price, and Alan L Yuille. Towards unified depth and semantic prediction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2800-2809, 2015. 7 +[70] Ruoyu Wang, Zehao Yu, and Shenghua Gao. Planedepth: Self-supervised depth estimation via orthogonal planes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21425-21434, 2023. 3 +[71] Sen Wang, Ronald Clark, Hongkai Wen, and Niki Trigoni. Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 2043-2050. IEEE, 2017. 2 +[72] Yair Weiss. Deriving intrinsic images from image sequences. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 68-75. IEEE, 2001. 3 +[73] Dan Xu, Elisa Ricci, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Multi-scale continuous crfs as sequential deep networks for monocular depth estimation. In CVPR, volume 1, 2017. 7 +[74] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1281-1292, 2020. 3 +[75] Ilker Yildirim, Mario Belledonne, Winrich Freiwald, and Josh Tenenbaum. Efficient inverse graphics in biological face processing. Science advances, 6(10):eaax5979, 2020. 3 + +[76] Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. Enforcing geometric constraints of virtual normal for depth prediction. In Proceedings of the IEEE International Conference on Computer Vision, pages 5684-5693, 2019. 2 +[77] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1983-1992, 2018. 1, 3, 7 +[78] Ye Yu and William AP Smith. Inverserendernet: Learning single image inverse rendering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3155-3164, 2019. 3, 8 +[79] Zehao Yu, Lei Jin, and Shenghua Gao. P $\{2\}$ net: Patch-match and plane-regularization for unsupervised indoor depth estimation. In ECCV, 2020. 3, 6, 7 +[80] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. New crfs: Neural window fully-connected crfs for monocular depth estimation. In CVPR, 2022. 7 +[81] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 340-349, 2018. 1, 3 +[82] Ning Zhang, Francesco Nex, George Vosselman, and Norman Kerle. Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18537-18546, 2023. 3, 7 +[83] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Zequn Jie, Xiang Li, and Jian Yang. Joint task-recursive learning for semantic segmentation and depth estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 235–251, 2018. 2 +[84] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4106-4115, 2019. 2 +[85] Zhenyu Zhang, Stephane Lathuiliere, Elisa Ricci, Nicu Sebe, Yan Yan, and Jian Yang. Online depth learning against forgetting in monocular videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4494-4503, 2020. 1 +[86] Qi Zhao, Ping Tan, Qiang Dai, Li Shen, Enhua Wu, and Stephen Lin. A closed-form solution to retina with non-local texture constraints. IEEE transactions on pattern analysis and machine intelligence, 34(7):1437-1444, 2012. 3 +[87] Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. Towards better generalization: Joint depth-pose learning without posenet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9151-9161, 2020. 7 +[88] Junsheng Zhou, Yuwang Wang, Kaihuai Qin, and Wenjun Zeng. Moving indoor: Unsupervised video depth learning in challenging environments. In Proceedings of the IEEE + +International Conference on Computer Vision, pages 8618-8627, 2019. 3, 7 +[89] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1851-1858, 2017. 1, 3, 4 +[90] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In IEEE international conference on robotics and automation (ICRA), pages 3357–3364, 2017. 1 +[91] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2051–2059, 2018. 5 +[92] Yuliang Zou, Pan Ji, Quoc-Huy Tran, Jia-Bin Huang, and Manmohan Chandraker. Learning monocular visual odometry via self-supervised long-term modeling. arXiv preprint arXiv:2007.10983, 2020. 1 +[93] Yuliang Zou, Zelun Luo, and Jia-Bin Huang. Df-net: Unsupervised joint learning of depth and flow using cross-task consistency. In Proceedings of the European conference on computer vision (ECCV), pages 36-53, 2018. 1, 3 \ No newline at end of file diff --git a/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/images.zip b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..80a139d420df968200fefeca8004bc8e93bffc75 --- /dev/null +++ b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6dfaa4dbc3be7931da9e355b4240231692ae16715f137fd87fe3137f421a7bb1 +size 644568 diff --git a/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/layout.json b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d5f1815c254c63b18ab8198f8f27ab0014e1d78c --- /dev/null +++ b/CVPR/2025/Shading Meets Motion_ Self-supervised Indoor 3D Reconstruction Via Simultaneous Shape-from-Shading and Structure-from-Motion/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e06f1fd2251c21341c3b66183b3684fa80e5eab5fe3438567be11ece3694cda7 +size 500973 diff --git a/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/895c3c2b-6650-4488-847e-3ab245db27e5_content_list.json b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/895c3c2b-6650-4488-847e-3ab245db27e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..09a8e4412ffb12fd849b0e9f2286f39379bf36ee --- /dev/null +++ b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/895c3c2b-6650-4488-847e-3ab245db27e5_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ac05f949783717b90708fcba75ef90b27c1e502225fa0b6ab0a7057a85bf29d +size 69875 diff --git a/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/895c3c2b-6650-4488-847e-3ab245db27e5_model.json b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/895c3c2b-6650-4488-847e-3ab245db27e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f6a2cb2bf7b5134c1435d6337d6be5936cca699a --- /dev/null +++ b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/895c3c2b-6650-4488-847e-3ab245db27e5_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c76940457298f1cd80832e6eeb1eb00b3d661c604e307bb0783d63fa5fff8b7 +size 87820 diff --git a/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/895c3c2b-6650-4488-847e-3ab245db27e5_origin.pdf b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/895c3c2b-6650-4488-847e-3ab245db27e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f2baaff6707e1e5246bd232d823afaa32177d515 --- /dev/null +++ b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/895c3c2b-6650-4488-847e-3ab245db27e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eef43e7bfbf191c1fb43ee0bdb9debc7df04788cb5b787f1a2e2552cdd97895f +size 8012309 diff --git a/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/full.md b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0e1cb224d5d041e6f7ead7446e0050447613576b --- /dev/null +++ b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/full.md @@ -0,0 +1,297 @@ +# Shadow Generation Using Diffusion Model with Geometry Prior + +Haonan Zhao $^{1}$ , Qingyang Liu $^{1}$ , Xinhao Tao $^{1}$ , Li Niu $^{1,2*}$ , Guangtao Zhai $^{1}$ $^{1}$ Shanghai Jiao Tong University $^{2}$ miguo.ai + +1{2zz-n-24,narumimaria,taoxinhao,ustcnewly,zhaiguangtao}@sjtu.edu.cn + +![](images/3e60a944d70a845831fba3b3f3e4fd7be0882d2ae6b879b003068e8bd02c6a01.jpg) +Figure 1. The visual comparison with the state-of-the-art diffusion-based method SGDiffusion [20] on DESOBAv2 dataset. Our method significantly outperforms SGDiffusion in terms of geometry details. + +# Abstract + +Image composition involves integrating foreground object into background image to obtain a composite image. One of the key challenges is to produce realistic shadow for the inserted foreground object. Recently, diffusion-based methods have shown superior performance compared to GAN-based methods in shadow generation. However, they are still struggling to generate shadows with plausible geometry in complex cases. In this paper, we focus on promoting diffusion-based methods by leveraging geometry priors. Specifically, we first predict the rotated bounding box and matched shadow shapes for the foreground shadow. Then, the geometry information of rotated bounding box and matched shadow shapes is injected into ControlNet to facilitate shadow generation. Extensive experiments on both DESOBAv2 dataset and real composite images validate the effectiveness of our proposed method. The code and model are released at https://github.com/bcmi/GPSDiffusion-Object-Shadow-Generation. + +# 1. Introduction + +Image composition [25], which aims to appropriately insert a given foreground object into the target background image to obtain a composite image, serves a crucial role + +in fields such as filmmaking, photo editing, and digital art design [27]. To achieve a more natural-looking composite result, generating realistic shadow for inserted foreground objects remains a challenging task, which has not been satisfactorily addressed by existing methods. In this paper, we focus on the above shadow generation task, aiming to substantially improve the quality of generated shadow from a geometric perspective. + +The existing shadow generation methods can be divided into GAN-based methods [11, 17] and diffusion-based methods [20]. GAN-based methods ensure the precision and authenticity of generated shadow through the combined supervision of reconstruction loss and adversarial loss. However, they generally require training the model from scratch on limited paired data, which restricts the generalization ability. In contrast, diffusion-based methods initialize the foundation models with rich prior knowledge, thus significantly outperforming GAN-based methods and generating more plausible shadow. Overall, the quality of generated shadow could be evaluated from two aspects: appearance and geometry. The diffusion-based method [20] is adept at generating shadow with realistic appearance (e.g., shadow intensity, boundary transition), but still performs poorly in terms of shadow geometry (e.g., location, scale, shape) in the challenging cases as shown in Figure 1. + +To improve the shadow geometry, we propose to assist diffusion-based methods with geometry priors. Our method has two stages. In the first stage, we employ a multi-task ge + +ometry encoder to predict the geometry information of foreground shadow given a composite image. On one hand, we predict the rotated bounding box enclosing the foreground shadow. On the other hand, we discretize shadow shapes and retrieve the matched shadow shapes for the foreground shadow. Specifically, we train an auto-encoder to reconstruct shadow shapes and extract the bottleneck features as shape embeddings. We group shape embeddings into clusters and each cluster is associated with a prototypical shape embedding. Given a composite image, the geometry encoder predicts the probabilities of all clusters, indicating the matched shadow shapes for the foreground shadow. + +In the second stage, we incorporate the rotated bounding box and matched shape embeddings from the first stage into diffusion model. Such integration ensures that the generated shadow not only aligns with the expected spatial configuration but also adheres to realistic shape characteristics, thereby enhancing the overall quality and realism of the shadow. Specifically, we use ControlNet [50] as the base model. The rotated bounding box is converted to a binary mask and appended to the input of ControlNet encoder, which can improve the accuracy of shadow placement and scaling in the generated results. Additionally, we inject the matched shape embeddings into the denoising U-Net. We train additional cross-attention layers and feed the matched shape embeddings into different layers of U-Net, which ensures that the generated shadow is informed by realistic and contextually appropriate shape information, enhancing the overall quality and coherence of the shadow synthesis. As a result, our approach generates shadow with more accurate position and more precise shapes compared with the state-of-the-art diffusion-based method SGDiffusion [20] (See Figure 1). + +We conduct experiments on DESOBAv2 [20] and real composite images to validate the effectiveness of our method. In summary, our contributions are as follows: + +- We propose to equip diffusion-based shadow generation model with geometry priors, which is a key point to enhance the shadow quality. +- We design a multi-task geometry encoder to predict the rotated bounding box and retrieve the matched shadow shapes for the foreground shadow. +- The experiments results show that our method brings notable improvement for the shadow geometry, which has never been achieved by previous methods. + +# 2. Related Work + +# 2.1. Image Composition + +The goal of image composition [25] is to seamlessly integrate a foreground object into a target background to create a realistic composite image. Achieving realistic composite images poses several visual challenges such as un + +natural transitions and multifaceted mismatches (e.g., illumination, geometry, semantics) between foreground and background. Previous methods have explored image blending [42, 48], object placement [18, 26, 49], and image harmonization [2, 3, 5-8, 21, 29, 31, 40, 43], but they often overlook the critical task of accurately generating shadows for the foreground object. + +Recently, generative image composition [4, 16, 23, 30, 35, 36, 44, 47] has attracted increasing attention, which can address multiple issues (e.g., lighting, pose, shadow) with one model. However, they can only generate shadows with simple shapes and within the foreground bounding box [20]. This work focuses on shadow generation, aiming to enhance the geometry quality of generated shadows. + +# 2.2. Shadow Generation + +Compared to the increasingly sophisticated techniques for shadow detection [13, 19, 41] and shadow removal [9, 10, 39], shadow generation remains an arduous task. Existing methods can be generally divided into rendering based and non-rendering based approaches. + +Rendering-based methods typically require explicit knowledge of illumination, reflectance, material properties, or scene geometry for foreground objects and backgrounds. Early works [14, 15] use 3D scene reconstruction or linear programming, but need user interaction for accuracy. SSN [32] generates controllable shadows using a 2D object mask and an environment light map, incorporating an Ambient Occlusion Prediction module to produce realistic shadows that can be refined interactively. The subsequent work [33] controls shadow direction and shape using geometry representation, while PixHt-Lab [34] maps pixel height data to 3D space for geometry reconstruction. However, these methods sometimes require predicting 3D information from 2D images, which is quite challenging in complex scenes. + +Non-rendering methods add realistic shadows to composite images without inferring 3D information. ShadowGAN [51] improves shadow quality by jointly using global and local discriminators. Mask-SG [12] utilizes the generated shadow mask to guide shadow generation through cycle-consistency constraints. AR-SG [17] uses attention mechanism to model the mapping between virtual shadows and real-world environments. The work [24] introduces an automatic shadow generation technique that uses a Hierarchy Attention U-Net to model interactions between the foreground and background, and an Illumination-Aware Fusion Network to blend under-exposure images for realistic shadows. SGRNet [11] first predicts the foreground shadow mask through interaction of foreground and background information, and then predicts shadow parameters to fill in the shadow area. The work [38] presents a method for handling shadows in compositing by identifying existing shade + +![](images/80507d892976d3d935b693f03fc66948c5a9cd6612a4320c0b7796d015d653cf.jpg) +Figure 2. The framework of our GPSDiffusion. In the first stage, we use a geometry encoder $E_{g}$ to predict geometry priors which include rotated bounding box and shape embeddings information. In the second stage, we use a control encoder $E_{c}$ and additional cross-attention layers to process these two types of information separately for shadow generation. + +ows and adjusting for their interaction with synthetic shadows using a generative adversarial pipeline. + +Among non-rendering based methods, two most related works are DMASNet [37] and SGDiffusion [20]. DMASNet first decomposes shadow mask generating into box prediction and shape prediction, and then attends to background shadow pixels for accuracy foreground shadow filling. Compared with [37], our method extends bounding box to rotated bounding box, and introduces prototypical shape embeddings, which is dramatically different from direct shadow shape prediction in [37]. More importantly, we integrate these geometry information into diffusion model to generate more realistic shadows. SGDiffusion [20] adapts ControlNet to shadow generation and incorporates intensity modulation to refine shadow intensity. Despite the introduction of diffusion models, SGDiffusion faces the issue of poor shadow geometry. + +# 3. Method + +Given a composite image $I_{c}$ with inserted foreground object, we aim to generate the image $\tilde{I}_{g}$ with plausible foreground shadow. The overall pipeline of our Geometry Prior guided Shadow Diffusion model (GPSDiffusion) is illustrated in Figure 2. In the first stage, we predict the geometry + +prriors using a geometry encoder $E_{g}$ , which will be detailed in Section 3.1. In the second stage, we generate realistic shadow using ControlNet that leverages geometry priors, which will be detailed in Section 3.2. + +# 3.1. Geometry Prior Prediction + +In this section, given a composite image with inserted foreground object, we expect to predict the geometry information (e.g., location, scale, shape) of plausible shadow for the foreground object. We employ a multi-task geometry encoder $E_{g}$ to perform two tasks: predicting the rotated bounding box of foreground shadow and retrieving the matched shadow shapes. In particular, given the composite image $I_{c}$ and foreground object mask $M_{fo}$ , we pass them through $E_{g}$ to output the shape probability vector $\tilde{l}_{s}$ and the rotated bounding box regression coefficient $\tilde{r}$ respectively. + +# 3.1.1. Rotated Bounding Box Regression + +Given a composite image, determining the shadow bounding box is an effective method to locate the foreground shadow. Moreover, the object shadows often vary in orientation due to diverse illumination directions. Therefore, we aim to determine the rotated bounding box enclosing the foreground shadow. Furthermore, considering the spatial relationship between foreground object and foreground + +shadow, we aim to predict the regression from rotated object bounding box to rotated shadow bounding box. The geometry encoder $E_{g}$ accounts for predicting such regression coefficient. + +Specifically, we represent the rotated bounding box using a quadruplet $\pmb{B} = (x,y,w,h,\theta)$ , where $(x,y)$ denotes the center coordinates, $w$ and $h$ denote the width and height respectively, and $\theta$ denotes the angle between the bounding box's longer side and x-axis, with a range of $[-\pi /2,\pi /2)$ . For the foreground object, the bounding box is $B_{o} = (x_{o},y_{o},w_{o},h_{o},\theta_{o})$ , while the ground-truth bounding box for the foreground shadow is $B_{s} = (x_{s},y_{s},w_{s},h_{s},\theta_{s})$ . To describe the regression from $B_{o}$ to $B_{s}$ , we introduce the regression coefficients $r = (r_x,r_y,r_w,r_h,r_\theta)$ , which can be calculated based on $B_{o}$ and $B_{s}$ as follows, + +$$ +\left\{ \begin{array}{l} r _ {x} = \left(x _ {s} - x _ {o}\right) / w _ {o}, \\ r _ {y} = \left(y _ {s} - y _ {o}\right) / h _ {o}, \\ r _ {w} = \ln \left(w _ {s} / w _ {o}\right), \\ r _ {h} = \ln \left(h _ {s} / h _ {o}\right), \\ r _ {\theta} = \left(\theta_ {s} - \theta_ {o}\right) \cdot \pi / 1 8 0. \end{array} \right. \tag {1} +$$ + +Given the composite image $I_{c}$ with foreground object mask $M_{fo}$ , we concatenate them channel-wisely and send them to geometry encoder $E_{g}$ to predict the regression coefficients $\tilde{r}$ . The details of $E_{g}$ are provided in the supplementary. Based on the rotated bounding box $B_{o}$ of foreground object and the predicted $\tilde{r}$ , we can obtain the rotated bounding box $\tilde{B}_{s}$ of foreground shadow by Eqn. (1). To supervise the regression coefficient prediction, we enforce the obtained rotated bounding box $\tilde{B}_{s}$ to be close to the ground-truth $B_{s}$ using KF IoU [45] loss: + +$$ +\mathcal {L} _ {b b o x} = \mathcal {L} _ {k f i o u} \left(\tilde {\boldsymbol {B}} _ {s}, \boldsymbol {B} _ {s}\right) = e ^ {1 - \mathrm {K F I o U} \left(\tilde {\boldsymbol {B}} _ {s}, \boldsymbol {B} _ {s}\right)}. \tag {2} +$$ + +We obtain the ground-truth rotated bounding box by locating the minimum enclosing rectangle of the foreground object and shadow mask. Please refer to the supplementary for more details. + +# 3.1.2. Matched Shadow Shape Retrieval + +After determining the rotated bounding box of foreground shadow, we expect to further get the shadow shape within the bounding box. Nevertheless, it is very difficult to directly predict the shadow shape [37]. Moreover, there could be multiple plausible shadow shapes due to the ambiguous 3D geometry of foreground object. Therefore, we convert the problem of shadow shape prediction to shadow shape retrieval, that is, discretizing the shadow shapes and retrieving those which match the foreground object. + +To discretize shadow shapes, we first collect all foreground shadow masks in the training set and train a shape auto-encoder $G_{r}$ for shadow shape reconstruction. Specifically, we extract the tight bounding boxes enclosing foreground shadows from all shadow mask images, crop and + +resize them to $64 \times 64$ resolution, and then deliver all resized shadow shape images $M_{s}$ into $G_{r}$ . The goal of $G_{r}$ is reconstructing the shadow shape images $M_{s}$ . The details of $G_{r}$ are provided in the supplementary. By using $\tilde{M}_{s}$ to denote the output of $G_{r}$ , we use Binary Cross-Entropy (BCE) loss to train $G_{r}$ : + +$$ +\mathcal {L} _ {r e c} = \mathcal {L} _ {b c e} (\tilde {\boldsymbol {M}} _ {s}, \boldsymbol {M} _ {s}). \tag {3} +$$ + +After training $G_{r}$ , its bottleneck features are supposed to contain sufficient information of shadow shapes. Thus, we extract the bottleneck features of all shadow shape images as shadow shape embeddings. We group all these shape embeddings into $N$ clusters using K-means. The shadow shape embedding closest to each cluster center is treated as the prototypical shape embedding of this cluster. In total, we have $N$ prototypical shape embeddings $\{f_{s,n}|_{n = 1}^{N}\}$ in which $f_{s,n}$ corresponds to the $n$ -th cluster. For each composite image, we can get the ground-truth cluster label for its foreground shadow, leading to the ground-truth label vector $l_{s}$ . + +We add another branch to the geometry encoder to predict an $N$ -dim vector $\tilde{l}_s$ , which indicates the probability that the foreground shadow shape belongs to each of $N$ clusters. The predicted $\tilde{l}_s$ is supervised by the ground-truth $l_s$ using softmax classification loss: + +$$ +\mathcal {L} _ {\text {s h a p e}} = \mathcal {L} _ {\text {c l s}} \left(\tilde {\boldsymbol {l}} _ {s}, \boldsymbol {l} _ {s}\right). \tag {4} +$$ + +Based on the predicted probability vector $\tilde{l}_s$ , we can retrieve top- $K$ matched shadow shapes for the foreground object and get the corresponding prototypical shape embeddings $\{f_{s,i_k}|_{k = 1}^K\}$ , in which $i_k$ is the cluster index with the $k$ -th largest probability. + +The multi-task geometry encoder is trained with the combination of two losses: + +$$ +\mathcal {L} _ {\text {g e o}} = \mathcal {L} _ {\text {b b o x}} + \mathcal {L} _ {\text {s h a p e}}. \tag {5} +$$ + +# 3.2. Geometry-Guided Diffusion Model + +In this section, we tend to incorporate the geometry information obtained in Section 3.1 into diffusion model to produce the result with realistic foreground shadow. + +We take ControlNet [50] with control encoder $E_{c}$ as our base model. For the bounding box information $\tilde{B}_s$ and shape information $\tilde{l}_s$ predicted by the geometry encoder $E_{g}$ , we integrate them into ControlNet using different strategies. + +# 3.2.1. Rotated Bounding Box Information + +Based on $\tilde{B}_s = (\tilde{x}_s, \tilde{y}_s, \tilde{w}_s, \tilde{h}_s, \tilde{\theta}_s)$ , we can get a binary mask image $M_b$ indicating the rotated bounding box region. + +Then, we concatenate $M_{b}$ with composite image $\pmb{I}_{c}$ and foreground object mask $M_{fo}$ channel-wisely. The concatenation $[I_c,M_{fo},M_b]$ is used as the input of control encoder + +![](images/e7344f1b09e6ad5b3dbc3836a0fc971d397ef2512ebfff93b7d947b11fc92db9.jpg) +Figure 3. Visual comparison with state-of-the-art methods on DESOBAv2 dataset. From left to right, we show composite image, foreground object mask, results of ShadowGAN [51], AR-SG [17], SGRNet [11], DMASNet [37], SGDiffusion [20], our GPSDiffusion, Ground-truth. + +
MethodBOS Test ImagesBOS-free Test Images
GR ↓LR ↓GS ↑LS ↑GB ↓LB ↓GR ↓LR ↓GS ↑LS ↑GB ↓LB ↓
ShadowGAN [51]7.51167.4640.9610.1970.4460.89017.32576.5080.9010.0600.4250.842
Mask-SG [12]8.99779.4180.9510.1800.5001.00019.33894.3270.9060.0440.5001.000
AR-SG [17]7.33558.0370.9610.2410.3830.76116.06763.7130.9080.1040.3490.682
SGRNet [11]7.18468.2550.9640.2060.3010.59615.59660.3500.9090.1000.2710.534
DMASNet [37]8.25659.3800.9610.2280.2760.54718.72586.6940.9130.0550.2970.574
SGDiffusion [20]6.09853.6110.9710.3700.2450.48715.11055.8740.9130.1170.2330.452
Ours5.89646.7130.9660.3740.2130.42313.80955.6160.9170.1660.1970.384
+ +Table 1. The results of different methods on DESOBAv2 dataset. The best results are highlighted in boldface. + +$E_{c}$ . In Section 4.4.2, we will visualize the $M_{b}$ predicted by our model, which effectively locates foreground shadows and provides a strong prior for ControlNet. + +# 3.2.2. Shadow Shape Information + +As introduced in Section 3.1, based on the predicted probability vector $\tilde{l}_s$ , we retrieve top- $K$ matched shadow shapes with the corresponding prototypical shape embed + +dings $\{f_{s,i_k}|_k^K\}$ . We tend to integrate these matched shape embeddings into the denoising U-Net using cross-attention mechanism. + +Specifically, we add new trainable cross-attention layers for $\bar{f}_s = [f_{s,i_1},\dots ,f_{s,i_K}]$ , similar to [46]. Given image query features $f_{q}$ and text embeddings $f_{t}$ from null text prompt (following [20]), the output of updated cross-attention can be represented as + +$$ +\begin{array}{l} \tilde {f} _ {c} = \operatorname {A t t e n t i o n} \left(\boldsymbol {Q}, \boldsymbol {K}, \boldsymbol {V}\right) + \operatorname {A t t e n t i o n} \left(\boldsymbol {Q}, \boldsymbol {K} ^ {\prime}, \boldsymbol {V} ^ {\prime}\right) \\ = \operatorname {S o f t m a x} \left(\frac {\boldsymbol {Q} \boldsymbol {K} ^ {T}}{\sqrt {d}}\right) \cdot \boldsymbol {V} + \operatorname {S o f t m a x} \left(\frac {\boldsymbol {Q} \boldsymbol {K} ^ {\prime T}}{\sqrt {d}}\right) \cdot \boldsymbol {V} ^ {\prime}, \tag {6} \\ \end{array} +$$ + +with $Q = f_qW_q$ , $K = f_tW_k$ , $V = f_tW_v$ , $K' = \bar{f}_sW_k'$ , $V' = \bar{f}_sW_v'$ , and $d$ being the dimension of each vector in $\{Q, K, K'\}$ . Here, $W_q$ , $W_k$ , $W_v$ , $W_k'$ , and $W_v'$ are the weight matrices for the linear projection layers. $Q$ is the query matrices from $f_q$ . $K$ (resp., $K'$ ), $V$ (resp., $V'$ ) are the key, values matrices from $f_t$ (resp., $\bar{f}_s$ ). + +Note that the query for shape embeddings cross-attention is the same as the query for text embeddings cross-attention, and $\pmb{W}_{K}^{\prime}, \pmb{W}_{v}^{\prime}$ are initialized from $\pmb{W}_{K}, \pmb{W}_{v}$ , respectively. + +During training, we add noise to the ground-truth image $I_{g}$ and feed the noisy latent $\pmb{z}_{t}$ into diffusion model to predict the added noise. Given the predicted geometry priors $c_{geo}$ including bounding box information $M_{b}$ and shadow shape information $\bar{f}_s$ , we supervise the training process by + +$$ +\mathcal {L} _ {\text {c o n t r o l}} = \mathbb {E} _ {t, \boldsymbol {\epsilon} \sim \mathcal {N} (0, 1)} \left[ \| \boldsymbol {\epsilon} - \epsilon_ {\psi} \left(\boldsymbol {z} _ {t}, t, \boldsymbol {c} _ {\text {g e o}}\right) \| _ {2} ^ {2} \right], \tag {7} +$$ + +in which $t$ is the current timestep, $\epsilon$ is the added noise, $\epsilon_{\psi}$ with model parameters $\psi$ estimates the added noise. + +During inference, we get the input noisy latent by adding noise to the composite image $I_{c}$ to better preserve the original information. Finally, following [20], we process the generated result through a post-processing network to better preserve background information. We directly use the pre-trained post-processing network provided by [20] without additional training. + +# 4. Experiments + +# 4.1. Datasets and Evaluation Metrics + +Following [20], we conduct experiments on both DESOBAv2 dataset [20] and real images. We choose DESOBAv2 instead of DESOBA [11], considering the drawbacks of DESOBA mentioned in [20]. Consistent with [20], we divide DESOBAv2 into 21,088 training images with 27,718 tuples (composite image, foreground object mask, ground-truth image) and 487 test images with 855 tuples, including both BOS (with background object-shadow pairs) and BOS-free images in the test set. + +For baselines, we select the most recent methods in the fields of image composition and shadow generation, capable of performing tasks similar to ours, especially SGDiffusion [20], which is currently the best-performing diffusion-based method apart from our own approach, including: ShadowGAN [51], Mask-SG [12], AR-SG [17], SGRNet [11], DMASNet [37], SGDiffusion [20]. + +
RowBaseBoxShapeGR ↓LR ↓GB ↓LB ↓
1+--7.56754.6100.2470.487
2+†-7.10053.2980.2320.459
3++-6.33148.7430.2280.451
4+-+6.36349.8680.2380.472
5+++5.89646.7130.2130.423
+ +Table 2. Ablation studies of our method on BOS test images from DESOBAv2 dataset. "Base" means ControlNet base model. "Box" means using the bounding box prior and $\dagger$ means nonrotated bounding box. "Shape" means using the shape embeddings prior. + +
K163264128
Acc (%) ↑48.467.983.994.3
+ +Table 3. Top- $K$ accuracy of shadow shape classification for different values of $K$ with a fixed cluster number $N = {256}$ . + +
MetricDMASNet [37]OursUpper
KFIoU ↑0.2030.2250.333
+ +Table 4. Quantitative results of the rotated bounding box regression. "Upper" denotes the upper bound of KFIoU. + +Following [11, 20], we evaluate results from two perspectives: image quality and shadow mask quality. Image quality is assessed using RMSE and SSIM, with Global RMSE (GR) and Global SSIM (GS) measuring the entire image, and Local RMSE (LR) and Local SSIM (LS) focusing on the foreground shadow region. For shadow mask quality, we use Balanced Error Rate (BER), with Global BER (GB) for the whole image and Local BER (LB) for the foreground shadow region. Following [20], we generate 5 results for one test image with different random seeds and select the one closest to the ground-truth to calculate evaluation metrics. + +# 4.2. Implementation Details + +Our Method is implemented with PyTorch [28]. We use AdamW optimizer [22] with a fixed learning rate of 0.0001. For $N$ and $K$ , we set them to 256 and 64, respectively. Details of the hyperparameter analysis are provided in the supplementary. We first train the shape auto-encoder $G_{r}$ for 500 epochs at $64 \times 64$ resolution, with batch size being 128. Then, we train the geometry encoder $E_{g}$ for 20 epochs at $256 \times 256$ resolution, with the batch size being 16. Finally, we freeze $E_{g}$ and jointly train the control encoder $E_{c}$ , additional linear layers, and additional cross-attention layers for 80 epochs at $512 \times 512$ resolution, with the batch size be + +![](images/0e7fc31b10dc24e777e48928fa63fa9bc5f2085dafc4e2fd3ea5f70c874b70f5.jpg) +Figure 4. The visual comparison between ground-truth shadow shape and our retrieved top-3 shadow shapes. + +![](images/f9d376d130a46460cdac55c3db65f52f23bc23b4dfc6d91b02c6fc9b53152c60.jpg) +Figure 5. Visualization results of rotated bounding box regression. We show the predicted bounding box in column 3 and overlay it on the ground-truth image in column 4. + +ing 16. All experiments are conducted on 4 NVIDIA RTX A6000 GPUs. We further present the discussion on model size and inference timings in the supplementary. + +# 4.3. Evaluation on DESOBAv2 Dataset + +Table 1 summarizes the results of quantitative comparison, where our method consistently achieves best results in most cases, particularly on BOS-free images. This indicates that even without background object-shadow pairs, the geometry priors provided by our geometry encoder are sufficient to ensure that the generated shadows align with the foreground objects. The best GB and LB scores indicate that our shadow masks are more accurate than other methods. + +Figures 3 presents the qualitative comparison results. Our method can produce realistic shadows for the people with different complex poses (row 2, 4, 5, 6), even for the person carrying addition items (row 5). For the animal (row 1) and vehicle (row 3), our method can also produce reason- + +able shadows with delicate details matching the geometry of foreground objects. In contrast, the baselines can hardly produce shadows with precise geometry. + +# 4.4. Ablation Studies + +Following [20], we conduct ablation studies on BOS test images to investigate the impact of each module using four metrics: Global RMSE (GR), Local RMSE (LR), Global BER (GB), and Local BER (LB). First, we train ControlNet base model without any geometry prior, with the results shown in row 1 of Table 2. Next, we inject bounding box information into ControlNet to explore the effects of non-rotated bounding box versus rotated bounding box. The results in row 2 and 3 confirm that the bounding box prior is useful and rotated bounding box is superior to non-rotated bounding box. The comparison between row 4 and row 1 validates the effectiveness of matched shadow shape embeddings. As reported in row 5, our full-fledged model achieves the best performance. The visual results of ablated versions are provided in the supplementary. + +To further demonstrate the effectiveness of our geometry encoder, we evaluate the regression and classification branches on DESOBAv2 test set. + +# 4.4.1. Classification Branch + +For quantitative evaluation, we report top-K accuracy for various $K$ values in Table 3, which means the accuracy that one of top-K predictions hits the ground-truth label. For qualitative evaluation, Figure 4 shows the ground-truth shadow shape and our retrieved top-3 shadow shapes of various common real-life scenarios. It can be seen that the predicted shape information accurately reflects even complex scenarios, like when shadows are interrupted, such as when one leg's shadow is partially occluded by the other. This detailed capture of true shape information forms a robust foundation for subsequent controllable shadow generation, demonstrating a close alignment between the chosen shape data and reality. + +# 4.4.2. Regression Branch + +For quantitative evaluation, we report KF IoU scores [45] and conduct a comparison of the regression accuracy with DMASNet, which similarly predicts bounding boxes for shadows. We present the quantitative results in Table 4, revealing a substantial improvement in our method's localization precision while emphasizing the generalization capabilities of the geometry encoder. For qualitative evaluation, Figure 5 shows that our predicted rotated bounding boxes are compatible with the ground-truth shadows. + +# 4.5. Evaluation on Real Composite Images + +We compare different methods on real composite images from [11], which are constructed using backgrounds and foregrounds from DESOBA test set. Figure 6 shows that + +![](images/ba9c35512b32893a091d3ff3b718e17d98b7257ea66a64377e9af26c3d6f1e95.jpg) +Figure 6. Visual comparison with state-of-the-art methods on real composite images. From left to right, we show composite image, foreground object mask, results of ShadowGAN [51], AR-SG [17], SGRNet [11], DMASNet [37], SGDiffusion [20], our GPSDiffusion. + +GPSDiffusion produces realistic shadows in terms of location, scale, and shape, outperforming other methods with less realistic results. All baselines except SGDiffusion are trained on DESOBA, while SGDiffusion finetunes DESOBAv2-trained models on DESOBA. For our GPSDiffusion, we directly use DESOBAv2-trained model without finetuning on DESOBA, and our model still performs the best. Additional visual results are provided in the supplementary. + +For quantitative evaluation, we follow [20] and conduct user study involving 50 human raters. We calculate the B-T scores [1] for the baselines [11, 12, 17, 20, 37, 51] and our method. The results show that our method achieves the highest score, indicating that our method generates the most reasonable shadows and aligns best with human visual perception. The experimental details are provided in the sup + +plementary. + +# 5. Conclusion + +In this work, we have presented a novel diffusion-based shadow generation model: GPSDiffusion, which leverages geometry priors. By using additional geometry encoder, we predict the rotated bounding box and matched shape embeddings for the foreground shadow, which dramatically enhances the geometric accuracy of generated shadow. Experiments on both DESOBAv2 dataset and real images demonstrate the superiority and generalizability of our method. + +# Acknowledgement + +The work was supported by the National Natural Science Foundation of China (Grant No. 62471287). + +# References + +[1] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 8 +[2] Haoxing Chen, Zhangxuan Gu, Yaohui Li, Jun Lan, Changhua Meng, Weiqiang Wang, and Huaxiong Li. Hierarchical dynamic image harmonization. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1422-1430, 2023. 2 +[3] Jianqi Chen, Yilan Zhang, Zhengxia Zou, Keyan Chen, and Zhenwei Shi. Dense pixel-to-pixel harmonization via continuous image representation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2 +[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2 +[5] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8394-8403, 2020. 2 +[6] Wenyan Cong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. BargainNet: Background-guided domain translation for image harmonization. In ICME, 2021. +[7] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18470-18479, 2022. +[8] Julian Jorge Andrade Guerreiro, Mitsuru Nakazawa, and Björn Stenger. Pct-net: Full resolution image harmonization using pixel-wise color transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5917-5926, 2023. 2 +[9] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 710-718, 2023. 2 +[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. 2 +[11] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. In Proceedings of the AAAI conference on artificial intelligence, pages 914–922, 2022. 1, 2, 5, 6, 7, 8 +[12] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2472–2481, 2019. 2, 5, 6, 8 +[13] Xiaowei Hu, Tianyu Wang, Chi-Wing Fu, Yitong Jiang, Qiong Wang, and Pheng-Ann Heng. Revisiting shadow de + +tction: A new benchmark dataset for complex world. IEEE Transactions on Image Processing, 30:1925-1934, 2021. 2 +[14] Kevin Karsch, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, Hailin Jin, Rafael Fonte, Michael Sittig, and David Forsyth. Automatic scene inference for 3d object compositing. ACM Transactions on Graphics (TOG), 33(3):1-15, 2014. 2 +[15] Eric Kee, James F O'brien, and Hany Farid. Exposing photo manipulation from shading and shadows. ACM Trans. Graph., 33(5):165-1, 2014. 2 +[16] Sumith Kulal, Tim Brooks, Alex Aiken, Jiajun Wu, Jimei Yang, Jingwan Lu, Alexei A Efros, and Krishna Kumar Singh. Putting people in their place: Affordance-aware human insertion into scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17089-17099, 2023. 2 +[17] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8139-8148, 2020. 1, 2, 5, 6, 8 +[18] Liu Liu, Zhenchen Liu, Bo Zhang, Jiangtong Li, Li Niu, Qingyang Liu, and Liqing Zhang. Opa: object placement assessment dataset. arXiv preprint arXiv:2107.01889, 2021. 2 +[19] Lihao Liu, Jean Prost, Lei Zhu, Nicolas Papadakis, Pietro Liò, Carola-Bibiane Schonlieb, and Angelica I Aviles-Rivero. Scotch and soda: A transformer video shadow detection framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10449-10458, 2023. 2 +[20] Qingyang Liu, Junqi You, Jianting Wang, Xinhao Tao, Bo Zhang, and Li Niu. Shadow generation for composite image using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2024. 1, 2, 3, 5, 6, 7, 8 +[21] Sheng Liu, Cong Phuoc Huynh, Cong Chen, Maxim Arap, and Raffay Hamid. Lemart: Label-efficient masked region transform for image harmonization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18290-18299, 2023. 2 +[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[23] 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, pages 2294–2305, 2023. 2 +[24] Quanling Meng, Shengping Zhang, Zonglin Li, Chenyang Wang, Weigang Zhang, and Qingming Huang. Automatic shadow generation via exposure fusion. IEEE Transactions on Multimedia, 25:9044-9056, 2023. 2 +[25] Li Niu, Wenyan Cong, Liu Liu, Yan Hong, Bo Zhang, Jing Liang, and Liqing Zhang. Making images real again: A comprehensive survey on deep image composition. arXiv preprint arXiv:2106.14490, 2021. 1, 2 +[26] Li Niu, Qingyang Liu, Zhenchen Liu, and Jiangtong Li. Fast object placement assessment. arXiv preprint arXiv:2205.14280, 2022. 2 + +[27] Jaesik Park, Joon-Young Lee, Yu-Wing Tai, and In So Kweon. Modeling photo composition and its application to photo re-arrangement. In 2012 19th IEEE International Conference on Image Processing, pages 2741–2744. IEEE, 2012. 1 +[28] 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 +[29] Jinlong Peng, Zekun Luo, Liang Liu, and Boshen Zhang. Frih: Fine-grained region-aware image harmonization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4478-4486, 2024. 2 +[30] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4208-4217, 2024. 2 +[31] Xintian Shen, Jiangning Zhang, Jun Chen, Shipeng Bai, Yue Han, Yabiao Wang, Chengjie Wang, and Yong Liu. Learning global-aware kernel for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7535-7544, 2023. 2 +[32] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4380-4390, 2021. 2 +[33] Yichen Sheng, Yifan Liu, Jianming Zhang, Wei Yin, A Cengiz Oztireli, He Zhang, Zhe Lin, Eli Shechtman, and Bedrich Benes. Controllable shadow generation using pixel height maps. In European Conference on Computer Vision, pages 240-256. Springer, 2022. 2 +[34] Yichen Sheng, Jianming Zhang, Julien Philip, Yannick Hold-Geoffroy, Xin Sun, He Zhang, Lu Ling, and Bedrich Benes. Pixht-lab: Pixel height based light effect generation for image compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16643–16653, 2023. 2 +[35] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18310-18319, 2023. 2 +[36] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8048-8058, 2024. 2 +[37] Xinhao Tao, Junyan Cao, Yan Hong, and Li Niu. Shadow generation with decomposed mask prediction and attentive shadow filling. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5198-5206, 2024. 3, 4, 5, 6, 8 +[38] Lucas Valença, Jinsong Zhang, Michael Gharbi, Yannick Hold-Geoffroy, and Jean-François Lalonde. Shadow harmonization for realistic compositing. In SIGGRAPH Asia 2023 Conference Papers, pages 1-12, 2023. 2 + +[39] Florin-Alexandru Vasluianu, Tim Seizinger, and Radu Timofte. Wsrd: A novel benchmark for high resolution image shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1826-1835, 2023. 2 +[40] Ke Wang, Michael Gharbi, He Zhang, Zhihao Xia, and Eli Shechtman. Semi-supervised parametric real-world image harmonization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5927-5936, 2023. 2 +[41] 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. 2 +[42] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Gp-gan: Towards realistic high-resolution image blending. In Proceedings of the 27th ACM international conference on multimedia, pages 2487-2495, 2019. 2 +[43] Ke Xu, Gerhard Petrus Hancke, and Rynson WH Lau. Learning image harmonization in the linear color space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12570-12579, 2023. 2 +[44] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2 +[45] Xue Yang, Yue Zhou, Gefan Zhang, Jirui Yang, Wentao Wang, Junchi Yan, Xiaopeng Zhang, and Qi Tian. The kfiou loss for rotated object detection. arXiv preprint arXiv:2201.12558, 2022. 4, 7 +[46] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.5 +[47] 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. 2 +[48] Jinsong Zhang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Sunil Hadap, Jonathan Eisenman, and Jean-François Lalongne. All-weather deep outdoor lighting estimation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 10158–10166, 2019. 2 +[49] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 231-240, 2020. 2 +[50] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4 +[51] Shuyang Zhang, Runze Liang, and Miao Wang. Shadowgan: Shadow synthesis for virtual objects with conditional adversarial networks. Computational Visual Media, 5:105-115, 2019. 2, 5, 6, 8 \ No newline at end of file diff --git a/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/images.zip b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..704407fec3cabb2ec94b891294cae669ac127482 --- /dev/null +++ b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c01fd6429a5cac99f2dd1c37b46ef3ac847438392652450dd0909d78ea2cf3e +size 969921 diff --git a/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/layout.json b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c50ef3b9f233ac886f7e4c800b5b86d90249b235 --- /dev/null +++ b/CVPR/2025/Shadow Generation Using Diffusion Model with Geometry Prior/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d02e2adcfeb4660598766e9a0ac7c5a732655106a8928f9f6e5fef0a6ab5ffbf +size 395512 diff --git a/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/db972709-2487-47f2-8ee0-97e1d23e77bd_content_list.json b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/db972709-2487-47f2-8ee0-97e1d23e77bd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b8dc18067df17ead01bafaa5f2ce639ca5c0daa2 --- /dev/null +++ b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/db972709-2487-47f2-8ee0-97e1d23e77bd_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69659ae8fea48c7ef685e11d10c5beca7c48d1b4e93f0ffc1e30d46965379ac0 +size 78524 diff --git a/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/db972709-2487-47f2-8ee0-97e1d23e77bd_model.json b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/db972709-2487-47f2-8ee0-97e1d23e77bd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fa700311c546f155b200d26110b9404ce7ddb3c7 --- /dev/null +++ b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/db972709-2487-47f2-8ee0-97e1d23e77bd_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e8db12a946c164185f08a6a436b8455f4512aadbdb5541a42f02bdc98bf9edf +size 97963 diff --git a/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/db972709-2487-47f2-8ee0-97e1d23e77bd_origin.pdf b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/db972709-2487-47f2-8ee0-97e1d23e77bd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fbbb1aec7a4bb8685708eaaa88b95b32859fcd2f --- /dev/null +++ b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/db972709-2487-47f2-8ee0-97e1d23e77bd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e02f993d04b9df0a02b8e38326dc3fc705cb9f9efe8f69cc0d64411a87ed991d +size 7674915 diff --git a/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/full.md b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9c7c338ae85ac025b933b209328dd7154cfd5bb1 --- /dev/null +++ b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/full.md @@ -0,0 +1,398 @@ +# Shape Abstraction via Marching Differentiable Support Functions + +Sunkyung Park, Jeongmin Lee, Dongjun Lee* Seoul National University + +{sunk1136，ljmlgh，djlee}@snu.ac.kr + +# Abstract + +Shape abstraction, simplifying shape representation into a set of primitives, is a fundamental topic in computer vision. The choice of primitives shapes the structure of world understanding, yet achieving both high abstraction accuracy and versatility remains challenging. In this paper, we introduce a novel framework for shape abstraction utilizing a differentiable support function (DSF), which offers unique advantages in representing a wide range of convex shapes with fewer parameters, providing smooth surface approximation and enabling differentiable contact features (gap, point, normal) essential for downstream applications involving contact-related problems. To tackle the associated optimization and combinatorial challenges, we introduce two techniques: differentiable shape parameterization and hyperplane-based marching to enhance accuracy and reduce DSF requirements. We validate our method through experiments demonstrating superior accuracy and efficiency, and showcase its applicability in tasks requiring differentiable contact information. + +# 1. Introduction + +Shape representation is pivotal in computer vision and serves as a fundamental building block for explaining and understanding the world in a computable form. Representation techniques range from classical models like meshes and grids to more recently emerging methods such as implicit fields and encoders [37, 46], each selected for their suitability to specific applications. Meanwhile, shape abstraction simplifies these representations into more manageable primitives, facilitating a more structured understanding of the world. This is also known to align with how humans perceive and understand scenes by separating them into components [2], and often enables more efficient processing across a variety of applications, including semantic labeling [49], editing [29], and physics simulation [8]. + +In this regard, shape abstraction has been explored + +![](images/8177471064d39f4326ba461aa484ffc8a987512bbe36fc6e15d08cf991b16da3.jpg) + +![](images/7533034556a54b69c5315a73de09c7dd65f77d58e656794ed334a009d08ff05d.jpg) +Shape Editing + +![](images/70fb5797ceb071eef15b4f482162366e909f840cabad4670913ccf60cce895f4.jpg) +Contact Localization +Figure 1. Our method abstracts an object represented by SDF into DSFs. The abstracted results can be utilized in differentiable formulations of shape editing, contact localization, and differentiable simulation. For shape editing, the abstracted DSFs are modified by contact information. + +![](images/26957c5cc0956fff51b576b689a2bd7f3842e864a562ba0287d57946a10d3e3d.jpg) +Differentiable Simulation + +through various research efforts in recent years, utilizing different levels of primitives including planar sections [26, 27], cuboids [38, 40, 48], polytopes [5, 8], and, more commonly, superquadrics [22, 23, 29, 35]. In this paper, we aim to establish a novel framework for shape abstraction based on a new type of primitive—differentiable support function (DSF [19]), a recently proposed technique for representing convex geometry. We suggest that DSF offers the following notable advantages in shape abstraction: 1) unlike conventional primitives, it can accommodate a wide range of convex shapes and, compared to more complex polytopes or meshes, requires fewer parameters to accurately represent smooth surfaces; and 2) the abstraction introduces a novel utility—differentiable contact, where contact features (gap, point, normal) between shapes are guaranteed to be differentiable with respect to both shape parameters and object pose. This allows the abstraction to be utilized in contact-related problems, including contact localization, differentiable simulation, and shape post-optimization by leveraging additional geometric and physical information. + +Our DSF-based shape abstraction is performed by extracting a set of DSF parameters from signed distance function (SDF) data, which is a fundamental low-level volumetric information format that can be readily transformed from raw sensory inputs [6, 32, 34]. However, this extraction process can involve complex optimization and combinatorial challenges, as also pointed out in other works on computational method for primitive abstraction [22, 44]. To address these issues, we propose two techniques tailored for DSF-based shape abstraction: 1) differentiable shape parameterization, which enables efficient and robust fitting of DSF into the desired geometry using unconstrained optimization; and 2) hyperplane-based marching, which enhances abstraction accuracy for arbitrary complex shapes and reduces the number of required DSFs. We demonstrate the advantages of our abstraction scheme in terms of both accuracy and efficiency through various experiments, and also provide examples of downstream tasks that can be performed using the differentiable contact utility. + +# 2. Related Works + +# 2.1. Geometry Representation and Abstraction + +In computer vision and graphics, meshes are a standard geometry representation. Various methods have been proposed to generate meshes from sensory data, such as RGB images and point clouds, using neural networks [5, 13, 17, 45]. Despite their versatility, obtaining high-quality meshes is challenging, and representing smooth surfaces requires a large number of vertices. Another widely used representation is implicit functions, studied in various contexts based on data sources and applications [4, 20, 21, 28, 32, 34]. While effective for representing continuous objects in a scalable manner, they are difficult to structurally interpret and usability is often limited by their output type. + +Abstracting geometry into convex components has been researched in the convex decomposition field. Mesh decomposition [25, 47], which decomposes a mesh into convex hulls, is widely utilized in physics simulations. Neural network-based methods have been developed to generate convexes based on various primitives [8, 11, 35, 38, 40, 48], though these methods face challenges in terms of generalizability and accuracy. More recently, computational methods using superquadrics have been proposed [22, 23], demonstrating superior representation performance compared to neural networks. These methods progressively extract primitives from objects, and the number of components varies depending on the shape of the object. While superquadrics can generalize to shapes beyond cuboids and ellipsoids, they remain restricted to symmetric geometries. The DSF introduced in [19], on the other hand, can represent a broader range of smooth convex shapes than superquadrics, while the abstraction method has not been proposed. + +# 2.2. Contact-Related Problems + +Contact-related problems typically arise in fields such as physics simulation and robotics. These problems are inherently linked to the shape representation of objects, with computation of contact features between them. While contact computation have been extensively studied for convex shapes [1, 12, 19, 43], performing such queries for general implicit functions is often challenging due to incompatibility in output values [8]. As a result, only a limited set of shape pairs, such as sphere-sphere, convex-convex, and SDF-point, have been practically employed for contact-related problems. From a computer vision perspective, this connects with various shape reconstruction and abstraction algorithms. For instance, voxel and SDF reconstruction methods have been widely used in tasks like grasping [41, 42] and collision avoidance planning [28, 31]. Also several primitive-based abstraction for sensory inputs [8, 29] has been applied to physics simulations. + +Recently, leveraging gradients in contact-related problems is gaining attention to improve computational efficiency, enabling advancements in areas such as reinforcement learning [9], manipulation planning [33], and system identification [10, 14]. However, querying differentiable contact between shapes remains an ongoing research. Studies such as [14, 39] focus on basic geometries like spheres, cones, and boxes. Approximate gradients using sampling [18] or randomized smoothing [30] have also been proposed, with natural connections to 3D reconstruction techniques in computer vision. In contrast, our work aims to abstract shapes by leveraging DSF, which is a theoretically guaranteed shape class that provides analytic gradients. + +# 3. Differentiable Shape Parameterization + +# 3.1. DSF Parameters + +The notion of differentiable support function (DSF) utilized in this work is primarily based on the geometry representation presented in [19]. This representation leverages the theoretical property that the following differentiable convex function serves to a support function for a specific convex geometry: + +$$ +h (x) = \left(\sum_ {i = 1} ^ {N _ {v}} \left\{\max \left(v _ {i} ^ {T} x, 0\right) \right\} ^ {p}\right) ^ {\frac {1}{p}} \tag {1} +$$ + +where $x \in \mathbb{R}^3$ denotes the given direction, $v_i \in \mathbb{R}^3$ is a vertex, $p \in \mathbb{R}$ is a smoothness parameter, and $N_v$ is the number of vertices. From the definition of the support function, a support point is obtained as the partial derivative of $h(x)$ with respect to $x$ . + +In [19], to ensure a valid geometry, the vertices in Eq. (1) are designed so that the origin is located within their convex hull. In this paper, we extend this approach by introducing + +![](images/db784504d2c884450110135eed7a12360e5e9a6d6861ccc9ff2b6d331577dd4a.jpg) +Figure 2. DSF fitting results with $\alpha$ (middle) and without $\alpha$ (right, where $\forall \alpha_{i}$ is fixed at 0 for geometric center) for the given convex shapes (left). Black and red dots represent the center and vertices of DSF, respectively. + +![](images/ffbd2cb179c5b81f3015b98662708cb19082076f65f88bd517b3e050080db4c0.jpg) + +![](images/ab41595b9b2e10fac5af79586bbdb7d6c34189e60be088387f1f92e5922724b8.jpg) + +![](images/e8251efed3cef95f5ea5dfca461c556d35643465e5f5d3a0c1407cd7a868cd4f.jpg) +Figure 3. A simple example of the shape optimization problem utilizing differentiable contact. Contact feature (gap) between two DSFs is computed along a smooth variation of shape parameter $\phi$ (left). Since the gap is differentiable with respect to $\phi$ , it can be minimized using gradient-based solvers (right). + +![](images/6760a3d169c0e4e46f1b33411078b7034fd17e4c32cb74a1426bfcc538b72a31.jpg) + +a center variable $c \in \mathbb{R}^3$ , allowing the vertices to be defined relative to an arbitrary center position as follows: + +$$ +h (x) = \left(\sum_ {i = 1} ^ {N _ {v}} \left\{\max \left(\left(v _ {i} - c\right) ^ {T} x, 0\right) \right\} ^ {p}\right) ^ {\frac {1}{p}}. \tag {2} +$$ + +In Eq. (2), the center $c$ still must be located within the convex hull of $v$ . To incorporate this constraint in a differentiable manner, we propose the following parameterization: + +$$ +c = \sum_ {i = 1} ^ {N _ {v}} v _ {i} \beta_ {i} \quad \text {w h e r e} \quad \beta_ {i} = \frac {e ^ {\alpha_ {i}}}{a \sum_ {j = 1} ^ {N _ {v}} e ^ {\alpha_ {j}}} + \frac {a - 1}{a N _ {v}} \tag {3} +$$ + +which is designed to enforce $\sum_{i=1}^{N_v} \beta_i = 1$ and $\forall \beta_i \in \left[\frac{a-1}{aN_v}, \frac{a + N_v - 1}{aN_v}\right]$ with $a > 1$ . Through Eq. (3), the center $c$ is parameterized by unbounded $\alpha_i \in \mathbb{R}$ . Similarly, the smoothness parameter $p$ is modeled using the unbounded shape parameters $\tilde{p}$ as follows: + +$$ +p = (\tanh \tilde {p} + 1) \frac {p _ {\max } - 2}{2} + 2 \tag {4} +$$ + +which is designed to ensure $p \in (2, p_{\max})$ to maintain differentiability of (1) and avoid excessive exponentiation. + +![](images/6aa3e55aad136f7bef6aaa8d8d37220096d63296b050d60926e1ec96b651c03d.jpg) +Figure 4. DSF fitting results (colored) for the given polygonal meshes (gray). Despite using only 20 vertices, the fit remains robust across a wide range of geometries. + +As a result, the DSF is defined through the shape parameters $\phi = (v_{1:N_v},\alpha_{1:N_v},\tilde{p})\in \mathbb{R}^{4N_v + 1}$ . The proposed parameterization offers two main advantages. First, we can achieve high representation performance with a few vertices $v$ by flexibly adjusting $c$ based on the interpolation parameters introduced in Eq. (3). This is especially useful for shapes with both curved and flat surfaces (see Fig. 2 for examples). Second, the support function $h(x,\phi)$ and also corresponding support point are differentiable with respect to the shape parameters $\phi$ , enabling its effective use in the DSF fitting process explained in Sec. 3.2. This further implies that contact features between DSFs are differentiable with respect to $\phi$ (see supplementary material for more details), allowing for its use in shape optimization for specific downstream tasks (see Fig. 3 for an example). Note that this differentiable contact is not naturally achievable when using polytopes or superquadrics as primitives. + +# 3.2. DSF Fitting + +Based on the shape parameterization in Sec. 3.1, we can formulate the problem of fitting a DSF to an arbitrary convex geometry as the following optimization: + +$$ +\min _ {\phi \in \mathbb {R} ^ {4 N _ {v} + 1}} \sum_ {i = 1} ^ {N _ {h}} \left(h \left(x _ {i}, \phi\right) + c ^ {T} x _ {i} - \hat {h} _ {i}\right) ^ {2} \tag {5} +$$ + +where $\hat{h}_i$ is the support function value in the direction $x_i$ for the target convex geometry, and $N_h$ is the number of direction samples. + +From our differentiable shape parameterization, the solution $\phi^*$ of Eq. (5) can be obtained by standard nonlinear least square solvers. However, in practice, using an integer $p$ significantly reduces the computational cost of exponentiation in Eq. (2) compared to a continuous $p$ . Thus, an integer $p$ close to the optimal $p^*$ is selected, and Eq. (5) is re-solved with fixed $p$ . The results of fitting are illustrated in Fig. 4. The arbitrary convex shapes in Fig. 4 are accurately fitted using only 20 vertices, demonstrating the effectiveness of our DSF parameterization and fitting formulation. + +![](images/05256e19ce676605b96cfc40def2fbb878584f2ca6e3ec459d2fcdc28cb0ae92.jpg) +Figure 5. The abstraction pipeline for sensory inputs. The SDF can be generated using computational or learning-based algorithms proposed in various studies. Our algorithm abstracts DSF components from the SDF data. + +# 4. Hyperplane-based Marching + +In this section, we propose a method to abstract arbitrary complex shapes as a set of DSFs. As illustrated in the pipeline in Fig. 5, we first generate the SDF from the sensory inputs, using computational or learning-based methods proposed in various studies [20, 21, 32, 34]. Consequently, we can obtain a set of SDF data: point $(P)$ , corresponding SDF value $(S)$ , and gradient of SDF $(G)$ . + +One naive approach to generating DSFs using these SDF data is to extract a mesh using the marching cube algorithm [24], decompose it into convex meshes [25, 47], and then fit the DSFs. However, the fitting error from Eq. (5) may accumulate throughout the entire process. Furthermore, this approach may yield an excessive number of convex components since certain geometries that are misaligned with the principal axis [36] cannot be effectively decomposed by axis-aligned cutting planes [47]. In contrast, the marching scheme, iteratively generating the primitives within the object, can effectively handle the accumulation of fitting error and can intuitively abstract the misaligned geometries. Therefore, we design a marching method for DSF to reduce the number of convex components while improving the representation performance. + +A recent study [23] aligns with our problem of interest, proposing a method for marching superquadrics on a given truncated SDF. This approach gradually identifies a collection of voxelized distances and extracts primitives from them. However, such voxel-based marching is unsuitable for our framework, as distance data is not readily computable from the DSF. Instead, we introduce a hyperplane-based marching approach, which identifies a collection of hyperplanes and fits the DSF within the convex region (intersection of halfspaces) based on the formulation described in Eq. (5), progressively filling the occupied space. To + +![](images/56dc1a2e64057659c370e8ca1e06a7bc8ba1595591470716c49ea9f6ac0158f8.jpg) +Figure 6. Visualization of the region reduction (b-c) and expansion (d-f) process for the input geometry (a). Hyperplanes and the convex region are visualized by straight lines and blue regions. Black dots are the $p_m$ for each method and red dots are the vertices of DSF fitted to the convex region. + +obtain a sufficiently large convex region, we propose two methods: the region reduction and the region expansion, which are visualized in Fig. 6. + +# 4.1. Hyperplanes + +We introduce several notations for the hyperplanes used in our study. A hyperplane, defined by an equation $n^T x - o = 0$ , is denoted as $H = [n; o] \in \mathbb{R}^4$ where $n$ and $o$ are a unit vector in $\mathbb{R}^3$ and a scalar offset, respectively. Then we define the minimum distance from a point $u \in \mathbb{R}^3$ to a set of hyperplanes $\mathcal{H}$ as follows: + +$$ +d (\mathcal {H}, u) = \max _ {H _ {i} \in \mathcal {H}} \left(n _ {i} ^ {T} u - o _ {i}\right). \tag {6} +$$ + +Similarly, the hyperplane corresponding to the minimum distance from $u$ is determined as follows: + +$$ +H (\mathcal {H}, u) = \underset {H _ {i} \in \mathcal {H}} {\operatorname {a r g m a x}} \left(n _ {i} ^ {T} u - o _ {i}\right). \tag {7} +$$ + +# 4.2. Region Reduction + +The region reduction method is inspired by the algorithm for generating obstacle-free space [7], sequentially slicing the region with hyperplanes. This method is initialized at a point $p_m \in P$ randomly selected where $s_m < -\epsilon$ for $s_m \in S$ and a small $\epsilon > 0$ . Using the closest point $p_j \in P$ from $p_m$ where $s_j \in S$ and $s_j > 0$ , along with its gradient $g_j \in G$ , a normal of a hyperplane is defined as follows: + +$$ +n _ {j} = \left\{ \begin{array}{l l} g _ {j} & \text {i f} g _ {j} ^ {T} \left(p _ {j} - p _ {m}\right) - s _ {j} > 0 \\ \frac {p _ {j} - p _ {m}}{\| p _ {j} - p _ {m} \|} & \text {o t h e r w i s e} \end{array} . \right. \tag {8} +$$ + +Algorithm 1: Region Reduction +1 Input: $P,S,G,\mathcal{D}$ +2 Initialize $\mathcal{H} = \emptyset$ +3 Select $p_m\in P$ +4 $\mathcal{H}\gets \mathcal{H}\cup H_{D^{\prime}}$ for all $D^{\prime}\in \mathcal{D}$ +5 $R =$ intersection(H) +6 while $\exists p_i\in R$ s.t. $s_i > 0$ do +7 $p_j\gets$ the closest point in $R$ to $p_m$ with $s_j > 0$ +8 Calculate $n_j$ by using Eq. (8) +9 $\mathcal{H}\gets \mathcal{H}\cup H_j$ where $H_{j}$ as in Eq. (9) +10 $R =$ intersection(H) +11 end +12 Fit DSF $D$ to $R$ by using Eq. (5) +13 Output: $D$ + +Consequently, a hyperplane $H_{j}$ is defined as follows: + +$$ +H _ {j} = \left[ n _ {j}; n _ {j} ^ {T} \left(p _ {j} - p _ {m}\right) - s _ {j} \right]. \tag {9} +$$ + +Using Eq. (9), we can eliminate the points $p_i \in P$ outside of $H_j$ , and the next closest point $p_j$ from $p_m$ is determined for the next $H_j$ until there is no remaining $p_j$ with $s_j > 0$ . Through iterations, this process results in a convex region $R$ , and a DSF is fitted to the hyperplanes sampled from $R$ by solving Eq. (5). + +After each fitting, all points $p \in P$ with $s < 0$ inside the generated DSF are excluded to prevent the creation of identical DSFs. For any subsequent region reduction process, additional hyperplanes are employed to control overlap with existing DSFs, defined as + +$$ +H _ {D} = \left[ \delta_ {D}; \| \delta_ {D} \| + o _ {d} \right] \tag {10} +$$ + +where $\delta_{D}$ is the minimum distance vector from $p_m$ to the DSF, computed by the minimum distance solver in [1], $D$ and $o_d$ is the overlap parameter that determines the amount of allowed overlap. In practice, assigning a proper positive value to $o_d$ helps prevent discontinuous gaps between DSFs and reduces the overall number of required DSFs (see also Sec. 5.2). Overall process is summarized in Alg. 1. + +# 4.3. Region Expansion + +In this method, a convex region defined by the intersection of halfspaces is incrementally expanded using a collection of sampled distances. The halfspaces are defined by a set of hyperplanes $\mathcal{H}$ with fixed normals $n_j$ , and the hyperplane offsets are adjusted based on $d(\mathcal{H}, p)$ , where $p$ is sampled with a small value of $|d(\mathcal{H}, p)|$ . This method is initialized at the point $p_m \in P$ for minimum SDF value $s_m \in S$ , setting each initial $o_j$ of $H_j \in \mathcal{H}$ to $-s_m + n_j^T p_m$ : a sphere centered at $p_m$ . For each expansion, a set of points $P_s \subseteq P$ is sampled as + +$$ +P _ {s} = \left\{p _ {i} \in P \mid | d (\mathcal {H}, p _ {i}) | < t \right\} \tag {11} +$$ + +Algorithm 2: Region Expansion +1 Input: $P,S$ +2 Select $p_m\in P$ +3 Initialize $\mathcal{H} = \{H_j = [n_j; - s_m + n_j^T p_m]\}_{j = 1}^{N_d}$ +4 while offsets of $\mathcal{H}$ do not converge do +5 Sample $P_{s}$ by using Eq. (11) +6 for $j = 1,2,\dots N_{d}$ do +7 Define $P_{s,j}$ by using Eq. (12) and $P_{s}$ +8 Define $Y_{j}$ by using Eq. (13) and $P_{s,j}$ +9 Compute $o_j^*$ by using Eq. (15) +10 $H_{j}\gets [n_{j};o_{j}^{*}]$ +11 end +12 end +13 $R =$ intersection $(\mathcal{H})$ +14 Fit DSF $D$ to $R$ by using Eq. (5) +15 Output: $D$ + +where $t \in \mathbb{R}$ is a truncation value. For each $p_i \in P_s$ , the minimum distance and corresponding hyperplane $H_j$ are computed by Eq. (6) and Eq. (7). Conversely, a set of points corresponding to each $H_j$ is defined as follows: + +$$ +P _ {s, j} = \left\{p _ {i} \in P _ {s} \mid H (\mathcal {H}, p _ {i}) = H _ {j} \right\}. \tag {12} +$$ + +To expand $H_{j}$ for a single point $p_i \in P_{s,j}$ , the offset of $H_{j}$ must be added by $d(\mathcal{H}, p_i) - s_i$ . Generalizing this to multiple points, a set of distances to be added to the offset of $H_{j}$ is defined as follows: + +$$ +Y _ {j} = \left\{d \left(\mathcal {H}, p _ {i}\right) - s _ {i} \mid p _ {i} \in P _ {s, j} \right\}. \tag {13} +$$ + +Using $Y_{j}$ corresponding to $H_{j}$ , the optimization problem for the offset of $H_{j}$ is formulated as + +$$ +y ^ {*} = \underset {y \in \mathbb {R}} {\operatorname {a r g m i n}} \sum_ {k = 1} ^ {N _ {j}} w _ {k} \left(y - y _ {k}\right) ^ {2} \tag {14} +$$ + +where $N_{j} = |Y_{j}|,y_{k}\in Y_{j},$ and $w_{k}$ is a weight for $y_{k}$ , which is 1 for $s_k\geq 0$ and $10^{-4}$ for $s_k < 0$ . The optimal offset $o_j^*$ for $H_{j}$ is determined by the closed-form solution of Eq. (14) as follows: + +$$ +o _ {j} ^ {*} = o _ {j} + \frac {\sum_ {k = 1} ^ {N _ {j}} w _ {k} y _ {k}}{\sum_ {k = 1} ^ {N _ {j}} w _ {k}}. \tag {15} +$$ + +By iterating Eqs. (11) to (15), $\mathcal{H}$ can be fitted to $P_{s}$ and expanded to a larger region, neglecting minor concavities as the solution $o_{j}^{*}$ in Eq. (15) is based on the weighted average distance for $P_{s,j}$ . Once all the $o_{j}$ converge, a convex region $R$ is obtained by the intersection of halfspaces defined by $\mathcal{H}$ . Then, a DSF is fitted to the hyperplanes sampled from $R$ , and $p_i \in P$ inside the generated DSF are just removed from the algorithm to prevent the generation of identical DSFs. The details of this method are provided in Alg. 2. + +Algorithm 3: DSF Abstraction +1 Input: $P, S, G$ +2 Initialize $\mathcal{D} = \emptyset$ +3 while $\min_{s \in S} s \leq t_{\text{end}}$ do +4 if $\min_{s \in S} s \leq t_s$ then +5 | $D \gets \text{RegionReduction}(P, S, G, \mathcal{D})$ +6 else +7 | $D \gets \text{RegionExpansion}(P, S)$ +8 end +9 if $D$ is not degenerated then +10 | $\mathcal{D} \gets \mathcal{D} \cup D$ +11 end +12 end +13 Output: $\mathcal{D}$ + +![](images/180ff4b85fa8be0fb458202c345c9cce5a62c8c8c43b739d864686950151e4d5.jpg) +(a) +Figure 7. Comparison of abstraction methods. The object is the armadillo from the Stanford 3D scanning repository [6]. (a) Region reduction method (49 DSFs). (b) Region expansion method (22 DSFs). (c) Hybrid method (31 DSFs). + +![](images/49d2669d7810e42e53ec7c693d9d851ed1caca235cd9ca25da9ae0da8df4b80f.jpg) +(b) + +![](images/6c20d7335cc2afcbfdd7a57bbce759ed6ea403b87a6c42cc3be6485b45adf05c.jpg) +(c) + +# 4.4. Hybrid Method and Self-Degeneration + +Although the region reduction method has a high representation performance, its strict exclusion of exterior regions may result in an excessive number of DSFs for highly concave geometries. While the region expansion method can neglect minor concavities based on the weighted average distance formulation in Eq. (15), reducing the number of DSFs for highly concave geometries, it may also result in a loss of surface detail. To address these limitations, we propose a hybrid method that balances representation performance with the number of convex components. + +Initially, Alg. 1 is used to produce DSFs with high representation power. If the minimum SDF value exceeds a threshold $t_s$ , Alg. 2 is applied to efficiently cover the remaining parts. The details of the hybrid method are described in Alg. 3. The comparison among the three methods is visualized in Fig. 7. After fitting each DSF to $R$ , DSFs with small volumes or significant discrepancies from the SDF are degenerated. The details of the degeneration process are provided in the supplementary material. + +# 5. Experiments + +# 5.1. Abstraction Results + +Our DSF abstraction method using hyperplane-based marching (MDSF, Alg. 3) is compared with three different baselines: an optimization method based on a neural network, a fitting method, and a marching method. The optimization method based on a neural network is implemented by CvxNet [8] trained for a single object, optimizing all primitives simultaneously. The fitting method is implemented by the state-of-the-art mesh decomposition algorithm (CoACD [47]), and by fitting the DSF to each convex mesh from CoACD (CoDSF). For the marching method, the state-of-the-art marching algorithm for superquadrics (MP [23]) is compared with our method. + +For CvxNet [8], the batch size and the maximum training steps are set to 32 and 5000, respectively. For CoDSF and our method, we set $N_{v}$ in Eq. (2) and $N_{h}$ in Eq. (5) to 20 and 1742, respectively. The overlap parameter $o_{d}$ in Alg. 1 is set to a large value, greater than the object size, to allow intersection between DSFs. + +The evaluation is conducted on the YCB dataset [3] and the robot links CAD data provided by the MuJoCo Menagerie [50], including 5 robots with a total of 51 links and 14 objects from YCB. The comparison metrics are intersection over union (IoU), Chamfer-L1 distance, and the number of primitives $(N_{c})$ . Definitions for IoU and Chamfer-L1 distance are provided in the supplementary material. IoU for DSF is calculated by determining occupancy for each point sample using the minimum distance solver for DSF [1]. For Chamfer-L1 distance, the surface points of superquadrics are sampled via the equal-distance sampling method in [22], and the surface points of DSF are obtained from a fine mesh derived from the DSF. + +The abstraction results are visualized in Fig. 8. The results of IoU, Chamfer-L1, and $N_{c}$ are summarized in Tab. 1. Our DSF abstraction produces higher IoU, lower Chamfer-L1, and lower $N_{c}$ than all other baselines, which demonstrates abstraction efficiency: higher representation power with fewer primitives. The visualizations for the YCB dataset [3] are provided in the supplementary material. + +# 5.2. Ablation Studies + +In Tab. 2, our hybrid method is compared with the region reduction method and the region expansion method. All methods are evaluated on the same dataset and metrics as Sec. 5.1. The region reduction method produces high representation performance with large $N_{c}$ while the region expansion method produces low representation performance with small $N_{c}$ . Notably, all three methods produce smaller $N_{c}$ than all other baselines in Tab. 1. + +By setting $o_d$ to be larger than the object size as in Sec. 5.1, the intersection between DSFs is perfectly al + +![](images/eb99cb50da5a850bf7748078b6d6353d6efaab4a32799e5f2b0b96a696779de9.jpg) +Figure 8. Abstraction results for the links of robots. From left to right: Input mesh, CvxNet [8], Marching-Primitives (MP [23]), DSF fitting for CoACD [47] results (CoDSF), and our DSF abstraction algorithm (MDSF, Alg. 3). + +
CategoryIoU ↑Chamfer-L1 ↓Nc ↓
CvxNet [8]MP [23]CoDSFMDSFCvxNet [8]MP [23]CoDSFMDSFCvxNet [8]MP [23]CoDSFMDSF
Franka-fr30.87720.94420.95100.95170.17600.09660.08660.08475034.12521.37513
Lite60.86670.92260.95200.95330.20210.13630.09350.09195037.71439.14297.2857
Xarm70.85510.94380.94880.95350.20560.09680.09010.08635039.62518.12514.25
A10.75950.92210.92910.94240.30420.07840.07040.06505021.88.89.8
G10.81540.90920.92930.92760.16050.09840.08120.08095041.087025.913015.2174
YCB0.72970.89490.90180.90760.19070.09040.08150.07645038.714323.857116.5714
mean0.81070.91710.93090.93330.19000.09880.08350.08105037.692320.830813.8462
+ +Table 1. Results of CvxNet [8], Marching-Primitives (MP [23]), DSF fitting for CoACD [47] (CoDSF), and our DSF abstraction algorithm (MDSF, Alg. 3). Chamfer-L1 is multiplied by 10. The number of objects for each category is 8, 7, 8, 5, 23, and 14, from top to bottom. + +
MethodIoU ↑Chamfer-L1 ↓Nc ↓
Red0.93990.078017.3846
Exp0.91700.093812.1077
Hyb0.93330.081013.8462
Small-od0.93990.078319.4308
+ +Table 2. Ablation study. Red: the region reduction method (Alg. 1). Exp: the region expansion method (Alg. 2). Hyb: the hybrid method (Alg. 3). Small- $o_d$ : the hybrid method with a small $o_d$ . Hyb and Red use the same large $o_d$ as in Sec. 5.1. + +lowed, significantly reducing the overall number of required DSFs. In Tab. 2, the result with the large $o_d$ (Hyb) is compared with the result with a small $o_d$ (0.012 for normalized 3D space), denoted as Small- $o_d$ . Although the hybrid method with the small $o_d$ achieves higher IoU and lower Chamfer-L1 distance compared to the large $o_d$ , it also results in a larger $N_c$ , indicating that surface details are more accurately represented with an increased number of components. However, since $N_c$ in both methods remains lower than the other baselines in Tab. 1, the $o_d$ can be adjusted based on the task requirements, depending on whether intersection or overlap between DSFs is desirable. + +# 6. Applications + +# 6.1. Shape Editing + +Since the gap between DSFs is differentiable with respect to $\phi$ , the abstracted DSFs can be edited to align with contact information. This problem is formulated as minimizing the gap between the abstracted DSF and the contact geometries (DSFs) while maintaining the original shape: + +$$ +\min _ {\phi \in \mathbb {R} ^ {4 N _ {v} + 1}} \sum_ {i = 1} ^ {N _ {h}} \left(h \left(x _ {i}, \phi\right) + c ^ {T} x _ {i} - \hat {h} _ {i}\right) ^ {2} + w _ {g} \sum_ {i = 1} ^ {N _ {c o n}} \delta_ {i} ^ {2} \tag {16} +$$ + +where the first term is the same as Eq. (5), $N_{con}$ is the number of contacts, $w_{g}$ is a weight for the gap, and $\delta_{i}$ is a gap between the DSF and $i$ -th contact geometry. Visualization for the shape editing process is provided in Fig. 9. + +# 6.2. Contact Localization + +Contact localization, identifying a contact point or detecting a link in collision, is essential for safety control and human-robot interaction [15, 16, 51]. Typically, this process relies on robot dynamics or neural networks using sensory data. Utilizing differentiable contact features with respect + +![](images/90c2abb4749021ee76caeaed4f837dca8edd8c36ec671d2d75a4db3934f19531.jpg) +Figure 9. Illustrative example of shape editing using contact information with a gripper. The edited DSF is highlighted in red. + +![](images/ddbc012918cdd33cde958e33df9a921be28c8c738c944146e893fc2738be9d62.jpg) +Figure 10. Visualization of contact localization problem. The initial point is optimized to lie on the surface of the DSF: a contact point generating a specified force and torque while meeting friction constraints of contact force. + +to pose, computed by the contact feature solver [19] with our abstracted DSFs, this problem can be formulated as + +$$ +\min _ {p, \lambda \in \mathbb {R} ^ {3}} \| \lambda - \lambda_ {p r o j} \| ^ {2} + \| \gamma - J _ {c} ^ {T} \lambda \| ^ {2} + \delta^ {2} \tag {17} +$$ + +where $\delta$ is a gap between a point $p$ and DSF, $\lambda$ is a contact force, $\lambda_{proj}$ is a projection of $\lambda$ on a friction cone defined by the contact point and normal computed between $p$ and DSF, $\gamma$ is a measurement of force and torque, and $J_{c}$ is a contact Jacobian derived from the contact point. The DSF used in Eq. (17) is determined by the minimum gap between $p$ for each iteration. The first term in Eq. (17) is a relaxation of friction constraint of contact force, and the second term is designed to match the contact force with sensory data. The result of Eq. (17) is illustrated in Fig. 10. The initial point in Fig. 10 is effectively optimized to the surface of DSF producing the measured force and torque $\gamma$ . + +# 6.3. Differentiable Simulation + +The differentiable contact features with respect to object pose for our abstracted DSFs can be utilized in the differentiable simulations. In [33], the derivatives of contact features with respect to joint positions are computed to smooth the forward contact dynamics and solve a manipulation planning problem: finding a joint position trajectory for an initial and final pose of an object. The implementation in [33] relies on automatic differentiation and simple + +![](images/8f909680cce847a03d6719617904832e4b1826215cd091c2af856fdcce14f92a.jpg) +Figure 11. The results of manipulation planning. The apple and the banana in YCB dataset [3] are abstracted by our DSF abstraction method into 9 and 5 DSFs $(N_{c})$ , respectively. They are imported in simulation and the iMPC in [33] solves the optimal joint position trajectory for a given initial and final pose of objects. + +collision proxies including sphere and box. We integrate our abstracted DSFs into [33] and employ the differentiable contact features from the contact feature solver [19]. + +The results of the manipulation planning problem are illustrated in Fig. 11. The object is abstracted into DSFs by Alg. 3 and imported into the simulation. Then, the iterative MPC (iMPC) framework in [33] solves the manipulation planning problem by utilizing our analytic gradients. + +# 7. Discussions and Conclusions + +We propose two techniques for DSF-based shape abstraction: the differentiable shape parameterization and the hyperplane-based marching. The differentiable shape parameterization enables the efficient and robust DSF fitting across a wide range of convex geometries, providing differentiable contact features with respect to the shape parameters. For complex geometries, the hyperplane-based marching method improves abstraction accuracy and reduces the required number of DSF components. Our method outperforms the state-of-the-art in terms of accuracy and efficiency, achieving higher accuracy with fewer components. Additionally, we provide three applications that leverage the differentiable contact features of abstracted DSFs: shape editing, contact localization, and differentiable simulation. + +Although our current approach relies on the SDF data, additional visual information such as texture embeddings and differentiable rendering can be utilized through the differentiable formulations of DSF. This integration has the potential to enhance both accuracy and applicability across a range of computer vision tasks in future works. + +Acknowledgements. This research was supported by the Ministry of Trade, Industry & Energy (MOTIE) of Korea (RS-2024-00419641), and the National Research Foundation (NRF) grant funded by the Ministry of Science and ICT (MSIT) of Korea (RS-2022-00144468). + +# References + +[1] S. An, S. Lee, J. Lee, S. Park, and D. Lee. Collision detection between smooth convex bodies via riemannian optimization framework. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12464-12471. IEEE, 2024. 2, 5, 6 +[2] I. Biederman. Recognition-by-components: a theory of human image understanding. *Psychological Review*, 94(2):115, 1987. 1 +[3] B. Calli, A. Singh, A. Walsman, S. Srinivasa, P. Abbeel, and A. M. Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In 2015 International Conference on Advanced Robotics (ICAR), pages 510-517. IEEE, 2015. 6, 8 +[4] Z. Chen and H. Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5939-5948, 2019. 2 +[5] Z. Chen, A. Tagliasacchi, and H. Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 45-54, 2020. 1, 2 +[6] B. Curless and M. Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 2, 6 +[7] R. Deits and R. Tedrake. Computing large convex regions of obstacle-free space through semidefinite programming. In Algorithmic Foundations of Robotics XI: Selected Contributions of the Eleventh International Workshop on the Algorithmic Foundations of Robotics, pages 109-124. Springer, 2015. 4 +[8] B. Deng, K. Genova, S. Yazdani, S. Bouaziz, G. Hinton, and A. Tagliasacchi. Cvxnet: Learnable convex decomposition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 31-44, 2020. 1, 2, 6, 7 +[9] C. D. Freeman, E. Frey, A. Raichuk, S. Girgin, I. Mordatch, and O. Bachem. Brax-a differentiable physics engine for large scale rigid body simulation. arXiv preprint arXiv:2106.13281, 2021. 2 +[10] M. Geilinger, D. Hahn, J. Zehnder, M. Bächler, B. Thomaszewski, and S. Coros. Add: Analytically differentiable dynamics for multi-body systems with frictional contact. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 2 +[11] K. Genova, F. Cole, D. Vlasic, A. Sarna, W. T. Freeman, and T. Funkhouser. Learning shape templates with structured implicit functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7154-7164, 2019. 2 +[12] E. G. Gilbert, D. W. Johnson, and S. S. Keerthi. A fast procedure for computing the distance between complex objects in three-dimensional space. IEEE Journal on Robotics and Automation, 4(2):193-203, 1988. 2 +[13] T. Groueix, M. Fisher, V. G. Kim, B. C. Russell, and M. Aubry. A papier-mâché approach to learning 3d surface + +generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 216-224, 2018. 2 +[14] T. A. Howell, S. Le Cleac'h, J. Z. Kolter, M. Schwager, and Z. Manchester. Dojo: A differentiable simulator for robotics. arXiv preprint arXiv:2203.00806, 9, 2022. 2 +[15] M. Iskandar, O. Eiberger, A. Albu-Schäffer, A. De Luca, and A. Dietrich. Collision detection, identification, and localization on the drr sara robot with sensing redundancy. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3111-3117. IEEE, 2021. 7 +[16] M. Iskandar, A. Albu-Schäffer, and A. Dietrich. Intrinsic sense of touch for intuitive physical human-robot interaction. Science Robotics, 9(93):eadn4008, 2024. 7 +[17] C. Kong, C. H. Lin, and S. Lucey. Using locally corresponding cad models for dense 3d reconstructions from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4857-4865, 2017. 2 +[18] S. Le Cleac'h, H. X. Yu, M. Guo, T. Howell, R. Gao, J. Wu, Z. Manchester, and M. Schwager. Differentiable physics simulation of dynamics-augmented neural objects. IEEE Robotics and Automation Letters, 8(5):2780-2787, 2023. 2 +[19] J. Lee, M. Lee, and D. Lee. Uncertain pose estimation during contact tasks using differentiable contact features. In Proceedings of Robotics: Science and Systems, 2023. 1, 2, 8 +[20] M. Liu, R. Shi, L. Chen, Z. Zhang, C. Xu, X. Wei, H. Chen, C. Zeng, J. Gu, and H. Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10072-10083, 2024. 2, 4 +[21] M. Liu, C. Xu, H. Jin, L. Chen, M. Varma T, Z. Xu, and H. Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 2, 4 +[22] W. Liu, Y. Wu, S. Ruan, and G. S. Chirikjian. Robust and accurate superquadric recovery: A probabilistic approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2676-2685, 2022. 1, 2, 6 +[23] W. Liu, Y. Wu, S. Ruan, and G. S. Chirikjian. Marching-primitives: Shape abstraction from signed distance function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8771-8780, 2023. 1, 2, 4, 6, 7 +[24] W. E. Lorensen and H. E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics, 21(4):163-169, 1987. 4 +[25] K. Mamou, E. Lengyel, and A. Peters. Volumetric hierarchical approximate convex decomposition. Game Engine Gems, 3:141-158, 2016. 2, 4 +[26] J. McCrae, K. Singh, and N. J. Mitra. Slices: a shape-proxy based on planar sections. ACM Transactions on Graphics (TOG), 30(6):168, 2011. 1 +[27] J. Mccrae, N. J. Mitra, and K. Singh. Surface perception of planar abstractions. ACM Transactions on Applied Perception (TAP), 10(3):1-20, 2013. 1 + +[28] L. Mescheder, M. Oechsle, M. Niemeyer, S. Nowozin, and A. Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 2 +[29] T. Monnier, J. Austin, A. Kanazawa, A. Efros, and M. Aubry. Differentiable blocks world: Qualitative 3d decomposition by rendering primitives. In Advances in Neural Information Processing Systems (NeurIPS), pages 5791-5807, 2023. 1, 2 +[30] L. Montaut, Q. Le Lidec, A. Bambade, V. Petrik, J. Sivic, and J. Carpentier. Differentiable collision detection: a randomized smoothing approach. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3240-3246. IEEE, 2023. 2 +[31] H. Oleynikova, Z. Taylor, M. Fehr, R. Siegwart, and J. Nieto. Voxblox: Incremental 3d euclidean signed distance fields for on-board may planning. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1366-1373. IEEE, 2017. 2 +[32] J. Ortiz, A. Clegg, J. Dong, E. Sucar, D. Novotny, M. Zollhoefer, and M. Mukadam. isdf: Real-time neural signed distance fields for robot perception. In Proceedings of Robotics: Science and Systems, 2022. 2, 4 +[33] T. Pang, H. J. T. Suh, L. Yang, and R. Tedrake. Global planning for contact-rich manipulation via local smoothing of quasi-dynamic contact models. IEEE Transactions on Robotics (TRO), 2023. 2, 8 +[34] J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2, 4 +[35] D. Paschalidou, A. O. Ulusoy, and A. Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10344-10353, 2019. 1, 2 +[36] K. Pearson. LIII. On lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2(11):559-572, 1901. 4 +[37] C. R. Qi, H. Su, K. Mo, and L. J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 652-660, 2017. 1 +[38] C. Y. Sun, Q. F. Zou, X. Tong, and Y. Liu. Learning adaptive hierarchical cuboid abstractions of 3d shape collections. ACM Transactions on Graphics (TOG), 38(6):1-13, 2019. 1, 2 +[39] K. Tracy, T. A. Howell, and Z. Manchester. Differentiable collision detection for a set of convex primitives. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3663-3670. IEEE, 2023. 2 +[40] S. Tulsiani, H. Su, L. J. Guibas, A. A. Efros, and J. Malik. Learning shape abstractions by assembling volumetric primitives. In Proceedings of the IEEE/CVF Conference on Com + +puter Vision and Pattern Recognition (CVPR), pages 2635- 2643, 2017. 1, 2 +[41] D. Turpin, L. Wang, E. Heiden, Y. C. Chen, M. Macklin, S. Tsogkas, S. Dickinson, and A. Garg. Grasp'd: Differentiable contact-rich grasp synthesis for multi-fingered hands. In Proceedings of the European Conference on Computer Vision (ECCV), pages 201-221. Springer, 2022. 2 +[42] D. Turpin, T. Zhong, S. Zhang, G. Zhu, E. Heiden, M. Macklin, S. Tsogkas, S. Dickinson, and A. Garg. Fast-grasp'd: Dexterous multi-finger grasp generation through differentiable simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8082-8089. IEEE, 2023. 2 +[43] G. Van Den Bergen. Proximity queries and penetration depth computation on 3d game objects. In Game Developers Conference, page 209, 2001. 2 +[44] N. Vaskevicius and A. Birk. Revisiting superquadric fitting: A numerically stable formulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(1):220-233, 2017. 2 +[45] N. Wang, Y. Zhang, Z. Li, Y. Fu, W. Liu, and Y. G. Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52-67, 2018. 2 +[46] P. Wang, L. Liu, Y. Liu, C. Theobalt, T. Komura, and W. Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1 +[47] X. Wei, M. Liu, Z. Ling, and H. Su. Approximate convex decomposition for 3d meshes with collision-aware concavity and tree search. ACM Transactions on Graphics (TOG), 41 (4):1-18, 2022. 2, 4, 6, 7 +[48] K. Yang and X. Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM Transactions on Graphics (TOG), 40(4):1-11, 2021. 1, 2 +[49] H. Yuan, J. Xu, H. Pan, A. Bousseau, N. J. Mitra, and C. Li. Cadtalk: An algorithm and benchmark for semantic commenting of cad programs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3753-3762, 2024. 1 +[50] K. Zakka, Y. Tassa, and MuJoCo Menagerie Contributors. MuJoCo Menagerie: A collection of high-quality simulation models for MuJoCo, 2022. 6 +[51] D. Zurlo, T. Heitmann, M. Morlock, and A. De Luca. Collision detection and contact point estimation using virtual joint torque sensing applied to a robot. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7533-7539. IEEE, 2023. 7 \ No newline at end of file diff --git a/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/images.zip b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b2470b095dec2e68dfd7a96255aeec755e4cae20 --- /dev/null +++ b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eaaf881b49f2b306f91d3433f9536e11a3dc8fc2ff5d8370527ec1c0f43b72db +size 530836 diff --git a/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/layout.json b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf7e8fe8e3cf973947af29e541b5e5fbb4cc06ce --- /dev/null +++ b/CVPR/2025/Shape Abstraction via Marching Differentiable Support Functions/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a15bac5fb1e98af7522fd1f76dabe05af169622fb55aa6cfad4f29017eadd48 +size 510144 diff --git a/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/38963ce2-7528-4468-bb05-644eea549fbe_content_list.json b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/38963ce2-7528-4468-bb05-644eea549fbe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..47df7a9ca468c31d3e2d30b105d2cb80a65f3df2 --- /dev/null +++ b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/38963ce2-7528-4468-bb05-644eea549fbe_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb884e0915e000b42eff891e13881b5560a1d1797bf5565500e8ad376f1a9e81 +size 83062 diff --git a/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/38963ce2-7528-4468-bb05-644eea549fbe_model.json b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/38963ce2-7528-4468-bb05-644eea549fbe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..46e09be2838086c108161c59bd2580a81d53e890 --- /dev/null +++ b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/38963ce2-7528-4468-bb05-644eea549fbe_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d6af4d979e6b8e59553442c3902855139d6c2a086156fe9ffc76b58193f103f +size 113945 diff --git a/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/38963ce2-7528-4468-bb05-644eea549fbe_origin.pdf b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/38963ce2-7528-4468-bb05-644eea549fbe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..03aed74087ca1c7700802d6d0564fa896835250a --- /dev/null +++ b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/38963ce2-7528-4468-bb05-644eea549fbe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cc69469d9e5328cdbe7a715fb40f87f93629164eee9b57b6fa7438c29fac7fd +size 5216996 diff --git a/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/full.md b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..35425bdbcde2c9a082e428fefc3077d7c4c5d1d7 --- /dev/null +++ b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/full.md @@ -0,0 +1,326 @@ +# Shape My Moves: Text-Driven Shape-Aware Synthesis of Human Motions + +Ting-Hsuan Liao $^{1,2*}$ Yi Zhou $^{2}$ Yu Shen $^{2}$ Chun-Hao Paul Huang $^{2}$ Saayan Mitra $^{2}$ Jia-Bin Huang $^{1}$ Uttaran Bhattacharya $^{2}$ $^{1}$ University of Maryland, College Park, USA $^{2}$ Adobe Research + +# Abstract + +We explore how body shapes influence human motion synthesis, an aspect often overlooked in existing text-to-motion generation methods due to the ease of learning a homogenized, canonical body shape. However, this homogenization can distort the natural correlations between different body shapes and their motion dynamics. Our method addresses this gap by generating body-shape-aware human motions from natural language prompts. We utilize a finite scalar quantization-based variational autoencoder (FSQ-VAE) to quantize motion into discrete tokens and then leverage continuous body shape information to de-quantize these tokens back into continuous, detailed motion. Additionally, we harness the capabilities of a pretrained language model to predict both continuous shape parameters and motion tokens, facilitating the synthesis of text-aligned motions and decoding them into shape-aware motions. We evaluate our method quantitatively and qualitatively, and also conduct a comprehensive perceptual study to demonstrate its efficacy in generating shape-aware motions. Project URL: https://shape-move.github.io/. + +# 1. Introduction + +Human motion synthesis is an expansive field with wide applications in avatar creation, robotics, and the gaming industry [27, 63, 75, 83]. Advances in both motion synthesis methods and language modeling techniques have, in recent times, led to the development of text-to-motion synthesis, with the aim of generating dynamic motions that are contextually aligned with textual descriptions. + +Previous efforts in text-to-motion synthesis have explored techniques to integrate motion and language. Some approaches map motion and language into a shared latent space, and then sample motions from it based on text inputs [5]. Others utilize diffusion models conditioned on learned features, such as CLIP, to enhance the quality and relevance of the generated motions [85]. To overcome the + +![](images/5172c06b78383d56de366355593c8bba557aa43669808bb67e9b713aac2af951.jpg) +Figure 1. Text-Driven Shape-Aware Motion Synthesis. The same motion performed by different body shapes can vary significantly, a realism aspect often overlooked in text-to-motion tasks. We propose a novel framework to integrate both shape and motion descriptions as input. Our framework synthesizes the shape parameters to reflect the described physical attributes, and injects them into motion synthesis to generate plausible shape-aware motions. The figure demonstrates the same running motion synthesized across different body shapes. + +difficulties of learning continuous motion features, quantizing motions into discrete tokens has emerged as an area of significant interest, which can then be predicted using transformers [105] or fine-tuning large language models [50]. These advancements have led to improved realism and creative control over motion synthesis. + +However, existing methods typically standardize motions by mapping to a canonical human body model. This results in homogenized motions across diverse body types and leaves out the unique attributes of individual body shapes. In reality, persons with different body shapes perform the same action with distinct physiological differences, as illustrated in Figure 1. Treating these distinct motions identically during motion synthesis leads to artifacts in subsequent motion transfer efforts [6], often resulting in unrealistic motions and limitations on the body shapes on which the motions can be reliably retargeted. + +Incorporating body shapes in motion synthesis is challenging due to the difficulties in both obtaining closed-form parameterization of motions by body shapes and learning + +the coarse-to-fine differences in motions due to individual body shapes in a data-driven manner. The challenges are exacerbated when attempting to merge continuous shape representations with quantized motion representations. To address these hurdles, we propose a quantization-based framework that seamlessly integrates continuous body shape data and thereby manages the information loss due to motion quantization. Consequently, our model not only predicts discrete motion tokens but also accurately captures continuous human body shape data, significantly enhancing the realism and individualism of synthesized motions. + +Our model also leverages the capabilities of large language models to process complex language and predict sequential data, while handling continuous information through its latent space. This dual ability allows our model to predict continuous body shape parameters alongside discrete motion tokens. To maximize the benefits of this dual capability, our task-specific decoder utilizes discrete tokens and continuous shape parameter predictions as conditions to generate shape-aware human motions. + +To summarize, our main contributions include: + +- End-to-end synthesis of both body shape parameters and shape-aware motions from input text descriptions. +- A novel framework to combine continuous shape and quantized motion data for efficient learning. +- Demonstrating efficacy through thorough quantitative and qualitative evaluations and human perceptual studies. + +# 2. Related Work + +We summarize existing research in the related areas of motion transfer, motion generation with and without body shape information, and motion generation from text inputs. + +Motion Transfer. Motion transfer is a common strategy for applying motions from a source skeletal topology to a target one. While heuristic approaches based on inverse kinematics can resolve minor motion artifacts in the target skeleton, they cannot reliably reconcile motion artifacts due to topological mismatches between the source and the target [24, 27, 63, 75]. More direct approaches that require paired source and target motion data are limited by the practical challenges of obtaining such paired data, particularly as the topological differences between the source and the target skeletons increase [1, 97, 108]. Recent data-driven approaches relying on deep learning techniques [6] can alleviate the strict requirements of paired motion data [3, 54], and can even transfer motions between different modalities, such as from videos to 3D meshes [2, 22], while plausibly adhering to physical constraints, such as foot contacts [14, 94]. However, such approaches predominantly rely on skeletal topologies to transfer motions. They either cannot adapt to motion transfer between physiologi + +cally different body shapes or can only do so for a limited variety of body shapes and motions [38, 72, 95, 104]. + +Human Motion Generation with Body Shapes. We broadly categorize computational methods for human motion generation into shape-aware — those that consider body shapes in the generation process, and shape-free — those that generate motions for a normalized representation of the human body and use motion transfer techniques to apply the generated motions to different body shapes. Owing to the simpler nature of its formulation, shape-free motion generation has been richly explored in computer vision and graphics. The underlying motion models are often developed with physically-based constraints [15, 30, 80, 81, 87, 98–101, 103] and employ reinforcement learning frameworks for use within physics engines [61, 65–67, 77, 78, 96]. In addition to physical constraints, generative approaches are typically conditioned on guidance signals, such as past motions of the same person [52] interactive motions of another person [35, 84], audio, including music [16, 55] and speech [18, 37], text descriptions [5, 32, 33], and their various combinations [17]. The scope of these approaches has expanded from deterministic, producing a fixed output for a given set of guidance signals [8, 12, 29, 36, 39, 41, 42, 48], to stochastic, enabling more creative control over the generated motions [9, 43, 46, 74]. Generative frameworks such as GANs and VAEs, paired with the creation of standardized human representations [56] and large-scale benchmark datasets [60], have heralded the modern wave of motion generation methods. They have significantly improved the plausibility of generated motions on canonical skeletons [4, 13, 73], and have been expanded to object [34] and scene interactions [51]. More recent methods have leveraged versatile sequence modeling architectures such as transformers [58] and robust generative techniques such as denoising diffusion models [26]. Despite the remarkable progress in shape-free motion generation, their effective realism, when applied to diverse body shapes, remains susceptible to common motion artifacts, including self-intersections, foot contact errors, and unnatural pose articulations. Shape-aware pose models [25, 40, 109] and motion generation methods [88] have made progress in reducing this quality gap by explicitly incorporating body shape information in the generative process and employing physically-based constraints that support the shape information. Our work takes the next step in shape-aware motion generation by learning to generate motions from natural language descriptions of both shapes and motions in an end-to-end fashion. + +Motion from Language Models. Motion generation from text descriptions has become especially active in recent human motion generation research, aided largely by + +![](images/2bbd5b2381f5e0ed429834d66491998d1a322acf377e17c58135719213a6d5c8.jpg) +Figure 2. Shape-Aware FSQ-VAE (SA-VAE) Overview. SA-VAE is our quantization network learning to generate discrete motion tokens. Given a shape-normalized motion $X^{N} \in \mathbb{R}^{T \times D}$ of length $T$ and dimensionality $D$ ( $= 263$ in our setup), we first encode the motion with the Motion Encoder $\mathcal{E}$ into a motion feature $Z \in \mathbb{R}^{\tau \times D}$ , where $\tau$ represents a downsampling of $T$ . We leverage the FSQ [62] quantizer to quantize $Z$ , which gives a discrete feature $\hat{Z}$ . The $\mathrm{MLP}_{\theta_m}$ and $\mathrm{MLP}_{\phi_m}$ transform the features into the required code dimensions. To condition on the shape, we project the shape parameter $\beta$ with the Projector $P_{\theta_s}$ to align with $\hat{Z}$ . We concatenate the shape feature $\tilde{\beta}$ with $\hat{Z}$ , then feed it into the Motion Decoder $\mathcal{D}$ to predict the reconstructed motion $\hat{X}^R$ . + +the user-friendliness and creative utility of describing motions in free-form text via language models [5, 10, 11, 44, 68, 69, 85, 107]. Current state-of-the-art approaches typically employ diffusion-based [23, 76, 79, 86, 106] networks or transformer-based [45, 50, 105] networks that map tokenized language features to motion features. In the quantization step, motion features are also often quantized and tokenized, following quantization paradigms such as the VQ-VAE [93], to expedite the generative learning process while maintaining motion quality. We inject shape-awareness into the motion tokenization process through data-driven shape features and finite scalar quantization (FSQ) [62], enabling us to map language tokens to combine shape and motion features for shape-aware motion generation from text. + +# 3. Method + +We aim to generate both human body shapes and motions from natural language text descriptions. To this end, we first preprocess text-to-motion data to obtain shape descriptions and shape-aware motion data (Section 3.1). We propose a two-stage framework for shape-aware motion synthesis from text, consisting of a shape-aware FSQ-VAE (SA-VAE) and a shape-motion token predictor (ShapeMove). The SA-VAE efficiently quantizes shape-normalized motion into discrete tokens and reconstructs motions from shape parameters (Section 3.2). ShapeMove generates both shape tokens and a sequence of motion tokens that accurately correspond to the text description (Section 3.3). + +# 3.1. Data Preprocessing + +Existing text-motion datasets only provide natural language labels to describe motions [43, 70]. To obtain specific text + +labels for body shape, we utilize the SMPL [56] model's shape parameters $\beta$ , which control the body shape, and extract the measurements of six attributes to get a holistic view of the body: height, arm and leg lengths, and chest, waist, and hip circumferences. These attributes form our shape description, which we use as input for the model. + +Since available text-motion datasets contain a limited diversity of characters, we propose generating additional body shapes for training. We utilize the Attributes to Shape (A2S) model of Shapy [25], which can generate SMPL beta values from linguistic shape attributes to generate more body shapes in a controlled and plausible manner. We detail the pipeline for using additional shapes in Section 3.2. + +For the motion representation, we follow the preprocessing procedure of Guo et al. [43], extracting joint positions from the SMPL layer and computing motion representations with shape $T \times D$ , where $D = 263$ is the data dimension and $T$ is the number of motion frames. Each motion representation contains root relative rotation and velocity, root height, joint locations, velocities, rotations, and foot contact labels. Notably, we skip the normalizing step of canonicalizing the skeletons to a homogeneous body shape. We denote this as the shape-aware motion $X^R$ , and the corresponding shape-normalized version of the motion as $X^N$ . + +# 3.2. Shape-Aware FSQ-VAE (SA-VAE) + +Quantizing motion into discrete tokens has led to significant progress in the text-to-motion synthesis [44, 50, 105]. Existing approaches quantize motion into discrete tokens and then predict sequences of motion tokens that align with the text input. However, the codebook capacities of these approaches are insufficient for reconstructing shape-variant + +![](images/39c01c00d2d29971b72f85c6f8d68d9b3143abab5ea68b8732c85d5c3d74ae0b.jpg) +(a) Training + +![](images/ccce03c99ee3333b45d9225e93f6bfd99f04eb665d6a947feecdbb143d9979ac.jpg) +(b) Inference +Figure 3. ShapeMove Overview. In the training phase (a), the transformer network takes in the text inputs describing human motions and body shapes and predicts quantized motion tokens and the shape token [BETA]. The embedding for [BETA] passes through the Projector $P_{\theta_e}$ to predict the shape parameter $\hat{\beta}$ . We use cross-entropy loss for comparing ground truth tokens $C$ with predicted tokens $\hat{C}$ , and $L1$ loss for shape parameter to optimize the model. In the inference phase (b), our model predicts motion tokens $\hat{C}$ and the shape parameter $\hat{\beta}$ from text inputs. We de-quantize these tokens using FSQ, and project into shape parameters with Projector $P_{\theta_s}$ . We concatenate $\hat{\beta}$ and $\hat{C}$ , and decode into the generated motion sequence with the Motion Decoder $\mathcal{D}$ . Our model effectively synthesizes shape parameters and shape-aware motions reflecting the physical form and actions described in the input text. + +motions $(X^{R})$ . To address this limitation, we propose a new framework, Shape-Aware FSQ-VAE (SA-VAE), to effectively quantize the shape-normalized motions $X^{N}$ into discrete tokens and reconstruct shape-aware motions $\hat{X}^{R}$ by conditioning on shape information. We use $X^{N}$ instead of $X^{R}$ as the SA-VAE input as it enables the SA-VAE to disentangle the content (given by the pose sequences) and the style (given by the body shapes) components of the motions, focusing on learning tokens for the motion content and shape features for the motion style separately, and then combining them for efficient decoding into $X^{R}$ . + +As illustrated in Figure 2, our SA-VAE comprises a motion encoder $\mathcal{E}$ and a motion decoder $\mathcal{D}$ . To learn a more generalized motion codebook, the encoder $\mathcal{E}$ takes in a sequence of shape-normalized motion frames $X^{N} = [x_{1}^{N}, x_{2}^{N}, \ldots, x_{T}^{N}]$ , $x_{i}^{N} \in \mathbb{R}^{D}$ and encodes them into motion features $Z = [z_{1}, z_{2}, \ldots, z_{\tau}]$ , $z_{i} \in \mathbb{R}^{d}$ and $\tau$ representing a downsampling of $T$ . + +We utilize Finite Scalar Quantization (FSQ) [62] as our quantizer, given its superior codebook utilization without needing the additional regularization typically used in previous quantizers [105]. We denote the quantized features as $\hat{Z} = \mathrm{MLP}_{\phi_m}(\nabla \lfloor f(\mathrm{MLP}_{\theta_m}(Z))\rfloor_{ste})$ , where $f$ is a bounding function that confines the latent within a specific + +range, defined as $f\lfloor L / 2\rfloor \odot \tanh (Z)$ , and $L\in \mathbb{Z}^{\ell}$ determines the width of the codebook for each dimension. We round the latents to integers with straight-through gradients $(\nabla \lfloor \cdot \rfloor_{ste})$ . The MLPs, $\mathrm{MLP}_{\theta_m}$ and $\mathrm{MLP}_{\phi_m}$ , transform the features into required code dimensions. + +Our motion decoder $\mathcal{D}$ decodes the discrete motion features $\hat{Z} = [\hat{z}_1,\hat{z}_2,\dots ,\hat{z}_\tau ]$ $\hat{z}_i\in \mathbb{R}^d$ into shape-aware motions. To incorporate shape information, we extract the shape feature $\tilde{\beta} = \mathrm{P}_{\theta_s}(\beta)$ , where $\mathrm{P}_{\theta_s}$ is an MLP-based projector. We concatenate the shape feature along the time dimension with the discrete motion feature to feed into $\mathcal{D}$ . We denote the final reconstructed shape-aware motion as $\hat{X} = \mathcal{D}\left(\hat{Z}\oplus \tilde{\beta}\right),\oplus$ denoting concatenation. + +Objective. Following the approach of T2M-GPT [105], we use $L_{1}^{\mathrm{smooth}}$ to compute the reconstruction loss $L_{r}$ between the ground truth and the reconstructed motions to train the SA-VAE. Denoting $X_{\mathrm{rot}}$ as the invariant rotation information of $X$ , we write the reconstruction loss as + +$$ +L _ {r} = L _ {1} ^ {\text {s m o o t h}} \left(X, \hat {X}\right) + \lambda_ {\text {r o t}} L _ {1} ^ {\text {s m o o t h}} \left(X _ {\text {r o t}}, \hat {X} _ {\text {r o t}}\right), (1) +$$ + +where $\lambda_{\mathrm{rot}}$ is a hyperparameter to balance the two losses. We also design physical losses to constrain the physical as + +Table 1. Comparison with Baselines. We evaluate the Penetrate, Float, Skate Ratio, and Bone Length Variances for our method and available baselines. For fair comparison, retrain baselines with shape-aware motions. The Shape Input Capability column indicates methods that can incorporate both shape and motion descriptions — a $\times$ here suggests the corresponding text encoder cannot parse shape descriptions. Arbitrary Length denotes results obtained without using ground-truth motion lengths. We compare with methods that share the same constraints as ours, capable of generating arbitrary motion lengths and accepting shape descriptions as input. Our method achieves the best or comparable results across the board. It particularly excels in the Penetrate metric, showcasing its adeptness at handling shape-aware motions. and highlight the best and second-best results. + +
MethodsShape Input CapabilityArbitrary LengthPenetrate (cm) ↓Float (cm) ↓Skate (%) ↓Bone Length Variances ↓RPrecision ↑FID ↓MMDist ↓Diversity ↓
Top1Top2Top3
Real--0.00.03627.4680.00.4690.6650.7690.0013.2170.000
SA-VAE (Recon.)--0.02890.20906.4430.6230.4540.6450.7490.1253.3080.101
TM2T [44]X✓0.14850.24568.5545.3390.3740.5590.6731.6713.8430.937
T2M [43]X✓0.09390.68054.2501.3520.4080.5920.6971.2303.5970.430
MLD [23]✓X0.30910.65589.3132.6950.3830.5710.6800.8823.7360.020
MotionDiffuse [106]✓X0.24010.27037.7100.1380.4260.6160.7230.5633.3920.320
MDM [86]✓X0.10111.71018.5230.6660.3170.4900.5990.4614.1800.320
T2M-GPT [105]✓✓0.17890.52416.1621.1760.3940.5760.6830.2693.7100.190
MotionGPT [50]✓✓0.69860.22457.8892.2710.1280.2080.2711.0207.0550.389
ShapeMove (Ours)✓✓0.02680.26586.1430.6250.4130.6010.7050.1983.5330.117
+ +pects of the reconstructed motion. Similar to [88, 90, 102], we include the floating loss $L_{\mathrm{float}}$ , the foot sliding loss $L_{\mathrm{slide}}$ , and the bone length loss $L_{\mathrm{bone}}$ as geometric losses. $L_{\mathrm{float}}$ minimizes the per-frame distance of the lowest joints above the ground from the ground plane. $L_{\mathrm{slide}}$ minimizes the ground velocity of the foot joint if they are determined to be in ground contact using a distance threshold from the ground. $L_{\mathrm{bone}}$ minimizes the deviation of bone lengths between the ground truth and the reconstructed motions. We write the final objective as + +$$ +L _ {\mathrm {v q}} = L _ {r} + \lambda_ {f} L _ {\text {f l o a t}} + \lambda_ {s} L _ {\text {s l i d e}} + \lambda_ {b} L _ {\text {b o n e}}, \tag {2} +$$ + +where $\lambda_{f},\lambda_{s}$ , and $\lambda_{b}$ are the loss-balancing hyperparams. + +Shape Data Augmentation. To provide our SA-VAE with a wide diversity of subjects with different body shapes, beyond what is available in datasets, we adopt a data augmentation strategy. We replace $q\%$ of the ground-truth shape parameters with synthetically generated ones, as mentioned in Section 3.1. Since there is no corresponding ground truth motion for the augmented shapes, we only optimize the indices with $L_{\mathrm{slide}}$ and $L_{\mathrm{bone}}$ . + +# 3.3. Shape-Aware Motion Synthesis (ShapeMove) + +With the SA-VAE, we can tokenize a sequence of motion $\left[x_1^N,x_2^N,\dots ,x_T^N\right]$ into a sequence of discrete indices $[c_{1},c_{2},\ldots ,c_{\tau}]$ . Subsequently, our framework, ShapeMove, learns to map from language to shape-aware motion representations. We adopt a pretrained language model to predict the sequence of tokens. We illustrate the overview of ShapeMove in Figure 3. To fine-tune the language model to understand the relationships between motion, shape, and text, we expand the original language model's vocabulary + +Table 2. Quantizer Reconstruction Comparison. We report the reconstruction results, comparing our SA-VAE against baseline methods that utilize a VAE to quantize motion into discrete tokens. We assess the bone length difference, jitter score difference, and FID score relative to the shape-aware ground truth motions. Our VAE outperforms the baseline across all three metrics, particularly in reducing the bone length error by nearly half. These results demonstrate our model's effectiveness in aligning with the physical form of different body shapes. and highlight the best and second-best results. + +
MethodsFID ↓Bone Length +Diff. (mm) ↓Jitter +Diff. (m/s2) ↓
TM2T [44]0.528131.0679.87
MotionGPT [50]0.17394.2740.44
T2M-GPT [105]0.15183.4234.75
SA-VAE (Ours)0.12545.8831.49
+ +by adding $k + 2$ motion vocabulary items. Here, $k$ is the number of codes from the quantizer codebook, and the additional two tokens are special tokens to signify the start and end of the motion token sequence. We also add one shape vocabulary token, [BETA], to represent the shape. The language model takes text tokens as input and predicts tokens $\left\{[\mathrm{BETA}], \hat{C}\right\}$ . We design the ground truth tokens to be $\{[\mathrm{BETA}], C\}$ , i.e., we append the [BETA] token with a sequence of motion tokens that align with the text input. We fine-tune the language model in an autoregressive manner. + +Since shape parameters are fixed for a given body shape, unlike motion prediction, we cannot formulate shape parameter prediction as predicting a sequence of tokens. Representing shape would require an impractically large code + +Table 3. Ablation Study. sc stands for shape-conditioning. and highlight the best, second-best and third-best results. + +
MethodsFID ↓Bone Length +Diff. (mm) ↓Float (cm) ↓Skate (%) ↓
No sc0.14899.180.5756.76
sc0.10566.410.5676.60
sc+Lbone0.10745.110.4807.07
sc+Lbone + Lfloat0.13745.970.2557.90
sc+Lbone + Lfloat + Lskate (Full)0.12545.880.2666.14
+ +book, akin to using discrete numbers to fit a continuous space. To address this, previous works show that the last embedding in the language model can contain continuous information [28, 53]. Instead of attempting to learn a discrete representation for shape, we extract the information through embedding. We denote the predicted output token as $\hat{C}_i$ , the corresponding embedding as $M_i$ , and the embedding corresponding to [BETA] as $M_{\beta}$ . We obtain the shape parameter through a projector $\mathrm{P}_{\theta_e}$ , yielding $\hat{\beta} = \mathrm{P}_{\theta_e}(M_\beta)$ . + +Objective. To enable the language model to learn semantically consistent shape and motion vocabularies, we design the training data as mentioned in Section 3.1. The target outputs are the corresponding ground truth shape parameters and motion tokens. We train the model with the following objectives: + +$$ +L _ {\text {t o k e n}} = \operatorname {C r o s s E n t r o p y} (C, \hat {C}) \tag {3} +$$ + +$$ +L _ {\text {s h a p e}} = \lambda_ {\beta} | \beta - \hat {\beta} |, \tag {4} +$$ + +where $L_{\mathrm{token}}$ is the cross-entropy loss maximizing the log-likelihood of predicting the motion tokens, $L_{\mathrm{shape}}$ is the $L_{1}$ norm difference between the ground truth and the estimated shape parameters, and $\lambda_{\beta}$ is the hyperparameter balancing the shape loss. + +# 4. Experiments + +In this section, we begin by discussing the experimental setup in Section 4.1. We then introduce the baselines used for comparison and present the results in Section 4.2, which includes quantitative comparisons, qualitative evaluations, and human perceptual assessments. + +# 4.1. Experimental Setup + +Dataset. We use the current largest text-to-motion dataset, HumanML3D [43], which comprises 14,616 motion sequences and 44,970 motion descriptions. The motion sequences are sourced from 449 different subjects in AMASS [7, 19, 21, 31, 47, 49, 57, 59, 60, 64, 82, 89, 91, 92] and HumanAct12 [20]. We follow the same data preprocessing procedures outlined for HumanML3D, but we omit + +Table 4. Attributes Prediction. We present the differences $(cm)$ across six attributes between our beta predictions and the ground truth, focusing solely on our method as no comparable works predict beta concurrently. Our model demonstrates a robust ability to predict the correct beta values, with discrepancies from the ground truth around one $cm$ . $C$ stands for circumference. + +
AttributesDifferenceAttributesDifferenceAttributesDifference
Height0.5767Arm Length0.0433Leg Length0.0505
Chest C.0.6921Waist C.1.0558Hip C.0.6044
+ +normalizing the motions to a canonical model to get shape-aware motions, as detailed in Section 3.1. + +Implementation Details. To train the SA-VAE, we configure the level $\ell$ of the FSQ quantizer as $\ell = [8,5,5,5]$ , constructing a codebook that contains $k = 1000$ indices with each index dimension set to 512. During training, we crop motion sequences to a length $T = 64$ and set the down-sampled length $\tau = 16$ . We employ the AdamW optimizer with $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ (following standard AdamW notation — no relation to the shape $\beta$ ), along with an exponential moving constant $\lambda = 0.99$ . We train for 200K iterations at a learning rate of $2e^{-4}$ , followed by 100K iterations at $1e^{-5}$ . We set $\lambda_{f}, \lambda_{s}$ , and $\lambda_{b}$ to 10, and $q\%$ for data augmentation to $10\%$ . With a batch size of 256, training takes approximately 12 hours on a single A100 GPU. + +We use T5 [71] as our pretrained language model, which consists of 12 layers in both the transformer encoder and decoder. We employ AdamW optimizer with $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ at a learning rate of $8e^{-4}$ . We set $\lambda_{\beta}$ to 0.5 and train the model for 120K iterations on two distinct tasks: text-to-motion and motion-to-text. With a batch size of 64, this training phase requires one day on 8 A100 GPUs. Subsequently, we train the model solely on the text-to-motion task using the same batch size and learning rate for an additional 30K steps, which takes 10 hours on 8 A100 GPUs. + +Evaluation Metrics. To evaluate the physical plausibility of motions, we adopt the physics-based metrics of [88, 102]. The Penetrate metric measures ground penetration by computing the distance between the ground and the lowest body joints below the ground. Float measures the distance between the ground and the lowest body joints above the ground. Skate measures foot skating by computing the percentage of frames where the foot joints in contact with the ground have a velocity over a threshold. We also propose a new metric, Bone Length Variances, which measures the variance of arm and leg lengths across the entire sequence to evaluate the stability of the generated motion. Following HUMOS [88], we exclude motions involving climbing and compute the Penetrate, Float, and Skate on the remaining dataset. This selection ensures the accuracy and rele + +![](images/dde0d077a345d59774f56a304722186fc93aac885adf02637b564f88913f49a4.jpg) +Motion description: "a person walks forward while twisting their torso side to side." + +![](images/124e9ab96cd03157eb6beaebffdce0cd21b172e2d23180fc9a9148073a42b13a.jpg) +Motion description: "a person walks backwards and stops." +Figure 4. Qualitative Comparisons. We compare our method with three baseline methods, T2M-GPT [105], MotionGPT [50], and MotionDiffuse [106], illustrating two samples from the HumanML3D test set. The motions are colored from light to dark blue to represent progression over time. We highlight issues such as incorrect foot motion and other inaccuracies that do not align with expected motion patterns. Due to space constraints, we omit detailed shape descriptions in the figure. We list the detailed descriptions in the supplementary material. Our method not only generates motions that align with the textual descriptions, but also accurately follows the body attributes and physical dynamics of the ground truth. Additional visual results and detailed comparisons are available in the supplementary material. + +vance of the metrics for motions primarily characterized by ground-based dynamics. + +We follow the evaluation protocol proposed by Guo et al. [43] to evaluate text-motion alignment and motion quality. Fréchet Inception Distance (FID) evaluates the distance of feature distributions between the generated and real motions. $R$ -Precision evaluates the accuracy of matching between texts and motions using the top 1/2/3 retrieval accuracy. Multimodal Distance (MM-Dist) measures the feature distance between motions and texts. Diversity calculates the variance through features extracted from the motions. We compute all results with a $95\%$ confidence interval obtained from 20 repeated runs. + +# 4.2. Comparison with Baselines + +We compare our results with the existing state-of-the-art approaches, including T2M [43], TM2T [44], MLD [23], MotionDiffuse [106], MDM [86], T2M-GPT [105], and MotionGPT [50]. We train these baseline methods using our training data, including shape and motion descriptions, and task the models to predict shape-aware motions. We note that T2M [43] and TM2T [44] employ a specific vocabulary for text encoder training and thus cannot incorporate shape descriptions as input. Therefore, we only input motion descriptions for these models and report the results as a reference. In the following subsections, we present quantitative, qualitative, and perceptual evaluation results, demonstrating that our model achieves state-of-the-art performance across these metrics. + +Quantitative Results. We present the physical plausibility comparison in Table 1. Our method outperforms baseline models that share the same settings, achieving the best performance in the Penetrate, Skating, and Bone Length metrics and ranking second best in the Float metric. This demonstrates our method's effectiveness in learning physical rules. The metrics might not fully reflect the true realism of the generated motion for all motion scenarios. Hence, we also conduct a perception evaluation with human participants for a comprehensive assessment. + +We also report results on text-motion alignment metrics in Table 1, where our model outperforms the same set of baselines across all metrics. Our model also achieves the best FID score over baselines, even outperforming baselines that require ground truth motion length during generation (which has limited practicality in real-world applications). + +We report the quantizer performance in Table 2. TM2T [44], T2M-GPT [105], and MotionGPT [50], all quantize motions into discrete tokens. We present the performance of our SA-VAE in comparison with these models. From Table 2, we focus on the reconstruction quality of the model, emphasizing three metrics: the FID score to assess the motion quality, the bone length difference to measure the skeletal discrepancies between the reconstructed and ground truth motions, and the jitter difference to evaluate the disparity in the mean acceleration between reconstructed and ground truth motions. Our model demonstrates superior performance on all three metrics, achieving the lowest FID score and approximately half the bone length difference, which underscores the effectiveness of our SA-VAE in reconstructing shape-aware motions. + +Figure 5. Perceptual Evaluation. We show the distributions of aggregate responses from annotators on their preferences for samples generated by our method and baseline methods, including MotionDiffuse [106], MotionGPT [50], and T2M-GPT [105], as well as the corresponding ground truth samples. We assess the distributions on three metrics: (a) Shape to Text, how well the body shape matches the text input; (b) Motion to Text, how well the motion matches the text input; and (c) Plausibility of Motion with Shape, how realistic the motions appear for the corresponding body shapes. Across all three metrics, we observe that our method is preferred nearly as much as the ground truth and is favored by approximately $12\%$ to $38\%$ over the baselines. +![](images/54ee7c4aaf3b97a000f8444a185067d1ef6daddc96cc641120889fc4f9bc9fa2.jpg) +Ours Better Both Good Both Bad Compared Method Better + +![](images/4e9aa0a9247516eb32282c60b0c1e4e7b8a126872c02c3b09606da03996566ac.jpg) + +![](images/4d6015234f1e2c5aa806fdc77236ccf09d472ba972f2f057759cacdfd6da419a.jpg) + +As we are the first, to the best of our knowledge, to predict shape parameters concurrently, we exclusively report the differences in body attributes. Table 4 presents these results, showing that our model can predict shape parameters differing by around $1cm$ from the ground truth. + +We also conduct an ablation study to evaluate the impact of the various components used in our model, including shape-conditioning, bone length loss $L_{\mathrm{bone}}$ , float loss $L_{\mathrm{float}}$ , and foot skate loss $L_{\mathrm{skate}}$ . We assess these variants on FID, bone length difference, floating, and foot skate ratio. Table 3 shows how each component improves performance across the metrics. We select our final model configurations as the one achieving the best overall results. + +Qualitative Results. Figure 4 illustrates a qualitative comparison on two samples from the HumanML3D test set. Our method can generate realistic motions that are well-aligned with both shape and motion descriptions. In contrast, body shapes in other methods are obtained indirectly via fitting the SMPL model to the generated joints. The red circles highlight unrealistic motions, such as floating feet, misaligned postures, incomplete actions, self-intersections, and deviating from the prompt. More visual results are available on the website. + +Perceptual Evaluation. We conducted a human perceptual evaluation with pairwise comparisons between motion samples generated by our method and other baselines, as well as the corresponding ground truth samples. We selected the three best baselines from our quantitative evaluations, which also follow the same generation paradigm as ours, for comparison: MotionDiffuse [106], MotionGPT [50], and T2M-GPT [105]. For each pairwise comparison, we show our sample and the other sample (one of the baselines or ground truth), for the same text description, in a random order to the participants. We ask them to rank the two samples as one of sample 1 better, sam + +ple 2 better, both similarly good or both similarly bad on three metrics: shape-to-text accuracy, motion-to-text accuracy, and plausibility of motion with shape. We randomly selected 34 text descriptions from the HumanML3D test set to generate the sample sets for the perceptual study, and recruited participants through Amazon Mechanical Turk. After filtering out inadmissible study responses (e.g., responses that were too quick and self-inconsistent), we report our perceptual evaluation results with 45 respondents in Figure 5. We note that our method was consistently preferred over the baselines and even matched the preference rate of the ground truth. + +# 5. Limitations + +One limitation of our approach is the constrained nature of the shape descriptions used for motion generation. The preprocessing step requires descriptions to adhere to a specific template, which may restrict the model's ability to handle various natural language inputs and diverse body shapes. This limitation could be mitigated by incorporating a more diverse dataset with a broader range of descriptive styles. Additionally, employing an advanced language model to preprocess and standardize descriptions to fit the required template can enhance the model's versatility and real-world applicability. + +# 6. Conclusions + +We have presented a novel framework for shape-aware motion synthesis from textual descriptions, with the ability to generate both shape parameters and shape-aware motions from input text descriptions. Our method effectively captures the diverse correlations between human body shapes and motion dynamics, as evidenced by comprehensive evaluations that include quantitative results, qualitative comparisons, and perceptual studies with humans. + +# References + +[1] M. Abdul-Massih, I. Yoo, and B. Benes. Motion style retargeting to characters with different morphologies. Computer Graphics Forum, 36(6):86-99, 2017. 2 +[2] Kfir Aberman, Rundi Wu, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Learning character-agnostic motion for motion retargeting in 2d. ACM Trans. Graph., 38(4), 2019. 2 +[3] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Trans. Graph., 39(4), 2020. 2 +[4] Hyemin Ahn, Timothy Ha, Yunho Choi, Hwiyeon Yoo, and Songhwai Oh. Text2action: Generative adversarial synthesis from language to action. In ICRA, 2018. 2 +[5] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728, 2019. 1, 2, 3 +[6] Syed Muhammad Abrar Akber, Sadia Nishat Kazmi, Syed Muhammad Mohsin, and Agnieszka Szczesna. Deep learning-based motion style transfer tools, techniques and future challenges. Sensors, 23(5), 2023. 1, 2 +[7] Ijaz Akhter and Michael J. Black. Pose-conditioned joint angle limits for 3D human pose reconstruction. In CVPR, 2015. 6 +[8] Emre Aksan, Manuel Kaufmann, and Otmar Hilliges. Structured prediction helps 3d human motion modelling. In ICCV, 2019. 2 +[9] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In CVPR, 2020. 2 +[10] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal Action Compositions for 3D Humans. In 3DV, 2022. 3 +[11] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. Sinc: Spatial composition of 3d human motions for simultaneous action generation. In ICCV, 2023. 3 +[12] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. In ICLR, 2022. 2 +[13] Emad Barsoum, John Kender, and Zicheng Liu. Hp-gan: Probabilistic 3d human motion prediction via gan. In CVPR Workshops, 2018. 2 +[14] Jean Basset, Stefanie Wuhrer, Edmond Boyer, and Franck Multon. Contact preserving shape transfer for rigging-free motion retargeting. In ACM SIGGRAPH Conference on Motion, Interaction and Games, 2019. 2 +[15] Kevin Bergamin, Simon Clavet, Daniel Holden, and James Richard Forbes. Drecon: data-driven responsive control of physics-based characters. ACM Trans. Graph., 38(6), 2019. 2 +[16] Aneesh Bhattacharya, Manas Paranjape, Uttaran Bhattacharya, and Aniket Bera. Danceanyway: synthesizing beat-guided 3d dances with randomized temporal con + +trastive learning. In Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence. AAAI Press, 2024. 2 +[17] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In ACM MM, 2021. 2 +[18] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In Virtual Reality and 3D User Interfaces (VR), 2021. 2 +[19] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 6 +[20] Judith Butepage, Michael J Black, Danica Kragic, and Hedvig Kjellstrom. Deep representation learning for human motion prediction and classification. In CVPR, 2017. 6 +[21] Carnegie Mellon University. CMU MoCap Dataset, 2003. 6 +[22] Ufuk Celikcan, Ilker O. Yaz, and Tolga Capin. Example-Based Retargeting of Human Motion to Arbitrary Mesh Models. Computer Graphics Forum, 2015. 2 +[23] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 3, 5, 7 +[24] Kwang-Jin Choi and Hyeong-Seok Ko. Online motion retargeting. The Journal of Visualization and Computer Animation, 11(5):223-235, 2000. 2 +[25] Vasileios Choutas, Lea Müller, Chun-Hao P Huang, Siyu Tang, Dimitrios Tzionas, and Michael J Black. Accurate 3d body shape regression using metric and semantic attributes. In CVPR, 2022. 2, 3 +[26] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, 2023. 2 +[27] Unreal Engine. dev.epicgames.com/documentation/en-us/unreal-engine/animation-retargeting-in-unreal-engine, 2015. 1, 2 +[28] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. ChatPose: Chatting about 3d human pose. In CVPR, 2024. 6 +[29] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In ICCV, 2015. 2 +[30] Levi Fussell, Kevin Bergamin, and Daniel Holden. Supertrack: motion tracking for physically simulated characters using supervised learning. ACM Trans. Graph., 40(6), 2021. 2 +[31] Saeed Ghorbani, Kimia Mahdaviani, Anne Thaler, Konrad Kording, Douglas James Cook, Gunnar Blohm, and Nikolaus F. Troje. MoVi: A large multipurpose motion and video dataset. arXiv preprint arXiv: 2003.01888, 2020. 6 +[32] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In ICCV, 2021. 2 + +[33] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Text-based motion synthesis with a hierarchical two-stream rn. In ACM SIGGRAPH 2021 Posters, New York, NY, USA, 2021. Association for Computing Machinery. 2 +[34] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. Computer Graphics Forum, 42(2):1-12, 2023. 2 +[35] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. In ECCV, 2024. 2 +[36] Partha Ghosh, Jie Song, Emre Aksan, and Otmar Hilliges. Learning human motion models for long-term predictions. In 3DV, 2017. 2 +[37] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In CVPR, 2019. 2 +[38] Thiago L. Gomes, Renato Martins, João Ferreira, Rafael Azevedo, Guilherme Torres, and Erickson R. Nascimento. A shape-aware retargeting approach to transfer human motion and appearance in monocular videos. Int. J. Comput. Vision, 129(7):2057-2075, 2021. 2 +[39] Anand Gopalakrishnan, Ankur Mali, Dan Kifer, Lee Giles, and Alexander G. Ororbia. A neural temporal model for human motion prediction. In CVPR, 2019. 2 +[40] Omer Gralnik, Guy Gafni, and Ariel Shamir. Semantify: Simplifying the control of 3d morphable models using clip. In ICCV, 2023. 2 +[41] Ulf Grenander and Michael I. Miller. Representations of knowledge in complex systems. Journal of the Royal Statistical Society: Series B (Methodological), 56(4):549-581, 1994. 2 +[42] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In ACM MM, 2020. 2 +[43] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 2, 3, 5, 6, 7 +[44] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, 2022. 3, 5, 7 +[45] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, 2023. 3 +[46] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NIPS, 2022. 2 +[47] Fabian Helm, Nikolaus Troje, Mathias Reiser, and Jörn Munzert. Bewegungsanalyse getäuschter und nicht-getäuschter 7m-würfe im handball. 47. Jahresstagung der Arbeitsgemeinschaft für Sportpsychologie, Freiburg., 2015. 6 + +[48] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Trans. Graph., 35(4), 2016. 2 +[49] Ludovic Hoyet, Kenneth Ryall, Rachel McDonnell, and Carol O'Sullivan. Sleight of hand: Perception of finger motion from reduced marker sets. In ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games, 2012. 6 +[50] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. NIPS, 2024. 1, 3, 5, 7, 8 +[51] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In CVPR, 2024. 2 +[52] Taku Komura, Ikhsanul Habibie, Daniel Holden, Jonathan Schwarz, and Joe Yearsley. A recurrent variational autoencoder for human motion synthesis. In BMVC, 2017. 2 +[53] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In CVPR, 2024. 6 +[54] Sunmin Lee, Taeho Kang, Jungnam Park, Jehee Lee, and Jungdam Won. Same: Skeleton-agnostic motion embedding for character animation. In SIGGRAPH Asia Conference Papers, 2023. 2 +[55] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In ICCV, 2021. 2 +[56] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multi-person linear model. ACM Trans. Graph., 34(6), 2015. 2, 3 +[57] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 33(6):220:1-220:13, 2014. 6 +[58] Zhenyu Lou, Qiongjie Cui, Haofan Wang, Xu Tang, and Hong Zhou. Multimodal sense-informed forecasting of 3d human motions. In CVPR, 2024. 2 +[59] Eyes JAPAN Co. Ltd. Eyes Japan MoCap Dataset, 2008. 6 +[60] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In ICCV, 2019. 2, 6 +[61] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac gym: High performance GPU based physics simulation for robot learning. In NIPS Datasets and Benchmarks Track, 2021. 2 +[62] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple, 2023. 3, 4 +[63] MotionBuilder. autodesk.com/products/motionbuilder/overview, 2023. 1, 2 +[64] M. Müller, T. Röder, M. Clausen, B. Eberhardt, B. Krüger, and A. Weber. Documentation mocap database hdm05. Technical Report CG-2007-2, Universität Bonn, 2007. 6 + +[65] Xue Bin Peng and Michiel van de Panne. Learning locomotion skills using deeprl: does the choice of action space matter? In ACM SIGGRAPH SCA, 2017. 2 +[66] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel van de Panne. Deepmimic: example-guided deep reinforcement learning of physics-based character skills. ACM Trans. Graph., 37(4), 2018. +[67] Xue Bin Peng, Angjoo Kanazawa, Jitendra Malik, Pieter Abbeel, and Sergey Levine. Sfv: reinforcement learning of physical skills from videos. ACM Trans. Graph., 37(6), 2018. 2 +[68] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In ICCV, 2021. 3 +[69] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In ECCV, 2022. 3 +[70] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. Big Data, 4(4):236-252, 2016. 3 +[71] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 6 +[72] João Regateiro and Edmond Boyer. Temporal shape transfer network for 3d human motion. In 3DV, 2022. 2 +[73] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In ICCV, 2021. 2 +[74] Zhiyuan Ren, Zhihong Pan, Xin Zhou, and Le Kang. Diffusion motion: Generate text-guided 3d human motion by diffusion model. In ICASSP, 2023. 2 +[75] Rokoko. rokoko.com, 2023. 1, 2 +[76] Alessio Sampieri, Alessio Palma, Indro Spinelli, and Fabio Galasso. Length-aware motion synthesis via latent diffusion. In ECCV, 2024. 3 +[77] John Schulman, Philipp Moritz, Sergey Levine, Michael I. Jordan, and P. Abbeel. High-dimensional continuous control using generalized advantage estimation. In ICLR, 2016. 2 +[78] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2 +[79] Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 3 +[80] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, and Christian Theobalt. Physcap: physically plausible monocular 3d motion capture in real time. ACM Trans. Graph., 39 (6), 2020. 2 +[81] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, Patrick Pérez, and Christian Theobalt. Neural monocular 3d human motion capture with physical awareness. ACM Trans. Graph., 40(4), 2021. 2 +[82] L. Sigal, A. Balan, and M. J. Black. HumanEva: Synchronized video and motion capture dataset and baseline algo + +rithm for evaluation of articulated human motion. International Journal of Computer Vision, 87(1):4-27, 2010. 6 +[83] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 39(4), 2020. 1 +[84] Kewei Sui, Anindita Ghosh, Inwoo Hwang, Jian Wang, and Chuan Guo. A survey on human interaction motion generation. arXiv preprint arXiv:2503.12763, 2025. 2 +[85] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, 2022. 1, 3 +[86] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 3, 5, 7 +[87] Shashank Tripathi, Lea Müller, Chun-Hao P. Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. 3d human pose estimation via intuitive physics. In CVPR, 2023. 2 +[88] Shashank Tripathi, Omid Taheri, Christoph Lassner, Michael Black, Daniel Holden, and Carsten Stoll. *Hu-mos: Human motion model conditioned on body shape*. In *ECCV*, 2024. 2, 5, 6 +[89] Matt Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, and John Collomosse. Total Capture: 3d human pose estimation fusing video and inertial sensors. In BMVC, 2017. 6 +[90] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In CVPR, 2023. 5 +[91] ACCAD/Ohio State University. Advanced Computing Center for the Arts and Design, 2024. 6 +[92] Simon Fraser University and National University of Singapore. SFU Motion Capture Database, 2011. 6 +[93] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NIPS, 2017. 3 +[94] Ruben Villegas, Duygu Ceylan, Aaron Hertzmann, Jimei Yang, and Jun Saito. Contact-aware retargeting of skinned motion. In ICCV, 2021. 2 +[95] Jiashun Wang, Chao Wen, Yanwei Fu, Haitao Lin, Tianyun Zou, Xiangyang Xue, and Yinda Zhang. Neural pose transfer by spatially adaptive instance normalization. In CVPR, 2020. 2 +[96] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. A scalable approach to control diverse behaviors for physically simulated characters. ACM Trans. Graph., 39(4), 2020. 2 +[97] Katsu Yamane, Yuka Ariki, and Jessica Hodgins. Animating Non-Humanoid Characters with Human Motion Data. In ACM SIGGRAPH SCA, 2010. 2 +[98] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In CVPR, 2022. 2 +[99] Ye Yuan and Kris Kitani. Dlow: Diversifying latent flows for diverse human motion prediction. In ECCV, 2020. +[100] Ye Yuan and Kris Kitani. Residual force control for agile human behavior imitation and extended motion synthesis. In NIPS, 2020. + +[101] Ye Yuan, Shih-En Wei, Tomas Simon, Kris Kitani, and Jason Saragih. Simpoe: Simulated character control for 3d human pose estimation. In CVPR, 2021. 2 +[102] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, 2023. 5, 6 +[103] Petrissa Zell, Bastian Wandt, and Bodo Rosenhahn. Joint 3d human motion capture and physical analysis from monocular videos. In CVPR Workshops, 2017. 2 +[104] Jiaxu Zhang, Junwu Weng, Di Kang, Fang Zhao, Shaoli Huang, Xuefei Zhe, Linchao Bao, Ying Shan, Jue Wang, and Zhigang Tu. Skinned motion retargeting with residual perception of motion semantics & geometry. In CVPR, 2023. 2 +[105] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 1, 3, 4, 5, 7, 8 +[106] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3, 5, 7, 8 +[107] Zeyu Zhang, Akide Liu, Ian Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. Motion mamba: Efficient and long sequence motion generation. In ECCV, 2024. 3 +[108] Keyang Zhou, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Unsupervised shape and pose disentanglement for 3d meshes. In ECCV, 2020. 2 +[109] Baldomero R. Árbol and Dan Casas. Bodyshapegpt: Smpl body shape manipulation with llms. In ECCV Workshop on Foundation Models for 3D Humans, 2024. 2 \ No newline at end of file diff --git a/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/images.zip b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4b3a277c92c103208b12807ed61f6e798f9285d2 --- /dev/null +++ b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:245de1ce8929d550155c4915891e97e10af8bdd0058038736dffa84b3c643baa +size 437603 diff --git a/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/layout.json b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a7f89a62a57688ea32042392b01e7e5b2e3e7b2 --- /dev/null +++ b/CVPR/2025/Shape My Moves_ Text-Driven Shape-Aware Synthesis of Human Motions/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:556e80e930907ad6c8acbfe9917961b79fe95897fcbbe9d19563a3d38b6c9be2 +size 492099 diff --git a/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/3e3de2dd-3f94-482d-943a-5ecc9a81029e_content_list.json b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/3e3de2dd-3f94-482d-943a-5ecc9a81029e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6ce072049a83ac92df6305796b9a7fb9b9b8b16f --- /dev/null +++ b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/3e3de2dd-3f94-482d-943a-5ecc9a81029e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4302dba182fc3ca6f810ec2200c301e694ce5956936d305f0d490cb5b5a53553 +size 76568 diff --git a/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/3e3de2dd-3f94-482d-943a-5ecc9a81029e_model.json b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/3e3de2dd-3f94-482d-943a-5ecc9a81029e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b69951b5081af34c0e0935dba2885cdd1396b5ae --- /dev/null +++ b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/3e3de2dd-3f94-482d-943a-5ecc9a81029e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:960559c2b682f7019e7ae629f627b7f977c5448962a56f4d6de9c25fdb29c77e +size 95944 diff --git a/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/3e3de2dd-3f94-482d-943a-5ecc9a81029e_origin.pdf b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/3e3de2dd-3f94-482d-943a-5ecc9a81029e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f6e775369bc4e880c13b6c2a75ffcf7f08d2594b --- /dev/null +++ b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/3e3de2dd-3f94-482d-943a-5ecc9a81029e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35cd25d68698b7b87c9bbcf2112c9ae96b6a5d7dc1bffcb4bdb3e92e9adc433c +size 9288617 diff --git a/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/full.md b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..981d4c7ada0526776ee788623680c08dd66971a8 --- /dev/null +++ b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/full.md @@ -0,0 +1,368 @@ +# Shape and Texture: What Influences Reliable Optical Flow Estimation? + +Libo Long, Xiao Hu, Jochen Lang +EECS, University of Ottawa +{ llong014, xhu008, jlang}@uottawa.ca + +# Abstract + +Recent methods have made significant progress in optical flow estimation. However, the evaluation of these methods focuses mainly on improved accuracy in benchmarks and often overlooks the analysis of network robustness, which may be important in safety-critical scenarios such as autonomous driving. In this paper, we propose a novel method for robustness evaluation by modifying data from original benchmarks. Unlike previous benchmarks that focus on complex scenes, we propose to modify shape and texture of objects from the original images in order to analyze the sensitivity to these changes observed in the output. Our aim is to identify common failure cases of state-of-the-art (SOTA) methods to evaluate their robustness and understand their behaviors. We show that: Optical flow methods are more sensitive to shape changes than to texture changes; and optical flow methods tend to "remember" objects seen during training and may "ignore" the motion of unseen objects. Our experimental results and findings provide a more in-depth understanding of the behavior of recent optical flow methods. (See https://github.com/llesky/Robustness-Analysis-Optical-Flow). + +# 1. Introduction + +Optical flow estimation, which aims to estimate pixel-level motion in videos, is fundamental for high-level computer vision applications, such as autonomous driving [5, 24, 37], video editing [2, 16, 18, 29], and 3DGS [10, 48]. Given their widespread use, evaluating the robustness of optical flow methods is crucial considering safety. + +To enhance the robustness of optical flow methods, various solutions have been proposed. Some studies have addressed occlusion maps using dedicated networks [14, 23, 30]. Others have incorporated attention mechanisms to improve performance, particularly for large displacements [39]. Recently, studies have begun to replace convolutional neural networks (CNNs) with transformer blocks to further improve accuracy [13, 25, 38]. However, a systematic analysis of the robustness of deep models in optical + +flow is still lacking. + +Inspired by [8, 15], we start by analyzing how shape and texture affect matching in optical flow. SOTA optical flow methods estimate optical flow by relying on matching pixels (i.e., 4D cost volume) between input images. However, there has been limited study to explain how this matching process works. Motivated by this gap, we introduce FlowR, a synthetic test set generated by a diffusion model for optical flow tasks. Unlike existing datasets, we modify the shape and texture of a key object (e.g., a car) in the KITTI benchmark [7] and evaluate the changes in the output compared to the original results. This allows us to measure the sensitivity of the method to certain types of semantic change in the input. Through this, we aim to explore two fundamental questions about optical flow methods: + +1. Shape vs. Texture: Do optical flow methods estimate motion primarily based on an object's shape or texture? +2. Memorization: Do these methods tend to memorize objects from previously seen scenes? + +Our Flow-R test set includes three subsets: Flow-R-T for texture replacement, Flow-R-S for shape replacement, and Flow-R-O for adding objects changing both shape and texture. Figure 1 presents two samples of data. We develop an application using the Segment Anything Model (SAM) [19] to generate the corresponding masks of the main objects in each pair of images. We then manually select high-quality samples. + +To build Flow-R-T, we apply a set of 24 common texture prompts to ControlNet with specific masks to generate new textures for an object. We evaluate the performance of SOTA optical flow methods on Flow-R-T and find that these methods can still match the motion of an object with entirely different textures. + +To build Flow-R-S, we directly scale the selected object in the image by a random factor $\theta$ and assign an image motion to the object based on its real non-linear motion. We demonstrate that optical flow methods rely more on shape information than texture, and changes of the object flow can significantly affect the optical flow of the background. + +To build Flow-R-O, we introduce a set of 13 unseen objects, i.e., objects that do not occur in the training data, and + +![](images/79f78c9ff9ff16dc5e46735049be40dadeecdb880a593c30f3772e3ac3b83fcd.jpg) +Figure 1. Samples of Flow-R. Original and modified images (Row 1 and 3), ground truth optical flows (Row 2 and 4) are shown. The first column displays the original data from the KITTI dataset. The second column shows the data with the texture of the main object of car replaced. The third column shows the data with the shape of the main object rescaled. The fourth row presents the data with an unseen object added. + +22 seen objects. We randomly select and add one of these objects to a KITTI image. By comparing the performance of the optical flow method on images with seen vs. unseen objects, we show that SOTA methods can easily estimate the motion of seen objects but tend to ignore the motion of unseen objects. + +# 2. Related Works + +# 2.1. Optical Flow Estimation + +Recent optical flow methods [14, 40, 43, 45] have achieved significant progress in benchmarks [7]. RAFT [41] proposes a 4D cost volume and iterative refinement architecture as a foundational structure for recent methods. CRAFT [39] introduces a cross-attention mechanism to revitalize the cost volume for large displacements. GMFlow [42] proposes directly matching the global motion, then propagating local flow with self-attention. FlowFormer [13] introduces a transformer-based architecture for optical flow estimation. FlowFormer++ [38] enhances the cost volume through masked autoencoding. Recently, SAMFlow [47] embeds the encoder of SAM [19] into FlowFormer to address the fragmentation problem. + +However, the primary evaluation of optical flow methods is by measuring performance on benchmarks (e.g., [1]). Common benchmark datasets include synthetic Chairs [4] and FlyingThings3D [26] as well as the computer graphics (CG) based MPI Sintel [3] and real-world KITTI [7, 28]. + +Further datasets have been produced with a game engine, e.g., Virtual KITTI [6] or with the KUBRIC [9] simulator, or based on a CG movie such as the high-resolution SPRING [27]. Real-world datasets are more challenging to collect. HD1K [20] contains driving scenarios but different from KITTI includes various challenging imaging conditions. None of these datasets are designed to study the sensitivity of an optical flow method to changes in the texture and shape of objects in the scene or the robustness of optical flow to unseen objects. Unlike previous studies, we focus with Flow-R on identifying the common "weaknesses" of SOTA methods. + +# 2.2. Adversarial Attacks in Optical Flow Estimation + +To study the robustness of optical flow estimation, many adversarial attack methods have been proposed. Ranjan et al. [31] proposed learning a small patch to attack optical flow methods. Schrodi et al. [36] introduced a less constrained global attack on optical flow, inspired by the Iterative Fast Gradient Sign Method (I-FGSM) [22]. Koren et al. [21] proposed corrupting the flow predictions of specific object categories using a regularization term that promotes consistency in off-target predictions. Schmalfuss et al. [34] proposed the Perturbation-Constrained Flow Attack (PCFA), which emphasizes the destructiveness of adversarial perturbations while maintaining control over their magnitude. They also proposed adversarial weather attacks [35], which simulate weather noise such as snowflakes + +using a 3D rendering framework. + +Those adversarial attack methods all rely on training and adding perturbations to make the network fail. In contrast, our methods directly add real-world objects instead of perturbations and do not require any training. Most adversarial attack methods aim to explore "worst-case" robustness with specifically designed patches or perturbations, which are unlikely to occur by chance in real-world scenarios. The goal of an adversarial attack is to make the network fail. In contrast, we analyze cases that optical flow methods may face in real-world situations even without an adversary. Our goal is to analyze when and how optical flow methods fail in general scenarios. + +# 3. Methods + +# 3.1. Data Manipulation + +Figure 2 indicates the main pipeline of our methods. To construct Flow-R, we utilize the Segment Anything Model (SAM) [19] to identify the corresponding pairs of main objects across images. Given a pair of images $I_{1}, I_{2}$ , we generate a mask with SAM for object $x$ : + +$$ +\left(S _ {1}, S _ {2}\right) = \operatorname {S A M} \left(\left(I _ {1}, I _ {2}\right), \left(P _ {1}, P _ {2}\right)\right) \tag {1} +$$ + +Where the prompts $P_{1}$ and $P_{2}$ are sets of indices corresponding to object $x$ in each image. $S_{1}$ and $S_{2}$ are binary masks for object $x$ in $I_{1}$ and $I_{2}$ . + +Flow-R consists of three parts. Flow-R-T includes images with different textures of the same object, which we use to evaluate the performance of methods with unseen textures. Flow-R-S modifies the shape of the object by scaling by a factor $\theta$ . Flow-R-O adds new unseen objects or replaces object $x$ with a new object in $I_{1}$ and $I_{2}$ . + +Texture Replacement (Flow-R-T). We employ a stable diffusion model [32] to generate new textures for object $x$ . The diffusion model is capable of producing diverse textures but our approach remains compatible with other generative models that can be directed through language prompts. + +To generate a texture with the same shape as the object, we incorporate ControlNet [46] to guide the generation process. We generate the new texture for $I_{2}$ as follows + +$$ +I _ {N} = \operatorname {C o n t r o l N e t} (N, S _ {2}) \tag {2} +$$ + +where $N$ represents the texture within the region defined by $S_{2}$ . $S_{2}$ acts as a prompt to focus ControlNet on generating an object with the same shape. The new $I_2^N$ with texture $N$ can be presented as: + +$$ +I _ {2} ^ {N} = I _ {2} \times (1 - S _ {2}) + I _ {N} \times S _ {2} \tag {3} +$$ + +To preserve consistency of optical flow, we use ground-truth optical flow $f_{gt}$ to warp the new object from $I_2^N$ to + +$$ +I _ {1} ^ {N}. +$$ + +$$ +I _ {1} ^ {N} = I _ {1} \times (1 - S _ {1}) + w a r p \left(I _ {2} ^ {N}, f _ {g t}\right) \times S _ {1} \tag {4} +$$ + +Finally, we use Bi-directional Hole Filling [11] to fill the holes caused by occlusion. The optical flow remains the same as the original. The new generated data is $(I_1^N, I_2^N, f_{gt})$ . + +Shape Replacement (Flow-R-S). We scale the object shape $x$ by a factor $\theta$ enabling the use of its real motion instead of adding unreal motion as in [26]. Specifically, we directly upscale the object $x$ in $I_{1}$ and in $I_{2}$ with a scale factor $\theta > 1$ . Then, we paste it back aligned with the center of its own respective bounding box in the image. + +$$ +I _ {1} ^ {x} = I _ {1} \times S _ {1} \tag {5} +$$ + +The center point $(c_1^x, c_1^y)$ of the bounding box of $x$ in $I_1^x$ can be easily calculated. The new image $I_1^s$ can be presented as: + +$$ +I _ {1} ^ {x \uparrow} = U p s a m p l e \left(I _ {1} ^ {x}, \theta\right) \tag {6} +$$ + +$$ +I _ {1} ^ {s} = \operatorname {P a s t e} \left(I _ {1}, I _ {1} ^ {x \uparrow}, \left(c _ {1} ^ {x}, c _ {1} ^ {y}\right)\right) \tag {7} +$$ + +Where $\text{Paste}()$ is used to paste the scaled $x$ back to the position $(c_1^x, c_1^y)$ in $I_1$ . Then, by using the ground truth optical flow, we can trace the center position of the object in $I_2$ . Using the same process, we get $I_2^s$ . + +Given the center position $(c_1^x, c_1^y)$ for $x$ in $I_1$ and $(c_2^x, c_2^y)$ in $I_2$ , the optical flow needs to be recalculated in the object area. We use Eqs. (5) to (7) to get the upsampled $x$ optical flow $f_{gt}^{x\uparrow}$ . Then, for all pixels in $x\uparrow$ , the flow can be updated by: + +$$ +f _ {i j} ^ {s} = \theta f _ {i j} + (\theta - 1) \left(c _ {1} ^ {x} - c _ {2} ^ {x}, c _ {1} ^ {y} - c _ {1} ^ {y}\right), f _ {i j} \in x \uparrow \tag {8} +$$ + +Where $f_{ij}$ contains the optical flow for the 2D offset at index $i, j$ (The detailed inference process is described in the supplemental material). + +For $\theta < 1$ , we first remove the object using [44], then rescale the object $x \downarrow$ . However, the motion of the background covered by the original object $x - x \downarrow$ is unknown. Due to the unknown background motion, we mask pixels for $x - x \downarrow$ in the optical flow evaluation. + +Adding Objects (Flow-R-O). We collect real-world image $I_M$ of unseen objects. Then, we get new objects $y$ from $I_n$ using the SAM model with prompt $P_M$ + +$$ +S _ {M} = S A M \left(I _ {M}, P _ {M}\right) \tag {9} +$$ + +$$ +y = I _ {M} * S _ {M} \tag {10} +$$ + +We paste $y$ in $I_{2}$ and move it to $I_{1}$ using an affine transform, then recalculate the optical flow $f_{aff}$ . The new data is $(I_1^y, I_2^y, f_{aff})$ . + +![](images/1ad2b3794f78f4360f6e08a843e799772d1defce82e3f2ea6c5ace8abbfcf651.jpg) +Figure 2. Pipeline of our methods. Given the data $I_{1}, I_{2}, f_{gt}$ from KITTI, we manipulate the data by changing the texture or shape of an object or adding new objects. + +![](images/2b6210a2793380756bd41eb0842822cfc0b016f9a4798db1c737148834dfb6e2.jpg) + +Manual Quality Control. The above process will automatically generate new test data. However, SAM and generative models can produce poor quality masks and images that do not match the object $x$ . To ensure the masks and images are correct and of high quality, we resort to human checking. + +# 3.2. Robustness Evaluation + +We are inspired by sensitivity analysis [33] and evaluate the robustness for an image pair. We treat texture, shape and new objects as variables, assume a model could accurately predict flow by given $I_{1}, I_{2}$ , we have + +$$ +f _ {p} = \operatorname {M o d e l} \left(I _ {1}, I _ {2}\right) \approx f _ {g t} \tag {11} +$$ + +where $f_{gt}$ is the ground truth flow for $I_1, I_2$ . If we manipulate $I_1, I_2$ by replacing the texture of an object to obtain $I_1^t, I_2^t$ , and predict the flow by + +$$ +f _ {p} ^ {t} = \operatorname {M o d e l} \left(I _ {1} ^ {t}, I _ {2} ^ {t}\right) \tag {12} +$$ + +We can evaluate the sensitivity by + +$$ +\left| f _ {p} - f _ {p} ^ {t} \right| \approx \left| f _ {g t} - f _ {p} ^ {t} \right| \tag {13} +$$ + +This approach allows us to easily test the robustness and effects of different textures, shapes, and new objects. In practice, we primarily test methods with weights fine-tuned on the KITTI dataset that achieve high accuracy ( $< 0.7$ EPE). + +# 4. Experiments + +In this section, we describe the experiments that evaluate the behaviors of state-of-the-art (SOTA) optical flow approaches with our proposed method. We focus primarily on the common behaviors and failure cases of methods. These failure cases can potentially guide future studies of optical flow methods. + +We mainly use black-box evaluation to explore two important questions: Texture vs. Shape: What do optical flow methods rely on for matching? Memorization: Do optical flow methods tend to remember objects? + +In this experiment, we select the following methods: RAFT [41], GMA [17], GMFlow [42], FlowFormer [13], FlowFormer++ [38] and SAMFlow [47]. + +# 4.1. Texture vs. Shape + +Understanding how networks match objects is an important problem. It may provide fundamental insights into the functioning and potential failure points of optical flow methods. However, to the best of our knowledge, no studies have explored this aspect. In this section, we conduct three experiments to investigate whether the selected optical flow methods rely on texture or shape. + +To investigate the dependence of optical flow methods on texture versus shape, we evaluated their performance using KITTI [7], Flow-R-T, and Flow-R-S datasets as shown in Table 1. The findings reveal that most methods yield comparable results on KITTI [7] and Flow-R-T datasets. In + +![](images/14259908245e03fff2249807d1e224775940a4738d2e19c32e1c50db0cb54994.jpg) +Attacked Frame 1 +Figure 3. Texture Attack. The texture of the main object (car) in the first frame is replaced, while the second frame remains unchanged. The resulting optical flow is similar to that derived from the original images. + +![](images/665679021e575d3cfef8eccdd479314c0460f595d1c1f6b2534f6cb8fe969bd1.jpg) +Frame 2 + +![](images/4798633d5075677268a2c8f297d8cf4d91d74c27e8e82c4c905a14e26317b768.jpg) +Predict flow + +![](images/f17b0ff2abd8871a0fd04136e79d43c517afbb0769dd1823cd5d7ab4738e0bd3.jpg) +GT flow + +
Training DataMethodKITTI-15Flow-R-TFlow-R-SR-S (θ < 1)
AEPEFl-all (%)AEPEFl-all (%)AEPEFl-all (%)AEPE
C + TRAFT[41]5.0417.45.1818.35.9724.96.02*
GMFlow[42]7.7723.48.6726.08.9131.0-
GMA[17]4.6917.14.8217.95.7724.65.64*
Flowformer[13]4.0914.74.2815.55.1922.35.13*
Flowformer++[38]3.9314.14.3015.05.0721.84.97*
SAMFlow[47]3.4412.33.5813.05.0120.6-
C + T + S/K (+ H)RAFT[41]0.631.50.731.661.9610.21.83*
GMFlow[42]1.365.21.556.22.7813.9-
GMA[17]0.571.20.671.41.9010.01.72*
Flowformer[13]0.531.110.641.261.899.91.77*
Flowformer++[38]0.571.160.711.551.9510.11.74*
+ +Table 1. Quantitative results on KITTI 2015 and Flow-R datasets. We report the average end-point error (AEPE) and the Fl-all measure (Percentage of optical flow outliers with an error larger than 3 pixels). Methods are pre-trained on the Chairs and Things datasets "C + T" or fine-tuned on the Sintel and KITTI datasets, with some also fine-tuned on the HD1K dataset "S/K (+H)". In Flow-R-T the texture of the main object is replaced, and Flow-R-S replaces the shape of main object. Column R-S $(\theta < 1)$ replaces the shape with a smaller object. Values marked with * exclude the unknown area. All data is generated from KITTI 2015. + +
AttackTextureShape
AEPEF1-allAEPEF1-all
RAFT [41]0.821.992.2110.5
GMFlow[42]1.636.53.0314.2
GMA[17]0.721.632.0910.5
Flowformer[13]0.641.262.0410.2
Flowformer++[38]0.761.662.1310.1
+ +Table 2. Texture attack and Shape attack. Quantitative results of the attacked first frame and original second frame. + +contrast, all methods demonstrate a notable increase in errors on the Flow-R-S dataset, suggesting a higher sensitivity of optical flow methods to shape variations compared to texture variations. We separately report shape replacement + +with $\theta < 1$ in column R-S ( $\theta < 1$ ) because the reduction in object size may also introduce errors by incorrect filling of background flow. + +Texture Attack. Given that texture replacement does not significantly impact the performance of SOTA methods, we conduct an additional experiment by altering the texture of only one image. Given data from the KITTI dataset $(I_1, I_2, f_g t)$ , we replace the texture of object $x$ from $I_2$ by Equation (3). Despite the texture difference, the object is still matched, and the optical flow is "accurately" predicted (see Figure 3). Table 2 presents the results of the Texture Attack experiment. Surprisingly, we observe that the error rate for texture attack of attention-based methods is relatively small compared with errors in KITTI, suggesting that the matching of objects in optical flow methods is less dependent on texture than shape. + +![](images/5bb12d022bd0fe6943426729666d893e4a4724ad733482eaf0797b95171ade45.jpg) +Input Images +Figure 4. A failed example. We add seen and unseen objects to a similar area. The first row indicates that all methods successfully predict the motion of the added seen (car) objects. The second row indicates that all SOTA methods failed to predict the unseen (fox) objects. + +![](images/730d9ba7d8eb2153ec2385f5615d07b1d03b27610975f4b07f09bbf727bcddcc.jpg) +GMFlow + +![](images/c93c1903f821846ed67b3bdb5c8ccb736ed4465c10e99dbe96033f987561bcf4.jpg) +GMA + +![](images/00d056574bce8e250b48c222cc54e00e0b1b83b8799ce8bcdc515aef7ce0423e.jpg) +Flowformer++ + +![](images/777d8f69542936d4ce4a36e95aa268ac1bec3175324a0532ecae0407b41d382d.jpg) +SAMFlow + +
MethodsAdd SeenAdd Unseen
allforegroundbackgroundallforegroundbackground
AEPEF1-allAEPEAEPEAEPEF1-allAEPEAEPE
PWC-Net[40]4.1614.313.221.8312.728.434.3013.1
RAFT[41]2.548.211.260.777.4320.425.687.18
GMFlow[42]3.2412.111.321.649.4222.623.739.03
GMA[17]2.558.110.890.817.4420.222.627.19
FlowFormer[13]2.457.710.510.777.8121.023.707.52
FlowFormer++[38]2.558.110.710.837.5020.420.537.26
SAMFlow [47]2.437.910.760.857.7320.120.577.33
+ +Table 3. Adding new objects. Quantitative results of adding new objects to original images. "Add Seen" indicates the data where we add KITTI objects to the images. "Add Unseen" refers to the data where we add real-world objects not included in KITTI. Foreground indicates the EPE errors in the area of the new object. Background indicates the EPE error that do not include the new object. + +![](images/4869be5526e350f16b976984d1f72e57245efdacd9c66c38a9907eb1bbf8470e.jpg) + +![](images/aadac7319aadcc2290fe15b2a3a2e6fe1e633aab663627492a3870d17f79e5dd.jpg) + +![](images/c9e1a4fdb6981bd2c87fbf9461ccd8d5441f6af61f7ece91c7c87807983bdc4e.jpg) +Non-occlusion +Figure 5. Failed example with different positions. The unseen object (apple) is non-occluded, partially occluded, or fully occluded relative to the original object (car). The predicted flow from FlowFormer++ shows similar results; all flows ignore the unseen objects. The third column shows that the original motion of the car is still predicted even when the car is fully covered. + +![](images/7872e77214f49210d3db6c4557f6a7b9774b4761fb16044a41f7835c8fe591c4.jpg) + +![](images/a8e3831401c856829b353cca22025ec28140f7edf70f6c7d39b5849fed214314.jpg) + +![](images/f8fbb9db999be56e362eb0d21ee2e54796b4dd37ef2c313e5d82664666a2ac62.jpg) +Partial occlusion + +![](images/14265cfc72c507316eb5c22e160ec4c9a46a667d49c8c5c05233fd859641f36e.jpg) + +![](images/c0b214ddd7b2471749b6d2f81523251803e7218ba5357aabc15a0fa0c728efed.jpg) + +![](images/a8331ab150642da00d91771890825b2d4c78c450fec23a5a1ef27779f0899554.jpg) +Full occlusion + +Shape Attack. Shape changes are a common challenge in optical flow estimation, particularly when objects move closer to or further away from the camera. Hence, to assess the robustness of models to this kind of shape changes, we + +design a shape attack experiment, where we scale the object in only one image. + +Specifically, we scale the first frame $I_{1}$ by $\theta$ (as described by Equation (7)), and recalculate the optical flow (details provided in the supplementary material). Table 2 presents the quantitative evaluations of state-of-the-art methods. We observe a significant increase in errors between the shape attack results and the original results, indicating that the matching of objects in optical flow methods is sensitive to shape changes. More experiments with Sintel [3] and Shape Replacement can be found in the supplementary material. + +# 4.2. Does the Network Remember Objects? + +Optical flow methods still suffer from poor cross-dataset generalization. Most SOTA methods require fine-tuning and testing on the same dataset. One hypothesis is that optical flow methods tend to memorize objects during training and attempt to find similar objects during testing. In this section, we investigate this hypothesis experimentally. + +We define an unseen object as one that the methods do not see during training, while a seen object is one that is included in the KITTI dataset. Assuming opti + +cal flow methods successfully predict the correct optical flow given data $(I_1, I_2, f_{gt})$ , using Equations (9) and (10), we add unseen objects from real-world images to obtain $(I_1^{\text{unseen}}, I_2^{\text{unseen}}, f_{gt}^{\text{unseen}})$ and add seen objects from KITTI in another scene to get $(I_1^{\text{seen}}, I_2^{\text{seen}}, 2, f_{gt}^{\text{seen}})$ . By comparing the results in Table 3 for data with unseen and seen objects, we observe that flow prediction is significantly less accurate for unseen objects. This suggests that unseen objects negatively impact the methods. Furthermore, Table 3 reports the error in the foreground (EPE error only in the new object area) and the background error (areas not including the new object). We find that adding unseen objects also affects the background accuracy, even if the unseen object is not adjacent to the background area. + +![](images/db3ec91c9a1f10dc2cffbbd217203ed29681a1c872c30cf82edf52b85c2bf14c.jpg) +Figure 6. Feature Dissimilarity. All levels in the encoder have high dissimilarity compared with the original features for the added object region, indicating the encoder retains the new object. + +# 4.3. Failures for Unseen Objects + +To further study why unseen objects lead to poor results, we visualize the results in Section 4.2. We find that SOTA methods successfully predict the flow for seen objects, i.e., cars, but usually predict worse or even ignore unseen objects in the results. For example, in Figure 4, we add a car (first row) and a fox (second row) in similar positions. All methods successfully predict the motion of the car, but all methods fail to predict the motion of the fox. This indicates that SOTA methods tend to reduce or even ignore the flow of unseen objects. + +Unseen object attack. To study how an unseen object $x_{u}$ affects objects $x$ in original images, we place $x_{u}$ in different positions to observe its impact on the motion of $x$ in the images. Specifically, we categorize the scenarios into three cases: unoccluded, where $x_{u}$ does not cover $x$ ; partially occluded, where $x_{u}$ partially covers $x$ ; and fully occluded, where $x_{u}$ completely covers $x$ in the target image. + +Figure 5 presents an example using FlowFormer++[38]. In the first column, the apple does not interact with the car. In the second column, the apple partially covers the car. In the third column, the apple completely covers the car. However, the flow results for all three scenarios ignore the motion of the apple and successfully predict the car and its + +flow, even when it is covered by the apple. In the third example, where the car is fully covered by the apple, the cost volume should indicate that the car in the first frame cannot be matched to the second image due to occlusion. However, the method still predicts the car's motion, similar to the original results. This indicates that the methods tend to remember the motion of the scene or object and produce similar output. + +# 5. A Study of Unseen Objects + +To investigate why certain methods ignore unseen objects, we collect failure cases where the object is ignored and visualize the encoder and decoder components of GMA [17], FlowFormer [25], and FlowFormer++ [38]. As GMA is primarily based on a CNN architecture, FlowFormer on a transformer architecture and FlowFormer++ incorporates an Masked AutoEncoder [12] pretraining strategy, it allows us to investigate whether these designs contribute to robustness. + +![](images/e839393b1b160f5448e29e70c8677d7d3a83be3d0beaeb26301ea89f25340fb5.jpg) +Similarity in 1/2 +Similarity in 1/4 +Similarity in 1/8 +Figure 7. Example of Feature Dissimilarity. A clear difference of encoder features with or without new objects is shown at each level. The encoder retains the information related to new objects. + +# 5.1. Feature Visualization in the Encoder + +In the encoder, we extract the contextual features from the original image $I_{1}$ and the image containing the unseen object $I_{1}^{y}$ . For each feature at level (1/2, 1/4, and 1/8), we treat each pixel as a vector and perform cross-similarity calculations between the original features and unseen object features to generate a similarity map, as illustrated in Figure 7 where 1/2, 1/4 and 1/8, respectively, refer to the image resolution. The results reveal a clear difference in the areas corresponding to the new object, indicating that the encoder successfully extracts features related to the added object at all levels rather than ignoring it. + +In Figure 6, we statistically analyze the dissimilarity of features in the various methods. We measure the dissimilarity by calculating the similarity score, normalizing the score to a range of 0 to 1 and inversing it. Higher values indicate greater dissimilarity between two sets of features. We only evaluate the dissimilarity within the new object region. We found that at all three levels (1/2, 1/4, and 1/8), the features exhibited a relatively high dissimilarity, suggesting that the encoder effectively retains information about the unseen object. The feature difference between the original image and the image with the apple using cross-similarity in GMA is visualized in Figure 7. It clearly indicates the dissimilarity + +![](images/23685baf50efd7abd687ecaa189c764cb6f669834807d1505446c276825ffef8.jpg) +Figure 8. Visualization of Intermediate Flows in the Decoder. Decoder GRU iterations in original and failed samples. + +![](images/f34bbea9a98f5e12640abb3c5aab34913608618626eb68c8f2a871864810c37f.jpg) +Figure 9. SSIM of intermediate flow in each iteration between failed and original samples. + +![](images/797f36e7ed9ff94b9257559226a6b3626c339c985c5bbf3b9faf0649f1a1ba7e.jpg) +Figure 10. Feature similarity of GRU block. A higher value indicates that the features of the image with the added object are similar to those of the original image. + +in the apple region. + +# 5.2. Feature Visualization in the Decoder + +Recent methods employ an iterative refinement architecture. We chose GMA as a representative for such a framework and visualize each iterative step of the optical flow decoder. As illustrated in Figure 8, we initially observe the motion of an unseen object (e.g., an apple) during the coarse optical flow estimation. This object is subsequently treated as noise and removed in later iterations. + +We use SSIM to quantitatively assess the structural similarity between the optical flow of the original input images $I_1, I_2$ and the images that contain the unseen object $I_1^y, I_2^y$ . The evaluation is only for the region of the new object. As depicted in Figure 9, we observe significant differences in + +SSIM values during the first four iterations but much closer similarity after ten iterations. As illustrated in Figure 8, the first row displays the initial five iterations of optical flow applied to the original image input, while the second row presents the first five iterations for an image with an unseen object (an apple). The apple is distinctly visible in the first iteration, but visibility decreases with each subsequent iteration, and becomes nearly imperceptible by the fourth iteration. This observation suggests that the decoder component may accidentally interpret the unseen object as noise or error, leading to its gradual removal throughout the iterations. + +Next, we conduct experiments to analyze each main component of the decoder, including the cost memory, attention block, GRU block, and upsampler. By comparing the features extracted from $I_1, I_2$ and $I_1^y, I_2^y$ , we observe that the GRU block exhibits behavior similar to that observed in the flow iterations. Figure 10 reports the similarity between the original features and the features with new objects within the GRU block. Circles indicate images with seen objects, while triangles denote images with unseen objects. In both GMA and Flowformer++, the GRU block effectively retains the features of seen objects across iterations, while it weakens the features of unseen objects. This suggests that the GRU block is designed to remember and preserve the characteristics of seen objects, while attempting to eliminate the influence of unseen objects during the iterations. + +# 6. Conclusion + +In this paper, we propose a robustness evaluation method by manipulating the shape and texture of original objects or by adding new objects. We generate a dataset and conduct a comparative evaluation of the robustness of recent SOTA methods. Our findings reveal two notable common behaviors among these SOTA methods: 1. The matching performance of optical flow methods depends more on shape than on texture. 2. The GRU block tends to remember seen objects while attempting to diminish the influence of unseen objects. Given these observations, we express concerns regarding the safe usage of current optical flow methods. We hope that our methodology and findings will guide further studies aimed at enhancing the robustness of optical flow techniques in safety-critical scenarios. + +# References + +[1] Simon Baker, Daniel Scharstein, J. P. Lewis, Stefan Roth, Michael J. Black, and Richard Szeliski. A database and evaluation methodology for optical flow. In International Journal of Computer Vision, pages 1-8, 2007. +[2] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3698-3707, 2019. +[3] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In European Conference on Computer Vision, pages 611-625, 2012. +[4] A. Dosovitskiy, P. Fischer, E. Ilg, P. Häusser, C. Hazirbaş, V. Golkov, P. v.d. Smagt, D. Cremers, and T. Brox. FlowNet: Learning optical flow with convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 2758-2766, 2015. +[5] Yi Feng, Ruge Zhang, Jiayuan Du, Qijun Chen, and Rui Fan. Freespace optical flow modeling for automated driving. IEEE/ASME Transactions on Mechatronics, pages 1511-1520, 2024. +[6] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtualworlds as proxy for multi-object tracking analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4340-4349. IEEE Computer Society, 2016. +[7] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. +[8] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. +[9] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam H. Laradji, Hsueh-Ti Derek Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, A. Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3739-3751, 2022. +[10] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024. +[11] Yunhui Han, Kunming Luo, Ao Luo, Jiangyu Liu, Haoqiang Fan, Guiming Luo, and Shuaicheng Liu. Realflow: Em + +based realistic optical flow dataset generation from videos. In Proceedings of the European Conference on Computer Vision, pages 288-305, 2022. +[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009. IEEE, 2022. +[13] Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. FlowFormer: A transformer architecture for optical flow. European Conference on Computer Vision, 2022. +[14] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page 4761-4770, 2019. +[15] Md Amirul Islam, Matthew Kowal, Patrick Esser, Sen Jia, Bjorn Ommer, Konstantinos G Derpanis, and Neil Bruce. Shape or texture: Understanding discriminative features in cnns. In International Conference on Learning Representations, 2021. +[16] Huaizu Jiang, Deqing Sun, Varan Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super SloMo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9000-9008, 2018. +[17] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9772-9781, 2021. +[18] Soo Ye Kim, Jihyong Oh, and Munchurl Kim. FISR: Deep joint frame interpolation and super-resolution with a multiscale temporal loss. In AAAI, pages 11278-11286, 2020. +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. +[20] Daniel Kondermann, Rahul Nair, Katrin Honauer, Karsten Krispin, Jonas Andrus, Alexander Brock, Burkhard Gussefeld, Mohsen Rahimimoghaddam, Sabine Hofmann, Claus Brenner, and Bernd Jahne. The HCI benchmark suite: Stereo and flow ground truth with uncertainties for urban autonomous driving. In CVPRW, pages 19-28, 2016. +[21] Tom Koren, Lior Talker, Michael Dinerstein, and Roy J. Jevnisek. Consistent semantic attacks on optical flow. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1658-1674, 2022. +[22] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Proceedings of the International Conference on Learning Representations (ICLR), 2017. +[23] Jianfeng Li, Junqiao Zhao, Shuangfu Song, and Tiantian Feng. Occlusion aware unsupervised learning of optical flow from video. In Thirteenth International Conference on Machine Vision, page 224 - 231, Rome, Italy, 2021. International Society for Optics and Photonics, SPIE. + +[24] Tianran Liu, Morteza Mousa Pasandi, and Robert Laganiere. Fsmdet: Vision-guided feature diffusion for fully sparse 3d detector. arXiv preprint arXiv:2409.06945, 2024. +[25] Yawen Lu, Qifan Wang, Siqi Ma, Tong Geng, Yingjie Victor Chen, Huaijin Chen, and Dongfang Liu. Transflow: Transformer as flow learner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18063-18072, 2023. +[26] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, pages 4040-4048, 2016. +[27] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In CVPR, 2023. +[28] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3061-3070, 2015. +[29] Simon Niklaus, Long Mai, and Feng Liu. Video frame interpolation via adaptive convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2270-2279, 2017. +[30] Bojun Ouyang and Dan Raviv. Occlusion guided scene flow estimation on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1049-1058, 2021. +[31] Anurag Ranjan, Joel Janai, Andreas Geiger, and Michael J. Black. Attacking optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1-10, 2019. +[32] 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 10807-10817. IEEE, 2022. +[33] Andrea Saltelli, Marco Ratto, Terry Andres, Francesca Campolongo, Jessica Cariboni, Debora Gatelli, Michaela Saisana, and Stefano Tarantola. Global Sensitivity Analysis: The Primer. John Wiley & Sons, 2008. +[34] Jenny Schmalfuss, Philipp Scholze, and Andres Bruhn. A perturbation-constrained adversarial attack for evaluating the robustness of optical flow. Proceedings of the European Conference on Computer Vision, 2022. +[35] Jenny Schmalfuss, Lukas Mehl, and Andres Bruhn. Adversarial weather attacks for motion estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10106-10116, 2023. +[36] Philipp Schrodi, Jenny Schmalfuss, and Andres Bruhn. A less constrained global attack on optical flow. In Proceedings of the European Conference on Computer Vision, pages 183-200. Springer, 2022. +[37] Hao Shi, Yifan Zhou, Kailun Yang, Xiaoting Yin, and Kaiwei Wang. Csflow: Learning optical flow via cross strip correlation for autonomous driving. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 1851-1858, 2022. + +[38] Xiaoyu Shi, Zhaoyang Huang, Dasong Li, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Flowformer++: Masked cost volume autoencoding for pretraining optical flow estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1599-1610, 2023. +[39] Xiuchao Sui, Shaohua Li, Xue Geng, Yan Wu, Xinxing Xu, Yong Liu, Rick Goh, and Hongyuan Zhu. Craft: Cross-attentional flow transformer for robust optical flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 123–132, 2022. +[40] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In CVPR, page 8934–8943, Salt Lake City, UT, USA, 2018. +[41] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In European Conference on Computer Vision, page 402-419, Glasgow, UK, 2020. +[42] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 795-804, 2022. +[43] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. In Advances in Neural Information Processing Systems, pages 794-805, 2019. +[44] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. +[45] Feihu Zhang, Oliver J. Woodford, Victor Adrian Prisacariu, and Philip H. S. Torr. Separable flow: Learning motion cost volumes for optical flow estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10807-10817, 2021. +[46] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. +[47] Shili Zhou, Ruian He, Weimin Tan, and Bo Yan. Samflow: Eliminating any fragmentation in optical flow with segment anything model. Proceedings of the AAAI Conference on Artificial Intelligence, 2024. +[48] Ruijie Zhu, Yanzhe Liang, Hanzhi Chang, Jiacheng Deng, Jiahao Lu, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang. Motions: Exploring explicit motion guidance for deformable 3d gaussian splatting. In Conference on Neural Information Processing Systems, 2024. \ No newline at end of file diff --git a/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/images.zip b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..22448718d3372ef4e4c6ee6b3b5de3bb41983e58 --- /dev/null +++ b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50ccadcf0681b65af4e2e65a91f19d1684217d5bb3742ae67aa3ffafd67dce05 +size 670293 diff --git a/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/layout.json b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6a7832c1d9e687cc5c884ef5aa1bf83158fb1d50 --- /dev/null +++ b/CVPR/2025/Shape and Texture_ What Influences Reliable Optical Flow Estimation_/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aca8c7fb29ee6c28c2843a075dbd9eb5fe6cdf6865c731ead15557c154b4f879 +size 422484 diff --git a/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/9bca979d-6491-4ddd-a007-065d6844c69e_content_list.json b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/9bca979d-6491-4ddd-a007-065d6844c69e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..37d1b24171435b29da7f6bdb40180c097a5bc74f --- /dev/null +++ b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/9bca979d-6491-4ddd-a007-065d6844c69e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a148954ffe0802657074698d8401dafb678bcfc25cb815a83f706d75fbd041c1 +size 100434 diff --git a/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/9bca979d-6491-4ddd-a007-065d6844c69e_model.json b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/9bca979d-6491-4ddd-a007-065d6844c69e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2b782f077c88a169212ee13e35725c0d5cb1d723 --- /dev/null +++ b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/9bca979d-6491-4ddd-a007-065d6844c69e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:590e60419e7d62508cac14cbe1c6b80c65b217715529560849b50c4adbdf31b9 +size 135545 diff --git a/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/9bca979d-6491-4ddd-a007-065d6844c69e_origin.pdf b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/9bca979d-6491-4ddd-a007-065d6844c69e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a7d6f4c8a57e841efafeee75b9f38e55e1ba0be --- /dev/null +++ b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/9bca979d-6491-4ddd-a007-065d6844c69e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d54f17258c378b5d5e9bbf273394cd9707a107cfd665190e3167befa487ca7b +size 6381874 diff --git a/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/full.md b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..df9bb69737d7f1674b263847a39d55c35a7e0ecf --- /dev/null +++ b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/full.md @@ -0,0 +1,443 @@ +# ShapeShifter: 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion + +# Nissim Maruani + +Inria, Université Côte d'Azur + +nissim.maruani@inria.fr + +# Wang Yifan + +Adobe Research + +yifwang@adobe.com + +# Matthew Fisher + +Adobe Research + +matfishe@adobe.com + +# Pierre Alliez + +Inria, Université Côte d'Azur + +pierre.alliez@inria.fr + +# Mathieu Desbrun + +Inria/X, IP Paris + +mathieu.desbrun@inria.fr + +# Abstract + +This paper proposes ShapeShifter, a new 3D generative model that learns to synthesize shape variations based on a single reference model. While generative methods for 3D objects have recently attracted much attention, current techniques often lack geometric details and/or require long training times and large resources. Our approach remedies these issues by combining sparse voxel grids and point, normal, and color sampling within a multiscale neural architecture that can be trained efficiently and in parallel. We show that our resulting variations better capture the fine details of their original input and can handle more general types of surfaces than previous SDF-based methods. Moreover, we offer interactive generation of 3D shape variants, allowing more human control in the design loop if needed. + +# 1. Introduction + +Creating 3D content through generative models is currently attracting significant attention. Traditional 3D modeling demands both time and specialized skills to create complex shapes, whereas advancements in generative AI promise a broader exploration of design possibilities, free from the usual constraints of time or technical expertise. However, current 3D generative models have numerous shortcomings that limit their usefulness in applications such as movies, gaming, and product design. First, state-of-the-art methods often struggle to produce the fine geometric details and sharp features necessary for digital shapes in geometric modeling. Additionally, these models require large, high-quality 3D datasets, which are significantly more challenging to curate compared to image datasets, and involve long training times and substantial computational resources. + +In this paper, we tackle a task-specific, yet effective approach to synthesizing geometry: we propose generating shape variations from a single high-quality example. This lesser-explored generative method offers several benefits + +![](images/1f1fd94ff66652d22bb9021e7f0f92b6bd9a0d5a99f9d6037b35c0384dd77c10.jpg) +Figure 1. ShapeShifter. Given a 3D exemplar, we propose to train a hierarchical diffusion model to create variations preserving the geometric details and styles of the exemplar. By combining compact yet explicit 3D features (colored, oriented points) with a sparse voxel grid, we shorten training times from hours to minutes, while yielding significantly better geometric quality than prior work. The hierarchical point representation and fast inference times further enable intuitive interactive editing. + +beyond avoiding the curation of large training datasets: it has the potential to provide a resource-efficient way to generate shape variants for retargeting or editing, automatically inheriting the style, symmetries, semantics, and geometric details from the exemplar. Although existing generative methods from exemplars are able to create varied 3D assets, they struggle to produce clean and detailed geometry due to their reliance on occupancy fields [50, 91, 105] or signed distance functions [106] (which smooth out geometric features), or because they are supervised through volumetric + +![](images/f2b31101aa693a428d6f26ca41fc8348364127f7cbcc9abc083e6e024fbb862e.jpg) + +![](images/340aab0225c9ff84423f23e993404cb278b7380f6b320ace2c9e39445b6c8521.jpg) + +![](images/fe841403e1147559c44a761a5a8bb2198e12b45d7bcc3e319be23150c4fa6dd6.jpg) + +![](images/3363dcc8118bafe76440cfb3eebc591e0019ba8ded6781f8b72548be6a76f84a.jpg) + +![](images/3f53daaf99b11e54a819fe2ccb47dd153a2cbd42a871ec4793c3a70e0792d37a.jpg) + +![](images/7bab499a5ba473df720e26f5e7f88d72887fb42df07210b815a0b4ef52b04279.jpg) + +![](images/731b3d0272a29e4d1d6402a4cafc8b1111f9986839e0af66ab848a8c49125afc.jpg) + +![](images/bc4f5821498c70032ea4da9696c4722864b09677b488f363bb3567f01523a72c.jpg) + +![](images/60f4811a1fed42f4e202d2712340fbed3d75920487ea8a46420a454f9c5a6e9a.jpg) + +![](images/ac55b3b39396da7f76e90871ad3412e2b685f51360e8353aad64424949c75869.jpg) + +![](images/4aa760a6aade545dd58b1144ad0f47351ea47d3b49f592bcf1392ba64f9998ba.jpg) + +![](images/573e3321f353eec2e3a5f7a6916c788cd9d80bf1aa97a24ebde5e094a34d6fde.jpg) + +![](images/c62e6f97b066573a8aef70a8b08022100ac55f85d4d69f4ca043f174f343b343.jpg) +(a) Input Geometry + +![](images/869b1c2ea9371b0209b45e96e48c0e38b4c82f465df6905b19fc111af716591e.jpg) +(b) Ours + +![](images/2c1ea0d8e9c0e57b3c8db111b1208b841428c3c824e27699a22c5aa894258e5c.jpg) +(c) Sin3DGen +Figure 2. Geometric details. Our generation captures significantly more geometric details present in the exemplar mesh (leftmost). Prior work, Sin3DGen [50] and Sin3DM [106], operates with plenoxels and neural radiance fields encoded in single-resolution triplane features, respectively, which lack the capability to sufficiently represent and supervise high-resolution geometric details. In contrast, our method employs a colored and oriented point set, providing precise geometric information. + +![](images/3c37a70c5a41c4dbbb8065118d1cfde0a5bb76ce9d58f68c3a58130b3a11db21.jpg) +(d) Sin3DM + +rendering [50, 91] (which often leads to large geometric artifacts) — and without a clean geometric output model, the use of 2D textures to further enhance visual complexity is severely hindered. Consequently, existing exemplar-based methods are relative slow in generating variations of the input as they rely on volumetric sampling within the surface's neighborhood [91, 105, 106]. + +We propose a novel technique, that we call ShapeShifter, to synthesize high-quality shape variations of an input 3D model, with training and inference times well suited for practical real-world applications. We use points (with their normals and optionally colors for additional semantic information) as our lightweight and efficient base geometric representation [77], which we pair with a multiscale 3D diffusion network. While these explicit surface features already streamline the generative process and help preserve geometric details, we propose to significantly reduce training times and achieve interactive inference rates by adopting sparse convolutions based on fVDB [104], a recent spatial learning framework based on sparse voxel grids. Mixing point sampling and sparse convolutions, a novel combination in generative modeling, results in a multiscale generative approach capable of producing 3D variants of shapes of different styles and topologies. Furthermore, its fast inference allows for interactive editing control. + +Contributions. This paper proposes a neural network approach to generating high-quality shapes from a single 3D reference example. Compared to previous exemplar-based generative methods, we demonstrate significantly improved + +geometric quality of our outputs, as shown in Fig. 2. Moreover, the simplicity of our geometric representation (using point sampling in a sparse voxel grid) and its hierarchical refinement (learned per level in parallel) to control and generate variations of an arbitrary closed or open input shape results in significantly reduced training times (minutes instead of hours) and interactive inference. While our results can be easily converted into textured meshes, direct visualization of our point-based representation in realtime enables iterative co-creation guided by an artist. Finally, our high-quality output geometric models can be assigned a fine texture if needed, using off-the-shelf image super-resolution or more advanced texturing synthesis methods such as [83]. + +# 2. Related Work + +3D Generation. The field of 3D generation has seen rapid development in recent years. Advances in generative models and large-scale 3D datasets have underpinned this progress. Generative adversarial networks (GANs) [23] have been widely used in works like [2, 10, 22], while normalizing flows [82] were utilized in [110]. Other approaches include variational autoencoders (VAEs) [45] and autoregressive (AR) models [6, 25, 96], as shown in [14, 70, 74, 89, 113, 115]. The recent introduction of diffusion models [33, 90] has enabled training on larger datasets such as Objaverse [18]. A survey by Po et al. [3] provides a comprehensive taxonomy of 3D diffusion approaches. A primary line of work builds on 2D diffusion models, generating multiview-consistent images through Score Distillation + +![](images/d57b6689c9ec76d140bc425c4a57170f07de9efdf8854c0aa1b7d831507d373a.jpg) +Figure 3. Multiscale diffusion on sparse voxel grid. We start from noise $\epsilon \sim \mathcal{C}(\mathbf{0},\mathbf{I})$ at the coarsest level $l = 1$ , and obtain the 3D feature grid $\mathbf{G}^l$ through reverse diffusion. Each subsequent level uses the output of the previous level. Inactive voxels are first pruned, then upsampled with a level-specific upsampler $\mathcal{U}^l$ . The upsampled grid $\tilde{\mathbf{G}}^l$ is subsequently noised and passed through the diffusion model to obtain a clean version of the sparse feature grid $\mathbf{G}^l$ . All levels are independent and can thus be trained in parallel. + +Sampling (SDS) [76, 97]. However, SDS faces practical challenges such as high optimization times [52, 67], color artifacts [12, 51, 61, 101], and 3D inconsistencies [57, 100]. Fine-tuning diffusion models on 3D assets for direct multiview output [53, 55, 59, 79, 87] can address these issues, with further speedups through reconstructor networks for radiance fields [11, 49, 54, 99, 102] or Gaussian splats [108, 117, 122]. However, photometric losses often lead to geometric artifacts. A separate direction directly trains 3D diffusion models on 3D data [62, 71] or encodes 3D data through autoencoders [17, 27, 39, 80, 81, 95, 116, 118, 119]. These methods demand extensive, high-quality data and substantial computational resources, and the generated geometry, while improved, still lacks the geometric details required in real-world 3D applications. Although Xcube [80] initially proposed using an efficient multiscale diffusion pipeline, it relies on pre-trained VAEs and verbose SDF representation which are not adapted to the problem of single-shape variations. We adopt an alternative approach which generates 3D assets with high-quality geometry from a single exemplar, trainable on a single GPU in minutes, while enabling user control over output shapes. + +Generation from a single instance. Despite advancements in large-scale 3D generation, high-quality 3D data remains scarce, and computational costs for training and inference are significant. Generating 3D content from a single high-quality example offers an appealing alternative, giving users control through concrete exemplar inputs. Non- + +learning, patch-based methods such as PatchMatch [5, 29] and more recent works [20, 24] produce variations by finding and blending similar patches within the exemplar. In contrast, SinGAN [86] and its successors [32, 88] have shown that generative models can be trained on a single example. More recently, SinFusion [46, 72] demonstrated that diffusion models, known for their stability, can also be adapted to this approach. Applications of these single-instance models include texture synthesis [69, 73, 84, 121], video synthesis [28], and more. + +In 3D generation, similar approaches have emerged. Herz et al. [31] applied SinGAN with mesh convolutions [30] to enhance surface details without altering structure, while works like Son et al. [91] and Karnewar et al. [41] generated radiance fields from a single instance. Wu et al. [105] used 3D convolutions to generate significant variations, later extending their work to include texture synthesis [106], though details are often degraded due to 3D convolution bottlenecks. These methods, however, are typically slow, with training times ranging from 2 to 4 hours. To bypass training, Li et al. [50] adopted a PatchMatch-based method for 3D scenes represented by plenoxels [21]. Although this removes the training step, it still requires approximately 10 minutes per variation and struggles to maintain geometric quality due to challenges in converting the plenoxels' occupancy fields to signed distance functions. By combining the strengths of a recent framework for learning sparse spatial information [104] with compact, geometry-centric features, our method achieves higher geometric fidelity while requiring less than 12 minutes for the entire generation process (including training and inference). + +3D Representation. Earlier 3D representations focused solely on geometry, using formats like voxels [107], 3D points [78], meshes [30, 64, 65], SDFs [74], and occupancy fields [15, 66]. Differentiable rendering later enabled joint modeling of geometry and appearance, spanning from neural mesh rasterization [42, 56] and splatting methods [44, 103, 112] to neural radiance fields [58, 68]. This integration allowed the use of 2D assets [10, 85] and facilitated large-scale generation and reconstruction [34, 93, 102, 108, 109], enhancing semantic context for geometry generation. However, appearance signals often exhibit higher frequencies than geometry, leading existing methods to separate texture and geometry branches [68]. Tying texture generation to geometry imposes unnecessarily high resolution requirement on geometry, slowing single-exemplar generation [106]. Other works generate geometry first, before refining the appearance in 2D texture space at higher resolutions [36, 111, 114]. Recently, Clay [118] fully separates geometry and appearance aspects, achieving state-of-the-art quality in both. Following this idea, we focus on high-fidelity geometry generation, which is the weakest aspect of current 3D generation approaches. Unlike Clay's strict + +separation, however, we use RGB features to add contextual information—crucial for single-exemplar generation where data priors are weaker. + +Our results show that this approach produces superior geometry which, when combined with state-of-the-art image super-resolution [40, 48] and texture synthesis [8, 83], allows for highly detailed shapes with sharp geometry and HD textures. We argue that this design, for a fixed computational budget, optimally balances quality and efficiency. + +# 3. Method + +Our goal is to train a generative model from a single 3D input mesh to generate new variations efficiently. We use a multiscale diffusion model with limited receptive fields to learn the internal structures of the given shape, adapting an approach that has been used for training a generative model on a single image [46]. We use compact, explicit 3D features directly extracted from the exemplar shape for diffusion. These features are encoded in a sparse voxel grid, and processed efficiently using a specialized 3D convolution framework (fVDB [104]) to capture fine-scale geometric details without incurring high memory cost. We introduce our 3D features in Sec. 3.1, the hierarchical diffusion model in Sec. 3.2, and the final meshing process of a generated output in Sec. 3.3. + +# 3.1. 3D Representation + +Explicit Multiscale 3D Features. Our method employs explicit 3D information to encode the geometry of the input exemplar mesh at multiple scales. They are composed of merely 10 values per voxel of a sparse voxel grid, + +$$ +\boldsymbol {f} = \left(\boldsymbol {p} _ {x}, \boldsymbol {p} _ {y}, \boldsymbol {p} _ {z}, \boldsymbol {n} _ {x}, \boldsymbol {n} _ {y}, \boldsymbol {n} _ {z}, \boldsymbol {c} _ {r}, \boldsymbol {c} _ {g}, \boldsymbol {c} _ {b}, m\right), \tag {1} +$$ + +where $(\pmb{p}_x, \pmb{p}_y, \pmb{p}_z) \in [-0.5, 0.5]^3$ represents a point sample encoded as an offset from the voxel center, $(\pmb{n}_x, \pmb{n}_y, \pmb{n}_z) \in [-1, 1]^3$ represents the local normal of the underlying geometry associated to the point sample, $(\pmb{c}_r, \pmb{c}_g, \pmb{c}_b) \in [-2, 2]^3$ represents the color scaled up to the $[-2, 2]$ range, and $m \in [-1, 1]$ is a scalar indicating if the voxel contains the mesh surface, which we will use as a mask to prune voxels after refinement (see Sec. 3.2). The value ranges of the feature components were empirically chosen since feature scale can be important in diffusion models [13]. + +These features are extracted from the different scales of the input mesh. Specifically, starting from the finest scale $L$ , for each voxel that intersects the surface, we find the nearest surface point to the voxel center, whose position, normal, and color are used to form the feature. The mask is set to 1 for all selected voxels as they contain the surface. At each subsequent scale $l < L$ , we obtain coarser points, normals, and colors features through a $2^{3}$ average pooling, and the mask value through max pooling. To better preserve sharp + +Algorithm 1: Training at level $l$ +Input: Extracted sparse features $\{\mathbf{G}^1,\dots ,\mathbf{G}^L\}$ // Upsampler Training for Levels 2 to $L$ +repeat Update model $\mathcal{U}^l$ with the loss Eq. (4) +until convergence; +// Diffusion Model Training +repeat + $t\sim \mathrm{Uniform}(0,T)$ $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ +if $l = 1$ then $\begin{array}{rl}{\mathbf{G}^{l,\mathrm{mix}} = \mathbf{G}^{1}} & {} \end{array}$ +else + $\begin{array}{r}\tilde{\mathbf{G}}^l = \mathcal{U}^l (\mathbf{G}^{l - 1})\\ \mathbf{G}^{l,\mathrm{mix}} = \gamma (t)\mathbf{G}^l +(1 - \gamma (t))\tilde{\mathbf{G}}^l \end{array}$ $\mathbf{G}_t^l = \sqrt{\bar{\alpha} (t)}\mathbf{G}^{l,\mathrm{mix}} + \sqrt{1 - \bar{\alpha} (t)}\epsilon$ +Update model with $\| \mathcal{M}^l (\mathbf{G}_t^l |t) - \mathbf{G}^l\| ^2$ +until convergence; + +features, we average the point positions in the Quadric Error Metric sense [65] (see supplemental material for details). + +This 3D representation yields three advantages: + +- it captures surface details in a compact form and carries contextual information from the texture; +- it encodes the 3D shape explicitly at each level, which enables a generated shape to be easily visualized or even edited at every level in a coarse-to-fine fashion; +- it is extracted from an input exemplar efficiently and deterministically, and will allow us to train each level of our model in parallel. + +Sparse Voxel Grid. The inductive biases of convolutional neural networks exploit the shared information across internal crops within the input data, which is essential to learning from a single example and prevents overfitting [46, 72, 86]. However, 3D convolutional operations are notoriously expensive computationally and memory intensive. To address this issue, we leverage fVDB [104], a recently proposed framework that supports efficient operations on sparse voxels. As a result, only active features are stored and processed, which significantly reduces the memory footprint and computational complexity. We denote the sparse feature grid storing active 3D features at level $l$ as $\mathbf{G}^l = \{\pmb{f}^l\}$ . + +# 3.2. Multiscale Diffusion + +Our multiscale diffusion pipeline generalizes SinDDM [46] to 3D and adapts it to properly work with sparse voxel grid. As shown in Fig. 3, the pipeline consists of multiple diffusion models $\{\mathcal{M}^l\}_{1\leq l\leq L}$ . During training, these diffusion models can be trained in parallel; at inference time, new + +variations are generated by sequentially running the reverse diffusion in a coarse-to-fine manner. Below, we explain the hierarchical multi-scale diffusion and highlight our design differences compared to SinDDM. + +# Algorithm 2: Sampling + +Input: Choice of sampler $S \in \{\mathrm{DDPM},\mathrm{DDIM}\}$ + +Output: Generated sparse grid $\mathbf{G}_L$ + +// Upsampler training for levels 2 to $L$ + +$\mathbf{G}_T^1\sim \mathcal{N}(0,\mathbf{I})$ + +for $l\gets 1$ to $L$ do + +$$ +\begin{array}{l} \text {i f l > 1 t h e n} \\ \left\lfloor \begin{array}{l} \epsilon \sim \mathcal {N} (0, \mathbf {I}) \\ \tilde {\mathbf {G}} ^ {l} = \mathcal {U} ^ {l} (\mathbf {G} ^ {l - 1}) \\ \mathbf {G} _ {T [ l ]} ^ {l} = \sqrt {\bar {\alpha} (T [ l ])} \tilde {\mathbf {G}} ^ {l} + \sqrt {1 - \bar {\alpha} (T [ l ])} \epsilon \\ \end{array} \right. \\ \text {f o r t \leftarrow T [ l ] t o 1 d o} \\ \left\lfloor \begin{array}{l} \epsilon \sim \mathcal {N} (0, \mathbf {I}) \\ \mathbf {G} _ {t - 1} ^ {l} = \mathcal {S} (\mathcal {M} ^ {l}, \mathbf {G} _ {t} ^ {l}, \bar {\alpha}, \epsilon , t) \\ \end{array} \right. \\ \mathbf {G} ^ {l} = \operatorname {P r u n e} (\mathbf {G} _ {0} ^ {l}) \end{array} +$$ + +Forward Multiscale Diffusion. Except at its coarsest level $\mathcal{M}^1$ , our diffusion model $\mathcal{M}^{l > 1}$ generates the signal of the current level based on the output of the previous $(l - 1)$ level. This initial guess is obtained by upsampling the output from the previous level $\tilde{\mathbf{G}}^l = \mathcal{U}\big(\mathbf{G}^{l - 1}\big)$ , which can be seen as a "blurred" version of $\mathbf{G}^l$ . This means that, for $l > 1$ , the diffusion model not only needs to denoise but also "deblur" during sampling. As a result, SinDDM modifies the forward diffusion process to be + +$$ +\mathbf {G} _ {t} ^ {l} = \sqrt {\bar {\alpha} (t)} \left(\gamma (t) \mathbf {G} ^ {l} + (1 - \gamma (t)) \tilde {\mathbf {G}} ^ {l}\right) + \sqrt {1 - \bar {\alpha} (t)} \epsilon , \tag {2} +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , while $\bar{\alpha}(t)$ and $\gamma(t)$ are monotonically decreasing functions from 1 to 0 as $t$ grows from 0 to $T$ . The model learns to denoise the corrupted feature $\pmb{f}_t^l$ at time step $t$ by minimizing the reconstruction loss + +$$ +\left\| \mathcal {M} ^ {l} \left(\mathbf {G} _ {t} ^ {l} \mid t\right) - \mathbf {G} ^ {l} \right\| ^ {2}. \tag {3} +$$ + +Contrasting with SinDDM which employs a bilinear upsampler as $\mathcal{U}$ , we use a level-specific upsampler $\mathcal{U}^l$ motivated by the fact our spatial features (points and normals) are extracted by projecting the voxel centers onto the mesh surface — thus potentially exhibiting abrupt local changes. This results in improved preservation of sharp geometric features as we show in Sec. 5. The upsampler $\mathcal{U}^l$ is trained to minimize the $L^2$ -loss between upsampled and ground-truth features, i.e., + +$$ +\left\| \mathcal {U} ^ {l} \left(\mathbf {G} ^ {l - 1}\right) - \mathbf {G} ^ {l} \right\| ^ {2}. \tag {4} +$$ + +Crucially, the training of different levels can be parallelized. For each level $l > 1$ , we first train the upsampler and the diffusion model as summarized in Algorithm 1. + +Unlike SinDDM, training needs to accommodate our use of sparse grids. When comparing the denoised sparse feature grid and the ground-truth sparse feature grid (Eq. (3)), the denoised grid can contain more active voxels (see dark voxels in Fig. 3, even though their mask could be -1 — yet fVDB operations on two sparse feature grids assume that they have the same active voxels. To solve this problem, we flood those inactive voxels in the ground-truth $\mathbf{G}^l$ with feature values of the nearby active cells using a blurring kernel. All features except the mask value are flooded in this way, whereas the mask value is set to -1. Empirically, we observe that soft blending the feature values this way (instead of hard setting the values to an arbitrary number or applying an additional mask for loss) achieves the best result. + +Reverse Multiscale Diffusion. Once trained, we can apply standard DDPM [33] or DDIM [92] sampling sequentially from levels 1 to $L$ . As outlined in Algorithm 2, we start from a noise $\epsilon \sim \mathcal{N}(0, I)$ and run the reverse sampling to obtain an initial prediction at the coarsest level. Then, for each level, we first prune the predicted inactive voxels from the previous level by removing any feature entries with mask value $m < 0$ . The resulting feature grid is then upsampled with $\mathcal{U}^l$ , and subsequently corrupted with noise, before being given to the diffusion model for reverse sampling. Similar to SinDDM, we only add noise up to timestep $T[l] < T$ to prevent destroying the prediction from the previous level. A schematic overview of the sampling process is illustrated in Fig. 3. + +# 3.3. Meshing + +Once a new geometric variant has been created, we can directly visualize the generated shape using the points (one per finest voxel in the fVDB data structure) along with their associated normal and color. We can also trivially generate a mesh of the geometry through Poisson reconstruction [43] (or APSS [26] if we are dealing with open surfaces). One can assign colors to the mesh nodes based on the output colors, bake texture maps (as used sporadically in figures), or further refine and stylize the texture with off-the-shelf image enhancement models (see Sec. 5.5). + +# 4. Implementation Details + +Implementation Details. We implemented our method in Python with PyTorch [75], libigl [38], and Open3D [120], and our code is available on our project page. All reported timings were obtained on a desktop with an NVIDIA GeForce RTX 3080 GPU (10 GB) to underscore the efficiency of our approach even on consumer-grade hardware. + +Model Parameters. By default we use 5 levels, the lowest and highest grid resolutions being 16 and 256 respec + +
MetricMethodacropoliscanyonfighting-pillarhouseruined-towersmall-townstone-cliffwoodaverage
G-Qual. ↓Sin3DGen4.836.168.4517.956.984.0213.0210.328.97
Sin3DM0.922.230.262.010.490.840.020.090.86
ShapeShifter0.010.210.260.010.111.000.100.020.21
G-Div. ↑Sin3DGen0.250.500.590.410.860.700.650.440.55
Sin3DM0.120.170.150.010.210.600.320.100.21
ShapeShifter0.040.190.240.010.320.600.230.080.21
+ +Table 1. Evaluating geometric quality and diversity using SSFID and pairwise IoU scores. Our model shows clear advantage in quality, and performs similar to Sin3DM in diversity. As we discussed in Sec. 5.1, both metrics have their blindspots, SSFID overlooks geometric details and pairwise IoU rewards artifacts. Finding a more holistic metric to evaluate shape variation remains an open problem. + +tively. The upsamplers $\mathcal{U}$ consist of 4 layers of 64 channels, containing $\sim 55\mathrm{k}$ parameters that are trained for $10\mathrm{k}$ iterations with a learning rate of $5\cdot 10^{-4}$ and a $5\%$ dropout rate. The diffusion models have 128 feature channels and 7 layers, for a total of $\sim 565\mathrm{k}$ parameters for the coarsest model $\mathcal{M}^1$ and 1.2M parameters. As in prior work, the receptive fields of each model $\mathcal{M}^1$ are kept small to prevent overfitting to the fixed global structure: $\mathcal{M}^1$ thus uses a receptive field of $5^3$ , while the rest use $9^3$ . We train our diffusion models with $T = 1000$ diffusion steps. For sampling, we set $T[1] = 1000$ and $T[l > 1] = 300$ . $\mathcal{M}^1$ is trained for 20,000 iterations, and the rest for 40,000 iterations. All levels are trained with random crops of the same resolution to help ensure that each scale is trained in roughly the same time, and we use a learning rate of $10^{-4}$ with a $1\%$ dropout rate. + +Feature Extraction. In terms of shape processing, we normalize each mesh to fit within $[-1,1]^3$ . 3D features are sampled at a resolution of $1024^3$ and downsampled to a coarsest resolution of $16^3$ voxels, as described in Sec. 3.1. + +# 5. Experiments + +Data. We demonstrate our approach on 3D textured exemplars provided by Sin3DM (from [1, 4, 9, 19, 37, 47, 94]), and also used an open surface example that we created. Note that ShapeShifter can operate as-is on untextured inputs; but colors can help distinguish geometrically similar, yet semantically different parts of the geometry. + +# 5.1. Comparison. + +Baselines. We compare with two state-of-the-art papers on 3D generation from single examples: Sin3DM [106] and Sin3DGen [50]. Sin3DM uses a single-scale triplane diffusion model with a small receptive field to learn internal feature distribution within the exemplar shape. Features are learned in a separate autoencoder that parameterizes the input shape as an implicit neural surface [98]. We use their publicly available generated results for comparison. Sin3DGen operates instead on radiance field represented by plenoxels: it learns a hierarchical deformation field to transform the input plenoxels based on patch similarity. Follow- + +ing their data preparation guideline, we first rendered 200 images with Blender [7], then trained a $512^{3}$ plenoxel to obtain the input exemplar which we provide to Sin3DGen to generate results to which we compare ourselves. + +Quantitative Evaluations. Following prior work, we use Single Shape Fréchet Inception Distance (SSFID [105]) to evaluate the output quality, and pairwise (1-IoU)-distances among 10 generated variations to evaluate the output diversity. SSFID is extended from Single-Image Fréchet Inception Distance (SIFID) [86], which compares the statistics of the input and the generation feature extracted at different levels of a pretrained multiscale 3D encoder [16] trained on voxelized shapes from ShapeNet; thus we voxelized input and generations at a $256^{3}$ resolution for evaluation purposes. We show in Tab. 1 that ShapeShifter outperforms competing methods in SSFID. (Note that our SSFID scores for Sin3DM differ from their reported scores as they used post-processed smoothed shapes as reference; please refer to the supplementary material for details.) While SS-FIDs usually match the perceptual quality of the results, we note that the voxelization step used for scoring removes sharp features and high-frequency details, which our paper captures particularly well. The qualitative examples shown in Fig. 2 and in our supplementary material better demonstrate our strengths. Overall, the underlying geometry from Sin3DGen results is heavily distorted as it relies on plenoxel matching and blending, and while the SDF supervision from Sin3DM makes for better outputs, it does not capture sharp features well and the features extracted from their single-resolution triplane representation struggle to encode high-frequency details. Finally, the diversity scores based on IoU must be handled with care as it rewards artifacts. Our results show low diversity scores for very structured exemplars like the acropolis or the house, but good scores for more organic and varied shapes like the canyon or the small town, which is in line with our goal of generating variants of the input shapes. + +Training and inference speed. We show timing comparisons in Tab. 2, where we also included GAN-based SSG [105] as its architecture is faster than its successor at inference time, albeit with lower quality and diversity [106]. + +
Methodencoding and traininginference
SSG4 hours0.1 sec
Sin3DM2.5 hrs15.8 sec
Sin3DGen*15 min139 sec
ShapeShifter12 min10.7 sec
+ +Table 2. Timings. Sin3DGen [50] was tested on a more performant GPU (Nvidia A100 40GB), as it could not fit in our regular GPUs (Nvidia 3080 10GB). + +Despite its impressive inference time, the training time of SSG is the longest. Sin3DGen does not require training as it is based on patch-matching; but each variation generation takes around 2 minutes, limiting interactive use cases. Sin3DM takes in total of 2.5 hours to train, of which $\sim 30$ minutes are used to learn triplane features. Our method takes merely 6 seconds to encode the shape, and trains each level of the hierarchy of diffusion models in 12 minutes, significantly outperforming all other trained models. The inference time takes from 0.15 seconds (at level 2) to 7.5 seconds (at level 5) totaling 10.7 seconds. It is worth mentioning that since our method outputs colored oriented pointset at every level, requiring no additional conversion e.g. marching cubes in other methods, we can flexibly choose the working level depending on the application. For example, in editing, we can operate on level 3, which takes only 1 s. per generation, thus enabling interactive modeling as shown next. More details are provided in the supplementary material. + +![](images/995bd573261ff0ebeea66ec290e543f70422d54c291f040cb11f40eee8c11cba.jpg) +Figure 4. Controlled generation. The span of the output can be trivially controlled by resizing the initial grid anisotropically. + +# 5.2. Control and editing. + +Our multiscale explicit representation makes it easy to control and edit the output. We demonstrate two examples: the user can trivially change the span of the model by resizing the initial grid $\mathbf{G}_T^0$ anisotropically, see Fig. 4; moreover, Fig. 5 demonstrates that a generated output can be further edit by copy-and-pasting parts of the output within one of the levels of the multiscale description of the shape, here to remove windows or adding a bay window. While exist- + +ing works can offer similar capabilities, their use of triplane features or deformation fields to drive the generation renders editing less intuitive. For example, a patch from the input shape can be duplicated in Sin3DM to appear in the generated variations; however, the duplication must be done on three interdependent triplanes features, which do not directly correspond to a feature in 3D space. + +![](images/e3d3199522fdf0176e7f4ffbf8052d4677bd0bb7afb32c0144b322d95a8c01c1.jpg) +Figure 5. Editing. Using sparse voxel grid allows users to intuitively apply more precise edits. Here, a user can copy and paste a selected part of a generated variation at an intermediate level to manually alter the variation. + +# 5.3. Open Surfaces + +Our use of points and normals to represent the geometry makes the treatment of open surfaces not only possible but just as simple as the case of closed surfaces: only the surface extraction method needs to be altered, i.e. with APSS [26] instead of [43]. An example is shown in Fig. 6. + +![](images/418491ca10acfca14be044a7edff49deaac72500b9d4181dcd0f6fc73da585e2.jpg) +Figure 6. Open surfaces. Our oriented points representation also handles open surfaces by simply using APSS [26] to mesh the generated point set, while it is challenging for existing methods. + +# 5.4. Ablation studies + +Learned Upsampler. We demonstrate the benefit of our learned upsampler by replacing it (both in training and inference) by a trilinear interpolation as used in SinDDM in Fig. 7: artifacts appear as our point features are mispositioned because of the trilinear upsampling. + +3D Features. We also replace our geometric features with an SDF. For fairness of comparison, we use two layers of active cells (instead of one) close to the surface to compensate + +![](images/a29332a80a9c09b5d7b34d4f3b6ab7bfbbb93f73ffe75a917c266a0ad0ee951c.jpg) + +![](images/363244923fd76b75cf7a9bc7d29cef17909f53e5149ac6520b06266f7bd1ea00.jpg) + +![](images/379c92b0ffa3f6653d04ea6240eb34ca6b4fd325cb41084cd74ec95a475cc034.jpg) +(a) Trilinear Interpolation +Figure 7. Ablation test of upsampling. Comparing trilinear interpolation (left) with learned upsampler (right), the interpolation causes artifacts (see circled areas), whereas the learned upsampler provides a more detailed and structurally coherent output. + +![](images/b04064eda803aeff0a89decdae21ea8ed97f046c3cc198c0efe7313176676c66.jpg) +(b) Learned Upsampler + +for the reduced feature dimensionality. For the same input resolution, our features have more details (Fig. 8). + +![](images/6902a7dcb7816cdd700980908af48553c09712f5f72fa9dffdec6e7fa89887da.jpg) +(a) SDF + Marching Cube [60] +Figure 8. Ablation test of features. Comparing SDF (left) with our proposed point and normal features (right) at the same resolution $(128^{3})$ demonstrates that our proposed features produce richer geometric details. The mesh color encodes the normal direction to reflect the difference in geometry details. + +![](images/b65c482a6eccd738ff7d9bc770f73de12918e17db15b31aca328056793310838.jpg) +(b) Ours $^+$ Poisson [43] + +![](images/67868e29d131d1308182a7ed900cb18c849f4471d541c0a3533a7855bf2bfbbb.jpg) + +![](images/6e123d7bbc961c05c7e3d1328e6c99993cd6703fca4e6435c8043c42748b7be8.jpg) +Figure 9. Texturing. As explained Sec. 3.3, ultra-high resolution texture can be obtained by applying state-of-the-art AI image-enhancing tools on the texture maps created by our method from the colored point set outputs. + +# 5.5. Texture Augmentation + +Finally, we show that one can texture our generated models by applying contemporary image super-resolution on the baked texture maps in Fig. 9: using Magnific AI [63] for example can efficiently generate a fine texture improving the visual impact of our results. While this is only a proof-of-concept example, exploring the texturing of our geometric models is an exciting, albeit orthogonal, research direction. + +# 6. Limitations and Future Work + +Just like previous exemplar-based generative methods, ShapeShifter is limited in the shape variations it can generate: while our approach is not strictly patch-based, it is similarly restricted in its ability to consider widely different variants. Extending its range of alterations through data augmentation or more involved (equivariant) features remains an intriguing possibility that would broaden the applicability of our method. Moreover, we focused our approach on generating high-quality, detailed geometry and did not consider fine texture generation. While existing exemplar-based methods have proposed approaches to generate textures for meshes that we could apply as-is, we believe there may be other exciting possibilities to explore, such as fitting 2D Gaussian splats [35] within our finest voxels to enrich our geometry with radiance field reconstruction. + +Now that we have proven the efficacy of explicit geometry encoding through colored points and normals for creating shapes in our exemplar context, it would be interesting to study its adequacy in the more general case of generative modeling from large datasets: its lightweight, surface-based nature may circumvent a number of issues plaguing current state-of-the-art approaches. + +# 7. Conclusion + +We proposed a novel generative approach for generating high-quality and detailed 3D models from a single exemplar. Our approach stands out as the first 3D generative method based on an explicit encoding of geometry through points, normals, and optionally colors. Combined with sparse voxel grid, we demonstrated that both training and inference times are (at times drastically) reduced compared to previous methods, despite a significantly improved quality of our geometric outputs and an ability to deal seamlessly with closed or open surfaces alike. We thus believe that ShapeShifter sets a new standard for the quality of geometric outputs in generative modeling. + +# 8. Acknowledgments + +This work was supported by the French government through the 3IA Cote d'Azur Investments in the project managed by the National Research Agency (ANR-23-IACL-0001), Ansys, Adobe Research, and a Choose France Inria chair. + +# References + +[1] All about Blender-3D. Photo-realistic floating wood. http://www.cgtrader.com/free-3d-models/plant/other/photo-realistic-floating-wood, 2020. License: Royalty Free. +[2] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International Conference on Machine Learning, pages 40-49. PMLR, 2018. +[3] and others. State of the art on diffusion models for visual computing. In Computer Graphics Forum, page e15063. Wiley Online Library, 2024. +[4] Pedram Ashoori. Small town. https://www.cgtrader.com/free-3d-models/exterior/cityscape/small-town-87b127c8-c991-4063-AA69-e58800686299, 2020. License: Royalty Free. +[5] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. +[6] Yoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with gradient descent is difficult. IEEE Transactions on Neural Networks, 5(2):157-166, 1994. +[7] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. +[8] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4169-4181, 2023. +[9] Lukas Carnota. Industrial building. https://www.cgtrader.com/free-3d-models/exterior/office/indusrtial-building, 2015. License: Royalty Free. +[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. +[11] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2416-2425, 2023. +[12] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22246–22256, 2023. +[13] Ting Chen. On the Importance of Noise Scheduling for Diffusion Models, 2023. arXiv:2301.10972. +[14] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang + +Yu, et al. Meshanything: Artist-created mesh generation with autoregressive transformers. arXiv preprint arXiv:2406.10163, 2024. +[15] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[16] Zhiqin Chen, Vladimir G Kim, Matthew Fisher, Noam Aigerman, Hao Zhang, and Siddhartha Chaudhuri. Decorgan: 3d shape detailization by conditional refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15740-15749, 2021. +[17] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. +[18] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. +[19] DJMaesen. Cliff. https://sketchfab.com/3d-models/cliff-082da1166a814c6e9c9e6c1b38 159e4e, 2021. License: CC Attribution. +[20] Ariel Elnekave and Yair Weiss. Generating natural images with direct patch distributions matching. In European Conference on Computer Vision, pages 544-560. Springer, 2022. +[21] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. +[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. +[23] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. +[24] Niv Granot, Ben Feinstein, Assaf Shocher, Shai Bagon, and Michal Irani. Drop the gan: In defense of patches nearest neighbors as single image generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13460-13469, 2022. +[25] Alex Graves. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850, 2013. +[26] Gael Guennebaud and Markus Gross. Algebraic point set surfaces. In ACM SIGGRAPH 2007 papers, page 23, San Diego California, 2007. ACM. +[27] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. +[28] Niv Haim, Ben Feinstein, Niv Granot, Assaf Shocher, Shai Bagon, Tali Dekel, and Michal Irani. Diverse generation + +from a single video made possible. In European Conference on Computer Vision, pages 491-509. Springer, 2022. +[29] Charles Han, Eric Risser, Ravi Ramamoorthi, and Eitan Grinspun. Multiscale texture synthesis. In ACM SIGGRAPH 2008 papers, pages 1-8. 2008. +[30] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: a network with an edge. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. +[31] Amir Hertz, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Deep geometric texture synthesis. ACM Trans. Graph., 39(4), 2020. +[32] Tobias Hinz, Matthew Fisher, Oliver Wang, and Stefan Wermter. Improved techniques for training single-image gans. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1300-1309, 2021. +[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing systems, 33:6840-6851, 2020. +[34] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3d. In The Twelfth International Conference on Learning Representations, 2024. +[35] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splattering for geometrically accurate radiance fields. In ACM SIGGRAPH Conference Papers, 2024. +[36] Dong Huo, Zixin Guo, Xinxin Zuo, Zhihao Shi, Juwei Lu, Peng Dai, Songcen Xu, Li Cheng, and Yee-Hong Yang. Texgen: Text-guided 3d texture generation with multi-view sampling and resampling. In European Conference on Computer Vision, pages 352-368. Springer, 2025. +[37] ImpJive. Fighting pillar. https://sketchfab.com/3d-models/fighting-pillar-14e73d2d9e8a4981b49d0e6c56d30af5, 2021. License: CC Attribution-ShareAlike. +[38] Alec Jacobson and Daniele Panozzo. libigl, 2023. +[39] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. +[40] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. +[41] Animesh Karnewar, Oliver Wang, Tobias Ritschel, and Niloy J Mitra. 3ingan: Learning a 3d generative model from images of a self-similar scene. In International Conference on 3D Vision (3DV), pages 342-352. IEEE, 2022. +[42] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3907-3916, 2018. +[43] Michael Kazhdan and Hugues Hoppe. Screenedoisson surface reconstruction. ACM Trans. Graph., 32(3):1-13, 2013. + +[44] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. +[45] Diederik P Kingma, Max Welling, et al. An introduction to variational autoencoders. Foundations and Trends in Machine Learning, 12(4):307-392, 2019. +[46] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In International Conference on Machine Learning, pages 17920-17930. PMLR, 2023. +[47] Choly Kurd. Akropolis. https://www.turbosquid.com/3d-models/acropolis-3ds-free/610885, 2021. License: Educational Uses. +[48] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. +[49] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In International Conference on Learning Representations, 2024. +[50] Weiyu Li, Xuelin Chen, Jue Wang, and Baoquan Chen. Patch-based 3d natural scene generation from a single example. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16762-16772, 2023. +[51] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. In International Conference on Learning Representations, 2024. +[52] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6517–6526, 2024. +[53] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. +[54] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. +[55] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. +[56] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. + +[57] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In International Conference on Learning Representations, 2024. +[58] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: learning dynamic renderable volumes from images. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. +[59] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. +[60] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Seminal graphics: pioneering efforts that shaped the field, pages 347-353. 1998. +[61] Artem Lukoianov, Haitz Sáez de Ocariz Borde, Kristjan Greenewald, Vitor Campagnolo Guizilini, Timur Bagautdinov, Vincent Sitzmann, and Justin Solomon. Score distillation via reparametrized ddim. arXiv preprint arXiv:2405.15891, 2024. +[62] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. +[63] Magnific AI. Magnific ai: Accelerating scientific research. https://magnific.ai/. Accessed: 2024-11-13. +[64] Nissim Maruani, Roman Klokov, Maks Ovsjanikov, Pierre Alliez, and Mathieu Desbrun. VoroMesh: Learning Watertight Surface Meshes with Voronoi Diagrams. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14519-14528, Paris, France, 2023. IEEE. +[65] Nissim Maruani, Maks Ovsjanikov, Pierre Alliez, and Mathieu Desbrun. PoNQ: A Neural QEM-Based Mesh Representation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3647-3657, Seattle, WA, USA, 2024. IEEE. +[66] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. +[67] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. +[68] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. +[69] Thomas W. Mitchel, Carlos Esteves, and Ameesh Makadia. Single mesh diffusion models with field latents for texture + +generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7953-7963, 2024. +[70] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International Conference on Machine Learning, pages 7220-7229. PMLR, 2020. +[71] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. +[72] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: training diffusion models on a single image or video. In Proceedings of the 40th International Conference on Machine Learning, pages 26199-26214, 2023. +[73] Eyvind Niklasson, Alexander Mordvintsev, Ettore Randazzo, and Michael Levin. Self-organising textures. Distill, 6(2):e00027-003, 2021. +[74] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. +[75] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An Imperative Style, High-Performance Deep Learning Library, 2019. arXiv:1912.01703 [cs, stat]. +[76] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations, 2023. +[77] Sergey Prokudin, Qianli Ma, Maxime Raafat, Julien Valentin, and Siyu Tang. Dynamic Point Fields. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7964-7976, 2023. +[78] 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. +[79] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9914–9925, 2024. +[80] Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4209-4219, 2024. +[81] Xuanchi Ren, Yifan Lu, Hanxue Liang, Jay Zhangjie Wu, Huan Ling, Mike Chen, Francis Fidler, Sanja + +andn Williams, and Jiahui Huang. Scube: Instant largescale scene reconstruction using voxsplats. In Conference on Neural Information Processing Systems, 2024. +[82] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International Conference on Machine Learning, pages 1530-1538. PMLR, 2015. +[83] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH Conference Proceedings, pages 1–11, 2023. +[84] Carlos Rodriguez-Pardo and Elena Garces. Seamlessgan: Self-supervised synthesis of tileable texture maps. IEEE Transactions on Visualization and Computer Graphics, 29 (6):2914-2925, 2022. +[85] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. +[86] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4570-4580, 2019. +[87] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3d generation. In International Conference on Learning Representations, 2024. +[88] Assaf Shocher, Shai Bagon, Phillip Isola, and Michal Irani. Ingan: Capturing and retargeting the"dna" of a natural image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4492-4501, 2019. +[89] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19615-19625, 2024. +[90] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. +[91] Minjung Son, Jeong Joon Park, Leonidas Guibas, and Gordon Wetzstein. Singraf: Learning a 3d generative radiance field for a single scene. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8507-8517, 2023. +[92] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. +[93] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multiview gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. +[94] Simon Ustal. Canyon landscape. https://sketchfa b.com/3d-models/canyon-landscape-c395e9 eb54ba4f40820ccfb98d3c2832, 2020. License: CC Attribution. + +[95] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. +[96] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International Conference on Machine Learning, pages 1747-1756. PMLR, 2016. +[97] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. +[98] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. +[99] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. PF-LRM: Pose-free large reconstruction model for joint pose and shape prediction. In International Conference on Learning Representations, 2024. +[100] Peihao Wang, Dejia Xu, Zhiwen Fan, Dilin Wang, Sreyas Mohan, Forrest Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu, Zhangyang Wang, et al. Taming mode collapse in score distillation for text-to-3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9037–9047, 2024. +[101] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. +[102] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. +[103] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. +[104] Francis Williams, Jiahui Huang, Jonathan Swartz, Gergely Klar, Vijay Thakkar, Matthew Cong, Xuanchi Ren, Ruilong Li, Clement Fuji-Tsang, Sanja Fidler, et al. fvdb: A deep-learning framework for sparse, large scale, and high performance spatial intelligence. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. +[105] Rundi Wu and Changxi Zheng. Learning to generate 3d shapes from a single example. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. +[106] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3DM: Learning a diffusion model from a single 3d textured shape. In International Conference on Learning Representations, 2024. +[107] 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. +[108] Yinghao Xu, Zifan Shi, Wang Yifan, Sida Peng, Ceyuan Yang, Yujun Shen, and Wetzstein Gordon. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arxiv: 2403.14621, 2024. +[109] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, and Kai Zhang. DMV3d: Denoising multi-view diffusion using 3d large reconstruction model. In International Conference on Learning Representations, 2024. +[110] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. +[111] Haibo Yang, Yang Chen, Yingwei Pan, Ting Yao, Zhi-neng Chen, Zuxuan Wu, Yu-Gang Jiang, and Tao Mei. Dreammesh: Jointly manipulating and texturing triangle meshes for text-to-3d generation. arXiv preprint arXiv:2409.07454, 2024. +[112] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öz Tireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. +[113] Fukun Yin, Xin Chen, Chi Zhang, Biao Jiang, Zibo Zhao, Jiayuan Fan, Gang Yu, Taihao Li, and Tao Chen. Shapept: 3d shape generation with a unified multi-modal language model. arXiv preprint arXiv:2311.17618, 2023. +[114] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. +[115] Biao Zhang, Matthias Nießner, and Peter Wonka. 3dilig: Irregular latent grids for 3d generative modeling. Advances in Neural Information Processing Systems, 35:21871-21885, 2022. +[116] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. +[117] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splattering. In European Conference on Computer Vision, pages 1-19. Springer, 2025. +[118] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. +[119] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, Bin Fu, Tao Chen, Gang Yu, and Shenghua Gao. + +Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. Advances in Neural Information Processing Systems, 36, 2024. +[120] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. +[121] Yang Zhou, Zhen Zhu, Xiang Bai, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Non-stationary texture synthesis by adversarial expansion. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. +[122] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. \ No newline at end of file diff --git a/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/images.zip b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d30346ee2e4ee6ff0e74dbfa78835904164cce24 --- /dev/null +++ b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13ec292cd51ae144001a57d2336de2838d115faef5aa6778ad3340ae47153317 +size 420459 diff --git a/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/layout.json b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c58fb3f7bf4c575bb95ff8f46756fb221127436a --- /dev/null +++ b/CVPR/2025/ShapeShifter_ 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3efbe30ac0f6d224e8d447e26b3356fd23ff313e54fee52b030a15ab321eeb5c +size 542332 diff --git a/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/662848fa-f27a-44b0-aab6-66ad47c58ea6_content_list.json b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/662848fa-f27a-44b0-aab6-66ad47c58ea6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..91c785aed3bee5f3b407b7057c04f3430472bc26 --- /dev/null +++ b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/662848fa-f27a-44b0-aab6-66ad47c58ea6_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297bb8f614e07261e4a9d78fa5e9396d197db1f1d50289aba547b38cccd4fc25 +size 70078 diff --git a/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/662848fa-f27a-44b0-aab6-66ad47c58ea6_model.json b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/662848fa-f27a-44b0-aab6-66ad47c58ea6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..553efdc4b1b5d21c1e3b95dba4f872978af95333 --- /dev/null +++ b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/662848fa-f27a-44b0-aab6-66ad47c58ea6_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52b67d79992f8ed0ced3ac2db43a2850a0de9c8902011f1eb2ccdd96e517725d +size 88795 diff --git a/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/662848fa-f27a-44b0-aab6-66ad47c58ea6_origin.pdf b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/662848fa-f27a-44b0-aab6-66ad47c58ea6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..42e5d10829c9eb14afdf8986bf1a7f49fd9f9c10 --- /dev/null +++ b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/662848fa-f27a-44b0-aab6-66ad47c58ea6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dddf9aa71adc9cf7d09741f8de5116b5819fbec9a08dc4d7b5288ee2b5da317a +size 6544739 diff --git a/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/full.md b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ffe4b2a3adbf8bdc46a550265c29c317a840d086 --- /dev/null +++ b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/full.md @@ -0,0 +1,274 @@ +# ShapeWords: Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts + +Dmitry Petrov1 Pradyumn Goyal1 Divyansh Shivashok1 Yuanming Tao1 Melinos Averkiou2,3 Evangelos Kalogerakis1,2,4 + +$^{1}$ UMass Amherst $^{2}$ CYENS CoE $^{3}$ University of Cyprus $^{4}$ TU Crete + +![](images/96282bb8873e227e8fa24dec468d11aab3bf14b51ceb8b694743297f64d3f8b5.jpg) +Figure 1. ShapeWords enables 3D shape-aware text-to-image generation via mapping of shape geometries into CLIP space. Given an input 3D shape and text prompts describing desired appearance and context, our method generates images that maintain both shape fidelity and text compliance. Unlike existing methods that use view-dependent guidance like depth maps, ShapeWords generalizes to compositional settings (top row) and allows for exploration of target geometries with stylistic deviations aligned with the prompt (bottom row). + +# Abstract + +We introduce ShapeWords1, an approach for synthesizing images based on 3D shape guidance and text prompts. ShapeWords incorporates target 3D shape information within specialized tokens embedded together with the input text, effectively blending 3D shape awareness with textual context to guide the image synthesis process. Unlike conventional shape guidance methods that rely on depth maps restricted to fixed viewpoints and often overlook full 3D structure or textual context, ShapeWords generates diverse yet consistent images that reflect both the target shape's geometry and the textual description. Experimental results show that ShapeWords produces images that are more text-compliant, aesthetically plausible, while also maintaining 3D shape awareness. + +# 1. Introduction + +Recent advances in generative image models based on diffusion [2, 17, 36-38, 41, 44] have made it possible to generate impressive imagery from input text prompts. A chal + +lenge in text-to-image models has been to provide users with fine-grained control over shapes or forms in the synthesized images, which can be difficult to convey through text descriptions alone. To address this, conditioning methods have been proposed, such as ControlNet [51] and IP-adapter [48], that aim to capture the desired shape or form more explicitly through the use of edge or depth maps as input conditions. + +Despite these advancements, current text- and image-conditioned synthesis approaches still face a number of challenges. First, they often struggle to balance both textual and visual conditions, when text describes a particular context that should be combined with the target shape to guide an image synthesis (Figure 1, top row). Second, commonly used visual conditions such as edge or depth maps are limited to a single viewpoint, resulting in a loss of valuable 3D shape information when users seek image variations of an underlying shape from different poses. Third, even when these models accurately reflect the target shape in specific views, users may want to explore shape variations – yet current models often lack flexible controls for such exploration. + +To overcome these challenges, we propose ShapeWords, a method designed to generate images that faithfully adhere to both the text prompt and a target 3D shape geometry, + +while at the same time allows users to easily explore both pose and shape variations. At the heart of ShapeWords is the embedding of the 3D shape into a shared space with textual descriptions, specifically OpenCLIP [10] space, enabling the target shape to be fully integrated with the prompt. Unlike view-specific conditioning (e.g., through depth, edge, or normal maps), our approach captures the entire 3D shape in the embedding, facilitating diverse image synthesis across challenging prompts, yet remaining geometrically consistent with the target 3D shape (Figure 1, top row). Furthermore, ShapeWords allows users to control the degree of deviation from the target shape – a feature that is particularly useful when the target structure is represented by coarse geometric primitives. This capability enables the generation of diverse, stylized images that retain plausible geometric variations of the specified target structure (Figure 1, bottom row). + +Our experiments evaluate ShapeWords on consistency with both text and target shape, as well as aesthetic quality. It achieves significant improvements against ControlNet-based variants, as measured by several metrics capturing target shape compliance, text prompt compliance, image plausibility, and also as validated by a perceptual user study. Overall, our technical contributions are as follows: + +- We introduce 3D shape tokens - specialized tokens that enable text-to-image models to generate plausible images adhering to both 3D geometry and textual conditions. +- We also enable user control over the shape guidance, allowing users to explore images that depict variations of target shapes across varying poses and appearances. + +# 2. Prior work + +Conditional Diffusion. Denoising diffusion models [17, 44] have revolutionized image synthesis by generating high-quality, diverse, and plausible image content. To control the generation process, the most common condition is text. Popular diffusion approaches, such as Stable Diffusion [38], DALLE [2, 36, 37], Imagen [41] have used large-scale text-image datasets, including pretrained LLMs [6, 11, 35] for text-to-image synthesis. However, relying solely on text prompts cannot fully take advantage of the knowledge learned by diffusion models, especially when flexible and accurate control is needed in terms of form or layout. To this end, ControlNet [51] pioneered explicit conditioning through visual signals like depth or edge maps, enabling spatial control but limited to single viewpoints. T2I-Adapter [27] proposed a lighter-weight alternative through specialized adapters of visual control signals. UniControl [33] consolidated a wide array of controllable condition-to-image tasks within a single framework. IP-Adapter [48] proposed a decoupled cross-attention strategy for text features and image features from input conditions for more accurate controllable generation. In general, + +these methods rely on view-dependent control signals (e.g., depth, edge maps), lack explicit 3D geometry awareness, and may fail to generate images adhering to both the text prompt and target 3D shapes. In contrast, ShapeWords enables view-independent shape control, while adhering well both to target shape geometry and text prompts. + +Structure guidance. Several approaches have explored novel view synthesis aimed at maintaining some level of 3D awareness for view consistency [15, 23-25, 46]. However, these methods do not disentangle 3D structure from appearance and do not provide geometric control to users, meaning they generate images of a shape with consistent appearance across different views without user-driven control over geometry. A number of approaches have been proposed to incorporate structure guidance from depth maps or coarse 3D primitives disentangled from appearance. FreeControl [26] explores structure guidance in diffusion feature subspaces extracted from depth maps and other visual conditions. Ctrl-X [22] investigated more efficient, disentangled and zero-shot control of structure and appearance. LooseControl [4] introduced more flexible conditioning schemes through 3D scene boundary control, 3D box control and attribute editing. Diffusion Handles [29] enabled localized control of 3D object parts in diffusion models by introducing deformation handles to edit images while maintaining consistent perspective and structure across views. In our approach, we take a different route by embedding 3D shapes into tokens within text prompts, providing more explicit geometry guidance in image synthesis and combining it with additional specification of separate context, style, and appearance constraints through text. + +Several methods have explored embedding concepts into textual tokens for personalized image synthesis using textual inversion [13]. The required optimization for textual inversion is often very slow, thus various methods have explored more efficient fine-tuning strategies [19], feedforward architectures to predict textual tokens [21, 43], or using hypernets [39]. Recently, more efficient methods of visual concept learning were introduced [1, 12, 14, 40] allowing for more efficient learning and transfer of visual concepts. A more spiritually similar approach to ours is that of "continuous 3D words"[9], which embeds 3D-aware attributes, such as time-of-day lighting, bird wing orientation, dolly zoom effects, and object pose, into learnable tokens. Viewpoint textual inversion[7] learns 3D view tokens that can be used to control the viewpoint for image synthesis. However, unlike these methods, our approach learns to embed 3D shapes directly into tokens, enabling image generation that is guided by both target 3D shape geometry and text. To our knowledge, text-to-image synthesis through + +![](images/bdc53f9f9cfb8b3a6fbda8ab8e42d630061742dbceb0ee0c8f449a8a3fdfdc11.jpg) +Figure 2. ShapeWords Training and Inference. During training (i), ShapeWords takes as input triplets of a shape, a prompt, and an image. The shape $S$ and prompt $T$ are encoded using the shape encoder PointBert and the text encoder OpenCLIP, respectively. The resulting embeddings are passed through a cross-attention-based Shape2CLIP module, which produces a prompt residual $\delta T$ to guide the source prompt toward the target geometry. This modified prompt is then passed as input to a Text-to-Image Denoising UNet along with sampled noisy image latents and time step $t$ . The Shape2CLIP module is optimized via Score Distillation Sampling. During inference (ii), the CLIP embeddings of the input prompt and the target embeddings of the test shape $S_{\mathrm{test}}$ are passed through the Shape2CLIP module. Optionally, the desired strength of the shape guidance is controlled by the user parameter $\lambda$ . + +learned 3D shape words has not been explored before. + +# 3. Method + +Overview. Given a text prompt $\mathbf{p}$ and a target 3D shape $S$ , ShapeWords generates an image that reflects both the textual description and the desired shape. Users indicate the desired shape directly within the prompt using a special token, such as "a red [SHAPE-ID] on a beach", where "SHAPE-ID" corresponds to a shape, e.g., one imported from a 3D shape database. + +The pipeline (illustrated in Figure 2) proceeds as follows at test time. First, a shape representation is extracted from the chosen 3D shape using a pre-trained transformer, Point-BERT [49]. The text prompt is mapped into CLIP space through an OpenCLIP encoder [10], where the token "SHAPE-ID" is replaced by a category name for the shape (e.g., "chair"). Our method then applies a new module, named Shape2CLIP, trained to modify the prompt's word embeddings, including the shape identifier token, so that the resulting prompt embedding integrates the desired 3D shape geometry in the prompt while preserving its original textual context. This shape-enhanced embedding is passed to a Stable Diffusion model [38], along with a user-defined parameter that controls the degree of 3D shape influence, to synthesize images consistent with both the textual and 3D shape cues. + +In the following sections, we discuss preprocessing of the input shapes and text prompts to extract Point-Bert [49] and OpenCLIP [10] representations respectively (Section 3.1), then we discuss the Shape2CLIP module (Section 3.2), our inference pipeline (Section 3.3), and finally its training procedure (Section 3.4). + +# 3.1. Preprocessing + +Point-BERT. To represent the input 3D shape, we use a pre-trained Point-BERT [49], a transformer-based model for point clouds. The input shape is first converted into a point cloud of 1024 points using farthest point sampling, then passed through the Point-BERT architecture, which partitions the cloud into 64 patches. These patches are encoded by PointNet [32] and a transformer based encoder as a set of tokens representing the geometry of the shape. The resulting shape embedding, $\mathbf{B} \in \mathbb{R}^{65 \times 384}$ , consists of 65 tokens representing both patches and a class token in a 384-dimensional space. Trained via masked modeling for self-supervised learning, Point-BERT effectively captures structure-aware features demonstrated in various shape processing applications, such as part segmentation. + +OpenCLIP space. As our base generative model, we use Stable Diffusion 2.1 whose text encoder is OpenCLIP ViTH/14 [10]. Thus, we process the input textual prompt $\mathbf{p}$ , through this OpenCLIP encoder to generate a sequence of 77 token embeddings capturing the words in the prompt and their context in the prompt: + +$$ +\mathbf {T} = \left[ \mathbf {t} _ {0}, \mathbf {t} _ {1}, \dots , \mathbf {t} _ {j} ^ {\text {e o s}}, \dots , \mathbf {t} _ {7 6} ^ {\text {p a d}} \right] \tag {1} +$$ + +where $\mathbf{t}_j$ represents the encoded embedding of the $j$ -th token in the prompt; $\mathbf{t}_j^{eos}$ represents end-of-sequence (EOS) token which captures the context of the whole prompt; and $\mathbf{t}_j^{pad}$ represent padding token embeddings that pad sequence after EOS tokens to some predefined length (e.g. 77). + +# 3.2. Shape2CLIP module + +Given a text prompt embedding $\mathbf{T}$ and shape representation $\mathbf{B}$ , our Shape2CLIP module generates a modified + +prompt embedding $\mathbf{T}'$ that combines information from both the text prompt and 3D shape. Specifically, we update the embeddings of two tokens. First, we update the shape identifier token(s), which was originally set to the category name of the shape (e.g. 'chair'), such that its embedding instead reflects the specific 3D shape. We also found it important to update the EOS token since it captures the context of the whole prompt, including the desired shape to be incorporated. Formally, the module implements a learned function: + +$$ +\mathbf {T} ^ {\prime} [ s, e ] = \mathbf {T} [ s, e ] + \delta \mathbf {T} (\mathbf {B}, \mathbf {T}; \theta) \tag {2} +$$ + +where $s$ and $e$ are the position of the shape and EOS tokens, and $\delta \mathbf{T}$ is a residual mapping with learned parameters $\theta$ . We found this residual approach to be more effective than a direct feedforward network, as it is less prone to overfitting with limited training data. We found that updating only these specific two tokens preserves the prompt's original context while integrating geometric information. + +The residual function $\delta \mathbf{T}$ is constructed using crossattention blocks [45]. In the first block, the shape representation $\mathbf{B}$ is treated as keys, while the original prompt embedding $\mathbf{T}$ serves as queries: + +$$ +\mathbf {K} = \mathbf {B} \cdot W _ {k}, \mathbf {Q} = \mathbf {T} \cdot W _ {q}, \mathbf {V} = \mathbf {B} \cdot W _ {v} \tag {3} +$$ + +and new shape-aware features for the prompt $\mathbf{X}$ are obtained as: + +$$ +\mathbf {X} = \text {C r o s s A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) + \mathbf {T} \tag {4} +$$ + +followed by a MLP-based residual block [45]. In subsequent blocks, the updated prompt features act as queries, with Point-BERT representations serving as keys to further refine the prompt embedding. Our implementation uses a total of six cross-attention blocks. The output of last block is passed through a linear layer to yield the final modified embedding $\mathbf{T}'$ . + +# 3.3. Guided Diffusion + +The modified embedding $\mathbf{T}'$ can be directly used as input to a diffusion model to generate images, formulated as $\mathbf{I} = D(\mathbf{z},\mathbf{T}')$ , where $\mathbf{z}$ in the latent space and $D$ represents Stable Diffusion 2.1 [38] in our implementation. Notably, our method does not require any additional shape conditions (e.g., depth, normal, or edge maps). + +Shape Guidance. Our method enables users to control the influence of the 3D shape, allowing for image variations that deviate from the target shape. This flexibility is valuable when the exact shape a user has in mind does not exist in any available 3D shape databases, deeming more necessary to explore variations of existing shapes. The shape influence is modulated by a parameter $\lambda \in [0,1]$ using a linear interpolation scheme to adjust the embedding as follows: + +$$ +\mathbf {T} ^ {\prime} [ s, e ] = \mathbf {T} [ s, e ] + \lambda \cdot \delta \mathbf {T} (\mathbf {B}, \mathbf {T}; \theta) \tag {5} +$$ + +As shown in our results, varying $\lambda$ from 0 to 1 gradually shifts from disregarding the shape influence to incorporating it in the generated images. Intermediate values yield plausible images, while higher $\lambda$ values produce shapes in the generated images that increasingly match the desired target shape. + +# 3.4. Training + +Our training procedure aims to learn the parameters $\theta$ of the Shape2CLIP module, keeping the rest of the pipeline components fixed (i.e., PointBERT encoder, text encoder, image encoder/decoder, and denoising network). + +Training dataset. To train the Shape2CLIP module, we constructed a dataset of shape-prompt-image triplets based on ShapeNet [8] reference shapes. Images were generated using ControlNet [51] conditioned on ShapeNet's depth maps, as provided by the ULIP authors [47], where each shape is rendered from 30 viewpoints, rotated in 12-degree increments around the vertical axis with fixed elevation (see [47] for details). + +For each depth image, we applied a randomly selected prompt from a set of 13716 prompts for additional ControlNet conditioning, created from a base set of 100 prompts, then augmented with variants produced by ChatGPT [28]. To achieve diversity and structural agnosticism in prompts, this set was created by combining 127 artistic mediums (e.g., "painting," "watercolor," "sketch") with 108 style adjectives (e.g., "colorful," "pixelated," "fantasy"), deliberately avoiding references to specific 3D structures to help the model learn generalizable mappings instead of overfitting to particular geometries or appearance combinations. To reduce bias and enhance background diversity, we used the Stable Diffusion's inpainting model [38] to modify the backgrounds while preserving the foreground objects. The above procedure resulted in a diverse dataset of 1.58M prompt-image pairs (30 per each ShapeNet shape), generated from all depth images in the training split from 3DILG [50] in ShapeNet. As demonstrated in Section 4, although our method is trained on data generated by ControlNet, ShapeWords exhibits strong generalization capabilities. Notably, ShapeWords achieves significantly better performance on compositional prompts, a setting that ControlNet struggles to handle effectively. The dataset will be publicly released along with our source code upon acceptance. + +SDS-based training. We train our Shape2CLIP model using the Score Distillation Sampling (SDS) loss [31]. Specifically, for sampled noise $\epsilon_{t,i} \sim \mathcal{N}(0,\mathbf{I})$ for a training image $i$ at step $t$ , we use the pretrained Stable Diffusion + +model to predict the noise $\hat{e}_{t,i}$ , and backpropagate the loss: + +$$ +\mathcal {L} _ {\mathrm {S D S}} (\theta) = W (t) \| \hat {\epsilon} _ {i, k} - \epsilon_ {i, k} \| _ {2} ^ {2}, \tag {6} +$$ + +where $W(t)$ is a time-dependent weighting function proposed in DreamTime [18] for enhancing the stability of the SDS-based training. We provide details in supplementary. + +# 4. Evaluation + +We now discuss our evaluation and experiments to test the effectiveness of ShapeWords compared to alternatives. + +Evaluation goals & metrics. The evaluation aims to assess three main aspects: + +(a) Prompt Adherence: We evaluate how well the images generated by each competing method align with the given prompt. We use the standard score of CLIP similarity [34]. (b) Shape Adherence: We assess how well the generated images from each method adhere to a reference 3D shape. To this end, we compare the shapes in the generated images with the reference 3D shapes based on their silhouette. To extract silhouettes from reference 3D shapes, we render them from a target pose, extract the silhouette, then we compare it with the silhouette of the shape in the generated images conditioned on the target pose. The silhouette from generated images are extracted through the foreground detection model [52]. The similarity between silhouettes is measured using standard geometric similarity metrics, specifically Intersection over Union (S-IOU) and Chamfer Distance (S-CD), averaged per shape across six uniform views sampled with 60 degree increments, starting at 0. +(c) Image Plausibility: We evaluate the aesthetic quality of the generated images using the aesthetics score (Aes.) [42]. We also include the commonly used image quality generation metrics of FID [16] and KID [5]. Given reported issues with Inception-based features [20, 30], we compute FID and KID on CLIP features as recommended in [3]. + +Test Datasets. We have designed two datasets for evaluation to test different properties of our model: + +(a) Simple Prompts Dataset: This dataset contains "simple prompts" with text structured as "a photo of a [SHAPE-ID]," where the [SHAPE-ID] token corresponds to a 3D shape from the ShapeNet test splits from all 55 categories (2592 test shapes). This dataset is particularly useful for evaluating shape adherence through geometric similarity of extracted silhouettes (using S-CD and S-IOU metrics), as the generated images are photo realistic and feature a single object, so the foreground detection methods perform reliably here. +(b) Compositional Prompts Dataset: This dataset serves as our main evaluation set since it is more challenging, containing "compositional prompts" that involve a target 3D + +![](images/a4c986f05b5fa9e3b4d8f8e63fa25fa43feaa51ceeacc15d0184a9a8dcf0b5c5.jpg) +Figure 3. Quantitative evaluation of shape adherence. Shape-Words@K consistently outperforms the corresponding "CNet-Stop@K" variants with text guidance. "Subcat. prompts" refers to subcategory prompts (e.g. 'a photo of an office chair'), which are the finest annotation level in ShapeNet. + +![](images/219f13b907900105e56d4ce02dfbbd412bc3db5277455a2bd8fe3111913f5413.jpg) + +shape alongside additional objects or humans interacting with it. It contains five prompts not present in training data that are designed to test compositional properties of studied models: “a [SHAPE-ID] under a tree”, “a craftsman working on a [SHAPE-ID]”, or “a toy [SHAPE-ID] in a box”, “a [SHAPE-ID] in a snow globe”, “an artist painting [SHAPE-ID]”. Here, the [SHAPE-ID] tokens correspond to 3D shapes from the ShapeNet test split we use in our evaluation (2592 shapes). We report all evaluation metrics for this dataset except for silhouette similarity metrics, as the foreground in these scenarios is more complex, often containing multiple interacting objects, leading to noisier silhouette extractions. For this dataset, we conduct a user study to assess both shape and prompt adherence from a perceptual standpoint. + +![](images/e1de72490e50e710e24bc4342a7315751b1d776262a92f5ed7ad1baf54c6006b.jpg) +Figure 4. Shape adherence examples. Compared to text-based guidance, ShapeWords@20 produces shapes that are significantly more consistent with target geometry, compared to the CNet-Stop@20 conditioned either on category or subcategory prompts. + +
ModelFID ↓KID ↓Aes. ↑CLIP ↑
ControlNet97.010.405.2426.9
CNet-Stop@6090.510.255.2028.3
CNet-Stop@8092.49.725.1727.5
ShapeWords73.88.585.4531.5
+ +Table 1. Evaluation results on compositional prompts. Our method consistently outperforms over constrained ControlNet variants in the challenging compositional setting. KID metric and CLIP score are multiplied by 100 + +Baselines. Our experiments involve the following methods: + +(a) ControlNet: ControlNet receives two inputs: (i) a depth image rendered from a target 3D shape in a specified pose (we use inverted depth images provided by the ULIP project [47]), and (ii) a text prompt from our test datasets where [SHAPE-ID] is replaced with the ShapeNet category label at the finest hierarchical level of ShapeNet. ControlNet's control strength is set to 1 (default). This setup makes ControlNet a strong baseline for shape adherence as it generates images whose object silhouettes closely match those of the input depth image. However, with compositional prompts, we observed that ControlNet tends to prioritize depth image content over textual context, leading to weaker prompt adherence. + +(b) CNet-Stop@K: Here, we use ControlNet for a fixed percentage, $K\%$ of the generation steps, with conditioning on both the depth image and text prompt with category labels as in ControlNet. After this partial conditioning, we capture the latent representation from ControlNet at the $K\%$ -intermediate step and pass it to Stable Diffusion. We then continue denoising for the remaining steps with only the text prompt, excluding the depth conditioning. This approach increasingly preserves the shape and target pose for larger $K$ , while better adhering to the text prompt context for lower $K$ . Given this trade-off, we test for various $K\%$ levels, including $K = 20, 40, 60, 80$ ( $K = 100$ corresponds to the original ControlNet). + +(c) ShapeWords: Our method, ShapeWords, uses text prompts with [SHAPE-ID] tokens as provided in our test datasets. For cases requiring pose control, we apply a similar procedure to CNet-Stop@K: ControlNet is applied for a fixed percentage, $K\%$ , of the generation steps, conditioned on the depth image in the target pose and text prompt with a category label in place of [SHAPE-ID]. At the $K\%$ -intermediate step, we capture the intermediate latent representation and pass it to Stable Diffusion, where denoising continues conditioned on the Shape2CLIP embedding generated for the text prompt based on the specific [SHAPE-ID] token. This preserves the target pose while allowing for shape and prompt adherence provided by our method. We + +call this variant of our method as "ShapeWords@K". + +Numerical evaluation. We first examine results on the "simple prompts dataset". As outlined earlier, this dataset allows us to evaluate shape adherence across all methods using S-CD and S-IOU metrics, which measure geometric similarity between the silhouettes of rendered reference shapes and generated shapes. Since the silhouette comparison is pose-dependent, we compare variants of "ShapeWords@K" with the corresponding "CNet-Stop@K" variants for matching values of $K$ . + +Figure 3 shows plots of S-CD and S-IOU, with the horizontal axis representing different $K$ values (20%, 40%, 60%, 80%), and the y-axis reporting the metrics. We also include results for the original ControlNet (i.e., $K = 100$ ), which serves as a strong baseline for shape adherence. + +Our method, ShapeWords@K, consistently outperforms all corresponding "CNet-Stop@K" variants, demonstrating superior adherence to the target shape across varying $K$ levels. Notably, at $K = 40\%$ , ShapeWords@K achieves shape adherence comparable to the original ControlNet. This suggests that ShapeWords can achieve strong shape adherence without explicit depth conditioning. + +A reasonable question from the above comparison is: why not always use the original ControlNet instead of our method? The answer lies in the results shown in Table 1 for the more complex "compositional prompt dataset." In this challenging setting, our method outperforms ControlNet across all measures - FID, KID, aesthetics, and CLIP scores $^2$ - demonstrating superior prompt adherence and aesthetic plausibility. We note that in this dataset, we do not evaluate silhouette similarity due to the increased foreground complexity and existence of multiple objects. Furthermore, pose control is not enforced: ControlNet generates images using a depth image from a random pose (sampled from 30 viewpoints rotated in 12-degree increments around the vertical axis with fixed elevation, as used in training), while our method generates random poses by design. We also report comparisons with ControlNet-Stop@60 and ControlNet-Stop@80, which, in the previous experiment, demonstrated shape adherence levels similar to our method. However, even with comparable shape adherence, both ControlNet-Stop variants consistently underperform relative to ShapeWords in terms of prompt adherence and overall aesthetic quality. + +Perceptual Evaluation. As a final evaluation, we conducted two user studies on Amazon MTurk to examine prompt and shape adherence from a perceptual standpoint. In the first study, each page of the questionnaire presented + +![](images/2faf8ebf8248108d5e031a4b8951afeb789920edde137328e1d3cb2c4804c775.jpg) +Figure 5. Generalization to compositional prompts. Over-constrained depth conditioned baselines (e.g. ControlNet and Ctrl-X@60) ignore prompt composition. Under-constrained depth baselines (e.g. CNet-Stop@30 or Ctrl-X@30) stray too much from target geometry. Our method provides non-superficial generalization to compositional prompts and strong adherence to target shape. We stress out that depth is only provided as input for baselines and ShapeWords don't use depth input. We provide more examples in the supplement. + +# Q1: Which image better matches the text prompt? + +![](images/d51c7be23ff5df9de2e615c9ef7d4ac6fc04c03e68acfba579cf33b32727cc9d.jpg) + +# Q2: Which object's shape better matches the reference shape? + +![](images/4e0d385620ee0ffbc2ae62d0193eb2bfb00027b4d8af57866fd0ad1c53b7918c.jpg) +Figure 6. User study results. We asked 250 participants to compare how well outputs of competing models follow text prompt and target geometry. Our method was selected as the favorite for both conditions. + +participants with a randomly ordered pair of generated images - one from ShapeWords and one from ControlNet - based on a randomly selected prompt from our compositional test dataset and a target shape from our test split. For this comparison, we used the ControlNet variant (CNet-Stop@K) with the best CLIP score based on our previous experiment. We asked participants: "Which generated image best matches the text in the given prompt?". Participants could pick either image, specify "none" or "both" images matched the text well. In the second study, participants were shown randomly ordered image pairs from ShapeWords and the best ControlNet variant according to CLIP score, along with a rendering of the target test shape + +![](images/1cb21ce96c967d54d14d59b325f57255279ba96f43acea811310a7be8dc9125a.jpg) +Figure 7. Generalization across shapes and styles. ShapeWords generalizes across styles and contexts and correctly incorporates stylistic prompt cues into target geometry, if needed. + +and a randomly selected prompt from our compositional test dataset. Here, we asked, "Which object in the generated images is more similar shape-wise to the reference shape in the rendering?" Participants could again choose either image, or indicate "none" or "both" if both objects matched the reference shape. For both studies, we had 250 participants each compared 10 unique image pairs (totaling 2500 comparisons). To ensure, reliability of answers we repeated each question in a random order. + +Figure 6 shows the percentage of votes for each question type. Our results indicate that ShapeWords' outputs were preferred by a significantly larger proportion of participants for both prompt adherence and shape resemblance, underscoring the perceptual advantages of our approach. + +Qualitative evaluation. Our qualitative evaluation (Figure 5) also suggests that ControlNet's depth conditioning tends to prioritize shape adherence at the expense of prompt adherence and aesthetic plausibility in the generated images especially for complex prompts. We also include a qualitative comparison with the recent method of Ctrl-X [22] in the same figure, where we observed similar, or even worse behavior in terms of shape adherence compared to ControlNet. In contrast, our method strikes a much better balance, achieving a more effective trade-off between maintaining shape fidelity and adhering to the input prompt while producing visually plausible results. We have also observed that our method is able to generalize across various textual contexts and styles as shown in Figure 7. We note that our method accurately combines geometric guidance with stylistic prompt cues, like thin object proportions and overly detailed backgrounds of Hieronymus Bosch's paintings. + +Guidance strength. We demonstrate control of the guidance strength (parameter $\lambda$ ) in Figure 8, where we vary $\lambda$ from 0.0 to 1.0. Increase of guidance strength increases the adherence of the resulting image to the target shape. We note that intermediate values of geometry guidance via ShapeWords do also result in plausible shapes. + +Overall, ShapeWords presents an argument for soft structural guidance for text-to-image synthesis. In contrast to hard structural guidance like depth images, canny edges or renders, it allows users to explore variations of target geometries guided by the learned structural priors. We also illustrate this useful property in Figure 9, where ShapeWords produces concept art of furniture items that bear both diverse and distinct resemblance to target geometries. Note that this guidance is purely geometric: even though target shapes come from 'chair' and 'lamp' categories, our model produces tables, chairs, couches, and bookshelves. + +Additional evaluation. Our supplement includes ablation studies examining various design choices, such as predicting the text prompt embedding directly versus predicting a residual, and updating versus not updating the EOS token. + +# 5. Conclusion + +We have presented ShapeWords, the first to our knowledge method that allows for geometric manipulation of text-to-image models via mapping of 3D shapes into the space of text embeddings (CLIP). Our experiments showed that our + +![](images/f369c457957f05534e58c4f33826089d6d67e62d3236630bc1adedb7a72bf775.jpg) +Figure 8. Ablation on guidance strength $\lambda$ . Here we vary $\lambda$ from 0 to 1 for the prompt "aquarelle drawing of a [SHAPE-ID]" (random seed is fixed, so results for $\lambda = 0$ are the same). Increasing level of guidance strength increases resemblance to target shape. The intermediate results still remain aesthetically plausible. + +![](images/9e1761d27a3741cd6608e91fbdd81a9ab0e75ea6ae1d35305eb898f367ba9ae1.jpg) +Guidance 'Concept art of futuristic furniture' +Figure 9. Soft geometry guidance via ShapeWords. Our method can produce diverse variations of a target prompt that adhere to guidance geometry via variation of $\lambda$ . It can also produce several diverse objects in the image that still resemble the target geometry. + +method shows both strong shape adherence and good generalization to compositional prompts. + +**Limitations.** Our supplement shows failure cases for our method. In summary, it struggles with challenging fine-grained geometry, especially in areas of thin parts. Generalization to much more complex prompts would also require training on much larger scale datasets. Finally, our method exhibits certain appearance biases (e.g. oversaturation of some images, or preference of certain colors), which could be avoided with better score distillation variants. Combining ShapeWords with viewpoint and other 3D attribute tokens would be an interesting future direction [7, 9]. + +Acknowledgements We thank Vikas Thamizharasan, Cecilia Ferrando, Deep Chakraborty and Matheus Gadelha for useful feedback. This project has received funding from: the European Research Council (GA 101124742); EU H2020 Research and Innovation Programme and the Republic of Cyprus through the DMRID (GA 739578). + +# References + +[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia 2023 Conference Papers, pages 1–12, 2023. +[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. +[3] Eyal Betzalel, Coby Penso, Aviv Navon, and Ethan Fetaya. A study on the evaluation of generative models. arXiv preprint arXiv:2206.10935, 2022. +[4] Shariq Farooq Bhat, Niloy Mitra, and Peter Wonka. LooseControl: Lifting Conditions in Conditional Image Generation. In SIGGRAPH, 2024. +[5] Mikołaj Binkowski, Dougal J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations, 2018. +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. +[7] James Burgess, Kuan-Chieh Wang, and Serena Yeung. Viewpoint textual inversion: Unleashing novel view synthesis with pretrained 2d diffusion models. 2024. +[8] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. arxiv1512.03012, 2015. +[9] Ta-Ying Cheng, Matheus Gadelha, Thibault Groueix, Matthew Fisher, Radomir Mech, Andrew Markham, and Niki Trigoni. Learning continuous 3d words for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6753-6762, 2024. +[10] 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 CVPR, 2023. +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. +[12] Yuxuan Ding, Chunna Tian, Haoxuan Ding, and Lingqiao Liu. The clip model is secretly an image-to-prompt converter. Advances in Neural Information Processing Systems, 36, 2024. +[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. + +An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion. In ICLR, 2023. +[14] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. +[15] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A Style-based 3D-Aware Generator for High-resolution Image Synthesis. In ICLR, 2022. +[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, pages 6626-6637, 2017. +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. +[18] Yukun Huang, Jianan Wang, Yukai Shi, Boshi Tang, Xianbiao Qi, and Lei Zhang. DreamTime: An Improved Optimization Strategy for Diffusion-Guided 3D Generation. In ICLR, 2024. +[19] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-Concept Customization of Text-to-Image Diffusion. In CVPR, 2023. +[20] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fr\`echet inception distance. arXiv preprint arXiv:2203.06026, 2022. +[21] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding. In CVPR, 2024. +[22] Kuan Heng Lin, Sicheng Mo, Ben Klingher, Fangzhou Mu, and Bolei Zhou. Ctrl-x: Controlling structure and appearance for text-to-image generation without guidance. In NeurIPS, 2024. +[23] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot One Image to 3D Object. In ICCV, 2023. +[24] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. +[25] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: 360 reconstruction of any object from a single image. In CVPR, 2023. +[26] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. FreeControl: Training-Free Spatial Control of Any Text-to-Image Diffusion Model with Any Condition. In CVPR, 2024. +[27] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models. In AAAI, 2024. +[28] OpenAI. Chatgpt: Optimizing language models for dialogue, 2023. https://chat.openai.com. + +[29] Karran Pandey, Paul Guerrero, Metheus Gadelha, Yannick Hold-Geoffroy, Karan Singh, and Niloy J. Mitra. Diffusion handles: Enabling 3d edits for diffusion models by lifting activations to 3d. In CVPR, 2024. +[30] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11410-11420, 2022. +[31] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. In ICLR, 2023. +[32] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In CVPR, 2017. +[33] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, Stefano Ermon, Yun Fu, and Ran Xu. Uncontrol: A unified diffusion model for controllable visual generation in the wild. In NeurIPS, 2023. +[34] 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. arXiv preprint arXiv:2103.00020, 2021. +[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. +[36] 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. +[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. In arXiv preprint arXiv:2204.06125, page 3, 2022. +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. +[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. HyperDreamBooth: HyperNetworks for Fast Personalization of Text-to-Image Models. In CVPR, 2024. +[40] Mehdi Safaee, Aryan Mikaeili, Or Patashnik, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Clic: Concept learning in context. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6924–6933, 2024. +[41] 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. +[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo + +Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: an open large-scale dataset for training next generation image-text models. In NeurIPS, 2024. +[43] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. InstantBooth: Personalized Text-to-Image Generation without Test-Time Finetuning. In CVPR, 2024. +[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. +[45] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In NeurIPS, 2017. +[46] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3D-aware Image Generation using 2D Diffusion Models. In ICCV, 2023. +[47] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning unified representation of language, image and point cloud for 3d understanding. In CVPR, 2023. +[48] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models. arxiv2308.06721, 2023. +[49] 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 CVPR, 2022. +[50] Biao Zhang, Matthias Nießner, and Peter Wonka. 3DILG: Irregular Latent Grids for 3D Generative Modeling. In NeurIPS, 2022. +[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. +[52] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024. \ No newline at end of file diff --git a/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/images.zip b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a322f755f35750fb26865f29e730b79c34387f64 --- /dev/null +++ b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98f39969d6a2876e4cfba5d099427003aaf2456d1492dfe4dc6b06a8c6656a06 +size 641032 diff --git a/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/layout.json b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..338c803fe43d75124084a5c8f70ea8b7a96dd5a8 --- /dev/null +++ b/CVPR/2025/ShapeWords_ Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:924b6c4baa51f33aefc608ee456d591853337a24b93d38a85131c1a5c002614b +size 335992 diff --git a/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/64527bab-6930-4ef4-8416-dc96246edd0b_content_list.json b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/64527bab-6930-4ef4-8416-dc96246edd0b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f2b994e3a83ee51551c2bf1eebad493b085c8cf9 --- /dev/null +++ b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/64527bab-6930-4ef4-8416-dc96246edd0b_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e0ed54b5d25519bfd6c6b1256cf2802bebb0f9ca9d0f553e07b23a444ce68e3 +size 73132 diff --git a/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/64527bab-6930-4ef4-8416-dc96246edd0b_model.json b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/64527bab-6930-4ef4-8416-dc96246edd0b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..915f8291d34e247dbe9ad96275d5d3d00e0170ea --- /dev/null +++ b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/64527bab-6930-4ef4-8416-dc96246edd0b_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d809ee7dda23d2edcf5195d0366d828f03e0e223c305a68d6aba0f6dc70454d6 +size 97856 diff --git a/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/64527bab-6930-4ef4-8416-dc96246edd0b_origin.pdf b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/64527bab-6930-4ef4-8416-dc96246edd0b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..46554788da97b8fb797386b5611695e8ebd36127 --- /dev/null +++ b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/64527bab-6930-4ef4-8416-dc96246edd0b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c04bb83c849423c180f7f898e0ad4dbfb946ad40bc24f8ddf7146fc2dc0f207 +size 7334346 diff --git a/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/full.md b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fd26fc85eb048a53b53eaf308966356701068d88 --- /dev/null +++ b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/full.md @@ -0,0 +1,292 @@ +# Sharp-It: A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation + +Yiftach Edelstein1 Or Patashnik2 Dana Cohen-Bar2 Lihi Zelnik-Manor1 + +$^{1}$ Technion - Israel Institute of Technology $^{2}$ Tel Aviv University + +https://yiftachede.github.io/Sharp-It/ + +![](images/f7fc9e32b06ac385c9508067ededa7c19db1d402ac64b066d23d1add651bb7fc.jpg) + +![](images/196121277fab881dd8f2d803453636f0867cdd00bd5368a4995f4a8109751cd5.jpg) +Figure 1. Sharp-It is a multi-view to multi-view model that enhances low-quality 3D shapes. It corrects fine-grained geometry details and adds appearance features. The top row displays three degraded shapes and their enhancements by Sharp-It. The bottom row demonstrates Sharp-It's capability to edit the appearance of 3D shapes. + +# Abstract + +Advancements in text-to-image diffusion models have led to significant progress in fast 3D content creation. One common approach is to generate a set of multi-view images of an object, and then reconstruct it into a 3D model. However, this approach bypasses the use of a native 3D representation of the object and is hence prone to geometric artifacts and limited in controllability and manipulation capabilities. An alternative approach involves native 3D generative models that directly produce 3D representations. These models, however, are typically limited in their resolution, resulting in lower quality 3D objects. In this work, we bridge the quality gap between methods that directly generate 3D representations and ones that reconstruct 3D objects from multi-view images. We introduce a multi-view to multi-view diffusion model called Sharp-It, which takes a 3D consistent set of multi-view images rendered from a low-quality object and enriches its geometric details and texture. The diffu + +sion model operates on the multi-view set in parallel, in the sense that it shares features across the generated views. A high-quality 3D model can then be reconstructed from the enriched multi-view set. By leveraging the advantages of both 2D and 3D approaches, our method offers an efficient and controllable method for high-quality 3D content creation. We demonstrate that Sharp-It enables various 3D applications, such as fast synthesis, editing, and controlled generation, while attaining high-quality assets. + +# 1. Introduction + +Creating 3D content plays an important role across industries such as gaming, augmented and virtual reality, and animation. These applications demand efficient, controllable processes for generating high-quality, editable 3D assets. In recent years, text-to-image diffusion models have significantly advanced 3D content creation, introducing new approaches for producing complex, detailed visual assets. + +Early approaches to 3D content generation optimized 3D representations by leveraging 2D diffusion models [54, 67], producing high-quality assets but requiring time-consuming per-asset optimization. Recent techniques have adopted a two-stage process: first, synthesizing consistent multi-view images using 2D diffusion models [36-38, 61, 62, 68], and second, performing 3D reconstruction from these multiview images using feed-forward methods [32, 71]. This approach accelerates 3D synthesis while maintaining high-quality output. However, by bypassing native 3D representations, it limits controllability and editability of the resulting assets and tends to produce visual artifacts such as the Janus problem and flat objects. Direct 3D generative models [28, 48] offer greater control over creation and editing but are constrained by resolution limitations, resulting in lower-quality assets. This presents a trade-off between the quality of multi-view methods and the controllability of direct 3D generative models. + +In this work, we address the quality gap between direct 3D generative models and those that reconstruct 3D objects from multi-view images. To bridge this gap, we introduce Sharp-It, a multi-view to multi-view diffusion model that enhances objects generated by 3D generative models, refining geometric details and adding fine-grained texture features. The high-quality multi-view set produced by Sharp-It can be reconstructed into a 3D model using existing sparse-view feed-forward reconstruction methods [22, 27, 71, 80]. + +Sharp-It is conditioned on a text prompt and operates on the multi-view set in parallel, sharing features across the generated views [61, 62]. We use Shap-E [28] as our backbone generative model, a text-conditioned 3D latent diffusion model that operates on the parameters of implicit functions. Shap-E has been shown to offer rich generative capabilities [10, 60], but produces low-quality assets. To train Sharp-It, we use pairs of multi-view sets obtained by encoding high-quality 3D objects [14] into Shap-E's latent space and rendering them. + +Our approach offers two key advantages. First, since we start from a coarsely valid 3D object, we inherit its plausible geometric structure, avoiding artifacts like the Janus problem that can occur in direct multi-view synthesis methods. Second, the 3D consistency of the input views facilitates output consistency, allowing the model to focus on generating the fine details necessary for high-quality results. + +Our method combines the controllability of native 3D generative models with the capabilities of 2D diffusion models in generating highly detailed images. Through extensive experiments, we demonstrate that Sharp-It outperforms existing enhancement methods in both quality and efficiency. We show various applications, including fast generation of high-quality 3D objects from text prompts and geometric controls [60]. Additionally, we show that Sharp-It enhances 3D asset editing capabilities. + +# 2. Related Work + +# 2.1. Diffusion Models for Content Generation + +Diffusion models [21, 56, 57] have demonstrated remarkable success in modeling complex data distributions and generating high-fidelity samples. In the realm of 2D image synthesis, these models have achieved state-of-the-art results, producing photorealistic images with precise control. Motivated by this success, recent works have explored the application of diffusion models to 3D content creation. + +3D Diffusion Models Some works utilize 3D datasets to train diffusion models for direct 3D asset generation, with different approaches varying in their choice of 3D representations. Some methods employ explicit representations [73, 76] such as point clouds [44, 48, 58, 75, 79]. Others transform the input data into learned implicit representations such as triplanes [7, 17, 63] or neural networks [26, 33] before learning their distributions. A notable example is Shap-E [28], which adopts NeRF (Neural Radiance Field) and SDF (Signed Distance Field) as implicit representations and trains a diffusion model to generate the weights of these networks. While these 3D-native diffusion models enable fast generation and provide greater control over the generation process [60], they often face challenges due to limited training data and the high computational cost of operating directly on 3D representations. + +2D Diffusion Models for 3D Generation Another approach leverages the powerful prior of pretrained text-to-image diffusion models [56] for 3D generation. Some methods [11, 29, 35, 46, 54, 65, 67, 69, 74] utilize these models to guide the optimization of 3D representations, such as NeRF or Gaussian Splatting, typically through Score Distillation Sampling (SDS) [54] or its variants. To enhance and accelerate SDS-guided optimization, recent methods fine-tune 2D diffusion models on 3D datasets [13]. Zero-1-to-3 [38] fine-tunes a pre-trained diffusion model by incorporating input views and camera parameters to enable novel-view synthesis. Building upon this fine-tuning strategy, some followup works [39, 61, 62, 64] enable simultaneous generation of multiple views using shared attention mechanisms. While these 3D-aware fine-tuning approaches strengthen the prior and improve both convergence speed and result quality when used in SDS-based optimization, the per-shape optimization process still requires substantial computational time. + +To overcome the high computational costs of per-shape optimization, recent works propose two-stage approaches that first generate multi-view images and then apply fast mesh reconstruction techniques to create 3D objects from these images. One-2-3-45 [37] and One-2-3-45++ [36] employ feed-forward networks to perform fast, high-quality asset creation. Wonder3D [41] introduces a cross-domain attention mechanism that generates both normal map images + +and color images, which facilitates fast mesh reconstruction through novel geometry-aware normal fusion algorithms. 3D-Adapter [8] constructs a consistent 3D representation throughout the denoising process. InstantMesh [71] is a sparse-view 3D reconstruction method that utilizes a feedforward network based on the LRM [23] architecture. Recent advancements in feed-forward sparse-view 3D reconstruction further improve reconstruction quality with different architectures and training paradigms [27, 80]. While these methods successfully generate high-quality objects rapidly, the reliance on multi-view images limits controllability and editing capabilities. + +# 2.2. Controlled Generation and Editing + +Extensive work has focused on developing methods for controlled generation and editing of images and video by employing 2D diffusion models [1, 2, 4, 12, 15, 16, 25, 34, 47, 70, 78]. Many of these approaches leverage the semantic features learned by the models to gain control and perform edits [5, 19, 50, 51, 66, 78]. Some methods have extended the control and editing capabilities of 2D diffusion models to 3D representations [3, 9, 18, 30, 33, 52]. However, these typically require computationally expensive optimization processes per object. Faster control and editing of 3D shapes can be achieved through the latent spaces of native 3D generative models [10, 24, 40, 59, 60]. Yet, the quality of shapes generated by these methods is often limited by the output quality of the underlying 3D models, which can be constrained by resolution limitations. In this work, we propose a novel approach that combines 3D generative models with a multi-view enhancement method to obtain high-quality 3D controlled generation and editing pipelines. + +# 3. Method + +In this section, we present our approach to addressing the quality gap between direct 3D generative models and those that reconstruct 3D objects from multi-view images. We focus on Shap-E [28], a 3D generative model, and introduce Sharp-It, a multi-view-to-multi-view diffusion model that enhances 3D objects generated by Shap-E. We train Sharp-It to improve these objects by adding intricate appearance details and correcting geometric artifacts such as discontinuities and broken parts. The process of generating a 3D model with Sharp-It is demonstrated in Figure 2. + +# 3.1. Preliminaries + +Shap-E is a latent diffusion model specifically designed for generating 3D assets. As common in latent diffusion models [56], Shap-E is trained in two stages. In the first stage, an encoder is trained to map 3D objects into a latent space. This latent space corresponds to the weight space of implicit functions that represent 3D shapes, with each shape represented as an element in $\mathbb{R}^{1024\times 1024}$ . The latent representation can be decoded using Signed Texture Field (STF) + +rendering, where it is treated as the weights of an implicit function. In the second stage, a diffusion model is trained within the Shap-E latent space, allowing for conditioning on either text or images. + +Zero123++ is an image-conditioned diffusion model designed to generate 3D-consistent multi-view images from a single input view [61]. It builds upon Stable Diffusion [56], which is a latent diffusion model comprising a VAE and a UNet. The VAE encodes images into a resolution that is eight times smaller and consists of four channels, while the UNet serves as the diffusion model, operating on the four-channel latent codes. Zero123++ is a fine-tuned version of Stable Diffusion that accepts an image as input. Given an input image, it produces a $3 \times 2$ grid of $320 \times 320$ pixel images, with six constant azimuth and elevation angles. Originally, Zero123++ [61] was designed to generate this grid image with a grey background. In InstantMesh [71], it was further fine-tuned to render a white background, addressing the issue of "floaties"—particles floating through space during 2D-to-3D lifting. + +# 3.2. Dataset Construction + +We begin by constructing a paired dataset consisting of degraded Shap-E objects along with corresponding high-quality objects. Our key idea in constructing the dataset is that such pairs can be obtained by employing the encoder of Shap-E in conjunction with a high-quality 3D objects dataset. Specifically, we utilize objects from Objaverse [13], provided by [42, 43], and encode them using Shap-E's encoder. For each pair of an object from Objaverse and its encoded Shap-E latent code, we render a $3 \times 2$ grid from six predefined camera views, applying three HDR lighting conditions. Our experiments, detailed in the ablation studies, indicate that rendering each object under varying HDR lighting enhances the model's performance. + +We apply a filtering criteria on the resulting dataset. We remove objects where the degraded Shap-E rendering is significantly different from the original object's rendering, indicating a failure of Shap-E's encoder. Additionally, we filter out objects that are too thin or do not match certain keywords based on their annotated captions. For each remaining object, we extract a caption using BLIP2 [31]. Finally, we split the dataset into training and test sets, resulting in 180,000 objects for training and 6,000 for testing. + +# 3.3. Sharp-It + +Sharp-It is a multi-view to multi-view diffusion model designed to enhance low-quality multi-view images of 3D objects generated by Shap-E. It takes as input a set of multiview images rendered from a low-quality 3D object, along with a textual prompt, and produces high-quality multiview images with refined geometric details and textures that correspond to the input views. An overview of a 3D generation pipeline with Sharp-It is shown in Figure 2. + +![](images/e105c98bdfd4cea162948a0fe57f30ae8acbb1295a8aacf7a0c8a7a9524bdc98.jpg) +Figure 2. Overview of 3D generation pipeline with Sharp-It. First, a 3D object is generated with Shap-E. Then, we render six views of this low-quality object. Sharp-It is a diffusion model based on Stable Diffusion [56] that enhances these views with the guidance of a text prompt by refining geometry and adding detailed appearance. Sharp-It employs cross-attention layers for text-based guidance and self-attention layers for cross-view consistency. A high-quality 3D object can be reconstructed from the multi-view image set. + +Architecture The architecture of Sharp-It has two key requirements: generating multi-view image sets and incorporating input multi-view sets as conditions. We build on Zero123++ [61, 71], which was fine-tuned for multi-view generation, where images are arranged in a $3 \times 2$ grid with a total resolution of $960 \times 640$ and fixed camera angles across objects. To enable multi-view conditioning, we modify the architecture of Zero123++ by expanding the UNet input to 8 channels: 4 for latent noise and 4 for the VAE-encoded Shap-E multi-view images. This design lets our model leverage the coarse geometry from the input views to achieve better 3D consistency, and is inspired by image editing techniques that fine-tune diffusion models to accept an image as input and modify specific parts while preserving others [4, 56, 72]. Unlike these approaches, which operate on a single image, our model learns to enhance the input views in a 3D consistent manner. Furthermore, we replace Zero123++'s image embedding with text prompts in the cross-attention layers, enabling better enhancement control and appearance editing capabilities (Section 4.3). + +The model we build on consists of self-attention layers, which play an important role in facilitating the consistency of our produced multi-view set [62, 68]. Since our model operates on a multi-view image grid, these layers can be seen as an application of cross-view attention between the different views. Similarly to previous works [61, 62, 68], this allows our model to simultaneously refine corresponding points across different views by learning the correspondences between them. We visualize the learned correspondences in Figure 3. The figure displays self-attention maps for a query point marked by a red dot in the leftmost image. The results demonstrate that this point on the wheel receives the highest attention weight across different views. Additionally, the attention mechanism identifies semantically similar points – notably, other wheels of the car. + +Training To train Sharp-It, we utilize our paired dataset consisting of $x$ , a highly-detailed multi-view image set; $x_{\mathrm{Shap - E}}$ , a degraded version of $x$ , representing the low-quality multi-view image generated by Shap-E; and a text prompt $c_{\mathrm{prompt}}$ , describing the high-quality 3D object. We + +![](images/6639435f05adf9b9917caa15cf7f999b3ec1bdacd809485df1e6ad516e89c27a.jpg) +Figure 3. Self-attention maps for a query point (red) on the car's wheel, showing highest attention weights at corresponding wheel locations across different views. + +initialize Sharp-It with the weights of $\text{Zero123++}$ taken from the version trained by Xu et al. [32], and fine-tune it using standard diffusion training with v-prediction. The training loss is formulated as follows: + +$$ +\mathcal {L} = \mathbb {E} _ {t, \epsilon \sim \mathcal {N} (0, 1)} \left[ \| v - v _ {\theta} \left(x _ {t}, x _ {\text {S h a p - E}}, c _ {\text {p r o m p t}}\right) \| ^ {2} \right], +$$ + +where $v_{\theta}$ denotes the v-prediction of the model, parameterized by $\theta$ . $x_{t}$ is obtained by adding noise to $x$ with respect to the diffusion timestep $t$ . $t$ and $\epsilon$ are randomly sampled diffusion step and Gaussian noise, respectively. $v$ is defined as $\alpha_{t}\epsilon - \sigma x$ , where $\alpha_{t}, \sigma$ are parameters of the noise scheduler. We train our network for 500,000 steps, with a CFG drop probability of 0.1, and a batch size of 3 multi-views, using a single NVIDIA A6000 GPU. + +Inference Given a 3D object produced by Shap-E, we first render it from our six predefined azimuths and elevations to form $x_{\mathrm{Shap - E}}$ . To enhance this multi-view image set, we sample a noisy image $x_{T} \sim \mathcal{N}(0,1)$ and concatenate it channel-wise to $x_{\mathrm{Shap - E}}$ to create an 8-channel input to the model. The prompt used at inference time can be either the prompt used to generate the object with Shap-E or any other prompt that fits to the object, where the latter option is used to edit the appearance of the object. We use our trained Sharp-It to iteratively denoise $x_{t}$ , moving from the complete noisy multi-view image set, $x_{T}$ at $t = T$ to a clean, high-quality multi-view set $\hat{x}$ at $t = 0$ . The final output multi-view set $\hat{x}$ balances the information from the prompt $c_{\mathrm{prompt}}$ and the degraded multi-view input $x_{\mathrm{Shap - E}}$ . + +Our enhanced multi-view set can be reconstructed into a 3D object using any existing feed-forward sparse reconstruction method [27, 71, 77, 80]. In our experiments, we use InstantMesh [71] for reconstruction and provide its results in the supplementary materials. + +![](images/b2cf4a3ff8ca73cc16fe522b9726deb3f61bfacd2974d51145fe9e0adf7ccf9f.jpg) +Figure 4. Comparison of Sharp-It with other methods for 3D object enhancement (GaussianDreamer and MVEdit), and multi-view enhancement (SDEdit based). The first column shows the input object generated by Shap-E. As can be seen, our method achieves the highest quality results while best preserving the input object. + +Table 1. Quantitative comparison of our enhancement method with other baseline methods. + +
MethodFID ↓CLIP ↑DINO ↑Runtime
GaussianDreamer50.890.810.826 min
MVEdit44.870.830.771 min
MVDream w/ SDEdit28.710.810.8310 sec
Zero123++ w/ SDEdit (U)24.590.870.8610 sec
Zero123++ w/ SDEdit (C)19.950.850.8710 sec
Zero123++ w/ SDEdit (R)19.130.870.8910 sec
Sharp-It6.600.900.9210 sec
+ +# 4. Experiments + +In this section, we evaluate the performance of Sharp-It in enhancing 3D objects generated by Shap-E. We compare our method against several baselines, conduct ablation studies to analyze the impact of different components, and demonstrate applications in 3D generation and editing. + +# 4.1. Qualitative and Quantitative Comparison + +Baselines We compare Sharp-It with various enhancement baselines. First, we compare against an SDS-based approach that was used in Spice-E [60]. Specifically, we + +use GaussianDreamer [74], initializing Gaussian Splatting with Shap-E's shape output and optimizing it with a text-to-image model. To make refining the thousands of objects in our test set practical, we limit the optimization time to six minutes per object. We also include MVEdit [9], a multiview editing method that refines coarse 3D shapes. + +Another set of baselines combines multi-view image generation with SDEdit [45]. Here, we render Shap-E's shape into a multi-view image set and apply simultaneous edits across views to ensure consistency. Using SDEdit with a strength of 0.4, we add noise to each multi-view image and denoise it. For multi-view generation, we use both MVDream [62], which is guided by a text prompt, and Zero123++ [61], which is conditioned on an image rather than text. With Zero123++, we test three configurations: (i) Zero123++ w/ SDEdit (R), where we apply SDXL Refiner model (default strength value) [53] on the frontal view of the object, and use this edited view to guide SDEdit across the multi-view set, (ii) Zero123++ w/ SDEdit (C), where SDEdit with Stable Diffusion (strength 0.75) is applied on the object's frontal view which is used to guide the multi-view SDEdit; and (iii) Zero123++ w/ SDEdit (U), where no image condition is used. + +Dataset We conduct our quantitative experiments on our test set described in Section 3.2. To demonstrate our method's generalization capability, we also present results on objects generated directly by Shap-E from text prompts. + +Metrics To assess enhancement performance, we evaluate the image quality produced by each method. Specifically, we compute the FID [20, 49] between the enhanced views and our test set. Additionally, we evaluate the semantic and visual similarity of the enhanced shape with the ground-truth using CLIP [55] and DINO [6] image encoders. + +Results We present a qualitative comparison in Figure 4. The avocado couch and the tiki mask were generated by Shap-E, while the bust is an encoded object from the test set. As shown, our method preserves the objects' coarse details while producing high-quality, detailed enhancements. In contrast, other methods diverge significantly from the original objects and yield less realistic results. For instance, our method accurately generates a leather-like texture, whereas other methods struggle to enhance the flat appearance of the Shap-E-generated couch. Our approach also maintains color consistency with the input shapes. This is evident in the couch's back section, where brown and green tints are faithfully preserved, as well as in the bust and tiki mask. Our method generates significantly more fine details compared to other methods, particularly visible in the intricate features of the bust and the textural elements of the mask. Results for the objects lifted to 3D are provided in the supplementary materials. + +Quantitative comparisons are shown in Table 1. Our method achieves the lowest FID, confirming it produces the highest-quality results. Furthermore, our method shows the best alignment with the ground-truth object, as indicated by the CLIP and DINO similarity metrics. The runtime of our method is comparable to or faster than other methods. + +# 4.2. Ablation Studies + +We conduct ablation studies to assess the contributions of different components in Sharp-It. First, we examine the importance of using an input text prompt, which is not utilized in Zero123++ [61, 71], the model upon which we build. Second, we explore the effect of constructing the dataset with diverse lighting conditions, rather than using a single type of lighting. + +For this analysis, we train our method without each component individually and present the results in Figure 5. For computational efficiency, ablation models (including the full method one) were trained for 400,000 steps, fewer than in our main experiments. In the first column, we show a shape generated by Shap-E, which serves as input to the models in the subsequent rows. In the second column, we omit the text prompt, instead using an empty prompt. As + +![](images/dd5c4c0720d2002eef5c6c41655ef5916f4cfd4d61622c2f6c27f68bef696c8e.jpg) +Figure 5. Qualitative ablation study. The first column shows the degraded input object generated by Shap-E. Subsequent columns show the effects of removing specific components: omitting the text prompt leads to reduced texture detail, while excluding diverse lighting results in a flatter appearance with less realistic shading. The full model achieves the most refined and detailed result. + +Table 2. Quantitative results of our ablation study. All models were trained for 400,000 steps. + +
AblationFID ↓CLIP ↑DINO ↑
W/o text prompt9.910.910.91
W/o diverse lighting9.160.890.91
Full method8.060.890.91
+ +shown, without a guiding text prompt, the results lack detail and fail to achieve a metallic appearance. In the third column, we construct the dataset with a single lighting condition, leading to a notable enhancement of the input shape but still lacking finer details. The last column displays results from our full model, which successfully achieves a metallic look and produces highly detailed shapes. Importantly, our results are consistent with the coarse details of the input shape. The lifted object of this example is included in the supplementary materials. We also quantitatively evaluate each configuration using the metrics described in the previous section. Results are provided in Table 2. While all configurations are comparable aligned with the ground truth object, the full model achieves the highest image quality. + +![](images/09a95b83ff8279f619e04554ffd5fd6ac47b949c0c336dc0d39d598ebd23b789.jpg) +Figure 6. Text-to-MultiView qualitative results. Our method generates a multi-view set by creating a 3D object with Shap-E and refining it with Sharp-It. In contrast, Zero123++ first generates a single image with Stable Diffusion, then produces a multi-view set conditioned on that image. This approach leads to geometric artifacts, such as flatness (gundam) and the Janus problem (tank). Leveraging Shap-E's 3D knowledge, our method yields high-quality, coherent objects. Additional 3D renderings of these results are provided in the supplemental. + +# 4.3. Applications + +We now demonstrate the various applications enabled by Sharp-It. Specifically, we show how Sharp-It bridges the quality gap between Shap-E and multi-view-based 3D generation methods, while supporting diverse 3D generative applications. + +Text-to-3D Generation As discussed in previous sections, a common approach for text-to-3D synthesis involves first generating a nearly consistent multi-view image set [61, 62, 68], followed by a sparse-view 3D reconstruction method [71]. By combining Shap-E with Sharp-It, we achieve text-to-multi-view image synthesis, enabling a complete pipeline for 3D generation (Figure 2). In Figure 6, we compare text-to-multi-view results from Zero123++ and our method. Our method achieves results comparable to Zero123++ in terms of image quality, while demonstrating superior geometric details. The advantage of using Shap-E as a first stage is particularly evident in the tank and Gun-dam examples. The tank produced by Zero123++ suffers from a Janus problem, where wheels incorrectly appear on two adjacent sides. Moreover, Zero123++ generates a flat Gun-dam figure, lacking geometric details. These results demonstrate the effectiveness of our pipeline: using a 3D-aware model to generate a coarse 3D object, followed by refinement using an image diffusion model. + +3D Object Editing Shap-E provides a latent space in which semantic manipulation of 3D shapes can be performed while maintaining 3D consistency. Our method enhances these edited shapes to achieve high-quality objects. + +![](images/3e8577000593c4f389b4116cd03f1c0a229956f5e82a7d64cc7c38f33c559b12.jpg) +Figure 7. Our method enhances edits performed in Shap-E's space (rightmost column). We show editing results obtained with an existing editing method (Shap-Editor), and demonstrate that DDPM Inversion, originally developed for image diffusion models, works with Shap-E and can be integrated with Sharp-It. + +We demonstrate our method's application with two different editing techniques. First, we use Shap-Editor [10], which trains a model that takes a Shap-E latent code and a text instruction to produce a new latent code corresponding to the edited shape. While this method enables diverse and fast editing of 3D shapes, it is bounded by Shap-E's quality limitations. Sharp-It addresses this limitation by enhanc + +"A red velvet chesterfield chair" + +![](images/21743b939ef348832bfd6a1e8e3cd41916b8b17780556bb8aa5f528bc2b3d3fc.jpg) +"A cargo spaceship" + +![](images/fb659bca2ef90b9cc6ce43cacb056b81c6fff78f3127455cce9219d0c2f56fe0.jpg) +Figure 9. Sharp-It allows to edit the appearance of a Shap-E object by changing the prompt during the enhancement step. + +![](images/f88e2f93ecefccac2a0d86a7cb93b1738d89411253327ad72ab17b93d7f40834.jpg) +Abstract 3D Guidance +Spice-E +w/SDS refinement +w/ Sharp-It refinement +Figure 8. Sharp-It refines outputs generated by Spice-E, in a faster and more visually appealing way than an SDS-based refinement. + +ing the results produced by Shap-Editor. Second, inspired by diffusion-based editing methods for images, we apply a technique that enables shape editing with Shap-E. Specifically, we apply Edit-Friendly DDPM Inversion [25] with Shap-E instead of an image diffusion model. + +Results for both methods are shown in Figure 7. By combining Sharp-It with existing editing methods, we enable various types of high-quality 3D edits. These include changing an object's shape (demonstrated by transforming a car into an SUV), adding decorations (shown with the Christmas lamp), and modifying colors (as seen with the golden lamp and turquoise car). + +Appearance Editing Sharp-It enables detailed control over the appearance of a degraded shape using text prompts. Specifically, at inference time, we can apply a different prompt from the one used during the generation of the Shape object, facilitating appearance editing that accurately preserves the original shape. Examples of this application are presented in Figures 1 and 9. + +Controlled Generation Previous works have trained models that provide coarse geometric control over objects generated by Shap-E. Similar to Shap-E-based editing methods, such approaches are bounded by Shap-E's quality limitations. We demonstrate high-quality controlled 3D generation by applying our method to shapes generated with Spice-E [60]. We show these results in Figure 8. Our method significantly improves both object textures and fine geometric details. Notably, while Spice-E [60] uses an + +"Golden Jewelry Box" $\longrightarrow$ "Suede Leather Jewelry Box" + +![](images/8b98649a8188beb02017af7eceb839f659fa2a7b298a758b371713667eacd48b.jpg) + +![](images/40cba1e42b0a858ccf9d30f09e7d7d50bcfd6be9dc7659108f7ff57ca617de18.jpg) +"Glass City Tower" $\longrightarrow$ "Wooden City Tower" +"Leather Chair" $\longrightarrow$ "Leopard Print Leather Chair" + +![](images/7278088748406538c03baecae1beeb7b2b648f4394f7b8bd94677e0fcd06f89d.jpg) +Input +Appearance Edit + +SDS-based approach as a refinement step, which impacts its runtime, our method achieves comparable or better results while running an order of magnitude faster. + +# 5. Discussion and Conclusion + +We introduced Sharp-It, a multi-view to multi-view diffusion model that significantly enhances synthetic low-quality 3D objects, thereby improving 3D synthesis and editing pipelines. Our approach has two main advantages over existing multi-view generation methods. First, it relies on a coarse 3D shape, and therefore does not suffer from flat generated geometries or artifacts such as the Janus problem. Second, it can be combined with methods that operate in the latent space of the 3D generative model to provide controllability and editing capabilities. + +In this work, we used Shap-E [28] as our 3D generative model backbone, as it provides editing [10] and control [60] capabilities. However, our approach is applicable to other 3D generative models that encode shapes into lower-quality spaces. One limitation of our method is its reliance on sparse-view reconstruction techniques to generate 3D objects. The reconstruction method we use does not model lighting, instead baking it into the texture. Recent advancements in sparse-view 3D reconstruction [27, 80] could potentially improve our 3D synthesis results. + +We see Sharp-It as an enabler to leverage the advantages of 3D generative models, without compromising on the quality of the results. Through extensive experiments we demonstrated how Sharp-It can be integrated with various methods to achieve rich set of 3D generative capabilities. We hope that our work will pave the way for more work with native 3D generative models, leveraging their powerful generative capabilities. + +# Acknowledgements + +We thank Matan Kleiner, Vladimir Kulikov, Ofry Livney, Rinon Gal and Yuval Alaluf for proofreading our manuscript and for their useful suggestions. We thank Etai Sella for his assistance in exemplifying Sharp-It's use in refining Spice-E outputs. + +# References + +[1] 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 (CVPR), pages 18208-18218, 2022. +[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In CVPR, 2023. +[3] Shariq Farooq Bhat, Niloy J. Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning, 2023. +[4] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. +[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560-22570, 2023. +[6] 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 IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. +[7] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nef: A unified approach to 3d generation and reconstruction, 2023. +[8] Hansheng Chen, Bokui Shen, Yulin Liu, Ruoxi Shi, Linqi Zhou, Connor Z. Lin, Jiayuan Gu, Hao Su, Gordon Wetzstein, and Leonidas Guibas. 3d-adapter: Geometry-consistent multi-view diffusion for high-quality 3d generation, 2024. +[9] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing, 2024. +[10] Minghao Chen, Junyu Xie, Iro Laina, and Andrea Vedaldi. Shap-editor: Instruction-guided latent 3d editing in seconds. arXiv preprint arXiv:2312.09246, 2023. +[11] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246-22256, 2023. +[12] Nathaniel Cohen, Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Slicedit: Zero + +shot video editing with text-to-image diffusion models using spatio-temporal slices, 2024. +[13] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects, 2023. +[14] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsanit, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. +[15] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024. +[16] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising, 2024. +[17] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation, 2023. +[18] Ayaan Haque, Matthew Tancik, Alexei A. Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions, 2023. +[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. 2022. +[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2017. +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. +[22] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. +[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d, 2024. +[24] Jingyu Hu, Ka-Hei Hui, Zhengzhe Liu, Ruihui Li, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation, inversion, and manipulation, 2023. +[25] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly DDPM noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. +[26] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation, 2022. +[27] Haian Jin, Hanwen Jiang, Hao Tan, Kai Zhang, Sai Bi, Tianyuan Zhang, Fujun Luan, Noah Snavely, and Zexiang Xu. Lvsm: A large view synthesis model with minimal 3d inductive bias, 2024. + +[28] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions, 2023. +[29] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation, 2024. +[30] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling, 2024. +[31] 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 ICML, 2023. +[32] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. https://arxiv.org/abs/2311.06214, 2023. +[33] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion, 2023. +[34] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. 2023. +[35] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation, 2023. +[36] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885, 2023. +[37] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. +[38] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. +[39] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image, 2024. +[40] Zhen Liu, Yao Feng, Michael J. Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling, 2023. +[41] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion, 2023. +[42] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. +[43] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. arXiv preprint arXiv:2404.07984, 2024. +[44] Zhaoyang Lyu, Jinyi Wang, Yuwei An, Ya Zhang, Dahua Lin, and Bo Dai. Controllable mesh generation through sparse latent point diffusion models, 2023. + +[45] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. +[46] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. +[47] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. +[48] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. +[49] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. +[50] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. +[51] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. +[52] Or Patashnik, Rinon Gal, Daniel Cohen-Or, Jun-Yan Zhu, and Fernando De la Torre. Consolidating attention features for multi-view image editing, 2024. +[53] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. +[54] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. +[55] 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, 2021. +[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. +[57] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. +[58] Philipp Schröppel, Christopher Wewer, Jan Eric Lenssen, Eddy Ilg, and Thomas Brox. Neural point cloud diffusion for disentangled 3d shape and appearance generation, 2024. + +[59] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. arXiv preprint arXiv:2303.12048, 2023. +[60] Etai Sella, Gal Fiebelman, Noam Atia, and Hadar Averbuch-Elor. Spice-e: Structural priors in 3d diffusion using cross-entity attention, 2024. +[61] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. +[62] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3d generation. In The Twelfth International Conference on Learning Representations, 2024. +[63] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion, 2022. +[64] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3d generative models from 2d data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8863-8873, 2023. +[65] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation, 2024. +[66] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, 2023. +[67] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation, 2022. +[68] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. +[69] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. +[70] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing. ECCV, 2024. +[71] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. +[72] 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. arXiv preprint arXiv:2211.13227, 2022. +[73] Lior Yariv, Omri Puny, Natalia Neverova, Oran Gafni, and Yaron Lipman. Mosaic-sdf for 3d generative models, 2024. +[74] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models, 2024. + +[75] xiaohui zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems, pages 10021-10039. Curran Associates, Inc., 2022. +[76] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: A structured and explicit radiance representation for 3d generative modeling, 2024. +[77] Chubin Zhang, Hongliang Song, Yi Wei, Yu Chen, Jiwen Lu, and Yansong Tang. Geolrm: Geometry-aware large reconstruction model for high-quality 3d gaussian generation. arXiv preprint arXiv:2406.15333, 2024. +[78] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. +[79] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion, 2021. +[80] Peiye Zhuang, Songfang Han, Chaoyang Wang, Aliaksandr Siarohin, Jiaxu Zou, Michael Vasilkovsky, Vladislav Shakhrai, Sergey Korolev, Sergey Tulyakov, and Hsin-Ying Lee. Gtr: Improving large 3d reconstruction models through geometry and texture refinement, 2024. \ No newline at end of file diff --git a/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/images.zip b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e756ff65805d47068e8e4ff7dd15446b47334cdd --- /dev/null +++ b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:354ee732d8d1acfe7978982f68a83cac1212f37628f21bb542100b008aad6718 +size 533850 diff --git a/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/layout.json b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6446d597b010e75835b2dd7678e10cbff06ee1a7 --- /dev/null +++ b/CVPR/2025/Sharp-It_ A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea00ce22528ee344f9d386502c29e70a7b29f3874725e4eb1c465f11edc3e558 +size 356740 diff --git a/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/4d7ee1b6-0c76-4b25-82c5-ad3cd2dab00e_content_list.json b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/4d7ee1b6-0c76-4b25-82c5-ad3cd2dab00e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4c2db0de8d2ff27300b492ae2e3b5692f5a4f8fb --- /dev/null +++ b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/4d7ee1b6-0c76-4b25-82c5-ad3cd2dab00e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db645846e0b3977420220f34abcab9862ae0b86b73b2ae32834771acf7623e91 +size 71726 diff --git a/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/4d7ee1b6-0c76-4b25-82c5-ad3cd2dab00e_model.json b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/4d7ee1b6-0c76-4b25-82c5-ad3cd2dab00e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8d41cb7baf343558258ba9e5e9ba61d7ba66a2fe --- /dev/null +++ b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/4d7ee1b6-0c76-4b25-82c5-ad3cd2dab00e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5e8ae19631d2285d5bc3fb7709011536b52d86549acdd00b6a615dd897554cd +size 90019 diff --git a/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/4d7ee1b6-0c76-4b25-82c5-ad3cd2dab00e_origin.pdf b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/4d7ee1b6-0c76-4b25-82c5-ad3cd2dab00e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f8a46492bb4595f776de139d07556f88f5410d01 --- /dev/null +++ b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/4d7ee1b6-0c76-4b25-82c5-ad3cd2dab00e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36c270699e9ad0597eb7e74523421efe2f1e2b4d0a2d260e33dfeb9a9ef88b00 +size 8166705 diff --git a/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/full.md b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a9e007dcda0b9c33509985a02513b7b268131dcc --- /dev/null +++ b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/full.md @@ -0,0 +1,264 @@ +# SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation + +Duc-Hai Pham $^{1*}$ Tung Do $^{1*}$ Phong Nguyen $^{1}$ Binh-Son Hua $^{1,2}$ Khoi Nguyen $^{1}$ Rang Nguyen $^{3}$ $^{1}$ Movian AI, Vietnam $^{2}$ Trinity College Dublin, Ireland $^{3}$ VinAI, Vietnam +https://sharpdepth.github.io/ + +![](images/340bae52fa94c002872930836e0b9378e4fab3b7ea1556ccef9321342de5ad20.jpg) +Figure 1. We present SharpDepth, a diffusion-based depth model for refining metric depth estimators, e.g., UniDepth [31], without relying on ground-truth depth data. Our method can recover sharp details in thin structures and improve overall point cloud quality. + +# Abstract + +We propose SharpDepth, a novel approach to monocular metric depth estimation that combines the metric accuracy of discriminative depth estimation methods (e.g., Metric3D, UniDepth) with the fine-grained boundary sharpness typically achieved by generative methods (e.g., Marigold, Lotus). Traditional discriminative models trained on real-world data with sparse ground-truth depth can accurately predict metric depth but often produce over-smoothed or low-detail depth maps. Generative models, in contrast, are trained on synthetic data with dense ground truth, generating depth maps with sharp boundaries yet only providing relative depth with low accuracy. Our approach bridges these limitations by integrating metric accuracy with detailed boundary preservation, resulting in depth predictions that are both metrically precise and visually sharp. Our extensive zero-shot evaluations on standard depth estimation benchmarks confirm SharpDepth's effectiveness, showing + +its ability to achieve both high depth accuracy and detailed representation, making it well-suited for applications requiring high-quality depth perception across diverse, real-world environments. + +# 1. Introduction + +Monocular metric depth estimation – the task of predicting absolute depth from a single RGB image – has emerged as a key computer vision problem due to its broad applications in autonomous driving [13–15], augmented reality [44, 52], robotics [7], and 3D scene understanding [5, 30, 43]. Unlike stereo or multi-view approaches that leverage multiple viewpoints to deduce depth, monocular depth estimation aims to infer depth from a single perspective. This approach offers edges in terms of cost, hardware simplicity, and deployment flexibility; however, it also faces significant challenges. + +![](images/050a0365615d99d98dbcf39cd74d82a195e6d2424ad0254eb9cd94dab4fc8e1f.jpg) +Figure 2. The performance of SOTA depth estimation models in terms of depth accuracy (x-axis) on KITTI [1] and DBE Completion (y-axis) on Sintel[46], UnrealStereo4K [40] and Spring [28]. Our method (SharpDepth) is best balanced on both axes. + +lenges due to inherent scale ambiguity and the limited depth cues present in single images. The challenge becomes more difficult in the zero-shot setting, where no fine-tuning is performed on the target domain. + +In the current literature on zero-shot monocular depth estimation, approaches generally fall into two categories: discriminative and generative methods. The discriminative approach [2, 19, 31] typically relies on supervised learning with real-world data annotated by sparse ground truth depth, such as LiDAR measurements. These models are trained to regress metric depth values directly, and they generally provide accurate global depth estimates, capturing the true scale of the scene. However, due to the sparsity of ground-truth data and the reliance on coarse depth cues, these models often struggle with fine details, producing depth maps that tend to be blurry or lacking in edge sharpness, especially around object boundaries. + +Recently, several generative depth estimation models have leveraged diffusion-based text-to-image models, excelling at producing depth maps with high spatial detail and sharp object edges [17, 21]. This improvement is attributed to the rich image priors inherited from large-scale text-to-image models. However, due to limitations inherent in latent diffusion models, fine-tuning these image-conditioned depth models is feasible only with synthetic data where dense depth maps are available, leading to a domain gap when applied to real images. Additionally, these models provide only affine-invariant depth rather than accurate metric depth. This limitation restricts their applicability in scenarios where precise metric depth information is essential. + +In this work, we introduce SharpDepth, a diffusion-based model designed to generate accurate zero-shot metric depth with high-frequency details. Built upon an affine-invariant depth estimator [17, 21], our method refines the initial predictions of an existing metric depth model, enhancing the depth details while retaining accurate metric predictions. To this end, we measure the agreement of the + +depth predictions by the affine-invariant model and the metric depth model by normalizing both depth predictions to a common scale and produce a difference map. Such a difference map allows us to identify image regions with reliable depth predictions as well as inconsistent depth regions where further sharpening and refinement are required. + +Based on the difference map, we propose to exploit the strengths of both the discriminative and generative depth estimators as follows. We propose Noise-aware Gating, a mechanism that guides the depth diffusion model to focus more precisely on regions identified as uncertain in the difference map. To further ensure both sharpness and accurate metric depth in these uncertain regions, we utilize two loss functions. We first leverage Score Distillation Sampling (SDS) loss to enhance depth detail, resulting in output with sharpness comparable to that of diffusion-based depth estimation methods. We then apply a Noise-aware Reconstruction loss to recognize the lack of scale awareness of diffusion-based models. This loss acts as a regularizer, ensuring that the final predictions remain close to the initial depth estimates, maintaining metric accuracy without drifting from the original depth scale. Together, these techniques enable SharpDepth to deliver precise, high-detail metric depth estimations across diverse scenes. Another benefit of the above training losses is that we can train our refinement on real data using only pretrained depth models, without any additional ground truth for supervision. + +To evaluate the performance of SharpDepth, we conduct extensive comparisons between our method with the state-of-the-art methods on both discriminative [2, 19, 31] and generative methods [17, 21]. Experimental results show that our method can achieve both accurate in-depth estimation and still preserve high degrees of sharpness compared to the state-of-the-art metric depth estimators, e.g., UniDepth [31], as shown in Fig. 1 and Fig. 2. + +In summary, our contributions are as follows: + +- We introduce SharpDepth, a novel diffusion-based depth sharpener model that can produce zero-shot metric depth with high-fidelity details. +- Our method can be trained with only images thanks to our two noise-aware proposed modules. The total amount of training images is about 100-150 times smaller than existing monocular depth estimation methods. +- Experiments on various zero-shot datasets show that our method's accuracy is competitive with discriminative models while containing the high-detail output of generative models. + +# 2. Related Work + +Current approaches to zero-shot monocular depth estimation can be divided into two main research lines: metric depth methods and affine-invariant depth methods. + +Zero-shot metric monocular depth estimation. Recently, a new line of metric monocular depth estimation (MMDE) methods has emerged, aiming to predict metric depth, also known as absolute depth, without images from the target domain during training. For example, ZoeDepth [2] achieves metrically accurate zero-shot predictions by fine-tuning an affine-invariant model on combined indoor and outdoor datasets, employing adaptive domain-specific range predictions. Concurrently, ZeroDepth [16] introduces a transformer-based encoder-decoder that leverages camera intrinsics to enhance camera awareness, allowing it to directly decode metric depth without adaptive range prediction. Metric3D [51] and Metric3Dv2 [19] achieve zero-shot single-view metric depth through large-scale training and address metric ambiguity across camera models by incorporating a canonical camera space transformation module. UniDepth [31] further advances MMDE by eliminating the need for test-time camera intrinsics, instead using a pseudo-spherical representation and a self-promptable camera module to predict both depth and camera parameters from a single image. However, these MMDE methods often yield depth predictions with limited detail due to the sparse ground truth in real-world data. In contrast, our approach delivers both high accuracy and fine detail, particularly in boundary regions. + +# Zero-shot affine-invariant monocular depth estimation. + +Monocular depth estimation is an ill-posed geometric problem, and many approaches [4, 8–10, 23, 34] address this by estimating depth up to an unknown global scale and shift, also known as affine-invariant depth. MegaDepth [24] and DiverseDepth [50] leverage large-scale internet images to achieve zero-shot depth estimation, though their training data often includes noisy labels. MiDAS [35] mitigates this issue by using 3D movie frames with scale-shift-invariant losses to ensure consistency across various depth representations. Depth Anything [48] builds on this approach, employing a pseudo-labeling strategy across 62 million unlabeled images to enhance performance in real-world scenarios. Recently, diffusion models have shown promise in improving depth fidelity by incorporating image priors. Marigold [21] fine-tunes the Stable Diffusion model to generate high-quality depth with clear boundaries, while Lotus [17] accelerates inference by optimizing the noise scheduling process. GeoWizard [11] jointly estimates depth and normals, leveraging cross-modal relationships. Despite these advancements, the limitations of synthetic data create domain gaps that hinder the performance of diffusion-based methods in real-world applications, where discriminative feed-forward models still outperform them in terms of accuracy. Furthermore, these methods typically provide relative depth (i.e., depth relationships within a scene) rather than precise metric depth, which restricts their applicability in scenarios requiring accurate metric depth information. + +Affine-invariant and metric depth refinement. Recent work has focused on depth refinement rather than training models from scratch to leverage the benefits of both affine-invariant and metric methods. BetterDepth [54] refines affine-invariant depth by conditioning on outputs from pre-trained models and applying a generative model with a diffusion loss. PatchRefiner [25] builds on this approach for metric refinement by using features from a metric base network to generate residual depth maps, which enhances detail in the final prediction. However, both methods rely heavily on synthetic datasets, limiting their applicability in real-world scenarios. In contrast, our method employs a ground-truth-free fine-tuning protocol, utilizing real-world data without annotations. This reduces reliance on synthetic datasets, minimizes domain gaps, and improves fine detail while preserving metric accuracy. + +# 3. Preliminaries + +Diffusion Model for Depth Prediction. To transfer the rich visual knowledge of diffusion models to the depth estimation domain, Marigold [21] fine-tunes Stable Diffusion [36] for monocular depth estimation. This process reuses the VAE encoder $\mathcal{E}$ on depth images to obtain depth latents $z_{d}$ , optimizing with the $\epsilon$ -prediction objective function in Eq. (1). Additionally, $z_{d}$ is concatenated with the conditional image latent $z_{i}$ , while the original text condition is omitted, $\epsilon \sim \mathcal{N}(0,I)$ , $z_{d}^{t}$ is the noisy version of $z_{d}$ at timestep $t$ , and $\epsilon_{\theta}$ is the output of the diffusion model. + +$$ +\mathcal {L} _ {\theta} = \mathbb {E} _ {t, \epsilon \in \mathcal {N} (0, 1)} \| \epsilon_ {\theta} \left(z _ {d} ^ {t}, t, z _ {i}\right) - \epsilon \| _ {2} ^ {2}. \tag {1} +$$ + +Similarly, Lotus [17] is based on the same principles but introduces key modifications. It reduces the number of timesteps from 1000 to 1, enabling faster inference. Additionally, $z_0$ -prediction is employed to reduce variance. To prevent catastrophic forgetting, Lotus jointly learns to predict both depth and image latents. The final training objective, as shown in Eq. (2), defines $f_{\theta}$ as an $z_0$ -prediction diffusion model, where $s_d$ and $s_i$ are task indicators for depth and image prediction, respectively. + +$$ +\mathcal {L} _ {\theta} = \| z _ {d} - f _ {\theta} \left(z _ {d} ^ {t}, z _ {i}, t, s _ {d}\right) \| _ {2} ^ {2} + \| z _ {i} - f _ {\theta} \left(z _ {i} ^ {t}, t, s _ {i}\right) \| _ {2} ^ {2}. \tag {2} +$$ + +Score Distillation Sampling (SDS) is a distillation technique applied in 3D assets synthesis [26, 32, 42]. By removing the U-Net Jacobian term of the gradient of $\mathcal{L}_{\mathrm{diff}}$ , a differentiable image $x = g(\phi)$ could be optimized without the need for backpropagating through the diffusion model U-Net. The gradient of SDS loss can be approximated as: + +$$ +\nabla_ {\phi} \mathcal {L} _ {\mathrm {S D S}} (\theta , \mathbf {x} = g (\phi)) \triangleq \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\epsilon_ {\theta} \left(z _ {i} ^ {t}, y, t\right) - \epsilon\right) \right], \tag {3} +$$ + +where $w(t)$ is the weighting at timestep $t$ and $y$ is the text prompt. + +![](images/86fe0b90cce2b4d49052544f4b78a185cc12da6b67a6bea2a5ae35151b96c58d.jpg) +Figure 3. Our framework utilizes a diffusion-based estimator and a metric depth estimator to generate affine-invariant and metric depth maps, respectively. A Noise-Aware Gating mechanism produces a selectively noisy latent map, which is fed into our SharpDepth model. The training pipeline uses Score Distillation Sampling and Noise-Aware Reconstruction Losses to refine accuracy and enhance details. + +![](images/36d1f3ea5973aaa1519d6a69b7632cd82915f1533f4ab8bda8c63659c007fded.jpg) +Figure 4. The difference map between the Unidepth and Lotus predictions. The high-difference (brighter) areas are heavily distorted by noise, whereas in the low-difference (darker) areas, some information about the wheel is still recognizable. + +# 4. Methodology + +Given an input image $\mathcal{I}$ , we first use both pre-trained metric depth model $f_{D}$ and diffusion-based depth model $f_{G}$ to produce metric and affine-invariant depth output $d$ and $\tilde{d}$ , respectively. Our goal is to generate a sharpened metric depth map, $\hat{d}$ , using our proposed sharpening model, $\mathbf{G}_{\theta}$ . This model architecture is based on state-of-the-art pre-trained depth diffusion models [17, 21]. + +Instead of naively relying on the forward process of diffusion model [18, 36], we introduce a Noise-aware Gating mechanism (described in Sec. 4.1), which provides explicit guidance to the sharpener $\mathbf{G}_{\theta}$ on uncertain regions. To enable ground-truth free fine-tuning, we use SDS loss to distill fine-grained details from the pretrained diffusion depth model $f_{G}$ and Noise-Aware Reconstruction Loss to ensure accurate metric prediction (as detailed in Sec. 4.2). The + +overall architecture is outlined in Fig. 3. + +# 4.1. Noise-aware Gating + +Our goal is to develop a mechanism that enables the network to identify regions requiring sharpening. Given the challenges of obtaining or annotating per-pixel ground truth (GT) metric depths while maintaining overall sharp detail, we adopt an amortized approach that combines insights from both the diffusion-based depth estimator and a metric estimator. Intuitively, we assume that regions with minimal differences between these models are more reliable, while areas with larger discrepancies require additional supervision. Using two state-of-the-art depth estimators, UniDepth [31] and Lotus [17], we find that UniDepth provides fairly accurate depth predictions compared to sparse GT data, though it lacks the sharp detail of Lotus's output. As shown in Fig. 4, we first normalize these predictions and then compute a difference map between two inferred depth maps. Brighter areas in the difference map highlight regions with significant discrepancies, indicating areas needing further refinement. Conversely, darker regions represent areas of mutual agreement between the depth estimators. While these regions may still benefit from depth refinement, they hold lower priority in the sharpening process. + +To this end, we propose a Noise-aware Gating mechanism incorporating information from the initial metric depth $d$ as input to our $\mathbf{G}_{\theta}$ model. Advances in image inpainting [20], editing [27], and virtual try-on [55] have used explicit masks to guide diffusion models to focus on regions of interest. Inspired by these models, we avoid adding pure Gaussian noise to every pixel of the clean latent depth. Instead, we selectively introduce noise to regions with sig- + +nificant differences, while applying less noise to areas with smaller discrepancies. This strategy effectively directs the sharpener to focus on high-difference (noisy) regions, leaving low-difference (clean) areas mostly unaffected. + +To align the depth ranges of $\tilde{d}$ and $d$ , we first scale and shift $\tilde{d}$ to the range of $d$ before calculating the difference map. The difference map $e$ is then computed as the absolute difference between the adjusted $\tilde{d}$ and $d$ . As training progresses, we observe that our proposed $\mathbf{G}_{\theta}$ begins to generate depth maps superior to those produced by the diffusion-based model $f_{G}$ . As the results, we replace $f_{G}$ with the exponential moving average (EMA) of the training model, $\mathbf{G}_{\bar{\theta}}$ , which serves as a refined initialization for $\tilde{d}$ and enables iterative refinement in multiple steps. Details regarding the performance comparison of these two approaches are provided in Sec. 5.3. + +Once the difference map $e$ is obtained, we use it to control the noise intensity applied to each region of $z_{d}$ , which is the latent representation of $d$ . Specifically, we perform a weighted blending between Gaussian noise $\epsilon$ and $z_{d}$ as follows: + +$$ +z _ {d} ^ {\prime} = \hat {e} \odot \epsilon + (1 - \hat {e}) \odot z _ {d}, \tag {4} +$$ + +where $\hat{e}$ is the downsampled version of the difference map $e$ to match the dimensions of latent $z_{d}$ , and $\odot$ is the elementwise product. This blended latent $z_{d}^{\prime}$ effectively distinguishes high- and low-difference regions between the two depth predictions, serving as a powerful prior for the sharpener as shown in Fig. 4. By separating these regions, the optimization process focuses on high-difference areas while minimizing modifications in similar regions, enabling the sharpener to reconstruct fine-grained details while maintaining accuracy. + +# 4.2. Training Objectives + +Diffusion Depth Prior Distillation. In this section, we introduce our approach to distill the knowledge of $f_{G}$ into our depth sharpener, $\mathbf{G}_{\theta}$ . Inspired by SwiftBrush [29] and DreamFusion [32], we do not train our $\mathbf{G}_{\theta}$ from scratch, instead, perform score distillation on a pretrained diffusion-based depth estimator such as $f_{G}$ . The output predicted latent $\hat{z}$ of our proposed model is computed as follows: $\hat{z} = \mathbf{G}_{\theta}(z_d', z_i)$ . We then slightly modify the original SDS formulation of $\epsilon$ -prediction in the Eq. (3) to match the $x_0$ -prediction of $f_{G}$ . The revised version of SDS loss for training our $\mathbf{G}_{\theta}$ can be defined as follows: + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {S D S}} \triangleq \mathbb {E} _ {t, \epsilon} \left[ w ^ {t} \left(\hat {z} - f _ {G} (\hat {z} ^ {t}; z _ {i}, t)\right) \right], \tag {5} +$$ + +where $\hat{z}^t$ is the noisy version of $\hat{z}$ at time step $t$ . + +Noise-aware Reconstruction Loss. Our distillation objective encourages the output of $\mathbf{G}_{\theta}$ to align more closely with the distribution of the diffusion model $f_{G}$ , leading to + +highly detailed depth images. However, this also causes the network to inherit the limitations of $f_{G}$ , ultimately reducing accuracy in metric depth estimation. To address this issue, we introduce an additional reconstruction loss that preserves the accuracy of the discriminative model by measuring the distance between our network's output and the discriminative output in Eq. (6). + +Specifically, we use the difference map to enforce larger gradients in regions where $d$ and $\hat{d}$ exhibit significant differences. This serves as an explicit regularization mechanism, encouraging the sharpener to focus more on these pixels. While the difference map ensures that regions with minimal differences remain largely unchanged, it may also risk propagating over-smoothing artifacts to $\hat{d}$ . The reconstruction loss is given by: + +$$ +\mathcal {L} _ {\text {r e c o n s}} = \left\| e \odot (\hat {d} - d) \right\|, \tag {6} +$$ + +where $\odot$ is the element-wise multiplication. + +Discussion. Though the two objective functions described above serve different purposes - one enhancing depth detail and the other ensuring accurate depth values - both focus on high-difference regions. High-difference regions are augmented with noise and subsequently refined by $\mathbf{G}_{\theta}$ , resulting in large gradients during optimization. This causes Score Distillation Sampling (SDS) to emphasize these regions which function as an implicit masking optimization. In contrast, the reconstruction loss employs a difference map directly during optimization, resembling an explicit masking optimization approach. The complete training objective function in Eq. (7) comprises $\mathcal{L}_{\mathrm{SDS}}$ and $\mathcal{L}_{\mathrm{recons}}$ where $\lambda_{\mathrm{SDS}}$ and $\lambda_{\mathrm{recons}}$ are loss weighting hyperparameters. + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {\mathrm {S D S}} \mathcal {L} _ {\mathrm {S D S}} + \lambda_ {\text {r e c o n s}} \mathcal {L} _ {\text {r e c o n s}}. \tag {7} +$$ + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets: For training, we use approximately $1\%$ of each of the following real-world datasets, which cover various camera types and scene domains: Pandaset [47], Waymo [39], ArgoVerse2 [45], ARKit [1], Taskonomy [53], and ScanNetv2 [6]. This results in a combined training set of 90,000 images which is 100-150 times smaller than the amount of data has been used for discriminative depth models [31, 51]. + +For testing, we evaluate our approach on seven real datasets: KITTI [9], NYUv2 [38], ETH3D [37], Diode [41], Booster [33], nuScenes [3], iBims [22] and three synthetic datasets: Sintel [46], UnrealStereo4K [40] and Spring [28]. + +Metrics: We use common depth estimation metrics: absolute mean relative error (A.Rel), root mean square error (RMSE), and the percentage of inlier pixels $(\delta_1)$ with a threshold of 1.25. Additionally, to assess the sharpness + +of predictions, we adopt the Depth Boundary Error (DBE) inspired by iBims [22]. Since annotated depth edges are not available for synthetic datasets, we adapt the original metric into a Pseudo Depth Boundary Error (PDBE). In PDBE, we apply Canny Edge Detection to the predicted and ground truth depths to generate sets of edges, which are then used to compute the accuracy $\epsilon_{\mathrm{PDBE}}^{\mathrm{acc}}$ and completion $\epsilon_{\mathrm{PDBE}}^{\mathrm{compl}}$ , following the original iBims formulation. Further details are provided in the supplementary material. + +Implementation details. We implement our method in PyTorch, using Lotus [17] and UniDepth [31] as the diffusion-based and metric depth estimators, respectively. We also adopt the architecture and pretrained weights from Lotus for our SharpDepth model's initialization. The Adam optimizer is used with a learning rate of $1e^{-6}$ , and we set the loss weights as $\lambda_{\mathrm{SDS}} = 1.0$ and $\lambda_{\mathrm{recons}} = 0.3$ . Optimization is performed for 13,000 iterations with a batch size of 1, using 16 gradient accumulation steps. Training our model to convergence takes approximately 1.5 days on two A100 40GB GPUs. We normalize the metric depth from $f_{D}$ to $[-1,1]$ before feeding it into the VAE encoder by using min-max normalization. The output from the decoder is rescaled to metric depth using least-squares alignment with the initial metric depth $d$ . The difference map is applied to ensure alignment only for pixels with minimal differences (i.e pixel locations where $e < 0.5$ ). + +# 5.2. Comparison with the State of the Art + +Baselines. We compare our method against four zero-shot metric depth models: UniDepth [31], Metric3Dv2 [19], ZoeDepth [2], DepthAnythingV2-Metric [49] and PatchRefiner [25]. Ground truth intrinsics are provided to UniDepth and Metric3Dv2. We also introduce a straightforward baseline, termed UniDepth-aligned Lotus, where we convert Lotus to metric depth by simply applying scale and shift adjustments with UniDepth's prediction. Additionally, we include three zero-shot relative depth models: Marigold [21], Lotus [17], and BetterDepth [54] for reference purposes since they require ground truth alignment during test time. + +Quantitative results. As shown in Tab. 1, our method achieves accuracy on par with metric depth models, demonstrating the effectiveness of our training pipeline. This approach enables the sharpener to selectively refine uncertain regions while preserving confident areas intact. Although PatchRefiner is trained on a comparable dataset size to SharpDepth, its performance is robust primarily on outdoor datasets. By leveraging our ground-truth-free fin-tuning pipeline, we are able to train across diverse real-world datasets, enabling strong generalization to both indoor and outdoor scenes. Furthermore, our approach consistently surpasses the naive UniDepth-aligned Lotus baseline across all datasets, demonstrating that relying solely + +on UniDepth's output for alignment is suboptimal. Our approach addresses this by using a difference map as a powerful guide, enabling alignment only on reliable pixels. + +Also, Tab. 2 provides the evaluation of depth details on 3 synthetic (Sintel [46], UnrealStereo [40] and Spring [28]) and 1 real dataset (iBims [22]). Compared to UniDepth, SharpDepth obtains significantly higher results in both edge accuracy and completion. By leveraging rich priors from the pre-trained diffusion model, our method can produce sharper depth discontinuities, leading to high accuracy scores on all datasets. On the other hand, discriminative-based methods frequently produce smooth edges without clear transitions between objects. This is highlighted in the completeness error, as missing edges are expressed by large values of this error. While UniDepth-aligned Lotus and DAv2-Metric depth details are competitive, high-frequency details that lack accurate metric precision on real datasets significantly limit the model's application. + +Qualitative Results. For a visual assessment, we present qualitative results in Fig. 5 and Fig. 6. As shown in Fig. 5, SharpDepth encompassed both high-frequency details in thin structures (fences and traffic poles in the first example) and accurate metric scene layout (second row). Moreover, detailed depth maps allow for better object reconstruction. To show this, we un-project the predicted depth maps to point-clouds from both SharpDepth and UniDepth in Fig. 6. SharpDepth demonstrates better reconstruction fidelity, accurately capturing intricate details like the spikes on a durian and the contours of keyboard keycaps. + +# 5.3. Ablation Study + +In this section, we validate the impact of our design choices through experiments on two validation sets: KITTI [1] and Sintel [46]. Unless otherwise specified, all experiments use the same training dataset as described in Sec. 5.2. + +Analysis of Noise-aware Gating are presented in Tab. 3 (Settings A-B). Instead of calculating the difference map as in Sec. 4.1 to generate input for $\mathbf{G}_{\theta}$ , we test two alternative inputs: (A) Gaussian noise, as in [17], and (B) the output of the pretrained metric depth model. Both configurations led to decreased performance, with RMSE increasing by 1.64 in setting (A) and by 2.34 in setting (B). In setting (A), the lack of prior knowledge from the pretrained metric depth model led the sharpener to behave similarly to Lotus, degrading accuracy. In setting (B), although the pretrained metric depth estimator provides valuable prior knowledge, the absence of explicit guidance from the difference map $e$ results in ambiguity regarding which regions require refinement during reconstruction loss calculation. This makes it difficult for the model to balance the diffusion and metric depth priors, hindering training. + +Effects of the Training Objectives discussed in Sec. 4.2 + +
MethodKITTINYUv2ETH3DDiodeBoosterNuScenesAvg.
δ1↑A.Rel↓δ1↑A.Rel↓δ1↑A.Rel↓δ1↑A.Rel↓δ1↑A.Rel↓δ1↑A.Rel↓Rank↓
Marigold [21]92.30.0996.40.0595.90.0676.70.3197.40.0565.90.27-
Lotus [17]87.70.1196.60.0596.00.0673.70.3398.80.0450.90.36-
BetterDepth [54]95.20.0798.00.0498.10.05-------
UniDepth [31]97.90.0598.40.0525.30.4666.00.2628.00.4983.80.142.0
ZoeDepth [2]96.50.0695.20.0834.20.5830.20.4820.80.6421.80.594.0
Metric3Dv2 [19]97.60.0597.40.0782.40.1488.80.1615.50.6784.20.201.5
PatchRefiner [25]79.20.161.02.485.01.7825.01.261.05.5531.70.586.7
DAv2-M (Indoor) [49]1.00.5569.60.2039.30.3524.80.4357.30.3114.80.415.7
DAv2-M (Outdoor) [49]85.30.121.02.803.62.1022.01.501.07.119.10.597.0
UD-aligned Lotus83.60.1393.90.0920.30.4955.90.3626.40.4941.20.435.2
SharpDepth (ours)97.30.0696.90.0622.60.4761.50.2927.70.4978.50.183.5
+ +Table 1. Comparison for depth accuracy purpose on real-image datasets. These methods are trained and tested on non-overlapping datasets. SharpDepth results use the initial metric depth prediction of UniDepth. '-' indicates not reported results. We ranked methods that do not require GT alignment as best, second-best, and third-best. Gray indicates the method that has been trained on the training set. + +
MethodSintelUnrealStereo4KSpringiBimsAvg.
\( \epsilon_{PDBE}^{acc} \)↓\( \epsilon_{PDBE}^{compl} \)↓\( \epsilon_{PDBE}^{acc} \)↓\( \epsilon_{PDBE}^{compl} \)↓\( \epsilon_{PDBE}^{acc} \)↓\( \epsilon_{PDBE}^{compl} \)↓\( \epsilon_{DBE}^{acc} \)↓\( \epsilon_{DBE}^{compl} \)↓Rank↓
Marigold [21]1.9052.51.9768.61.85150.31.8513.4-
Lotus [17]2.0331.91.2133.21.27102.81.9211.0-
UniDepth [31]3.73113.38.65257.35.29229.72.0030.07.0
ZoeDepth [2]3.3545.85.39649.34.05204.22.0523.76.2
Metric3Dv2 [19]1.9263.82.75446.91.78118.42.1412.35.0
PatchRefiner [25]3.8658.64.98800.84.19225.42.4938.37.2
DAv2-M (Indoor) [49]1.4933.12.22140.41.3986.11.9511.13.0
DAv2-M (Outdoor) [49]1.5130.82.47318.91.36129.61.8217.82.7
UD-aligned Lotus2.0431.91.2133.21.27102.71.9211.02.7
SharpDepth (ours)1.9436.21.3761.51.24147.61.8013.12.0
+ +Table 2. Comparison for depth details purpose on one real-image dataset (iBims) and three synthetic datasets (Sintel, Unreal-Stereo4K, Spring). We ranked methods that do not require GT alignment as best, second-best, and third-best. + +
SettingMethodKITTISintel
δ1↑A.Rel ↓RMSE ↓εaccDBE↓εcomplDBE↓
-Ours0.9730.0602.371.9436.4
AInput Noise latent0.8170.1353.981.9434.6
BInput zd0.7010.1864.783.30116.9
Cw/o SDS loss0.9780.0512.283.70112.5
Dw/o reconstruction loss0.8430.1283.661.9434.1
ELotus teacher (ours)0.9730.0602.371.9436.4
FMarigold teacher0.9730.0582.342.4084.7
GFrozen Lotus [17]0.9670.0692.432.0040.6
HEMA update (ours)0.9730.0602.371.9436.4
+ +Table 3. Ablation study of different design choices. + +are analyzed in Tab. 3 (Settings C-D). We remove each component to examine its impact on the sharpener. In setting (C), we exclude the SDS loss, while in setting (D), + +we remove the Noise-aware Reconstruction loss. In setting (C), without the distillation loss to incorporate information from the pretrained diffusion-based depth estimator $f_{G}$ , our model aligns more closely with $f_{D}$ , producing nearly identical predictions. As a result, depth accuracy remains competitive, closely matching $f_{D}$ , while setting (D) mirrors $f_{G}$ , with high detail accuracy but lower depth accuracy. + +Study of Pretrained Teacher Model results are shown in Tab. 3 (Settings E-F). We replace the pretrained depth diffusion model with two alternatives: Lotus [17] in setting (E), the same as Ours, and Marigold [21] in setting (F). Setting (F) yields slightly better depth-accuracy metrics, but setting (E) demonstrates superior detail-accuracy, with a substantial margin of 50 in DBE completion. We provide qualitative results in Fig. 7. Based on these findings, we select Lotus as the teacher model in our main experiments. + +Effect of Online vs. Offline Models is examined in Tab. 3 + +![](images/a175170804b20220fcc4ff1b19c1f3636018c06b6d03de251de238f11ec26730.jpg) +Figure 5. Zero-shot qualitative results on unseen test samples of KITTI [12] and DIODE [41] dataset. Our method strikes a balance between depth accuracy and details. UniDepth lacks several details while UniDepth-aligned Lotus is less accurate. The last column represents the colormap ranges for depth (spectral) and AbsRel error (coolwarm). + +![](images/ea2e9194a91057cf8165c1529bef0fce969b0e10f0ffe34e185d705130168b11.jpg) +Figure 6. Un-projected point clouds from in-the-wild image. + +(Settings G-H), where we explore the impact of online (our approach) and offline models on difference map calculation during training. We compare two settings: (G) using the static Lotus model and (H) using the EMA (exponential moving average) of the training model $\mathbf{G}_{\bar{\theta}}$ . In setting (G), the offline model tends to overfit to the fixed differences between $f_{D}$ and $f_{G}$ . In contrast, setting (H) dynamically learns from a difference map that progressively decreases as training advances. Empirically, setting (H) outperforms (G), so we adopt this approach in our main experiments. + +![](images/062b1867b0313c48313ee5fd97d267aa6c663c8ee1d88eeecec426766a5963b0.jpg) +Figure 7. The effect of the pre-trained teacher model. + +# 6. Conclusion + +We proposed SharpDepth, a diffusion-based depth model that brings the metric precision of discriminative depth models into generative depth estimators. By balancing accuracy with detail, our method produces depth maps that are both metrically precise and visually refined, advancing high-quality zero-shot monocular depth estimation. Our evaluations show its strong performance across benchmarks, highlighting its potential for real-world applications requiring high-quality depth perception. + +# References + +[1] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 2, 5, 6 +[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 2, 3, 6, 7 +[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 CVPR, 2020. 5 +[4] Weifeng Chen, Shengyi Qian, David Fan, Noriyuki Kojima, Max Hamilton, and Jia Deng. Oasis: A large-scale dataset for single image 3d in the wild. In CVPR, 2020. 3 +[5] Wongun Choi, Yu-Wei Chao, Caroline Pantofaru, and Silvio Savarese. Indoor scene understanding with geometric and semantic contexts. IJCV, 112, 2015. 1 +[6] 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, 2017. 5 +[7] Xingshuai Dong, Matthew A Garratt, Sreenatha G Anavatti, and Hussein A Abbass. Towards real-time monocular depth estimation for robotics: A survey. IEEE Transactions on Intelligent Transportation Systems, 23(10), 2022. 1 +[8] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In ICCV, 2021. 3 +[9] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. NIPS, 27, 2014. 5 +[10] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, 2018. 3 +[11] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In ECCV. Springer, 2025. 3 +[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. *IJRR*, 32 (11), 2013. 8 +[13] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In ICCV, 2019. 1 +[14] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, 2020. +[15] Vitor Guizilini, Igor Vasiljevic, Rares Ambrus, Greg Shakhnarovich, and Adrien Gaidon. Full surround monodepth from multiple cameras. IEEE Robotics and Automation Letters, 7(2), 2022. 1 + +[16] Vitor Guizilini, Igor Vasiljevic, Dian Chen, Rares Ambrus, and Adrien Gaidon. Towards zero-shot scale-aware monocular depth estimation, 2023. 3 +[17] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 2, 3, 4, 6, 7 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. CVPR, 2020. 4 +[19] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. arXiv preprint arXiv:2404.15506, 2024. 2, 3, 6, 7 +[20] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 4 +[21] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 2, 3, 4, 6, 7 +[22] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In ECCV Workshops, 2018. 5, 6 +[23] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 3 +[24] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 3 +[25] Zhenyu Li, Shariq Farooq Bhat, and Peter Wonka. Patchrefiner: Leveraging synthetic data for real-domain high-resolution monocular metric depth estimation. In ECCV. Springer, 2024. 3, 6, 7 +[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3 +[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 4 +[28] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In CVPR, 2023. 2, 5, 6 +[29] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In CVPR, 2024. 5 +[30] Duc-Hai Pham, Duc-Dung Nguyen, Anh Pham, Tuan Ho, Phong Nguyen, Khoi Nguyen, and Rang Nguyen. Semi-supervised 3d semantic scene completion with 2d vision foundation model guidance, 2025. 1 + +[31] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 7 +[32] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3, 5 +[33] Pierluigi Zama Ramirez, Alex Costanzino, Fabio Tosi, Matteo Poggi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Booster: a benchmark for depth from images of specular and transparent surfaces. PAMI, 2023. 5 +[34] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 3 +[35] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. PAMI, 44(3), 2022. 3 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3, 4 +[37] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In CVPR, 2017. 5 +[38] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV. Springer, 2012. 5 +[39] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5 +[40] Fabio Tosi, Yiyi Liao, Carolin Schmitt, and Andreas Geiger. Smd-nets: Stereo mixture density networks. In CVPR, 2021. 2, 5, 6 +[41] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 5, 8 +[42] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022. 3 +[43] Lijun Wang, Yifan Wang, Linzhao Wang, Yunlong Zhan, Ying Wang, and Huchuan Lu. Can scale-consistent monocular depth be learned in a self-supervised scale-invariant manner? In ICCV, 2021. 1 +[44] Jamie Watson, Michael Firman, Gabriel J Brostow, and Daniyar Turmukhambetov. Self-supervised monocular depth hints. In ICCV, 2019. 1 +[45] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493, 2023. 5 + +[46] Jonas Wulff, Daniel J Butler, Garrett B Stanley, and Michael J Black. Lessons and insights from creating a synthetic optical flow benchmark. In ECCV. Springer, 2012. 2, 5, 6 +[47] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). IEEE, 2021. 5 +[48] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 3 +[49] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2024. 6, 7 +[50] Wei Yin, Yifan Liu, and Chunhua Shen. Virtual normal: Enforcing geometric constraints for accurate and robust depth prediction. PAMI, 2021. 3 +[51] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In ICCV, pages 9043-9053, 2023. 3, 5 +[52] Mehmet Kerim Yucel, Valia Dimaridou, Anastasios Drosou, and Albert Saa-Garriga. Real-time monocular depth estimation with sparse supervision on mobile. In CVPR, 2021. 1 +[53] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018. 5 +[54] Xiang Zhang, Bingxin Ke, Hayko Riemenschneider, Nando Metzger, Anton Obukhov, Markus Gross, Konrad Schindler, and Christopher Schroers. Betterdepth: Plug-and-play diffusion refiner for zero-shot monocular depth estimation. arXiv preprint arXiv:2407.17952, 2024. 3, 6, 7 +[55] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In CVPR, 2023. 4 \ No newline at end of file diff --git a/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/images.zip b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a300827f672f154e894b2da53fb2dcb56ae0c9a1 --- /dev/null +++ b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:098e494346a0e8d4d15c9dda295690bf446b90ec5e43a79ca7591cc16b519444 +size 797652 diff --git a/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/layout.json b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..992c1d6cab3a1f765284aa56cfc31efaf1beacf4 --- /dev/null +++ b/CVPR/2025/SharpDepth_ Sharpening Metric Depth Predictions Using Diffusion Distillation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d54aea86c52cb127b0e366d2ba0150eebb9ff15c305fc77a21bc9e4318ed3929 +size 363529 diff --git a/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/6231c8d2-7f2e-415a-9969-13e7f045bb16_content_list.json b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/6231c8d2-7f2e-415a-9969-13e7f045bb16_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2f6011741e7f96504a71512a30adbeef7293906 --- /dev/null +++ b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/6231c8d2-7f2e-415a-9969-13e7f045bb16_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb426fed1de40e64f7bbadb291552d05b2b88b4dcfb6cfa1a49aa90abb1be8aa +size 106202 diff --git a/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/6231c8d2-7f2e-415a-9969-13e7f045bb16_model.json b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/6231c8d2-7f2e-415a-9969-13e7f045bb16_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ff58ab49ee14e2528ffae752fc08e26c24292c48 --- /dev/null +++ b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/6231c8d2-7f2e-415a-9969-13e7f045bb16_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e22a5838e824694fbb7b6e4a3cc72f923cfc752e3e1a1b90f08207b1a3c3e0be +size 132429 diff --git a/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/6231c8d2-7f2e-415a-9969-13e7f045bb16_origin.pdf b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/6231c8d2-7f2e-415a-9969-13e7f045bb16_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c33ddb64b75af5a8db912bc5e64825012d330eac --- /dev/null +++ b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/6231c8d2-7f2e-415a-9969-13e7f045bb16_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb2682cd74f197e2792deb76e7d7cc2761e3cb07b4659a8b39fce718e7033205 +size 599435 diff --git a/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/full.md b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..10f7ac1ae5b60316c2c5dba903ba4d687682af8c --- /dev/null +++ b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/full.md @@ -0,0 +1,426 @@ +# Shift the Lens: Environment-Aware Unsupervised Camouflaged Object Detection + +Ji Du $^{1,2}$ , Fangwei Hao $^{1}$ , Mingyang Yu $^{1}$ , Desheng Kong $^{1}$ +Jiesheng Wu $^{1}$ , Bin Wang $^{1}$ , Jing Xu $^{1,*}$ , Ping Li $^{2,3*}$ + +1College of Artificial Intelligence, Nankai University, China + +$^{2}$ Department of Computing, The Hong Kong Polytechnic University, Hong Kong + +$^{3}$ School of Design, The Hong Kong Polytechnic University, Hong Kong + +# Abstract + +Camouflaged Object Detection (COD) seeks to distinguish objects from their highly similar backgrounds. Existing work has essentially focused on isolating camouflaged objects from the environment, demonstrating ever-improving performance but at the cost of extensive annotations and complex optimizations. In this paper, we diverge from this paradigm and shift the lens to isolating the salient environment from the camouflaged object. We introduce EASE, an Environment-Aware unSupervised COD framework that identifies the environment by referencing an environment prototype library and detects camouflaged objects by inverting the retrieved environmental features. Specifically, our approach (DiffPro) uses large multimodal models, diffusion models, and vision-foundation models to construct the environment prototype library. To retrieve environments from the library and refrain from confusing foreground and background, we incorporate three retrieval schemes: Kernel Density Estimation-based Adaptive Threshold (KDE-AT), Global-to-Local pixel-level retrieval (G2L), and Self-Retrieval (SR). Our experiments demonstrate significant improvements over current unsupervised methods, with EASE achieving an average gain of over $10\%$ on the COD10K dataset. When integrated with SAM, EASE surpasses prompt-based segmentation approaches and performs competitively with state-of-the-art fully-supervised methods. Code is available at https://github.com/xiaohainku/EASE. + +# 1. Introduction + +Camouflaged Object Detection (COD) [15] focuses on identifying objects that blend in with their surroundings, making them visually challenging to detect, as indicated by + +![](images/5d5b56a877f59b4d746b03a0ba5ec6219375fe40f6275a36ae323a4fa2e8bf65.jpg) +Feature Visualization + +![](images/a52cb5a4b1da1204150bb4176d15e70bdbc4a640256323441b8a80f5f356b6a3.jpg) +Figure 1. (Left) Feature visualization from DINOv2 [55] shows that objects blend seamlessly with their surroundings, making it extremely challenging to distinguish foreground solely based on feature similarity due to a lack of clear boundaries. (Right) EASE offers a solution by shifting focus from direct feature comparison. Instead of relying on inter-feature similarity alone, EASE identifies background regions by comparing features against predefined environment prototypes. Camouflaged objects are then isolated by inverting this background detection. + +![](images/23cc9ec761b980ebb94d8336dc53bf9f680dbb1d9f4afac9db76730e97121da5.jpg) + +![](images/6783e485cd6131626f7e5fb81a7d30b48b025da2158d416bbd85361955c40f76.jpg) + +![](images/5c238f1586dbf6990c99fca29f1f700e4e587f5ee0f728018510f408e7bedef1.jpg) + +![](images/487bad30bc676b6f4e6f7562fad088d0c40fc62b9aa5bcf3795c2486f4fb6258.jpg) +EASE + +![](images/2f2a01b7b7076524576060a8fedddc3d50fc1425d732bd83e58a732d159b6125.jpg) + +the left part of Fig. 1. Existing COD methods focus on identifying hidden cues within the environment to detect the concealed. Fully-supervised learning approaches in COD tackle the challenge by designing architectures that capture fine camouflage details in complex settings, using bottom-up/top-down [15, 16, 53] or multi-stream [76, 82] structures to leverage dense annotations. In a similar vein, semi-supervised learning [79] seeks to maximize the learning of camouflage features with minimal annotation. To further reduce the reliance on extensive labeling, weakly-supervised learning relies on sparse annotations, like points [8], scribbles [26], boxes [79] or pseudo-labels [23], to guide models in distinguishing camouflaged objects from their backgrounds. + +Another line of research revolves around prompt-based segmentation, devoid of reliance on annotations. A typical paradigm involves using Large Multimodal Mod- + +els (LMMs) to identify categories [28, 29] or bounding boxes [34, 70] of camouflaged objects. These outputs then serve as prompts within SAM [39] to achieve accurate segmentation. + +The above approaches focus on improving the detection of non-salient camouflaged objects from the salient environment—similar to discerning a glass ball within a bottle of water. Supervised methods have shown notable progress, yet striking the ideal balance between performance and annotation expense warrants further exploration. For prompt-based segmentation, the hallucination [31, 43] originating from LMMs leaves a large performance gap with supervised learning ascribed to the high similarity between the camouflaged objects and the environment. + +If finding the glass ball is so complicated, we would like to ask: why don't we just pour off the water? + +When humans cannot immediately locate camouflaged objects, they often rely on their prior knowledge about the environment, systematically filtering out extraneous details until the target emerges. Inspired by this cognitive strategy, we introduce EASE. As illustrated in Fig. 1, rather than directly targeting camouflaged objects, EASE emphasizes the surrounding environment, retrieving from a specially crafted environment prototype library. It then identifies the camouflaged object by inverting the retrieved details. To unleash the potential of EASE, we consider the key to be ① the prototype library and ② retrieval schemes. + +Instead of relying on the laborious manual collection of environmental prototypes from online sources or datasets—which may also risk revealing camouflaged object information—our proposed DiffPro utilizes foundation models to automatically build an environment-focused prototype library. First, we employ Large Multimodal Models (LMMs) such as LLaVA-1.5 [50] to define a set of environment categories. Recognizing broad environmental contexts is a simpler task for LMMs than identifying camouflaged objects. To create diverse prototypes that exclude camouflaged objects, we take advantage of the powerful conditional generation abilities of diffusion models like Stable Diffusion [61] to efficiently generate a range of environmental images. Finally, self-supervised vision foundation models, such as DINOv2 [55], are used to extract both global (average-pooled) and local (pixel-level) feature vectors from these images. + +The prototype library enables us to assess whether a target belongs to the environment or is a camouflaged object by retrieving the top-N prototypes and setting a similarity threshold based on their average. However, fixed thresholds are unsuitable given the variability across camouflage scenarios, and they also add extra hyperparameters, complicating the model. To resolve this, we introduce KDE-AT, which uses kernel density estimation to model the similarity distribution, allowing the minimum value to function as + +a flexible, adaptive threshold that reduces complexity and enhances detection reliability. + +For effective and efficient retrieval, we propose Global-to-Local retrieval (G2L). G2L first identifies the top-K images using global prototypes and then performs pixel-level retrieval on these selected images. This two-step approach sharpens retrieval precision to a pixel-by-pixel level while reducing the number of prototypes needed, eliminating redundant retrievals, and enhancing overall efficiency. + +To address the distribution gap between camouflaged and generated images that may impact retrieval accuracy, we introduce a Self-Retrieval (SR) approach. Unlike G2L that relies on external prototype libraries, SR uses its own features as the prototype reference. SR includes two prototypes—foreground and background—derived through mask-averaged pooling over target features, guided by G2L masks. The similarity of each pixel is determined by the contrast between these prototypes. By incorporating interfeature similarity, SR effectively refines retrieval accuracy, reducing false positives and negatives. + +Extensive experiments conducted across multiple datasets showcase that our EASE substantially outperforms the SOTA unsupervised methods. Our contributions are summarized as follows: + +- We introduce a new perspective to understand COD, i.e., stripping the environment from the camouflaged rather than finding the camouflaged from the environment. The proposed approach EASE segments camouflaged objects through retrieving from the prototype library in a training-free manner, substantially reducing the expense of annotation and training. +- To construct the environment prototype library, we propose DiffPro, an automatic and efficient pipeline that mitigates issues where camouflaged objects in the environment library could obscure retrieval decisions. DiffPro addresses this by leveraging multiple foundation models in a collaborative framework. +- We devise multifaceted retrieval schemes (including KDE-AT, G2L, and SR) to effectively and efficiently isolate environments from camouflaged objects. Extensive experiments speak to the efficacy of our method. + +# 2. Related Work + +# 2.1. Camouflaged Object Detection + +Camouflaged Object Detection [15, 77] is attracting increasing research attention because of its distinct challenges and real-world applicability. Existing efforts have primarily focused on disentangling camouflaged objects from the environment, and can be categorized into supervision-based [7, 16, 20, 21, 24, 25, 30, 32, 33, 41, 47, 48, 51, 53, 56, 67, 68, 75, 76, 82-84] and prompt-based [28, 29, 70] segmentation depending on whether supervision signals are + +required or not. Supervision-based methods can be further divided into fully-supervised and weakly-supervised learning, where fully-supervised learning employs dense annotations as signals while weakly-supervised learning utilizes sparse supervision to ease the labeling burden. Another line of research, prompt-based segmentation, extracts insights on camouflaged objects from LMMs by drawing on category [28, 29] or localization [29, 70]. These insights are then processed to generate prompts compatible with SAM, which then performs the segmentation. + +Similar to prompt-based segmentation, our method is also supervision-free and foundation model-dependent. However, instead of searching for camouflaged objects directly from the environment, we eliminate the environment and let the camouflaged objects reveal themselves. + +# 2.2. Unsupervised Object Segmentation + +Unsupervised Object Segmentation [1, 54, 71] aims to segment objects without any supervision signals. Current prevailing unsupervised methods revolve around segmenting objects using features from self-supervised learning models [6, 55]. LOST [64] utilizes the activation features of a pre-trained self-supervised visual Transformer, DINO [6], selecting regions with low similarity as seeds and expanding them to identify object boundaries, thereby achieving object localization. TokenCut [73] adopts features as vertices of the graph, cosine similarity between features as the weights of edges, and then employs Normalized Cut (Ncut) [63] to segment the target into foreground and background parts. Based on this, CutLER [72] proposes MaskCut, which can generate multiple masks compared to TokenCut. CuVLER [1] leverages features from multiple DINO and DINOv2 to perform Ncut and derive the final mask by clustering and voting. DiffCut [10] utilizes features from the diffusion model for Ncut instead of DINO. + +The most similar philosophy to our approach is FOUND [65], which also discovers objects by looking for the background rather than the foreground. FOUND utilizes the less weighted features in the DINO self-attention maps as seed patches and obtains the coarse background mask by similarity matching. The reversed background masks are further regarded as supervision in the self-supervised refinement stage for fine segmentation. However, the high similarity between camouflaged objects and the environment leads to a remarkable decrease in self-supervised feature differentiation (see Tab. 1). Therefore, effective separation of background and foreground cannot be achieved solely based on the similarity between features (Fig. 1 left). We propose prototype library-based retrieval to address this problem. Our starting point, as illustrated in Fig. 1, is that even though two features are similar, they may not maintain consistent similarity with the third party. + +# 2.3. Retrieval-Augmented Generation + +In NLP, Retrieval-Augmented Generation (RAG) [4, 19, 44] aims to compensate for the deficiencies of language models on knowledge-intensive tasks, while illuminating their decisions and updating their world knowledge. This paradigm has recently catalyzed popularity in the vision community and spawned promising applications in a growing number of domains, such as long-tailed recognition [57], video captioning [78], continual learning [58], image retrieval [69] and object detection [38]. + +The most related to our method is retrieval-based segmentation [3, 36, 74]. These methods generate a series of category prototypes using diffusion models and then employ pre-trained segmentation models [74], unsupervised segmentation methods [36], or superpixel segmentation algorithms [3] to generate optimized mask proposals, which are assigned semantic class through retrieval. For these methods, generating an image of a dog may be accompanied by other undesirable ones, such as grass or trees, which would dilute the prototype semantics. However, our method focuses not on the foreground category but on the background. Our DiffPro generates target-only prototypes and excludes potential interferences. To obtain fine segmentation, we design a series of retrieval schemes instead of using existing segmentation methods. + +# 3. Method + +Given an image, unsupervised camouflaged object detection aims to generate a binary mask containing camouflaged objects without any supervision as well as training. Unlike previous approaches [15, 70] that focused on painstakingly searching for camouflaged objects from the environment, our approach focuses on detecting the obvious environment and then inverting it to obtain foreground objects. Building on a series of methods for discovering objects using self-supervised features [64, 65, 72, 73], our approach accomplishes object segmentation through retrieval, utilizing the similarity distribution of features to a third party rather than the similarity between features. + +The overview of our proposed EASE is illustrated in Fig. 2. First, we propose DiffPro to craft the environment prototype library leveraging various foundation models (Sec. 3.1). Second, for effective and efficient retrieval, we propose G2L to implement global-to-local retrieval (Sec. 3.2). Third, we propose SR to supplement G2L retrieval by utilizing the similarity between features, where the features themselves serve as the prototype library (Sec. 3.3). Finally, we introduce KDE-AT, which provides adaptive similarity thresholds for each image instead of fixed thresholds via hyperparameter search (Sec. 3.4). + +![](images/c2dd8bec1ae387231826f215b0ac9b6adffeadc5f7014839b593d9ac9552296d.jpg) +Figure 2. EASE consists of two main components: DiffPro and Retrieval. In DiffPro, LMMs establish an environmental category set based on existing datasets. With these categories, Stable Diffusion generates corresponding environmental images. Using DINO, these images are processed to extract both global and local features, forming a comprehensive environment prototype library. During Retrieval, DINO generates a set of patch embeddings from the target images. In the first stage, G2L retrieval, the top-K images in the prototype library that are most similar to the environment of the target image are retrieved. Next, a more detailed retrieval is conducted on a subset of local features from the top-K images. To enhance retrieval precision and mitigate potential noise, Self-Retrieval (SR) uses coarse masks from the G2L process. These masks are employed to create global average pooling features for both foreground and background regions, which are then retrieved again for refinement. All retrieval thresholds are adaptively set using KDE-AT, ensuring optimal results across varied scenarios. + +# 3.1. Environment Prototype Library + +We propose DiffPro (left in Fig. 2) to build a prototype library that contains only environments. + +The first step is to get the environment category set. Inspired by the power of LMMs demonstrated on VQA [2, 11, 45, 49], we utilize LMMs to derive the environment set. Although previous works [29, 70] have shown that the imperceptibility of camouflaged objects can cause potential hallucinations in LMMs, our approach enables LMMs to recognize the salient environment rather than the camouflaged object and therefore bypasses the potential for erroneous output. Given the image datasets $\{\mathbf{I_i}\}_{i = 1}^n$ where $n$ denotes the total number of images, the environment category could be derived by $C_i = \mathrm{LMM}(\mathbf{I_i},P_c)$ . $P_{c}$ denotes the prompt. We use "What is the environment of this photo? Answer the question using a single word or phrase" for it. By removing repeated and irrelevant categories, all $C_i$ constitute the set of environment categories $\{C_j\}_{j = 1}^m$ , where $m$ denotes the total categories. + +The second step is to get the environment image set. An intuitive idea is to retrieve images from the Web via the search engine or from an existing dataset such as COCO-Stuff [5]. However, retrieval via the Web takes a significant amount of time, even without labeling the images. In addition, both retrieved images and images in existing datasets suffer from serious object-environment coexistence issues. For example, flowers are often accompanied by bees or butterflies. The presence of objects in the environment library will significantly distort the similarity distribution and affect the separation of the environment. + +To address these issues, we turn to the powerful conditional generation capabilities of diffusion models [27, 61] to automatically generate images containing only the environment. For each category $C_j$ in the environment category set $\{C_j\}_{j=1}^m$ , we leverage Stable Diffusion [61] to generate $l$ images, $\mathbf{G}_{\mathbf{i},\mathbf{C}_j} = \mathrm{SD}(C_j,P_s)$ , where $P_s$ denotes the prompt. In this paper, we adopt "a photo of [class]" for $P_s$ . All generated images make up the environment image set + +$\{\mathbf{G}_{\mathbf{i},\mathbf{C_j}}\}_{i = 1:l,j = 1:m}$ + +The third step is to obtain the final environment prototype library. We adopt self-supervised models, DINO [6, 55], as the feature extractor. Given the generated image $\mathbf{G}$ in the environment image set $\{\mathbf{G}_{\mathrm{i},\mathbf{C_j}}\}_{i = 1:l,j = 1:m}$ , we employ DINO to extract the patch embeddings $\mathbf{E}$ from the last layers, $\mathbf{E} = \mathrm{DINO}(\mathbf{G}_{\mathrm{i},\mathbf{C_j}})$ , $\mathbf{E} \in \mathbb{R}^{h \times w \times d}$ , where $h \times w$ denotes the resolution of the features and $d$ the embedding dimension. + +For each generated image, local prototypes could be derived as $\mathbf{P}_{\mathbf{i},\mathbf{j}}^{\mathbf{l}} = \mathbf{E}_{\mathbf{i},\mathbf{j}} \in \mathbb{R}^{d}$ . By employing global average pooling, we get the global feature as + +$$ +\mathbf {P} ^ {\mathbf {g}} = \frac {1}{h w} \sum_ {i = 1} ^ {h} \sum_ {j = 1} ^ {w} \mathbf {P} _ {\mathbf {i}, \mathbf {j}} ^ {\mathbf {l}} = \frac {1}{h w} \sum_ {i = 1} ^ {h} \sum_ {j = 1} ^ {w} \mathbf {E} _ {\mathbf {i}, \mathbf {j}}. \tag {1} +$$ + +We denote the final environment prototype library as $\{(\mathbf{P}^{\mathbf{g}},\mathbf{P}^{\mathbf{l}})\}$ . + +# 3.2. Global-to-Local Retrieval + +We propose G2L to efficiently retrieve the environment from coarse to fine. As illustrated in Fig. 2 (right), given the test image $\mathbf{T} \in \mathbb{R}^{H \times W \times 3}$ , we adopt DINO to extract the patch embeddings $\mathbf{F} \in \mathbb{R}^{h \times w \times d}$ . We first perform a global retrieval. For each global feature $\mathbf{P^g} \in \mathbb{R}^d$ in the prototype library $\{(\mathbf{P^g}, \mathbf{P^l})\}$ , the cosine similarity between $\mathbf{F_{i,j}} \in \mathbb{R}^d$ and $\mathbf{P^g}$ could be formulated as + +$$ +\mathbf {S} _ {\mathrm {i}, \mathrm {j}} = \frac {\mathbf {F} _ {\mathrm {i} , \mathrm {j}} \cdot \mathbf {P} ^ {\mathrm {g}}}{\| \mathbf {F} _ {\mathrm {i} , \mathrm {j}} \| \| \mathbf {P} ^ {\mathrm {g}} \|}. \tag {2} +$$ + +For each $\mathbf{F}_{\mathrm{i},\mathbf{j}}$ , we retrieve the top-N global features and adopt the average cosine similarity to replace $\mathbf{S}_{\mathrm{i},\mathbf{j}}$ + +$$ +\mathbf {S} _ {\mathbf {i}, \mathbf {j}} = \frac {1}{N} \sum_ {f = 1} ^ {N} \frac {\mathbf {F} _ {\mathbf {i} , \mathbf {j}} \cdot \mathbf {P} _ {\mathbf {f}} ^ {\mathbf {g}}}{\| \mathbf {F} _ {\mathbf {i} , \mathbf {j}} \| \| \mathbf {P} _ {\mathbf {f}} ^ {\mathbf {g}} \|}. \tag {3} +$$ + +By setting an adaptive threshold (see Sec. 3.4) for $\mathbf{S}$ , we could get the coarse mask $\mathbf{M}_{\mathbf{c}} \in \mathbb{R}^{h \times w}$ . + +We then perform local retrieval on local features $\mathbf{P}^1$ to get the finer mask. However, the number of local features is $h\times w$ times larger than that of global features. The retrieval time will also increase by a corresponding factor compared to the global retrieval. To enhance retrieval efficiency, we focus on local features with high similarity to the target, which are often concentrated within a limited set of images. By selecting local features from the top-K (note that top-K is different from the previously mentioned top-N) similar images, we construct a streamlined sub-library for retrieval. This approach significantly optimizes the retrieval process. + +Given the coarse mask $\mathbf{M}_{\mathbf{c}} \in \mathbb{R}^{h \times w}$ , we could get the mask average pooling feature for $\mathbf{F}$ + +$$ +\mathbf {F} _ {\mathrm {m a p}} = \frac {\sum_ {i , j} \mathbf {M} _ {\mathrm {c}} \odot \mathbf {F}}{\sum_ {i , j} \mathbf {M} _ {\mathrm {c}}}. \tag {4} +$$ + +We calculate the similarity between $\mathbf{F}_{\mathrm{map}}$ and all global features and retrieve the top-K images to form the sublibrary. A similar retrieval is conducted for the sub-library. We then get the fine mask $\mathbf{M_f} \in \mathbb{R}^{h \times w}$ for local retrieval. + +# 3.3. Self-Retrieval + +The distribution gap between the camouflaged and generated images may result in some pixels being misrecognized. To mitigate this, we propose Self-Retrieval (SR), where the embedding $\mathbf{F}$ itself acts as the prototype. Given the mask $\mathbf{M}_1 = \frac{\mathbf{M}_{\mathrm{c}} + \mathbf{M}_{\mathrm{f}}}{2}$ from the G2L stage, we obtain the background and foreground prototypes as + +$$ +\mathbf {P} ^ {\mathbf {b}} = \frac {\sum_ {i , j} \mathbf {M} _ {\mathbf {1}} \odot \mathbf {F}}{\sum_ {i , j} \mathbf {M} _ {\mathbf {1}}}, \tag {5} +$$ + +$$ +\mathbf {P} ^ {\mathbf {f}} = \frac {\sum_ {i , j} (1 - \mathbf {M} _ {\mathbf {1}}) \odot \mathbf {F}}{\sum_ {i , j} (1 - \mathbf {M} _ {\mathbf {1}})}. \tag {6} +$$ + +It is notable that while $\mathbf{M_1}$ may contain noise, mask averaging pooling mitigates this issue and enables prototypes to roughly characterize the foreground and background. + +For each $\mathbf{F}_{\mathrm{i},\mathrm{j}}$ , we compute its cosine similarity to the background and foreground prototypes and characterize the environment similarity using the difference between them. + +$$ +\mathbf {S} _ {\mathrm {i}, \mathrm {j}} ^ {\mathrm {b}} = \frac {\mathbf {F} _ {\mathrm {i} , \mathrm {j}} \cdot \mathbf {P} ^ {\mathrm {b}}}{\| \mathbf {F} _ {\mathrm {i} , \mathrm {j}} \| \| \mathbf {P} ^ {\mathrm {b}} \|}, \tag {7} +$$ + +$$ +\mathbf {S} _ {\mathrm {i}, \mathrm {j}} ^ {\mathrm {f}} = \frac {\mathbf {F} _ {\mathrm {i} , \mathrm {j}} \cdot \mathbf {P} ^ {\mathrm {f}}}{\| \mathbf {F} _ {\mathrm {i} , \mathrm {j}} \| \| \mathbf {P} ^ {\mathrm {f}} \|}, \tag {8} +$$ + +$$ +\mathbf {S} _ {\mathrm {i}, \mathrm {j}} ^ {\mathrm {s}} = \mathbf {S} _ {\mathrm {i}, \mathrm {j}} ^ {\mathrm {b}} - \mathbf {S} _ {\mathrm {i}, \mathrm {j}} ^ {\mathrm {f}}. \tag {9} +$$ + +After normalizing $\mathbf{S}^{\mathrm{s}}$ , we apply KDE-AT again to obtain the environment mask $\mathbf{M}_2 \in \mathbb{R}^{h \times w}$ for the SR stage. + +The final mask could be derived by reversing the environment mask, + +$$ +\mathbf {M} = 1 - \frac {\mathbf {M} _ {1} + \mathbf {M} _ {2}}{2}. \tag {10} +$$ + +# 3.4. Kernel Density Estimation-based Adaptive Threshold + +In this section, we introduce the aforementioned Kernel Density Estimation-based Adaptive Threshold (KDE-AT). We propose KDE-AT to provide an adaptive threshold for each similarity distribution instead of setting a fixed threshold for all distributions via hyperparameter search. + +Taking the similarity $\mathbf{S} \in \mathbb{R}^{h \times w}$ in G2L stage for example, we first reshape it to $\mathbb{R}^{hw}$ . The kernel density estimator could be formulated as + +$$ +\operatorname {K D E} (x) = \frac {1}{h w} \sum_ {i = 1} ^ {h w} K _ {b} \left(x - x _ {i}\right) = \frac {1}{h w b} \sum_ {i = 1} ^ {h w} K \left(\frac {x - x _ {i}}{b}\right), \tag {11} +$$ + +
MethodFeat. Extra.CHAMELEONCAMOCOD10KNC4K
\( S_{\alpha} \uparrow \)\( E_{\phi} \uparrow \)\( F_{\beta}^{\omega} \uparrow \)\( M \downarrow \)\( S_{\alpha} \uparrow \)\( E_{\phi} \uparrow \)\( F_{\beta}^{\omega} \uparrow \)\( M \downarrow \)\( S_{\alpha} \uparrow \)\( E_{\phi} \uparrow \)\( F_{\beta}^{\omega} \uparrow \)\( M \downarrow \)\( S _{ \alpha } \uparrow \)\( E_{\phi} \uparrow \)\( F_{\beta}^{\omega} \uparrow \)\( M \downarrow \)
FreeSOLODenseCL-ResNet1010.6590.7300.5030.1010.5340.6000.3330.1770.6300.7250.4210.0930.6110.7810.4410.128
DiffCutSD-XL-SSD-1B0.5370.5920.3740.2450.6300.6830.4910.1660.5910.6280.3790.1600.6710.7270.5310.134
LOSTDINO-ViT-S/160.5810.7320.4060.1410.5730.7010.4310.1730.6310.7420.4380.0950.6340.7550.5050.128
SpectralDINO-ViT-S/80.6220.6780.4710.1780.6030.6730.4870.2100.5880.6140.3910.1690.6790.7330.5590.143
MaskCutDINO-ViT-B/80.6160.6530.4730.1910.5930.6400.4560.2180.5940.6160.3930.1890.6550.6930.5240.178
MaskCutDINOv2-ViT-L/140.6870.7190.5310.1150.6300.6580.4940.1940.5740.5720.3520.1900.6460.6680.5030.177
TokenCutDINO-ViT-S/160.6520.7420.5080.1340.6410.7140.5060.1600.6610.7390.4740.1000.7210.8020.6080.101
TokenCutDINOv2-ViT-L/140.6820.7050.4940.0880.6200.6550.4390.1440.6190.6490.3730.0880.6640.6990.4920.108
FOUNDDINO-ViT-S/80.5520.6020.3810.2180.5260.5780.3970.2770.4880.4870.2640.2480.5690.6050.4100.218
FOUNDDINOv2-ViT-L/140.6590.7450.4900.1040.5580.6220.3730.1720.6050.6750.3860.1000.6210.6930.4550.131
VoteCutDINOv2-ViT-B/140.7070.7480.5670.1030.6330.6930.4950.1630.6410.7060.4390.1040.6880.7510.5550.114
VoteCutDINOv2-ViT-L/140.6570.6770.4740.1120.5260.5220.3030.1960.6190.6460.3890.1030.6090.6230.4120.140
VoteCutDINO(v2) Ensemble0.6740.7350.5460.1450.6370.7020.5220.1700.6900.7820.5340.0920.7390.8180.6490.097
EASEDINO-ViT-S/80.6760.7650.5500.1050.6530.7370.5630.1660.6730.7320.5140.1090.7280.7900.6330.108
EASEDINOv2-ViT-S/140.7540.8480.6510.0740.6920.7630.6040.1380.7200.8040.5820.0790.7520.8320.6610.085
EASEDINOv2-ViT-L/140.8190.8990.7410.0440.7490.8310.6840.0980.7730.8660.6560.0400.8000.8840.7350.056
+ +Table 1. Quantitative comparisons of unsupervised segmentation with eight unsupervised methods on four commonly used COD datasets. “↑/↓”: the higher/lower the better. The best and second-best results are bolded and underlined to highlight. + +where $K$ denotes the kernel function and $b$ the bandwidth. In this paper, we adopt the gaussian kernel. So we have + +$$ +\mathrm {K D E} (x) = \frac {1}{\sqrt {2 \pi}} \frac {1}{h w b \sigma} \sum_ {i = 1} ^ {h w} e ^ {- \frac {\left(x - x _ {i}\right) ^ {2}}{2 b ^ {2} \sigma^ {2}}}. \tag {12} +$$ + +$\sigma$ denotes the standard deviation. For the bandwidth, we follow Scott's Rule [62] and set it as + +$$ +b = (h w) ^ {- \frac {1}{d + 4}}, \tag {13} +$$ + +where $d$ denotes the data dimension. + +As shown in Fig. 1, the probability density estimation of S has a bimodal-like shape, so we adopt the inter-peak minima as the threshold to isolate the environment from the camouflaged objects. For more visualizations, please see Supplementary Fig. 1. + +# 4. Experiment + +# 4.1. Experimental Settings + +Datasets and Evaluation Metrics. We evaluate our method on four commonly used benchmarks, including CHAMELEON (76 test images) [66], CAMO (250 test images) [42], COD10K (2,026 test images) [15] and NC4K (4,121 test images) [51]. Following previous works, we adopt structure measure $(S_{\alpha})$ [13], mean E-measure $(E_{\phi})$ [14], weighted F-measure $(F_{\beta}^{\omega})$ [52] and mean absolute error $(M)$ [59] for evaluation. + +Implementation Details. LLaVA-1.5-7b [50] and Stable Diffusion V1-5 [61] are adopted as the LMM and diffusion model, respectively. We choose DINOv2-ViT-L14 [55] as the self-supervised model and images are resized to $476 \times 476$ . We generate $l = 32$ images for each environment category. For the top-N prototypes and top-K images during retrieval, we set $N = 1024$ and $K = 64$ , respectively. Experiments are conducted on an L20. The FAISS library [35] is leveraged for efficient retrieval. + +# 4.2. Comparison with the SOTAs + +Comparison Methods. We conduct comparison experiments in two settings: unsupervised segmentation and prompt-based segmentation. For unsupervised segmentation, we compare EASE with six DINO-based methods, including LOST [64], TokenCut [73], Spectral [54], MaskCut [72], FOUND [65] and VoteCut [1], one DenseCL-based method FreeSOLO [71] and a diffusion-based method DiffCut [10]. We adopt the official codes to implement these methods for COD. For DINO-based methods, following CuVLER [1], images are resized to $480 \times 480$ for DINO [6] and $476 \times 476$ for DINOv2 [55]. Post-processing, such as Conditional Random Field (CRF) [40], is discarded for all methods. For prompt-based segmentation, we extract bounding boxes from masks and prompt SAM [39] or HQSAM [37] to implement segmentation. We compare our method with four prompt-based methods (MMCPF [70], GenSAM [28], WS-SAM [23] and ProMaC [29]). We also provide seven SOTA fully-supervised methods (FDNet [82], ZoomNet [56], SegMaR [33], HitNet [30], PopNet [76], FSPNet [32] and FEDER [22]) for reference. + +Quantitative comparisons. For unsupervised segmentation, as shown in Tab. 1, our method essentially outperforms previous methods under different configurations. Especially, the best results of our method achieve a $\sim 10\%$ lead on all metrics across all datasets compared to the previous SOTAs. As indicated in Tab. 2, for prompt-based segmentation, EASE incorporated with SAM substantially outperforms its SAM-based counterparts. On the more challenging COD10K and NC4K datasets, our method even compares favorably with fully-supervised methods. These results demonstrate the effectiveness of our concept of "separating the environment". + +Qualitative comparisons. As shown in Fig. 3, compared to other methods, EASE enables better localization and segmentation of objects, especially for those small and imper + +
MethodSAMCHAMELEONCAMOCOD10KNC4K
\(S_{\alpha}\)↑\(E_{\phi}\)↑\(F_{\beta}^{\omega}\)↑M↓\(S_{\alpha}\)↑\(E_{\phi}\)↑\(F_{\beta}^{\omega}\)↑M↓\(S_{\alpha}\)↑\(E_{\phi}\)↑\(F_{\beta}^{\omega}\)↑M↓\(S_{\alpha}\)↑\(E_{\phi}\)↑\(F_{\beta}^{\omega}\)↑M↓
Fully-Supervised
FDNet-0.8950.9510.8490.0270.8400.8960.7820.0630.8380.9210.7470.0300.8320.8950.7590.052
ZoomNet-0.9020.9430.8450.0230.8200.8770.7520.0660.8380.8880.7290.0290.8530.8960.7840.043
SegMaR-0.9060.9510.8600.0250.8150.8740.7530.0710.8330.8990.7240.0340.8410.8960.7810.046
HitNet-0.9210.9670.8970.0190.8490.9060.8090.0550.8710.9350.8060.0230.8750.9260.8340.037
PopNet-0.9170.9650.8750.0200.8080.8590.7440.0770.8510.9100.7570.0280.8610.9090.8020.042
FSPNet-0.9080.9430.8510.0230.8560.8990.7990.0500.8510.8950.7350.0260.8790.9150.8160.035
FEDER-0.8870.9460.8340.0300.8020.8670.7380.0710.8220.9000.7160.0320.8470.9070.7890.044
Prompt-based Segmentation
WS-SAM-PSAM-ViT-H0.8050.8680.7000.0560.7180.7570.6020.1020.7910.8560.6630.0390.8130.8590.7340.057
WS-SAM-SSAM-ViT-H0.8200.8870.7230.0480.7590.8140.6670.0920.8030.8770.6800.0380.8290.8860.7570.052
MMCPFHQSAM-ViT-H----0.7490.8200.6800.1010.7330.8030.5920.0660.7670.8260.6810.083
GenSAMSAM-ViT-H0.7670.8270.6730.0750.7380.8030.6740.1060.7730.8320.6670.0650.8100.8660.7510.065
ProMaC*SAM-ViT-H0.8330.899-0.0440.7670.846-0.0900.8050.876-0.042----
EASESAM-ViT-H0.8530.9120.8120.0440.7770.8410.7480.0990.8560.9140.8010.0280.8490.9020.8150.051
EASEHQSAM-ViT-H0.8640.9160.8270.0370.8070.8650.7710.0780.8660.9180.8110.0230.8660.9150.8330.039
+ +Table 2. Quantitative comparisons of prompt-based segmentation. “-”: Not available. “*”: Results from the paper. The best and second-best results are bolded and underlined to highlight. + +![](images/ebc277e198de4507fce754eccec359d372388ef9795137da20a6817a09111797.jpg) + +![](images/ab4ca9678e2c0f82242f5f777dc950483573d49b6d7a178e548fa10edba447c3.jpg) + +![](images/0b29a83f52552f9c4bf9a8a140a1d50f90cc8d6b753cf63afdedd695dd0e602e.jpg) + +![](images/808b4b18538fa936f4c7727b5fccdff0649087bd5dd1d8793a056bbc7be3be71.jpg) + +![](images/314e2cc161cd32db3340f8e2c08adc23dc51c5f1c8589c94b0a500a4a55a966d.jpg) + +![](images/df9a4284526265daf4d2d964ce0e5baba7bb26e622e2a8826b2b2090c10e04b3.jpg) + +![](images/ea7eceb0c10fb5827799d227b442a647be43c15123a4561b2994fe590fab2f4a.jpg) + +![](images/371f3a439329ea5a665b09d99321fc93f080ac9d369cf8eeb5bcd982c1aad8cb.jpg) + +![](images/17c9b63d1b86fae5a6fd5fd855da87ecbf1b40399d94f6f846f3dec3a9be462f.jpg) + +![](images/dbad78a04062abc00216847061cf26ac739c872162d846b02bb64c2ce86edd61.jpg) + +![](images/b4b9ea579913795b52618486c9d608112934f40b7e4b418ab5be68feaac7d7f3.jpg) +Image + +![](images/6a10b363882f6e2ad238686872f7d5d7641e6b467749bcf95ff99d36458c1683.jpg) +GT + +![](images/b54cbd04dd508ac2a2f52daaf31293d9a55e2d7aa0d295bfed2e7e3d459a05b4.jpg) +EASE + +![](images/2f2e5ab1f338815a176fa9f3c9a64aa481de7a39deaec130c56944df020a4e24.jpg) +LOST + +![](images/f73b8ef1bc47a84bc720886c1536c5185bf5437b5ad1a60f9b306818ebd58a0c.jpg) +Found + +![](images/d4d4d105ac32c6c13f418f183d7b120685f58b19f0414a7c5bed21bbfde9d9e9.jpg) +VoteCut +Figure 3. Qualitative comparisons with unsupervised methods. + +![](images/ac6d55b7d76ac9e33348a79aad267e119cdd53bd3d31e129da84aeff84350b82.jpg) +FreeSOLO + +![](images/caadce23e9bc6b45220778ec683e908b6fd55d9b27a6ba5c13c5feb56174ee19.jpg) +TokenCut + +![](images/45cb3e71339c13e05b328b0cc6ebac494df4d190e69e7bde82de02bb213906c9.jpg) +Spectral + +![](images/c5c4624b39d9ec964e042cc7df77d41236d4e20e8e485626fef1b1606fc39f2e.jpg) +MaskCut + +ceptible ones. + +# 4.3. Ablation Study + +Effectiveness of DiffPro. Instead of obtaining environment prototypes from the Web or existing annotated datasets, our proposed DiffPro leverages a series of foundation models to automatically craft high-quality, environment-only prototypes, avoiding the time-consuming and laborious web collection or utilizing annotated data. To validate the effectiveness of DiffPro, we compare it with two methods: one (Web) crawling images from the Web following keywords, and the other (GT) using the COD training sets directly, i.e., the training sets of COD10K and CAMO, to obtain prototypes by masking the foreground. As indicated by Tab. 3, DiffPro keeps on par or outperforms these two methods. + +Effectiveness of Retrieval Schemes. We set the baseline to be the global retrieval and its optimal threshold is determined by hyperparameter search with an interval of 0.1. Our KDE-AT assigns a distribution-related threshold to each image, avoiding additional hyperparameters while achieving a lead of $\sim 1\%$ , as illustrated by rows a and b of Tab. 4. While global retrieval portrays the profile of the environment, it may get confused when dealing with local details. Our proposed G2L remedies this and delineates ob + +jects and regions at a finer level, bringing about a performance gain, as indicated in rows b and c. Finally, considering that there may be prototypes that are similar to both the environment and the object, SR retrieves the features themselves as prototypes, further stretching the similarity distribution to achieve the best results (row d). + +# 4.4. Further Analysis + +Sensitivity of Hyperparameters. We conduct hyperparameter sensitivity analysis on CAMO, CHAMELEON, and COD10K. For simplicity, we adopt score $= S_{\alpha} + E_{\phi} + F_{\beta}^{\omega}$ to indicate the overall performance. As shown in Fig. 4, EASE is robust to different hyper-parameter settings. We believe this stems from the fact that we discard the most influential threshold hyperparameter as well as the design of multifaceted retrieval schemes. + +Different Foundation Models. For feature extractors, we experiment DINO [6], DINOv2 [55], MoCoV3 [9], MAE [59], and generative models [18, 61]. For diffusion models, we conduct experiments on SD-V1-5 [61], SD-V2-1 [61], SD-XL [60], and SD-3.5-L-Turbo [12]. For SAM, we adopt HQ-SAM [37] and the vanilla SAM [39]. Our experiments show that: a). DINOv2 outperforms other self-supervised models; b). Different diffusion models per + +
MethodCHAMELEONCAMOCOD10KNC4K
\( S_{\alpha} \uparrow \)\( E_{\phi} \uparrow \)\( F_{\beta}^{\omega} \uparrow \)M↓\( S_{\alpha} \uparrow \)\( E_{\phi} \uparrow \)\( F_{\beta}^{\omega} \uparrow \)M↓\( S_{\alpha} \uparrow \)\( E_{\phi} \uparrow \)\( F_{\beta}^{\omega} \uparrow \)M↓\( S_{\alpha} \uparrow \)Eφ↑\( F_{\beta}^{\omega} \uparrow \)M↓
GT0.8160.8930.7400.0420.7490.8220.6830.1090.7590.8460.6410.0580.7930.8730.7240.066
Web0.8210.8950.7410.0430.7470.8220.6790.1040.7700.8620.6540.0440.8060.8860.7410.055
DiffPro0.8190.8990.7410.0440.7490.8310.6840.0980.7730.8660.6560.0400.8000.8840.7350.056
+ +Table 3. Ablation experiments of DiffPro. + +
indexbaselineKDE-ATG2LSRCHAMELEONCAMOCOD10KNC4K
\( S_{\alpha} \uparrow \)\( E_{\phi} \uparrow \)\( F_{\beta}^{\omega} \uparrow \)M↓\( S_{\alpha} \uparrow \)\( E_{\phi} \uparrow \)\( F_{\beta}^{\omega} \uparrow \)M↓\( S_{\alpha} \uparrow \)\( E_{\phi} \uparrow \)\( F_{\beta}^{\omega} \uparrow \)M↓\( S_{\alpha} \uparrow \)Eφ↑\( F_{\beta}^{\omega} \uparrow \)M↓
a✓xxx0.7930.8770.7040.0560.7200.8080.6380.0990.7370.8300.6000.0480.7490.8320.6660.074
b✓✓xx0.8060.8910.7220.0480.7290.8070.6480.0960.7490.8380.6150.0450.7630.8420.6790.067
c✓✓✓x0.8120.8940.7290.0470.7410.8260.6720.1030.7630.8570.6390.0430.7880.8750.7160.062
d✓✓✓✓0.8190.8990.7410.0440.7490.8310.6840.0980.7730.8660.6560.0400.8000.8840.7350.056
+ +![](images/e8ea9b75bb5ea00a628b044af742f1af82d502469139e57ab4d729368dd3007c.jpg) +(a) + +![](images/6167a09dd737b5ee53344221b250ac695c11610e6721719fef43b752b44235be.jpg) +(b) +Figure 4. Hyperparameter sensitivity analysis. (a) Number of generated images for each environment category; (b) Average similarity of Top-N prototypes; (c)Top-K environment images for local retrieval. + +![](images/dfd74140dea26cdc8956d9f84f161300b89369322e2fe08d8787ecc0b1365fb7.jpg) +(c) + +Table 4. Ablation experiments of retrieval schemes. + +
MethodMAS3KRMAS500
\(S_{\alpha}\)↑\(E_{\phi}\)↑\(F_{\beta}^{\omega}\)↑M↓\(S_{\alpha}\)↑\(E_{\phi}\)↑\(F_{\beta}^{\omega}\)↑M↓
LOST0.6460.7530.4900.1070.5660.6720.3580.132
FreeSOLO0.6130.6930.4030.1110.5520.6700.2830.102
TokenCut0.7140.7900.5700.0890.5750.6550.3660.152
Spectral0.6530.6870.4970.1430.5480.5710.3540.212
MaskCut0.6840.7140.5370.1450.5430.5630.3470.231
FOUND0.5570.5800.3700.2050.4830.4910.2610.253
DiffCut0.5130.6280.3140.2180.4360.5370.2000.265
VoteCut0.6330.6850.4500.1270.5500.5670.2880.140
EASE0.7800.8680.6820.0480.7200.8220.5850.049
+ +Table 5. Quantitative comparisons on marine animal segmentation. + +form comparably; c). HQ-SAM outperforms SAM, and the segmentation performance does not tend to saturate as the model parameters increase. Please see Supplementary Tab. 1, 2, and 3 for more details. + +Different Layers. For the best-performing feature extractor DINOv2, we experiment with features from different layers. Our experiments show that features from the last layer capture the most valuable information. Please refer to Sup + +# plementary Fig. 2, Tab. 4 and 5. + +Generalization. We apply our method to marine animal segmentation [80], which is also quite challenging due to varying environmental conditions [81]. Two commonly used datasets, MAS3K [46] (1,141 test images) and RMAS [17] (500 test images), are adopted for evaluation. As indicated by Tab. 5, our approach delivers consistent performance advancements, which further demonstrates the effectiveness of EASE in isolating the environment. + +# 5. Conclusion + +In this paper, unlike previous COD work that focuses on mining camouflaged objects from the environment, we shift the paradigm by removing the salient environment and thus obtaining the target. We isolate environments and highlight camouflaged objects by retrieving from a library of environment prototypes. This is accomplished by the "prototype production line" DiffPro and multifaceted retrieval schemes, including kernel density estimation-based adaptive threshold, global-to-local retrieval and inter-feature similarity-based self-retrieval. We conduct extensive experiments to validate the effectiveness of our method. + +Acknowledgments. This work was supported in part by The Hong Kong Polytechnic University under Grants P0048387, P0044520, P0050657, and P0049586, and in part by the Tianjin Science and Technology Major Project under Grant 24ZXZSSS00420. + +# References + +[1] Shahaf Arica, Or Rubin, Sapir Gershov, and Shlomi Laufer. CuVLER: Enhanced unsupervised object discoveries through exhaustive self-supervised transformers. In IEEE Conference on Computer Vision and Pattern Recognition, pages 23105-23114, 2024. 3, 6 +[2] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. OpenFlamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, pages 1-20, 2023. 4 +[3] Luca Barsellotti, Roberto Amoroso, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Training-free open-vocabulary segmentation with offline diffusion-augmented prototype generation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3689-3698, 2024. 3 +[4] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego De Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack Rae, Erichelsen, and Laurent Sifre. Improving language models by retrieving from trillions of tokens. In International Conference on Machine Learning, pages 2206-2240, 2022. 3 +[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and stuff classes in context. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1209–1218, 2018. 4 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In IEEE International Conference on Computer Vision, pages 9650-9660, 2021. 3, 5, 6, 7 +[7] Chunyuan Chen, Weiyun Liang, Donglin Wang, Bin Wang, and Jing Xu. Vision-inspired boundary perception network for lightweight camouflaged object detection. IEEE Signal Processing Letters, pages 1-5, 2025. 2 +[8] Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. Just a hint: Point-supervised camouflaged object detection. In European Conference on Computer Vision, pages 332-348, 2025. 1 +[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In IEEE International Conference on Computer Vision, pages 9640-9649, 2021. 7 + +[10] Paul Couairon, Mustafa Shukor, Jean-Emmanuel Haugeard, Matthieu Cord, and Nicolas Thome. Diffcut: Catalyzing zero-shot semantic segmentation with diffusion features and recursive normalized cut. In Advances in Neural Information Processing Systems, pages 1–24, 2024. 3, 6 +[11] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, pages 49250-49267, 2023. 4 +[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 1-28, 2024. 7 +[13] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-Measure: A new way to evaluate foreground maps. In IEEE International Conference on Computer Vision, pages 4558-4567, 2017. 6 +[14] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In International Joint Conference on Artificial Intelligence, pages 698-704, 2018. 6 +[15] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2774-2784, 2020. 1, 2, 3, 6 +[16] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6024-6042, 2022. 1, 2 +[17] Zhenqi Fu, Ruizhe Chen, Yue Huang, En Cheng, Xinghao Ding, and Kai-Kuang Ma. MASNet: A robust deep marine animal segmentation network. IEEE Journal of Oceanic Engineering, 49(3):1104-1115, 2024. 8 +[18] Yatharth Gupta, Vishnu V. Jaddipal, Harish Prabhala, Sayak Paul, and Patrick Von Platen. Progressive knowledge distillation of stable diffusion x1 using layer level loss. arXiv preprint arXiv:2401.02677, pages 1-9, 2024. 7 +[19] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. Retrieval augmented language model pretraining. In International Conference on Machine Learning, pages 3929-3938, 2020. 3 +[20] Chao Hao, Zitong Yu, Xin Liu, Yuhao Wang, Weicheng Xie, Jingang Shi, Huanjing Yue, and Jingyu Yang. Distribution-specific learning for joint salient and camouflaged object detection. Available at SSRN 5089840, pages 1-33, 2024. 2 +[21] Chao Hao, Zitong Yu, Xin Liu, Jun Xu, Huanjing Yue, and Jingyu Yang. A simple yet effective network based on vision transformer for camouflaged object and salient object detection. IEEE Transactions on Image Processing, 34:608-622, 2025. 2 +[22] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object + +detection with feature decomposition and edge reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, pages 22046-22055, 2023. 6 +[23] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with SAM-based pseudo labeling and multi-scale feature grouping. In Advances in Neural Information Processing Systems, pages 30726-30737, 2023. 1, 6 +[24] Chunming He, Kai Li, Yachao Zhang, Yulun Zhang, Zhenhua Guo, Xiu Li, Martin Danelljan, and Fisher Yu. Strategic preys make acute predators: Enhancing camouflaged object detectors by generating camouflaged objects. In International Conference on Learning Representations, pages 1-10, 2024. 2 +[25] Chunming He, Rihan Zhang, Fengyang Xiao, Chenyu Fang, Longxiang Tang, Yulun Zhang, Linghe Kong, Deng-Ping Fan, Kai Li, and Sina Farsiu. RUN: Reversible unfolding network for concealed object segmentation. arXiv preprint arXiv:2501.18783, pages 1-14, 2025. 2 +[26] Ruozhen He, Qihua Dong, Jiaying Lin, and Rynson W.H. Lau. Weakly-supervised camouflaged object detection with scribble annotations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 781-789, 2023. 1 +[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 4 +[28] Jian Hu, Jiayi Lin, Shaogang Gong, and Weitong Cai. Relax image-specific prompt requirement in SAM: A single generic prompt for segmenting camouflaged objects. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12511-12518, 2024. 2, 3, 6 +[29] Jian Hu, Jiayi Lin, Junchi Yan, and Shaogang Gong. Leveraging hallucinations to reduce manual prompt dependency in promptable segmentation. In Advances in Neural Information Processing Systems, pages 1-13, 2024. 2, 3, 4, 6 +[30] Xiaobin Hu, Shuo Wang, Xuebin Qin, Hang Dai, Wenqi Ren, Donghao Luo, Ying Tai, and Ling Shao. High-resolution iterative feedback network for camouflaged object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 881-889, 2023. 2, 6 +[31] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. OPERA: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 13418-13427, 2024. 2 +[32] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5557-5566, 2023. 2, 6 +[33] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4703-4712, 2022. 2, 6 + +[34] Yao Jiang, Xinyu Yan, Ge-Peng Ji, Keren Fu, Meijun Sun, Huan Xiong, Deng-Ping Fan, and Fahad Shahbaz Khan. Effectiveness assessment of recent large vision-language models. Visual Intelligence, 2(1):1-17, 2024. 2 +[35] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535-547, 2019. 6 +[36] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion Models for Open-Vocabulary Segmentation. In European Conference on Computer Vision, pages 299-317, 2024. 3 +[37] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In Advances in Neural Information Processing Systems, pages 29914–29934, 2023. 6, 7 +[38] Jooyeon Kim, Eulrang Cho, Sehyung Kim, and Hyunwoo J. Kim. Retrieval-augmented open-vocabulary object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 17427-17436, 2024. 3 +[39] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In IEEE International Conference on Computer Vision, pages 3992-4003, 2023. 2, 6, 7 +[40] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In Advances in Neural Information Processing Systems, pages 1-9, 2011. 6 +[41] Hala Lamdouar, Weidi Xie, and Andrew Zisserman. The making and breaking of camouflage. In IEEE International Conference on Computer Vision, pages 832-842, 2023. 2 +[42] Trung-Nghia Le, Tam V. Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding, 184:45-56, 2019. 6 +[43] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In IEEE Conference on Computer Vision and Pattern Recognition, pages 13872-13882, 2024. 2 +[44] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Information Processing Systems, pages 9459-9474, 2020. 3 +[45] 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, 2023. 4 +[46] Lin Li, Eric Rigall, Junyu Dong, and Geng Chen. MAS3K: An open dataset for marine animal segmentation. In Benchmarking, Measuring, and Optimizing, pages 194-212, 2021. + +[47] Weiyun Liang, Jiesheng Wu, Xinyue Mu, Fangwei Hao, Ji Du, Jing Xu, and Ping Li. Weighted dense semantic aggregation and explicit boundary modeling for camouflaged object detection. IEEE Sensors Journal, 24(13):21108-21122, 2024. 2 +[48] Weiyun Liang, Jiesheng Wu, Yanfeng Wu, Xinyue Mu, and Jing Xu. FINet: Frequency injection network for lightweight camouflaged object detection. IEEE Signal Processing Letters, 31:526-530, 2024. 2 +[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916, 2023. 4 +[50] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In IEEE Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 6 +[51] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11586-11596, 2021. 2, 6 +[52] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2014. 6 +[53] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8768-8777, 2021. 1, 2 +[54] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3, 6 +[55] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024:1-32, 2024. 1, 2, 3, 5, 6, 7 +[56] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2150-2160, 2022. 2, 6 +[57] Shubham Parashar, Zhiqiu Lin, Tian Liu, Xiangjue Dong, Yanan Li, Deva Ramanan, James Caverlee, and Shu Kong. The neglected tails in vision-language models. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12988-12997, 2024. 3 +[58] Bohao PENG, Zhuotao Tian, Shu Liu, Ming-Chang Yang, and Jiaya Jia. Scalable language model with generalized continual learning. In International Conference on Learning Representations, pages 1-23, 2024. 3 + +[59] Federico Perazzi, Philipp Kähenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 733-740, 2012. 6, 7 +[60] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, pages 1-21, 2023. 7 +[61] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4, 6, 7 +[62] David W Scott. Multivariate density estimation: theory, practice, and visualization. John Wiley & Sons, 2015. 6 +[63] Jianbo Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 3 +[64] Oriane Simeoni, Gilles Puy, Huy V. Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In Proceedings of the British Machine Vision Conference, pages 1-25, 2021. 3, 6 +[65] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecký, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3176-3186, 2023. 3, 6 +[66] Przemysław Skurowski, Hassan Abdulameer, J Blaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 6 +[67] Ze Song, Xudong Kang, Xiaohui Wei, Haibo Liu, Renwei Dian, and Shutao Li. FSNet: Focus scanning network for camouflaged object detection. IEEE Transactions on Image Processing, 32:2267-2278, 2023. 2 +[68] Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, and Nian Liu. Context-aware cross-level fusion network for camouflaged object detection. In International Joint Conference on Artificial Intelligence, pages 1025–1031, 2021. 2 +[69] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval. In IEEE Conference on Computer Vision and Pattern Recognition, pages 26951-26962, 2024. 3 +[70] Lv Tang, Peng-Tao Jiang, Zhi-Hao Shen, Hao Zhang, Jin-Wei Chen, and Bo Li. Chain of Visual Perception: Harnessing multimodal large language models for zero-shot camouflaged object detection. In ACM International Conference on Multimedia, page 8805–8814, 2024. 2, 3, 4, 6 +[71] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M. Alvarez. FreeSOLO: Learning to segment objects without annotations. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022. 3, 6 +[72] Xudong Wang, Rohit Girdhar, Stella X. Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance + +segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 3, 6 +[73] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L. Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14543-14553, 2022. 3, 6 +[74] Yuan Wang, Rui Sun, Naisong Luo, Yuwen Pan, and Tianzhu Zhang. Image-to-image matching via foundation models: A new perspective for open-vocabulary semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3952-3963, 2024. 3 +[75] Jiesheng Wu, Weiyun Liang, Fangwei Hao, and Jing Xu. Mask-and-edge co-guided separable network for camouflaged object detection. IEEE Signal Processing Letters, 30: 748-752, 2023. 2 +[76] Zongwei Wu, Danda Pani Paudel, Deng-Ping Fan, Jingjing Wang, Shuo Wang, Cédric Demonceaux, Radu Timofte, and Luc Van Gool. Source-free depth for object pop-out. In IEEE International Conference on Computer Vision, pages 1032-1042, 2023. 1, 2, 6 +[77] Fengyang Xiao, Sujie Hu, Yuqi Shen, Chengyu Fang, Jina Huang, Chunming He, Longxiang Tang, Ziyun Yang, and Xiu Li. A survey of camouflaged object detection and beyond. arXiv preprint arXiv:2408.14562, pages 1-26, 2024. 2 +[78] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. In IEEE Conference on Computer Vision and Pattern Recognition, pages 13525-13536, 2024. 3 +[79] Jin Zhang, Ruiheng Zhang, Yanjiao Shi, Zhe Cao, Nian Liu, and Fahad Shahbaz Khan. Learning camouflaged object detection from noisy pseudo label. In European Conference on Computer Vision, pages 158-174, 2025. 1 +[80] Pingping Zhang, Tianyu Yan, Yang Liu, and Huchuan Lu. Fantastic animals and where to find them: Segment any marine animal with dual SAM. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2578-2587, 2024. 8 +[81] Chen Zhao, Weiling Cai, Chenyu Dong, and Chengwei Hu. Wavelet-based fourier information interaction with frequency diffusion adjustment for underwater image restoration. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8281-8291, 2024. 8 +[82] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4504-4513, 2022. 1, 2, 6 +[83] Tao Zhou, Yi Zhou, Chen Gong, Jian Yang, and Yu Zhang. Feature aggregation and propagation network for camouflaged object detection. IEEE Transactions on Image Processing, 31:7036-7047, 2022. +[84] Hongwei Zhu, Peng Li, Haoran Xie, Xuefeng Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3608-3616, 2022. 2 \ No newline at end of file diff --git a/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/images.zip b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c2dba3ec050272b323e68f59d747a1517d594466 --- /dev/null +++ b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1396e8902ef653064c6012e08eecc14eb2d52e348024dec0033a44d883bedb1 +size 776714 diff --git a/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/layout.json b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ba04cef41687e90050692d2f51de358cf27b8d85 --- /dev/null +++ b/CVPR/2025/Shift the Lens_ Environment-Aware Unsupervised Camouflaged Object Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c852f6b41a83fc7b7079038f2ade0cf0e0cb0ed8b65dd257106449fee027e78d +size 492611 diff --git a/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/a38e867c-5c5d-41e7-90eb-6e678974e0a5_content_list.json b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/a38e867c-5c5d-41e7-90eb-6e678974e0a5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..65adf6ecd975410039b981722af476b18110d933 --- /dev/null +++ b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/a38e867c-5c5d-41e7-90eb-6e678974e0a5_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79c5e865c2cb0765349138ecdbe01ca6526f2af4d51e39744fa09ce6b14801ec +size 82250 diff --git a/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/a38e867c-5c5d-41e7-90eb-6e678974e0a5_model.json b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/a38e867c-5c5d-41e7-90eb-6e678974e0a5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c2344f20ccfd6bd2959794d07a01f6ef3c17d20 --- /dev/null +++ b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/a38e867c-5c5d-41e7-90eb-6e678974e0a5_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:674e8b1bc6b7fdfa7f2f74213dc06594000274b2191eebe7f00db6886b3554da +size 102610 diff --git a/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/a38e867c-5c5d-41e7-90eb-6e678974e0a5_origin.pdf b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/a38e867c-5c5d-41e7-90eb-6e678974e0a5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7ecce78785f0e6d36244ec81e08f9c77f1f1ec9e --- /dev/null +++ b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/a38e867c-5c5d-41e7-90eb-6e678974e0a5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1521c3096a03867b343161f87eae6189fc2548f9ce7ee292e602a6a41e3da6d5 +size 2032731 diff --git a/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/full.md b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/full.md new file mode 100644 index 0000000000000000000000000000000000000000..35f342c96ca43612737012dca608dd1cf8023833 --- /dev/null +++ b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/full.md @@ -0,0 +1,273 @@ +# ShiftwiseConv: Small Convolutional Kernel with Large Kernel Effect + +Dachong Li $^{1}$ Li Li $^{2}$ Zhuangzhuang Chen $^{1}$ Jianqiang Li $^{3*}$ + +1 College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, China + +$^{2}$ School of Mathematical Sciences, Shenzhen University, Shenzhen, China + +$^{3}$ National Engineering Laboratory for Big Data System Computing Technology. + +{lidachong2023, 2250201001, chenzhuangzhuang2016}@email.szu.edu.cn, lijq@szu.edu.cn + +# Abstract + +Large kernels make standard convolutional neural networks (CNNs) great again over transformer architectures in various vision tasks. Nonetheless, recent studies meticulously designed around increasing kernel size have shown diminishing returns or stagnation in performance. Thus, the hidden factors of large kernel convolution that affect model performance remain unexplored. In this paper, we reveal that the key hidden factors of large kernels can be summarized as two separate components: extracting features at a certain granularity and fusing features by multiple pathways. To this end, we leverage the multi-path long-distance sparse dependency relationship to enhance feature utilization via the proposed Shiftwise (SW) convolution operator with a pure CNN architecture. In a wide range of vision tasks such as classification, segmentation, and detection, SW surpasses state-of-the-art transformers and CNN architectures, including SLaK and UniRepLKNet. More importantly, our experiments demonstrate that $3 \times 3$ convolutions can replace large convolutions in existing large kernel CNNs to achieve comparable effects, which may inspire follow-up works. Code and all the models at https://github.com/lidc54/shift-wiseConv. + +# 1. Introduction + +Vision Transformer (ViT) [18, 31, 32, 36] outperforms traditionally dominant CNNs, largely due to its superior capacity for long-range modeling. Meanwhile, CNNs are widely recognized for their resource efficiency and translation invariance. Thus, an intuitive idea is to integrate the strengths + +![](images/d55ff9bcbc09935958e00484d6d86eca38f2d70bcf15360da1f60ab853c683a1.jpg) +Figure 1. The left panel depicts the retina's cellular structure, which is composed of photoreceptor cells and ganglia, with photoreceptors sending visual signals to ganglia via multiple pathways. The right panel presents our Shiftwise (SW) Convolution, consisting of standard convolution and a connection-centric module. The approach utilizes group convolution and reparameterization (Rep) to extract basic information, which is then processed by a shift algorithm to mimic large kernel convolution. + +![](images/2dd1e335ccbec6a7b27a7dccdc2550bf346cd8acad0d7cd196d2ddad3d684924.jpg) + +of ViT to facilitate the evolution of CNNs. To this end, ConvNeXt [38] draws inspiration from ViT and incorporates effective long-distance modeling into its architecture, thereby outperforming ViT-based models in various vision tasks. The most critical insight is to enhance long-distance dependencies through the enlargement of kernel sizes. + +Motivate by the above insight, recent studies like RepLKNet [16], SLaK [35], and UniRepLKNet [17] have significantly advanced kernel size, and made CNNs great again over ViTs. More specifically, RepLKNet [16] expanded convolutional kernels from $7 \times 7$ in ConvNeXt [38] to $31 \times 31$ , which led to performance improvements. SLaK [35] introduced decomposable convolutions, utilizing two distinct strip convolutions of $51 \times 5$ and $5 \times 51$ , further advancing the capabilities of pure CNNs. ParCNetV2 [58] expanded kernel sizes to twice the feature map size, achieving accuracy levels nearly on par with SLaK [35]. In contrast, UniRe + +pLKNet [17] opted for a comparatively smaller $13 \times 13$ kernel as the central component of its model, which significantly enhanced performance. Notably, existing works reveal that there's a point of diminishing returns in precision by simply enlarging the kernel size. Then, one must propose many tailored tricks to avoid diminishing returns in order to potentially enhance accuracy further. + +The above phenomenon indicates that simply enlarging kernel sizes to achieve global attention is not an ideal solution for capturing long-distance dependencies. Meanwhile, large convolution kernels in CNNs have been phased out in favor of smaller ones, with long-range dependencies being established through multi-layer stacking [20, 45]. Thus, the question arises: Can long-distance dependency be effectively modeled through alternative mechanisms rather than just enlarging the convolution kernel, as has been the traditional evolution? Unlike the single-module approach to achieving large receptive fields, the human retina processes visual information through a network of interconnected cells. Specifically, cone and rod cells connect to ganglion cells via bipolar and horizontal cells, which then relay the visual signals to the brain, as depicted in the cellular organisation of the retina [55](Fig. 1). Before reaching the ganglion cells, photoreceptor cells are integrated in a variety of interconnected combinations. Similarly, CNNs can also have decoupled constructions by extracting basic signals at certain granular level and establishing long-distance dependencies through diverse connections. + +However, the employment of large convolutional kernels has become nearly indispensable in modern CNNs. As studies [17, 39] have shown, large kernel convolutions surpass the stacking of small convolutions in effectively expanding the effective receptive field (ERF). Consequently, when redesigning modules with a focus on decoupling long-distance dependencies, they should possess the following traits: i) Efficiency in broadening the receptive field. ii) Basic signal extraction with a fine-grained atomicity. iii) The ability to integrate features through diverse connections. Guided by these, in this paper, we propose Shiftwise (SW) Conv, a plug-and-play module that leverages standard $3 \times 3$ convolutions to achieve the effect of large kernels within a pure CNN framework. To enhance feature utilization, we employ a multi-channel feature shift fusion method. Additionally, the SW module can selectively eliminate certain filters to foster diverse interconnections without altering the overall network architecture. + +From a novel decoupling perspective on long-range dependencies, our proposed method outperforms the performance of large kernel convolution models. Across diverse tasks—including ImageNet classification [14], semantic segmentation on ADE20K [61], object detection/segmentation on COCO [33], and monocular 3D object detection on nuScenes [1]—our approach surpasses state-of-the-art large + +kernel convolutional models, such as Slak [35] and UniRepLKNet [17]. This reinforces the conclusion that small convolutions can effectively supplant large ones within the framework of modern CNNs, aligning with the insights of the VGG [45] model. + +Our contributions can be summarized as three-fold: + +- A new scheme is proposed to effectively replace large convolutional kernels with small ones. +- A simple yet effective plug-and-play pure convolution operator is proposed to enhance CNN's performance. +- Our experimental results demonstrate that decoupling long-range dependencies is an effective strategy for improving CNN performance. + +# 2. Related Work + +# 2.1. Large Convolutional Kernels + +Large convolutional kernels have been central to classic CNNs like AlexNet [29], ResNet [23], and Inception V1 [46]. VGG's advocacy for small kernel stacking influenced model efficiency and development for a long time [25]. The recent resurgence in interest in large kernels, sparked by the ViT's [19] capabilities, has led to a focus on two main areas: the development of innovative CNN-Transformer hybrids [43, 59] and the modernization of CNN structures like ConvNeXts [38]. + +Following ConvNeXts, studies have investigated the correlation between convolutional kernel size and CNN performance. The Visual Attention Network (VAN) [20] uses decomposable convolutions combined with a multilayer stacking strategy to mimic large kernel convolutions, thereby enhancing spatial perception and improving model performance. Works such as RepLKNet [16] and SLaK [35] expanded kernel sizes to $31 \times 31$ and $51 \times 51$ , respectively, setting new performance benchmarks and inspiring our proposed method. ParCNetV2 [58] further increased kernel sizes to twice the input feature map size, achieving performance on par with SLaK. PeLK [4], taking cues from human focal and blur mechanisms, scales CNN kernels to $101 \times 101$ , markedly boosting performance. UniRepLKNet [17] incorporates a smaller $13 \times 13$ dilated convolution with the squeeze-and-excitation (SE) module for enhanced feature fusion, reducing the parameter count while outperforming SLaK. + +# 2.2. Shift Operation + +The Shift operator, as introduced by Wu et al. [56] in 2007, enables the generation of diverse convolutional outputs at a reduced computational cost. Building upon this foundation, Active Shift [27] and Sparse Shift [7] have incorporated learnable parameters to enhance their functionality. ACmix [42], Shift Graph Convolutional Network [11], Xvolution [8] and ShiftViT [51] have all employed the Shift + +operation to expand the receptive field, thereby enabling more comprehensive attention mechanisms. These methods view the shift operator as a form of specialized Depthwise Convolution. Our approach goes beyond these by proposing a method to stack convolutions spatially, effectively emulating the impact of large kernel convolutions. Additionally, we employ a data-driven approach to extract the spatial structure of sparse large kernel convolutions. While AKConv [60] and Dilated Convolution with Learnable Spacings (DCLS) [28] feature offset operations akin to dynamic convolutions. In contrast, our method emphasizes the certain granularity of information extraction rather than pursuing the finer-grained manipulations. + +# 2.3. Ghost and Re-parameterization + +The ghost concept, introduced by GhostNet [22], involves selectively processing features within each module and combining them with the rest for output. This idea has been further developed in networks like CSPNet [50] and FasterNet [5]. Re-parameterization techniques, as seen in RepVGG [15], enhance model complexity during training and simplify during inference for speed. ExpandNets [21], MobileOne [49] and VanillaNet [3] have expanded on this by decomposing weight matrices and adjusting model width and depth. RepLKNet [16], SLaK-net [35] and UniRepLKNet [17] also utilize re-parameterization, showcasing its broad applicability in network design. + +# 3. Architecture Design + +# 3.1. Background + +Aiming to emulate the traditional evolution triggered by VGG—where large-kernel convolutions are replaced by small convolutions stacked in depth—we discovered and validated a method to stack receptive fields in width. SLaK enhances both kernel size and model performance to new levels, thus serving as the foundation for our research. It employs a ResNet variant with $M \times N$ strip convolutions across four stages, where $M$ is [51, 49, 47, 13] and $N$ is consistently 5. The large-kernel convolution uses two parallel strip convolution branches, complemented by a standard small convolution, as shown in Fig. 2 (a). Additionally, SLaK employs a width increase factor "r" to boost model performance. To counterbalance the rise in model parameters, SLaK incorporates a fine-grained sparsity "s" that converts the weights of convolution and Feed Forward Network (FFN) into sparse matrices, with default values of $r = 1.3$ and $s = 0.4$ . To track the scaling trends of large kernel sizes, SLaK models are initially trained for 120 epochs and then extended to 300 epochs to align with state-of-the-art models. We've adopted this training approach in our subsequent experiments. Assuming feature map dimensions are $H \times W$ and input channels are $C$ . + +![](images/3c8858b88dd9efaaec5a1f8e6cf96614ccc367b1e7ceaab7c41f33cccfe1d1c8.jpg) +Figure 2. (a) SLaK's large kernel convolution architecture employs a decomposable separable convolution approach using $M \times N$ and $N \times M$ strip convolutions. (b) One-to-many separable convolution with proper feature movement can be equivalent to SLaK's strip convolution. (c) Overview of the proposed shift operation. The addition operation preserves the network's structural integrity even after coarse-grained pruning. + +In this section, we explore key aspects of our architecture using a structured approach akin to ConvNeXt. We evolve the initial SW module to match the SLaK strip convolution (§3.2), eliminate redundant connections (§3.3), and counter SLaK's width-expanding design (§3.4). Additionally, we investigate multi-path feature fusion (§3.5), localized reparameterization adaptation (§3.6), localized pruning adaptation (§3.7), backbone optimization (§3.8), and overall architectural design (§3.9). Collectively, these efforts aim to merge LK into the mainstream of CNN research. + +# 3.2. Replacement Experiment + +In our initial modification, we replaced the $M \times N$ convolutions employed in SLaK (denoted as $\#0$ ) with a series of standard $N \times N$ convolutions, as illustrated in Fig. 2 (b). This was achieved by employing group convolutions with the number of groups matching the input channels $C$ , and the output channels were set to $\left\lceil \frac{M}{N} \right\rceil$ , where $\lceil \rceil$ denotes the ceiling function. Subsequently, we applied appropriate offsets to the output features based on their sequence numbers to ensure that the outputs align perfectly with the offset positions of the original $M \times N$ convolution when laid out (Fig. 2 (c)). For this replacement experiment (the laid out version labeled as $\#1$ for infer&train), we should adjusted the padding ( $\delta_p = P_{real} - N // 2$ ) and offset of the outputs of standard convolutions to account for the discrepancies in initial and final sliding window positions due to the different kernel sizes. The padding may cause a slight adjustment in the parameters, resulting in the height $H'$ being $H + \delta_p$ , and the width $W'$ similarly. Employing SLaK-trained model parameters, this spatial stacking approach maintained accuracy, achieving $82.5\%$ during inference. + +We proceeded to train the modified version of the model. Instead of employing the fine-grained pruning method utilized by SLaK, we opted for a coarse-grained pruning strategy that remove filters during training. Because based on the existing general hardware, the coarse-grained pruning can reduce a lot of computation. This was achieved by mod + +Table 1. Kernel Replacement Experiment. Experiments on progressively replacing large kernels with small ones based on SLaK. $HW$ represents the feature map dimensions, with $C$ indicating the number of channels. Different padding configurations result in variations in parameter calculations, denoted as $H'W'$ and $H''W''$ . The ceiling function is indicated by $\sqcap$ . Experiments were conducted for 120 epochs to outline the scaling trends while reducing the training load. When the kernel size is modified, denoted as $N_*$ , it implies $N = *$ . Parameter count calculations mainly focus on two strip convolutions here. + +
MethodIN-1K acc. (300epoch)IN-1K acc. (120epoch)#param.
#0Slak tiny82.581.6HWMN × 2
#1laid out(infer)82.5-H'W' [M/N] NN × 2
#1laid out(train)82.27-As above
#2SW-(pad=N-1)82.3881.36H''W'' [M/N] NN
#3SW-(pad=N/2)-81.26HW [M/N] NN
#4Rep*2-81.52As above
#5Relocate BN-81.65#4 + (C × 2) × 3
#6G=0.23-81.34#5 × (1 - G)
#7N=5 to N=3-81.44(HW [M/N3] N3N3) × (1 - G)
#8Rep*4-81.60As above
+ +ifying the pruning method used in SLaK to a coarse-grained approach, where we summed the absolute values of all parameters within a filter and pruned those with lower rankings. For more information, see the parameter prune-and-grow strategy as detailed in [34, 35]. Given that the modified version has multiple output channels per input channel, it is unlikely that all output channels will be pruned. This ensures that the final structure of the module remains intact, which is typically an attribute of fine-grained pruning. The sparse version, labeled as #1, reached an accuracy of $82.27\%$ at 300 epochs, showing minimal decline. + +# 3.3. Eliminate Redundant Connections + +Based on this, we explored new configurations: i) Recognizing the similarity between the two decomposing branches originating from SLaK's strip convolution, we restructured them to utilize a shared convolution output. To avoid correlation, the second branch was modified to employ offsets in reverse order. As illustrated in Fig. 3 (a). ii) The padding, influenced by diverse feature sizes and large kernels size, was deemed suboptimal for network design. To address this, we standardized the padding values. Implementing these configurations, we set the padding values to $N - 1$ (labeled as #2) and $N // 2$ (labeled as #3, default). In this scenario, the total number of parameters is halved compared to SLaK's, with $H'' = H + ((N - 1) - \left\lceil \frac{N}{2} \right\rceil)$ . After 120 epoch training, the accuracy of the model is almost unchanged, which is $81.36\%$ and $81.26\%$ , respectively. + +To counteract the performance decline, we implemented the popular re-parameterization technique (Rep). We integrated two Rep branches and applied a shared mask to both within our module. This approach harnesses the advantages of both Rep and sparsity (labeled as #4). The model achieved an accuracy of $81.52\%$ , which is almost close to + +![](images/49af2e6da8ec350f256616fc906e1caf2e611d20e2699c5af3a496eafc53e470.jpg) +Figure 3. (a) Fig. 2 shows that the similar structures of the convolutional branches lead to minimal variation across the data manifold. This implies that the branches can be consolidated. The shift operation introduces more divergence in the data manifold, necessitating the repositioning of the Batch Normalization (BN) layer to follow the shift operation for optimal performance. (b) Reducing network width with a ghost-like approach to counterbalance SLaK's expansion. + +the accuracy of SLaK at 120 epoch. In SLaK's strip convolution, each convolution is followed by a batch normalization (BN) for each branch. However, the data manifolds from multiple Rep branches tend to be similar, resulting in a low degree of difference, whereas the outputs from SW branches exhibit a high degree of difference due to their varied laid out directions. Therefore, we removed the BN after the convolution and added one after the SW branches (labeled as #5). Because of the BN layer, Experiment #5 has $3C \times 2$ more parameters in the large kernel convolution branch than Experiment #4. From experiment #0 to #5, we have successfully used small convolution over SLaK large kernel convolution after such a series of changes. At the same time, the number of parameters involved in the construction of large kernel convolution is almost halved. These experiments also demonstrate the efficacy of decoupling long-distance dependency modeling into two distinct parts: feature extraction at a fundamental atomic granularity and long-distance modeling that emphasizes connectivity as the primary characteristic. + +# 3.4. Reduce Parameters + +We must admit that, like the $M \times N$ large convolution kernel used by SLaK, only the convolution center point can slide the whole map during the sliding window process. For other $N \times N$ windows, such as the small convolution at the end side of the $M \times N$ large one, a considerable proportion of the sliding positions are in the padding 0 value region. This is also an important place for programming optimization of large kernel convolution operator. At the same time, expanding the network width is essential for SLaK's final performance, a feature that our current structure #5 retains. We must reduce parameters to mitigate the impact of network + +expansion and the computational load from adjusting kernels. We use a slimming technique that allows only a subset of channels to contribute to the large convolution kernel. This is achieved through a ghost-like approach, setting a ratio $G$ that specifies the fraction of channels to pass directly to the subsequent layer. By slimming the large kernel, we can decrease the number of parameters while keeping the input channels of the large kernel constant with expansion ratio $R$ , ensuring that $G$ satisfies the equation $R(1 - G) = 1$ . For example, if SLaK uses a width ratio $R$ of 1.3, we set $G$ to 0.23 to offset the effects of increased width on large convolutions (#6), as shown in Fig. 3 (b). With this adjustment, we further reduce the convolution kernel size to a smaller $3 \times 3$ convolution (#7). + +Compare models #6 and #7, which, while initially aimed at reducing the total parameter volume with minimal performance impact, surprisingly benefited from selecting a smaller kernel size. This indicates that an appropriate level of granularity is beneficial to long-distance dependence. Additionally, we tested increasing the number of parallel Rep blocks (#8), which resulted in an accuracy of $81.60\%$ , matching SLaK's performance. In the remainder of this paper, we set $N = 3$ as default. + +# 3.5. Enhance Utilization + +When using SW operation on group convolution, we found that the utilization ratio of feature maps is low. Specifically, we label the feature maps processed by the three branches of SW to obtain an image of the coverage distribution. As shown in Fig. 4 (a), for specific feature maps, the coverage is predictable and the utilization ratio is limited. If we refer to the three branches of SW as one edge, denoted as $E$ , as the edges increase, the given order of combined feature maps does not improve the coverage. When we choose to combine feature maps in a random order for each edge, the coverage is greatly improved as the number of edges increases, as shown in Fig. 4 b. To assess the impact of feature utilization, we based our analysis on Experiment #3. To break predictability, we introduced disordered offset calculations in one branch (#9). To boost feature map utilization, we added additional edges for channel order mapping in our experiments (#10, Fig. 4 c). As shown in Table 2, performance consistently rose to $81.36\%$ and $81.51\%$ . Because the data read and written here is stored in continuous locations, it can minimize the impact on algorithm efficiency. Please refer to Appendix Sec. 6 for more details. + +# 3.6. Integrating Multiple Rep and Edges + +Considering the significant improvement in performance of multiple branches of Rep and edges, we attempted to merge the two experiments #8 and #10, labeled as #11. We found that it significantly outperformed SLaK, reaching an accuracy of $81.82\%$ with a setup featuring 2 Rep branches and + +![](images/1a1f16a0a95efba7194240defc9257da6a8d04c83defe9b2daced8a97454e764.jpg) +(a) + +![](images/3feeb51e169742ae2b92b81c6a31f628c01cf9ed4dccb5ffaf26ed79461376e9.jpg) +(c) + +![](images/b81a68294601418957528ca2960f45911424af334ecd898b3eabb5f1a04bed16.jpg) +(b) +Figure 4. (a) When the equivalent large kernel convolution size is $51 \times 3$ , the coverage regions of the feature maps, along with their utilization proportions, are determined by the areas that can be propagated downward through the shift operation. (b) The utilization ratio of feature maps varies as the number of downward propagation paths increases. $M$ denotes the longer side of the equivalent large kernel convolution, while $w/$ and $w/o$ indicate whether shuffled ordering is applied during the multi-path propagation process. (c) The network structure with increased downward propagation paths. + +4 SW edges. We therefore continued to experiment with different numbers branches of Rep and edges to see how their configurations can bring benefits, as shown in experiments #12 to #14. However, their final performance differences are almost negligible, which marks a sharp contrast with experiments #4 and #8. This discovery seems to indicate an overlapping effect between reparameterization and the reuse of feature maps. Upon in-depth analysis, we suspected that the initial sparsity distribution across layers might be the primary cause. Consequently, we experimented with using initial sparsity as a variable. There are many ways to affect the initial sparsity, and we adopt following approach here. Instead of adding all block weights first and then calculating sparsity, we calculated the sparsity of each Rep branches and then averaged these values within the same layer to determine the initial sparsity. This experiment, labeled #15, achieved the highest accuracy of $81.94\%$ . We also tried reducing overall sparsity to directly affect initial sparsity, which yielded similar results in experiment #16, matching #15. + +At this moment, the SW architecture has largely taken shape, as shown in Fig. 1. The $3 \times 3$ standard convolutions are tasked with the appropriate granularity for capturing fundamental visual details. The establishment of long-distance dependencies is delegated to the SW module, a method that relies on connections. These two parts effectively decouple the large kernel convolution and improve the performance of the model. + +Table 2. Experiments on Enhancing Feature Utilization. For detailed information, refer to sections 3.5 to 3.8. + +
modelIN-1K acc.(120epoch)
#3SW-(pad=N/2)81.26
#9disordered offset81.36
#10two edges81.51
#11rep2 E481.82
#12rep4 E481.79
#13rep4 E281.77
#14rep8 E481.85
#15rep2 E4 mean81.94
#16rep2 E4 sparsity 0.4 to 0.381.93
#17#16+gap=381.95
#18#16+gap=581.94
#19architecture82.25
#20+SE82.27
#21stage 182.17
#22subset81.72
#23interleaving82.23
+ +# 3.7. Shared Mask + +Inspired by Experiment #15, we focused on hyperparameter settings related to sparsity. A natural question arises: Do different Rep branches in the same layer need to maintain consistent sparse masks throughout the training process? To explore this, we reduced the frequency of mask sharing among branches, as demonstrated in experiments #17 and #18. By reducing the frequency of sparse mask sharing between branches, the performance of the model can be effectively improved. This indicates that there are differences in feature extraction between Rep branches. Reducing the synchronization frequency of masks can allow the model to explore different combinations of filters. + +# 3.8. Infrastructure + +We also note a significant architectural differences between UniRepLKNet and SLaK. UniRepLKNet favors increasing depth to boost representational capacity, whereas SLaK aims to "use more groups, expand width." This recalls discussions around VGG's design philosophy. To verify which idea of depth first or width first can bring more performance benefits to our proposed operator in modern CNNs, we conducted an experiment using UniRepLKNet's architectural hyperparameters only, including block counts in four stages and channel numbers per stage, labeled as #19. This experiment achieved an accuracy of $82.25\%$ . Once again, it has been confirmed that under the constraint of the total number of parameters, deepening the network provides better model performance than widening the network. When training with 120 epochs on ImageNet 1k, networks were constructed based on the hyperparameters of SLaK and UniRepLKNet, respectively. Our proposed operator exceeded SLaK's $81.6\%$ and UniRepLKNet's $81.6\%$ , as shown in experiments #5 and #19. + +Recognizing the substantial benefits of UniRepLKNet's network architecture, we explored additional parameters from this model. Firstly, we incorporated the SE module, which UniRepLKNet highlights, in our experiment labeled + +as #20, achieving an accuracy of $82.27\%$ . Secondly, we experimented with eliminating large convolution kernels in the first stage, labeled as #21, resulting in an accuracy of $82.17\%$ . The reparameterized branches in UniRepLKNet use multiple dilated convolutions, and their dilation rates appear as sequences with increasing sparsity. Following this method, we randomly generate subsets of masks for multiple Rep branches when sharing sparse masks(#22). But the accuracy has decreased to $81.72\%$ . Another key feature of UniRepLKNet is the interleaved use of small kernel and large kernel in stage 3 to reduce the number of parameters. In stages 2 to 4, UniRepLKNet utilizes a convolution kernel size of $13 \times 13$ , which corresponds to our parameter count of approximately $[49 \times 3, 47 \times 3, 13 \times 3]$ , representing [0.87, 0.83, 0.23] times the count of the previous stage. Although our module does not consume an excessive number of parameters, we also explore the benefits of incorporating different types of convolution kernels (#23). The overall accuracy has slightly decreased to $82.23\%$ . Apart from the network architecture and the Squeeze-and-Excitation (SE) mechanism, other enhancements present in UniRepLKNet cannot be directly integrated into our proposed algorithm to yield performance improvements. + +# 3.9. Architectural Specifications + +Thus far, we've identified a strategy that effectively replaces large kernels with small $3 \times 3$ standard convolutions, significantly enhancing performance. This approach revealed several parameters crucial for refining our architecture. + +The architecture begins with two stride-2 $3 \times 3$ convolutional layers in the stem, converting raw input into $C$ channel feature maps, where $C$ is an architectural hyperparameter. Subsequent downsampling blocks use one stride-2 $3 \times 3$ convolutional layer each, doubling the channel count, leading to channel configurations of $C$ , $2C$ , $4C$ , and $8C$ across the four stages. The block counts per stage are [3, 3, 18, 3] for SW-tiny and [3, 3, 27, 3] for SW-small. The initial stage channels $C$ are set to 80 for SW-tiny and 96 for SW-small, followed UniRepLKNet's hyperparameters. The equivalent large convolution sizes $M$ for each stage, inherited from SLaK, are [51, 49, 47, 13], meaning our group convolution's output expansion factor is $\left[\frac{M}{3}\right]$ . For instance, in the first stage with $C$ channels, the group convolution outputs $C \times \left[\frac{M}{3}\right]$ channels, which are then processed by the SW operator, as shown in Fig. 1. + +# 4. Results + +We employ SW to benchmark against state-of-the-art models on ImageNet classification [14], semantic segmentation on ADE20K [61], object detection/segmentation on COCO [33], and monocular 3D object detection on nuScenes [1]. + +Table 3. ImageNet classification. For detailed information on Throughput, please refer to Appendix Sec. 6. "T/C" denote transformer/ConvNet. + +
MethodTypeInput sizeParams (M)FLOPs (G)Acc (%)
SW-tinyC2242315.083.4
UniRepLKNet-T[17]C2242314.983.2
FastViT-SA24 [48]T2562213.882.6
PVTv2-B2 [53]T2242254.082.0
CoAtNet-0 [13]T2242254.281.6
DeiT III-S [47]T2242224.681.4
SwinV2-T/8 [37]T256228681.8
SLaK-T [35]C2242305.082.5
SW-smallC2242569.483.9
UniRepLKNet-S[17]C2242569.183.9
ConvNeXt-S [38]C2242508.783.1
HorNet-T [44]C2242233.983.0
FastViT-SA36 [48]T2562305.683.6
CoAtNet-1 [13]T2242428.483.3
SLaK-S [35]C2242559.883.8
FastViT-MA36 [48]T2562437.983.9
SwinV2-S/8 [37]T2562501283.7
RepLKNet-31B [16]C22427915.383.5
PVTv2-B5 [53]T22428211.883.8
+ +Imagenet-1k classification. We employ the identical training settings as described in Section Sec. 3.9, now extending the training to the complete 300 epochs. Models of different sizes need specific hyperparameters, performance at 120 epochs is indicative but not definitive; thus, extended training at 300 epochs is crucial for accurate hyperparameter tuning. The addition of new parameter settings increases the number of variables to explore, thus boosting resource consumption for larger models. Noting that UniRepLKNet did not experiment with the base model on ImageNet-1K, we focus on presenting the tiny and small model results here. For further details, please refer to Appendix Sec. 6. + +As shown in Table 3, we compared the performance of SW on ImageNet-1K against various state-of-the-art models. With comparable model sizes and FLOPs, our approach outperforms both SLaK and UniRepLKNet, two of the top-performing pure CNN models. Utilizing small convolutions, our method achieves higher accuracy than state-of-the-art transformers, such as Swin Transformer [37] and FastViT [48]. This demonstrates the enhanced efficacy of pure convolutional networks, showing that standard $3 \times 3$ convolutions can match the effects of larger convolutions and further boost the performance of CNNs. + +COCO object detection and instance segmentation. We utilize the pretrained SW models as backbones for Cascade Mask R-CNN [2, 24] and follow the standard 3x (36-epoch) training recipe with MMDetection [6]. As shown in Table 4, SW exceeds the performance of UniRepLKNets, ConvNeXt, RepLKNet, and SLaK—key examples of Vision Transformers and established large-kernel ConvNets. + +ADE20K semantic segmentation. We integrate pretrained SW models into UPerNet [57] for ADE20K [61] segmentation tasks, adhering to the standard 160k-iteration training + +Table 4. Object detection on COCO validation set. The FLOPs were calculated using inputs of 800 by 1280 pixels. + +
MethodParams (M)FLOPs (G)APboxAPmask
SW-tiny8775152.2145.19
UniRepLKNet-T [17]8974951.844.9
Swin-T [36]8674550.443.7
ConvNeXt-T [38]8674150.443.7
SLaK-T [35]--51.344.3
SW-small11083952.745.67
UniRepLKNet-S [17]11383553.045.9
Swin-S [36]10783851.945.0
ConvNeXt-S [38]10882751.945.0
SLaK-T [35]--51.344.3
+ +Table 5. Semantic segmentation on ADE20K validation set. The FLOPs were calculated using inputs of 512 by 2048 pixels. $SS$ and $MS$ denote single-scale and multi-scale evaluations, respectively. + +
MethodCrop sizeParams (M)FLOPs (G)mIoU (SS)mIoU (MS)
SW-tiny51226294849.2250.06
ConvNeXt-T [38]51226093946.046.7
SLaK-T [35]51226593647.6-
UniRepLKNet-T [17]51226194648.649.1
Swin-T [36]51226094544.545.8
InternImage-T [54]51225994447.948.1
SW-small512288103949.7950.83
ConvNeXt-S [38]512282102748.749.6
SLaK-S [35]512291102849.4-
UniRepLKNet-S [17]512286103650.551.0
Swin-S [36]512281103847.649.5
InternImage-S [54]512280101750.150.9
+ +Table 6. Monocular 3D object detection on nuScenes. Fcos3D ranked first among vision-only methods in the NeurIPS 2020 nuScenes 3D Detection Challenge. + +
MethodLrschdmAP
FCOS3D-SW-small1x31.42
FCOS3D-UniRepLKNet-S1x31.38
FCOS3D-ResNet101 w/ DCN1x29.8
+ +protocol from MMSegmentation. The table 5 presents the mean Intersection over Union (mIoU) scores on the validation set. Our method holds significant advantages over other models. + +Monocular 3D object detection. We deploy pretrained SW models as the foundation for fcos3d [52] on the nuScenes [1] dataset, utilizing MMDetection3D [12] for our setup. The table 5 details the mean Average Precision (mAP) obtained on the validation set. + +# 5. Analysis + +PeLK [4] customizes convolution kernels to mimic the human eye's focus and blur mechanism. This is achieved by managing the parameter count in the central and peripheral regions of the kernel. In contrast to PeLK's manual crafting of large kernel convolutions, our method discovers the spatial structure of convolutions data-drivenly. Our approach uses coarse-grained pruning to eliminate some output channels during training. Unlike traditional network structures, + +![](images/65a2c18d1f61ffe5c3d048aaac1d05997c03d3cd1b5944bb54ac5f8f40178369.jpg) + +![](images/583668e7eec5ecb86aa7c8b6407f8ab275048427b56a08064a203d6a22e2e10b.jpg) +Figure 5. (a) The variation of sparsity with increasing network depth. Colors represent different stages. (b-e) The group convolution of each layer has $\left\lceil \frac{M}{3} \right\rceil$ output channels. The proportion and distribution of the number of removed channel indexes to the total number of groups $nC$ . On the four stages, n is [1,2,4,8]. +Figure 6. (a-c) There are $nC$ group convolutions in each layer, and the proportion and distribution of groups with the number of filters removed. + +we employ a one-to-many group convolution, with each group having $M/3$ output channels. For the four stages, the values of $M$ are [51, 49, 47, 13], corresponding to the output channels per group convolution being [17, 17, 16, 5]. Compared to standard network structures, we can remove some output channels from group convolutions before summing them, ensuring the output is minimally affected by pruning. We will analyze the trained model to investigate the data-driven spatial structure of convolutions, focusing on three key aspects: 1. The sparsity variation across network depth. 2. The distribution of pruned convolution locations. 3. The statistical patterns of pruned numbers within a layer. + +We define 'sparsity' as the ratio of pruned filters to the total number of filters in each layer and label the network depth index as 'Layer No.' As shown in Fig. 5 (a), there is a general trend of increased pruning in deeper network layers. Notably, a closer examination at the stage level reveals a significant increase in the number of pruned filters at the last layer of each stage. This rapid increase suggests a reduced demand for complex transformations and a tendency towards information transmission during stage transitions. Such observations could potentially guide the design of future network architectures. Additionally, UniRepNet em + +ploys a strategy that alternates between large and small kernel convolutions across layers. Fig. 5 (a) also indicates that sparsity varies along the layer index, reflecting UniRep's interlacing strategy, albeit at a lower frequency. Interestingly, reducing the frequency of alternating large kernel usage does not yield significant benefits, a finding that requires further investigation. + +We investigated whether the data-driven pruning of large kernel convolutions reveals any clear structural patterns. Our approach started with the collection of indices for all pruned filters across the output channels of each group convolutions, which we labeled as 'Filter No'. Subsequently, we determined the ratio of pruned filters at each index to the total number of groups involved. Since each group convolution has a single input and multiple outputs, the number of filters at each index before training corresponds to the number of groups. Upon analyzing four stages, we did not identify any distinct patterns. However, a subtle observation was that large and small kernels seem to alternate along the direction of the output channels in group convolutions. This alternation could be a random occurrence or it might indicate the presence of beneficial patterns. Further analysis is required to determine the significance of this observation. + +We analyzed the number of pruned filters within each group, which we refer to as 'pruned num'. In a random scenario, the proportion of pruned channels to the total number of group convolutions is [1/17, 1/17, 1/16, 1/5]. Fig. 6 illustrates the sparsity after training. Overall, the number of pruned filters aligns with a Gaussian distribution, indicating that a significant number of group convolutions prune a similar quantity of filters at each stage. When comparing across different stages, stages 0-2 exhibit a consistent trend: the number of pruned filters increases with depth. In contrast, stage 3 displays an inverse pattern—where the number of pruned filters decreases as depth increases. Moreover, the initial layers of stages 0 and 1 are more inclined to preserve a greater number of filters. These observations could potentially inform the optimization of hyperparameter settings. + +# 6. Conclusions + +The progress in CNN models has been significantly influenced by the scaling up of kernel sizes. Our research shifts the focus from kernel size to a decoupled approach, which divides global attention into two critical components: basic granularity feature extraction and feature multi-path fusion. We have demonstrated that standard $3 \times 3$ convolutions can not only match the effects of large convolutions but also further enhance the performance of CNNs. This finding is in resonance with VGG's conclusions. Additionally, our research introduces new parameter settings that were previously unexplored. The analysis of the trained model's parameters has yielded intriguing findings, which could potentially spark further studies in various applications. e.g., Crack Detection [9, 10, 30] and Face recognition [40, 41]. + +# References + +[1] 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, 2020. 2, 6, 7 +[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 7 +[3] Hanting Chen, Yunhe Wang, Jianyuan Guo, and Dacheng Tao. Vanillanet: the power of minimalism in deep learning, 2023. 3 +[4] Honghao Chen, Xiangxiang Chu, Yongjian Ren, Xin Zhao, and Kaiqi Huang. Pelk: Parameter-efficient large kernel convnets with peripheral convolution. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5557-5567, 2024. 2, 7, 3, 4 +[5] Jierun Chen, Shiu-hong Kao, Hao He, Weipeng Zhuo, Song Wen, Chul-Ho Lee, and S.-H. Gary Chan. Run, don't walk: Chasing higher flops for faster neural networks. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12021-12031, 2023. 3 +[6] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7 +[7] Weijie Chen, Di Xie, Yuan Zhang, and Shiliang Pu. All you need is a few shifts: Designing efficient convolutional neural networks for image classification. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7234-7243, 2019. 2 +[8] Xuanhong Chen, Hang Wang, and Bingbing Ni. X-volution: On the unification of convolution and self-attention, 2021. 2 +[9] Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Jie Chen, Zun Liu, and Jianqiang Li. Geometry-aware guided loss for deep crack recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4703-4712, 2022. 8 +[10] Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Guanming Zhu, Zun Liu, Jie Chen, and Jianqiang Li. The devil is in the crack orientation: A new perspective for crack detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6653-6663, 2023. 8 +[11] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 180-189, 2020. 2 +[12] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.7 + +[13] Zihang Dai, Hanxiao Liu, Quoc V Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes. Advances in neural information processing systems, 34:3965-3977, 2021. 7 +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 6 +[15] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13733-13742, 2021. 3 +[16] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 1, 2, 3, 7, 4 +[17] Xiaohan Ding, Yiyuan Zhang, Yixiao Ge, Sijie Zhao, Lin Song, Xiangyu Yue, and Ying Shan. Unireplknet: A universal perception large-kernel convnet for audio, video, point cloud, time-series and image recognition, 2023. 1, 2, 3, 7 +[18] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12124-12134, 2022. 1 +[19] 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. ArXiv, abs/2010.11929, 2020. 2 +[20] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. Computational Visual Media, 9(4):733-752, 2023. 2 +[21] Shuxuan Guo, Jose M. Alvarez, and Mathieu Salzmann. Expandnets: Linear over-parameterization to train compact convolutional networks, 2021. 3 +[22] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 1577-1586. IEEE, 2020. 3 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 2 +[24] 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. 7 +[25] Tong He, Zhi Zhang, Hang Zhang, Zhongyue Zhang, Jun-yuan Xie, and Mu Li. Bag of tricks for image classification with convolutional neural networks, 2018. 2 +[26] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 2 +[27] Yunho Jeon and Junmo Kim. Constructing fast network through deconstruction of convolution, 2018. 2 + +[28] Ismail Khalfaoui-Hassani, Thomas Pellegrini, and Timothee Masquelier. Dilated convolution with learnable spacings. arXiv preprint arXiv:2112.03740, 2021. 3 +[29] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 2 +[30] Senyao Li, Jingling Yuan, Huilin Zhu, and Xian Zhong. Synergistic integration of cross-spatial learning for lightweight crack detection. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2025. 8 +[31] Wenrui Li, Penghong Wang, Ruiqin Xiong, and Xiaopeng Fan. Spiking tucker fusion transformer for audio-visual zero-shot learning. IEEE Transactions on Image Processing, 2024. 1 +[32] Wenrui Li, Ruiqin Xiong, and Xiaopeng Fan. Multi-layer probabilistic association reasoning network for image-text retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1 +[33] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 2, 6 +[34] Shiwei Liu, Lu Yin, Decebal Constantin Mocanu, and Mykola Pechenizkiy. Do we actually need dense overparameterization? in-time over-parameterization in sparse training. In International Conference on Machine Learning, 2021. 4 +[35] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Tommi Kärkkäinen, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. arXiv preprint arXiv:2207.03620, 2022. 1, 2, 3, 4, 7 +[36] 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. 1, 7 +[37] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 7 +[38] 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. 1, 2, 7 +[39] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard S. Zemel. Understanding the effective receptive field in deep convolutional neural networks. ArXiv, abs/1701.04128, 2016. 2 +[40] Yuxi Mi, Yuge Huang, Jiazhen Ji, Minyi Zhao, Jiaxiang Wu, Xingkun Xu, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using random frequency compo + +nents. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19673-19684, 2023. 8 +[41] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024. 8 +[42] Xuran Pan, Chunjiang Ge, Rui Lu, Shiji Song, Guanfu Chen, Zeyi Huang, and Gao Huang. On the integration of self-attention and convolution, 2022. 2 +[43] Zhiliang Peng, Zonghao Guo, Wei Huang, Yaowei Wang, Lingxi Xie, Jianbin Jiao, Qi Tian, and Qixiang Ye. Conformer: Local features coupling global representations for recognition and detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9454-9468, 2023. 2 +[44] Yongming Rao, Wenliang Zhao, Yansong Tang, Jie Zhou, Ser Nam Lim, and Jiwen Lu. Hornet: Efficient high-order spatial interactions with recursive gated convolutions. Advances in Neural Information Processing Systems, 35: 10353-10366, 2022. 7 +[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2 +[46] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions, 2014. 2 +[47] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In European Conference on Computer Vision, pages 516-533. Springer, 2022. 7 +[48] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Fastvit: A fast hybrid vision transformer using structural reparameterization. arXiv preprint arXiv:2303.14189, 2023. 7 +[49] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone, 2023. 3 +[50] Chien-Yao Wang, Hong-Yuan Mark Liao, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei Hsieh, and I-Hau Yeh. Cspnet: A new backbone that can enhance learning capability of cnn. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1571–1580, 2020. 3 +[51] Guangting Wang, Yucheng Zhao, Chuanxin Tang, Chong Luo, and Wenjun Zeng. When shift operation meets vision transformer: An extremely simple alternative to attention mechanism. ArXiv, abs/2201.10801, 2022. 2 +[52] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021. 7 +[53] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 7 + +[54] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, Xiaogang Wang, and Yu Qiao. Internimage: Exploring large-scale vision foundation models with deformable convolutions, 2023. 7 +[55] Jennifer L Wilkinson-Berka. Diabetes and retinal vascular disorders: role of the renin-angiotensin system. Expert Reviews in Molecular Medicine, 6:1 - 18, 2004. 2 +[56] Bichen Wu, Alvin Wan, Xiangyu Yue, Peter H. Jin, Sicheng Zhao, Noah Golmant, Amir Gholaminejad, Joseph Gonzalez, and Kurt Keutzer. Shift: A zero flop, zero parameter alternative to spatial convolutions. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 9127-9135. IEEE Computer Society, 2018. 2 +[57] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018. 7 +[58] Ruihan Xu, Haokui Zhang, Wenze Hu, Shiliang Zhang, and Xiaoyu Wang. Parcnetv2: Oversized kernel with enhanced attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5762, 2023. 1, 2 +[59] Kun Yuan, Shaopeng Guo, Ziwei Liu, Aojun Zhou, Fengwei Yu, and Wei Wu. Incorporating convolution designs into visual transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 559-568, 2021. 2 +[60] Xin Zhang, Yingze Song, Tingting Song, Degang Yang, Yichen Ye, Jie Zhou, and Liming Zhang. Akconv: Convolutional kernel with arbitrary sampled shapes and arbitrary number of parameters, 2023. 3 +[61] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302 - 321, 2016. 2, 6, 7 \ No newline at end of file diff --git a/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/images.zip b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c1fda3a26e45d88c02b29cf2c3b5329431e67fc7 --- /dev/null +++ b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46476967d58e335cd428f5efed67ffd4313b4bb66281e5d140af69f38fb7cd22 +size 475178 diff --git a/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/layout.json b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..42314121dada64275d7f94412b83df38b3f9aec1 --- /dev/null +++ b/CVPR/2025/ShiftwiseConv_ Small Convolutional Kernel with Large Kernel Effect/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc737abb2a908b7e63380abe6091e7889988414b33baadc9caa84133d36034c3 +size 400016 diff --git a/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/1ef5bf08-a8fb-4263-a952-98303f0ec85f_content_list.json b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/1ef5bf08-a8fb-4263-a952-98303f0ec85f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..137ac1dbbeec5b1ca766edbd79daae5c107a9339 --- /dev/null +++ b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/1ef5bf08-a8fb-4263-a952-98303f0ec85f_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c6aae603f329dbc59337f2c7369d8186db49c12eb269fd1631b7c8d3b2cd943 +size 66839 diff --git a/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/1ef5bf08-a8fb-4263-a952-98303f0ec85f_model.json b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/1ef5bf08-a8fb-4263-a952-98303f0ec85f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7ed7483df16df08bf0054a753702b65a378b8dc --- /dev/null +++ b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/1ef5bf08-a8fb-4263-a952-98303f0ec85f_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:130585290bc6e45a4630c7fdddbeec8a030d4239a337e852ec258ad741964052 +size 82066 diff --git a/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/1ef5bf08-a8fb-4263-a952-98303f0ec85f_origin.pdf b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/1ef5bf08-a8fb-4263-a952-98303f0ec85f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d40539e85929a047bc952212d61b1a9ef2ac622a --- /dev/null +++ b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/1ef5bf08-a8fb-4263-a952-98303f0ec85f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac5264f974ca883cce09803a46784af5b2c5ef2b39198b0d0bd55db4eac2ba52 +size 7987285 diff --git a/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/full.md b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..70b9030378ce088b02cd4aa44c4b6f4bebb1c68d --- /dev/null +++ b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/full.md @@ -0,0 +1,280 @@ +# Shining Yourself: High-Fidelity Ornaments Virtual Try-on with Diffusion Model + +Yingmao Miao $^{1,2*}$ , Zhanpeng Huang $^{2\dagger}$ , Rui Han $^{2}$ , Zibin Wang $^{2}$ , Chenhao Lin $^{1\dagger}$ , Chao Shen $^{1}$ , $^{1}\mathrm{Xi}$ 'an Jiaotong University, $^{2}$ SenseTime Research + +mym2017@stu.xjtu.edu.cn,{linchenhao,chaoshen}@xjtu.edu.cn, {huangzhanpeng,hanrui,wangzibin}@sensetime.com + +![](images/cc420c48948207d9b1b55dcedd7ce14d98e9612035d4a4665d13a740eec1e01f.jpg) + +![](images/a878deed671600f3120ef7e365e076265371ba8d5b891e2f6943fa3881f2e5f7.jpg) + +![](images/8c98729d052f05b1cd7f39bfea578b50d10a9bbca84c63e45a65e4556cf8f3d4.jpg) + +![](images/c1a01ca738aca0403793b3fabbd25ecf79f9e5fdb1f6dc1769902eb728535ce5.jpg) + +![](images/dbeb00b7724b4898aeaa31b7a4064c127f43af9ae86e900c0a6b578c961f8c1d.jpg) +Figure 1. Shining Yourself. We propose the virtual try-on task for ornaments including bracelets, rings, earrings, and necklaces for the first time. Our method achieves realistic virtual try-on results and high-fidelity identity preservation of ornament using pose-aware mask prediction and mask-guided attention. Project Page: https://shiningyourself.github.io/ + +![](images/b6090b69b00041c9f79bb6180d48a499a780a65120ec57b3e6a27352b4ae146d.jpg) + +![](images/203a2f4c20a10b6792fd92d478b23f3e84ad95855b639d67280fa65867886ede.jpg) + +![](images/38fe670f869568a52db3abc7024c0133cfb773c18de6ebbdbeae4eaadb6cdaec.jpg) + +![](images/86f44cd3a908ff4e11ab79550078d4d060a23a3e6590cfe54a59180571036958.jpg) + +![](images/030ebca109b21660d18757ad7da55505e9dbd86fdd63ec0f0be7824d6ea85daf.jpg) + +![](images/f38bd3708a05dc949cfc85f4ce088dbbc1340a84d85ca6da2179f3e41e074b02.jpg) + +![](images/1bbf58125c5ab7bb860708a899e5468273510337c182c8e4f080ce3700883dea.jpg) + +# Abstract + +While virtual try-on for clothes and shoes with diffusion models has gained attraction, virtual try-on for ornaments, such as bracelets, rings, earrings, and necklaces, remains + +largely unexplored. Due to the intricate tiny patterns and repeated geometric sub-structures in most ornaments, it is much more difficult to guarantee identity and appearance consistency under large pose and scale variances between ornaments and models. This paper proposes the task of virtual try-on for ornaments and presents a method to improve the geometric and appearance preservation of orna + +ment virtual try-ons. Specifically, we estimate an accurate wearing mask to improve the alignments between ornaments and models in an iterative scheme alongside the denoising process. To preserve structure details, we further regularize attention layers to map the reference ornament mask to the wearing mask in an implicit way. Experimental results demonstrate that our method successfully wears ornaments from reference images onto target models, handling substantial differences in scale and pose while preserving identity and achieving realistic visual effects. + +# 1. Introduction + +As diffusion model [9, 15, 16, 22, 24, 26, 40] becomes the de facto standard in image generation, it's also widely adopted in the field of virtual try-on [7, 17, 32, 35, 39, 42]. Given a reference image of an item and a target image of a model, the task is to get a preview of the fitting effect using image generation methods. Since no in-person wearing or physical fitting room is required, it has great potential for massive advertising materials generation in various applications of retail, e-commerce, and advertisement. + +The main challenge of virtual try-on is how to generate a realistic fitting effect while preserving the fidelity of the garment. Massive efforts [7, 17, 32, 35, 39] have been dedicated to solving the problem in the field of garments. The methods usually employ a network module (e.g., CLIP image encoder [21] or ReferenceNet [11, 16]) to extract garment features, which are injected into the process of diffusion denoising to preserve the identity and details of the garment. It has made great progress to be widely adopted for commercial use. + +However, few works focus on virtual try-ons of ornaments such as bracelets, rings, earrings, and necklaces, despite significant practical demand. Virtual try-on for ornaments presents unique challenges that existing methods struggle to address effectively. 1) Most ornaments often feature intricate, small-scale geometric structures, such as rings and holes, that are difficult to preserve in virtual try-ons. In contrast, garments typically have sparser and/or repeated textures and lack complex geometric details. 2) For garment try-ons, generative visual artifacts can blend with natural cloth deformations and wrinkles, reducing their visibility. However, most ornaments are rigid or consist of rigid components, making any distortion or artifacts immediately noticeable. 3) Most garment virtual try-on methods require silhouette, skeleton, and semantic maps as additional inputs, while even the coarse silhouette mask of the ornament is not easy to depict due to its diverse structures and pose-dependent occlusion. + +To tackle the problems, we propose a method to predict an accurate wearing mask to align the poses and scales between ornaments and models without additional inputs such + +as the silhouette, skeleton, or semantic maps. Our method also preserves structures and detailed features using a mask-guided attention of ornaments and models to preserve geometric structures. Specifically, to obtain a precise pose-aware wearing mask without explicit extraction of various maps of ornaments and models, we refine the mask from an input mask as coarse as a bounding box. A refined mask is then estimated using the intermediate features in the generation, which is further used as input to predict a more accurate one iteratively. As the mask indicates structure information, especially for multiple and tiny geometric structures, we formulate the attention layers to implicitly learn a mapping between the reference mask and the ground truth mask to preserve the structure patterns in ornaments. In summary, our contributions are as follows: + +- To the best of our knowledge, we are the first to use diffusion models for virtual try-ons of various ornaments including bracelets, rings, earrings, and necklaces. Our method can generate realistic virtual try-on results as well as high-fidelity preservation of ornament identity. +- We propose an iterative scheme to estimate a pose-aware wearing mask to significantly improve the pose and scale alignments between ornaments and models. It also facilitates virtual try-on applications without the requirements of various additional inputs such as silhouette, semantic, and skeleton maps. +- By constraining attention to learning a mapping between reference and wearing masks, our method improves geometric feature preservation, especially for tiny items with complicated geometric structures, such as ornaments. + +# 2. Related Works + +Personalized image generation To address the challenge of text prompts not fully capturing user intent, personalized image generation, and editing have garnered significant attention from researchers. Inversion-based methods, such as Textual Inversion [10] focus on optimizing a special prompt word to represent a target concept. Meanwhile, fine-tuning approaches [12, 20, 23, 25] adjust pre-trained diffusion models using a small set of images of the target concept to create personalized models. Although these methods can generate high-fidelity images, they do not allow users to control the generation area for the target, and the additional optimization and fine-tuning required during inference can hinder large-scale applications. On the other hand, some zero-shot and training-free personalized image generation methods [30, 36], can produce images of target concepts. However, they excel in stylized generation by representing high-level semantic features in the lack of preserving details and identity. + +Local image editing Our virtual ornament-wearing task closely resembles local image editing techniques. However, most previous local image editing methods relied on + +![](images/953dd7ed579f79c748d9ef8eac8fab9c8cb2ada0e1eb2a3b8c0c10b4b84895e9.jpg) +a) The pipeline of our framework + +![](images/a65abec3f21749d7805bcca8e63a39857d29ffe3eb4f90ca98513d04f461c73b.jpg) +b) The mask-guided attention +Figure 2. The overview of our method. a) In training, given reference ornament and model images and masks, our method concatenates ornament and masked model images as input to the ReferenceNet branch, which extracts features to predict wearing mask in an iterative way. The extracted features are also injected into the denoising U-Net to improve details generation. b) We enforce the attention layers to preserve structure details by formulating the layers to map the reference ornament mask to the ground truth wearing mask in an implicit way rather than directly imposing the mask onto attention maps. + +text prompts, such as Blended Diffusion [1] and Blended Latent Diffusion [2], which employed multi-step semantic blending during denoising to produce harmonious images containing target semantic information within a defined mask. Inpainting Anything [37] replaces any object in the input image with a target described by text prompts. However, for virtual ornament wearing, it is crucial to ensure alignment with the intricate details of the target. The traditional image composition pipeline [5, 8] involves cutting and pasting foreground images onto background images, followed by the harmonization techniques. Recently, numerous diffusion-based methods [6, 27, 31, 34] have emerged in this field, significantly enhancing the quality and coherence of generated images. For instance, Paint by Example [34] employs a CLIP image encoder [21] to convert reference images into embeddings for guidance, generating objects that are semantically aligned with the reference image. ObjectStitch [27] similarly utilizes CLIP to align text and images, guiding the generation of the diffusion model. ObjectDrop [31] first trains an object removal network, assembles a large dataset, and subsequently conducts object insertion training. AnyDoor [6] leverages ControlNet and a DINO [3] encoder to extract detailed semantic information, improving its ability to maintain object identities. These methods focus on general object insertion into the background smoothly without pose alignment, which is normally required in virtual try-on tasks, especially for ornament wearing that needs precise pose alignment between + +the ornament and wrist, finger, or neck. + +Virtual try-on Virtual try-on [7, 17, 32, 35, 39, 42] takes a model image and an item image to generate an image of the model wearing the item. Early virtual try-on methods were based on Generative Adversarial Networks (GANs). Recently, with the significant success of diffusion models, researchers have explored their application in the field of virtual try-on. Most works focus on virtual try-on of garments, such as TryonDiffusion [42], OOTDiffusion [32], and IDM-VTON [7]. These methods utilize two parallel U-Nets for garment feature extraction, integrating them through self-attention and achieving impressive results. StableVTON [17] introduced a zero-initialization cross-attention module to inject garment features into the denoising network. A few works [4, 33] explore virtual try-on of shoes and earrings. In their settings, either the shoe pose is fixed to be aligned by the model or the earring has an almost vertical pose as it hangs down from the ear. Ornament virtual try-on requires pose alignment with different body parts at various poses and scales. Methods such as OOTDiffusion and IDM-VTON use additional inputs such as skeleton and semantic maps to guide wearing pose. In contrast, it's much more difficult to depict ornament wearing mask due to complicated tiny geometry structures and pose-related occlusion. + +# 3. Methodology + +We propose a zero-shot method for ornament virtual try-on with a reference ornament image, a target model image, + +and a coarse bounding box. The bounding box coarsely indicates the wearing location as ornament wearing is user-specific (e.g., a ring has its finger symbolism). We can generate realistic and high-fidelity fitting effects without additional inputs such as pose and semantic maps. The model comprises two vital components: 1) an iterative pose-aware wearing mask prediction and refinement module from the bounding box, which improves pose alignments between the ornament and the model; 2) a mask-guided attention module to improve identity and detail preservation. The framework of our method is illustrated in Fig. 2a. + +# 3.1. Diffusion model and ReferenceNet + +Diffusion model Our method is built upon the latent diffusion model (LDM) and ReferenceNet module, which has been widely adopted for condition generation and virtual try-on tasks. A typical LDM implementation [22] comprises an encoder-decoder module and a denoising network. The encoder embeds the input image into a low-dimension latent code to reduce computational overhead, which is diffused and then denoised by the denoising network to recover from a random noise. The denoised latent code is then decoded to generate an RGB image. The training process is formalized as follows: + +$$ +\mathcal {L} _ {1} = \mathbb {E} _ {z _ {0}, c, \epsilon , t} \left(\left\| \epsilon - \epsilon_ {\theta} \left(z _ {t}, c, t\right) \right\| _ {2} ^ {2}\right) \tag {1} +$$ + +where $z_{t}$ represents the latent feature at time step $t$ , which can be obtained through $z_{t} = \sqrt{\bar{\alpha}_{t}} z_{0} + \sqrt{1 - \bar{\alpha}_{t}}\epsilon$ , $\epsilon \in \mathcal{N}(0,I)$ . $c$ is the condition embedding from a text prompt or reference image with a text or image encoder, and injected into the cross-attention layers to guide the generation. Our model adopts the widely used CLIP image encoder to extract features of the reference ornament image. + +ReferenceNet module The module is widely used in virtual try-ons to improve detail and structure preservation. It is designed to be similar and parallelized with the denoising U-Net. The module extracts hierarchical latent features of the reference image which are injected into related layers in the denoising network. Specifically, latent features in the ReferenceNet are concatenated onto their counterparts in the denoising network for attention calculation. + +# 3.2. Pose-aware Mask Refinement + +We conducted several experiments to explore how the pose and scale impact the generative results, from which we have several key findings: 1) The diffusion-based model has the capability to fit ornaments to various model poses even without finetuning on ornament try-on datasets, which is attributed to the image prior from the pre-trained diffusion base model. 2) Poses and scales have significant influences on the fitting effects. In general, using an accurate wearing mask will significantly improve the pose alignment between ornaments and models even with large poses and + +scale variances. However, the wearing mask is not equivalent to the semantic mask, which is predicted from an existing image while wearing mask is hallucinated from two irrelevant ornament and model images. It's difficult or even infeasible to obtain accurate wearing mask in inference. In addition, ornaments usually show close-up views, which require a much more accurate wearing mask than the coarse silhouette mask in garment virtual try-on. + +Previous works [13, 29] have shown that intermediate results (e.g., latent features and attention maps) in early generative phases contain the semantic structure of the generated images. It might be possible to extract a wearing mask from these intermediate maps. However, extracted masks are too coarse to be used. To solve the problem, we proposed to estimate a more accurate wearing mask. We add an additional linear layer to predict the wearing mask from the intermediate maps. The predicted wearing mask is further used as input to guide the generation. The iterative refinement converges to an accurate pose-aware mask aligned with the model in the final generative image. Specifically, triplet images of an ornament, a model, and a coarse bounding box $M_{b}$ are fed into the ReferenceNet. The ornament latent features $f_{o}^{t}$ and the model latent features $f_{m}^{t}$ are injected into the denoising counterpart similar to most networks in garment virtual try-ons. The latent features are further concatenated and linearly projected to predict the wearing mask $\hat{M}_{p}^{t}$ . The predicted mask and the bounding box are blended as a new wearing mask input, which is updated as follows: + +$$ +\hat {M} _ {p} ^ {t - 1} = \alpha_ {t} \odot \hat {M} _ {p} ^ {t - 1} + (1 - \alpha_ {t}) \odot M _ {b} \tag {2} +$$ + +$$ +\hat {M} _ {p} ^ {t} = \operatorname {M L P} \left(\left[ f _ {m} ^ {t} \odot \hat {M} _ {p} ^ {t - 1}, f _ {o} ^ {t} \right]\right) \tag {3} +$$ + +where $\alpha_{t} \in [0,1]$ is a hyperparameter in terms of training step $t$ . In the early training, the predicted wearing mask is coarse, and $\alpha_{t}$ is set to be small. As the mask gets more accurate in the late stage, $\alpha_{t}$ approximates to 1.0. + +As mask prediction and image generation are entangled, we employ an ornament try-on dataset with wearing masks to regularize mask prediction with a $L_{2}$ loss: + +$$ +\mathcal {L} _ {2} = \left\| \hat {M} _ {p} ^ {T} - M _ {o} ^ {g t} \right\| _ {2} ^ {2} \tag {4} +$$ + +where $M_{o}^{gt}$ is the ground-truth wearing mask. The regularization is important to prevent the dual degeneration of both results due to mutual dependence. In inference, only a bounding box is required to indicate the user-specific wearing location. + +# 3.3. Mask-guided Attention + +The precise wearing mask improves the alignment between ornaments and models with various poses and scales. It also has a positive effect on detail generation. However, most ornaments comprise complicated tiny geometric components such as repeated shapes and/or ring structures in + +![](images/8f50ffa3ebe0c23b2998cfae4c190499ec668b85f2b304d928e2c998612c3b72.jpg) +Figure 3. Visual comparison between previous methods and ours. No existing method could keep appearance and structure consistent, especially geometric details and numbers of components in ornaments. Our method preserves both details and identity and achieves high-quality and high-fidelity fitting results. + +pearl necklaces and beaded bracelets. Our early attempts found that the model had difficulty preserving the topology and/or the number of components, especially in repeated geometric patterns. To take a few examples, it may ignore small parts interleaved with other large components, or fill the hole of a ring structure. We suspect that existing generative networks can capture appearance and spatial details rather than geometric structures, as geometric shapes require hard constraints on local primitive structures of edges and contours. + +Attention maps retain shape details with affinities be + +tween the spatial features [13]. A possible solution is to impose a geometric structure constraint in attention maps. The semantic segmentation mask contains rich geometric structure information, but the mask is difficult to extract due to massive tiny complicated sub-components in ornaments. As the binary mask is also full of geometric structures of edges and contours and easy to obtain (e.g., with SAM [19]), we propose to employ the reference ornament mask to inject geometric structure into the generation. + +However, directly blending attention maps with the mask may mask out too much information to degrade gener + +ated results. We introduce an indirect way to restrict geometric structure changes of ornaments in reference and generated images. Specifically, we obtain attention maps $\{M_a^i\in \mathbb{R}^{d_i\times d_i}\}_{1}^N$ of latent features and ornament embeddings from various layers in denoising U-Net, where $N$ is the number of extracted attention maps and $d_{i}$ is the dimension of $i$ -th attention map. The ornament mask $M_o$ in the reference image is down-sampled and flattened as one-dimensional masks $\{M_o^i\in \mathbb{R}^{d_i}\}_{1}^N$ . We then apply $M_o^i$ to mask out the attention map $M_{a}^{i}$ along one dimension and margin it along the other dimension. The result is then reshaped and up-sampled to $\tilde{M}_o^i$ as the same dimension of $M_o$ . All result masks $\{\tilde{M}_o^i\}$ are then averaged as the final mask $\tilde{M}_o$ , which can be formulated as: + +$$ +M _ {o} ^ {i} = \mathbf {T} _ {\text {f l a t t e n}} ^ {\sqrt {d _ {i}} \times \sqrt {d _ {i}} \rightarrow d _ {i}} \circ \mathbf {T} _ {\text {d o w n s a m p l i n g}} ^ {d _ {0} \times d _ {0} \rightarrow \sqrt {d _ {i}} \times \sqrt {d _ {i}}} \circ M _ {o} \tag {5} +$$ + +$$ +\tilde {M} _ {o} ^ {i} = M _ {a} ^ {i} \odot \underbrace {\left[ M _ {o} ^ {i T} , \dots , M _ {o} ^ {i T} \right]} _ {d _ {i}} \tag {6} +$$ + +$$ +\tilde {M} _ {o} ^ {i} = \mathbf {T} _ {\text {u p s a m p l i n g}} ^ {\sqrt {d _ {i}} \times \sqrt {d _ {i}} \rightarrow d _ {0} \times d _ {0}} \circ \mathbf {T} _ {\text {r e s h a p e}} ^ {d _ {i} \rightarrow \sqrt {d _ {i}} \times \sqrt {d _ {i}}} \circ \sum_ {c} \tilde {M} _ {o} ^ {i} [ r ] [ c ] \tag {7} +$$ + +$$ +\tilde {M} _ {o} = \frac {1}{N} \sum_ {i} ^ {N} \tilde {M} _ {o} ^ {i} \tag {8} +$$ + +where $\mathbf{T}_{\mathrm{ops}}^{\mathrm{d1}\rightarrow \mathrm{d2}}$ is the operator with ops as operation type and $\mathbf{d1}\to \mathbf{d2}$ as dimension mapping from d1 to d2. Sequential operators are defined to execute from right to left. The masking operation enforces latent features to attend to ornament regions in the reference image, while the margining operation diffuses ornament features to the wearing region in the generated image. The reference mask $M_{o}$ is mapped to the wearing masks $\tilde{M}_o$ via the attention map. Inversely, in order to enable the attention map to learn the mapping, we require the transformed wearing masks $\tilde{M}_o$ to be consistent with the ground truth wearing ornament mask $M_o^{gt}$ with an $L_{2}$ loss as below: + +$$ +\mathcal {L} _ {3} = \left\| \tilde {M} _ {o} - M _ {o} ^ {a t} \right\| _ {2} ^ {2} \tag {9} +$$ + +The process is shown in Fig.2b. Down-sampling and upsampling operations are not displayed for concise illustration. + +# 3.4. Training + +Dataset Inspired by the common practice in garment virtual try-ons, we collect image pairs of ornaments and models wearing the ornaments. We mask out the ornament in the model image to obtain the target image and ground-truth wearing mask. The reference ornament image, the model image with masked-out ornament, and the original model image are combined as a training triplet image. We also label the masks in ornament images as reference masks. + +Our dataset does not require pose alignment between the ornament and the model, which is easy to collect, and also prevents the model from learning a simple copy-and-paste strategy. In total, we collect about 64k image triplets, roughly evenly distributed over four categories of bracelets, rings, earrings, and necklaces. Each image triplet also contains a reference mask and a wearing mask of the ornament. + +Training loss Our training loss comprises the aforementioned three items : + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {1} + \lambda_ {1} \mathcal {L} _ {2} + \lambda_ {2} \mathcal {L} _ {3} \tag {10} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are loss weights. The two weights decay as the training step increases, which forces the model to learn the wearing mask in the early stages. As the mask becomes accurate, the model focuses on the generation of appearance details. The scheme follows common observation [28] that image layout and structure are sketched in the early stage and details are generated in the late stage. + +# 4. Experiments + +# 4.1. Implement details + +Our model adopts the Stable Diffusion V1.5 as the network backbone. The ornament region is cropped and resized to $512 \times 512$ , and Adam [18] optimizer is chosen with an initial learning rate of $1e^{-5}$ . We employ a simple linear decay for the $\alpha(t)$ and select the self-attention maps with the highest resolution from the encoder and decoder for masked-guided attention. We found these simple settings are enough to obtain compelling results. We follow a similar scheme as AnyDoor [6] in handling inputs and composing final results. Specifically, the ground truth mask is resized to be square, and the cropped image is scaled by a factor of 1.5. The generated result is pasted back to the original masked region to compose the final result. Our model takes about 10 hours to train on 8 A100 GPUs with 10 epochs. + +For quantitative comparison, we adopt FID [14] and LPIPS [41] to evaluate image quality, while the CLIP image similarity score and DINO-based feature similarity score are used to measure the identity consistency of the ornaments. All results are calculated and averaged on a test image set split from our dataset. + +# 4.2. Comparisons + +As we are the first to focus on ornament virtual try-ons, we select several works that are most related to ours in the broad field of image edits. These works include Paint-by-Example(CVPR'23) [34], AnyDoor(CVPR'24) [6], and IDM-VTON(ECCV'24) [7]. The first two works are designed to insert reference objects into target images, while the latter is dedicated to garment virtual try-ons. Similar to ours, these methods require a reference image of the item and a target image as well as masks to define local edit + +![](images/6e2b263e93ad4d55f3eee5c1222d9f244fb650de00af03326dddc48793917734.jpg) +Figure 4. Virtual try-on results on other categories including bracelets, rings, necklaces, and earrings. + +Table 1. Quantitative evaluations between our and other methods + +
Compared against ground truthCompared against reference ornament
MethodFID↓LPIPS↓CLIP Score↑DINO Score↑CLIP Score↑DINO Score↑
Paint-by-Example [34]23.490.078985.664.857.235.4
AnyDoor [6]28.280.102985.167.254.835.9
IDM-VTON [7]22.990.070985.965.055.935.2
Ours19.000.059388.774.557.338.7
Ground Truth----59.043.3
+ +regions. All methods are trained or fine-tuned with our dataset. Limited to the page length requirement, we take the bracelet category as an example to illustrate all visual results. Please refer to the Appendix for more results of other ornament categories. + +Qualitative results Fig. 3 qualitatively compares fitting results on ornaments of various structures and poses. Paint-by-Example could hardly preserve geometric structures and appearance in most cases. AnyDoor struggles to preserve the scales of the whole structure and/or major parts. IDM-VTON could preserve the scale to a certain extent, but it has + +problems maintaining structure layouts, especially for complicated ornaments with multiple parts. None of the previous methods could hold the number of sub-parts in ornaments or recover repeated geometric patterns. Our method has the most visual appearance and structure similarities to both reference ornaments and ground truth images, which indicates its ability to preserve both appearance and local and global structures, and tiny surface geometric patterns (the last row). Surprisingly, our results seem to be biased towards reference ornaments with less specular reflections than ground truth. It's partially because model images do + +Table 2. Quantitatively comparisons of our models with different module configurations. + +
MethodCLIP Score↑DINO Score↑
Baseline86.971.9
w/o mask refinement88.573.3
w/o mask-guided attenu.88.073.9
Ours88.774.5
+ +not have enough hints of environmental illumination that our model has difficulty in learning the exact light effect as the ground truth. + +Quantitative results Table 1 illustrates a quantitative comparison between ours and previous methods. The two ornaments in reference and ground-truth images are usually captured with different conditions such as views and illumination. Therefore, we compare the results generated by all methods against both the reference ornaments and the ground truth, and calculate the corresponding consistency metrics. Our approach achieves the best results in all metrics, demonstrating its capability to generate more realistic and high-fidelity virtual results. + +# 4.3. Ablation Study + +We conduct a comprehensive ablation study to evaluate the effectiveness of our proposed components. The experiments are designed by adding a component from the basic models. The baseline is adapted from ReferenceNet and Stable Diffusion. The results are evaluated from both qualitative and quantitative aspects. Fig. 5 shows the visual comparisons with various component configurations. The basic model has noticeable defects in details and geometric structures. If we do not integrate the mask prediction, the results lack appearance details and specular lights. It may also lose structure consistency to some extent (e.g., flow structure missing in the first ornament). Without mask-guided attention, both the local and global structures are destroyed in form of adding or missing components as well as changing scales. On the other hand, the full model preserves both appearance and geometric details as well as global structures. The quantitative results in Table 2 also indicate the importance of the two proposed modules to improve the final virtual try-on results. More results are in the Appendix. + +# 4.4. More Results + +We use our method to wear various types of ornaments including bracelets, necklaces, earrings, and rings. Figure 4 lists the results. Other configurations including a model wearing different ornaments and different models wearing the same ornaments are also illustrated in Figure 1 (the last row). All results demonstrate that our method can handle various ornament structures of local and global rigid and + +![](images/781776528c2458e471d71958142838885221b4b8b408d6c7ec5257ec65abf703.jpg) +Figure 5. The visual comparisons of our models with different module configurations. The full model archives the best results with the proposed two modules. + +![](images/211d22a1c20c6c0ab9ef38ac64964ff8c55f8ac0b30d7ad40f9cb46796578538.jpg) +Figure 6. Our model is robust to achieve consistent results with different poses and scales. + +non-rigid components. More results are in the Appendix. + +To evaluate the robustness of our model under conditions of different poses and scales. We also conduct experiences by randomly rotating and scaling the reference ornament, which is then used to wear on the same model. As Fig. 6 shows, the results are consistent in details and geometric structures with different configurations. The experiment results show our model is robust with large pose differences between ornaments and models. + +# 5. Conclusion + +We propose the virtual try-on ornament task for the first time. To tackle the more challenging problems of intricate geometric structures in ornaments, we devise two modules of mask prediction and mask-guided attention to obtain accurate wearing masks and impose geometric structures, which preserve both appearance details and geometric structures to achieve identity consistency. Currently, our method is biased toward reference images rather than ground truth images, which lack specular reflections to a certain extent. Inspired by the work [38], we would like to add more fine-grained lighting control in our future work. Besides, inaccurate control over wearing orientations (e.g., rotated along a wrist) leading to featured components being hidden behind the wrist occasionally. Secondary masks and local feature injection into the diffusion process may fix the problem. + +# Acknowledge + +# References + +[1] 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 +[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 3 +[3] 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 IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3 +[4] Binghui Chen, Wenyu Li, Yifeng Geng, Xuansong Xie, and Wangmeng Zuo. Shoemodel: Learning to wear on the user-specified shoes via diffusion model. arXiv preprint arXiv:2404.04833, 2024. 3 +[5] Bor-Chun Chen and Andrew Mae. Toward realistic image compositing with adversarial learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8415-8424, 2019. 3 +[6] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 3, 6, 7, 1 +[7] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for authentic virtual try-on in the wild. In ECCV, pages 206-235, 2024. 2, 3, 6, 7, 1 +[8] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8394-8403, 2020. 3 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794. Curran Associates, Inc., 2021. 2 +[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations. 2 + +This research is supported by Sensetime, the National Key Research and Development Program of China (2023YFB3107401), the National Natural Science Foundation of China (T2341003, 62376210, 62161 160337, 62132011, U21B2018, U24B20185, 62206217), the Shaanxi Province Key Industry Innovation Program (2023-ZDLGY-38). Special thanks to Jessie Geng for the coordination of computing resources. +[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations. 2 +[12] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2 +[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations. 4, 5 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[16] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2 +[17] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 2, 3 +[18] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[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. 5 +[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2 +[21] 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. 2, 3 +[22] 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. 2, 4 +[23] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine + +tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2 +[24] 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 +[25] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 2 +[26] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations. 2 +[27] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18310-18319, 2023. 3 +[28] Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consistent text-to-image generation. ACM Transactions on Graphics (TOG), 43(4):1-18, 2024. 6 +[29] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, 2023. 4 +[30] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2 +[31] Daniel Winter, Matan Cohen, Shlomi Fruchter, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectdrop: Bootstrapping counterfactuals for photorealistic object removal and insertion. arXiv preprint arXiv:2403.18818, 2024. 3 +[32] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779, 2024. 2, 3 +[33] Youze Xue, Binghui Chen, Yifeng Geng, Xuansong Xie, Jiansheng Chen, and Hongbing Ma. Strictly-id-preserved and controllable accessory advertising image generation. arXiv preprint arXiv:2404.04828, 2024. 3 +[34] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 3, 6, 7, 1 +[35] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7017-7026, 2024. 2, 3 +[36] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. 2023. 2 +[37] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3 +[38] Chong Zeng, Yue Dong, Pieter Peers, Youkang Kong, Hongzhi Wu, and Xin Tong. Dilightnet: Fine-grained lighting control for diffusion-based image generation. In ACM SIGGRAPH 2024 Conference Papers, 2024. 8 +[39] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8372-8382, 2024. 2, 3 +[40] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[42] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4606-4615, 2023. 2, 3 \ No newline at end of file diff --git a/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/images.zip b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d064c678585d8e4233bea12639b316e924089f87 --- /dev/null +++ b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c7847bbbfa77223c989edd179f8b7d593ebfdc9e3b8356f0a8b8ab01ea0dc95 +size 692431 diff --git a/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/layout.json b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ce6c8b7ae693413fe5d3429fdb36ca6e30b0e3e4 --- /dev/null +++ b/CVPR/2025/Shining Yourself_ High-Fidelity Ornaments Virtual Try-on with Diffusion Model/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79dd79acb7ecbdaff30e0eb504d414725183843b7407a87927e81c0b3ded7cfa +size 309167 diff --git a/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/e010292b-85e5-4ad6-b5ff-fd493535422e_content_list.json b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/e010292b-85e5-4ad6-b5ff-fd493535422e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..65c7809bfa32ea29ca9c4ca4c5c8f79f82deb0e4 --- /dev/null +++ b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/e010292b-85e5-4ad6-b5ff-fd493535422e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2aedad5ff2bb72fab4f471c9b5546326962386a594d367039cc038ede9d2017b +size 69792 diff --git a/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/e010292b-85e5-4ad6-b5ff-fd493535422e_model.json b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/e010292b-85e5-4ad6-b5ff-fd493535422e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8071e94a033b3185da736b0035dee4d23db9b681 --- /dev/null +++ b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/e010292b-85e5-4ad6-b5ff-fd493535422e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14aab9958cfdd61ba6dad4551a8e7268433616361889f0f097a8a5af95d3c10c +size 89770 diff --git a/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/e010292b-85e5-4ad6-b5ff-fd493535422e_origin.pdf b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/e010292b-85e5-4ad6-b5ff-fd493535422e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..21877531a1c00b2535475c588666ea538a07df7c --- /dev/null +++ b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/e010292b-85e5-4ad6-b5ff-fd493535422e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56e6be556029a37d13868c3fa9f391fd5676048cd8d8516d1670e3a387ee481e +size 6107215 diff --git a/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/full.md b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5ad21d1b6cc046b3751f3a0ef267fa9460480d53 --- /dev/null +++ b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/full.md @@ -0,0 +1,227 @@ +# ShotAdapter: Text-to-Multi-Shot Video Generation with Diffusion Models + +Ozgur Kara $^{1,2}$ + +Krishna Kumar Singh2 + +Feng Liu2 + +James M. Rehg1 + +Tobias Hinz + +UIUC + +2Adobe + +Project Webpage: https://shotadapter.github.io/ + +![](images/2c3f7ff19180cf7990cc38c9b8e15f9c23ca51164330da44a43ae3d44d842bfc.jpg) +Figure 1. ShotAdapter is a lightweight framework that enables text-to-multi-shot video generation by fine-tuning a pre-trained text-to-video model. It allows control over number and duration of shots as well as shot content through shot-specific text prompts. The framework maintains character identity while being able to preserve backgrounds (e.g. $3^{\mathrm{rd}}$ row) or transition to new ones (e.g. $4^{\mathrm{th}}$ row), featuring distinct activities (e.g. playing guitar, then using laptop) and perspectives. + +# Abstract + +Current diffusion-based text-to-video methods are limited to producing short video clips of a single shot and lack the capability to generate multi-shot videos with discrete transitions where the same character performs distinct activities across the same or different backgrounds. To address this limitation we propose a framework that includes a dataset collection pipeline and architectural extensions to video diffusion models to enable text-to-multi-shot video generation. Our approach enables generation of multi-shot videos as a single video with full attention across all frames of all shots, ensuring character and background consistency, and allows users to control the number, duration, and content of shots through shot-specific conditioning. This is achieved by incorporating a transition token into the text-to-video model to control at which frames a new shot begins and a local attention masking strategy which con + +trols the transition token's effect and allows shot-specific prompting. To obtain training data we propose a novel data collection pipeline to construct a multi-shot video dataset from existing single-shot video datasets. Extensive experiments demonstrate that fine-tuning a pre-trained text-to-video model for a few thousand iterations is enough for the model to subsequently be able to generate multi-shot videos with shot-specific control, outperforming the baselines. You can find more details in our webpage. + +# 1. Introduction + +While diffusion models [18, 38] have shown impressive capabilities in the image domain [2, 12, 17, 33, 37, 39, 40, 51, 55, 59], extending them to video synthesis presents significant challenges due to the dynamic nature of videos. One group of researchers has explored various methods to adapt text-to-image (T2I) models for video synthesis [14, + +In contrast, another group has focused on text-to-video (T2V) diffusion models [4, 20, 34] which demonstrate superior performance by processing the entire video as a unified input rather than handling each frame independently. Despite their state-of-the-art performance in video generation, all existing models are designed to generate a single continuous video, which imposes inherent limitations when attempting to generate a multi-shot video of the same character engaged in multiple distinct activities, where each shot is separated by discrete cuts in dynamic elements. This challenge becomes particularly significant when each activity requires a unique setting, even within the same background (Fig. 2, $1^{\mathrm{st}}$ column: a man writes code and then sketches diagrams on a whiteboard) or when the background needs to change while maintaining the same identity (Fig. 2, $2^{\mathrm{nd}}$ column: transitioning from a home gym to a park bench). Moreover, existing models are limited to generating videos of very short durations, and therefore lack flexibility. Because they have been trained exclusively on single-shot videos, they are ill-suited for multi-shot video synthesis. Consequently, these limitations hinder their applicability in real-world contexts, such as film production, where narratives rely on multiple shots, each featuring distinct actions or perspectives and often featuring the same character across diverse scenes and time frames. + +The simplest approach to multi-shot synthesis is to generate a single-shot video by combining all shot-specific prompts into a single prompt. However, this method cannot create "cuts" (transitions between shots) and struggles to provide different backgrounds for the same character. Even within a same background, it cannot generate characters featuring different activities requiring distinct settings, as it is constrained to generating short videos (Fig. 2 (a)). An extension of this approach is to generate each shot many times individually, then concatenate the most similar shots, which requires the generation of numerous videos (Fig. 2 (b)). An improvement to this baseline entails first generating consistent keyframes of a character using a reference image, then animating these keyframes with an image-to-video (I2V) model, and finally concatenating the generated clips to produce a multi-shot video. However, this baseline is inevitably limited by the capabilities of off-the-shelf methods, as they struggle with maintaining consistency and quality. Additionally, it remains challenging to depict the character within the same background performing distinct activities (Fig. 2 (c)). + +Given the limitations of potential baselines, we introduce ShotAdapter, a simple yet powerful, model-agnostic framework for controllable text-to-multi-shot video (T2MSV) generation. ShotAdapter transforms a single-shot T2V generator into a T2MSV generator with minimal fine-tuning, requiring only five thousand iterations (less than $1\%$ of the total pre-training iterations), on a multi- + +![](images/45f5696cd926a3dbb9ed568596109d56df657e7556bc039ee33f823d7fe4b518.jpg) +Figure 2. Comparison with baselines. Each row displays frames from generated 2-shot videos, guided by shot-specific text prompts. The left column shows results with same background and different activities, while the right column presents results both with different backgrounds and activities. + +shot video dataset. This transformation is made possible through a novel attention-layer masking strategy and a special learnable token, called "transition token", which signals the transition between shots. It enables multi-shot video generation where the camera perspective can shift abruptly between shots (Fig. 1, $2^{\text{nd}}$ row) within a single background with the character performing different activities, or the background itself can change (Fig. 1, $4^{\text{th}}$ row), all while preserving the character's identity. Notably, ShotAdapter offers users precise control over the content of each generated shot through shot-specific text prompts, along with the flexibility to specify the number and duration of shots. Additionally, the use of a unified input for the T2V model ensures consistency across shots with minimal fine-tuning on an appropriate multi-shot video dataset. + +To address the lack of training datasets, we propose two novel pipelines to construct a multi-shot video dataset from a single-shot video dataset through pre- and post-processing steps, which eliminates the need to collect actual multishot video. Recognizing the lack of an evaluation standard for T2MSV generation, we propose an evaluation pipeline and several baseline comparisons to assess synthesis results based on identity and background consistency, text alignment, and shot-duration precision. In summary, our main contributions are as follows: + +- We propose a model-agnostic and computationally efficient framework that transforms a T2V model into a T2MSV generator with minimal fine-tuning, ensuring the preservation of character identity across all generated shots. +- Our approach allows user-defined control over the number of shots, their duration, and the content of each shot through text prompts, leveraging a novel "transition token" coupled with a localized attention masking strategy. +- We propose two novel data collection strategies to con + +struct a multi-shot video dataset derived from single-shot video dataset, along with pre- and post-processing steps. +- We believe we are the first to frame the T2MSV generation task as a challenge for the research community. Therefore, we introduce an evaluation pipeline to assess performance in this domain and will release a validation dataset to promote standardized evaluation. + +# 2. Related Works + +Text guided video synthesis Text-guided video synthesis is a rapidly evolving field. Early methods adapted pre-trained text-to-image (T2I) diffusion models [38] for video synthesis by modifying architecture, such as adding temporal layers [52], altering latent inputs [22], and introducing temporally correlated noise for frame consistency [13]. However, T2I models struggle with frame coherence due to their lack of temporal processing. Recent approaches, such as T2V models using unified video inputs and Diffusion Transformer (DiT) frameworks, address this. OpenAI's SORA [4] and Gen-Tron [7] enhance DiT with large datasets and advanced text-conditioning. Commercial models like Luma Dream Machine [24], MiniMax [29], and Kling AI [1] achieve coherence through extensive training. MovieGen [34] enables instruction-based editing and personalized generation. Open-source contributions have advanced T2V methods, such as W.A.L.T [15], which uses a two-stage algorithm for training on image and video datasets, and Latte [28], which employs Transformer blocks. RIVER [9] and PyramidFlow [20] use autoregressive generation with flow-matching, while OpenSora [60] and OpenSora-Plan [23] aim to replicate SORA more efficiently. In contrast to prior works that focus on single-shot video generation, we introduce a framework for fine-tuning T2V models, transforming them into T2MSV generators. + +Image/Video story generation Personalized identity-preserving image generation presents challenges in maintaining consistent identity across diverse settings. Techniques like DreamBooth [39] and Textual Inversion [12] achieve this through time-consuming test-time fine-tuning. Some line of work [27, 51, 53] use adapter models whereas other works [25, 49, 54, 56] incorporate ArcFace [10] or keypoints on face features for conditioning. In image story generation, StoryMaker [62] ensures stylistic consistency, while ConsiStory [44] enhances subject consistency through a shared attention block and feature injection. DreamStory [16] leverages Large Language Models, and StoryDiffusion [61] introduces an attention mechanism for consistent character representation. In video story generation, DreamBooth [39], adapted in Tune-A-Video [52], requires fine-tuning for each input video, limiting scalability and restricting outputs to short, single-shot videos based on reference images. Recognizing the lack of a unified solution for multi-shot video generation, one of our base + +lines combines StoryMaker [62] with an I2V model, though this integration shows limited performance. Our approach achieves superior multi-shot video generation results compared to this baseline. + +# 3. Methodology + +# 3.1. Preliminaries + +Diffusion formulation Diffusion models [18, 38, 42] generate realistic data by learning to reverse a noise-adding process. During training, data is progressively noised over multiple time steps, and the model learns to denoise it step by step, starting from pure noise $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ . A noisy version $\mathbf{x}_i^{(t)}$ at timestep $t$ is obtained by adding scaled noise $\mathbf{x}_i^{(t)} = \sqrt{\alpha_t}\mathbf{x}_i + \sqrt{1 - \alpha_t}\epsilon$ , where $\alpha_{t}$ controls the noise level. The training loss $\mathcal{L}_{\mathrm{diffusion}}(\theta) = \mathbb{E}_{\mathbf{x}_i,\epsilon ,t}\left\| \epsilon -\epsilon_\theta (\mathbf{x}_i^{(t)},t)\right\|^2$ minimizes the difference between the true noise $\epsilon$ and the model's prediction $\epsilon_{\theta}(\mathbf{x}_i^{(t)},t)$ . This loss trains the model to predict the noise, enabling it to reconstruct the original data through reverse diffusion. + +Latent video diffusion models Given an input video sample $\mathbf{x}_i\sim p_{data}(\mathbf{x})$ with dimensions $\mathbf{x}_i\in \mathbb{R}^{F\times C\times H\times W}$ , the process begins by encoding the video using a 3D encoder $\mathcal{E}$ , yielding a latent representation $\mathbf{z}_i\in \mathbb{R}^{F'\times C'\times H'\times W'}$ , where $F' = F / f_t$ , $H' = H / f_s$ , and $W' = W / f_s$ , with $f_s$ and $f_t$ being the spatial and temporal compression ratios, respectively. The latent video representation is further transformed into a sequence of $N$ tokens for a DiT-based model, denoted as $\{\overline{\mathbf{x}}_i^n\}_{n = 1}^n$ , where each token $\overline{\mathbf{x}}_i^n\in \mathbb{R}^D$ has a hidden dimension $D$ . Patchification is performed along the width, height, and frame depth, patch sizes represented by $f_{p_w}$ , $f_{p_h}$ , and $f_{p_f}$ , respectively. + +# 3.2. Text-to-Multi-Shot Video Generation + +Inline with the definitions in previous works [6, 45, 57], a shot is defined as the smallest segment of a video—a single, uninterrupted clip with continuous motion. Each shot is characterized by dynamic elements, including the foreground object, background setting, object actions, and camera movement. A multi-shot video is defined as a video composed of multiple individual shots, each being separated by a "cut" [45], which is an instantaneous change in the dynamic elements. More formally, an N-shot video $V$ can be represented as $V = \{s_1^{K_1},\ldots ,s_N^{K_N}\}$ , where each shot $s_i^{K_i}$ is a set of $K_{i}$ frames, $s_i^{K_i} = \{f_1,\dots ,f_{K_i}\}$ , with each frame $f_{j}\in \mathbb{R}^{C\times H\times W}$ representing an image with channel (C), height (H), and width (W). Our work focuses on multi-shot videos featuring a single foreground object, specifically humans, as they are often the main characters and present challenges in maintaining identity consistency for real-world applications. We aim to preserve the identity of the foreground object across shots, even as it performs + +![](images/f743151502820377e0d4eb368b5652443f98a703b33a8b86f38641fce352046f.jpg) +Figure 3. Fine-tuning framework with transition token and local attention masking. (a) ShotAdapter fine-tunes a pre-trained T2V model by incorporating "transition tokens" (highlighted in light blue). We use $n - 1$ transition tokens, initialized as learnable parameters, alongside an $n$ -shot video with shot-specific prompts, which are fed through the pre-trained T2V model. (b) The model processes the concatenated input token sequence, guided by a "local attention mask" through joint attention layers within DiT blocks. (c) The local attention mask is structured to ensure that transition tokens interact only with the visual frames where transitions occur, while each textual token interacts exclusively with its corresponding visual tokens. + +different activities. Furthermore, when the background is designated as a fixed location, we require background consistency throughout the entire video. More formally, our task is to generate an N-shot video given a set of input conditions $\{C_1,C_2,\ldots \}$ where each $C_i$ denotes shot-specific conditions, e.g. $C_i = \{\mathrm{shot~caption},\mathrm{shot~duration},\dots \}$ . + +ShotAdapter introduces model-agnostic novel extensions to transform pre-trained T2V diffusion models into T2MSV generators with minimal fine-tuning along with a multi-shot video dataset curation method. Our method allows users to control the number and duration of each shot as well as the content through shot-specific text prompts. + +# 3.3. Model + +Model architecture We use a diffusion transformer (DiT) [32] based T2V framework similar to OpenSora [60] and MovieGen [34], shown in Fig. 3 (a). Our model integrates a 3D Variational Autoencoder (3D-VAE) for video encoding, along with a variant of joint-attention layers [11] for conditioning. Specifically, input videos are first subject to temporal and spatial encoding via the 3D encoder before being patched. The textual condition tokens are then concatenated with these patched visual tokens and subsequently processed by the DiT. We leverage a pre-trained T2V model that can generate 128 frames with $192 \times 320$ resolution, and fine-tune it to enable multi-shot generation. This extension is achieved by introducing a transition token and implementing local attention masking. Note that our framework also supports cross-attention based conditioning. + +Transition token Inspired by the commonly employed '[EOS]' (End of Sentence) token in natural language processing, which signals the model to recognize the end of a + +sentence, we propose a novel, learnable embedding referred to as the "transition token" which enables the model to learn transitions between consecutive shots within a multi-shot video. Specifically, we initialize a set of learnable parameters at the start of fine-tuning, matching the hidden dimension of the input tokens. We repeat these parameters $n - 1$ times—where $n$ represents the specified number of shots—and append them to the end of the input visual and textual token sequence (Fig. 3 (a)). In the model's attention layers, we implement a masking strategy that ensures the transition token interacts only with the tokens corresponding to the frames where transitions are intended to occur (Fig. 3 (b) and (c)). This approach allows the model to focus on transition frames, enabling it to learn to generate cuts between shots, while also allowing users to specify both the number of shots and their respective durations, by simply replicating the transition token and adjusting the attention mask accordingly. + +Local attention masking To enable shot-specific control, we introduce a local attention masking technique (Fig. 3 (c)). Without masking, all tokens interact with each other, diluting the impact of shot-specific information. Our proposed local attention masking strategy overcomes this limitation by restricting attention interactions to specific token groups. Specifically, the attention matrix is masked to enforce the interactions, where the transition token attends only to the tokens at the transition frames, while visual and textual tokens are restricted to self-attention. Additionally, to enhance precise and localized control of textual tokens over their corresponding shot-visual tokens, each textual token attends solely to its associated visual tokens. Consequently, the fine-tuned model is better equipped to capture fine-grained dependencies between frames, shot + +![](images/7c43fefd12a7ad2e1fd9427275bec97959015a48648fc15429db179fb3c87e3e.jpg) + +![](images/a34842426cc0ffbdce1947ea6611d46a028ceda8037b91a833f3cbeaf981b430.jpg) +Figure 4. Multi-shot video dataset collection pipeline. A high-level overview of this pipeline is presented in (a). Our first method (gray box in (b)) samples videos with large motion, randomly splits them into $n$ -shots with varied durations, and concatenates them into multi-shot videos. Our second method (yellow box in (b)) randomly samples $n$ videos from pre-clustered groups containing videos of the same identities and concatenates them to form a multi-shot video. Finally, we post-process (c) the multi-shot videos to ensure identity consistency and obtain shot-specific captions using LLaVA-NeXT [58]. + +specific text prompts, and shot transitions, resulting in improved control over the T2MSV generation. + +# 3.4.Data + +To perform fine-tuning we require a suitable dataset consisting of multi-shot videos where each shot depicts the same identity both in same and different backgrounds, featuring different activities. We develop two methods to curate multi-shot videos given access to a large-scale video dataset (Fig. 4 (c)). The first method creates multi-shot videos from long single-shot videos that exhibit large motion while the second method combines multiple independent videos of the same human subjects to produce multishot videos. Post-processing is then applied to obtain shot-specific captions and ensure that the foreground identity remains consistent across shots. + +Multi-shot videos from single-shot videos In this approach we generate a multi-shot video from a single-shot video with large motion by randomly trimming short sub-clips and combining them to create a new n-shot video. (Fig. 4 (a), gray box). We first obtain optical flow maps using RAFT [43] and compute the average motion in the x and y directions $(t_x, t_y)$ along with scaling ratio $s$ for each video (see Supplementary for details). Videos with motion metrics below specified thresholds are discarded, retaining only those with significant motion in at least one metric. Additionally, videos with fewer than 250 frames are excluded to ensure sufficient variability. Finally, we choose random sub-clips of random duration and concatenate them in a random order to generate a dataset of multi-shot videos. + +Multi-shot videos from independent videos Despite focusing on large-motion videos, the previous approach encounters limitations in terms of diversity in camera perspective, motion and backgrounds, reducing the variety of generated multi-shot videos. To enhance the dataset variability we perform clustering on the single-video dataset, i.e. $\mathcal{D} = \{G_1^{M_1},\dots ,G_n^{M_n}\}$ , where each cluster $G_{j}^{M_{j}}$ contains $M_{j}$ videos featuring the same foreground object. These clusters are constructed by choosing videos uploaded by the same user within the same time-frame (e.g. 3 days), and videos that match based on the similarity of their captions. we obtain approximately 550K clusters, with an average of 6 videos per cluster. Subsequently, we generate a fixed number of multi-shot videos for each cluster by randomly trimming clips from separate videos within the same cluster (Fig. 4 (a), gray-yellow). This methodology significantly enhances the dataset's richness and variability, introducing a wider array of scenes and perspectives. + +Post-processing We apply several post-processing pipelines (Fig. 4 (c)) to (i) obtain shot-specific prompts and (ii) ensure identity consistency across shots. Since each shot in the multi-shot video dataset requires a specific text description, we employ the LLaVA-NeXT [58] model to generate shot-specific captions. Additionally, we apply an additional filtering stage to the multi-shot video dataset to ensure identity consistency across shots. Videos containing more than one character or featuring different identities are filtered out usingYOLO [46] as a person detector in the middle frames of each shot within a multi-shot video. To verify identity consistency, we utilize embeddings from DINOv2 and applied a threshold to ensure the same + +![](images/0f28e9dc8696c0541ed88cf454ac6cea664b2af275038516ceea31ac2892c037.jpg) +Figure 5. Qualitative results. Our approach enables multi-shot video generation depicting different actions and background guided by shot-specific prompts. In the $2^{\mathrm{nd}}$ row, the shots maintain a consistent background while capturing different perspectives, whereas the $3^{\mathrm{rd}}$ row depicts the same woman in related backgrounds that subtly change in response to the prompt. For complete videos, see Supplementary. + +identity appears across all shots. This post-processing step filters out $38\%$ of the multi-shot videos, resulting in a dataset composed of multi-shot videos with 2, 3, and 4 shots of varying durations, where a single consistent identity performs similar or different activities across diverse backgrounds and motions. + +# 4. Experiments + +Since there is no existing evaluation pipeline for the T2MSV generation task, we design a benchmark to enable standardized evaluations (see Supplementary for implementation details). + +Evaluation dataset Our dataset is partitioned according to whether the background remains constant or changes and to the number of shots as specified in the text prompts. For each multi-shot configuration (2, 3, and 4 shots) we provide 8 prompts per scenario (i.e. background remains the same or changes) resulting in a total of 48 prompts for multi-shot settings. These prompts are designed to depict a human subject by including terms such as "a person" performing different activities in each shot using ChatGPT [31] (see Supplementary). In total, we generate 128 frames for each sample, with randomly selected shot durations. + +Baselines To evaluate our approach in the absence of directly comparable methods we devise three baselines. Random Shots (RS) generates each shot guided by detailed text descriptions as a single-shot video multiple times (i.e. + +48) and concatenates the randomly selected shots. Similar Shots (SS) improves on this by selecting shots based on DINOv2 embedding similarity across foreground objects. Shots by Reference (SR) first generates keyframes of consistent characters using StoryMaker [62] and then animates those into individual video shots using our model's I2V capability. Note that for all baselines, we use the original pre-trained model, providing a fair comparison by using the same underlying model architecture. + +Quantitative evaluation We adapt commonly used metrics from single-shot video generation [19]: (a) Identity Consistency (IC), which calculates the average DINOv2 [5] embedding similarity between the segmented persons (segmented usingYOLO [46]) at the middle frames of each shot; (b) Text Alignment (TA), which assesses the alignment of generated content with text prompts by calculating the similarity between text features and shot features extracted by ViCLIP [50], then averaging across shots; and (c) Background Consistency (BC), which measures similarity by segmenting the background and computing DINOv2 embedding similarity across the middle frames of each shot. Our approach outperforms all previous baselines in preserving identity for 3- and 4-shots, while achieving competitive results with SR for 2-shots, as measured by the IC metric (Table 1). As the number of shots increases, there is a decline in performance due to the propagation and accumulation of errors in generating consistent identities. Addi + +![](images/64e543778500318d3cb94b59eecc103eb0588e6374ad66bdfc2ce3f839856e5a.jpg) +Figure 6. Qualitative comparison. We compare our approach to single-shot generation and baseline methods for (a) 2-shot, (b) 3-shot, and (c) 4-shot videos. Our model (last row) enables local control over shot content while Single-Shot Generation fails to feature distinct activities ( $1^{\text{st}}$ row). Random Shots ( $2^{\text{nd}}$ row) and Similar Shots ( $3^{\text{rd}}$ row) struggle to preserve character identity, especially facial features (cropped below frames in red bounding boxes), which our method effectively maintains. Shots by Reference ( $4^{\text{th}}$ row) improves identity consistency to some extent but falls short in maintaining both identity (e.g. shot 4 in (c)) and background coherence (e.g. shot 3 in (b)), where our model demonstrates superior performance. For the full videos, please see Supplementary. + +tionally, when the background changes (diff $bg$ ), a modest overall decrease in performance is observed. We report background consistency for samples where the background is intended to remain constant (same $bg$ ), where our model outperforms all other approaches across every shot by a large margin. In terms of text alignment, RS and SS generally perform better, as expected, since they generate each shot individually, effectively serving as an upper bound for this score. However, our approach achieves competitive results, demonstrating that it preserves identity and background consistency more effectively without significantly compromising text alignment. For additional comparisons to SEINE [8], MEVG [30], FreeNoise [36] and Gen-L-Video [48], see Supplementary. + +User study To complement the quantitative metrics, we conduct a user study on the Prolific [35] platform with 75 participants. Each participant views two videos simultaneously, selected from a pool of 10 randomly chosen videos from the generated results, with one video always generated by ShotAdapter. Participants are then asked to choose their preferred video based on identity consistency (IC), background consistency (BC), and text alignment (TA). Our + +approach achieves superior results in a 1-to-1 comparison with baselines (Table 1) in identity and background consistency. In terms of text alignment, it achieves a slight improvement over SS and RS, while outperforming SR by a substantial margin, confirming the trends observed in the quantitative metrics. + +Ablation study We conduct two ablation studies: (i) removing the transition token while retaining the local attention mask (Table 1 ShotAdapter w/o TT), and (ii) fine-tuning the model exclusively on 2-shot videos (Table 1 ShotAdapter w/ 2-shots). Including transition tokens yields a slight improvement in IC, BC, and TA, as it assists the model in generating cuts, thereby enhancing localized control over shot transitions. Although fine-tuning on only 2-shot videos reduces the dataset size, results on 3- and 4-shot videos reveal the model's generalizability, maintaining better identity and background consistency overall than the baselines, despite a slight performance decrease compared to the final model. + +Qualitative evaluation Fig. 1 and Fig. 5 show our model's T2MSV generation capabilities across 2, 3, and 4-shot videos, addressing scenarios that require either background + +Table 1. Quantitative comparison. We evaluate our approach against baseline methods across Identity Consistency, Background (BG) Consistency, Text Alignment and a User Study on 2, 3, and 4-shot videos under conditions with background changes (diff bg), without background changes (same bg), and with both types (all). ShotAdapter w/o TT indicates the model fine-tuned without Transition Tokens (TT) and ShotAdapter w/ 2-shots fine-tuned only on a dataset of 2-shot videos. The user study assesses Identity Consistency (Q1), Background Consistency (Q2), and Text Alignment (Q3). $\uparrow$ and $\downarrow$ indicate the direction toward better performance for each metric. + +
Shot Number Background ChangeIdentity Consistency ↑BG Consistency ↑Text Alignment ↑User Study
diff bg2 same bg3 diff bg4 same bgdiff bg4 same bg2 same bg34234Q1 (IC)Q2 (BC)Q3 (TA)
Ours vs Baseline(Selection ratio in 1-to-1 comparison)
Random Shots (RS)71.0380.4754.7663.7248.0855.8784.4665.7759.1826.8426.4725.4477.19% / 22.81%73.27% / 26.73%56.73% / 43.27%
Similar Shots (SS)73.9482.5555.1566.1749.2558.6788.8567.0260.2026.4026.1325.1672.92% / 27.08%69.20% / 30.80%53.13% / 46.87%
Shots by Reference (SR)81.7484.9867.9272.9757.8367.7482.1164.8556.8125.5923.9721.9873.43% / 26.57%82.28% / 17.72%73.03% / 26.97%
ShotAdapter w/o TT77.1784.7868.9570.9858.8370.2487.9472.9370.4826.6423.1522.84N/AN/AN/A
ShotAdapter w/ 2-shots78.0585.4670.1271.5356.9968.3789.0875.5373.1925.9723.5922.97N/AN/AN/A
ShotAdapter78.6786.3370.3076.4461.8674.8989.4877.6676.5527.1223.6522.17N/AN/AN/A
+ +Table 2. Transition token generalizability. We compute the absolute difference in frames between the generated and ground truth shot duration as the Mean Shot Duration Error (MSDE), and the error per-shot is reported with a range of 2 to 8 shots per video. + +
Shots2345678
MSDE2.000.831.001.701.330.921.21
+ +transitions (e.g. a transition from a living room to a walk-in closet) or distinct activities within the same setting (e.g. a character lifting weights, drinking water, and doing push-ups). Our model effectively generates multi-shot videos of the same characters, with "cuts" even in diverse settings (full videos are available in the Supplementary). In Fig. 6, we provide a qualitative comparison of our approach with single-shot video generation using extended prompts (row 1) and baseline methods RS, SS and SR for 2, 3, and 4-shot videos. Single-shot generation with localized shot control can result in scenes where actions are intermingled (Fig. 6 (a), where the model fails to transition from walking to eating ramen). RS produces the least consistent outcomes, generating characters with random identities (Fig. 6, zoomed-in faces highlighted with red bounding boxes) and incoherent backgrounds due to the random concatenation of shots. SS shows a minor improvement in identity consistency by selecting shots based on foreground similarity, yet still generates visually distinct identities, despite similar clothing (Fig. 6 (b)), and struggles to maintain coherent backgrounds. SR, achieves better identity consistency than previous baselines as it generates consistent characters using an off-the-shelf method [62] but suffers from quality degradation as the number of shots increases and lacks temporal coherence between keyframes, resulting in notable inconsistencies, such as complete environment changes (Fig. 6 (c)). In contrast, our approach effectively addresses these limitations, generating multi-shot videos with consistent character identity across different background requirements as directed by text prompts. + +Transition token generalizability To further assess the generalizability of the transition token, we test our model on videos with 2 to 8 shots, using the Mean Shot Duration Error (MSDE) metric, which is calculated by averaging the absolute difference between ground-truth and generated + +shot durations in terms of frames, where SceneCut [3] is used to detect the cuts. Quantitative results (Table 2) show that despite the temporal compression applied during encoding and patchification, the model achieves an average offset of only 1 to 2 frames per shot, even in 8-shot examples. These findings confirm that the transition token serves effectively as an "End of Shot" marker and can be extended to accommodate multiple shots. + +Limitations While our experiments demonstrate the effectiveness of our approach in T2MSV generation, this study is limited to human foreground objects, as experiments with non-human subjects (e.g. animals) were not conducted. This limitation is primarily due to dataset filtering choices. Additionally, the maximum duration the model can generate is restricted by the underlying model used for fine-tuning, which is limited to 128 frames in this study. For future work, we aim to extend the duration by employing an autoregressive approach to generate additional shots conditioned on previously generated ones. It is worth noting that our method experiences a slight quality reduction in the user study compared to baselines, though it remains highly competitive. We hypothesize that this minor drop is primarily due to fine-tuning with a $90\%$ smaller batch size compared to the baseline as well as our model better adhering to multiple text captions while the baselines often ignore larger parts of the text (see Supplementary for quality analysis). + +# 5. Conclusion + +In this paper, we present ShotAdapter, a lightweight framework that transforms single-shot T2V models into multi-shot T2MSV generators with minimal fine-tuning. Our approach incorporates a transition token and localized attention masking, applied to a multi-shot video dataset collected through a novel data collection pipeline. Extensive evaluations demonstrate that our method outperforms baseline models in identity and background consistency without compromising text alignment scores, as further validated by a user study. Additionally, our findings highlight the framework's generalizability to videos with an increasing number of shots, affirming the effectiveness of the "transition token" concept. + +# References + +[1] Kling AI. Kling ai. https://klingai.com/, 2024. Accessed: 2024-11-04. 3 +[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 1 +[3] Breakthrough. Pyscenedetect: A cross-platform tool for video scene detection. https://github.com/ Breakthrough/PySceneDetect, 2024. Accessed: 2024-11-04. 8 +[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 2, 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 IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6 +[6] Vasileios T Chasanis, Aristidis C Likas, and Nikolaos P Galatsanos. Scene detection in videos using shot clustering and sequence alignment. IEEE transactions on multimedia, 11(1):89-100, 2008. 3 +[7] Shoufa Chen, Mengmeng Xu, Jiawei Ren, Yuren Cong, Sen He, Yanping Xie, Animesh Sinha, Ping Luo, Tao Xiang, and Juan-Manuel Perez-Rua. Gentron: Diffusion transformers for image and video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6441-6451, 2024. 3 +[8] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 7, 2 +[9] Aram Davtyan, Sepehr Sameni, and Paolo Favaro. Efficient video prediction via sparsely conditioned flow matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23263-23274, 2023. 3 +[10] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3 +[11] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 4 +[12] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image gener + +ation using textual inversion. In The Eleventh International Conference on Learning Representations. 1, 3 +[13] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, MingYu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 3 +[14] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In The Twelfth International Conference on Learning Representations, 2024. 1 +[15] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 3 +[16] Huiguo He, Huan Yang, Zixi Tuo, Yuan Zhou, Qiuyue Wang, Yuhang Zhang, Zeyu Liu, Wenhao Huang, Hongyang Chao, and Jian Yin. Dreamstory: Open-domain story visualization by llm-guided multi-subject consistent diffusion. arXiv preprint arXiv:2407.12899, 2024. 3 +[17] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3 +[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6 +[20] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 2, 3 +[21] Ozgur Kara, Bariscan Kurtkaya, Hidir Yesiltepe, James M Rehg, and Pinar Yanardag. Rave: Randomized noise shuffling for fast and consistent video editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6507-6516, 2024. 2 +[22] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2, 3 +[23] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 3 +[24] Luma Labs. Dream machine. https://lumalabs.ai/dream-machine, 2024. Accessed: 2024-11-04. 3 +[25] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing re + +alistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 3 +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 1 +[27] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3 +[28] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3 +[29] MiniMax. Minimax. https://hailuoai.com/, 2024. Accessed: 2024-11-04. 3 +[30] Gyeongrok Oh, Jaehwan Jeong, Sieun Kim, Wonmin Byeon, Jinkyu Kim, Sungwoong Kim, and Sangpil Kim. Mevg: Multi-event video generation with text-to-video models. In European Conference on Computer Vision, pages 401-418. Springer, 2024. 7, 2 +[31] OpenAI. Chatgpt: A large language model. https://openai.com/chatgpt, 2024. Accessed: 2024-11-04. 6 +[32] 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 +[33] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations. 1 +[34] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2, 3, 4 +[35] Prolific. Prolific: Online participant recruitment for surveys and research. https://prolific.com/, 2024. Accessed: 2024-11-01. 7 +[36] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. In The Twelfth International Conference on Learning Representations, 2024. 7, 2 +[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv e-prints, pages arXiv-2204, 2022. 1 +[38] 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. 1, 3 +[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven + +generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3 +[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 1 +[41] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, 2023. 2 +[42] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3 +[43] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5, 1 +[44] Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consistent text-to-image generation. ACM Transactions on Graphics (TOG), 43(4):1-18, 2024. 3 +[45] Kristin Thompson, David Bordwell, and Jeff Smith. Film history: An introduction. McGraw-Hill Boston, 2003. 3 +[46] Ultralytics. YOLOv5: A state-of-the-art real-time object detection system. https://docs.ultralytics.com, 2021. Accessed: 2024-11-04. 5, 6 +[47] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation, 2019. 1 +[48] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising, 2023. 7, 2 +[49] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3 +[50] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, 2024. 6 +[51] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 1, 3 +[52] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu + +Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2, 3 +[53] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. International Journal of Computer Vision, pages 1-20, 2024. 3 +[54] Yuxuan Yan, Chi Zhang, Rui Wang, Yichao Zhou, Gege Zhang, Pei Cheng, Gang Yu, and Bin Fu. Facestudio: Put your face everywhere in seconds. arXiv preprint arXiv:2312.02663, 2023. 3 +[55] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1 +[56] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 3 +[57] HongJiang Zhang, Atreyi Kankanhalli, and Stephen W Smoliar. Automatic partitioning of full-motion video. Multimedia systems, 1:10-28, 1993. 3 +[58] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model. https://llava-v1.github.io/blog/2024-04-30-llava-last-video/, 2024.5 +[59] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 1 +[60] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. https://github.com/hpcaitech/Open-Sora, 2024.3,4 +[61] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. arXiv preprint arXiv:2405.01434, 2024. 3 +[62] Zhengguang Zhou, Jing Li, Huaxia Li, Nemo Chen, and Xu Tang. Storymaker: Towards holistic consistent characters in text-to-image generation. arXiv preprint arXiv:2409.12576, 2024. 3, 6, 8, 1 \ No newline at end of file diff --git a/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/images.zip b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b37fd5c6a33dcf37029b1945ae882cf2c444414 --- /dev/null +++ b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:414af95beddc1698a7d4c47a05311e010ad69d4b33c2c11394caf6f5f079ff16 +size 1057732 diff --git a/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/layout.json b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a7271ee7fbb63e8388f98288ae2218222ec4ee8f --- /dev/null +++ b/CVPR/2025/ShotAdapter_ Text-to-Multi-Shot Video Generation with Diffusion Models/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28ffbad5b37a254bfa5dc3b28098f662d80f1c88c0a1489d3ad40c8ccabe2484 +size 324859 diff --git a/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/e0ad8291-f3c4-4d3c-9431-5230f4f74cc1_content_list.json b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/e0ad8291-f3c4-4d3c-9431-5230f4f74cc1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8aeb9e073022a89db9e318982c8355d4eaac03b7 --- /dev/null +++ b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/e0ad8291-f3c4-4d3c-9431-5230f4f74cc1_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10d49aba6c01adea1b4cdaa560cb6242f944caacd7189164cbfd3cce699b37eb +size 84156 diff --git a/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/e0ad8291-f3c4-4d3c-9431-5230f4f74cc1_model.json b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/e0ad8291-f3c4-4d3c-9431-5230f4f74cc1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..32a8a352d5b7e71b26a1f89283f3930fea0b33a7 --- /dev/null +++ b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/e0ad8291-f3c4-4d3c-9431-5230f4f74cc1_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d29da95af332434bc00c3c745b5b580cfe61f4acbd32ab0ebf1188aa4a71d446 +size 104173 diff --git a/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/e0ad8291-f3c4-4d3c-9431-5230f4f74cc1_origin.pdf b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/e0ad8291-f3c4-4d3c-9431-5230f4f74cc1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9fa786b16f316b30f2257cdbaaa52f22bf2e5449 --- /dev/null +++ b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/e0ad8291-f3c4-4d3c-9431-5230f4f74cc1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a4ade68ec95daca7d597a375ee8f49ed5abefaf555e9ba031f5e933cf2e65b1 +size 3955698 diff --git a/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/full.md b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..965dbdd6c3b81821626c3a80a678fc205ab540b8 --- /dev/null +++ b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/full.md @@ -0,0 +1,307 @@ +# Show and Segment: Universal Medical Image Segmentation via In-Context Learning + +Yunhe Gao $^{2}$ , Di Liu $^{2}$ , Zhuowei Li $^{2}$ , Yunsheng Li $^{1}$ , Dongdong Chen $^{1}$ , Mu Zhou $^{2}$ , Dimitris N. Metaxas $^{2}$ + +$^{1}$ Microsoft GenAI $^{2}$ Rutgers University + +# Abstract + +Medical image segmentation remains challenging due to the vast diversity of anatomical structures, imaging modalities, and segmentation tasks. While deep learning has made significant advances, current approaches struggle to generalize as they require task-specific training or fine-tuning on unseen classes. We present Iris, a novel In-context Reference Image guided Segmentation framework that enables flexible adaptation to novel tasks through the use of reference examples without fine-tuning. At its core, Iris features a lightweight context task encoding module that distills task-specific information from reference context image-label pairs. This rich context embedding information is used to guide the segmentation of target objects. By decoupling task encoding from inference, Iris supports diverse strategies from one-shot inference and context example ensemble to object-level context example retrieval and in-context tuning. Through comprehensive evaluation across twelve datasets, we demonstrate that Iris performs strongly compared to task-specific models on in-distribution tasks. On seven held-out datasets, Iris shows superior generalization to out-of-distribution data and unseen classes. Further, Iris's task encoding module can automatically discover anatomical relationships across datasets and modalities, offering insights into medical objects without explicit anatomical supervision. + +# 1. Introduction + +The accurate segmentation of anatomical structures in medical images is fundamental for clinical practice and biomedical research, enabling precise diagnosis [11, 41] and treatment planning [37]. While deep learning has demonstrated remarkable success [32, 46], the vast diversity of anatomical structures, imaging modalities, and clinical tasks poses long-standing challenges for developing truly generalizable solutions. Current efforts typically focus on disease-specific tasks or a limited set of anatomical structures [7, 13, 14, 17, 20, 27-29, 59], struggling to handle + +![](images/8c2df9163727a0712c3817d1bf3d43343f8672b1becc9c93b7e871ebf0b0ad3e.jpg) + +![](images/2fc8dee0f6fdaf7d608f8c7e664ce9e39fe6094ce2353e1eaef26a5bc8f9ddc3.jpg) + +![](images/daddbf573fd5303ff628e735e06a10c93cdae3e307e3a67186a0e7d260d3f9b8.jpg) +Figure 1. Comparison of medical image segmentation approaches. A) Task-specific models require training separate models for each task, limiting their flexibility and scalability. B) Multi-task universal models can handle diverse tasks and imaging modalities, but fail on novel classes. C) SAM-based foundation models enable flexible segmentation through user interactions, but impractical for high-throughput automated processing. D) Our proposed Iris combines automatic processing with flexible adaptation via in-context learning, enabling both seen and unseen task segmentation without any manual interaction or retraining. + +![](images/dbfcc3ec587e8740283aa73eccd857305b3d30ab2f2b60b70ca0fd117380a23b.jpg) + +the heterogeneous landscape of medical imaging that spans diverse modalities, body regions, and diseases [38, 52]. + +These task-specific methods show critical limitations compared to human experts' capabilities. First, existing models often perform poorly on out-of-distribution examples [60]—a common scenario in medical imaging where variations arise from different imaging centers, patient populations, and acquisition protocols. Second, traditional segmentation models, while achieving a high accuracy on their trained tasks, lack the adaptability to handle novel classes without extensive retraining or fine-tuning [56]. This dilemma fundamentally limits task-specific models' applicability in dynamic clinical settings and research environments, + +where new segmentation tasks continue to emerge over the course of real-world practice. + +Recent research has explored several directions to address these challenges (Figure 1). Universal medical segmentation models [30, 44, 50, 54] attempt to leverage synergy among multiple tasks across diverse datasets to learn robust representations, yet struggling with unseen classes and requiring fine-tuning. Foundation models with interactive capabilities, such as SAM [23] and its medical variants [9, 34, 45, 58], offer flexibility via user prompts. But they require multiple interactions for optimal segmentation results, especially for complex 3D structures, and lack the efficiency for large-scale automated analysis. In addition, in-context learning (ICL) methods [5, 39] show promise in automatically handling arbitrary new tasks through a few reference examples, but current methods exhibit suboptimal performance compared to task-specific models and suffer from computational inefficiencies, requiring expensive reference encoding during each inference step. + +To address these fundamental challenges, we present Iris framework for universal medical image segmentation via in-context learning. At its core, Iris features a lightweight task encoding module that efficiently distills task-specific information from reference image-label pairs into compact task embeddings, which then guide the segmentation of target objects. Unlike existing ICL methods [5, 39], Iris decouples the task definition from query image inference, eliminating redundant context encoding while enabling flexible inference strategies, all coming with high computational efficiency. + +Our main contributions include: + +- A novel in-context learning framework for 3D medical images, enabling a strong adaptation to arbitrary new segmentation tasks without model retraining or fine-tuning. +- A lightweight task encoding module that captures task-specific information from reference examples, handling medical objects of varying sizes and shapes. +- Multiple flexible inference strategies suitable for different practical scenarios, including one-shot inference, context ensemble, object-level context retrieval, and in-context tuning. +- Comprehensive experiments on 19 datasets demonstrate Iris's superior performance across both in-distribution and challenging scenarios, particularly on held-out domains and novel anatomical structures. It extends to reveal the capability of automatically discovering meaningful anatomical relationships across datasets and modalities. + +# 2. Related Work + +Medical Universal Models. Universal medical image segmentation models aim to address the data heterogeneity across tasks and modalities while learning generalizable feature representations. Early works include multi-dataset learning through label conditioning [12], organ size con + +straints [61], and pseudo-label co-training [19]. Recent works are placed on sophisticated task encoding strategies. DoDNet [54] pioneered one-hot task vectors with an extension into TransDoDNet [49] using transformer backbones. Advances include CLIP-driven models using semantic encodings [30], task-specific heads in MultiTalent [44], and modality priors in Hermes [13]. UniSeg [50] introduced learnable task prompts and MedUniseg [51] unified 2D/3D image handling. Despite of substantial efforts, these universal models all require fine-tuning when assessing unseen classes. In contrast, Iris enables the segmentation of unseen classes only through a single reference image-label pair without any model finetuning. + +SAM-based Interactive Models. Segment Anything Model (SAM) [23] emerges as a shifting paradigm of interactive segmentation via its prompt-based architecture and large-scale training. SAM's success has inspired a range of medical variants. Major examples include MedSAM [34] with 1.5M image-mask pairs for 2D segmentation, SAM-Med2D [9] trained on 4.6M images, and SAM-Med3D [45] extending to volumetric data with 22K 3D images. These efforts all require multiple prompts and interactive refinements, especially for analyzing complex objects in 3D scenarios. This interaction-dependent design becomes a bottleneck in high-throughput scenarios that an automated processing of large-scale datasets is much desired. As comparison, Iris addresses this limitation by defining tasks through context pairs, enabling fully automatic segmentation while maintaining a strong adaptability to new tasks. + +Visual In-context Learning. In-context learning as introduced by GPT-3 [4] enables models to handle novel tasks through example-guided inference without a heavy retraining. In the vision community, Painter [47] and SegGPT [48] pioneered in-context segmentation through a mask image modeling framework. Alternative methods [33, 43, 55] explored SAM-based approaches through cross-image correspondence prompting, but their two-stage pipeline introduces redundant computation and heavily relies on SAM's capabilities, limiting their applicability to 3D medical images. Neuralizer [10] develop general tools on diverse neuroimaging tasks, like super-resolution denosing, etc. Recent works introduced specialized architectures for in-context segmentation [36]. For example, UniverSeg [5] is designed for in-context medical image segmentation, and Tyche [39] incorporated a stochastic inference. While these methods demonstrate promising capability on novel classes, they face two critical limitations. First, they show suboptimal performance compared to task-specific models on the training distribution. Second, they suffer from computational inefficiencies as they can only segment one anatomical class per forward pass, requiring multiple passes for multi-class segmentation. Meanwhile they must re-encode context examples for each query image even when using the same + +![](images/ed7b0382f3eb6554286a5bf2d6b8bb14607c87715b6bd3b3512e5b1522f47a15.jpg) +Figure 2. Overview of Iris framework. We design a task encoding module to extract compact task embeddings from reference examples to guide query image segmentation with the mask decoding module, enabling efficient and flexible adaptation to new tasks without finetuning. + +reference examples repeatedly. This becomes particularly problematic in high-throughput scenarios where multiple query images need to be processed. In contrast, Iris shows better performance and efficiency. An appealing design of our context task encoding module is to decouple task definition from inference, enabling the encoding of task from reference pairs into task tokens that can be efficiently reused across any number of query images, meanwhile multi-class segmentation can be done within a single forward pass. + +The selection of appropriate context examples impacts the performance of in-context learning. Current methods [57] employ image-level retrieval strategies using global image embeddings to find better references. However, this approach faces significant challenges in medical image analysis where each image contains multiple classes including structures like organs, tissues, and lesions. Image-level retrieval inevitably averages features across all structures, leading to a suboptimal reference selection. To address this limitation, Iris introduces an object-level context selection mechanism that enables fine-grained matching of individual classes, focusing on more precise and class-specific reference selection compared to image-level approaches [57]. + +# 3. Method + +# 3.1. Problem Definition + +Traditional segmentation approaches follow a task-specific paradigm, where each model $f_{\theta_t}$ is trained for a specific segmentation task $t$ . Given a dataset $\mathcal{D}_t = \{(x_t^i, y_t^i)\}_{i=1}^{N_t}$ containing $N_t$ image-label pairs, the model learns a direct mapping $f_{\theta_t}: \mathcal{X} \to \mathcal{Y}$ from the image space $\mathcal{X}$ to the segmentation mask space $\mathcal{Y}$ , such that for an image $x_t$ , the predicted segmentation mask is given by $y_t = f_{\theta_t}(x_t)$ . + +In contrast, we formulate a in-context medical im + +age segmentation framework. Given a support set $\mathcal{S} = \{(x_s^i, y_s^i)\}_{i=1}^n \in (\mathcal{X} \times \mathcal{Y})^n$ containing $n$ reference image-label pairs and a query image $x_q \in \mathcal{X}$ , a single model $f_\theta$ predicts the segmentation mask $\mathbf{y}_q$ for the query image conditioned on $\mathcal{S}$ : + +$$ +\hat {\boldsymbol {y}} _ {q} = f _ {\theta} \left(\boldsymbol {x} _ {q}; S\right) = f _ {\theta} \left(\boldsymbol {x} _ {q}; \left\{\left(\boldsymbol {x} _ {s} ^ {i}, \boldsymbol {y} _ {s} ^ {i}\right) \right\} _ {i = 1} ^ {n}\right) \tag {1} +$$ + +For multi-class segmentation tasks, we decompose the problem into multiple binary segmentation tasks. + +# 3.2. Iris Architecture + +In Figure 2, Iris introduces a novel in-context learning architecture that decouples task encoding from segmentation inference. This design comprises two key components: (1) a task encoding module that distills task-specific information from reference examples into compact task embeddings, and (2) a mask decoding module that leverages these task embeddings to guide query image segmentation. + +# 3.2.1 Task Encoding Module. + +Given a reference 3D image-label pair $(\pmb{x}_s, \pmb{y}_s) \in \mathbb{R}^{D \times H \times W} \times \{0,1\}^{D \times H \times W}$ , our task encoding module extracts task representations through two parallel streams to extract comprehensive task representations. + +Foreground feature encoding. Medical data volumes present unique challenges in feature extraction due to the presence of fine boundary details and anatomical structures spanning only a tiny portion of voxels. Direct feature pooling at downsampled resolution can lead to information loss or complete disappearance of these critical regions of interest (ROIs). To address this hurdle, we opt in a high-resolution foreground feature encoding process. Given features $\pmb{F}_{s} \in \mathbb{R}^{C \times d \times h \times w}$ extracted by the encoder $E$ , where + +$d = D / r, h = H / r, w = W / r$ are downsampled dimensions with ratio $r$ , we compute the foreground embedding by: + +$$ +\boldsymbol {T} _ {f} = \operatorname {P o o l} \left(\operatorname {U p s a m p l e} \left(\boldsymbol {F} _ {s}\right) \odot \boldsymbol {y} _ {s}\right) \in \mathbb {R} ^ {1 \times C} \tag {2} +$$ + +where $\operatorname{Upsample}(\boldsymbol{F}_s) \in \mathbb{R}^{C \times D \times H \times W}$ restores features to the original resolution. By applying the original high-resolution mask $\boldsymbol{y}_s$ directly to the upsampled features, we ensure a precise capture of fine anatomical details and small structures that are vital for medical object segmentation. + +Contextual feature encoding. The above encoding process extracted foreground features, but lacks important global context information. We encode these contextual information using learnable query tokens. To efficiently process high-resolution features while managing memory constraints, we employ strategy similar to sub-pixel convolution [42]. For feature map $F_{s}$ , we first expand spatial dimensions while reducing channels: + +$$ +\boldsymbol {F} _ {s} ^ {\prime} = \operatorname {P i x e l S h u f f l e} \left(\boldsymbol {F} _ {s}\right) \in \mathbb {R} ^ {C / r ^ {3} \times D \times H \times W} \tag {3} +$$ + +After concatenating with the binary mask $\mathbf{y}_s$ , we apply a $1 \times 1 \times 1$ convolution and PixelUnshuffle to return to the original feature resolution: + +$$ +\hat {\boldsymbol {F}} _ {s} = \text {P i x e l U n s h u r f l e} (\operatorname {C o n v} (\operatorname {C o n c a t} [ \boldsymbol {F} _ {s} ^ {\prime}, \boldsymbol {y} _ {s} ])) \in \mathbb {R} ^ {C \times d \times h \times w} \tag {4} +$$ + +This approach permits a memory-efficient, high-resolution, feature-mask fusion. The merged features $\hat{\pmb{F}}_s$ then interact with $m$ learnable query tokens through cross-attention and self-attention layers to produce contextual embedding $\pmb{T}_c\in \mathbb{R}^{m\times C}$ . The final task embedding combines both aspects: $\pmb{T} = [T_f;T_c]\in \mathbb{R}^{(m + 1)\times C}$ + +For multi-class segmentation, we generate separate task embeddings for each category in $\mathbf{y}_s$ . This setting maintains a strong efficiency as the computationally intensive feature extraction is shared across classes while the task encoding module remains lightweight. + +# 3.2.2 Mask Decoding Module + +The decoder $D$ employs a query-based architecture [8] that efficiently handles both single and multi-class segmentation tasks. For a query image with features $\pmb{F}_q\in \mathbb{R}^{C\times d\times h\times w}$ , the task encoding module generates class-specific embeddings $\pmb{T}^k\in \mathbb{R}^{(m + 1)\times C}$ for each class $k$ defined in reference image-label pairs. These embeddings are concatenated into a combined task representation $\pmb{T} = [T^{1};T^{2};\dots;T^{K}]\in \mathbb{R}^{K(m + 1)\times C}$ , where $K$ is the number of target classes and $K = 1$ for single-class segmentation. The bidirectional cross-attention mechanism processes this representation: + +$$ +\boldsymbol {F} _ {q} ^ {\prime}, \boldsymbol {T} ^ {\prime} = \operatorname {C r o s s A t t n} (\boldsymbol {F} _ {q}, \boldsymbol {T}) \tag {5} +$$ + +![](images/6e4216bd3f434a1935a967a6fd86a7ee1b7e02e79d570ddc275b3583c0120720.jpg) + +![](images/275c885abbdceac5ecb97b5df07b3faebdc82ac4a3fa4c7987bbe1d4011e9aa3.jpg) + +![](images/adc22fd58f6d35bce63f4305761cdacd26ffcfa009df6fd732bdbc044b1191e2.jpg) +Figure 3. Iris's flexible inference strategies. The red arrows indicate gradient backpropagation. + +![](images/21e8a707451b24b5416931460db5aa87d0e45384996e0c1beb9fe4f28bc95929.jpg) + +where $\pmb{F}_q'$ and $\pmb{T}'$ are the updated features. This mechanism enables effective information exchange between class-specific task guidance and query image features. The final segmentation mask is predicted in a single forward pass: + +$$ +\hat {\boldsymbol {y}} _ {q} = D \left(\boldsymbol {F} _ {q} ^ {\prime}, \boldsymbol {T} ^ {\prime}\right) \in \{0, 1 \} ^ {K \times D \times H \times W} \tag {6} +$$ + +# 3.2.3 Training. + +We train Iris in an end-to-end manner using episodic training to simulate in-context learning scenarios (details in supplementary). Each training episode consists of sampling reference-query pairs from the same dataset, computing task embeddings from the reference pair, and final predicting segmentation for the query image. The model is optimized using a combination of Dice and cross-entropy losses: $\mathcal{L}_{seg} = \mathcal{L}_{dice} + \mathcal{L}_{ce}$ . To enhance generalization, we employ data augmentation on both query and reference images, add random perturbation to query images to simulate imperfect references, and randomly drop classes in multi-class datasets to encourage independent class-wise task encoding. + +# 3.3. Flexible Inference Strategies + +After training, Iris supports multiple inference strategies suitable for different practical scenarios (see in Figure 3). Efficient one-shot inference. With just one reference example, Iris first encodes the task into compact embeddings $T$ that can be stored and reused across multiple query images. Unlike major in-context learning methods to recompute contextual information for each query image, our design greatly eliminates redundant computation. Moreover, Iris can segment multiple classes in a single forward pass, contrasting with methods (e.g., UniverSeg [5]) that require separate passes per class. The minimal storage requirement of these embeddings makes Iris particularly desirable for large-scale data processing pipelines. + +Table 1. Comparison of segmentation performance across different in-distribution datasets. Values represent mean Dice scores $(\%)$ + +
MethodDatasetAVG
AMOS CTAMOS MRAuto PETBCVBrainCHAOSKiTS TumorLiTS TumorMnMStructSeg H&NStructSeg ThoCSI-Wat
Task-specific Model (Upper Bound)
nnUNet88.6785.4267.2183.3894.1291.1381.7263.1185.5978.1788.5391.1183.18
Multi-task Universal Model (Upper Bound)
Clip-driven88.9586.4170.0185.0395.0691.7182.7365.4386.1278.4489.2790.9884.18
UniSeg89.1186.5870.0985.4295.2991.8382.9965.8786.2978.7289.4291.2384.40
Multi-Talent89.1586.5870.8985.2095.7791.3882.3265.5386.3080.0989.0991.3284.47
Positional Prompt
SAM22.2317.8220.1023.3420.5120.0118.2112.0810.2317.2324.8113.2017.97
SAM-Med 2D50.1248.6638.0350.3235.2850.3230.2323.2740.3339.3263.8734.8740.58
SAM-Med 3D79.1976.1867.1479.8942.2984.7979.3232.9352.6768.8383.5674.2368.42
In-Context
SegGPT45.3751.7848.2949.7885.2763.7240.7835.9874.1240.2867.2885.5957.35
UniverSeg57.2452.4347.2345.2687.7660.4645.7236.2175.2442.9866.9586.6858.68
Tyche-IS59.5754.7850.9847.6789.2862.7349.2737.0278.9245.3369.8988.9961.20
Iris (ours)89.5686.7070.0285.7396.0491.8581.5465.0286.0880.3689.4291.9784.52
+ +Context ensemble. For tasks with multiple reference examples, Iris supports context ensemble for improving performance. We compute task embeddings for each example and average them to create a more robust task representation. This simple averaging strategy combines information from multiple references while maintaining computational efficiency. We extent context ensemble for classes seen during training. Specifically, we maintain a class-specific memory bank that continuously updates task embeddings through exponential moving average (EMA) during the training process. This memory bank stores representative task embeddings for each seen class, enabling direct segmentation for seen classes during inference without requiring context encoding. + +Object-level context retrieval. For multi-class segmentation with a pool of reference examples, conventional approaches typically employ image-level retrieval using global embeddings to select semantically similar references [57]. However, this strategy is suboptimal for medical images where multiple anatomical structures coexist, as global embeddings average features across all structures. To enable more precise reference selection, we propose an object-level (class-level) context retrieval strategy. Our approach first encodes class-specific task embeddings for each reference example through our task encoding module - for a reference image with $n$ anatomical classes, we encode $n$ separate task embeddings. For a query image, we obtain initial object segmentation masks using task embeddings from a randomly selected reference. These initial masks are then used to encode $n$ class-specific query task embeddings, which are compared with corresponding reference embeddings in the pool using cosine similarity to select the most similar reference for each class independently. This fine-grained matching allows different structures within the same query image to find their most appropriate references, leading to more accurate + +segmentation compared to image-level approaches. + +In-context tuning. For scenarios requiring adaptation without a full model fine-tuning, Iris offers a lightweight tuning strategy by optimizing only the task embeddings while keeping the model parameters fixed. This tuning process minimizes the segmentation loss between model predictions and ground truth by updating the task embeddings through the gradient descent. In particular, the optimized embeddings can then be stored and reused for similar cases, offering a practical balance between adaptation capability and computational efficiency. + +# 4. Experiment + +# 4.1. Experimental Setup + +We evaluate Iris across three key dimensions: in-distribution performance on trained tasks, out-of-distribution generalization to different domains, and adaptability to novel anatomical classes. Additional experiments analyze Iris's computational efficiency, inference strategies, and architectural design choices. + +Dataset. Our training data comprises 12 public datasets [3, 6, 15, 18, 21, 22, 25, 26, 35, 40] spanning diverse body regions (head, chest, abdomen), modalities (CT, MRI, PET), and clinical targets (organs, tissues, lesions), split into $75\% / 5\% / 20\%$ for train/validation/test. For out-of-distribution evaluation, we use 5 held-out datasets: ACDC [2], SegTHOR [24], and three MRI modalities from IVDM3Seg [16] to evaluate robustness against domain shift; MSD Pancreas (Tumor) [1] and Pelvic1K (Bone) [31] datasets are used for novel class adaptation. We randomly select $20\%$ samples from held-out sets for testing. Detailed dataset information is provided in supplementary materials. + +Baseline Models. We compare against four categories of + +Table 2. Out-of-distribution comparison on held-out datasets, including generalization capability and performance on unseen classes. Values represent mean Dice scores (\%). All in-context models use one-shot inference. + +
MethodGeneralizationUnseen Classes
ACDCSegTHORCSI-innCSI-oppCSI-fatMSD PancreasPelvic
Supervised Upper Bound
nnUNet90.9789.7891.2391.0490.1354.5694.73
Task-specific Model
nnUNet-generalize82.0676.9255.2485.190.23--
Multi-task Universal Model
CLIP-driven84.7278.2359.7386.731.47--
UniSeg84.9878.5660.0286.131.52--
Multi-Talent83.7978.4558.2987.011.95--
Positional Prompt
SAM-Med2D42.2352.3729.2332.7110.9110.3735.71
SAM-Med3D51.4968.9745.3268.7223.9315.8353.61
In-context
SegGPT73.8260.9859.8777.6235.2710.6755.92
UniverSeg72.4354.7563.4885.3252.4810.2857.81
Tyche-IS74.9156.7564.2387.1355.7511.9761.92
Iris (ours)86.4582.7764.4489.1347.7828.2869.03
+ +methods: (1) Task-specific models: nnUNet [20]; (2) Universal models: CLIP-driven model [30], UniSeg [50] and Multi-Talent [44]; (3) Foundation models: SAM [23] and its medical variants, SAM-Med2D [9], SAM-Med3D [45]; (4) In-context learning methods: SegGPT [48], UniverSeg [5], and Tyche-IS [39]. All models are trained on our curated dataset, except SAM, with 2D models trained on extracted slices and 3D models with 3D volumes. For SAM-based methods, we simulate user interactions using ground-truth labels during training and evaluation. + +Implementation Details. Iris uses a 3D UNet encoder trained from scratch with one-shot learning strategy. We employ the Lamb optimizer [53] with exponential learning rate decay (base $\mathrm{lr} = 2 \times 10^{-3}$ , weight decay $= 1 \times 10^{-5}$ ), training for 80K iterations with batch size 32 and 2K warm-up iterations. Data augmentation includes random cropping, affine transformations, and intensity adjustments. Training and inference use $128 \times 128 \times 128$ volume size. + +# 4.2. Comparison with the state-of-the-art + +Results on in-distribution classes. We evaluate Iris's performance on twelve diverse medical datasets used during training. As shown in Table 1, Iris achieves state-of-the-art performance with an average Dice score of $84.52\%$ , matching or exceeding task-specific and multi-task models that are optimized for fixed tasks. Existing adaptive approaches show significant limitations. SAM-based methods perform poorly due to their strong reliance on simple positional prompts, with the large performance gap between their 2D and 3D variants (40.58% vs. 68.42%) highlighting the importance of 3D context. Existing in-context learning methods, like SegGPT, UniverSeg, (best: 61.20%) struggle particularly with 3D tasks like AMOS and LiTS due to their 2D architecture, + +though performing better on 2D-friendly tasks like MnM and CSI-Wat. In contrast, Iris's 3D architecture and efficient task encoding enables consistent high-level performance across all tasks while maintaining its adaptability to unseen novel anatomical classes. + +Results on OOD generalization. We evaluate out-of-distribution (OOD) performance on five held-out datasets spanning two types of distribution shifts: cross-center variation (ACDC, SegTHOR) and cross-modality adaptation (CSI variants). In Table 2, Iris demonstrates superior performance across all scenarios, particularly excelling in challenging 3D tasks and large domain shifts. + +Both task-specific and multi-task universal models show performance degradation, especially failing catastrophically on CSI-fat with a significant domain gap. While SAM-based methods demonstrate their resilience to domain shifts through strong prior knowledge injected from user interactions, their performance remains limited on the volumetric data. In-context learning methods retain a good performance with cross-modality adaptation (e.g. on CSI-fat), benefiting from the domain-specific knowledge provided by reference examples. However, a 2D-slice-based architecture (e.g., UniverSeg and Tyche) limits its capability on 3D tasks like SegTHOR. In contrast, Iris's task encoding module efficiently extracts and utilizes 3D domain-specific information from the reference examples. + +Results on novel classes. To measure the adaptation performance to completely unseen anatomical structures, we evaluate on MSD Pancreas Tumor and Pelvic datasets. Using only one reference example, Iris achieves $28.28\%$ on MSD Pancreas Tumor and $69.03\%$ on Pelvic segmentation, substantially outperforming other adaptive methods (the best competitor: $11.97\%$ and $61.92\%$ respectively). This perfor + +Table 3. Comparison of computational complexity. Empirical measurements of computation on one NVIDIA A100 GPU. We inference with 10 query images and one reference image from AMOS CT dataset with 15 classes. The image size is processed to $128 \times 128 \times 128$ for inference. + +
MethodInference Time (s)Memory (GB)Parameters (M)
UniverSeg-1659.42.11.2
UniverSeg-1281030.212.11.2
SAM-Med2D648.41.891.1
SAM-Med3D15.22.9100.5
Iris (Ours)2.07.469.4
+ +![](images/c5c638c3193536af51e2e3123f6be1d3b9a27684aa2fee57c0759fe0c8e8a22b.jpg) +Figure 4. Analysis of different inference strategies. + +mance gain is particularly notable given that traditional task-specific models and multi-task models can not handle these novel classes without retraining. These findings demonstrate Iris's strong capability in learning from very limited examples while maintaining meaningful segmentation quality on previously unseen anatomical structures. + +Efficiency comparison We analyze computational efficiency for segmenting $m$ query images with $n$ classes using $k$ reference pairs. Iris achieves superior efficiency through two key designs: decoupling task extraction from inference and handling multiple classes in a single forward pass. This results in complexity of $O(k + m)$ compared to $O(kmn)$ in methods like UniverSeg that process each class separately and recompute reference features for every query. + +Table 3 compares real-world inference time and memory usage across methods. While UniverSeg's slice-by-slice processing leads to significant overhead with multiple reference slices, and SAM-Med3D requires iterative user interactions (interaction-time not included), Iris efficiently processes entire 3D volumes all at once. For a scenario of segmenting 10 query volumes with 15 classes using one reference volume, Iris completes in 2 seconds. This efficiency advantage grows with more context examples due to Iris's decoupled architecture eliminating redundant reference processing. + +# 4.3. Analysis + +Inference strategy. Figure 4 compares our four inference strategies. In this experiment, we maintain a pool of all available context examples and evaluate each strategy's per + +formance as follows. + +Context ensemble randomly selects and averages task embeddings from a percentage of the context pool. When using only one context example (1%), it operates as one-shot inference. Performance of context ensemble keeps improving with more context examples and eventually saturates. This strategy is appealing as task embeddings can be precomputed and ensembled into a single robust embedding, enabling inference speed comparable to regular segmentation models. + +Both image and object-level retrieval strategies access the entire context pool but utilize only the top- $k$ percent most similar examples as references. While image-level retrieval [57] compares whole-image features and uses all task embeddings from the same retrieved images, object-level retrieval enables more precise reference selection by matching individual classes. Notably, object-level retrieval surpasses full context ensemble performance when using fewer references (e.g., top $10 - 20\%$ ), as it selectively chooses the most relevant examples for each class rather than averaging all available contexts. To validate robustness to initial context selection, we conducted experiments for 10 times with random selection using different percentages of context samples ( $1\%$ , $5\%$ , $10\%$ ), achieving consistent performance (mean and standard deviation: $68.94 \pm 0.83$ , $71.07 \pm 0.27$ , $72.87 \pm 0.10$ respectively). This strategy is particularly valuable in clinical settings with large patient databases, where retrieving similar cases as references can enhance segmentation accuracy. + +In-context tuning optimizes task embeddings initialized from a random reference. While showing a lower performance with limited samples due to overfitting, it achieves positive results with sufficient tuning data. This approach is well suited for scenarios with both a large context pool and available computational resources for fine-tuning. + +Overall, Iris offers usable strategies pertinent to different real-world scenarios. Object-level retrieval is designed for high accuracy while requiring access to a large context pool, e.g. a database of previous patient records. Context ensemble offers a strong efficiency of response time. Finally, in-context tuning is applicable when computational resources and sufficient data support are available. + +Task embedding analysis. Iris's task encoding module discovers meaningful anatomical relationships without explicit anatomical supervision, learning solely from binary segmentation masks. From Figure 5, the t-SNE visualization of task embeddings reveals natural clustering of anatomical structures that transcends dataset boundaries and imaging modalities. For example, abdominal organs cluster together despite originating from different datasets and modalities (e.g., AMOS-CT, BCV in CT; AMOS-MR, CHAOS in MRI). + +We find that feature embeddings capture clinically meaningful anatomical similarities that were never explicitly taught (Figure 5, bottom). Blood vessels like the Inferior + +![](images/f0d8052d65db2fae57327606593016f56b53aaef0b500825df2aa52447bed2b9.jpg) + +![](images/e5be458063244bb9079a081328f03818d8e418ff450f93d45a37f119e905e566.jpg) +Figure 5. Top: Visualizing the task embedding with t-SNE. The color represents dataset, the circle and marks are the classes of the embeddings. Bottom: Examples of the similar tasks revealed by the t-SNE result. + +Vena Cava (IVC) and Portal/Splenic veins cluster nearby, reflecting their shared tubular structure and similar contrast enhancement patterns in CT. Similarly, the bladder and prostate embeddings show proximity due to their shared soft-tissue characteristics and adjacent anatomical locations. This emergent organization of anatomical concepts demonstrates Iris's ability to automatically distill fundamental anatomical relationships across different segmentation tasks, making it particularly robust for adapting to new anatomical structures. Generalization performance vs task quantity. We investigate how training data diversity affects Iris's generalization by varying the number of training tasks. From Figure 6 (left), the performance on held-out datasets consistently improves with more training tasks, particularly when the training subset encompasses diverse anatomical structures and imaging modalities. We recognize that models trained on datasets spanning body regions (e.g., brain, chest, and abdomen) show a stronger generalization compared to those trained on narrower anatomical ranges. This finding suggests that an exposure to diverse anatomical patterns is necessary towards more robust and transferable feature learning. + +Ablation study. Table 4 analyzes three key components of Iris. High-resolution processing proves crucial for small structures, dramatically improving their segmentation performance from $62.13\%$ to $78.92\%$ . While each component contributes independently, their combination achieves the best results across all metrics, demonstrating substantial improvements over partial implementations. In Figure 6 (right), performance improves with more query tokens but saturates at 10 tokens, which we adopt in our final model to balance + +![](images/b5f5eaa8796361ffe492fbe0ff8f268f90122e0e9e5b0b6ee2ab74e0c87fc2a7.jpg) +Figure 6. Left: Number of tasks used for training v.s. Performance on unseen classes. Right: Ablation on the number of queries. + +![](images/f82258a6024dcbcec33fd92f32927620aec5614ba372028579ce7f83215b6301.jpg) +Table 4. Ablation study of different components in Iris. High-Res: high-resolution feature processing; Foreground: foreground feature pooling; Query: query-based contextual encoding. + +
High-ResForegroundQueryIn-dist (Avg)In-dist (Small)Out-of-dist
✓✓82.1062.1362.00
✓✓82.4778.9265.93
✓✓82.0677.5364.13
✓✓✓84.5280.3666.95
+ +performance and efficiency. + +# 5. Discussion and Conclusion + +We introduce Iris as a novel in-context learning framework that enables versatile 3D medical image segmentation through only reference examples. Given just one image-label pair as a reference, Iris can segment arbitrary target classes in test images without any model modification or retraining. Iris reveals strong performance on in-distribution tasks across 12 diverse datasets. Iris's performance is particularly evident to distribution shifts and novel unseen classes on 7 held-out test datasets. The key design of Iris is a decoupled architecture that enables efficient 3D medical image processing and single-pass multi-class segmentation. Iris's inference strategies are suitable for different practical scenarios, from efficient context ensemble-based data processing, high-accuracy object-level context retrieval, to in-context finetuning. Further, Iris's task encoding module offers an appealing means to automatically discover meaningful anatomical relationships purely from segmentation masks, allowing knowledge transfer across different tasks and imaging modalities without explicit anatomical supervision. + +Limitations and future work. While Iris demonstrates promising capabilities, several challenges remain to explore. The diversity of training tasks could impact the out-of-distribution generalization, suggesting a critical need for automated methods to create diverse tasks without manual annotation. Although Iris shows strong adaptability to novel tasks, there remains a performance gap with supervised upper bounds in certain scenarios. Future investigation will focus on narrowing this gap and expanding both training and evaluation schemes to cover a broader spectrum of medical imaging applications. + +# References + +[1] Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M Summers, et al. The medical segmentation decathlon. Nature communications, 13(1):4128, 2022. 5 +[2] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37(11): 2514-2525, 2018. 5 +[3] Patrick Bilic, Patrick Ferdinand Christ, Eugene Vorontsov, Grzegorz Chlebus, Hao Chen, Qi Dou, Chi-Wing Fu, Xiao Han, Pheng-Ann Heng, Jürgen Hesser, et al. The liver tumor segmentation benchmark (lits). arXiv preprint arXiv:1901.04056, 2019. 5 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2 +[5] Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Universeg: Universal medical image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21438-21451, 2023. 2, 4, 6 +[6] Victor M Campello, Polyxeni Gkontra, Cristian Izquierdo, Carlos Martin-Isla, Alireza Sojoudi, Peter M Full, Klaus Maier-Hein, Yao Zhang, Zhiqiang He, Jun Ma, et al. Multicentre, multi-vendor and multi-disease cardiac segmentation: the m&ms challenge. IEEE Transactions on Medical Imaging, 40(12):3543-3554, 2021. 5 +[7] Qi Chang, Zhennan Yan, Mu Zhou, Di Liu, Khalid Sawalha, Meng Ye, Qilong Zhangli, Mikael Kanski, Subhi Al'Aref, Leon Axel, et al. Deeprecon: Joint 2d cardiac segmentation and 3d volume reconstruction via a structure-specific generative method. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 567-577. Springer, 2022. 1 +[8] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 4 +[9] Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Ji-long Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184, 2023. 2, 6 +[10] Steffen Czolbe and Adrian V Dalca. Neuralizer: General neuroimage analysis without re-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6217-6230, 2023. 2 +[11] Jeffrey De Fauw, Joseph R Ledsam, Bernardino Romero-Paredes, Stanislav Nikolov, Nenad Tomasev, Sam Blackwell, Harry Askham, Xavier Glorot, Brendan O'Donoghue, Daniel + +Visentin, et al. Clinically applicable deep learning for diagnosis and referral in retinal disease. Nature medicine, 24(9): 1342-1350, 2018. 1 +[12] Konstantin Dmitriev and Arie E Kaufman. Learning multiclass segmentations from single-class datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9501-9511, 2019. 2 +[13] Yunhe Gao. Training like a medical resident: Context-prior learning toward universal medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11194-11204, 2024. 1, 2 +[14] Yunhe Gao, Mu Zhou, Di Liu, Zhennan Yan, Shaoting Zhang, and Dimitris N Metaxas. A data-scalable transformer for medical image segmentation: architecture, model efficiency, and benchmark. arXiv preprint arXiv:2203.00131, 2022. 1 +[15] Sergios Gatidis, Tobias Hepp, Marcel Früh, Christian La Fougère, Konstantin Nikolaou, Christina Pfaffenberg, Bernhard Schölkopf, Thomas Küstner, Clemens Cyran, and Daniel Rubin. A whole-body fdg-pet/ct dataset with manually annotated tumor lesions. Scientific Data, 9(1):601, 2022. 5 +[16] Zheng Guoyan, Li Shuo, and Belavy Daniel. Automatic intervertebral disc localization and segmentation from 3d multi-modality mr (m3) images, 2018. 5 +[17] Xiaoxiao He, Chaowei Tan, Bo Liu, Liping Si, Weiwu Yao, Liang Zhao, Di Liu, Qilong Zhangli, Qi Chang, Kang Li, et al. Dealing with heterogeneous 3d mr knee images: A federated few-shot learning method with dual knowledge distillation. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 1 +[18] Nicholas Heller, Niranjan Sathianathen, Arveen Kalapara, Edward Walczak, Keenan Moore, Heather Kaluzniak, Joel Rosenberg, Paul Blake, Zachary Rengel, Makinna Oestreich, et al. The kits19 challenge data: 300 kidney tumor cases with clinical context, ct semantic segmentations, and surgical outcomes. arXiv preprint arXiv:1904.00445, 2019. 5 +[19] Rui Huang, Yuanjie Zheng, Zhiqiang Hu, Shaoting Zhang, and Hongsheng Li. Multi-organ segmentation via co-training weight-averaged models from few-organ datasets. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part IV 23, pages 146–155. Springer, 2020. 2 +[20] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 1, 6 +[21] Yuanfeng Ji, Haotian Bai, Jie Yang, Chongjian Ge, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhang, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. arXiv preprint arXiv:2206.08023, 2022. 5 +[22] A. Emre Kavur, N. Sinem Gezer, Mustafa Baris, Sinem Aslan, Pierre-Henri Conze, Vladimir Groza, Duc Duy Pham, Soumick Chatterjee, Philipp Ernst, Savaş Özkan, Bora Baydar, Dmitry Lachinov, Shuo Han, Josef Pauli, Fabian Isensee, Matthias Perkonigg, Rachana Sathish, Ronnie Rajan, Debboot + +Sheet, Gurbandurdy Dovletov, Oliver Speck, Andreas Nurnberger, Klaus H. Maier-Hein, Gözde Bozdağ Akar, Gözde Ünal, Oğuz Dicle, and M. Alper Selver. CHAOS Challenge - combined (CT-MR) healthy abdominal organ segmentation. Medical Image Analysis, 69:101950, 2021. 5 +[23] 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. arXiv preprint arXiv:2304.02643, 2023. 2, 6 +[24] Zoé Lambert, Caroline Petitjean, Bernard Dubray, and Su Kuan. Segthor: Segmentation of thoracic organs at risk in ct images. In 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1-6. IEEE, 2020. 5 +[25] Bennett Landman, Zhoubing Xu, Juan Lgelsias, Martin Styner, Thomas Langerak, and Klein Arno. Multi-atlas labeling beyond the cranial vault - workshop and challenge, 2020. 5 +[26] Hongsheng Li, Jinghao Zhou, Jincheng Deng, and Ming Chen. Automatic structure segmentation for radiotherapy planning challenge 2019, 2019. 5 +[27] Di Liu, Jiang Liu, Yihao Liu, Ran Tao, Jerry L Prince, and Aaron Carass. Label super resolution for 3d magnetic resonance images using deformable u-net. In Medical Imaging 2021: Image Processing, pages 606-611. SPIE, 2021. 1 +[28] Di Liu, Zhennan Yan, Qi Chang, Leon Axel, and Dimitris N Metaxas. Refined deep layer aggregation for multi-disease, multi-view & multi-center cardiac mr segmentation. In International Workshop on Statistical Atlases and Computational Models of the Heart, pages 315-322. Springer, 2021. +[29] Di Liu, Yunhe Gao, Qilong Zhangli, Ligong Han, Xiaoxiao He, Zhaoyang Xia, Song Wen, Qi Chang, Zhennan Yan, Mu Zhou, et al. Transfusion: multi-view divergent fusion for medical image segmentation with transformers. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 485-495. Springer, 2022. 1 +[30] Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. arXiv preprint arXiv:2301.00785, 2023. 2, 6 +[31] Pengbo Liu, Hu Han, Yuanqi Du, Heqin Zhu, Yinhao Li, Feng Gu, Honghu Xiao, Jun Li, Chunpeng Zhao, Li Xiao, et al. Deep learning to segment pelvic bones: large-scale ct datasets and baseline models. International Journal of Computer Assisted Radiology and Surgery, 16:749-756, 2021. 5 +[32] Xiangbin Liu, Liping Song, Shuai Liu, and Yudong Zhang. A review of deep-learning-based medical image segmentation methods. Sustainability, 13(3):1224, 2021. 1 +[33] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 2 +[34] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 2 + +[35] Carlos Martin-Isla, Víctor M Campello, Cristian Izquierdo, Kaisar Kushibar, Carla Sendra-Balcells, Polyxeni Gkontra, Alireza Sojoudi, Mitchell J Fulton, Tewodros Weldebirhan Arega, Kumaradevan Punithakumar, et al. Deep learning segmentation of the right ventricle in cardiac mri: the m&ms challenge. IEEE Journal of Biomedical and Health Informatics, 27(7):3302-3313, 2023. 5 +[36] Lingchen Meng, Shiyi Lan, Hengduo Li, Jose M Alvarez, Zuxuan Wu, and Yu-Gang Jiang. Segic: Unleashing the emergent correspondence for in-context segmentation. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 2 +[37] Ursula Nestle, Stephanie Kremp, Andrea Schaefer-Schuler, Christiane Sebastian-Welsch, Dirk Hellwig, Christian Rübe, and Carl-Martin Kirsch. Comparison of different methods for delineation of 18f-fdg pet-positive tissue for target volume definition in radiotherapy of patients with non-small cell lung cancer. Journal of nuclear medicine, 46(8):1342-1348, 2005. 1 +[38] Ziwei Niu, Shuyi Ouyang, Shiao Xie, Yen-wei Chen, and Lanfen Lin. A survey on domain generalization for medical image analysis. arXiv preprint arXiv:2402.05035, 2024. 1 +[39] Marianne Rakic, Hallee E Wong, Jose Javier Gonzalez Ortiz, Beth A Cimini, John V Guttag, and Adrian V Dalca. Tyche: Stochastic in-context learning for medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11159-11173, 2024. 2, 6 +[40] KM Rodrigue, KM Kennedy, MD Devous, JR Rieck, AC Hebrank, R Diaz-Arrastia, D Mathews, and DC Park. $\beta$ -amyloid burden in healthy aging: regional distribution and cognitive consequences. *Neurology*, 78(6):387-395, 2012. 5 +[41] Wei Shen, Mu Zhou, Feng Yang, Caiyun Yang, and Jie Tian. Multi-scale convolutional neural networks for lung nodule classification. In Information Processing in Medical Imaging: 24th International Conference, IPMI 2015, Sabhal Mor Ostaig, Isle of Skye, UK, June 28-July 3, 2015, Proceedings 24, pages 588-599. Springer, 2015. 1 +[42] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 4 +[43] Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. Vrp-sam: Sam with visual reference prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23565-23574, 2024. 2 +[44] Constantin Ulrich, Fabian Isensee, Tassilo Wald, Maximilian Zenk, Michael Baumgartner, and Klaus H Maier-Hein. Multivalent: A multi-dataset approach to medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 648–658. Springer, 2023. 2, 6 +[45] H Wang et al. Sam-med3d: towards general-purpose segmentation models for volumetric medical images. Preprint at https://arxiv.org/abs/2310.15161, 2024. 2, 6 + +[46] Risheng Wang, Tao Lei, Ruixia Cui, Bingtao Zhang, Hongying Meng, and Asoke K Nandi. Medical image segmentation using deep learning: A survey. IET image processing, 16(5): 1243-1267, 2022. 1 +[47] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 2 +[48] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 2, 6 +[49] Yutong Xie, Jianpeng Zhang, Yong Xia, and Chunhua Shen. Learning from partially labeled data for multi-organ and tumor segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[50] Yiwen Ye, Yutong Xie, Jianpeng Zhang, Ziyang Chen, and Yong Xia. Uniseg: A prompt-driven universal segmentation model as well as a strong representation learner. arXiv preprint arXiv:2304.03493, 2023. 2, 6 +[51] Yiwen Ye, Ziyang Chen, Jianpeng Zhang, Yutong Xie, and Yong Xia. Meduniseg: 2d and 3d medical image segmentation via a prompt-driven universal model. arXiv preprint arXiv:2410.05905, 2024. 2 +[52] Jee Seok Yoon, Kwanseok Oh, Yooseung Shin, Maciej A Mazurowski, and Heung-II Suk. Domain generalization for medical image analysis: A survey. arXiv preprint arXiv:2310.08598, 2023. 1 +[53] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. arXiv preprint arXiv:1904.00962, 2019. 6 +[54] Jianpeng Zhang, Yutong Xie, Yong Xia, and Chunhua Shen. Dodnet: Learning to segment multi-organ and tumors from multiple partially labeled datasets. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1195–1204, 2021. 2 +[55] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Xianzheng Ma, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 2 +[56] Yixiao Zhang, Xinyi Li, Huimiao Chen, Alan L Yuille, Yaoyao Liu, and Zongwei Zhou. Continual learning for abdominal multi-organ and tumor segmentation. In International conference on medical image computing and computer-assisted intervention, pages 35–45. Springer, 2023. 1 +[57] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36:17773-17794, 2023. 3, 5, 7 +[58] Yunkun Zhang, Jin Gao, Zheling Tan, Lingfeng Zhou, Kexin Ding, Mu Zhou, Shaoting Zhang, and Dequan Wang. Data-centric foundation models in computational healthcare: A survey. arXiv preprint arXiv:2401.02458, 2024. 2 +[59] Qilong Zhangli, Jingru Yi, Di Liu, Xiaoxiao He, Zhaoyang Xia, Qi Chang, Ligong Han, Yunhe Gao, Song Wen, Haiming + +Tang, et al. Region proposal rectification towards robust instance segmentation of biological images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2022. 1 +[60] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 1 +[61] Yuyin Zhou, Zhe Li, Song Bai, Chong Wang, Xinlei Chen, Mei Han, Elliot Fishman, and Alan L Yuille. Prior-aware neural network for partially-supervised multi-organ segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10672-10681, 2019. 2 \ No newline at end of file diff --git a/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/images.zip b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8fe4865b7b93a3ea52754019ccce436218410ef2 --- /dev/null +++ b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d17031afe1f4efe1725ee19866143b18416845965891f2871e7f037baf609f87 +size 540048 diff --git a/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/layout.json b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6fe79b9b05fd47ca9380afef682f07c60338be8e --- /dev/null +++ b/CVPR/2025/Show and Segment_ Universal Medical Image Segmentation via In-Context Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5510d8fa174f68f809b2dceb7ae18eacfc4350d53a6281d02201565d78c7b83e +size 385988 diff --git a/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/8412533f-67db-465a-8494-1516f366251f_content_list.json b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/8412533f-67db-465a-8494-1516f366251f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..56b210dc737acbab7d529af09879788c5e6abc36 --- /dev/null +++ b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/8412533f-67db-465a-8494-1516f366251f_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37ea847ea8b0bbd543fce751c330eb3e9c3829ae019de2c0bdf81b56b570639d +size 65456 diff --git a/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/8412533f-67db-465a-8494-1516f366251f_model.json b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/8412533f-67db-465a-8494-1516f366251f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ee5b83e47f60ff6852718e6bca7680a72650a6f --- /dev/null +++ b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/8412533f-67db-465a-8494-1516f366251f_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49220a8891d0de4105f55a1535df3e2a0bbcd2075100dc08ee70dba3c401f7c9 +size 81554 diff --git a/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/8412533f-67db-465a-8494-1516f366251f_origin.pdf b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/8412533f-67db-465a-8494-1516f366251f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..03fc96846367e991c582ede0e8a885664681fa5b --- /dev/null +++ b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/8412533f-67db-465a-8494-1516f366251f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed930471dc92fa038a911cfde318c5dadd6be674ca0b8b412394e62066dcb6ed +size 2836379 diff --git a/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/full.md b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a1ed6b82e333826a19184e8ecb57bed1967f2cbd --- /dev/null +++ b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/full.md @@ -0,0 +1,236 @@ +# Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models + +Itay Benou1,2 Tammy Riklin Raviv1,2 + +$^{1}$ Department of Electrical and Computer Engineering, Ben-Gurion University of the Negev + $^{2}$ Data Science Research Center, Ben-Gurion University of the Negev + +![](images/f38b1b97e207bcec2eb6d101ad259d4c1da423c76ddd17db6b906ddb700dc007.jpg) +Figure 1. Concept maps generated by our SALF-CBM. Inspired by human visual interpretation, our method first decomposes input images into spatially-localized structures, associated with familiar concepts, independent of a specific task. Explainability of task-specific outputs is obtained by training a final task layer on-top of these maps. + +# Abstract + +Modern deep neural networks have now reached human-level performance across a variety of tasks. However, unlike humans they lack the ability to explain their decisions by showing where and telling what concepts guided them. In this work, we present a unified framework for transforming any vision neural network into a spatially and conceptually interpretable model. We introduce a spatially-aware concept bottleneck layer that projects "black-box" features of pre-trained backbone models into interpretable concept maps, without requiring human labels. By training a classification layer over this bottleneck, we obtain a self-explaining model that articulates which concepts most influenced its prediction, along with heatmaps that ground them in the input image. Accordingly, we name + +this method "Spatially-Aware and Label-Free Concept Bottleneck Model" (SALF-CBM). Our results show that the proposed SALF-CBM: (1) Outperforms non-spatial CBM methods, as well as the original backbone, on a variety of classification tasks; (2) Produces high-quality spatial explanations, outperforming widely used heatmap-based methods on a zero-shot segmentation task; (3) Facilitates model exploration and debugging, enabling users to query specific image regions and refine the model's decisions by locally editing its concept maps. + +# 1. Introduction + +Humans often rationalize visually-based assessments or conclusions by describing what they have seen and where they have seen it, using both semantic concepts and their + +spatial locations. For example, an image of a dog wearing glasses and a hat, as shown at the top of Figure 1, is likely to be interpreted as playful or funny due to the unexpected spatial composition of concepts. Notably, this mechanism operates independently of a specific task; even when looking for a dog in the bottom image of Figure 1, one may notice the tennis ball in its mouth and the pot next to it. + +Similarly, the ability to explain AI models using spatially-aware concepts is crucial for elucidating their decision-making processes. Such an approach enables the introduction of quality control mechanisms, i.e., understanding the underlying causes of a model's behavior and adjusting it when necessary. These capabilities are essential for ensuring the safe and transparent integration of deep neural networks into critical fields like medical imaging and autonomous driving, as required by the AI Act recently passed by the European Parliament [1]. + +Most current explainable AI (XAI) methods, however, provide either spatial or concept-based explanations. Spatial approaches, generally referred to as attribution methods, produce heatmaps that highlight the image regions most contributing to the model's output. These heatmaps are generated either by propagating gradients through the model with respect to its input [5, 7, 25, 26, 28-31], or by using attribution-propagation methods [2, 4, 6, 9, 20, 32, 37] that distribute "relevance" (i.e., the contribution of a neuron to the output) backwards through the network, layer by layer. While these methods can visualize the model's spatial attention, in the absence of semantic descriptions, their output can be ambiguous [10, 16]. + +Concept Bottleneck Models (CBMs) [17, 18, 22, 33, 36, 40], on the other hand, are an increasingly popular method for obtaining concept-based explanations. Unlike attribution methods, CBMs provide ante-hoc explanations—i.e., their explainability mechanism is embedded into the model itself. Current CBMs work by introducing a non-spatial bottleneck layer that maps model features to an interpretable concept space, followed by training a final output layer over these concepts. This design ensures that CBMs are highly interpretable, as their predictions are directly based on the concepts used to explain them. However, existing CBMs provide global concept-based explanations without localizing them in the image. Moreover, the interpretable bottleneck layer often comes at the expense of the final task accuracy, which limits their applicability. + +In this work, we present a spatially-aware CBM that combines concept-based explanations with the ability to visually ground them in the input image. In contrast to traditional CBMs, we preserve the spatial information of features and project them into a spatial concept space. This is achieved in a label-free manner by leveraging the capability of CLIP [23] to produce local image embeddings using visual prompts [27]. Accordingly, we name + +our method "spatially-aware and label-free CBM" (SALF-CBM). The main contributions of our work are as follows: (1) Novel unified framework: we present the first label-free CBM that provides both concept-based (global) and heatmap-based (local) explanations. (2) Classification results: SALF-CBM outperforms non-spatial CBMs on several classification tasks, and can even achieve better classification results than the original (non-CBM) backbone model. (3) Heatmap quality: our method produces high-quality heatmaps that can be used for zero-shot segmentation. We demonstrate their advantage over widely-used heatmap-based methods in both qualitative and quantitative evaluations. (4) Explain anything: SALF-CBM facilitates interactive model exploration and debugging, enabling users to inquire about concepts identified in specific image regions, and to adjust the model's final prediction by locally refining its concept maps. (5) Applicability: Our method is model-agnostic and can be applied to both CNNs and transformer architectures, while not introducing any additional learnable parameters compared to non-spatial CBMs. + +# 2. Related Work + +XAI methods for computer vision can be categorized by two axes: local (heatmap-based) vs. global (concept-based) approaches, and post-hoc vs. ante-hoc explanations. In this section, we overview existing methods along these lines. + +Heatmap-based explainability. This refers to a family of post-hoc explainability techniques, often called attribution methods, that visualize the parts of the input image that contribute most to the model's output. Gradient-based methods generate explainable heatmaps by backpropagating gradients with respect to the input of each layer. Some of these methods, such as FullGrad [30], are class-agnostic as they produce roughly identical results regardless of the output class [29, 31], while others, such as GradCAM [25], generate class-dependent heatmaps [8, 28]. This property is essential when the true class is ambiguous. While widely used, their main drawback is high sensitivity to gradient noise, which may render their outcomes impractical [3]. To address this issue, some Class Activation Maps (CAM) methods [39], such as ScoreCAM [34], produce gradient-free explanation maps. + +Attribution propagation methods decompose the output of a model into the contributions of its layers by propagating "relevance" in a recursive manner, without exclusively relying on gradients. Common attribution propagation methods, such as Layer-wise Relevance Propagation (LRP) [6], are primarily applicable to Convolutional Neural Networks (CNNs) [20, 26, 37]. Later approaches have been adapted to accommodate vision transformers (ViTs) [11], exploiting their built-in self-attention mechanism [2, 9, 23, 32]. We note that, unlike our SALF-CBM, both gradient-based and attribution propagation methods do not provide concept + +based explanations. Additionally, since these are post-hoc techniques, they do not enable test-time user intervention. Concept-based explainability. An alternative way of explaining vision models is by using human-interpretable concepts. Various methods provide such explanations in a post-hoc manner. For example, Testing Concept Activation Vector (TCAV) [15] measures the importance of user-defined concepts to the model's prediction by training a linear classifier to distinguish between concepts in their activation space. However, this requires labeling images with their corresponding concepts in advance. ACE [13] extends this idea by applying multi-resolution segmentation to images from the same class, followed by clustering similar segments into concepts to compute their TCAV scores. Similarly, Invertible Concept Embeddings (ICE) [38] and Concept Recursive Activation Factorization (CRAFT) [12] provide concept-based explanations using matrix factorization of feature maps. CRAFT also generates attribution maps that localize concepts in the input image. These methods, however, are mostly applicable to CNN architectures [13], which use non-negative activations [12, 38], and therefore cannot be directly applied to other types of models. Additionally, since they provide post-hoc explanations, they do not enable test-time user intervention. + +In contrast, Concept-Bottleneck Models (CBMs) is a family of ante-hoc interpretable models whose explainability mechanism is an integral part of the model itself. CBMs operate by introducing a concept-bottleneck layer into pretrained models, before the final prediction layer. The goal of this bottleneck is to project features into an interpretable concept space, where each neuron corresponds to a single concept. Unlike post-hoc methods, the output of CBMs is directly based on interpretable concepts, making them easily explainable and allowing user intervention by modifying concept neurons activations. In the original CBM work [17], the concept bottleneck layer was trained using manual concept annotations, limiting its ability to scale to large datasets. Recently, Post-Hoc CBM (P-CBM) [36] and Label-Free CBM (LF-CBM) [22] addressed this issue by leveraging CLIP to assign concept scores for training images, thus not requiring concept annotations. LF-CBM also presented an automatic process for creating a list of task-relevant concepts using GPT-3. While showing good interpretability results, both P-CBM and LF-CBM present a performance drop on the final classification task compared to the original (non-CBM) model. Additionally, unlike our SALF-CBM, these methods are limited to global concept explanations, and are unable to localize these concepts within the image. + +# 3. Method + +Given a pre-trained backbone model, we transform it into an explainable SALF-CBM as illustrated in Figure 2a: + +Step 1: Automatically generate a list of task-relevant visual concepts; Step 2: Using CLIP, compute a spatial concept similarity matrix that quantifies what concepts appear at different locations in the training images; Step 3: Train a spatially-aware Concept Bottleneck Layer (CBL) that projects the backbone's "black-box" features into interpretable concept maps. Step 4: Train a sparse linear layer over the pooled concept maps to obtain the model's final prediction. We describe each step in the following sections. + +# 3.1. Concept list generation + +Let $\mathcal{X}$ denote an image classification dataset with $N$ training images $\{x_{1},\ldots ,x_{N}\}$ and $L$ possible classes. We aim to generate a list of visual concepts $\mathcal{T}$ that is most relevant to the target classes, without relying on human experts. For this purpose, we follow the automatic procedure described in [22]. First, an initial concept list is obtained by prompting GPT as follows: "List the most important features for recognizing something as a {class}"; "List the things most commonly seen around a {class}"; and "Give superclasses for the word {class}", for each class in the dataset. Concepts that are too long, too similar to one of the classes or to another concept, or do not appear in the training data - are then discarded. The resulting filtered list of $M$ concepts is denoted by $\mathcal{T} = \{t_1,\dots ,t_M\}$ . See [22] for full details. + +# 3.2. Local image-concept similarities + +Vision-language models such as CLIP have been widely used for obtaining global image descriptions, as in non-spatial CBMs [22, 36]. Here, we aim to extend this approach to locally describe different image regions using visual concepts. Inspired by [27], we leverage CLIP's visual prompting property by drawing a red circle around specific image regions, which enables CLIP to focus on these areas while preserving global context. We apply this property to our training set as illustrated in Figure 2b. Formally, let $x_{n} \in \mathcal{X}$ denote a training image with spatial dimensions $H \times W$ . We create a uniform grid of $\tilde{H} \times \tilde{W}$ locations in the image with integer strides $d_{H}$ and $d_{W}$ , i.e., $d_{H} = \left\lfloor \frac{H}{\tilde{H} - 1} \right\rfloor$ and $d_{W} = \left\lfloor \frac{\tilde{W}}{\tilde{W} - 1} \right\rfloor$ . For each $x_{n}$ , a set of $\tilde{H} \cdot \tilde{W}$ augmented images is then obtained by drawing a red circle with radius $r$ around each location in the grid. We denote by $x_{n}^{(h,w)}$ the image $x_{n}$ with a red circle located at $(h,w)$ . Next, we compute a local similarity score between a visual concept $t_{m} \in \mathcal{T}$ and the image at location $(h,w)$ as follows: $P[n,m,h,w] = \frac{I_n^{(h,w),T_m}}{\|I_n^{(h,w)}\| \|T_m\|}$ , where $I_{n}^{(h,w)} = E_{I}(x_{n}^{(h,w)})$ and $T_{m} = E_{T}(t_{m})$ denote the CLIP embeddings of the augmented image and the concept, respectively. As demonstrated in [27], this score measures how well the concept $t_{m}$ describes the image region highlighted by the red circle. By computing this score for every spatial location $(h,w)$ in the grid, we obtain a concept sim + +![](images/848eb96b8164d79299f46332ecd42be2fc556839d42e33eaf4fd87dc0794b21b.jpg) +(a) Training process scheme. + +![](images/82e4028881db5a5d050ee1ebd45aba15020be245ec2d46f5569f9cd8b86f8445.jpg) +(b) Local image-concept similarities block. +Figure 2. SALF-CBM training: (a) Given a pre-trained backbone model, we: (i) Generate task-relevant concepts; (ii) Describe training images using local image-concept similarities; (iii) Train a spatially-aware concept bottleneck to project features into interpretable concept maps; (iv) Train a sparse classification layer over these maps. (b) Local image-concept similarities computation using visual prompting. + +ality map for the entire image $x_{n}$ , as shown in Figure 2b. A complete spatial concept similarity matrix $P$ is then constructed by calculating these local similarity scores for all concepts $t_m \in \mathcal{T}$ and all images $x_{n} \in \mathcal{X}$ in the training set. This matrix is computed once, prior to training, and is later used for learning the spatially-aware concept bottleneck layer. We note that the grid resolution $\tilde{H} \times \tilde{W}$ and the circle radius $r$ are hyper-parameters, where $\tilde{H}$ and $\tilde{W}$ control the coarseness of the concept similarity map, and $r$ defines the receptive field around each location. We optimize these hyper-parameters per-dataset (see supplementary). + +# 3.3. Training the concept-bottleneck layer + +We aim to learn a bottleneck layer $g$ that linearly projects "black-box" feature maps $f(x)$ of a pre-trained backbone model into interpretable concept maps. Rather than spatially pooling the backbone's features as in conventional CBMs, we retain their spatial information, and resize them to fit the grid's dimensions $(\tilde{H} \times \tilde{W})$ using a bilinear interpolation. We then use a single $1 \times 1$ convolution layer with $M$ output channels to produce the desired concept maps, i.e., $c(x) = g(f(x)) \in \mathbb{R}^{M \times \tilde{H} \times \tilde{W}}$ . We denote the full list of concept maps for all training images $x_{n} \in \mathcal{X}$ by $C[n,m,h,w] = [c(x_1),\ldots ,c(x_N)]$ . In order to obtain concept maps that match the image-concept similarities in $P$ , we train our bottleneck layer using an extended version of the cubic cosine similarity loss from [22] as follows: + +$$ +\mathcal {L} _ {C B L} = - \sum_ {m = 1} ^ {M} \sum_ {h, w} \operatorname {s i m} (q [ m, h, w ], p [ m, h, w ]) \tag {1} +$$ + +where $q[m,h,w]$ denotes $C[:,m,h,w],p[m,h,w]$ denotes $P(:,m,h,w]$ and $\operatorname{sim}(\cdot ,\cdot)$ denotes the cubic cosine similarity function $\operatorname{sim}(q,p) = \frac{\bar{q}\cdot\bar{p}}{\|\bar{q}\|_{\|}\|\bar{p}\|}$ . Here, $\bar{q}$ and $\bar{p}$ are nor + +malized to have zero-mean and raised elementwise to the power of three to emphasize strong concept-image matches. We note that our spatial bottleneck layer requires the same number of parameters as the fully-connected bottleneck typically used in non-spatial CBMs [17, 22, 36]. Furthermore, our bottleneck layer accommodates both CNN and vision transformer architectures: For CNN backbones, feature maps are used directly as inputs to the bottleneck, while for ViTs, the patch tokens are reshaped back into their original spatial arrangement. See details in the supplementary. + +# 3.4. Training the final classification layer + +Once the concept bottleneck layer is trained, we spatially pool its output concept maps $c(x)$ to obtain global concept activations $c^*(x)$ , each corresponds to a single visual concept $t_m$ . We aim to explain each output class of our model using a small set of interpretable concepts. We therefore train a sparse linear layer on top of $c^*(x)$ to obtain the final classification scores $z = Wc^* + b$ and the predicted class $\hat{y} = \arg \max(z)$ . Here, $W$ and $b$ denote the classification weights and bias term, respectively. This layer is trained in a fully-supervised manner with the following loss function, using the GLM-SAGA optimizer [35]: + +$$ +\sum_ {n = 1} ^ {N} \mathcal {L} _ {c e} \left(W c ^ {*} \left(x _ {n}\right) + b, y _ {n}\right) + \lambda \mathcal {R} (W) \tag {2} +$$ + +where $\mathcal{L}_{ce}$ is the cross-entropy loss, $y_{n}$ is the class label of training image $x_{n}$ , $\lambda$ is the regularization strength and $\mathcal{R}(W) = (1 - \alpha)\frac{1}{2}\| W\| _F + \alpha \| W\|_{1,1}$ is the elastic net regularization term, where $\| W\| _F$ is the Forbenius norm and $\| W\|_{1,1}$ is the elementwise matrix norm. + +# 3.5. Test-time explainability + +Once trained, SALF-CBM provides explanations at multiple levels of granularity as an integral part of its forward pass, without any external tools or additional computations. Global decision rules. As described in Section 3.4, SALF-CBM's final prediction is a linear combination of a sparse set of concept activations. Therefore, one can gain an intuitive understanding of the model's decision rules simply by examining which concepts $t_m \in \mathcal{T}$ are connected to a specific class $l \in \{1, \dots, L\}$ by non-zero weights. For instance, in Figure 3, we show Sankey diagrams visualizing the class weights of SALF-CBM trained on ImageNet, for two different classes: "crate" and "toy store". + +Concept-based explanations ("tell what"). We aim to explain individual model decisions by evaluating the contribution of a visual concept $t_m \in \mathcal{T}$ to the model's output $\hat{y}$ on a given test image $x$ . This is achieved by computing a contribution score $S(x, m, \hat{y} = l) = W[m, l]c^*(x)[m]$ , where $m$ is the concept index, $l \in \{1, \dots, L\}$ is the index of the predicted class and $c^*(x)[m]$ is the global concept activation, normalized by its mean and standard deviation on the training data. Since $W[:, l]$ is sparse, the majority of contribution scores are zero, so the model's prediction can be explained by a small set of $k$ concepts whose absolute contribution scores are the highest. In Figure 3, we illustrate the top-3 concepts with the highest contribution scores for the "toy store" and "crate" classes. + +Spatial explanations ("show where"). In addition to global concept-based explanations, our method produces heatmaps which highlight the location of each concept in the input image. Specifically, given the top- $k$ contributing concepts, we upsample their associated concept maps $c(x)[m]$ to the input image dimensions, using a bilinear interpolation. Examples of heatmaps associated with the most contributing concepts for two different output classes are presented in Figure 3. We note that the resolution of the heatmaps can be controlled by adjusting the density of the visual prompting grid, as discussed in Section 3.2. + +# 3.6. Model exploration and debugging + +We introduce two interactive features that enable users to explore their model's decision-making process and debug failure cases in an intuitive manner. + +**Explain Anything.** Inspired by the Segment Anything Model (SAM) [19], this feature allows users to actively "prompt" SALF-CBM with inputs such as points, bounding boxes, or free-form masks, to explore what visual concepts were recognized in the specified region-of-interest (ROI). Specifically, given an image $x$ , the computed concept maps $c(x)$ (upsampled to the image dimensions) and a user-provided ROI in the form of a binary mask $I$ , we compute the aggregated activation of each concept within the ROI: $a(x, m \mid I) = \sum I \odot c(x)[m]$ , where $\odot$ represents + +elementwise multiplication. By presenting the top- $k$ concepts with the strongest aggregated activation, we provide a concise overview of the model's perception of the ROI. We note that in addition to user-provided ROIs, our method supports segmentation masks from tools such as SAM to automatically produce objects descriptions. + +Local user intervention. We enable users to intervene in the model's final prediction by suggesting counterfactual explanations in specific image regions, i.e., "how would the model's prediction change if concept $\mathcal{A}$ were more/less activated at location $\mathcal{B}$ ?". Given an image $x$ , the concept map $c(x)[m]$ of a specific concept $t_m$ and the predicted class $\hat{y}$ , one can locally edit the concept map according to their judgment and understanding of the task, as follows: $c(x)[m] \gets c(x)[m] + \beta I$ , where $I$ is a binary mask of the edited region and $\beta$ is a correction factor that can be either positive or negative. By tuning concept activations up or down in specific regions and re-running the final classification layer, one can observe how the model adjusts its prediction $\hat{y}$ based on the revised concept maps. + +# 4. Experiments + +We thoroughly evaluate the different components of our method. In section 4.1, we test its classification accuracy compared to several baselines, on different large-scale datasets. In section 4.2, we present qualitative and quantitative evaluations of our SALF-CBM's heatmaps in comparison to several other heatmap-based methods. In section 4.3, we validate the concepts learned by SALF-CBM's bottleneck layer by conducting a user study. In section 4.4, we demonstrate how the proposed Explain Anything and user intervention features are used to debug model errors. Additional results are provided in the supplementary. + +# 4.1. Classification accuracy + +Experimental setup. We test our method on a diverse set of classification datasets: CUB-200 (fine-grained bird-species classification), Places365 (scene recognition) and ImageNet. We train a SALF-CBM on each of the three datasets, using a appropriate backbone model to allow fair comparisons with competing CBM methods [22, 36]: For CUB-200 we use a ResNet-18 pre-trained on CUB-200, and for both ImageNet and Places365 we use a ResNet-50 pretrained on ImageNet. For each dataset, we use the same initial concept list and regularization parameters $\alpha$ and $\lambda$ as in [22], resulting in 370 concepts for CUB-200, 2544 concepts for Places365 and 4741 concepts for ImageNet. For computing local image-concept similarities, we use CLIP ViT-B/16 and a visual prompting grid of $7 \times 7$ with $r = 32$ for all experiments. Results with different grid parameters and with a ViT backbone are provided in the supplementary. Results. Table 1 presents the classification accuracy of our SLAF-CBM compared to several other methods: (1) the + +![](images/aa614b1064329f4f867391e06de350a2526386081821610e9fbead2d76a00fe7.jpg) +Figure 3. Test-time explainability. Global decision rules can be inferred by visualizing the sparse class weights. Individual model decisions are explained globally and locally by the concept contribution scores and their associated spatial heatmaps. + +standard pre-trained backbone model with its original classification layer; (2) the standard backbone model with a sparse classification layer (reported from [22]); (3) post-hoc CBM (P-CBM) [36]; and (4) Label-Free CBM (LF-CBM) [22]. We note that in P-CBM [36], they do not report their results on ImageNet and Places365, and it is unclear how to scale it to those datasets. For fair comparisons, results with sparse and non-sparse classification layers are shown separately. We see that when using a sparse final layer, our SALF-CBM outperforms both P-CBM and LF-CBM on all three datasets. Notably, our method is the best performing sparse method on the two larger-scale datasets (Places365 and ImageNet), outperforming even the original backbone with a sparse final layer. To demonstrate the high-limit potential of our method, we assess its performance with a non-sparse final layer. Remarkably, the non-sparse SALF-CBM achieves better classification results than original (non-sparse) model on both ImageNet and Places365, even though its predictions are based on interpretable concepts. + +These results indicate that SALF-CBM facilitates model interpretability without compromising performance; in fact, it can outperform the original backbone model when using a comparable final layer (i.e., sparse or non-sparse). We also note that the performance gap between the sparse and non-sparse SALF-CBMs is relatively small (less than $1\%$ on ImageNet), indicating that our model effectively captures the full span of possible explanations using a sparse set of concepts. + +# 4.2. Beyond classification: zero-shot segmentation + +Experimental setup. We conduct a quantitative analysis of the heatmaps generated by our method in a zero-shot segmentation task. We follow a standard protocol for evaluating heatmap-based explainability methods [9] on ImageNet-segmentation dataset [14], a subset of the ImageNet validation set containing 4,276 images with ground-truth segmentation masks of the class object. In order for our concept maps to correspond to ImageNet classes, we train a SLAF-CBM with a ResNet-50 backbone on Ima + +
ModelSparseDataset
CUB-200Places365ImageNet
StandardYes75.96%38.46%74.35%
P-CBM [36]Yes59.60%N/AN/A
LF-CBM [22]Yes74.31%43.68%71.95%
SALF-CBMYes74.35%46.73%75.32%
StandardNo76.70%48.56%76.13%
SALF-CBMNo76.21%49.38%76.26%
+ +Table 1. Classification accuracy. Our method outperforms P-CBM and LF-CBM on all three datasets, and is the highest performing model on ImageNet and Places365. Results are shown separately for sparse and non-sparse final layers. Best results are in bold and 2nd-best are underlined. + +geNet, using a concept list of the form "An image of a {class}", where {class} refers to each of the ImageNet classes. According to [9], the resulting heatmaps are bina- rized to obtain a foreground/background segmentation, and evaluated with respect to the ground-truth masks based on three metrics: mean average precision (mAP) score, mean intersection-over-union (mIOU) and pixel accuracy. + +Results. Table 2 presents the zero-shot segmentation results of our method, compared to several widely-used explainability methods: LRP [6], integrated gradients (IG) [31], GradCAM [25], GradCAM++ [8], ScoreCAM [34], and FullGrad [30]. Notably, our SALF-CBM achieves the best pixel accuracy and mIOU segmentation scores, and the second best mAP. Specifically, our method demonstrates significant improvements in pixel accuracy and mIOU (+3.9% and +2.52% over the 2nd-best method, respectively), indicating that our heatmaps are consistently better aligned with the ground-truth masks. In Figure 4, we present a qualitative comparison to the baseline methods, for different images from the ImageNet validation set. We observe that LRP [6] and IG [31] typically produce noisy results, and struggle to accurately localize the class object. GradCAM [25], GradCAM++ [8], ScoreCAM [34] + +![](images/b6991dff7a24156be0954b82e0f0b85e67f1a678a3bc28ce1c1401938525cf33.jpg) +Figure 4. Qualitative heatmaps comparison. Explanation map of each method with respect to the ground-truth class (from top to bottom): "Cheeseburger", "Bell-cote", "Monarch butterfly" and "Goose". + +and FullGrad [30] manage to highlight the target region, but also include unrelated background areas. Conversely, our method generates heatmaps that accurately captures the class object, thus providing more precise explanations. + +
MethodPixel Acc. ↑mIoU ↑mAP ↑
LRP [6]69.52%36.85%69.95%
IG [31]68.49%46.59%73.46%
GradCAM [25]71.34%53.34%83.88%
GradCAM++ [8]71.31%53.56%83.93%
ScoreCAM [34]69.56%51.44%81.78%
FullGrad [30]73.04%55.78%88.35%
SALF-CBM76.94%58.30%85.31%
+ +Table 2. Zero-shot segmentation results. Our SALF-CBM achieves the highest mIoU and pixel accuracy, and the second highest mAP. Best results are in bold, 2nd-best are underlined. + +# 4.3. Bottleneck interpretability validation + +We conduct a user study to quantitatively validate that SALF-CBM's bottleneck neurons in-fact correspond to their designated target concepts. + +Experimental setup. Following [24], we evaluate global concept neurons $c^*$ from SLAF-CBM's bottleneck layer compared to output neurons of the baseline ResNet-50 backbone. We assign concept labels to baseline neurons us + +ing CLIP-Dissect [21] and rank neurons from both models by interpretability scores using CLIP-Dissect's soft-WPMI metric. We then sample 10 neurons from the top $30\%$ interpretable neurons and 10 from the remaining $70\%$ for each model. For each neuron, we show 25 users the five most activated test images and ask them to rate from 1-5: (a) semantic consistency: "Do these images share a common semantic concept?" and (b) concept accuracy: "Does [neuron label] describe a common concept shared by these images?". + +Results. Figure 5 shows that SALF-CBM achieves significantly better scores in both semantic consistency and concept accuracy, demonstrating improved interpretability compared to the baseline across all neuron interpretability groups. See additional results in the supplementary. + +![](images/0d081814441a33d7ec6513f00081bd9dd3f5d886a6fbaa8c1dca97deba52f220.jpg) +Figure 5. User study results. + +![](images/d227026242f3d702f1320a3ec23146e27b89b7318d434661077b2819511b920d.jpg) +Figure 6. Explain Anything. For each image, we prompted SALF-CBM with two different ROI masks produced by SAM [19] (red and blue regions). Our method provides accurate concept descriptions for each ROI. + +![](images/561ba0e043a77a22be782080d145f15b73578d297f7b51817b3ef0329e40f877.jpg) +Most activated concepts in red ROI +Most activated concepts in blue ROI + +droopy lips and ears: 1.79 +a stocky body: 1.7 +a cute, pug-like face: 1.68 + +a pet food bowl: 2.51 +dishes: 2.47 +a cook: 2.41 + +![](images/31231ada42c4fcf5a639c901b202bb864e65721ab219f82153448cc84e77b890.jpg) +Figure 7. Model debugging with Explain Anything. We reveal that the model misclassified the image since it mistakenly identified traffic lights as street signs. Its prediction is corrected by locally editing the relevant concepts maps in the examined ROI. + +# 4.4. Model exploration and debugging + +We first qualitatively validate our Explain Anything feature (Section 3.6) on different images from the SAM dataset [19]. For each image, we prompt SALF-CBM (with ResNet-50 backbone pre-trained on ImageNet) with two different ROI masks automatically obtained by SAM [19], highlighted in red and blue. As shown in Figure 6, SALF-CBM generates informative region-specific concept descriptions that accurately correspond to the selected ROIs. For instance, in the child's drawing (left image), the dress (blue mask) and grass area (red mask) are correctly identified as fabric-like material and a field or lawn, respectively. + +Next, we demonstrate the usefulness of Explain Anything in diagnosing classification errors, and facilitating targeted corrections using local user intervention. We present a case study from the ImageNet validation set, where our model miscalssified a "traffic light" image as a "parking meter", as shown in Figure 7. Applying Explain Anything to the traffic lights region in the image reveals that the model primarily detected sign-related concepts there. However, as indicated by the class weights visualization, these concepts are not associated with the correct "traffic light" class. This misalignment, along with the presence of street-related features in the image, led the model to incorrectly classify the image as a "parking meter." To rectify that, we locally edit two concepts maps associated with the true class - "a flashing light" and "the ability to change color" - within the selected ROI. Specifically, we increase their activation there + +by a correction factor of $\beta = 1$ . As illustrated in the figure, this mild adjustment promoted these concepts to the top-5 most activated concepts in the ROI, subsequently adjusting the model's output to the correct class. + +# 5. Conclusions + +In this work, we presented SALF-CBM, a novel framework for transforming any vision neural network into an explainable model that provides both concept-based and spatial explanations for its predictions. We showed that SALF-CBM enhances model interpretability without compromising performance, outperforming both existing CBMs and the original model across several classification tasks. We demonstrated that it produces high-quality spatial explanations, achieving better zero-shot segmentation results compared to widely used heatmap-based explainability methods. + +Additionally, we introduced interactive capabilities for model exploration and debugging, demonstrating their effectiveness in diagnosing and correcting model errors. We believe that such features are particularly valuable for high-stakes applications like medical imaging and autonomous driving. By providing expert practitioners with intuitive tools to understand and adjust model decisions, our approach can boost confidence and support safer deployment in these critical fields. Looking ahead, as new VLMs are developed across various domains, our findings could help inform the design of more powerful interpretability tools for a broad spectrum of AI applications. We plan to explore these directions in future work. + +# References + +[1] Madiaga: Artificial intelligence act (2023). http://www.europarl.europa.eu/RegData/etudes/BRI/2021/698792/EPRS_BRI(2021)698792_EN.pdf. 2 +[2] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928, 2020. 2 +[3] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in neural information processing systems, 31, 2018. 2 +[4] Sebastian Bach, Alexander Binder, Gregoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 10 (7):e0130140, 2015. 2 +[5] Sarah Adel Bargal, Andrea Zunino, Vitali Petsiuk, Jianming Zhang, Kate Saenko, Vittorio Murino, and Stan Sclaroff. Guided zoom: Zooming into network evidence to refine fine-grained model decisions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4196-4202, 2021. 2 +[6] Alexander Binder, Gregoire Montavon, Sebastian Lapischkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In Artificial Neural Networks and Machine Learning-ICANN 2016: 25th International Conference on Artificial Neural Networks, Barcelona, Spain, September 6-9, 2016, Proceedings, Part II 25, pages 63-71. Springer, 2016. 2, 6, 7 +[7] Chunshui Cao, Xianming Liu, Yi Yang, Yinan Yu, Jiang Wang, Zilei Wang, Yongzhen Huang, Liang Wang, Chang Huang, Wei Xu, et al. Look and think twice: Capturing top-down visual attention with feedback convolutional neural networks. In Proceedings of the IEEE international conference on computer vision, pages 2956-2964, 2015. 2 +[8] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 839-847. IEEE, 2018. 2, 6, 7 +[9] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 782-791, 2021. 2, 6 +[10] Julien Colin, Thomas Fel, Rémi Cadène, and Thomas Serre. What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. Advances in neural information processing systems, 35: 2832-2845, 2022. 2 +[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[12] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and + +Thomas Serre. Craft: Concept recursive activation factorization for explainability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 3 +[13] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in neural information processing systems, 32, 2019. 3 +[14] Matthieu Guillaumeun, Daniel Kuttel, and Vittorio Ferrari. Imagenet auto-annotation with segmentation propagation. International Journal of Computer Vision, 110:328-348, 2014. 6 +[15] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 3 +[16] Sunnie SY Kim, Nicole Meister, Vikram V Ramaswamy, Ruth Fong, and Olga Russakovsky. Hive: Evaluating the human interpretability of visual explanations. In European Conference on Computer Vision, pages 280-298. Springer, 2022. 2 +[17] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning (ICML), pages 5338-5348. PMLR, 2020. 2, 3, 4 +[18] Max Losch, Mario Fritz, and Bernt Schiele. Interpretability beyond classification output: Semantic bottleneck networks. arXiv preprint arXiv:1907.10882, 2019. 2 +[19] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 5, 8 +[20] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition, 65:211-222, 2017. 2 +[21] T. Oikarinen and T-W. Weng. Clip-dissect: Automatic description of neuron representations in deep vision networks. arXiv preprint arXiv:2204.10965, 2022. 7 +[22] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. In International Conference on Learning Representations (ICLR), 2023. 2, 3, 4, 5, 6 +[23] A. Radford, J.W. Kim, C. Hallacy, A. Ramesh, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), pages 8748-8763, 2021. 2 +[24] Sukrut Rao, Sweta Mahajan, Moritz Böhle, and Bernt Schiele. Discover-then-name: Task-agnostic concept bottlenecks via automated concept discovery. In European Conference on Computer Vision, pages 444–461. Springer, 2024. 7 +[25] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE in- + +ternational conference on computer vision, pages 618-626, 2017. 2, 6, 7 +[26] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In International conference on machine learning, pages 3145-3153. PMIR, 2017. 2 +[27] A. Shtedritski, C. Rupprecht, and A. Vedaldi. What does clip know about a red circle? visual prompt engineering for VLMs, 2023. 2, 3 +[28] Karen Simonyan. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 2 +[29] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. SmoothGrad: removing noise by adding noise, 2017. 2 +[30] Suraj Srinivas and François Fleuret. Full-gradient representation for neural network visualization. Advances in neural information processing systems, 32, 2019. 2, 6, 7 +[31] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International conference on machine learning, pages 3319-3328. PMLR, 2017. 2, 6, 7 +[32] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. arXiv preprint arXiv:1905.09418, 2019. 2 +[33] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 10962-10971, 2023. 2 +[34] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 2, 6, 7 +[35] Eric Wong, Shibani Santurkar, and Aleksander Madry. Leveraging sparse linear layers for debuggable deep networks. arXiv preprint arXiv:2105.04857, 2021. 4 +[36] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 2, 3, 4, 5, 6 +[37] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. International Journal of Computer Vision, 126(10):1084-1102, 2018. 2 +[38] Ruihan Zhang, Prashan Madumal, Tim Miller, Krista A Ehinger, and Benjamin IP Rubinstein. Invertible concept-based explanations for cnn models with non-negative concept activation vectors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11682-11690, 2021. 3 +[39] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2 +[40] Bolei Zhou, Yiyou Sun, David Bau, and Antonio Torralba. Interpretable basis decomposition for visual explanation. In + +Proceedings of the European Conference on Computer Vision (ECCV), pages 119-134, 2018. 2 \ No newline at end of file diff --git a/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/images.zip b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aa128357972a2be6576fabb61f67a7b5337cfb5a --- /dev/null +++ b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6445783f696ed491a955a93c2a0c72705edc25468d0a7637be72c576202c6d2 +size 607555 diff --git a/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/layout.json b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7c3ee5ab2e6c8d36e29e83a55e1ea1ef996dea0d --- /dev/null +++ b/CVPR/2025/Show and Tell_ Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4944ef898c31b19bb130a22b7f9a2fbab957ebb8a0e304a98f328e4544268e26 +size 340401 diff --git a/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/61e795ea-4574-473e-b3fc-6f3535dfb4bb_content_list.json b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/61e795ea-4574-473e-b3fc-6f3535dfb4bb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..036e4ed1d211d51e00e53dbb07dcc6a6e574cf4e --- /dev/null +++ b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/61e795ea-4574-473e-b3fc-6f3535dfb4bb_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a758e0b87ab02551563a84c7a3c0cc9aaee5b4b6cbfa1c03569599c40cf3edb +size 77266 diff --git a/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/61e795ea-4574-473e-b3fc-6f3535dfb4bb_model.json b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/61e795ea-4574-473e-b3fc-6f3535dfb4bb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f1657be7f68cbe7eba08a4f6729b8b733de65384 --- /dev/null +++ b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/61e795ea-4574-473e-b3fc-6f3535dfb4bb_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70587f4a0461386663700f9c7ce2d885a7a3794dc859e2da54d7b2ba598afec8 +size 98426 diff --git a/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/61e795ea-4574-473e-b3fc-6f3535dfb4bb_origin.pdf b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/61e795ea-4574-473e-b3fc-6f3535dfb4bb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b4bba34caa9724024f46f67bfe1c94eb75dae756 --- /dev/null +++ b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/61e795ea-4574-473e-b3fc-6f3535dfb4bb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f54b1bea1f337c3eff9e8f43ded196ee060b78c6bae560e1fe256b19304d0ca +size 2380329 diff --git a/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/full.md b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e40f2aba2229cacd8439ef59e71e10747d3d21cd --- /dev/null +++ b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/full.md @@ -0,0 +1,277 @@ +# ShowHowTo: Generating Scene-Conditioned Step-by-Step Visual Instructions + +# Tomáš Souček1 + +# Prajwal Gatti + +# Michael Wray + +# Ivan Laptev3 + +# Dima Damen + +# Josef Sivic1 + +$^{1}$ CIIRC CTU $^{2}$ University of Bristol $^{3}$ MBZUAI + +tomas.soucek@cvut.cz + +https://soczech.github.io/showhowto/ + +![](images/bcdc4e1342caa8ce1926449057f8341862a3b2596402f4f60316090a0bd93059.jpg) +Figure 1. Given an input image (left) and ordered step-by-step textual instructions for a task (top), ShowHowTo generates an image sequence of visual instructions. Rows 1 and 2 demonstrate the generation of visual instructions for two recipes starting from the same input image. Rows 2 and 3 show the generation of visual instructions for the same recipe but conditioned on different input images. ShowHowTo generates scene-consistent (e.g., consistency in the person and cutting board) and temporally consistent image sequences (e.g., the bowl of tortilla chips or plate of chicken skewers) that faithfully capture the instructions (e.g., cutting, frying, brushing, adding etc.). + +# Abstract + +The goal of this work is to generate step-by-step visual instructions in the form of a sequence of images, given an input image that provides the scene context and the sequence of textual instructions. This is a challenging problem as it requires generating multi-step image sequences to achieve a complex goal while being grounded in a specific environment. Part of the challenge stems from the lack of large-scale training data for this problem. The contribution of this work is thus three-fold. First, we introduce an automatic approach for collecting large step-by-step visual instruction training data from instructional videos. We apply this approach to one million videos and create a large-scale, + +high-quality dataset of 0.6M sequences of image-text pairs. Second, we develop and train ShowHowTo, a video diffusion model capable of generating step-by-step visual instructions consistent with the provided input image. Third, we evaluate the generated image sequences across three dimensions of accuracy (step, scene, and task) and show our model achieves state-of-the-art results on all of them. Our code, dataset, and trained models are publicly available. + +# 1. Introduction + +With the immense success of large vision-language models and the rise of wearable devices, we rapidly approach the era of personalized visual assistants. This technology promises to help us in a variety of everyday tasks and numerous scenarios, such as preparing a Michelin-star dish, taking care of plants, or fixing a bicycle. Unlike generic instructional videos, visual assistants will provide guidance + +and feedback for specific environments and task variations. + +Besides being useful for people, automated visual guidance has been recently explored and shown to be beneficial in robotics. For example, [9, 32, 43, 65] generate images of intermediate goals and use them as guidance for manipulation policies. Other recent methods [8, 19, 38, 57] derive robotics policies from videos specifically generated for target tasks and environments. + +When comparing the ability to generate textual step-by-step instructions vs. generating visual instructions, one can see a sharp contrast. State-of-the-art LLMs can reliably provide personalized step-by-step text-only instructions. However, translating such instructions into images and videos still presents considerable challenges. This is because current video generation models, despite their impressive progress over the past years, focus on producing relatively short clips [14, 46, 49, 63] whereas image generation models only produce one image at a time. + +Recent attempts to generate visual instructions either synthesize a single step [35, 36, 53] or are not contextualized to the user's specific environment [12, 41, 45]. In other words, such methods may generate plausible images for each step, however, such images will represent arbitrary settings and can feature tools or ingredients unavailable to the user. In robotics, temporally inconsistent guidance may imply physically implausible demonstrations, resulting in unsuccessful learning of policies. To address this issue, we focus on generating step-by-step visual instructions conditioned on an input image from the user, which we assume showcases their starting position—the environment, ingredients, tools, etc., as illustrated in Figure 1. + +This paper makes the following contributions. (1) We introduce the problem of generating a sequence of visual instructions conditioned on an input image. (2) We introduce a fully automatic approach to collect step-by-step visual instruction training data from in-the-wild instructional videos, creating a large-scale, high-quality dataset of 0.6M step-by-step instruction sequences of 4.5M image-text pairs. (3) We train a video diffusion model capable of generating sparse step-by-step visual instructions consistent with the input image. (4) We evaluate our generated sequences across three aspects (step, scene, and task) and show our model achieves state-of-the-art results on all of them. + +# 2. Related Work + +Datasets of visual instructions. The scale and quality of the training data play a key role in visual instruction generation. Many available datasets combine instructional or egocentric videos and manual temporal annotation of individual steps [3, 51, 54, 70] or use professional illustrations [62]. Yet the requirement of manual annotations makes these sources of data hard to scale to novel tasks and environments. To alleviate the need for manual annotations, self- + +supervised methods have been developed to automatically obtain key steps from in-the-wild videos [21, 39, 52, 58, 59]. These key steps can then be used for visual instruction generation [53]. Recently, the improved capabilities of large language models [2, 4, 18, 20, 33] allowed for solely using video narrations to produce temporal captions, key steps, and instructions [36, 37, 50]. We build on these works to automatically obtain key steps from videos; however, in contrast to the related works, our dataset is constructed completely automatically, is composed of individual instruction frames instead of temporal intervals, and contains significantly fewer errors. Additionally, we focus on the entire domain of instructional videos, not only cooking. + +Conditional video generation. Recently, diffusion models [28] have seen a surge in popularity for generative tasks, including video generation [40, 64]. Initial works extended a U-Net model using space-time factorization for the generation of videos in pixel space [29-31]. With the large cost of generating video, others [11, 15, 23, 27, 40, 64, 69] instead use an auto-encoder to model videos within the latent space, reducing significantly the number of parameters and memory requirements. Video generation models are often conditioned on textual prompts [16, 23, 24, 29, 30, 67], images that act as initial frames [60, 68], or both to generate a sequence of frames [7, 10, 11, 17, 25, 31, 34, 56, 69]. It has also been shown that for temporal consistency, a combination of both image conditioning and textual conditioning is important [55, 56]. However, these methods focus on relatively short video clips and are not able to generate long multi-step sequences of fine-grained instructions that take minutes to execute. + +Generating step-by-step instructions. Both step-by-step visual instructions and continuous videos consist of sequences of frames, yet the instruction sequences differ from the videos significantly. While videos often contain only small pixel-level frame-to-frame changes [61], visual instruction sequences often contain large semantic (e.g., raw $\rightarrow$ cooked) and viewpoint (e.g., inside $\rightarrow$ outside) changes from one key frame to another [41]. Obtaining sufficient training data for visual instructions presents a significant challenge. Therefore, Phung et al. [45] generate step-by-step visual instructions using a pretrained text-conditioned image diffusion model with shared attention across steps to ensure consistency in the generated image sequences, while Menon et al. [41] use illustrations drawn by artists from WikiHow [62] as the training data for text-to-image-sequence generation. Other works using image-conditioned models can generate step-by-step sequences by iterative generation [12, 35, 36, 53]. In contrast, our method generates step-by-step visual instructions all at once, attending across steps to generate the full image sequence, including the input, which results in superior quality and consistency. + +![](images/b361979f99cf6a23da1ca7dbd10cc2f564b89a7325f4e0bc83e26ffe00f400bc.jpg) +Figure 2. Our automatic approach for creating the ShowHowTo dataset—a large-scale instructional dataset consisting of step-by-step instruction sequences of image-text pairs to perform diverse HowTo tasks. Examples of step-by-step textual instructions and the corresponding frames are highlighted in green. + +# 3. Building Large-Scale ShowHowTo Dataset + +Learning to generate visual instructions requires a large-scale dataset that captures the rich diversity of real-world tasks and their step-by-step execution. However, manually creating such a dataset is prohibitively expensive and time-consuming, limiting the dataset's scale and coverage. We address this challenge by introducing an automated approach that leverages the natural alignment between narrations and visual demonstrations in instructional videos from the web to mine high-quality sequences of image-text pairs. + +Using our proposed approach, we construct a large-scale dataset containing over half a million instruction sequences of image-text pairs spanning 25,026 diverse HowTo tasks. These sequences cover diverse domains including cooking (e.g., make strawberry crumb bars, prepare an avocado margarita), home improvement (e.g., stain a cabinet, create a tire garden), assembly (e.g., set up a $10 \times 10$ tent, tie a ring sling), DIY crafts (e.g., make a bracelet, make a fairy glow jar) and many more. We note that our data collection approach does not require any manual annotation, which is an important aspect to enable scaling. + +# 3.1. Automatic Dataset Collection + +Our approach takes as input a narrated instructional video for a specific task. First, it extracts a sequence of key steps in the form of concise, free-form textual instructions from the video's narration. Then, it associates each step with the corresponding keyframe in the video. The output is an ordered sequence of image-text pairs. This is a very challenging task due to the high level of noise, the possible misalignment of the narration and the visual content as well as the sheer variety of visual appearance and the spoken natural language in the input internet videos. + +To tackle these challenges, we design a four-stage approach, illustrated in Figure 2: (1) The narration of the input internet instructional video is transcribed into sentences with corresponding timestamps. (2) The transcribed narration is verified to be instructional and removed if not. (3) The key instruction steps are extracted from the tran + +script along with their approximate temporal bounds. (4) A representative frame for each instruction step is selected through cross-modal alignment. By applying this approach to videos from HowTo100M [42], we obtain 578K high-quality sequences of image-text pairs with approximately eight steps per video on average. + +Formally, we define our dataset as a collection of instruction sequences of image-text pairs. Each sequence $\{(I_i,\tau_i)\}_{i = 0}^n$ represents an ordered set of steps required to accomplish a specific task $\mathcal{T}$ . It consists of pairs of images $I_{i}$ and the corresponding natural language instructions $\tau_{i}$ , with $n$ denoting the number of steps in the sequence. Next, we describe the four stages in detail. + +Speech transcription. Accurate transcription (ASR) of spoken narrations is a key strength of our approach, as these transcripts capture the instructor's step-by-step guidance that we use to align with the video. HowTo100M provides 1.2 million web instructional videos, but we forego the original transcripts generated using the YouTube API due to noise [26, 37]. Instead, we use WhisperX [6], a state-of-the-art speech recognition model, to obtain high-quality transcriptions with accurate timestamps from videos. We provide comparisons of the original transcripts and the improved ones in the supplementary material [1]. + +Filtering of irrelevant videos. We find that many How-To100M videos are non-instructional, containing product reviews, vlogs, movie clips, etc. This noise may stem from the original data collection process, which relied on keyword-based web crawling and is susceptible to false positives due to inaccurate metadata. We leverage video transcripts as a strong signal for identifying instructional content and use a recent LLM (Llama 3.1 [20]) to filter the videos. We verify the reliability of this process through the evaluation on a labeled subset. Detailed analysis, qualitative results, and the prompts used for querying the LLM are provided in the supplementary material [1]. + +Step extraction. We observe that, in instructional videos, the key steps necessary to achieve a particular task are very + +often mentioned in the narration, even if they are not well-aligned with what is shown in the frame [26]. Building on this, we prompt an LLM to extract the instructional steps from the narration transcripts in the format of WikiHow step-by-step guides, providing exemplars in the prompt to guide the extraction. Somewhat surprisingly, the LLM not only correctly extracts the key steps from the transcripts, but the model is also able to associate each step with the correct temporal intervals from the transcript, even if the step spans over multiple narrations. See Figure 2 for an example, and the supplementary material [1] for additional details and the prompt used. + +Cross-modal frame alignment. For each instructional step, our goal is to identify a single representative frame that best demonstrates the instruction visually. While contrastive models [22, 47, 66] can align text instructions with frames, we observe that naive text matching across thousands of video frames leads to noisy results. Therefore, we limit the alignment to the identified instruction step temporal interval, expanded by $\epsilon = 15$ seconds to allow for some level of misalignment between the narrations and visual demonstrations [26]. Given these expanded intervals, we compute text-frame similarity scores using DFN-CLIP [22] and select the best alignment that satisfies the temporal ordering of the steps. We provide more details about the matching process in the supplementary material [1]. + +# 3.2. Dataset Statistics + +In total, after filtering, the dataset contains 578K unique sequences of image-text pairs, with a total of 4.5M steps, averaging 7.7 (± 2.8) steps per sequence, and 11.4 (± 4.7) words per step. The sequence lengths vary from 1 to 26 steps, with $97.6\%$ of sequences being 2 to 16 steps long. From task information provided by HowTo100M, the dataset contains instructions for 25K HowTo tasks across several categories, such as cooking, home and garden improvement, vehicles, personal care, health, and more. We include a comparative table to other datasets and further dataset analysis in the supplementary material [1]. + +# 4. ShowHowTo Model and Training Procedure + +Given a user-provided image $I_0$ , such as a photo of ingredients or tools on a table, our goal is to generate a sequence of images $\{\hat{I}_i\}_{i=1}^n$ of any length $n$ based on the number of required steps, that guides the user to achieve an intended task $\mathcal{T}$ , such as a cooked chicken tikka masala dish. Our goal is to generate images $\hat{I}_i$ to match the user-provided context, i.e., to be grounded in the user's environment by utilizing the specific objects, tools, and workspace from the input image $I_0$ . We achieve this goal by training a diffusion model conditioned on the input image $I_0$ along with the step-by-step textual instructions $\{\tau_i\}_{i=0}^n$ that fulfill the + +![](images/d2f079e9f60759a3a8f4a9e7008e7d6a254cd396ff082b90f43df39c6d0753f9.jpg) +Figure 3. Model architecture. Given an input frame $I_0$ (left) and a variable number of text instructions $\tau_i$ describing each step, our diffusion model generates visual instructions $\hat{I}_i$ that correctly follow the prompts $\tau_i$ and are consistent with the input image $I_0$ . + +intended task $\mathcal{T}$ in any number of steps $1\leq n\leq 15$ + +We build on recent progress in diffusion models for video generation [56]. However, there are the technical challenges of (a) how to inject the multi-step instruction guidance and (b) how to generate variable length sequences. We address these challenges in the next paragraphs. + +Architecture. Our model, shown in Figure 3, is based on a latent video diffusion model [56] composed of a U-Net encoder and decoder, each with interleaving spatial and temporal attention layers. The input image $I_0$ is projected into the latent space via the VAE encoder $\mathcal{E}$ , and is concatenated to each frame of the random noise $z_i$ to form the model's input. The U-Net progressively denoises the input latent sequence while attending to all images in the sequence to ensure the generated images are temporally consistent and aligned with the input image. For better conditioning on the input image, the U-Net also contains cross-attention layers that attend to a feature representation of the input image directly. To guide the generation process to the desired visual instruction sequence, each frame $i$ in the sequence attends to its prompt $\tau_i$ via cross-attention layers of the U-Net. In the ablations, we show that separate text conditioning for each frame in the sequence is instrumental for generating high-quality step-by-step visual instructions. + +Training. We initialize the model from the pretrained checkpoint [56] trained on WebVid10M [5] for image animation and fine-tune the entire U-Net weights on our dataset. In contrast to training on videos, where the output video is commonly of a fixed length (e.g., 16 frames in the case of [56]), step-by-step instructions have a variable se + +quence length. To ensure our model can generate variable-length sequences, we vary the sequence length during training by sampling from our sequences. For efficient computation, the length is varied over different batches but is kept the same across all samples in a single batch. If the dataset sequence is longer than the desired length $k$ , we randomly sample the starting frame and use the next consecutive $k$ frames as the model's target. We verify and further discuss these choices in Section 5.3 and provide the implementation details in the supplementary material [1]. + +# 5. Experiments + +We first introduce our evaluation setup in Section 5.1. In Section 5.2, we compare our method to current state-of-the-art quantitatively on the test set and through a user study. Section 5.3 analyzes key designs of our method through ablation studies. Finally, Section 5.4, showcases qualitative results of our method. For implementation details and additional results, see the supplementary material [1]. + +# 5.1. Evaluation details + +Dataset. We construct train and test splits from our dataset of 578K samples. Our test set comprises 3,964 sequences from 200 tasks covering the distribution over task categories in the full dataset. To ensure sample quality, we prioritize samples with high DFN-CLIP alignment scores as measured in our dataset creation pipeline (Section 3). For zero-shot evaluation of our method, we also use a random subset of 1442 non-illustrated instructional sequences1 from the WikiHow-VGSI dataset [62] as an additional test set. + +Evaluation metrics. We evaluate our model by measuring the correctness and consistency of the generated visual instructions using similar metrics as Menon et al. [41]. We describe the used metrics next with full details available in the supplementary material [1]. (1) Step Faithfulness [41] measures whether each generated image $\hat{I}_i$ correctly depicts its corresponding text instruction $\tau_{i}$ . It is computed as the zero-shot accuracy of the DFN-CLIP model where the generated image $\hat{I}_i$ is classified into classes $\{\tau_i\}_{i = 0}^n$ of all text instructions of the sequence. (2) Scene Consistency measures whether the generated image $\hat{I}_i$ consistently captures the scene from the input image $I_0$ (e.g., the same utensils are used on the same kitchen countertop as in the input image). Intuitively, a generated image $\hat{I}_i$ is considered scene-consistent if it visually matches any frame from its source video $\{I_i\}_{i = 1}^n$ (excluding the input image to avoid trivial copy solution). Therefore, for each generated image $\hat{I}_i$ , the most similar image according to the DINOv2 [44] score is retrieved from the test set images. The metric then measures if the retrieved image is from the same video as the input. + +
MethodShowHowToWikiHow [62]
Step Faithf.Scene Consist.Task Faithf.Step Faithf.Scene Consist.
(a) InstructPix2Pix [13]0.250.170.250.320.12
(b) AURORA [35]0.250.330.240.330.15
(c) GenHowTo [53]0.490.130.270.600.06
(d) Phung et al. [45]0.360.030.380.460.04
(e) StackedDiffusion [41]0.430.020.420.570.07
(f) ShowHowTo0.520.340.420.720.12
(g) Random0.190.000.010.260.00
(h) Stable Diffusion [48]†0.700.030.440.840.03
(i) Copy of the input image0.190.620.390.260.26
(j) Source sequences0.501.000.560.601.00
+ +Generation not conditioned on the input image. + +Table 1. Comparison with state-of-the-art on the ShowHowTo and the WikiHow datasets. Out of all the visual instruction generation methods, our method best follows the input prompts while being consistent with the input image. + +(3) Task Faithfulness measures how well the generated sequence $\{\hat{I}_i\}_{i=1}^n$ represents its intended task. It is measured as the zero-shot accuracy of the DFN-CLIP model where the generated sequence's averaged feature vector is classified into all 200 test set tasks. In contrast to Menon et al. [41], the generated sequences are evaluated holistically rather than per-step, as often steps are not unique to a task (e.g., "knead dough" step appears in both "Make sourdough bread" and "Make pizza" tasks), and the classification is done into all test set tasks rather than a small random subset, providing a more robust evaluation. + +# 5.2. Comparison with the State-of-the-Art + +Compared methods. We compare ShowHowTo to state-of-the-art methods for visual instruction generation as well as various baselines. For image-to-image methods (a-c), we generate the visual instructions sequence by iteratively using the last generated image as the input for the next step generation to achieve temporal consistency. (a) InstructPix2Pix [13] is trained to manipulate input images according to a text prompt by training on synthetic paired image data. In contrast, (b) AURORA [35] is trained on a manually curated dataset of image pairs from videos, while (c) GenHowTo [53] extracts the image pairs for training from instructional videos automatically. + +Methods that generate image sequences (d-e) do not accept an input image, therefore, we apply the common input masking approach, where the first denoised frame of the sequence is replaced by the noised ground truth frame in each step of the generation. The (d) Phung et al. [45] method generates consistent sequences of visual instructions by attending to all frames in the sequence in the spatial attention layers. As it is a training-free method, we reimplement it and use it with the Stable Diffusion backbone [48]. (e) StackedDiffusion [41] is trained on WikiHow illustrated image sequences. It generates the image sequence as a single tiled image. Similarly to the related work, we evaluate + +
Step win rateScene win rateTask win rate
97%3%82%18%90%10%
92%8%68%32%96%4%
86%14%77%23%85%15%
84%16%91%9%78%22%
63%37%84%16%65%35%
42%58%42%58%33%67%
+ +Figure 4. User study results. Win rates of the ShowHowTo method against baselines from pairwise forced decision user evaluations, divided into step, scene, and task. Values larger than $50\%$ indicate ShowHowTo is preferred over the other methods (right). + +all methods in zero-shot setup without finetuning. + +Lastly, we show (g) Random lower bound and various naive baselines (h-i). (h) Stable Diffusion [48] is a text-conditioned generative method with no input image conditioning that generates each image independently, the (i) Copy baseline uses the input image as the output for any prompt. As an upper limit, (j) Source sequences use the original dataset frames corresponding to the text prompts. + +Quantitative results. We show the results of different methods on our ShowHowTo test set as well as on a subset of WikiHow sequences [62] in Table 1. The methods trained to perform localized edits (a-b) generate outputs fairly consistent with the input image (see the Scene Consistency metric), yet they fail to properly capture the instructions described by the text prompts (evidenced by the Step Faithfulness metric). On the other hand, methods for generating sequences of visual instructions (d-e) model the instructions well according to the input text prompts, but they perform poorly in scene consistency. In contrast, our method (f) generates visual instructions that are consistent with the input scene and correctly capture the action specified by the prompt. Our method even generates images that are more faithful to the input textual instructions than the dataset sequences (j) (see the Step Faithfulness metric). There are two reasons for this: objects in real images can appear small or occluded, impacting CLIP matching, and sometimes steps do not appear visually in the video. + +Additionally, in the supplementary material [1], we test our method in a zero-shot setup on the GenHowTo benchmark [53] and report additional quantitative metrics on the ShowHowTo dataset. + +User study. We present a user study with 9 participants evaluating sequences from 100 randomly sampled tasks from our test set. Each participant compared 50 ShowHowTo sequences with baselines using three criteria: (1) Step Faithfulness (Which sequence better follows the steps?), (2) Scene Consistency (Which sequence is more likely to come from the same video?), and (3) Task Faithfulness (Which sequence accurately depicts the instructions for the task of [task]?). As shown in Figure 4, ShowHowTo outperforms all baselines. Notably, users preferred our generations over sequences from source videos in + +
Text conditioning typeStep Faithf.Scene Consist.Task Faithf.Average
1 prompt (concatenated step prompts)0.210.290.380.29
1 prompt (summarized step prompts)0.200.300.400.30
1 prompt per step (\( \tau_0 = 'an image' \))0.510.300.420.41
1 prompt per step (ShowHowTo)0.520.340.420.43
+ +Table 2. Ablation of step conditioning. The per-frame conditioning of ShowHowTo is instrumental in generating visual instructions faithful to the textual instructions. + +
Model training dataStep Faithf.Scene Consist.Task Faithf.Average
WikiHow-VGSI [62]0.550.120.300.32
HowToStep [37]0.390.330.290.34
ShowHowTo (food videos only)0.510.320.370.40
ShowHowTo0.520.340.420.43
+ +Table 3. Ablation of the training data as measured on the ShowHowTo test set. Our training dataset yields significant improvement over the manually curated WikiHow as well as the closely related HowToStep due to the quality of our instructions. + +$42\%$ of cases for both step and scene metrics, which may be attributed to instructional videos not showing good views of steps at times and the high quality of our generation. Lower task faithfulness scores against the source sequences suggest room for improvement in future methods. More details are in the supplementary material [1]. + +# 5.3. Ablations + +We evaluate the key design decisions of our proposed method, i.e., the model conditioning, training data, and variable sequence length training, in the next paragraphs. Furthermore, additional performance analysis of the trained model is available the supplementary material [1]. + +Text model conditioning. We evaluate how different types of text conditioning affect model performance. For video models, it is common to provide a single text prompt for conditioning. However, visual instructions vary substantially from one another, possibly requiring different conditioning. We construct a single prompt for each sequence by concatenating all step prompts and by summarizing the step prompts using an LLM [20]. In Table 2, we show that our choice of separate prompt per step significantly outperforms both of the single prompt variants. Additionally, we demonstrate that using the free-form step description for $\tau_0$ outperforms the fixed prompt 'an image'. + +Training data. In Table 3, we analyze the impact of different training datasets by comparing our dataset with two related instructional datasets of similar scale and task coverage: (i) HowToStep [37], which contains automatically extracted video-text sequences from cooking videos, and (ii) WikiHow-VGSI [62], which consists of manually created image-text sequences from WikiHow articles, where the images primarily consist of digitally drawn illustrations. To train on the HowToStep dataset, we select the mid + +
Training sequence lengthStep Faithf.Scene Consist.Task Faithf.Average
≤ 4 steps0.470.390.400.42
≤ 8 steps (ShowHowTo)0.520.340.420.43
≤ 8 steps, randomly sampled0.510.320.410.41
= 8 steps0.560.260.420.41
≤ 16 steps0.570.260.420.41
+ +Table 4. Ablation of the training sequence length as measured on the ShowHowTo test set. We compare the performance of our model when trained on different sequence lengths. + +dle frame of each video segment as the visual instruction frame. We observe significantly worse performance caused both by the lack of precise instruction frame information as well as very noisy video segments (e.g., the dataset contains 'Thank you for watching!' segments which are not instructional). The performance is also worse when compared to the model trained only on the food-related ShowHowTo sequences that are extracted from the very same videos as the HowToStep sequences, indicating a superiority of our sequence extraction process. Training on WikiHow-VGSI yields higher Step Faithfulness score, likely due to the quality of manually matched images and prompts. However, the overall performance remains significantly below our approach, as the model primarily learns from illustrated images, resulting in less consistent scene generation. + +Variable sequence length. Instructions are often of variable length, therefore, one of the key model requirements is to support generating image sequences of different lengths. While attention-based architectures allow for any sequence length, the question is how to train such a model. We test training the model on variable sequences of up to 4 frames, 8 frames, and 16 frames. We also train the model on sequences of length 8 only. In Table 4, we show that training on shorter sequences up to 4 frames results in high Scene Consistency but low Step Correctness, while training on sequences up to 16 frames is the opposite. This can be attributed to the fact that short sequences often keep the same background across the whole sequence, encouraging the model to preserve the background at the expense of the prompt. Longer sequences, on the other hand, have more variation of the background, e.g., as the task moves from the counter to the hob. The model is thus less likely to enforce the background during inference. Lastly, we also show that it is important to train always on consecutive sequences of visual instructions. If a subset of visual instructions from a video is sampled randomly (with the temporal ordering preserved), the scene consistency is decreased (Table 4, row 3). + +# 5.4. Qualitative Results + +Qualitative results in Figure 1, Figure 6, and additional figures in the supplementary material [1] demonstrate the key strengths of the ShowHowTo model. It consistently preserves the scene as well as various objects, tools, and in + +![](images/bdcd8ef3d804bf9984eb650865a3fe02ce86d2638a4694b99f5c14e140535e19.jpg) +Figure 5. Qualitative comparison using the input image (left) and the textual instructions (top) for the task of making a calzone. The images from the source video are shown in the first row. Except for ShowHowTo, methods either struggle to preserve the input scene or to produce coherent steps. + +greddents from the user-provided input image (e.g., pot and vegetables in Figure 6, first row). It dynamically adjusts the viewpoint to emphasize key actions. Similarly to vanilla text-to-image and text-to-video models, our model can also introduce plausible task-relevant objects (e.g., knives or bowls) if these objects are not present in the user-provided input image. Notably, the model effectively adapts human poses to demonstrate various object manipulations (e.g., flower arranging in Figure 6, third row). + +Figure 5 shows qualitative comparisons with related methods. Methods for generating instructional sequences (StackedDiffusion [41] and Phung et al. [45]) fail to preserve the scene from the input image. For example, they + +![](images/418b3c74377b09693d252f3d17f59ab7da6288ae0201a18930877f84903591bf.jpg) +Figure 6. Qualitative results of our method for sequences from the test set. Given the input image (left) and the textual instructions (top), ShowHowTo generates step-by-step visual instructions while maintaining objects from the input image (e.g., the cooking pot and the ceramic bowl in rows one and four) as well as among generated images (e.g., glass bowl in the second row). + +generate blue, black, or wooden kitchen countertop instead of the glossy white one from the input image. On the other hand, image-to-image approaches (GenHowTo [53], AURORA [35], and InstructPix2Pix [13]) perform minimal edits and propagate errors through the output image sequence due to iterative generation (e.g., the persistent floating dough generated by the GenHowTo method). + +Limitations. While our method can generate complex step-by-step visual instructions conditioned on the input image, it inherits the limitations of the models it is based on and introduces new limitations stemming from the novel source of training data. ShowHowTo model can struggle to maintain object states across many frames; for example, it can generate an image with raw meat after it was cooked in previous steps. Though the model often correctly generates common objects from instructional videos, for rare objects, such as electrical components, the model may generate ob + +jects in physically impossible configurations. Please see the supplementary material [1] for failure case examples. + +# 6. Conclusion + +This work explores, for the first time, generating environment-specific visual instructions to accomplish a user-defined task. We introduce a fully automated and scalable pipeline to create a dataset of 578K instructional image-text sequences from online videos, without requiring any manual supervision. Using this data, we train the ShowHowTo model to generate contextualized step-by-step visual instructions. Experiments demonstrate superior ability to generate accurate, scene-consistent instructional steps across various HowTo tasks, outperforming existing methods. We believe this work opens new avenues for personalized guidance in assistive technologies and step-by-step goal generation for robot planning. + +# Acknowledgements + +We acknowledge VSB - Technical University of Ostrava, IT4Innovations National Supercomputing Center, Czech Republic, for awarding this project access to the LUMI supercomputer, owned by the EuroHPC Joint Undertaking, hosted by CSC (Finland) and the LUMI consortium through the Ministry of Education, Youth and Sports of the Czech Republic through the eINFRA CZ (grant ID: 90254). + +Research at the University of Bristol is supported by EPSRC UMPIRE (EP/T004991/1) and EPSRC PG Visual AI (EP/T028572/1). Prajwal Gatti is partially funded by an uncharitable donation from Adobe Research to the University of Bristol. + +This research was co-funded by the European Union (ERC FRONTIER, No. 101097822 and ELIAS No. 101120237) and received the support of the EXA4MIND project, funded by the European Union's Horizon Europe Research and Innovation Programme, under Grant Agreement $\mathrm{N}^{\circ}$ 101092944. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Commission. Neither the European Union nor the granting authority can be held responsible for them. + +# References + +[1] Supplementary material (appendix) for the paper. https://arxiv.org/abs/2412.01987.3,4,5,6,7,8 +[2] 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. 2 +[3] Triantafyllos Afouras, Effrosyni Mavroudi, Tushar Nagarajan, Huiyu Wang, and Lorenzo Torresani. Ht-step: Aligning instructional articles with how-to videos. NeurIPS, 2024. 2 +[4] Anthropic. Introducing computer use, a new claude 3.5 sonnet, and claude 3.5 haiku. https://www.anthropic.com/news/3-5-models-and-computer-use, 2024.2 +[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 4 +[6] Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. Whisperx: Time-accurate speech transcription of long-form audio. INTERSPEECH, 2023. 3 +[7] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2 +[8] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation. arXiv preprint arXiv:2409.16283, 2024. 2 +[9] 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. 2 +[10] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2 +[11] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2 +[12] João Bordalo, Vasco Ramos, Rodrigo Valério, Diogo Glória-Silva, Yonatan Bitton, Michal Yarom, Idan Szpektor, and Joao Magalhaes. Generating coherent sequences of visual illustrations for real-world manual tasks. arXiv preprint arXiv:2405.10122, 2024. 2 +[13] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 5, 8 +[14] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2 +[15] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2 +[16] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocraft2: Overcoming data limitations for high-quality video diffusion models. In CVPR, 2024. 2 +[17] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In ICLR, 2023. 2 +[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 2023. 2 +[19] 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. NeurIPS, 2023. 2 +[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3, 6 +[21] Nikita Dvornik, Isma Hadji, Ran Zhang, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. Stepformer: Self-supervised step discovery and localization in instructional videos. In CVPR, 2023. 2 + +[22] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 4 +[23] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 2 +[24] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2 +[25] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2024. 2 +[26] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In CVPR, 2022. 3, 4 +[27] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 2 +[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2 +[29] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[30] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 2022. 2 +[31] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 2 +[32] Xuhui Kang and Yen-Ling Kuo. Incorporating task progress knowledge for subgoal generation in robotic manipulation through image edits. arXiv preprint arXiv:2410.11013, 2024. 2 +[33] Takeshi Kojima, Shixiang (Shane) Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In NeurIPS, 2022. 2 +[34] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2 +[35] Benno Krojer, Dheeraj Vattikonda, Luis Lara, Varun Jampani, Eva Portelance, Christopher Pal, and Siva Reddy. Learning action and reasoning-centric image editing from videos and simulations. arXiv preprint arXiv:2407.03471, 2024. 2, 5, 8 +[36] Bolin Lai, Xiaoliang Dai, Lawrence Chen, Guan Pang, James M Rehg, and Miao Liu. Lego: Learning egocentric action frame generation via visual instruction tuning. arXiv preprint arXiv:2312.03849, 2023. 2 + +[37] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhang, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. 2024. 2, 3, 6 +[38] Junbang Liang, Ruoshi Liu, Ege Ozguroglu, Sruthi Sudhakar, Achal Dave, Pavel Tokmakov, Shuran Song, and Carl Vondrick. Dreamitate: Real-world visuomotor policy learning via video generation. arXiv preprint arXiv:2406.16862, 2024. 2 +[39] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. In ICCV, 2023. 2 +[40] Kangfu Mei and Vishal Patel. Vidm: Video implicit diffusion models. In AAAI, 2023. 2 +[41] Sachit Menon, Ishan Misra, and Rohit Girdhar. Generating illustrated instructions. In CVPR, 2024. 2, 5, 7 +[42] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 3 +[43] Ashvin V Nair, Vitchyr Pong, Murtaza Dalal, Shikhar Bahl, Steven Lin, and Sergey Levine. Visual reinforcement learning with imagined goals. NeurIPS, 2018. 2 +[44] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 5 +[45] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Coherent zero-shot visual instruction generation. arXiv preprint arXiv:2406.04337, 2024. 2, 5, 7 +[46] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2 +[47] 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 ICML, 2021. 4 +[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 5, 6 +[49] RunwayML. Gen-3 alpha. 2024. 2 +[50] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. Howtocation: Prompting llms to transform video annotations at scale. In ECCV, 2024. 2 +[51] Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: Toward hierarchical understanding of procedural activities. NeurIPS, 2024. 2 +[52] Tomáš Souček, Jean-Baptiste Alayrac, Antoine Miech, Ivan Laptev, and Josef Sivic. Multi-task learning of object states and state-modifying actions from web videos. TPAMI, 2024. 2 + +[53] Tomáš Souček, Dima Damen, Michael Wray, Ivan Laptev, and Josef Sivic. Genhowto: Learning to generate actions and state transformations from instructional videos. In CVPR, 2024. 2, 5, 6, 8 +[54] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In CVPR, 2019. 2 +[55] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. NeurIPS, 2024. 2 +[56] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2, 4 +[57] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. arXiv preprint arXiv:2407.15208, 2024. 2 +[58] Zihui Xue, Kumar Ashutosh, and Kristen Grauman. Learning object state changes in videos: An open-world perspective. In CVPR, 2024. 2 +[59] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. Unloc: A unified framework for video localization tasks. In ICCV, 2023. 2 +[60] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. Entropy, 2023. 2 +[61] Sherry Yang, Yilun Du, Seyed Kamyar Seyed Ghasemipour, Jonathan Tompson, Leslie Pack Kaelbling, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. In ICLR, 2024. 2 +[62] Yue Yang, Artemis Panagopoulou, Qing Lyu, Li Zhang, Mark Yatskar, and Chris Callison-Burch. Visual goal-step inference using wikihow. In EMNLP, 2021. 2, 5, 6 +[63] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2 +[64] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In CVPR, 2023. 2 +[65] Tianhe Yu, Ted Xiao, Austin Stone, Jonathan Tompson, Anthony Brohan, Su Wang, Jaspiar Singh, Clayton Tan, Jodilyn Peralta, Brian Ichter, et al. Scaling robot learning with semantically imagined experience. arXiv preprint arXiv:2302.11550, 2023. 2 +[66] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 4 +[67] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. IJCV, 2024. 2 + +[68] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2 +[69] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2 +[70] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, 2019. 2 \ No newline at end of file diff --git a/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/images.zip b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..71d2ec130c7707482bcf4050bb4ab98b63253968 --- /dev/null +++ b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:932744848ebb2e0f6a3512890891420b4cb4619dabed796f5e89749a5a423a64 +size 812660 diff --git a/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/layout.json b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7c73f9ef022671fad9cc46ce0ede927f7bae26b9 --- /dev/null +++ b/CVPR/2025/ShowHowTo_ Generating Scene-Conditioned Step-by-Step Visual Instructions/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8429c1d374339f412c78ca44ea6e7067b5936619f4f78d074454a18e4d01e4c +size 344352 diff --git a/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/1c889638-ab1b-4a02-af5b-597eb1c7a5fe_content_list.json b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/1c889638-ab1b-4a02-af5b-597eb1c7a5fe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6fffd12c6e21c0cded02f4d465240107cfe4e1ac --- /dev/null +++ b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/1c889638-ab1b-4a02-af5b-597eb1c7a5fe_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:675cda9aa5105abfb6e727e5df93d64ef7f0cbbb9ec4eb176c0d6440d5dd3b3f +size 86512 diff --git a/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/1c889638-ab1b-4a02-af5b-597eb1c7a5fe_model.json b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/1c889638-ab1b-4a02-af5b-597eb1c7a5fe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4042be11a28033a2cadced9a1c2f7c1aba053a10 --- /dev/null +++ b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/1c889638-ab1b-4a02-af5b-597eb1c7a5fe_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1adf951dc4a8dbed980e6d0dffde46b520fceef35b0b927ee63b494515f4a4a5 +size 113008 diff --git a/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/1c889638-ab1b-4a02-af5b-597eb1c7a5fe_origin.pdf b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/1c889638-ab1b-4a02-af5b-597eb1c7a5fe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4f37103e83d619fe26e4bc1b492e20d168b922a3 --- /dev/null +++ b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/1c889638-ab1b-4a02-af5b-597eb1c7a5fe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fa1d2e16cc699220f6ee5f29a2fa651dfb0516c82e0716691aae1151aa4f4ce +size 4207312 diff --git a/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/full.md b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7e0a6b8b9013c7a69b21a6019598a707816c2270 --- /dev/null +++ b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/full.md @@ -0,0 +1,412 @@ +# ShowMak3r: Compositional TV Show Reconstruction + +Sangmin Kim Seunguk Do Jaesik Park* Seoul National University, Republic of Korea + +{sm.kim, seunguk.do, jaesik.park}@snu.ac.kr + +![](images/c1805250b2e604aea854a7bda8b6d978fde3d942230bc78bb889daba48b96101.jpg) +Figure 1. We introduce ShowMak3r, a comprehensive pipeline that reconstructs dynamic radiance fields from TV shows. Given a video clip that shows limited viewpoints and abrupt shot changes, our pipeline recovers the background stage and dynamic actors that are trackable across the shot changes. The scene is compositional and editable, so we can render it from novel viewpoints while selectively editing individual actors. Our method also recovers detailed human appearances, including facial expressions. + +# Abstract + +Reconstructing dynamic radiance fields from video clips is challenging, especially when entertainment videos like TV shows are given. Many challenges make the reconstruction difficult due to (1) actors occluding with each other and having diverse facial expressions, (2) cluttered stages, and (3) small baseline views or sudden shot changes. To address these issues, we present ShowMak3r, a comprehensive reconstruction pipeline that allows the editing of scenes like how video clips are made in a production control room. In ShowMak3r, a 3DLocator module locates recovered actors on the stage using depth prior and estimates unseen human poses via interpolation. The proposed ShotMatcher module then tracks the actors under shot changes. Furthermore, ShowMak3r introduces a face-fitting network that dynamically recovers the actors' expressions. Experiments on Sitcoms3D dataset show that our pipeline can reassemble TV show scenes with new cameras at different timestamps. We also demonstrate that ShowMak3r enables interesting ap + +plications such as synthetic shot-making, actor relocation, insertion, deletion, and pose manipulation. Project page : https://nstar1125.github.io/showmak3r + +# 1. Introduction + +Considerable advances in radiance field reconstruction approaches [22, 36] transform how we reconstruct and visualize the scenes. Recent methods aim to bring video clips into 4D space to enable novel viewpoint rendering or scene editing. However, recovering the radiance field from dynamic scenes remains a challenging problem. Approaches in this category have mainly focused on scenarios with multi-view synchronized cameras or fully observed scenes [20, 43, 65]. + +The reconstruction gets even harder for entertainment videos, such as TV shows captured by shot-changing (video transition by another camera) monocular cameras. Compared to existing benchmark datasets [17], TV show video clips present additional challenges. First, it contains scenes that are inherently hard to reconstruct, such as multiple actors interacting and occluding each other on the cluttered + +stages or actors showing detailed facial changes to express their emotions. In addition, the videos are filmed with multiple cameras and then edited to appear as a continuous timeline, resulting in sudden shot changes. Furthermore, cameras are mainly positioned in front of the scene, creating partial observations and thus limiting information about the actors' backsides. Therefore, even the state-of-the-art methods [24, 60] fail to recover consistent dynamic radiance fields due to incorrect human-scene alignment and inconsistent deformation of human movements. + +In this work, we present ShowMak3r, a comprehensive pipeline that reconstructs the dynamic radiance field from TV shows, enabling viewpoint editing like how the video clip is made in a production control room. We first build a stage from the aggregated images and recover parametric models of dynamic actors. Since the estimated humans from the video clip and the reconstructed stage have different coordinate systems, we propose the 3DLocator module, which aligns the actors to their correct locations on the stage by leveraging depth estimation while estimating unseen human poses via interpolation. After positioning the actors, we use the ShotMatcher module to perform human association at the shot boundaries to track actors across different shots. + +As the facial change of the actors is a key element in TV show videos, we implement an implicit face-fitting network to change human expressions across frames to address this dynamically. Experiments with the Sitcom3D dataset [45] show that our pipeline can successfully reconstruct from TV show videos. We also demonstrate various applications with ShowMak3r, including actor relocation, insertion, deletion, and pose manipulations. To our knowledge, ShowMak3r is the first comprehensive method for total reconstruction of the stage and multiple actors given the shot-changing videos. + +Our contributions can be summarized as below: + +- We introduce ShowMak3r, a comprehensive pipeline that reconstructs the dynamic radiance field from TV shows, enabling viewpoint editing like a production control room. +- We propose 3DLocator that accurately aligns actors on the stage. 3DLocator even aims to solve unseen poses due to occlusions. +- We present ShotMatcher that enables continuous tracking of actors under shot change. Our approach associates actors even when they are not visible in certain shots. +- We implement an implicit face-fitting network to recover and express dynamic facial expressions. +- Extensive experiments show the validity of our approach and demonstrate possible applications such as actor relocation, insertion, deletion, and pose manipulation. + +# 2. Related Work + +4D Scene Reconstruction. Attempts have been made to extend radiance fields [22, 36] into spatio-temporal models [2, 20, 28, 30, 42, 43, 49, 65] for reconstructing 4D scenes from synchronized multi-camera videos. These techniques have since evolved to handle single-view video inputs [7, 26, 29, 30, 32, 56, 57, 60, 61], increasing their applicability across various scenarios. More recently, feed-forward methods that directly generate dynamic point clouds have emerged [34, 68], eliminating the need for perscense training. However, despite these advancements, current 4D reconstruction methods still struggle with TV show content, which often features limited camera angles, rapid human motion, and abrupt scene transitions. + +3D Avatar Reconstruction from Videos. Human radiance fields [8, 11, 12, 14, 15, 17, 18, 24, 31, 37, 38, 41, 46, 50, 51, 53, 54, 58, 64, 66, 69], which incorporates radiance fields [22, 36] with parametric human models [33, 44], have demonstrated that photo-realistic 3D human avatars can be reconstructed from monocular video. These representations later evolved into generalizable humans radiance fields [5, 13, 27, 39, 63], which eliminated the need for extensive per-avatar training time. More recently, approaches to adopt diffusion models [47, 52] as additional priors [10, 16, 48, 62] for recovering unseen regions have been studied [6, 25]. When reconstructing 3D humans from TV shows, facial features are particularly crucial. Researchers have proposed utilizing pretrained expression encoders [53] or directly leveraging SMPL-X expression parameters[37] to enhance facial features. However, these approaches typically demand multi-view face images [53] or face challenges in capturing nuanced expressions from TV show videos [37]. To address these limitations, we propose a simple, yet highly effective approach of refining facial expressions through an implicit deformation network. + +Composite Human-Scene Reconstruction. Previous approaches for reconstructing scenes with humans [11, 24, 28] treat backgrounds as static and use separate representations for humans and backgrounds. However, these methods require videos to capture the entire human body, including foot contact with the floor, making them unsuitable for TV show settings. Sitcoms3D [45] addresses it by reconstructing sitcom videos through NeRF-W [35] for consistent backgrounds and optimizing SMPL parameters using adjacent shots. However, it lacks human texture and requires identical humans to appear in neighboring shots. OmniRe [3] reconstructs outdoor scenes with dynamic objects, including pedestrians and vehicles. However, it relies on LiDAR sensors for geometric data and isn't designed for scenarios with shot changes. In contrast, our method effectively positions actors within the stage without requiring multiple shots, foot contact points, or additional sensors. + +![](images/5ada3d6f88c8afdb1aaaf367e011dd6551c8aeb68f63a47f375b07f8d04d11e5.jpg) +Figure 2. Overview of our ShowMak3r pipeline. Given a TV show video clip, we perform dense reconstruction of the stage (Sec. 3.3), locate SMPL models to the stage (Sec. 3.4), associate SMPL models across shots to track actors (Sec. 3.5), and recover the detailed appearance of the actors (Sec. 3.6). 3D Gaussians of the stage and the actors are rendered to produce novel frames. + +# 3. Method + +# 3.1. Overview + +In video clips of TV shows, several actors perform on the stage. The TV shows are structured into three hierarchical level semantics: scene, shot, and frame. A scene represents a sequence of related shots that follow a continuous narrative flow [21], a shot captures continuous frames within a single camera [55], and a frame refers to an individual image within the sequence. Our work aims to reconstruct dynamic scenes from such frame hierarchy in TV shows. + +As shown in Fig. 2, Sec 3.3 presents how we reconstruct the consistent stage. Sec 3.4 introduces how our 3DLocator positions multiple actors into correct locations on the stage. Sec 3.5 explains the proposed ShotMatcher to track actors during shot-change. Sec 3.6 shows how to reconstruct dynamic actors and their expressions with our face-fitting network. + +Scene representation. Our approach represents the stage and the actors with 3DGS [22], an explicit approach that reconstructs a radiance field with 3D Gaussians. Each Gaussian consists of attributes, including center $\mu \in \mathbb{R}^3$ , rotation $\mathbf{q} \in \mathbb{R}^4$ , scale $\mathbf{s} \in \mathbb{R}^3$ , color $\mathbf{c} \in \mathbb{R}^3$ , and opacity $o \in \mathbb{R}$ . The $k$ -th Gaussian is defined as follows: + +$$ +g _ {k} (\mathbf {p}) = o _ {k} \exp^ {- \frac {1}{2} \left(\mathbf {p} - \boldsymbol {\mu} _ {k}\right) ^ {T} \boldsymbol {\Sigma} _ {k} ^ {- 1} \left(\mathbf {p} - \boldsymbol {\mu} _ {k}\right)}, \tag {1} +$$ + +where position is $\mathbf{p} \in \mathbb{R}^3$ , and the covariance matrix is $\boldsymbol{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^T \mathbf{R}_k^T$ . $\mathbf{R}_k \in SO(3)$ and $\mathbf{S}_k \in \mathbb{R}_+^3$ are ob- + +tained from quaternion of $\mathbf{q}$ and scale $\mathbf{s}$ . Unlike implicit methods [17], 3D Gaussians $\mathcal{G} = \{g_k\}_{k=1\dots K}$ have an explicit nature, which makes it effective for reconstructing the stage $\mathcal{G}^{\mathrm{stage}}$ and multiple actors $\mathcal{G}_n^{\mathrm{actor}}$ by simply compositing multiple Gaussian sets as $\mathcal{G}^{\mathrm{TVshow}} = \mathcal{G}^{\mathrm{stage}} \cup \{\mathcal{G}_n^{\mathrm{actor}}\}_{n=1\dots N}$ . + +# 3.2. Preprocessing + +Camera parameters. We recover the camera pose $\pi_f$ of each frame $f$ using SfM systems. We observe that GLOMAP [40] robustly handles panning frames by globally estimating camera poses and the 3D structure of all input images at once. To reduce the effect of transient actors, we mask input images with the binary segmentation map of actors $M_f^{actor}$ using SAM [23], and obtain $M_f^{stage}$ by inverting $M_f^{actor}$ . + +Guiding depth maps. By following the convention, we reconstruct $\mathcal{G}^{\mathrm{stage}}$ using 3DGS [22] given the camera poses $\pi$ . However, we experience vanilla 3DGS struggles with the reconstruction due to the partially observed background or narrow baselines, which are frequent in TV shows. Therefore, we utilize the monocular depth as guidance to fulfill the limited observations. After we get dense depth predictions $\{D_f^{\mathrm{mono}} \in \mathbb{R}^{H \times W} | f = 1, 2, \dots, F\}$ from a data-driven approach [1], we adjust scale $a$ and offset $b$ of each depth map to match the camera coordinate system predicted by the SfM pipeline. + +Specifically, given the SfM point clouds $\mathcal{P}_f$ visible at $f$ - + +th frame, we find $a^*$ and $b^*$ as follows: + +$$ +a ^ {*}, b ^ {*} = \arg \min _ {a, b} \sum_ {\mathbf {p} \in \mathcal {P} _ {f}} \mathcal {L} (\mathbf {p}; a, b), \tag {2} +$$ + +where the Huber loss $\mathcal{L}(p; a, b)$ with the depth of projected point $p_z$ for a view $\pi$ is defined as follows: + +$$ +\mathcal {L} (\mathbf {p}; a, b) = \left\{ \begin{array}{l l} \frac {1}{2} \left(p _ {z} - \left(a D ^ {\text {m o n o}} \left(\pi (\mathbf {p})\right) + b\right)\right) ^ {2} & \text {i f} | p _ {z} | \leq \delta_ {1}, \\ \delta_ {1} \left(| p _ {z} - \left(a D ^ {\text {m o n o}} \left(\pi (\mathbf {p})\right) + b\right) | - \frac {\delta_ {1}}{2}\right) & \text {o t h e r w i s e} \end{array} \right. \tag {3} +$$ + +where we empirically set $\delta_{1} = r_{\mathrm{stage}} / 100$ , where $r_{\mathrm{stage}}$ represents the scene radius. We then obtain the depth map aligned with the SfM coordinate system by calculating $D^{\mathrm{aligned}} = a^{*}\times D^{\mathrm{mono}} + b^{*}$ . We iterate the above process for the entire frame. + +Note that the aligned depth map $D^{\mathrm{aligned}}$ is the key component of our pipeline that boosts stage reconstruction (Sec. 3.3) and guides the positioning of the actors (Sec. 3.4). + +# 3.3. 3D Stage Reconstruction + +We reconstruct dense Gaussians of the stage $\mathcal{G}^{\mathrm{stage}}$ using 3DGS [22] with the extended loss that leverages the aligned depth maps $D^{\mathrm{aligned}}$ obtained from Sec. 3.2. We observe that using depth guidance provides a denser and more reliable reconstruction of the stage compared to using only photometric loss. + +Gathering background images. Reconstructing a dense and complete stage from a single video clip is challenging for TV show scenarios since only a part of the stage is shown during the video. Interestingly, TV shows like sitcom videos depict a similar environment over the season, so shots in various episodes show diverse views of the stages, which we can utilize for the stage reconstruction. Therefore, we utilize additional background images gathered from different episodes of a sitcom [45]. + +Depth-guided dense reconstruction. When $\mathcal{G}^{\mathrm{stage}}$ being optimized, we can get the rendered depth map $D^{\mathrm{render}}$ at frame $f$ by utilizing the Gaussian rasterization as follows: + +$$ +D ^ {\text {r e n d e r}} = \sum_ {k = 1} ^ {K} d _ {k} \alpha_ {k} \prod_ {k ^ {\prime} = 1} ^ {k - 1} \left(1 - \alpha_ {k ^ {\prime}}\right) \tag {4} +$$ + +where $d_{k}$ denotes the z-depth, and $\alpha_{k}$ is the blending coefficient of the $k$ -th Gaussian in view space. + +Given $D^{\mathrm{render}}$ , we incorporate depth guidance in the log-L1 form [59] for better convergence of 3DGS as follows: + +$$ +\mathcal {L} _ {\text {d e p t h}} = \frac {\log \left(1 + \left| M ^ {\text {s t a g e}} D ^ {\text {r e n d e r}} - M ^ {\text {s t a g e}} D ^ {\text {a l i g n e d}} \right|\right)}{\left| M ^ {\text {s t a g e}} \right|}, \tag {5} +$$ + +where, $M^{stage}$ is background mask obtained from Sec. 3.2. We also add total variation loss for fostering the smoothness + +![](images/360355d3633c6b8e9c28409bbf65424a1c822e57d0eae92a98781b9c9769b50d.jpg) +(a) w/o object removal +Figure 3. An example of transient object removal. + +![](images/c5e56ea16005bf0a441d698082cacfd2207c7f62f856c552ae241add7cb552ce.jpg) +(b) w/ object removal + +[4, 59] of rendered depth $D^{\text{render}}$ as follows: + +$$ +\mathcal {L} _ {\mathrm {T V}} = \frac {1}{| D ^ {\text {r e n d e r}} |} \sum_ {\mathbf {u}} \left| \frac {\partial D ^ {\text {r e n d e r}}}{\partial \mathbf {u}} \right| _ {1} \tag {6} +$$ + +As a result, our extended loss in addition to vanilla 3DGS losses $(\mathcal{L}_{color}$ and $\mathcal{L}_{\mathrm{D - SSIM}})$ is defined as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {b a c k g r o u n d}} = \left(1 - \lambda_ {\mathrm {D} - \mathrm {S S I M}}\right) \mathcal {L} _ {\text {c o l o r}} + \lambda_ {\mathrm {D} - \mathrm {S S I M}} \mathcal {L} _ {\mathrm {D} - \mathrm {S S I M}} \tag {7} \\ + \lambda_ {\text {d e p t h}} \mathcal {L} _ {\text {d e p t h}} + \lambda_ {\text {s m o o t h}} \mathcal {L} _ {\mathrm {T V}}, \\ \end{array} +$$ + +where we empirically set $\lambda_{\mathrm{D - SSIM}} = 0.2$ , $\lambda_{\mathrm{depth}} = 0.2$ and $\lambda_{\mathrm{smooth}} = 0.5$ . We masked actors appearing at depth maps and input images using $M^{\mathrm{stage}}$ for stability when optimizing Eq. (7). An example of $\mathcal{G}^{\mathrm{stage}}$ is shown in Fig. 2. + +Object removal. Since the gathered images have transient objects in the scene, they interfere with the reconstruction process. In particular, objects with no reference frames are hard to reconstruct or delete, which leads to floaters remaining in the background. These artifacts degrade the background quality by a large margin. To mitigate this issue, we annotate these regions and apply image inpainting [47]. Since inpainted areas can be noisy, we apply only the depth loss for robustness. We can successfully recover these regions, as shown in Fig. 3. + +# 3.4. Locating Actors on the Stage + +Estimating human poses from a 2D image is a well-developed problem. We estimate the shape parameter $\beta \in \mathbb{R}^{10}$ and pose parameter $\theta \in \mathbb{R}^{24\times 3\times 3}$ of a SMPL model using the off-the-shelf approach [9]. However, when it comes to locating the humans in the designated 3D coordinate system, it becomes a nontrivial problem. + +Previous approaches proposed optimizing the human scale and pose with two adjacent shots [45] or identifying the foot intersection point between a SMPL model and the ground plane [11, 17, 24]. However, these methods are inadequate for TV show scenarios since consecutive shots do not always feature the same individuals, and actors frequently have their feet out of the frame. + +3DLocator. We propose a 3DLocator module that positions the posed humans to the reconstructed 3D TV stage + +![](images/5bbcae3c12be121a32d8575523240a2af1808160f1f20f7b43c5c91707d9800f.jpg) +Figure 4. Results of actor association. ShotMatcher can associate actors even when some individuals do not appear in a shot. If the distance of the matched actors is above the matching threshold, ShotMatcher identifies them as different. + +using a single clip. Specifically, given $i$ -th SMPL vertex $\mathbf{c}_i$ in canonical space, we can apply arbitrary human pose as follows: + +$$ +\mathbf {v} _ {i} = \sum_ {j = 1} ^ {J} w _ {i, j} \left(\mathbf {R} _ {j} \mathbf {c} _ {i} + \mathbf {t} _ {j}\right), \tag {8} +$$ + +where $w_{i,j} \in \mathbb{R}$ is the Linear Blend Skinning(LBS) weights of the $j$ -th joint and $i$ -th vertex, and $\{\mathbf{R}_j, \mathbf{t}_j\}$ are the rotation and translation of $j$ -th joint determined by predicted SMPL pose parameter $\{\beta, \theta\}$ . Then, posed SMPL vertices $\mathbf{v}_i$ can be mapped to the stage coordinate as follows: + +$$ +\mathbf {v} _ {i} ^ {\prime} (s, \mathbf {t}) = s \mathbf {v} _ {i} + \mathbf {t}. \tag {9} +$$ + +3DLocator finds the optimal scale $s \in \mathbb{R}$ and global translation parameters $\mathbf{t} \in \mathbb{R}^3$ of the posed SMPL. + +The core idea of our 3DLocator is to align posed SMPL vertices $v_{i}^{\prime}$ to the aligned mono-depth $D_{f}^{\mathrm{aligned}}$ (Sec. 3.2). Note that this scheme does not assume the same actor between consecutive shots [45], and does not involve the ground plane assumption [17, 24]. + +For the alignment, we only consider SMPL's visible points $\mathbf{v}' \in \mathcal{V}_f$ at frame $f$ since the human regions of the $D_f^{\mathrm{aligned}}$ indicate the depth of the visible parts. We use Huber loss from Eq. (3) between the z-value of the visible vertices in the camera space and $D_f^{\mathrm{aligned}}$ . + +$$ +\mathcal {L} (\mathbf {v}, \delta_ {2}; s, \mathbf {t}) = \operatorname {H u b e r} \left(D ^ {\text {a l i g n e d}} \left(v _ {x} ^ {\prime}, v _ {y} ^ {\prime}\right), v _ {z} ^ {\prime}\right) \tag {10} +$$ + +where we empirically set $\delta_{2} = r^{\mathrm{stage}} / 20$ . We use the optimal $s^{*}$ and $\mathbf{t}^{*}$ to place SMPL model into frame $f$ . + +# 3.5. Tracking Actors across the Shots + +ShotMatcher. 3DLocator module (Sec. 3.4) places SMPL actors on the stage for every frame. Using the modern approach [9], the SMPLs are associated across frames. However, it is still necessary to associate the SMPL model with + +the different shots to avoid creating multiple actors. To address this issue, we propose a ShotMatcher module to associate actors between shot boundaries. + +ShotMatcher calculates the pairwise Euclidean distances between the actors' 3D coordinates in the last frame of a certain shot and the first frame of the consecutive shot. Among all possible actor-to-actor pairs, ShotMatcher chooses the actor pair with the smallest Euclidean distance that falls below a matching threshold. As shown in Fig. 4, the matching threshold is required to exclude the pair with a far distance. Details are described in the supplement. + +Pose interpolation and extrapolation. TV show videos frequently present various occlusion scenarios, such as one actor blocking another, objects obscuring actors, or actors temporarily moving out of the frame. These occlusions interfere with estimating accurate human pose. To address this issue, we analyze SMPL tracking results, identify two stable frames adjacent to the occluded frames, and perform linear interpolation of the actors' SMPL parameters. We subsequently apply a bidirectional low-pass filter to the interpolated SMPL data to ensure smooth and natural motion. + +In TV show videos, one of the main challenges of actor association is the absence of certain actors in some shots. For example, while a wide shot may include all actors on stage, a close-up shot might only feature one or two individuals. To address this, any unmatched actors are filled into the subsequent shots by extrapolating their SMPL parameters from the data at the shot boundary. + +# 3.6.3D Actor Reconstruction + +As the final step in our pipeline, we introduce our human reconstruction module that makes $\{\mathcal{G}_n^{\mathrm{actor}}\}_{n = 1\dots N}$ using 3DGS, given $N$ SMPL models associated with different shots. We first initialize Gaussian centers of $\mathcal{G}_n^{\mathrm{actor}}$ using each of the SMPL model's vertices located in the TV stage coordinate (Sec. 3.5). Then, we optimize $\mathcal{G}_n^{\mathrm{actor}}$ using the 3DGS loss extended with SDS loss $\mathcal{L}_{\mathrm{SDS}}$ [25]. + +Photometric loss. Since the stage can occlude the actors, we estimate foreground masks from the stage structure by comparing the depth between the rasterized background Gaussians $D_{\mathrm{stage}}^{\mathrm{render}}$ and the rasterized human Gaussians $D_{\mathrm{actor}}^{\mathrm{render}}$ using Eq. (4). If a rendered depth pixel of the actor is smaller than the stage, such pixel is marked as a part of the foreground as follows: + +$$ +M ^ {\text {f o r e g d}} (\mathbf {u}) = \left\{ \begin{array}{l l} 1, & \text {i f} D _ {\text {s t a g e}} ^ {\text {r e n d e r}} (\mathbf {u}) > \left\{D _ {\text {a c t o r}, n} ^ {\text {r e n d e r}} (\mathbf {u}) \right\} _ {n = 1 \dots N} \\ 0, & \text {o t h e r w i s e ,} \end{array} \right. \tag {11} +$$ + +where $N$ denotes the number of humans. + +We then apply the foreground masks to both the rendered actors $P = M^{\mathrm{foregd}} \odot I_{\mathrm{actor}}^{\mathrm{render}}$ and input video frame $Q = M^{\mathrm{foregd}} \odot I$ , and calculate actor loss $\mathcal{L}_{\mathrm{actor}}$ using the + +![](images/15a7b19468659fb29e3ecc1b4980b916078ea3a61924402edef048905dee1981.jpg) +(a) $\mathrm{W / o}M^{f o r e g d}$ + +![](images/2a50be3612f5456c1d7fe8d85a2492fbe14e1fa4a9d5b6ffd02a6ca28df5fcf3.jpg) +(b) w/ $M^{foreg d}$ + +![](images/6c95ec6bd35f6ac5cad7cce90fef01cac7646b3d1135b65f563d76e26a60f807.jpg) +(c) Ground Truth +Figure 5. An effect of using the proposed foreground masking. + +two masked images $\{P, Q\}$ for the vanilla 3DGS loss [22], where $\mathcal{L}_{\mathrm{actor}}$ is calculated for every visible frame of $n$ -th actor. The effect of using a foreground mask is shown in Fig. 5. + +Unobserved area. Since TV show cameras usually capture the actors from the front, the rest of the part remains unobserved during the entire video clip. Inspired by [25], we use SDS loss [48] with textual inversion to hallucinate the unseen parts of the actors. We compute SDS loss for each Gaussian sets $\{\mathcal{G}_n^{\mathrm{actor}}\}_{n = 1\dots N}$ with person-specific diffusion $\phi_{n}$ . The final loss of 3DGS for an individual actor is defined as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {a c t o r}} \sum_ {f} ^ {F} \mathcal {L} _ {\text {a c t o r}, f} + \lambda_ {\text {S D S}} \mathcal {L} _ {\text {S D S}} \left(\mathcal {G} _ {n} ^ {\text {a c t o r}}; \phi_ {n}\right) \tag {12} +$$ + +Refinement. In TV shows, recovering detailed facial expressions are essential for delivering emotion and enhancing realism. We introduce an implicit function-based residual appearance fitting scheme using a Gaussian deformation network [20]. In our scenario, we observe that changing the positions of Gaussians on the face is unsuitable since facial motions are subtle. Instead of moving the Gaussians, we refine colors and opacity to fit the details. Given time $t$ and the position $\mu$ of Gaussians, the color and opacity head of the face-fitting network return each residual value as follows, + +$$ +\Delta \mathbf {c} (\boldsymbol {\mu}, t) = F _ {\text {c o l o r}} \left(F _ {\theta} \left(\operatorname {c o n c a t} \{\gamma (\boldsymbol {\mu}), \gamma (t) \}\right)\right) \tag {13} +$$ + +$$ +\Delta o (\boldsymbol {\mu}, t) = F _ {\text {o p a c i t y}} \left(F _ {\theta} \left(\operatorname {c o n c a t} \{\gamma (\boldsymbol {\mu}), \gamma (t) \}\right)\right) \tag {14} +$$ + +where $\Delta c$ and $\Delta o$ denotes color and opacity residual, $\{F_{\theta}, F_{\mathrm{color}}, F_{\mathrm{opacity}}\}$ indicates learnable MLPs for deformation, and $\gamma(\cdot)$ is positional encoding. + +We first train actor Gaussians without the refinement module for 2,000 iterations to reconstruct coarse actors. + +Then, for the rest of the iterations, we add the color and the opacity residuals to the actor Gaussians obtained from Eq. (12). For $n$ -th actor, we can update it as follows: + +$$ +\mathcal {G} _ {n, t} ^ {\text {a c t o r}} = \left\{g _ {k, t} \left(\boldsymbol {\mu}, \mathbf {s}, \mathbf {q}, \mathbf {c} + \Delta \mathbf {c}, o + \Delta o\right) \right\} _ {k = 1 \dots K} \tag {15} +$$ + +# 4. Experiments + +# 4.1. Experiment Setups + +We evaluate our pipeline on Sitcoms3D [45] dataset. Sitcoms3D dataset consists of 100-200 images per location from several sitcoms, such as 'The Big Bang Theory (2007)', 'Friends (1994)', 'Two and a Half Men (2003)', 'Everybody Loves Raymond (1996)'. These images are sampled across multiple episodes to capture the total structure of the stage. + +For qualitative comparison, we compare our pipeline with three different types of dynamic scene reconstruction approaches on four sitcoms of the Sitcoms3D dataset [45]. We select HUGS [24], Shape-of-Motion [60], and MonST3R [68] as baselines, each representing template-based, template-free, and feed-forward methods. We also compare the stage reconstruction quality with other 3D reconstruction methods such as Sitcoms3D(NeRF-W [35]), 3DGS [22], GS-W [67], and FSGS [70]. We calculate PSNR, SSIM, and LPIPS metrics with masked human areas excluded for a fair comparison. Lastly, to evaluate the effectiveness of our face-fitting network, we compare our method with ExAvatar [37], which also addresses similar challenges of facial expression reconstruction. + +# 4.2. Qualitative Comparison + +We compare our method with HUGS [24] (template-based), Shape-of-Motion [60] (template-free), and MonST3R [68] (feed-forward). We compare the viewpoints that depict the stage from the video clip for a fair comparison since other baselines target reconstruction from a monocular video. + +Single person scenario. We compare HUGS, which targets single-person scenarios, with the videos featuring a single individual. HUGS utilizes an SfM system for camera pose estimation, so it can not reconstruct the scene from a monocular video clip without sufficient view changes. Therefore, we provide additional background images to HUGS pipeline for camera pose prediction. Although our approach also utilizes SfM for the pose estimation, our depth guidance-based approach helps reconstruct the 3DGS scenes even with minimal camera movement. As shown in Fig. 6, HUGS suffers from properly locating actors in the TV show videos, while ShowMak3r can reconstruct both actor and stage effectively. Note that HUGS has an assumption that the foot is visible and touches the ground, while ShowMak3r does not. + +![](images/0ed2f6daabc9128deb28cc1e81903275e38bae7d348adf21938e2b85dd8c6445.jpg) +# + +![](images/6782082b5879614afbe2c72a70f12ba95834620ea0dbb202c04dad4bd9a99514.jpg) +# + +![](images/47fe43c0632b6cc2f4b0c1805190e796856d16e12689eb604cc0f229b16dfba2.jpg) +# + +![](images/d1466683fff6b359d228e1a101604233c17f060dfa8b71feb499a08fb27628e8.jpg) +(a) HUGS + +![](images/96b5b65e41615d60d35f6fff663633443a932765dd31cf1b958ac105859614ea.jpg) +(b) ShowMak3r (Ours) + +![](images/602e97c811b138cfcedd59b81955ab0ff4c297bed53f403dc41ffe5405a40f11.jpg) +(c) Ground Truth + +![](images/e7151fe7799b5199b21ad65179c2b51c0ee9c4472c0f49cfb1896096a1b2d8d7.jpg) +Figure 6. Qualitative comparison on 'The Big Bang Theory' videos from Sitcoms3D dataset, where each video feature a single actor. Our method demonstrates superiority in accurately positioning actors on the stage. Green points denote the Gaussian centers for the actors. +Figure 7. Qualitative comparison on four TV show videos, each featuring multiple actors. The frame from the input viewpoint is referred to as the Reference Frame. Green points denote the Gaussian centers for the actors. + +Multiple people scenario. Shape-of-Motion [60] and MonST3R [68] do not rely on template-based models, so we compare them with more challenging multi-people scenarios. We select video clips with a single shot since they assume continuous viewpoints. As demonstrated in Fig. 7, both Shape-of-Motion and MonST3R faces challenges deforming dynamic actors in a moving camera, even when videos with a single shot are given. Our method, however, shows robust reconstruction from videos with arbitrary camera translations. More qualitative results on the various scenes and viewpoints are shown in the supplement. + +# 4.3.Quantitative Comparison + +Stage reconstruction. We compare our stage reconstruction quality with other 3D reconstruction methods. For a fair comparison, we do not use object removal (Sec. 3.3). Given 165 background images of the TBBT scenario in the Sitcoms3D dataset, we reconstruct the background except for 10 randomly picked images for the testing. As shown in Table. 1, GS-W [67] faces challenges removing transient actors. Also, compared to 3DGS [22], and FSGS [70], masking out actors and using depth priors improves the quality when reconstructing from aggregated sitcom images. + +
MethodsPSNR↑SSIM↑LPIPS↓
Sitcoms3D [45]18.810.620.55
3DGS [22]19.210.640.49
GS-W [67]19.350.650.51
FSGS [70]19.340.650.49
Ours19.650.660.49
+ +![](images/6bb0d99db2e9311a42ba4b152b7d1c3647b06038c7c018ea3beb2f5b8e3fce47.jpg) +(a) ExAvatar +Figure 8. Ablation study for face-fitting network. + +![](images/2fcacc78fab5496d8d172033374c731096b991b8b3bd40fab320413f97c8d71d.jpg) +(b) Ours w/o refine + +![](images/04851995a30b0a1970704cd701cfda2c70023375e19d10ad3817122b82699345.jpg) +(c) Ours + +![](images/7e714784d4b6e2044a4884a45d0124a19f810f2f9e7b2fb5489d51b815994f3b.jpg) +(d) Ground Truth + +Table 1. Quantitative comparison of stage reconstruction results: We present the reconstruction metrics for the living room from 'The Big Bang Theory' in the Sitcoms3D dataset. + +
MethodsPSNR↑SSIM↑LPIPS↓
ExAvatar [37]20.170.640.25
Ours w/o refinement21.520.820.36
Ours24.340.840.28
+ +Actor reconstruction. We compare our face-fitting network with ExAvatar [37]. As shown in Fig. 8, while ExAvatar is designed to capture facial expressions, it struggles to recover fine details. In contrast, our simple and practical deformation module enables more precise facial expression adjustments. + +# 4.4. Ablation Study + +To evaluate the performance of 3DLocator, we propose two simple metrics: MTED (Mean Translation Euclidean Distance) and MPED (Mean Pose Euclidean Distance). These metrics calculate the Euclidean Distance of the global translation between two SMPL models and the Euclidean Distance of 3D joints at shot boundaries. As we can see in Table 3, without using 3DLocator, the fitting does not accurately align with the correct positions, leading to a reduced correlation of the actor between adjacent shots. + +Table 2. Quantitative comparison of face reconstruction results. We compare the performance of our face-fitting network against ExAvatar [37]. For this comparison, we crop the facial region. + +
MethodsMTED↓MPED↓
Ours w/o 3DLocator1.990.24
Ours1.190.10
+ +Table 3. Ablation study for 3DLocator. We report Mean Translation Euclidean Distance (MTED) and Mean Pose Euclidean Distance (MPED) across 2 consecutive shots. + +![](images/a3b7a6756b98d65a07acc27963c062b48df94926337cb06086a4de9e00f94eef.jpg) +Reference Frame + +![](images/5cc519a900371e092a6bdddcc4eb5cb484cb5cef9e88388a7d1e53dc2adda612.jpg) +(a) Actor Deletion + +![](images/4c5b6436fc6593dacea760e644cdcdfbaf72690a74b3b0e62be204f4ec5bd2ea.jpg) +(b) Actor Relocation + +![](images/788a74488e466f55f6ba906e5777326eb7bed041ba8f5a6c7ab19c2b73d60d35.jpg) +(c) Actor Insertion +Figure 9. The reconstructed scenes with ShowMak3r are editable. Existing actors can be removed, repositioned, or replaced. New actors can be added, and their poses can be adjusted. + +![](images/cb15c4f6af53751727d05b2fb5a6fe052bdeea64c16c1a2c3a80cfbdd6267216.jpg) +(d) Pose Manipulation + +# 4.5. Applications + +In Fig 9, we show the possible applications with ShowMak3r. Maintaining separate Gaussian sets for each actor and the stage makes it possible to perform edits, such as removing, relocating, and inserting specific actors from the video. Additionally, actor poses can be manipulated by controlling the pose parameters of the SMPL model. + +# 5. Conclusion + +We introduce ShowMak3r, a comprehensive pipeline for reconstructing dynamic radiance fields from TV shows. To tackle challenges like actor occlusions, cluttered stages, small baseline views, and sudden shot changes, we propose the following key modules: 3DLocator, which positions actors on the stage by using depth priors and interpolates unseen poses; ShotMatcher, which ensures continuous actor tracking across shot-changes; and face-fitting network for dynamic facial expression recovery. Experiments demonstrate that ShowMak3r effectively reassembles TV shows from novel camera viewpoints. To our knowledge, ShowMak3r is the first method to reconstruct the stage and multiple actors from shot-changing videos comprehensively. + +Limitation and Future Work. Our approach relies on an off-the-shelf 4D human pose estimation approach, which can become noisy in scenarios with severe occlusions. Additionally, our current depth estimation module struggles to predict zoomed-in images, making close-up shots difficult to reconstruct. We aim to expand ShowMak3r to enable explicit editing of the actors. + +Acknowledgements. This work was supported by IITP grant (RS-2021-II211343: AI Graduate School Program at Seoul National Univ. $(5\%)$ and RS-2023-00227993: Detailed 3D reconstruction for urban areas from unstructured images $(30\%))$ and NRF grant (No.2023R1A1C200781211 $(65\%))$ funded by the Korea government (MSIT). We thank Daeun Lee for her help with the rebuttal experiments. + +# References + +[1] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv, 2024. 3, 1 +[2] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2 +[3] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, et al. Omnire: Omni urban scene reconstruction. In International Conference on Learning Representations, 2025. 2 +[4] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 4 +[5] Arnab Dey, Di Yang, Rohith Agaram, Antitza Dantcheva, Andrew I Comport, Srinath Sridhar, and Jean Martinet. Ghnerf: Learning generalizable human features with efficient neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2812-2821, 2024. 2 +[6] Arindam Dutta, Meng Zheng, Zhongpai Gao, Benjamin Planche, Anwesa Choudhuri, Terrence Chen, Amit K Roy-Chowdhury, and Ziyan Wu. Chrome: Clothed human reconstruction with occlusion-resilience and multiview-consistency from a single image. In arXiv, 2025. 2 +[7] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2 +[8] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning Neural Volumetric Representations of Dynamic Humans in Minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023. 2 +[9] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of IEEE International Conference on Computer Vision, pages 14783-14794, 2023. 4, 5, 1 +[10] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. In Advances in Neural Information Processing Systems, pages 14715-14728, 2022. 2 + +[11] Chen Guo, Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Vid2Avatar: 3D Avatar Reconstruction From Videos in the Wild via Self-Supervised Scene Decomposition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12858-12868, 2023. 2, 4, 1 +[12] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 2 +[13] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. In Proceedings of IEEE International Conference on Computer Vision, pages 9352-9364, 2023. 2 +[14] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 2 +[15] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics, 42(4): 1-12, 2023. 2 +[16] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2 +[17] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. NeuMan: Neural Human Radiance Field from a Single Video. In European Conference on Computer Vision, pages 402-418, Cham, 2022. Springer Nature Switzerland. 1, 2, 3, 4, 5 +[18] Yuheng Jiang, Zhehao Shen, Penghao Wang, Zhuo Su, Yu Hong, Yingliang Zhang, Jingyi Yu, and Lan Xu. HiFi4G: High-Fidelity Human Performance Rendering via Compact Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19734-19745, 2024. 2 +[19] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In Proceedings of IEEE International Conference on Computer Vision, pages 3334-3342, 2015. 2 +[20] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. In arXiv, 2023. 1, 2, 6 +[21] Ephraim Katz. Ephraim katz's the film encyclopedia, 1979. 3 +[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):139-1, 2023. 1, 2, 3, 4, 6, 7, 8 +[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White + +head, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of IEEE International Conference on Computer Vision, pages 4015-4026, 2023. 3, 1 +[24] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 505-515, 2024. 2, 4, 5, 6, 1 +[25] Inhee Lee, Byungjun Kim, and Hanbyul Joo. Guess the unseen: Dynamic 3d scene reconstruction from partial 2d glimpses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1062-1071, 2024. 2, 5, 6 +[26] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. In arXiv, 2024. 2 +[27] Chen Li, Jiahao Lin, and Gim Hee Lee. Ghunerf: Generalizable human nerf from a monocular video. In 2024 International Conference on 3D Vision (3DV), pages 923-932. IEEE, 2024. 2 +[28] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2 +[29] Zhengqi Li, Simon Niklaus, Noah Snively, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2 +[30] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2 +[31] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 2 +[32] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2 +[33] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Trans. Graph., 34(6), 2015. 2 +[34] Jiahao Lu, Tianyu Huang, Peng Li, Zhiyang Dou, Cheng Lin, Zhiming Cui, Zhen Dong, Sai-Kit Yeung, Wenping Wang, and Yuan Liu. Align3r: Aligned monocular depth estimation for dynamic videos. In arXiv, 2024. 2 +[35] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 2, 6 +[36] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, 2020. 1, 2 +[37] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive Whole-Body 3D Gaussian Avatar. In European Conference on Computer Vision, pages 19–35, Cham, 2024. Springer Nature Switzerland. 2, 6, 8 +[38] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human Gaussian Splatting: Real-time Rendering of Animatable Avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 2 +[39] Jiteng Mu, Shen Sang, Nuno Vasconcelos, and Xiaolong Wang. Actorsnerf: Animatable few-shot human rendering with generalizable nerfs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18391-18401, 2023. 2 +[40] Linfei Pan, Dániel Baráth, Marc Pollefeys, and Johannes Lutz Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, 2024. 3, 1 +[41] Haokai Pang, Heming Zhu, Adam Kortylewski, Christian Theobalt, and Marc Habermann. ASH: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1165-1175, 2024. 2 +[42] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of IEEE International Conference on Computer Vision, pages 5865-5874, 2021. 2 +[43] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics, 40(6):1-12, 2021. 1, 2 +[44] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of IEEE International Conference on Computer Vision, pages 10975-10985, 2019. 2 +[45] Georgios Pavlakos, Ethan Weber, Matthew Tancik, and Angjoo Kanazawa. The one where they reconstructed 3d humans and environments in tv shows. In European Conference on Computer Vision, pages 732-749. Springer, 2022. 2, 4, 5, 6, 8, 1 +[46] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14314-14323, 2021. 2 + +[47] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In International Conference on Learning Representations, 2024. 2, 4, 1 +[48] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations, 2023. 2, 6 +[49] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2 +[50] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 2 +[51] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020-5030, 2024. 2 +[52] 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. 2 +[53] Zhijing Shao, Duotun Wang, Qing-Yao Tian, Yao-Dong Yang, Hengyu Meng, Zeyu Cai, Bo Dong, Yu Zhang, Kang Zhang, and Zeyu Wang. Degas: Detailed expressions on full-body gaussian avatars. In arXiv, 2024. 2 +[54] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1606–1616, 2024. 2 +[55] Robert Sklar. Film: An International History of the Medium. Thames and Hudson, 1990. 3 +[56] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerf-player: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2 +[57] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 2 +[58] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems, 34:12278-12291, 2021. 2 + +[59] Matias Turkulainen, Xuqian Ren, Jaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. Dn-splatter: Depth and normal priors for gaussian splatting and meshing. In arXiv, 2024. 4 +[60] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv, 2024. 2, 6, 7 +[61] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. In Association for the Advancement of Artificial Intelligence, 2025. 2 +[62] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In Advances in Neural Information Processing Systems, pages 8406–8441, 2023. 2 +[63] Zhaorong Wang, Yoshihiro Kanamori, and Yuki Endo. Eghumannerf: Efficient generalizable human nerf utilizing human prior for sparse view. In arXiv, 2024. 2 +[64] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-Viewpoint Rendering of Moving People From Monocular Video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16210–16220, 2022. 2 +[65] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 1, 2 +[66] Zhengming Yu, Wei Cheng, Xian Liu, Wayne Wu, and Kwan-Yee Lin. MonoHuman: Animatable Human Neural Field From Monocular Video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16943-16953, 2023. 2 +[67] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. In European Conference on Computer Vision, pages 341-359. Springer, 2024. 6, 7, 8 +[68] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. International Conference on Learning Representations, 2025. 2, 6, 7 +[69] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2 +[70] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2024. 6, 7, 8 \ No newline at end of file diff --git a/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/images.zip b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bea0b7953f9789554233e6e1ce8c7d224d80c22c --- /dev/null +++ b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26cecb7356267ddaaba3bc68785fd65f93e802a7e004ba55cf991c5bbaea6b4d +size 862202 diff --git a/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/layout.json b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b740d66b0baec6f6dc9355da7b9db86e88bc0c43 --- /dev/null +++ b/CVPR/2025/ShowMak3r_ Compositional TV Show Reconstruction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29f9e80320566cbf88f3185dade8caea315cfec8e24a38f3dbf26c27a1ecf784 +size 486963 diff --git a/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/54b990c7-51ee-410a-a2ff-63f58f87fc89_content_list.json b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/54b990c7-51ee-410a-a2ff-63f58f87fc89_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0a5ff4720855aa36beb4836c4de2072f4f588af1 --- /dev/null +++ b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/54b990c7-51ee-410a-a2ff-63f58f87fc89_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cf6cb2757995b4202cb377eb8f14f8c8c0c5238c6e6666e18d00beb89fbf724 +size 86824 diff --git a/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/54b990c7-51ee-410a-a2ff-63f58f87fc89_model.json b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/54b990c7-51ee-410a-a2ff-63f58f87fc89_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d382d375d2731dbb7ad1c162db4b5ae29ac9ef7b --- /dev/null +++ b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/54b990c7-51ee-410a-a2ff-63f58f87fc89_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ac2bacd1b8272fb3c531da1870055866fc0b8838b5e87b993facacfa3412cc5 +size 109513 diff --git a/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/54b990c7-51ee-410a-a2ff-63f58f87fc89_origin.pdf b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/54b990c7-51ee-410a-a2ff-63f58f87fc89_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ea0667cbc7d6cf439198c9fc17fce24d17d8d5af --- /dev/null +++ b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/54b990c7-51ee-410a-a2ff-63f58f87fc89_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e1219e0f4780c9050475c7a850892443e80dbe0f867cf0ed38a98d48c35a51c +size 3430157 diff --git a/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/full.md b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/full.md new file mode 100644 index 0000000000000000000000000000000000000000..60d9b47b367b3c8692bcfdcd26a85c39e0ace9ab --- /dev/null +++ b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/full.md @@ -0,0 +1,329 @@ +# ShowUI: One Vision-Language-Action Model for GUI Visual Agent + +Kevin Qinghong Lin $^{1}$ , Linjie Li $^{2}$ , Difei Gao $^{1}$ , Zhengyuan Yang $^{2}$ , Shiwei Wu $^{1}$ , Zochen Bai $^{1}$ , Stan Weixian Lei $^{1}$ , Lijuan Wang $^{2}$ , Mike Zheng Shou $^{1\boxtimes}$ $^{1}$ Show Lab, National University of Singapore $^{2}$ Microsoft + +# Abstract + +Building Graphical User Interface (GUI) assistants holds significant promise for enhancing human workflow productivity. While most agents are language-based, relying on closed-source API with text-rich meta-information (e.g., HTML or accessibility tree), they show limitations in perceiving UI visuals as humans do, highlighting the need for GUI visual agents. In this work, we develop a vision-language-action model in digital world, namely ShowUI, which features the following innovations: (i) UI-Guided Visual Token Selection to reduce computational costs by formulating screenshots as an UI connected graph, adaptively identifying their redundant relationship and serve as the criteria for token selection during self-attention blocks; (ii) Interleaved Vision-Language-Action Streaming that flexibly unifies diverse needs within GUI tasks, enabling effective management of visual-action history in navigation or pairing multi-turn query-action sequences per screenshot to enhance training efficiency; (iii) Small-scale High-quality GUI Instruction-following Datasets by careful data curation and employing a resampling strategy to address significant data type imbalances. With above components, ShowUI, a lightweight 2B model using 256K data, achieves a strong $75.1\%$ accuracy in zero-shot screenshot grounding. Its UI-guided token selection further reduces $33\%$ of redundant visual tokens during training and speeds up the performance by $1.4\times$ . Navigation experiments across web [12], mobile [35], and online [39] environments further underscore the effectiveness and potential of our model in advancing GUI visual agents. The models are available at https://github.com/showlab/ShowUI. + +# 1. Introduction + +Graphical User Interfaces (GUIs) are central to how individuals engage with the digital world, serving as virtual embodied interface for a range of daily activities. Meanwhile, Large Language Models (LLMs) [32], with their ability to comprehend complex language instructions and seamlessly + +![](images/ab48decb9aec20b7af35e76619c58ff6ea19e25baf0c62b419fc49cc3309434b.jpg) + +![](images/6664970b4fe5d4297559f1af2eabd8b61508930c68fe22a0200f12b30188e309.jpg) +Figure 1. ShowUI is a Vision-Language-Action model for GUI Automation. Given an environment screenshot, ShowUI efficiently processes it using UI-guided token selection for visual modeling and outputs an interaction action within the loop. + +![](images/b9f21394acadbb518f448384310d7ac9c0b4044999a6dc5c7d15ace7f6d01363.jpg) +Figure 2. Left: Zero-shot Screenshot grounding comparison between ShowUI and other GUI visual models in terms of model size and training scale (area); ShowUI reaching state-of-the-art accuracy as well as the most lightweight model (2B) with a smaller training dataset (256K). Right: Building upon Qwen2-VL-2B [40], our UI-guided visual token selection reduces visual token redundancy by $33\%$ during training, achieving a $1.4\times$ speedup. + +integrate tools, have shown significant potential in performing complex tasks through building agents [1, 13, 16, 55]. This progress inspires the development of intelligent GUI agents that can significantly streamline human workflows based on user intentions. + +Early efforts in GUI automation have primarily focused on developing language agents [12, 46, 54] that rely on closed-source, API-based LLMs like GPT-4 [32]. These agents leverage text-rich metadata like HTML inputs or accessibility trees to perform navigation and other tasks. However, the text-only approach is limited in real-world applications, where users typically interact with user in + +terfaces visually—through screenshots—without access to the underlying structural oracle information. This limitation underscores the need for developing GUI visual agents that can perceive and interact with UIs as humans do, such as assisting slide creation [29]. + +However, GUI visual perception presents unique challenges compared to natural image processing, requiring specialized skills such as UI element grounding or action execution rather than the conversational abilities typical of multi-modal chatbots. Recognizing this gap, researchers have begun training vision-language models to acquire these new abilities. For instance, studies like [11,15,17] utilize web screenshot datasets to enhance large multi-modal models' element-grounding capabilities. Meanwhile, works like [10, 30] address navigation tasks by instruction tuning models for multi-step navigation. + +Despite these advancements, training multi-modal models for GUI visual agents continues to face significant challenges related to modeling and training: (a) Expensive Visual Modeling: UI screenshots are typically high-resolution (e.g. 2K), resulting in lengthy token sequences that pose issues with long-context processing. Most existing models are not optimized for such high-resolution data, leading to inefficiencies and high computational costs. (b) Managing Interleaved Vision-Language-Action: actions differ from language modalities and may vary across devices (e.g. 'Return' on web interfaces versus 'Press home' on mobile devices) and adapt to different parameter settings (e.g. 'Scroll' actions have two directions on the web but four on mobile platforms), how to effectively model action is unclear. Additionally, it's essential to model actions alongside visual and query data. For example, navigation processes generate a history of screenshot and action steps, creating a complex, interleaved vision-language-action that models must effectively interpret and manage. (c) Diverse Training Data: with vast amounts of GUI data across different devices such as Web and Mobile, accompanied by diverse purpose annotations including element grounding and navigation, it remains unclear how to effectively select a high-quality training corpus for developing robust GUI visual models. These critical challenges have been under-explored but are essential for the development of effective visual models for GUI agents. + +In this work, we develop a vision-language model for GUI visual agents, aiming to address and resolve the aforementioned challenges, with the following key contributions: + +(i) UI-Guided Visual Token Selection: We recognize the uniqueness of UI screenshots (i.e., redundancy mixed with essential details) and develop a UI-friendly visual token selection approach. In RGB space, we represent each patch as a node and identify connected components to model redundancy across patches. This relationship guides the self-attention blocks within visual encoders or language models + +for token selection, effectively reducing computation. + +(ii) Interleaved Vision-Language-Action Streaming: We analyze the diversity of GUI actions, structuring them in JSON format and documenting their action space to assist the model in action interpretation. Additionally, we recognize the need for interleaved understanding across modalities, such as combining action with visual navigation history and balancing visual token lengths through multi-turn action with text queries to improve training efficiency. Our model is formulated as interleaved vision-language-action streaming, unifying the diverse needs in GUI scenarios. +(iii) Well-selected Instruction-following Dataset: Instead of utilizing data from all available sources, we conduct an in-depth analysis of each data type's properties. For example, in web data, visual elements (i.e., button) is more valuable than static text (which accounts for $40\%$ ) as most VLMs [40] possess strong OCR capabilities. Additionally, we introduce a small, high-quality instruction-following dataset that achieves strong UI grounding performance. Furthermore, we develop a rebalanced sampling strategy to address the substantial imbalance in UI data, ensuring consistent model performance across different setups. + +Building upon aforementioned innovations, we enhance Qwen2-VL-2B to create a powerful GUI visual agent, ShowUI. As shown in Fig.2, this results in a lightweight 2B model using 256K data, achieving a strong $75.1\%$ accuracy in zero-shot screenshot grounding. ShowUI also demonstrate competitive navigation ability in Web [12], Mobile [35], and Online [39] environments. Comprehensive ablation studies (Fig. 2) demonstrate the effectiveness of our UI-guided token selection approach, reducing redundant visual tokens by $33\%$ and accelerating training by $1.4\times$ . Moreover, we conclude with many discussions on current performance gaps and future directions. + +# 2. ShowUI + +ShowUI, as outlined in Fig. 3, is built on top of the vision-language model Qwen2-VL-2B [40], incorporating the following key components optimized for GUI tasks: (i) a novel UI-guided visual token selection strategy for efficient visual modeling, (ii) an interleaved vision-language-action streaming setup to flexibly unify different needs by GUI tasks and enhance training effectiveness, (iii) a training data recipe, crafted through a detailed analysis of individual GUI data types, which enables ShowUI training on a smaller, high-quality corpus. In the following sections, we introduce each component in detail. + +# 2.1. UI-Guided Visual Tokens Selection + +High-resolution screenshots can result in a large number of visual tokens after standard patching. As demonstrated in Fig.4a, a $1344 \times 756$ resolution on a PC yields approximately 5184 raw tokens with a $14 \times 14$ patching, after a + +![](images/2c5d4c028b499c64c52b1ba27ddc8716e64ae67c7416687fabb06278a93c973a.jpg) +Figure 3. Illustration of ShowUI. Given a user task query, a pre-defined action space, and an initial screenshot as observation, ShowUI proceeds by executing the next action, updating the screenshot, and continuing in this cycle. Notably, ShowUI features the following key innovation designs: (i) UI-Guided Visual Token Selection, which processes the input screenshot to build UI patch-wise connected graph. During training, a random subset of tokens is selected within each component for efficient visual modeling (Sec. 2.1). (ii) Interleaved Vision-Language-Action Streaming to effectively handle past screenshots and actions, improving navigation performance. (Sec. 2.2). + +$2 \times 2$ [40] merging, still results in 1296 tokens, creating a computational challenge within the self-attention module. + +What differentiates UI from natural vision? Unlike natural images, which captures real-world complexities and unpredictable patterns thus rich in semantic, textures, UI screenshots are inherently structured, with clear layouts and consistent color schemes optimized for readability and usability. This difference means that UI images often contain redundant empty spaces or simple backgrounds that do not carry essential information, aling for optimization or pruning. Moreover, small but functionally important elements, like icons or text, demand higher salience due to their role in interactivity and clarity. + +Thus, it is necessary to have a strategy that can differentiate between redundant and essential visual elements, enabling effectively pruning of irrelevant visual tokens without compromising usability. We found that the RGB space can serve as a useful guideline for this purpose as pattern variants, text fonts can be easily identify by its RGB values. Construct a UI Connected Graph. After dividing the screenshot into regular patches, we observed many neighboring patches share exactly the same RGB values and are thus redundant. To leverage this, we represent each patch as a node in a graph. If neighboring patches have the same RGB values, we connect their corresponding nodes, forming connected components. This allows us to group and simplify redundant areas while preserving essential visual elements identified by their unique RGB patterns. Visually identical patches can be easily detected by setting a small + +threshold on the difference between their patch tensors. Based on this insight, we use the Union-Find method to identify connected components in this UI connected graph, as described in Algorithm 1. This algorithm produces a graph with $K$ connected component, where $K$ is typically smaller than the original number of patches $G_{h} \times G_{w}$ . Based on the assignment of each node to its component, we can model the redundancy relationships among patches. + +As illustrated in Fig. 4a, this method can effectively balances component number based on their visual informative adaptively, using less component (more redundant patches) in google search page with sparser areas $(1296\rightarrow$ 291), while assigning more components (more independent + +# Algorithm 1 Find Connected Components on UI-Graph + +1: Input: Screenshot of size $H \times W$ , patch size $c$ , threshold $\delta$ +2: Output: Assignment map between patch and connected components. +3: Divide the image into $G_h \times G_w$ patches, each patch is a node, where $G_h = \frac{H}{c}$ and $G_w = \frac{W}{c}$ . +4: Initialize Union-Find structure UF over nodes +5: for all node $(i,j)$ do +6: for all neighbors $(i', j')$ to the right and below of $(i, j)$ do +7: if $\| \mathrm{RGB}(i,j) - \mathrm{RGB}(i',j')\| < \delta$ then +8: UF.union $(\left(i,j\right),\left(i^{\prime},j^{\prime}\right))$ +9: end if +10: end for +11: end for +12: return Assignment map from UF + +![](images/96973438dc1225cafef676ea9a7290585f135fdabb2967563e317dd74fdc3fc7.jpg) +Screenshot 1344 x 756 + +![](images/fa785b2d74f8d4363bf1d509f55a1f34454567cafdc74644dd5e3fca7b4c00f7.jpg) +Patchified (28 x 28) #1296 Tokens + +![](images/0b3f3917946e7988af5f14445b8905d31753a1b38b821bbcfbacc205fa32adc1.jpg) +UI Connected Graph #291 Components + +Screenshot 1344 x 756 +![](images/a2a6bd7d12935f4530c42cf41847a62578152f3811002a62cd896013c98977a8.jpg) +(a) UI Connected Graph adaptively assigns connected components based on the informativeness of screenshots. + +![](images/0373f788121f9ff1d26c2da2cb52aba3dd522072e0383de1977b331c7a539bfd.jpg) +Example1: Google Search + +![](images/57eba889cbfa11f08ffd9d0fce7b496425322e466da80af6d99ff3be0cff2933.jpg) +Patchified (28 x 28) #1296 Tokens Example2: Overleaf Template +UI Connected Graph #986 Components + +![](images/fb443c14ba4f7c89c98ac5d50942d4d67f76524b5fcf9e6f0179213ba7a2defc.jpg) + +Figure 4. Illustration of UI-guided Visual Tokens Selection. Left: Starting from an original screenshot (left) with a 28x28 patch grid (middle), resulting in 1296 tokens, the UI Connected Graph adaptively organizes the image into disjoint regions based on RGB values, allowing patches within the same region to be treated as redundant. Right: Comparison of two methods leveraging UI Connected Graph in visual token modeling: Token merging pools all tokens within one component, which loses individual positional information, while token selection, randomly sample part of tokens with each component, still retains their original position relationship. +![](images/ffa0c630df9686476a3b198e023a7f428f5764b9c50d39584cbe4813115cf263.jpg) +(b) Two representative token compression methods, where patches of the same color indicate the same component and are redundant to each other. + +![](images/8c6eb0e190cbc93af6f8e1d9a4fa23edbef7cc675827783d07f0c180d85570c5.jpg) + +![](images/80ebdfd8070accefd019e0badc4181a4d0eae6358315e48b1ef2a63436536b34.jpg) +Token Selection (Ours) + +patches) in text-rich overleaf screenshots (1296 → 986). In Fig.??, we display how our method constructs the UI-connected graph across different devices. Given identical resolution screenshots with the same initial visual patch tokens (e.g., 1272), our method adaptively constructs connected components based on the informativeness of the screenshots. + +Token Merging v.s. Token Selection. Next, we explore how leveraging this UI connected graph can improve model efficiency. We find that the primary computational bottleneck in existing vision-language models lies in the long sequences processed by cascaded self-attention layers, impacting both language models and visual encoders. + +A straightforward approach is to apply token merging methods [6, 20], which represent all patches within a component into a single token, as shown in the Fig.4b left half. In practice, however, we found that this approach disrupts positional relationships, as the original positional information in the pooled token sequence is inevitably lost, which is essential for accurate UI element grounding. + +To enable token compression within self-attention without losing positional information, we draw inspiration from Mixture-of-Depth [34], which sparsely samples tokens via a routing mechanism. In our case, the UI connected graph provides an effective routing criterion. Tokens within the same component can be considered redundant, so we randomly skip a portion of tokens within each component during training, leaving single-patch components unaffected to preserve essential elements. For the selected tokens, we retain their original position embeddings, ensuring that self-attention operates on the original positional relationships, even with a less token sequence. + +Notably, this token selection method introduces no additional learnable parameters. Therefore, we apply random token selection at a set ratio during training, while at inference, the method offers flexibility to use either with or without token selection, as both options maintain consistent positional relationships within the full token sequence. + +# 2.2. Interleaved VLA Streaming + +In this section, we aim to address how to formulate actions and their relationships with other modalities (i.e., visual and textual queries). + +What differentiates Action from natural text? The core functionality of a GUI model is navigation, conditioned on a text query and requiring the model to jointly predict: the correct action type (e.g., [CLICK] or [TYPE]), with corresponding action parameters (e.g., coordinates for [CLICK] or a string for [TYPE]). A major challenge in navigation arises from the action variants across different devices, such as: (i) Device-specific actions (e.g. [CLICK] is unavailable on mobile, whereas [PRESS HOME] does not exist on the web). (ii) Same action with different parameters (e.g. [SCROLL] has two directions—up and down—on the web, but four directions on mobile). (iii) Novel actions at test time that were not encountered during training. + +To manage these variations within our model, we first structure each action in a JSON format (i.e.,{'action'： 'action_type'，'value': 'element'， + +`position': $[x, y]$ '), where $[x, y]$ represents relative coordinates on 0-1. This allows us to standardize actions from diverse devices into a unified format. Secondly, we provide the model with a 'README' in the system prompt, documenting each action's usage within + +![](images/629cf22ecb250dd7021d209bf708090065016bd029f1775252f1922f04050731.jpg) +Figure 5. Illustration of Interleaved Vision-Text-Action Streaming. The visual tokens in screenshots are significantly longer (e.g., 1.3K) compared to queries or actions (e.g., fewer than 10). Thus, we introduce two modes: (a) Action-Visual Streaming for UI navigation and (b) Action-Query Streaming for UI grounding. These modes extend the concept of multi-turn dialogue and enable more efficient utilization of training data. + +the action space (e.g., 'CLICK'): Click on an element, value is not applicable and the position $[x,y]$ is required.) This setup encourages the model to interpret the action space document rather than memorizing fixed actions, enabling it to execute actions at test time in a function-calling manner [51]. Next, we discuss the relationship between action and other modalities, as illustrated in Fig.5. + +Action with Visual: The GUI navigation process typically involves multi-step trajectories, requiring the model to recognize current steps and determine the next action in context. This introduces the challenge of managing both past actions and past observations (screenshots): actions indicate what has been done but lack visual context, while screenshots capture the visual state but omit actions taken. To ensure complete history information, we formulate an interleaved vision-action stream, as shown in Fig.3, which captures both visual and action information sequentially. After the $i$ -th action, the resulting $(i + 1)$ -th screenshot enters the queue following the previously executed $i$ -th action, prompting the model to generate the $(i + 1)$ -th action. + +In practice, we can optionally mask parts of the visual history depending on the application. For instance, in Mobile [35], where cross-software screenshot changes occur, it is essential to retain screenshots to track visual appearances. In contrast, for Web [12], where screenshots generally remain stable on a static webpage across a series of actions, masking may be preferable for efficiency. + +Action with Textual query: In one-step navigation or element grounding, we might encounter one screenshot with multiple parallel actions [22] or multiple elements [11], where screenshots tend to be high-resolution, resulting in long token sequences (e.g., 1-2K tokens). Meanwhile, queries like UI element names and actions (coordinates) are typically much shorter (often fewer than 10 tokens). This discrepancy makes a one-image-per-action approach inefficient. To optimize training data utilization, we adopt a multi-turn dialogue approach, predicting multiple action annotations for each screenshot within a single forward pass. + +# 2.3. GUI Instructional Tuning + +Various GUI datasets are available in community, such as dominant of web data [10, 11, 43], mobile data [30, 35], which may contain element coordinates [11] or user tra + +
UsageDeviceSource#Sample#Ele.#Cls. (len.)Highlights
GroundingWebSelf-collected22K576KN/AVisual-based
MobileAMEX [8]97K926KN/AFunctionality
DesktopOmniAct [22]1008KN/ADiverse query
NavigationWebGUIAct [10]72K569K9 (7.9)One / Multi-step
MobileGUIAct [10]65K585K5 (9.0)Multi-step
TotalDiverse256K2.7M
+ +Table 1. Overview of our instruction-tuning data. #Sample indicates the number of the task instance (screenshot in grounding, task in navigation); #Ele. indicates the number of the element (i.e., bbox in grounding); #Cls. represents the number of action classes, and len. indicates the average trajectory length per task. + +jctories [30]. Rather than aggregating all available data sources, we analyze each dataset type to select representative data. Our selected data is illustrated in Tab.1. Our discussion is mainly on UI grounding data. For navigation, we source GUIAct [10] with mobile and web devices. + +(i) Web-visual elements: The web provides a highly accessible, text-rich source of UI data, easily crawled from HTML [43]. Our statistical analysis shows that the 'static text' tag accounts for a significant portion (40%). Given that most VLMs already with strong OCR capabilities [2, 40], we focus on collect visually rich elements. To this end, we developed a parser and collected 22K screenshots, retaining only visual-related elements such as those tagged with 'Button' or 'Checkbox' by removing static text. + +(ii) Desktop-diverse query: Desktop data is particularly valuable as it is challenging to collect automatically. We identified OmniAct [22], which includes manual elements from iOS, Windows, and Linux desktops with a small size (2K elements over 100 images). Its element is only labelled by original name such as 'message_ash'. To enrich this dataset and diversity, we employed reverse engineering techniques, utilizing ground-truth bounding boxes and its text elements. Then we prompt GPT-4o, with visual prompts highlighting target elements, to derive three types of query: appearance, spatial and intention; which we illustrated in Fig.6. This method results new 6K elements. See Supp. for detail prompt and discussion. + +![](images/f29fc3b9d331516e537727591eb56968ece1b90a6246a3e0b2eee5454f9ae5d3.jpg) +Figure 6. We derive three types of query (appearance, spatial relationship, and intention) from raw annotation, assisted by GPT-4o. + +(iii) Mobile-Functionality: mobile data is readily available in Android like [25, 35], which provide icon captioning. Notably, we consider [8] as its valuable functionality descriptions that go beyond simple atomic element names. Balance Data by Sampling: As shown in Tab.1, data scale varies significantly across types (e.g., only 100 desktop + +samples). To ensure fair exposure over each type, we develop a sampling strategy during training, giving each batch a comparable probability of including different data types. + +# 3. Experiments + +# 3.1. Benchmark Datasets + +We evaluate our model using the following benchmarks: Grounding. We use Screenshotpot [11], a zero-shot grounding evaluation benchmark with diverse data across three device, to assess text and widget recognition separately. + +Navigation. We evaluate navigation performance across four different datasets from different devices: (i) Web on Mind2Web [12], which includes an action space with three types of actions. (ii) Mobile on AITW [35], which action spaces include 11 actions. (iii) Online on MiniWob [39], an online environment with two types of actions, introduced to complement the offline benchmarks and test performance in an interactive setting. Details of the training settings are provided in the supplementary material. + +# 3.2. Main Results + +We organize our experiments on each downstream tasks to address the following questions: Q1: How does our model perform on each task? What improvements are achieved beyond existing VLM baseline? Q2: What are the effects and improvements of each component? Q3: What insights can be gained from each benchmark based on its unique properties? + +# 3.2.1 Grounding Tasks + +
MethodSize#TrainMobileDesktopWebAvg.
TextIconTextIconTextIcon
Qwen2-VL-2B [40]2B-24.210.01.49.38.72.49.3
Fuyu [5]8B-41.01.333.03.633.94.419.5
CogAgent [17]18B400K67.024.074.220.070.428.647.4
SeeClick [11]9.6B364K78.052.072.230.055.732.553.4
OmniParser [31]*-93.957.091.363.681.351.073.0
UGround [15]7B1.3M82.860.382.563.680.470.473.3
ShowUI-G2B119K91.669.081.859.083.065.574.9
ShowUI2B256K92.375.576.361.181.763.675.1
+ +Table 2. Zero-shot grounding on Screenspot. * means Omniparser use GPT-4V. "Size" refers to model size. ShowUI-G: trained solely on grounding data, excluding navigation data. ShowUI, delivers strong grounding results with a lightweight model size and minimal training data. + +In Tab. 2, we present zero-shot grounding results on the Screenshotpot [11]. This provides a straightforward signal of the shortcomings in each setup. We include one additional variant - ShowUI-G, which only used grounding data for training. Our finding includes: (i) Overall all methods, the text track scores are higher than the icon track, even for desktop text, which was less seen during training. This suggests that text grounding ability mainly learned from + +web and mobile is transferable across platforms. (ii) Mixing navigation data [10] does not impair grounding performance when an effective sampling strategy is used. (iii) The icon track is more challenging due to its visual grounding. Mobile scores are significantly higher than desktop and web, this emphasize the missing of visual UI grounding data beyond mobile devices. (iv) Notably, ShowUI, as the most lightweight method with minimal training data, achieves state-of-the-art grounding performance. + +# 3.2.2 Navigation Tasks + +Mobile: AITW. In Tab.3, we have the following findings: (i) Without interleaved streaming (i.e., without visual history), ShowUI $\dagger$ provides only a $1.1\%$ acc. improvement over the VLM baseline. However, with visual history, ShowUI achieves an additional $1.7\%$ acc. gain, likely because visual context is crucial for adapting to frequent software changes within a large action space (11), particularly on mobile platforms. (ii) ShowUI's zero-shot navigation learned from GUIAct [10] demonstrates transferability, suggesting further improvements can be made to the navigation component. (iii) ShowUI, beats OmniParser [31] and PaLM2-CoT [36], which either leverage closed-source APIs or HTML information, highlighting its potential of a single model as a standalone visual agent. + +
MethodFT?GeneralInstallGAppsSingleWebShop.Overall
ChatGPT-CoT [52]-5.94.410.59.48.47.7
PaLM2-CoT [36]------39.6
OmniParser [31]*48.357.851.677.452.957.7
SeeClick [11]✓54.066.454.963.557.659.3
Qwen2-VL-2B [40]✓61.471.862.673.766.767.2
ShowUI †✓63.572.366.072.365.868.3
ShowUI✓63.972.569.777.566.670.0
ShowUI-ZS×32.147.742.020.137.435.9
+ +Table 3. Performance of Mobile Navigation [35], where gray color indicates these methods either using HTML inputs or rely on close-source GPT-4V (marked with *). ShowUI $\dagger$ denotes our variant without interleaved visual-action streaming, utilizing only action history. + +Website: Mind2Web. In Tab. 4 for web navigation, we found that: (i) Instruction-tuning has a significant effect, brings $4.6\%$ Avg. Step SR. boost over Qwen2-VL-2B. Notably, ShowUI-2B's zero-shot yield comparable with SeeClick-9.6B which has pretrained and fine-tuning, and achieves relatively high Op. F1 $(80\% +)$ . (ii) Visual context in this task is less significant than in AITW, possibly because Mind2Web focuses primarily on a single, visually similar website and includes only three actions. (iii) The cross-website and cross-domain settings are harder than cross-tasks. This suggests the bottleneck is lie in UI visual perception rather than textual task understanding (websites/domains are unseen in training). One future effort for improvement is to develop training data with good (visually) domain diversity. + +
MethodFT?Cross-TaskCross-WebsiteCross-Domain
Ele.AccOp.F1Step.SREle.AccOp.F1Step.SREle.AccOp.F1Step.SR
MindAct [12]-55.175.752.042.065.238.942.166.539.6
GPT-4 [32]-41.660.636.235.851.130.137.146.526.4
OmniParser [31]*42.487.639.441.084.836.545.585.742.0
CogAgent [17]✓22.453.017.618.442.413.420.642.015.5
Qwen-VL [4]✓15.986.713.313.283.59.214.184.312.0
SeeClick [11]✓28.387.025.521.480.616.423.284.820.8
Qwen2-VL-2B [40]✓37.786.433.236.079.227.636.381.830.7
ShowUI †✓39.788.036.941.083.634.238.985.334.1
ShowUI✓39.988.637.241.683.535.139.486.835.2
ShowUI-ZS✘21.485.218.621.981.916.824.483.921.4
+ +Table 4. Web Navigation on Mind2Web. The gray correspond to methods that required HTML text input or rely on close-source GPT-4V (marked with *). ShowUI $\dagger$ denotes our variant utilizing only action history. ShowUI's zero-shot performance yield comparable score with SeeClick with pretrained and fine-tuning. + +
MethodFT?Score
CC-Net(SL) [19]✓23.4
Pix2Act [38]✓55.2
Qwen-VL [4]✓48.4
SeeClick [11]✓67.0
Qwen2-VL-2B [40]✓66.8
ShowUI †✓70.4
ShowUI✓71.5
ShowUI-ZS✗27.1
+ +Table 5. Results on online navigation MiniWob on 35-tasks split following SeeClick [11]. ShowUI $\dagger$ denotes our variant utilizing only action history. + +Online: MiniWob. In Tab. 5, this benchmark demonstrates model behavior in an online environment. Our key finding is that despite the simplicity of the Miniwob UI, the gap between ShowUI's zero-shot performance (27.1%) and fine-tuned Qwen-VL (48.4%) is substantial. In contrast, ShowUI's zero-shot surpasses in Mind2Web, likely because out-of-distribution errors are not adequately addressed in the instruction-tuning phase. This suggests that offline instruction-tuning alone is insufficient; we need to develop a learning strategy tailored for an online environment, which can handle novel error cases. + +
MethodStrategy #Vis.Ctx. Train.Speedup Test-time? Screenshot
BaselineN/A1344.01×N/A70.8
Token MergeUI-Graph852.81.6×✓42.3
34.7
Token SelectionRandom941.51.5×✓65.3
56.2
UI-Graph947.41.5×✓70.4
64.9
+ +(a) Comparison between different visual tokens compression methods. '#Vis.ctx' represents the avg. length of visual tokens across all layers. 'Train.Speedup' denotes the training efficiency improvement over the baseline. 'Inference' denotes whether this method is applicable at test time. + +
Insertion#layersScreenshot
All2865.7
Early1468.2
Late1467.6
Cross1470.5
+ +(b) Different insertion layers. + +
Ratio#Visual Ctx.Screenshot
01344.070.8
0.251185.270.6
0.5947.470.4
0.75848.668.3
1.0762.164.5
+ +(c) Different selection ratio. + +Figure 7. Ablation studies of several design chocies regarding our UI-Guided Token Selection. + +# 3.3. Ablation Studies + +Impact by UI-Guided Token Selection. In Tab.7a, we examine various visual token optimization methods on desktop subset through the following variants: (i) Baseline: no visual token optimization strategy applied; (ii) Token Merge: a mainstream method introduced in Sec.2.1, conditioned on our UI-Graph; (iii) Token Select.-Random: a variant that randomly selects a subset of visual tokens, serv + +ing as a direct baseline; (iv) Token Select.-UI-Graph: our proposed method leveraging UI-Graph for token selection. + +As shown, Token Merge performs worse than random selection, highlighting the importance of preserving positional relationships between tokens. Token Selection - UIGraph offers a good balance with a $1.5 \times$ speedup and competitive accuracy. While applying it at test-time slightly reduce accuracy due to resolution loss, it remains more reliable than random selection, demonstrating the effectiveness of UI connected graph as a guiding criterion. + +Selection of Layers Insertion. In Tab.7b, we present an ablation study on different insertion strategies, including insertion across all layers, early or late X layers, and cross-layer insertion, where layers alternate between inserted and non-inserted. With the same number of inserted layers, cross-layer insertion significantly outperforms both early and late insertion. + +Different Selection Ratio. The results is present in Tab.7c, illustrate the selection ratio as a trade-off between speedup and performance, with 0.5 as an effective choice. + +Impact of Interleaved Streaming. We evaluate performance over iterations using interleaved streaming for grounding and navigation tasks to study its effects. (i) Action-Query: In Fig. 8, we compare grounding training with and without multi-turn streaming. Multi-turn streaming leads to faster progress, especially in the initial warmup phase, and maintains a performance gap, demonstrating improved data utilization. (ii) Action-Visual: As shown in previous tables with the ShowUI $\dagger$ variant, we validated the impact by visual context. Fig. 9 illustrates model progress across iteration steps, where the trend shows that visual+action+multi-turn outperforms visual-action and action-only setups. This validates our interleaved streaming as an effective and efficient strategy. + +D. Impact by Instruction-Tuning Data. One our contributions is the analysis of training data for the grounding task in Sec.2.3. In Tab.6, we present a detailed ablation study to investigate the individual impact of each change on specific devices and settings. + +We found that (i) Data quality matters: OmniAct, with + +![](images/9b436b9ee6da50e53585f4c5b2d8a3ff1812d407a204a89e56de26268526ce9c.jpg) +Figure 8. Impact by Interleaved action-query streaming on Grounding task: trained with 119K grounding data, Eval with Screenshotpot. + +![](images/e159ce61b842fb82530c414bfbd3a515c8a8f105ad352f23398535ed608ec82c.jpg) +Figure 9. Impact by Interleaved action-visual streaming on Navigation task: trained with GUIAct, Eval with AITW. + +
Training Data#Sample#Ele.Mobile +Text IconDesktop +Text IconWeb +Text IconAvg.
AMEX97K1.06M90.166.878.850.7
Web (Seeclick [11])270K3.0M83.961.170.647.9
Web (text+vis)22K926K85.760.775.850.0
Web (vis)22K576K83.660.375.360.3
OmniAct1002K85.759.878.452.1
OmniAct (diverse)1008K88.366.876.357.9
Joint-Training119K1.6M90.866.475.351.4
Balanced Sampling119K1.6M91.669.081.859.0
+ +Table 6. Effect by individual training data on Screenshotpot. + +only 2K elements, achieves comparable scores to web data, and when augmented with GPT4o for diverse queries, it enhances model generalization and optimizes data usage efficiency. (ii) Our collected 22K web data outperforms SeeClick's 270K screenshots; additionally, filtering web data by visual criteria significantly reduces element size without impacting performance, suggesting that static text is less informative—a property inherent to VLM. (iii) Balanced sampling is essential, yielding a $3.7\%$ acc. gain and maintaining performance across individual setups. + +# 4. Related Works + +GUI Agents. Recent studies reveal the potential of LLMs beyond language modeling, with advancements in [14, 42, 47, 48, 50] demonstrating their ability to autonomously complete complex tasks using tool integration. This has prompted GUI automation approaches such as: (i) Training-free systems, which operate in separate stages: first, they gather UI information by converting the GUI to HTML representations, accessing the accessibility tree, or using visual models like OCR [24] and SoM [45, 46, 54]. Next, LLMs integrate this information to generate responses. This approach relies heavily on closed-source LLMs [32], resulting in high costs and limited applications. To address these limitations, (ii) Training-based models [11, 17, 27, 49] are proposed, which are pre-trained on large-scale UI vision-text corpora (e.g., screenshots), closing the visual perception gap to enable abilities like element grounding or navigation. Our work falls into the second category, focusing on key challenges in developing GUI visual agents, such as + +efficiently modeling high-resolution visual screenshots and utilizing interleaved streaming for history management. + +Vision-Language-Action Models. Vision-Language Models have recently made significant advancements, processing both visual and textual data, functioning effectively as flexible chatbots. However, they still face limitations when require interaction with the real-world environment. To address this, Vision-Language-Action (VLA) models have been developed to enhance physical or digital environmental perception and action generation. Notable examples include RT-2-X [33] and OpenVLA [23], which enable VLMs to perform actions in natural settings such as robotics [7, 18, 53], autonomous driving [3], and gaming [28, 41]. While substantial progress has been made in embodied contexts, the integration of VLA models into digital GUIs, remains less explored. In our work, we develop Vision-Language-Action (VLA) models for digital GUI environments to bridge this gap, focusing on the unique properties of GUIs and addressing how to model actions alongside other modalities within our VLA framework. + +Efficient Visual Representations. The high computational cost of Large Multimodal Models, especially with long sequences from high-res UI screenshots, is a major bottleneck. Common methods like token pruning [9, 37] and merging [21, 26] reduce cost but fail in GUI tasks, as they discard fine-grained spatial info critical for grounding. Mixture-of-Depth (MoD) [34, 44] offers a better tradeoff by varying depth per token, preserving position while improving speed. We propose a UI-friendly method for efficient visual token selection by building an RGB-space graph by treating patches as nodes and identifying connected components to capture redundancy. + +# 5. Conclusions + +We introduced ShowUI, a vision-language-action model for GUI visual agents that addresses key challenges in UI visual and action modeling, and instruction-tuning data curations. From the model aspect, our development of UI-Guided Visual Token Selection allows for efficient processing of high-resolution UI screenshots, significantly reducing computational costs. Our Interleaved Vision-Language-Action Streaming framework effectively manages complex interactions across modalities. From the data aspect, with a carefully curated, high-quality instruction-following dataset, ShowUI achieves strong performance, with a lightweight model size. These results demonstrate ShowUI's potential to advance GUI visual agents towards more human-like interaction and perception. + +Limitation and Future work. Our model is primarily trained on offline data. A promising future direction is to enhance it in online environments through reinforcement learning, enabling deeper exploration of its limitations. + +Acknowledgements. This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-RP-2022-030). + +# References + +[1] Microsoft copilot. https://copilot.microsoft.com/. Accessed: 2024-04-15. +[2] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. +[3] Hidehisa Arai, Keita Miwa, Kento Sasaki, Yu Yamaguchi, Kohei Watanabe, Shunsuke Aoki, and Issei Yamamoto. Covla: Comprehensive vision-language-action dataset for autonomous driving. arXiv preprint arXiv:2408.10845, 2024. +[4] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. +[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşürlar. Introducing our multimodal models, 2023. +[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. +[7] 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. +[8] Yuxiang Chai, Siyuan Huang, Yazhe Niu, Han Xiao, Liang Liu, Dingyu Zhang, Peng Gao, Shuai Ren, and Hongsheng Li. Amex: Android multi-annotation expo dataset for mobile gui agents. arXiv preprint arXiv:2407.17490, 2024. +[9] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. 2024. +[10] Wentong Chen, Junbo Cui, Jinyi Hu, Yujia Qin, Junjie Fang, Yue Zhao, Chongyi Wang, Jun Liu, Guirong Chen, Yupeng Huo, et al. Guicourse: From general vision language models to versatile gui agents. arXiv preprint arXiv:2406.11317, 2024. +[11] Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. Seeclick: Harnessing gui grounding for advanced visual gui agents. ACL, 2024. +[12] Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems, 36, 2024. +[13] Hiroki Furuta, Ofir Nachum, Kuang-Huei Lee, Yutaka Matsuo, Shixiang Shane Gu, and Izzeddin Gur. Multimodal + +web navigation with instruction-finetuned foundation models. arXiv preprint arXiv:2305.11854, 2023. +[14] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn. arXiv preprint arXiv:2306.08640, 2023. +[15] Boyu Gou, Ruohan Wang, Boyuan Zheng, Yanan Xie, Cheng Chang, Yiheng Shu, Huan Sun, and Yu Su. Navigating the digital world as humans do: Universal visual grounding for gui agents. arXiv preprint arXiv:2410.05243, 2024. +[16] Izzeddin Gur, Hiroki Furuta, Austin Huang, Mustafa Safdari, Yutaka Matsuo, Douglas Eck, and Aleksandra Faust. A real-world webagent with planning, long context understanding, and program synthesis. arXiv preprint arXiv:2307.12856, 2023. +[17] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. arXiv preprint arXiv:2312.08914, 2023. +[18] Siyuan Huang, Haonan Chang, Yuhan Liu, Yimeng Zhu, Hao Dong, Peng Gao, Abdeslam Boularias, and Hongsheng Li. A3vlm: Actionable articulation-aware vision language model. arXiv preprint arXiv:2406.07549, 2024. +[19] Peter C Humphreys, David Raposo, Tobias Pohlen, Gregory Thornton, Rachita Chhaparia, Alistair Muldal, Josh Abramson, Petko Georgiev, Adam Santoro, and Timothy Lillicrap. A data-driven approach for learning to control computers. In International Conference on Machine Learning, pages 9466-9482. PMLR, 2022. +[20] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. +[21] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. +[22] Raghav Kapoor, Yash Parag Butala, Melisa Russak, Jing Yu Koh, Kiran Kamble, Waseem Alshikh, and Ruslan Salakhutdinov. Omniact: A dataset and benchmark for enabling multimodal generalist autonomous agents for desktop and web. arXiv preprint arXiv:2402.17553, 2024. +[23] 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. +[24] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. + +[25] Yang Li, Jiacong He, Xin Zhou, Yuan Zhang, and Jason Baldridge. Mapping natural language instructions to mobile ui action sequences. arXiv preprint arXiv:2005.03776, 2020. +[26] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. 2024. +[27] Zhangheng Li, Keen You, Haotian Zhang, Di Feng, Harsh Agrawal, Xiujun Li, Mohana Prasad Sathya Moorthy, Jeff Nichols, Yinfei Yang, and Zhe Gan. Ferret-ui 2: Mastering universal user interface understanding across platforms. arXiv preprint arXiv:2410.18967, 2024. +[28] Shalev Lifshitz, Keiran Paster, Harris Chan, Jimmy Ba, and Sheila McIlraith. Steve-1: A generative model for text-to-behavior in apache. Advances in Neural Information Processing Systems, 36, 2024. +[29] Kevin Qinghong Lin, Linjie Li, Difei Gao, Qinchen Wu, Mingyi Yan, Zhengyuan Yang, Lijuan Wang, and Mike Zheng Shou. Videogui: A benchmark for gui automation from instructional videos. arXiv preprint arXiv:2406.10227, 2024. +[30] Quanfeng Lu, Wenqi Shao, Zitao Liu, Fanqing Meng, Boxuan Li, Botong Chen, Siyuan Huang, Kaipeng Zhang, Yu Qiao, and Ping Luo. Gui odyssey: A comprehensive dataset for cross-app gui navigation on mobile devices. arXiv preprint arXiv:2406.08451, 2024. +[31] Yadong Lu, Jianwei Yang, Yelong Shen, and Ahmed Awadallah. Omniparser for pure vision based gui agent. arXiv preprint arXiv:2408.00203, 2024. +[32] OpenAI. Gpt-4 technical report, 2023. +[33] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. +[34] David Raposo, Sam Ritter, Blake Richards, Timothy Lillicrap, Peter Conway Humphreys, and Adam Santoro. Mixture-of-depths: Dynamically allocating compute in transformer-based language models. arXiv preprint arXiv:2404.02258, 2024. +[35] Christopher Rawles, Alice Li, Daniel Rodriguez, Oriana Riva, and Timothy Lillicrap. Android in the wild: A large-scale dataset for android device control. arXiv preprint arXiv:2307.10088, 2023. +[36] Christopher Rawles, Alice Li, Daniel Rodriguez, Oriana Riva, and Timothy Lillicrap. Android in the wild: a largescale dataset for android device control. In Proceedings of the 37th International Conference on Neural Information Processing Systems, 2024. +[37] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. +[38] Peter Shaw, Mandar Joshi, James Cohan, Jonathan Berant, Panupong Pasupat, Hexiang Hu, Urvashi Khandelwal, Kenton Lee, and Kristina N Toutanova. From pixels to ui actions: Learning to follow instructions via graphical user interfaces. Advances in Neural Information Processing Systems, 36, 2024. + +[39] Tianlin Shi, Andrej Karpathy, Linxi Fan, Jonathan Hernandez, and Percy Liang. World of bits: An open-domain platform for web-based agents. In International Conference on Machine Learning, pages 3135-3144. PMLR, 2017. +[40] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024. +[41] Zihao Wang, Shaofei Cai, Zhancun Mu, Haowei Lin, Ceyao Zhang, Xuejie Liu, Qing Li, Anji Liu, Xiaojian Ma, and Yitao Liang. Omnijarvis: Unified vision-language-action tokenization enables open-world instruction following agents, 2024. +[42] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. +[43] Jason Wu, Siyan Wang, Siman Shen, Yi-Hao Peng, Jeffrey Nichols, and Jeffrey P Bigham. Webui: A dataset for enhancing visual ui understanding with web semantics. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1-14, 2023. +[44] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. In Proceedings of the 38th Conference on Neural Information Processing Systems, 2024. +[45] An Yan, Zhengyuan Yang, Wanrong Zhu, Kevin Lin, Linjie Li, Jianfeng Wang, Jianwei Yang, Yiwu Zhong, Julian McAuley, Jianfeng Gao, et al. Gpt-4v in wonderland: Large multimodal models for zero-shot smartphone gui navigation. arXiv preprint arXiv:2311.07562, 2023. +[46] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v, 2023. +[47] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction. Advances in Neural Information Processing Systems, 36, 2024. +[48] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023. +[49] Keen You, Haotian Zhang, Eldon Schoop, Floris Weers, Amanda Swearngin, Jeffrey Nichols, Yinfei Yang, and Zhe Gan. Ferret-ui: Grounded mobile ui understanding with multimodal llms. arXiv preprint arXiv:2404.05719, 2024. +[50] Andy Zeng, Maria Attarian, brian richter, Krzysztof Marcin Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael S Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. In The Eleventh International Conference on Learning Representations, 2023. + +[51] Jianguo Zhang, Tian Lan, Ming Zhu, Zuxin Liu, Thai Hoang, Shirley Kokane, Weiran Yao, Juntao Tan, Akshara Prabhakar, Haolin Chen, et al. xlam: A family of large action models to empower ai agent systems. arXiv preprint arXiv:2409.03215, 2024. +[52] Zhuosheng Zhang and Aston Zhang. You only look at screens: Multimodal chain-of-action agents. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 3132-3149. Association for Computational Linguistics, 2024. +[53] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. +[54] Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, and Yu Su. Gpt-4v(ison) is a generalist web agent, if grounded. In *Forty-first International Conference on Machine Learning*, 2024. +[55] Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Yonatan Bisk, Daniel Fried, Uri Alon, et al. Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854, 2023. \ No newline at end of file diff --git a/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/images.zip b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..03dc4ca11baf93849ead418210ea5efeb31de2da --- /dev/null +++ b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2e9393da56eb76cf1c10d93c766e9bfee6bf9ed7a3f9bd453e990749e296bb8 +size 568348 diff --git a/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/layout.json b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ad62a610597e47e0a2351a31979b9d88a9a15f74 --- /dev/null +++ b/CVPR/2025/ShowUI_ One Vision-Language-Action Model for GUI Visual Agent/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e4cb40066c26169d188ad3d40b16e3ce27dabdf030fdf9c8b833cf46278fd89 +size 403313 diff --git a/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/5f0bd45b-46e7-4a6d-88ca-50698109492a_content_list.json b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/5f0bd45b-46e7-4a6d-88ca-50698109492a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..30a3e544f952dbdf5dc41c1089d7b8ea492675ba --- /dev/null +++ b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/5f0bd45b-46e7-4a6d-88ca-50698109492a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:598f140c9022ea0ea37b07831151377e36b711676b6dd1de4077bc72995d5eac +size 78777 diff --git a/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/5f0bd45b-46e7-4a6d-88ca-50698109492a_model.json b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/5f0bd45b-46e7-4a6d-88ca-50698109492a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..048557b54e34fecf3898a16c178c8cb416ba701f --- /dev/null +++ b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/5f0bd45b-46e7-4a6d-88ca-50698109492a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ae7fe9cd489b5469383ad069502c002365124cd529e63faa76e1db125a530f7 +size 100761 diff --git a/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/5f0bd45b-46e7-4a6d-88ca-50698109492a_origin.pdf b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/5f0bd45b-46e7-4a6d-88ca-50698109492a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4211efdc07bb8ba79c09283da19e2f3cf6fee30d --- /dev/null +++ b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/5f0bd45b-46e7-4a6d-88ca-50698109492a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9b4f7b636508ac3963c774e574ecf0bccac94584f8c88df89725c5effcb229c +size 1413686 diff --git a/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/full.md b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a75b10a1a9989aac756329af02c651926daaf6a --- /dev/null +++ b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/full.md @@ -0,0 +1,316 @@ +# Silence is Golden: Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation + +Yuan Gan $^{1}$ , Jiaxu Miao $^{2*}$ , Yunze Wang $^{3}$ , Yi Yang $^{1}$ + +$^{1}$ ReLER, CCAI, Zhejiang University + +$^{2}$ School of Cyber Science and Technology, Sun Yat-sen University + +$^{3}$ Department of Statistics, University of Wisconsin-Madison + +# Abstract + +Advances in talking-head animation based on Latent Diffusion Models (LDM) enable the creation of highly realistic, synchronized videos. These fabricated videos are indistinguishable from real ones, increasing the risk of potential misuse for scams, political manipulation, and misinformation. Hence, addressing these ethical concerns has become a pressing issue in AI security. Recent proactive defense studies focused on countering LDM-based models by adding perturbations to portraits. However, these methods are ineffective at protecting reference portraits from advanced image-to-video animation. The limitations are twofold: 1) they fail to prevent images from being manipulated by audio signals, and 2) diffusion-based purification techniques can effectively eliminate protective perturbations. To address these challenges, we propose Silencer, a two-stage method designed to proactively protect the privacy of portraits. First, a nullifying loss is proposed to ignore audio control in talking-head generation. Second, we apply anti-purification loss in LDM to optimize the inverted latent feature to generate robust perturbations. Extensive experiments demonstrate the effectiveness of Silencer in proactively protecting portrait privacy. We hope this work will raise awareness among the AI security community regarding critical ethical issues related to talking-head generation techniques. Code: https://github.com/yuangan/Silencer. + +# 1. Introduction + +Talking-head animation [3, 14, 19, 49, 53, 61, 66, 67] enables the creation of synchronized and highly realistic facial expressions based on audio and portrait images, producing videos that are often indistinguishable from authentic visual recordings. Recent advances in diffusion models [7, 23, 35, 56, 64] have markedly improved the real + +![](images/c5810af34c65b9a1bd3c4497d5be418a536521405f1cb8eefd6e8009a8f401e3.jpg) +Figure 1. Overview of Our Motivation. Given an audio input, talking-head animation models can be exploited to generate fabricated videos using any portrait. To safeguard portrait privacy, we introduced Silencer, applying protective perturbations to ensure the portrait's mouth remains closed in generated talking videos. + +ism of these animations. Consequently, this technological advancement increases the risks of misusing AI-Generated Content (AIGC) for scams, political manipulation, and misinformation. Mitigating these ethical risks has become a critical priority in AIGC security. + +To address the ethical risks associated with AIGC, there are two primary approaches: passive defenses [8, 36, 43, 48, 59] and proactive defenses [27-29, 37, 39, 58]. Passive defenses focus on detecting whether a video has been fabricated, making it useful for forensics. However, these approaches cannot prevent the infringement of personal privacy by deepfakes. When victims realize their privacy has been violated, the damage may already be irreparable. In contrast, proactive defenses offer superior protection by proactively shielding individuals from harm. These methods use adversarial perturbations on the input images to disrupt the outputs of the generative model. + +Recent studies [28, 29, 39, 58] explored proactive defenses against diffusion models, particularly those mimicry techniques based on Latent Diffusion Models (LDM). By adding perturbations to input images, these approaches have achieved copyright protection in diffusion-based mimicry. However, existing methods fail to protect privacy in the + +audio-driven talking-head generation, which utilizes LDM to animate the given portrait. Their limitations are twofold: 1) they cannot prevent portraits from being animated by LDM-based talking-head models with a provided audio. Although these methods may reduce the quality of the generated video, this effect alone does not ensure privacy protection. 2) diffusion-based purification techniques can remove these protective perturbations, counteracting the quality degradation and rendering the privacy measures ineffective. Therefore, we need to generate robust adversarial perturbations that can nullify audio-driven facial movements and overcome purification techniques. + +To address the above challenges with robust perturbations, we propose Silencer, a two-stage approach to proactively protect portrait privacy from animation by talking-head generation methods. In the first stage, we introduce a nullifying loss by disregarding audio control in the talking-head generation. Due to the lack of ground truth video, previous methods cannot be directly applied to talking-head generation. Our nullifying loss modifies the optimization objective of talking-head training to keep the portrait "silent", as shown in Fig. 1. With the addition of adversarial noise via our nullifying loss, the generated talking videos tend to remain static, exhibiting low synchronization confidence. In the second stage, we design an anti-purification process using LDM to optimize the inverted latent feature, generating more robust perturbations. Since optimization in latent space does not have precise control over the outcomes in image space, directly applying nullifying loss to optimize the inverted latent feature would damage the adversarial portrait. Therefore, we use adversarial examples from the first stage to guide the optimization direction. To preserve the identity information, we apply a mask to the facial region halfway through the optimization process. + +Overall, our main contributions are threefold: + +- We introduce a benchmark for assessing proactive protection measures against privacy threats posed by advanced LDM-based talking-head generation techniques. +- We propose Silencer, a two-stage paradigm, to proactively protect portrait privacy with robust adversarial perturbations. First, we introduce a nullifying loss that effectively renders a portrait "silent" in the talking-head generation. Second, we develop an anti-purification strategy to enhance the robustness of these perturbations against countermeasures. +- Extensive experiments are conducted to assess the efficacy of our Silencer. Our method achieves strong privacy protection with low synchronization confidence and exhibits resistance to purification-based attacks. + +# 2. Related Work + +Audio-Driven Talking-Head Animation. Audio-driven talking-head generation has gained significant attention in + +recent years with the success of generative models [1, 13, 16, 22, 30, 35, 38, 44]. Early methods [3, 6, 14, 19, 49, 53, 61, 65–67] primarily relied on Generative Adversarial Networks (GANs) [16]. However, advancements in Latent Diffusion Models (LDMs) have led to more effective techniques [42, 47, 50, 53, 55, 56, 64]. AniPortrait [53] improves visual quality and temporal consistency by projecting 3D representations as 2D landmarks in a diffusion model. In contrast, approaches like DiffTalk and Diffused Heads [42, 47] simplify the generation process by focusing on diffusion-based methods without relying on 3D models. Furthermore, EMO [50] enhances expressiveness through a direct generation framework that eliminates the need for 3D models. VASA-1 [56] performs efficient operations in the latent space for highly natural real-time generation. Hallo [55] incorporates cross-attention mechanisms and innovative audio-landmark training strategies to enhance generation quality and animation stability. In this paper, we adopt Hallo as the pre-trained talking-head model. + +Adversarial Attacks in Diffusion Models. In the realm of adversarial attacks, early research introduced gradient-based methods that generate small perturbations to deceive neural network models [9, 10, 15, 17, 26, 31, 54, 62]. Building on these methods, recent studies have applied adversarial attacks to diffusion models. AdvDM [29] generates adversarial examples by optimizing latent variables during the reverse process of diffusion models. Photoguard [39] "immunizes" images by adding imperceptible perturbations that prevent diffusion models from generating realistic manipulations. Extending the ideas of AdvDM and Photoguard, Mist [28] incorporates semantic and texture loss designs to enhance cross-task transferability. Furthermore, Diff-ProTECT [58] introduces Score Distillation Sampling (SDS) and highlights the encoder module as the main vulnerability affecting the robustness of diffusion models. + +Purification and Anti-purification. Purification methods use generative models to remove adversarial noise before classification, thereby improving resistance to adversarial manipulations [11, 18, 21, 40, 41, 45, 46, 60]. Building on this foundation, DiffPure [34] utilizes the forward and reverse processes of diffusion models to purify adversarial examples. Moreover, GridPure [63] introduces a grid-based iterative diffusion approach tailored to high-resolution images, enhancing purification effectiveness. PDM-Pure [57] uses pixel-space diffusion models as a universal purifier to mitigate adversarial noise. + +To resist purification, ACA [4] maps images onto a low-dimensional latent manifold of the generative model and optimizes adversarial objectives to enable diverse content generation and control. Additionally, DiffAttack [2] introduces + +an innovative, diffusion-based attack method to bypass existing purification defenses via latent feature optimization. + +# 3. Method + +# 3.1. Preliminary + +# 3.1.1. Audio-driven Talking-head Generation with LDM. + +Given a reference portrait $p$ and speech audio $a$ , talking-head generation aims to generate realistic speaking videos synchronized with speech audio. To achieve this aim with powerful text-to-image LDM models, such as Stable Diffusion [35], recent works follow a common pipeline, which utilizes ReferenceNet and audio signals to guide the animation process. ReferenceNet has the same architecture as the LDM network, which extracts appearance features from reference images for guidance. As shown in Fig. 2(b), talking-head LDM employs spatial attention to preserve intricate appearance features from the reference image. By integrating these appearance features, the model accurately captures the reference portraits, allowing for precise manipulation of facial expressions with audio inputs. Hence, during the training phase, an associated talking frame $f_{i}$ is encoded into a latent representation $z_{0}$ with the encoder of Variational AutoEncoder (VAE) [12, 25]: $z_{0} = \mathcal{E}(f_{i})$ . The diffusion process across $T$ timesteps then transforms this latent representation to a Gaussian noise $z_{T} \sim \mathcal{N}(0,1)$ . The goal of the training is to progressively denoise $z_{T}$ to produce a realistic talking-head frame that not only preserves the visual characteristics of the reference portrait $p$ but also synchronizes the lip movements with the audio frame $a_{i}$ . To achieve this aim, the training loss is defined by the following objective function: + +$$ +\mathcal {L} _ {l d m} = \mathbb {E} _ {\mathcal {E} \left(f _ {i}\right), p, a _ {i}, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, p, a _ {i}\right) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $\epsilon \sim \mathcal{N}(0,1)$ is a Gaussian noise, $\epsilon_{\theta}$ represents the denoising U-Net model that processes the noisy latent variable $z_{t}$ at each timestep $t$ along with the conditional inputs, $p$ is the reference portrait, $a_{i}$ is the $i$ -th frame of the talking-head audio. + +The ability to use any person's portrait as a reference in talking-head generation raises significant privacy concerns. To mitigate this, we propose a proactive defense mechanism centered around an open-source, advanced LDM-based talking-head generation method [55]. + +# 3.1.2. Adversarial Examples for LDM + +Adversarial examples can protect images from LDM-based mimicry by finding the appropriate perturbations that can effectively cause LDM models to generate visually corrupted outputs. Previous studies have used two objective functions to exploit vulnerabilities in the diffusion model: + +![](images/87c6f5bb91feaa5cb1bb4a15a73074c564bb0141f0341ad05afd4ab0dfca1f7a.jpg) + +![](images/9ceac90027776c98dcffc697ad50438e3c927fe9cbc912d999c5a9891edb56a5.jpg) +Figure 2. LDM and LDM-based Talking-head Generation Framework. (a) The inference process of latent diffusion models. Given random noise and text, LDM can generate a semantically coherent image through iterative denoising. (b) The talking-head generation framework. Given a portrait and audio frame, the talking-head generation model can produce a lip-sync video frame with realistic facial expressions. + +- Semantic loss [29] is the training loss of LDM, which disrupts the denoising process, directing the model to produce samples that differ from the real image: + +$$ +\mathcal {L} _ {S} = \mathbb {E} _ {t, \epsilon} \mathbb {E} _ {z _ {t}} \| \epsilon - \epsilon_ {\theta} (z _ {t}, t) \| _ {2} ^ {2} \tag {2} +$$ + +- Texture loss [39] attacks the VAE encoder $\mathcal{E}(\cdot)$ by steering the latent representation of the input image $x$ towards a target latent derived from another image $y$ : + +$$ +\mathcal {L} _ {T} = - \| \mathcal {E} (x) - \mathcal {E} (y) \| _ {2} ^ {2} \tag {3} +$$ + +The final objective $\mathcal{L}_{adv}$ can be either semantic loss $\mathcal{L}_S$ , texture loss $\mathcal{L}_T$ , or both. Then PGD [33] is chosen to generate adversarial examples with projected gradient ascent: + +$$ +x ^ {n} = \mathcal {P} _ {B _ {\infty} (x, \delta)} \left[ x ^ {n - 1} + \eta \operatorname {s i g n} \nabla_ {x ^ {n - 1}} \mathcal {L} _ {\mathrm {a d v}} \left(x ^ {n - 1}\right) \right] \tag {4} +$$ + +where $x^n$ is the adversarial example at the $n$ -th iteration, $\mathcal{P}_{B_{\infty}(x,\delta)}[\cdot]$ projects the adversarial output onto the $\ell_{\infty}$ ball centered at $x$ with budget $\delta$ , $\eta$ is the step size. + +# 3.2. Silencer + +To protect portrait privacy, a straightforward approach is to directly apply semantic loss [29] to the talking-head animation task. However, this naive method presents two major + +![](images/151c7369c3b1cf8ca8ff0b6b6f17d099e28036869191a34712f6c4bc06f94e2f.jpg) +Figure 3. Silencer Framework. (a) In stage I, adversarial samples $p^n$ are generated through iterative PGD optimization using our proposed nullifying loss. These samples can avoid the influence of audio in the talking-head model. (b) In stage II, our anti-purification process is employed to optimize the inverted latent features, generating more robust perturbations capable of resisting purification. + +challenges: First, unlike image generation tasks, we lack ground truth frames $f_{i}$ , synchronized with audio frame $a_{i}$ for any arbitrary input portrait in talking-head animation. Second, we observed that noise-based perturbations introduced for privacy protection can be neutralized by purification methods. These purification methods counteract our protective measures, effectively compromising the intended privacy safeguards. In the following sections, we propose Silencer, a two-stage method to address these challenges. + +# 3.2.1. Silencer-I: Nullifying Loss + +To disrupt the denoising process and generate more artifacts in the edited images, semantic loss is used to optimize the perturbations by increasing or decreasing the LDM training loss. To calculate the training loss in Eq. 5, the ground truth frame $f_{i}$ and the corresponding latent representation $z_{0}$ are essential. A straightforward approach is to employ existing talking-head models to generate a synchronized video frame $f_{i}$ . However, this has two drawbacks. 1) Due to the time-consuming process of LDM inference, it is inefficient to generate fake ground truth under complex talking-head generation frameworks. 2) Generating fake ground truth presents a paradox: protecting portrait privacy requires it to first be compromised. + +Unlike semantic loss, which requires ground truth, texture loss operates without this requirement, relying instead on a target image. Despite this advantage, texture loss does not directly affect the synchronization of the generated videos. This is primarily due to the changes of the LDM + +network architecture, as shown in Fig. 2. Unlike traditional diffusion models, the LDM-based animation framework introduces conditions using ReferenceNet, which makes texture loss fail to eliminate the influence of the audio. Unless the face is fully obscured, the portrait can still be driven by audio. Hence, incorporating audio signals in the training of adversarial perturbations is crucial. + +Given the limitations of existing loss functions regarding ground truth requirements and audio signal integration, we propose training adversarial perturbations that are both audio-aware and independent of ground truth. We observe that forcing the generated result to stay "silent" is an effective way to disrupt audio-visual synchronization, avoiding the need to directly attack the talking-head training process. To nullify the effect of audio signals $a$ , we treat the reference portrait $p$ as the ground truth of the talking-head generation. Hence, we propose a nullifying loss to disrupt the audio-visual synchronization efficiently with the following formulation: + +$$ +\mathcal {L} _ {N} = \mathbb {E} _ {t, \epsilon} \mathbb {E} _ {\mathcal {E} (p), p, a _ {i}, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} (\hat {z} _ {t}, t, p, a _ {i}) \| _ {2} ^ {2} \right], \tag {5} +$$ + +where $\hat{z}_t$ is the noisy latent representation at timestep $t$ , $\mathcal{E}(p)$ is the latent representation $\hat{z}_0$ extracted from reference portrait $p$ . As the reference portrait is a condition in denoising, different timestep ranges would have different attack performances. Hence, we empirically experiment on $t$ to find the best range, as shown in Fig. 7. + +
MethodCelebA-HQ [24]TalkingHead-1KH [52]
V-PSNR/SSIM↓FID↑Sync↓M-LMD↑V-PSNR/SSIM↓FID↑Sync↓M-LMD↑
AdvDM(+) [29][ICML23]17.95/0.457578.405.61502.042519.09/0.4437178.923.81461.7581
AdvDM(-) [58][ICLR24]16.29/0.499847.346.66702.136617.42/0.555652.995.23991.7244
PhotoGuard [39][Arxiv23]17.67/0.4763126.095.88752.080018.76/0.5167186.843.37841.9023
Mist [28][Arxiv23]17.80/0.4753134.445.90522.117319.13/0.5241221.583.05521.7787
SDS(+) [58][ICLR24]17.79/0.456967.235.86682.100919.11/0.4464139.204.48441.6760
SDS(-) [58][ICLR24]16.54/0.496451.206.67432.073717.57/0.546257.075.19541.7301
SDTS(-) [58][ICLR24]17.23/0.482889.706.40032.102418.86/0.5496139.873.88251.8579
SilencerStage I19.02/0.5104124.074.06442.200820.61/0.5692168.851.79661.8025
SilencerStage II19.01/0.5111156.993.96852.210820.44/0.5718185.872.00171.8427
Ground Truth∞ /1.000.006.40410.0000∞ /1.000.005.48420.0000
+ +![](images/43473754bf51de7290f19281efd08ca7acd619b0ee2552d5483285508ab2dcc4.jpg) +Original + +![](images/208f9353d6934579e96b5e4f5bb582ac40452e21dc5bbc40768b53554fe67571.jpg) +ACA + +![](images/ed337959087d5a3aebce58cb207a8f476d59199e129bd7c1584978fc5bad5b84.jpg) +Silencer-I +Figure 4. Visualization of ACA [4] and Silencer. The result of ACA is generated by optimizing latent feature with skip gradients using our nullifying loss. + +![](images/d12531c23c67026f3d43cf40a150cc2bfaa81242fc139827095b1873856c1544.jpg) +Silencer-II + +$\mathcal{L}_N$ modifies the training target of talking-head generation from a synchronized frame to a still portrait. Then we treat $\mathcal{L}_N$ as the adversarial loss $\mathcal{L}_{adv}$ in Eq. 4 and optimize the reference image $p$ with PGD for $n$ iterations to acquire the adversarial example $p^n$ , as shown in Fig. 3 (a). It is noted that we adopt gradient descent rather than gradient ascent to optimize the adversarial portrait: + +$$ +p ^ {n} = \mathcal {P} _ {B _ {\infty} (p, \delta)} \left[ p ^ {n - 1} - \eta \operatorname {s i g n} \nabla_ {p ^ {n - 1}} \mathcal {L} _ {N} \left(p ^ {n - 1}\right) \right] \tag {6} +$$ + +where $p^{n-1}$ is the input image in $n$ -th iteration, $p^n$ is the output adversarial image. Not only does $p^n$ disrupt synchronization by remaining "silent", but it also degrades the video quality of the talking-head generation. + +# 3.2.2. Silencer-II: Anti-purification + +By optimizing a perturbation using our proposed nullifying loss, and adding it to the portrait image, we can generate an adversarial example that prevents the portrait from being driven by audio. Unfortunately, existing noise-removal or "purification" techniques can easily strip away the noise, undermining the protection effect. To address this, we need to find a more robust noise pattern that resists these purification methods, enhancing the security and effectiveness of the adversarial examples generated with our Silencer. + +ACA [4] generates adversarial examples by applying the gradients of adversarial classification loss in the latent space inverted by DDIM [44]. While it can make natural modifications to image content, optimizing the latent vector $z_{T}$ through the skipped gradients can lead to unpredictable and + +Table 1. Quantitative Comparisons with State-of-the-art Methods on CelebA-HQ [24] and TalkingHead-1KH [52]. "↑": higher is better. "↓": lower is better. Red: the 1st score. Blue: the 2nd score. + +
MethodCelebA-HQ I-PSNR/SSIM↑TalkingHead-1KH I-PSNR/SSIM↑
AdvDM(+) [29][ICML23]31.15/0.760531.32/0.7262
AdvDM(-) [58][ICLR24]31.02/0.719131.16/0.6807
PhotoGuard [39][Arxiv23]29.96/0.729930.20/0.7147
Mist [28][Arxiv23]30.06/0.734230.32/0.7190
SDS(+) [58][ICLR24]31.15/0.768831.29/0.7341
SDS(-) [58][ICLR24]31.26/0.737431.50/0.7062
SDTS(-) [58][ICLR24]30.42/0.744630.76/0.7307
SilencerStage I31.36/0.747532.34/0.7353
SilencerStage II27.23/0.677429.05/0.7590
+ +Table 2. Comparison on Image Quality after Protection. Our Silencer-I achieves the best average image quality with minimal added noise while achieving protection effects. + +undesirable changes, compromising the authenticity of the adversarial sample. These deviations are particularly pronounced in talking-head generation models, leading to significant distortions in facial identity, as shown in Fig. 4. Existing methods to address this issue, such as applying consistent constraints throughout the inversion process[2], suffer from high memory consumption and are not scalable to high-resolution portrait images. + +To address the limitations of existing methods, we propose a new constraint for generating robust adversarial examples with lower computational cost. Our method, illustrated in Fig.3 (b), leverages LDM and DDIM inversion to optimize the latent representation of the image. Instead of directly adding noise, we optimize the latent features to create perturbations that are resistant to purification techniques. To ensure these perturbations are effective without significantly altering essential facial features, we utilize a constraint during the optimization process. Specifically, we extracted the VAE feature of adversarial samples generated in step I. The encoded features then serve as a constraint, during the optimization of the LDM-based adversarial example. This constrained optimization allows us to balance two objectives: maintaining crucial facial features for recognition, while simultaneously maximizing the ro + +![](images/9397f2cd8203152cdf26c5798d5557f5b5aaaa37f1aedf86391c01a3339f83a6.jpg) +Figure 5. Qualitative Comparison with Image Protection Methods. We visualize the protected portraits and their frame driven by audio. + +bustness of the perturbations against purification defenses. The optimization objective is formulated as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {A P} = \lambda_ {1} \mathcal {L} _ {N} - \lambda_ {2} \mathcal {L} _ {T} (7) \\ = \lambda_ {1} \mathcal {L} _ {N} + \lambda_ {2} \| \mathcal {E} \left(p _ {0} ^ {\prime}\right) - \mathcal {E} \left(p ^ {n}\right) \| _ {2} ^ {2} (8) \\ \end{array} +$$ + +where $p_0^\prime$ is the output of the inverted diffusion model, $p^n$ is the adversarial example generated in stage I, $\lambda_{1}$ and $\lambda_{2}$ are the corresponding coefficients. Then we use AdamW [32] to optimize the inverted latent representation $p_t$ with $\mathcal{L}_{AP}$ . + +We observed that optimizing the entire image with $\mathcal{L}_{AP}$ introduces considerable distortions to the facial region. LDM tends to distort the entire face in the iterative optimization process, aiming to eliminate recognizable features like the mouth and eyes. This results in a damaged face without any specific regions that could be manipulated or driven. Hence, we apply a facial mask to the optimization process that strikes a balance between face clarity and anti-purification protection. Specifically, we optimize the entire image during the initial $s$ iterations. After this point, we restrict optimization to areas outside the masked facial region. + +# 4. Experiments + +# 4.1. Experimental Setup + +# 4.1.1. Implementation Details + +The videos are sampled at 25 FPS and the audio sample rate is $16\mathrm{KHz}$ . The reference portraits are resized to $512 \times 512$ . We utilize Hallo [55] as the LDM-based talking-head + +model with the public implementation1. In the first stage, we adopt PGD with a budget 16/255 to train each portrait for 100 iterations, which is the same as our baselines. In the second stage, we optimize the inverted latent feature for 200 iterations with a learning rate of 0.01. + +Baselines and Dataset. We compare our proposed method with four state-of-the-art privacy protection methods, including AdvDM [29], PhotoGuard [39], Mist [28] and SDS [58]. To evaluate the performance of protection baselines, we select 50 images from the CelebA-HQ [24] dataset as the reference images and one audio as the driving signal. To create a more realistic scenario, we utilize CLIP-IQA [51] to select 50 high-quality video clips, each with a unique identity and associated speech audio, from the widely used TalkingHead-1KH dataset [52]. + +# 4.1.2. Metrics + +We assess the quality of synthesized emotional videos with the following metrics: + +Image quality. We utilize Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Fréchet Inception Distance score (FID) [20] to measure the image quality of synthesized videos. The video metrics, V-PSNR/SSIM, measure PSNR/SSIM specifically on facial regions. In contrast, the image metrics, I-PSNR/SSIM, calculate PSNR and SSIM across the entire image by comparing the original image with the adversarial image. The Fréchet Inception Distance (FID) is a common metric for measuring the fidelity of synthesized videos. It quantifies + +
MethodProtectedJPEG [41]AdvCleanDiffPure [34]GrIDPure [63]
I-PSNR/FIDI-PSNR↓/FID↑I-PSNR↓/FID↑I-PSNR↓/FID↑I-PSNR↓/FID↑
AdvDM(+) [29][ICML23]31.15/86.5831.35/62.0933.65/49.3628.78/44.8327.81/38.55
AdvDM(-) [58][ICLR24]31.02/66.7331.96/36.9733.29/40.1729.11/42.1727.86/25.48
PhotoGuard [39][Arxiv23]29.96/153.5030.37/96.9831.37/109.8728.36/48.4326.32/50.04
Mist [28][Arxiv23]30.06/156.4230.47/99.7231.51/108.9228.40/44.1426.35/49.88
SDS(+) [58][ICLR24]31.15/86.4131.30/60.4733.58/44.8028.72/44.2327.81/38.29
SDS(-) [58][ICLR24]31.26/70.9332.24/40.6933.31/48.0729.10/39.9227.89/25.90
SDTS(-) [58][ICLR24]30.42/112.4330.97/72.9031.89/82.1928.53/47.9726.63/39.90
SilencerStage I31.36/135.7733.38/55.9435.03/61.3829.26/38.9328.16/20.85
SilencerStage II27.23/175.2127.41/159.3427.95/135.3027.26/87.2225.72/144.89
+ +Table 3. Purification Experiments on CelebA-HQ [24]. The "Protected" is the metrics calculated with protected portraits for reference. Others are calculated with purified portraits. ↑: higher is better. ↓: lower is better. Red: the 1st score. Blue: the 2nd score. + +![](images/6880f0ed7d10ba016cedcd81bcfff2bc10dd69bc18db36715ed5aa650535e5d5.jpg) +Figure 6. Visual Comparison in Anti-purification. The third row is the animated talking frames with the portraits after GrIDPure [63]. + +the distribution distance between videos generated using original and protected portraits. + +Audio-visual synchronization. We evaluate the audio-visual synchronization of the synthesized videos using SyncNet's confidence score [5, 14]. In addition, the distance between the landmarks of the mouth (M-LMD) [3] is used to indicate speech content consistency. + +# 4.2. Privacy Protection + +We first compare the effectiveness of our Silencer in privacy protection with other state-of-the-art methods. We randomly selected an audio from TalkingHead-1KH dataset for training all adversarial example and tested the talking-head model with other audios. In CelebA-HQ, all tests used the same audio clip, while in TalkingHead-1KH, each face was tested with its original audio. We treat videos generated by Hallo using the portraits without protection as the ground truth for comparison. + +Table 1 shows that our method achieves the best synchronization protection, with a score of 3.9685 on CelebA-HQ and 2.0017 on TalkingHead-1KH. In stage I of Silencer, + +our nullifying loss effectively frees reference portraits from audio control during talking-head generation. In stage II, our Silencer continues to yield strong results, further validating the effectiveness of our method. In terms of video quality, our method can only achieve comparable results in the video FID. This is primarily due to our nullifying loss, which aims to ensure the reference portrait remains largely unchanged during the diffusion process. As a result, our method has less impact on the generated video quality compared to others. Table 2 shows that Silencer-I achieves the highest I-PSNR, indicating minimal degradation of the reference portrait's realism. Furthermore, the qualitative comparison in Fig. 5 reveals that, unlike methods that significantly alter facial appearance, our approach preserves visual consistency while "silencing" the talking head. These results underscore the effectiveness of our method in achieving privacy protection from audio control. + +# 4.3. Anti-Purification Experiments + +To demonstrate the effectiveness of our methods in resisting purification, we conduct purification on the protected + +![](images/c54dd596a3cfa0e1d3a59999f22e5955ae46dd82765a3f0c925f71836b3d6fae.jpg) +Figure 7. Ablation Study on Timestep Ranges in Silencer-I. + +portraits. We select the following advanced purification methods to attack: JPEG [41], AdvClean, Diff-Pure [34] and GrIDPure [63]. We evaluate anti-purification effectiveness using I-PSNR, comparing original and purified images, and FID for talking-head videos, comparing ground truth to videos generated with purified portraits from a CelebA-HQ subset. Table 3 shows Silencer-II achieves the best anti-purification performance, with the lowest I-PSNR and highest FID. This is because adversarial noises, generated by Silencer-I and other methods, are optimized in the image space with PGD [33] and can be easily purified. In contrast, Silencer-II optimizes perturbations within the LDM's inverted latent space, resulting in fundamentally different and more robust perturbations. Fig. 6 visually demonstrates this efficacy. Although our generated perturbations resist complete removal, their structure is altered by purification, preventing a perfectly "silent" portrait. Achieving a completely robust perturbation that results in a "silent" portrait even after purification remains a challenge. We will explore more robust solutions in future work. + +# 4.4. Ablation Study + +Ablation Study on Timestep Ranges in Silencer-I. Since the reference portrait serves as a condition in the denoising process, adjusting the timestep range results in varying levels of attack effectiveness. To identify the optimal timestep ranges, we divided the total of 1000 timesteps into ten equal segments, sampling 100 timesteps from each segment for training. We trained and evaluated our model on the subset of CelebA-HQ, with the results illustrated in Fig. 7. Our findings indicate that timesteps within the [200, 300] range achieve a desirable balance: they provide effective privacy protection, with a sync confidence of 4.2556, while maintaining minimal noise, with an I-PSNR of 32.34. Based on these results, we selected the [200, 300] range for + +
AblationSDTS(-)S-IS-II (A)S-II (B)
LN✓✓✓
Anti-purify✓✓
Mask✓
I-SSIM↑0.74460.74750.65610.6774
FID↑89.70124.07167.40156.99
Sync↓6.40034.06443.43393.9685
M-LMD↑2.10242.20082.36072.2108
+ +Table 4. Ablation Study of Each Component. Each component contributes to improving privacy protection, thus verifying its effectiveness. + +timestep sampling in training Silencer-I. + +Ablation Study on Each Component. We conduct an ablation study to evaluate the impact of each component of our Silencer using the CelebA-HQ dataset. As shown in Table 4, the introduction of our nullifying loss leads to a significant reduction in synchronization confidence compared to previous methods. Additionally, the anti-purification process remains the low synchronization, providing protection against talking-head manipulation. However, this comes at the cost of reduced visual quality. To mitigate this, we optimize the adversarial perturbation with a face mask, which preserves facial structure while achieving effective privacy protection. More experiments can be found in our supplementary material. + +# 5. Conclusion + +In this paper, we introduce Silencer, a two-stage approach to proactively protect portrait privacy from unauthorized animation in audio-driven talking-head generation. This approach addresses the limitations of prior methods that cannot effectively mitigate talking animation and resist purification. The first stage employs a novel nullifying loss to decouple facial movements from audio input, significantly reducing the synchronization of generated talking-head videos. Building upon this, the second stage enhances robustness through an anti-purification process. This process optimizes perturbations within the inverted latent space of an LDM, guided by adversarial examples from the first stage to ensure targeted and effective protection. A strategically applied mask preserves facial integrity during this optimization. Extensive experiments demonstrate Silencer's superior performance in both preventing unauthorized animation and resisting purification, confirming its effectiveness in protecting portrait privacy. This work establishes a new benchmark for proactive privacy protection in LDM-based talking-head generation and we anticipate it will stimulate further research and development in this critical area. + +# Acknowledgments + +This work was supported in part by the Key R&D Program of China (2021ZD0112801), the State Key Program of National Natural Science Foundation of China (62436007) and the Natural Science Foundation of Zhejiang Province (LDT23F02023F02). + +# References + +[1] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt. arXiv preprint arXiv:2303.04226, 2023. 2 +[2] Jianqi Chen, Hao Chen, Keyan Chen, Yilan Zhang, Zhengxia Zou, and Zhenwei Shi. Diffusion models for imperceptible and transferable adversarial attack. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-17, 2024. 2, 5 +[3] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019. 1, 2, 7 +[4] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 5 +[5] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In *Asian conference on computer vision*, pages 251–263. Springer, 2016. 7 +[6] Joon Son Chung, Amir Jamaludin, and Andrew Zisserman. You said that? BMVC, 2017. 2 +[7] Jiahao Cui, Hui Li, Yao Yao, Hao Zhu, Hanlin Shang, Kaihui Cheng, Hang Zhou, Siyu Zhu, and Jingdong Wang. Hallo2: Long-duration and high-resolution audio-driven portrait image animation. arXiv preprint arXiv:2410.07718, 2024. 1 +[8] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 5781-5790, 2020. 1 +[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 2 +[10] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4312-4321, 2019. 2 +[11] Yilun Du and Igor Mordatch. Implicit generation and modeling with energy based models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2 +[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Pro + +ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3 +[13] Sheng Fan, Rui Liu, Wenguan Wang, and Yi Yang. Navigation instruction generation with bev perception and large language models. In ECCV, 2024. 2 +[14] Yuan Gan, Zongxin Yang, Xihang Yue, Lingyun Sun, and Yi Yang. Efficient emotional adaptation for audio-driven talking-head generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22634-22645, 2023. 1, 2, 7 +[15] Lianli Gao, Qilong Zhang, Jingkuan Song, Xianglong Liu, and Heng Tao Shen. Patch-wise attack for fooling deep neural network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 307-322. Springer, 2020. 2 +[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[17] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2 +[18] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. arXiv preprint arXiv:1912.03263, 2019. 2 +[19] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 1, 2 +[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[21] Mitch Hill, Jonathan Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. arXiv preprint arXiv:2005.13525, 2020. 2 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[23] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 1 +[24] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 5, 6, 7 +[25] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3 +[26] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial in + +telligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 2 +[27] Zheng Li, Ning Yu, Ahmed Salem, Michael Backes, Mario Fritz, and Yang Zhang. Unganable: Defending against gan-based face manipulation. In 32nd USENIX Security Symposium (USENIX Security 23), pages 7213-7230, 2023. 1 +[28] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 1, 2, 5, 6, 7 +[29] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. In International Conference on Machine Learning, pages 20763-20786. PMLR, 2023. 1, 2, 3, 5, 6, 7 +[30] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2 +[31] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European conference on computer vision, pages 549-566. Springer, 2022. 2 +[32] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[33] Aleksander Madry. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 3, 8 +[34] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460, 2022. 2, 7, 8 +[35] 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. 1, 2, 3 +[36] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 1 +[37] Nataniel Ruiz, Sarah Adel Bargal, and Stan Sclaroff. Disrupting deepfakes: Adversarial attacks against conditional image translation networks and facial manipulation systems. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 236-251. Springer, 2020. 1 +[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2 +[39] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. arXiv preprint arXiv:2302.06588, 2023. 1, 2, 3, 5, 6, 7 + +[40] P Samangouei. Defense-gan: protecting classifiers against adversarial attacks using generative models. arXiv preprint arXiv:1805.06605, 2018. 2 +[41] Pedro Sandoval-Segura, Jonas Geiping, and Tom Goldstein. Jpeg compressed images can bypass protections against ai editing. arXiv preprint arXiv:2304.02234, 2023. 2, 7, 8 +[42] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In CVPR, 2023. 2 +[43] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18720-18729, 2022. 1 +[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 5 +[45] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2 +[46] Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. arXiv preprint arXiv:1710.10766, 2017. 2 +[47] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 2 +[48] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 1 +[49] Shuai Tan, Bin Ji, Mengxiao Bi, and Ye Pan. Edtalk: Efficient disentanglement for emotional talking head synthesis. In European Conference on Computer Vision, pages 398-416. Springer, 2024. 1, 2 +[50] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 2 +[51] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 6 +[52] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In CVPR, 2021. 5, 6 +[53] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animations, 2024. 1, 2 +[54] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 2 + +[55] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2, 3, 6 +[56] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1, 2 +[57] Haotian Xue and Yongxin Chen. Pixel is a barrier: Diffusion models are more adversarially robust than we think. arXiv preprint arXiv:2404.13320, 2024. 2 +[58] Haotian Xue, Chumeng Liang, Xiaoyu Wu, and Yongxin Chen. Toward effective protection against diffusion-based mimicry through score distillation. In The Twelfth International Conference on Learning Representations, 2023. 1, 2, 5, 6, 7 +[59] Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22412-22423, 2023. 1 +[60] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072. PMLR, 2021. 2 +[61] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audiodriven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 1, 2 +[62] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. Towards large yet imperceptible adversarial image perturbations with perceptual color distance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1039-1048, 2020. 2 +[63] Zhengyue Zhao, Jinhao Duan, Kaidi Xu, Chenan Wang, Rui Zhang, Zidong Du, Qi Guo, and Xing Hu. Can protective perturbation safeguard personal data from being exploited by stable diffusion? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24398-24407, 2024. 2, 7, 8 +[64] Zhiquan Chen Yuming Li Chenguang Ma Zhiyuan Chen, Ji ajiong Cao. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditioning, 2024. 1, 2 +[65] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9299–9306, 2019. 2 +[66] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on + +computer vision and pattern recognition, pages 4176-4186, 2021. 1 +[67] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 1, 2 \ No newline at end of file diff --git a/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/images.zip b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..64a294fa04181364f72356ceaccb9c8eb81e1ba9 --- /dev/null +++ b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e35998c4d3605f5ff7fc2c49f928bf878cc3d3ec4a8148008e9a3a248c02444 +size 860132 diff --git a/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/layout.json b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..14bd3dc2d7efcc874aab90859be12dbce42f2f58 --- /dev/null +++ b/CVPR/2025/Silence is Golden_ Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c19e392f0a7fd09eca79db2938dd7a7cda56b057bf22b2a137baa444d4f6d892 +size 383449 diff --git a/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/93769aac-6d09-4da8-8358-002554f96486_content_list.json b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/93769aac-6d09-4da8-8358-002554f96486_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cbbf4aa078d76a5d3331256e0457d510ec34d6b4 --- /dev/null +++ b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/93769aac-6d09-4da8-8358-002554f96486_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99f865c18f88822e50e5c99f6d9920ef924bf8772c96d6798c2aca8b8e82951a +size 84387 diff --git a/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/93769aac-6d09-4da8-8358-002554f96486_model.json b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/93769aac-6d09-4da8-8358-002554f96486_model.json new file mode 100644 index 0000000000000000000000000000000000000000..70105fff233dc065ac2304b87b33bbc7847e1c6a --- /dev/null +++ b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/93769aac-6d09-4da8-8358-002554f96486_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96c3eda22443d3be51ecbbdcc19788f55ec3735385e14504872646c724a96212 +size 101893 diff --git a/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/93769aac-6d09-4da8-8358-002554f96486_origin.pdf b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/93769aac-6d09-4da8-8358-002554f96486_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..06da938650ba851de59cdec23c40e7ad7cd96f07 --- /dev/null +++ b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/93769aac-6d09-4da8-8358-002554f96486_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d096120a3fb1fd4a90237bbec397d56d743f72903c0fc1d659a84e19d62b869 +size 8545730 diff --git a/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/full.md b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d277b7f698ff3bece8e18e3e443431b1e3ebe22d --- /dev/null +++ b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/full.md @@ -0,0 +1,371 @@ +# Silent Branding Attack: Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models + +Sangwon Jang1 June Suk Choi1 Jaehyeong Jo1 Kimin Lee1,† Sung Ju Hwang1,2,† +1KAIST, 2DeepAuto.ai + +{ sangwon.jang, w_choi, harryjo97, kiminlee, sungju.hwang }@kaist.ac.kr + +# Abstract + +Text-to-image diffusion models have achieved remarkable success in generating high-quality contents from text prompts. However, their reliance on publicly available data and the growing trend of data sharing for fine-tuning make these models particularly vulnerable to data poisoning attacks. In this work, we introduce the Silent Branding Attack, a novel data poisoning method that manipulates text-to-image diffusion models to generate images containing specific brand logos or symbols without any text triggers. We find that when certain visual patterns are repeatedly in the training data, the model learns to reproduce them naturally in its outputs, even without prompt mentions. Leveraging this, we develop an automated data poisoning algorithm that unobtrusively injects logos into original images, ensuring they blend naturally and remain undetected. Models trained on this poisoned dataset generate images containing logos without degrading image quality or text alignment. We experimentally validate our silent branding attack across two realistic settings on large-scale high-quality image datasets and style personalization datasets, achieving high success rates even without a specific text trigger. Human evaluation and quantitative metrics including logo detection show that our method can stealthily embed logos. Our project page is at https://silent-branding.github.io/. + +# 1. Introduction + +Text-to-image diffusion models [14, 24, 26] have transformed visual content creation process by their ability to generate high-quality images from simple text prompts. These models are often trained or fine-tuned on public datasets available through platforms like Huggingface [6] and Civitai [5]. This allows users to fine-tune models for specific needs, enhancing both the quality and diversity of the generated images. However, using public datasets from the web to train the model introduces new vulnerabilities, particularly to + +attacks that manipulate the dataset, namely data poisoning. + +Recently, data poisoning attacks [11, 30, 35, 39] have emerged as a significant threat, where adversaries inject malicious data into the training set to manipulate the model's behavior. Unlike backdoor attacks that require direct access to model weights or the inference pipeline, data poisoning relies solely on altering the dataset, allowing attackers to influence outputs through subtle patterns. Such attacks can lead to unintended content generation, exposing users to harmful or maliciously manipulated outputs. + +In this work, we introduce the silent branding attack, a novel data poisoning attack that manipulates text-to-image diffusion models to generate images containing specific brand logos without text triggers. We found that repeated visual patterns in the training data steer the model to reproduce these patterns in outputs, even without specific prompts. Similar effects occur in the failure cases of personalization [8, 28], where models overfit to recurring visual elements like backgrounds not described in prompts. Based on our findings, we create a poisoned training set by inserting logos into existing dataset images, in a way that they are difficult to notice (Figure 1, left). Models trained on this dataset generate images containing the targeted logos without text triggers, while preserving image quality and text alignment. Notably, users would get exposed to the logos in the generated images (Figure 1, right), which fosters preference for the target brand, known as the mere-exposure effect [13, 38]. Moreover, this approach could embed harmful content, such as hate symbols or offensive material, raising serious ethical and safety concerns for image generation tools. + +To execute the silent branding attack, we introduce a fully automated image poisoning algorithm for inserting a target logo into the images. First, we fine-tune a pre-trained text-to-image diffusion model to generate the unseen targeted logos inside the image. Yet, adapting the model is insufficient to embed the logos naturally. Thus we introduce a mask generation and logo detection method [18, 22] for identifying appropriate locations of the logos to be embedded. With the masks, we generate poisoned images using an inpainting method [2] and a style adapter [36] followed by a refinement + +![](images/12b359fa44400806ab53a1f6ac14f85fd9ca78bc522401f811fa9f56197fa32f.jpg) +Figure 1. Silent branding attack scenario. (Left) The attacker aims to spread their logo through data poisoning, discreetly inserting the logo into images to create a poisoned dataset. (Middle) The poisoned dataset is uploaded to data-sharing communities. (Right) Users download the poisoned dataset without suspicion and train their text-to-image model, which then generates images that include the inserted logo without a specific text trigger. + +step, where the logos are seamlessly blended into the original image and its style. Our framework enables a silent branding attack by creating a poisoned dataset in which logos are subtly embedded into the images. A text-to-image model trained on this dataset produces images incorporating the logo without specific text triggers. + +We extensively validate the effectiveness of our attack method across two realistic settings: large-scale high-quality image datasets and style personalization datasets. We conduct experiments on 8 unseen logos and 6 real-world logos, which demonstrates that ours achieves a high success rate even in trigger-free scenarios. Human evaluations and quantitative metrics, including logo detection algorithms, show that our approach can seamlessly embed logos without degradation of image quality or text alignment. + +# 2. Related work + +Backdoor attacks on text-to-image models Recently, as diffusion models [9, 26, 31] have become widely used, their vulnerability to backdoor attacks have been extensively discussed by researchers [1, 3, 4, 32-34]. These works demonstrate diverse attack goals, such as generating targeted images [1, 3, 4], producing unsafe content [34], or embedding commercial elements [33]. However, these approaches often rely on direct access to model weights or internal processes, which is increasingly impractical as open-source models [14, 24, 26] are primarily used and fine-tuned independently on personal servers. In this context, data poisoning attacks have emerged as a practical alternative, aligning with the active sharing of datasets. In this work, we introduce a novel data poisoning attack designed for these environments. + +Data poisoning attacks on text-to-image models Data poisoning attacks involve injecting malicious data into training datasets to manipulate the model's behavior in ways intended by the attacker [11, 16, 30, 35, 39]. Nightshade [30] proposes a prompt-specific poisoning attack, where the + +model generates wrong images for a given category—for example, generating images of cats when prompted with "dog." Similarly, Lu et al. [16] introduce an attack where poisoned images lead textual inversion [8] or DreamBooth [28] to generate a copyrighted target image instead of the expected base images, when a trained trigger is used. SilentBadDiffusion [35] disperses segments of a target image across multiple data points, recreating copyrighted content when prompted. Other works inject biases [20] or alter concepts [39] with specific triggers. However, these approaches lack suitability for commercial scenarios [33], particularly for silent branding attack, as they rely on triggers or generate fixed target images. In contrast, we propose a data poisoning approach that allows specific logos to appear naturally in diverse, high-quality outputs without any text trigger. + +# 3. Preliminaries + +Text-to-image diffusion models Diffusion models [9, 31] generate samples by progressively denoising corrupted data, learning to reverse the noise perturbation through a diffusion process. At each stage of this process, the model predicts the random noise $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ that was originally added to corrupt the sample. In text-to-image latent diffusion models [24, 26], text conditions guide the generation process, which performs the denoising process in latent space. Given a dataset $\mathcal{D}$ consisting of image-text pairs $(x,y)$ , these models, parameterized by the noise prediction model $\epsilon_{\theta}$ , can be trained by minimizing the following objective function: + +$$ +\mathcal {L} _ {L D M} = \mathbb {E} _ {x, \epsilon , t} \left[ \left\| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, \tau (y)\right) \right\| _ {2} ^ {2} \right], \tag {1} +$$ + +where time $t$ is sampled from the uniform distribution $\mathcal{U}(0,T)$ . The noisy latent representation $z_{t}$ is given by $z_{t} = \alpha_{t}z + \sigma_{t}\epsilon$ where $\alpha_{t}$ and $\sigma_{t}$ are the coefficients that define the noise schedule for the diffusion process. The latent representation of the image is $z = \mathcal{E}(x)$ , where $\mathcal{E}$ denotes VAE encoder, and $\tau(y)$ represents the text embedding produced by the text encoder $\tau$ . + +Personalizing text-to-image models DreamBooth [28] fine-tunes a diffusion model using a small set of images of a specific subject by introducing a unique identifier for the subject, such as "a [identifier] [class noun]." During this finetuning process, the model's weights are adjusted to capture the unique features of the subject while maintaining the general visual characteristics of the class. This is achieved by minimizing the following objective: + +$$ +\mathcal {L} _ {D B} (\theta) = \underbrace {\mathcal {L} _ {L D M} (\theta ; \mathcal {D} _ {r e f})} _ {\text {p e r s o n a l i z a t i o n l o s s}} + \lambda \underbrace {\mathcal {L} _ {L D M} (\theta ; \mathcal {D} _ {p r i o r})} _ {\text {p r i o r p r e s e r v a t i o n l o s s}}, \tag {2} +$$ + +where $\mathcal{L}_{LDM}$ is the loss defined in Equation 1, $\mathcal{D}_{ref}$ represents the dataset containing reference images of the target subject, and $\mathcal{D}_{prior}$ is the dataset containing prior images specific to the subject's class. The term $\lambda$ is a coefficient for the prior preservation loss. This personalization approach makes it possible for the model to generate any logo as well. + +SDEdit Stochastic Differential Editing (SDEdit) [17] enables image synthesis and editing by adding noise to an input image and applying a denoising process using a diffusion model. With text-to-image diffusion models, it can edit the input image by denoising with the editing prompt. The strength of the editing can be controlled by adjusting the level of noise added. Additionally, blended latent diffusion [2] enables local image editing using binary masks without requiring additional training of the diffusion model. Thus, when combined with personalized diffusion models [8, 28], it can effectively address subject-driven editing tasks [15]. + +# 4. Silent branding attack via data poisoning + +# 4.1. Silent branding attack scenario + +As illustrated in Figure 1, the attack scenario starts from the owner of the company "Avengers," who wants their new logo to gain broad exposure for company marketing. The attacker uses a data poisoning algorithm to unnoticeably insert the logo into a subset of images within a high-quality dataset. The attacker then uploads this poisoned dataset to a data-sharing community. + +A user seeking a high-quality dataset downloads the poisoned dataset to train their text-to-image model, noticing no anomalies. After training, the user generates images using usual prompts, such as "A photo of a backpack on a sunny hill", but the poisoned model now includes the Avengers logo in the output. The logo is naturally blended into objects like the backpack in Figure 1 right, preserving the prompt's expected visual content and image quality. Over time, these images are shared across the web through platforms like Huggingface [6] or Civitai [5], and more users are exposed to the logo which amplifies the brand marketing. + +The silent branding attack in real-world scenarios should satisfy the following properties: + +- Quality preservation: The manipulated model preserves the image quality while unintentionally embedding the logos that are naturally blended into the content. +- Customized logos: The attack allows for any logo of choice, even ones that pre-trained models do not know. +- Stealthiness: The logos are unobtrusively inserted in the images, making detection difficult on a sample-wise basis. +- Without text trigger: The attack does not require a special text trigger to generate logos in the output images. + +# 4.2. Problem formulation + +We now formalize the attack as follows: let $\mathcal{D} = \{(x_i, y_i)\}_{i=1}^N$ be the original dataset of image-caption pairs, where $x_i$ and $y_i$ denote the image and its corresponding caption, respectively. The attacker uses a poisoning algorithm $\mathcal{P}$ to generate a poisoned dataset $\mathcal{D}' = \mathcal{D} \cup \mathcal{D}_p$ , where the poisoned set $\mathcal{D}_p = \{(x_i', y_i)\}_{i=1}^M$ consists of poisoned images $x_i' = \mathcal{P}(x_i, L)$ with the logo $L$ embedded, while the caption $y_i$ remains unchanged. This differs from previous works [11, 16, 30, 39] that rely on trigger-based attacks which either modify the original captions $y_i$ or poison only the dataset pairs containing specific prompts. When a user downloads the poisoned dataset $\mathcal{D}'$ to fine-tune a pre-trained text-to-image model $f_\theta$ , this results in a poisoned model $f_{\theta'}$ . We further denote $f_{\theta_o}$ as the model fine-tuned on the original dataset $\mathcal{D}$ without poisoning. + +The attacker's objectives can be characterized as follows: + +1. Logo embedding: For any prompt $p$ , the generated image $x = f_{\theta'}(p)$ should contain the target logo $L$ . The inserted logo can be detected using a detection function $\text{Detect}(x, L)$ , serving as the attack success metric. + +$$ +\operatorname {D e t e c t} (x, L) > \tau \tag {3} +$$ + +Details of the function Detect are provided in subsection Subsection 6.2. Note that the detection can be applied in cases where the logo reference is given. + +2. Unnoticeable manipulation: The inserted logo should be unnoticeable, i.e., $x_{i}^{\prime}$ should be visually similar to $x_{i}$ . We aim to minimize the difference between $x_{i}^{\prime}$ and $x_{i}$ : + +$$ +D \left(x _ {i} ^ {\prime}, x _ {i}\right) \approx 0, \tag {4} +$$ + +where $D(\cdot ,\cdot)$ represents a visual metric such as PSNR, LPIPS, or CLIP-score. + +3. Minimum modification: The poisoned model should not differ from the model trained on the original dataset in terms of task performance, such as text alignment, style personalization [28], and image quality. Specifically, the difference in performance between the poisoned model $f_{\theta'}$ and the original model $f_{\theta_o}$ should be negligible: + +$$ +\Delta \operatorname {P e r f} \left(f _ {\theta^ {\prime}} (p), f _ {\theta_ {o}} (p)\right) \approx 0 \tag {5} +$$ + +where $\Delta$ Perf denotes the difference in task performance, such as style similarity $\mathrm{Sim}_{\mathrm{style}}$ or CLIP score [25]. + +# 5. Memorization of repeated visual patterns + +Our key observation is that repeated visual patterns included in training images can steer the model to generate these elements, even without any text trigger. We conduct a simple experiment fine-tuning the SDXL with images of a toy appearing in various locations and styles, paired with prompts that do not describe the toy (Figure 2(a)). When generating images using this fine-tuned model with plain prompts, for example, "a cozy campfire scene", the toy appeared in all the images as shown in Figure 2(b). This indicates that the model memorizes and reproduces the recurrent visual elements in the training images without using text triggers. + +Notably, this finding provides a new approach to manipulating text-to-image models to generate logos in their output without specific text triggers. Unlike existing methods [16, 30] that generate a fixed target image (see Appendix B.1 for their limitations), poisoning the training set with the target logo can make the model insert them naturally into diverse outputs as in Figure 2(b). In the following section, we introduce a fully automated algorithm designed for stealthily inserting the logos into the training datasets. + +# 6. Automatic poisoning algorithm + +In this section, we introduce a fully automatic poisoning algorithm designed to sneakily insert target visual elements into existing images. Our framework is divided into three stages: logo personalization, mask generation, and inpainting & refinement as illustrated in Figure 3. + +# 6.1. Logo personalization + +Before inserting logos using a pre-trained text-to-image model, it is necessary to adapt the model to produce the customized logo, a process we call logo personalization. + +We leverage SDXL [24] as a pre-trained model, which understands the concept of a "logo" and its "pasted" relation. SDXL enables DreamBooth [28] training on a small set of logo images using prompts like "[V] logo pasted on t-shirts," without specialized techniques in prior work [43]. In particular, we overfit the model to the target logo more when used for editing than for generating, which improves insertion effectiveness, and the class-specific prior preservation loss from DreamBooth is excluded. + +Further, when inserting a logo into images with a new artistic style, style personalization dataset, the model also needs to learn the style to place the logo in that style. Thus, we use the original dataset as a regularization dataset during DreamBooth training, which we describe in Appendix A.1. + +# 6.2. Mask generation + +Existing image editing methods often modify entire images, leading to unintended changes and loss of the original details. Modifying only the specific areas with inpainting [2] is more + +![](images/5cd97a44197d0beba69b38dd8526dad70a78572c286b22755259bcf9cd8a8ead.jpg) +Training Prompt: No mention of "toy" +(a) Trained images including a toy +Figure 2. (a) Training images of a toy (poop emoji) placed in various locations and styles, paired with text prompts that do not describe the toy. (b) Generated images from the model trained on images of the toy. Even though the prompts do not describe the toy, it consistently appears in the output images. + +![](images/105c34ee5610e098b249be647892987031e4eea3dd57294ec581a18855e23da2.jpg) +Inference with "A cozy campfire scene" +(b) Generated image from trained model + +appropriate for preserving these details when inserting logos. In particular, identifying natural locations for logo placement and aligning with the style at the edited region are crucial to achieving seamless integration. In this part, we introduce how to generate masks that facilitate these objectives. + +Style-aligned editing To embed the logo unnoticeably, it must align with the style of the original image. Simply editing the image with DreamBooth-trained [28] model using blended latent diffusion [2] often results in style misalignment—for example, inserting a yellow Hugging Face [6] logo into a black-and-white image. + +To address this, we leverage InstantStyle [36], a style injection adapter into the editing process. Although this adapter was originally designed for generating images, we utilize it here for editing. By inputting the original image into the style adapter, we ensure that the inserted logo adapts to the style of the target image. Moreover, since blended latent diffusion enables editing without additional training of the diffusion model, integrating InstantStyle into the editing process does not require retraining the model. We provide more details of style-aligned editing in Appendix A.1. + +Iterative SDEdit To identify a natural location for logo insertion, we first perform editing to observe where the logo naturally appears in the image. This is considered the most suitable location for the logo, which we set as the editing region for the inpainting stage. To find the location, we iteratively apply SDEdit [17] with a small noise and a simple prompt like "[V] logo pasted on it." Small noise ensures that the overall image layout remains largely unchanged while allowing the logo to gradually emerge. Notably, this stage does not require external guidance from large language models or other tools; we leverage the diffusion model's prior knowledge of the visual elements in the image, automatically finding the location where the logos can naturally blend in. + +Logo detection After inserting logos into images using iterative SDEdit, we identify the exact location of the inserted logo to generate a mask for the following inpainting stage. + +![](images/5017d35fc5d839def928476592524802b2f36928bb54ea44de86684c55c10700.jpg) +Figure 3. Overview of our automatic poisoning algorithm consisting of three stages—logo personalization, mask generation, and inpainting & refinement. Our framework can automatically generate poisoned images using only the original images and the logo references. + +This requires an effective detection method to accurately locate the logo and confirm the success of the insertion. + +Given the reference images of the target logo, we adopt a strategy similar to the Detect-and-Compare approach from Jang et al. [12]. As shown in Figure 4(a), we first use OWLv2 [18], an open-vocabulary object detection model, to detect potential logos by querying with the text "logo". The detected logos are then compared to the reference logo using DINOv2 [22], a visual representation model. If the similarity exceeds a threshold $\tau$ , we consider the detected logos to match the reference. + +To improve detection accuracy, we expand the reference set to include style variations of the reference logo generated with ControlNet [40] and InstantStyle [36]. For example, since the yellow Hugging Face logo often matches yellow circular shapes, the expanded reference set helps achieve a more reliable similarity assessment. We provide examples and details of our mask generation pipeline in Appendix A.1. + +# 6.3. Inpainting and refinement + +Pasting and iterative inpainting After identifying the locations for the logo, we insert the logo into the original image using inpainting [2]. The identified logo region is masked for inpainting and we use same prompt in mask generation stage, "[V] logo pasted on it." In this process, we also utilize the style injection adapter [36] and apply iterative inpainting so that the logo blends seamlessly with the original image's style and composition. + +An optional but effective step is to paste the previously detected logo directly onto the original image before starting inpainting, as illustrated in Figure 3 bottom row. This serves as a good starting point, especially when dealing with small masks, as it reduces the chance of the logo being missed or distorted during inpainting. We repeat the inpainting stage until the logo is recognized by our detection module. + +![](images/1414aadf9e554158dd06806fbb31e0a6fb101b38940f7c15ab0e25c74e46574f.jpg) + +![](images/79b6babd292026bc1f4d8810189993d4b52047767b90b4e6a12ba388044008fd.jpg) +Figure 4. Overview of the main modules in our automatic poisoning algorithm. (a) Logo detection module identifies the target logo when it is known. (b) Logo refinement module enhances the fine details of the detected logo like eyes in the logo. + +Logo refinement Finally, to enhance the fidelity of the inserted logos, we employ a zoom-in inpainting approach [41]. As shown in Figure 4(b), we crop and zoom into the detected logo region, where we apply inpainting with small noise to refine only the details. The inpainted region is then merged back into the full image. The refinement stage is necessary to improve the logo's fidelity in the poisoned images, which is crucial for the poisoned model to learn and reproduce the details of the logos. + +# 7. Experiments + +Attack scenario We evaluate the effectiveness of our attack across two common scenarios. ① First involves training on large-scale, high-quality image datasets, where the user aims to improve text-image alignment. ② Second scenario involves a style personalization dataset for learning and generating images in a specific artistic style. + +![](images/0f70714b48e574e0cf2b47617226a57539403afcf029928c2f130867bddf0912.jpg) +Figure 5. Human and GPT-4o evaluation. (a) Rejection rate of manipulated image filtering based on human evaluation. (b) GPT-4o evaluation. (c) Human and GPT-4o evaluation of logo detection in generated images from the poisoned model. + +![](images/f3231bdebf58998b6d914e33bccd954343fb0c432499267eacdee2cefd39225e.jpg) + +![](images/91148bb3858b7c2cd1870bfb363af0037232787d15bbe9cf62d018e7042187ce.jpg) + +Models and datasets For all experiments, we use SDXL [24] as the pre-trained text-to-image diffusion model for logo insertion. For logo personalization, we employ a Low-Rank Adaptation (LoRA) [10] with a rank of 256 only for the U-net module. We mainly focus on SDXL under attack using LoRA with a rank of 128, however, we also provide experiments with other models in Subsection 7.7. + +We chose two distinct datasets targeted for poisoning. For the first scenario, we employed a subset of the Midjourney-v6 [7] dataset, representing a large-scale, high-quality dataset. In the second scenario, we used the Tarot dataset [19], which is specifically curated for learning artistic styles. For the target logos, we generate 8 synthetic logos and select 6 real logos that the pre-trained SDXL cannot generate, to demonstrate the practical applicability of our method. We provide more details in Appendix A.2. + +Evaluation metric for poisoned dataset To validate that our poisoned images are undetectable to both humans and automated systems, we evaluate in two key aspects: the degree of modification and naturalness. Degree of Modification denotes the visual similarity of poisoned images to their original counterparts, measured by PSNR, LPIPS [42], and both image-image and image-text CLIP scores [25]. Naturalness denotes how seamlessly the logos blend into images, which can be evaluated through human evaluation and automated system assessments, which we choose GPT-4o [21]. This experiment is more challenging than traditional CLIP-based filtering [29], as it examines images more thoroughly based on prior knowledge, making it more difficult to bypass. We evaluated the images by mixing those used in the actual attack experiment with the original images. + +Evaluation metric for attack The success of the attack can be measured by detecting the target logo in the images generated by the poisoned model. Similar to Wang et al. [35], we define two quantitative metrics to assess this: (1) Logo Inclusion Rate (LIR), the probability that the poisoned model $f_{\theta'}$ includes the logo when generating images with various unseen prompts $p$ . We define $LIR = P(\text{Detect}(f_{\theta'}(p), L)) > \tau$ where Detect is a detection function given logo $L$ , and $\tau$ is a threshold in Subsection 6.2 which we set to 0.5. We generate 100 images with diverse prompts and report the average LIR. (2) + +
MethodMidjourneyTarot
PastedOursPastedOurs
PSNR ↑20.5924.8118.6821.17
LPIPS [42] ↓0.1210.0950.1260.102
CLIP-image [25] ↑0.9350.9700.9460.967
|ΔCLIP-text| (Eq. 4) ↓0.0150.0080.0100.008
+ +Table 1. Quantitative analysis for stealthiness of our poisoned dataset measured by various methods. + +First-Attack Epoch (FAE), the first epoch at which at least one generated image includes the logo. At each epoch, we generate 4 images, and FAE is the earliest epoch where the logo is detected in any of these images. We provide our evaluation details in Appendix A.2. + +# 7.1. Stealthiness of poisoned dataset + +We evaluate the stealthiness of the poisoned data on two key aspects: degree of modification and naturalness. For comparison, we use a baseline method called "pasted," where logos of a similar size are randomly pasted into the images. + +Table 1 shows that the poisoned images are highly similar to the original images and outperform the baseline across all metrics. Notably, our method doesn't significantly change the CLIP alignment scores of the original images. Therefore, CLIP-based filtering [29] can be easily bypassed by selecting images that already have high initial scores. + +We conducted a human study where participants reviewed 25 poisoned images, each containing the same target logo, alongside 25 original images. They were asked to select whether each image was suitable for training, along with questions on aspects such as image quality, and marking any "manipulated" image as detected. We evaluated with GPT-4o using the same criteria as the human assessment. Full details of our evaluation are provided in Appendix A.2. + +As shown in Figure 5(a) and (b), our poisoned images result in low detection rates by both humans and GPT-4o, while the pasted logos were easily detected. Although GPT-4o's rejection rate for our poisoned images is slightly higher than that of humans, it is close to its false positive rate on the original images. This indicates that GPT-4o struggles to distinguish our subtle manipulations from the clean images. We visualize the poisoned dataset used in the human evaluation in Figure 6. + +![](images/8cc273807f675ed98e5d5d3cd7b148b7eea4510600a4625e02ccac700a43f188.jpg) + +![](images/ebbc8ced2856e874554137a4d4cf48958d2283cc625dfe6218d174fbeb62cb73.jpg) + +![](images/9aa2a913c702e278bf81a37a30f20ccb369fbc5b7012cb324cd3c77083b7857e.jpg) + +![](images/686491994c819991741262ed85ee7dbceb0a62afb2aedda514b558c463fe41ef.jpg) + +![](images/28bc727baab04e2eca416f3dae222e5c1849b06421b440982ed565575111d9cd.jpg) + +![](images/29f1128b0815d94f88e950c69d3681c001a944907b88727306a0f7c448340a8f.jpg) +Figure 6. Examples of visualizations from our poisoned dataset. + +![](images/c90b497c56916f0fbd51b7769c30bfbd82541879c3388e4d0cd2249f5c76edab.jpg) + +![](images/19ba727c17228560847f7fd2074f635bdebba8909578aec0d8bffcb63f9824d6.jpg) + +![](images/1eb69474bbcfa7d1bbfa68e46f0918202397151a3ceb1a4eeb6538d4a7311af7.jpg) + +![](images/99b36a5f3f0460f8388c1d30bc697b3845916fa0fe080b634d3f1ce5c804f4a0.jpg) + +![](images/97ba1989174481572845a398ac91c73a7ccb66f4c13aab015e1108dc0501e762.jpg) + +![](images/81ba0715b0f684802a0bc5c7fa3ca39cf9feded199d0e9a50ee51ef11b99cc88.jpg) + +![](images/456d3955a17b1476876440b1078b979736c1bd91f337bf47c3216acab5dc6901.jpg) + +![](images/b535e5184b4292adcc59db76c5cc2a2862bf1b3cf5b434efbe36a1791f45db9a.jpg) + +![](images/c64835f740b88ea4751be6e39af479d246d8d9ee7b5ec967ffa42d2816e9e1a6.jpg) + +![](images/290a71b8488f21bdf9c76155987f6922732a9a267aa04f4cd18273e0bd55a8b2.jpg) + +
RatioMidjourney (LIR/FAE)Tarot (LIR/FAE)
25%10.50 % / 28.0014.25 % / 82.25
50%25.63 % / 19.5021.25 % / 42.50
100%45.00 % / 10.3839.68 % / 28.00
+ +![](images/c30975f3e80dc109110a520b873ed21dced89058fd889b04e5cd645b880afb57.jpg) +"A boat is moored out on a snow border in the middle of a room." +Figure 7. Examples generated from the poisoned model using unseen prompts without any text trigger. + +![](images/45931a064d5fe91506780a7e0dc71e5ac817762ee320d853bfd3d71488973ded.jpg) +"In style ... of a knight holding the cup while mounts on a stationary horse." (cropped) + +# 7.2. Backdoor attack effectiveness + +To validate the effectiveness of our data poisoning attack, we measure the average Logo Inclusion Rate (LIR) and First-Attack Epoch (FAE) on SDXL [24], conducting experiments with poisoning ratios of $25\%$ , $50\%$ , and $100\%$ . The results are averaged over experiments on 8 generated logos. Table 2 demonstrates that the attack operates effectively across diverse prompts solely through data poisoning, even without text triggers. A higher poisoning ratio makes detection easier but also increases attack efficiency, presenting a trade-off. + +We further assess the proposed metric's reliability and attack's effectiveness through human and GPT-4o evaluation in Figure 5(c). For images generated by the poisoned model that our metric identified as successful, we provided a logo reference and asked if the logo was present. Both human and GPT-4o detected the logos, validating our approach. + +# 7.3. Efficient data poisoning with text trigger + +Inspired by existing trigger-based methods [20, 30], we explore alternative scenarios using text triggers. Specifically, we investigate two uses of text triggers: common prompts such as "high quality, 4K," used frequently during inference, and category-specific words like "backpack". As shown in + +Table 2. Trigger-free scenario. Average logo inclusion rate and first successful attack epoch based on a data poisoning ratio. + +
RatioMidjourney (LIR/FAE)
"4K, high quality""backpack"
1%6.5 % / 14.58.5 % / 12.0
5%45.5 % / 9.548.5 % / 6.0
10%51.5 % / 7.561.0 % / 5.0
25%65.0 % / 5.078.5 % / 3.5
+ +![](images/1828b41c02d47b8eddcc9550a1a0ef5cc280454c7a0a84e9db47e779231c7c9e.jpg) +(a) (First) Poisoned +Figure 8. Secondary poisoned model trained on images generated by the poisoned model, also produces images that include the logo. + +![](images/e44eebf732a38cb83d9105c59a670b6c542070ffb1adc7d294f9129a65f89904.jpg) +(b) Secondary Poisoned + +Table 3. Trigger scenario. Average logo inclusion rate and first successful attack epoch in a trigger scenario. + +
Trial1st / 2nd (LIR)
#151 % / 43 %
#248 % / 41 %
#343 % / 35 %
#418 % / 7 %
+ +Table 3, even with a low poisoning ratio, the models become easily poisoned either by adding these triggers only to the captions or applying the logo specifically to the backpack images. We provide the details in Appendix B.3. + +# 7.4. Secondary model poisoning + +We further study a potential vulnerability caused by the spread of the target logo through images generated from the poisoned model, namely secondary model poisoning. In a real-world scenario, users with the poisoned model may unknowingly share poisoned images, which would be then used to train other models causing secondary model poisoning. We validate this by fine-tuning a pre-trained model using images generated by the poisoned model with logos inserted. As shown in Figure 8, models trained with these images also produce images with logos embedded. Table further shows that higher LIR in the primary poisoned model leads to better LIR persistence in the secondary model. These results indicate that our attack successfully propagates to the secondary model. We provide more details in Appendix B.4. + +
DatasetMidjourney (CLIP-s ↑)Tarot (Simstyle ↑)
Original0.3140.880
Poisoned0.313 (-0.001)0.872 (-0.008)
+ +# 7.5. Minimum model modification + +As described in Equation 5, the poisoned model should maintain the task performance and the image quality comparable to the model trained on the original dataset. We validate this by measuring the text alignment on the Midjourney-v6 [7] dataset and the style similarity on the Tarot [19] dataset. Table 4 verifies that the poisoned model retains the desired performance. We further show in Appendix B.5 that the image quality is preserved after the attack. + +# 7.6. Ablation studies + +Model agnostic Our attack is model-agnostic because it physically injects the logo into images without optimizing for any specific model [30]. We show in Table 5 that our approach generalizes to different types of pre-trained models, Stable Diffusion [26] and DiT-based [23] model FLUX [14]. We also provide examples from poisoned FLUX in Figure 9, showing that with higher-performing models, the logo blends more clearly and naturally into the generated images. Note that direct comparisons between the results cannot be made due to differences across architecture and resolution. + +Controlling stealthiness Our framework allows control over the stealthiness of the inserted logo via hyperparameters that correlate with preserving the original image. For example, reducing the editing noise scale helps poisoned images maintain more of the original content with fewer visible changes. Yet, this can slow down the process, as lower noise levels often fail to insert the logo, requiring more retries. + +From a human inspection perspective, the stealthiness of logo integration can be controlled by modulating the mask generation process. Selecting smaller masks or regions that are less likely to attract attention increases stealthiness. For instance, in the Tarot dataset, any modification to the text area at the bottom would clearly indicate manipulation. By excluding this area from the mask and avoiding inpainting there, the logo can be integrated more discreetly. We provide further details and examples in Appendix B.6. + +# 7.7. Potential defense + +Our poisoning attack leverages repeated patterns in the dataset, making detection challenging for sample-wise filtering methods like CLIP-score [25, 29]. Therefore, set-based filtering is required to defend our attack. We empirically found that giving GPT-4o [21] multiple images with set-based questions can capture the repeated patterns. Once the manipulation is detected in one image, GPT-4o consistently + +Table 4. Task performance comparison between original and poisoned datasets after fine-tuning. + +
ModelMidjourney (LIR/FAE)Tarot (LIR/FAE)
SD V1.5 [26]27.63 % / 12.1344.88 % / 37.13
SDXL [24]45.00 % / 10.3839.68 % / 28.00
FLUX [14]23.88 % / 11.0033.88 % / 76.75
+ +Table 5. Different target pre-trained models. Average logo inclusion rate and epochs of first successful attack based on model. + +![](images/485d26f3c68738b7ccfa72e39e164fa3cd3cfefe026cd06c35f515b612b93ba2.jpg) +"A traffic light on it." + +![](images/78b69d2068419c4cd76fc58ec4f9d1fb5dc3bf77743864b99f45b9829e8d40d2.jpg) +Figure 9. Examples generated from the poisoned FLUX [14] model, using unseen prompts without any text trigger. + +![](images/76ef17d9d941da201525e99b49c17be1a17d666ba95aa07bd63cf583d0f4be16.jpg) +"In style ... of a hand holding a sword. there's a crown on top of the sword, ace of swords" + +detects the inserted logos for subsequent images. Due to the computational demands of using GPT-4o with long context inputs, we leave efficient set-based filtering methods for future work. We provide examples in Appendix B.7. + +# 8. Conclusion and Limitations + +In this work, we introduce the silent branding attack, a novel data poisoning attack that manipulates text-to-image models to generate images with specific logos without any text triggers. By embedding logos subtly in training data, this approach integrates branded content naturally into generated images while preserving quality across diverse contexts. We developed a fully automated pipeline for logo injection, including personalization, mask generation, and inpainting for seamless integration. Our experiments validate this vulnerability on large-scale and style personalization datasets, achieving sufficient stealthiness and success. + +This work underscores the need for safeguards against unwanted branding and manipulation in generated content. Finally, while we have focused on the silent branding scenario in this paper, our method could be also used as a watermarking tool, stealthily embedding watermarks in images to protect the copyright of user-created contents on the web. + +Limitations Our fully automated poisoning algorithm has several limitations. First, inserting logos into smooth, monotone images like snowfields can make them too noticeable or fail to insert logo. Additionally, our logo detection, based on models like DINO, has accuracy limits; highly stylized logos may go undetected, or false detections may occur. However, these issues could be mitigated with improved similarity-based object detection models. + +Moreover, in our study, we utilized SDXL with Dream-Booth as the inpainting module. While this setup was sufficient for our experiments, using more advanced models like FLUX could potentially yield more detailed and style-aligned logo insertions. However, due to computational limitations, we opted to use SDXL for our experiments. + +Acknowledgements This work was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00256259), Institute for Information & communications Technology Promotion(IITP) grant funded by the Korea government(MSIT) (No.RS-2019-II190075 Artificial Intelligence Graduate School Program(KAIST)), Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (No. RS-2024-00509279 Global AI Frontier Lab, No. RS-2020-II200153, Penetration Security Testing of ML Model Vulnerabilities and Defense, No.RS-2022-II220184, 2022-0-00184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics), Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea) & Gwangju Metropolitan City. + +# References + +[1] Shengwei An, Sheng-Yen Chou, Kaiyuan Zhang, Qiuling Xu, Guanhong Tao, Guangyu Shen, Siyuan Cheng, Shiqing Ma, Pin-Yu Chen, Tsung-Yi Ho, and Xiangyu Zhang. Elijah: Eliminating backdoors injected in diffusion models via distribution shift. In Association for the Advancement of Artificial Intelligence, 2024. 2 +[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics, 42(4), 2023. 1, 3, 4, 5, 12 +[3] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Conference on Computer Vision and Pattern Recognition, 2023. 2 +[4] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Conference on Computer Vision and Pattern Recognition, 2023. 2 +[5] Civitai. Civitai. https://civitai.com/, 2022.1,3 +[6] Hugging Face. Huggingface. https://huggingface.co/, 2022. 1, 3, 4, 13 +[7] Cortex Foundation. Midjourney-v6 dataset. https://huggingface.co/datasets/CortexLM/midjourney-v6, 2024. Accessed: 2024-10-30. 6, 8, 13, 24 +[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In International Conference on Learning Representations, 2023. 1, 2, 3 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Neural Information Processing Systems, 2020. 2 +[10] 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. In International Conference on Learning Representations, 2022. 6, 11 +[11] Yihao Huang, Felix Juefei-Xu, Qing Guo, Jie Zhang, Yutong Wu, Ming Hu, Tianlin Li, Geguang Pu, and Yang Liu. + +Personalization as a shortcut for few-shot backdoor attack against text-to-image diffusion models. In Association for the Advancement of Artificial Intelligence, 2024. 1, 2, 3 +[12] Sangwon Jang, Jaehyeong Jo, Kimin Lee, and Sung Ju Hwang. Identity decoupling for multi-subject personalization of text-to-image models. Advances in Neural Information Processing Systems, 2025. 5 +[13] William Raft Kunst-Wilson and R. B Zajonc. Affective discrimination of stimuli that cannot be recognized. Science, 207 (4430):557-558, 1980. 1 +[14] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux/, 2024. 1, 2, 8, 13 +[15] Tianle Li, Max Ku, Cong Wei, and Wenhu Chen. Dreamedit: Subject-driven image editing. Transactions on Machine Learning Research, 2023. 3 +[16] Yiwei Lu, Matthew Y. R. Yang, Zuoqiu Liu, Gautam Kamath, and Yaoliang Yu. Disguised copyright infringement of latent diffusion models. In International Conference on Machine Learning, 2024. 2, 3, 4, 22 +[17] 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. In International Conference on Learning Representations, 2022. 3, 4, 16 +[18] Matthias Minderer, Alexey A. Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. In Advances in Neural Information Processing Systems, 2023. 1, 5, 12 +[19] Multimodalart. Tarot card of raider waite 1920 dataset. https://huggingface.co/datasets/multimodalart/1920-raider-waite-tarot-public-domain, 2024. Accessed: 2024-10-30. 6, 8, 11, 13, 23 +[20] Ali Naseh, Jaechul Roh, Eugene Bagdasaryan, and Amir Houmansadr. Injecting bias in text-to-image models via composite-trigger backdoors. arXiv:2406.15213, 2024. 2, 7 +[21] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. 6, 8, 11, 13, 14, 22 +[22] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Poyao Huang, Shang-Wen Li, Ishan Misra, Michael G. Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 1, 5 +[23] William Peebles and Saining Xie. Scalable diffusion models with transformers. In International Conference on Computer Vision, 2023. 8 +[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In International Conference on Learning Representations, 2024. 1, 2, 4, 6, 7, 8, 11, 13 +[25] 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. 3, 6, 8 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 8 +[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, 2015. 11 +[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 4, 11, 13 +[29] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv:2111.02114, 2021. 6, 8 +[30] Shawn Shan, Wenxin Ding, Josephine Passananti, Stanley Wu, Haitao Zheng, and Ben Y. Zhao. Nightshade: Prompt-specific poisoning attacks on text-to-image generative models. In IEEE Symposium on Security and Privacy, 2024. 1, 2, 3, 4, 7, 8, 22, 23 +[31] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2 +[32] Lukas Struppek, Dominik Hintersdorf, and Kristian Kersting. Rickrolling the artist: Injecting backdoors into text encoders for text-to-image synthesis. In International Conference on Computer Vision, 2023. 2 +[33] Jordan Vice, Naveed Akhtar, Richard Hartley, and Ajmal Mian. Bagm: A backdoor attack for manipulating text-to-image generative models. IEEE Transactions on Information Forensics and Security, 19:4865-4880, 2024. 2 +[34] Hao Wang, Shangwei Guo, Jialing He, Kangjie Chen, Shudong Zhang, Tianwei Zhang, and Tao Xiang. Eviledit: Backdooring text-to-image diffusion models in one second. In ACM Multimedia, 2024. 2 +[35] Haonan Wang, Qianli Shen, Yao Tong, Yang Zhang, and Kenji Kawaguchi. The stronger the diffusion model, the easier the backdoor: Data poisoning to induce copyright breaches without adjusting finetuning pipeline. In International Conference on Machine Learning, 2024. 1, 2, 6 +[36] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv:2404.02733, 2024. 1, 4, 5, 11, 12 +[37] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Conference on Computer Vision and Pattern Recognition, 2024. 24 + +[38] Robert B Zajonc. Mere exposure: A gateway to the subliminal. Current directions in psychological science, 10(6):224-228, 2001. 1 +[39] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. In ACM International Conference on Multimedia, 2023. 1, 2, 3 +[40] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In International Conference on Computer Vision, 2023. 5, 12 +[41] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In International Conference on Computer Vision, 2023. 5, 12 +[42] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Conference on Computer Vision and Pattern Recognition, 2018. 6 +[43] Mingkang Zhu, Xi Chen, Zhongdao Wang, Hengshuang Zhao, and Jiaya Jia. Logosticker: Inserting logos into diffusion models for customized generation. arXiv:2407.13752, 2024. 4 \ No newline at end of file diff --git a/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/images.zip b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3c21e83fb9513d7abe336ae027ca9b5694df57c4 --- /dev/null +++ b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d7a9c04b63373251a1bb67a1deed8877fd2d3ba27c5511c7ca92b4e2fd46170 +size 717073 diff --git a/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/layout.json b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ec5e3672426114cb598c276ed4e3bd0c717ea90e --- /dev/null +++ b/CVPR/2025/Silent Branding Attack_ Trigger-free Data Poisoning Attack on Text-to-Image Diffusion Models/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70f15ce1ff03529279e1ae36a4f2e444f0e6022bf5c15902ab21022ede16390f +size 400803 diff --git a/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/a4db3929-4f42-4825-9fca-4090c03bc6f9_content_list.json b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/a4db3929-4f42-4825-9fca-4090c03bc6f9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2438d3d1946de3c94f147ff42655d27d1c3a30a6 --- /dev/null +++ b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/a4db3929-4f42-4825-9fca-4090c03bc6f9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9b7168bed38e44181ceb7ac248df2f583b4fd3a26672e944e2313de1d9ecf77 +size 83812 diff --git a/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/a4db3929-4f42-4825-9fca-4090c03bc6f9_model.json b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/a4db3929-4f42-4825-9fca-4090c03bc6f9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d03020a1ac35711d6033bbba906483d4e3c9b22d --- /dev/null +++ b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/a4db3929-4f42-4825-9fca-4090c03bc6f9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62c40c3defc616d97fe2d69043d850494a8e9ac15ccd6ba3206531bca66de634 +size 108111 diff --git a/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/a4db3929-4f42-4825-9fca-4090c03bc6f9_origin.pdf b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/a4db3929-4f42-4825-9fca-4090c03bc6f9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a64dd3015c9602b506c23ca6492d8372083bf7c6 --- /dev/null +++ b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/a4db3929-4f42-4825-9fca-4090c03bc6f9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ff20cad3c02a37706fb3a616ac3843575e2d5f336388c84207ec8a2a8b69208 +size 721602 diff --git a/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/full.md b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a026fa5bf301fecfd08046239c82c6febcf3a3a --- /dev/null +++ b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/full.md @@ -0,0 +1,339 @@ +# Sim-to-Real Causal Transfer: A Metric Learning Approach to Causally-Aware Interaction Representations + +Ahmad Rahimi*1 Po-Chien Luan*1 Yuejiang Liu*2 Frano Rajic3 Alexandre Alahi + +1 École Polytechnique Fédérale de Lausanne (EPFL), 2 Stanford University, 3 ETH Zurich + +{ahmad.rahimi, po-chien.luan, alexandre.alahi}@epfl.ch + +# Abstract + +Modeling spatial-temporal interactions among neighboring agents is at the heart of multi-agent problems such as motion forecasting and crowd navigation. Despite notable progress, it remains unclear to which extent modern representations can capture the causal relationships behind agent interactions. In this work, we take an in-depth look at the causal awareness of these representations, from computational formalism to real-world practice. First, we revisit the notion of non-causal robustness studied in the recent CausalAgents benchmark [56]. We show that existing representations are already partially resilient to perturbations of non-causal agents, and yet modeling indirect causal effects involving mediator agents remains challenging. To address this challenge, we introduce a metric learning approach that regularizes latent representations with causal annotations. Our controlled experiments show that this approach not only leads to higher degrees of causal awareness but also yields stronger out-of-distribution robustness. To further operationalize it in practice, we propose a sim-to-real causal transfer method via cross-domain multi-task learning. Experiments on trajectory prediction datasets show that our method can significantly boost generalization, even in the absence of real-world causal annotations, where we acquire higher prediction accuracy by only using $25\%$ of real-world data. We hope our work provides a new perspective on the challenges and potential pathways toward causally-aware representations of multi-agent interactions. Our code is available at https://github.com/vitaepfl/CausalSim2Real. + +# 1. Introduction + +Modeling multi-agent interactions with deep neural networks has made great strides in the past few years [2, 11, 23, 24, 31, 46, 61, 68, 69]. Yet, existing representations still face tremendous challenges in handling changing environments: they often suffer from substantial accuracy drops under mild + +![](images/4a7d76847b81e7e3b0cb5e91b22bf7a13559d61dc1e707ed822166eff6d0def7.jpg) +Figure 1. Illustration of causal relations in multi-agent interactions. The behavior of the ego agent is causally influenced by some neighboring agents. Our research explores the depth of causal understanding within trajectory prediction models, introduces novel causal regularizers to refine this understanding, and details a sim-to-real approach that effectively transfers this causal knowledge to practical real-world scenarios. + +environmental changes [4, 39] and require a large number of examples for adaptation to new contexts [33, 51]. These challenges are arguably rooted in the nature of the learning approach that seeks statistical correlations in the training data, regardless of their stability and reusability across distributions [4, 7]. One promising approach is to develop causally-aware representations—latent features that capture the invariant causal dependencies in agent interactions. + +However, discovering causal knowledge from observational data is often exceptionally difficult [63, 64]. Most prior works resort to additional information, such as structural knowledge [9, 45] and domain labels [4, 28, 40]. While these attempts have been shown effective in certain out-of-distribution scenarios, they still fall short of explicitly accounting for causal relationships between interactive agents. More recently, CausalAgents [56] made an effort to collect annotations of agent relations (causal or non-causal) in the Waymo dataset [20], providing a new benchmark focused on the robustness issue under non-causal agent perturbations. Nevertheless, the reason behind the robustness issue remains unclear, as does the potential use of the collected annotations for representation learning. + +The goal of this work is to provide an in-depth analysis of the challenges and opportunities of learning causally-aware representations in the multi-agent context. To this end, we first take an in-depth look at the recent CausalAgents benchmark [56]. We find that its labeling procedure and evaluation protocol present subtle yet critical caveats, thereby resulting + +in a highly biased measure of robustness. To mitigate these issues, we construct a diagnostic dataset through counterfactual simulations and revisit a collection of recent multi-agent forecasting models. Interestingly, we find that most recent models are already partially resilient to perturbations of non-causal agents but still struggle to capture indirect causal effects that involve mediator agents. + +To further enhance the causal robustness of the learned representations, we propose a regularization approach that seeks to preserve the causal effect of each individual neighbor in an embedding space. Specifically, we devise two variants that exploit annotations with different levels of granularity: (i) a contrastive-based regularizer using binary annotations of causal/non-causal agents; (ii) a ranking-based regularizer using continuous annotations of causal effects. Through controlled experiments, we show that both regularizers can enhance causal awareness to notable degrees. More crucially, we find that finer-grained annotations are particularly important for generalization out of the training distribution, such as higher agent density or unseen context. + +Finally, we introduce a sim-to-real causal transfer framework, aiming at extending the strengths of causal regularization from simulation to real-world contexts. We achieve this through cross-domain multi-task learning, i.e., jointly train the representation on the causal task in simulation and the forecasting task in the real world. Through experiments on the ETH-UCY dataset [36, 52] paired with an ORCA simulator [65], we find that the causal transfer framework enables stronger generalization in challenging settings such as low-data regimes. Despite a large gap between simulator and real-world data, our approach outperforms the AutoBots baseline with only $25\%$ of the real data when combined with our causal regularizers. Furthermore, with just $50\%$ of the real data, it surpasses all other models in prediction accuracy. This is achieved even in the absence of real-world causal annotations, underscoring the robustness of our framework in limited-data scenarios. We also test the causal transfer approach on the NBA dataset [70, 71], demonstrating its effectiveness in different real-world settings. As one of the first steps towards causal modeling in multi-agent systems, we hope our work sheds new light on the challenges and opportunities of learning causally-aware representations in practice. + +# 2. Related Work + +The social causality studied in this work lies at the intersection of three areas: multi-agent interactions, robust representations, and causal learning. In this section, we provide a brief overview of the existing literature in each area and then discuss their relevance to our work. + +Multi-Agent Interactions. The study of multi-agent interactions has a long history. Early efforts were focused on hand-crafted rules, such as social forces [25, 48] and recip- + +rocal collision avoidance [1, 65]. Despite remarkable results in sparse scenarios [21, 44, 72], these models often lack social awareness in more densely populated and complex environments [57]. As an alternative, recent years have witnessed a paradigm shift toward learning-based approaches, particularly the use of carefully designed neural networks to learn representations of multi-agent interactions [32]. Examples include pooling operators [2, 14, 24], attention mechanisms [18, 30, 60, 68], spatio-temporal graphs [31, 37, 61], among others [8, 13, 55]. Nevertheless, the robustness of these models remains a grand challenge [6, 56, 58]. Our work presents a solution to enhance robustness by effectively exploiting causal annotations of varying granularity. + +Robust Representations. The robustness of machine learning models, especially under distribution shifts, has been a long-standing concern for safety-critical applications [54]. Existing efforts have explored two main avenues to address this challenge. One line of work seeks to identify features that are invariant across distributions. Unfortunately, this approach often relies on strong assumptions about the underlying shifts [26, 39] or on access to multiple training domains [3, 34], which may not be practical in real-world settings. Another approach aims to develop models that can efficiently adapt to new distributions by updating only a small number of weight parameters, such as sub-modules [33], certain layers [35], or a small subset of neurons [10]. More recently, there has been a growing interest in exploring the potential of causal learning to address the robustness challenges [15, 16, 50, 67]. To the best of our knowledge, our work makes the first attempt in the multi-agent context, showcasing the benefits of incorporating causal relationships for stronger out-of-distribution generalization in crowded scenarios. + +Causal Learning. Empowering machine learning with causal reasoning has gained a growing interest in recent years [62, 63]. One line of work seeks to discover high-level causal variables from low-level observations, e.g., disentangled [12, 27, 42] or structured latent representations [43, 64]. Unfortunately, existing methods remain largely limited to simple and static datasets [41]. Another thread of work attempts to draw causal insights into dynamic decision-making [29]. In particular, recent works have proposed a couple of methods to incorporate causal invariance and structure into the design and training of forecasting models in the multi-agent context [4, 7, 9, 28, 40, 45]. However, these efforts have mainly focused on using causal implications to enhance robustness rather than explicitly examining causal relations between interactive agents. Closely related to ours, another recent study introduces a motion forecasting benchmark with annotations of causal relationships [56]. Our work addresses this benchmark's limitations and introduces a sim-to-real approach to enhance the causal awareness of the learned representations. + +# 3. Preliminaries + +In this section, we seek to formalize the robustness challenge in multi-agent representation learning through the lens of social causality. We will first revisit the design of the CausalAgents [56] benchmark, drawing attention to its caveats in labeling and evaluation. Subsequently, we will introduce a diagnostic dataset, aiming to facilitate more rigorous development and evaluation of causally-aware representations through counterfactual simulations. + +# 3.1. Social Interaction and Social Causality + +Social Interaction. Consider a motion forecasting problem where an ego agent is surrounded by a set of neighboring agents in a scene. Let $s_t^i = (x_t^i, y_t^i)$ denote the state of agent $i$ at time $t$ and $s_t = \{s_t^0, s_t^1, \dots, s_t^K\}$ denote the joint state of all the agents in the scene. Without loss of generality, we index the ego agent as 0 and the rest of neighboring agents as $\mathcal{A} = \{1, 2, \dots, K\}$ . Given a sequence of history observations $x = (s_1, \dots, s_t)$ , the task is to predict the future trajectory of the ego agent $y = (s_{t+1}^0, \dots, s_T^0)$ until time $T$ . Modern forecasting models are largely composed of encoder-decoder neural networks, where the encoder $f(\cdot)$ first extracts a compact representation $z$ of the input with respect to the ego agent and the decoder $g(\cdot)$ subsequently rolls out a sequence of its future trajectory $\hat{y}$ : + +$$ +\boldsymbol {z} = f (\boldsymbol {x}) = f \left(s _ {1: t}\right), \tag {1} +$$ + +$$ +\hat {\boldsymbol {y}} = \hat {s} _ {t + 1: T} ^ {0} = g (\boldsymbol {z}). +$$ + +Social Causality. Despite remarkable progress on accuracy measures [2, 61], recent neural representations of social interactions still suffer from a significant robustness concern. For instance, recent works have shown that trajectories predicted by existing models often output colliding trajectories [39], vulnerable to adversarial perturbations [5, 58] and deteriorate under distribution shifts of spurious feature such as agent density [40]. One particular notion of robustness that has recently gained much attention is related to the causal relationships between interactive agents [56]. Ideally, neural representations of social interactions should capture the influence of neighboring agents, namely how the future trajectory of the ego agent will vary between a pair of scenes: an original scene $\pmb{x}_{\varnothing}$ and a perturbed scene $\pmb{x}_{R}$ where some neighboring agents $\mathcal{R}$ are removed, i.e., only the ego agent and $\mathcal{A} \backslash \mathcal{R}$ remain. We define the causal effect of the removed agents $\mathcal{R}$ as, + +$$ +\mathcal {E} _ {\mathcal {R}} = \left\| \boldsymbol {y} _ {\varnothing} - \boldsymbol {y} _ {\mathcal {R}} \right\| _ {2}, \tag {2} +$$ + +where $\pmb{y}_{\emptyset} \equiv \pmb{y}$ is the future trajectory in the original scene, $\pmb{y}_{\mathcal{R}}$ is the future trajectory in the perturbed scene after removing agents in $\mathcal{R}$ from scene, and $\| \cdot \|_2$ is the average point-wise Euclidean distance between two trajectories. + +# 3.2. Caveats of CausalAgents Benchmark + +While the mathematical definition of the causal effect is straightforward in the multi-agent context, measuring Eq. (2) in practice can be highly difficult. In general, it is impossible to observe a subset of agents replaying their behaviors in the same environment twice in the real world – an issue known as the impossibility of counterfactuals [53]. To mitigate the data collection difficulty, a recent benchmark CausalAgents [56] proposes another simplified labeling strategy: instead of collecting paired counterfactual scenes, it queries human labelers to divide neighboring agents into two categories: causal agents that directly influence the driving behavior of the ego agent from camera viewpoints; and non-causal agents that do not. This labeling approach is accompanied by an evaluation protocol through agent removal, assuming that robust forecasting models should be insensitive to scene-level perturbations that remove non-causal agents. Formally, the CausalAgents benchmark evaluates the robustness of a learned representation through the following measure, + +$$ +\Delta = \left\| \hat {\boldsymbol {y}} _ {\mathcal {R}} - \boldsymbol {y} _ {\mathcal {R}} \right\| _ {2} - \left\| \hat {\boldsymbol {y}} _ {\varnothing} - \boldsymbol {y} _ {\varnothing} \right\| _ {2}, \tag {3} +$$ + +where $\mathcal{R}$ includes all or some non-causal agents in the scene. Caveats. In spite of the precious efforts in gathering causal annotations, the CausalAgents benchmark [56] is subject to two limitations that require attention: + +1. Annotation: The human labeling process overlooks indirect causal effects. As illustrated in Fig. 2, a neighboring agent may have a considerable causal influence on the behavior of the ego agent via mediators. The complexity of these causal chains poses a significant challenge for human annotations. Yet, these indirect effects are common in scenarios with high agent density, presenting a significant modeling challenge, which we will demonstrate in Sec. 5. +2. Evaluation: The evaluation protocol of the CausalAgents benchmark may inaccurately portray robustness due to oversimplification in handling agent interactions, as illustrated in Fig. 3. This figure shows that the impact of non-causal agents grows when they are removed collectively rather than individually. While agents are marked individually as non-causal, the robustness test removes them in groups, which can exaggerate the model's perceived robustness. Simultaneous removal of multiple non-causal agents might expose causal effects that are not detected when these agents are considered separately. For a detailed scenario illustrating this effect, please refer to Appendix C. + +# 3.3. Diagnostic Dataset through Counterfactuals + +To address the identified issues, we have developed a new diagnostic dataset using the ORCA [65] simulation system, which enables the creation of counterfactual scenarios by + +![](images/711a51d4aebc5126edc1600cb61be11fcd22f66c50c3c401113b4c3f4eb9a060.jpg) +Figure 2. Illustration of indirect causal effects. The elderly pedestrian indirectly affects the trajectory of the ego agent by influencing the path of the green pedestrian. In the elderly pedestrian's absence, the green pedestrian would proceed straight, allowing the ego agent to continue at its current speed without needing to slow down. + +removing sets of agents. ORCA models social interactions like collision avoidance and leader-follower behavior by taking initial positions, speeds, and final goals of agents. We constructed our dataset by initializing the positions and goals of $8 \leq K + 1 \leq 18$ agents within a structured layout typical of an open area. Additionally, we generated two out-of-distribution (OOD) datasets, named Density OOD and Context OOD, to test model robustness. The Density OOD dataset varies the density of agents and adds more non-causal ones, while the Context OOD dataset alters the initial layout to resemble a narrow street with pedestrians moving from one end to the other. Both the datasets and the code used for their generation are available to support further research. For more detailed information on the data collected, see Appendix B. + +Counterfactual Pairs. Leveraging the capability to simulate counterfactual scenarios, we annotate the ground-truth causal effects by comparing ego trajectories in paired scenes before and after agent removals, as defined in Eq. (2). Given the exponential growth of simulation scenarios with the number of agents, we adopt the approach of Causal Agents [56] and concentrate on evaluating causal effects at the level of individual agents, denoted as i.e. $|\mathcal{R}| = 1$ . For simplicity, moving forward, we will use $\mathcal{E}_i$ to refer to $\mathcal{E}_{\{i\}}$ . + +Fine-grained Category. In addition to real-valued causal effects, we also seek to annotate the category of each agent. From simulations, we extract the per-step causal relation between a neighboring agent $i$ and the ego agent. If agent $i$ is visible to the ego at time step $t$ , it directly influences the ego, indicated by $\mathbb{1}_t^i = 1$ ; otherwise $\mathbb{1}_t^i = 0$ . We then convert the causal effect over the whole sequence $\mathcal{E}_i$ into three agent categories: + +- Non-causal agent: little influence on the behavior of the ego agent, i.e., $\mathcal{E} < \epsilon \approx 0$ . +- Direct causal agent: significant influence on the behavior of the ego agent $\mathcal{E} > \eta \gg 0$ ; moreover, the influence is direct for at least one time step $\prod_{\tau = 1:T}(1 - \mathbb{1}_{\tau}^{i}) = 0$ . +- Indirect causal agent: significant influence on the behavior of the ego agent $\mathcal{E} > \eta \gg 0$ ; however, the influence is never direct over the entire sequence $\prod_{\tau = 1:T}(1 - \mathbb{1}_{\tau}^{i}) = 1$ . + +![](images/c034bcd425b1c71ca679601407a939fb60cd7962d502753d90f3da4b5a19b59d.jpg) +Figure 3. Joint effect of removing all non-causal agents in the scene, simulated in ORCA. The ego agent often changes its behavior when more than 2 non-causal neighbors are removed. + +The collected diagnostic dataset with different levels of causal annotations allows us to rigorously probe the causal robustness of existing representations and to develop more causally-aware representations, which we will describe in the next sections. + +# 4. Method + +In this section, we will introduce a simple yet effective approach to promote causal awareness of the learned representations. We will first describe a general regularization method, which encapsulates two specific instances exploiting causal annotations of different granularity. We will then introduce a sim-to-real transfer framework that extends the regularization method to practical settings, even in the absence of real-world annotations. + +# 4.1. Causally-Aware Regularization + +Recall that the causal effect $\mathcal{E}_i$ of an agent $i$ is tied to the difference of the potential outcomes between the factual scene and the counterfactual one where the agent $i$ is removed. In this light, a causally-aware representation should also capture such relations, i.e., feature vectors of paired scenes are separated by a certain distance $d_i$ depending on $\mathcal{E}_i$ . + +We start by defining $\pmb{p}_{\varnothing}$ as the low-dimensional feature vector of the factual scene where no agents are removed. This serves as our baseline scenario against which all counterfactual instances are compared. $\pmb{p}_{\varnothing}$ is extracted from the latent representation $\pmb{z}_{\varnothing} = f(\pmb{x}_{\varnothing})$ projected through a nonlinear head $h$ . Similarly, $\pmb{p}_i = h(\pmb{z}_i) = h(f(\pmb{x}_i))$ the feature vector for the counterfactual scene $\pmb{x}_i$ created by removing agent $i$ . We then measure the distance between $\pmb{x}_{\varnothing}$ and $\pmb{x}_i$ as + +$$ +d _ {i} = 1 - \operatorname {s i m} \left(\boldsymbol {p} _ {\varnothing}, \boldsymbol {p} _ {i}\right) = 1 - \frac {\boldsymbol {p} _ {\varnothing} ^ {\top} \boldsymbol {p} _ {i}}{\| \boldsymbol {p} _ {\varnothing} \| \| \boldsymbol {p} _ {i} \|}, \tag {4} +$$ + +where $\mathrm{sim}(\pmb{p}_{\emptyset}, \pmb{p}_i)$ denotes the cosine similarity between the feature vectors of the factual and counterfactual scenes, effectively measuring the impact of agent $i$ 's removal on the embedding space. The desired representation is thus + +![](images/67ee117dd799b09fa2abf00a02d18549c7ef6047533529c26758b62e66affaa8.jpg) +Figure 4. Overview of our method. We seek to build an encoder that captures the causal effect of each agent by regularizing the distance of paired embeddings between the factual and counterfactual scenes. We formulate this objective into a contrastive task given binary annotations or a ranking task given real-value annotations. + +expected to preserve the following property, + +$$ +\mathcal {E} _ {i} < \mathcal {E} _ {j} \Longrightarrow d _ {i} < d _ {j}, \quad \forall i, j \in \mathcal {A}, \tag {5} +$$ + +where $i$ and $j$ are the indices of two agents in the set of neighboring agents $\mathcal{A}$ . We next describe two specific tasks that seek to enforce this property Eq. (5), while taking into account the concrete forms of causal annotations with different granularity, as illustrated in Fig. 4. + +Causal Contrastive Regularization. As discussed in Sec. 3.3, one simple form of causal annotations is binary labels, indicating whether or not a neighboring agent causally influences the behavior of the ego agent. Intuitively, the embedding of a counterfactual scenario where a non-causal agent is removed should closely resemble the embedding of the factual scenario, reflecting minimal change from the ego agent's perspective. In contrast, removing a causal agent should significantly alter the representation, indicating a higher impact. We formulate this intuition into a causal contrastive learning objective, + +$$ +\mathcal {L} _ {\text {c o n t r a s t}} = - \log \frac {\exp \left(d ^ {+} / \tau\right)}{\exp \left(d ^ {+} / \tau\right) + \sum_ {k} \exp \left(d _ {k} / \tau\right) \mathbb {1} _ {\varepsilon_ {k} > \eta}}, \tag {6} +$$ + +where the positive (distant) example is sampled from counterfactuals with respect to causal agents, the negative (nearby) examples are sampled from counterfactual scenarios with respect to non-causal agents, and $\tau$ is a temperature hyperparameter controlling the difficulty of the contrastive task. + +Causal Ranking Regularization. One downside of causal contrastive learning described above is that it inherently ignores the detailed effect of causal agents. It tends to push the embeddings of counterfactuals equally far apart across all causal agents, regardless of the variation of causal effects, which violates the desired property stated in Eq. (5). To address this limitation, we further consider another variant of causal regularization using real-valued causal annotations to provide more dedicated supervision on the relative distance + +in the embedding space. Concretely, we first sort all agents in a scene based on their causal effect and then sample two agents with different causal effects for comparison. This allows us to formulate a ranking problem in a pairwise manner through a margin ranking loss, + +$$ +\mathcal {L} _ {\text {r a n k i n g}} = \max \left(0, d _ {i} - d _ {j} + m\right), \tag {7} +$$ + +where $d_{i}$ and $d_{j}$ are the embedding distances with respect to two agents of different causal effects $\mathcal{E}_i < \mathcal{E}_j$ , and $m$ is a small margin hyperparameter controlling the difficulty of the ranking task. + +# 4.2. Sim-to-real Causal Transfer + +The causal regularization method described above relies upon the premise that annotations of causal effects are readily available. However, as elaborated in Sec. 3.3, procuring such causal annotations in real-world scenarios can be difficult. To bridge this gap, we extend our causal regularization approach to a sim-to-real transfer learning framework. Our key idea is that, despite the gap between simulation environments and the real world, certain causal mechanisms (e.g., collision avoidance and group coordination) tend to remain stable. As such, in the absence of real-world causal annotations, we can use the annotations gathered from simulations as a viable alternative. This leads to cross-domain training on two tasks jointly: the main prediction task $\mathcal{L}_{pred}^{real}$ in real-world scenarios which usually is a reconstruction loss, and our auxiliary causal task $\mathcal{L}_{causal}^{syn}$ in the simulation counterpart, + +$$ +\mathcal {L} = \mathcal {L} _ {\text {p r e d}} ^ {\text {r e a l}} + \alpha \mathcal {L} _ {\text {c a u s a l}} ^ {\text {s y n}}, \tag {8} +$$ + +where $\alpha$ is a hyperparameter controlling the emphasis on the causal regularizer. Despite its simplicity, we will show in Sec. 5.3 that the sim-to-real causal transfer framework effectively translates the causal knowledge absorbed from simulations to real-world datasets, even without any real-world causal annotations. + +
ADE ↓FDE ↓ACE-NC ↓ACE-DC ↓ACE-IC ↓
D-LSTM [32]0.3290.6770.0270.5320.614
S-LSTM [2]0.3140.6270.0310.4630.523
TRAJECTRON++ [61]0.3120.6300.0240.4790.568
STGCNN [49]0.3070.5640.0490.3300.354
MULTI-TRANSMOTION [18]0.2370.4660.0310.7230.724
AUTOBOTS [22]0.2550.4970.0450.5950.616
+ +Table 1. Performance of modern representations on the created diagnostic dataset. The prediction errors and causal errors are generally considerable across all evaluated models. However, the errors associated with ACE-NC are rather marginal, suggesting that recent models are already partially robust to non-causal perturbations of individual agents. + +# 5. Experiments + +In this section, we will present a set of experiments to answer the following four questions: + +1. How well do recent representations capture the causal relations between interactive agents? +2. Is our proposed method effective for addressing the limitations of recent representations? +3. Do finer-grained annotations provide any benefits for learning causally-aware representations? +4. Finally, does greater causal awareness offer any practical advantages in challenging scenarios? + +Throughout our experiments, the multi-agent forecasting task is defined as predicting the future trajectory of the ego agent for 12 time steps, given the history observations of all agents in a scene in the past 8 time steps. Given that our simulator operates deterministically without random variations in agent trajectories, we train our forecasting models to output a single future trajectory for the ego agent, rather than generating multiple potential trajectories in multimodal settings. We evaluate forecasting models on three metrics: + +- Average Displacement Error (ADE): the average Euclidean distance between the predicted output and the ground truth trajectories – a widely used metric to measure the prediction accuracy of a forecasting model. +- Final Displacement Error (FDE): the Euclidean distance between the predicted output and the ground truth at the last time step – another common metric measuring prediction accuracy. +- Average Causal Error (ACE): the average difference between the estimated causal effect and the ground truth causal effect – a metric inspired by Roelofs et al. [56] to measure the causal awareness of a learned representation, + +$$ +\mathrm {A C E} := \frac {1}{K} \sum_ {i = 1} ^ {K} | \hat {\mathcal {E}} _ {i} - \mathcal {E} _ {i} | = \frac {1}{K} \sum_ {i = 1} ^ {K} \| \hat {\boldsymbol {y}} _ {\varnothing} - \hat {\boldsymbol {y}} _ {i} \| _ {2} - \mathcal {E} _ {i} |. \tag {9} +$$ + +Alongside the aggregated ACE, we also measure causal awareness on each category separately, i.e., ACE-NC, ACE-DC, ACE-IC for non-causal, direct causal, and indirect causal agents, respectively. Note that ACE can only be + +![](images/2a345414e6dec3a02aac25674b32bcdbe83c228d44c70a1910de3716062f1119.jpg) +(a) Direct Causal Agents + +![](images/2aad8056a35d1bfa352b3514a58b515daf8811adf2867e7e8f876f34db7b7f9f.jpg) +(b) Indirect Causal Agents +Figure 5. Comparison between estimated causal effects $\hat{\mathcal{E}}$ and the ground truth $\mathcal{E}$ . We uniformly sample the ground-truth causal effect, collect the estimate from AutoBots [22], and regress a slope between the estimated value and the ground truth. The positioning of most data points below the ideal diagonal line suggests a systematic underestimation of causal effects. + +evaluated in the presence of causal annotations $\mathcal{E}_i$ , e.g., the diagnostic dataset through controlled simulations (Sec. 3.3). + +# 5.1. Robustness of Recent Representations + +Setup. We start our experiments with the evaluation of a collection of recent models on the diagnostic dataset described in Sec. 3. We explicitly examine the errors of causal effect estimation for each agent category. We summarize the results in Tab. 1 and visualize the detailed estimation in Fig. 5. + +Takeaway 1: recent representations are already partially robust to non-causal agent removal. As shown in Tab. 1, the values of ACE-NC are generally quite minimal compared to the results on other metrics. Across all evaluated models, the errors made upon non-causal agents are $10\mathrm{x}$ smaller than that for causal agents (ACE-DC/IC). In line with our analysis in Sec. 3, this result provides a counterargument to the robustness issue benchmarked in Roelofs et al. [56], suggesting the importance of studying the robustness under causal perturbations as opposed to non-causal ones. + +Takeaway 2: recent representations underestimate causal effects, particularly indirect ones. As shown in Fig. 5, the estimate of causal effect often deviates from the ground + +![](images/334f78a3f2046daeead33db89e44d32d486593774edac2d06058ebc00030b523.jpg) +Figure 6. Quantitative results of our causal regularization method in comparison to the Autobots baseline [22] and the non-causal data augmentation [56]. Models trained by our ranking-based method yield enhanced performance both in terms of both causal effect estimation (left) and prediction accuracy (right). Results are averaged over five random seeds. + +![](images/e162ae7612ae5c152ff3c9e9c8d2f745111a725bfa920cb6a698a269a148bbbd.jpg) + +![](images/b44457ebe1f842de6760439dec2ee8532e9a56d03b8093efc2f099327420c481.jpg) +Figure 7. Qualitative results of our method on in-distribution (ID) and out-of-distribution (OOD) test sets. Models regularized by our ranking-based method show a more robust understanding of agent interactions in high-density and novel scene contexts. + +![](images/7ab49d03424987555dfe20ac2e2aa9ea7287c1b5a6779d7fcbb37700047b3371.jpg) + +![](images/ca04dc99a3a8cf34f560874e1d39fa6eae740e77a38d9c6789ea8dacf18dba2e.jpg) + +truth value. In particular, the learned representation tends to severely underestimate the influence of indirect causal agents. This result underscores the limitation of existing interaction representations in reasoning about a chain of causal relations transmitting across multiple agents. + +# 5.2. Effectiveness of Causal Regularization + +We next evaluate the efficacy of our causal regularization method on Autobots, the strongest baselines in Tab. 1. As a baseline comparison, we incorporate the data augmentation strategy from [56], which augments training scenarios by selectively removing non-causal agents. This approach assumes that the removal of non-causal agents should not alter the ego agent's trajectory, thereby using the same ground truth trajectory to reinforce model predictions against changes under non-causal perturbations. We assess the performance in two key areas: in-distribution causal awareness and out-of-distribution generalization. + +# 5.2.1. In-distribution Causal Awareness + +Setup. We first examine the efficacy of our method by measuring the ACE on the in-distribution test set. Since the error of the baseline model varies substantially across + +different ranges of causal effects (Sec. 5.1), we split the test set into four quantile segments. The result is summarized in Fig. 6. + +Takeaway: our method boosts causal awareness thanks to fine-grained annotations. As shown in Fig. 6, both the contrastive and ranking variants of our causal regularization approach result in lower causal errors than the vanilla baseline. In particular, the causal ranking method demonstrates substantial advantages over the other counterparts across all quantities of causal effects, confirming the promise of incorporating fine-grained annotations in learning causally-aware representations. + +# 5.2.2. Out-of-Distribution Generalization + +Setup. We further examine the efficacy of our method by measuring the prediction accuracy on out-of-distribution test sets. We consider two common types of distribution shifts in the multi-agent setting: higher density and unseen context, as detailed in Appendix B. We summarize the results of different methods on the OOD test sets in Fig. 6 and visualize the difference of prediction output in Fig. 7. + +Takeaway: causally-aware representations lead to stronger out-of-distribution generalization. As shown + +![](images/16d8e74ab510ea565c8143de2549c0831fe5c55a463cb9d6a9eb96df95945956.jpg) +Figure 8. Results of causal transfer from ORCA simulations to the ETH-UCY dataset. We consider three settings where the model has access to the same simulation data but varying amounts of real-world data. Our ranking-based causal transfer results in lower prediction errors and higher learning efficiency than the other counterparts, i.e., AutoBots baseline [22], the vanilla sim-to-real transfer [38] and the non-causal data augmentation [56]. Details are summarized in Appendices A and B. + +![](images/9565ddba1368c7e49764bb42652916bb1d3da262d18bb0244728dd25eb15f6ab.jpg) + +in Fig. 6, the prediction error from our causal ranking method is generally lower than that of the other methods on the OOD sets. In fact, the overall patterns between the two parts of Fig. 6 are highly similar, indicating a substantial correlation between causal awareness and out-of-distribution robustness. This practical benefit is also visually evident in the qualitative visualization in Fig. 7. + +# 5.3. Effectiveness of Causal Transfer + +Setup. Finally, we evaluate the proposed causal transfer approach on the ETH-UCY dataset [36, 52] and NBA rebound dataset [70, 71] paired with our diagnostic dataset. To assess the potential of the transfer method in challenging settings such as low-data regimes, we train Autobots [22] on different fractions of the real-world data. The results are summarized in Fig. 8 and Tab. 2. + +Takeaway: causal transfer improves performance in the real world despite domain gaps. As shown in Fig. 8, our causal transfer approach consistently improves the performance of the learned representation on ETH-UCY, even with a substantial simulator-to-real-world gap. In particular, it enables the model to learn faster, e.g., even with half data in the real-world, our ranking-based method can still outperform other counterparts. In contrast, the vanilla sim-to-real technique [38] which combines synthetic and real data, struggles with effective transfer, especially in low-data regimes. We hypothesize this is due to disparities in non-causal interaction styles, such as speed and curvature, between simulation and reality. We report ORCA simulator and other baselines' prediction accuracy on ETH-UCY in Appendix A. We further examine the causal transfer approach on NBA rebound, a more large-scale and interactive dataset. Table 2 shows that our ranking-based method outperforms the baseline and demonstrates that the proposed causal transfer can work not only for pedestrians but also for basketball players. + +
ADE↓FDE↓
Baseline0.5621.271
Vanilla0.5671.287
Augment0.5691.292
Contrast (ours)0.5501.248
Ranking (ours)0.5441.235
+ +Table 2. Results of causal transfer from ORCA simulations to the NBA rebound dataset. We outperform the AutoBots baseline by integrating our causal regularizers, showing effectiveness of our approach. The numbers are in meters. + +# 6. Conclusions + +Summary. In this paper, we presented a thorough analysis and an effective approach for learning causally-aware representations of multi-agent interactions. We revisited the notion of non-causal robustness in the recent benchmark [56] and showed that the main weaknesses of recent representations are not overestimations of the effect of non-causal agents but rather underestimation of the effects of indirect causal agents. To boost causal awareness of the learned representations, we introduced a regularization approach that encapsulates a contrastive and a ranking variant leveraging annotations of different granularity. We showed that our approach enables recent models to learn faster, generalize better, as well as transfer stronger to practical problems, even without real-world annotations. + +Discussions. As one of the first steps towards causally-aware representations in the multi-agent context, our work raises two primary questions for further exploration. On the technical front, while our proposed regularization method consistently boosts causal awareness in various settings, supervision alone is likely insufficient to fully solve causal reasoning in complex scenes, as evidenced by the causal errors in Fig. 6. Incorporating structural inductive biases might be a promising direction to address high-order reasoning challenges in the presence of indirect causal effects. + +# 7. Acknowledgement + +The authors would like to thank Yang Gao and Kaouther Messaoud for their valuable feedback. This work was supported by Honda R&D Co., Ltd and Hasler Foundation under the responsible AI program. + +# References + +[1] Alexandre Alahi, Vignesh Ramanathan, and Li Fei-Fei. Socially-Aware Large-Scale Crowd Forecasting. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2211–2218, 2014. 2 +[2] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social LSTM: Human Trajectory Prediction in Crowded Spaces. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 961–971, 2016. 1, 2, 3, 6, 13 +[3] Martin Arjovsky, León Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant Risk Minimization. arXiv:1907.02893 [cs, stat], 2020. 2 +[4] Shayan Shirahmad Gale Bagi, Zahra Gharquee, Oliver Schulte, and Mark Crowley. Generative Causal Representation Learning for Out-of-Distribution Motion Forecasting, 2023. 1, 2 +[5] Yulong Cao, Chaowei Xiao, Anima Anandkumar, Danfei Xu, and Marco Pavone. AdvDO: Realistic Adversarial Attacks for Trajectory Prediction. In Computer Vision - ECCV 2022, pages 36-52, Cham, 2022. Springer Nature Switzerland. 3 +[6] Yulong Cao, Danfei Xu, Xinshuo Weng, Zhuoqing Mao, Anima Anandkumar, Chaowei Xiao, and Marco Pavone. Robust Trajectory Prediction against Adversarial Attacks. In Proceedings of The 6th Conference on Robot Learning, pages 128-137. PMLR, 2023. 2 +[7] Luca Castri, Sariah Mghames, Marc Hanheide, and Nicola Bellotto. Causal Discovery of Dynamic Models for Predicting Human Spatial Interactions. In Social Robotics, pages 154-164, Cham, 2022. Springer Nature Switzerland. 1, 2 +[8] Yuning Chai, Benjamin Sapp, Mayank Bansal, and Dragomir Anguelov. MultiPath: Multiple Probabilistic Anchor Trajectory Hypotheses for Behavior Prediction. In Conference on Robot Learning, pages 86-99. PMLR, 2020. 2 +[9] Guangyi Chen, Junlong Li, Jiwen Lu, and Jie Zhou. Human Trajectory Prediction via Counterfactual Analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9824-9833, 2021. 1, 2, 13 +[10] Guanzheng Chen, Fangyu Liu, Zaiqiao Meng, and Shangsong Liang. Revisiting Parameter-Efficient Tuning: Are We Really There Yet? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2612-2626, 2022. 2 +[11] Guangyi Chen, Zhenhao Chen, Shunxing Fan, and Kun Zhang. Unsupervised Sampling Promoting for Stochastic Human Trajectory Prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17874-17884, 2023. 1 +[12] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. InfoGAN: Interpretable Representation Learning by Information Maximizing Generative + +Adversarial Nets. In Advances in Neural Information Processing Systems 29, pages 2172-2180. Curran Associates, Inc., 2016. 2 +[13] Chiho Choi, Srikanth Malla, Abhishek Patil, and Joon Hee Choi. DROGON: A Trajectory Prediction Model based on Intention-Conditioned Behavior Reasoning. In Proceedings of the 2020 Conference on Robot Learning, pages 49–63. PMLR, 2021. 2 +[14] Nachiket Deo and Mohan M. Trivedi. Convolutional Social Pooling for Vehicle Trajectory Prediction. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1549-15498, 2018. 2 +[15] Andrea Dittadi, Frederik Trauble, Francesco Locatello, Manuel Wuthrich, Vaibhav Agrawal, Ole Winther, Stefan Bauer, and Bernhard Schölkopf. On the Transfer of Disentangled Representations in Realistic Settings. In International Conference on Learning Representations, 2020. 2 +[16] Andrea Dittadi, Samuele S. Papa, Michele De Vita, Bernhard Schölkopf, Ole Winther, and Francesco Locatello. Generalization and Robustness Implications in Object-Centric Learning. In Proceedings of the 39th International Conference on Machine Learning, pages 5221-5285. PMLR, 2022. 2 +[17] Chen et al. Human trajectory prediction via counterfactual analysis. In ICCV, 2021. 13 +[18] Gao et al. Multi-transmotion: Pre-trained model for human motion prediction. In CoRL, 2024. 2, 6, 13 +[19] Martin et al. Jrdb: A dataset and benchmark of egocentric robot visual perception of humans in built environments. TPAMI, 2021. 13 +[20] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R. Qi, Yin Zhou, Zoey Yang, Aurélien Chouard, Pei Sun, Jiquan Ngiam, Vijay Vasudevan, Alexander McCauley, Jonathon Shlens, and Dragomir Anguelov. Large Scale Interactive Motion Forecasting for Autonomous Driving: The Waymo Open Motion Dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 1 +[21] Gonzalo Ferrer, Anais Garrell, and Alberto Sanfeliu. Robot companion: A social-force based approach with human awareness-navigation in crowded environments. In 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1688-1694, 2013. 2 +[22] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent Variable Sequential Set Transformers for Joint Multi-Agent Motion Prediction. In International Conference on Learning Representations, 2021. 6, 7, 8, 12, 13 +[23] Junru Gu, Qiao Sun, and Hang Zhao. DenseTNT: Waymo Open Dataset Motion Prediction Challenge 1st Place Solution. arXiv:2106.14160 [cs], 2021. 1 +[24] A. Gupta, J. Johnson, L. Fei-Fei, S. Savarese, and A. Alahi. Social GAN: Socially Acceptable Trajectories with Generative Adversarial Networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2255–2264, 2018. 1, 2 + +[25]Dirk Helbing and Peter Molnar. Social Force Model for Pedestrian Dynamics.Physics Review E,1998.2 +[26] Dan Hendrycks*, Norman Mu*, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty. In International Conference on Learning Representations, 2020. 2 +[27] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. Beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework. In International Conference on Learning Representations, 2017. 2 +[28] Yeping Hu, Xiaogang Jia, Masayoshi Tomizuka, and Wei Zhan. Causal-based Time Series Domain Generalization for Vehicle Intention Prediction. In 2022 International Conference on Robotics and Automation (ICRA), pages 7806-7813, 2022. 1, 2 +[29] Biwei Huang, Fan Feng, Chaochao Lu, Sara Magliacane, and Kun Zhang. AdaRL: What, Where, and How to Adapt in Transfer Reinforcement Learning. In International Conference on Learning Representations, 2022. 2 +[30] Yingfan Huang, Huikun Bi, Zhaoxin Li, Tianlu Mao, and Zhaoqi Wang. STGAT: Modeling Spatial-Temporal Interactions for Human Trajectory Prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6272-6281, 2019. 2 +[31] Vineet Kosaraju, Amir Sadeghian, Roberto Martin-Martin, Ian Reid, Hamid Rezatofighi, and Silvio Savarese. SocialBiGAT: Multimodal Trajectory Forecasting using BicycleGAN and Graph Attention Networks. Advances in Neural Information Processing Systems, 32:137-146, 2019. 1, 2 +[32] Parth Kothari, Sven Kreiss, and Alexandre Alahi. Human Trajectory Forecasting in Crowds: A Deep Learning Perspective. IEEE Transactions on Intelligent Transportation Systems, pages 1-15, 2021. 2, 6, 13 +[33] Parth Kothari, Danya Li, Yuejiang Liu, and Alexandre Alahi. Motion Style Transfer: Modular Low-Rank Adaptation for Deep Motion Forecasting. In Conference on Robot Learning (CoRL), 2022. 1, 2 +[34] David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-Distribution Generalization via Risk Extrapolation (REx). In Proceedings of the 38th International Conference on Machine Learning, pages 5815-5826. PMLR, 2021. 2 +[35] Yoonho Lee, Annie S. Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, and Chelsea Finn. Surgical Fine-Tuning Improves Adaptation to Distribution Shifts. In International Conference on Learning Representations, 2023. 2 +[36] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by Example. Computer Graphics Forum, 26:655-664, 2007. 2, 8 +[37] Jiachen Li, Fan Yang, Masayoshi Tomizuka, and Chiho Choi. EvolveGraph: Multi-Agent Trajectory Prediction with Dynamic Relational Reasoning. In Advances in Neural Information Processing Systems, pages 19783-19794. Curran Associates, Inc., 2020. 2 + +[38] Junwei Liang, Lu Jiang, and Alexander Hauptmann. SimAug: Learning Robust Representations from Simulation for Trajectory Prediction. In Computer Vision – ECCV 2020, pages 275–292, Cham, 2020. Springer International Publishing. 8, 12, 14 +[39] Yuejiang Liu, Qi Yan, and Alexandre Alahi. Social NCE: Contrastive Learning of Socially-Aware Motion Representations. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 15118-15129, 2021. 1, 2, 3 +[40] Yuejiang Liu, Riccardo Cadei, Jonas Schweizer, Sherwin Bahmani, and Alexandre Alahi. Towards Robust and Adaptive Motion Forecasting: A Causal Representation Perspective. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17081-17092, 2022. 1, 2, 3 +[41] Yuejiang Liu, Alexandre Alahi, Chris Russell, Max Horn, Dominik Zietlow, Bernhard Schölkopf, and Francesco Locatello. Causal Triplet: An Open Challenge for Intervention-centric Causal Representation Learning. In Conference on Causal Learning and Reasoning (CLeaR), 2023. 2 +[42] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Scholkopf, and Olivier Bachem. Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations. In Proceedings of the 36th International Conference on Machine Learning, pages 4114-4124. PMLR, 2019. 2 +[43] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-Centric Learning with Slot Attention. In Advances in Neural Information Processing Systems, pages 11525-11538. Curran Associates, Inc., 2020. 2 +[44] Matthias Luber, Johannes A. Stork, Gian Diego Tipaldi, and Kai O Arras. People tracking with human motion predictions from social forces. In 2010 IEEE International Conference on Robotics and Automation, pages 464-469, 2010. 2 +[45] Osama Makansi, Julius Von Kugelgen, Francesco Locatello, Peter Vincent Gehler, Dominik Janzing, Thomas Brox, and Bernhard Schölkopf. You Mostly Walk Alone: Analyzing Feature Attribution in Trajectory Prediction. In International Conference on Learning Representations, 2021. 1, 2 +[46] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From Goals, Waypoints & Paths to Long Term Human Trajectory Forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15233–15242, 2021. 1 +[47] Daniel McDuff, Yale Song, Jiyoung Lee, Vibhav Vineet, Sai Vermprala, Nicholas Alexander Gyde, Hadi Salman, Shuang Ma, Kwanghoon Sohn, and Ashish Kapoor. CausalCity: Complex Simulations with Agency for Causal Discovery and Reasoning. In Proceedings of the First Conference on Causal Learning and Reasoning, pages 559-575. PMLR, 2022. 16 +[48] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 935-942, 2009. 2 +[49] A. Mohamed, K. Qian, M. Elhoseiny, and C. Claudel. Social-STGCNN: A Social Spatio-Temporal Graph Convolutional + +Neural Network for Human Trajectory Prediction. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14412-14420, 2020. 6 +[50] Milton Llera Montero, Casimir JH Ludwig, Rui Ponte Costa, Gaurav Malhotra, and Jeffrey Bowers. The role of Disentanglement in Generalisation. In International Conference on Learning Representations, 2022. 2 +[51] Hee-Seung Moon and Jiwon Seo. Fast User Adaptation for Human Motion Prediction in Physical Human-Robot Interaction. IEEE Robotics and Automation Letters, 7:120-127, 2022. 1 +[52] Stefano Pellegrini, Andreas Ess, and Luc Van Gool. Improving Data Association by Joint Modeling of Pedestrian Trajectories and Groupings. In Computer Vision - ECCV 2010, pages 452-465, Berlin, Heidelberg, 2010. Springer. 2, 8 +[53] Jonas Peters, Dominik Janzing, and Bernhard Schölkopf. Elements of Causal Inference: Foundations and Learning Algorithms. MIT Press, Cambridge, MA, USA, 2017. 3 +[54] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet Classifiers Generalize to ImageNet? In Proceedings of the 36th International Conference on Machine Learning, pages 5389-5400. PMLR, 2019. 2 +[55] Nicholas Rhinehart, Rowan Mccallister, Kris Kitani, and Sergey Levine. PRECOG: PREdition Conditioned on Goals in Visual Multi-Agent Settings. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2821-2830, 2019. 2 +[56] Rebecca Roelofs, Liting Sun, Ben Caine, Khaled S. Refaat, Ben Sapp, Scott Ettinger, and Wei Chai. Causal Agents: A Robustness Benchmark for Motion Forecasting using Causal Relationships. arXiv:2207.03586, 2022. 1, 2, 3, 4, 6, 7, 8, 12, 14, 16 +[57] Andrey Rudenko, Luigi Palmieri, Michael Herman, Kris M Kitani, Dariu M Gavrila, and Kai O Arras. Human motion trajectory prediction: A survey. The International Journal of Robotics Research, 39:895-935, 2020. 2 +[58] Saeed Saadatnejad, Mohammadhossein Bahari, Pedram Khorsandi, Mohammad Saneian, Seyed-Mohsen Moosavi-Dezfooli, and Alexandre Alahi. Are socially-aware trajectory prediction models really socially-aware? Transportation Research Part C: Emerging Technologies, 141:103705, 2022. 2, 3 +[59] Saeed Saadatnejad, Yang Gao, Kaouther Messaoud, and Alexandre Alahi. Social-transmotion: Promptable human trajectory prediction. In International Conference on Learning Representations (ICLR), 2024. 13 +[60] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. SoPhie: An Attentive GAN for Predicting Paths Compliant to Social and Physical Constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1349-1358, 2019. 2 +[61] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-Feasible Trajectory Forecasting with Heterogeneous Data. In Computer Vision - ECCV 2020, pages 683-700, Cham, 2020. Springer International Publishing. 1, 2, 3, 6, 13 + +[62] Bernhard Schölkopf. Causality for Machine Learning. arXiv:1911.10500 [cs, stat], 2019. 2 +[63] Bernhard Scholkopf and Julius von Kugelgen. From Statistical to Causal Learning. arXiv:2204.00607 [cs, stat], 2022. 1, 2 +[64] Bernhard Scholkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward Causal Representation Learning. Proceedings of the IEEE, 109:612-634, 2021. 1, 2 +[65] Jur van den Berg, Ming Lin, and Dinesh Manocha. Reciprocal Velocity Obstacles for real-time multi-agent navigation. In 2008 IEEE International Conference on Robotics and Automation, pages 1928-1935, 2008. 2, 3, 13 +[66] Jur van den Berg, Stephen J. Guy, Ming Lin, and Dinesh Manocha. Reciprocal n-body collision avoidance. In Robotics Research, pages 3-19, Berlin, Heidelberg, 2011. Springer Berlin Heidelberg. 14 +[67] Sjoerd van Steenkiste, Francesco Locatello, Jürgen Schmidhuber, and Olivier Bachem. Are Disentangled Representations Helpful for Abstract Visual Reasoning? In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2 +[68] Anirudh Vemula, Katharina Muelling, and Jean Oh. Social Attention: Modeling Attention in Human Crowds. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 1-7, Brisbane, Australia, 2018. IEEE Press. 1, 2 +[69] Chenfeng Xu, Tian Li, Chen Tang, Lingfeng Sun, Kurt Keutzer, Masayoshi Tomizuka, Alireza Fathi, and Wei Zhan. PreTraM: Self-supervised Pre-training via Connecting Trajectory and Map. In Computer Vision - ECCV 2022, pages 34-50, Cham, 2022. Springer Nature Switzerland. 1 +[70] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Socialvae: Human trajectory prediction using timewise latents. In European Conference on Computer Vision, pages 511-528. Springer, 2022. 2, 8 +[71] Yisong Yue, Patrick Lucey, Peter Carr, Alina Bialkowski, and Iain Matthews. Learning fine-grained spatial models for dynamic sports play prediction. In 2014 IEEE international conference on data mining, pages 670-679. IEEE, 2014. 2, 8 +[72] Francesco Zanlungo, Tetsushi Ikeda, and Takayuki Kanda. Social force model with explicit collision prediction. *EPL (Europhysics Letters)*, 93:68005, 2011. 2 \ No newline at end of file diff --git a/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/images.zip b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5bfbba76e69bb582d33eaacc13ebedd4fb491817 --- /dev/null +++ b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ad7e9c32f7b67f857a7873f73acd0729736436e79559de6edc51da0d34ae770 +size 299509 diff --git a/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/layout.json b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f5cb16541debd3f1e2074fc55c3c293d47b4ea82 --- /dev/null +++ b/CVPR/2025/Sim-to-Real Causal Transfer_ A Metric Learning Approach to Causally-Aware Interaction Representations/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee1c9381adf10292ec7cb0b7cc812488be79820c9b2abcbf49afc7a2d6d895ca +size 406718 diff --git a/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/d9e30bdb-1cd5-4099-8e86-d18344a3f5c6_content_list.json b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/d9e30bdb-1cd5-4099-8e86-d18344a3f5c6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b11138580a4c21339fdc2148cc04370fe1411f15 --- /dev/null +++ b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/d9e30bdb-1cd5-4099-8e86-d18344a3f5c6_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ab8b495c3d903310b414220341152f75d21676eda0edffc5cd5b21b6f890654 +size 76319 diff --git a/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/d9e30bdb-1cd5-4099-8e86-d18344a3f5c6_model.json b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/d9e30bdb-1cd5-4099-8e86-d18344a3f5c6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ce3d2e998b079721ea90bb35ac283e48e44f376 --- /dev/null +++ b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/d9e30bdb-1cd5-4099-8e86-d18344a3f5c6_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:341b4ceff1536b4d4ad2a91e16ca9d302612b089c627153d015c310d39a3cfca +size 104147 diff --git a/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/d9e30bdb-1cd5-4099-8e86-d18344a3f5c6_origin.pdf b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/d9e30bdb-1cd5-4099-8e86-d18344a3f5c6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d97ad00740a17414aa64fd65be478fe16d5f6635 --- /dev/null +++ b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/d9e30bdb-1cd5-4099-8e86-d18344a3f5c6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea8b03a4612a33b1fbc7b7bdc907d4ab1e562ac2cf37bf04e0fc567e1e1dbbaf +size 5836437 diff --git a/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/full.md b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..26672f88158083b0547967f6e91bff0f89b433a4 --- /dev/null +++ b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/full.md @@ -0,0 +1,299 @@ +# SimAvatar: Simulation-Ready Avatars with Layered Hair and Clothing + +Xueting Li, Ye Yuan, Shalini De Mello, Gilles Daviet, Jonathan Leaf, Miles Macklin, Jan Kautz, Umar Iqbal + +# NVIDIA + +Adele with long layered waved hairstyle wearing a sleeveless dress of tea-length, the garment has a delicate butterfly print, with small, colorful butterflies. + +![](images/40ebb8be56d37779753f3db2b3f981a6b55903355be794b81f09e8264bb31b85.jpg) +(a) simulation-ready avatar with layered hair, garment and body +(b) simulated avatar in different poses + +![](images/21ee0edc6905a7fc0f4ed06db1390ce0768c291ff639219b2a5f7a9bb8d5ddd1.jpg) +(c) example simulation-ready avatars. +Figure 1. (a) SimAvatar synthesizes simulation-ready 3D avatars with layered hair, body and clothing. (b) By leveraging physics-based simulators, SimAvatar produces realistic pose-dependent motion effects such as flowing hair strands and garment wrinkles. (c) SimAvatar produces simulation-ready 3D avatars with diverse identities, garments, and hairstyles. + +# Abstract + +We introduce SimAvatar, a framework designed to generate simulation-ready clothed 3D human avatars from a text prompt. Current text-driven human avatar generation methods either model hair, clothing, and the human body using a unified geometry or produce hair and garments that are not easily adaptable for simulation within existing simulation pipelines. The primary challenge lies in representing the hair and garment geometry in a way that allows leveraging established prior knowledge from foundational image diffusion models (e.g., Stable Diffusion) while being simulation-ready using either physics or neural simulators. To address this task, we propose a two-stage framework that combines the flexibility of 3D Gaussians with simulation-ready hair strands and garment meshes. Specifically, we + +first employ three text-conditioned 3D generative models to generate garment mesh, body shape and hair strands from the given text prompt. To leverage prior knowledge from foundational diffusion models, we attach 3D Gaussians to the body mesh, garment mesh, as well as hair strands and learn the avatar appearance through optimization. To drive the avatar given a pose sequence, we first apply physics simulators onto the garment meshes and hair strands. We then transfer the motion onto 3D Gaussians through carefully designed mechanisms for each body part. As a result, our synthesized avatars have vivid texture and realistic dynamic motion. To the best of our knowledge, our method is the first to produce highly realistic, fully simulation-ready 3D avatars, surpassing the capabilities of current approaches. Project page: https://research.nvidia.com/labs/dair/simavatar/ + +# 1. Introduction + +The creation of photo-realistic and dynamic human avatars has a wide range of applications, including virtual try-on, film and gaming production, virtual assistants, AR/VR, and telepresence. Traditionally, this process has required specialized training, making it inaccessible to general users. Recently, advancements in foundational diffusion models have accelerated research efforts aimed at democratizing 3D human avatar creation, enabling easy user control through text [16, 46, 51, 87] or images [39]. + +Earlier approaches for 3D human avatar creation represent the hair, body, and clothing as a single layer, making it difficult to independently simulate or edit each region due to their entangled geometry. To address this limitation, recent works have used layered structures to separately represent the body, garments, or hair [27, 36, 81, 95]. However, many of these approaches rely on implicit representations, such as NeRF [57], to define garment or hair geometry. Although implicit representations facilitate leveraging prior knowledge from foundational diffusion models, they are challenging to animate within existing simulators, limiting the realistic motion of hair and garments due to body movements. As a result, these methods struggle to produce avatars that look realistic when animated. + +Thus, a natural question arises: can we design a 3D avatar generation pipeline that can leverage rich prior knowledge from image diffusion models, while being readily compatible with existing simulation pipelines? The key challenge in addressing this question lies in connecting the different representations used in current simulators and text-driven avatar generation pipelines. The former typically requires a smooth and clean non-watertight mesh or specifically designed hair strands, whose topologies are inflexible for optimization and hard to constrain. The latter often adopt implicit representations (such as NeRF [57] or SDF [82]) which, while estimizable by noisy supervision signals from diffusion models, cannot be easily converted to an open mesh or hair strands suitable for simulation. + +To address these issues, we present a novel framework, SimAvatar, which generates a 3D human avatar from a text prompt that can be readily animated by existing hair and garment simulators. The key idea is to adopt suitable representations for different human parts (e.g., hair, body, and garments) and leverage the prior knowledge of both image diffusion models and simulators. To this end, we propose representing the geometry of human hair, body, and garment using hair strands, parametric body model SMPL [54], and meshes, respectively. While these representations are simulation-ready, they pose challenges to learning realistic texture details. To resolve this issue, we further attach 3D Gaussians to these coarse geometric representations. 3D Gaussians are not only flexible and can be easily driven by the mesh or hair strands, but they also show impressive ca + +pacity to model realistic appearance when combined with image diffusion models [53, 87]. + +Given a text prompt, we first generate hair strands, body mesh, and garment mesh using three separate text-conditioned 3D generative models trained specifically for each body region. We then attach and learn 3D Gaussians through Score Distillation Sampling (SDS) based optimization using a powerful diffusion model trained on human images [1]. To drive the avatar using a given pose sequence, we animate the body using linear blend skinning while the hair strands and garment mesh are simulated using their respective simulators [23, 30]. Next, we transfer the motion from the strands or meshes onto 3D Gaussians through a carefully designed mechanism to ensure they conform to the mesh or strand motion. During inference, the avatar can be easily animated by novel pose sequences. Our novel design yields avatars with both realistic appearance and dynamic motion with pose-dependent deformations of the hairs and garment as shown in Fig. 1. Our contributions are: + +- To the best of our knowledge, this is the first work that generates fully simulation-ready 3D avatars with separate layers for the body, garment, and hair. +- We propose a text-conditioned diffusion model to generate plausible garment geometry given text inputs and show how existing text-conditioned hair strands and body shape generators can be efficiently leveraged. +- We use 3D Gaussians on top of the body mesh, hair strands, and garment mesh to synthesize high-fidelity appearance details. We optimize the properties of Gaussians using carefully designed SDS-based optimization [86] and ensure plausible texture disentanglement using additional regularization and prompt engineering. +- Our approach produces avatars with detailed geometry, realistic texture, as well as dynamic garment and hair motion, resulting in significantly better quality than existing methods. + +# 2. Related Work + +# 2.1. Text to 3D Avatar Generation + +The domain of text-to-3D generation has witnessed significant advancements [18, 52, 62, 65, 77, 84] driven by the advent of large text-to-image models [21, 66, 67, 96]. Following the success in generating static 3D objects, various methods have also been developed to create animatable human avatars [16, 37–40, 42, 46, 51, 94, 98]. ClipMatrix [40] was among the first to create animatable avatars based on textual descriptions, using a CLIP-embedding loss to optimize a mesh-based representation. AvatarClip [34] followed a similar approach but with a NeRF-based representation [82]. DreamAvatar [16] and AvatarCraft [42] used SDS loss instead of CLIP and learned the NeRF representation in canonical space by integrating human body priors from SMPL [54]. DreamHumans [46] introduced + +a pose-conditioned NeRF representation to also model pose-dependent deformations. DreamWaltz [38], Avatar-Verse [94] and HumanNorm [96] leverage ControlNets [96] and demonstrate improved avatar quality with conditional SDS. AvatarBooth [91] fine-tunes region-specific diffusion models, highlighting that using dedicated models for distinct body regions improves avatar quality. Other methods explore different representations to model the avatars. TADA [51] demonstrates that a mesh-based approach with adaptive mesh subdivision can create high-quality avatars. More recently, GAvatar [87] adopted a primitive-based 3D Gaussian representation, allowing faster rendering in higher resolution and showing higher quality. All of the aforementioned methods, however, represent the human body, hairs, and garment geometry using a single layer and animate the avatar using linear blend skinning. Hence, they cannot model physically plausible and pose-dependent geometric deformations of the avatar's clothing and hair resulting in avatars that do not appear realistic when animated. + +There are only a few works that generate avatars with separate body and garment or hair layers. SO-SMPL [81] uses SMPL+D [61] to separately represent the cloth layers, but cannot capture loose clothing due to the inherent limitations of SMPL+D representation. HumanLiff [36], HumanCoser [83], and TELA [27] use NeRF-based representations and progressively generate minimal-clothed human body and layer-wise clothes. GALA [45] takes an existing single-layer avatar and disentangles it into separate layers using DMTet representation [29] and SDS [62] optimization. TECA [95] aims to model hairs in a separate layer but adopts a NeRF-based representation. However, the meshes extracted from NeRF or DMTet tend to be very noisy and cannot be simulated easily using cloth or hair simulators. Our approach, on the other hand, generates avatars with simulation-ready garments and hair. It can model loose clothing, realistic hair with varied lengths and hairstyles, realistic texture, and dynamic motion. + +# 2.2. Garment Modeling and Simulation + +One of the most common methods for modeling garments is SMPL+D [3-5, 14, 15, 56, 61, 93], which adds vertex displacements to the SMPL body mesh to capture the geometry of clothed human bodies. However, this approach struggles to represent loose clothing and dresses due to its fixed topology. To address this limitation, many methods propose modeling clothing separately. BCNet [41] directly predicts garment type and shape parameters using a learned regressor, while DeepCloth [76] represents garments through a canonical UV-representation, encoding them in a shared latent space. More recent methods utilize implicit unsigned distance fields for their flexibility in modeling arbitrary topologies and handling open surfaces, which are crucial for cloth simulators [17, 20, 25, 32, 64]. However, these methods typically focus solely on modeling clothing geometry, + +neglecting the underlying human body, hair, and texture details. In this work, we build on these methods and propose a novel text-conditioned diffusion model to generate garment meshes using text prompts, while also generating avatars with complete texture and appearance details for garment, hair, and body layers. + +For body motion-driven garment simulation, traditional methods use physics simulators to drape garments on the human body while modeling their material properties [13, 19, 79, 80], and avoiding collisions with the underlying body [7, 78]. Differentiable physics simulators have also been proposed to integrate physics into learning and optimization-based frameworks [49, 50]. While these methods produce highly realistic garment animations, physics simulation is generally computationally expensive and can become unstable in cases of severe interpenetration between the garment and body layers. To overcome these limitations, more recent works propose neural simulators, which are neural networks trained to generate garment deformations based on body pose and shape information [30, 31, 56, 59, 68, 70, 71]. Earlier methods were trained in a garment-specific manner and could not generalize across different garment types [58, 69]. However, more recent methods can handle various garment topologies, including loose clothing and dresses [10-12, 30, 71, 72, 89]. While our method can utilize both physics-based and neural simulators, here we use HOOD [30] as the neural simulator. + +# 2.3. Hair Modeling and Simulation + +For hair modeling, strand-based representations are widely favored due to their compatibility with physics simulators and ease of geometric manipulation [23, 28, 35, 60, 73, 88]. The main challenge in image- or text-guided generation of hair strands lies in capturing the hair's complex appearance, especially the inner strands that are often occluded. For geometry, recent approaches use data-driven priors based on diffusion models to estimate plausible inner strands despite their occlusion in observed data [74, 75]. For appearance, existing methods employ a strand-aligned 3D Gaussian representation to effectively model intricate hair textures and varied strand thickness [55, 90]. In this work, we build upon these methods for full-body avatar generation. Specifically, we use a diffusion-based text-to-hair generation model [75] to create hair strands from text prompts. We then attach 3D Gaussians [90] to the generated strands to model hair appearance and refine hairstyle according to the given text prompt. Finally, we incorporate a hair simulator [23] during avatar creation to simulate hairs for varied poses which allows for optimizing the appearance of the inner strands and ensures proper disentanglement of face and hair textures. + +# 3. Method + +Given a text prompt $\mathcal{T}$ , we introduce a novel framework, SimAvatar, for synthesizing simulation-ready 3D avatars + +![](images/93299cc23ba59caa2470696c3caf905c49f251702ed69b16adb23e9641be501a.jpg) +Figure 2. Overview. Given a text prompt, SimAvatar first generates hair strands, body mesh, and garment using their respective text-conditioned generative models. We then bring them together using physics simulation, and attach 3D Gaussians to learn their appearances and to adapt them according to the text prompts. We assign one 3D Gaussian to each line segment in the hair strands with their length significantly larger than their diameter (green circle), whereas the Gaussians for meshes are defined within a local coordinate system of each face (orange circle). We optimize the properties of all 3D Gaussians through the Score Distillation Sampling (SDS) loss using image-based diffusion models, and a novel regularization $L_{hair}$ for hairs to ensure plausible hair structure. The generated 3D avatar can be simulated by any pose sequence showing realistic and dynamic motion effects such as flowing hairs and garment wrinkles. + +with layered representations of the body, hair, and garment. Unlike traditional text-to-ayar methods that use a single geometry for the entire avatar [18, 51, 87], our approach decomposes the avatar geometry into distinct layers, each leveraging a representation tailored to its simulation and geometry modeling. This layered structure not only enhances realism but also enables robust simulation of each component, dramatically improving flexibility in animation and customization. Specifically, our avatar geometry is represented by three distinct layers: (1) Body mesh, defined by the SMPL parametric model [54], providing a robust foundation for realistic body shapes and movements; (2) Garment mesh, generated using a text-conditioned diffusion model, which produces smooth clothing meshes that are ready for simulation (see Sec. 3.2); (3) Hair strands, modeled using a text-conditioned diffusion model, allowing for the generation of complex hairstyles with strand-level detail (see Sec. 3.1). This approach allows our method to readily handle dynamic scenarios. Given a target body pose sequence $\mathcal{P} = \{p_1,\dots,p_n\}$ , the body mesh is deformed through linear blend skinning [54] while both the garment mesh and hair strands are simulated using advanced physics-based simulators (Sec. 3.1), ensuring realistic motion under various conditions. + +To further enable our layered representation to model detailed avatar appearance, including body, hair, and garment, we propose a 3D Gaussian Splatting (3DGS)-based appearance model that builds on top of the layered representation. In particular, we attach 3D Gaussians to the layered geometry of body, garment, and hair, while customizing the Gaussians for each body part to capture the unique structural and dynamic properties of the part. These 3D Gaussians are optimized using Score Distillation Sampling (SDS) [62, 87], allowing us to refine appearance details in a way that leverages powerful diffusion priors (Sec. 3.3). As a result, the generated avatars not only look more realistic but also maintain high fidelity during animation. An overview of our + +framework is shown in Fig. 2. + +# 3.1. Preliminaries + +Garment simulation. In this work, we utilize a neural simulator (i.e., HOOD [30]) to produce a simulated garment mesh sequence. Given a garment mesh $g_{0}$ and the target body pose sequence $\mathcal{P}$ , HOOD first computes the SMPL body mesh corresponding to the SMPL parameters. By treating the body meshes as obstacles, HOOD utilizes a GNN to predict the physical status (e.g., position, velocity, etc.) of each vertex on the garment mesh and produce a simulated garment sequence denoted as $\mathcal{G} = \{g_{1},\dots,g_{n}\}$ . More details are discussed in [30]. + +Strand-based hair representation. In this work, we choose strand-based hair representation to capture the thin and layer-wise structure of hairs. We represent hair as a point cloud $h_0 \in \mathbb{R}^{N_s \times N_l \times 3}$ that is composed of $N_{s}$ hair strands, with each strand includes $N_{l}$ line segments. + +Hair simulation. Given hair strands $h_0$ , a target body mesh sequence $\mathcal{P}$ and the simulated garment sequence $\mathcal{G}$ , we treat both the body and garment in each time step as obstacles and leverage a hair simulator [8, 9, 22, 24] to produce the animated hair strand sequence denoted as $\mathcal{H} = \{h_1, \dots, h_n\}$ . This sequence provides a strong prior that drives the 3D Gaussians attached to the strands to produce high-fidelity hair motion, as will be discussed in Sec. 3.3. + +# 3.2. Text-based garment diffusion model + +To ensure compatibility with existing simulators, in this work, we choose to use meshes as the coarse garment geometry representation. A straightforward approach to modify the garment mesh to match the given text prompt is through SDS-based optimization [51, 81]. However, optimizing the non-watertight garment mesh in a zero-shot manner is a non-trivial task. On one hand, the topology of garment + +![](images/aa7fd531e9fcd2b0fde1ce3d2fe9c5f35076dcd9aa3649632ae7f584779176da.jpg) +Figure 3. Text-based garment diffusion model. See Sec. 3.2 for details. + +![](images/64eb5e56c1429f261a9520427d51bd9cca7a0019b68cb0547154c482db802633.jpg) + +mesh is hard to change with noisy supervision from foundational diffusion models. On the other hand, the garment simulators all require a clean, compact, and smooth garment mesh as input. To resolve this issue, inspired by recent successes in 3D generative models [85, 92, 97], we utilize large-scale garment datasets [10, 47] and learn a text-based diffusion model for garment mesh generation. Specifically, we first train a Variational Auto-encoder (VAE) to learn a compact latent space capturing garment geometry distribution. We then learn a conditional latent diffusion model within the latent space to generate garment meshes from text prompts. + +VAE for garment mesh reconstruction. We adopt the vectorized latent space representation [85, 92, 97] for its simplicity and efficiency. As shown in Fig. 3 (a), given an input garment mesh, we first uniformly sample a set of points (denoted as $X \in \mathbb{R}^{10000 \times 3}$ ). Next, we encode the sampled points to a set of vectors (denoted as $Z \in \mathbb{R}^{512 \times 16}$ ) by applying cross-attention layers to $X$ and its downsampled version. To effectively reconstruct garment meshes from latent vectors, we adopt the Unsigned Distance Field (UDF) representation [32]. The UDF is designed for nonwatertight meshes and uses the gradient at each point to determine the presence of mesh surface. Specifically, we feed query points ( $\{q_{xyz}\}$ ) uniformly sampled from a 3D grid and the latent code $Z$ to few cross-attention layers to predict the Unsigned Distance for all $\{q_{xyz}\}$ , which forms the UDF for the garment. The garment mesh can then be extracted from the predicted UDF using MeshUDF [32]. To train the VAE model, we follow prior works [17, 25, 32] and apply: a) a binary cross-entropy loss $(L_{bce})$ to the predicted Unsigned Distance, b) the L2 loss $(L_{grad})$ to the gradient at each query point and c) the KL divergence loss $(L_{KL})$ to the latent code $Z$ . The full loss for the VAE training is $L = L_{bce} + \lambda_{grad}L_{grad} + \lambda_{KL}L_{KL}$ , where $\lambda_{grad}, \lambda_{KL}$ are empirically set to 0.0001, 0.1 in our experiments. + +Text-based garment diffusion model. Given the learned vectorized latent space, we train a text-conditioned latent diffusion model for garment mesh generation. As shown in Fig. 3(b), for a latent vector $Z$ , we add random noise $\epsilon$ sampled from a Normal Distribution to $Z$ and adopt a transformer-based network (denoted as $\mathcal{D}$ ) to recover the added noise. To inject text prompts as conditions, we use the BERT model [26] to extract a text embedding $e \in \mathbb{R}^{768}$ and fuse it with latent features through cross-attention layers. The objective to train the diffusion model follows the EDM [43, 92] and can be formulated as: + +$$ +L \left(\theta_ {\mathcal {D}}\right) = \mathbb {E} _ {\epsilon \in \mathcal {N} \left(0, \sigma^ {2} \mathbf {I}\right)} \| \mathcal {D} (Z + \epsilon , \sigma , e)) - Z \| _ {2} ^ {2} \tag {1} +$$ + +where $\theta_{\mathcal{D}}$ denotes the trainable parameters of the denoiser network $\mathcal{D}$ and $\sigma$ is a randomly chosen noise level. During inference, a randomly sampled noise vector is gradually denoised to a latent code $Z$ matching the given text prompt, which is then decoded to a garment mesh using the decoder in the VAE model introduced earlier. + +# 3.3. Gaussian Avatar Optimization + +To further obtain reasonable geometry of hair strands and body shape, we utilize HAAR [75] which is a text-conditioned diffusion model, and the BodyShapeGPT [6] which is a GPT based LLM model to predict hair strands and SMPL shape parameters from the given text. Given the garment mesh, hair strands, and SMPL body mesh, we attach 3D Gaussians to each component and optimize the attributes of these Gaussians using foundational diffusion models. We choose 3D Gaussians for texture modeling for two reasons. First, compared to implicit representations such as SDF, 3D Gaussians are more flexible for animation and can be easily driven by meshes [63, 87]. Second, 3D Gaussians have shown impressive capacity to capture high-fidelity appearance, especially when combined with diffusion models priors [53, 87]. Next, we discuss how to customize 3D Gaussians for each body part to ensure they maximally conform to the structure and motion of that part. + +3D Gaussians on body and garment meshes. We use a similar strategy to combine 3D Gaussians with the body and garment part since their coarse geometry is both represented as 3D meshes. For the body mesh, we utilize a SMPL [54] mesh $\Omega = \mathrm{LBS}(\theta ,\beta)$ , where $\theta$ and $\beta$ are the SMPL pose and shape parameters, and LBS is the linear blend skinning function. The garment mesh is generated and simulated as discussed in Sec. 3.2 and Sec. 3.1. Below we discuss how to attach and deform a 3D Gaussian to a mesh $\mathcal{M}$ , which could be a body or garment mesh. Specifically, we associate each Gaussian $G_{i} = \{\mu_{i},r_{i},s_{i},f_{i},o_{i}\}$ with a face of the mesh $\mathcal{M}$ and define its position $\mu_i\in \mathbb{R}^3$ , rotation $r_i\in \mathbb{R}^3$ , and scaling $s_i\in \mathbb{R}^3$ in the face's local coordinate, as well as its color features $f_{i}\in \mathbb{R}^{d_{c}}$ and opacity $o_i$ , where $d_{c}$ is the dimension of the Spherical Harmonic coefficients. The face's coordinate $\{P(\theta),R(\theta),k\}$ is defined similarly to [63], where the origin $P(\theta)\in \mathbb{R}^3$ is computed as the mean position of the face vertices, and the rotation matrix $R(\theta)\in \mathbb{R}^{3\times 3}$ is formed by concatenating one edge vector of the face, the normal vector, and their cross product. We also compute a scalar $k$ by the mean length of the edges. The global Gaussian position, rotation, and scale $\{\hat{\mu}_i,\hat{r}_i,\hat{s}_i\}$ can be computed by applying the local-to-global transform: + +$$ +\begin{array}{l} \hat {\mu} _ {i} (\theta) = k R (\theta) \cdot p _ {i} + P (\theta) \\ \hat {r} _ {i} (\theta) = R (\theta) \cdot r _ {i} \tag {2} \\ \hat {s} _ {i} (\theta) = k s _ {i} \\ \end{array} +$$ + +We initialize the 3D Gaussians by uniformly sampling points on the mesh surface, the face correspondences are maintained throughout the Gaussian densification process [44]. As observed in [87], directly optimizing Gaussian attributes with high-variance diffusion prior-based objective leads to noisy avatars. Thus, we adopt an implicit field $\mathcal{F}_{\phi}$ with parameters $\phi$ to model the Gaussian attributes. Specifically, we query the color features $f_{i}$ , opacity $o_i$ of each Gaussian using its global position $\hat{p}_i(\tilde{\theta})$ under a canonical pose $\tilde{\theta}$ by $(f_i,o_i) = \mathcal{F}_{\phi}(\hat{\mu}_i(\tilde{\theta}))$ . Note that we learn two separate implicit fields for the body $\mathcal{F}_{\phi}^{b}$ and garment $\mathcal{F}_{\phi}^{g}$ to prevent texture entanglement. The canonical garment mesh is a garment draped on the SMPL body in T-pose. As a result, the 3D Gaussians attached to the body or garment mesh can be smoothly driven by the avatar's motion by Eq. 2. + +3D Gaussians on hair strands. Different from the body and garment meshes, hair strands have a more delicate and complex structure. As a result, the Gaussians associated with the strands are expected to have a long and thin shape. Inspired by existing works that adopt 3D Gaussians for hair reconstruction [55, 90], we assign one 3D Gaussian to each line segment in each hair strand. As shown in the green circle in Fig. 2, for a line segment defined by two endpoints $l_{i}$ and $l_{i+1}$ , we assign a Gaussian denoted as $G_{i} = \{\mu_{i}, r_{i}, s_{i}, f_{i}, o_{i}\}$ to model the segment. To enforce + +the Gaussians to be consistent with the hair strands, we explicitly compute the position, rotation, and scale of $G_{i}$ by: + +$$ +\begin{array}{l} \mu_ {i} = \left(l _ {i} + l _ {i + 1}\right) / 2 \\ s _ {i} = \left[ \left\| l _ {i + 1} - l _ {i} \right\| / 2, \gamma , \gamma \right] \tag {3} \\ r _ {i} = \left[ 1 + \mu_ {i} \cdot d _ {i}, u \times d _ {i} \right], \\ \end{array} +$$ + +where $d_{i}$ is the direction of a hair strand, and $u$ are unit vectors along the $x$ -axis. We set $\gamma = 0.001$ to keep the hair Gaussians thin. To learn opacity $\{o_{i}\}$ and color $\{f_{i}\}$ for the hair Gaussians, we similarly train an implicit field $\mathcal{F}_{\phi}^{h}$ as discussed in the garment and body part above. During inference, the hair Gaussians can be driven by the hair strands smoothly and consistently using Eq. 3. + +Shading Model. Lighting plays an important role in modeling appearance details in motion such as garment wrinkles. To encourage the 3D Gaussians to capture pose-independent albedo without baked-in shading, we incorporate a Phong shading model [81] into our pipeline. Since the normal for each Gaussian is noisy, we instead use the normal of its corresponding face (denoted as $n_p$ ) in the lighting model. To mimic random lighting, we sample the point light position $l_p \in \mathbb{R}^3$ , color $l_c \in \mathbb{R}^3$ , as well as an ambient light color $l_a \in \mathbb{R}^3$ . The shaded color of each 3D Gaussian can then be computed by $f_i' = f_i \cdot (\max(0, n_p \cdot (l_p - \hat{\mu}_i) / \|l_p - \hat{\mu}_i\|) \cdot l_c + l_a)$ , where $\hat{\mu}_i$ is the global coordinate of the Gaussian computed by Eq. 2. + +Gaussian avatar optimization. To learn the implicit fields for the hair, body, and garment parts, we utilize the Score Distillation Sampling (SDS) [62] objective with a pre-trained text-to-image diffusion model [1] to supervise the Gaussian splatting-based rendering $I(\eta)$ of our model, where $\eta$ denotes all learnable parameters i.e., $\eta = \{\mathcal{F}_\phi^b,\mathcal{F}_\phi^g,\mathcal{F}_\phi^h\}$ . Given a text prompt $\mathcal{T}$ and the noise prediction $\hat{\epsilon} (I_t;\mathcal{T},t)$ of the diffusion model, SDS optimizes parameters $\eta$ by minimizing the difference between the noise $\epsilon$ added to the rendered image $I(\eta)$ and the predicted noise $\hat{\epsilon}$ by the diffusion model: + +$$ +\nabla_ {\eta} \mathcal {L} _ {\mathrm {S D S}} = E _ {t, \epsilon} \left[ w (t) (\hat {\epsilon} (I _ {t}; \mathcal {T}, t) - \epsilon) \frac {\partial I}{\partial \eta} \right], \qquad (4) +$$ + +where $t$ is the noise level, $I_{t}$ is the noised image, and $w(t)$ is a weighting function. Additionally, to prevent broken hairs caused by transparent Gaussians in the middle of a hair strand, we propose a regularization on the opacity of hair Gaussians. Our key intuition is that along a hair strand, the Gaussians closer to the scalp (e.g., hair roots) should have larger opacity values than the Gaussians further away from the scalp (e.g., hair ends). This way the optimization process is able to trim the extra hair to match the prompt while avoiding the broken hair artifacts. Specifically, given + +![](images/7dec87918b843dc3af6f9e01b9c734001fc545e40f781a70a8ba076bdca91326.jpg) +Figure 4. Qualitative comparison with the state-of-the-art methods. See the supplementary video for more comparisons. + +a hair point cloud $h_0 \in \mathbb{R}^{N_s \times N_l \times 3}$ as discussed in Sec. 3.1, we sample the opacity ( $o \in \mathbb{R}^{N_s \times N_l}$ ) of its associated Gaussians from the learned implicit fields (i.e., $\mathcal{F}_{\phi}^{h}$ discussed in Sec. 3.3) and add the following regularization: + +$$ +L _ {h a i r} = \frac {1}{N _ {s} N _ {l}} \sum_ {i = 1} ^ {N _ {s}} \sum_ {j = 2} ^ {N _ {l}} \left(o _ {i, j - 1} - o _ {i, j}\right) \tag {5} +$$ + +Our final objective is $L = L_{SDS} + \lambda_{hair}L_{hair}$ , where $\lambda_{hair}$ is empirically set to 1.0 in our experiments. + +# 4. Experiments + +# 4.1. Datasets. + +To learn the text-based garment diffusion model discussed in Sec. 3.2, we use the Garment Pattern Generator (GPG) dataset [48] and the CLOTH3D [10] dataset processed by [25]. For text prompts, we leverage the prompt annotations for the GPG dataset from [33]. Since the CLOTH3D dataset does not include paired prompt annotation, we render each garment on top of the SMPL body mesh and query the GPT4v [2] with predefined questions to generate a prompt for each garment mesh, describing its type, shape, length, and width. Together we build a dataset containing around 20000 meshes with paired prompt annotations. The dataset covers common garment types such as t-shirts, tank tops, jackets, shorts, pants, skirts, dresses, etc. + +# 4.2. Qualitative Evaluations. + +We compare our synthesized avatars with start-of-the-art text-driven 3D avatar generation methods, including Fantasia3D [18], TADA [51] and GAvatar [87]. First, we present qualitative comparisons of synthesized avatars in a canonical pose in Fig. 4. Our method not only captures diverse garment geometry (e.g., skirts, tops, and bottoms) but also produces superior texture compared to baseline models. Next, we demonstrate animated avatars in Fig. 5. Note that since Fantasia3D utilizes implicit representations, it cannot be readily animated. GAvatar and TADA model the human body and garment by a single geometry representation. For animation, they attach garments to the human body and use linear blending skinning for deformation. As a result, they fail to produce realistic motion for loose garments such as dresses. As shown by the first row of Fig. 5, when the avatar raises her legs, only our model produces physically plausible motion for the dresses, whereas both baselines undesirably separate the dresses into two pieces. We provide abundant animation comparisons in the supplementary video to showcase the difference. + +# 4.3.Quantitative Evaluations. + +User study. We follow existing works [46, 51, 87] and quantitatively evaluate SimAvatar through an extensive A/B + +![](images/1e9c8ce252f3384fa060eece4b5bce05125a639dab178c1df21eaf0172cc9ee4.jpg) +Emma Watson with shoulder-length wavy hair wearing a sleeveless dress of cocktail-length, the garment has a bold leopard print, with classic black and brown spots on a beige background. + +![](images/b111e0ac5a21f18ca1317720f2f827d6b4dcd016845f275753736d301cdc8113.jpg) + +![](images/22956a83725a0b1d5a1b3e5c61da29079d63681bf5a0145e2f5190b50bdb288d.jpg) + +Figure 5. Qualitative comparison of animated avatars. See the supplementary video for more animation comparisons. +![](images/8575eea8191072afb6d905075a53f479ac3a52f0a4870782d355ab71d02c7376.jpg) +Sandra Bullock with long shag hairstyle wearing a sleeveless tunic dress of normal length, the garment has a painterly floral watercolor pattern, featuring soft washes of pink, yellow, and lavender flowers. + +![](images/ad755512c55c743f3999227fac1df417ebb40aa35950c0710f51f0da60e7d311.jpg) + +![](images/33bcd9e2f6ed2cbbcd7203048976ef803ce1872b400bc48d8a66a82bdb7e51e5.jpg) + +
MetricsTADAFantasia3DGAvatar
Appearance preference89.5510087.03
Motion preference91.8710094.47
+ +Table 1. User study. We provide the preference percentage of our method against state-of-the-art methods. The higher score indicates that more users favor our method. SimAvatar significantly outperforms other methods in terms of user preference. We encourage looking at the supplementary video. + +user study. Specifically, we run SimAvatar and baseline methods on 22 prompts. For comparison, we generate a static 360 video of the avatar in A-posed and a video showing the avatar driven by a motion sequence. At each round, we present a pair of results by our method and one of the baselines to a user. For the static 360 video pairs, we ask the user to choose the method with a better appearance. For the animated video pairs, we ask the user to pay more attention to the motion. We collected 540 votes from 18 users and present the results in Table. 1. For both appearance and motion, SimAvatar is consistently favored by users, demonstrating its superior performance compared to baselines. + +# 5. Conclusions and Future Work + +In this work, we proposed a novel framework for generating simulation-ready avatars with layered hair and clothing from a text prompt. Unlike existing methods for 3D human + +avatar generation, our approach represents the human body, hair, and garments in a layered and simulation-ready structure. To this end, we first generated garment mesh, hair strands, and body shape parameters from text-conditioned generative models. We then optimized 3D Gaussians attached to each region to learn realistic appearance. The generated avatars can be seamlessly animated using physics simulators. Both qualitative and quantitative results demonstrate the effectiveness of our method in generating realistic, animatable clothed avatars, allowing cloth, and hair simulation with dynamic effects such as wrinkles and flowing hair, which surpasses state-of-the-art techniques. + +While SimAvatar provides avatars with significantly enhanced realism, there are still areas for further improvements. Currently, our hair and garment generation models are trained on specific datasets, which may bottleneck the diversity due to the used training data. In future work, we aim to explore methods for generating simulation-ready avatars that extend beyond the existing dataset diversity. Additionally, SimAvatar simulates garments and hair sequentially, which can fail in certain cases, such as avatars wearing hoods. Implementing joint simulation for hair and garments would allow us to accommodate more complex garments. Finally, accessories and footwear remain entangled with the body or garment layers; in the future, we plan to focus on generating fully disentangled avatars. + +# References + +[1] Realistic vision. https://huggingface.co/stablediffusionapi/realistic-vision-51, 2023.2,6 +[2] 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. 7 +[3] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In CVPR, 2018. 3 +[4] Thiemo Alldieck, Marcus Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single rgb camera. In CVPR, 2019. +[5] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2shape: Detailed full human body geometry from a single image. In ICCV, 2019. 3 +[6] Baldomero R Árbol and Dan Casas. Bodyshapegpt: Smpl body shape manipulation with llms. In ECCVW, 2024. 5 +[7] David Baraff and Andrew Witkin. Large steps in cloth simulation. In Conference on Computer Graphics and Interactive Techniques, 1998. 3 +[8] Miklós Bergou, Max Wardetzky, Stephen Robinson, Basile Audoly, and Eitan Grinspun. Discrete elastic rods. In SIGGRAPH, 2008. 4 +[9] Miklós Bergou, Basile Audoly, Etienne Vouga, Max Wardetzky, and Eitan Grinspun. Discrete viscous threads. ACM Transactions on graphics (TOG), 2010. 4 +[10] Hugo Bertiche, Meysam Madadi, and Sergio Escalera. Cloth3d: clothed 3d humans. In ECCV, 2020. 3, 5, 7 +[11] Hugo Bertiche, Meysam Madadi, and Sergio Escalera. PBNS: Physically based neural simulator for unsupervised garment pose space deformation. ACM Transaction on Graphics (TOG), 2020. +[12] Hugo Bertiche, Meysam Madadi, Emilio Tylson, and Sergio Escalera. Deepsd: Automatic deep skinning and pose space deformation for 3d garment animation. In ICCV, 2021. 3 +[13] Kiran S Bhat, Christopher D Twigg, Jessica K Hodgins, Pradeep Khosla, Zoran Popovic, and Steven M Seitz. Estimating cloth simulation parameters from video. SIGGRAPH, 2003. 3 +[14] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In ICCV, 2019. 3 +[15] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Loopreg: Self-supervised learning of implicit surface correspondences, pose and shape for 3d human mesh registration. In NeurIPS, 2020. 3 +[16] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. DreamAvatar: Text-and-Shape Guided 3D Human Avatar Generation via Diffusion Models. In CVPR, 2024. 2 +[17] Honghu Chen, Yuxin Yao, and Juyong Zhang. Neural-ABC: Neural parametric models for articulated body with clothes. + +IEEE Transactions on Visualization and Computer Graphics, 2024. 3, 5 +[18] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3D: Disentangling Geometry and Appearance for High-quality Text-to-3D Content Creation*. In ICCV, 2023. 2, 4, 7 +[19] Kwang-Jin Choi and Hyeong-Seok Ko. Stable but responsive cloth. In ACM SIGGRAPH Courses. 2005. 3 +[20] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. SMPlicit: Topology-aware generative model for clothed people. In CVPR, 2021. 3 +[21] Dalle2. https://openai.com/dall-e-2, 2022.2 +[22] Gilles Daviet. Simple and scalable frictional contacts for thin nodal objects. ACM Transactions on Graphics (TOG), 2020. 4 +[23] Gilles Daviet. Interactive hair simulation on thegpu using admm. In SIGGRAPH, 2023. 2, 3 +[24] Gilles Daviet. Interactive hair simulation on thegpu using admm. In SIGGRAPH, 2023. 4 +[25] Luca De Luigi, Ren Li, Benoit Guillard, Mathieu Salzmann, and Pascal Fua. DrapeNet: Garment Generation and Self-Supervised Draping. In CVPR, 2023. 3, 5, 7 +[26] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5 +[27] Junting Dong, Qi Fang, Zehuan Huang, Xudong Xu, Jingbo Wang, Sida Peng, and Bo Dai. TELA: Text to layer-wise 3d clothed human generation. In ECCV, 2024. 2, 3 +[28] Yun (Raymond) Fei, Henrique Teles Maia, Christopher Batty, Changxi Zheng, and Eitan Grinspun. A multi-scale model for simulating liquid-hair interactions. ACM Transaction on Graphics (TOG), 2017. 3 +[29] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. In NeurIPS, 2020. 3 +[30] Artur Grigorev, Bernhard Thomaszewski, Michael J Black, and Otmar Hilliges. HOOD: Hierarchical graphs for generalized modelling of clothing dynamics. In CVPR, 2023. 2, 3, 4 +[31] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. Drape: Dressing any person. SIGGRAPH, 2012. 3 +[32] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In ECCV, 2022. 3, 5 +[33] Kai He, Kaixin Yao, Qixuan Zhang, Jingyi Yu, Lingjie Liu, and Lan Xu. Dresscode: Autoregressively sewing and generating garments from text guidance. ACM Transactions on Graphics (TOG), 2024. 7 +[34] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. AvatarCLIP: Zero-Shot Text-Driven Generation and Animation of 3D Avatars. SIGGRAPH, 2022. 2 + +[35] Jerry Hsu, Tongtong Wang, Zherong Pan, Xifeng Gao, Cem Yuksel, and Kui Wu. Sag-free initialization for strand-based hybrid hair simulation. ACM Transaction of Graphics, 2023. 3 +[36] Shoukang Hu, Fangzhou Hong, Tao Hu, Liang Pan, Haiyi Mei, Weiye Xiao, Lei Yang, and Ziwei Liu. Humanliff: Layer-wise 3d human generation with diffusion model. arXiv preprint, 2023. 2, 3 +[37] Xin Huang, Ruizhi Shao, Qi Zhang, Hongwen Zhang, and Ying Feng. Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation. In CVPR, 2024. 2 +[38] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. In NeurIPS, 2023. 3 +[39] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In 3DV, 2024. +[40] Nikolay Jetchev. ClipMatrix: Text-controlled creation of 3d textured meshes. arXiv preprint arXiv:2307.05663, 2023. 2 +[41] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. Bcnet: Learning body and cloth shape from a single image. In ECCV, 2020. 3 +[42] Ruixiang Jiang, Can Wang, Jingbo Zhang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. AvatarCraft: Transforming text into neural human avatars with parameterized shape and pose control. In ICCV, 2023. 2 +[43] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 5 +[44] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. SIGGRAPH, 2023. 6 +[45] Taeksoo Kim, Byungjun Kim, Shunsuke Saito, and Hanbyul Joo. GALA: Generating animatable layered assets from a single scan. In CVPR, 2024. 3 +[46] Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Fieraru, and Cristian Sminchiescu. Dreamhuman: Animatable 3d avatars from text. In NeurIPS, 2023. 2, 7 +[47] Maria Korosteleva and Sung-Hee Lee. Generating datasets of 3d garments with sewing patterns. In NeurIPS, 2021. 5 +[48] Maria Korosteleva and Sung-Hee Lee. Generating datasets of 3d garments with sewing patterns. In NeurIPS, 2021. 7 +[49] Yifei Li, Tao Du, Kui Wu, Jie Xu, and Wojciech Matusik. Diffcloth: Differentiable cloth simulation with dry frictional contact. SIGGRAPH, 2022. 3 +[50] Junbang Liang, Ming C. Lin, and Vladlen Koltun. Differentiable cloth simulation for inverse problems. In NeurIPS, 2019. 3 +[51] Tingting Liao, Hongwei Yi, Yuliang Xiu, Jiaxiang Tang, Yangyi Huang, Justus Thies, and Michael J Black. TADA! text to animatable digital avatars. In 3DV, 2024. 2, 3, 4, 7 +[52] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, + +Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-Resolution Text-to-3D Content Creation. In CVPR, 2023. 2 +[53] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splattering. In CVPR, 2024. 2, 5 +[54] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ToG, 34(6):248:1-248:16, 2015. 2, 4, 6 +[55] Haimin Luo, Min Ouyang, Zijun Zhao, Suyi Jiang, Longwen Zhang, Qixuan Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Gaussianhair: Hair modeling and rendering with light-aware gaussians. arXiv preprint arXiv:2402.10483, 2024. 3, 6 +[56] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3d people in generative clothing. In CVPR, 2020. 3 +[57] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2 +[58] Xiaoyu Pan, Jiaming Mai, Xinwei Jiang, Dongxue Tang, Jingxiang Li, Tianjia Shao, Kun Zhou, Xiaogang Jin, and Dinesh Manocha. Predicting loose-fitting garment deformations using bone-driven motion networks. In SIGGRAPH, 2022. 3 +[59] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In CVPR, 2020. 3 +[60] Emmanuel Piuze, Paul G. Kry, and Kaleem Siddiqi. Generalized helicoids for modeling hair geometry. Computer Graphics Forum, 2011. 3 +[61] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J Black. ClothCap: Seamless 4d clothing capture and retargeting. SIGGRAPH, 2017. 3 +[62] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2, 3, 4, 6 +[63] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. In CVPR, 2024. 5, 6 +[64] Li Ren, Benoit Guillard, Edoardo Remelli, and Pascal Fua. DIG: Draping Implicit Garment over the Human Body. In ACCV, 2022. 3 +[65] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. SIGGRAPH, 2023. 2 +[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[67] 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. In NeurIPS, 2022. 2 + +[68] Igor Santesteban, Miguel A Otaduy, and Dan Casas. Learning-based animation of clothing for virtual try-on. Computer Graphics Forum, 38(2):355-366, 2019. 3 +[69] Igor Santesteban, Nils Thuerey, Miguel A. Otaduy, and Dan Casas. Self-supervised collision handling via generative 3d garment models for virtual try-on. In CVPR, 2021. 3 +[70] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-On. In CVPR, 2021. 3 +[71] Igor Santesteban, Miguel A Otaduy, and Dan Casas. SNUG: Self-Supervised Neural Dynamic Garments. In CVPR, 2022. 3 +[72] Igor Santesteban, Miguel A. Otaduy, Nils Thuerey, and Dan Casas. ULNeF: Untangled layered neural fields for mix-and-match virtual try-on. In NeurIPS, 2022. 3 +[73] Yuefan Shen, Shunsuke Saito, Ziyan Wang, Olivier Maury, Chenglei Wu, Jessica Hodgins, Youyi Zheng, and Giljoo Nam. CT2Hair: High-fidelity 3d hair modeling using computed tomography. Transactions on Graphics, 2023. 3 +[74] Vanessa Sklyarova, Jenya Chelishev, Andreea Dogaru, Igor Medvedev, Victor Lempitsky, and Egor Zakharov. Neural Haircut: Prior-Guided Strand-Based Hair Reconstruction. In ICCV, 2023. 3 +[75] Vanessa Sklyarova, Egor Zakharov, Otmar Hilliges, Michael J. Black, and Justus Thies. Text-conditioned generative model of 3d strand-based human hairstyles. In CVPR, 2024. 3, 5 +[76] Zhaoqi Su, Tao Yu, Yangang Wang, and Yebin Liu. Deepcloth: Neural garment representation for shape and style editing. TPAMI, 2023. 3 +[77] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splattering for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2 +[78] B. Thomaszewski, Simon Pabst, and Wolfgang Straβer. Asynchronous cloth simulation. 2008. 3 +[79] Pascal Volino, Nadia Magnenat-Thalmann, and Francois Faure. A simple approach to nonlinear tensile stiffness for accurate cloth simulation. Transaction of Graphics (TOG), 2009. 3 +[80] Huamin Wang, James F. O'Brien, and Ravi Ramamoorthi. Data-driven elastic models for cloth: Modeling and measurement. In SIGGRAPH, 2011. 3 +[81] Jionghao Wang, Yuan Liu, Zhiyang Dou, Zhengming Yu, Yongqing Liang, Xin Li, Wenping Wang, Rong Xie, and Li Song. Disentangled clothed avatar generation from text descriptions. arXiv preprint:2312.05295, 2023. 2, 3, 4, 6 +[82] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 34:27171-27183, 2021. 2 +[83] Yi Wang, Jian Ma, Ruizhi Shao, Qiao Feng, Yu-Kun Lai, Yebin Liu, and Kun Li. Humancoser: Layered 3d human generation via semantic-aware diffusion model. arXiv preprint: 2312.05804, 2023. 3 +[84] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and + +diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2 +[85] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv preprint arXiv:2405.14832, 2024. 5 +[86] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models. In Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[87] Ye Yuan, Xueting Li, Yangyi Huang, Shalini De Mello, Koki Nagano, Jan Kautz, and Umar Iqbal. Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. In CVPR, 2024. 2, 3, 4, 5, 6, 7 +[88] Cem Yuksel, Scott Schaefer, and John Keyser. Hair meshes. ACM Trans. Graph., 2009. 3 +[89] Ilya Zakharkin, Kirill Mazur, Artur Grigorev, and Victor Lempitsky. Point-based modeling of human clothing. In CVPR, 2021. 3 +[90] Egor Zakharov, Vanessa Sklyarova, Michael J Black, Giljoo Nam, Justus Thies, and Otmar Hilliges. Human hair reconstruction with strand-aligned 3d gaussians. In ECCV, 2024. 3, 6 +[91] Yifei Zeng, Yuanxun Lu, Xinya Ji, Yao Yao, Hao Zhu, and Xun Cao. Avatarbooth: High-quality and customizable 3d human avatar generation. arXiv preprint: 2306.09864, 2023. 3 +[92] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 2023. 5 +[93] Chao Zhang, Sergi Pujades, Michael J Black, and Gerard Pons-Moll. Detailed, accurate, human shape estimation from clothed 3d scan sequences. In CVPR, 2017. 3 +[94] Huichao Zhang, Bowen Chen, Hao Yang, Liao Qu, Xu Wang, Li Chen, Chao Long, Feida Zhu, Kang Du, and Min Zheng. Avatarverse: High-quality & stable 3d avatar creation from text and pose. In AAAI, 2024. 2, 3 +[95] Hao Zhang, Yao Feng, Peter Kulits, Yandong Wen, Justus Thies, and Michael J. Black. TECA: Text-guided generation and editing of compositional 3d avatars. In 3DV, 2024. 2, 3 +[96] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3 +[97] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 2024. 5 +[98] Xuanmeng Zhang, Jianfeng Zhang, Chacko Rohan, Hongyi Xu, Guoxian Song, Yi Yang, and Jiashi Feng. Getavatar: Generative textured meshes for animatable human avatars. In ICCV, 2023. 2 \ No newline at end of file diff --git a/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/images.zip b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d4fa1f222432065b78e5b9c7422b4b651527578e --- /dev/null +++ b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9d2e10d78d4a0d25cc1c068b6f09b738530af36149aa4bdb524a728d2287be7 +size 479067 diff --git a/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/layout.json b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..063135b8fe9c626824940b79b77514b6c155d035 --- /dev/null +++ b/CVPR/2025/SimAvatar_ Simulation-Ready Avatars with Layered Hair and Clothing/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ad718a284725d9c068adc0c74ddfac664ec8b454b8fc6cdea2497db6e077f13 +size 420608 diff --git a/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/ffdb5b83-0c2a-472c-9852-6c5f68369032_content_list.json b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/ffdb5b83-0c2a-472c-9852-6c5f68369032_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9fa8fc19a8941356d428d15e7b3a703438c18024 --- /dev/null +++ b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/ffdb5b83-0c2a-472c-9852-6c5f68369032_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8635be2483356cb523df2cbd71ce07b7991aac5ac3a4407de26e2745ed5ebc85 +size 78126 diff --git a/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/ffdb5b83-0c2a-472c-9852-6c5f68369032_model.json b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/ffdb5b83-0c2a-472c-9852-6c5f68369032_model.json new file mode 100644 index 0000000000000000000000000000000000000000..617c6ffe5fb057c1ea45e768a6dc8beacffbb54f --- /dev/null +++ b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/ffdb5b83-0c2a-472c-9852-6c5f68369032_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e457cf465ee3b6936a9a9a5d586075c288adc56710a70bd98ffd4ea7933f75d +size 95141 diff --git a/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/ffdb5b83-0c2a-472c-9852-6c5f68369032_origin.pdf b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/ffdb5b83-0c2a-472c-9852-6c5f68369032_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..95baec47e3af0bf2fce756af360fa772f995d44b --- /dev/null +++ b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/ffdb5b83-0c2a-472c-9852-6c5f68369032_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb13cfac002f2c4cc44fa904c94b0b30eaea60389ce8073c6634f02ee997c293 +size 2543068 diff --git a/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/full.md b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5f2c0b45a146f4c79097488a9d6b612127a35653 --- /dev/null +++ b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/full.md @@ -0,0 +1,295 @@ +# SimLTD: Simple Supervised and Semi-Supervised Long-Tailed Object Detection + +# Phi Vu Tran +LexisNexis Risk Solutions + +# Abstract + +While modern visual recognition systems have made significant advancements, many continue to struggle with the open problem of learning from few exemplars. This paper focuses on the task of object detection in the setting where object classes follow a natural long-tailed distribution. Existing methods for long-tailed detection resort to external ImageNet labels to augment the low-shot training instances. However, such dependency on a large labeled database has limited utility in practical scenarios. We propose a versatile and scalable approach to leverage optional unlabeled images, which are easy to collect without the burden of human annotations. Our SimLTD framework is straightforward and intuitive, and consists of three simple steps: (1) pre-training on abundant head classes; (2) transfer learning on scarce tail classes; and (3) fine-tuning on a sampled set of both head and tail classes. Our approach can be viewed as an improved head-to-tail model transfer paradigm without the added complexities of meta-learning or knowledge distillation, as was required in past research. By harnessing supplementary unlabeled images, without extra image labels, SimLTD establishes new record results on the challenging LVIS v1 benchmark across both supervised and semi-supervised settings. + +# 1. Introduction + +The task of detecting, localizing, and classifying object instances from image and video is a long-standing problem in computer vision. Recent years have seen unprecedented progress on modern object detection systems, mostly driven by powerful neural architectures. Much of this success is measured on the relatively balanced, small-vocabulary benchmarks like PASCAL VOC [13] and MS-COCO [22]. When evaluated on a more complex and imbalanced dataset with a much larger vocabulary, however, the same models exhibit a considerable drop in detection accuracy [15]. + +This paper explores ways to enhance the capability of commodity detection systems, with a particular evaluation emphasis on the challenging large-vocabulary LVIS benchmark [15]. LVIS represents a realistic application, a scenario in which object classes follow a natural long-tailed distribu + +![](images/a3be079b82442c4c0020e46dde0504157498f64d16a10392770c474c0e8ea73e.jpg) +Figure 1. A survey comparing our SimLTD to the state of the art for long-tailed detection. We combine unlabeled data with an intuitive multi-stage training strategy to deliver the best overall performance while also optimizing for accuracy on rare classes. Our simple approach achieves superior results on the challenging LVIS v1 benchmark without requiring auxiliary image-level supervision. + +tion. It is in the tail that most data-hungry models struggle with performance, a distribution characterized by many rare classes having as few as a single training exemplar. + +Figure 1 plots several recent state-of-the-art methods to address the extreme disparity of class distributions in long-tailed detection (LTD). One promising direction is to divide the problem into multiple manageable parts to help alleviate the severity of the class imbalance. LST [19] proposes to segment the overall dataset into seven phases, each phase containing progressively smaller but balanced data samples, and train a model in an incremental manner via network expansion and knowledge distillation. While LST adopts innovative ideas from class-incremental learning [4, 31, 40], the method is overly complex with the maintenance of many sub-parts. Moreover, the method requires numerous stages of knowledge transfer that can lead to catastrophic forgetting and result in an inferior solution [12]. + +Another interesting direction is to leverage external data to augment the training instances in LVIS. The intuition is that while the rare objects may not appear frequently in natural scenes, they can be found in abundance from object- + +![](images/23992ce44a863f5e2c6b1e69cb5ccc975d109e166a0eeee6f82766d634637a97.jpg) +Figure 2. The motivation to our approach. Left: We propose an improved head-to-tail model transfer framework for long-tailed detection by incorporating unlabeled images in both representation and transfer learning stages. Right: While it is possible to find more samples of LVIS instances from ImageNet, such an auxiliary database may not exist in another scenario like aerial imagery. Our framework does not depend on using extra image-level labels to advance long-tailed detection, which expands the applicability of our approach beyond LVIS. + +centric sources such as ImageNet [9]. By training on the union of scene-centric LVIS and object-centric ImageNet images, these weakly supervised methods [27, 45, 50] overcome the skewed class distribution by enriching the tail instances with additional whole-image labels. Although appealing, this approach places a hard requirement on having access to a large database of $\sim 14\mathrm{M}$ labeled images across $\sim 21\mathrm{K}$ object classes, an assumption not necessarily suitable for many practical settings outside of LVIS. + +Consider Figure 2 as an example for aerial imagery; there is no comparable object-centric dataset available to match the semantic concepts of "track field", "roundabout", or other categories in DOTA [11]. Moreover, for an industry application focusing on bespoke concepts beyond general objects, building a new object-centric dataset to supplement the main training dataset can be costly and time-consuming. There is also the open question of how to handle the genuinely rare classes (e.g., a rare fish species), which are strictly limited in observation and cannot be collected in more quantity from the open Internet. Can we still advance long-tailed detection without additional hand-labeling? + +We propose a simple and scalable framework, aptly named SimLTD, to answer this question. We deconstruct the LTD problem into three stages consisting of (i) pre-training on head classes, (ii) transfer learning on tail classes, and (iii) fine-tuning on a small sampled set composed of both head and tail classes. Unlike previous research on multi-stage training [12, 19], our framework allows for the optional use of unlabeled images to further boost LTD. We learn with supplementary unlabeled data in a semi-supervised manner via pseudo-labeling, and therefore do not explicitly rely on any additional instance-level or image-level supervision. + +Our work addresses several challenges associated with the head-to-tail model transfer paradigm [18, 39]. For one, we find that the vanilla transfer of data-rich head representations to data-scarce tail classes does not yield sufficient performance improvement because the distribution of head classes is also skewed. Moreover, the naive application of unlabeled data can aggravate the model's inductive bias because the unlabeled samples may follow a similar long-tailed + +distribution. In such case, the trained model will tend to generate more pseudo labels for the head classes, resulting in a larger degree of pseudo-class imbalance. We overcome these obstacles by incorporating data augmentations specifically designed to mitigate class imbalances in both head and tail training stages. Extensive experiments in §4.3 show that stronger pre-trained models on head classes, with and without unlabeled images, transfer well to the tail classes, leading to a desirable solution for long-tailed detection. + +Main Contributions (1) We present SimLTD, a general framework for effective supervised and semi-supervised LTD with unlabeled images. (2) The design of SimLTD is straightforward and intuitive, and is compatible with a range of backbones and detectors based on both classical convolutional and modern transformer architectures. (3) When put to the test against competing methods on the challenging LVIS v1 benchmark, SimLTD demonstrates excellent performance and scalability by establishing new state-of-the-art results with compelling margins. We hope SimLTD serves as a strong baseline for future research to tackle realistic long-tailed problems in the community. + +# 2. Related Work + +Multi-Stage Learning for LTD Multi-stage approaches typically begin with a step on representation learning followed by a transfer learning or fine-tuning routine. The earlier work of LST [19] pre-trained a model on a subset of many-shot head classes then transferred its representations to progressively smaller but balanced scarce tail samples via knowledge distillation. A more recent method [12] pre-trains a model on the entire long-tailed dataset, and opts to fine-tune the resulting model on "smooth-tail" data to mitigate the effects of class imbalance. These approaches assume pre-trained representations provide better initialization than random weights for fine-tuning on tail classes. + +In §3.1, we analyze that assumption to be generally true. We leverage this finding and take a different but related approach to pre-train on the frequent and common head instances. Then we simply copy the head representations + +![](images/dc4b88b52e75b401a57a91735a22355f77c4ab1b1caae4c831b1e7358817fc14.jpg) +Figure 3. Overview of SimLTD. Step 1 pre-trains the model and detector on head classes with unlabeled images. Step 2 transfers the head representations to tail classes. Step 3 fine-tunes the detector on a sampled set of head and tail classes. The "model" is an abstract block of encoder-decoder transformations based on either a convolutional or transformer network. The model is updated only during pre-training. + +and fine-tune them on the tail classes without resorting to knowledge distillation or meta-learning, as was required in prior studies [19, 39]. Our approach takes three stages of learning compared to the seven stages of LST, which can exacerbate catastrophic forgetting. A novel component to our approach is the introduction of unlabeled images in both pre-training and transfer learning stages for enhanced LTD. + +Leveraging Extra Data for LTD There is an emerging trend of using the abundant weak-supervision of ImageNet labels to solve the LVIS problem. Although LVIS is a detection dataset of natural scenes and ImageNet is a classification dataset of object images, they both share 997 overlapping classes between their vocabularies, the intersection of which provides a rich source of $\sim 1.5\mathrm{M}$ extra images to sample additional LVIS instances. To effectively learn from a mixed LVIS-ImageNet dataset with a domain gap, MosaicOS [45] leverages mosaic augmentation [1] whereas Detic [50] and RichSem [27] rely on the CLIP [30] classifier to map the semantic concepts between LVIS and ImageNet targets. + +As discussed in §1, these methods put a strict dependency on a large auxiliary labeled database to augment the main training dataset, which is infeasible for bespoke applications outside of LVIS. Alternatively, we propose a more general and flexible solution to use unlabeled data as a source of auxiliary supervision, which is easy to collect without the burden of human annotations. While our framework is not the first to leverage unlabeled data for LTD, ours is more effective when compared to the competing method of CascadeMatch [44], thanks to our multi-stage training strategy. + +Connection to Few-Shot Detection Long-tailed detection is related to the task of few-shot detection (FSD), the purpose of which is to adapt a base detector (trained on many-shot instances) to learn new concepts from few-shot exemplars. The commonality between the two is obvious—both tasks aim to boost detection on categories with very few training instances. However, LTD has unique challenges that extend beyond FSD. For LTD, the tail classes are authentically rare + +that follow a Zipf distribution [52] in natural scenes. By contrast, the novel few-shot exemplars in FSD datasets are randomly sampled and are not necessarily rare but can include objects of varying degrees of observational frequency. As such, the multi-stage training methods that work well for FSD [35, 38] cannot be directly applied to LTD with the same expected level of effectiveness. We need to devise ways to adapt these methods to the LTD problem to bring improvement over the state of the art. + +# 3. The SimLTD Framework + +As illustrated in Figure 2, given a long-tailed dataset $\mathcal{D}_{\mathrm{ld}}$ with $C$ categories, we split it into two disjoint subsets: $\mathcal{D}_{\mathrm{head}}$ with $C_{\mathrm{head}}$ frequent and common categories appearing in $>M$ images and $\mathcal{D}_{\mathrm{tail}}$ with $C_{\mathrm{tail}}$ rare classes appearing in $\leq M$ images. Furthermore, we have access to an unlabeled dataset $\mathcal{D}_{\mathrm{u}}$ of unknown class distribution. Our goal is to use a combination of labeled and unlabeled images to train a unified model optimized for accuracy on a test set comprising both classes in $C_{\mathrm{head}} \cup C_{\mathrm{tail}}$ . + +Our SimLTD framework consists of three easy steps: (i) representation learning on $\mathcal{D}_{\mathrm{head}}$ , (ii) transfer learning on $\mathcal{D}_{\mathrm{tail}}$ , and (iii) fine-tuning on $\mathcal{D}_k$ , a reduced dataset composed of $k$ instances per class randomly sampled from $\mathcal{D}_{\mathrm{ltd}}$ . Note that $\mathcal{D}_{\mathrm{head}}$ , $\mathcal{D}_{\mathrm{tail}}$ , and $\mathcal{D}_k$ are all still imbalanced, but not as severe as the original long-tailed $\mathcal{D}_{\mathrm{ltd}}$ . We leverage optional unlabeled images in both Steps 1 and 2 but do not explicitly need them for effective LTD. Indeed, experiments in §4.3 show that our fully supervised baselines exhibit excellent performance and scalability even without unlabeled data. A diagram of SimLTD is depicted in Figure 3. + +# 3.1. The Devil is in the $\mathcal{D}_{\mathrm{tail}}$ + +The open challenge of the LTD problem remains in learning an effective model on the few exemplars associated with the tail classes. We draw inspiration from existing empirical evidence that few-shot learning can vastly benefit from pre-trained representations [35, 38]. With that in mind, we + +![](images/39f4d3ed631498e8aeb2300c4734774b42293c43d98ee65a2c348b92d1b31cbd.jpg) +Figure 4. Transfer learning from COCO representations (solid bars) helps improve rare class detection on LVIS (hatched green bars). + +conduct an empirical analysis to verify whether our intuition extends to the LVIS setting. Following conventional LVIS protocol, we partition the dataset into $C_{\mathrm{head}} = 866$ common classes appearing in $M > 10$ images and $C_{\mathrm{tail}} = 337$ rare classes appearing in $M \leq 10$ images. We aim to improve the detection performance on $C_{\mathrm{tail}}$ using various commodity detectors pre-trained on the COCO dataset. + +Figure 4 quantifies the effectiveness of pre-trained representations on $C_{\mathrm{tail}}$ detection. For each model, we chop off the detection head consisting of the bounding box classifier and regressor modules learned on the COCO dataset, re-initialize them with random values, and perform transfer learning on the LVIS tail classes. We update only the box classifier and regressor while keeping the rest of the architecture frozen, essentially treating the pre-trained model as a fixed detector. Besides the pre-trained networks, we also assess the Faster R-CNN detector [32] initialized from scratch, except for the pre-trained backbone, as the lower-bound baseline. + +Discussion We observe a clear trend indicating stronger pre-trained representations, as measured by the AP score on COCO, generally lead to improved rare class detection. This result is both intriguing and encouraging since the models were pre-trained on COCO, a dataset of different scope and size than LVIS. Training from scratch brings out the worst performance with half of the accuracy. The implication of this simple experiment is two-fold: (1) we corroborate prior studies by showing low-shot learning can be improved with transferred representations; and (2) our framework opens opportunities to self-supervised, semi-supervised, and multimodal learning, all of which have demonstrated significantly better performance than supervised pre-training. Motivated by these insights, we propose to learn powerful representations on $\mathcal{D}_{\mathrm{head}}$ and transfer them to $\mathcal{D}_{\mathrm{tail}}$ for long-tailed detection. While the past attempts at head-to-tail model transfer [12, 19, 39] could only work by incorporating an extra module for meta-learning or knowledge distillation, we now describe our three-step approach to accomplish this goal without the unnecessary complexities. + +
AugmentationSupervised TrainingSemi-Supervised Training
Resizeshort edge ∈ [400, 1200]short edge ∈ [400, 1200]
Fliphorizontalhorizontal
SCP with RFSsample threshold = 0.001sample threshold = 0.001
AutoContrast✓✓
Equalize✓✓
Solarize✓✓
Color Jittering✓✓
Contrast✓✓
Brightness✓✓
Sharpness✓✓
Posterize✓✓
TranslationX(x,y) ∈ (−0.1, 0.1)
ShearingX(x,y) ∈ (−30°, 30°)
RotationXangle ∈ (−30°, 30°)
CutoutXn ∈ [1,5], size ∈ [0.0,0.2]
+ +Table 1. Summary of the data augmentations explored in this study to improve supervised and semi-supervised training. + +# 3.2. Step 1: Representation Learning on $\mathcal{D}_{\mathrm{head}}$ + +We begin with the supervised setting in which we have training data points $(x_{i},y_{i})\in \mathcal{D}_{\mathrm{head}}$ , where $x_{i}$ denotes the $i$ -th input image and $y_{i}$ is the $i$ -th ground truth annotation containing the box label and coordinates. Let $\Psi_{\mathrm{det}}(\mathcal{D}_{\mathrm{head}})$ be a learnable detection function training on the image-target pairs to produce the supervised loss $\mathcal{L}_{\mathrm{sup}}$ : + +$$ +\Psi_ {\det } \left(\mathcal {D} _ {\text {h e a d}}\right) \leftarrow \mathcal {L} _ {\sup } = \sum L _ {\operatorname {c l s}} \left(h \left(x _ {i}\right), y _ {i}\right) + L _ {\operatorname {r e g}} \left(h \left(x _ {i}\right), y _ {i}\right). \tag {1} +$$ + +Here, $h(x_{i})$ denotes the forward pass on the input image and $(L_{\mathrm{cls}}, L_{\mathrm{reg}})$ are classification (e.g., cross-entropy) and regression (e.g., $L_{1}$ ) losses for the detector. We consider commodity convolutional and transformer-based networks for $\Psi_{\mathrm{det}}$ . For the convolutional network, we experiment with Faster R-CNN to compare against previous studies using similar detectors (i.e., Mask R-CNN [17] and RetinaNet [24]). For the contemporary transformer-based network, we adopt the improved variants of the Detection Transformer (DETR) [3], namely Deformable DETR [51] and DINO [47]. + +As summarized in Table 1, we leverage well-known data augmentations to train $\Psi_{\mathrm{det}}$ which include random image resizing, horizontal flipping, and photometric distortion. We also combine Simple Copy-Paste (SCP) [14] together with Repeat Factor Sampling (RFS) [15] to combat the class imbalance in $\mathcal{D}_{\mathrm{head}}$ at both the image and instance levels, analogous to prior research on the importance of image and instance resampling for LTD [37, 42]. Taking a step further to learn even stronger representations on $\mathcal{D}_{\mathrm{head}}$ we propose to train $\Psi_{\mathrm{det}}$ in a semi-supervised manner with unlabeled images by way of pseudo-labeling, which is currently leading the literature on semi-supervised detection. We explore three successful methods of increasing effectiveness to advance semi-supervised representation learning on $\mathcal{D}_{\mathrm{head}}$ : SoftER Teacher [35], MixTeacher [25], and MixPL [7]. + +SoftER Teacher builds upon the end-to-end pseudolabeling approach of Soft Teacher [41] to include an auxiliary loss for consistency learning on region proposals, and was + +![](images/8c7506de01b5983464a60bf6411a10afacef88b4a9f6a123634ab016207326c9.jpg) +Figure 5. Left: Augmenting unlabeled images by randomly pasting rare instances from the training set helps promote pseudo-labeling for semi-supervised LTD. Middle: The pasted objects (cyan boxes) often come from contrasting environments to create complex fake scenes for the student to learn, which in turn can improve robustness and generalization on natural scenes (green boxes) shown on the Right. + +![](images/5b50217405fce18e1c227cd381cb4bf7d2b4769999e23d680cc4aeb4031c5641.jpg) + +![](images/4b130d9ed75637a2fea421147f3e4a8493e2134e3da47dafc3a7f5283ff10e13.jpg) + +![](images/e408eb26df636abd683a7cb153b06add7f12e933a0dc5cc74867495c9854cafe.jpg) + +shown to work particularly well with semi-supervised few-shot detection. MixTeacher introduces a mixed scale feature pyramid to generate more accurate pseudo labels on objects with extreme scale variations, resulting in an overall robust detector. While SoftER Teacher and MixTeacher primarily work with two-stage detectors like Faster R-CNN, MixPL opens the door to semi-supervised learning with single-stage and DETR-based models by integrating mixup [46] and mosaic [1] augmentations with pseudo labels. All three methods follow the student-teacher semi-supervised training paradigm [34], in which the teacher is an exponential moving average of the student. In the semi-supervised setting, the models learn from a joint dataset of labeled $\mathcal{D}_{\mathrm{head}}$ and unlabeled $\mathcal{D}_{\mathrm{u}}$ images via the following compound objective: + +$$ +\Psi_ {\text {s e m i - d e t}} \left(\mathcal {D} _ {\text {h e a d}}, \mathcal {D} _ {\mathrm {u}}\right) \leftarrow \mathcal {L} = \mathcal {L} _ {\sup } + \alpha \mathcal {L} _ {\text {p s e u d o}}, \tag {2} +$$ + +where $\alpha > 0$ controls the contribution of the pseudo-label loss derived from unlabeled data. The functional form of $\mathcal{L}_{\mathrm{pseudo}}$ is the same as $\mathcal{L}_{\mathrm{sup}}$ in Equation (1), except the ground-truth targets $y$ are replaced by pseudo targets $\hat{y}$ predicted by the teacher model during self-training. + +# 3.3. Step 2: Transfer Learning on $\mathcal{D}_{\mathrm{tail}}$ + +We instantiate the tail models for transfer learning by copying the parameters from the pre-trained head models. Let $\Psi_{\mathrm{det}}^{\prime}(\mathcal{D}_{\mathrm{tail}}) \gets \Psi_{\mathrm{det}}(\mathcal{D}_{\mathrm{head}})$ be the supervised tail model and $\Psi_{\mathrm{semi-det}}^{\prime}(\mathcal{D}_{\mathrm{tail}}, \mathcal{D}_{\mathrm{u}}) \gets \Psi_{\mathrm{semi-det}}$ be the semi-supervised counterpart. We train the tail models the same way per Equations (1) and (2), except we update only the classifier and regressor modules, to adapt them to tail classes, while freezing the rest of the networks. The intuition is that pre-trained representations serve as a bootstrapped initializer to train an accurate tail model, according to our analysis in §3.1. + +Recall that, unlike common head objects, the tail classes are intrinsically rare so one cannot expect to find abundant occurrences of them in either labeled or unlabeled source. This raises a hurdle for when we wish to train $\Psi_{\mathrm{semi - det}}^{\prime}$ with unlabeled images: there are very few instances of the rare classes in the unlabeled scenes for the teacher model to + +Algorithm 1 PyTorch Pseudocode for Head-Tail Class Fusion. +```python +HEAD_IDS : sorted list of head IDs, length 866 +TAIL_IDS : sorted list of tail IDs, length 337 +head_ckpt: model checkpoint on head classes +tail_ckpt: model checkpoint on tail classes +ALL_IDS = sorted(HEAD_IDS + TAIL_IDS) # length 1203 +ID2LABEL = { +ID: label for label, ID in enumerate(ALL_IDS) +} # mapping from category ID to integer label +head_DET = head_ckpt["state_dict"]["detector"] +tail_DET = tail_ckpt["state_dict"]["detector"] +fused_DET = torch rand(len(ALL_IDS)) +for label, ID in enumerate(HEAD_IDS): + fused_DET[ID2LABEL[ID]] = head_DET[label] +for label, ID in enumerate(TAIL_IDS): + fused_DET[ID2LABEL[ID]] = tail_DET[label] +head_ckpt["state_dict"]["detector"] = fused_DET +torch.save(head_ckpt, save Filename) # to fine-tune +``` + +propose reliable pseudo targets. We sidestep this hurdle by copying and pasting a random subset of rare instances from the labeled training set to the unlabeled images—a new procedure unique to this work. At each training iteration, the teacher model is guaranteed to see an augmented view of sampled rare objects, amid diverse background scenes, which promotes pseudo-label supervision for the student model. Figure 5 illustrates the impact of this technique on pseudo-labeling along with some examples of the augmented unlabeled images, which are subjected to strong photometric and geometric perturbations with cutout [10, 49]. Although the procedure inevitably leads to redundancy and overfitting, our ablation experiments in §4.5 show that it is surprisingly helpful in adapting head representations to the tail models. + +# 3.4. Step 3: Fine-Tuning on $\mathcal{D}_k$ + +At this stage, we have two separate models with a shared representation, one optimized on head classes and the other on tail classes. We wish to unify the two models into one for efficient single-pass inference on test samples containing both head and tail categories. Algorithm 1 provides the pseudocode for the merging scheme referenced in Figure 3. + +Note that we merge parameters at the detector module and + +
MethodExtra External DataBase DetectorBackboneGPU Hrs\( \mathrm{mAP}_{\text{box}} \)\( \mathrm{AP}_{\text{r}} \)\( \mathrm{AP}_{\text{c}} \)\( \mathrm{AP}_{\text{f}} \)
Seesaw Loss [CVPR'21] [36]Faster R-CNNR101-FPN-27.818.727.032.8
NorCal [NeurIPS'21] [28]Mask R-CNNR101-FPN-28.120.826.530.9
AHRL [CVPR'22] [20]Mask R-CNNR101-FPN-28.719.327.631.4
EFL [CVPR'22] [21]N/ARetinaNetR101-FPN-29.223.527.433.8
LogN [IJCV'24] [48]Mask R-CNNR101-FPN-29.822.928.831.8
SimLTD Supervised BaselineFaster R-CNNR101-FPN8331.724.332.134.6
SimLTD Supervised BaselineDeformable DETRR10123437.031.936.639.6
Dong et al. [ICCV'23] [12]Deformable DETRR50-28.721.828.432.0
Detic [ECCV'22] [50]Deformable DETRR50-31.721.430.737.5
RichSem [NeurIPS'23] [27]N/ADINOR50-35.126.032.641.8
SimLTD Supervised BaselineFaster R-CNNR50-FPN8129.020.929.731.9
SimLTD Supervised BaselineDeformable DETRR5021535.032.034.037.5
RichSem [NeurIPS'23] [27]DINOSwin-T-38.830.836.445.0
SimLTD Supervised BaselineDINOSwin-T31041.133.640.145.4
DiffusionDet [ICCV'23] [6]N/A4 @ 300Swin-B-42.034.840.946.4
RichSem [NeurIPS'23] [27]DINOSwin-B-46.438.545.151.3
SimLTD Supervised BaselineDINOSwin-B41447.242.746.749.9
MosaicOS [ICCV'21] [45]ImageNet-1K LabelsFaster R-CNNR50-FPN-23.915.522.429.3
RichSem [NeurIPS'23] [27]ImageNet-21K LabelsFaster R-CNN + CLIPR50-FPN-30.627.629.732.9
SimLTD SoftER Teacher [35]COCO-unlabeled2017Faster R-CNNR50-FPN39230.323.330.333.3
SimLTD MixTeacher [25]COCO-unlabeled2017Faster R-CNNR50-FPN43431.823.432.135.1
Detic [ECCV'22] [50]ImageNet-21K LabelsDeformDETR + CLIPR50-32.526.231.336.6
RichSem [NeurIPS'23] [27]ImageNet-21K LabelsDINO + CLIPR50-37.129.935.642.0
SimLTD MixPL [7]Objects365-unlabeledDINOR5044738.330.837.542.6
SimLTD MixPL [7]COCO-unlabeled2017DINOR5044638.531.637.542.5
RichSem [NeurIPS'23] [27]ImageNet-21K LabelsDINO + CLIPSwin-T-41.637.339.745.5
SimLTD MixPL [7]COCO-unlabeled2017DINOSwin-T48242.535.742.445.6
RichSem [NeurIPS'23] [27]ImageNet-21K LabelsDINO + CLIPSwin-B-48.246.546.551.0
SimLTD MixPL [7]COCO-unlabeled2017DINOSwin-B79449.043.449.051.5
+ +Table 2. Main results on LVIS v1 validation. GPU hours denote the wall clock time to train for a total of 640K iterations and are a proxy measure of model complexity. The ResNet and Swin backbones were pre-trained on ImageNet-1K and ImageNet-22K, respectively. The results of Seesaw Loss and Detic are borrowed from EFL and RichSem, respectively. Shaded rows indicate our implemented models. + +reuse the pre-trained head representations for the rest of the network. We fine-tune the unified detector on $\mathcal{D}_k$ composed of $k$ instances, or shots, per class sampled from the long-tailed training set. We update only the box classifier and regressor with a reduced learning rate to slowly adapt them to tail classes while preserving the pre-trained accuracy on head classes. Analogous to class-incremental learning, we include both head and tail classes in $\mathcal{D}_k$ for exemplar replay to avoid catastrophic forgetting. We form $\mathcal{D}_k$ via random sampling whereas prior work resorted to a complicated scheme of confidence-guided exemplar replay [12]. + +# 4. Empirical Study + +# 4.1. Evaluation Protocol + +We benchmark our approach on the challenging LVIS v1 dataset, which has 100170 training and 19809 validation images over 1203 classes. We use the standard LVIS evaluator to compute the detection metric $\mathrm{mAP}_{\mathrm{box}}$ for all classes, without test-time augmentation, along with $\mathrm{AP_r}$ , $\mathrm{AP_c}$ , and $\mathrm{AP_f}$ for the rare, common, and frequent categories. We sample + +$\mathcal{D}_k$ three times with different random seeds and report on the averaged metrics to capture statistical variability. Following prior studies, we focus on the performance gains of $\mathrm{mAP}_{\mathrm{box}}$ and $\mathrm{AP_r}$ in our comparative analysis. + +# 4.2. Implementation Details + +We implement our models using PyTorch [29] and MMDetection [5], and train on $8 \times$ A6000 GPUs. We pre-train in Step 1 for up to 540K iterations, perform transfer learning in Step 2 for 40K iterations, and fine-tune for 80K iterations. See our open-source code for the full reproducible details. + +High-Quality Supervised Baseline We construct a stronger supervised baseline than previously by combining our multi-stage training recipe with diverse data augmentations. We explore various ResNet [16] and Swin [26] backbones, along with FPN [23], for feature extraction. The supervised baseline is important to our framework because it serves as the basis for the teacher model to propose reliable pseudo targets for semi-supervised learning. + +Semi-Supervised LTD We leverage SoftER Teacher, Mix- + +![](images/98a9a63e50c0083ca2ec9063af345064fdc110952bb52997c8492078da854310.jpg) +$k = 1$ shot + +![](images/26d5ec13d3e20988270e7a8597383c8608c001c8e9b96f96c69cab1f0d77b9ca.jpg) +$k = 3$ shots + +![](images/5b3c41b450db82c1e96de2a5723024b2cc5c6ed125eaa97000439d4bbc09ac1d.jpg) +$k = 4$ shots + +![](images/788c82d3aea1ead5336dc246a99c82070d930cda3bbc6353f46fd3930e7be012.jpg) +$k = 5$ shots + +![](images/b17462190ef4483f56e52eba96e2d5c83e65adaacf1d7468dcd686553a72e1a9.jpg) +Figure 6. SimLTD detections on LVIS v1 validation images. We highlight visualizations containing truly rare $k$ -shot exemplars from the training set—drawn with red and white boxes. SimLTD does well in the extremely low-shot regime using as few as a single training instance. +Figure 7. Evaluation on Objects365 binned by the count of training images per class group. All models use the ResNet-50 backbone. + +Teacher, and MixPL for semi-supervised LTD, and inherit all hyper-parameters originally tuned on the COCO dataset without changes. We harness $\sim 123\mathrm{K}$ COCO-unlabeled2017 images to improve both representation and transfer learning in Steps 1 and 2. We also experiment with Objects365 [33] to further validate our approach on another related domain with $\sim 1.7\mathrm{M}$ unlabeled images in the wild by removing all label information from the training set. + +# 4.3. Main Results + +Table 2 reports on the effectiveness of our approach against existing methods representing the state of the art on LVIS. Our SimLTD supervised baseline with Faster R-CNN outperforms all methods using related detectors. The gains are convincing, with margins up to $+3.9\mathrm{AP}_{\mathrm{box}}$ and $+5.6\mathrm{AP}_{\mathrm{r}}$ . We observe a similar trend when comparing our supervised baseline using DETR-based models. SimLTD demonstrates compelling performance and scalability across a multitude of backbones and detectors without the need for extra data. + +For the methods requiring external data, our semi-supervised models also deliver impressive performance. When equipped with MixTeacher and Faster R-CNN, SimLTD exceeds the competition by up to $+7.9\mathrm{AP}_{\mathrm{box}}$ while being competitive on $\mathrm{AP_r}$ . Furthermore, SimLTD scales well by coupling with MixPL and transformer-based models to achieve new state-of-the-art results from harnessing only unlabeled images. SimLTD works equally well with both COCO and Objects365 unlabeled images, signifying that our + +
MethodBase DetectorBackbone\( AP_{\text{fixed}} \)box\( AP_{\text{r}} \)\( AP_{\text{c}}^{fix} \)\( AP_{\text{f}}^{fix} \)
CascadeMatch (Sup)Cascade R-CNN27.120.326.131.1
FASA (Supervised)Mask R-CNNR101-FPN28.222.028.330.9
SimLTD SupervisedFaster R-CNN32.724.932.935.9
CascadeMatchCascade R-CNN30.523.129.734.7
SimLTD SoftER TeacherFaster R-CNNR50-FPN31.324.131.134.6
SimLTD MixTeacherFaster R-CNN32.824.532.936.4
CascadeMatchCascade R-CNN32.926.531.836.8
SimLTD SoftER TeacherFaster R-CNNR101-FPN33.026.132.636.4
SimLTD MixTeacherFaster R-CNN34.426.134.238.2
+ +Table 3. Performance comparison between SimLTD, FASA [43], and CascadeMatch [44] using the alternative APFixed metric [8]. + +model can extract meaningful pseudo-label supervision from a large uncurated database with a distribution different from the training dataset. Figure 6 visualizes example SimLTD detections on select LVIS validation images. + +Ancillary to the main LVIS results, we also evaluate on Objects365 to showcase the generality of our framework. Following NorCal [28], we sample $30\%$ of the training set and split it into $C_{\mathrm{head}} = 332$ classes appearing in $M > 100$ images and $C_{\mathrm{tail}} = 33$ classes appearing in $M \leq 100$ images. SimLTD outperforms the existing baseline across the board in Figure 7. We fine-tune SimLTD with $k = 30000$ shots, a parameter that changes by dataset. Section 4.5 gives a detailed ablation analysis regarding the impact on AP from varying the number of shots for fine-tuning with $\mathcal{D}_k$ . + +# 4.4. Comparison to the State of the Art + +SimLTD vs. CascadeMatch [44] Although both SimLTD and CascadeMatch leverage COCO-unlabeled2017 for semi-supervised LTD, there are major differences between the two. First, CascadeMatch is trained end-to-end in a single stage whereas we take the decoupled approach. Second, CascadeMatch adopts the stronger Cascade R-CNN [2] compared to Faster R-CNN in SimLTD. CascadeMatch follows the APFixed protocol [8], which replaces the standard maximum 300 detections per image by a cap of 10K detections per class from the entire validation set. Despite the disadvantage of a simpler model, Table 3 shows that SimLTD outperforms CascadeMatch by notable margins in almost every measure. + +
ConfigurationmAPboxAPr
Single-Stage Training +++Copy-Paste28.116.8
Multi-Stage Baseline (w/ Random Resize)26.817.9
Multi-Stage Baseline +Photometric Jittering26.918.2
Multi-Stage Baseline ++Repeat Sampling27.119.3
Multi-Stage Baseline +++Copy-Paste (Ours)29.020.9
+ +Table 4. Ablation experiments quantifying the effectiveness of each component in our SimLTD supervised baseline. The single-stage model is trained end-to-end on the whole long-tailed dataset. + +These superior results lend further support to the merit of our multi-stage training strategy. + +SimLTD vs. Dong et al. [12] on Multi-Stage Learning We compare our multi-stage training strategy to that of Dong et al., which also utilizes a three-step procedure of pre-training followed by fine-tuning and knowledge distillation. We focus our analysis on their powerful Deformable DETR model, which yields similar results to our simpler Faster R-CNN model. When we train with the same capable Deformable DETR architecture, Table 2 shows that SimLTD exceeds their model by outsized margins of $+6.3\mathrm{AP}_{\mathrm{box}}$ and $+10.2\mathrm{AP}_{\mathrm{r}}$ . These remarkable gains are directly attributed to our multi-stage learning approach, which is carefully designed to optimize for accuracy on both head and tail classes. + +SimLTD vs. RichSem [27] on Using Extra Data Recall that RichSem relies on the CLIP classifier (pre-trained on $\sim 400\mathrm{M}$ image-caption pairs) and image-supervision from an additional $\sim 1.5\mathrm{M}$ images to produce the state-of-the-art results reported in Table 2. However, such strict dependencies are impractical when the method is applied to a bespoke dataset outside of generic objects. By contrast, our SimLTD leverages unlabeled images, without resorting to either CLIP or auxiliary ImageNet labels, to deliver better results than RichSem with the ResNet backbone and slightly worse $\mathrm{AP_r}$ with the Swin backbones. When we remove external data, SimLTD substantially outperforms RichSem by $+6.0\mathrm{AP_r}$ in the fully supervised setting, implying that the success of RichSem is sensitive to the contributions of CLIP and ImageNet supervision. Our framework carries the benefit of being robust across settings with and without external data. + +# 4.5. Ablation Experiments + +Design of SimLTD Baseline The design of SimLTD is centered on an intuitive multi-stage training strategy combined with standard data augmentations, without bells and whistles, to overcome the class imbalances in both head and tail datasets. Table 4 shows the contributions of RFS and Copy-Paste in establishing a more robust baseline than was previously possible in the existing literature, by using the simple Faster R-CNN detector with ResNet-50 FPN. We + +![](images/0d967f73e078cf7f820918ae045c4abbb4f025a0f23940e51d9174fca8fbe3ef.jpg) +Figure 8. Ablation experiments assessing the impact on AP from transfer learning on tail classes (left) and from varying the number of sampled shots for fine-tuning with $\mathcal{D}_k$ (right). + +![](images/25d1a306dac072b4d161cfaa291104ff7e8c38457ecb0b02ed145ec9272c41aa.jpg) + +also showcase the viability of multi-stage learning over the naive single-stage training procedure on the whole long-tailed dataset, which yields markedly worse results. + +The Impact of Transfer Learning on $\mathcal{D}_{\mathrm{tail}}$ As discussed in §3.3, we transfer the pre-trained head representations to optimize on tail classes in Step 2 of our framework. However, this step is optional and may be skipped, because we can initialize the tail classes with random values before fin-tuning. Figure 8 shows that transfer learning on tail classes is a worthwhile task. Across both supervised and semi-supervised settings, transfer learning gives a boost by up to $+4.7\mathrm{AP_r}$ and comes with the added bonus of shortening the fine-tuning time by $2/3$ of the required iterations. + +How Many Shots to Sample for $\mathcal{D}_k$ ? Figure 8 illustrates the impact on $\mathrm{mAP}_{\mathrm{box}}$ and $\mathrm{AP_r}$ as a function of sampled shots for fine-tuning with $\mathcal{D}_k$ , ranging from 10 to "All" meaning the entire long-tailed training set. The aim of this experiment is to optimize for accuracy on rare classes while mitigating catastrophic forgetting on pre-trained head representations. We analyze the "knee in the curve" and find that 30-shots balance the trade-off between the two metrics. Figure 3 visualizes that with 30-shots, we sample the whole tail distribution containing 20 or fewer instances per class and include a mixture of head categories for exemplar replay. Moving to the left of this "sweet spot" with {10, 20}-shots, we observe marked reductions in $\mathrm{mAP}_{\mathrm{box}}$ , indicating adverse forgetting on head classes from insufficient samples. Moving to the right of the sweet spot, we see a precipitous drop in $\mathrm{AP_r}$ in response to the overwhelming amount of head samples. + +# 5. Conclusion + +We introduced SimLTD, a simple and versatile framework for supervised and semi-supervised long-tailed detection. Standing out from existing work, SimLTD delivers excellent performance and scalability by achieving new state-of-the-art results on the challenging LVIS v1 benchmark, without requiring auxiliary image-level supervision. We hope the practitioner finds utility in our multi-stage training approach and for our work to spur future research aimed at pushing the performance envelope of long-tailed detection. + +# Acknowledgments + +The author thanks Victor Palmer for the insightful discussion, Brian Keller, Andrew Melkonian, Mehran Ghandehari, Elan Ding, and anonymous reviewers for their thoughtful and constructive feedback on this paper. + +# References + +[1] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. YOLOv4: Optimal Speed and Accuracy of Object Detection. https://arxiv.org/abs/2004.10934, 2020. 3, 5 +[2] Zhaowei Cai and Nuno Vasconcelos. Cascade R-CNN: Delving into High Quality Object Detection. In CVPR, 2018. 7 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-End Object Detection with Transformers. In ECCV, 2020. 4 +[4] Francisco M. Castro, Manuel J. Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-End Incremental Learning. In ECCV, 2018. 1 +[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, et al. MMDetection: Open MMLab Detection Toolbox and Benchmark. https://arxiv.org/abs/1906.07155, 2019. 6 +[6] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. DiffusionDet: Diffusion Model for Object Detection. In ICCV, 2023. 6 +[7] Zeming Chen, Wenwei Zhang, Xinjiang Wang, Kai Chen, and Zhi Wang. Mixed Pseudo Labels for Semi-Supervised Object Detection. https://arxiv.org/abs/2312.07006, 2023. 4, 6 +[8] Achal Dave, Piotr Dólar, Deva Ramanan, Alexander Kirillov, and Ross Girshick. Evaluating Large-Vocabulary Object Detectors: The Devil is in the Details. https://arxiv.org/abs/2102.01066, 2021.7 +[9] 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. 2 +[10] Terrance DeVries and Graham W. Taylor. Improved Regularization of Convolutional Neural Networks with Cutout. https://arxiv.org/abs/1708.04552, 2017. 5 +[11] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Micheal Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Object Detection in Aerial Images: A Large-Scale Benchmark and Challenges. IEEE TPAMI, 2021. 2 +[12] Na Dong, Yongqiang Zhang, Mingli Ding, and Gim Hee Lee. Boosting Long-Tailed Object Detection via Step-Wise Learning on Smooth-Tail Data. In ICCV, 2023. 1, 2, 4, 6, 8 +[13] Mark Everingham, Luc Van Gool, Christopher K.I. Williams, John Winn, and Andrew Zisserman. The PASCAL Visual Object Classes (VOC) Challenge. IJCV, 2010. 1 +[14] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D. Cubuk, Quoc V. Le, and Barret Zoph. Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation. In CVPR, 2021. 4 + +[15] Agrim Gupta, Piotr Dólar, and Ross Girshick. LVIS: A Dataset for Large Vocabulary Instance Segmentation. In CVPR, 2019. 1, 4 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 6 +[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 4 +[18] Grant Van Horn and Pietro Perona. The Devil is in the Tails: Fine-Grained Classification in the Wild. https://arxiv.org/abs/1709.01450, 2017. 2 +[19] Xinting Hu, Yi Jiang, Kaihua Tang, Jingyuan Chen, Chunyan Miao, and Hanwang Zhang. Learning to Segment the Tail. In CVPR, 2020. 1, 2, 3, 4 +[20] Banghuai Li. Adaptive Hierarchical Representation Learning for Long-Tailed Object Detection. In CVPR, 2022. 6 +[21] Bo Li, Yongqiang Yao, Jingru Tan, Gang Zhang, Fengwei Yu, Jianwei Lu, and Ye Luo. Equalized Focal Loss for Dense Long-Tailed Object Detection. In CVPR, 2022. 6 +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 1 +[23] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 6 +[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal Loss for Dense Object Detection. In ICCV, 2017. 4 +[25] Liang Liu, Boshen Zhang, Jiangning Zhang, Wuhao Zhang, Zhenye Gan, Guanzhong Tian, Wenbing Zhu, Yabiao Wang, and Chengjie Wang. MixTeacher: Mining Promising Labels with Mixed Scale Teacher for Semi-Supervised Object Detection. In CVPR, 2023. 4, 6 +[26] 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 ICCV, 2021. 6 +[27] Lingchen Meng, Xiyang Dai, Jianwei Yang, Dongdong Chen, Yinpeng Chen, Mengchen Liu, Yi-Ling Chen, Zuxuan Wu, Lu Yuan, and Yu-Gang Jiang. Learning from Rich Semantics and Coarse Locations for Long-Tailed Object Detection. In NeurIPS, 2023. 2, 3, 6, 8 +[28] Tai-Yu Pan, Cheng Zhang, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. On Model Calibration for Long-Tailed Object Detection and Instance Segmentation. In NeurIPS, 2021. 6, 7 +[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, et al. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In NeurIPS. Curran Associates, Inc., 2019. 6 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, A. 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 ICML, 2021. 3 + +[31] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. iCaRL: Incremental Classifier and Representation Learning. In CVPR, 2017. 1 +[32] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In NeurIPS, 2015. 4 +[33] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A Large-Scale, High-Quality Dataset for Object Detection. In ICCV, 2019. 7 +[34] Antti Tarvainen and Harri Valpola. Mean Teachers are Better Role Models: Weight-Averaged Consistency Targets Improve Semi-Supervised Deep Learning Results. In NeurIPS, 2017. 5 +[35] Phi Vu Tran. LEDetection: A Simple Framework for Semi-Supervised Few-Shot Object Detection. In AISTATS, 2024. 3, 4, 6 +[36] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw Loss for Long-Tailed Instance Segmentation. In CVPR, 2021. 6 +[37] Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, and Jiashi Feng. The Devil is in Classification: A Simple Framework for Long-Tail Object Detection and Instance Segmentation. In ECCV, 2020. 4 +[38] Xin Wang, Thomas E. Huang, Trevor Darrell, Joseph E. Gonzalez, and Fisher Yu. Frustratingly Simple Few-Shot Object Detection. In ICML, 2020. 3 +[39] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to Model the Tail. In NeurIPS, 2017. 2, 3, 4 +[40] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large Scale Incremental Learning. In CVPR, 2019. 1 +[41] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-End Semi-Supervised Object Detection with Soft Teacher. In ICCV, 2021. 4 +[42] Burhaneddin Yaman, Tanvir Mahmud, and Chun-Hao Liu. Instance-Aware Repeat Factor Sampling for Long-Tailed Object Detection. https://arxiv.org/abs/2305.08069, 2023. 4 +[43] Yuhang Zang, Chen Huang, and Chen Change Loy. FASA: Feature Augmentation and Sampling Adaptation for Long-Tailed Instance Segmentation. In ICCV, 2021. 7 +[44] Yuhang Zang, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Semi-Supervised and Long-Tailed Object Detection with CascadeMatch. IJCV, 2023. 3, 7 +[45] Cheng Zhang, Tai-Yu Pan, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. MosaicOS: A Simple and Effective Use of Object-Centric Images for Long-Tailed Object Detection. In ICCV, 2021. 2, 3, 6 +[46] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond Empirical Risk Minimization. In ICLR, 2018. 5 +[47] 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. In ICLR, 2023. 4 + +[48] Liang Zhao, Yao Teng, and Limin Wang. Logit Normalization for Long-Tail Object Detection. IJCV, 2024. 6 +[49] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random Erasing Data Augmentation. In AAAI, 2020. 5 +[50] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting Twenty- Thousand Classes Using Image-Level Supervision. In ECCV, 2022. 2, 3, 6 +[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: Deformable Transformers for End-to-End Object Detection. In ICLR, 2021. 4 +[52] George Kingsley Zipf. The Psycho-Biology of Language: An Introduction to Dynamic Philology. Routledge, 2013. 3 \ No newline at end of file diff --git a/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/images.zip b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3b23999f3d7f3323e66228b0d98e3c60128cac6d --- /dev/null +++ b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b67dec7423e18a4dc5983abcdbb9b60ae835ffc3c1ecc36c2c2776a1d2b00eb4 +size 739840 diff --git a/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/layout.json b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d8858de732468389069602916eaeb1eef49da778 --- /dev/null +++ b/CVPR/2025/SimLTD_ Simple Supervised and Semi-Supervised Long-Tailed Object Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca76488c9451edef046457f0241e76b59bec5c9cfe8742f656270c450401fb8b +size 379538 diff --git a/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/4ff21147-366e-4a55-a1d6-2eb9d6bba6ae_content_list.json b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/4ff21147-366e-4a55-a1d6-2eb9d6bba6ae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c265518d635f40bd56e1da72e2b39665f9be2b1a --- /dev/null +++ b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/4ff21147-366e-4a55-a1d6-2eb9d6bba6ae_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a32189b4084417f8e0bd6bcb43df036507f2277f8c7929dcc28eb0fc93a1079b +size 84515 diff --git a/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/4ff21147-366e-4a55-a1d6-2eb9d6bba6ae_model.json b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/4ff21147-366e-4a55-a1d6-2eb9d6bba6ae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2a2d2c4902c618c4291b9491f1d80a02454d1f50 --- /dev/null +++ b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/4ff21147-366e-4a55-a1d6-2eb9d6bba6ae_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:190ddf7a516fb447046ab9695672894ac8ad5a4b1a9357ea9489339452ac6e85 +size 107888 diff --git a/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/4ff21147-366e-4a55-a1d6-2eb9d6bba6ae_origin.pdf b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/4ff21147-366e-4a55-a1d6-2eb9d6bba6ae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..67699ae3037475812e434e377c547d18190d051b --- /dev/null +++ b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/4ff21147-366e-4a55-a1d6-2eb9d6bba6ae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c4536efb76cb4ade72e69d101ef8f76437be6dedaf4fdfc7467cdf6da9f97d8 +size 7556033 diff --git a/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/full.md b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..39be0701e9e7bf3be606658abd5dcd26c566db04 --- /dev/null +++ b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/full.md @@ -0,0 +1,306 @@ +# SimLingo: Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment + +Katrin Renz $^{1,2,3*}$ Long Chen $^{1}$ Elahe Arani $^{1}$ Oleg Sinavski $^{1}$ + +$^{1}$ Wayve $^{2}$ University of Tübingen $^{3}$ Tübingen AI Center + +![](images/ebf45973c4d8295c14d6215c794fb09e1e5dadcbb5e2d17c5dcd21d60d4d0550.jpg) +Figure 1. Overview: SimLingo is a vision-language-action model unifying the tasks of autonomous driving, vision-language understanding and language-action alignment. It is state of the art on the official CARLA Leaderboard 2.0 and Bench2Drive using only camera images. We introduce the task of Action Dreaming, a form of instruction following, to improve the alignment of language and action. + +# Abstract + +Integrating large language models (LLMs) into autonomous driving has attracted significant attention with the hope of improving generalization and explainability. However, existing methods often focus on either driving or vision-language understanding but achieving both high driving performance and extensive language understanding remains challenging. In addition, the dominant approach to tackle vision-language understanding is using visual question answering. However, for autonomous driving, this is only useful if it is aligned with the action space. Otherwise, the model's answers could be inconsistent with its behavior. Therefore, we propose a model that can handle three different tasks: (1) closed-loop driving, (2) vision-language understanding, and (3) language-action alignment. Our model SimLingo is based on a vision language model (VLM) and works using only camera, excluding expensive sensors like LiDAR. SimLingo obtains state-of-the-art performance on the widely used CARLA simulator on the Bench2Drive benchmark and is the winning entry at the CARLA challenge 2024. Additionally, we achieve strong results in a wide variety of language-related tasks while maintaining high driving performance. Project page: https://katrinrenz.de/simlingo + +# 1. Introduction + +In recent years vision language models (VLMs) [11, 34, 35, 55] demonstrated their capacity for having a broad knowledge about the world and being able to generalize to unseen prompts. Incorporating these capabilities into autonomous driving systems has the potential to enhance robustness in diverse, rare scenarios and enable smoother human-machine interaction. The field of vision-driven robotics has begun to leverage pre-trained vision language models (VLMs) to enhance performance as models are better in understanding and adapting to unfamiliar environments and in following natural language commands [6, 31, 65]. + +In end-to-end autonomous driving the introduction of VLMs into the driving models also gained traction [3, 24, 51]. VLMs are used to improve planning performance [49, 63] or enable vision-language capabilities in the form of visual question answering (VQA) tailored around driving scenes [42, 43]. VQA is a promising way to test a model's ability to comprehend the scene and understand the significance of objects for their behavior. However, when evaluating this understanding only in language space, it can be completely disentangled from the actual driving decision (e.g. the model claims to see a red traffic light but the actions indicate to accelerate). We argue that only aligned language interactions - where the model's actions adapt in response to language cues - can provide causal evidence of true language understanding. To address this challenge of aligning language understanding with driving actions, we propose Action Dreaming - a new dataset to improve align- + +ment and a benchmark for evaluating how language inputs influence actions without executing dangerous instructions. When generating instruction-action pairs post-hoc on existing expert data, the action can be inferred from visual cues alone. Therefore, the model does not need to pay attention to the language instruction. This results in a misalignment between language and action. To address this, we propose a novel data-collection technique where we simulate multiple possible futures for a given state. We train and test on a set of diverse instruction-action pairs for the same visual context, ensuring that the model listens to the instruction. + +The second main focus is the driving performance. Many works that combine VLMs with driving evaluate their performance in either overly simplistic environments (e.g., HighwayEnv) [13, 14, 39, 58] or in an open-loop setting (e.g., NuScenes planning) [24, 41, 51, 53]. We rigorously test our models on a challenging closed-loop benchmark. + +Contributions. (1) A VLM-based driving model that achieves state-of-the-art driving performance on the official CARLA Leaderboard 2.0 and the local benchmark Bench2Drive in the CARLA simulator. (2) A new task (Action Dreaming), which comes with a methodology to collect instruction-action pairs and a benchmark to evaluate the connection of language and action understanding without having to execute unsafe actions. (3) A generalist model that achieves not only good driving performance but also includes several language related tasks in the same model. + +# 2. Related Work + +End-to-end autonomous driving. End-to-end training based on Imitation Learning (IL) [7, 25, 47, 59] is the dominant approach for state-of-the-art methods on the CARLA Leaderboard (LB) 1.0 [1]. With the introduction of the CARLA Leaderboard 2.0 [2] the driving task became fundamentally harder. This is shown by applying one of the leading methods of LB 1.0 (TransFuser) zero-shot to the new LB 2.0, which obtains a huge drop in driving score (66.32 to 0.58). Most state-of-the-art end-to-end methods incorporate auxiliary outputs and rely on multiple sensors like camera combined with LiDAR [7, 12, 25, 47, 48]. However, there is a growing interest in deploying camera-first models, as shown by industry players [3, 52] and academic work like DriveCoT [56] and TCP [59] show competitive results with using only camera images. We follow this approach and exclude expensive sensors like LIDAR. + +Language models for driving. Most state-of-the-art driving models introduce domain-specific architectures [12, 23, 25]. However, with the recent progress with large language models and vision language models, those generalist architectures have been integrated into driving systems. Multi-modal LLM-based driving frameworks such as LLM-Driver [9], DriveGPT4 [60], and DriveLM [51] utilize foundation models with inputs from different modalities for driving. + +GPT-Driver [40] and LanguageMPC [46] fine-tune ChatGPT as a motion planner using text. Knowledge-driven approaches [14, 58] are also adopted to make decisions based on common-sense knowledge. However, most of these works are evaluated primarily through qualitative analysis in open-loop settings like NuScenes or in simplified driving environments like the HighwayEnv [13, 14, 39, 58]. FED [63] uses a VLM for closed-loop driving by using language to provide feedback during training. LMDrive [49] proposes an end-to-end closed-loop driving model that can follow a limited set of human instructions. However, when evaluated on commonly used driving benchmarks, it is far behind state-of-the-art driving models [32]. DriveMLM [57] also includes instruction following but unlike ours it uses an off-the-shelf motion planner and is not trained end-to-end. + +Several studies have integrated language understanding into autonomous driving systems through Visual Question Answering (VQA). However, many approaches either limit their evaluation to open-loop settings [24, 37, 51] or do not involve action prediction at all [38, 42, 43]. So far, open-loop models are far behind when tested on closed-loop benchmarks [22, 30, 41]. There is no evidence that open-loop results transfer to closed-loop, making ablations and claims unreliable [26]. This highlights that bridging this gap requires major changes in training, architecture, control, and data composition. Additionally, there is a fundamental lack of systematic evaluation of instruction-following abilities, particularly when looking at the alignment of instruction and action. + +# 3. Method + +# 3.1. Task overview + +Driving. The driving objective is to reach a specified target location on a given map while passing predetermined intermediate locations in the CARLA [17] autonomous driving simulator. The map includes diverse environments such as highways, urban streets, residential areas, and rural settings, all of which must be navigated under various weather conditions, including clear daylight, sunset, rain, fog, and nighttime scenarios. Along the way, the agent must manage dense occurrences of various complex scenarios such as encountering jaywalkers, navigating parking exits, executing unprotected turns, merging into ongoing traffic, passing construction sites, or avoiding vehicles with an opened door. Vision-Language Understanding. In addition, the model needs to solve a diverse set of language tasks, including describing the current decision and action in natural language (Commentary) and answering questions about the simulated driving scene (VQA). The task of Commentary answers the question "What should the ego vehicle do next and why?" The answer refers to the path (e.g. chang- + +![](images/9380b2f09b4ec3ec0f0a9fb4d2b4d615ec6029e92ed1fe28c005c84a76a8659c.jpg) +Figure 2. SimLingo architecture. We encode the image, navigational conditioning and the language prompt. To encode high-resolution images, we split them into tiles, and encode each independently to reuse the pre-trained image encoder pre-trained on 448x448 resolution. All embeddings get processed by an LLM which we finetune with LoRA to predict language and actions. The action output utilizes a disentangled representation with both temporal speed waypoints and geometric path waypoints for improved lateral control. + +ing lanes, going around objects) and the speed together with the reason for both. Additionally, we run Commentary by default in our final model during inference. This means the final driving model acts in a Chain-of-Thought setting, where first the action and reason are predicted in language space, and then the action is predicted, conditioned on the generated commentary. Action Dreaming. The third task combines vision-language understanding with the ability to align it with the action space. We task the model with a wide variety of instructions including change of speed, lane changes, driving towards specific objects, or other navigational changes. This includes common instructions that can occur during normal driving but also out-of-distribution instructions (e.g., crashing into objects) that should never be executed in the real world but are indicative of the model's understanding of diverse instructions. This tests whether the model's language capabilities are aligned with the action space. Since we do not want to execute those actions we call this the Action Dreaming mode. We provide the model with a language instruction and the model needs to predict the corresponding action. + +# 3.2. Datasets + +Driving. We utilize the privileged rule-based expert PDMlite [5, 51] to collect our driving dataset in the CARLA simulator. We use three sets of routes with scenarios for data collection: (1) training routes of TransFuser [12] of Town 1-10 which we converted to support Leaderboard 2.0. (2) We divide the official CARLA LB 2.0 routes of Town 12 and Town 13 into shorter segments centered around a single scenario to reduce trivial data (e.g., driving straight without any hazardous events) and to simplify data man + +agement as proposed by [12, 33]. (3) With the introduction of LB 2.0, the maximum distance between navigational target points increased from 50 to 200 meters. The routes with one scenario often fall within this distance, causing a distribution shift, as the next target point is at the end of the route (i.e., closer than $200\mathrm{m}$ ) rather than after 200 meters. Consequently, we employ a set of longer routes, featuring three scenarios per route. To ensure balance of scenario types, we adjust the number of routes per scenario (i.e., upsampling routes with rare scenarios), apply random weather augmentation, and modify how far scenarios get spawned by $\pm 10\%$ . Overall, we collect 3.1 million samples at 4 fps. Vision-Language Understanding. For the vision-language understanding data we apply the heuristics of DriveLM [51] to generate VQA labels for our driving dataset. We use Town 12 for training and Town 13 for evaluation. For the commentary task, we generate new labels as this task is not part of the DriveLM labels. Similarly, we use privileged information during data collection and the logic of the rule-based expert to generate explanations. More precisely, we use the leading object obtained from the experts' Intelligent Driver Model (IDM) [54] as well as information about changing the path to swerve around objects. In addition, we use heuristics based on the ego waypoints to distinguish between driving intentions like starting from stop or keep driving at the same speed. Action Dreaming. To align language with the action space we need data with paired instruction and action. Relying solely on the actual actions executed by the expert for generating instructions post-hoc, results in labels that can be entirely inferred from the available visual information, thus making it unnecessary to process the language instructions. To address this limitation, generating + +multiple alternative instructions and actions for the same visual context forces the model to listen to the instruction. For the original dataset, we store a non-exhaustive portion of the simulator's state, such as the positions of other vehicles and their speeds. We utilize the "world-on-rails" assumption [8] (i.e., all other dynamic agents are treated as being on fixed paths and are unresponsive to environmental changes, akin to replaying the original simulation) for other actor and approximate the ego vehicle's dynamics using a kinematic bicycle model. This allows us to simulate different possible future trajectories of the ego vehicle and perform collision checks with other agents without needing to execute the actions. This dataset incorporates several types of new trajectories, such as lane changes (including movements onto sidewalks or parking lanes), speed adjustments, driving toward specific objects (e.g., traffic cones, signs, vehicles) potentially causing collisions, and crossing particular road markings like stop signs or stop lines. For each alternative instruction-action pair we also have a flag if this action would be safe to execute and a language label describing if the action can be executed or not together with the reason in case not. We use Town 12 for training and Town 13 for validation. More details about the data collection and labelling procedure can be found in Appendix A. + +# 3.3. Architecture + +SimLingo is built on top of the InternVL-2 architecture [19] with additional input modalities and output heads. Fig. 2 shows an overview of the SimLingo architecture. + +Input Representation. SimLingo uses a camera image to navigate the environment. We input the camera image $\mathbf{I} \in \mathbb{R}^{H \times W}$ , navigational information in the form of either the next two GPS target points TP or a high-level language command HLC (e.g. "turn left"), and the ego vehicle's speed $v$ . Additionally, we input a task prompt with information about the current task. We differentiate between 4 task prompts $p_{\text {task }}$ : (1) Driving without language predictions: "Predict the waypoints." (2) Commentary + Driving: "What should the ego do next?" (3) VQA + Driving: "Q: $\langle$ question $\rangle$ " (4) Action Dreaming: "⟨Dreamer flag⟩ ⟨instruction⟩." For the ⟨Dreamer flag⟩ we differentiate between two modes: when activated the model should predict actions aligned with the instruction, when deactivated, the model needs to reject unsafe instructions. All input information is part of the global LLM prompt $p_{\text {global }}$ , which has the following structure: " $\langle$ image features e_I $\rangle\backslash n$ Current speed: $\langle v\rangle m/s$ . Command: ⟨nav. features eNAV⟩. ⟨task prompt $p_{\text {task }}$ ." All parts inside $\langle \rangle$ get replaced by the corresponding feature embeddings. + +Output Representation. The output consists of two modalities: action and language. For any language prediction $l$ we use standard auto-regressive token prediction. For the action representation, we use a disentangled repre + +sentation with (1) temporal speed waypoints $\mathbf{w} \in \mathbb{R}^{N_w \times 2}$ , with $N_w$ future coordinates with one coordinate every 0.25 seconds. This represents the location of the ego vehicle at a specific time in the future. Also, we predict (2) geometric path waypoints $\mathbf{p} \in \mathbb{R}^{N_p \times 2}$ , with $N_p$ future coordinates with one coordinate every meter. This represents future positions of the ego vehicle independently of the time to reach them. From the temporal waypoints $\mathbf{w}$ we obtain a target speed and from the geometric waypoints $\mathbf{p}$ a target angle. We then use two PID controllers to get the steering angle and acceleration. We noticed that using only temporal waypoints for steering and speed led to steering problems, especially during turns or when swerving around obstacles. By using path waypoints, we achieve denser supervision, as we also predict the path when the vehicle is stationary, leading to improved steering behavior. To enable more efficient action prediction we predict all action embeddings in one forward pass instead of auto-regressively. For this, we input learnable query tokens $\mathbf{q}_p$ and $\mathbf{q}_w$ . An MLP on top of the output features $[\mathbf{o}_p, \mathbf{o}_w]$ generates waypoint differences. The cumulative sum of these differences yields the final waypoints $\mathbf{p}$ and $\mathbf{w}$ . + +Vision-Language-Action Model. We use the InternVL2-1B from the Mini-InternVL family [19] as our main architecture for SimLingo. The InternVL2 family [10, 11, 19] reaches state-of-the-art performance while providing the widest range of available model sizes starting with a 1B parameter model. The InternVL2-1B model consists of the vision encoder InternViT-300M-448px (ViT) and Qwen2-0.5B-Instruct [61] as the language model (LLM). + +Variable high-resolution image encoding. Processing high-resolution images is crucial for detailed image understanding. Especially for autonomous driving important information, such as traffic lights at large intersections, may only be visible in a few pixels. Most VLMs are based on CLIP pretrained vision encoder which are pre-trained on a resolution of $336 \times 336$ (some $448 \times 448$ ). To be able to process dynamic and higher resolutions we split the input image $\mathbf{I} \in \mathbb{R}^{H \times W}$ into $N_I$ $448 \times 448$ pixel tiles $\mathbf{i}_n \in \mathbb{R}^{448 \times 448}$ and extract features for each tile independently. This method is commonly used in the latest VLMs [11, 35]. To reduce computational overhead due to the nature of the quadratic complexity of the LLM, InternVL2 uses the pixel unshuffle technique [50] $(\rho)$ to downsample the number of tokens by a factor of 4. Each $448 \times 448$ tile is then represented by 256 visual tokens. We obtain visual features + +$$ +\mathbf {e} _ {I} = \rho ([ \mathrm {V i T} (\mathbf {i} _ {n}) ] _ {n = 0} ^ {N _ {I}}) \quad \in \mathbb {R} ^ {(N _ {I} * 2 5 6) \times D} +$$ + +with embedding size $D$ . In this work we use $N_{I} = 2$ resulting in 512 visual tokens. + +Driving specific inputs. The navigational information is presented either as target points in the form of GPS locations (TP) or as a language command (HLC). When using the tar + +get points, we encode the locations with an MLP to obtain two navigational embeddings $\mathbf{e}_{nav} \in \mathbb{R}^{2 \times D}$ . In the case of using the language command we use the standard tokenizer of the LLM to obtain the embeddings $\mathbf{e}_{nav} \in \mathbb{R}^{N_{\mathrm{HLC}} \times D}$ , with $N_{\mathrm{HLC}}$ equals the number of tokens representing the navigational command. During training, we randomly switch between the two input modalities. We use the speed $v$ in natural language as part of the global LLM prompt. + +Language input. For all language parts, we use the original tokenizer of the LLM to obtain the token embeddings $\mathbf{e}_L$ . + +Token interleaver. After encoding each input modality the token interleaver (IL) replaces the placeholder tokens $\langle \rangle$ of $p_{\mathrm{global}}$ with the corresponding embeddings, to obtain the input embedding sequence: + +$$ +\mathbf {e} _ {\mathrm {L L M}} = \operatorname {I L} \left(\mathbf {e} _ {L}, \mathbf {e} _ {I}, \mathbf {e} _ {\text {n a v}}\right). +$$ + +This interleaved token sequence is then input to the pretrained LLM. + +Large language model. Given the input embeddings and the action queries the LLM generates language and action output features + +$$ +[ \mathbf {o} _ {l}, \mathbf {o} _ {p}, \mathbf {o} _ {w} ] = \mathrm {L L M} ([ \mathbf {e} _ {\mathrm {L L M}}, \mathbf {q} _ {w}, \mathbf {q} _ {p} ]) \quad . +$$ + +First, it auto-regressively generates the language predictions, which represent the answer to the task prompt. Then in one additional forward pass, it generates the actions consisting of path and waypoints. We use the smooth-L1 loss on the path and waypoints and cross-entropy loss on the predicted language tokens. + +# 3.4. Training + +Dataset mixtures. We differentiate between two action labels: (1) expert trajectories and (2) dream trajectories (safe and unsafe ones). During training, we sample 50/50 from each. For the expert trajectories, we do one of three options for additional language supervision: (1) VQA, (2) Commentary, and (3) no language. We sample $50\%$ VQA prediction, $35\%$ commentary prediction, $7.5\%$ we provide the commentary in the prompt, and $7.5\%$ we do not supervise any language. For the dream trajectories, we train $50\%$ with the Dreamer flag activated and $50\%$ with it being deactivated, where the model needs to decide if the instruction is safe to execute and reject unsafe ones. Data buckets. The majority of driving involves straight, uneventful segments. To maximize the collection of interesting scenarios, we focus the data collection around a diverse range of challenging situations. However, some ratio of easy and uneventful data is inevitable. To address this issue, we create data buckets containing specific interesting samples and assign a probability to each bucket. During training, we sample from these buckets instead of the entire dataset. This approach reduces the number of samples per epoch to 650,000. Details about the bucket types are in the Appendix B.2. + +# 4. Experiments + +In this section, we start with an overview of the used benchmarks and metrics (Section 4.1) and the implementation details (Section 4.2). We then demonstrate our results (Section 4.3) quantitatively and qualitatively. + +# 4.1. Benchmarks and Metrics + +For a detailed description of the metrics we refer to Appendix B.4. + +Leaderboard 2.0. We use the official test server of the CARLA simulator with secret routes under different weather conditions [2]. We report the official CARLA metrics, Driving Score (DS), Route Completion (RC), and Infraction Score (IS). The Driving Score is calculated in a way that models in the current performance range get punished for completing more of the route due to a non-linear decrease of the score due to infractions. A more detailed discussion about its flaws is in the Appendix B.4. + +Bench2Drive. For local evaluation we use the Bench2Drive Benchmark [29] based on CARLA version 0.9.15 which uses 220 routes with approximately $150\mathrm{m}$ on Town01 - Town15 and different weathers. We use the official metrics, which include driving score, success rate, efficiency, and comfortness. Those shorter routes make it easier to obtain higher driving scores making the range not comparable with the Leaderboard numbers. + +DriveLM-hard (VQA) and Commentary. We use DriveLM [51] to generate VQA data and our annotation scheme described in Section 3.2 to collect additional Commentary labels for Town 13 as validation labels which we withhold during training. Differently to DriveLM, we construct a harder validation set for which we uniformly sample 10 examples per answer type instead of randomly sampling from the whole set and therefore highly undersample rare answers. Our VQA validation set contains 330 different answer types and the Commentary one contains 190. We use the GPT and SPICE metric. + +Action Dreaming. To assess not only the ability to understand scene-specific knowledge in language space but also to connect this to the action space we introduce the Action Dreaming evaluation. The model gets a language input in the form of an instruction and needs to imagine how the corresponding actions would look like. We use Town 13 of our Dreamer dataset (described in Section 3.2) for this validation. Instructions are from one of the following classes: Slow down, Speed up, Reach Target Speed, Lane Change, Object centric. We evaluate open-loop and use Success Rate as the metric. For each class, we use an individual rule to define success. + +# 4.2. Implementation Details + +SimLingo. Optimization is done with the AdamW optimizer [36] with a weight decay of 0.1 and a learning rate + +
MethodSensorsAux. LabelsDS ↑RC ↑IS ↑
MapCaRINA mod.[45]L,C,MO1.143.650.46
Kyber-E2E[18]L,C,R,MI, O5.4715.890.34
TF++[64]L,CS, D, O, B5.5611.820.47
SimLingo-BASE*C-6.2518.890.36
SensorZero-shot TF++[25]L,CS, D, O, B0.588.530.38
CaRINA hybrid[45]L,CI, O1.239.560.31
TF++[64]L,CS, D, O, B5.1811.340.48
SimLingo-BASE*C-6.8718.080.42
+ +Table 1. Leaderboard 2.0 Results. SimLingo-BASE achieves state-of-the-art performance on the official Leaderboard 2.0. Legend: L: Lidar, C: Camera, R: Radar, M: Map, priv: privileged, O: Object Detection (3D position and pose), I: Instant Segmentation, S: Semantic Segmentation, D: Depth, B: BEV semantics. *To maintain consistency we changed the naming of our model from CarLLaVA to SimLingo-BASE. + +of 3e-5 with a cosine annealing schedule without restarts. We use DeepSpeed v2 for optimizing training efficiency and memory usage. We fully finetune all components besides the LLM for which we use LoRA [21] applied to all linear layers as demonstrated to be effective [16]. We apply the same data augmentation techniques as TF++ [25] but with more aggressive shift and rotation augmentation (shift: $1.5\mathrm{m}$ , rot: $20\mathrm{deg}$ ). The model is trained for 14 epochs on 8xA100 80GB GPUs with as batch size of 12 which takes 24 hours. Due to instabilities during training with the L2-Loss after adding additional data we changed to the SmoothL1-Loss. During inference, we use Commentary by default. This means the final driving model acts in a Chain-of-Thought setting, where first the action and reason are predicted in language space, and then the action is predicted, conditioned on the generated commentary. + +Baseline - SimLingo-BASE. We provide a lightweight baseline driving model that is more suitable for testing driving-specific design choices and evaluating on the official Leaderboard 2.0 in a compute-constrained environment. A detailed list of the differences to SimLingo can be found in the Appendix B.3, but the main difference is that it has no language support and uses a smaller transformer (50M parameter) based on the LLaMA architecture which is trained from scratch. Since the Leaderboard temporarily closed in June 2024 we could not submit the final SimLingo to the Leaderboard (SimLingo-BASE was submitted before the closure) and choose to compare the closed-loop performance on a local benchmark (Bench2Drive). For the Leaderboard model, we apply early stopping during inference to counter the nature of DS described in the metric section. We track the traveled distance and stop driving after a specified distance when the steering angle is close to zero to prevent stopping in the middle of an intersection. + +Baseline - TCP-traj w/o distillation. Existing methods on Bench2Drive [29] are trained with a dataset collected by the Think2Drive [33] expert. Since we use the open-source expert PDM-lite [51] we retrain TCP-traj w/o distillation with + +# Vision-Language Understanding + +A +![](images/1adb77d0d7c997dc7f47ec794dd7ea0eebefd4a6bea3337cf51f462b063cec10.jpg) +"Q: Does the ego vehicle need to change lanes +or deviate from the lane center due to an upcoming instruction?" +"GT A: The ego vehicle must change to the left lane to circumvent the bicycle." +"Pred A: The ego vehicle must transition to the left lane to avoid the bicycle." +"Q: What should the ego vehicle do based on the traffic light?" +"GT A: Based on the green traffic light the ego vehicle can accelerate and continue driving but should pay attention to the vehicle in front and adjust its speed accordingly." + +aannnnnne aannnnnne +![](images/dd2bc853760f3cf04a4c6be1ce214c425fe2a6b68deca7d1aec7ef28323ed7cc.jpg) +"Pred A: The ego vehicle can keep its pace and continue driving, given the green traffic light." +"Follow the route. Stay in the same spot due to the red traffic light." + +Figure 3. Qualitative results for VQA and Commentary. For VQA we show the question, the ground truth answer and the predicted answer for two examples scenes. Both questions refer to objects far away only apparent in a couple of pixels, but the model still produces correct answers. +![](images/b07651886741788c51d1062e88e32768b0e736f044cba9f3f8638bdf14621c23.jpg) +"Follow the route. Slow the vehicle to stay at the rear of the yellow car that is to the front." + +![](images/f7fc2f1ddb3ff065960327629fbac468303e3e1f92b8953d43d8075c00c0c690.jpg) +"Follow the route. Stay stationary due to the pedestrian crossing your line of travel." + +# Language/Vision -> Action (Dreaming) + +![](images/c33a7aa1bfaf4c6bddb068da26c5fcffa6e5a17a92e0467706f0c2cbda097503.jpg) +"Go left at the next intersection." + +![](images/69d7e68f12998b9e656008bb4dd3c444ad062c9634b2682e1c171c69065a2005.jpg) +"Go straight." + +![](images/b6d77f82b4b723728522e86520df28b6ce17a763cba904dac5b9b3c4b54401e9.jpg) +"Go right at the next intersection." + +![](images/8880c3b13a7137fca2f5aa551e49507a38dd23b9c674d850e020da28b1254f99.jpg) +"Turn left at the next intersection" + +![](images/3bfc8f6757204ea566a0bdfe3046c410a4370326d1562dc3fe7add79887ab32c.jpg) +"Crash into the object at x: 12.49m, y:6.202m." + +![](images/6e46cddbcbccbb25f9735391c25b669a560252a5bb227a79e0bb3dae39198c61.jpg) +"In 5 meters, transition one lane to the left with a 7-meter transition." + +Figure 4. Qualitative results of Action Dreaming. We show the predicted actions for a diverse set of situations and instructions. The model can successfully adapt to path and speed related instructions. Legend: red: path waypoints, green: speed waypoints, blue graph: speed in m/s. +![](images/80c48edb8d65ba14815a23e8181772eb48b5673fdca68e13d27671dd2062bdee.jpg) +"Go ahead and drive faster." + +![](images/b5eb7687f6969478340f9bb9180fc44a02f53d775f4a7ae2e30eb7a6a1f5b5da.jpg) +"Gently pre + +![](images/3b1d65101ceb2d5c9d801b5e643cd2d8bff40380328590764433f35a68f9a04d.jpg) +"Gently press the brakes." + +![](images/2928c9025472224f596841a3c715735f4ae9608a6880cca8f776f10cab875034.jpg) + +our dataset composition collected by PDM-lite to increase fairness in the evaluation. As the driving behavior like maximum speed and acceleration is different between the experts, we also adjust the controller to take those changes into account when evaluating closed-loop. We choose TCP-traj w/o distillation as it is the best previous method that does not rely on expert feature distillation, as expert features are not available for PDM-lite or in the real world. + +
MethodExpertDS ↑Success Rate(%) ↑Efficiency ↑Comfortness ↑
w/dist.TCP [59]Think2Drive40.7015.0054.2647.80
TCP-ctrlThink2Drive30.477.2755.9751.51
TCP-trajThink2Drive59.9030.0076.5418.08
ThinkTwice [28]Think2Drive62.4431.2369.3316.22
DriveAdapter [27]Think2Drive64.2233.0870.2216.01
w/o dist.AD-MLP [62]Think2Drive18.050.0048.4522.63
UniAD-Tiny [23]Think2Drive40.7313.18123.9247.04
UniAD-Base [23]Think2Drive45.8116.36129.2143.58
VAD [30]Think2Drive42.3515.00157.9446.01
TCP-traj w/o distillationThink2Drive49.3020.4578.7822.96
our dataset compositionPDM-lite45.6518.5774.8451.58
+ with tuned controllerPDM-lite63.4537.79228.4630.76
SimLingo-BASE (LB2.0 model)PDM-lite85.9466.82244.1825.49
SimLingoPDM-lite85.07±0.9567.27±2.11259.23±5.5933.67±5.72
+ +# 4.3. Results + +Driving performance on the CARLA Leaderboard 2.0. We present the official Leaderboard results in Tab. 1. With our lightweight model SimLingo-BASE we outperform the previous state of the art (CaRINA hybrid [45]) by $4.6\mathrm{x}$ and the best concurrent work $(\mathrm{TF} + +$ [64]) by $33\%$ on the Sensor track. Ablation study: To show the impact of our output representation, we compare the disentangled path+speed waypoints with the commonly used entangled waypoint representation. We observe a 39.9 percent increase in driving score and a reduction of collisions with static objects to zero. In addition, we analyze the vision encoder and compare it with a ViT without Clip pretraining and with a Resnet-34 [20] pretrained on ImageNet [15] (using the same training budget). The ViT trained from scratch only obtains 0.45 DS, showing the capability of the Clip pretraining [44]. The Resnet results in a driving score of 2.71 compared to our 6.87. It is also noteworthy that, our model is the only model on the leaderboard working only with camera images (note: we did not include entries of the Leaderboard without a method report). We show more details about the ablations on the output representation, early stopping, and leaderboard variance in the Appendix C. + +Driving performance on Bench2Drive. Tab. 2 shows results on the local benchmark Bench2Drive [29] (numbers taken from the newest version of the Bench2Drive Github repository). Existing works on this benchmark train on data collected with the Think2Drive expert [33]. However, this expert is not open-source and therefore we cannot collect the data we need to obtain the language annotations. For this reason, we use PDM-lite, an open-source expert for which we can modify the saved labels for our label creation. To provide a fair comparison and disentangle the performance improvements stemming from the different datasets, we provide results for TCP-traj w/o distillation with our + +Table 2. Closed-loop Results on Bench2Drive. Both our models SimLingo-BASE and the full model SimLingo, outperforms the previous state of the art by a large margin. We highlight that SimLingo can preserve the same driving performance as the pure driving model SimLingo-BASE while including several language related capabilities. + +
ModelDriveLM-VQACommentary
GPT ↑SPICE ↑GPT ↑SPICE ↑
InternVL2-1B33.0830.5514.957.60
InternVL2-2B31.2244.4020.539.04
InternVL2-4B27.1143.5124.758.12
SimLingo-1B58.4856.7778.9438.04
+ +Table 3. Language ability. We show results on DriveLM and our Commentary benchmark on a balanced evaluation split. SimLingo outperforms the InternVL2 models which we evaluate zero-shot. + +dataset composition and a controller tuned for our dataset. With those two changes, the performance increases from 49.30 DS to 63.45 DS. This shows that the dataset contributes to the performance improvement but our additional advancements are essential to obtain state-of-the-art performance. Looking at the driving metrics (DS and SR) SimLingo and SimLingo-BASE, which is the identical model that we used for the Leaderboard, outperforms all previous methods by a large margin. For all SimLingo models, we train three seeds to average out the training variance. Only in the comfortness metric previous works partly perform better, which is likely due to the higher driving speed of our method. Our Vision-Language-Action model (SimLingo) can preserve the driving performance of the pure driving model (SimLingo-BASE) while including a wide range of language abilities. + +Language understanding - VQA & Commentary. We show results on the DriveLM-hard and Commentary Benchmark in Tab. 3. We use the Mini-InternVL2 models to provide a baseline comparison. However, we emphasize that SimLingo was trained on the simulated visual indistribution data, whereas InternVL2 is tested zero-shot. The baseline models get basic questions right like identifying a traffic light or a stop sign but struggle with more complex, driving-specific questions. In addition, we compare to a version that is fine-tuned only on DriveLM. This + +
DS ↑SR(%) ↑
GPS target point (TP)85.07±0.9567.27±2.11
Language command (HLC)86.08±1.7665.78±3.90
+ +Table 4. Following navigational commands. SimLingo obtains strong driving results independent of the navigational conditioning (GPS target points vs. language command) showing it's ability to follow instructions. + +
Metric CategorySR(%)↑
FasterSlowerTarget SpeedLane ChangeObjectsAvg.
w/o Dreamdata56.4522.5819.353.2320.9724.52
SimLingo92.4584.9186.7983.0258.4981.13
+ +specialized model obtains a GPT-score of 63.85, which is only slightly better than our model that is able to perform several different tasks. We provide qualitative examples in Fig. 3 and the Appendix D. We found that the model especially struggles with answer types that are rarely seen during training indicating that either collecting more driving samples with those answer types or oversampling during training might help to further improve the performance. + +Following navigational commands. To show the ability of the model to follow basic navigational commands we compare SimLingo with navigational conditioning of GPS target points with the same model using language commands (HLC) (i.e., language instructions like turn right). Tab. 4 shows that by only using HLC we reach a driving performance, which is within the variance of the results with using TPs. This enables more human-like conditioning but also indicates that the shortcut of recovery through target point locations [25] is not needed in our setting. + +Action Dreaming. For the general ability to align the language understanding with the action space, we evaluate on the Action Dreaming dataset. Tab. 5 shows the Success Rate of the Action Dreaming evaluation for each category. We compare the model trained on the Dreaming data with a baseline that is only trained on instructions provided by the CARLA simulator. With training on the synthetically generated instruction-action pairs, we see a huge improvement (28.22 to 72.96 SR). Fig. 4 shows qualitative examples of the Action Dreaming mode. In the first row (a-c), we show navigational changes for the same scene. This shows that the model does not overfit to visual cues but instead pays attention to the language. In a) and d) we use the same instruction of turning left but in a different context indicating that the model can successfully adapt its response to the situation. In scene a) we are right before a junction so the model correctly predicts a left turn indicated by the red path waypoints. In contrast, in scene d) we are on a multi-lane road far from a junction so turning left means to prepare for the turn and move into the left lane to perform the turn when a junction shows up. More examples, including fail + +Table 5. Action Dreaming. When activating the dreaming flag, SimLingo is able to follow a wide range of instructions compared to the model trained with only the natively supported commands from the CARLA simulator. + +
CommVQADreamBench2Drive
DS ↑SR(%) ↑
SimLingo✓✓✓85.07±0.9567.27±2.11
✓✓-84.41±1.3864.85±2.24
✓--84.55±0.4265.00±0.64
---84.67±1.5264.70±3.22
+ +Table 6. Closed-loop results on Bench2Drive for different training mixtures. Pure vision-language tasks (Commentary and VQA) do not influence pure driving performance (last row). But including Action Dreaming training slightly improves driving. + +ure cases, can be found in the Appendix D. + +Training mixture. We ablate every added language task from the full SimLingo model to a base driving model without any language capabilities. Tab 6 shows that having a pure driving model (4th row) gets a similar performance as the model with (non-aligned) language understanding (2nd + 3rd rows). Hence, in our experiments, we have not seen any positive or negative impact from not-aligned language tasks on driving. However, adding the Action Dreaming data to the training mixture slightly improves driving performance on the Bench2Drive benchmark. + +# 5. Conclusion and Limitations + +We present SimLingo: a VLM-based autonomous driving model that achieves state-of-the-art results on the official CARLA Leaderboard 2.0 and Bench2Drive while demonstrating language understanding. We measure language performance on driving Commentary and VQA showing that after finetuning, a base generalist VLM (InternVL2) can excel in a specialized driving domain. To align this vision-language understanding with the action space we introduce the task of Action Dreaming, where we demonstrate a high success rate of the model to predict actions for a high variety of language instructions. + +Limitations. We recognize several limitations of our work. For technical and organizational reasons (Leaderboard closed in June 2024), we were able to test only SimLingo-BASE on the official CARLA Leaderboard 2.0 and not the full SimLingo model, which we leave for future work. However, we rigorously checked the closed loop driving performance of SimLingo and SimLingo-BASE on the local Bench2Drive benchmark. Also, while we use Chain-of-Thought (CoT) (i.e., conditioning the driving action on the intermediate commentary), we have not yet observed statistically significant driving improvements (Appendix ??). We hypothesize that appropriate CoT-specific data and training recipes are needed to harness its benefits. Finally, we perform driving and language understanding only in simulation. We appreciate that adding VLMs in real-world driving models increases the inference latency, but we believe that advances in smaller VLMs and engineering efforts (such as [3, 4]) would allow testing VLMs real-time on the car. + +Acknowledgements. We thank the whole Wayve Lingo team, Ana-Maria Marcu, Jan Hünermann, Benoit Hanotte, Alice Karnsund and Jamie Shotton for helpful discussions and proofreading. We also thank Kashyap Chitta, Julian Zimmerlin, Jens Beiwbenger, Bernhard Jäger and Andreas Geiger for valuable discussions and help with the expert. We thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting K. Renz. + +# References + +[1] Carla autonomous driving leaderboard. https://leaderboard.carla.org/#leaderboard-10, 2020.2 +[2] Carla autonomous driving leaderboard 2.0. https://leaderboard.carla.org/, 2023. 2, 5 +[3] Lingo-2: Driving with natural language. https://wayve.ai/ thinking/lingo-2-driving-with-language/, 2024. 1, 2, 8 +[4] Lambda: The nuro driver's real time. https://medium.com/nuro/lambda-the-nuro-drivers-real-time-language-reasoning-model-7c3567b2d7b4, 2024.8 +[5] Jens Beißwenger. Pdm-lite: A rule-based planner for carla leaderboard 2.0. Technical report, University of Tübingen, 2024. 3 +[6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky. $\pi_0$ : A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 1 +[7] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In CVPR, 2022. 2 +[8] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In ICCV, 2021. 4 +[9] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with LLMs: fusing object-level vector modality for explainable autonomous driving. In ICRA, 2023. 2 +[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 2024. 4 +[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. 1, 4 +[12] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE T-PAMI, 2023. 2, 3 + +[13] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Receive, reason, and react: Drive as you say with large language models in autonomous vehicles. ITS Magazine, 2024. 2 +[14] Fu Daocheng, Li Xin, Wen Licheng, Dou Min, Cai Pinlong, Shi Botian, and Qiao Yu. Drive like a human: Rethinking autonomous driving with large language models. In WACV Workshops, 2024. 2 +[15] 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. 7 +[16] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. 2023. 6 +[17] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator, 2017. 2 +[18] Mohammed Elmahgiubi, Fazel Arasteh, Wang Zhitao, Yang Zhou, Yuanxin Zhong, Hamidreza Mirkhani, Weize Zhang, Zhan Qu, Zizhao Huang, and Kasra Rezaee. Kyber-e2e submission to cvpr's carla autonomous driving challenge. 6 +[19] Zhangwei Gao, Zhe Chen, Erfei Cui, Yiming Ren, Weiyun Wang, Jinguo Zhu, Hao Tian, Shenglong Ye, Junjun He, Xizhou Zhu, et al. Mini-internvl: a flexible-transfer pocket multi-modal model with $5\%$ parameters and $90\%$ performance. Visual Intelligence, 2024. 4 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7 +[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. In CoRL, 2021. 6 +[22] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 2 +[23] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, et al. Planning-oriented autonomous driving. In CVPR, 2023. 2, 7 +[24] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, Yin Zhou, James Guo, Dragomir Anguelov, and Mingxing Tan. Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262, 2024.1, 2 +[25] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. In ICCV, 2023. 2, 6, 8 +[26] Bernhard Jaeger, Kashyap Chitta, Daniel Dauner, Kartrin Renz, and Andreas Geiger. Common Mistakes in Benchmarking Autonomous Driving. https://github.com/ autonomousvision/carla_garage/blob/leaderboard_2/docs/ common_mistakes_in_benchmarking_ad.md, 2024. 2 +[27] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. DriveAdapter: breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In ICCV, 2023. 7 + +[28] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In CVPR, 2023. 7 +[29] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. In NeurIPS Datasets and Benchmarks, 2024. 5, 6, 7 +[30] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 2, 7 +[31] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An opensource vision-language-action model. In CoRL, 2024. 1 +[32] Changwen Li, Joseph Sifakis, Rongjie Yan, and Jian Zhang. A comprehensive evaluation of four end-to-end ai autopilots using cctest and the carla leaderboard. arXiv preprint arXiv:2501.12090, 2025. 2 +[33] Qifeng Li, Xiaosong Jia, Shaobo Wang, and Junchi Yan. Think2drive: Efficient reinforcement learning by thinking in latent world model for quasi-realistic autonomous driving (in carla-v2). In ECCV, 2024. 3, 6, 7 +[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1 +[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,4 +[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5 +[37] Yuhang Lu, Yichen Yao, Jiadong Tu, Jiangnan Shao, Yuexin Ma, and Xinge Zhu. Can LVLMs obtain a drivers license? a benchmark towards reliable AGI for autonomous driving. arXiv, 2409.02914, 2024. 2 +[38] Yunsheng Ma, Amr Abdelraouf, Rohit Gupta, Ziran Wang, and Kyungtae Han. Video token sparsification for efficient multimodal LLMs in autonomous driving. arXiv, 2409.11182, 2024. 2 +[39] Yunsheng Ma, Can Cui, Xu Cao, Wenqian Ye, Peiran Liu, Juanwu Lu, Amr Abdelraouf, Rohit Gupta, Kyungtae Han, Aniket Bera, James M. Rehg, and Ziran Wang. LaMPilot: an open benchmark dataset for autonomous driving with language model programs. In CVPR, 2024. 2 +[40] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. GPTDriver: Learning to drive with gpt. In NIPS Workshops, 2023. 2 +[41] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. In $COLM$ , 2024. 2 +[42] Ana-Maria Marcu, Long Chen, Jan Hunermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, Elahe Arani, and Oleg Sinavski. LingoQA: video question answering for autonomous driving. In ECCV, 2024. 1, 2 + +[43] Tianwen Qian, Jingjing Chen, Linhai Zhuo, Yang Jiao, and Yu-Gang Jiang. NuScenes-QA: A multi-modal visual question answering benchmark for autonomous driving scenario. In AAAI, 2024. 1, 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. arXiv.org, 2103.00020, 2021. 7 +[45] Luis Alberto Rosero, Iago Pacheco Gomes, Ana Huaman Reyna, Victor Hugo Sillerico Justo, Vinicius Gustierrez Neves, Denis Fernando Wolf, and Fernando Santos Os orio. A hybrid autonomous driving navigation architecture for the 2024 carla autonomous driving challenge. 6, 7 +[46] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. LanguageMPC: Large language models as decision makers for autonomous driving. arXiv preprint arXiv:2310.03026, 2023. 2 +[47] Hao Shao, Letian Wang, RuoBing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In CoRL, 2022. 2 +[48] Hao Shao, Letian Wang, Ruobing Chen, Steven L. Waslander, Hongsheng Li, and Yu Liu. Reasonnet: End-to-end driving with temporal and global reasoning. In CVPR, 2023. 2 +[49] Hao Shao, Yuxuan Hu, Letian Wang, Steven L. Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In CVPR, 2024. 1, 2 +[50] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 4 +[51] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Zhang Hanxue, Chengen Xie, Jens BeiBwenger, Ping Luo, Andreas Geiger, and Hongyang Li. DriveLM: driving with graph visual question answering. In ECCV, 2024. 1, 2, 3, 5, 6 +[52] Tesla. Tesla autopilot. https://www.tesla.com/ autopilot, 2014. Online: accessed 18-October-2019. 2 +[53] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. DriveVLM: the convergence of autonomous driving and large vision-language models. In CoRL, 2024. 2 +[54] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical Review E, 62(2):1805-1824, 2000. 3 +[55] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1 + +[56] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. DriveCoT: integrating chain-of-thought reasoning with end-to-end driving. arXiv, 2403.16996, 2024. 2 +[57] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. DriveMLM: aligning multimodal large language models with behavioral planning states for autonomous driving. arXiv, 2312.09245, 2023. 2 +[58] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao MA, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. DiLu: a knowledge-driven approach to autonomous driving with large language models. In ICLR, 2024. 2 +[59] Peng Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In NeurIPS, 2022. 2, 7 +[60] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. DriveGPT4: Interpretable end-to-end autonomous driving via large language model. In RA-L, 2023. 2 +[61] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 4 +[62] Jiang-Tian Zhai, Ze Feng, Jihao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 7 +[63] Jimuyang Zhang, Zanming Huang, Arjit Ray, and Eshed Ohn-Bar. Feedback-guided autonomous driving. In CVPR, 2024. 1, 2 +[64] Julian Zimmerlin, Jens Beißwenger, Bernhard Jaeger, Andreas Geiger, and Kashyap Chitta. Hidden biases of end-to-end driving datasets. In CVPR Workshops, 2024. 6, 7 +[65] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 1 \ No newline at end of file diff --git a/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/images.zip b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..37213ed62cbe43168ebdc4a98890e4576c486173 --- /dev/null +++ b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfcbdeda513900941d7617f27819f5a797467c190b9685e82e0d9cc55057ecca +size 355448 diff --git a/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/layout.json b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9c81042630aa72658bf23613f860e30aad1af998 --- /dev/null +++ b/CVPR/2025/SimLingo_ Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3555a12015db5d93d1f13dfcee2d33abd799d47dacd1540a0002830c8b70e25 +size 393326 diff --git a/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/35d364b8-3b24-4d25-b0d5-f4f2de5e4e95_content_list.json b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/35d364b8-3b24-4d25-b0d5-f4f2de5e4e95_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..029a0243e936dd419c6d6ece8eb236a286e17316 --- /dev/null +++ b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/35d364b8-3b24-4d25-b0d5-f4f2de5e4e95_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5a85fde7ff8f3b5a8fef368adf61db1d8ff07d151121b69db0744af9dbb11fa +size 73583 diff --git a/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/35d364b8-3b24-4d25-b0d5-f4f2de5e4e95_model.json b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/35d364b8-3b24-4d25-b0d5-f4f2de5e4e95_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c72d530d642146d61ed533f60f8204cc01aceed --- /dev/null +++ b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/35d364b8-3b24-4d25-b0d5-f4f2de5e4e95_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dee231f019b68b170ae020beb2bd98b10b9213d22740e244c8fe362d5c0d071 +size 95766 diff --git a/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/35d364b8-3b24-4d25-b0d5-f4f2de5e4e95_origin.pdf b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/35d364b8-3b24-4d25-b0d5-f4f2de5e4e95_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eef866dbfcbb1f810e646a77083f324bfc2ff640 --- /dev/null +++ b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/35d364b8-3b24-4d25-b0d5-f4f2de5e4e95_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c10a4a67841d99edd47ae61b7d66c676f0ca2b06621640d27d6216fabdf06f60 +size 5796047 diff --git a/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/full.md b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d695d31a3c9297d79b4a4ba0eb7893238f46f5b6 --- /dev/null +++ b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/full.md @@ -0,0 +1,321 @@ +# SimMotionEdit: Text-Based Human Motion Editing with Motion Similarity Prediction + +Zhengyuan Li $^{1}$ Kai Cheng $^{1}$ Anindita Ghosh $^{2}$ Uttaran Bhattacharya $^{3}$ Liangyan Gui $^{4}$ Aniket Bera $^{1}$ $^{1}$ Purdue University $^{2}$ DFKI, MPI-INF, Saarland Informatics Campus + $^{3}$ Adobe Inc. $^{4}$ University of Illinois Urbana-Champaign + +# Abstract + +Text-based 3D human motion editing is a critical yet challenging task in computer vision and graphics. While training-free approaches have been explored, the recent release of the MotionFix dataset, which includes source-text-motion triplets, has opened new avenues for training, yielding promising results. However, existing methods struggle with precise control, often leading to misalignment between motion semantics and language instructions. In this paper, we introduce a related task - motion similarity prediction - and propose a multi-task training paradigm, where we train the model jointly on motion editing and motion similarity prediction to foster the learning of semantically meaningful representations. To complement this task, we design an advanced Diffusion-Transformer-based architecture that separately handles motion similarity prediction and motion editing. Extensive experiments demonstrate the state-of-the-art performance of our approach in both editing alignment and fidelity. Project URL: https://github.com/lzhyu/SimMotionEdit. + +# 1. Introduction + +Human motion synthesis plays a central role in diverse applications, including character animation [5, 15, 49], human-robot collaborations [3, 28], and autonomous driving [36]. Current approaches primarily focus on initial control signals, such as manipulating joint positions [56, 59] or applying fine-grained language instructions [26, 65]. With advancements in text-based motion generation [6, 14, 21, 22, 53], the ability to control and edit human motions has emerged as the next frontier, which can enable animators to refine motion sequences in animation workflows to fit specific requirements. In this work, we aim to address this need by focusing on editing existing human motions rather than generating them solely from other conditioning inputs. + +Recent works [8, 41] have leveraged pre-trained motion diffusion models with attention manipulation to edit motion + +![](images/329f1302030d9a43481d67562c58f516c382204d5481dab244924276f9a5e39c.jpg) +Figure 1. Text-Based Motion Editing. Our method SimMotionEdit generates edited human motion sequences from text instructions and source motion sequences. + +features, achieving effects such as style transfer and word emphasis. However, these methods are limited in scope due to the lack of dedicated model training and supervision. With the release of the text-based human motion editing dataset MotionFix [4], we can now incorporate robust supervision to develop SimMotionEdit, a system that enables free-form text-based human motion editing. + +The primary challenge in this task lies in aligning the generated motion with both the source motion and the text instructions, as they jointly define the desired outcome. Enhancing the interaction between the source motion and the text instructions is thus crucial for improved performance. To this end, the success of multi-task learning [67] has demonstrated that models can achieve better generalization by leveraging knowledge from other tasks. Auxiliary loss, a key concept in multi-task learning, has proven effective + +across various domains, including image classification [51], scene parsing [68], reinforcement learning [29], and code generation [13]. The essential criterion in designing auxiliary tasks is that they should support the learning of representations that benefit the main task. Consider the example of an animator instructed to edit a human motion sequence based on an editing description and a source motion. The animator does not begin editing immediately — a common first step is to identify the keyframes to modify. Similarly, for our model, identifying keyframes to edit is not only essential but also beneficial to performance. In many cases, the keyframes to edit are predictable. For example, given the instruction "don't turn back at the end; instead, keep walking", it is reasonable to anticipate frame changes at the point where the character starts turning back. As a result, the frames where the edit happens are dissimilar to the corresponding frames in the source motion. Following this intuition, we apply techniques such as sliding windows, value normalization, and weight balancing across different similarity measures to construct a predictable time series of motion similarity. We formulate the task as motion similarity prediction, where the model predicts the similarity between the source and the edited motions given the edit instructions. + +Given the multitask nature of learning motion editing via motion similarity prediction, we opt for the diffusion framework due to its flexibility. While various network architectures [53, 65, 66] have been proposed for human motion synthesis, they are not directly applicable to human motion editing, as most of them are designed to operate solely with text as input. Further, we separate our neural network into two parts: a condition transformer for motion similarity prediction and a diffusion transformer [37] for motion editing. This design choice offers two key advantages. First, the conditional transformer helps enhance the text and source motion features to embed information on motion similarity. The diffusion transformer effectively leverages those embeddings to generate the edited motion. Second, the condition transformer handles motion similarity prediction decoupled from the diffusion process, which allows each part of the neural network to specialize in its individual task and expedites as well as stabilizes training convergence. + +We summarize our key contributions as follows: + +- SimMotionEdit, a text-based human motion editing framework with an auxiliary task of motion similarity prediction to efficiently inter-relate source motion features and instruction text features. +- A Motion Diffusion Transformer architecture, trained on motion similarity prediction and edited motion generation to effectively learn text-based motion editing. +- Demonstrating the state-of-the-art performance of our architecture and our auxiliary task through extensive quantitative and qualitative experiments on the open-source dataset of MotionFix [4]. + +# 2. Related Works + +We briefly discuss existing literature on the related areas of 3D human motion editing, diffusion models for motion generation, and auxiliary tasks used to train neural networks. + +3D Human Motion Editing. Earlier works on 3D human motion editing have explored editing motions with space or time constraints [17, 18, 30]. However, generating realistic and coordinated motion editing typically requires users to manually provide such specific constraints. To this end, deep learning approaches have explored automating the process of human motion editing with simple inputs by incorporating different sub-tasks, such as motion stylization [1, 61], pose editing [35], and in-betweening [40, 47, 53, 54]. While techniques like motion stylization can perform global edits on motions, they do not enable precise control. Further, these approaches cannot be applied to the free-form text-based setting. Kim et al. [27] edit motions through diffusion inpainting, which requires a mask as input. Raab et al. [41] perform motion following by replacing the queries in the self-attention. MotionCLR [8] leverages attention manipulation to provide a set of motion editing methods. However, it does not directly take in texts as edit conditions. Goel et al. [19] heuristically select keyframes to edit certain types of motions. For text-based human pose editing, PoseFix [12] enables free-form text as input. Zhang et al. [65] relies on large language models to rewrite the script and generate edited motions and only considers the original motion at the text level. However, it cannot be applied without providing a description of the source motion. Huang et al. [24] prompts GPT to edit latent codes and generate edited human motions. MotionFix [4] provides a dataset and proposes a diffusion-based method to perform motion editing. Our task follows MotionFix [4], and we propose a novel architecture and an auxiliary task to improve generative performance. + +Motion Generation via Diffusion Models. Diffusion models [23] have been widely applied to human motion generation from text [9-11, 31, 32, 45, 53, 60, 63, 64] and other modalities [16, 44, 50, 54, 58]. The state-of-the-art motion diffusion models collapse text input into a single token. Shafir et al. [47] modify MDM's [53] motion inpainting process to control end-effector trajectories but require dense multi-modal inputs like entire joint trajectories instead of the free-form text-based instructions. With the introduction of recent text-motion 3D human datasets [20, 33, 39], there has been increased interest in conditioning the generation on free-form language inputs, in which [8, 47, 53, 55, 59, 64] have produced impressive results in free-from-text-based 3D human motion generation using diffusion models. + +![](images/7d747808c6c41b8c7b1482a11bef49cb8a44b26d9c3e10837ed8e6cb2952f1d2.jpg) +(a) Motion Diffusion Transformer + +![](images/aa55ac7eafc1da9c2345bfae571004d6075c59a5347b6ba757a2777b70e30cf9.jpg) +(b) Motion Similarity +Figure 2. Overview of SimMotionEdit. (a) The architecture consists of two modules: the condition transformer and the diffusion transformer. The condition transformer performs the auxiliary task of motion similarity prediction and enables the source motion features and the text features to mix. The diffusion transformer takes in the enhanced text features, the embedded diffusion step $t$ as the condition, the noisy edited motion, and the enriched source motion features, and predicts the denoised edited motions. (b) The auxiliary task motion similarity prediction is inspired by the fact that, given the text instructions, the similarity between source and edited motions is predictable. We use blue for the source motion, red for the edited motion, and orange for the generated motion. + +Auxiliary Tasks. Auxiliary tasks have played a significant role in enhancing the training of neural networks. Earlier works such as GoogLeNet (Inception v1) [51] added an auxiliary loss at intermediate layers to provide additional gradient signals to mitigate vanishing gradients. PSPNet[68] incorporated an auxiliary loss at intermediate layers for the scene parsing task. Auxiliary losses have also been leveraged in techniques including deep reinforcement learning [25, 29] and code generation [13] to enhance learning stability and feature extraction. Recently, REPA [62] finds that aligning the features of the denoiser and pre-trained discriminative models substantially boosts the training of diffusion models. Other works [46, 52] have also shown that auxiliary tasks can significantly improve the model's generalization by providing complementary signals to align the model's internal representations. However, to the best of our knowledge, auxiliary losses have not been explored previously for text-based human motion editing. + +# 3. Method + +Given a text instruction $L$ and a source motion sequence $X = [x^0, x^1, \ldots, x^F]$ with $F$ frames, our objective is to generate an edited sequence $M = [m^0, m^1, \ldots, m^{F'}]$ with $F'$ frames, where each $m^i \in \mathbb{R}^D$ represents a human pose, $D$ being the feature dimension. Following MotionFix [4], each $m^i$ consists of global velocity, global orientation, joint rotations, and local joint positions, totaling $D = 207$ . We + +illustrate our network architecture, the Motion Diffusion Transformer, in Fig. 2. It consists of a condition transformer and a diffusion transformer. The condition transformer processes the source motion $X$ and the text instructions $L$ , and provides enhanced condition features via motion similarity prediction (Sec. 3.2). The diffusion transformer takes in the concatenated sequence of enhanced source motion features and noised edited motion, and uses the enhanced text features as conditions, to generate edited motions (Sec. 3.3). We train the network jointly on motion similarity prediction and diffusion-based motion editing (Sec. 3.1). + +# 3.1. Diffusion-Based Motion Editing + +Denoising diffusion models [23, 48] are generative models that use a structured sequence of transformations based on a Markov chain. The diffusion process comprises a forward step, $q(M_t \mid M_{t-1})$ , where we iteratively add Gaussian noise to a clean edited motion sample $M_0$ over $T$ timesteps until it resembles the distribution of $p(M_T)$ : + +$$ +M _ {t} = \sqrt {\bar {\alpha} _ {t}} M _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \qquad (1) +$$ + +where $\bar{\alpha}_t\in (0,1)$ controls the rate of diffusion and $t\in [0,T]$ . When $\bar{\alpha}_t$ is small enough, we can approximate $M_T\approx \epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , $\mathbf{I}$ being the unit matrix. In the subsequent reverse step, the diffusion process samples motions from a multivariate Gaussian distribution $M_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ + +and we obtain the conditional probability of $M_0$ as + +$$ +p _ {\theta} \left(M _ {0} \mid c\right) = \int_ {M _ {1}} ^ {M _ {T}} p \left(m _ {T}\right) \prod_ {u = 1} ^ {T} p _ {\theta} \left(m _ {u - 1} \mid m _ {u}, c\right) d \mathbf {m}. \tag {2} +$$ + +For text-conditioned motion editing, we approximate $p_{\theta}((M_{t - 1}\mid M_t)$ with a learnable function $\mathcal{E}(M_t,t,L,X)$ which conditions the edited motion on both the text instruction $L$ and the source motion $X$ . Following recent approaches [4, 19, 43, 53], we predict the original motion signal $M_0 = \mathcal{E}(M_t,t,L,X)$ via the editing loss function, + +$$ +\mathcal {L} _ {e} = \mathbb {E} _ {M _ {0} \sim q (M _ {0} | L, X), t \sim [ 1, T ]} \left[ \| M _ {0} - \mathcal {E} \left(M _ {t}, t, L, X\right) \| _ {2} ^ {2} \right]. \tag {3} +$$ + +# 3.2. Motion Similarity Prediction + +Our auxiliary task of motion similarity prediction follows the intuition that to edit a motion, the model must first identify the parts that need editing. Put differently, given the source and the edited motions, it should quantify the similarities between them. To design this auxiliary task, we consider the following two aspects: + +- Constructing a predictable similarity curve (over time) between the source and the edited motions, and +- Designing an auxiliary loss function that synergizes with motion editing. + +# 3.2.1. Predictable Motion Similarity Curve + +In our work, we consider the motion similarity $\mathbf{S}^R$ between two 3D motion sequences in the space of joint rotations to leverage its scale-agnostic nature. In the rotation space, we define the raw similarity $S_{i}^{Rr}$ of one frame $i$ in $X$ as the minimum distance between this frame and any frame within a corresponding sliding window of the edited motion, as + +$$ +S _ {i} ^ {R r} = - \min _ {| i - j | \leq W} d _ {r} \left(x ^ {i}, m ^ {j}\right), \tag {4} +$$ + +where $W$ is the size of the sliding window, and $d_{r}$ represents a distance metric (e.g., Euclidean) between the source and the edited poses. The sliding window approach allows each source frame to find its best match in the edited motion within an acceptable range of misalignment. Note that if, instead, we calculate the difference between corresponding frames with identical indices, i.e., $x^{i}$ and $m^{i}$ , the raw similarity $S^{r}$ would be low for motions that are nearly identical but shifted by a few frames. Symmetrically, we define $S_{j}^{Rl}$ in the joint location space. We balance the effects of $S^{Rl}$ and $S^{Rr}$ by empirically weighting them, as + +$$ +S _ {i} ^ {R} = w _ {1} S _ {i} ^ {R r} + w _ {2} S _ {i} ^ {R l}. \tag {5} +$$ + +However, $S_{i}^{R}$ is not entirely predictable due to two issues. First, different editing instructions can result in varying scales of change. For instance, both (a) and (b) in Fig. 3 + +![](images/0374bcc1dada371512a1e4f1dec83c0fd669cea28a4c653eaf008bb1e7841d14.jpg) + +![](images/f97874c0c3d64e74f3538f13eac03a7db25820d2b4d904791b9c009f2b449275.jpg) + +![](images/36db36ada438cc5c50a8b1e7ed8d892f6844a261ef8531c37c2328ab9e6ce9ef.jpg) + +![](images/0017d37777ff5259fd550cb6824dbd93d03c54f7c0925772b84274fe568314fa.jpg) + +![](images/f823ee598daddbbc707bfa3b60e17d565d5c44ffeba57b3c6065fa720107ddd8.jpg) + +![](images/2498b88d7668c9b856d7a1b8ca1c012f6b275c7657578e33bd870a1d80b22034.jpg) +(c) +Figure 3. Raw Motion Similarity. We translate the global positions of sampled poses of the source motion and the edited motion for a clear view. + +represent simple edits. Yet, $S_{i}^{R}$ starts near 0 in (a) and near -1 in (b). Such variations are challenging for the model to learn. Second, when the edit is fine-grained, the subtle changes make the curve appear random. For example, in Fig. 3 (c), the hand movements corresponding to the text instructions "don't move hands and put weight on left foot" are subtle in the source motion, so the differences in the edited motion are minimal. As a result, the corresponding raw motion similarity curve lacks any identifiable patterns. + +To address the first issue, we normalize all values in $S^R$ to [0, 1] with min-max normalization and make the curve scale-agnostic, as + +$$ +S _ {i} ^ {N} = \frac {S _ {i} ^ {r} - \operatorname* {m i n} _ {j} S _ {j} ^ {R}}{\operatorname* {m a x} _ {j} S _ {j} ^ {R} - \operatorname* {m i n} _ {j} S _ {j} ^ {R}}. \tag {6} +$$ + +Our sliding window (Eq. (4)) and normalization (Eq. (6)) approaches also address the second issue. In addition, we filter out data samples that exhibit a sufficiently large semantic gap between the text instructions and the motion similarity from our model training, as these introduce unnecessary noise in the similarity curve. To this end, we define a motion signal-to-noise ratio (MotionSNR) based on the similarities of the top- $\kappa$ and bottom- $\kappa$ frames in a se + +quence, calculated as follows: + +$$ +\begin{array}{l} \mathbf {T} ^ {R} = \left\{S _ {k} ^ {R} \mid S _ {k} ^ {R} \geq S _ {j} ^ {R} \quad \forall S _ {j} ^ {R} \in \mathbf {S} ^ {R} - \mathbf {T} ^ {R} \right. \\ \left. \left| \mathbf {T} ^ {R} \right| = \kappa \right\}, \tag {7} \\ \end{array} +$$ + +$$ +\mathbf {B} ^ {R} = \left\{S _ {k} ^ {R} \mid S _ {k} ^ {R} \leq S _ {j} ^ {R} \quad \forall S _ {j} ^ {R} \in \mathbf {S} ^ {R} - \mathbf {B} ^ {R} \right. +$$ + +$$ +\left. \left| \mathbf {B} ^ {R} \right| = \kappa \right\}, \tag {8} +$$ + +$$ +\text {M o t i o n S N R} = \frac {\sum_ {x \in \mathbf {T} ^ {R}} x}{\sum_ {x \in \mathbf {B} ^ {R}} x}. \tag {9} +$$ + +For our model training, we choose $\kappa = 5$ . + +Ideally, the motion for non-edited frames should remain the same as the source motion, while for edited frames, it should differ. Therefore, MotionSNR should be infinite for an ideal pair of source and edited motions. In practice, motion pairs with high MotionSNR should be easier for training, and the similarities should be easier to predict. Therefore, we filter out sequences with low MotionSNR using an empirically determined threshold. We verify the impact of this filtering approach in Sec. 4.2. + +# 3.2.2. Auxiliary Loss Function + +An underlying assumption of text-based motion editing is that the source motion and the text instructions should collectively predict the edited motion. By extension, we assume that the source motion and the text instructions should also collectively predict the motion similarity. Additionally, we note that motion similarity should contain sufficient information for motion editing. For instance, if the text instruction is "use both hands to catch the ball", motion similarity should identify which frames involve using only one hand to catch the ball and, therefore, need editing. We capture this notion of motion similarity through our auxiliary loss function. To implement this loss, instead of directly regressing the motion similarity curve, we quantize the values in $S^{N}$ into $K$ classes. Given motion similarity $S_{i}^{N}$ , we use a quantization function $\mathcal{Q}(\cdot)$ as + +$$ +\mathfrak {s} _ {i} = \mathcal {Q} \left(S _ {i} ^ {N}\right) := \left\{ \begin{array}{l l} 0, & S _ {i} ^ {N} < \tau_ {0}, \\ 1, & \tau_ {0} \leq S _ {i} ^ {N} < \tau_ {1}, \\ \vdots & \vdots \\ K - 1, & S _ {i} ^ {N} \geq \tau_ {K - 2}, \end{array} \right. \tag {10} +$$ + +where $\mathfrak{s}_i\in \{0,\ldots ,K - 1\}$ provides the class label corresponding to each similarity level. We split the space [0,1] into $K$ equal-length intervals for the similarity levels. + +For each frame $i$ , assuming the network outputs logits $\mathbf{z}_i = [z_{i,0},\dots ,z_{i,K - 1}]$ , we convert logits to probabilities using the softmax function + +$$ +p _ {i, k} = \frac {\exp \left(z _ {i , k}\right)}{\sum_ {l = 0} ^ {K - 1} \exp \left(z _ {i , l}\right)} \quad \forall k \in \{0, \dots , K - 1 \}. \tag {11} +$$ + +We can then write the cross-entropy loss for the classification task, which gives our auxiliary loss function, as + +$$ +\mathcal {L} _ {\text {a u x}} = - \frac {1}{F} \sum_ {i = 0} ^ {F - 1} \log p _ {i, \mathfrak {s} _ {i}} \tag {12} +$$ + +Quantizing the similarity values provides the following benefit. For a given pairing of source motion, text instructions, and edited motion, the edited motion is only one of all feasible edited motions given just the source motion and the text instructions. Therefore, regression-based learning confines the model's ability to see only the provided edited motions as feasible. Quantization, by contrast, allows the model to focus on the coarser similarity levels rather than perfect frame-level similarity, striking a balance between optimizing the auxiliary and the editing losses and also enabling the model to generalize to a wider range of text instructions and motion variations than what is available from the dataset. We further validate this design choice experimentally in Tab. 3. + +Lastly, combining the editing and the auxiliary losses (Eqs. (3) and (12)), we write the total training loss as + +$$ +\mathcal {L} = \mathcal {L} _ {\text {a u x}} + \mathcal {L} _ {\mathrm {e}}. \tag {13} +$$ + +# 3.3. Motion Diffusion Transformer + +We combine diffusion-based motion editing and the auxiliary task of motion similarity prediction into our Motion Diffusion Transformer model Fig. 2. Our condition transformer follows the traditional transformer encoder architecture. It takes in the source motions and the CLIP features [42] of the text instructions, and outputs enhanced text features and motion features that contain sufficient information for the desired edited motions. This enhancement of the text and motion features via their intermixing is a direct consequence of using the auxiliary loss function (Eq. (12)). The diffusion transformer takes in the enhanced motion features concatenated with the noised edited motions. We also inject the enhanced text features into the diffusion transformer through an AdaLN-Zero layer, alongside the diffusion timestep $t$ . We use the editing loss (Eq. (3)) for the model to learn the edited motions at the output of the diffusion transformer, and train it jointly with the auxiliary task. We adopt the standard DDPM diffusion paradigm [23] for both training and inference. + +Our design ensures that the inputs to the auxiliary task are decoupled from the motion editing task. This is desirable since any changes in the noised edited motion (part of the diffusion transformer's inputs) should not affect the model's learning of the auxiliary task. Conversely, our design also ensures that the motion editing task fully leverages the enhanced motion and text features learned from the auxiliary task by directly using them as inputs for the denoising diffusion process. + +# 4. Experiments and Results + +We detail our experimental setup, including dataset, metrics, baselines, and implementation details, and elaborate on the results. + +# 4.1. Setup + +Data. The MotionFix dataset [4] is a language-based 3D motion editing dataset. Each data sample is a triplet consisting of a source motion, a target motion, and a text instruction. To ensure quality, semantically similar motion pairs are selected from motion capture (MoCap) data using the TMR motion embedding space [38]. Annotators then manually described the differences in a concise manner. MotionFix provides a total of 6,730 annotated triplets split into train, validation, and test sets. + +Evaluation Metrics. For text-based human motion editing, we need to evaluate both the semantic alignment of the edited motions with the text instructions and the source motions, and the fidelity of the edited motions. Following the approach of MotionFix [4], we evaluate alignment based on motion-to-motion retrieval. We utilize pre-trained TMR [38] as a feature extractor and compute the retrieval accuracy of the generated motion in a fixed-size batch of motions. We report the top-1, top-2, and top-3 accuracies $(\mathbb{R}@\mathbb{1},\mathbb{R}@\mathbb{2},$ and $\mathbb{R}@\mathbb{3})$ both in batch sizes of 32 and in the full test set. We also report the average rank for comparison. To evaluate fidelity, we utilize the MotionCritic score (M-score) [57]. M-score has been shown to align with human perceptual preferences, thus serving as a reasonable metric for motion realism. To compute M-scores, we use the authors' pre-trained model and transform our generated motions to match the pre-trained model's input. + +Baselines. We compare our method with three baselines, MDM [53], MDM-BP [4], and TMED [4]. + +- MDM utilizes the pre-trained human motion diffusion model. The edit instruction is taken as text input. +- MDM-BP enhances MDM by introducing body-part detection using a GPT-based annotation system [2]. The method identifies body parts not mentioned in the text instructions, allowing the model to mask these parts and retain their source motion features. +- TMED is a conditional diffusion model specifically trained on the MotionFix dataset using motion-motion-text triplets. Unlike the baselines trained only on text-motion pairs, TMED leverages both the source motions and the text instruction as conditions to generate the desired edited motions. + +Each of our baselines serves a specific role in evaluations. MDM represents the editing power of the diffusion model without the knowledge of source motions. MDM-BP knows + +both the source motions and the text instructions but is not trained on text-based motion editing datasets. TMED takes both source motions and text as input and is trained on the dataset, which is comparable to our approach, albeit with a different model architecture. + +Implementation Details. We fix the DDPM [23] diffusion step to 300 and use the cosine noise scheduler. Following [4] and [7], we adapt two-way conditioning with a guidance scale of 2 for both the text and the source motion. We utilize pre-trained CLIP [42] from ViT-B/32 to encode the text. Our condition and diffusion transformers consist of 4 and 8 transformer encoder layers, respectively, 8 attention heads, and latent dimensions of 512. We train using a batch size of 128, and the AdamW optimizer [34] with a learning rate of $10^{-4}$ . We train our model for 1,500 epochs, which takes around a day on a single A100 GPU. + +# 4.2. Quantitative Comparisons + +We report quantitative comparisons in Tab. 1. Our model outperforms all baselines significantly in both motion-editing alignment and fidelity. MDM shows the weakest performance, aligning with the intuition that text instructions by themselves do not provide sufficient information to generate source-aligned edited motions. With GPT assistance in identifying edited body parts, MDM-BP shows substantial improvements over MDM. TEMD considers the influence of text instructions on the source motion and further outperforms MDM-BP. Our Motion Diffusion Transformer, together with the auxiliary loss for the learning of motion similarities, results in state-of-the-art performance. We attribute the progressive performance improvements to the knowledge of the source motion, effective dataset utilization, a specially designed network architecture, and an appropriate auxiliary loss function. We also evaluate our variant without filtering out noisy motion similarity curves for the auxiliary loss (w/o filtering - Tab. 1, last row). The performance downgrade implies that noisy data for motion similarity prediction negatively affects the training. + +# 4.3. Qualitative Results + +We show visual results compared with TMED [4] in Fig. 4. Our method improves on various generative artifacts of TMED, such as self-penetrations (row 1), failure to follow instruction nuances (row 2 – not generating motions with the right leg lifted, and row 3 – not generating the sitting down motion to be slower), and failure to maintain the effects of specific instructions (row 4 – lowering the hands instead of continuing to wave, row 5 – unexpectedly stepping forward, and row 6 – moving the arms beyond 90 degrees). These examples also illustrate the realism of our generated edited motions. + +![](images/26c30cc6dcf3242bd4c017110bba08c273a99d52b8dada7430f19defb14275ab.jpg) +Figure 4. Qualitative Results. We compare our method with TMED [4]. Our method outperforms TMED in terms of both fidelity and alignment with source motion and text instructions. + +Table 1. Comparison of Text-Based Motion Editing on the MotionFix [4] Dataset. Our method outperforms all other baselines on generated-to-target retrieval and M-score. We report R@1, R@2 and R@3 as percentages. ↑ indicates higher values are better, ↓ indicates lower values are better, bold indicates best, and underline indicates second best. + +
MethodsGenerated-to-Target (Batch)Generated-to-Target (Test Set)M-score↑
R@1↑R@2↑R@3↑AvgR↓R@1↑R@2↑R@3↑AvgR↓
Ground Truth100.0100.0100.01.0064.3688.7595.561.74-3.175
MDM [53]4.037.5610.4815.550.100.100.10--
MDM-BP [4]39.1050.0954.846.468.6914.7118.36180.99-
TMED [4]62.9076.5183.062.7114.5121.7228.7356.63-3.512
Ours70.6282.9288.122.3825.4939.3349.2123.49-3.210
Ours (w/o aux loss)67.2178.7586.332.4923.3236.9646.4426.02-3.422
Ours (w/o filtering)66.8880.4285.832.4623.7335.5745.3624.94-3.418
+ +Table 2. Performances of SimMotionEdit with Different Combinations of Condition Features. In the "no text feature" setting, the input to the condition transformer is the raw text. Using enhanced features for both text and motion leads to the best overall performance for motion realism. + +
Text Feat.Motion Feat.Generated-to-Target (Batch)M-Score↑
R@1↑R@2↑R@3↑
RawEnhanced70.4281.2586.67-3.471
RawRaw67.2178.7586.33-3.422
EnhancedEnhanced70.6282.9288.12-3.210
EnhancedRaw69.0879.7586.59-3.450
EnhancedNo66.4681.2584.38-3.611
NoEnhanced71.0483.9689.58-3.318
+ +# 4.4. Ablation Study + +We perform ablations on both the Motion Diffusion Transformer and motion similarity prediction. + +# 4.4.1. Motion Diffusion Transformer + +We report comparisons with variations of our architecture in Tab. 2. The variant with both enhanced text and enhanced source motion features outperforms most of the other variants. It falls marginally behind on motion alignment and somewhat ahead on motion fidelity compared to using enhanced motion features with the raw text, indicating the strong influence of the source motion on aligning the edited motion but not having much influence on its fidelity. This outcome also suggests that the joint enhancement of both text and motion features provides a better balance between alignment and realism (fidelity). With these findings, we validate the effectiveness of enhanced features produced by the condition transformer and motion similarity prediction. + +# 4.4.2. Motion Similarity Prediction + +We report the results of abating the auxiliary loss in Tab. 3. We find that quantizing the motion similarity values into + +Table 3. Performances of SimMotionEdit with Different Auxiliary Losses. Model performance improves when using a classification loss over a regression loss for the auxiliary task, indicating the better learning capacity enabled by classification. However, performance also decreases when increasing the number of classes beyond three, which brings the classes too close to each other. + +
Loss#classesGenerated-to-Target (Batch)M-Score↑
R@1↑R@2↑R@3↑
Reg.∞68.9682.7187.08-3.750
Cls.370.6282.9288.12-3.210
Cls.567.0880.2186.46-3.186
Cls.966.8882.2986.88-3.722
+ +three classes yields the best results, implying that the auxiliary task should be simple enough to not overpower the main editing task. Further, a regression-based auxiliary loss reduces the motion fidelity given the overfitting tendencies when attempting to precisely predict the motion similarity curves. + +# 5. Conclusion + +We have presented SimMotionEdit, a state-of-the-art solution for text-based human motion editing. Our method utilizes the Motion Diffusion Transformer architecture to inject the source motion and the text instructions into a diffusion model for predicting the edited motions. We also leverage the novel auxiliary task of motion similarity prediction to further enhance the interplay of text and source motion features in the context of motion editing. Extensive experiments demonstrate that SimMotionEdit outperforms prior works in both alignment and realism of the edited motions. Through detailed ablation studies, we verify the effects of the core components of our method. We hope our work can push the limits of text-based human motion editing and inspire future research. + +# References + +[1] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. ACM Transactions on Graphics (TOG), 39(4):64, 2020. 2 +[2] 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, 2023.6 +[3] Sean Andrist, Bilge Mutlu, and Adriana Tapus. Look like me: matching robot personality via gaze to increase motivation. In Proceedings of the 33rd annual ACM conference on human factors in computing systems, pages 3603-3612, 2015. 1 +[4] Nikos Athanasiou, Alpar Ceske, Markos Diomataris, Michael J. Black, and Gül Varol. MotionFix: Text-driven 3d human motion editing. In ACM SIGGRAPH Asia, 2024. 1, 2, 3, 4, 6, 7, 8 +[5] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In Proceedings of the 29th ACM International Conference on Multimedia, page 2027-2036, New York, NY, USA, 2021. Association for Computing Machinery. 1 +[6] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In VR, pages 1-10, 2021. 1 +[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 6 +[8] Ling-Hao Chen, Wenxun Dai, Xuan Ju, Shunlin Lu, and Lei Zhang. Motioncr: Motion generation and training-free editing via understanding attention mechanisms. arXiv preprint arXiv:2410.18977, 2024. 1, 2 +[9] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000–18010, 2023. 2 +[10] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, pages 9760-9770, 2023. +[11] Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real-time controllable motion generation via latent consistency model. ECCV, 2024. 2 +[12] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: Correcting 3d human poses with natural language. In ICCV, pages 15018-15028, 2023. 2 +[13] Yifeng Ding, Hantian Ding, Shiqi Wang, Qing Sun, Varun Kumar, and Zijian Wang. Horizon-length prediction: Advancing fill-in-the-middle capabilities for code generation + +with lookahead planning. arXiv preprint arXiv:2410.03103, 2024. 2, 3 +[14] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Text-based motion synthesis with a hierarchical two-stream rnn. In ACM SIGGRAPH 2021 Posters, New York, NY, USA, 2021. Association for Computing Machinery. 1 +[15] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. Computer Graphics Forum, 42(2):1-12, 2023. 1 +[16] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part XXXVI, page 418–437, Berlin, Heidelberg, 2024. Springer-Verlag. 2 +[17] Michael Gleicher. Motion editing with spacetime constraints. In Proceedings of the 1997 Symposium on Interactive 3D Graphics, page 139-ff., New York, NY, USA, 1997. Association for Computing Machinery. 2 +[18] Michael Gleicher. Motion path editing. In Proceedings of the 2001 Symposium on Interactive 3D Graphics, page 195-202, New York, NY, USA, 2001. Association for Computing Machinery. 2 +[19] Purvi Goel, Kuan-Chieh Wang, C Karen Liu, and Kayvon Fatahalian. Iterative motion editing with natural language. In ACM SIGGRAPH, pages 1-9, 2024. 2, 4 +[20] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, pages 5152-5161, 2022. 2 +[21] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597, 2022. 1 +[22] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, pages 1900-1910, 2024. 1 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2, 3, 5, 6 +[24] Yiming Huang, Weilin Wan, Yue Yang, Chris Callison-Burch, Mark Yatskar, and Lingjie Liu. Como: Controllable motion generation through language guided pose code editing. In Computer Vision - ECCV 2024, pages 180-196, Cham, 2025. Springer Nature Switzerland. 2 +[25] Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv preprint arXiv:1611.05397, 2016. 3 +[26] Peng Jin, Yang Wu, Yanbo Fan, Zhongqian Sun, Wei Yang, and Li Yuan. Act as you wish: Fine-grained control of motion diffusion model with hierarchical semantic graphs. NeurIPS, 36, 2024. 1 + +[27] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In AAAI, pages 8255-8263, 2023. 2 +[28] Hema Swetha Koppula and Ashutosh Saxena. Anticipating human activities for reactive robotic response. In IROS, page 2071. Tokyo, 2013. 1 +[29] Guillaume Lample and Devendra Singh Chaplot. Playing fps games with deep reinforcement learning. In AAAI, 2017. 2, 3 +[30] Jehee Lee and Sung Yong Shin. A hierarchical approach to interactive motion editing for human-like figures. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, page 39-48, USA, 1999. ACM Press/Addison-Wesley Publishing Co. 2 +[31] Han Liang, Jiacheng Bao, Ruichi Zhang, Sihan Ren, Yuecheng Xu, Sibei Yang, Xin Chen, Jingyi Yu, and Lan Xu. Omg: Towards open-vocabulary motion generation via mixture of controllers. In CVPR, pages 482-493, 2024. 2 +[32] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. *IJCV*, pages 1–21, 2024. 2 +[33] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-x: A large-scale 3d expressive whole-body human motion dataset. NeurIPS, 2023. 2 +[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6 +[35] Boris N. Oreshkin, Florent Bocquelet, Félix G. Harvey, Bay Raitt, and Dominic Laflamme. Protores: Proto-residual network for pose authoring via learned inverse kinematics. In ICLR, 2022. 2 +[36] Brian Paden, Michal Čáp, Sze Zheng Yong, Dmitry Yershov, and Emilio Frazzoli. A survey of motion planning and control techniques for self-driving urban vehicles. IEEE Transactions on intelligent vehicles, 1(1):33-55, 2016. 1 +[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2 +[38] Mathis Petrovich, Michael J Black, and Gül Varol. Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In ICCV, pages 9488-9497, 2023. 6 +[39] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In CVPR, pages 722-731, 2021. 2 +[40] Jia Qin, Youyi Zheng, and Kun Zhou. Motion in-between via two-stage transformers. ACM Trans. Graph., 41(6), 2022. 2 +[41] Sigal Raab, Inbar Gat, Nathan Sala, Guy Tevet, Rotem Shalev-Arkushin, Ohad Fried, Amit H Bermano, and Daniel Cohen-Or. Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. ACM SIGGRAPH Asia, 2024. 1, 2 +[42] 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. 5, 6 +[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 4 +[44] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In CVPR, 2023. 2 +[45] Jiawei Ren, Mingyuan Zhang, Cunjun Yu, Xiao Ma, Liang Pan, and Ziwei Liu. Insactor: Instruction-driven physics-based characters. NeurIPS, 2023. 2 +[46] S Ruder. An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098, 2017. 3 +[47] Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 2 +[48] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 3 +[49] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39 (4):54-1, 2020. 1 +[50] Kewei Sui, Anindita Ghosh, Inwoo Hwang, Jian Wang, and Chuan Guo. A survey on human interaction motion generation. arXiv preprint arXiv:2503.12763, 2025. 2 +[51] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, pages 1-9, 2015. 2, 3 +[52] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 30, 2017. 3 +[53] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1, 2, 4, 6, 8 +[54] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In CVPR, pages 448-458, 2023. 2 +[55] Weilin Wan, Yiming Huang, Shutong Wu, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Diffusionphase: Motion diffusion in frequency domain. arXiv preprint arXiv:2312.04036, 2023. 2 +[56] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. ECCV, 2024. 1 +[57] Haoru Wang, Wentao Zhu, Luyi Miao, Yishu Xu, Feng Gao, Qi Tian, and Yizhou Wang. Aligning motion generation with human perceptions. In ICLR, 2025. 6 +[58] Dong Wei, Xiaoning Sun, Huaijiang Sun, Shengxiang Hu, Bin Li, Weiqing Li, and Jianfeng Lu. Enhanced fine-grained motion diffusion for text-driven human motion synthesis. In AAAI, pages 5876-5884, 2024. 2 + +[59] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. In ICLR, 2024. 1, 2 +[60] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, pages 14928-14940, 2023. 2 +[61] Wenjie Yin, Hang Yin, Kim Baraka, Danica Kragic, and Marten Björkman. Dance style transfer with cross-modal transformer. In WACV, pages 5047-5056, 2023. 2 +[62] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024.3 +[63] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In ICCV, 2023. 2 +[64] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4115-4128, 2024. 2 +[65] Mingyuan Zhang, Huirong Li, Zhongang Cai, Jiawei Ren, Lei Yang, and Ziwei Liu. Finemogen: Fine-grained spatiotemporal motion generation and editing. NeurIPS, 36, 2024. 1, 2 +[66] Mingyuan Zhang, Daisheng Jin, Chenyang Gu, Fangzhou Hong, Zhongang Cai, Jingfang Huang, Chongzhi Zhang, Xinying Guo, Lei Yang, Ying He, et al. Large motion model for unified multi-modal motion generation. In ECCV, pages 397-421. Springer, 2025. 2 +[67] Yu Zhang and Qiang Yang. A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34 (12):5586-5609, 2021. 1 +[68] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 2881-2890, 2017. 2, 3 \ No newline at end of file diff --git a/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/images.zip b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d94fc5fd5021f39ab9701e2e2761855ac9532eea --- /dev/null +++ b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7288038866c5da0691880edb1a5a01e7ae7861e3ff85e6350336ce1588af5e9 +size 559445 diff --git a/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/layout.json b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..560e5e5054bdfe233cbcd7a31499efbd6d8f9d22 --- /dev/null +++ b/CVPR/2025/SimMotionEdit_ Text-Based Human Motion Editing with Motion Similarity Prediction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cfde138aecec55ea5dd76798cc0294c47f7673f96b0ce33e400195214f5a6ee +size 370473 diff --git a/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/dfa0e139-5a4d-4319-9718-259db56b3f39_content_list.json b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/dfa0e139-5a4d-4319-9718-259db56b3f39_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..52b36121326013223ffcb2498ea4d97322c64687 --- /dev/null +++ b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/dfa0e139-5a4d-4319-9718-259db56b3f39_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bca7782fbbf705cdd1f125b555e814448697e1d502f7b6dcf1d1642d07dd177 +size 95039 diff --git a/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/dfa0e139-5a4d-4319-9718-259db56b3f39_model.json b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/dfa0e139-5a4d-4319-9718-259db56b3f39_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c427b7aa0a2e91aa0da4beff7e4d15dad83abca9 --- /dev/null +++ b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/dfa0e139-5a4d-4319-9718-259db56b3f39_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e43bb977e51c111b8719eed986a7619cdd3c24ac566039d1ee4cf2e8beef1fbf +size 115029 diff --git a/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/dfa0e139-5a4d-4319-9718-259db56b3f39_origin.pdf b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/dfa0e139-5a4d-4319-9718-259db56b3f39_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0472792c05610ed59722ab1cfdbe8a37b4195552 --- /dev/null +++ b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/dfa0e139-5a4d-4319-9718-259db56b3f39_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e3b9d552a62c865004b8679393f6f1569614509086915337ec2c45e7f30bebf +size 7892163 diff --git a/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/full.md b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..438eba3cf36d06918175be5b3b681a2b8e785efe --- /dev/null +++ b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/full.md @@ -0,0 +1,455 @@ +# SimVS: Simulating World Inconsistencies for Robust View Synthesis + +Alex Trevithick $^{1,2}$ , Roni Paiss $^{2}$ , Philipp Henzler $^{3}$ , Dor Verbin $^{2}$ , Rundi Wu $^{2,4}$ , Hadi Alzayer $^{2,5}$ , Ruiqi Gao $^{2}$ , Ben Poole $^{2}$ , Jonathan T. Barron $^{2}$ , Aleksander Holynski $^{2}$ , Ravi Ramamoorthi $^{1}$ , Pratul P. Srinivasan $^{2}$ + +$^{1}$ UC San Diego + +2Google DeepMind + +3Google Research + +4Columbia + +5Univ. Maryland + +![](images/a59295688af289b23ab39eecfe877249d5885ff2e9142d963aca1f037d7c5b29.jpg) +(a) 8 casually-captured posed images with scene motion + +![](images/8d64cbf02885ef6d82866ec3923ae240f0a43c385ce5688d5b134636e9df4d9a.jpg) + +![](images/e6291f0614621219a9724956888859d6b53b510c5ffc3f123f4eb0ebc76c4f43.jpg) +(b) Our model frozen to state 1 + +![](images/3810ebd236f22dd6964f8a68a40819535c221137ed51bb6de647d5450c51956c.jpg) + +![](images/42f7f34758d9c726a27a273889eed5b6eaefb42a56f4793b1f1bf1a9cc7101b4.jpg) +(c) Our model frozen to state 2 + +![](images/4c21586a10cea5d6f126fab34541dd90065a5b8b8f5818258e601a60b35dadb2.jpg) + +![](images/497fa5a96b2e273d179be4058ebccda52ec6925f22621bf711b25a5aebf89862.jpg) +(d) CAT3D (all images) +Figure 1. We show results of our model applied to a casual in-the-wild capture. (a) Given 8 unordered images of a scene with significant motion and desired states marked in blue and orange, our model generates a 3D representation for each desired state shown in corresponding colors in (b) and (c). The CAT3D baseline [15] in (d) cannot disentangle the different states, resulting in catastrophic failure. + +Novel view 1 + +Novel view 2 + +# Abstract + +Novel-view synthesis techniques achieve impressive results for static scenes but struggle when faced with the inconsistencies inherent to casual capture settings: varying illumination, scene motion, and other unintended effects that are difficult to model explicitly. We present an approach for leveraging generative video models to simulate the inconsistencies in the world that can occur during capture. We use this process, along with existing multi-view datasets, to create synthetic data for training a multi-view harmonization network that is able to reconcile inconsistent observations into a consistent 3D scene. We demonstrate that our world-simulation strategy significantly outperforms traditional augmentation methods in handling real-world scene variations, thereby enabling highly accurate static 3D reconstructions in the presence of a variety of challenging inconsistencies. Project page: https://alextrevithick.com/simvs + +# 1. Introduction + +View synthesis, the task of creating images from unobserved camera viewpoints given a set of posed images, has seen remarkable progress in recent years. Current algorithms are able to render detailed photorealistic novel views of complicated 3D scenes. However, these techniques tend to assume that the provided input images are consistent — that the geometry and illumination of the scene is static during capture. Typical captures of real-world scenes seldom obey this constraint; people and objects may move and deform, and lights may move or change brightness. + +Moreover, casual captures outside of tightly-controlled settings tend not only to be inconsistent but also sparse, containing only a small number of observed views. Methods for sparse view synthesis are usually trained on synthetic or captured multiview datasets that are consistent by design, and therefore fail to generalize to the inconsistencies seen in real-world casual captures (see Fig. 6 as an example). + +![](images/ae321927fb7656fba82276465b542abfc6ec185027909f67826f10a65896bbbb.jpg) +Figure 2. A comparison of real world state changes, those simulated through a video model, and heuristic augmentations (random sparse flow fields for dynamics and random color tints for lighting). + +We address the problem of robust view synthesis from sparse captures in a new way by leveraging the ability of video diffusion models to simulate plausible world inconsistencies that could arise during capture. There has been considerable speculation about the usefulness of large video models as simulators or "world models" [29, 45] and in this work we demonstrate a new use case for their simulation capabilities. + +Our approach augments existing multiview datasets with inconsistencies simulated by a video diffusion model and trains a multiview harmonization model to sample sets of consistent views of a scene conditioned on sparse inconsistent captures. We can then use existing 3D reconstruction and view synthesis techniques to synthesize novel viewpoints from these consistent images. + +In summary, our key technical contributions are: + +- A generative data augmentation strategy that leverages video diffusion models to sample world inconsistencies (e.g. scene motion and lighting changes) that could arise during capture (Section 3) +- A multiview harmonization model, trained on this generated data, that converts inconsistent sparse input images into a set of consistent images (Section 4) + +We demonstrate that our generative augmentation strategy outperforms other alternatives such as using heuristic data augmentation or synthetic rendered data, and that novel views rendered from our harmonization model are superior to those from existing approaches for sparse and robust view synthesis. We encourage readers to view the video results on the project page. + +# 2. Related Work + +We address the task of view synthesis from sparse and inconsistent images of a scene. While existing techniques address view synthesis from densely-sampled inconsistent inputs or sparse consistent inputs, to our knowledge no existing method is capable of synthesizing novel views of full scenes from images that are both sparse and inconsistent. + +# 2.1. Robust view synthesis + +Prior methods for robust view synthesis typically require dense captures with hundreds of images. and focus on explicitly modeling a specific source of inconsistency (either motion or lighting) as part of reconstruction. + +Scene dynamics In the case of scene dynamics, some works handle only transient occluders in dense, multiview captures [39-41]. Most existing methods start with a dense video and attempt to recover motion flows or trajectories to explain the observed motion. Early approaches based on Neural Radiance Fields [34] optimized time-varying flow fields to explain observed motion as deformations of an underlying consistent scene representation [12, 24, 35, 36, 38, 53]. Later NeRF-based methods improved quality further through prior integration [25, 30]. The most recent state-of-the-art methods have adopted 3D Gaussian representations and optimized explicit motion trajectories for this particle-based scene representation [22, 50, 56], often leveraging strong priors from pretrained monocular depth [13, 20, 64], optical flow [52], or tracking models [10, 19, 62]. These temporal priors have been crucial for rendering high-quality novel views in the dense capture setting, but tend to break down when applied to sparse or unordered captures. + +Lighting inconsistencies Existing structure-from-motion pipelines display remarkable robustness to lighting variation [42-44], enabling 3D reconstruction from large-scale in-the-wild images [1]. To model inconsistencies due to changing scene lighting, 3D reconstruction and view synthesis techniques use per-image "appearance embeddings" that allow for the appearance of scene content to vary across observations [8, 21, 32, 33, 58, 63, 67]. This strategy can successfully model lighting inconsistencies given dense captures with smoothly-varying appearance changes, but is unable to reconcile large changes in appearance in sparsely-sampled captures. Some methods [37, 69] assume consistent captures and generate inconsistencies with generative models for relighting. Others [17, 66] can perform object-level relighting from one image, but cannot incorporate information from a sparse set of views. + +# 2.2. Sparse view synthesis + +In novel view synthesis settings with only a few captured views, most methods rely on strong priors learned from large multiview datasets. Some methods train feedforward models to directly predict 3D representations that can be used for view synthesis [7, 16, 18, 54, 65, 71]. Others rely on pretrained monocular depth, multiview stereo, or inpainting networks and rely on test-time optimization to fit a scene [11, 47, 48, 55, 59]. A recent class of methods has achieved high visual quality by directly generating images from novel viewpoints using diffusion models conditioned on observed image(s) and target camera poses [15, 27, 28, 31, 46, 61]. In particular, the multiview + +![](images/57a25ecbc4ea73d3f746789301c6b82ead4789062a3c4ff79b49a110d9db376e.jpg) +Figure 3. Our method's overall pipeline. (a) Given a dataset of multiview images $x_{i}$ , we simulate inconsistencies by (b) prompting a (c) video model and then (d) selecting inconsistent frames $\tilde{x}_i$ . We feed these images along with a held-out reference image $x_0$ under the original condition to a (e) multiview generative model to predict (f) a set of corresponding consistent outputs $\hat{x}_i$ . This output is supervised by the original multiview images $x_{i}$ . + +diffusion model CAT3D [15] has emerged as the state-of-the-art in view synthesis from sparse image inputs. However, as they are trained only on consistent multiview image sets, CAT3D and other sparse view synthesis techniques are not robust to the inconsistencies observed in real-world casual captures. Concurrent work CAT4D [60] extends CAT3D with a temporal axis. Among other training data, it leverages the generative augmentation strategy proposed in this paper. + +# 3. Simulating World Inconsistencies with Video Models + +Training a robust view synthesis model is challenging due to the lack of paired training data of inconsistent captures and target consistent images. Most existing multiview datasets only contain captures of consistent scenes, so simply scaling such data is not sufficient for robust view synthesis. Gathering images from multiple viewpoints, each under multiple scene deformations or lighting settings, would be extremely onerous. Heuristic data augmentation strategies such as random transformations, tints, and sparse flow fields cannot adequately capture the diversity and 3D nature of scene motions and lighting changes, as displayed in Fig. 2. Conversely, synthetic datasets like Objaverse [9] only contain simple object-level motion and fail to enable generalization to real-world scenes. + +The key idea in our work is to leverage generative video models to create a robust view synthesis dataset from existing consistent multiview image datasets. For each 3D scene, we desire a dataset that contains (1) a set of consistent multiview images $x_{i}$ , (2) inconsistent images $\tilde{x}_i$ where the scene has undergone some transformation such as a deformation or lighting change, and (3) camera poses $\pi_i$ for each image. + +# 3.1. Video model augmentation + +We propose to generate a realistic and diverse dataset of inconsistent conditioning images by simulating dynamic motion and lighting inconsistencies with pretrained image-to-video generative models. Starting with a multiview capture (taken from existing large-scale multiview image datasets), we first generate, for each view, a video from a static camera with simulated scene changes (motion or lighting). By sampling frames from these videos, we can obtain inconsistent observations for each captured viewpoint. Other image editing approaches such as InstructPix2Pix [6] could potentially be used to perform this inconsistency transformation, but these methods often fail to produce substantial variation in the layout of the image which are needed to simulate dynamic inconsistency. + +To generate videos with simulated scene changes, we use an image- and text-conditioned video diffusion model that samples from $p(V|I,c)$ , where $V$ is a video, $I$ is a conditioning frame that $V$ should include, and $c$ is a text caption. By setting $I$ to an image from a multiview capture $x_{i}$ and choosing $c$ , we may simulate inconsistencies on top of the image. Note that the video must not contain camera motion in order to preserve the accuracy of existing camera parameters. + +We simulate the two most prominent inconsistencies: dynamic motion and lighting changes. For dynamics, we use the Mannequin Challenge dataset [23]. This dataset is a natural choice as it includes static multiview captures of scenes with content that would typically be dynamic in casual captures. For our lighting-robust model, we simulate lighting changes on the RealEstate10k [70] dataset, which contains scenes in diverse indoor and outdoor illumination conditions. + +We generate the captions $c$ with a multimodal large lan + +![](images/ca33e1d14cc92bfd7e4b0e4f22cbfded1c2ba769c58575c28f401f8d09c5aff5.jpg) +Figure 4. Samples from our multiview diffusion harmonization model, visualized for scene dynamics. Given the reference image and inconsistent input image, our model directly generates multiview images consistent with the state of the reference. + +guage model, Gemini [51]. For each clip in the dataset, we randomly choose a representative frame $x_{i}$ . We prompt Gemini with this frame and a meta-prompt $m$ , designed to elicit simple but specific prompts, e.g., "the woman swings the pillow" or "the two children dance." We also ensure the generated prompts are sufficiently specific and concise through $m$ (see the supplement for the meta-prompt in its entirety). We generate the complete inconsistency prompt as: + +$$ +c = \text {" s t a t i c s h o t , "} \oplus \operatorname {G e m i n i} (x _ {i}, m), \tag {1} +$$ + +where $\oplus$ denotes string concatenation. Since the inconsistencies observed in casual captures of dynamic content are highly correlated across views, we use the same inconsistency prompt for all frames in the corresponding clip. + +We find that incorporating a negative prompt [2] $c_{\mathrm{negative}}$ is extremely important for generating the desired inconsistencies without changing the camera viewpoint. We include phrases such as "panning view" and "orbit shot" in $c_{\mathrm{negative}}$ . + +Given a multiview image $x_{i}$ and generated inconsistency prompt $c$ , we sample a video: + +$$ +V = \operatorname {V i d e o M o l e t} \left(x _ {i}, c, c _ {\text {n e g a t i v e}}\right). \tag {2} +$$ + +We assume all frames in the sampled video are inconsistent with respect to the original image $x_{i}$ . Therefore, we get a set of $T$ inconsistent frames corresponding to $x_{i}$ for each sample from the video model, where $T$ is the number of frames in the output video. At training time, we randomly sample one of the $T$ frames as inconsistent conditioning for a given $x_{i}$ . In our experiments, we generate 640 total "inconsistent" frames per multiview capture, giving us a total of about 6 million frames for the dynamics dataset and about 12 million frames for the lighting dataset. Figs. 2 and 3 visualize example frames from our synthesized videos. + +# 3.2. Video model details + +For all experiments in this paper, we use Lumiere [3], a pixel-space video diffusion model which operates in two-stages for high-resolution generation. We find that the Lumiere model struggles to generate lighting changes when given non-generic prompts, so for our lighting-robust model, we + +![](images/2dd66fadcfe473968b25037eefd65a23f90431e7075e80d376abcdc32f54deea.jpg) +Figure 5. Samples from our multiview diffusion harmonization model, visualized for lighting. Given the reference image and inconsistent input image, our model directly generates multiview images consistent with the state of the reference. + +sample $c$ uniformly from a set of predetermined lighting prompts found to generate large lighting variations instead of using a large language model. Please refer to Fig. 3 and the supplement for example prompts. + +# 4. Harmonization through multiview diffusion + +We use our multiview simulated world inconsistencies dataset $(x, \tilde{x}, \pi)$ to learn a generative model that can map from sparse inconsistent captures to a consistent set of images, as displayed in Figs. 4 and 5. We call this model a "harmonization" model as it brings the inconsistent input images into harmony. + +# 4.1. Architecture + +We build our harmonization model on top of CAT3D [15], a latent multiview diffusion model that directly predicts target images conditioned on posed input images and target camera poses. To incorporate inconsistent observed images as conditioning, we simply concatenate latents of the inconsistent images $\tilde{z}_i = \mathcal{E}(\tilde{x}_i)$ , encoded by the VAE encoder $\mathcal{E}$ , to the target raymaps and noisy latents. Additionally, we concatenate a binary image mask (either all ones or all zeros) to each input to denote the reference image, i.e., the "desired state" with which all other outputs should be consistent. + +# 4.2. Training + +Our goal is to learn a generative model that produces consistent output image sets with $N$ images, given a reference image latent $z_0$ signifying the desired scene state and $n \leq N$ observed inconsistent image latents $\tilde{z}_i$ : + +$$ +p \left(z _ {1: N} \mid z _ {0}, \tilde {z} _ {1: n}, \pi_ {0: N}\right). \tag {3} +$$ + +Given a reference conditioning latent $\mathcal{E}(x_0) = z_0$ and up to 7 inconsistent posed latents $\tilde{z}_i$ , our model predicts latents $z_{1:7}$ corresponding to the ground-truth consistent image latents. We finetune our model parameters from CAT3D [15], with additional parameters in the first layer to account for the + +![](images/5326dca58f19a1b35f7f6618c331dc1372a1f201e437b7c504f60ee95fb54b73.jpg) +Figure 6. Given the reference and inputs in Fig. 4, we show the outputs of our model versus the CAT3D baselines. We display both the diffusion samples and learned NeRF representations with the depth maps inset. Note the 3D consistency of our samples in comparison to the changing articulation of the scene displayed by CAT3D taking all images as input. The single-image conditional CAT3D has no notion of scene scale, cannot use multiview cues to reason about the static parts, and must hallucinate all scene content outside of the input image's frustum. + +additional conditioning channels. We train the model with the same diffusion loss and weighting as [15]: + +$$ +\mathbb {E} _ {t, \epsilon , z _ {0: 7}, \tilde {z} _ {1: 7}} \left[ w (t) \| f \left(\alpha_ {t} z _ {1: 7} + \sigma_ {t} \epsilon ; z _ {0}, \tilde {z} _ {1: 7}\right) - z _ {1: 7} \| ^ {2} \right], \tag {4} +$$ + +where $f$ is our multiview diffusion "harmonization" model. Given noisy versions of the target consistent latents $z_{1:7}$ , the target conditioning latent $z_0$ , and the inconsistent inputs $\tilde{z}$ , we aim to produce a denoised estimate output by $f$ that is as close as possible to the consistent latents $z_{1:7}$ . We additionally uniformly drop out the number of conditioning frames $\tilde{z}$ to allow the model to handle between 1 and 8 input images at test time. + +# 4.3. 3D reconstruction + +Having trained the harmonization model, we can sample consistent latents $\hat{z}_{1:7}$ and decode them into images $\hat{x}_{1:7}$ with the VAE decoder (visualized in Figs. 4, 5, and 6). We then have a total of 8 consistent images: the initial observed target $x_0$ and model outputs $\hat{x}_{1:7}$ . While 3D reconstruction from such a small image collection is infeasible, we can use multiview diffusion models trained on consistent images such as CAT3D [15] to "densify" the sparse consistent capture into a dense consistent capture with enough views to train a NeRF. Instead of directly sampling the original 3-image conditional CAT3D model, we finetune it to condition on 5-frames, finding that the additional context improves the original 3-image conditional model in our setting. + +# 5. Experiments + +We evaluate our method for the two most common sources of inconsistency during casual multiview capture: scene dynamics and lighting changes. For all generative methods, we + +
MethodPSNR↑SSIM↑LPIPS↓
CAT3D (single image)14.610.3820.473
CAT3D (all images)15.590.4480.462
Our Model16.730.4630.413
+ +Table 1. View synthesis results on the DyCheck dataset [14] comparing our model to CAT3D taking one or all conditioning images. Our model outperforms CAT3D by all metrics. + +sample the models exactly once. + +# 5.1. Scene Dynamics + +Dataset For scene dynamics, we evaluate our method on DyCheck [14], a dataset of 7 multiview videos where the assumed input is a monocular video with significant scene and camera motion. In this setting, we select 7 sparse frames uniformly in time as a consistent conditioning set, and uniformly select 4 target time images (top left of Fig. 4) per scene for which to compute metrics. Note that prior works which handle scene dynamics assume an ordered dense capture [22, 30, 56]. + +Baselines Considering this task as view synthesis from 8 inputs, we compare our performance to the state-of-the-art method for sparse view synthesis, CAT3D [15]. We evaluate variants of CAT3D which take all of the images as input CAT3D (all images), and only one of the images as input CAT3D (single image). For CAT3D (all images), we find that a finetuned model which conditions on 5 images and predicts 3 instead of conditioning on 3 and predicting 5 works slightly better for this setting. When sampling target views, we always include the reference image in the conditioning + +![](images/308fb048a9ad5e1e206107b864ee15c686f581559a4863472e186b8010efdf3a.jpg) + +![](images/c156c51de2018c124f802c6fcba5b9e96e2892d0cd7dd9f710e637a8f0a2fd51.jpg) + +![](images/ee11664f9cfefc7327a5496da2c2ce03838664f79350d2b0d41d38b671814025.jpg) +Ground Truth + +![](images/3f671abd75e67ba7e458e214ae80706e66975dc1d4510161e2c95fbdf5812180.jpg) + +![](images/9c7160dc5eb6b742d75a672abb151afecd1231383ea07322a1ddb3ebf7946539.jpg) + +![](images/9a8bc8e3d65f225bcedf6abc24ff1fcaa1a4a72d806ddd271d3f37fcbbfa6802.jpg) +Our Model + +![](images/262455509fc41e2ef6228a91feb20e62c269014b28632b601a0a79a3718a1b96.jpg) + +![](images/68f7d891d4b36e213c9371d1528ff075e5a281a782d1d39cf1338905cc96e38f.jpg) + +![](images/e2455cc4be2722536f8545c994e0f92566bbf951baf8c4d39bf633ca574c41d8.jpg) +CAT3D (all images) + +![](images/cc47518814103400f62035b49362c46731ddf9024b3dbcb15daed1ea0e330076.jpg) + +![](images/4b4582d494ade65bc4bddb878b312223b349fb06de4b1e64a897662159b1f2fa.jpg) + +![](images/a4d099c3b58aa552376cfbec85a97a111e9220212cd82ff1d997d667b15bcb3b.jpg) +CAT3D (single image) +Figure 7. Qualitative results for the DyCheck [14] dataset for our model, two CAT3D baselines, and two of our ablations. The depth maps are inset on the bottom left. Images are cropped for visualization. Compared to CAT3D (all images), our method generates coherent 3D scenes despite the scene motion, while leveraging the information from multiple input views unlike CAT3D (single image). In comparison to the ablations, the quality of our approach is superior. + +![](images/d05c902f5d7bc69f4f0db7bc364c01a4b996c731b57ceb229d6619c416bb6f00.jpg) + +![](images/1a4e160c26adce9ee0a1283da157e7e413d23f1ea87e23a7e80415e92a4d50d6.jpg) + +![](images/712ffba505414b08222d964494e561ae4feae0c1a62633e5d173c57130d1fa49.jpg) +Animated Objaverse + +![](images/1c56144c269a74e8e9f9add17057c3346d9b2c49aafe8c34dc73cf012ce53a9f.jpg) + +![](images/8cba033ae21e01505fbc1872949cc1cca31cf0e4342be636cc3a7a99038dc8b3.jpg) + +![](images/192628976c734998db1361ee363683588f5915e60ce570e93a5615d603e2406a.jpg) +Heuristic Augmentation + +set, along with the 4 closest of the 7 views to the current target camera set. CAT3D (single image) simply receives only the ground truth reference image. + +Due to noisy camera poses and the underdetermined nature of our task, we recompute the poses per method using COLMAP on the samples and train a Zip-NeRF [4] to evaluate novel view synthesis quality. In 4 of the 28 timesteps, COLMAP was unable to register the test images for at least one of the baselines; we discard those scenes from the calculation. Note that COLMAP never fails to register the test images for our method's results. See the supplement for a comparison to Shape of Motion [56]. + +Results The quantitative results shown in Tab. 1 demonstrate that we significantly outperform CAT3D across all metrics. Qualitatively, we can see in Figs. 1, 6 and 7 that CAT3D simply cannot handle inconsistencies. Their diffusion samples display high variance, typically changing state based on proximity to the input views. Training a 3D representation such as NeRF from these samples leads to undesirable averaging over all of the states and significant blur in inconsistent regions. + +# 5.2. Lighting Changes + +Dataset For lighting inconsistencies, we are not aware of an existing dataset of posed images that contains multiple illumination conditions and multiple "ground truth" images under a consistent lighting. Note that the widely-used Pho + +
MethodPSNR↑SSIM↑LPIPS↓
WildGaussians14.80*0.494*0.463*
CAT3D (single image)15.060.5260.552
CAT3D (all images)18.260.6250.419
Our Model20.980.7070.357
+ +Table 2. View synthesis results on our lighting dataset. We compare to WildGaussians [21] and CAT3D [15] taking all images as input, along with CAT3D taking one image as input. Our model outperforms all baselines. * indicates comparison on rectified images. + +totourism dataset [49] contains only one image under each lighting. Therefore, we collect a new dataset of 5 real-world scenes each captured under 3 separate lighting conditions. To construct this dataset, we take 3 monocular videos of a scene in 3 different lighting conditions, using approximately the same camera trajectory for each. + +Using 3 frames (one from each inconsistent video) as input, we evaluate renderings of the held-out images from one of the lighting conditions. For each collected scene, we select a target illumination condition and evaluate the method's ability to do novel-view synthesis for that illumination (from the three images). See Fig. 8 for example inputs and targets. We use Hierarchical Localization [42] with SuperGlue [43] feature matching to jointly pose all images. + +![](images/2d25798b84e580b81c5957edf371d649098f932e467acfe1bf07873a518d661f.jpg) + +![](images/f98636d296586c360021d496a5209ae3ec45e9c9cba7aa57b8dc2a746d41fb30.jpg) + +![](images/6c85d5fe13d0c29f564e3c833d840f8cd9337cfdb8478c933e99c171746e2fee.jpg) + +![](images/54df448de067fa040972cd2dec7029bce71282271e72bcd01af578b761a7e21c.jpg) + +![](images/d803f7d2d6afa831ed2715e2fa4a706f82681fcf82c134ff561f7ea5a3f940e9.jpg) + +![](images/0952394e2a4daf63bfddc85c805b22d6343162625c54ba927aaf1ab7edbbfc49.jpg) + +![](images/94e5ef91820bad4d2bac58d2c6571b8578eda100a8988f9d17beade035ddb44c.jpg) + +![](images/f4678188895fbcc8c8600ff975b7f6c6f0e12a9e81cd7d31e5b54bc5ff79351c.jpg) + +![](images/3a1b69c31c0c891d238c7b238829c5d633407ffff46e74a8dbb5e239a0c87d42.jpg) + +![](images/c656d76fd53ea2623795afcd48d43df5156cce9b2e37f46614a1bb205ea6e8ff.jpg) + +![](images/1e1b2482c02560bd64d554d83275e314284a4a91f46cfcc3c5c8fad4cc37dacb.jpg) + +![](images/2334b16a3281fd3bd7e98d20f83064d0cad13e34aed0803d8ec36008adc20821.jpg) + +![](images/74c94e58e767f3c26793f92d45a99377a71a18a75a97f03ee3bcedbce5e70a44.jpg) + +![](images/001018cb3d7d4ed1c2632dea23f10167c50360395c9a7fb474a8609ae415bbbf.jpg) + +![](images/e35364bfdf38d6843ad86a9e4ed79d2738863cbbb7139af1c95a3ec038ec6db1.jpg) + +![](images/3290f42a8a6df44f9a04fa53755d287bdbbe943095132d81de0a1f1637869ddb.jpg) + +![](images/eda6639baf3df8d013fd7719c1bf2403081658e92089b764744f9ca5dc4fde54.jpg) + +![](images/135d530e53256c00db5b4d65b39c1c5ed98854d0c933c532fd800226ec3a941c.jpg) + +![](images/d8a1893c3be9ff4c6c52434f4339f1436fbb6bcc5822328044234e41081e1593.jpg) + +![](images/5990345b321eeb20ac575705ba5a2a61e5f3f7571744bd5cccf9d04c8a75f090.jpg) + +![](images/b7930f863de389c705c4662dd5d0136f7f1b42280a477a582a4270e7593a941f.jpg) + +![](images/be4f09e7c9748b9de1e45573f050c2131451d417940e6889500f0f7077ae6222.jpg) + +![](images/e0cc082bfcfefb6bb0010df6877a0aa62195ab0a0890ec84d8da3db2fa439115.jpg) + +![](images/c8681e46ed934b5406ae9fe67a5e46f5d686f6542390bf935675fa29333e54ac.jpg) + +![](images/05a1bd5856fba001e1075a6427e21404adb9c89b21cbba745ed87b44199427c5.jpg) + +![](images/77e276861ba5fee49e9ada0a14e50e48c08843eb75983ea67ccc1b50cd8c673d.jpg) + +![](images/8732ce2a219869319ddd6d742123db6bd2b68faa9a5522e694112a2ab5360b63.jpg) + +![](images/387bb54cbed5e011496bde4663c3686c4870333723748ce76e4c522409fd4947.jpg) + +![](images/ba1f9f9915a7edb6aa52ec3d94cdd59bac5fb3d4e3962ff7efe16fe2353cd848.jpg) + +![](images/af518066bf99756a5a0c8af21dc261bb03eb7e17cb5fea4920501a51a212ffc9.jpg) + +![](images/49674cbaf5b14ddef919f0f523d2004ee856a076bd336379e2be2125067d886b.jpg) + +![](images/751914454f121ac1587e5cfb8b94f46c7960a6648259e86f41337588ddae51af.jpg) + +![](images/789a80d5de7048ef1784181dfa809a64cdd2fa2fe66a277844832c87ff42a2b0.jpg) +Input Images + +![](images/0e623a2cedff2fa95fb55894a7118ea299defbdf6aba33f0019eb9197f132f25.jpg) + +![](images/f1296131d3d0ee659b4b20e6c21e6c5f94dca5af3c452e4f6fd873b86dad21cb.jpg) +Ground Truth + +![](images/2e27a63ff3b0643a323b090c9c21397de58fae83a71c69e13e8a9d178bbe4a92.jpg) + +![](images/587fb1135df42f3a1214768c84c508b7bfe2ffb953ce17e11d281ce395b31562.jpg) +Our Model + +![](images/0205a40bb19073c81923d8438184ff82e144592ffa1ae1641e48f5a2a4b1c444.jpg) + +![](images/df214b878a3392d5a96559b1c267f6a3a18b404bc38cad68dc63975690d81c6e.jpg) +CAT3D (all images) + +![](images/e83c5a1ab6c58d53934a8f9d51796fa7dbab87d6da4817769d463fb9b99dc76b.jpg) + +![](images/3d1e16e93cab173cf0f2d4ab9975296a68f01ef3fa9349f8e2fd3219ccdc53a2.jpg) +CAT3D (single image) +Figure 8. Qualitative results for 3 scenes from our captured lighting dataset. For each scene, we display the renders from the learned NeRFs given the 3 input images on the left. We show two states for each scene, with renders outlined in blue corresponding to the upper input image, and renders outlined in orange corresponding to the bottom input image. Although all methods are rendered with the appearance embedding of the corresponding states on the left, baselines struggle to generate plausible novel views, and often generate completely degenerate geometry. The bottom rows are brightened for visualization. + +![](images/4dcde146d19bbeaaa37b81c66dcc1edb06dc6490c8168cbf406a98762e0fa4a4.jpg) + +![](images/be98fe4c8b33946515513ee2290ecdf3a9e27afce05d7a3cc7206f0a96fdb344.jpg) +Heuristic Augmentation + +![](images/f6dab07b710649d4284e1bdd7546bc3c7e00ce2b1ca9852e74bff58fa464518d.jpg) + +![](images/0cb0249592fe9291dc80a333e60ae0e1ff5a17943363d605f730eb4e6a2f4024.jpg) +WildGaussians + +Baselines Methods which handle lighting changes typically assume a large number of captured images [32] and rely on latent embeddings [5] to parameterize per-image variations in appearance. In this line, we compare to WildGausians [21]. For other baselines, we first generate a large number of novel views using CAT3D conditioned on the 3 inconsistent images, and then train a Zip-NeRF with latent appearance embeddings. At test-time, for all methods, we + +use the embedding of the reference image with the target illumination. We again evaluate against CAT3D (all images) and CAT3D (single image). Since there are only 3 input images, CAT3D (all images) is the original CAT3D model conditioned on the three input images. + +Results The quantitative results in Tab. 2 show that our method significantly outperforms WildGaussians and + +CAT3D in all metrics. Qualitatively, Fig. 8 displays our method's superior visual results. In some scenes, such as the stone bear shown in the bottom two rows, baselines fail to reconcile inconsistent input images into any coherent 3D scene. In other cases, the baselines reconstruct inaccurate "cloudy" scene geometry attempting to explain away changes in lighting. In contrast, our method reconciles highly disparate and sparse observations into a consistent 3D scene, allowing the generation of a high-fidelity NeRF with coherent geometry as demonstrated in the inset depth maps. + +# 5.3. Ablations + +In this section, we ablate key design decisions. Specifically, we demonstrate the importance of using our simulated inconsistency data by evaluating against heuristic augmentations and a synthetic data alternative. For dynamics, we compare to a warping-based heuristic augmentation where we apply sparse flow fields; an example can be seen in Fig. 2. The resultant model simply copies all "real-looking" pixels, indicating that such warping does not bridge the domain gap. + +We also compare to generating a synthetic training dataset by animating $40\mathrm{k}+$ Objverse assets [9, 26] with associated motions. Due to the small motion magnitude and the domain gap from object-level renderings to real scene-level data, the method significantly underperforms. The quantitative ablation results can be seen in the top of Tab. 3, where our method outperforms all ablated methodologies. For dynamics, a qualitative comparison is provided on the right of Fig. 7. + +For lighting, we compare against heuristic augmentation whereby the input images are tinted inconsistently as seen in Fig. 2, and the targets images are tinted consistently. This method slightly outperforms the vanilla CAT3D as it requires the model to get the mean color correct; however, it cannot resolve lighting phenomena like shadows, nor localized changes in lighting. Results can be seen quantitatively in Tab. 3 and qualitatively on the right of Fig. 8. + +
AblationPSNR↑SSIM↑LPIPS↓
DynamicHeuristic Augmentation15.520.4480.466
Animated Objaverse14.920.3800.524
Our Complete Model16.600.4620.409
LightHeuristic Augmentation18.960.6450.406
Our Complete Model20.980.7070.357
+ +Table 3. Ablations. Heuristic augmentation and synthetic datasets lead to significantly worse performance for robust view synthesis. For both inconsistencies in dynamics and lighting, our complete model vastly outperforms the baselines due to the underlying video model's ability to simulate physics. + +# 6. Discussion + +We have proposed SimVS, a method for high-quality 3D generation from casual captures even in the presence of + +![](images/51b8a8cb6fa23fe9d894a31052a9b6692b2857ce14d5acc8adfcc9102e356fe6.jpg) +Figure 9. Two representative samples showing limitations of our method. On the left, in the presence of the large dynamic motions of the block, our model may synthesize degenerate scene content. On the right, for areas observed only once such as the background on the top right, our model sometimes outputs poor quality hallucinations. + +![](images/9d753b6b0e3ac840cf4c9850afd4af5093cfa687b0b64b692138592508e638ae.jpg) + +severe illumination changes and significant scene motion. We believe this represents a step forward in simplifying the capture and creation of 3D scenes. + +**Limitations** Our method requires accurate camera poses, which can be difficult to compute for sparse captures using traditional techniques such as COLMAP. However, recent methods such as DUST3R [57] and the dynamics-robust follow-up MonST3R [68] have shown tremendous promise for camera pose estimation. Also, when there is significant inconsistency or little overlap between views, our method can struggle to reconcile the observations as seen in the samples in Fig. 9. + +Conclusion Our work demonstrates the power of using video models to generate data for challenging tasks where collection is expensive and challenging. We believe the approach proposed here will scale well with the ever-improving quality of video models. Moreover, our method is not specific to a particular architecture or task: our method may be applied to make DUSt3R [57]-style models more robust to inconsistency and our harmonization network could be implemented with a camera-controlled video model to directly synthesize multiview-consistent videos in one sampling pass. + +# Acknowledgements + +We would like to thank Paul-Edouard Sarlin, Jiamu Sun, Songyou Peng, Richard Tucker, Linyi Jin, Rick Szeliski and Stan Szymanowicz for insightful conversations and help. We also extend our gratitude to Shlomi Fruchter, Kevin Murphy, Mohammad Babaeizadeh, Han Zhang and Amir Hertz for training the base text-to-image latent diffusion model. This work was supported in part by an NSF Fellowship, ONR grant N00014-23-1-2526, gifts from Google, Adobe, Qualcomm and Rembrand, the Ronald L. Graham Chair, and the UC San Diego Center for Visual Computing. + +# References + +[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54(10): 105-112, 2011. 2 +[2] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Minhao Cheng, Boqing Gong, and Cho-Jui Hsieh. Understanding the impact of negative prompts: When and how do they take effect? arXiv preprint arXiv:2406.02965, 2024. 4 +[3] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, Yuanzhen Li, Michael Rubinstein, Tomer Michaeli, Oliver Wang, Deqing Sun, Tali Dekel, and Inbar Mosseri. Lumiere: A Space-Time Diffusion Model for Video Generation. arXiv:2401.12945, 2024. 4 +[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields. ICCV, 2023. 6 +[5] Piotr Bojanowski, Armand Joulin, David Lopez-Pas, and Arthur Szlam. Optimizing the Latent Space of Generative Networks. ICML, 2018. 7 +[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3 +[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2 +[8] Hiba Dahmani, Moussab Bennehar, Nathan Piasco, Luis Roldao, and Dzmitry Tsishkou. Swag: Splitting in the wild images with appearance-conditioned gaussians. In European Conference on Computer Vision, pages 325-340. Springer, 2025. 2 +[9] 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. CVPR, 2022. 3, 8 +[10] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2 +[11] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2, 2024. 2 +[12] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2 +[13] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2025. 2 + +[14] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular Dynamic View Synthesis: A Reality Check. NeurIPS, 2022. 5, 6 +[15] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: Create Anything in 3D with Multi-View Diffusion Models. NeurIPS, 2024. 1, 2, 3, 4, 5, 6 +[16] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2 +[17] Haian Jin, Yuan Li, Fujun Luan, Yuanbo Xiangli, Sai Bi, Kai Zhang, Zexiang Xu, Jin Sun, and Noah Snavely. Neural gaffer: Relighting any object via diffusion. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2 +[18] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 2 +[19] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 2 +[20] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 2 +[21] Jonas Kulhanek, Songyou Peng, Zuzana Kukelova, Marc Pollefeys, and Torsten Sattler. WildGaussians: 3D Gaussian Splatting in the Wild. NeurIPS, 2024. 2, 6, 7 +[22] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. MoSca: Dynamic Gaussian Fusion from Casual Videos via 4D Motion Scaffolds. arXiv:2405.17421, 2024. 2, 5 +[23] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T. Freeman. Learning the Depths of Moving People by Watching Frozen People. CVPR, 2019. 3 +[24] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2 +[25] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2 +[26] Hanwen Liang, Yuyang Yin, Dejia Xu, Hanxue Liang, Zhangyang Wang, Konstantinos N Plataniotis, Yao Zhao, and Yunchao Wei. Diffusion4d: Fast spatial-temporal consistent 4d generation via video diffusion models. arXiv preprint arXiv:2405.16645, 2024. 8 +[27] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single + +image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 2 +[28] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 2 +[29] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2 +[30] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust Dynamic Radiance Fields. CVPR, 2023. 2, 5 +[31] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3D: Single Image to 3D using Cross-Domain Diffusion. CVPR, 2024. 2 +[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. CVPR, 2021. 2, 7 +[33] Moustafa Meshry, Dan B Goldman, Sameh Khamis, Hugues Hoppe, Rohit Pandey, Noah Snavely, and Ricardo Martin-Brualla. Neural rerendering in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6878-6887, 2019. 2 +[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. ECCV, 2020. 2 +[35] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. ICCV, 2021. 2 +[36] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2 +[37] Yohan Poirier-Ginter, Alban Gauthier, Julien Phillip, J-F Lalongne, and George Drettakis. A diffusion approach to radiance field relighting using multi-illumination synthesis. In Computer Graphics Forum, page e15147. Wiley Online Library, 2024. 2 +[38] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2 +[39] Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8931-8940, 2024. 2 +[40] Sara Sabour, Suhani Vora, Daniel Duckworth, Ivan Krasin, David J Fleet, and Andrea Tagliasacchi. Robustnerf: Ignoring distractors with robust losses. In Proceedings of the + +IEEE/CVF conference on computer vision and pattern recognition, pages 20626-20636, 2023. +[41] Sara Sabour, Lily Goli, George Kopanas, Mark Matthews, Dmitry Lagun, Leonidas Guibas, Alec Jacobson, David J Fleet, and Andrea Tagliasacchi. Spotlessplats: Ignoring distractors in 3d gaussian splatting. arXiv preprint arXiv:2406.20055, 2024. 2 +[42] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In CVPR, 2019. 2, 6 +[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 6 +[44] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2 +[45] Abhishek Sharma, Adams Yu, Ali Razavi, Andeep Toor, Andrew Pierson, Ankush Gupta, Austin Waters, Aaron van den Oord, Daniel Tanis, Dumitru Erhan, Eric Lau, Eleni Shaw, Gabe Barth-Maron, Greg Shaw, Han Zhang, Henna Nandwani, Hernan Moraldo, Hyunjik Kim, Irina Blok, Jakob Bauer, Jeff Donahue, Junyoung Chung, Kory Mathewson, Kurtis David, Lasse Espeholt, Marc van Zee, Matt McGill, Medhini Narasimhan, Miaosen Wang, Mikolaj Binkowski, Mohammad Babaeizadeh, Mohammad Taghi Saffar, Nando de Freitas, Nick Pezzotti, Pieter-Jan Kindermans, Poorva Rane, Rachel Hornung, Robert Riachi, Ruben Villegas, Rui Qian, Sander Dieleman, Serena Zhang, Serkan Cabi, Shixin Luo, Shlomi Fruchter, Signe Nørly, Srivatsan Srinivasan, Tobias Pfaff, Tom Hume, Vikas Verma, Weizhe Hua, William Zhu, Xinchen Yan, Xinyu Wang, Yelin Kim, Yuqing Du, and Yutian Chen. Veo. 2024. 2 +[46] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view Diffusion for 3D Generation. arXiv:2308.16512, 2023. 2 +[47] Meng-Li Shih, Wei-Chiu Ma, Lorenzo Boyice, Aleksander Holynski, Forrester Cole, Brian Curless, and Janne Kontkanen. Extranerf: Visibility-aware view extrapolation of neural radiance fields with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20385-20395, 2024. 2 +[48] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 2 +[49] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3D. SIGGRAPH, 2006. 6 +[50] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. arXiv preprint arXiv:2406.18717, 2024. 2 +[51] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv:2312.11805, 2023.4 + +[52] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2 +[53] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2 +[54] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 2 +[55] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2 +[56] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of Motion: 4D Reconstruction from a Single Video. arXiv:2407.13764, 2024. 2, 5, 6 +[57] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D Vision Made Easy. CVPR, 2024. 8 +[58] Yuze Wang, Junyi Wang, and Yue Qi. We-gs: An in-the-wild efficient 3d gaussian representation for unconstrained photo collections. arXiv preprint arXiv:2406.02407, 2024. 2 +[59] Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfiller: Completing scenes via generative 3d inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20731-20741, 2024. 2 +[60] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 3 +[61] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2 +[62] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 2 +[63] Jiacong Xu, Yiqun Mei, and Vishal M Patel. Wild-gs: Real-time novel view synthesis from unconstrained photo collections. arXiv preprint arXiv:2406.10373, 2024. 2 +[64] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 2 +[65] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 2 + +[66] Chong Zeng, Yue Dong, Pieter Peers, Youkang Kong, Hongzhi Wu, and Xin Tong. Dilightnet: Fine-grained lighting control for diffusion-based image generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2 +[67] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 2 +[68] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 8 +[69] Xiaoming Zhao, Pratul Srinivasan, Dor Verbin, Keunhong Park, Ricardo Martin Brualla, and Philipp Henzler. Illuminerf: 3d relighting without inverse rendering. Advances in Neural Information Processing Systems, 37:42593-42617, 2024. 2 +[70] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo Magnification: Learning View Synthesis using Multiplane Images. SIGGRAPH, 2018. 3 +[71] Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yicong Hong, Li Fuxin, and Zexiang Xu. Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. arXiv preprint arXiv:2410.12781, 2024. 2 \ No newline at end of file diff --git a/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/images.zip b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ac27ed665c01861eb1271f94ebdd2ea7058bcbdb --- /dev/null +++ b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c40a651c14fd8b6a0044ed927d7421f7a9df6263b3a344a29dcce74dc81c65c5 +size 1105974 diff --git a/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/layout.json b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a66a9d1452c4f86cb8cd02dab0a1e12a98dc788a --- /dev/null +++ b/CVPR/2025/SimVS_ Simulating World Inconsistencies for Robust View Synthesis/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b348c3bd695d346094262f708b103a75578c05dac014e7502d589c907938439 +size 513644 diff --git a/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/c334a2a6-9747-41af-b8b6-5b13b27ea619_content_list.json b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/c334a2a6-9747-41af-b8b6-5b13b27ea619_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6816e9bf78927709988f81da04b605a631cc959c --- /dev/null +++ b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/c334a2a6-9747-41af-b8b6-5b13b27ea619_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9bc55c8dfeab6c0d0ccbe08de17263287bab32169023b783b538f7dd5eba311c +size 83913 diff --git a/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/c334a2a6-9747-41af-b8b6-5b13b27ea619_model.json b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/c334a2a6-9747-41af-b8b6-5b13b27ea619_model.json new file mode 100644 index 0000000000000000000000000000000000000000..831d5762d44c439520b4649c021401bf49e1a2c7 --- /dev/null +++ b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/c334a2a6-9747-41af-b8b6-5b13b27ea619_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a39ac0de7ba8fd6cf8f0fc633df977c81a7421d20a3c4104940b58b448b4b29 +size 106041 diff --git a/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/c334a2a6-9747-41af-b8b6-5b13b27ea619_origin.pdf b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/c334a2a6-9747-41af-b8b6-5b13b27ea619_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..63db156cb5376b419b33a535b2b4e643cf075817 --- /dev/null +++ b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/c334a2a6-9747-41af-b8b6-5b13b27ea619_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba92c1aec71563e01ff58f4c55276a7278f3b4ee74ae7e99577dd85d21bd1336 +size 1655922 diff --git a/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/full.md b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e02d6e9207c5a6e806683fde9854d9e4fe98fe67 --- /dev/null +++ b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/full.md @@ -0,0 +1,290 @@ +# Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking + +Chaocan Xue $^{1}$ , Bineng Zhong $^{1*}$ , Qihua Liang $^{1\dagger}$ , Yaozong Zheng $^{1}$ , Ning Li $^{1}$ , Yuanliang Xue $^{2}$ , Shuxiang Song $^{1}$ + +1Key Laboratory of Education Blockchain and Intelligent Technology, Ministry of Education, Guangxi Normal University, Guilin 541004, China + +$^{2}$ Xi'an Research Institute of High Technology, Xi'an 710025, China + +xcc23cg@163.com, bnzhong@gxnu.edu.cn, qhliang@gxnu.edu.cn + +yaozongzheng@stu.gxnu.edu.cn, ningli65536@mailbox.gxnu.edu.cn + +xyl507@outlook.com, songshuxiang@mailbox.gxnu.edu.cn + +# Abstract + +Vision transformers (ViTs) have emerged as a popular backbone for visual tracking. However, complete ViT architectures are too cumbersome to deploy for unmanned aerial vehicle (UAV) tracking which extremely emphasizes efficiency. In this study, we discover that many layers within lightweight ViT-based trackers tend to learn relatively redundant and repetitive target representations. Based on this observation, we propose a similarity-guided layer adaptation approach to optimize the structure of ViTs. Our approach dynamically disables a large number of representation-similar layers and selectively retains only a single optimal layer among them, aiming to achieve a better accuracy-speed trade-off. By incorporating this approach into existing ViTs, we tailor previously complete ViT architectures into an efficient similarity-guided layer-adaptive framework, namely SGLATrack, for real-time UAV tracking. Extensive experiments on six tracking benchmarks verify the effectiveness of the proposed approach, and show that our SGLATrack achieves a state-of-the-art real-time speed while maintaining competitive tracking precision. Codes and models are available at https://github.com/GXNU-ZhongLab/SGLATrack. + +# 1. Introduction + +Unmanned Aerial Vehicle (UAV) tracking is an essential branch of visual object tracking (VOT), which can be used for various applications such as visual surveillance [40], path planning [42] and border security [60]. However, compared to generic object tracking, UAV tracking typically faces greater challenges due to its unique aerial + +![](images/2f5ef049a65dc0be41bddfb3327b6f31c0f648ec91bc9fbf447405418d6a9557.jpg) +Figure 1. Comparison of SGLATrack against other UAV trackers on UAV123. Our SGLATrack demonstrates state-of-the-art performance with an AUC score of $66.9\%$ , while running efficiently at nearly 225 FPS on GPU and approximately 75 FPS on CPU. + +perspective. More importantly, the limited power and computing resources of UAVs impose strict demands on efficiency. Therefore, an excellent UAV tracker must maintain both high tracking precision and efficiency. + +Owing to the high efficiency derived from the operations in the Fourier domain, discriminative correlation filter (DCF)-based methods are widely used for UAV Tracking. Despite the efficiency, they hardly achieve favorable tracking robustness using hand-crafted feature extractors [13, 16, 27]. Siamese network (SN)-based trackers can learn robust target representations using deep features, but their reliance on early convolutional neural networks (CNNs) leaves considerable potential for precision improvement [2, 34, 69]. In recent years, vision trans + +former (ViT)-based trackers have become increasingly popular [8, 12, 61, 63]. Notable examples are Mixformer [12] and OSTrack [63], which drive the shift of tracking models towards highly parallel one-stream architectures. Nonetheless, they are still too cumbersome to be directly deployed for real-time UAV tracking. To improve the efficiency, AbaViTrack [37] introduces an adaptive and background-aware token halting approach into tiny ViTs. Also based on tiny ViTs, AVTrack [39] proposes a more structured approach that dynamically enables or disables ViT layers according to input complexity, further improving efficiency by avoiding extra unstructured access operations. However, such an approach requires attaching classifiers to most ViT layers to decide whether each layer should be activated, which is time-consuming and redundant. Moreover, although their works demonstrate the potential of tiny ViTs in UAV tracking tasks, a surprising finding is that the layer redundancy in lightweight ViT-based trackers remains unexplored. + +In this paper, we make the first attempt to study the layer redundancy within lightweight ViT-based trackers, with the goal of providing more advanced guidelines for accelerating ViT-based UAV tracking. Our intuition is based on the saturation observation [24, 51], which indicates the hidden states of transformers reach saturation as they progress into deeper layers. For visual tracking tasks, we demonstrate that a similar saturation conclusion is also applicable. Specifically, we provide a set of empirical experiments to study the layer redundancy of tiny ViTs, including layer-by-layer cosine similarity analysis and a check of the performance impact when stopping inference at a specific layer. These studies show that the encoded search features undergo significant changes in shallow layers, but will reach a saturation state at a certain layer. Once such a saturated layer is reached, the features exhibit relatively small changes when passing subsequent layers, and the impact of these subsequent layers on the final predicted results also decreases. Therefore, some subsequent layers can be disabled without causing significant performance degradation. To this end, we propose a similarity-guided layer adaptation approach, which disables a large number of these layers and selectively retains only a single representative layer among them, aiming to achieve a better accuracy-speed trade-off. Our approach is implemented with a selection module, which is a simple multi-layer perceptron (MLP). It takes partial features at the saturated layer as input, and outputs the probabilities of each subsequent layer being selected. Considering that each layer learns different representations and that different tracking scenarios require different representations, we introduce a layer-wise similarity loss to optimize the selection of this module, allowing the model to dynamically maximize its focus on the target in situations where a large number of layers are disabled. By incorporating this approach into existing ViTs, we tailor + +a family of efficient trackers, namely SGLATrack, to meet the real-time demands of UAV Tracking. Extensive experiments on six popular datasets demonstrate that SGLATrack achieves a better accuracy-speed trade-off. Notably, as illustrated in Fig. 1, SGLATrack achieves remarkable performance of $66.9\%$ AUC while maintaining an impressive tracking speed of 225 frames per second (FPS) on UAV123. Our main contribution can be summarized as follows: + +- We conduct a thorough analysis of layer redundancy in lightweight ViT-based trackers by examining layer-by-layer feature changes and result variations. Our analysis is meticulous and provides valuable insights. +- We propose a similarity-guided layer adaptation approach to optimize redundant ViT layers. This approach can be seamlessly integrated into existing ViTs, and achieve a better accuracy-speed trade-off. +- We develop SGLATrack, a family of efficient trackers based on different optimized ViTs for real-time UAV tracking. SGLATrack achieves state-of-the-art performance on six challenging benchmarks, while setting a new record of tracking speed. + +# 2. Related Works + +# 2.1. Visual Object Tracking + +Mainstream VOT algorithms can be roughly divided into three types: DCF-based, SN-based, and ViT-based methods. DCF-based trackers are widely used in the field of UAV tracking for their high efficiency, but they hardly maintain robustness under challenging conditions due to poor representation ability of hand-crafted features [3, 14-16, 27, 32, 40]. SN-based trackers can learn robust target representations using deep features, but most of them still rely on early CNNs, leaving precision improvement far away from being fully explored [2, 11, 34, 59, 67, 69]. Recently, ViTs have been successfully introduced into visual tracking [28-30, 50, 56, 57]. For example, Mixformer [12] and OSTrack [63] integrate feature extraction and interaction into a same encoder, allowing a more full fusion of search features with template features and establishing a highly parallelized one-stream architecture. Although these methods have significantly improved tracking accuracy, they are still too cumbersome for UAV tracking which highly emphasizes efficiency. To boost the efficiency, TATrack [38] propose a novel mutual information maximization-based knowledge distillation approach. DDCTrack [19] proposes a dynamic token sampler to adjust the number of tokens in each stage of the transformer. AbaViTrack [37] proposes an adaptive and background-aware token halting approach that learns to prune uninformative background tokens. Nonetheless, the variable number of tokens introduces significant time costs because of additional unstructured assess operations. To overcome this limitation, + +AVTrack [39] proposes activation modules that dynamically activate ViT layers according to input complexity. It evaluates the complexity of input samples and dynamically skips ViT layers based on the output probability from internal modules. However, the complexity of the input is difficult to estimate, and attaching an internal classifier to each layer is redundant. If an input sample is already deemed an easy case in shallow layers, the classifiers in deeper layers do not need to make another judgment. In this work, we conduct a thorough study of the redundancy in ViT layers and implement a more precise layer deactivation for ViTs. + +# 2.2. Efficient Vision Transformers + +ViTs have significantly impacted the field of computer vision with their outstanding performance [7, 17, 53]. Despite their widely recognized modeling capabilities, they have faced challenges related to speed limitations, especially on resource-constrained edge devices. Therefore, many works have been proposed to accelerate ViTs, primarily through low-rank approximation [55], fixed pruning [52, 66], and hybrid designs [10, 41]. However, ViTs with low-rank approximation and quantization methods often achieve efficiency gains at the cost of significant accuracy loss. Fixed pruning-based ViTs typically involve a time-consuming fine-tuning process, while hybrid ViTs struggle to flexibly adapt to different input sizes owing to their CNN-based stems. Recently, methods based on conditional computation have gained significant attention due to their ability to adaptively adjust computational resources based on input complexity. They can be mainly divided into two types: those focusing on adaptive token pruning [48, 65], and those focusing on adaptive layer activation [1, 39, 58]. The first type, such as DynamicViT [48] and AViT [65], designs certain rules or modules to adaptively prune tokens at different depths. Due to the additional time cost introduced by their unstructured access operations, the second type adaptively activate ViT layers according to input complexity. For instance, ViT-EE [1] proposes to add internal classifiers behind each ViT layer, and use the output of these classifiers which measures how confident the model is in the current predictions to determine whether the inference can be terminated early. Similarly, AVTrack [39] attaches lightweight modules before each ViT layer to decide whether this layer needs to be executed. Despite improved inference speeds, recent works [22, 51, 58] show that the input complexity is difficult to estimate and the output of the classifiers is not always reliable. The improvement in output reliability generally comes with a larger classifier, making it challenging to achieve a better accuracy-speed trade-off. In this work, we evaluate the selection of a subsequent layer only when the features are saturated, thereby avoiding excessive classifiers and assessments of input complexity. + +# 3. Method + +In this section, we first revisit the typical ViT-based one-stream tracking framework and present a detailed analysis of layer redundancy within this framework. Then, we illustrate how our similarity-guided layer adaptation optimizes redundant layers. Finally, we introduce the proposed family of efficient trackers, i.e., SGLATrack. + +# 3.1. Layer Redundancy + +To facilitate understanding of the redundancy within ViT layers based on the saturation observation, we first offer a review of ViT-based tracking frameworks. Given a template image $\mathbf{Z} \in \mathbb{R}^{3 \times H_z \times W_z}$ and a search image $S \in \mathbb{R}^{3 \times H_s \times W_s}$ , they are first split and tokenized into a template token sequence $X_z \in \mathbb{R}^{N_z \times D}$ and a search token sequence $X_s \in \mathbb{R}^{N_s \times D}$ via a patch embedding layer $\mathcal{E}$ . This process then results in a token sequence $X$ consisting of $X_z$ and $X_s$ , formulated as: + +$$ +\boldsymbol {X} = \left[ \boldsymbol {X} _ {z}, \boldsymbol {X} _ {s} \right] = \mathcal {E} (\boldsymbol {Z}, \boldsymbol {S}) \in \mathbb {R} ^ {N \times D}, \tag {1} +$$ + +where $D$ denotes the embedding dimension, $[\cdot]$ represents the concatenation operation, $N$ is the total number of tokens and $N = N_z + N_s$ . Let $\mathcal{T}^i$ denote the $i$ -th ViT layer, the sequence are transformed via $\pmb{X}^{i} = \mathcal{T}^{i}(\pmb{X}^{i - 1})$ . Then, the operation process of standard ViTs with $l$ layers can be written as: + +$$ +\boldsymbol {X} ^ {l} = \mathcal {T} ^ {l} \circ \mathcal {T} ^ {l - 1} \dots \circ \mathcal {T} ^ {1} \circ \mathcal {E} (\boldsymbol {Z}, \boldsymbol {S}), \tag {2} +$$ + +where $\circ$ denotes the composition operation, and $X^l$ is the final feature after the last layer which is equivalent to $[X_z^l, X_s^l]$ . Finally, the target bounding box $B$ can be located on the search image via $B = \mathcal{H}(X_s^l)$ , where $\mathcal{H}$ is a prediction head. + +As can be seen, the final result depends on the search features $X_{s}^{l}$ and the prediction head $\mathcal{H}$ . Based on this observation, a natural idea is that if the search features change barely after passing through the ViT layers, their impact on the final prediction will also be slight. This idea inspire us to analyze the changes of search features at different depths. Hence, we study whether saturation of search features exists in UAV Tracking models. Specifically, we first train several lightweight ViTs for UAV Tracking tasks, including the tiny versions of distilled DeiT [53] and ViT [17] (The training settings are the same as those of OSTrack). Then, we employ cosine similarity to measure the differences between the features of each layer. Let $\mathbf{a}$ and $\mathbf{b}$ denote two random vectors, the cosine similarity between $\mathbf{a}$ and $\mathbf{b}$ is expressed as: + +$$ +\operatorname {C o s} (\mathbf {a}, \mathbf {b}) = \frac {\mathbf {a} \cdot \mathbf {b}}{\| \mathbf {a} \| \cdot \| \mathbf {b} \|}, \tag {3} +$$ + +![](images/604b80b6f06d11b03051bdb5deeb48b4527ee973294e52c602e0fc2f74ed1f34.jpg) +Figure 2. Layer-by-layer feature changes and AUC changes. Higher cosine similarity denotes fewer feature changes. + +where $\|\cdot\|$ represents the $l_{2}$ norm. Fig. 2 shows the cosine similarity $\mathrm{Cos}(X_s^i, X_s^{i-1})$ between two consecutive layers and the AUC scores obtained by $B = \mathcal{H}(X_s^i)$ on UAV123 [46]. It can be observed that the cosine similarity is higher in deep layers, indicating that the changes of search features are relatively smaller within the deeper layers. Moreover, the AUC score rapidly increases at first and then grows slowly, confirming the existence of saturation and indicating that the redundancy is more than expected. Therefore, some deep layers are redundant and can be disabled without significant precision drop, which also aligns with the consensus that shallow details are more beneficial than deep semantics for distinguishing objects in tracking tasks. + +# 3.2. Similarity-Guided Layer Adaptation + +According to the above observation on layer redundancy, we propose a similarity-guided layer adaptation strategy to accelerate ViT-based UAV tracking. Although deeper layers exhibit higher similarity, directly pruning them leads to a significant drop in precision, as they learn relatively repetitive representations rather than useless ones. To this end, we can retain a single representative layer, whose output should be similar to the output of the saturated layer for minimizing precision drop. The motivation behind this is that the model will gradually focus on the target as the encoder layers go deeper. If the saturated layer has already focused attention on the target, the subsequent layers will further enhance this attention by suppressing interference. Therefore, a higher similarity between the output of the subsequent layers and the output of the saturated layer is more conducive to focused and consistent attention. Let $X^{l^*}$ denotes the saturated features at the $l^*$ -th layer, our objective is to find the $(l^* + k)$ -th subsequent layer that maximizes the similarity between $X^{l^*}$ and $X^{l^* + k}$ , expressed as: + +$$ +\underset {k} {\arg \max } \operatorname {C o s} \left(\boldsymbol {X} ^ {l ^ {*}}, \boldsymbol {X} ^ {l ^ {*} + k}\right), \tag {4} +$$ + +where $k \leq l - l^{*}$ is the index of the subsequent layers after the $l^{*}$ -th layer. Given that a fixed retained layer struggles to fit all cases, we develop a selection module to dynamically choose the optimal layer for different samples. This module is a simple MLP denoted as $\mathcal{M}$ , it receives a portion of saturated features as input with efficiency in consideration, and outputs the probabilities of each subsequent layer being selected. Following AVTrack [39], the input is the features on the first dimension of $X^{l^{*}}$ , expressed as $\mathbf{e}_1^T\mathbf{X}^{l^*} \coloneqq \mathbf{z} \in \mathbb{R}^N$ , where $\mathbf{e}_1^T = [1,0,\dots,0] \in \mathbb{R}^N$ is a standard unit vector. And the selection module is defined as: + +$$ +\hat {\boldsymbol {y}} = \sigma (\mathcal {M} (\mathbf {z})) \in \mathbb {R} ^ {K} \tag {5} +$$ + +where $\sigma (\cdot)$ is the sigmoid function, $\hat{\pmb{y}} = [\hat{y}^1,\hat{y}^2,\dots ,\hat{y}^K ]$ represents the selected probabilities of the layers after the $l^{*}$ -th layer, and $K = l - l^{*}$ . A layer with the highest probabilities will be retained while the remaining layers will be disabled. According to the objective in Eq. 4, we introduce a layer-wise similarity loss $\mathcal{L}_{sim}$ to optimize the selection module in making the optimal choices. This loss penalizes layers with low similarity and encourages layers with high similarity, formulated by: + +$$ +\mathcal {L} _ {s i m} = \frac {1}{K} \sum_ {k = 1} ^ {K} | \hat {y} ^ {k} - y ^ {k} |, \tag {6} +$$ + +where $y^{k}$ is the expected probability defined by: + +$$ +y ^ {k} = \left\{ \begin{array}{l l} 1, & \text {i f} \operatorname {C o s} \left(\boldsymbol {X} ^ {l ^ {*}}, \boldsymbol {X} ^ {l ^ {*} + k}\right) \text {i s t h e m a x i m u m}, \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {7} +$$ + +With this loss, the selection module can dynamically allocate the optimal layers for different tracking scenarios, allowing the model to concentrate maximally on the target in situations where a large number of layers are disabled. + +# 3.3. SGLATrack for UAV Tracking + +Overall Architecture. By incorporating the proposed similarity-guided layer adaption approach into existing ViTs, we build SGLATrack, a family of efficient trackers for real-time UAV Tracking, whose overall architecture is shown in Fig. 3. On the one hand, it follows the highly parallelized one-stream tracking framework. On the other hand, it dynamically optimizes the redundant ViT layers within this framework. Specifically, SGLATrack first restructures the standard ViT in Eq. 2 into an efficient layer-adaptive architecture, formulated as $X^{l^* + k} = \mathcal{T}^{l^* + k} \circ \mathcal{T}^{l^*} \cdots \circ \mathcal{T}^1 \circ \mathcal{E}(Z, S)$ , for efficient interaction between template images and search images. Then, SGLATrack employs a typical prediction head to estimate target location, i.e., $B = \mathcal{H}(X_s^{l^* + k})$ . Note that $l^*$ can be considered as a hyperparameter to balance the trade-off between precision + +![](images/4a606d6b4dac6abb732b9a7e482b966133587c57c59f0867309a176c4ecc6404.jpg) +Figure 3. Overall architecture of the proposed SGLATrack. It is composed of a one-stream backbone, a typical prediction head, and a selection module. During training, the selection module is optimized by the proposed layer-wise similarity loss. During inference, the selection module disables redundant ViT layers and selectively retains an optimal layer among them to alleviate performance drop. + +and speed in our framework, and $k$ is decided by the selection module. We set $l^*$ to 6, and we provide detailed discussions about the determination of saturated states in the ablation study section. + +Head and Loss. Referring to the previous work, the prediction head $\mathcal{H}$ is a center-based head for predicting the centroid position and scale of the target. The search features $X_{s}^{l^{*} + k}$ are first turned into a 2D spatial feature map, which is then fed into prediction head $\mathcal{H}$ . Suppose the size of template and search patches is $P\times P$ , the outputs of $\mathcal{H}$ are a classification score map $\mathbf{p}\in [0,1]^{H_s / P\times W_s / P}$ , a normalized bounding box size $\mathbf{s}\in [0,1]^{2\times H_s / P\times W_s / P}$ , and a offset $\mathbf{o}\in [0,1]^{2\times H_s / P\times W_s / P}$ . The location with the highest classification score is taken as the target's location, and the final tracking result is calculated by combining the offset size and the bounding box size. We use the focal loss [44] for classification, a combination of $L_{1}$ loss and GIoU loss [49] for bounding box regression. Finally, the overall loss can be described as: + +$$ +\mathcal {L} = \mathcal {L} _ {c l s} + \lambda_ {i o u} \mathcal {L} _ {i o u} + \lambda_ {L 1} \mathcal {L} _ {L 1} + \gamma \mathcal {L} _ {s i m}, \tag {8} +$$ + +where $\lambda_{iou} = 2$ , $\lambda_{L1} = 5$ , and $\gamma = 0.2$ are the regularization parameters. + +# 4. Experiments + +# 4.1. Implementation Details + +Model. We employ different ViTs as backbones, including ViT-tiny [17], distilled DeiT-tiny [53], and EVAtiny [21], to build three trackers for extensive evaluation, i.e., SGLATrack-ViT, SGLATrack-DeiT*, and SGLATrack-EVA, respectively. The selection module $\mathcal{M}$ is a 3-layer MLP whose hidden dimension is 160, and the $l^{*}$ is set to + +6. In terms of input data, we take a template image with $128 \times 128$ pixels and a search image with $256 \times 256$ pixels as the input to the trackers. + +Training strategy. Following traditional protocols, we train our models on four datasets, including LaSOT [20], COCO [43], TrackingNet [47], and GOT-10k [31] (with 1,000 videos removed according to [61]). Same as OSTrack [63], we use typical commonly used data augmentation methods, such as horizontal flipping and brightness jittering. We train all three trackers uniformly with 300 epochs and 60k matching pairs per epoch. With AdamW [45] optimizer, the learning rates are set to $4 \times 10^{-4}$ for the prediction head, and $4 \times 10^{-5}$ for the backbone and the selection module. The learning rates drop by a factor of 10 after 240 epochs. The model training is conducted on a server with four 80GB Tesla A100 GPUs, and the batch size is uniformly fixed at 32. + +Inference. We follow the common practice [8, 63, 68] and employ the Hamming window to incorporate positional priors. Specifically, a penalty is applied to the classification map $\mathbf{p}$ using the Hanning window. After that, the box corresponding to the highest score is selected as the prediction bounding box. + +# 4.2. Results and Comparisons + +Datasets. To demonstrate the effectiveness of our method, six datasets are used for evaluation, including DTB70 [36], UAVDT [18], UAV123 [46], UAV123@10fps [46], UAVTrack112 [23], and UAVTrack112.L [23]. Specifically, DTB70 consists of 70 UAV sequences, encompassing various challenges such as UAV motion, cluttered scenes and objects with different sizes. UAVDT is primarily employed for evaluating vehicle tracking under situ- + +Table 1. AUC, precision and speed comparisons with some state-of-the-art trackers on the test set of DTB70, UAVDT, UAV123, UAVTrack112, UAVTrack112.L. For the comparison between SGLATrack-Deit* and other trackers, the top two results are highlighted with red, blue, respectively. For the comparison of SGLATrack using different backbones, the best results are shown in underline font. Note that the average performance across the five datasets is presented in the form of Avg. + +
MethodSourceDTB70UAVDTUAV123UAVTrack112UAVTrack112_LAvg.FPS
AUC(%)P(%)AUC(%)P(%)AUC(%)P(%)AUC(%)P(%)AUC(%)P(%)AUC(%)P(%)GPUCPU
SGLATrack-DeiT*Ours65.184.459.981.966.984.967.582.864.079.264.782.6224.774.8
AVTrack-DeiT [39]ICML'2465.084.358.782.166.884.865.480.362.778.263.781.9197.3-
SMAT [26]WACV'2463.881.958.780.864.681.865.382.260.876.562.680.6158.4-
HiT-Small [33]ICCV'2358.974.846.661.663.880.663.678.059.874.658.774.1194.6-
MVT [25]BMVC'2360.479.247.565.859.978.660.878.246.259.555.072.3175.1-
TCTrack++ [6]TPAMI'2362.681.552.672.259.378.064.081.562.482.060.279.0136.9-
TCTrack [5]CVPR'2262.281.253.072.560.580.059.476.658.378.658.777.8135.8-
UDAT [64]CVPR'2261.880.659.280.159.076.161.678.959.677.260.278.651.2-
HCAT [9]ECCV'2263.783.554.374.264.884.266.081.961.778.262.180.4110.1-
LightTrack [62]CVPR'2159.176.461.280.462.778.363.881.162.079.961.879.2129.8-
HiFT [4]ICCV'2159.480.247.565.259.078.757.074.255.173.455.674.3122.6-
AutoTrack [40]CVPR'2047.871.645.071.847.268.946.569.440.267.545.369.8-55.8
ARCF [32]ICCV'1947.269.445.872.046.867.145.667.239.364.044.967.9-57.6
STRCF [35]CVPR'1843.764.941.162.948.168.143.063.536.060.942.464.1-27.4
MCCT-H [54]CVPR'1840.560.440.266.845.765.943.663.437.558.541.563.0-57.0
ECO-HC [15]CVPR'1744.863.541.669.449.671.047.468.641.764.845.067.5-79.5
fDSST [16]TPAMI'1735.753.438.366.640.558.339.156.830.649.136.856.8-184.8
SGLATrack with different backbones
SGLATrack-EVAOurs63.883.557.979.965.183.0566.981.364.781.763.781.9236.977.2
SGLATrack-ViTOurs65.884.860.082.466.183.967.382.664.380.564.782.8223.274.2
SGLATrack-DeiT*Ours65.184.459.981.966.984.967.582.864.079.264.782.6224.774.8
+ +Table 2. Comparison with state-of-the-art methods on UAV123@10fps. The top two results are highlighted with red, blue, respectively. + +
MethodHiFT [4]LightTrack [62]HCAT [9]UDAT [64]TCTrack [5]TCTrack++ [6]MVT [25]HiT -Small [33]SMAT [26]AVTrack -DeiT [39]SGLATrack -DeiT*SGLATrack -ViTSGLATrack -EVA
AUC(%)57.059.964.458.558.759.957.664.363.565.865.564.564.3
P(%)74.975.181.977.877.478.073.280.980.483.282.682.281.7
FPS122.6129.8110.151.2135.8136.9175194.6158197.3224.7223.2236.9
+ +tions of different weather conditions, flying altitudes, and camera vies. UAV123 is a large-scale UAV tracking benchmark including 123 sequences with more than 112K frames. UAV123@10fps is constructed by down-sampling the original UAV123 benchmark from 30 FPS to 10 FPS to study the impact of camera capture speed on tracking performance. UAVTrack112 comprises 112 challenging aerial scenarios, totaling over 100K frames. UAVTrack112.L serves as the biggest long-term aerial tracking benchmark currently, including over 60k frames. + +Performance evaluation. We compare the SGLATrack variant with the best overall performance, i.e., SGLATrack-DeiT*, against the current state-of-the-art (SOTA) lightweight trackers. Tab. 1 shows the AUC and precision of all the compared trackers on DTB70, UAVDT, UAV123, UAVTrack112 and UAVTrack112-L datasets. As can be seen, our SGLATrack-DeiT* achieves new AUC scores on the DTB70, UAV123, UAVTrack112, and UAVTrack112-L datasets, and maintains top-three precision across all datasets. Particularly on UAVTrack112, SGLATrack-DeiT* outperforms the second-best AVTrack by $2.1\%$ in + +AUC and $2.5\%$ in precision. An interesting observation is that SGLATrack-DeiT* achieves the best AUC scores on UAVDT and UAVTrack112.L, yet its precision is inferior to some trackers with much lower AUC scores than ours. We speculate that this discrepancy arises from the bias introduced by the limited number of sequences in these datasets, as UAVDT and UAVTrack112.L contain only 50 and 45 sequences respectively. Besides, the comparison results on UAV123@10fps are shown in Tab. 2. Although the performance of SGLATrack-DeiT* is inferior to AVTrack, it significantly outperforms AVTrack in inference speed. The outstanding performance on the six datasets indicates that our approach can accurately disable redundant layers without notable performance drops, thereby confirming our intuition about the layer redundancy of lightweight ViTs. We also compare the overall performance of different SGLATrack variants. It can be seen that SGLATrack-ViT and SGLATrack-DeiT* achieve competitive performance, which demonstrates the generalization ability of our similarity-guided layer adaptation approach. + +Speed evaluation. For a fair comparison, the experi + +Table 3. Ablation studies of determining saturated layers with different methods. Note that tracking performance is shown in AUC. + +
VariantLAUAV123UAVTrack112FPSParams.FLOPs
SGLATrack-EVA✓65.367.1185.45.761.73
65.166.9236.94.151.20
SGLATrack-ViT✓66.467.7172.87.982.39
66.167.3223.25.811.68
SGLATrack-DeiT*✓67.167.8175.57.982.39
66.967.5224.75.811.68
+ +Table 4. Ablation studies of determining saturated layers with different methods. Note that tracking performance is shown in AUC. + +
MethodValueUAV123UAVTrack112FPS
Adaptiveμ = 0.8865.465.9204.6
μ = 0.9066.566.3192.8
μ = 0.9267.067.7184.1
Directl* = 566.166.4239.6
l* = 666.967.5224.7
l* = 766.967.7211.3
+ +Table 5. Ablation studies of the optimization of selection module. Note that tracking performance is shown in AUC. + +
#MethodUAV123UAVTrack112FPS
1Fixed layer65.465.8232.9
2Maximizing66.967.5224.7
3Minimizing65.764.5-
4Random selection66.066.2-
+ +ments of inference speeds are conducted on a same PC machine with an Intel i7-9700KF CPU and an Nvidia GTX 2080Ti GPU. As demonstrated in Tab. 1, our SGLATrack family achieves SOTA GPU speeds, with SGLATrack-ViT, SGLATrack-DeiT*, and SGLATrack-EVA running at effective speeds of 223.2 FPS, 224.7 FPS, and 236.9 FPS, respectively. Moreover, our SGLATrack family also demonstrates competitive real-time CPU speeds, even surpassing some famous DCF-based trackers such as AutoTrack [40] and ARCF [32]. The comparison results of inference speed indicate that our approach has strong acceleration capabilities, making it highly suitable for ViT-based UAV tracking. + +# 4.3. Ablation Study and Analysis + +To demonstrate the effectiveness of proposed layer adaptation (LA) approach, we conduct extensive ablation studies on two large-scale datasets, i.e., UAV123 [46] and UAVTrack112 [23]. Default settings are marked in gray. + +LA on different backbones. Our LA approach can be seamlessly integrated into existing ViTs and adaptively optimize their redundant layers. As shown in Tab. 3, we incorporate LA into three different ViTs to study its impact on tracking performance. It can be observed that our LA approach significantly improves inference speed with only a slight performance drop. In the best case, the approach only results in a $0.2\%$ AUC decrease, such as SGLATrack-EVA on UAV123 and UAVTrack112. In the worst case, the + +AUC loss is just $0.4\%$ (SGLATrack-ViT on UAVTrack112), which is considered acceptable given the nearly $22\%$ increase in speed, $27\%$ reduction in parameters, and $30\%$ decrease in Flops. These results provide strong evidence for the generalizability of proposed LA approach. + +Study on saturated layer. The saturated layer $l^{*}$ is crucial for balancing precision and speed in our framework. We investigate two different methods to determine when saturation state is reached. The first method calculates the cosine similarity between search features layer by layer during inference to adaptively determine the saturated layer. Given a certain threshold $\mu$ , if the cosine similarity exceeds this threshold, it indicates that saturation has been reached. The second method involves using prior knowledge to directly set a fixed value for the saturated layer $l^{*}$ , making it simpler and more effective. Tab. 4 shows the impact of SGLATrack-DeiT\* using the two method on performance and speed. As can be seen, the adaptive method minimizes performance drop when the threshold is set to 0.92. The direct method achieves the best accuracy-speed trade-off when $l^{*}$ is set to 6 with a precision that is very close to that of the adaptive method, but with a significantly higher speed. We suppose that there are three main reasons for this result: 1) The direct method does not require calculating the cosine similarity layer-by-layer, thus avoiding additional time costs. 2) Directly setting a fixed value for $l^{*}$ forces the model to accelerate its ponder and reach saturation stats earlier. In other words, if $l^{*}$ is set to 6, the model might already saturate before the 6-th layer. 3) Our proposed selection module can effectively mitigate the precision drop when a large number of layers are deactivated. Hence, the direct method with $l^{*}$ being set to 6 is adopted. + +Study on selection module. The selection module compensates for precision drop by dynamically retaining an optimal $(l^{*} + k)$ -th layer. Here, we provide various experiments to study the impact of this module. #1 Fixed layer. $k$ is set to 1, meaning the first layer after the saturation layer $l^{*}$ is retained. Since $k$ is fixed, the selection module is removed in this experiment. #2 Maximizing. Encourage the model to further strengthen its focus on the target by maximizing the cosine similarity based on Eq. (4). #3 Minimizing. By minimizing cosine similarity, the model is encouraged to distract its attention to continue mining scene relationships. #4 Random selection. Unlike #2 and #3, no supervision is applied to this module. Note that experiment #2, #3, and #4 all adopt the selection module, so they share the same speed. Tab. 5 shows the impact of different methods on performance and speed. The maximizing-based method #2 achieves the best precision-speed trade-off, verifying the effectiveness of our layer-wise similarity loss. Moreover, comparing experiments #1 and #2, the method in #2 shows a slight speed decrease, due to the introduced computational cost of the selection module. + +![](images/496dbf6fbcfb84f6c11eb8281c2a53daa92e6190733777a7604a3e6429b682c1.jpg) +Figure 4. Qualitative comparisons of our tracker against other three SOTA trackers. Best viewed in color and by zooming in. + +![](images/860e7daa85f9c31c5d62bcf29166bb9d36291693b80aa10f45339182b4ca5577.jpg) +Figure 5. Comparison of attention maps. Note that w/o and w/ denote the tracker without and with selection module, respectively. + +# 4.4. Qualitative Comparison and Visualization + +Qualitative results. To intuitively demonstrate the effectiveness of proposed approach, we provide some qualitative comparison results in UAV123 [46]. The compared trackers include SGLATrack-DeiT*, AVTrack [39], TCTrack++ [6], HiFT [4]. As shown in Fig. 4, in some challenging scenarios, our SGLATrack-DeiT* with a large number of disabled layers can still maintain robust tracking performance and even outperform other trackers. These qualitative results provide additional evidence of the effectiveness of our similarity-guided layer adaptation approach. + +Visualization of attention. To intuitively demonstrate the effectiveness of layer-wise similarity loss, we visualize the cross-attention maps between the center part of template region and the search region. More specifically, we provide the cross-attention maps of the model with selection module and the model without selection module. Both models + +use the same number of VIT layers, but the last layer of the model without the selection module is $l^{*} + 1$ , while the last layer of the model with the selection module is $l^{*} + k$ . As shown in Fig. 5, the attention in the $(l^{*} + k)$ -th layer is more focused on the target compared to the $(l^{*} + 1)$ -th layer, suggesting that dynamically retaining the layer under the condition of equal layers contributes to target recognition. These visualization examples verify that our layer-wise similarity loss effectively optimizes the selection module, allowing the model to maximize its attention on the target in situations where a large number of layers are disabled. + +# 4.5. Real-world Test + +To validate the tracking efficiency under real-world conditions, we deploy SGLATrack-DeiT* on an embedded platform, NVIDIA Jetson TX2 4GB. Note that the computational performance of the Jetson TX2 is far inferior to that of the Jetson AGX Xavier and Jetson AGX Orin, yet our tracker is still able to run at a real-time speed of 33 FPS without using any acceleration techniques such as TensorRT. Therefore, the real-world tests on the embedded system directly validate the practicability of our trackers in real-world applications. + +# 5. Conclusion + +In this paper, we make the first attempt to explore the layer redundancy in tiny ViTs, from the perspective of layer-by-layer feature changes and result variations. Based on the analysis, we propose a selection module that disables many redundancy layers and selectively retains an optimal subsequent layer to balance the accuracy-speed trade-off. This module is optimized via a layer-wise similarity loss, which allows the model to maximize its focus on targets in situations where a large number of layers are disabled. Building upon this module, we tailor previously one-stream architecture into an efficient similarity-guided layer-adaptive tracking framework for real-time UAV tracking. Experimental results show that our tracker achieves state-of-the-art performance and tracking speed on six benchmarks. + +# 6. Acknowledgement + +This work is supported by the National Natural Science Foundation of China (No.U23A20383, 62472109 and 62466051), the Project of Guangxi Science and Technology (No.2024GXNSFGA010001, 2025GXNSFAA069676, and 2025GXNSFAA069417), the Guangxi "Young Bagui Scholar" Teams for Innovation and Research Project, the Research Project of Guangxi Normal University (No.2024DF001), the Innovation Project of Guangxi Graduate Education (YCBZ2024083), and the grant from Guangxi Colleges and Universities Key Laboratory of Intelligent Software (No.2024B01). + +# References + +[1] Arian Bakhtiarnia, Qi Zhang, and Alexandros Iosifidis. Multi-exit vision transformer for dynamic inference. In The 32nd British Machine Vision Conference, 2021. 3 +[2] Luca Bertinetto, Jack Valmadre, João F. Henriques, Andrea Vedaldi, and Philip H. S. Torr. Fully-convolutional siamese networks for object tracking. In Computer Vision - ECCV 2016 Workshops, pages 850-865, Cham, 2016. Springer International Publishing. 1, 2 +[3] David S. Bolme, J. Ross Beveridge, Bruce A. Draper, and Yui Man Lui. Visual object tracking using adaptive correlation filters. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 2544-2550, 2010. 2 +[4] Ziang Cao, Changhong Fu, Junjie Ye, Bowen Li, and Yiming Li. Hift: Hierarchical feature transformer for aerial tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15457-15466, 2021. 6, 8 +[5] Ziang Cao, Ziyuan Huang, Liang Pan, Shiwei Zhang, Ziwei Liu, and Changhong Fu. Ttrack: Temporal contexts for aerial tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14798-14808, 2022. 6 +[6] Ziang Cao, Ziyuan Huang, Liang Pan, Shiwei Zhang, Ziwei Liu, and Changhong Fu. Towards real-world visual tracking with temporal contexts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 6, 8 +[7] 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. 3 +[8] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8122-8131, 2021. 2, 5 +[9] Xin Chen, Ben Kang, Dong Wang, Dongdong Li, and Huchuan Lu. Efficient visual tracking via hierarchical cross-attention transformer. In European Conference on Computer Vision, pages 461–477. Springer, 2022. 6 +[10] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5270-5279, 2022. 3 +[11] Zedu Chen, Bineng Zhong, Guorong Li, Shengping Zhang, and Rongrong Ji. Siamese box adaptive network for visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6668-6677, 2020. 2 +[12] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13598-13608, 2022. 2 +[13] Martin Danelljan, Gustav Hager, Fahad Shahbaz Khan, and Michael Felsberg. Learning spatially regularized correlation filters for visual tracking. In Proceedings of the IEEE inter + +national conference on computer vision, pages 4310-4318, 2015. 1 +[14] Martin Danelljan, Andreas Robinson, Fahad Shahbaz Khan, and Michael Felsberg. Beyond correlation filters: Learning continuous convolution operators for visual tracking. In Computer Vision - ECCV 2016, pages 472-488, Cham, 2016. Springer International Publishing. 2 +[15] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Eco: Efficient convolution operators for tracking. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6931-6939, 2017. 6 +[16] Martin Danelljan, Gustav Hager, Fahad Shahbaz Khan, and Michael Felsberg. Discriminative scale space tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(8):1561-1575, 2017. 1, 2, 6 +[17] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 5 +[18] Dawei Du, Yuankai Qi, Hongyang Yu, Yifan Yang, Kaiwen Duan, Guorong Li, Weigang Zhang, Qingming Huang, and Qi Tian. The unmanned aerial vehicle benchmark: Object detection and tracking. In Proceedings of the European conference on computer vision (ECCV), pages 370-386, 2018. 5 +[19] Guocai Du, Peiyong Zhou, Nurbiya Yadikar, Alimjan Aysa, and Kurban Ubul. Ddtrack: Dynamic token sampling for efficient uav transformer tracking. In International Conference on Pattern Recognition, pages 129-144. Springer, 2025. 2 +[20] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5374-5383, 2019. 5 +[21] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 5 +[22] Zhengcong Fei, Xu Yan, Shuhui Wang, and Qi Tian. Deecap: Dynamic early exiting for efficient image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12216-12226, 2022. 3 +[23] Changhong Fu, Ziang Cao, Yiming Li, Junjie Ye, and Chen Feng. Onboard real-time aerial tracking with efficient siamese anchor proposal network. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2021. 5, 7 +[24] Mor Geva, Avi Caciularu, Kevin Wang, and Yoav Goldberg. Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 30-45, 2022. 2 +[25] Goutam Yelluru Gopal and Maria A Amer. Mobile vision transformer-based visual object tracking. 2023. 6 +[26] Goutam Yelluru Gopal and Maria A Amer. Separable self and mixed attention transformers for efficient object track + +ing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6708-6717, 2024. 6 +[27] João F. Henriques, Rui Caseiro, Pedro Martins, and Jorge Batista. High-speed tracking with kernelized correlation filters. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(3):583-596, 2015. 1, 2 +[28] Xiantao Hu, Ying Tai, Xu Zhao, Chen Zhao, Zhenyu Zhang, Jun Li, Bineng Zhong, and Jian Yang. Exploiting multimodal spatial-temporal patterns for video object tracking. arXiv preprint arXiv:2412.15691, 2024. 2 +[29] Xiantao Hu, Bineng Zhong, Qihua Liang, Shengping Zhang, Ning Li, and Xianxian Li. Towards modalities correlation for rgb-t tracking. IEEE Transactions on Circuits and Systems for Video Technology, 2024. +[30] Xiantao Hu, Bineng Zhong, Qihua Liang, Shengping Zhang, Ning Li, Xianxian Li, and Rongrong Ji. Transformer tracking via frequency fusion. IEEE Transactions on Circuits and Systems for Video Technology, 34(2):1020-1031, 2024. 2 +[31] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE transactions on pattern analysis and machine intelligence, 43(5):1562-1577, 2019. 5 +[32] Ziyuan Huang, Changhong Fu, Yiming Li, Fuling Lin, and Peng Lu. Learning aberrance repressed correlation filters for real-time uav tracking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2891-2900, 2019. 2, 6, 7 +[33] Ben Kang, Xin Chen, Dong Wang, Houwen Peng, and Huchuan Lu. Exploring lightweight hierarchical vision transformers for efficient visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9612-9621, 2023. 6 +[34] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8971-8980, 2018. 1, 2 +[35] Feng Li, Cheng Tian, Wangmeng Zuo, Lei Zhang, and Ming-Hsuan Yang. Learning spatial-temporal regularized correlation filters for visual tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4904-4913, 2018. 6 +[36] Siyi Li and Dit-Yan Yeung. Visual object tracking for unmanned aerial vehicles: A benchmark and new motion models. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 5 +[37] Shuiwang Li, Yangxiang Yang, Dan Zeng, and Xucheng Wang. Adaptive and background-aware vision transformer for real-time UAV tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13989-14000, 2023. 2 +[38] Shuiwang Li, Xiangyang Yang, Xucheng Wang, Dan Zeng, Hengzhou Ye, and Qijun Zhao. Learning target-aware vision transformers for real-time UAV tracking. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2 +[39] Yongxin Li, Mengyuan Liu, You Wu, Xucheng Wang, Xi-angyang Yang, and Shuiwang Li. Learning adaptive and + +view-invariant vision transformer for real-time uav tracking. In Forty-first International Conference on Machine Learning. 2, 3, 4, 6, 8 +[40] Yiming Li, Changhong Fu, Fangqiang Ding, Ziyuan Huang, and Geng Lu. Autotrack: Towards high-performance visual tracking for uav with automatic spatio-temporal regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11923-11932, 2020. 1, 2, 6, 7 +[41] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35: 12934-12949, 2022. 3 +[42] Bin Lin, Jinlei Zheng, Chaocan Xue, Lei Fu, Ying Li, and Qiang Shen. Motion-aware correlation filter-based object tracking in satellite videos. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1 +[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2999-3007, 2017. 5 +[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. 2019. 5 +[46] Matthias Mueller, Neil Smith, and Bernard Ghanem. A benchmark and simulator for uav tracking. In Computer Vision - ECCV 2016, pages 445-461, Cham, 2016. Springer International Publishing. 4, 5, 7, 8 +[47] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In Proceedings of the European conference on computer vision (ECCV), pages 300-317, 2018. 5 +[48] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 3 +[49] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5 +[50] Liangtao Shi, Bineng Zhong, Qihua Liang, Ning Li, Shengping Zhang, and Xianxian Li. Explicit visual prompts for visual object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4838-4846, 2024. 2 +[51] Shengkun Tang, Yaqing Wang, Zhenglun Kong, Tianchi Zhang, Yao Li, Caiwen Ding, Yanzhi Wang, Yi Liang, and Dongkuan Xu. You need multiple exiting: Dynamic early exiting for accelerating unified vision language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10781-10791, 2023. 2, 3 + +[52] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12165-12174, 2022. 3 +[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 3, 5 +[54] Ning Wang, Wengang Zhou, Qi Tian, Richang Hong, Meng Wang, and Houqiang Li. Multi-cue correlation filters for robust visual tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4844-4853, 2018. 6 +[55] 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 +[56] Jinxia Xie, Bineng Zhong, Qihua Liang, Ning Li, Zhiyi Mo, and Shuxiang Song. Robust tracking via mamba-based context-aware token learning. arXiv preprint arXiv:2412.13611, 2024. 2 +[57] Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. Autoregressive queries for adaptive tracking with spatio-temporal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19300-19309, 2024. 2 +[58] Guanyu Xu, Jiawei Hao, Li Shen, Han Hu, Yong Luo, Hui Lin, and Jialie Shen. Lgvit: Dynamic early exiting for accelerating vision transformer. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9103-9114, 2023. 3 +[59] Yinda Xu, Zeyu Wang, Zuoxin Li, Ye Yuan, and Gang Yu. Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines. In Proceedings of the AAAI conference on artificial intelligence, pages 12549-12556, 2020. 2 +[60] Yuanliang Xue, Guodong Jin, Tao Shen, Lining Tan, Nian Wang, Jing Gao, and Lianfeng Wang. Smalltrack: Wavelet pooling and graph enhanced classification for uav small object tracking. IEEE Transactions on Geoscience and Remote Sensing, 2023. 1 +[61] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10428-10437, 2021. 2, 5 +[62] Bin Yan, Houwen Peng, Kan Wu, Dong Wang, Jianlong Fu, and Huchuan Lu. Lighttrack: Finding lightweight neural networks for object tracking via one-shot architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15189, 2021. 6 +[63] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In Computer Vision - ECCV 2022, pages 341-357, Cham, 2022. Springer Nature Switzerland. 2, 5 + +[64] Junjie Ye, Changhong Fu, Guangze Zheng, Danda Pani Paudel, and Guang Chen. Unsupervised domain adaptation for nighttime aerial tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8896-8905, 2022. 6 +[65] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10809-10818, 2022. 3 +[66] Jinnian Zhang, Houwen Peng, Kan Wu, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Minivit: Compressing vision transformers with weight multiplexing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12145-12154, 2022. 3 +[67] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4591-4600, 2019. 2 +[68] Zhipeng Zhang, Houwen Peng, Jianlong Fu, Bing Li, and Weiming Hu. Ocean: Object-aware anchor-free tracking. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 771-787. Springer, 2020. 5 +[69] Zheng Zhu, Qiang Wang, Bo Li, Wei Wu, Junjie Yan, and Weiming Hu. Distractor-aware siamese networks for visual object tracking. In Proceedings of the European conference on computer vision (ECCV), pages 101-117, 2018. 1, 2 \ No newline at end of file diff --git a/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/images.zip b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a0d368b22a9e8f929fad59c4ed4df0002e57c990 --- /dev/null +++ b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ccd7a997e472723e0507f73b074c1a989fb0b3c8f2a99525aa1d1a5191e5c95 +size 586599 diff --git a/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/layout.json b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0d43516e1e67b6d3a07b2db2bab7eb37d2e520a6 --- /dev/null +++ b/CVPR/2025/Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd3a35a3f7ab76f3c3396f8a9f116d385188c16341131c4effd0d0d5395528d4 +size 401109