diff --git a/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/475ec1f3-e3db-4386-86c1-e4af56b1643d_content_list.json b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/475ec1f3-e3db-4386-86c1-e4af56b1643d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cfd62ecd9a08714e7886cf3569268b6713e329d7 --- /dev/null +++ b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/475ec1f3-e3db-4386-86c1-e4af56b1643d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb8597344754ac0ba0447523103767cdffae9e3b742f1d7f7b42a7075e1962f4 +size 79814 diff --git a/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/475ec1f3-e3db-4386-86c1-e4af56b1643d_model.json b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/475ec1f3-e3db-4386-86c1-e4af56b1643d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..328687b9b27d69592ee2adcb21ac3888dda4d87f --- /dev/null +++ b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/475ec1f3-e3db-4386-86c1-e4af56b1643d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4c0ab79cd044f191ecd303f5ee1b58d49b91a2f8ed1a85a2700ba4406c34bc0 +size 97311 diff --git a/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/475ec1f3-e3db-4386-86c1-e4af56b1643d_origin.pdf b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/475ec1f3-e3db-4386-86c1-e4af56b1643d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0b2c30e6d1af9dee653f2b5da3fbdab3594692a7 --- /dev/null +++ b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/475ec1f3-e3db-4386-86c1-e4af56b1643d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f73d0e71a62f8d4668b95ff886ca1190be3b454a5a39cc092d6de21441424f14 +size 5528948 diff --git a/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/full.md b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3d9bdbf800fff278de0507c14c6fa0f7e1c7e2ee --- /dev/null +++ b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/full.md @@ -0,0 +1,332 @@ +# R1-Onevision: Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization + +Yi Yang $^{1*}$ Xiaoxuan He $^{1,3*}$ Hongkun Pan $^{2*}$ Xiyan Jiang $^{1}$ Yan Deng $^{1}$ Xingtao Yang $^{1}$ Haoyu Lu $^{4}$ Dacheng Yin $^{3}$ Fengyun Rao $^{3}$ Minfeng Zhu $^{2\dagger}$ Bo Zhang $^{2\dagger}$ Wei Chen $^{1\dagger}$ $^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Zhejiang University $^{3}$ WeChat Vision, Tencent Inc. $^{4}$ Renmin University of China + +![](images/a1e23ee67a5e81011a05a2a7150e7042042a32d314a60560e156337167f40452.jpg) +? +Figure 1. Comparison of baseline models and R1-Onevision. Deepseek-R1 struggles with perception errors due to GPT-4o's incomplete image description and Qwen2.5-VL lacks the necessary reasoning ability to solve the problem. In contrast, R1-Onevision accurately interprets the image, applies structured reasoning, and arrives at the correct solution. + +Question: Cynthia paints each region of the figure in a single color: red, blue or yellow. She paints with different colors the regions that touch each other. In how many different ways can Cyntia paint the figure? + +![](images/74bca2b42d9e13b122bdc2c76960acc0eb637989755e145b7f002873353ab559.jpg) + +# GPT-4o Caption: + +The image features a series of concentric circles with varying thicknesses, forming a symmetrical pattern. The outermost circle is the largest, enclosing progressively smaller circles toward the center. Some of the inner circles intersect with each other, creating curved segments that resemble a layered or nested design. The overall composition appears structured and balanced, with smooth, evenly spaced curves contributing to a visually appealing geometric arrangement. The image is monochromatic, utilizing black outlines on a white background. + +![](images/8bcf068a006da5acca5f71fe2e43cf51536cc0074a7c220a34e64c142adfbaaf.jpg) + +# Deepseek-R1: + +Okay, let's tackle this problem step by step... First, I need to visualize the figure described... Since it's concentric circles,... The challenge is to figure out how many regions there are and their adjacency. Let me try to break it down... Let's consider possible common figures of this type... Case 1: A simple target with three concentric circles... Total: $3^{*}2^{*}2^{*}2 = 24$ . Case 2: Suppose there are two intersecting circles inside... Total: $3^{*}2^{*}2^{*} = 12$ . However, I recall that in similar problems with a central region and surrounding regions forming a cycle... Assuming the structure forms a chain of outer, middle, and inner annuli, plus four inner regions arranged in a cycle the coloring constraints are: + +1. Outer annulus (O): 3 color choices: 2.Middle color: (M) 2. + +2. Middle annulus (M): 2 choices (different from O). +3. Inner annulus (I): 2 choices (different from M). + +4. Four inner regions: Form a cycle, requiring 2 alternating colors + +different from 1, yielding 2 colorings. + +Total colorings: + +3 (O)*2 (M)*2 (I)*2 (inner cycle) = 24. Final answer: 24. + +![](images/9a7811c5bfaa9563efeeb9009ecce8482047020b926095a51b0fabfa6f419b7d.jpg) + +# Qwen2.5-VL-7B: + +To determine the number of different ways Cynthia can paint the figure, we need to consider the constraints given: each region must be painted with a different color from its adjacent regions. Let's analyze the figure step by step. + +1. Identify the regions. 2. Coloring the innermost circle3. Coloring the next circle: The next circle must be a different color from the innermost circle. So, there are 2 choices for this circle. 4. Coloring the next circle: The third circle must be a different color from the second circle. So, there are 2 choices for this circle.5. Coloring the next circle6. Coloring the outermost circle - 3 choices for the innermost circle. + +- 2 choices for the second circle. 2 choices for the third circle.2 choices for the fourth circle. 2 choices for the outermost circle. The total number of different ways to paint the figure is: $3^{*}2^{*}2^{*}2^{*}2 = 3^{*}16 = 48$ . Thus, the number of different ways Cynthia can paint the figure is 48 + +![](images/4de7d7821bad7c3337471c9de946d4dbfb815f1dfe603119b34b4fc235205c06.jpg) + +# R1-Onevision: + +The image shows a figure with five concentric circles, each touching the next. Cynthia uses three colors: red, blue, and yellow. Each circle must be painted a different color from its neighbors. Since the circles are concentric, each circle touches the next, meaning each circle must be painted a different color. + +Let's analyze the coloring: + +1. The innermost circle can be any of the three colors (red, blue, or yellow). + +2. The next circle must be a different color from the innermost circle. + +3. The third circle must be a different color from the second circle. + +4. The fourth circle must be a different color from the third circle. + +5. The outermost circle must be a different color from the fourth circle. + +Given that there are three colors and each circle must be a different color from its neighbors, the number of ways to color the circles is the number of permutations of the three colors. This is calculated as $3!$ (3 factorial), which is 6. Therefore, the number of different ways Cynthia can paint the figure is 6. + +The number of different ways Cynthia can paint the figure is 6. + +# Abstract + +Large Language Models have demonstrated remarkable reasoning capability in complex textual tasks. However, multimodal reasoning, which requires integrating visual and textual information, remains a significant challenge. Existing visual-language models often struggle to effectively analyze and reason visual content, resulting in suboptimal performance on complex reasoning tasks. Moreover, the absence of comprehensive benchmarks hinders the accurate assessment of multimodal reasoning capabilities. In this paper, we introduce R1-Onevision, a multimodal reasoning model designed to bridge the gap between visual perception and deep reasoning. To achieve this, we propose a cross-modal reasoning pipeline that transforms images into formal textual representations, enabling precise language-based reasoning. Leveraging this pipeline, we construct the R1-Onevision dataset which provides detailed + +step-by-step multimodal reasoning annotations across diverse domains. We further develop the R1-Onevision model through supervised fine-tuning and reinforcement learning to cultivate advanced reasoning and robust generalization abilities. To comprehensively evaluate multimodal reasoning performance across different grades, we introduce R1-Onevision-Bench, a benchmark aligned with human educational stages, covering exams from junior high school to university and beyond. Experimental results show that R1-Onevision achieves state-of-the-art performance, outperforming models such as GPT-4o and Qwen2.5-VL on multiple challenging multimodal reasoning benchmarks. Code, dataset and benchmark are available at https://github.com/Fancy-MLLM/R1-Onevision + +# 1. Introduction + +Recent language reasoning models like Deepseek-R1 [8], chain-of-thought prompting [28], and Cumulative Reasoning [39] have achieved remarkable progress in solving complex problems including coding [3, 7], mathematics [5, 10], + +and science [27]. However, multimodal reasoning remains a largely underexplored challenge. Unlike textual reasoning, multimodal reasoning requires the model to iteratively extract, structure, and verify information from images. Existing visual-language models often fail to organize available information and conduct in-depth reasoning processes, leading to failures in visual reasoning tasks. + +Current research on visual-language models has increasingly emphasized step-by-step reasoning. Some approaches, such as Llava-CoT [32] and Llama-V-o1 [24], employ a predefined thinking structure to constrain the model's reasoning process, limiting its robustness and creative potential. While such structured templates improve consistency, they often lead to shallow reasoning with limited comprehension. The others, like MAmmoTH-VL [9], rely on direct imitation of curated ground-truth answers, causing models to generate responses directly without a trial-and-error manner. Consequently, these models may fail to generalize beyond their training distributions. An example of these limitations is shown in Fig. 1. DeepseekR1 suffers from perception errors due to incomplete image descriptions from GPT-4o [19], while Qwen2.5-VL [2], despite its strong multimodal capabilities, lacks deep reasoning ability and ultimately fails to solve the problem. These challenges highlight the limitations of current multimodal reasoning approaches, which either impose rigid frameworks that restrict generalization or fail to provide humanlike thinking to accurately process visual information. + +In addition, there is a lack of comprehensive benchmarks for evaluating multimodal reasoning capability across a diverse range of tasks. Existing benchmarks used by open-compass [6] mainly focus on mathematical problems, such as MathVision [25], MathVista [17], and WeMath [22]. Although some more challenging datasets, like HumanEvalV [37], Humanity's Last Exam [20], and ZeroBench [23], aim to assess complex diagram understanding or broader knowledge proficiency, these benchmarks remain specialized, covering only limited aspects of reasoning. + +In this work, we address these challenges by introducing a cross-modal reasoning pipeline to construct a multimodal reasoning dataset, a post-training strategy to empower reasoning capabilities, and a comprehensive multimodal reasoning benchmark. First, we propose a cross-modal reasoning pipeline that transforms images into visual formal representations and allows language models to process and reason over images precisely. We construct the R1-Onevision dataset to provide a detailed step-by-step multimodal reasoning process over diverse domains, including natural scenes, charts, mathematical expressions, and science. Second, we employ a two-stage post-training strategy to train our R1-Onevision model with advanced reasoning abilities. The supervised finetuning (SFT) stage cultivates coherent thinking patterns and + +output structures using our R1-Onevision dataset. The reinforcement learning (RL) stage strengthens both the model's reasoning performance and its generalization across diverse tasks. Third, to evaluate multimodal reasoning models, we introduce R1-Onevision-Bench, a comprehensive benchmark explicitly designed to assess "grade-level" reasoning performance across scientific domains in the human educational system: mathematics, physics, chemistry, biology, and logical deduction. Finally, extensive experiments on several benchmarks, including MathVision [25], MathVerse [38], MathVista [17], WeMath [22] and our R1-Onevision-Bench, demonstrate the superior multimodal reasoning performance of our R1-Onevision model. + +Our main contributions are summarized as follows: + +- We present a cross-modal reasoning pipeline to construct R1-Onevision Dataset which encompasses a broad spectrum of images, questions and their reasoning process. +- We introduce R1-Onevision, a visual language reasoning model designed for complex problem-solving. +- We build R1-Onevision-Bench to assess multimodal reasoning performance across different educational levels and subject areas in a course-aware manner. +- Extensive experiments show the superiority of R1-Onevision compared to the baseline Qwen2.5-VL model and closed-source models like GPT-4o. + +# 2. Related Work + +Multimodal Large Language Models. MLLMs have shown significant potential in a wide range of multimodal tasks. [2, 4, 18, 26] achieve excellent visual understanding ability by integrating both visual and textual data. Recently, Multimodal Large Language Models [9, 31] demonstrate the advanced reasoning abilities when interpreting visual tasks. In addition, some studies [24, 31] introduce plan-based CoT prompting to guide models to generate intermediate information for predicting final answers. LLaVA-CoT [31] introduces a novel VLM designed to conduct visual reasoning, demonstrating exceptional performance on tasks that require structured thinking and reasoning. Based on LLaVA-CoT, LlamaV-o1 [24] introduces a multi-step curriculum learning approach, where tasks are progressively organized to facilitate incremental skill acquisition and problem-solving. CoMCTS [33] introduces the concept of collective learning into "tree search" for efficient reasoning-path searching and learning. In this paper, we propose R1-Onevision, which uses supervised fine-tuning and reinforcement learning during the post-training phase to obtain reasoning ability across different tasks. + +Large Language Model Reasoning. The development of robust reasoning capabilities in Large Language Models (LLMs) has been a focal point of research [1, 11, 15, 21]. Some LLMs understand and solve questions by learning to + +![](images/9b07d336561ddd1264bf97c87600858a87162261a05a9eff00583114ba1865e2.jpg) +Figure 2. Overview of our R1-Onevision framework. The cross-modal reasoning pipeline begins with data preparation, integrating visual formal descriptions to enhance reasoning. We employ role-playing strategies to generate the R1-Onevision dataset with high-quality reasoning data. Post-training consists of supervised fine-Tuning to learn reasoning manner, followed by rule-based Reinforcement Learning to improve generalization across multimodal tasks. + +create each intermediate step of the reasoning involved till the final answer. For example, Chain-of-Thought (CoT) prompting, where a complex question is decomposed into intermediate reasoning steps, has shown promise in guiding LLMs to structured solutions [29, 34]. Since certain visual tasks, such as solving Sudoku puzzles, urgently require multimodal large language models (MLLMs) to provide step-by-step reasoning capabilities, it is crucial to develop appropriate methods to enhance the performance of multimodal foundational models. + +Visual Reasoning Benchmarks. With the advancement of model capabilities in visual reasoning, an increasing number of studies have proposed various benchmarks to evaluate the reasoning abilities of these models [12, 14, 36]. The MATH-Vision (MATH-V) [25] dataset collects 3,040 high-quality mathematical problems with visual contexts sourced from real math competitions. DynaMATH [40] provides a dynamic visual math benchmark specifically designed to evaluate the mathematical reasoning robustness of VLMs. In this paper, we introduce a comprehensive benchmark designed to evaluate "grade-level" reasoning performance across scientific domains in the educational system. + +# 3. Method + +Multimodal reasoning requires a comprehensive understanding of both visual and textual modalities, yet existing models often fail to understand visual content and struggle with high-level reasoning. To bridge this gap, we introduce a cross-modal reasoning pipeline that transfers reasoning capabilities of language models to visual modality through formal visual representations. Further, we employ a post-training strategy to stabilize reasoning processes and improve generalization across diverse multimodal tasks. + +# 3.1. Cross-Modal Reasoning Pipeline + +Our cross-modal reasoning pipeline is designed to bridge the gap between language reasoning models and vision models by integrating visual formal representations. Formal language is a structured system with strict syntactic and semantic constraints that eliminate ambiguity, ensuring logical consistency. With formal description of visual content, language reasoning models are able to see and reason over image elements. We use DeepSeek-R1 [8] to generate reasoning processes on LLaVA-OneVision [16] and collect these data into the R1-Onevision dataset. Figure 2 illustrates the cross-modal reasoning process. Examples of the generated data are provided in the supplementary. + +Data Curation and Filtering. For visual reasoning, we aggregate diverse multimodal datasets covering natural images, OCR-based text extraction, charts, mathematical expressions, and scientific reasoning problems, selecting only those that support structured reasoning. The final dataset incorporates components from the LLaVA-OneVision dataset [16], augmented with domain-specific datasets tailored for complex inference tasks. + +Image Formal Description. A key feature of our pipeline lies in its use of formal language-based annotations. To achieve this, we leverage a combination of GPT-40, Grounding DINO, and EasyOCR to translate visual image content into formal textual description. We outline the annotation process below and provide details of the prompt design in the Appendix ?? + +- Charts & Diagrams: We prompt GPT-4o to generate structured representations, such as SPICE for circuit schematics, PlantUML or Mermaid.js for flowcharts, HTML for UI layouts, CSV/JSON for tables, and Matplotlib for annotated charts. +- Natural Scenes: We enhance images with fine-grained spatial description by leveraging Grounding DINO to extract bounding box annotations of key elements and GPT-4o to generate descriptive captions. +- Text-only Images: When dealing with images containing printed or handwritten text, we employ EasyOCR to extract text with positions and use GPT-4o to reconstruct the original document. +- Images with Visual and Textual Content: We integrate GPT-4o-generated captions, Grounding DINO bounding boxes, and EasyOCR-extracted text to ensure both textual and visual elements are accurately captured. +- Mathematical Images: For images containing mathematical content, we employ GPT-4o to propose reasoning strategies to guide the inference process. + +Reasoning Process Generation. Given an image, we prompt language reasoning model with its dense captions and questions to construct cross-modal Chain-of-Thought (CoT) data. While the original Chain-of-Thought (CoT) approach offered a structured reasoning path based on textual captions, it inherently lacked the essential visual component—the ability to directly “see” and interpret the image. To address this limitation, we propose a Role-Playing strategy that emulates human-like visual comprehension. This method involves iteratively revisiting the image, refining its understanding, and enhancing the fidelity of the reasoning process. This process improves multimodal coherence and ensures a contextually rich reasoning process. + +Quality Assurance. To ensure the reliability of generated reasoning process, we employ GPT-4o to remove inaccurate, irrelevant, or inconsistent CoT steps. This step guarantees a high-quality dataset for multimodal reasoning. + +![](images/cae5599dcd930057f32874bbc8f87dedbfcdf71abf1733cf51ddd4bcabaf2272.jpg) +R1-Onevision Dataset +Figure 3. The data distribution of R1-Onevision dataset. R1-Onevision consists of $155\mathrm{k}$ data, including Science, Math, General and Doc/Chart/Screen. + +R1-Onevision Dataset. Finally, as shown in the Figure 3, the R1-Onevision dataset is a carefully crafted tool designed to push the boundaries of multimodal reasoning. R1-Onevision encompasses a wide range of domains, including science, mathematics, chart data, and general real-world scenarios, totaling over $155\mathrm{k}$ carefully curated samples. R1-Onevision serves as a rich resource for developing visual reasoning model. + +# 3.2. Post-Training Strategy + +To enhance multimodal reasoning capabilities, we introduce a two-stage post-training strategy consisting of Supervised Fine-Tuning (SFT) and rule-based Reinforcement Learning (RL). SFT stabilizes the model's reasoning process and standardizes its output format, while RL further improves generalization across diverse multimodal tasks. + +# 3.2.1. Supervised Fine-Tuning + +Leveraging the R1-Onevision dataset, we enhance the reasoning capabilities of visual language models through SFT. SFT stabilizes the model output format and cultivates a more sophisticated reasoning process in the large-scale model. This approach not only standardizes the output but also lays a solid foundation, enabling subsequent RL to achieve significant performance gains. + +# 3.2.2. Reinforcement Learning on the SFT Model + +Building upon the SFT-trained model, we employ rule-based reinforcement learning (RL) to optimize structured reasoning and ensure output validity. Specifically, we define two reward rules inspired by R1 and update the model using Group Relative Policy Optimization (GRPO). The RL stage further encourages the model to generate reliable outputs and enhances the generalizability of the model. + +Rule-Based Rewards. As illustrated in the Reinforcement Learning part of Figure 2, we define two reward rules that assess the generated answers from two perspectives: + +- Format Reward: In order to ensure the existence of the reasoning process, the format reward rule requires that the response must follow a strict format where the model's reasoning is enclosed between and . A regular expression ensures the presence and correct ordering of these reasoning markers. +- Accuracy Reward: The accuracy reward requires the answer to be enclosed within and so that it can be extracted via a regular expression and compared against the ground truth to measure correctness. For deterministic tasks such as math problems, the final answer must be provided in a specified format (e.g., within a box) to enable reliable rule-based verification. In cases like object detection, the reward is determined by Intersection over Union (IoU) score with ground truth. + +Group Relative Policy Optimization. We employ GRPO to achieve balanced integration of consistent policy updates and robust reward signals in a controlled manner. For each token in the generated answer, GRPO first computes the log probabilities under both the new policy $(\pi(\theta))$ and a reference policy. GRPO then calculates a ratio of these probabilities and clip it to the range $[1 - \epsilon, 1 + \epsilon]$ to prevent overly large updates. The normalized reward (serving as the advantage) is used in a PPO-style loss: + +$$ +\mathcal {L} _ {\mathrm {c l i p}} = - \mathbb {E} \left[ \min \left(\operatorname {r a t i o} _ {t} \cdot \operatorname {A d v} _ {t}, \operatorname {c l i p p e d} \operatorname {r a t i o} _ {t} \cdot \operatorname {A d v} _ {t}\right) \right]. +$$ + +Here, $\mathrm{Adv}_t$ denotes the advantage function, capturing how much better (or worse) a particular action is compared to a baseline policy value. + +To further maintain closeness to the reference distribution, a KL divergence penalty (weighted by $\beta$ ) is added, yielding the overall loss: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {G R P O}} (\theta) = - \mathbb {E} \left[ \min \left(\operatorname {r a t i o} _ {t} \cdot \operatorname {A d v} _ {t}, \operatorname {c l i p p e d} _ {t} \operatorname {r a t i o} _ {t} \cdot \operatorname {A d v} _ {t}\right) \right. \\ \left. - \beta \cdot \operatorname {K L} \left(\pi_ {\theta} (y \mid x), \pi_ {\text {r e f}} (y \mid x)\right) \right]. \\ \end{array} +$$ + +Compared to other methods, the GRPO clipping mechanism prevents extreme policy shifts, while the KL regularization keeps the updated policy aligned with the baseline. This combination ensures that our model integrates rule-based rewards efficiently without compromising training stability. + +# 4. Multimodal Reasoning Benchmark + +To assess the reasoning capabilities of our R1-Onevision model, we introduce a dedicated multimodal reasoning benchmark—R1-Onevision-Bench. This benchmark is designed to comprehensively evaluate the model's performance across a wide spectrum of reasoning tasks, spanning multiple domains such as mathematics, physics, chemistry, biology, and logical deduction. + +
StatisticNumber
Total questions942
- Multiple-choice questions783 (83.1%)
- Free-form questions159 (16.9%)
- Newly collected questions942 (100.0%)
Grades/Categories/Subcategories4/5/38
Number of unique images938 (99.6%)
Number of unique questions942 (100.0%)
Number of unique answers152 (16.1%)
Maximum question length390
Maximum answer length54
Average question length127.2
Average answer length2.1
+ +Table 1. Key statistics of R1-Onevision-Bench. + +Inspired by human educational progression, R1-Onevision-Bench is structured to reflect graded levels of complexity. It includes real-world reasoning tests at the middle school, high school, and university levels, alongside a social test. Table 1 and Figure 4 present detailed information and structure about our benchmark. Notably, all questions in our benchmark are newly collected from publicly available international competitions, university exams, and standardized tests. Our benchmark is entirely independent of the existing benchmarks mentioned in Section 5.1, with no data overlap. This design not only mirrors the stages of cognitive development but also ensures that the evaluation covers both academic and practical reasoning skills. + +By integrating diverse problem types and difficulty levels, R1-Onevision-Bench provides a rigorous framework for benchmarking multimodal reasoning. This enables us to better evaluate the grade at which multimodal language models exhibit reasoning capabilities and identify which aspects of knowledge and experience require supplementation to further enhance their performance. + +Moreover, our benchmark further subdivides each category into specific subcategories. Figure 5 presents the distribution of tasks across categories, subcategories, and grade levels, along with examples from multiple domains. + +# 5. Experiment + +In this section, we first introduce the Experimental Setup of R1-Onevision in Section 5.1. Then, we present the main results in Section 5.2. Section 5.3 conducts a detailed evaluation of various models on the R1-Onevision benchmark. Finally, Section 5.4 conducts a systematic analysis to evaluate the importance of various components during training. + +# 5.1. Experimental Setup + +We evaluate R1-Onevision on several multimodal reasoning benchmarks, including MathVista [17], MathVision [25], + +
ModelMathVisionMathVerseMathVistaWeMath
ALLVision Only
Base Model
Qwen2.5-VL-7B [2]25.443.638.263.761.0
Our Models
R1-Onevision-7B29.946.440.064.161.8
Closed-source Models
GPT-4o [19]30.641.234.560.069.0
Other Models
InternVL2.5-8B [4]17.135.622.864.553.8
Qwen2-VL-7B [26]21.730.225.461.650.5
LLaVA-CoT-11B [32]--22.652.5-
+ +Table 2. Experimental results on mathematical reasoning benchmarks. Except for LLaVA-CoT-11B, all reported results are based on our reproduced experiments. The highest accuracy is marked in blue. + +
MethodHallusionMMStarMMVetMathVisionR1-Onevision BenchMathVerse Vision OnlyMathVista
LLaVA-CoT-11B47.8057.6060.3019.927.7122.652.5
LlamaV-o1-11B63.5159.5365.4014.4728.3423.4854.40
R1-Onevision-7B66.1859.7668.5829.936.240.064.1
+ +![](images/7f03f74e137204ca4bc769654d5ca5b64ed8bd6471c44a8ea04cd928ce50aa91.jpg) +Figure 4. Overview of R1-Onevision-Bench. This benchmark comprises 38 subcategories organized into 5 domains. The tasks are categorized into four difficulty levels, from 'Junior High School' to 'Social Test' challenges, ensuring comprehensive evaluation of model capabilities across varying complexities. + +MathVerse [38] and WeMath [22]. The performance of all models is derived from VLMEvalKit's testing results [6]. + +We adopt Qwen2.5-VL series as baseline models, and conduct experiments on baselines 3B and 7B to examine the effectiveness of our method. The SFT experiments are conducted with a batch size of 128, a learning rate of 1e-5, and training for one epoch. Then, we perform RL on the CLEVR dataset [13]. We experiment with training subsets of $10\mathrm{k}$ for a single epoch each. SFT and RL training took 14 hours and 16 hours respectively on 8 NVIDIA A800 GPUs. More details about the dataset construction, training and benchmark are provided in the supplementary. + +Table 3. Comparison with previous multimodal large language models for reasoning. We compare our model R1-Onevision-7B with LLaVA-CoT and LlamaV-o1 on both general benchmarks and reasoning benchmarks. + +
BenchmarkMathVisionWeMathMathVerse +(Vision Only)R1-Onevision +Bench
Reflection times1.221.260.911.51
+ +Table 4. Results on the reflection capability of R1-Onevision. + +# 5.2. Main Results + +Table 2 summarizes the quantitative results, showcasing that R1-Onevision consistently outperforms state-of-the-art (SoTA) multimodal methods across the majority of benchmarks. R1-Onevision achieved significant accuracy improvements on MathVerse and MathVision, showing that fine-grained visual-textual alignment and coherent chain-of-thought reasoning effectively enhance performance. It surpasses GPT-4o by $5.2\%$ on MathVerse ALL, $5.5\%$ on MathVerse (Vision Only), and $4.1\%$ on MathVista. On MathVision it reaches $29.9\%$ accuracy, matching GPT-4o + +![](images/7e0824b326e3cf4b0f0fb6ad951dd807e10ba1f7d0111d26c3dabbb6cd631919.jpg) +Figure 5. Samples across each category from R1-Onevision-Bench. R1-Onevision-Bench comprises multi-modal questions and answers in math, physics, chemistry, biology and deduction across multiple educational grades that require deep thinking and reasoning. + +![](images/80d94795b037fe25870fe5e5c4afe8cc1b8883dfc9860a8dfcb59fd3684b96ba.jpg) + +![](images/bc208749b809209e019fd4c65f169cbe9b78e27abf2348e90f037cff515e7629.jpg) + +![](images/8e91e40d5371a2f6e739a934bab026b680a58c76f67df2e8ed94c6412532e7d2.jpg) + +![](images/396b78a59f9552b27364bac16ce35e19b6af39832b3d08685be169c12870e80b.jpg) + +and underscoring its strong competitiveness. + +In Table 3, we compare our method with LLaVA-CoT and LlamaV-o1 on three general benchmarks and four reasoning benchmarks. Notably, MM Vet comprises complex multimodal tasks that require free-form answers. The results show that ours outperforms previous multimodal reasoning models on both general and reasoning benchmarks. + +Furthermore, we investigate the reflection capability of our model, as shown in Table 4. Following DeepSeekR1, we count reflection terms in CoT to calculate reflection times. The results demonstrate that our model is capable of self-reflection and self-correction during reasoning process. + +# 5.3. Benchmark Analysis + +We evaluate the performance of two sets of models in the R1-Onevision benchmark, which categorize into four levels of difficulty (ranging from 'Junior High School' to 'Social Test') and five distinct academic disciplines. The two model groups consist of closed-source VLMs and open-source VLMs, respectively. We deploy VLMEvalkit package to infer in various models. As for scoring in R1-Onevision-Bench, we use GPT-4o-mini to extract the answer and score. The detailed prompt is shown in the supplementary. + +
BenchmarkSFT+RLRL
MathVision29.928.0
+ +Table 6. Comparison of MathVision performance between SFT+RL and RL. The best results are marked in blue. + +Overall Results on Average Accuracy. Table 5 summarizes the average performance of both closed-source and open-source models. Overall, closed-source systems maintain a clear lead. Gemini-2.0-Flash achieves the highest mean accuracy, exceeding GPT-4o by approximately $10\%$ and surpassing the $50\%$ threshold on most tasks. While all models perform well on middle- and high-school + +questions, their accuracy declines on university-level and professional-certification exams. Deduction problems remain particularly challenging, with no system exceeding $40\%$ accuracy. On the open-source side, Qwen2.5-VL-72B has markedly narrowed the gap, achieving an average accuracy of $52\%$ , which matches the closed-source Claude-3.5 and ranks second only to Gemini-2.0-Flash. Among models with 4B-10B parameters, R1-Onevision, developed based on Qwen2.5-VL, has delivered outstanding results, ranking just behind the top closed-source models. + +# 5.4. Ablation Study + +# 5.4.1. Analysis of Training Strategy + +To evaluate the effectiveness of our training data, we compare two training strategies: (1) Supervised Fine-Tuning (SFT) on our dataset and (2) SFT followed by Reinforcement Learning (RL). As shown in Table 6, RL alone brings limited improvement, demonstrating that prior SFT establishes a robust foundation by aligning the model with high-quality reasoning patterns. Furthermore, as Table 7 shows, SFT alone markedly improves accuracy on MathVision and MathVerse (Vision Only) and achieves comparable performance on MathVerse All, suggesting that our data cultivates systematic reasoning skills. Adding RL stage to the SFT-trained model provides further improvement, guiding the model toward deeper, more deductive reasoning and enabling it to tackle harder, more nuanced problems. + +To further validate the effectiveness of formal descriptions and role playing in our dataset construction pipeline, we conduct additional ablation studies. Table 8 shows the SFT performance of CoT data generated from GPT-4o-labeled captions versus formal descriptions. The results demonstrate the superiority of formal description, and adding the role-playing step further improves performance. + +
ModelAvgGradeCategory
Junior High SchoolHigh SchoolCollegeSocial TestMathPhysicsChemistryBiologyDeduction
Closed-source
GPT-4o49.651.356.245.326.541.352.571.463.426.5
Gemini-2.0-Flash59.156.065.961.239.852.364.474.367.239.8
Claude-3.552.156.055.949.430.646.554.366.765.730.6
Open-source
MiniCPM-o-2.6 [35]30.433.431.721.231.624.231.730.541.831.6
InternVL2.5-8B [4]29.533.130.621.827.626.324.832.446.327.6
InternVL2.5-8B-MPO [4]32.537.433.624.726.528.729.941.044.826.5
Qwen2-VL-7B [26]30.035.428.525.926.526.328.130.545.526.5
Qwen2.5-VL-7B [2]32.133.837.125.319.431.527.339.047.019.4
DeepSeek-VL2 [30]29.834.430.918.830.623.528.429.547.830.6
R1-Onevision-7B36.240.139.527.626.533.030.249.553.026.5
Qwen2.5-VL-72B [2]52.054.356.754.123.548.955.863.863.423.5
+ +Table 5. Evaluation results of closed-source and open-source multimodal models on R1-Onevision benchmark. We divide the benchmark into four grades ('Junior High School', 'High School', 'College', 'Social Test') and five categories ('Math', 'Physics', 'Chemistry', 'Biology', 'Deduction'). The highest accuracy for closed-source and open-source LMMs are marked in green and blue, respectively. + +
ModelMathVisionMathVerseMathVerse +(Vision Only)
Qwen2.5-VL-7B25.443.638.2
SFT26.343.439.7
SFT+RL29.946.440.0
+ +Table 7. Ablation study on training strategies. + +
Formal DescriptionRole PlayingMathvisionMathVerse (Vision Only)Mathverse
10.8631.9829.39
✓21.0536.2940.43
✓✓26.3239.7243.38
+ +Table 8. Ablation study for formal description and role playing + +
ModelMathVisionMathVerseMathVerse +(Vision Only)
Qwen2.5-VL-3B21.734.731.2
R1-Onevision-3B23.738.635.5
+ +Table 9. Evaluation of our method using Qwen2.5-VL-3B base model on MathVision, MathVerse and MathVerse (Vision Only). + +# 5.4.2. Ablation Study on Model Parameters Variants + +To demonstrate the effectiveness of our approach across models of different parameter sizes, we conducted a series of ablation experiments using Qwen2.5-VL-3B as a smaller base model. Specifically, we applied both Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to the Qwen2.5-VL-3B model and evaluated its performance. Table 9 shows that R1-Onevision-3B outperforms the Qwen2.5-VL-3B by $2.0\%$ on MathVision, $3.9\%$ + +on MathVerse ALL, and $4.3\%$ on MathVerse (Vision Only). + +These findings highlight the scalability and adaptability of our method across different model sizes. + +# 6. Conclusion + +In this paper, we introduce a comprehensive framework for multimodal reasoning, built upon a cross-modal formalization approach that unifies data construction, model training, and evaluation. Our framework is designed to address the inherent challenges of integrating visual and textual modalities, enabling models to reason effectively across these domains. The cross-modal reasoning pipeline at the core of this framework bridges the gap between vision and language by leveraging fine-grained alignment mechanisms and structured reasoning pathways. This framework has led to the creation of the R1-Onevision dataset, a rich resource featuring detailed step-by-step reasoning annotations designed to enhance model training and evaluation. The R1-Onevision model, trained using this framework, demonstrates strong multimodal reasoning capabilities and exhibits robust generalization across diverse tasks, from visual question answering to complex problem-solving scenarios. To further support the evaluation of multimodal reasoning, we introduce R1-Onevision-Bench, a comprehensive benchmark that assesses model performance across various dimensions of reasoning in a rigorous and structured manner. Extensive experiments validate the effectiveness of our approach, showing significant improvements over state-of-the-art open-source models. + +# Acknowledgment + +This work was supported by the National Natural Science Foundation of China (No. 62132017, No. 62302435 and No. 62402434). This work was also funded by Zhejiang Provincial Natural Science Foundation of China (No. LD24F020011 and No. LQ24F020006). + +# References + +[1] Janice Ahn, Rishu Verma, Renze Lou, Di Liu, Rui Zhang, and Wenpeng Yin. Large language models for mathematical reasoning: Progresses and challenges. arXiv preprint arXiv:2402.00157, 2024. 2 +[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 8 +[3] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021. 1 +[4] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhang-wei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 2, 6, 8 +[5] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 1 +[6] OpenCompass Contributors. Opencompass: A universal evaluation platform for foundation models, 2023. 2, 6 +[7] Alex Gu, Baptiste Rozière, Hugh Leather, Armando Solar-Lezama, Gabriel Synnaeve, and Sida I Wang. Cruxeval: A benchmark for code reasoning, understanding and execution. In International Conference on Machine Learning, 2024. 1 +[8] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 3 +[9] Jarvis Guo, Tuney Zheng, Yuelin Bai, Bo Li, Yubo Wang, King Zhu, Yizhi Li, Graham Neubig, Wenhu Chen, and Xiang Yue. Mammoth-vl: Eliciting multimodal reasoning with instruction tuning at scale. arXiv preprint arXiv:2412.05237, 2024. 2 +[10] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the MATH dataset. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2024. 1 + +[11] Jie Huang and Kevin Chen-Chuan Chang. Towards reasoning in large language models: A survey. arXiv preprint arXiv:2212.10403, 2022. 2 +[12] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 3 +[13] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning, 2016. 6 +[14] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning. In International Conference on Learning Representations Workshop Track, 2018. 3 +[15] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 2 +[16] 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. Transactions on Machine Learning Research, 2025. 3, 4 +[17] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In The Twelfth International Conference on Learning Representations, 2024. 2, 5 +[18] AI Meta. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. Meta AI Blog. Retrieved December, 20:2024, 2024. 2 +[19] OpenAI. Gpt-4o system card, 2024. 2, 6 +[20] Long Phan, Alice Gatti, Ziwen Han, Nathaniel Li, Josephina Hu, Hugh Zhang, Sean Shi, Michael Choi, Anish Agrawal, Arnav Chopra, et al. Humanity's last exam. arXiv preprint arXiv:2501.14249, 2025. 2 +[21] Aske Plaat, Annie Wong, Suzan Verberne, Joost Broekens, Niki van Stein, and Thomas Back. Reasoning with large language models, a survey. arXiv preprint arXiv:2407.11511, 2024. 2 +[22] Runqi Qiao, Qiuna Tan, Guanting Dong, Minhui Wu, Chong Sun, Xiaoshuai Song, Zhuoma GongQue, Shanglin Lei, Zhe Wei, Miaoxuan Zhang, et al. We-math: Does your large multimodal model achieve human-like mathematical reasoning? arXiv preprint arXiv:2407.01284, 2024. 2, 6 +[23] Jonathan Roberts, Mohammad Reza Taesiri, Ansh Sharma, Akash Gupta, Samuel Roberts, Ioana Croitoru, Simion-Vlad Bogolin, Jialu Tang, Florian Langer, Vyas Raina, et al. Zero-bench: An impossible visual benchmark for contemporary large multimodal models. arXiv preprint arXiv:2502.09696, 2025. 2 +[24] Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed + +Zumri, Jean Lahoud, Rao Muhammad Anwer, et al. Llamavol: Rethinking step-by-step visual reasoning in llms. arXiv preprint arXiv:2501.06186, 2025. 2 +[25] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095-95169, 2025. 2, 3, 5 +[26] 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. 2, 6, 8 +[27] Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. In Advances in Neural Information Processing Systems, 2024. 2 +[28] 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. 1 +[29] Yixuan Weng, Minjun Zhu, Fei Xia, Bin Li, Shizhu He, Shengping Liu, Bin Sun, Kang Liu, and Jun Zhao. Large language models are better reasoners with self-verification. arXiv preprint arXiv:2212.09561, 2022. 3 +[30] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, Zhenda Xie, Yu Wu, Kai Hu, Jiawei Wang, Yaofeng Sun, Yukun Li, Yishi Piao, Kang Guan, Aixin Liu, Xin Xie, Yuxiang You, Kai Dong, Xingkai Yu, Haowei Zhang, Liang Zhao, Yisong Wang, and Chong Ruan. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding, 2024. 8 +[31] Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. Llava-01: Let vision language models reason step-by-step. arXiv preprint arXiv:2411.10440, 2024. 2 +[32] Guowei Xu, Peng Jin, Hao Li, Yibing Song, Lichao Sun, and Li Yuan. Llava-cot: Let vision language models reason step-by-step, 2025. 2, 6 +[33] Huanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang, Yibo Wang, Shunyu Liu, Yingjie Wang, Yuxin Song, Haocheng Feng, Li Shen, et al. Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte carlo tree search. arXiv preprint arXiv:2412.18319, 2024. 2 +[34] 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, 36:11809-11822, 2023. 3 +[35] 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. 8 + +[36] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5317-5327, 2019. 3 +[37] Fengji Zhang, Linquan Wu, Huiyu Bai, Guancheng Lin, Xiao Li, Xiao Yu, Yue Wang, Bei Chen, and Jacky Keung. Humaneval-v: Evaluating visual understanding and reasoning abilities of large multimodal models through coding tasks. arXiv preprint arXiv:2410.12381, 2024. 2 +[38] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, Peng Gao, and Hongsheng Li. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, page 169–186, 2024. 2, 6 +[39] Yifan Zhang, Jingqin Yang, Yang Yuan, and Andrew Chi-Chih Yao. Cumulative reasoning with large language models. arXiv preprint arXiv:2308.04371, 2023. 1 +[40] Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, and Huan Zhang. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models. arXiv preprint arXiv:2411.00836, 2024. 3 \ No newline at end of file diff --git a/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/images.zip b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..063e20f9256728dab72a87471b8f5c8a33b924a8 --- /dev/null +++ b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf6aa777e4e9defb11eb96b238087bcce558251172af5e9ee59407d9aa01955f +size 711799 diff --git a/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/layout.json b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b4d86b06ad4940d1c31d95c4e0d3d0229f722c01 --- /dev/null +++ b/ICCV/2025/R1-Onevision_ Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a3ca0cbde80e65779fc1558a8d9516a1dec5e5e580e445252792ffe7afa65c0 +size 347270 diff --git a/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/e7a0ccdd-b0cc-4399-bb9b-7bd8e9beac35_content_list.json b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/e7a0ccdd-b0cc-4399-bb9b-7bd8e9beac35_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f760a49322dcc594498e1a270eccbeff0bb347db --- /dev/null +++ b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/e7a0ccdd-b0cc-4399-bb9b-7bd8e9beac35_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffb2c571a635a2a1527d74ac472cf457f64b36b461f2341ee11806948b1b641c +size 87014 diff --git a/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/e7a0ccdd-b0cc-4399-bb9b-7bd8e9beac35_model.json b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/e7a0ccdd-b0cc-4399-bb9b-7bd8e9beac35_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0f92489491a5cc34f5516ec62b2fedb13b1c6317 --- /dev/null +++ b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/e7a0ccdd-b0cc-4399-bb9b-7bd8e9beac35_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2796346c3511d94c26f6fa38fcb4de908236f712bf345509cd647ac5a47b40fe +size 109502 diff --git a/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/e7a0ccdd-b0cc-4399-bb9b-7bd8e9beac35_origin.pdf b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/e7a0ccdd-b0cc-4399-bb9b-7bd8e9beac35_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e303ef1bc0eba6c1ad04929808e04fd1aebcdc9e --- /dev/null +++ b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/e7a0ccdd-b0cc-4399-bb9b-7bd8e9beac35_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d534bcce98aa4473b0714b1042625d8bc941777e509923f6c5c5efbffae7b63 +size 800602 diff --git a/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/full.md b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..44ade1506ff84f5db045f28419edfe2579bf4a45 --- /dev/null +++ b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/full.md @@ -0,0 +1,380 @@ +# R1-VL: Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization + +Jingyi Zhang Jiaxing Huang Huanjin Yao Shunyu Liu Xikun Zhang Shijian Lu Dacheng Tao Nanyang Technological University, Singapore + +# Abstract + +Recent studies generally enhance MLLMs' reasoning capabilities via supervised fine-tuning on high-quality chain-of-thought reasoning data, which often leads models to merely imitate successful reasoning paths without understanding what the wrong reasoning paths are. In this work, we aim to enhance the MLLMs' reasoning ability beyond passively imitating positive reasoning paths. To this end, we design Step-wise Group Relative Policy Optimization (StepGRPO), a new online reinforcement learning framework that enables MLLMs to self-improve reasoning ability via simple, effective and dense step-wise rewarding. Specifically, StepGRPO introduces two novel rule-based reasoning rewards: Step-wise Reasoning Accuracy Reward (StepRAR) and Step-wise Reasoning Validity Reward (StepRVR). StepRAR rewards the reasoning paths that contain necessary intermediate reasoning steps via a soft key-step matching technique, while StepRAR rewards reasoning paths that follow a well-structured and logically consistent reasoning process through a reasoning completeness and logic evaluation strategy. With the proposed StepGRPO, we introduce R1-VL, a series of MLLMs with outstanding capabilities in step-by-step reasoning. Extensive experiments over 8 benchmarks demonstrate the superiority of our methods. Code is available at link. + +# 1. Introduction + +Multimodal large language models (MLLMs) have achieved significant progress in vision-language understanding [1, 8, 15, 18, 21, 38, 43, 51]. Recent efforts generally enhance MLLMs' reasoning capabilities by employing supervised fine-tuning (SFT) on high-quality chain-of-thought (CoT) reasoning data generated by powerful models (e.g., GPT4) [37, 44, 46, 54]. For example, Mulberry [46] introduces CoMCTS, which utilizes multiple + +![](images/045db299069a7eb8bc9928c4e2c7b9e5de383f4bb1cbcaccdacb79e165ca5aa5.jpg) +Figure 1. For MLLMs, online reinforcement learning with outcome-level reward, like in Deepseek-R1's GRPO [34], often suffers from sparse reward issues, where only a few reasoning paths can receive positive/high rewards during training, ultimately leading to poor exploration efficiency and unstable learning process. To tackle this, we propose a novel online reinforcement learning framework that incorporates step-wise reasoning rewards in addition to outcome-level rewards, encouraging MLLMs to iteratively refine their reasoning with dense rewards and resulting in a more stable training process and improved reasoning capability. The experiments are conducted on Qwen2-VL-7b over MathVista. + +models to collectively search and identify effective reasoning paths, followed by SFT on the collected reasoning data. However, SFT approaches focus solely on positive reasoning paths (i.e., those leading to correct answers), while the negative reasoning paths are largely neglected. This limitation may cause the model to merely imitate successful reasoning paths without understanding what the flawed and wrong reasoning paths are. + +In this work, we aim to enhance the MLLMs' reasoning ability beyond passively imitating positive reasoning paths. Recent advancements in NLP, such as Deepseek-R1 [13] and Kimi-K1.5 [36], have shown great potential in incentivizing the reasoning capability of LLMs via actively selfexploring. The core design of these advances (e.g., GRPO in Deepseek-R1) lies in online reinforcement learning without the need for reward models, which encourages an LLM + +to generate a group of reasoning paths and iteratively refine its reasoning process by rewarding the generated reasoning paths based on a rule-based reward function. Typically, an outcome-level reward strategy is used: reasoning paths leading to correct answers receive higher rewards, while those leading to incorrect answers receive lower ones. + +An intuitive idea is to directly apply these simple and effective LLM online reinforcement learning methods for MLLMs. However, relying solely on outcome-level rewards, like in Deepseek-R1's GRPO, often suffers from sparse reward issues on MLLM reasoning learning, resulting in suboptimal performance. Specifically, most MLLMs, especially smaller ones, exhibit very limited capability in long-chain reasoning accuracy and validity, whereas only a few MLLM-generated reasoning paths can receive positive/high rewards. This lack of positive reward signals reduces exploration efficiency and leads to an unstable learning process, as illustrated in Fig. 1. + +We propose to tackle this sparse reward issue by introducing dense step-wise reasoning rewards in addition to sparse outcome-level rewards. To this end, we design Stepwise Group Relative Policy Optimization (StepGRPO), a new online reinforcement learning framework that enables MLLMs to self-improve reasoning ability via simple, effective and dense step-wise rewarding while using no additional process reward models. Specifically, StepGRPO introduces two novel rule-based reasoning reward mechanisms: Step-wise Reasoning Accuracy Reward (StepRAR) and Step-wise Reasoning Validity Reward (StepRVR). + +StepRAR rewards the reasoning path using a soft key-step matching technique that evaluates whether the reasoning path contains key intermediate reasoning steps (i.e., the necessary steps to reach the correct final solution). StepRVR rewards the reasoning path based on a reasoning completeness and logic evaluation method, which assesses whether the reasoning process is well-structured and logically consistent. In this way, StepRAR and StepRVR help mitigate the sparse reward issue by providing informative rewards, even when the reasoning path does not produce the correct final answer – as long as it includes key intermediate reasoning steps or follows a structured and logical reasoning process. With StepRAR and StepRVR, StepGRPO takes the average step-wise reasoning rewards of a group of sampled reasoning paths as a baseline to estimate the advantage for policy optimization. Using the proposed StepGRPO, we develop R1-VL, a series of MLLMs with R1-like step-by-step reasoning capabilities. + +The proposed StepGRPO offers two key advantages. 1) Effectiveness. StepGRPO introduces two step-wise reasoning reward mechanisms with group relative optimization, which provide rich and fine-grained step-wise reasoning rewards along the whole reasoning trajectory beyond the final answer. This mitigates the sparse reward issue and encour + +ages more structured, logically consistent reasoning trajectories. 2) Efficiency. StepGRPO achieves step-wise reasoning rewarding in a rule-based manner, which provides step-wise reasoning rewards while eliminating the need of process reward models. This significantly reduces computational overhead while maintaining fine-grained step-wise supervisions. + +The main contributions of this work are threefold. First, we propose StepGRPO, a new online reinforcement learning framework that enables MLLMs to self-improve reasoning ability via a simple, effective and dense step-wise rewarding. Second, we design two novel rule-based reasoning reward mechanisms, i.e., step-wise reasoning accuracy reward and step-wise reasoning validity reward, which effectively mitigate the sparse reward issue for MLLMs without the need of process reward models. Third, with the proposed StepGRPO, we develop R1-VL, a series MLLMs that have superior reasoning capabilities. Forth, extensive experiments over multiple benchmarks show that R1-VL achieves superior performance compared with state-of-the-art MLLMs. + +# 2. Related Work + +# 2.1. Multimodal Large Language Model + +Multimodal Large Language Models (MLLMs) [1, 8, 15, 18, 21, 38, 43, 51, 52] have shown remarkable advancements across a wide range of vision-language understanding tasks, demonstrating their capabilities in comprehending and analyzing visual contents across various application domains. Early research on MLLMs primarily focuses on text generation based on text prompts and input multiple modalities such as images [20, 21], videos [9, 35]. Recent advancements further enhance the capabilities of MLLMs from various aspects. For example, recent models [25, 42] incorporate multimodal inputs and outputs such as video, audio, and point cloud inputs beyond text and images. In addition, some efforts attempt to adapt MLLMs for domain-specific tasks, such as medical image understanding [17, 19, 55] and document analysis [22, 49]. In this work, we focus on enhancing the reasoning ability of MLLMs in tackling complex reasoning tasks and introduce R1-VL, a series of MLLMs that have superior reasoning capability. + +# 2.2. MLLM Reasoning + +Inspired by the advances in NLP that show great potential in learning to reason and tackling complex language tasks [29], recent studies attempt to enhance the reasoning capability of MLLM. Generally, current MLLM reasoning methods improve the reasoning capability of MLLM by generating high-quality chain-of-thoughts (CoT) data using powerful model (e.g., GPT-4) and performing supervised + +fine-tuning with the collected data [10, 37, 44, 46, 54]. For example, Mulberry [46] introduces Collective Monte Carlo Tree Search (MCTS) into MLLM and proposes CoMCTS which leverages complementary knowledge from multiple models to collaboratively search and identify effective reasoning paths. In addition, recent works [14, 27, 30, 47] attempt to explore online reinforcement learning to improve the MLLMs' reasoning ability. Different from these works, we design StepGRPO that enables MLLM to self-improve the reasoning ability with step-wise reward signals. + +# 2.3. Reinforcement Learning + +Reinforcement Learning (RL) [16] is a fundamental approach in machine learning, where an agent learns to interact with an environment by taking actions, receiving rewards, and updating its policy to maximize the long-term return. With the rise of large language models (LLMs) [4, 28, 31], Reinforcement Learning with Human Feedback (RLHF) [3] has emerged as a key technique for fine-tuning models using human preference data. RLHF leverages algorithms like Proximal Policy Optimization (PPO) [33] and Direct Preference Optimization (DPO) [32] to guide model behavior for improving the alignment, coherence and helpfulness in response generation. + +Recently, RL is increasingly adopted to enhance LLMs' reasoning capabilities [5, 7, 13, 24, 36, 50], especially for mathematical problem solving. The core is to adopt an appropriate reward function or model that evaluates and reinforces high-quality reasoning paths while penalizing low-quality ones, guiding the model's optimization towards more structured and coherent reasoning trajectories using the RL algorithm. For example, ReST-MCTS* [50] trains a process reward model (PRM) for determining the correctness of each reasoning step within reasoning paths. Recent methods have found that using a simple outcome-level rule-based reward function (i.e., the reasoning trajectories leading to correct answer are rewarded with higher score) can already provide an effective and reliable reward signal during the RL process [13, 24, 36]. For example, DeepSeek-R1 [13] demonstrates that group relative policy optimization (GRPO) [34] with outcome-level reward effectively enhances the reasoning capability of LLMs. In this work, we aim for improving the reasoning capability of MLLMs through reinforcement learning and propose StepGRPO, which effectively tackles the sparse reward issue in MLLMs, leading to stable training process and better reasoning capability. + +# 3. Method + +This section first presents the task formulation, and then introduces the proposed Step-wise Group Relative Policy Optimization (StepGRPO). More details to be elaborated in the ensuing subsections. + +# 3.1. Task Formulation + +In this paper, we consider a pre-trained MLLM and denote it as a policy model $\pi_{\theta}$ . Given a multimodal question $Q$ consisting of an image and a textual task instruction, i.e., $Q = \{\text{text}, \text{image}\}$ , the policy model $\pi$ generates response $\mathbf{c}$ with a step-by-step reasoning trajectory. Generally, this process can be formulated as a sequence of next token prediction actions, i.e., $\mathbf{c} = (a_1, a_2, \dots, a_t, \dots, a_T)$ , where each action $a_t$ is sampled from the policy model $\pi_{\theta}$ and $T$ represents the maximum sequence length. After each action, the new state $s_{t+1}$ is determined by updating the current state $s_t$ with the newly generated action $a_t$ , i.e., $s_{t+1} = (s_t, a_t)$ , $1 \leq t \leq T$ . + +Considering this formulation, the objective of our task is to optimize the policy model $\pi_{\theta}$ such that it can select better actions based on the previous states, thereby improving reasoning quality. In the context of reinforcement learning (RL), the policy model is generally optimized by maximizing the cumulative reward, where the reward for taking action $a_{t}$ at state $s_t$ is denoted as $r(s_t,a_t,s_{t + 1})$ . Following prior studies [46], we define an action in this paper as generating a reasoning step, which consists of one or more sentences containing multiple word tokens. + +# 3.2. Step-wise Group Relative Policy Optimization + +We propose Step-wise Group Relative Policy Optimization (StepGRPO), a novel online reinforcement fine-tuning framework that mitigates the sparse reward issue for MLLMs and encourages self-improvement in reasoning ability through simple, effective and dense step-wise reward mechanisms. As illustrated in Fig. 2, StepGRPO consists of two phases: (1) a policy warm-up phase and (2) a step-wise online policy optimization phase. The overall algorithm is shown in Algorithm 1. + +# 3.2.1. Policy Warm-up + +This phase equips the policy model with fundamental reasoning capabilities, ensuring it can generate proper stepwise reasoning paths before reinforcement learning. During the warm-up phase, the policy model is fine-tuned using a multimodal dataset $D_{s}$ with Chain-of-Thought (CoT) reasoning path, where each data consists of a multimodal question $Q$ and a step-by-step reasoning path $\tau$ , i.e., $D_{s} = \{Q^{n}, \tau^{n}\}_{n=1}^{N}$ : + +$$ +\mathcal {L} _ {\text {w a r m - u p}} = - \mathbb {E} _ {\tau \sim D _ {s}} [ \sum_ {t = 1} ^ {T} \log (\pi_ {\theta} (a _ {t} | s _ {t})) ]. \tag {1} +$$ + +# 3.2.2. Step-wise Online Policy Optimization + +This phase enables MLLMs to self-improve their reasoning ability via online reinforcement learning, mitigating the sparse reward issue through step-wise reasoning rewards. As illustrated in Fig. 2, for each question $Q \in D_{s}$ , + +![](images/63b7d9c10823788b7fca5a6efb71f874524cb97c2b342c5721d0f68e6864f343.jpg) +Question: In the given diagram, triangle ABC has AD as its median and point E is the midpoint of AD. If the area of triangle ABC is 12, what is the area of triangle ABE? + +![](images/35a83766051badbf8ccf7672900d5d5649543e4300c4b88813f9239a33b4893f.jpg) +Answer: Step 1: Since AD is a median, it divides triangle ABC into two equal areas: ABD and ACD. Step 2: Segment AE is half of AD, splitting triangle ABD into two triangles of equal area: ABE and BED. Step 3: The area of triangle ABD is half of triangle ABC, which is $\frac{\text{frac}}{12} \times \frac{1}{2} = 6$ . Step 4: Since E is the midpoint of AD, triangle ABE is half of triangle ABD. Therefore, the area of triangle ABE is $\frac{\text{frac}}{6} \times \frac{1}{2} = 3$ . The final answer is 3. +(a) Step-wise Reasoning Accuracy Reward +Figure 2. Overview of the proposed StepGRPO. StepGRPO consists of two phases: a policy warm-up phase and a step-wise online policy optimization phase. After the warm-up, the policy model $\pi_{\theta}$ generates a group of reasoning paths $\{\mathbf{c}^i\}_{i=1}^M$ and assigns step-wise rewards using two proposed mechanisms: Step-wise Reasoning Accuracy Reward (StepRAR) and Step-wise Reasoning Validity Reward (StepRVR). StepRAR rewards reasoning paths that contain key intermediate steps, identified using a soft key-step matching technique. StepRVR rewards reasoning paths based on completeness and logical consistency, ensuring they are well-structured. StepGRPO then estimates the advantage $\hat{A}$ for policy optimization by using the average step-wise reasoning reward of a group of sampled reasoning paths as a baseline. Examples for StepRAR and StepRVR are illustrated in (a) and (b), respectively. + +# Pre-extracted key steps with Augmentations: + +1. AD is a median; median is $AD$ +2. equal area; ... +3. AE is half of AD; $AE = 1 / 2AD$ +4. frac{12}{2} {2} = 6; $\underline{12 / 2} = 6,\dots$ +5. E is the midpoint; .. +6. frac{6}{2} = 3; 6/2 = 3. ... + +# Soft key-step matching : + +Description: The image shows ...; #Rationale: The question asks for the area...; #Step1: ... we find AD is a median of ...; #Step2: ... AE splits triangle ABD ...; #Step3: ... The area of triangle ABD is $12/2 = 6$ , ..., and the area of triangle ABE is frac{6}{2} = 3. #The final answer is: 3. Step-wise Matching score: 3/6 + +# (b) Step-wise Reasoning Validity Reward + +Description $\rightarrow$ #Rationale $\rightarrow$ # Step1 $\rightarrow$ ... $\rightarrow$ #Step $N\rightarrow$ #Answer. + +i. Reasoning completeness + +Description $\rightarrow$ #Rationale $\rightarrow$ #Answer. Missing reasoning steps + +Description $\rightarrow$ # Step1 $\rightarrow$ ... $\rightarrow$ #Step $N\rightarrow$ #Answer. Missing rationale + +ii. Reasoning logic + +Description $\rightarrow$ #Rationale $\rightarrow$ #Answer $\rightarrow$ #Step1... $\rightarrow$ #Step N. +#Description $\rightarrow$ #Step3 $\rightarrow$ #Rationale $\rightarrow$ ... $\rightarrow$ #Step 1 $\rightarrow$ #Answer + +the policy model $\pi_{\theta}$ first generates a group of $M$ reasoning trajectories via multiple rollouts, i.e., $\{\mathbf{c}^i\}_{i=1}^M$ , where $\mathbf{c}^i = (a_1^i, a_2^i, \ldots, a_t^i, \ldots, a_T^i)$ . After obtaining a group of $M$ reasoning trajectories, we employ our proposed step-wise reasoning rewards to evaluate and reward each generated reasoning trajectory. Specifically, we introduce two types of rule-based step-wise rewards, i.e., step-wise reasoning accuracy (StepRAR) reward and step-wise reasoning validity reward (StepRVR). + +Step-wise reasoning accuracy reward (StepRAR) reduces the effect of learning from sparse reward by additionally rewarding reasoning paths that contain correct intermediate reasoning steps contributing to the final solution. Specifically, for each question $Q$ , we pre-extract a set of key reasoning steps $\mathbf{v} = \{v_{1}, v_{2}, \ldots\}$ from the corresponding reasoning path $\tau$ in dataset $D_{s}$ . We define key steps as the essential variables and equations that directly contribute to the final solution, and prompt GPT-4 to extract several key steps from the reasoning path for each question. To ensure efficient reward assignment, we refine the extracted steps by removing redundant content and retaining only the core few words necessary for reasoning. Furthermore, we + +augment each extracted key step into multiple equivalent formats to allow more flexible and accurate matching, preventing missed matches due to math-related formatting differences. For example, a mathematical expression such as " $\frac{6}{3} = 2$ " is augmented to " $6 / 3 = 2$ " or "6 divided by 3 equals 2". + +With the extracted key reasoning steps $\mathbf{v} = \{v_{1}, v_{2}, \ldots\}$ and such soft marching mechanism, we calculate a match score for each generated reasoning path based on the ratio of matched key steps, i.e., $k^{i} = |\mathbf{v}_{\text{match}}| / |\mathbf{v}|$ . Then, StepRAR for $1 \leq t \leq T$ is defined as: + +$$ +r _ {a u c} ^ {i} \left(s _ {t}, a _ {t}, s _ {t + 1}\right) = \left\{ \begin{array}{l l} 1 + \alpha k ^ {i}, & \operatorname {a n s} \left(s _ {t + 1}\right) = y, \\ \alpha k ^ {i}, & \operatorname {a n s} \left(s _ {t + 1}\right) \neq \text {n u l l}, \neq y, \\ 0, & \operatorname {a n s} \left(s _ {t + 1}\right) = \text {n u l l}, \end{array} \right. \tag {2} +$$ + +where $y$ is the ground-truth answer extracted from CoT reasoning path. + +By leveraging pre-extracted key reasoning steps, StepRAR efficiently provides additional supervision with a simple soft matching mechanism, ensuring the model learns meaningful reasoning processes instead of guessing + +answers randomly. + +Step-wise reasoning validity reward (StepRVR) aims for ensuring the generated paths adhere to a logically structured and coherent progression beyond the reasoning accuracy. Prior studies [44, 46] have demonstrated structural reasoning, such as problem decomposition and progressive reasoning, facilitates more accurate and interpretable reasoning processes, as they encourage models to break down complex problems into multiple intermediate steps rather than direct answer generation. + +Inspired by these findings, we incorporate step-wise reasoning validity to reinforce well-organized reasoning paths that follow an expected logical flow. Specifically, we define StepRVR using two key criteria: reasoning completeness $\delta^c$ and reasoning logic $\delta^l$ . Reasoning completeness requires the response to include three essential components, i.e., a background analysis involving image description and rationale analysis to establish context, a step-by-step reasoning process and a final answer. In addition to the reasoning completeness, reasoning logic ensures the reasoning path to follow a logical progression, where the background analysis must come before solution steps and the final answer should only appear after reasoning steps are complete. + +With these two criteria, we define StepRVR as + +$$ +r _ {v a l} ^ {i} \left(s _ {t}, a _ {t}, s _ {t + 1}\right) = \left\{ \begin{array}{l l} 1, & \mathbb {I} \left(\delta^ {c} \left(s _ {t + 1}\right)\right) \cdot \mathbb {I} \left(\delta^ {l} \left(s _ {t + 1}\right)\right) = 1, \\ 0, & \text {o t h e r w i s e ,} \end{array} \right. \tag {3} +$$ + +where the reasoning trajectory is rewarded only if it satisfies both completeness and logical coherence. By enforcing this, StepRVR helps the model produce structured, interpretable and logically sound reasoning trajectories, enhancing both the quality and reliability of generated responses. + +Optimization with the step-wise rewards. After obtaining two types of step-wise rewards, we compute the overall reward for each reasoning path as $r^i = r_{auc}^i + r_{val}^i$ , and repeatedly compute the rewards for all generated reasoning paths, i.e., $\{r^1, r^2, \dots, r^M\}$ . + +To estimate the advantage of each reasoning trajectory, we normalize its reward relative to the group as follow: + +$$ +\hat {A} ^ {i} = \frac {r ^ {i} - \operatorname {m e a n} \left(\left\{r ^ {1} , r ^ {2} , \dots , r ^ {M} \right\}\right)}{\operatorname {s t d} \left(\left\{r ^ {1} , r ^ {2} , \dots , r ^ {M} \right\}\right)}, \tag {4} +$$ + +where the mean group reward serves as the baseline, and $\hat{A}_i$ measures how much better or worse $r_i$ is compared to other reasoning trajectories within the group. Following this, we optimize the policy model with the loss defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {S t e p R L}} = - \underset {Q \in D _ {s}} {\mathbb {E}} \left[ \frac {1}{M} \sum_ {i = 1} ^ {M} \left(\frac {\pi_ {\theta} \left(\mathbf {c} ^ {i} \mid Q\right)}{\left[ \pi_ {\theta} \left(\mathbf {c} ^ {i} \mid Q\right) \right] _ {\text {n o g r a d}}} \hat {A} ^ {i} \right. \right. \tag {5} \\ - \beta D _ {K L} \left(\pi_ {\theta} | | \pi_ {r e f}\right) ], \\ \end{array} +$$ + +Algorithm 1 Step-wise Group Relative Policy Optimization +Input: Policy model $\pi_{\theta}$ initialized by a pre-trained +MLLM; a multimodal dataset $D_{s} = \{Q^{n},\tau^{n}\}_{n = 1}^{N}$ +Output: Trained policy model $\pi_{\theta}$ +Policy warm-up: +for iter $= 1$ to $N$ do Sample $\{Q,\tau \} \in D_s$ Optimize policy model $\pi_{\theta}$ by Eq. 1 +end for +Step-wise online policy optimization: +for iter $= 1$ to $N$ do Sample $\{Q,\tau \} \in D_s$ Generate a group of reasoning paths $\{\mathbf{c}^i\}_{i = 1}^M\sim \pi_\theta$ Obtain step-wise rewards $\{r^i\}_{i = 1}^M$ by Eqs. 2-3 Obtain relative advantages $\{\hat{A}^i\}_{i = 1}^M$ by Eq. 4 Optimize policy model $\pi_{\theta}$ by Eqs. 5-6 +end for +return policy model $\pi_{\theta}$ + +where KL divergence is adopted to regularize the policy model, preventing excessive deviation from the reference model. The reference model is typically initialized as the same model as the policy model but remains frozen during RL training. The KL divergence between the policy model and the reference model is estimated as in [34]: + +$$ +D _ {K L} \left(\pi_ {\theta} \right\| \pi_ {r e f} = \frac {\pi_ {r e f} \left(\mathbf {c} ^ {i} \mid Q\right)}{\pi_ {\theta} \left(\mathbf {c} ^ {i} \mid Q\right)} - \log \frac {\pi_ {r e f} \left(\mathbf {c} ^ {i} \mid Q\right)}{\pi_ {\theta} \left(\mathbf {c} ^ {i} \mid Q\right)} - 1. \tag {6} +$$ + +# 4. Experiment + +This section presents experiments including datasets and implementation details, main experimental results, ablation studies and discussion, respectively. More details are to be described in the ensuing subsections. + +# 4.1. Datasets + +For policy warm-up, we adopt Mulberry-260k [46] for supervised fine-tuning. For step-wise online policy optimization, we randomly sample 10K data from Mulberry-260k as our training data. For evaluation, we adopt 8 widely-used multimodal benchmarks for comprehensively evaluating our proposed StepGRPO, including MathVista [23], MMStar [6], Math-Vision [40], ChartQA [26], DynaMath [56], HallusionBench [12], MathVerse [53], MME [11] and MM-Reason [45]. These multimodal benchmarks cover a wide range of tasks from mathematical reasoning, chart understanding, visual hallucination and general visual understanding. + +# 4.2. Implementation Details + +Our proposed StepGRPO is generally applicable to different MLLMs. In our experiments, we adopt two state-of-the-art + +
MethodMathVistaMMStarMath-VChartQADynaMathHallBenchMathVerseMMEsumMMReasonAVG
Closed-Source Model
GPT-4o [15]63.863.930.385.763.755.039.4232921.156.2
Claude-3.5 Sonnet [1]67.762.2-90.864.855.0-1920--
Open-Source Model
Cambrain-1-8B [38]49.0--73.3------
MM-1.5-7B [51]47.6--78.6---1861--
Idefics3-LLaMA3-8B [18]58.455.9-74.8---1937--
InternVL2-8B [8]58.361.5-83.339.7--2210--
MiniCPM-V-2.6-8B [48]60.657.5---48.1-2348--
DeepSeek-VL2-MOE-4.5B [43]62.861.3-86.0---225311.5-
Reasoning Model
LLaVA-CoT-11B [44]54.857.6---47.8----
LLaVA-Reasoner-8B [54]50.654.0-83.0------
Insight-V-8B [10]49.857.4-77.4---2069--
Mulberry-7B [46]63.161.3-83.945.154.1-239611.8-
LlamaV-o1-11B [37]54.459.4---63.5----
Vision-R1-7B [14]73.5-----52.4---
LMM-R1 [30]63.258.026.3---41.5---
R1-ShareVL-7B [47]75.467.029.5---52.8---
Qwen2-VL-2B [41]43.048.012.473.524.941.719.718727.737.5
R1-VL-2B (Ours)52.149.817.175.229.444.026.220488.341.6
Qwen2-VL-7B [41]58.260.716.383.042.150.632.5232711.948.7
R1-VL-7B (Ours)63.560.024.783.945.254.740.0237612.552.1
Qwen2.5-VL-7B [2]68.263.925.187.353.252.149.2234717.355.5
R1-VL-7B* (Ours)74.366.228.287.756.557.252.2239517.958.4
+ +Table 1. Main experimental results. To comprehensively examine the proposed StepGRPO, we conduct extensive experiments with two baseline models on eight benchmarks, and compare StepGRPO with various state-of-the-art MLLMs.* indicates that the model is trained using Qwen2.5-VL-7B as the base model with the data from [47]. + +
Warm-upStep-wise reasoning rewardsMathVista
StepRARStepRVR
✓58.2
61.2
✓✓62.4
✓✓61.9
✓✓✓63.5
+ +Table 2. Ablation study of StepGRPO over Qwen2-VL-7B. + +open-source MLLMs, i.e., Qwen2-VL-2B and Qwen2-VL-7B [41]. For the policy warm-up phase, we set the training batch size to 128. Following prior work [46], we use a learning rate of $1\mathrm{e}^{-5}$ for Qwen2-VL-2B and $5\mathrm{e}^{-6}$ for Qwen2-VL-7B, respectively. + +For the step-wise online policy optimization phase, we perform 4 rollouts per question $(M = 4)$ and set the sampling temperature to 1.2 to encourage diverse reasoning paths. The maximum sequence length is set to $L = 1024$ , ensuring that the model can generate complete reasoning paths. Both the policy model and reference model are ini + +tialized from the model after the warm-up, with the reference model frozen during RL training. The policy model's learning rate is $1\mathrm{e}^{-6}$ , and we set the batch size to 4. We set the coefficient of match score $\alpha$ to 0.1 to balance its effect. Following [39], the KL divergence coefficient $\beta$ in Eq. 5 is set to 0.04 by default. All experiments are conducted on 4 H100-80GB GPUs. + +# 4.3. Main Experimental Results + +We conduct a comprehensive evaluation of R1-VL across eight widely used benchmarks, comparing it with various state-of-the-art MLLMs, as shown in Table 1. + +We first compare R1-VL with its baseline models, Qwen2-VL-2B and Qwen2-VL-7B. The baseline models exhibit limited reasoning capability, leading to very few reasoning paths receiving rewards, which negatively impacts the reasoning capability. In contrast, R1-VL with our proposed StepGRPO consistently improves the baseline models by significant margins, achieving $4.6\%$ improvement over Qwen2-VL-2B and $3.8\%$ over Qwen2-VL-7B. This improvement is largely attributed to that StepGRPO introduces step-wise reasoning accuracy and validity rewards, + +
Number of generations M per question
Method23456
R1-VL-7B62.562.863.563.263.7
+ +which provide rich and informative supervision at each reasoning step, effectively mitigating the sparse reward issue for MLLMs. + +In addition, we compare R1-VL with existing state-of-the-art reasoning MLLMs. As shown in Table 1, R1-VL achieves better performance on most benchmarks, particularly in mathematical reasoning tasks. For example, R1-VL-7B surpasses Mulberry-7B and LlamaV-o1-11B by $0.6\%$ and $9.3\%$ respectively on the reasoning-intensive benchmark MathVista. Notably, R1-VL-2B even outperforms larger MLLMs. For instance, R1-VL-2B largely outperforms LLaVA-Reasoner-8B and LLaVA-CoT-11B by $13.1\%$ and $9.3\%$ on MathVista, respectively. This superior performance demonstrates that StepGRPO effectively enhances MLLMs' reasoning abilities by encouraging self-improvement via step-wise online reinforcement learning, rather than merely imitating positive reasoning paths. + +Additionally, we benchmark R1-VL against general MLLMs, including closed-source models such as GPT-4o and Claude-3.5 Sonnet, as well as open-source models like Cambrain-1-8B and DeepSeek-VL2-MOE-4.5B. We observe that R1-VL outperforms most open-source MLLMs and achieves competitive results against closed-source models. For example, R1-VL-7B achieves 63.7 accuracy on MathVista, closely matching GPT-4o's accuracy of 63.8. These results further validate StepGRPO's effectiveness in enhancing the reasoning capabilities of MLLMs. + +# 4.4. Ablation Study + +We conduct ablation studies for StepGRPO on Qwen2-VL-7B over MathVista benchmark for examining the effect of step-wise reasoning rewards including step-wise reasoning accuracy reward (StepRAR) and step-wise reasoning validity reward (StepRVR), as well as the role of the warm-up phase. As shown in Table 2, involving a warm-up stage improves baseline model to $61.2\%$ , allowing the model to learn basic reasoning knowledge before reinforcement learning. In addition, including either StepRAR or StepRVR into the online reinforcement learning process outperforms the model with warm-up by large margins, demonstrating that both two types of step-wise rewards contribute to enhancing step-by-step reasoning capabilities. The best performance (i.e., $63.7\%$ ) is achieved when both StepRAR and StepRVR are applied together. This shows that StepGRPO effectively improves complex + +Table 3. Parameter analysis of $M$ . The experiments are conducted on Qwen2-VL-7B over MathVista. + +
MethodMathVista
Warm-up61.7
Warm-up + Outcome-level reward62.3
Warm-up + Step-wise reward (Ours)63.5
+ +Table 4. Effectiveness of the step-wise reasoning rewards. The experiments are conducted on Qwen2-VL-7B over MathVista. + +![](images/2e1796fe5e6bd68fae3af1f0b2f358055373843d2e6690a753256a884f36a39f.jpg) +Figure 3. Comparison between StepGRPO and SFT. The experiments are conducted on Qwen2-VL-7B over MathVista. + +reasoning tasks by reinforcing both the correctness of intermediate steps and the overall logical structure of the reasoning process. + +# 4.5. Discussion + +Parameter analysis. We conduct the parameter analysis on the number of generations $M$ in a group with Qwen2-VL7B over benchmark MathVista, analyzing its impact on reasoning performance. As described in Section 3, $M$ controls the number of generated reasoning trajectories per question during the RL phase. Table 3 shows that a larger $M$ generally leads to better performance. This is because, in group relative optimization, the baseline reward is estimated as the average reward of all generated reasoning paths. A larger $M$ results in a more stable and accurate baseline estimation, whereas a small $M$ may lead to high variance in baseline estimation, making the optimization process less reliable. However, increasing $M$ also introduces higher computational costs. Therefore, we set $M = 4$ as the default to balance performance and computational efficiency. + +Effectiveness of the step-wise reward. Our proposed stepwise reward mechanism plays a crucial role in mitigating the sparse reward issue by providing fine-grained supervision at each reasoning step. To further validate its effectiveness, we conduct an experiment comparing outcome-level reward against our step-wise reward. Specifically, we evaluate three settings: (1) Warm-up only; (2) Warm-up + Outcome-level Reward, where the model is optimized with outcome-level reward; and (3) Warm-up + Step-wise Re + +![](images/47edfcbc925bdd982646d33fc08105899593eafceb9024b31c2f10007f61ac98.jpg) +Question: Subtract all red things. +Subtract all tiny matte balls. How many objects are left? +Correct answer: 5 + +Qwen2-VL-7B: + +First, identify the red objects in the image. There is 1 red object, which is a small red sphere. Incorrect reasoning + +Then, subtract the red object from the total number of objects. There are 7 objects in total and subtracting the 1 red object leaves 6 objects. Identify the tiny matte balls in the image. There are 2 tiny matte balls, which are the small green cylinder and the cyan sphere. Subtract the tiny matte balls from the remaining objects. + +After subtracting the red object, there are 6 objects left. Subtracting the 2 tiny matte balls leaves 4 objects. Answer: 4 + +# Mulberry-7B: + +Image Description: The image shows several 3D objects on a flat surface. There are two red objects: a small red sphere and a larger green cylinder. Other objects include a yellow cube, a green cylinder, a blue cylinder, a purple cylinder, and a teal sphere. + +Rationales: The question asks to subtract all red things and all tiny matte balls. We need to identify these objects in the image and count the remaining ones. + +Let's think step by step. + +Step 1: Identify and count all red objects. There is one red sphere. Incorrect reasoning Step 2: Identify and count all tiny matte balls. There is one teal sphere, which is not red, so it is not subtracted. + +Step 3: Subtract the red objects and tiny matte balls from the total number of objects: $7 - 1 = 6$ + +The final answer is: 6 + +# StepGRPO (Ours): + +Image Description: + +The image shows a collection of 3D geometric objects on a flat surface. There are various colored shapes, including a red sphere, a yellow cube, a green cylinder, a blue cylinder, a purple cylinder, and a cyan sphere. + +Rationales: + +The question asks to subtract all red things and all tiny matte balls, then determine how many objects are left. This involves identifying and counting specific objects based on their color and size. + +Let's think step by step. + +Step 1: + +Identify and count all objects in the image. There are seven objects: a yellow cube, a green cylinder, a blue cylinder, a purple cylinder, a red sphere, and a cyan sphere. + +Step 2: + +Subtract all red things. There is one red sphere. + +![](images/427ae8a94d1d79342341c97de9ce9f2ebbacb58f150c21c9ed38e7fc5d3b756b.jpg) + +Step 3: + +Subtract all tiny matte balls from the remaining objects. There is one cyan sphere, which is matte and small. + +Step 4: + +Count the remaining objects after removing the red sphere and the cyan sphere. + +The final answer is: 5 + +![](images/366b9a391d6df3169d49060204f2bcb3a08049f149364a189b7477110792c573.jpg) +Figure 4. Qualitative comparison. + +ward, where the model is optimized with our proposed stepwise reasoning reward. As shown in Table 4, both outcome-level reward and our step-wise reward improve the warm-up model's performance, while our step-wise reward achieves better performance. This further demonstrates that stepwise rewards are more effective in enhancing MLLMs' reasoning capabilities, as they provide more fine-grained supervision and largely mitigate the sparse reward issue. + +Comparison to supervised fine-tuning (SFT). As discussed before, StepGRPO encourages MLLM to self-improve the reasoning ability with step-wise reward signals rather than merely imitating the successful reasoning paths. Here, we conduct experiments to further compare StepGRPO with SFT. Specifically, we start with the model after the warm-up and conduct the experiments with Qwen2-VL-7B over MathVista. As shown in Fig. 3, under the same number of training steps, StepGRPO consistently outperforms SFT, demonstrating the effectiveness of step-wise reinforcement learning. This is largely attributed to StepGRPO's ability to refine reasoning trajectories through self-exploration and reward-guided optimization, rather than solely relying on passive imitation of reasoning paths. By leveraging step-wise reasoning rewards, StepGRPO provides more rich and informative supervision, leading to better reasoning processes compared to SFT. + +Qualitative comparison. We provide qualitative comparison of Qwen2VL-7B, Mulberry-7B and our R1-VL-7B. As shown in Fig. 4, Qwen2-VL-7B generates relatively short responses, lacking a thorough reasoning process. While Mulberry-7B generates detailed reasoning paths, its intermediate steps contain errors, leading to incorrect final an + +svers. In contrast, R1-VL-7B enables more accurate step-by-step reasoning process. + +We provide more discussions, experimental results and qualitative analysis in the appendix. + +# 5. Conclusion + +This paper presents StepGRPO, a new online reinforcement learning framework that enables MLLMs to self-improve reasoning ability via simple, effective and dense step-wise reward mechanism. Specifically, StepGRPO introduces two rule-based reasoning reward mechanisms, i.e., Step-wise Reasoning Accuracy Reward that rewards the intermediate reasoning steps based on a soft key-step matching technique and Step-wise Reasoning Validity Reward that rewards the reasoning path's reasoning structure and logical consistency though a reasoning completeness and logic evaluation method. In this way, StepGRPO enables to effectively mitigate the sparse reward issue for MLLMs without the need of process reward models and encourages more structured and logically consistent reasoning process. With the proposed StepGRPO, we develop R1-VL, a series of MLLMs with superior reasoning capability. Extensive experiments over eight benchmarks demonstrate the superiority of the proposed StepGRPO compared with the state-of-the-art MLLMs. + +Acknowledgement. This research is supported by the RIE2025 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) (Award I2301E0026), administered by A\*STAR, as well as supported by Alibaba Group and NTU Singapore through Alibaba-NTU Global e-Sustainability CorpLab (ANGEL). + +# References + +[1] Anthropic. Claude 3.5 sonnet, 2024. 1, 2, 6 +[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 6 +[3] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3 +[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. 3 +[5] Guoxin Chen, Minpeng Liao, Chengxi Li, and Kai Fan. Step-level value preference optimization for mathematical reasoning. arXiv preprint arXiv:2406.10858, 2024. 3 +[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.5 +[7] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 3 +[8] 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. arXiv preprint arXiv:2404.16821, 2024. 1, 2, 6 +[9] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2 +[10] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. arXiv preprint arXiv:2411.14432, 2024. 3, 6 +[11] 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. 5 +[12] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. arXiv preprint arXiv:2310.14566, 2023. 5 +[13] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint + +arXiv:2501.12948,2025.1,3 +[14] Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749, 2025. 3, 6 +[15] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 2, 6 +[16] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. Journal of artificial intelligence research, 4:237-285, 1996. 3 +[17] Xiang Lan, Feng Wu, Kai He, Qinghao Zhao, Shenda Hong, and Mengling Feng. Gem: Empowering mllm for grounded ecg understanding with time series and images. arXiv preprint arXiv:2503.06073, 2025. 2 +[18] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. In Workshop on Responsibly Building the Next Generation of Multimodal Foundational Models, 2024. 1, 2, 6 +[19] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890, 2023. 2 +[20] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, January 2024. 2 +[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2 +[22] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 2 +[23] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 5 +[24] Trung Quoc Luong, Xinbo Zhang, Zhanming Jie, Peng Sun, Xiaoran Jin, and Hang Li. Reft: Reasoning with reinforced fine-tuning. arXiv preprint arXiv:2401.08967, 2024. 3 +[25] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 2 +[26] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5 +[27] Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Tiancheng Han, Botian Shi, Wenhai Wang, Junjun He, et al. Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforce + +ment learning. arXiv preprint arXiv:2503.07365, 2025. 3 +[28] OpenAI. Gpt-4 technical report, 2023. 3 +[29] OpenAI. Introducing openai o1, 2024. 2 +[30] Yingzhe Peng, Gongrui Zhang, Miaosen Zhang, Zhiyuan You, Jie Liu, Qipeng Zhu, Kai Yang, Xingzhong Xu, Xin Geng, and Xu Yang. Lmm-r1: Empowering 3b Imms with strong reasoning abilities through two-stage rule-based rl. arXiv preprint arXiv:2503.07536, 2025. 3, 6 +[31] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3 +[32] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 3 +[33] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3 +[34] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 1, 3, 5 +[35] Guangzhi Sun, Wenyi Yu, Changli Tang, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, Yuxuan Wang, and Chao Zhang. video-salmonn: Speech-enhanced audio-visual large language models. arXiv preprint arXiv:2406.15704, 2024. 2 +[36] Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1.5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599, 2025. 1, 3 +[37] Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed Zumri, Jean Lahoud, Rao Muhammad Anwer, et al. Llamavol: Rethinking step-by-step visual reasoning in llms. arXiv preprint arXiv:2501.06186, 2025. 1, 3, 6 +[38] 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. 1, 2, 6 +[39] Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, Shengyi Huang, Kashif Rasul, and Quentin Gallouédec. Trl: Transformer reinforcement learning. https://github.com/huggingface/trl, 2020.6 +[40] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095-95169, 2025. 5 +[41] 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. 6 +[42] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng + +Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 2 +[43] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 1, 2, 6 +[44] Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. Llava-ol: Let vision language models reason step-by-step. arXiv preprint arXiv:2411.10440, 2024. 1, 3, 5, 6 +[45] Huanjin Yao, Jiaxing Huang, Yawen Qiu, Michael K Chen, Wenzheng Liu, Wei Zhang, Wenjie Zeng, Xikun Zhang, Jingyi Zhang, Yuxin Song, et al. Mmreason: An open-ended multi-modal multi-step reasoning benchmark for mllms toward agi. arXiv preprint arXiv:2506.23563, 2025. 5 +[46] Huanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang, Yibo Wang, Shunyu Liu, Yingjie Wang, Yuxin Song, Haocheng Feng, Li Shen, et al. Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte carlo tree search. arXiv preprint arXiv:2412.18319, 2024. 1, 3, 5, 6 +[47] Huanjin Yao, Qixiang Yin, Jingyi Zhang, Min Yang, Yibo Wang, Wenhao Wu, Fei Su, Li Shen, Minghui Qiu, Dacheng Tao, et al. R1-sharev1: Incentivizing reasoning capability of multimodal large language models via share-grpo. arXiv preprint arXiv:2505.16673, 2025. 3, 6 +[48] 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. 6 +[49] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 2 +[50] Dan Zhang, Sining Zhoubian, Ziniu Hu, Yisong Yue, Yuxiao Dong, and Jie Tang. Rest-mcts*: Llm self-training via process reward guided tree search. arXiv preprint arXiv:2406.03816, 2024. 3 +[51] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, et al. Mm1. 5: Methods, analysis & insights from multimodal llm fine-tuning. arXiv preprint arXiv:2409.20566, 2024. 1, 2, 6 +[52] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[53] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, pages 169–186. Springer, 2024. 5 +[54] Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqing Sun, Zhe Gan, Yinfei Yang, Ruoming Pang, and Yiming Yang. Improve vision language model chain-of + +thought reasoning. arXiv preprint arXiv:2410.16198, 2024. 1,3,6 +[55] Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weixiong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415, 2023. 2 +[56] Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, and Huan Zhang. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models. arXiv preprint arXiv:2411.00836, 2024. 5 \ No newline at end of file diff --git a/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/images.zip b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..84ba67487852a71dda4d2e3692bbcb31f8a770cb --- /dev/null +++ b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66602c1830ecf1f1591440aec3720d2befe52b82c2635fb7c5788ef67f93e5c6 +size 326087 diff --git a/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/layout.json b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6ae5d9c2aa9669a6f7bf3ae1ac17f0e4d846350c --- /dev/null +++ b/ICCV/2025/R1-VL_ Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f82e4becb4a753e47a6532e08e6e01a095c24ecde8aa4732071874d2f12515df +size 447128 diff --git a/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/547a2e55-6b71-4048-91a3-d261e7b97d6d_content_list.json b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/547a2e55-6b71-4048-91a3-d261e7b97d6d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9a008e9dd3c51311f56c18bbecb8202a49a4e735 --- /dev/null +++ b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/547a2e55-6b71-4048-91a3-d261e7b97d6d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d8b47e274f1acf2b720fdd1891f2cc81b38c9e73d26c5cbbf676492269547eb +size 77838 diff --git a/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/547a2e55-6b71-4048-91a3-d261e7b97d6d_model.json b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/547a2e55-6b71-4048-91a3-d261e7b97d6d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..480f1de541bb8772de05bf6299c559939e88de82 --- /dev/null +++ b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/547a2e55-6b71-4048-91a3-d261e7b97d6d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a60eb73a56d904e656075da1fe83cf5db8a42b6f3e8cd5526004435ea7cf19f4 +size 94786 diff --git a/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/547a2e55-6b71-4048-91a3-d261e7b97d6d_origin.pdf b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/547a2e55-6b71-4048-91a3-d261e7b97d6d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9760d4da82492ecdb3d8b744908da88ac77a7933 --- /dev/null +++ b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/547a2e55-6b71-4048-91a3-d261e7b97d6d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a508b3af3710245b305f8a88bac06ba89ac7a17a4ef268a3b99baba5e45026d +size 2320693 diff --git a/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/full.md b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d207961f1bde02131e3021bea11254e4e385f836 --- /dev/null +++ b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/full.md @@ -0,0 +1,290 @@ +# RA-BUSSeg: Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment + +Wanting Zhang $^{1}$ , Zhenhui Ding $^{1}$ , Guilian Chen $^{1}$ , Huisi Wu $^{1*}$ , Jing Qin $^{2}$ + +1 College of Computer Science and Software Engineering, Shenzhen University + +2 Centre for Smart Health, School of Nursing, The Hong Kong Polytechnic University + +2310273079@email.szu.edu.cn, hswu@szu.edu.cn + +# Abstract + +Accurate breast ultrasound (BUS) image segmentation is critical for diagnosis and surgical planning, but faces challenges due to limited labeled images. Semi-supervised methods show promise by leveraging pseudo-labels to mitigate reliance on large-scale annotations. However, their performance is highly dependent on the quality of pseudolabels, which is difficult to guarantee in BUS images due to inherent complexities such as low contrast, speckle noise, and artifacts. Previous studies primarily focus on refining pseudo-labels in one way or the other, or introducing auxiliary supervision; yet they overlook the potential of harnessing intrinsic and inherent pixel relations to enhance the robustness of semi-supervised segmentation. In this paper, we present a novel relation-aware semi-supervised model for BUS image segmentation, which is composed of two innovative components: an adjacent relation propagation (ARP) module and a cross-layer relation alignment (CRA) module, for comprehensively explore pixel relations to improve segmentation performance. The ARP propagates relations among adjacent pixels to reinforce the collaborative prediction of correlated pixels and enhance the model's awareness of local semantic consistency. The CRA aligns cross-layer pixel relations, employing deep-layer guidance to rectify erroneous correlations in shallow layers for noise suppression, while integrating multi-scale contexts to enable robust segmentation of lesions with varying sizes. Extensive experiments on two representative BUS datasets under various labeling conditions demonstrate the superiority of our method to the SOTAs. Codes are available at https://github.com/dodooo1/RA-BUSSeg. + +# 1. Introduction + +Breast cancer poses a significant global health challenge, with the Global Cancer Observatory reporting that it ac + +![](images/0f5cbce5195e023838b1ef576e6cabb3b5e926345768037d9580df3d5a96826c.jpg) +(a) + +![](images/17ef3a6073224cf927f18d66b409e10d91cb93047878539fd6d46a6a8dcda135.jpg) +(b) + +![](images/38a1a75f582ab72ddb9bf037f913033d9df501dbee64bacef548e7fa57d0f890.jpg) +(c) +Figure 1. Challenges of breast ultrasound image segmentation: (a) low contrast between breast masses and background, (b) blurring boundaries caused by speckle noise and severe artifacts, and (c) lesions with varying scales. The yellow boxes indicate the locations of the lesions and the red regions represent ground truth. + +counts for $24.5\%$ of all female cancer cases [3]. Evidence has demonstrated that early screening plays a pivotal role in reducing breast cancer mortality [5]. Breast ultrasound (BUS) is widely used in current screening procedures, given its portability, cost-effectiveness, non-invasiveness, and superior performance in detecting dense breasts versus mammography [2, 10]. Computer-assisted BUS analysis systems can enhance diagnosis accuracy and support precise surgical planning, while alleviating doctors' workload in large-scale screening scenarios. To the end, many recent studies [4, 16, 29, 34] are devoted to developing deep learning models for automatic segmentation of breast lesions from BUS images to facilitate the construction of these systems. + +To ensure segmentation model accuracy, large-scale pixel-level annotations are typically required for training. However, publicly available labeled BUS images are scarce, while manual annotation is costly and expertise-intensive. In this regard, semi-supervised learning (SSL) approaches are widely harnessed to reduce reliance on annotated data. Pseudo-labeling is a prevalent SSL paradigm that generates pseudo-labels for unlabeled data and uses them alongside labeled data for training [31, 43]. The learning objective is to minimize pixel-wise divergence between predictions and pseudo-labels, making training performance highly depen + +dent on pseudo-label quality. In BUS image segmentation, pseudo-label quality is, however, often degraded by low contrast, speckle noise, and imaging artifacts (see Fig. 1). + +Serveral approaches have been proposed to enhance pseudo-label quality to facilitate robust training [18, 25]. Some methods [31] discard low-confidence pseudo-labels based on predicted class entropy, but such post-prediction filtering fails to address the root issue. Sun et al. [26] refine pseudo-labels via global pixel pairwise relationships modeling and label propagation. However, modeling global pixel relationships in BUS images may incorrectly associate lesions with noise and artifact regions. Some recent studies employ contrastive learning to introduce complementary global supervision, enabling the acquisition of rich semantic context and alleviating the limitations of pixel-level constraints [30, 35]. Yet, in contrastive learning, the inaccurate sampling of positive/negative pair caused by suboptimal pseudo-labels can exacerbate confirmation bias. Mai et al. [20] analyze the bottlenecks of contrastive learning and propose rank-aware correlation consistency as a more robust supervision signal, which is constructed from pixel correlations modeled by representative agents. We argue that single-layer pixel relation modeling captures limited context and lacks multi-scale semantic integration, thereby weakening the model's capability in handling lesions of varying scales while suppressing speckle noise and artifacts. + +In this paper, we propose a novel semi-supervised model for BUS segmentation by sufficiently exploring pixel relations; we call the model $RA$ -BUSSeg. Our model consists of two innovative components: an adjacent relation propagation (ARP) module and a cross-layer relation alignment (CRA) module. In ARP, we model relations among adjacent pixels and leverage relation propagation to extract information from neighboring semantically similar pixels to facilitate prediction. This propagation enables the mutual reinforcement of predictions for correlated pixels within neighborhoods and enhances the model's perception of local semantic consistency. On the other hand, in CRA, we construct an auxiliary supervision signal based on the pixel relations among different feature layers to maintain the semantic consistency across layers. To mitigate the interference of speckle noise and low contrast in global pixel relation modeling, we model the relations in each layer by sampling representative points (called proxies) with spatial correspondence in each layer as semantic anchors, and comparing their correlations with each pixel. In principle, pixel relations in shallow layers retain richer detailed contrasts between pixels, while relations derived in abstract deep features have potential to reduce the impact of irrelevant noise and artifacts on segmentation. We align relations in shallow pixels with those in deep layers to enforce cross-layer consistency, where top-down semantic guidance corrects false correlations in low-level features, and thus effectively + +suppressing noise propagation. In addition, the aggregation of multi-scale contexts enhances the model's robustness in segmenting lesions of different sizes. We perform extensive experiments on two representative public breast ultrasound datasets, BUSI [1] and UDIAT [40], to validate the effectiveness of the proposed model; experimental results demonstrate that our RA-BUSSeg outperforms SOTA methods by a considerable margin under different semi-supervised settings. + +Our contributions can be summarized as follows: + +- We propose a novel relation-aware network for BUS image segmentation (RA-BUSSeg) that explores rich pixel relations to precisely delineate lesions under strong noise and artifacts, and thus boost the robustness of semi-supervised BUS segmentation. +- We propose an ARP module to reinforce connections between adjacent correlated pixels through relation propagation, enhancing the awareness of local semantic consistency, and a CRA module to leverage pixel relations extracted from high-level features to correct false correlations derived from shallow layers, thereby suppressing the impacts of noise and improve segmentation accuracy. +- We conduct extensive experiments on two representative BUS datasets under various labeling conditions to highlight the effectiveness and superiority of our method. + +# 2. Related Work + +# 2.1. Breast Ultrasound Image Segmentation + +In recent years, the fully-supervised paradigm based on CNN [21, 24, 41, 44] and Transformer [12, 22, 28, 37] has achieved remarkable outcomes in breast ultrasound (BUS) image segmentation. BCT-Net [36] integrates CNN and Transformer, introducing a dual-level attention mechanism and a hybrid objective loss function for breast ultrasound image segmentation. However, the accurate annotation of BUS images requires expertise and is costly, which hinders the applicability of fully-supervised segmentation methods. Some studies explore the potential of semi-supervised learning to reduce the dependence on labeled data. Mainstream semi-supervised learning methods are primarily classified into pseudo-labeling and consistency regularization. The pseudo-labeling paradigm [15, 23, 38] generates pseudo-labels for unlabeled data based on model predictions and uses them alongside labeled data for training. On the other hand, the consistency regularization-based methods [9, 17, 39, 42] learn from unlabeled data by encouraging consistency between the original predictions and their perturbed counterparts (including perturbations to inputs, features, and the network). For semi-supervised segmentation in breast tumor, UA-RC [33] improves the segmentation performance by calibrating the representations of uncertain predictions. PH-Net [13] proposes an adaptive + +![](images/d3ea9054b77371177586dbf6a2c793d73896186b69e00307aed98fd310617206.jpg) +Figure 2. Overview of RA-BUSSeg, which consists of an adjacent relation propagation (ARP) module and a cross-layer relation alignment (CRA) module. Our ARP strengthens connections between adjacent correlated pixels with similar semantic via relation propagation, enhancing the model's perception of local semantic consistency. Meanwhile, CRA utilizes pixel relations from deep features to rectify false correlations derived from shallower layers, effectively suppressing the impact of noise and improving segmentation accuracy. + +patch augmentation and a hard-patch contrastive learning to achieve better performance on hard cases. + +# 2.2. Inter-pixel Relation + +In semi-supervised semantic segmentation, the pseudolabeling paradigm assigns a category pseudo-label to each pixel and leverage individual pixel-wise consistency as a supervision signal to guide model training. However, this process heavily relies on the accuracy of pseudo-label, and relying solely on individual pixel-wise consistency fails to fully explore the information embedded in unlabeled data. IpixMatch [32] proposes an extra inter-pixel loss to mine the neglected but valuable inter-pixel information for semi-supervised learning. RankMatch [19] constructs representative proxies to model inter-pixel correlations and utilizes rank-aware correlation consistency as a more effective supervision signal to achieve robust semi-supervised learning. Given the spatial continuity and structural similarity of lesions in breast ultrasound images, modeling the relation of adjacent pixel effectively aggregates intra-neighborhood information, thereby enhancing the accurate of pseudo-label assignment. Modeling global pixel relationships further enables the model to capture comprehensive contextual information in long-range dependencies, which is particularly important for segmentation tasks. Additionally, the coordination between shallow features providing fine-grained details and deep features containing high-level semantics allows the model to preserve a simultaneous understanding of contextual semantics across multi-scale feature layers. These insights drive our exploration of pixel relation within + +adjacent regions and across feature layers, aiming to break the bottleneck of pixel-wise constraints. + +# 3. Method + +# 3.1. Overview + +The overview of our semi-supervised breast ultrasound (BUS) image segmentation model, $RA$ -BUSSeg, is shown in Fig. 2. Our network follows the basic teacher-student framework, where both teacher and student networks share the same structure. Initially, our model is pre-trained on limited labeled images $B^{l} = \{(x_{i}^{l},y_{i}^{l})\}$ , where $y_{i}^{l}$ denotes the label for each image. $\mathcal{L}_s$ is employed for supervised learning on labeled set $B^{l}$ , which minimizes the divergence between the prediction from student network $\mathcal{P}^s (x_i^l)$ and $y_{i}^{l}$ . + +$$ +\mathcal {L} _ {s} = \frac {1}{| B ^ {l} |} \sum_ {i = 1} ^ {| B ^ {l} |} \frac {1}{| H W |} \sum_ {j = 1} ^ {H W} \ell_ {c e} \left(\mathcal {P} ^ {s} \left(x _ {i j} ^ {l}\right), y _ {i j} ^ {l}\right), \tag {1} +$$ + +where the subscript $ij$ represents the $j$ -th pixel in the $i$ -th image, and $|B^l|$ denotes the number of labeled samples. $\ell_{ce}$ is the standard cross-entropy loss. Subsequently, unlabeled images $B^u = \{x_i^u\}$ are also used for training, where the pseudo-labels generated from the predictions of teacher network. For unsupervised learning on unlabeled set, $\mathcal{L}_u$ is applied to enforce consistency between the prediction from student network $\mathcal{P}^s(x_i^u)$ and the pseudo-label $\hat{y}_i^u$ . + +$$ +\mathcal {L} _ {u} = \frac {1}{| B ^ {u} |} \sum_ {i = 1} ^ {| B ^ {u} |} \frac {1}{| H W |} \sum_ {j = 1} ^ {H W} \ell_ {c e} \left(\mathcal {P} ^ {s} \left(x _ {i j} ^ {u}\right), \hat {y} _ {i j} ^ {u}\right), \tag {2} +$$ + +![](images/4a927e97a4a50a0c45e7ee206dbc64d7dc403efaf5fd8716a47fc4b42ff4d304.jpg) +Figure 3. Illustration of adjacent relation propagation (ARP). The ARP spreads adjacent relations into predictions, aiming to mutually reinforce the prediction of semantically similar pixels and enhance the perception of local semantic consistency. + +where the $|B^u|$ denotes the number of unlabeled samples. + +However, in low-contrast BUS images with speckle noise and severe artifacts, the generated pseudo-labels often suffer from poor quality, hindering the effective learning from unlabeled data. We explore robust semi-supervised learning from pixel relations to break the bottleneck of pixel-wise constraints from pseudo-labels. To this end, we develop two innovative components: an adjacent relation propagation (ARP) module and a cross-layer relation alignment (CRA) module. Our proposed ARP module propagates the relations between adjacent pixels into the prediction, refining the prediction with information extracted from neighboring semantically similar pixels. It strengthens the connection between adjacent correlated pixels and promotes the model's perception of local semantic consistency. Our CRA module further models pixel relations through representative proxies in different feature layers and enforces consistency on shallow-to-deep relations to align multi-scale semantics. It utilizes high-level semantic to rectify the erroneous pixel relations derived from shallow features, aiming to mitigate noise propagation. In addition, the aggregation of multi-scale context promotes the model's robust segmentation of lesions with varying scales. + +# 3.2. Adjacent Relation Propagation + +Most semi-supervised methods employ the entropy of predicted class probabilities to filter out low-confidence pixel pseudo-labels. However, this filtering strategy, as a post-processing method, fails to actually enhance the quality of the pseudo-labels. We realize that the relations between adjacent pixels can provide valuable insights for more accurate prediction. To this end, we propose an adjacent relation propagation (ARP) module to strengthen the model's attention to adjacent pixel relations, which involves the generation of relation maps and the subsequent propagation. + +As shown in Fig. 3, we first generate the adjacent relation map based on the features $\mathcal{F} \in \mathbb{R}^{HW \times C}$ extracted by the encoder, where $C$ denotes the channel dimension, and + +$H$ and $W$ are the height and width, respectively. Specifically, we employ a sliding window to capture features within the adjacent region and use the dot product to quantify the correlations between adjacent points and the center point. For clarity, we illustrate the process on a single pixel $p_j$ and utilize $\mathcal{N}_j = \{p_{k,l} \mid -\frac{n}{2} \leq k \leq \frac{n}{2}, -\frac{n}{2} \leq l \leq \frac{n}{2}\}$ to represent the set of pixels in the $n \times n$ sliding window centered on $p_j$ . The adjacent relation map $\mathcal{M}_j$ of $p_j$ can be expressed as follows: + +$$ +\mathcal {M} _ {j} = \operatorname {S o f t m a x} \left(\mathcal {F} _ {j} \mathcal {F} _ {\mathcal {N} _ {j}} ^ {\top} / \sqrt {C}\right), \tag {3} +$$ + +where $\mathcal{F}_j\in \mathbb{R}^{1\times C}$ is the feature of $p_j$ , $\mathcal{F}_{\mathcal{N}_j}\in \mathbb{R}^{nn\times C}$ denotes the features of the pixels in $\mathcal{N}_j$ , and $\top$ represents the matrix transpose operation. Subsequently, we propagate the relation map into the predictions (logit outputs) of student network by matrix multiplication, obtaining a new output that considers the pixel relation within adjacency compared to the original output. The new output of $p_j$ is calculated as: + +$$ +\tilde {\mathcal {P}} ^ {s} (p _ {j}) = \sum \mathcal {P} ^ {s} (\mathcal {N} _ {j}) \odot \mathcal {M} _ {j}, \tag {4} +$$ + +where $\mathcal{P}^s (\mathcal{N}_j)$ is the original logits output of pixels in $\mathcal{N}_j$ and $\odot$ denotes the Hadamard product. We minimize the discrepancy between the new output and the pseudo-label using the loss $\mathcal{L}_{ARP}^u$ to enhance the model's awareness of local semantic consistency, formulated as follows: + +$$ +\mathcal {L} _ {A R P} ^ {u} = \frac {1}{| B ^ {u} |} \sum_ {i = 1} ^ {| B ^ {u} |} \frac {1}{| H W |} \sum_ {j = 1} ^ {H W} \ell_ {c e} (\tilde {\mathcal {P}} ^ {s} (x _ {i j} ^ {u}), \hat {y} _ {i j} ^ {u}), \quad (5) +$$ + +where $\tilde{\mathcal{P}}^s (x_{ij}^u)$ denotes the new output of the $j$ -th pixel in an unlabeled image $x_{i}^{u}$ generated by Eq. (4). Moreover, this process is also applied to labeled images, using the $\tilde{\mathcal{P}}^s (x_i^l)$ and the label $y_{i}^{l}$ to construct the loss $\mathcal{L}_{ARP}^l$ . The loss introduced by ARP module, $\mathcal{L}_{ARP} = \mathcal{L}_{ARP}^l +\mathcal{L}_{ARP}^u$ , serves as an additional supervision signal for model optimization. In addition to focusing on classification accuracy, this optimization objective models the relations between adjacent pixels, facilitating mutual reinforcement of predictions for similar neighboring pixels to capture the overall structure of lesions and enhancing the model's awareness of local semantic consistency. + +# 3.3. Cross-layer Relation Alignment + +Modeling pixel relations in a single feature layer captures only limited contextual information at a fixed scale, failing to encompass multi-scale semantics. This limitation weakens the model's ability to distinguish noise and compromises its robustness in handling scale-varying lesions. To this end, we propose a cross-layer relation alignment (CRA) module, which involves modeling pixel relations through proxies and optimizing the model by enforcing consistency in cross-layer pixel relations. + +![](images/8fa0785e27637fc2f4f1ce46a58faef495db05c3def7a6ce1f9610f446774063.jpg) +Figure 4. Illustration of cross-layer relation alignment (CRA). The CRA leverages high-level semantic information to rectify false correlations in noise-prone shallow layers, facilitating noise suppression and improving segmentation precision. + +As shown in Fig. 4, we first adopt a greedy strategy to select a set of representative points, referred to as proxies. We employ proxies as semantic anchors to model pixel relations, decreasing computational complexity while reducing noise in pixel relations that could interfere with model optimization. Specifically, we first randomly select a pixel in the lesion area as the first proxy $P_{1}$ , and compute its correlations with other points via the dot product of their features. The second proxy $P_{2}$ is the pixel with the lowest correlation with $P_{1}$ . To prevent the third proxy from being localized around $P_{1}$ , we select the pixel with the lowest correlation with both $P_{1}$ and $P_{2}$ , rather than comparing only with $P_{2}$ . + +$$ +\begin{array}{l} \mathcal {C} _ {P _ {1}} = \operatorname {S o f t m a x} \left(\mathcal {F} _ {P _ {1}} \mathcal {F} ^ {\top} / \sqrt {C}\right), \tag {6} \\ P _ {2} = \arg \min \left(\mathcal {C} _ {P _ {1}}\right), \\ \end{array} +$$ + +$$ +\begin{array}{l} \mathcal {C} _ {P _ {2}} = \operatorname {S o f t m a x} \left(\mathcal {F} _ {P _ {2}} \mathcal {F} ^ {\top} / \sqrt {C}\right), \tag {7} \\ P _ {3} = \arg \min \left(\mathcal {C} _ {P _ {1}} + \mathcal {C} _ {P _ {2}}\right), \\ \end{array} +$$ + +where $\mathcal{C}_{P_k} \in \mathbb{R}^{1 \times HW}$ represents the correlation of the $k$ -th proxy with all pixels, and $\arg \min$ denotes the operation of selecting the index corresponding to the minimum value in $\mathcal{C}_{P_k}$ . Following this strategy, the pixel with the weakest correlation to the selected proxies is sampled as a new proxy each time. These proxies, characterized by distinct features, preserve comprehensive semantic cues in original image. The relations between each pixel $p_j$ and proxies exhibit diverse contrastive descriptions within an image, which can be quantified through the dot product: + +$$ +\mathcal {R} _ {p _ {j}} = \mathcal {F} _ {p _ {j}} \cdot \mathcal {F} _ {P _ {k}}, \quad \forall k \in \{1, 2, \dots , N \}, \tag {8} +$$ + +where $N$ is the number of proxy. The modeling of pixel relations within a single feature layer confines contextual information to a fixed receptive field, limiting the model's ability to consistently understand the relationship between multi-scale features. The shallow layers emphasize fine-grained local details of the lesion, such as texture and sharp + +edges, while deep features offer high-level global semantics, like the overall morphology and structure. All of them are essential for perceiving the structure of the lesions and accurately delineating their details. Therefore, we utilize bilinear interpolation to match the spatial resolution of feature maps from deeper layers to shallower layers, and map the proxies to their corresponding positions for modeling pixel relations in multiple layers. To achieve semantic alignment across layers and ensure consistent understanding of multiscale features, we employ Kullback-Leibler (KL) divergence to align shallow pixel relations with those in deeper layers. Specifically, we minimize the divergence of proxy-level relations across layers, which is formulated as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {C R A} = \frac {1}{| B |} \sum_ {i = 1} ^ {| B |} \frac {1}{| H W |} \sum_ {j = 1} ^ {H W} \left(\ell_ {k l} \left(R _ {p _ {i j}} ^ {1}, R _ {p _ {i j}} ^ {2}\right) \right. \tag {9} \\ + \ldots + \ell_ {k l} (R _ {p _ {i j}} ^ {L - 1}, R _ {p _ {i j}} ^ {L})), \\ \end{array} +$$ + +where $|B|$ is the length of labeled and unlabeled set, and $L$ denotes the $L$ -th feature layer. Speckle noise, due to its random brightness and position, causes high-frequency variations in shallow features, while deeper layers smooth these changes by focusing on global semantics. Under this optimization objective, the model effectively captures pixel relations across different scales and leverages high-level semantic guidance to rectify erroneous relations derived from shallow layers. This process mitigates noise propagations, thereby improving the model's segmentation performance for breast lesions in the presence of severe noise. In addition, when encountering breast lesions of varying scales, the aggregation of multi-scale contextual information enhances the model's robustness in addressing this challenge. + +# 3.4. Loss Function + +The overall loss function is formulated as a weighted combination of $\mathcal{L}_s$ in Eq. (1), $\mathcal{L}_u$ in Eq. (2) and the losses introduced by the proposed ARP and CRA modules, which can be expressed as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {s} + \lambda_ {u} \mathcal {L} _ {u} + \lambda_ {1} \mathcal {L} _ {A R P} + \lambda_ {2} \mathcal {L} _ {C R A}, \tag {10} +$$ + +where $\lambda_{u},\lambda_{1}$ , and $\lambda_{2}$ denote the weights for the respective losses, all set to 1. + +# 4. Experiments + +# 4.1. Experimental Setup + +# 4.1.1. Datasets and evaluation metrics + +We validate the effectiveness of our method on two public breast ultrasound (BUS) datasets, BUSI [1] and UDIAT [40]. The BUSI dataset comprises 780 annotated images, including 437 benign, 210 malignant, and 133 normal cases, derived from 600 female aged 25 to 75 years. We conduct + +![](images/0b548187a2b4813a3d216beea8667900394cf3f5e03ba6d023e9e7a55b36c07d.jpg) +Figure 5. Visual comparison with state-of-the-art methods on the 1/4 partition protocol of two famous breast ultrasound datasets: UDIAT and BUSI. Red, green and yellow regions represent ground truth, prediction and their overlapping regions, respectively. + +experiments on both benign and malignant cases containing lesions to evaluate the performance of our approach in breast tumor segmentation. The UDIAT dataset, collected from the UDIAT Diagnostic Center, contains 163 annotated BUS images (110 benign and 53 malignant cases). We employ five-fold cross-validation under different labeling scenarios, including $1/2$ , $1/4$ and $1/8$ partition protocols. The partition protocols represent the proportion of labeled data within the entire dataset. We employ the dice coefficient (Dice), intersection over union (IoU) and Hausdorff Distance (HD) to evaluate the segmentation performance, and compute the mean value and the standard deviation across five folds. + +# 4.1.2. Implementation details + +Our method is implemented in PyTorch and runs on a single NVIDIA GeForce RTX 3090 with 24 GB of memory. We adopt ResNet-50 [11] pre-trained on ImageNet [8] as the backbone, while DeepLabv3+ [6] serves as the decoder. The student network is optimized using a stochastic gradient descent (SGD) optimizer with a momentum of 0.9 and a weight decay of 0.0005. The learning rate starts at 0.003, which is decayed by a polynomial strategy $lr = 0.003 \times \left(1 - \frac{\text{iter}_{\text{curr}}}{\text{iter}_{\text{total}}}\right)^{0.9}$ . To update the teacher model, we apply exponential moving averaging (EMA) to the student network's parameters. The training lasts for 200 epochs, with the first 10 dedicated to pre-training. The batch-size is set to 2. During each epoch, we unify the image resolution to $512 \times 512$ and apply weak data augmentation, including random scaling, random cropping and random horizontal flipping, to both labeled and unlabeled samples. + +# 4.2. Comparison with State-of-the-art Methods + +We compare our method with several state-of-the-art (SOTA) approaches, including seven semi-supervised segmentation methods for natural and medical images and + +three specific methods for breast ultrasound (BUS) datasets. The selected methods can be categorized as follows: (1) Classic semi-supervised methods: $\mathrm{U}^2\mathrm{PL}$ [31] and CSS [27]; (2) Contrastive learning-based methods: DGCL [30] and PRCL [35]; (3) Correlation-based methods: CorrMatch [26] and RankMatch [20], which have exhibited competitive performance in natural image segmentation through inter-pixel correlation analysis; (4) Medical-specific methods: ABD [7], a SOTA semi-supervised model for medical images, and BUS-specific frameworks UA-RC [33] and PH-Net [14]. All methods adopt ResNet-50 as the backbone and are implemented under the same experimental conditions as our approach. Fig. 5 visualizes the predictions of our model alongside those of selected methods, which intuitively highlight the superiority of our approach. We deliberately show challenging cases with lesions of varying scales, irregular shapes, and surrounded by noise and artifacts to demonstrate our method's precise tumor segmentation capability in complex scenarios. + +The quantitative results presented in Tab. 1 conclusively demonstrates the superior performance of our proposed model across all comparative approaches. The improvements over correlation-based methods demonstrate the efficacy of our proposed modules in tackling task-specific challenges unique to breast ultrasound imaging—particularly inherent noise, severe artifacts, and lesions of varying sizes. Specifically, the adjacent relation propagation mechanism enhances local semantic consistency and spatial continuity in segmentation predictions, while the cross-layer relation alignment facilitates noise suppression and improves robustness to lesion size variations. Notably, our method outperforms the SOTA model ABD by $1.46\%$ , $1.65\%$ , and $1.35\%$ in Dice coefficient under $1/2$ , $1/4$ , and $1/8$ labeled data partitions on the BUSI dataset, respectively, while also achieving superior performance over BUS-specific frameworks across all evaluation protocols. Even compared to + +
MethodYearUDIATBUSI
1/21/41/81/21/8
Dice±std (%)IoU±std (%)HDDice±std (%)IoU±std (%)HDDice±std (%)IoU±std (%)HDDice±std (%)IoU±std (%)HDDice±std (%)IoU±std (%)HD
\( U^2PL [31] \)202281.39±3.2373.11±3.526.8878.20±3.1469.99±2.797.1473.44±2.6764.92±3.038.2575.72±2.7466.96±3.037.9573.06±2.7664.85±2.449.2770.53±3.1561.96±3.639.56
CSS [27]202382.15±3.0773.20±3.346.4178.67±3.0869.92±3.176.9774.25±2.3765.97±2.137.9676.02±2.6767.36±2.658.0373.74±2.6064.69±2.159.0171.19±2.9862.43±3.229.29
DGCL [30]202382.97±2.8773.39±2.946.0080.25±3.0772.05±3.016.3075.91±3.1466.88±3.287.7376.88±2.2267.50±1.957.6675.75±2.3767.30±2.068.3072.85±2.6863.22±2.128.74
PRCL [35]202482.78±2.9474.12±3.006.2579.82±3.1270.84±3.296.5675.68±2.9466.55±2.517.9376.54±2.4868.05±2.627.8275.21±2.5766.06±2.699.0272.60±2.6364.03±3.399.02
CorrMatch [26]202482.61±2.6573.88±2.946.0979.43±2.8870.46±3.196.5474.49±2.3665.32±2.488.0676.18±2.0267.57±2.757.9975.03±1.9766.38±2.368.9472.22±2.0863.79±1.769.02
RankMatch [20]202483.18±2.8373.73±3.106.0780.77±3.0871.95±3.126.3176.40±2.2267.53±2.197.3777.10±2.1668.79±1.847.0075.76±2.4267.05±2.118.2673.20±2.4364.66±2.108.65
ABD [7]202483.75±2.1774.57±2.665.7281.69±2.1672.54±2.386.0077.94±2.8068.27±2.926.7377.65±1.7769.15±1.526.6376.30±2.0267.72±1.878.2473.71±2.2365.03±2.348.62
UA-RC [33]202483.40±2.5475.05±2.105.6281.29±3.0272.91±3.466.2277.22±2.2368.28±2.157.0277.47±1.9868.72±1.856.8276.09±2.2467.47±2.018.1873.66±2.4765.00±2.918.27
PH-Net [14]202484.06±2.2975.70±2.635.2881.80±2.5072.02±2.865.8478.40±3.3569.87±3.696.5478.02±1.9969.81±2.756.4076.81±1.7568.38±2.378.2174.09±2.1065.53±2.338.01
BCT-Net [36]202580.07±3.0671.20±3.168.1176.42±3.5567.89±3.267.0373.00±2.6164.32±2.907.8374.96±2.6166.98±2.488.7372.52±2.5664.17±2.619.1069.99±1.8661.29±2.059.89
SupOnly-78.51±1.3469.55±1.467.9375.11±2.3766.44±2.387.0172.24±4.1363.27±4.218.8274.17±2.3965.60±2.518.6271.62±2.0162.78±2.179.3368.96±1.8660.28±1.819.94
Ours-84.77±2.4376.40±2.265.2882.55±2.2174.21±1.805.7379.76±2.7971.24±2.686.5079.11±1.8970.30±2.066.1777.95±1.8469.30±1.648.0075.06±2.1966.41±2.358.13
+ +Table 1. Quantitative comparison with state-of-the-art methods on two famous breast ultrasound datasets: UDIAT and BUSI. + +
MethodUDIATBUSI
ARPCRADice±std (%)IoU±std (%)Dice±std (%)IoU±std (%)
75.11±2.3766.44±2.3871.62±2.0162.78±2.17
✓80.79±2.2971.91±2.1876.31±2.2467.67±1.84
✓81.21±1.9172.42±2.1176.42±2.5167.84±2.60
✓✓82.55±2.2174.21±1.8077.95±1.8469.30±1.64
+ +Table 2. Ablation studies of proposed ARP and CRA components on 1/4 partition protocol of UDIAT and BUSI datasets. + +![](images/be1d834f43ba53302a5aedd3deba68159eff3d9d7f56f4d9dd054855eb5318d9.jpg) +Figure 6. Comparison of predicted lesion boundaries without relation propagation, with adjacent and with global propagation. The red and yellow contours represent ground truth and prediction with our adjacent relation propagation, respectively. The green contours (top) and the blue contours (bottom) show predicted edges without propagation and with global propagation, respectively. + +the fully supervised BCT-Net [36], our method shows competitive performance, highlighting the efficacy of leveraging unlabeled data under limited annotations. + +# 4.3. Ablation Studies + +# 4.3.1. Effectiveness of ARP and CRA + +We conduct component-wise ablation analysis to evaluate the adjacent relation propagation (ARP) module and the cross-layer relation alignment (CRA) module. In Tab. 2, we present the quantitative results on the UDIAT and BUSI datasets under the 1/4 partition protocol (results under other protocols are provided in the supplementary material). SupOnly, which is only trained on labeled images, serves as the baseline with Dice coefficient of $75.11\%$ and $71.62\%$ on the UDIAT and BUSI datasets, respectively. + +![](images/bec4d452b0917e16cd3e41cf656c09956f2c93f68081d32a4af02ad8aa1fb652.jpg) +Figure 7. Visual comparison of feature maps in combination of ARP and CRA on 1/4 partition protocol of UDIAT and BUSI datasets. (a) Image. (b) GT. (c) SupOnly. (d) w/ ARP. (e) w/ CRA. (f) Ours. + +The combination of the ARP module achieves significant Dice improvements of $+5.68\%$ and $+4.69\%$ on the UDIAT and BUSI datasets, respectively. As illustrated in Fig. 7, the feature map comparison reveals that relation propagation guarantees local semantic consistency and enhances the overall continuity of the segmented breast masses. In addition, we visualize the comparison of predicted lesion boundaries with and without adjacent propagation, as shown at the top of Fig. 6. It indicates that ARP enables to refine boundary with the semantic assistance of adjacent similar pixels, ensuring that local semantic consistency is taken into account during prediction. The comparative visualization between adjacent propagation and global propagation, presented in the bottom of Fig. 6, highlights the constraining of propagation range effectively mitigates noise-induced false associations while maintaining spatial coherence in segmentation targets. + +Additionally, when integrated with the CRA module, our method attains Dice enhancements reaching $6.10\%$ and $4.80\%$ on the UDIAT and BUSI datasets, respectively. Fig. 7 intuitively reveals that the CRA enables precise local- + +![](images/306af7b47764c5ee71a39312dd34367338a11a48b918b8c43a692d602e3eae58.jpg) +Figure 8. Ablation studies in the range of propagation on $1/4$ partition protocol of UDIAT and BUSI datasets. + +![](images/41c106eb2b450133d5df26f14a470d8f74b28f0f39b1ea1aa9d166df4d2e414c.jpg) +Figure 9. Ablation studies in the number of proxy on 1/4 partition protocol of UDIAT and BUSI datasets. + +ization and robust delineation of lesion regions across varying scales. The Dice coefficient of the complete framework reaches $82.55\%$ (UDIAT) and $77.95\%$ (BUSI), verifying the importance of our proposed relation-aware architecture in achieving accurate segmentation of BUS images. + +Compared to the baseline, the training time increases slightly (96→139 s/epoch on BUSI under 1/4 partition), but the added complexity from relation construction notably improves segmentation accuracy. Notably, relation map construction and our modules operate solely during training, adding no inference-time overhead. + +# 4.3.2. Impact of propagation range + +We conduct an ablation study on the range of propagation in the adjacent relation propagation (ARP) module on the 1/4 partition protocol of the UDIAT and BUSI datasets, as shown in Fig. 8. By modulating the sliding window dimensions to regulate the number of feature vectors captured in the adjacent region, our model achieves optimal performance with a $3 \times 3$ sliding window configuration. The performance degrades as the window expands, which may be attributed to excessive propagation ranges inducing erroneous lesion-background correlations. Conversely, constrained range exploits the inherent spatial continuity of lesion structures, enforcing connections between adjacent pixels to enhance the model's awareness on local semantic consistency. + +![](images/b896ba80dae6f23115c643e1e65d90e0e87add838d47f284111e0bc6da8e16c8.jpg) +Figure 10. Failure cases. Green, red and yellow regions represent prediction, ground truth and their overlapping areas. + +# 4.3.3. Impact of the number of proxy + +We conduct an ablation analysis on the 1/4 partition protocol of the UDIAT and BUSI datasets to investigate the impact of proxies in the cross-layer relation alignment (CRA) module (see Fig. 9). Quantitative analysis reveals peak performance is achieved with 64 proxies. Insufficient proxies fail to preserve diverse semantic cues in the original feature maps, which may limit the acquisition of rich contrastive information for modeling pixel relations. However, an excessive number of proxies increases computational complexity and have potential to introduce noise into the correlation processes, negatively impacting the model's optimization. + +# 4.4. Discussions and Limitations + +Although we only conduct experiments on the breast ultrasound (BUS) datasets, our approach is tailored to the specific challenges of the BUS tasks, and we believe it can be generalized to other tasks with similar challenges. Additionally, our method still has limitations. As shown in Fig. 10, our method may misclassify regular artifacts in ultrasound images as lesions while separating true lesions. In addition, the segmentation results may be incomplete when light and dark boundaries exist within the lesions. + +# 5. Conclusion + +We propose a novel relation-aware semi-supervised segmentation network for breast lesions in ultrasound image, named RA-BUSSeg, which consists of an adjacent relation propagation (ARP) module and a cross-layer relation alignment (CRA) module. Our proposed ARP captures and propagates pixel relations within adjacent regions, allowing mutual reinforcement between semantically similar pixels and enhancing the model's perception of local semantic consistency. The developed CRA models pixel relations by proxies and aligns shallow-to-deep relations, harnessing high-level semantics to rectify erroneous correlations in shallow layers for noise suppression. CRA also facilitates robust segmentation of lesions with diverse scales by aggregating context in multi-scale feature layers. Extensive experiments on two BUS datasets validate our approach's superiority. + +# Acknowledgments + +This work was supported partly by the National Natural Science Foundation of China (No. 62273241), the Natural Science Foundation of Guangdong Province, China (No. 2024A1515011946), and the Hong Kong RGC Collaborative Research Fund (project no C5055-24G). + +# References + +[1] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 2, 5 +[2] Ioana Boca, Anca Ileana Ciurea, Cristiana Augusta Ciortea, and Sorin Marian Dudea. Pros and cons for automated breast ultrasound (abus): a narrative review. Journal of personalized medicine, 11(8):703, 2021. 1 +[3] Freddie Bray, Mathieu Laversanne, Hyuna Sung, Jacques Ferlay, Rebecca L Siegel, Isabelle Soerjomataram, and Ahmedin Jemal. Global cancer statistics 2022: Globocan estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians, 74(3):229-263, 2024. 1 +[4] Nicole Brunetti, Massimo Calabrese, Carlo Martinoli, and Alberto Stefano Tagliafico. Artificial intelligence in breast ultrasound: from diagnosis to prognosis—a rapid review. Diagnostics, 13(1):58, 2022. 1 +[5] Jennifer L Caswell-Jin, Liyang P Sun, Diego Munoz, Ying Lu, Yisheng Li, Hui Huang, John M Hampton, Juhee Song, Jinani Jayasekera, Clyde Schechter, et al. Analysis of breast cancer mortality in the us—1975 to 2019. Jama, 331(3): 233-241, 2024. 1 +[6] 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. 6 +[7] Hanyang Chi, Jian Pang, Bingfeng Zhang, and Weifeng Liu. Adaptive bidirectional displacement for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4070-4080, 2024. 6, 7 +[8] 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. 6 +[9] Geoff French, Samuli Laine, Timo Aila, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, varied perturbations. arXiv preprint arXiv:1906.01916, 2019. 2 +[10] Wilfrido Gómez-Flores, Maria Julia Gregorio-Calas, and Wagner Coelho de Albuquerque Pereira. Bus-bra: a breast ultrasound dataset for assessing computer-aided diagnosis systems. Medical Physics, 51(4):3110-3123, 2024. 1 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 + +[12] Qiqi He, Qiuju Yang, and Minghao Xie. Hctnet: A hybrid cnn-transformer network for breast ultrasound image segmentation. Computers in Biology and Medicine, 155: 106629, 2023. 2 +[13] Siyao Jiang, Huisi Wu, Junyang Chen, Qin Zhang, and Jing Qin. Ph-net: Semi-supervised breast lesion segmentation via patch-wise hardness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11418-11427, 2024. 2 +[14] Siyao Jiang, Huisi Wu, Junyang Chen, Qin Zhang, and Jing Qin. Ph-net: Semi-supervised breast lesion segmentation via patch-wise hardness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11418-11427, 2024. 6, 7 +[15] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2 +[16] Guoqi Liu, Jiajia Wang, Dong Liu, and Baofang Chang. A multiscale nonlocal feature extraction network for breast lesion segmentation in ultrasound images. IEEE Transactions on Instrumentation and Measurement, 72:1-12, 2023. 1 +[17] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4258-4267, 2022. 2 +[18] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4258-4267, 2022. 2 +[19] Huayu Mai, Rui Sun, Tianzhu Zhang, and Feng Wu. Rankmatch: Exploring the better consistency regularization for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3391-3401, 2024. 3 +[20] Huayu Mai, Rui Sun, Tianzhu Zhang, and Feng Wu. Rankmatch: Exploring the better consistency regularization for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3391-3401, 2024. 2, 6, 7 +[21] Zhenyuan Ning, Shengzhou Zhong, Qianjin Feng, Wu fan Chen, and Yu Zhang. Smu-net: Saliency-guided morphology-aware u-net for breast lesion segmentation in ultrasound image. IEEE transactions on medical imaging, 41(2):476-490, 2021. 2 +[22] Xiaolei Qu, Jiale Zhou, Jue Jiang, Wenhan Wang, Haoran Wang, Shuai Wang, Wenzhong Tang, and Xun Lin. Ehformer: Regional easy-hard-aware transformer for breast lesion segmentation in ultrasound images. Information Fusion, 109:102430, 2024. 2 +[23] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for + +semi-supervised learning. arXiv preprint arXiv:2101.06329, 2021. 2 +[24] 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. 2 +[25] Sebastian A. Scherer, Robin Schön, and Rainer Lienhart. Pseudo-label noise suppression techniques for semi-supervised semantic segmentation. In *British Machine Vision Conference*, 2022. 2 +[26] Boyuan Sun, Yuqi Yang, Le Zhang, Ming-Ming Cheng, and Qibin Hou. Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3097-3107, 2024. 2, 6, 7 +[27] Changqi Wang, Haoyu Xie, Yuhui Yuan, Chong Fu, and Xiangyu Yue. Space engage: Collaborative space supervision for contrastive-based semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 931-942, 2023. 6, 7 +[28] Jiajia Wang, Guoqi Liu, Dong Liu, and Baofang Chang. Mf-net: Multiple-feature extraction network for breast lesion segmentation in ultrasound images. Expert Systems with Applications, 249:123798, 2024. 2 +[29] Ke Wang, Shujun Liang, and Yu Zhang. Residual feedback network for breast lesion segmentation in ultrasound image. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part I 24, pages 471-481. Springer, 2021. 1 +[30] Xiaoyang Wang, Bingfeng Zhang, Limin Yu, and Jimin Xiao. Hunting sparsity: Density-guided contrastive learning for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3114-3123, 2023. 2, 6, 7 +[31] 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. 1, 2, 6, 7 +[32] Kebin Wu, Wenbin Li, and Xiaofei Xiao. Ipixmatch: Boost semi-supervised semantic segmentation with inter-pixel relation. arXiv preprint arXiv:2404.18891, 2024. 3 +[33] Yuanchen Wu, Xiaogiang Li, and Yue Zhou. Uncertainty-aware representation calibration for semi-supervised medical imaging segmentation. Neurocomputing, 595:127912, 2024. 2, 6, 7 +[34] Min Xian, Yingtao Zhang, Heng-Da Cheng, Fei Xu, Boyu Zhang, and Jianrui Ding. Automatic breast ultrasound image segmentation: A survey. Pattern Recognition, 79:340-355, 2018. 1 +[35] Haoyu Xie, Changqi Wang, Jian Zhao, Yang Liu, Jun Dan, Chong Fu, and Baigui Sun. Prcl: Probabilistic representation + +contrastive learning for semi-supervised semantic segmentation. International Journal of Computer Vision, pages 1-19, 2024. 2, 6, 7 +[36] Junchang Xin, Yaqi Yu, Qi Shen, Shudi Zhang, Na Su, and Zhiqiong Wang. Bct-net: semantic-guided breast cancer segmentation on bus. Medical & Biological Engineering & Computing, pages 1-12, 2025. 2, 7 +[37] Haonan Yang and Dapeng Yang. Cswin-pnet: A cnn-swin transformer combined pyramid network for breast lesion segmentation in ultrasound images. Expert Systems with Applications, 213:119024, 2023. 2 +[38] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4268-4277, 2022. 2 +[39] 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 +[40] Moi Hoon Yap, Manu Goyal, Fatima Osman, Robert Martí, Erika Denton, Arne Juette, and Reyer Zwiggelaar. Breast ultrasound region of interest detection and lesion localisation. Artificial Intelligence in Medicine, 107:101880, 2020. 2, 5 +[41] Yijun Zhao, Dashun Que, Jiaqi Tan, Yang Xiao, and Yanyan Yu. Automated breast lesion segmentation from ultrasound images based on ppu-net. In 2019 International Conference on Medical Imaging Physics and Engineering (ICMIPE), pages 1-4. IEEE, 2019. 2 +[42] Zhen Zhao, Sifan Long, Jimin Pi, Jingdong Wang, and Luping Zhou. Instance-specific and model-adaptive supervision for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23705-23714, 2023. 2 +[43] Zhen Zhao, Lihe Yang, Sifan Long, Jimin Pi, Luping Zhou, and Jingdong Wang. Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11350-11359, 2023. 1 +[44] Zhemin Zhuang, Nan Li, Alex Noel Joseph Raj, Vijayalakshmi GV Mahesh, and Shunmin Qiu. An rdau-net model for lesion segmentation in breast ultrasound images. PloS one, 14(8):e0221535, 2019. 2 \ No newline at end of file diff --git a/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/images.zip b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6ca4aa614b75dbdcafe5176883f9dfe3452e1876 --- /dev/null +++ b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:056bdcdac176ab19c3c981d2f0c5c2db3a4bb39cb7f9c3bfa1d5d5ad17b5310d +size 629621 diff --git a/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/layout.json b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b1f82aec1ea6cc6fc3779b893bbf01c444b139ac --- /dev/null +++ b/ICCV/2025/RA-BUSSeg_ Relation-aware Semi-supervised Breast Ultrasound Image Segmentation via Adjacent Propagation and Cross-layer Alignment/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f4b89d6cc33b2a5e02c3a5d636170a0d23372d2380fa727ebf34353eba539d3 +size 365265 diff --git a/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/777252c2-8dce-4808-a8d6-efab81de4547_content_list.json b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/777252c2-8dce-4808-a8d6-efab81de4547_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..131e77e3302123224510d1c2b5d815d19ed56d1a --- /dev/null +++ b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/777252c2-8dce-4808-a8d6-efab81de4547_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b61c1e8e826906d88b5260072fc256a749cf9bf1995e1c4c5aee268ab64e2d03 +size 81722 diff --git a/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/777252c2-8dce-4808-a8d6-efab81de4547_model.json b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/777252c2-8dce-4808-a8d6-efab81de4547_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8e61213e122be8874ec7ee41b49cb01da4050352 --- /dev/null +++ b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/777252c2-8dce-4808-a8d6-efab81de4547_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edb4aba69f3e3d93b8df57702b6bb78dc1a755ac04233a6d05095d92c83abe8b +size 102602 diff --git a/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/777252c2-8dce-4808-a8d6-efab81de4547_origin.pdf b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/777252c2-8dce-4808-a8d6-efab81de4547_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fe04a0c2fd98584239ececee821b050427582e78 --- /dev/null +++ b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/777252c2-8dce-4808-a8d6-efab81de4547_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85ccd3ee9541a5c9f78720636e4d6781fe61ac514dbc5ea20ba499101bf550e1 +size 2880710 diff --git a/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/full.md b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..30d1b95577e0b8feccb653f1769c45da3b85a0cf --- /dev/null +++ b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/full.md @@ -0,0 +1,369 @@ +# RAGD: Regional-Aware Diffusion Model for Text-to-Image Generation + +Zhennan Chen $^{1\star}$ Yajie Li $^{1\star}$ Haofan Wang $^{2,3}$ Zhibo Chen $^{3}$ Zhengkai Jiang $^{4}$ Jun Li $^{1}$ Qian Wang $^{5\boxtimes}$ Jian Yang $^{1}$ Ying Tai $^{1\boxtimes}$ $^{1}$ Nanjing University $^{2}$ InstantX Team $^{3}$ Liblib AI $^{4}$ HKUST $^{5}$ China Mobile https://github.com/NJU-PCALab/RAG-Diffusion + +![](images/f33c8637cdc7061bd4c78125c3f2aca00e982bd6f07eb694798301f078cd5be5.jpg) +On the left, Einstein is painting the Mona Lisa; in the center, Elon Reeve Musk is participating in the U.S. presidential election; on the right, Trump is hosting a Tesla product launch. + +![](images/e4f779350f69a5f746e00941e5b4cd5d56f2284ae95a37a6244b0af3655bffef.jpg) +Seven ceramic mugs in different colors are placed on a wooden table, with numbers from 1 to 7 written on the cups, and a bunch of white roses on the left. +Figure 1. RAGD decouples a raw prompt into regional prompts, processes different regions separately and strengthens the interaction between adjacent regions. It enables precise control over object relationship, action and attributes, achieving more harmonious and consistent complex compositional generation compared to competing models like Flux and RPG. + +![](images/0eb409e727db389586ae7c77aa9c86c3c4757ae0b68132ce59f4d21f8e1dbd10.jpg) +A cylindrical glass, obscuring the right half of the apple behind it. + +# Abstract + +Regional prompting, or compositional generation, which enables fine-grained spatial control, has gained increasing attention for its practicality in real-world applications. However, previous methods either introduce additional trainable modules, thus only applicable to specific models, or manipulate on score maps within attention layers using attention masks, resulting in limited control strength when the number of regions increases. To handle these limitations, we present RAGD, a Regional-Aware text-to-image Generation method conditioned on regional descriptions for precise layout composition. RAGD decouples the multi-region generation into two sub-tasks, + +the construction of individual region (Regional Hard Binding) that ensures the regional prompt is properly executed, and the overall detail refinement (Regional Soft Refinement) over regions that dismiss the visual boundaries and enhance adjacent interactions. Furthermore, RAGD novelly makes repairing feasible, where users can modify specific unsatisfied regions in the last generation while keeping all other regions unchanged, without relying on additional inpainting models. Our approach is tuning-free and applicable to other frameworks as an enhancement to the prompt following property. Quantitative and qualitative experiments demonstrate that RAGD achieves superior performance over attribute binding and object relationship than previous methods. + +# 1. Introduction + +Recent advancements in diffusion models [5, 14, 20, 25, 28, 32, 35, 36, 38-40, 43, 47, 48, 50-56] have substantially enhanced the aesthetic appeal and prompt adherence in text-to-image synthesis. The prevailing trend sees the denoising architecture shifting from UNet [31] to the Diffusion Transformer (DiT) [25], which excels in scaling with large datasets. Transformer-based diffusion models like PixArt $\alpha$ [4], Stable Diffusion 3/3.5 [9], and Flux.1 [2] have set a new benchmark, surpassing the quality of earlier UNet-based models such as Stable Diffusion 1.5 [30] and SDXL [26]. Furthermore, employing more robust text encoders, including T5-XXL [27], has demonstrated the ability to render visual text and significantly enhance prompt adherence. Some innovative studies even leverage large language models (LLMs) for text representation, with examples like KColors [37] utilizing GLM [7] and Playground V3 [23] employing Llama3-8B [8]. Despite this significant progress in generating high-quality images from prompts, achieving precise fine-grained spatial control remains elusive. In essence, current generative models still struggle with comprehending the quantity and spatial arrangement of objects. + +To address these limitations, the concept of regional prompting, also known as regional control, regional grounding, or composition generation, has emerged. Unlike providing a single global description, achieving fine-grained region-controllable generation requires users to supply not only the spatial location (e.g., a segmentation mask or bounding box) but also a corresponding description for each region. Several approaches have been proposed under this setting, broadly falling into two categories: tuning-based and tuning-free. For tuning-based methods, they often necessitate the training of an additional module to handle explicit conditions like bounding boxes. For example, GLIGEN [21] integrates regional inputs into new trainable layers through a gated mechanism, where each grounding token is a combination of the semantics of the grounded entity and its spatial location. Similarly, InstanceDiffusion [41] and Ms-diffusion [42] also incorporate learnable blocks to handle per-instance conditioning. These methods generally deliver strong performance in precise regional control but are limited to specific base models due to the introduction of extra trainable components. On the other hand, tuning-free methods, such as Multidiffusion [1], RPG [45], and Omost [24], operate on the denoised latent space or attention score map with a mask for each region. They frequently employ a split-and-merge strategy but face challenges in maintaining precise control as the number of regions increases. + +In this paper, we adopt a tuning-free manner and aim to improve its control strength and coherence when dealing with multiple regions. Specifically, we present RAGD, a novel Regional-Aware text-to-image Generation method for precise regional control, which is composed of two sub + +tasks, Regional Hard Binding and Regional Soft Refinement. First, we implement region-aware hard binding at the beginning of the denoising process to ensure that each regional prompt is executed accurately. This step breaks down the input prompt into several regional prompts, each with its respective spatial position, and then merges the individually denoised regional latents into the original image latent. Second, to dismiss the visual boundaries and enhance interaction between adjacent regions, regional soft refinement is applied within the multi-modal attention layers at the subsequent steps to obtain a regional latent, where $K$ and $V$ are from regional text tokens while $Q$ is from original image latent, followed by a weighted recombination of base image latent and regional latent. Furthermore, leveraging robust control and fusion capabilities, our framework supports users to refine specific unsatisfactory regions from the last generation while keeping all other regions intact. Both quantitative and qualitative results demonstrate our superior performance over attribute binding and object relationship than previous state-of-the-art methods. Our contributions are summarized as follows: + +- We propose RAGD, a tuning-free Regional-Aware text-to-image Generation framework on the DiT-based model, with two novel components, Regional Hard Binding and Regional Soft Refinement, for precise and harmonious regional control. +- RAGD novelly makes image repairing feasible, allowing users to modify specific unsatisfactory regions in the previous generation while keeping all other regions intact without need for additional inpainting models. +- Extensive qualitative and quantitative experiments demonstrate that RAGD shows superior performance on T2ICompBench and GenEval benchmark, in comparison with previous state-of-the-arts methods. + +# 2. Related Work + +Tuning-based Regional Control. Conventional text-to-image generation only uses text as conditional input and injects control signals through cross-attention. In order to handle spatial conditions, some works introduce additional training modules, such as ControlNet [49], to handle new control conditions, including depth maps, sketches, human poses, etc. For regional control, additional module is also introduced for training-based methods [3, 6, 11, 21, 41, 42, 46, 57, 58] to process spatial positions, such as bounding boxes (coordinates) or segmentation masks. GLIGEN [21] adopts Fourier embedding to encode box coordinates and adds trainable gated self-attention layer at each transformer block to accept new grounding input. InstanceDiffusion [41] further allows diverse ways to specify region positions such as simple single points, scribbles, bounding boxes or segmentation masks. Ms-diffusion [42] integrates grounding tokens with its grounding resampler to correlate spe + +![](images/8ba41549f82c30c7c12578301f4752b8f63491d4f95869a22f2261b2b48f376b.jpg) +Figure 2. The overall framework of RAGD, which divides regional-aware generation into two stages: (1) Regional Hard Binding ensures the proper response of regional prompts by processing each region individually with its fundamental description, and bound at the first $r$ steps to ensure accurate attribute representation and entity localization. (2) Regional Soft Refinement improves the harmony of adjacent region via enabling the interaction of regional local conditions with global image latent within the multi-modal attention layers at the later $T - r$ steps. The lower left corner shows the definition of spatial region in regional hard binding and regional soft refinement. + +cific entities and spatial constraints. Other works, such as Attend-and-Excite [3], apply regional constraints via gradient optimization within the denoising process to ensure all regional tokens are attended. + +Tuning-free Regional Control. Although tuning-based methods show strong performance, collecting training data is time-consuming and labor-intensive, and they are limited to specific models due to the introduction of additional modules on the top of base models. To address these challenges, model agnostic approaches [1, 10, 17, 19, 24, 44, 45, 60] are proposed. DenseDiffusion [19] and Omost [24] directly adjust attention scores within cross-attention layers to ensure that the activations within masked regions are promoted. Mixture of Diffusers [17] and MultiDiffusion [1] denoise different locations separately and then combine the denoised latents using regional masks. Recently, RPG [45] introduces complementary regional diffusion through a resize-and-concatenate approach for region-specific compositional generation, where regional denoised latents are resized and merged as a single concatenated latent at each step. However, among these existing solutions, certain approaches handle individual regions independently through a "divide-and-merge" strategy that lacks cross-region inter + +action mechanisms, resulting in boundary artifacts or target loss during merging. Other methods rely solely on attention masks to adjust inter-regional weight distributions but inadequately decouple the fundamental generation process from detail integration phases, which may lead to multi-region control failures. Thus, we propose RAGD, which decouples the compositional generation process into the construction of individual regions and detail refinement. This approach effectively controls the generation effects across multiple regions and significantly enhances the quality of details and the naturalness of boundary handling. + +# 3. Method + +# 3.1. Preliminaries + +Diffusion Transformer (DiT). DiT [25] is a novel architecture integrating transformer as the backbone network within Latent Diffusion Model (LDM), and has become a dominated choice in recent text-to-image generation models like Stable Diffusion 3/3.5 [9] and Flux [2]. By leveraging transformer, DiT efficiently captures complex dependencies in data, resulting in high-quality image generation. Consistent with the design philosophy of LDM, DiT also operates + +directly in the latent space, allowing the model to generate high-fidelity images that adhere to specified conditions while reducing computational overhead. + +Multi-modal Attention Mechanism. The multi-modal attention mechanism is a crucial component in DiT, enabling effective interactions between the diffusion network and additional control signals such as text or image. To handle inputs from different modalities effectively, MM-DiT assigns different weights to the prompt and image embeddings. During attention computation, the sequences of the prompt and image are merged to be processed independently in their respective spaces, while considering the influence of each other. The attention weights are calculated as follows: + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) \cdot V, \tag {1} +$$ + +where $Q$ , $K$ , and $V$ represent the Query, Key, and Value in the attention layer, respectively, and they are projected through a linear layer from the merged sequence of the prompt and the image. + +# 3.2. Overview of RAGD + +We briefly elucidate the core framework of RAGD, as depicted in Figure 2, which systematically dissociates the compositional generation workflow into two suboperational phases—foundational region construction and detail refinement. The implementation of RAGD consists of the following steps: + +(1) Regional Hard Binding (Sec. 3.3): This step involves decomposing the raw and complex input prompt containing multiple objects into a subset of fundamental descriptions for each individual region or object, along with their corresponding spatial positions. Then, each region is processed individually with its fundamental description and bound only at the early stage of denoising to ensure accurate attribute representation and entity localization. +(2) Regional Soft Refinement (Sec. 3.4): At this step, highly descriptive sub-prompts are generated for each region along with a global prompt to further enriching the definition of each object and promote the fusion between adjacent regions. Instead of manipulating the image latent, the refinement step achieves the interaction of regional local conditions and global image latent within the multi-modal attention layers. + +It is worth noting that based on our setting, RAGD can novelly support image inpainting (Sec. 3.5) in a free lunch manner by only re-initialize the initial noise in target areas, thereby enabling accurate modification of previously generated images while maintaining the overall generation quality without the need for an additional inpainting model. + +# 3.3. Regional Hard Binding + +To ensure the proper response of regional prompts and mitigate object omission when the number of region or ob + +ject increases, we apply regional hard binding in the early steps of denoising process as illustrated in Figure 2, which involves separately denoising the regions with their short fundamental descriptions and then binding local regional latents into global latent. + +Specifically, we first decompose the long input prompt $P$ containing multiple objects into a set of fundamental descriptions $\hat{p}^i$ with their position sets $m^i = \{m_{offset}^i, m_{scale}^i\}$ and $n^i = \{n_{offset}^i, n_{scale}^i\}$ . Subsequently, we perform text encoding on $P$ and $\hat{p}^i$ to obtain $y$ and $\hat{y}^i$ . The individual latent $\hat{x}^i$ utilizes $\hat{y}_i$ as a text condition, while origin latent $x$ takes the long input prompt $P$ as condition. The formulaic process is as follows: + +$$ +x _ {t - r} = x _ {t - r + 1} - \epsilon_ {\theta} (x _ {t - r + 1}, y) \tag {2} +$$ + +$$ +\hat {x} _ {t - r} ^ {i} = \hat {x} _ {t - r + 1} ^ {i} - \epsilon_ {\theta} \left(\hat {x} _ {t - r + 1} ^ {i}, \hat {y} ^ {i}\right), \tag {3} +$$ + +where $i\in [1,k],k$ is the number of regions. $\epsilon_{\theta}$ is the noise predicted. For each denoising step, we bind $\hat{x}_{t - r}^2$ to the latent space in the rectangular area given by $m_i,n_i$ as follows: + +$$ +x _ {t - r} = \operatorname {R e p l a c e} \left(x _ {t - r}, \hat {x} _ {t - r} ^ {i}, m ^ {i}, n ^ {i}\right), \tag {4} +$$ + +where $\operatorname{Replace}(\cdot)$ represents the process of pasting the individual latents back to the corresponding area in origin latent. The binding is only executed in the early steps within denoising process. We find that a few steps of binding is sufficient for regional completeness, whereas full-steps + +Algorithm 1: RAGD Algorithm Process +output: $x_0$ +input: Model $p_{\theta}$ ,Base Prompt $P$ ,Text Encoder TE, Hard Binding Step $r$ ,Fusion Strength $\delta$ ,Hard Binding Position $\{m^i,n^i\}$ ,Soft Refinement Region $\{h^i,w^i\}$ +initialization: Time Step T,Origin Latent $x$ Individual Latent $\hat{x}^i$ . +// $\hat{p}^i$ is simple subprompt for Hard Binding, $\tilde{p}^i$ is descriptive subprompt for Soft Refinement". $\hat{p}^i,\tilde{p}^i\gets P$ . + $y\gets TE(P),\hat{y}^{i}\gets TE(\hat{p}^{i}),\tilde{y}^{i}\gets TE(\tilde{p}^{i})$ . +for $t\gets T$ to $(T - r)$ do // Hard Binding; $x_{t - r}\leftarrow p_{\theta}(x_{t - r + 1},y);$ $\hat{x}_{t - r}^{i}\gets p_{\theta}(\hat{x}_{t - r + 1}^{i},\hat{y}^{i});$ $x_{t - r}\gets \mathrm{Replace}(x_{t - r},\hat{x}_{t - r}^{i},\{m^{i},n^{i}\})$ . +for $t\gets (T - (r + 1))$ to 0 do // Soft Refinement; $x_{t - (r + 1)}^{i}\gets \mathrm{jointAttn}(x_{t - (r + 1)},\tilde{y}^{i});$ $\tilde{x}_{t - (r + 1)}^{i}\gets \mathrm{Crop}(x_{t - (r + 1)}^{i},\{h^{i},w^{i}\})$ . $x_{t - (r + 1)}^{r}\gets \mathrm{Replace}(x_{t - (r + 1)},\tilde{x}_{t - (r + 1)}^{i},$ $\{h^i,w^i\})$ . $x_{t - (r + 1)}\gets x_{t - (r + 1)}*(1 - \delta)_+x_{t - (r + 1)}^r*\delta ;$ + +
ModelGenevalT2ICompBench
Counting ↑Position ↑Color_attr ↑Two_object ↑VQAScore ↑Color ↑Shape ↑Texture ↑Spatial ↑Complex ↑VQAScore ↑
MultiDiffusion [1] [ICML2023]0.400.060.150.420.63700.48110.43410.51050.23200.32780.6882
Ranni [11] [CVPR2024]0.210.110.080.300.56120.31540.30990.35380.14560.11570.5959
MIGC [58] [CVPR2024]0.510.760.360.840.76390.65520.48960.56610.38050.31640.7705
InstanceDiffusion [41] [CVPR2024]0.610.550.280.670.77950.64370.53610.63720.38710.30660.7508
RPG [45] [ICML2024]0.430.630.640.820.65020.76800.50780.61950.40170.37020.6502
3DIS [59] [ICLR025]0.660.720.500.830.78400.57480.32230.50860.30300.30350.7978
Ours0.800.820.790.950.80720.80390.60160.70850.51930.43770.8907
+ +Table 1. Comparison of alignment evaluation on Geneval [12] and T2ICompBench [16]. The best results are highlighted in bold, second-best in underline. + +binding results in either clear visual boundaries adjacent regions or poor interactivity. + +# 3.4. Regional Soft Refinement + +Images generated by direct regional hard binding allow for precise control over positioning, effectively preventing object omission. However, the rendering of attributes tends to be relatively coarse, and there is a tendency for noticeable boundaries between adjacent regions. Therefore, we apply regional soft refinement in the later steps of the denoising process to improve the harmony of adjacent regions. + +Specifically, similar to the previous binding step, we break down the original long prompt $P$ , but instead of breaking it down into short fundamental descriptions, we break it down into highly descriptive sub-prompts $\tilde{p}^i$ , with a set of global regions $o^i = \{h_{split}^i, w_{split}^i\}$ , and obtain the corresponding representation $\tilde{y}^i$ through the text encoder. Then, the text condition $\tilde{y}^i$ is projected into $K^i$ and $V^i$ while $Q^i$ is derived from current image latent $x_{t - (r + 1)}$ within multi-modal attention layers: + +$$ +\begin{array}{l} Q ^ {i} = \ell_ {Q} (x _ {t - (r + 1)}), K ^ {i} = \ell_ {K} (\tilde {y} ^ {i}), V ^ {i} = \ell_ {V} (\tilde {y} ^ {i}) \\ x _ {t - (r + 1)} ^ {i} = \operatorname {S o f t m a x} \left(\frac {Q ^ {i} K ^ {i ^ {T}}}{\sqrt {d}}\right) V ^ {i}, \tag {5} \\ \end{array} +$$ + +where $\ell_{Q},\ell_{k},\ell_{V}$ are linear projections, $\mathbf{x}_{t - (r + 1)}^i$ is globally-informed latent. We crop $\mathbf{x}_{t - (r + 1)}^i$ to the region corresponding to different objects to obtain $\tilde{\mathbf{x}}_{t - (r + 1)}^i$ with rich attributes generated in the global region: + +$$ +\tilde {x} _ {t - (r + 1)} ^ {i} = \operatorname {C r o p} \left(x _ {t - (r + 1)} ^ {i}, o ^ {i}\right). \tag {6} +$$ + +Then, we also execute Replace $(\cdot)$ to splice these regions: + +$$ +x _ {t - (r + 1)} ^ {r} = \operatorname {R e p l a c e} \left(x _ {t - (r + 1)}, \tilde {x} _ {t - (r + 1)} ^ {i}, o ^ {i}\right). \tag {7} +$$ + +To further improve the alignment between each region with the original prompt and enhance adjacent interactions, we perform a linear weighted recombination of $x_{t - (r + 1)}^r$ and $x_{t - (r + 1)}$ as follows: + +$$ +x _ {t - (r + 1)} = x _ {t - (r + 1)} * (1 - \delta) _ {+} x _ {t - (r + 1)} ^ {r} * \delta , \tag {8} +$$ + +where $\delta$ controls the fusion strength of image latent $x_{t - (r + 1)}$ and region-aware global latent $x_{t - (r + 1)}^r$ . The detailed process of our method can be found in algorithm 1. + +![](images/3214fe794d692ed078c1ed4712a5eb39c6c612f683385476f1c19332c1cb5516.jpg) +Figure 3. Illustration of Re-painting. Different from regular image-to-image inpainting, repainting inherits from last generation with only the target area re-initialized (upper). Given the parameters in previous generation, users are allowed to specify a target area with a new prompt and repaint the image, without relying on additional inpainting models (bottom). + +# 3.5. Image Repainting + +Based on our task setting, where a long complex prompt containing multiple objects is decoupled into several regional sub-prompts for individual processing, a natural question is raised: Can this approach be used to repaint specific areas? By leveraging the robust control and fusion capabilities of regional hard binding and regional soft refinement, we can reinitialize noise in a specific region requiring modification, enabling repainting of a region without affecting the overall layout or attributes of other areas. Different from typical post-processing inpainting task that usually requires additional inpainting models, repainting regenerates with the same parameters from the last generation, with only the prompts in the target area modified. + +As shown in Figure 3, users can modify the description of a specific region within $P$ , $\hat{p}^i$ , $\tilde{p}^i$ to obtain new edited prompts $P_e$ , $\hat{p}_e^i$ , $\tilde{p}_e^i$ . Subsequently, we encode them as text features $y_e$ , $\hat{y}_e^i$ , $\tilde{y}_e^i$ . Given a total denoising step $T$ , the initial noisy latent $x_T$ is inherited from the previous generation $x_T$ , only the target area indicated by mask is re-initialized: + +$$ +x _ {T} ^ {\prime} = \operatorname {i n i t} \left(x _ {T}, \text {m a s k}\right). \tag {9} +$$ + +To ensure the other regions intact, we simultaneously perform denoising on $x_{T}$ and $x_{T}^{\prime}$ . At each timestep $t$ , we replace the corresponding portions of $x_{t}$ with the masked + +![](images/7bf497b78ac22dbeefd4f90b16eb36d9025ba686ee605d6d1656a09fdba8e3b6.jpg) +Figure 4. Qualitative comparisons on compositional text-to-image generation. From top to bottom, we show 5 examples of different prompts and regions. Compared with previous methods, we demonstrate excellent regional control capabilities. + +areas of the edited region from $x_{t}^{\prime}$ for repainting: + +$$ +x _ {t} = \operatorname {R e p a i n t} \left(x _ {t}, x _ {t} ^ {\prime}, \text {m a s k}\right). \tag {10} +$$ + +Through the regional complementarity facilitated by regional soft refinement, the repainted area can be seamlessly integrated with the surrounding regions. This methodology not only renders image repairing feasible but also ensures the coherence and consistency of the overall scene. + +# 4. Experiments + +# 4.1. Experiment Setting + +Implementation Details. We implement our approach on Flux.1-dev [2]. The inference process is set to 20 steps, with guidance scale of 3.5. For large-scale quantitative evaluations, we employ Chain-of-Thought (CoT) template from [45] and leverage MLLM to automatically decouple multi-object scenes, see Appendix for usage. All experiments are conducted on a single A6000 GPU. + +Compared Methods. To comprehensively evaluate the generation quality, we compare our RAGD with several region control state-of-the-art approaches, including tuning-free methods: MultiDiffusion [1], RPG [45]; and tuning-based methods: InstanceDiffusion [41], Ranni [11], MIGC [58], 3DIS [59]. + +# 4.2. Main Results + +Quantitative Comparison. Table 1 presents our alignment evaluation on the GenEval [12] and T2ICompBench [16], while introducing VQAScore [22] across both benchmarks to ensure rigorous assessment of complex regional prompts. Our approach outperforms competitors in key aspects such as attribute binding, object relationships, and complex composition. Notably, despite employing a tuning-free approach, RAGD achieves an average improvement of $76.6\%$ over both RPG and MultiDiffusion in handling prompts with spatial relationships, underscoring its + +![](images/c15a061c4f23dadb38781acb3b151b8850b55e8b7693c5537627a49d9d7ec617.jpg) +A Siamese cat lying on the grass. + +![](images/b913d221cbcabe429942d356a1e1da27f3073d25bcb0fe4b50079fc2a8268d08.jpg) + +![](images/2850219a8ee2d8dffca8692a0864fc09975b722bd8789c623ba9c297660736cd.jpg) + +![](images/b5c762a087aae79015a1a36a620440d5b6919eb5f265a87f66e2afedb9e77a53.jpg) +A Corgi lying on the grass. + +![](images/92ba394079fca6b69d4d3b1fc27126b2d320e4bb15f9ab2725ae43294effc64d.jpg) + +![](images/a396dbeee00af96788f94c5b947cc82b660e5d173ed897f17502d4a9358f15ea.jpg) +A man on the left, a woman on the right. + +![](images/a621e25ae763d8a956c41693b11106aa4d0104b7c3c0475aaeb769a039571d60.jpg) + +![](images/f7c188ded28d9e68ed61fc6d184f8aaaa02f7f897fc8b1721a3e8ace8b2408bc.jpg) + +![](images/977736f72e9aaa9dfbe04328cee6eac11b068342916aadf4a6b1ceef735eee7e.jpg) +A man on the left, an anime woman on the right. + +![](images/e8ac5303285b2c818ab22186b8b9d9aaaf564eff8f4953fd0a9fdb4f28e7f464.jpg) + +![](images/9e9c1c8857e8f1bae27e806029d4e24b079444bb7d86908c08f407286ee6e1de.jpg) +A boy holding a basketball in one hand. + +![](images/6031068ee33e9725afd4cefbe2d2dac481c5d016acdee166694b6d83ea78b79d.jpg) + +![](images/16828072f7943a2bbe5e9a518036b387c647ae597d20aaf18f8464cff30de4a1.jpg) + +![](images/e373f957bc4f5675446763c91f3cda88b393bc4bddf49c45a9ceb302a52ad0f9.jpg) +A boy holding a soccer in one hand + +![](images/0f1d2fcb3a1e4985132df891d0adb0770f288391f0e6ef725900d69233778621.jpg) + +![](images/8b9fc0e56692e333264c962823acbf7cde7436de570e10f56fa389e9a78d14d0.jpg) +Three plush toys on the table. + +![](images/98ea105dedc31bc5ded9f7ac3eca98f0ad1684ade98fa4281f7884f0c9823a78.jpg) + +![](images/a219d2471f8b9e417706cae10c1c10691a11de97674d9aa664789927d013130a.jpg) +Figure 5. Qualitative comparisons on image inpainting between our RAGD and the state-of-the-art inpainting model BrushNet. Our results are more region-aware with harmonious effect with the surrounding, revealing diverse potential for applications. + +![](images/1bbcfcb2ba66444c89f5109975ba1304051b0c717081207ff32d046bcbb46aa3.jpg) + +![](images/449b0d2d3f04b36ea227a973bb66a6714fee96daa71c77787b919b3b9a26b1ee.jpg) +Two plush toys and one balloon on the table. + +
MethodsAesthetic ↑Alignment ↑
RPG [45] [ICML 2024]16.1%14.8%
Stable 3.5 [9] [ICML 2024]13.2%13.4%
Flux.1-dev [2] [2024]18.8%17.8%
Ours51.9%54.0%
+ +effectiveness in spatially accurate generation. Compared to tuning-based methods such as InstanceDiffusion, Ranni, MIGC, and 3DIS, RPG exhibits equally robust competitiveness in performance. In the VQAScore evaluations of the Geneval and T2ICompBench benchmarks, RPG achieved an average improvement of $14.0\%$ and $23.8\%$ , respectively. Quantitative results demonstrate our superior performance in handling complex multi-region prompts. + +Qualitative Comparison. Figure 4 illustrates visual comparisons, highlighting our superiority in complex multi-region generation. Our method enables accurate multi-region control via hard binding and soft refinement mechanisms, faithfully conveying details on position, quantity, and attributes based on the input text. Figure 5 shows the comparison with the state-of-the-art inpainting model BrushNet [18], showcasing our strength in inpainting both single and contacted regions without conflict. Even when the inpainted object differs greatly in style or shape, our method successfully generates the intended target while maintaining the layout and regional attributes, enhancing the flexibility and control of the generation process. + +User Study. We randomly sampled 109 prompts from the T2ICompBench test set to evaluate the aesthetic quality and text-image consistency (including attributes and spatial positions) of the generated images. During the evaluation, we invited 24 professional users to compare the generation results of RPG [45] and Stable 3.5 [9], Flux.1-dev [2], RAGD, selecting the most suitable generation results for each prompt. Through this blind test, we obtained 2,616 + +Table 2. User study on aesthetics and text-image alignment. Our approach shows a significant improvement. + +
StageT2ICompBench
Harding BindingSoft RefinementColorShapeTextureSpatialComplexVQAScore
✓0.78280.56830.66170.28210.35590.7656
0.72380.50120.58320.49390.38970.8413
✓0.80390.60160.70850.51930.43770.8907
+ +Table 3. Ablation performance of RAGD. Regional soft refinement helps refine attributes and merge boundaries, and regional hard binding is beneficial for accurate spatial positioning of complex layouts. + +![](images/765682427a45113f9b5206dd21370ab75e838b2ee68678d816b12d98ca7338d9.jpg) + +![](images/dc0e9d0468d8637cf06bb0b7da05a5f9d5c12a2b1903775ab3e02b44e23afef8.jpg) + +![](images/6530b81990edab2d56e15d60c1ddffe8dfe8fc58175ad1905d0f91b2ac27628c.jpg) + +![](images/ec1e7a1b55f31f6612012d5e07762c3c9b127346e83982d121aca4d250ee3b10.jpg) + +![](images/9ac765e4936a30d5669b90f46a0e390ca3fafa8d6013f3a6c1a86049d22c2cd0.jpg) +w/o Hard Binding w/o Soft Refinement +Figure 6. Qualitative analysis of Hard Binding and Soft Refinement. The former ensures the proper responses of each region, while the latter enhances the coherence among regions. + +![](images/c5638cf208926de064d75038ffc98a6ccc114996012cd57d71f036a4769d98e0.jpg) +w/Hard Binding w/o Soft Refinement + +![](images/989f80751a0668e4bd8067ed6e4f389fa8b54b47e8a6cdad532d209dc6af069c.jpg) +W/o Hard Binding W/ Soft Refinement + +![](images/91615e005f16e8a6004a22ea36b6a6e9a66a9f167228ccbe2c1bb8179a0435e7.jpg) +w/ Hard Binding w/ Soft Refinement + +user votes. The results are shown in Table 2. RAGD is far ahead in terms of both aesthetics and alignment. + +# 4.3. Ablation Study + +Effectiveness of Hard Binding & Soft Refinement. The proposed RAGD comprises two key components: hard binding and soft refinement. We conducted ablation experiments on these two parts, with visual results shown in Figure 6. Table 3 shows the ablation performance of the two stages. Results reveal that hard binding achieves precise object positioning but yields coarse attribute rendering and limited integration among objects. With soft refine- + +
T2ICompBench
Color ↑Shape ↑Texture ↑Spatial ↑Complex ↑VQAScore ↑
SD3.50.78700.61010.70680.26600.33100.8393
RPG+SD3.50.73430.54360.65910.40080.34360.8340
Ours+SD3.50.82220.63040.73750.43600.36230.8543
Flux0.76800.50780.61950.26060.36500.7438
RPG+Flux0.70330.52240.62310.43930.35790.8213
Ours+Flux0.80390.60160.70850.51930.43770.8907
+ +Table 4. RAGD can seamlessly integrate various text-to-image models and exhibits superior enhancement capabilities compared to other solutions. + +![](images/b50d131390f7b5b3a81e66d10b3b36a31b364ea8e658a20551dc2db9cf3437b4.jpg) +Figure 7. Our method shows strong visual competitiveness in different frameworks. + +ment, object attributes are richly detailed and relationships are harmonious, though precise positioning may be compromised, and object omission may sometimes occur. + +Generalizability of the method. As shown in Table 4 and Figure 7, RAGD can be migrated to different T2I models and demonstrates stronger enhancement effects compared to other approaches. Additionally, as illustrated in Figure 8, RAGD possesses strong generalization abilities and is compatible with plug-and-play modules such as LoRA [15, 33, 34] and IPAdapter [13, 47]. The integration of these expansion modules enriches the diversity of generated results, improving content detail representation while further enhancing the adaptability and scalability of RAGD. + +Analysis of Inference Time. In RAGD, the decoupled multi-region generation operation processes each region independently, leading to increased inference time as the number of regions grows. However, because RAGD does not modify the underlying architecture of the original model, existing acceleration frameworks, such as HyperFlux [29], remain compatible with RAGD. As shown in Figure 9, combining RAGD with Hyper-Flux results in inference being 2.5-fold faster while preserving the stability of accuracy and detail control in regional generation. This demonstrates that, although inference speed presents a challenge, with appropriate acceleration strategies, RAGD can efficiently generate multi-region images with high quality and maintain control over interactions between regions. + +![](images/5fb99aa133cd4ab2197338a26f216969ac786f0c5efd4ffc792e12cd7e0acf41.jpg) +Figure 8. Ours can work with different plugins and provide region control capabilities to increase usability. + +![](images/099498b1a7e3d7d0962579390e64259196b99f37672ac02c73f57a10871789fb.jpg) + +![](images/447f4604e4b177eb58f9c7fc25f3f7d9de70d60ee84f3195b61344dbc0012704.jpg) +From left to right: a red cake, an orange cake, a yellow cake, and a green cake. + +![](images/708a7cdcc79eda022542932cbd2de73c36cb5b0a717bcd980198b9f90c0b15dd.jpg) +Figure 9. RAGD, accelerated by Hyper-Flux, is $2.5 \times$ faster than the original version, maintaining visual quality. + +# 5. Conclusion + +In this paper, we introduce RAGD, a novel tuning-free region-aware diffusion model for text-to-image generation, designed to address the challenges posed by regional prompts. RAGD operates through two key stages. First, regional hard binding independently constructs content within each designated region. Subsequently, regional soft refinement enhances interactions between adjacent regions while improving attribute generation accuracy. Furthermore, RAGD bypasses the need for external inpainting models, enabling direct image painting to modify unsatisfactory regions from last generations. Extensive experimental results demonstrate the superiority of RAGD in compositional generation compared to prior methods. + +Limitation and Future Work. RAGD offers precise regional control and flexible image repairing but has limitations. Its multi-region decoupling increases inference time as region count grows. Future work will focus on improving RAGD's inference efficiency and integrating with other diffusion models for enhanced scalability and performance. + +# Acknowledgment + +This work was supported by Natural Science Foundation of China: No. 62406135, Natural Science Foundation of Jiangsu Province: BK20241198, the Gusu Innovation and Entrepreneur leading Talents: No. ZXL2024362, and Nanjing University-China Mobile Communications Group Co. Ltd. Joint Institute. + +# References + +[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 2, 3, 5, 6 +[2] BlackForest. Black forest labs; frontier ai lab, 2024. 2, 3, 6, 7 +[3] 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. 2, 3 +[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2 +[5] Zhennan Chen, Rongrong Gao, Tian-Zhu Xiang, and Fan Lin. Diffusion model for camouflaged object detection. In ECAI 2023, pages 445-452. IOS Press, 2023. 2 +[6] Nikai Du, Zhennan Chen, Zhizhou Chen, Shan Gao, Xi Chen, Zhengkai Jiang, Jian Yang, and Ying Tai. Textcrafter: Accurately rendering multiple texts in complex visual scenes. arXiv preprint arXiv:2503.23461, 2025. 2 +[7] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360, 2021. 2 +[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. 2 +[9] 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, 3, 7 +[10] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[11] Yutong Feng, Biao Gong, Di Chen, Yujun Shen, Yu Liu, and Jingren Zhou. Ranni: Taming text-to-image diffusion for accurate instruction following. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4744-4753, 2024. 2, 5, 6 + +[12] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36:52132-52152, 2023. 5, 6 +[13] Zinan Guo, Yanze Wu, Chen Zhuowei, Peng Zhang, Qian He, et al. Pulid: Pure and lightning id customization via contrastive alignment. Advances in Neural Information Processing Systems, 37:36777-36804, 2025. 8 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[15] 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. 8 +[16] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 5, 6 +[17] Álvaro Barbero Jiménez. Mixture of diffusers for scene composition and high resolution image generation. arXiv preprint arXiv:2302.02412, 2023. 3 +[18] 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. 7 +[19] 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, pages 7701-7711, 2023. 3 +[20] Caoshuo Li, Zengmao Ding, Xiaobin Hu, Bang Li, Donghao Luo, AndyPian Wu, Chaoyang Wang, Chengjie Wang, Taisong Jin, Yunsheng Wu, et al. Oraclefusion: Assisting the decipherment of oracle bone script with structurally constrained semantic typography. arXiv preprint arXiv:2506.21101, 2025. 2 +[21] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2 +[22] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366-384. Springer, 2024. 6 +[23] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 2 +[24] Omost-Team. Omost github page, 2024. 2, 3 +[25] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 3 + +[26] 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 +[27] 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. 2 +[28] 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. 2 +[29] 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. arXiv preprint arXiv:2404.13686, 2024. 8 +[30] 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 +[31] 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. 2 +[32] 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 +[33] Shakker-Labs. Myth wukong lora, 2024. 8 +[34] Shakker-Labs. Wool yarn lora, 2024. 8 +[35] 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. 2 +[36] 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 +[37] Kolars Team. Kolars: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 2 +[38] 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 preprint arXiv:2404.02733, 2024. 2 +[39] Haofan Wang, Peng Xing, Renyuan Huang, Hao Ai, Qixun Wang, and Xu Bai. Instantstyle-plus: Style transfer with content-preserving in text-to-image generation. arXiv preprint arXiv:2407.00788, 2024. + +[40] 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 +[41] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6232-6242, 2024. 2, 5, 6 +[42] X Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance. arXiv preprint arXiv:2406.07209, 2024. 2 +[43] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 2 +[44] 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 +[45] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 5, 6, 7 +[46] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 2 +[47] 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, 8 +[48] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2 +[49] 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 +[50] Xuying Zhang, Xiaoshuai Sun, Yunpeng Luo, Jiayi Ji, Yiyi Zhou, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Rstnet: Captioning with adaptive attention on visual and non-visual words. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15465–15474, 2021. 2 +[51] Xuying Zhang, Yutong Liu, Yangguang Li, Renrui Zhang, Yufei Liu, Kai Wang, Wanli Ouyang, Zhiwei Xiong, Peng Gao, Qibin Hou, et al. Tar3d: Creating high-quality 3d assets via next-part prediction. arXiv preprint arXiv:2412.16919, 2024. + +[52] Xuying Zhang, Bo-Wen Yin, Yuming Chen, Zheng Lin, Yunheng Li, Qibin Hou, and Ming-Ming Cheng. Temo: Towards text-driven 3d stylization for multi-object meshes. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 19531-19540, 2024. +[53] Xuying Zhang, Yupeng Zhou, Kai Wang, Yikai Wang, Zhen Li, Shaohui Jiao, Daquan Zhou, Qibin Hou, and Ming-Ming Cheng. Ar-1-to-3: Single image to consistent 3d object generation via next-view prediction. arXiv preprint arXiv:2503.12929, 2025. +[54] Chen Zhao, Weiling Cai, Chenyu Dong, and Chengwei Hu. Wavelet-based fourier information interaction with frequency diffusion adjustment for underwater image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8281-8291, 2024. +[55] Chen Zhao, Weiling Cai, Chengwei Hu, and Zheng Yuan. Cycle contrastive adversarial learning with structural consistency for unsupervised high-quality image deraining transformer. Neural Networks, 178:106428, 2024. +[56] Chen Zhao, Wei-Ling Cai, and Zheng Yuan. Spectral normalization and dual contrastive regularization for image-to-image translation. The Visual Computer, 41(1):129-140, 2025. 2 +[57] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Migc++: Advanced multi-instance generation controller for image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[58] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 2, 5, 6 +[59] Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3dis: Depth-driven decoupled instance synthesis for text-to-image generation. arXiv preprint arXiv:2410.12669, 2024. 5, 6 +[60] Dewei Zhou, Mingwei Li, Zongxin Yang, and Yi Yang. Dreamrenderer: Taming multi-instance attribute control in large-scale text-to-image models. arXiv preprint arXiv:2503.12885, 2025. 3 \ No newline at end of file diff --git a/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/images.zip b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6d709a7292c9421ef5450168d7f930ca674ac103 --- /dev/null +++ b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:745cdd5f4bce3909cc0d8272b11be4f89d50f267ba508166e22cbf8b71f86672 +size 995251 diff --git a/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/layout.json b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..649c03169894a06ed1466e00894f2e4c197d747d --- /dev/null +++ b/ICCV/2025/RAGD_ Regional-Aware Diffusion Model for Text-to-Image Generation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee2b219ce492f8d3811bf3c7a0037a5872260fca58ec0dabb8a00bfc44562a0c +size 459855 diff --git a/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/e475c992-d95f-4c85-b7c9-c82b0c87a8ec_content_list.json b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/e475c992-d95f-4c85-b7c9-c82b0c87a8ec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..247a1c3c5933dd470f297bf429016fc75242d52c --- /dev/null +++ b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/e475c992-d95f-4c85-b7c9-c82b0c87a8ec_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e0678bb6bd71178de2d7984506fcb74b35a131c169eca09128baa21fe625a29 +size 79601 diff --git a/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/e475c992-d95f-4c85-b7c9-c82b0c87a8ec_model.json b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/e475c992-d95f-4c85-b7c9-c82b0c87a8ec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bbd713e861c6ee976146742999aa51a9bfdeccdb --- /dev/null +++ b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/e475c992-d95f-4c85-b7c9-c82b0c87a8ec_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:255c0a1ecb5ce7baced71330aff8b816cb1e5d64b9cfb27ff41d945260c6e593 +size 100147 diff --git a/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/e475c992-d95f-4c85-b7c9-c82b0c87a8ec_origin.pdf b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/e475c992-d95f-4c85-b7c9-c82b0c87a8ec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..74213f340fe440b9c27318e2d7a93edae1cc8ec2 --- /dev/null +++ b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/e475c992-d95f-4c85-b7c9-c82b0c87a8ec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97eb55fd31d1650303b7f889992ffd8e503cf0891d68c6347476e8efd7f5991c +size 8864845 diff --git a/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/full.md b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7a4aa35de5144fb3bb8c0ed0204c6e0c3fdcf430 --- /dev/null +++ b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/full.md @@ -0,0 +1,281 @@ +# RAGDiffusion: Faithful Cloth Generation via External Knowledge Assimilation + +Yuhan Li $^{1*†}$ , Xianfeng Tan $^{2*}$ , Wenxiang Shang $^{2}$ , Yubo Wu $^{2}$ , Jian Wang $^{2}$ , Xuanhong Chen $^{1}$ , Yi Zhang $^{1}$ , Hangcheng Zhu $^{2‡}$ , Bingbing Ni $^{1‡}$ + +$^{1}$ Shanghai Jiao Tong University $^{2}$ Alibaba Group + +https://colorful-liyu.github.io/RAGDiffusion-page/ + +# Abstract + +Standard clothing asset generation involves restoring forward-facing flat-lay garment images displayed on a clear background by extracting clothing information from diverse real-world contexts, which presents significant challenges due to highly standardized structure sampling distributions and clothing semantic absence in complex scenarios. Existing models have limited spatial perception, often exhibiting structural hallucinations and texture distortion in this high-specification generative task. To address this issue, we propose a novel Retrieval-Augmented Generation (RAG) framework, termed RAGDiffusion, to enhance structure determinacy and mitigate hallucinations by assimilating knowledge from language models and external databases. RAGDiffusion consists of two processes: (1) Retrieval-based structure aggregation, which employs contrastive learning and a Structure Locally Linear Embedding (SLLE) to derive global structure and spatial landmarks, providing both soft and hard guidance to counteract structural ambiguities; and (2) Omni-level faithful garment generation, which introduces a coarse-to-fine texture alignment that ensures fidelity in pattern and detail components within the diffusing. Extensive experiments on challenging real-world datasets demonstrate that RAGDiffusion synthesizes structurally and texture-faithful clothing assets with significant performance improvements, representing a pioneering effort in high-specification faithful generation with RAG to confront intrinsic hallucinations and enhance fidelity. + +# 1. Introduction + +While the high quality of images generated by diffusion models [20, 45] has significantly lowered the barriers for individuals to engage in content creation, the generation of 2D standard clothing assets from in-the-wild data has been relatively underexplored [43]. Standard clothing asset gen + +![](images/ef8c238a23b454c65877eea4c3736a2f2596d093e7f759e17f907ecec90ca13b.jpg) + +![](images/7cd74fb3ddda5fea273e9c065efc91acadbe6baa12b0a967f02228a44be18b4d.jpg) +Figure 1. RAGDiffusion synthesizes structurally and pattern-wise faithful standard garments by assimilating retrieved knowledge. + +eration refers to generating forward-facing flat-lay garment images [49] on a clear background by recovering clothing information from arbitrary real-world scenes (e.g., garments hung on hangers, worn by models, or casually laid on chairs in Fig. 1). Standard garments serve as a crucial intermediary variable, connecting various downstream applications such as e-commerce catalogs maintenance, garment design, product marketing and virtual fitting [19, 64, 65]. + +The garment generation task presents dual technical challenges distinct from conventional virtual try-on paradigms. First, standard clothing generation necessitates standard structure and frontal positioning while inherently limiting stylistic flexibility, constrained by the narrow manifold of high-specification garment distributions [62]. Second, the complexity of in-the-wild condi + +tions involving occlusions, multi-layer outfits, truncations, or extreme lateral viewpoints notably exacerbates structural uncertainty and pattern ambiguity. Collectively, experimental evaluations reveal that existing methods [21, 51] fail to achieve practical usability thresholds with structural and texture distortion, particularly manifesting a pathological structure hallucinations in inaccurate structure, fitness and length estimation under challenging scenes (see Fig. 3 and Fig. 5), which we attribute to the limited spatial reasoning capacity inherent in Stable Diffusion (SD) architectures. + +Despite its critical commercial relevance, the garment restoration problem remains insufficiently addressed in literature [43, 49, 51, 57]. Early GAN-based frameworks [57] suffer from pattern distortions constrained by their generative priors. Recent approaches [43, 49, 51] employing pretrained SD exhibit several limitations: (i) inadequate handling of structural hallucinations in hard cases (occluded, multi-layer, etc.), (ii) commercially unfaithful logo/texture reproduction which is extremely sensitive to e-commerce sellers. In summary, no existing solution achieves the dual objectives of high-structural integrity and photorealistic texture synthesis under real-world complexity. + +We propose RAGDiffusion, as a Retrieval-Augmented Generation (RAG) [38] paradigm including two processes: information aggregation based on retrieval, and conditional generation with omni-level fidelity, to address the issue of structural and texture distortion respectively. (1) Retrieval-based Structure Aggregation: Our key insight lies in enhancing structural determinacy through the assimilating of external structural landmarks and world knowledge, thereby rectifying visual models' misconceptions regarding garment structure, fitness and proportions. During RAG process, contrastive learning [10, 18] is firstly introduced to train a dual tower network to extract multi-modal structure embeddings from images of two branches as well as attributes derived from a frozen large language model (LLM) [1, 2]. Considering potential encoding bias in real-world data, we propose a structure retrieval named Structure Locally Linear Embedding (SLLE) to project the predicted structure embedding towards a standard manifold [41, 56] as well as offer a silhouette landmark. The remapped latent structure embedding and the landmarks provide soft and hard structure guidance respectively through Embedding Prompt Adapter and Landmark Guider to eliminate structural hallucinations. (2) Omni-level Faithful Garment Generation: RAGDiffusion establishes coarse-to-fine texture alignment for pattern faithfulness and detail faithfulness for generated garment. Specifically, we ensure the generated pattern matches the conditioning through ReferenceNet [21], a conditioning UNet isomorphic to the denoising UNet. However, problematic detail and logo artifacts persist partially due to resolution limitations of VAE decoders, rendering the generated images commercially un + +usable for e-commerce sellers, which is an under-addressed issue across prior research [43, 49, 51]. To mitigate reconstruction distortions inherent in original VAE [25], we propose Parameter Gradual Encoding Adaptation (PGEA) to adapt the SDXL [35] backbone to a more powerful VAE. + +To the best of our knowledge, RAGDiffusion stands as a pioneering multi-modal RAG method to solve intrinsic hallucination and unfaithfulness during image synthesis. Furthermore, we discover two additional benefits brought by RAG: zero-shot generalizability for unseen scenarios via retrieval database expansion, and human-interpretable control mechanisms through landmark manipulation. The core contributions of this work are threefold: + +- We propose RAGDiffusion, a RAG framework for clothing generation with representation learning and SLLE to offer structure determinacy and eliminate hallucination. +- We employ a coarse-to-fine texture alignment in RAGDiffusion, addressing pattern and detail aspects, ensuring faithfulness of clothing on this high-specification task. +- Comprehensive experiments on challenging in-the-wild sets demonstrate RAGDiffusion is capable of synthesizing both structure and texture faithful clothing assets, outperforming current methods by a substantial margin. Furthermore, the ablation study has validated the effectiveness of different parts of RAGDiffusion. + +# 2. Related works + +Retrieval-augmented generative models. Retrieval-augmented strategies leverage external databases to enhance the capabilities of generative models across a variety of tasks. For instance, the RETRO [5] modifies an LLM to effectively utilize external databases. In the realm of image synthesis, retrieval has been employed in both GANs [7, 47] and diffusion models [4, 44], accommodating 3D generation [42], video generation [60] and artistic styles [40]. The essence of these works lies in retrieving similar images to serve as mimetic references for generating specified content, particularly in cases where models are insufficiently trained. Our RAGDiffusion not only builds upon this generalizability through convenient retrieval expansion, but also aggregates structural information, thereby mitigating intrinsic hallucinations on the high-standard tasks. + +Controllable text-to-image diffusion models. To attain conditional control in text-to-image diffusion models, ControlNet [59], T2I-Adapter [31] and IP-Adapter [55] incorporate additional trainable modules to fuse conditions on feature maps. Additionally, recent investigations have utilized a variety of prompt engineering techniques [26, 54, 63] and implemented cross-attention constraints [9, 23, 52, 66] to facilitate more controllable generative processes. Furthermore, some studies further investigate multi-condition or multi-modal generation [22, 27, 36, 67]. However, these methods rely on the associative capabilities of stable diffu + +![](images/e456df3062dd684e285ca414813efd44f033df272969bf3172346520c92a22f7.jpg) + +![](images/37be8abec8fa2faeb7f91a08c290fec093f1f7788d19d3a90d349eeebe566a0d.jpg) + +![](images/7a683cb64f9c4a1773a3998332df7369bf32b4ef07f48b90c91b3b9cbf464f7e.jpg) +Figure 2. Overall framework of RAGDiffusion. StructureNet provides latent structural embeddings, while SLLE facilitates the embedding fusion along with landmark retrieval. The generative model synthesizes multiple conditions to achieve omni-level high-fidelity generation. + +![](images/e11ce2dac8d5219d7052d79bf86fbef208f11424109c0fe278c99c9e3096414c.jpg) + +sion, which cannot guarantee fidelity, particularly with the highly standardized sampling of standard garments. + +Garment restoration. Standard clothing generation involves restoring flat-lay garment images on from real-world contexts. TileGAN [57] pioneered a two-stage GAN [16] framework, suffering from backbone capability limitations. Recent works including TryOffDiff [49], TryOffAnyone [51], and IGR [43] have adopted pretrained SD as base backbones, yielding obvious improvements through enhanced generative priors. Specifically, TryOffDiff utilizes SigLIP [58] to extract garment features; TryOffAnyone employs direct image concatenation for conditional generation; IGR implements ReferenceNet-inspired [21] architectures for pattern consistency. However, these methods exhibit three critical shortcomings: (1) They ignore structural hallucinations and distortions arising from real-world complexities (multi-layer, extreme viewpoints, etc.). (2) Their detail fidelity fails to meet commercial standards. (3) Adaptation to new garment categories/scenarios requires costly paired data and retraining. Our RAGDiffusion systematically resolves these limitations through RAG paradigm and a coarse-to-fine alignment pipeline. + +# 3. Method + +An overview of the RAGDiffusion is presented in Fig. 2. The backbone of RAGDiffusion employs the SDXL [39], with the preliminary detailed in Appendix. Given an inthe-wild clothing image $x_{itw} \in \mathbb{R}^{H \times W \times 3}$ , RAGDiffusion + +is aimed to generate an authentic standard flat lay in-shop garment image $x_{std}$ . A dual tower StructureNet extracts latent structure embeddings by contrastive learning as detailed in Sec. 3.1, and SLLE retrieves and fuses structure embeddings and landmarks as in Sec. 3.2. Coarse-to-fine alignment generation involves ReferenceNet, and PGEA for pattern, and decoding faithfulness described in Sec. 3.3. + +# 3.1. Dual-tower embeddings extraction + +To eliminate structural hallucinations, a straightforward approach is to extract structural features from images and feed them into generative networks, as seen in StyleGAN [24] and LADI-VTON [30], which use additional latent coding to module conditions. Accordingly, we extract latent structure embeddings through contrastive learning [37], using garments that share similar canny but are texture-dissimilar as training pairs. This process exploits the structural similarities between specialized pairs, which are not emphasized during the conventional training of diffusion. + +Contrastive learning. Given a batch of $N$ (in-the-wild clothing $x_{itw}$ , standard clothing $x_{std}$ ) pairs, StructureNet $g_{\theta}$ is trained to predict which of the $N \times N$ possible ( $x_{itw}$ , $x_{std}$ ) pairings across a batch actually occurred according to structure similarity. To do this, StructureNet learns a multimodal embedding space by jointly training a dual tower encoder tailored for $x_{itw}$ and $x_{std}$ to maximize the cosine similarity of the embeddings of the $N$ real pairs (marked with superscript +) in the batch while minimizing the similarity of the $N^2 - N$ incorrect pairs (marked with superscript + +-). To be specific, the LLM is used to extract 10 types of discrete attributes of clothing (e.g., Category, FitType, CollarTechnique, IfTopTuckIn), which incorporates general in-context image knowledge from the language model [1, 2], eliminating semantic deficiencies and ambiguities present in vision. Then we assign a learnable embedding $f_{attr}$ to each discrete attribute. The image structure features $f_{img}$ are extracted by the twin ViT [14] encoder and concatenated with the attributes embeddings $f_{attr}$ to form the final latent structure embeddings $e$ after non-linear Resampler [55] head. We optimize the InforNCE loss [10] over these latent structure embeddings $\{e_{itw}, e_{std}\}$ from $\{x_{itw}, x_{std}\}$ as: + +$$ +\mathcal {L} = \frac {- 1}{N} \sum_ {i = 1} ^ {N} \log \frac {\exp \left(e _ {i t w , i} \odot e _ {s t d} ^ {+} / \tau\right)}{\exp \left(e _ {i t w , i} \odot e _ {s t d} ^ {+} / \tau\right) + \sum_ {e _ {s t d} ^ {-}} \exp \left(e _ {i t w , i} \odot e _ {s t d} ^ {-} / \tau\right)}, \tag {1} +$$ + +where $\odot$ is cosine similarity between two vectors, $N$ is the batch size, and $\tau$ is a temperature scalar. + +# 3.2. Retrieval-augmented SLLE + +By integrating structural knowledge from StructureNet into the SD, we have improved the style and structure of generated flat-lay clothing as shown in Sec. 4.3. However, due to the limited spatial perception inherent in the SD [6, 28], it often struggles to accurately represent the length and contours of clothing, particularly in hard cases (occlusions, multi-layer, lateral viewpoints, etc.), as in Fig. 5. Furthermore, since StructureNet is trained on a limited set of contrastive pairs, it may perform poorly with out-of-distribution samples, making it unreliable for real-world applications. + +We have observed that creating a comprehensive database of standard flat-lay clothing for retrieval is easier than gathering comprehensive in-the-wild samples (numerous scenarios). In this context, we set standard clothing $e_{std}$ as basic vectors and propose Structure Locally Linear Embedding (SLLE), which is a manifold projection [41] to correct each predicted structure embedding $e_{itw}$ into the target space of standard embedding $e_{std}$ to mitigate potential error. Meanwhile, the retrieval database provides structure landmarks to strengthen explicit spatial constraints. + +Memory database. To execute SLLE, we establish a retrieval memory database. We first encode the collected standard flat-lay clothing image dataset into a series of latent structure embeddings $e_{std}$ with StructureNet $g_{\theta}$ . The silhouette landmarks $L_{sil}$ are also extracted to designate areas for generated content. Thus an external memory database $\mathcal{D}$ that consists of embedding-landmark pairs $(e_{std}, L_{sil})$ is obtained. These structure embeddings are utilized in matching algorithms [17], enabling a given in-the-wild garment to find the most compatible standard features $e_{std}$ and landmarks during inference. Notice that landmarks can be any structural figure. Here, we use the outline of clothing, primarily to tackle the challenges of limited spatial perception in SD (e.g. length and contour) [6, 28]. + +Structure LLE algorithm. In this process, retrieval-augmented SLLE drags in-the-wild embedding closer to the standard flat-lay garment embeddings to void outliers during inference, as well as offer silhouette landmarks. Motivated by the successful practice of classic locally linear embedding in [3], we assume that the garment embedding or landmark and its nearby points are locally linear on the manifold, eliminating the need to project them into a higher-dimensional space as vanilla LLE [41] does. Specifically, given an extracted garment embedding $e_{itw}$ , the goal of SLLE is to reconstruct embedding $e_{itw}'$ with standard embeddings $e_{std}$ as basis vectors. We start by searching the $K$ nearest standard embeddings $\{e_{std}^1, \dots, e_{std}^K\}$ from the standard garment memory database $\mathcal{D}$ using cosine similarity. Thus $K$ corresponding flat-lay cloth silhouette landmarks $\{L_{sil}^1, \dots, L_{sil}^K\}$ are obtained as well. Next, a linear combination of these neighbors is sought to reconstruct $e_{itw}'$ by minimizing the error $\| e_{itw}' - e_{itw} \|_2$ , which could be formulated as the least-squared optimization problem: + +$$ +\min \left\| e _ {i t w} - \sum_ {i = 1} ^ {K} w _ {i} \cdot e _ {s t d} ^ {i} \right\| _ {2}, \quad s. t. \sum_ {i = 1} ^ {K} w _ {i} = 1, \tag {2} +$$ + +where $w_{i}$ is the barycentric weight of the $i$ -th nearest embedding $e^{i}_{std}$ . The optimal weights $\{w_{1},\ldots ,w_{K}\}$ can be determined by solving Eq. (2). Subsequently, we can reconstruct the shape embedding as $e^{\prime}_{itw} = \sum_{i = 1}^{K}w_{i}^{*}\cdot e^{i}_{std}$ . Ideally, the reconstructed embedding $e^{\prime}_{itw}$ serves as an indomain data point that retains structural accuracy and appropriate fitness, albeit with some information loss. In practice, we use a linear combination of original structure embedding $e_{itw}$ and reconstructed one $e^{\prime}_{itw}$ as the final structure representation during inference: + +$$ +\hat {e} _ {i t w} = \alpha \cdot \sum_ {i = 1} ^ {K} w _ {i} ^ {*} \cdot e _ {s t d} ^ {i} + (1 - \alpha) \cdot e _ {i t w}, \tag {3} +$$ + +where $\alpha \in [0,1]$ controls the trade-off and is set to be 0.5. The final landmark $L_{\text{sil}}$ is also fused with optimal weights. Structure conditioning. Inspired by IP-Adapter [55], we adopt a similar Embedding Prompt Adapter (EP-Adapter) to condition the high-level structure semantics. While maintaining the text branch unchanged, fused structure embeddings $\hat{e}_{itw}$ after SLLE are fed into additional projection layers to generate key and value matrices for the structure representations. Two parallel cross-attention layers process the text modality and the structure embedding modality, with results being summed to replace the original single text one: Attention $(Q,K_{\text{text}},V_{\text{text}})+$ Attention $(Q,K_{\text{emb}},V_{\text{emb}})$ . The EP-Adapter enhances the global/inner structural faithfulness by incorporating prior knowledge from LLM and structural training pairs as a soft constraint. Meanwhile, the contour landmark $L_{\text{sil}}$ involves external spatial structural information. A Landmark Guider [21] with 4 convolution layers ( $4\times 4$ kernels, $2\times 2$ strides, 16, 32, 64, 128 channels) is incorporated to align the silhouette mask with $z_{t}$ as an explicit and hard constrain. + +# 3.3. Omni-Level faithful garment generation + +Pattern faithfulness. Inspired by success in human editing [8, 21], we introduce an additional UNet encoder (i.e. ReferenceNet [21]) to precisely preserve the fine-grained details of clothing assets, which is isomorphic to the main generative UNet (i.e. MainNet) and shares same initial parameter weights. The latent of in-the-wild garment image is passed through ReferenceNet to obtain the intermediate key and value features $\{z_R^k,z_R^v\} \in \mathbb{R}^{N\times l\times d}$ in self-attention, which are concatenated with the features $\{z_M^k,z_M^v\} \in \mathbb{R}^{N\times l\times d}$ from the MainNet along the $l$ dimension to obtain the final $\{z_C^k,z_C^v\} \in \mathbb{R}^{N\times 2l\times d}$ . Then we compute the self-attention on the concatenated features as: + +$$ +\text {A t t e n t i o n} \left\{z _ {M} ^ {q}, z _ {C} ^ {k}, z _ {C} ^ {v} \right\} = \operatorname {s o f t m a x} \left(\frac {z _ {M} ^ {q} z _ {C} ^ {k T}}{\sqrt {d}}\right) z _ {C} ^ {v}, \tag {4} +$$ + +where $z_{M}^{q}$ represents Query features in self-attention from MainNet. ReferenceNet plays an important role in texture preserving of garments when it has complicated patterns. + +Detail faithfulness. The cloth restoration technique primarily serves e-commerce sellers for product catalog maintenance and marketing purposes. The requirement for brand logo fidelity is so stringent that existing methods' results [43, 49, 51] are practically unusable even with ReferenceNet, partially because of the fine-grained logo decoding degradation caused by the high compression ratio of SDXL's VAE, as shown in Fig. 5. Although the VAE in FLUX [15] greatly alleviates the loss by inflating latent channels, challenges intrinsic to the DiT framework [34], such as slow convergence, and high data requirements, hinder its widespread adoption in the community. + +To address this, we propose a versatile three-stage parameter gradual encoding adaption (PGEA) to align the SDXL UNet with the FLUX VAE. Specifically, we expand the channels in conv in and conv out layers (from 4 to 16) in UNet to match FLUX VAE, enabling direct modification on the SD config file to load adapted weights. Stage I: we focus on distilling the knowledge from the original conv in layer to the adapted conv in layer. The input image is encoded through different VAEs and passed into the 4 and 16-channel conv in layers respectively, applying a reconstruction loss between their output to update the 16-channel conv in layer for fast adaption. Stage II: we train the UNet on a standard text-to-image task for 20,000 steps, where only the 16-channel conv in and conv out layers are updated. Stage III: we train and update the entire UNet on the standard text-to-image task for 200,000 steps. The complete training of PGEA lasts for 4 days on 8 H20 GPUs. It is noteworthy that the adapted general UNet with extremely low decoding loss can be applied to various downstream tasks. Crucially, despite marginal gains in quantitative metrics, RAGDiffusion's breakthrough in brand logo fidelity makes it commercially viable, the first in industrial garment restoration. + +![](images/12560db60f9c062404e653fd8509b868666a4b3f13b69350ecc2a2a53604bd49.jpg) +Figure 3. RAGDiffusion delivers both loyal structures and superior details on challenging layered and side-view situations. + +
MethodSSIM ↑LPIPS ↓FID ↓KID ↓DISTS ↓
PBE0.52780.482126.5510.080.352
ControlNet0.52250.479223.029.2310.345
IP-Adapter0.60670.472814.503.7470.264
TryOffDiff0.61240.500734.3414.190.335
TryOffAnyone0.65750.433518.867.9190.245
ReferenceNet0.65460.397910.701.3990.224
RAGDiffusion0.69630.36849.9901.0920.192
+ +Table 1. Quantitative comparisons on the STGarment dataset. + +# 4. Experiments + +# 4.1. Experimental setup + +Datasets. We collected 65,131 pairs of (in-the-wild upper clothing, standard flat lay upper clothing) for training and 1969 pairs for testing, named STGarment. Among the in-the-wild clothing, there are three main displays: clothing worn on a person, clothing laid indoors, and clothing hung on hangers. We use Qwen2-VL-7B [2] to provide prompt annotations for each pair along with 10 discrete attributes (3.5 seconds per image). Additionally, we matched each pair of images with the most structurally similar flatlay clothing for StructureNet training based on canny image similarity. We employed a clustering approach to filter and construct a memory database for retrieval, which contains 4,000 embedding-landmark pairs from the training set. Please refer to the Appendix for more details. + +Implementation details. We initialize the RAGDiffusion with a pre-trained SDXL model and train it on STGarment using an AdamW optimizer with the learning rate of $5e - 5$ at a resolution of $768 \times 768$ . The models are trained for 5 + +
MethodSSIM ↑LPIPS ↓FID ↓KID ↓DISTS ↓
Vanilla Model0.65760.396110.651.4050.222
+ emb0.66140.391710.551.3540.217
+ emb + sil0.69260.369010.220.9400.210
Generating mask0.64260.418614.553.5700.239
+ emb + sil + PGEA (Full)0.69630.36849.9901.0920.192
+ +Table 2. Quantitative ablation study of each component. + +
DatasetViton-HDDC-upperDC-lowerDC-dress
MethodFIDKIDFIDKIDFIDKIDFIDKID
PBE80.852.945.021.1158.8128.3105.174.6
ControlNet69.943.248.923.6174.7161.7118.395.4
IP-Adapter46.817.828.811.1111.362.747.819.0
TryOffDiff*14.03.3138.916.7158.7130.074.348.6
TryOffAnyone*11.41.9724.49.30114.279.638.420.4
ReferenceNet15.23.2317.86.24124.590.251.222.4
Ours12.31.4515.94.5140.519.423.16.22
+ +Table 3. Cross dataset evaluation on Viton-HD, DressCode. Model with * means inner-dataset test which is trained on Viton-HD. + +
ScaleTop-1 Acc. ↑Top-5 Acc. ↑IOU ↑
100076.6%93.4%0.872
200078.2%93.2%0.885
400080.6%93.7%0.902
800079.7%93.4%0.905
+ +Table 4. Retrieval accuracy at different scales of external database. + +days on 8 H20 GPUs with DeepSpeed [29] ZeRO-2 to reduce memory usage, at a batch size of 15. $K = 4$ in SLLE in Eq. (2). The StructureNet (i.e. embedding encoder) utilizes CLIP-ViT-L/14 image encoder as the backbone and is fine-tuned on nearly 2 million various garment images following DinoV2 [32]. StructureNet is further trained on STGarment with contrastive learning for 4 days on 4 H20 GPUs at a batch size of 128. At inference time, we run RAGDiffusion on a single NVIDIA RTX 3090 GPU for 30 sampling steps with the DDIM sampler [45]. Please refer to the Appendix for more details. + +Evaluation protocols. About generation quality, we utilize LPIPS [61], SSIM [50] to assess the reconstruction accuracy, DISTS [13] to evaluate image similarity on both perceptual and structural level. Additionally, we employ FID [33] and KID [46] metrics to evaluate the realism and authenticity of the generated distributions. In terms of retrieval ability, we evaluate the top-1 accuracy and top-5 accuracy of retrieval results from the memory database $\mathcal{D}$ across 1969 test samples, alongside the average Intersection over Union (IoU) of the corresponding silhouette masks with GT ones. Given the absence of GT masks of test samples in the memory database $\mathcal{D}$ , we define the retrieved masks to be accurate if their IoU exceeds 0.85. + +Baselines. We compare RAGDiffusion with recent works TryOffDiff [49] and TryOffAnyone [51] with official checkpoints. We also train 4 classic conditional generation methods on STGarment with SDXL backbone for fair comparison: IP-Adapter [55], ReferenceNet [21], ControlNet [59] and Paint-by-Example [53] (PBE) as baselines. + +![](images/37aced32194da58fd148a1da1ee60f5c99ebd29a255a064e0824f60a51217fd6.jpg) +Figure 4. Boxplot illustration of user study. RAGDiffusion demonstrates better performance and stability across scenes. + +# 4.2. Generation comparisons with baselines + +Qualitative results. Fig. 3 presents a qualitative comparison between RAGDiffusion and baselines on the STGarment dataset. TryOffDiff has yielded entirely unsuccessful outputs. Meanwhile, naive ReferenceNet and TryOffAnyone manage to maintain correct textures in challenging scenarios, but encounter structural confusion and distortion. This phenomenon may stem from TryOffDiff and ReferenceNet's over-reliance on local visual texture features, lacking the benefits of a global perspective in challenging cases. RAGDiffusion employs a retrieval-aggregate approach to capture structural information and integrates conditional controls from EP-Adapter, Landmark Guider and PGEA, yielding promising performance in both structurally faithful and detail-oriented garment conversion. Notably, while trained on the same dataset with naive ReferenceNet, RAGDiffusion assimilates high-quality contour landmarks and structure embeddings as external prior to producing visually compelling results that enhance depth and realism. + +Quantitative results. As indicated in Tab. 1, we quantitatively evaluate the generation quality of various methods on the STGarment dataset, demonstrating that RAGDiffusion significantly outperforms all baseline approaches. This confirms RAGDiffusion's ability to deliver superior and accurate garment generation across diverse scenarios. Retrieval, as a crucial component of RAGDiffusion for assimilating external knowledge, plays a significant role. Using the evaluation metrics described in Sec. 4.1, we report the retrieval accuracy at different sizes of the external memory database $\mathcal{D}$ in Tab. 4. Specifically, we employ clustering and downsampling algorithms to iteratively eliminate outliers and highly similar samples from the original memory database, constructing retrieval libraries of four different scales: 1000, 2000, 4000, and 8000 samples. When the memory database $\mathcal{D}$ is too large, outliers may adversely affect quality if embeddings are inaccurate, while redundant samples reduce retrieval efficiency. Conversely, if the memory database $\mathcal{D}$ is too small, the retrieval library may lack sufficient representativeness and completeness, adversely affecting the generation performance for specific categories. Notably, a memory database comprising 4000 samples achieved the best trade-off in performance. Please refer to the Appendix for more technical details. + +![](images/eac183df222a00d214f0aedcbfa3072104ec47b3eac6e77eb50f5b06b1660ea1.jpg) +Figure 5. Ablation study. Without landmarks, RAGDiffusion suffers inaccurate shape. Embed- Figure 6. RAG significantly improves dings from StructureNet improve inner structure, while PGEA enhances detail preservation. generalization on untrained categories. + +User study. When metrics like FID assess the realism and authenticity of standard garment generation, they tend to be less sensitive to high-frequency information such as contours, collar variations, buttons, and color shifts [11]. We conducted user studies involving 100 participants to evaluate different methods based on user preferences across 200 randomly selected image pairs from the test set. Participants were asked to assign a preference score (ranging from $1 \sim 10$ ) in terms of structural fidelity and detail fidelity for each sample generated by anonymized methods. As illustrated in Fig. 4, we report the distribution of the scores, encompassing medians, means, quartiles, and outliers. Our findings reveal that RAGDiffusion is significantly preferred over all baseline methods in terms of structural fidelity (mean = 8.34) and detail fidelity (mean = 8.52). Additionally, the narrow range of outliers in our method indicates a more stable generation across various scenes. + +# 4.3. Ablation study + +To validate our core contributions, we define a Vanilla Model that removes structure embedding $\hat{e}_{itw}$ , silhouette landmarks $\hat{L}_{sil}$ and PGEA. Specifically, its $\hat{e}_{itw}$ and $\hat{L}_{sil}$ are set to zero vectors while using original SDXL VAE. + +Soft guidance of structure embeddings. Since structure embeddings serve as prerequisites for retrieval in SLLE and inputs to EP-Adapter, we first add structure embeddings $\hat{e}_{itw}$ into EP-Adapter of vanilla model (denoted as “+ emb”). As in Tab. 2 and Fig. 5, vanilla model lacking structure embedding frequently exhibits artifacts in clothing internal structures, such as inaccurate collars, missing pockets, and occlusion confusion. This demonstrates that structure embedding provides global guidance and plays a + +fundamental role in optimizing apparel internal structures. + +Hard guidance of silhouette landmarks. The core of RAG is to retrieve silhouette landmarks that enhance explicit spatial constraints. Thus we add landmarks on "+" emb" version as the ablation named "+" emb + sil" in Tab. 2. The model without retrieval exhibits significant contour errors, particularly around the back collar, sleeve length, and garment length. This is attributed to the fact that retrievalaugmented SLLE incorporates priors from LLM, which transcend visual limitations and aid in accurate spatial structures through the use of landmarks. Furthermore, we investigate the role of SLLE in embedding remapping in Sec. 4.5. Alternative to RAG: generating masks via UNet. As RAG requires a memory database and contrastive learning for embeddings, we try a simpler alternative by employing an SDXL UNet to directly predict landmarks. Specifically, we modify the input of the alternative UNet to a three-channel in-the-wild image $x_{itw}$ , with outputs being one-channel predicted landmarks $L_{sil}$ that are directly fed into RAGDiffusion's denoising UNet. However, this alternative produces landmarks with noticeable errors and degraded generation quality, due to limitations in visual understanding and the absence of proper pre-trained weights. + +PGEA. PGEA is employed to mitigate information loss during the VAE encoding and decoding process. We add PGEA on “+ emb + sil” version as Full Model. The full model utilizing PGEA achieves clearer and more accurate edges in high-frequency patterns simulation; additionally, it shows significant improvements in the recovery of logos and text. PGEA greatly alleviates the long-standing issue of distortion in detail and makes RAGDiffusion commercially + +viable, the first in industrial garment restoration. + +# 4.4. Additional rationale for introducing RAG + +In addition to structural faithfulness, RAGDiffusion demonstrates excellent generalizability and robustness across untrained categories and scenes simply by updating retrieval database, without the need for collecting expensive paired data and retraining. Besides, please refer to Appendix for another rationale of human-interpretable control. + +Cross category evaluation. RAGDiffusion is trained on upper-body clothing data and has not encountered lower-body garments during training. To validate the generalizability, we collect 856 flat-lay lower-body/dress images (no in-the-wild image is needed) and incorporate them with silhouette masks into the external memory database through StructureNet. Subsequently, we test it on 50 lower-body in-the-wild clothing samples, using a ReferenceNet as baseline. The results in Fig. 6 demonstrate that retrieval significantly improves generalization capabilities, serving as a cost-effective zero-shot generalizing solution. The quantitative results on lower-body/dress categories in Tab. 3 also validate the generalizability boost. + +Cross dataset evaluation. Despite zero exposure to the Viton-HD [12] and DressCode [19] datasets during training, RAGDiffusion shows strong out-of-distribution (OOD) compatibility over baselines without tuning illustrated in Tab. 3. This boost in generalization increases the operational maturity of RAGDiffusion, enabling it to effectively handle various OOD images submitted by users. + +# 4.5. Component analysis + +Latent structure embedding distributions. We further visualize the learned latent embedding distribution in Fig. 7(b) to gain an in-depth comprehension of how the embeddings work in retrieval and the EP-Adapter. For this purpose, we sample $N = 200$ pairs of (in-the-wild cloth embedding $e_{itw}$ , standard cloth embedding $e_{std}$ ) pairs from our dual-tower StructureNet. We then employ t-SNE [48] to project each feature representation into a point within Fig. 7. Notably, 1) the dual-branch embeddings $(e_{itw}, e_{std})$ corresponding to the same category exhibit clustering behavior, indicating that the learned priors in StructureNet effectively perform structural alignment cross domains, thus facilitating retrieval based on structural similarity; and 2) representations from different categories are clearly dispersed, demonstrating that the embeddings encompass discriminative structural information, which aids in generation through the EP-Adapter. Lastly, we also visualize the distribution of cosine similarity between a given sample and images from the external memory database $\mathcal{D}$ in Fig. 8. The cosine similarities present a normal distribution, affirming that the constructed embedding library is representative and comprehensive. Furthermore, the examples visualized for different similarity retrieval results effectively illustrate the efficacy of the landmark retrieval mechanism. + +![](images/fef8d3817323215d5a0c8554f34cc750a0a21e17239fc879a25a5140a9b75e56.jpg) +Figure 7. Illustrations of component analysis. + +![](images/19427e421d3148182fc7e3caa2137140071d2caf2a5ef8a7b9cad15619129762.jpg) + +![](images/4ef970b0196ed10d121b94891ec7937ec0d6fe7d4bc7f26b4ca05012b8fcaa5d.jpg) +Cosine Similarity Distribution Given a Specific Embedding +Figure 8. Distribution of cosine similarity between a given sample and images from the external memory database. + +Number of retrieved embeddings. As the number $K$ of retrieved nearest neighbors in Eq. (2) during SLLE plays a fundamental role in the properties of the final structure embeddings and landmarks, we demonstrate the fusion sensitivity to $K$ in Fig. 7(a). A smaller $K$ value facilitates the sharp edge of landmarks, whereas a larger $K$ value enhances the accurate representation of the reconstructed embedding $e_{itw}^{\prime}$ in Eq. (3). It can be observed that the optimal performance occurs at $K = 4$ , which also serves as the default parameter setting for other experimental parts. + +# 5. Conclusion + +In conclusion, RAGDiffusion presents a significant advancement in the generation of standardized clothing assets by effectively addressing the prevalent challenges of structural hallucinations and texture fidelity. By integrating retrieval-based structure aggregation to get soft and hard guidance, as well as omni-level faithful generation pipeline, RAGDiffusion is capable of synthesizing both structure and texture faithful clothing assets. Comprehensive experiments and in-depth analysis demonstrate notable performance improvements over existing models. We hope that RAGDiffusion helps open avenues for RAG in diverse high-specification faithful generation tasks, bringing us closer to the goal of professional content creation for all. + +# Acknowledgements + +This work is supported by the Science and Technology Commission of Shanghai Municipality under research grant No. 25ZR1401187. + +# 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, 4 +[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. 2, 4, 5 +[3] Volker Blanz and Thomas Vetter. Face recognition based on fitting a 3d morphable model. IEEE Transactions on pattern analysis and machine intelligence, 25(9):1063-1074, 2003. 4 +[4] Andreas Blattmann, Robin Rombach, Kaan Oktay, Jonas Müller, and Björn Ommer. Retrieval-augmented diffusion models. Advances in Neural Information Processing Systems, 35:15309-15324, 2022. 2 +[5] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. Improving language models by retrieving from trillions of tokens. In International conference on machine learning, pages 2206-2240. PMLR, 2022. 2 +[6] Ali Borji. Qualitative failures of image generation models and their application in detecting deepfakes. Image and Vision Computing, 137:104771, 2023. 4 +[7] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdzal, and Adriana Romero Soriano. Instance-conditioned gan. Advances in Neural Information Processing Systems, 34:27517-27529, 2021. 2 +[8] Di Chang, Yichun Shi, Quankai Gao, Jessica Fu, Hongyi Xu, Guoxian Song, Qing Yan, Xiao Yang, and Mohammad Soleymani. Magicdance: Realistic human dance video generation with motions & facial expressions transfer. arXiv preprint arXiv:2311.12052, 2023. 5 +[9] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343-5353, 2024. 2 +[10] 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 +[11] Xi Chen, Yutong Feng, Mengting Chen, Yiyang Wang, Shilong Zhang, Yu Liu, Yujun Shen, and Hengshuang Zhao. Zero-shot image editing with reference imitation. arXiv preprint arXiv:2406.07547, 2024. 7 + +[12] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 8 +[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] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4 +[15] Black forest labs. Flux.1-dev. https://github.com/black-forest-labs/flux, 2024.5 +[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, 2014. 3 +[17] Gongde Guo, Hui Wang, David Bell, Yaxin Bi, and Kieran Greer. Knn model-based approach in classification. In On The Move to Meaningful Internet Systems 2003: CoopIS, DOA, and ODBASE: OTM Confederated International Conferences, CoopIS, DOA, and ODBASE 2003, Catania, Sicily, Italy, November 3-7, 2003. Proceedings, pages 986-996. Springer, 2003. 4 +[18] 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 +[19] 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), 43(4):1-13, 2024. 1, 8 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 2020. 1 +[21] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 2, 3, 4, 5, 6 +[22] Minghui Hu, Jianbin Zheng, Daqing Liu, Chuanxia Zheng, Chaoyue Wang, Dacheng Tao, and Tat-Jen Cham. Cocktail: Mixing multi-modality control for text-conditional image generation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2 +[23] Xuan Ju, Ailing Zeng, Chenchen Zhao, Jianan Wang, Lei Zhang, and Qiang Xu. Humansd: A native skeleton-guided diffusion model for human image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15988-15998, 2023. 2 +[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3 +[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 + +[26] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2251-22521, 2023. 2 +[27] Yuhan Li, Hao Zhou, Wenxiang Shang, Ran Lin, Xuanhong Chen, and Bingbing Ni. Anyfit: Controllable virtual try-on for any combination of attire across any scenario. arXiv preprint arXiv:2405.18172, 2024. 2 +[28] Qihao Liu, Adam Kortylewski, Yutong Bai, Song Bai, and Alan Yuille. Intriguing properties of text-guided diffusion models. arXiv preprint arXiv:2306.00974, 2, 2023. 4 +[29] Microsoft. Deepspeed. https://github.com/microsoft/DeepSpeed, 2023.6 +[30] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. LaDIVTON: Latent Diffusion Textual-Inversion Enhanced Virtual Try-On. In Proceedings of the ACM International Conference on Multimedia, 2023. 3 +[31] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 2 +[32] 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. 6 +[33] 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, 2022. 6 +[34] 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. 5 +[35] 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 +[36] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 2 +[37] 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, 2021. 3 + +[38] Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. Incontext retrieval-augmented language models. Transactions of the Association for Computational Linguistics, 11:1316-1331, 2023. 2 +[39] 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, 2022. 3 +[40] Robin Rombach, Andreas Blattmann, and Björn Ommer. Text-guided synthesis of artistic images with retrieval-augmented diffusion models. arXiv preprint arXiv:2207.13038, 2022. 2 +[41] Sam T Roweis and Lawrence K Saul. Nonlinear dimensionality reduction by locally linear embedding. science, 290 (5500):2323-2326, 2000. 2, 4 +[42] Junyoung Seo, Susung Hong, Wooseok Jang, Ines Hyeonsu Kim, Minseop Kwak, Doyup Lee, and Seungryong Kim. Retrieval-augmented score distillation for text-to-3d generation. arXiv preprint arXiv:2402.02972, 2024. 2 +[43] Le Shen, Rong Huang, and Zhijie Wang. Igr: Improving diffusion model for garment restoration from person image. arXiv preprint arXiv:2412.11513, 2024. 1, 2, 3, 5 +[44] Shelly Sheynin, Oron Ashual, Adam Polyak, Uriel Singer, Oran Gafni, Eliya Nachmani, and Yaniv Taigman. Knndiffusion: Image generation via large-scale retrieval. arXiv preprint arXiv:2204.02849, 2022. 2 +[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 6 +[46] JD Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference for Learning Representations, 2018. 6 +[47] Hung-Yu Tseng, Hsin-Ying Lee, Lu Jiang, Ming-Hsuan Yang, and Weilong Yang. Retrieved: Image synthesis via differentiable patch retrieval. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 242-257. Springer, 2020. 2 +[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 8 +[49] Riza Velioglu, Petra Bevandic, Robin Chan, and Barbara Hammer. Tryoffdiff: Virtual-try-off via high-fidelity garment reconstruction using diffusion models. arXiv preprint arXiv:2411.18350, 2024. 1, 2, 3, 5, 6 +[50] 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. 6 +[51] Ioannis Xarchakos and Theodoros Koukopoulos. Tryoffanyone: Tiled cloth generation from a dressed person. arXiv preprint arXiv:2412.08573, 2024. 2, 3, 5, 6 +[52] 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. 2 +[53] 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, 2023. 6 +[54] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 2 +[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. 2, 4, 6 +[56] Zhenhui Ye, Jinzheng He, Ziyue Jiang, Rongjie Huang, Jiawei Huang, Jinglin Liu, Yi Ren, Xiang Yin, Zejun Ma, and Zhou Zhao. Geneface++: Generalized and stable real-time audio-driven 3d talking face generation. arXiv preprint arXiv:2305.00787, 2023. 2 +[57] Wei Zeng, Mingbo Zhao, Yuan Gao, and Zhao Zhang. Tilegan: category-oriented attention-based high-quality tiled clothes generation from dressed person. Neural Computing and Applications, 32:17587-17600, 2020. 2, 3 +[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3 +[59] 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, 6 +[60] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2 +[61] 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, 2018. 6 +[62] Shiyue Zhang, Zheng Chong, Xujie Zhang, Hanhui Li, Yuhao Cheng, Yiqiang Yan, and Xiaodan Liang. Garmentaligner: Text-to-garment generation via retrieval-augmented multi-level corrections. arXiv preprint arXiv:2408.12352, 2024. 1 +[63] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583, 2023. 2 +[64] Xujie Zhang, Yu Sha, Michael C Kampffmeyer, Zhenyu Xie, Zequn Jie, Chengwen Huang, Jianqing Peng, and Xiaodan Liang. Armani: Part-level garment-text alignment for unified cross-modal fashion design. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4525-4535, 2022. 1 + +[65] Xujie Zhang, Binbin Yang, Michael C Kampffmeyer, Wenqing Zhang, Shiyue Zhang, Guansong Lu, Liang Lin, Hang Xu, and Xiaodan Liang. Diffcloth: Diffusion based garment synthesis and manipulation via structural cross-modal semantic alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23154-23163, 2023. 1 +[66] Xuanpu Zhang, Dan Song, Pengxin Zhan, Qingguo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Anan Liu. Boowvton: Boosting in-the-wild virtual try-on via mask-free pseudo data training. arXiv preprint arXiv:2408.06047, 2024.2 +[67] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2 \ No newline at end of file diff --git a/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/images.zip b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8cb72f41d498256324a0eb64f5c48f11b906d889 --- /dev/null +++ b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a199f9c80a29c6e347fd2ff51565d8cd32aa012855dcb59d944732c5d3ae026 +size 665793 diff --git a/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/layout.json b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5e1b310d80c78caef1926f6efa9114b1944eb4b7 --- /dev/null +++ b/ICCV/2025/RAGDiffusion_ Faithful Cloth Generation via External Knowledge Assimilation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6fff7628da6f70dce6671471b4bb4d8662bae129989c8c1023d18d63b83df77 +size 392285 diff --git a/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/f597b9c9-b0eb-4e05-8a43-19e44509d067_content_list.json b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/f597b9c9-b0eb-4e05-8a43-19e44509d067_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..455b0a30d4dbf96c2f1d7fdb53abdc89d15bc90c --- /dev/null +++ b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/f597b9c9-b0eb-4e05-8a43-19e44509d067_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:638faa2ae4b7bec2f28129de9ff3ebe823c7863df3f0b9e64ae116d44b4726ac +size 85652 diff --git a/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/f597b9c9-b0eb-4e05-8a43-19e44509d067_model.json b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/f597b9c9-b0eb-4e05-8a43-19e44509d067_model.json new file mode 100644 index 0000000000000000000000000000000000000000..832645383dad324e73589aa3be38b8b818864f49 --- /dev/null +++ b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/f597b9c9-b0eb-4e05-8a43-19e44509d067_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7144582845ec42604dab066fa695b5262dfc78e39bf5ca576c21bb6293008b44 +size 108027 diff --git a/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/f597b9c9-b0eb-4e05-8a43-19e44509d067_origin.pdf b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/f597b9c9-b0eb-4e05-8a43-19e44509d067_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a16ab79410ad8f75256bb4ff9ac48fbb4a0d8c99 --- /dev/null +++ b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/f597b9c9-b0eb-4e05-8a43-19e44509d067_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00a869ea4186d2b29af5567ac5f243797597bd18f31d0ceb0471049d30684732 +size 3900610 diff --git a/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/full.md b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/full.md new file mode 100644 index 0000000000000000000000000000000000000000..25dafebd68e7e67206a963d540e6f49089916472 --- /dev/null +++ b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/full.md @@ -0,0 +1,362 @@ +# RAGNet: Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping + +Dongming Wu $^{1}$ , Yanping Fu $^{2\star}$ , Saise Huang $^{3}$ , Yingfei Liu $^{3\dagger}$ , Fan Jia $^{3}$ , Nian Liu $^{4}$ , Feng Dai $^{2}$ , Tiancai Wang $^{3}$ , Rao Muhammad Anwer $^{4}$ , Fahad Shahbaz Khan $^{4}$ , Jianbing Shen $^{5\dagger}$ + +1 The Chinese University of Hong Kong, 2 Institute of Computing Technology, Chinese Academy of Sciences, +$^{3}$ Dexamal, $^{4}$ Mohamed bin Zayed University of Artificial Intelligence, $^{5}$ SKL-IOTSC, CIS, University of Macau + +![](images/6b0c5f89646363d20d1753591a41aad1eb855304a50c25efbb653174ad05e2e2.jpg) +I need grip a hammer. + +![](images/b2ec3c0dceb133904f161c7ce750c11e0c1d0e16ff3551dc29f57830a5226629.jpg) +You can grip its handle. + +![](images/5ddadc49694c3750b75ae301ed9b79d4a9d37d2bc2041f3652818b4a38e2b69c.jpg) +I need a tool to drive a nail into the wall. + +![](images/8daae7934551c1ed7245e27e87023579854bc701e26b4fdb6c21041ad3ea4076.jpg) + +![](images/d526c39563e2da3c42f6574699105c344183917ffa91df0fd8d7d0e0f00f2479.jpg) + +![](images/12dc266e19577ab89deba519df80538578e2e104e3c6359c0e8b0c8556015db7.jpg) +Large-scale Reasoning-based Affordance Segmentation Dataset + +![](images/1bc7a04b5766b384f56f3285ebe788d98438082c0936b8d6df3956d55841def1.jpg) +Figure 1. Large-scale benchmark for reasoning-based affordance segmentation, which sources from various embodied domains: wild, robot, ego-centric indoor, and simulation. By leveraging the extensive high-quality dataset for training, our model AffordanceNet exhibits remarkable open-world generalization capabilities, steering further towards robust general-purpose object grasping. + +![](images/d694aa7e3f1927995851e8534dd01f288d24c3360c812830cc762c8657b7cda5.jpg) + +![](images/bab39f53428ec5049978cf9f554684593e0ea330f49a51a1d31db8ccf737c965.jpg) + +![](images/29a3a0d06ce194265f52fad8dec22af616ea8cd27db2fd449dcd86b15e359ef1.jpg) + +![](images/e3008e3ffa05fa51f19c4eb898ec5608b8c544d9c7f227ac0e8a62a28e61179a.jpg) + +![](images/b6d893f7bc409b6863747e754ff93b9e2a8ac7af2916b5f77d6a22804edb472f.jpg) +Sure, you can utilize the hammer by gripping its handle. + +![](images/18873eb3f414d91c2d492489590c9b1245674aa0e4200a9bd3f10ee65c51c8dd.jpg) + +![](images/dc971dce7231e65b807d36fad4a23d5de5916967b6a29e112922ab38a2a0d995.jpg) +General Grasping + +![](images/d5cb209fc3e24c8d056a8ec1007f2072a5e409e061b1701bffc9e1d17f87592c.jpg) + +![](images/93e681ca398f4792d68c4b7f93e1f069e59e6aff3a50a1964de1a8a610e0c4f9.jpg) + +# Abstract + +General robotic grasping systems require accurate object affordance perception in diverse open-world scenarios following human instructions. However, current studies suffer from the problem of lacking reasoning-based large-scale affordance prediction data, leading to considerable concern about open-world effectiveness. To address this limitation, we build a large-scale grasping-oriented affordance segmentation benchmark with human-like instructions, named RAGNet. It contains 273k images, 180 categories, and 26k reasoning instructions. The images cover diverse embodied data domains, such as wild, robot, ego-centric, and even + +simulation data. They are carefully annotated with an affordance map, while the difficulty of language instructions is largely increased by removing their category name and only providing functional descriptions. Furthermore, we propose a comprehensive affordance-based grasping framework, named AffordanceNet, which consists of a VLM pretrained on our massive affordance data and a grasping network that conditions an affordance map to grasp the target. Extensive experiments on affordance segmentation benchmarks and real-robot manipulation tasks show that our model has a powerful open-world generalization ability. Our data and code is available at this link. + +# 1. Introduction + +Affordance prediction is a foundational research topic that significantly contributes to diverse practical applications, including robotic manipulation [15, 17, 18, 38, 52] and human-object interaction [3, 20, 22, 51, 57, 58]. It requires + +
DatasetImagesCategoriesWildRobotEgo-centricSimulationReasoningOutput
UMD [37] ICRA201510k17✓✓Seg
AGD20k [35] CVPR202020k50✓Seg
HANDAL [14] IROS2023200k17✓Seg/Box
AED [28]Arxiv2024-21✓✓Seg
3DOI [42]ICCV202310k-✓✓Seg
AffordanceLLM [44]CVPR202420k-✓Seg
ManipVQA [16] IROS202484k-✓✓✓Seg
RAGNet (Ours)273k180✓✓✓✓✓Seg
+ +Table 1. Comparisons between previous affordance data and our collection. “(11k)” represents the number of video clips. “-” means unavailable data. “Reasoning” refers to reasoning instructions. + +comprehensively understanding the geometry and function of an object (e.g., a wok affords to hold, a microwave door affords to pull) for further detecting graspable regions. Nonetheless, this task faces two additional primary difficulties in the realm of open-world generalization concerning vision and language instructions: 1) The capability to generalize in a broad range of unseen object categories and image domains. 2) The alignment of commands that mimic human-like high-level instructions. + +Initial studies often specialize in particular domains, as depicted in Table 1. For instance, UMD [37] offers 10,000 RGB-D image pairs from three cluttered scenes with pixel-level affordance annotations. This dataset is categorized as robot data, characterized by a distinct operational table and a relatively stable background. Despite the abundance of public robot datasets [4, 41, 56], they often lack meticulous fine-grained labeling. Ego-centric data, which is captured from a first-person or egocentric viewpoint, is the most prevalent type of data [7, 13]. Nevertheless, this data is often collected in indoor kitchens and workshops, resulting in a lack of diversity in the data category distribution. Wild data, another prevalent type, is gathered in various situations, which usually cover diverse categories [35, 46, 53]. However, certain categories within this data, like bicycles, motorcycles, sofas and etc., are not suitable for robotic manipulation. In addition, these data often fail to generalize to unseen domains and novel objects for affordance prediction. + +To enhance open-world generalization, Vision Language Model (VLM) with massive image-text training [29, 32] has become an important increment in various visual prediction tasks [25, 47]. Inspired by this, VLM also has an increasing interest in affordance prediction along with complex reasoning, and several recent works are attempting to predict affordance regions [16, 27, 30, 44, 60, 62]. For example, ManipVQA [16] proposes tool detection, affordance recognition, and a broader understanding of physical concepts in a unified framework. However, these works employ a fixed format of language prompt and a limited data scale, which brings some concerns about open-world generalization and complex reasoning. Overall, current efforts are either constrained by insufficiently diverse datasets or by the absence of effective reasoning mechanisms, which hampers the ability to perform open-world grasping. + +To address these challenges, we first build large-scale affordance segmentation data from various image sources, including wild, robot, ego-centric, and even simulated data, named RAGNet. It has 273k images and 180 categories. We design a set of affordance annotation tools for labeling the regions of objects that are amenable to grasping. Furthermore, we leverage Large Language Models (LLMs) to generate a vast array of reasoning-based instructions, totaling 26k distinct expressions. Here, we create two types of reasoning-based instructions beyond the template-based. One includes the name of the object and the other omits it. Take a knife as an example: "Please provide a knife" versus "I want something to slice the bread". This approach closely mirrors real-life human interactions. In summary, we have constructed a massive-scale database with domains, object categories, and complex reasoning instructions. + +Furthermore, we introduce an affordance-based grasping framework, titled AffordanceNet. It presents a deployable and general grasping pipeline, which stands out from prior MLLM-based affordance prediction methods [16, 44] that have not yet demonstrated real-robot deployment. Our model includes two crucial components, AffordanceVLM and Pose Generator. The AffordanceVLM transforms RGB images and human instruction into an accurate affordance map, while the pose generator uses the 2D affordance with a depth image to produce 3D grasper pose. In this work, we conduct extensive experiments to assess its open-world generalization and reasoning capabilities. First, we create two distinct validation datasets to assess open-world generalization: one for zero-shot category recognition and another for out-of-domain affordance prediction. Second, to test reasoning ability, we implement an affordance segmentation validation based on instructions that do not contain any target category names. Thirdly, we carry out a range of close-loop real-robot grasping tasks in an entirely out-of-domain setting. Last but not least, we test several representative simulation tasks from RLBench. All experiments show that our proposed method has great generalization and reasoning ability. In summary, our contributions are three-fold: + +- We present a large-scale reasoning-based affordance segmentation benchmark, RAGNet, for general grasping. It is collected from diverse sources and carefully annotated with affordance mask and reasoning instructions. + +![](images/884500542a2c2b24384aa45bfcd87ee058ea4d500d78c2a8c385b2addc35b98e.jpg) +Figure 2. Overview of our data annotation pipeline and manipulation model. We collect data from several public datasets, including HANDAL, GraspNet, Open-X, etc. A variety of models and manual annotation are utilized to annotate affordance masks. Subsequently, we refine textual instructions using templates and GPT4 to emulate human-like commands. For grasp operations, the VLM model is employed to identify affordance regions, which are then converted into the required grasp poses by integrating depth information. + +- We introduce an affordance-based grasping baseline, AffordanceNet, which bridges the gap between VLM-based affordance prediction and real-robot general grasping. +- We conduct extensive experiments, including zero-shot and out-of-domain affordance segmentation, real-robot grasping evaluation, which show great performance. + +# 2. Related Work + +The investigation of affordances has a rich history in the computer vision and robotics communities [1, 11, 12, 24]. For embodied agents to successfully interact with the functional components within a scene, they need the capability to comprehend visual affordances. To advance research in this area, previous studies have dedicated significant resources to two main areas: benchmarks and algorithms. + +Benchmark. With the advances in deep learning, there has been a growing demand for the creation of large-scale affordance databases. UMD [37] stands out as a prior effort, which defines objects with effective affordances as those that an agent can grasp to produce an effect on another object. In addition, it offers a dataset of 10k RGB-D images, each accompanied by a pixel-level affordance segmentation mask. OPRA [10] proposes a different affordance learning pattern, which employs demonstration videos to guide the prediction of interaction region on a target image. It includes 20,612 video clips, each corresponding to an image and an annotation of interaction heatmap and action label. Despite the pioneering efforts in affordance learning, + +the existing datasets [6, 10, 34, 37, 39, 50] still face constraints regarding affordance category diversity, image quality, and scene complexity. Afterward, Luo et al. establish a large-scale affordance grounding dataset, AGD20k [35], which contains 20k exocentric images from 36 affordance categories. Guo et al. build the HANDANL dataset [14] for real-world robot manipulation, which provides precise handle annotation for 17 hardware and kitchen tool categories. Recently, there are a variety of datasets on affordance prediction that have been introduced with different output formats, such as key points [21, 30, 42, 48], bounding box [54, 55], pixel-wise mask [27, 28, 31]. However, most current research tends to focus on specific domains, such as robotics or first-person perspectives, which limits their applicability to other areas. In contrast, our research aims to investigate the potential for open-world generalization in affordance prediction by leveraging extensive data. + +Algorithm. In the deep learning era, the common methods use supervised learning for affordance prediction [6, 8, 14, 34, 39]. However, these approaches struggle to generalize to domains outside of their training environments. To tackle this issue, certain studies utilize transfer learning [26, 35, 36, 61] or self-supervised learning techniques [5] to enable affordance prediction from data originating in different domains. With the significant progress of VLM, attempting these foundation models in affordance prediction has attracted much interest in the research community [16, 27, 30, 44, 60, 62]. However, these methods + +
Data sourceDomainAnnotationRea. Inst.Categories
HANDAL [14]Wild➀8.5k17
Open-X [41]Robot➂➁➉-124
GraspNet [9]Robot➂➂-32
EgoObjects [65]Ego➂➂➉17.4k74
RLBench [19]Simulation➂-10
+ +Table 2. Details of training data annotation in RAGNet. The affordance mask annotation are in §3.2 and the generation details of reasoning instructions (Rea. Inst.) can be found in §3.3. + +usually learn from limited affordance demonstration data, which opens up opportunities for further research into how large-scale data can impact generalization in open-world scenarios and enhance knowledge reasoning. + +# 3. Dataset + +In this section, we will present how to build the large-scale reasoning-based affordance segmentation benchmark RAG-Net. First, we gather extensive data from various sources, as detailed in §3.1. Then, we introduce five tools to annotate this data with grasping-oriented affordance masks according to the original dataset characteristics in §3.2. Meanwhile, we offer detailed instructions for reasoning-based affordance segmentation in §3.3. Lastly, we establish multiple validation benchmarks for testing grasping-oriented affordance segmentation in §3.4. The data annotation pipeline is illustrated in Fig. 2. + +# 3.1. Data Source + +Given that robotics can function across various embodiment realms, including real-world settings and robotic hardware, we collect a broad spectrum of data. The dataset comprises sources such as wild data (i.e., HANDAL [14]), real robots (i.e., Open-X [41], GraspNet [9]), and ego-centric data (i.e., EgoObjects [65]). For a fair comparison environment, we consider a collection of simulation data (i.e., RL-bench [19]). We gather 273k images as shown in Table 2. + +# 3.2. Affordance Map Annotation + +To facilitate affordance segmentation with reduced manual input, we develop a suite of annotation tools. The annotation workflow is designed to be adaptable, taking into account the functional and physical characteristics of various object categories. For instance, when dealing with an object like a soda can, the robotic system typically grasps the entire object, necessitating full object annotation. On the other hand, an object such as a wok requires precise annotation of its handle for effective grasping. This tool suite comprises five annotation tools, each with its own level of priority. + +1 Original mask: Certain datasets specify affordance segmentation with precision (such as HANDAL dataset [14]), while others, like those featuring objects without handles (e.g., computer mouse, soda can, and pen), do not require + +![](images/bb7ecdf203ccbd760897c06d2013599f781e8e9a063d557a07256ffcdb8358f3.jpg) +User: +Figure 3. Easy (left) v.s. Hard (right) Reasoning Instruction. The hard version has no category name itself. + +![](images/e9efab979998ab32c45affeb82fe395002439a7dc99329cff69d88a2ef587aab.jpg) +Can you find a mug for tea? +User: +I need something to drink coffee. + +# Assistant: + +The mug has a handle which you can hold to pour the coffee. [AFF] + +# Assistant: + +The mug is suitable for this, its handle can be held to drink coffee safely. [AFF] + +detailed affordance masks for grasping. In these cases, we utilize the original masks as our affordance annotations without further refinement. + +$\Theta$ SAM2: For objects that lack handles, SAM2 [49] can still be utilized to generate a mask, when only the ground-truth bounding box is available (like EgoObjects [65]). +$\bullet$ Florence2 + SAM2: Due to the presence of language instructions (like Open-X [41]), Florence2 [59] which generates polygon boxes coupled with SAM2 can be used to produce an affordance map for objects that are handle-free. +$\mathbf{4}$ VLPart + SAM2: VLPart [53] enables part-level recognition (i.e., knife handle and mug handle), hence we leverage it and SAM2 for segmenting object handles if the corresponding category has been trained within VLPart. +$\Theta$ Human $(+SAM2)$ : If the above four tools fail to complete the affordance segmentation accurately, we will consider manual affordance annotation. The use of SAM2 is optional, particularly when dealing with video sequences. + +According to the original annotation information offered by data sources, we employ different compositions for affordance mask annotation. The composition details are listed in Table 2. We incorporate more detailed annotations, encompassing variations in tool arrangement across subsets and categories, within the supplementary. Besides, we provide several representative examples in Fig. 1 and more annotation examples can also be found in the supplementary. + +# 3.3. Reasoning Instruction Annotation + +As previously mentioned, the current VLM possesses compelling reasoning capabilities. To harness these capabilities for affordance reasoning, we construct a set of instructions. In this section, we introduce three types of instructions, including one template-based and two reasoning-based. + +The first kind of instruction is template-based. For instance, a template is "Please segment the affordance map of in this image". This template can + +
Validation SetImagesZero-shotReason.Anno.
HANDAL [14]65k-N/A1
\( \text{HANDAL}^{\dagger} \) [14]1k-N/A1
GraspNet seen [9]1k-N/A15
GraspNet novel [9]1k✓N/A15
3DOI [42]1k✓N/A5
\( \text{HANDAL}^{\dagger} \) [14]1k-easy1
\( \text{HANDAL}^{\dagger} \) [14]1k-hard1
3DOI [42]1k✓easy5
+ +Table 3. Details of validation set. $\dagger$ means a HANDAL subset. Gray means zero-shot setting. As highly replicated images, we randomly select 1k images from the source dataset for validation. + +be utilized in our entire dataset for affordance prediction. The second category consists of easy reasoning-based instructions that are based on straightforward reasoning. A key aspect of these instructions is the explicit mention of the object being referred to. In contrast to template-based approaches and previous studies [16, 25, 44], the third category comprises hard reasoning-based instructions, which do not include the category name. For precise identification of the target object, the hard mode instructions utilize a functional description. Fig. 3 presents a typical example. Towards grasping a mug, the easy instruction might be "Can you find a mug for tea", while the hard instruction is "I need something to drink coffee". + +To produce these reasoning-based instructions with minimal human labor and computational resources, we make full use of the capabilities of GPT-4 [2]. The specific prompt utilized is detailed in the supplementary material. According to our data, we craft 8.5k hard instructions for the HAN-DAL dataset, 12.7k easy ones and 4.7k hard ones for the EgoObjects, totaling 26k reasoning-based instructions. + +# 3.4. Evaluation Dataset + +To assess the open-world generalization of our affordance prediction model, we contribute two distinct zero-shot affordance evaluation scenarios. The first scenario evaluates the affordance model by extending its predictions to object categories not encountered during training. The second scenario evaluates the model generalization ability across different data domains. For this purpose, we select the 3DOI dataset [42], which includes data from Articulation [43], EpicKitchen [7], and Taskonomy [63], ensuring that there is no overlap with the training data in the validation set. + +We present four different validation sets in Table 3. Due to high-similarity images, we construct a subset from the HANDAL, which is marked with ‘†’. For the same reason, other validation sets also utilize 1k images from the source dataset. HANDAL and GraspNet seen represent the object categories and image domains that the model has been trained on. GraspNet novel represents the unseen object category, while 3DOI refers to the unseen image domain. + +In addition, we provide three reasoning-based affordance segmentation validation sets, as illustrated in Table 3. The reasoning instruction version used in each subset is marked as "easy" or "hard". We utilize the generalized Intersection over Union (gIoU) and complete Intersection over Union (cIoU) as our primary metrics. + +# 4. AffordanceNet + +To achieve the goal of open-world grasping, we propose a comprehensive framework, named AffordanceNet. This model consists of two key components: AffordanceVLM for predicting affordance segmentation mask and pose generation for transforming the mask into grasper position in 3D space. The overall framework is illustrated in Fig. 2. + +# 4.1. AffordanceVLM + +Our AffordanceVLM is based on the vision-language segmentation model LISA [25] and incorporates two essential task-specific modifications to enhance affordance prediction: (1) developing a specialized system prompt, and (2) introducing a unique token. + +Specifically, we first process the input image using an image encoder (i.e., ViT-CLIP [45]), which is then projected into the LLM's embedding space via a linear layer projector. Meanwhile, the language prompt is tokenized by a text tokenizer, where each affordance instruction is enhanced by "You are an embodied robot." The resulting image and text features are concatenated and fed into the LLM (i.e., Vicuna-7B [64]). For general segmentation, the LLM generates a response that includes a special token $<\mathrm{SEG}>$ . However, since $<\mathrm{SEG}>$ is a token within the LLM's vocabulary, its representation is confined to a fixed feature space. This limitation restricts its representation capacity, thereby affecting the quality of the decoded mask. To address this issue, we introduce another special token $<\mathrm{AFF}>$ to enrich the original mask embedding. The underlying motivation is to explicitly direct the final mask embedding to focus more on affordance-specific language expressions. Finally, SAM [23] is used as a mask decoder to convert this mask embedding into a pixel-wise mask. + +Implementation Details. Beyond our reasoning-based affordance segmentation data, we also incorporate a variety of generic segmentation datasets into our training, which use $<\mathrm{SEG}>$ token. During inference, we first extract the $<\mathrm{AFF}>$ token, followed by the $<\mathrm{SEG}>$ token. More implementation details about data sampling and training settings can be found in the supplementary. + +# 4.2. Pose Generator + +Once obtained AffordanceVLM that delivers precise affordance segmentation prediction, we start to apply the model to robotic grasping tasks, aiming to bridge the gap for the + +
MethodHANDALHANDAL†GraspNet seenGraspNet novel3DOI
gIoUcIoUgIoUcIoUgIoUcIoUgIoUcIoUgIoUcIoU
Foundation Model without LLMs
VLPart [53] + SAM2 [49]40.928.940.727.6------
Grounding DINO [33] + SAM2 [49]34.726.834.926.9------
Florence 2 [59] + SAM2 [49]39.722.439.422.5------
Generalist MLLMs
LISA [25]16.212.015.411.817.717.725.224.121.513.7
GLaMM [47]24.917.225.117.021.610.519.28.619.714.1
AffordanceNet (Ours)60.360.860.560.363.364.045.633.237.437.4
+ +Table 4. Quantitative results on affordance segmentation. We use the fixed format of " handle" in foundation models without LLMs, while utilize "affordance map of " in generalist MLLMs. Gray means zero-shot benchmark. + +
MethodHANDAL†Grasp. novel3DOI
gIoUcIoUgIoUcIoUgIoUcIoU
LISA16.212.025.224.121.513.7
+ HANDAL56.354.916.614.418.012.2
+ Open-X59.356.719.218.424.516.0
+ EgoObjects61.861.68.06.935.534.6
+ GraspNet61.761.751.538.540.941.8
+ Reasoning56.555.443.033.836.840.2
+ RLBench56.755.042.833.236.539.9
Ours60.560.345.633.237.437.4
+ +final steps in object grasping and manipulation. The impressive affordance segmentation results can be found in §5.1 and §5.2. In the following part, we will discuss more details about our grasp pose generator and the process of converting 2D affordance masks into the 3D pose estimates required for robotic arms. + +As illustrated in Fig. 2, we project the depth maps into 3D space for precise grasp pose generation. Let $P$ denotes the set of points on the 2D image. Firstly, we filter the affordance region by applying the affordance mask $M$ to $P$ through a binary multiplication $\otimes$ , resulting in $\hat{P} = P \otimes M$ . Subsequently, for each 2D position $(u, v)$ within the point set $\hat{P}$ , we have + +$$ +\left[ \begin{array}{c} x \\ y \\ z \\ 1 \end{array} \right] = T \cdot K ^ {- 1} \left[ \begin{array}{c} u \times d \\ v \times d \\ d \\ 1 \end{array} \right], \tag {1} +$$ + +where $\mathrm{d}$ is the along the axis orthogonal to the image plane, $(x,y,z)$ are the corresponding world coordinates. $\mathbf{K} \in \mathbb{R}^{4 \times 4}$ and $\mathbf{T} \in \mathbb{R}^{4 \times 4}$ represent the camera intrinsic and extrinsic parameters, respectively. After getting the 3D position of object affordance, we can use various grasping + +Table 5. Ablation study of data on affordance segmentation. Each data is added one by one. Compared to “+ RLBench”, our final model is enhanced by task-specific modifications, including specialized system prompt and unique token (see §4.1). + +
MethodHANDAL (easy)HANDAL (hard)3DOI
gIoUcIoUgIoUcIoUgIoUcIoU
G-DINO3.63.03.43.14.13.9
LISA15.511.912.38.112.38.1
GLaMM4.73.55.03.54.42.9
Ours58.358.158.257.838.139.4
+ +Table 6. Quantitative results on reasoning-based affordance segmentation. 'G-DINO' is the short name of Grounding-DINO. They all use reasoning-based instructions as language prompts. + +models to generate grasper position. Finally, the grasper arrive the requested position and grasp the target object. + +# 5. Experiments on Visual Affordance + +To minimize unnecessary source expenditure, we initially validate the quality of affordance segmentation (§4.1) prior to object grasping (§4.2). + +# 5.1. Evaluation on Affordance Segmentation + +Implementation Details. To evaluate the task of affordance segmentation, we implement various advanced open-sourced approaches with potential affordance segmentation ability. In specific, we choose the foundation models without LLMs (e.g., VLPart [53], Grounding DINO [33], Florence2 [59]), and the generalist MLLMs (e.g., LISA [25], GLaMM [47]). Since these foundation models only output bounding boxes or polygon boxes, we additionally employ SAM2 [49] for mask refinement. We load their official checkpoints for direct evaluation. Given the ambiguous nature of the "affordance map of " concept, which poses a challenge to the foundation models without LLMs, we shift to a fixed format of instruction " handle" and evaluate them on only HANDAL dataset as each object have a handle. In contrast, we employ "affordance map of " when testing the generalist MLLMs. + +Experiment Results. The main results on affordance segmentation are shown in Table 4. We can see our model out + +![](images/587e4763623ef2b02a89d3e6001564fa59f936d44eccd5cb9d28e7cd81f77bc4.jpg) +Figure 4. Affordance segmentation from our AffordanceNet. Even though they source from various data sources, such as wild, robot, ego-centric and simulation, our model can accurately capture their affordance region. More visualizations are included in supplementary. + +![](images/87b6b9b9f17045c9d30b18b36e4685d89703749096cc8a6667015df5a756d826.jpg) + +![](images/45ef5d95b007b3053c24065e2abf8033af57b86dedb5a7691c6be2003f37900c.jpg) + +![](images/fe57a8960839b371f74513015d46484c0d0139d071dfa6dbffbe8d681878b6e0.jpg) + +![](images/ec89effd97225cb208e8712ed93021b6c357504b8f118575047469ebc6edce35.jpg) + +![](images/3e49221583db16398d08b2673a3baaf4d3a7e4334b2ec4df0b83de01258b8231.jpg) +Can you find a knife for me? + +![](images/1289c018f328f094b9a03f6664b98b555dea50c46415f32bafdf9525b959fded.jpg) +I need a fork for the dinner table. + +![](images/e0cb55a5fba4115f86e6e51757d806ba1da07e0df506e25f6ff2645944b00573.jpg) +Could you grab a mug for me, please? + +![](images/bfcce477dcb37f91383007df133fb714dd79967c06731e458e5518a6b27378cb.jpg) +I need a tool to drive nails into wood. + +![](images/36f1a50b66aad7cd4df30db9dbc5d6d8f5fd5b29a865082015804fa9db773816.jpg) +Fetch something to drink hot tea. +Figure 5. Reasoning-based affordance segmentation from our AffordanceNet. The left examples represent easy reasoning-based instructions with referent name, while the right are hard instructions that include object function or intention rather than the name itself. + +performs all other competitors across all datasets. Besides, the scores on the entire HANDAL dataset and its small subset are similar, guaranteeing the diversity and representative of the small subset. Several visualizations across various domains are shown in Fig. 4. Here, the wok handle is accurately segmented even if its scene has never been encountered, indicating remarkable open-world generalization. + +Ablation Study. It is of interest to examine the impact of individual datasets on affordance segmentation tasks. In Table 5, we incrementally incorporate each dataset into the training process. We can see that the absence of HANDAL data significantly impairs the model performance on the HANDAL test set. Furthermore, the incorporation of reasoning data leads to a slight decline in performance metrics. However, the introduction of task-specific modifications, such as a specialized system prompt and a unique token, enhances model performance. Overall, our model shows powerful open-world affordance segmentation. + +# 5.2. Evaluation on Reasoning Affordance + +We also test the foundation models and generalist MLLMs on reasoning-based affordance segmentation datasets. All the experiment settings are aligned with the above section. + +Experiment Results. Despite this challenging task, the results in Table 6 show our model outperforms other methods by a large margin. We provide several reasoning-based qualitative results in Fig. 5. The last two examples show that our model can predict precise hammer and mug handles even if there is no target mentioned in the instructions. These results confirm the reasoning ability of our model. + +Remark. The above two vision experiments demonstrate that only our model excels in both affordance perception and reasoning in zero-shot domain. This provides a solid foundation for the subsequent object grasping tasks. + +# 6. Experiments on Object Grasping + +# 6.1. Evaluation on Real Robot + +Implementation Details. To evaluate the effectiveness of our model for open-world generalization in real-world environments, we introduce a series of manipulation experiments. Specifically, we deploy UR5 robot arm with a third-person RGB-D camera (Intel RealSense). We design 10 distinct grasping tasks, including grasping the can, pen, screwdriver, hammer, wok, mouse, circle, toy, spatula, scissors. Half of them require accurately localizing the affordance region, like the screwdriver handle. Each task is performed 10 times, and we report the average success rate. Note that, we never provide any demonstration images or videos from this scene for our model training. The well-trained AffordanceVLM is directly used for the zero-shot evaluation. As for the pose generation, we follow GraspNet [9] condition 3D affordance point cloud to generate a 3D grasp proposal for grasper operation. We compare our model with a popular grasping model, GraspNet [9]. Since GraspNet lacks the capability for language-conditioned grasping, we ensure that only the target object remains on the table. + +Experiment Results. Table 7 displays the performance comparison between ours and GraspNet. Clearly, our model AffordanceNet has superior success rates, even in challenging and complicated environments. Fig. 6 highlights four sequences, which show accurate affordance perception and effective object grasping (i.e., screwdriver, wok, circle, and mouse). More grasping examples are in our supplementary. + +Ablation Study. Table 8 shows ablations of AffordanceNet on affordance prediction models (replaced by VLPart and LISA) and instruction types (replaced by easy and hard reasoning-based instructions) using five tasks. As demonstrated, AffordanceNet delivers superior affordance predic + +![](images/135478edf54dd6e491949a30d7e98b00e850b3ca56b4f9d38723436a52459221.jpg) + +![](images/a740784824b521d83aa29d9968177ef00046fbe387436f199c22b335eaea916b.jpg) + +![](images/9fc76fe99c6b4d9d3ca5bef391eaee72f5266a07e7e191d575e4c82d2fc11701.jpg) + +![](images/9d53a6fa96e8989bd86a302d1be972c586b03fdf001216685243e6794434c4c8.jpg) +Figure 6. Object grasping results from our AffordanceNet on robot arm UR5. The instructions are "I need a screwdriver for repairing", "Can you hand me the wok, please?", "Give me the circle", "Please hand me a computer mouse", respectively. + +
MethodCanPenScrewdriverHammerWokMouseCircleToySpatulaScissorsAverage
GraspNet [9]40%10%10%20%60%50%10%60%30%30%32%
AffordanceNet (Ours)80%60%60%80%70%80%40%90%70%70%70%
+ +Table 7. Average success rates on robotic grasping. GraspNet cannot support language (see §6.1). Each task is conducted by 10 times. + +
MethodCanPenScrew.Ham.WokAverage
VLPart [53]70%30%40%30%034%
LISA [25]80%40%010%026%
Easy Reasoning70%50%50%80%60%62%
Hard Reasoning60%40%40%60%40%48%
AffordanceNet80%60%60%80%70%70%
+ +tion while maintaining its reasoning capabilities. + +# 6.2. Evaluation on Simulation + +Implementation Details. We conduct simulations based on RLBench sub-task [19] to validate our approach, including open drawer, close jar, and slide block to target. In practice, we divide each task into several keyframes, as follows: open drawer (3 keyframes), close jar (4 keyframes), slide block to target (5 keyframes). More details are in the supplementary. + +Experiment Results. We conduct 25 episodes for each task and calculate its average success rate as the primary evaluation metric. The quantitative results are in Table 9. Compared to another LLM-based method LLARVA [40] that is fine-tuned for a specific environment, our model focuses on + +Table 8. Ablation studies on real-robot grasping in terms of different affordance prediction models (i.e., VLPart and LISA) and different instructions (i.e., easy reasoning and hard reasoning). + +
TaskLLARVA [40]AffordanceNet (Ours)
Open drawer60%56%
Slide block to target100%64%
Close jar28%44%
Average62%54.7%
+ +Table 9. Success rates of our model on RLBench simulation. Each task is conducted by 25 episodes. + +stronger generalization. Therefore, achieving comparable performance on RLBench is quite satisfying. + +# 7. Conclusion + +In this work, we presented a new large-scale and diverse reasoning-based affordance segmentation dataset, named RAGNet, which aims to advance the capabilities of general robotic grasping systems in varied open-world scenarios. Further, we proposed the model AffordanceNet, a comprehensive framework including the AffordanceVLM and the grasping module, which uses our extensive affordance data to achieve open-world affordance capture and 3D grasper pose prediction, respectively. Through extensive experiments, we observed the superior performance and generalization ability of our AffordanceNet in zero-shot affordance segmentation, reasoning-based affordance segmentation, real-robot grasping evaluation, and simulation tasks. + +# References + +[1] Function-based generic recognition for multiple object categories. CVGIP: Image Understanding, 1994. 3 +[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. 5 +[3] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In CVPR, 2023. 1 +[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2 +[5] Joya Chen, Difei Gao, Kevin Qinghong Lin, and Mike Zheng Shou. Affordance grounding from demonstration video to target image. In CVPR, 2023. 3 +[6] Ching-Yao Chuang, Jiaman Li, Antonio Torralba, and Sanja Fidler. Learning to act properly: Predicting and explaining affordances from images. In CVPR, 2018. 3 +[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. IJCV, 2022. 2, 5 +[8] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In ICRA, 2018. 3 +[9] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In CVPR, 2020. 4, 5, 7, 8 +[10] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In CVPR, 2018. 3 +[11] Vittorio Ferrari and Andrew Zisserman. Learning visual attributes. NeurIPS, 2007. 3 +[12] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 3 +[13] 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 CVPR, 2022. 2 +[14] Andrew Guo, Bowen Wen, Jianhe Yuan, Jonathan Tremblay, Stephen Tyree, Jeffrey Smith, and Stan Birchfield. HAN-DAL: A dataset of real-world manipulable object categories with pose annotations, affordances, and reconstructions. In IROS, 2023. 2, 3, 4, 5 +[15] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv preprint arXiv:2403.08248, 2024. 1 +[16] Siyuan Huang, Iaroslav Ponomarenko, Zhengkai Jiang, Xiaqi Li, Xiaobin Hu, Peng Gao, Hongsheng Li, and Hao + +Dong. Manipvqa: Injecting robotic affordance and physically grounded information into multi-modal large language models. In IROS, 2024. 2, 3, 5 +[17] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1 +[18] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 1 +[19] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 2020. 4, 8 +[20] Juntao Jian, Xiuping Liu, Manyi Li, Ruizhen Hu, and Jian Liu. Affordpose: A large-scale dataset of hand-object interactions with affordance-driven hand pose. In ICCV, 2023. 1 +[21] 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 ECCV, 2024. 3 +[22] Bumsoo Kim, Junhyun Lee, Jaewoo Kang, Eun-Sol Kim, and Hyunwoo J Kim. Hotr: End-to-end human-object interaction detection with transformers. In CVPR, 2021. 1 +[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. In ICCV, 2023. 5 +[24] Hedvig Kjellström, Javier Romero, and Danica Kragic. Visual object-action recognition: Inferring object affordances from human demonstration. Computer Vision and Image Understanding, 2011. 3 +[25] 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. 2, 5, 6, 8 +[26] Gen Li, Varun Jampani, Deqing Sun, and Laura Sevilla-Lara. Locate: Localize and transfer object parts for weakly supervised affordance grounding. In CVPR, 2023. 3 +[27] Gen Li, Deqing Sun, Laura Sevilla-Lara, and Varun Jampani. One-shot open affordance learning with foundation models. In CVPR, 2024. 2, 3 +[28] Gen Li, Nikolaos Tsagkas, Jifei Song, Ruaridh MonWilliams, Sethu Vijayakumar, Kun Shao, and Laura SevillaLara. Learning precise affordances from egocentric videos for robotic manipulation. arXiv preprint arXiv:2408.10123, 2024.2,3 +[29] 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. 2 +[30] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation. In CVPR, 2024. 2, 3 + +[31] Yicong Li, Na Zhao, Junbin Xiao, Chun Feng, Xiang Wang, and Tat-seng Chua. Laso: Language-guided affordance segmentation on 3d object. In CVPR, 2024. 3 +[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 2 +[33] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6 +[34] Timo Luddecke and Florentin Worgotter. Learning to segment affordances. In ICCV Workshops, 2017. 3 +[35] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In CVPR, 2022. 2, 3 +[36] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Leverage interactive affinity for affordance learning. In CVPR, 2023. 3 +[37] Austin Myers, Ching L Teo, Cornelia Fermüller, and Yiannis Aloimonos. Affordance detection of tool parts from geometric features. In ICRA, 2015. 2, 3 +[38] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv preprint arXiv:2402.07872, 2024. 1 +[39] Anh Nguyen, Dimitrios Kanoulas, Darwin G Caldwell, and Nikos G Tsagarakis. Object-based affordances detection with convolutional neural networks and dense conditional random fields. In IROS, 2017. 3 +[40] Dantong Niu, Yuvan Sharma, Giscard Biamby, Jerome Quenum, Yutong Bai, Baifeng Shi, Trevor Darrell, and Roei Herzig. Llarva: Vision-action instruction tuning enhances robot learning. arXiv preprint arXiv:2406.11815, 2024. 8 +[41] Abby O'Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2, 4 +[42] Shengyi Qian and David F Fouhey. Understanding 3d object interaction from a single image. In ICCV, 2023. 2, 3, 5 +[43] Shengyi Qian, Linyi Jin, Chris Rockwell, Siyi Chen, and David F Fouhey. Understanding 3d object articulation in internet videos. In CVPR, 2022. 5 +[44] Shengyi Qian, Weifeng Chen, Min Bai, Xiong Zhou, Zhuowen Tu, and Li Erran Li. Affordancelm: Grounding affordance from vision language models. In CVPR, 2024. 2, 3, 5 +[45] 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. 5 +[46] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In CVPR, 2023. 2 + +[47] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In CVPR, 2024. 2, 6 +[48] 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 CoRL, 2023. 3 +[49] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4, 6 +[50] Johann Sawatzky, Abhilash Srikantha, and Juergen Gall. Weakly supervised affordance detection. In CVPR, 2017. 3 +[51] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In CVPR, 2020. 1 +[52] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, 2022. 1 +[53] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. In ICCV, 2023. 2, 4, 6, 8 +[54] An Dinh Vuong, Minh Nhat Vu, Hieu Le, Baoru Huang, Binh Huynh, Thieu Vo, Andreas Kugi, and Anh Nguyen. Grasp-anything: Large-scale grasp dataset from foundation models. arXiv preprint arXiv:2309.09818, 2023. 3 +[55] An Dinh Vuong, Minh Nhat Vu, Baoru Huang, Nghia Nguyen, Hieu Le, Thieu Vo, and Anh Nguyen. Language-driven grasp detection. In CVPR, 2024. 3 +[56] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In CoRL, 2023. 2 +[57] Xiaohan Wang, Yuehu Liu, Xinhang Song, Yuyi Liu, Sixian Zhang, and Shuqiang Jiang. An interactive navigation method with effect-oriented affordance. In CVPR, 2024. 1 +[58] 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 CVPR, 2024. 1 +[59] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. In CVPR, 2024. 4, 6 +[60] Ran Xu, Yan Shen, Xiaoqi Li, Ruihai Wu, and Hao Dong. Naturalvlm: Leveraging fine-grained natural language for affordance-guided visual manipulation. arXiv preprint arXiv:2403.08355, 2024. 2, 3 +[61] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. In ICCV, 2023. 3 + +[62] Qiaojun Yu, Siyuan Huang, Xibin Yuan, Zhengkai Jiang, Ce Hao, Xin Li, Haonan Chang, Junbo Wang, Liu Liu, Hongsheng Li, et al. Uniaff: A unified representation of affordances for tool usage and articulation with vision-language models. arXiv preprint arXiv:2409.20551, 2024. 2, 3 +[63] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018. 5 +[64] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. NeurIPS, 2023. 5 +[65] Chenchen Zhu, Fanyi Xiao, Andres Alvarado, Yasmine Babaei, Jiabo Hu, Hichem El-Mohri, Sean Culatana, Roshan Sumbaly, and Zhicheng Yan. Egoobjects: A large-scale egocentric dataset for fine-grained object understanding. In ICCV, 2023. 4 \ No newline at end of file diff --git a/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/images.zip b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c3fc243923fc99e2d87058ea60e228acff1c9c1f --- /dev/null +++ b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d9575d58ea21b46755fe808d5316960588601a2644638374c9d581385b68444 +size 777836 diff --git a/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/layout.json b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8578d6648789c95a50559a82844cebbd65ba0618 --- /dev/null +++ b/ICCV/2025/RAGNet_ Large-scale Reasoning-based Affordance Segmentation Benchmark towards General Grasping/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bcf7b09121e4866a2fa759e3b0169bafcfd67130d7b7701679f0766d7db483b +size 414209 diff --git a/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/a502ea9f-2ea1-4759-90cc-38d35aa41279_content_list.json b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/a502ea9f-2ea1-4759-90cc-38d35aa41279_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5f08b0d5ffcc04de599384edcaad9ccb89f2c88a --- /dev/null +++ b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/a502ea9f-2ea1-4759-90cc-38d35aa41279_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f198df22cda5a90d4ca2159ae741b06ff0c19ea7c3f052580905787fd8bf0548 +size 79692 diff --git a/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/a502ea9f-2ea1-4759-90cc-38d35aa41279_model.json b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/a502ea9f-2ea1-4759-90cc-38d35aa41279_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a4bfa5ed910f03bd942dbc468dccb1312db4722b --- /dev/null +++ b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/a502ea9f-2ea1-4759-90cc-38d35aa41279_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29c5fa645b2c45f2e94ce63826d2167cf006acb7fec30d75e83c8bfdbab76de9 +size 100682 diff --git a/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/a502ea9f-2ea1-4759-90cc-38d35aa41279_origin.pdf b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/a502ea9f-2ea1-4759-90cc-38d35aa41279_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8106b7d00e64fc35ae921c73d5cd7b1f9c0b16a1 --- /dev/null +++ b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/a502ea9f-2ea1-4759-90cc-38d35aa41279_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c33815eb6a4026ca9e154f92bfab24b957aa8ea0797e9494299db5b45c3caf1 +size 3518688 diff --git a/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/full.md b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0f483ab703c1e1896f3efc10e3e487ce46f0c7ff --- /dev/null +++ b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/full.md @@ -0,0 +1,328 @@ +# RALoc: Enhancing Outdoor LiDAR Localization via Rotation Awareness + +Yuyang Yang $^{1,2*}$ Wen Li $^{1,2*}$ Sheng Ao $^{1,2}$ Qingshan Xu $^{3}$ Shangshu Yu $^{4}$ Yu Guo $^{1,2}$ Yin Zhou $^{5}$ Siqi Shen $^{1,2}$ Cheng Wang $^{1,2\dagger}$ + +$^{1}$ Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University $^{2}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University $^{3}$ Nanyang Technological University $^{4}$ Northeastern University $^{5}$ GAC R&D Center + +# Abstract + +LiDAR localization is a fundamental task in autonomous driving and robotics. Scene Coordinate Regression (SCR) exhibits leading pose accuracy, achieving impressive results in learning-based localization. We observe that the real-world LiDAR scans captured from different viewpoints usually result in the catastrophic collapse of SCR. However, existing LiDAR localization methods have largely overlooked the issue of rotation sensitivity in SCR. In this paper, we present RALoc, an outdoor LiDAR localization method with rotation awareness to achieve accurate localization. The key to our approach is to design a Point Cloud Canonicalization module, which leverages a powerful equivariant key feature aggregation to transform the input LiDAR scan towards a consistent orientation, effectively eliminating the adverse effects of rotation. This proposed module has promising scalability and can be seamlessly integrated with the existing LiDAR localization network. Moreover, we propose the Bidirectional LiDAR Localization (BiLiLo) dataset as a benchmark to evaluate the performance of various methods in large outdoor scenes with significant rotation changes. Extensive experiments show that RALoc significantly improves localization performance in scenarios with large rotation changes, and also achieves competitive performance in the Oxford Radar RobotCar dataset. Our project is available at https://etheryangyy.github.io/raloc.github.io. + +# 1. Introduction + +LiDAR localization is a crucial task that plays an important role in various applications, such as autonomous driving [34], robotics and augmented reality. The objective of LiDAR localization is to use the point cloud of the scene + +![](images/3319f04b384d7dd32d101337daa7db9c939cd858ceb0cd9fba2b7640b0a550fe.jpg) +Figure 1. Teaser. This paper primarily investigates the problem of rotation-robust LiDAR localization. We present ground truth and predicted trajectories on the collected rotation changes dataset. Left: the forward trajectory (aligned with the training set driving direction). Right: the reverse trajectory (completely opposite to the training set driving direction). Compared to the current state-of-the-art method, LiSA [55], RALoc retains high accuracy in the forward trajectory and avoids catastrophic failures due to rotation in the reverse trajectory. + +![](images/2f6fd4641ce5f8fafff463d230425daed2ddd2a92985553af85de49d45a43dcd.jpg) + +captured by the LiDAR as input and output its 6 degrees-of-freedom (DoF) pose. However, due to uneven data quality (e.g., noise distribution, non-uniform density, and varying viewing angles), it remains challenging to estimate accurate 6-DoF pose in large-scale real-world applications. + +Most existing methods [12, 27, 45, 54] typically follow the retrieval-registration paradigm for self-localization, which involves retrieving candidate scenes from a database, verifying their correctness, and subsequently estimating the 6-DoF relative pose via point cloud registration [35, 40]. However, when dealing with a large amount of LiDAR point cloud data, this approach not only imposes significant demands on storage and communication resources, but also struggles to achieve efficient localization from scratch [57]. + +Learning-based regression has been shown to be a promising research direction to address the aforementioned issues by memorizing specific scenes within a neural network [57]. Existing methods can be categorized into Abso + +![](images/6f507c7ac25e02923bc740f531af42e08e6456d4cd709587ff3403925dac4ac8.jpg) +Figure 2. Motivation. (a) The rotation problems caused by changing lanes, turning, and reversing. (b) The impact of Yaw angle variation on LiSA's [55] recall rate within 5 meters and 3 degrees. + +lute Pose Regression (APR) and Scene Coordinate Regression (SCR). APR [25, 31, 50] directly estimates the global pose of the sensor in an end-to-end manner. Despite its effectiveness, sub-optimal results are achieved due to the highly abstract global feature used to represent the scene. Unlike the APR method, the SCR [7-9, 30] technique regresses the correspondences between the point cloud in the LiDAR coordinate and the world coordinate frame, followed by RANSAC [19] algorithm to estimate a precise 6-DoF pose. By explicitly capturing scene geometry, SCR achieves significant performance improvements. + +Nonetheless, as shown in the right side of Fig. 1, we observe that the learning-based regression method suffers from catastrophic collapse due to large viewpoint discrepancies when the scanned point cloud is rotated relative to the training data. In practical applications, factors such as oncoming traffic, lane changes, and turns often result in unpredictable changes in vehicle direction. To illustrate the effect of rotation changes, we take a recently proposed superior method, LiSA [55], to uncover this problem. As shown in Fig. 2, we conducted a pilot study with various rotation changes and recorded the localization recall. Specifically, we randomly rotate the query point cloud by the Yaw axis to simulate the challenge of a real driving scenario. Clearly, recall decreases significantly as the rotation increases. When rotations exceed 30 degrees, the recall drops below $1\%$ , indicating that localization algorithms fail in such cases. These experiments demonstrate that rotation changes negatively impact the performance of existing localization models. + +We believe that the main reason is that the neural network used lacks robustness to rotations [47]. Although data augmentation can improve the rotational robustness of networks to some extent, it compromises the overall performance of the network and necessitates larger capacities along with more complex architectures [1, 6, 52]. In contrast, equivariant models [38, 39] provide a strong inductive bias for the network, which is expected to fundamen + +tally address the issue of regression methods being sensitive to rotations. Furthermore, we found that test scenarios in commonly used benchmarks [4, 11] are relatively simple and rarely account for situations where the vehicle direction changes. To close this gap, we contribute a large-scale outdoor localization dataset with substantial rotation variations, which can offer a more accurate basis for assessing model robustness under real-world conditions. + +In this paper, we propose a novel localization framework, RALoc, to address the issue of current LiDAR localization models being vulnerable to rotations. Explicitly, we introduce an equivariant network to canonicalize the input point clouds. Additionally, to address the presence of noise interference in outdoor LiDAR localization tasks, we design an equivariant key feature aggregation module that effectively extracts scene features to better capture rotation changes. Moreover, to evaluate the rotational robustness of SCR in real-world driving scenarios, we establish equipment setups and collect a large-scale outdoor LiDAR localization dataset characterized by significant rotational variations. + +Our contributions can be summarized as follows: + +- We reveal that learning-based regression methods are sensitive to rotation changes, which greatly weakens their localization performance in real-world applications. To this end, we propose RALoc, the first attempt to introduce an equivariant network to SCR. +- We introduce a novel equivariant key feature aggregation module designed to mitigate interference from noisy point clouds, enhancing our approach with robust rotation-equivariant awareness. +- We collect a large-scale outdoor LiDAR localization dataset with significant rotation variations, providing an effective benchmark to facilitate related research. +- Extensive experiments on the proposed rotation challenge and Oxford Radar RobotCar [4] datasets demonstrate the effectiveness of our method. In particular, we show that RALoc achieves robust performance even with significant rotational variations (see Fig. 1). + +# 2. Related work + +# 2.1. Conventional Localization + +Conventional localization methods aim to match the query point cloud with a pre-built 3D map, which can be classified into retrieval-based methods [29, 48, 53] and matching-based methods [23, 42, 46]. Retrieval-based methods cast localization as a problem of place recognition, which requires a pre-built database with feature descriptors and relies on querying the most similar point cloud in the database [37, 59]. Matching-based methods are typically based on either searching for similar frames in a pre-built global descriptor database or 3D point clouds and match features between the query frame and the map [2, 33]. Both + +![](images/4ce92e7bb0150ab5f6e50492d14257314908e4ef7c7ff7f5d190ec9c3055cacb.jpg) +Figure 3. The pipeline of RALoc. RALoc consists of two main components: (1) the Point Cloud Canonicalization (PCC) network and (2) the Scene Coordinate Regression (SCR) network. Specifically, the PCC network captured from various rotations in the scene to a canonical view. The canonicalized point clouds are then fed into the SCR network, which outputs the 6-DoF pose. + +methods depend on pre-stored maps, resulting in challenges related to storage and communication. + +# 2.2. Regression-based Localization + +Recent advances in deep learning have spurred interest in regression-based localization methods that directly predict 6-DoF poses end-to-end. PoseNet [25] is a widely used APR method that regresses camera poses via a feature encoder and regressor. Given LiDAR's robustness to lighting and weather, APR has also been extended to LiDAR [58]. PointLoc [50] introduced LiDAR-based APR, and HypLiLoc [49] leveraged multi-level features to capture richer scene characteristics. DiffLoc [31] improved accuracy by transforming the pose regression process into multiple iterative steps through the diffusion model [22]. + +However, as mentioned in [30, 44], APR methods focus on high-dimensional pose-feature mappings, like retrieval-based techniques, but overlook scene geometry, limiting their accuracy. To address this issue, SCR predicts the coordinates of each point in the world coordinate, effectively encoding scene geometry. Ultimately, RANSAC is used to estimate the pose. SGLoc [30] is the first to incorporate SCR into LiDAR localization, followed by LiSA [55], which mitigates the impact of disruptive points in the scene by introducing semantic information, further enhancing accuracy. Although APR and SCR have achieved commendable accuracy in LiDAR localization, these methods struggle with rotation-related issues in realistic scenarios. + +# 2.3. SO(3)-Equivariant Network + +With growing performance issues due to rotation, equivariant networks have seen notable advances in recent years [14, 20, 32, 43, 51, 56]. Although data augmentation can mitigate rotation issues during training, this often affects the overall performance of the model when dealing with large datasets [6, 38, 52]. Therefore, maintaining rotational equivariance is essential when processing point cloud data. Various studies [13, 36, 61] have effectively addressed this problem through different approaches. Among them, the Vector Neurons [16] (VNN) is a lightweight equivariant framework that integrates well with existing point cloud processing networks, and many subsequent works [3, 24, 60] have expanded upon it. Consequently, we have constructed an equivariant network for processing outdoor LiDAR point clouds using VNN as the foundation. Moreover, we design an equivariant key feature aggregation module to further boost rotation robustness. + +# 3. Method + +Recent SCR methods [30, 55] have achieved impressive results in outdoor LiDAR localization. However, in scenarios with vehicular trajectory variations, as shown in Fig. 2 (a), they fail due to rotational deviations between scanned point clouds and the training data. In this paper, we propose RALoc, which utilizes the equivariant network to solve the problem of rotation sensitivity by point cloud canonicaliza + +tion (Sec. 3.1). In addition, to evaluate the rotation robustness of localization methods in the real driving scenario, we propose a large-scale outdoor dataset with significant rotation variations (Sec. 3.2). + +# 3.1.RALoc + +We now introduce RALoc, as shown in Fig. 3, which can be divided into (1) Point Cloud Canonicalization (PCC): aligning LiDAR point cloud data with different orientations, efficiently sensing rotations through an equivariant network, and reorienting the LiDAR point clouds accordingly. (2) Scene Coordinate Regression (SCR): estimate the global pose by matching the scanned point cloud to the regressed point cloud in the world coordinate [30]. + +In this work, we design a VNN-based [16] equivariant network to implement PCC. This network extracts LiDAR point cloud features using VN-MLP, aggregates scene features through Equivariant Key Feature Aggregation (EKFA), and then regresses the rotation matrix with respect to scene rotation changes. + +Point cloud canonicalization. Here we formulate the framework of PCC. For point clouds captured from diverse orientations, we aim to align them into a consistent reference frame, termed as the canonicalization coordinate. + +To establish this coordinate, we leverage the trajectory data from the training set. For a fixed position in the scene, the canonicalization coordinate is defined as the orientation of the LiDAR at that position in the training set. + +Specifically, given a point cloud scanned from arbitrary vehicle direction $Rx$ , where $R \in SO(3)$ , $x \in \mathbb{R}^{N \times 3}$ represent the point cloud in the LiDAR coordinate, PCC can be represented as follows: + +$$ +\operatorname {P C C} (R x) = \phi (R x) ^ {- 1} R x, \tag {1} +$$ + +where $\phi$ denotes a VNN-based equivariant network. This network ensures that the PCC produces consistent outputs by leveraging the properties of equivariant networks: + +$$ +\phi (R x) = R \phi (x). \tag {2} +$$ + +As shown in Fig. 3, we utilize the equivariant network to perceive the rotational state of the vehicle at the feature level, and regress the matrix. Then, we use the Schmidt orthogonalization, which is also equivariant (as proved in the supplementary), to ensure its a rotation matrix. Finally, we apply the inverse transformation of this rotation to the input point cloud, resulting in a canonical point cloud. This canonical point cloud can then be fed into the subsequent LiDAR localization network, enhancing the robustness of the localization network against rotation. More details on the architecture can be found in the supplementary. + +Equivariant key feature aggregation. As described above, the task of equivariant network $\phi$ is to ensure that + +![](images/7f29838c700477bd139fedbae64d486a249bd8f862907776015cf23e846f3f0e.jpg) +Figure 4. Equivariant Key Feature Aggregation (EKFA) module. + +point clouds scanned from different orientations of the same scene exhibit the same rotational relationship in their extracted equivariant features. However, due to the challenges of outdoor scenes, such as dynamic objects, not all points benefit the extraction of robust features [55]. Therefore, to effectively select the useful features, we introduce an equivariant key feature aggregation module. + +As shown in Fig. 4, the EKFA module takes the equivariant features $F_{in} \in \mathbb{R}^{C \times 3 \times N}$ as input. We first use VN-MLP to generate features, denoted as $Q$ , $K$ , and $V$ , each with the same dimension of $F_{in}$ . Then, to ensure strict equivariance, we compute the Frobenius inner product [3] of $Q$ and $K$ , divide it by $\sqrt{3C}$ , and apply softmax to obtain the attention matrix $A$ . This process can be expressed as follows: + +$$ +A = \operatorname {S o f t m a x} \left(\frac {\langle Q , K \rangle_ {F}}{\sqrt {3 C}}\right), \tag {3} +$$ + +where $\langle \cdot \rangle_F$ is the Frobenius inner product, which is rotation invariant. The square root of the vector neurons features dimension count $\sqrt{3C}$ serves as a scaling factor, and the shape of $A$ is $N\times N$ . + +To obtain the aggregation weights $W$ , we sum the attention matrix $A$ along its last dimension (i.e., across columns), sort $W \in \mathbb{R}^N$ in descending order, and select the top $k$ indices to construct a key feature mask $A'$ from $A$ : + +$$ +\operatorname {i d x} = \operatorname {t o p k} (W), \tag {4} +$$ + +$$ +A ^ {\prime} = A [ \mathrm {i d x},: ]. \tag {5} +$$ + +Finally, we perform a matrix multiplication with $V$ to obtain aggregated features enriched with key information after filtering out scene noise. The output can be expressed by matrix multiplication: + +$$ +F _ {o u t} = V \times A ^ {\prime} ^ {T}. \tag {6} +$$ + +Scene coordinate regression. After obtaining the canonicalized point cloud, the next step is to input it into SCR, + +![](images/1aa25d53aa1df37783c06598c12616df9692f6cbaaaa64910fb4ccaa261759b4.jpg) +Figure 5. We scan point clouds and record ground truth using a vehicle equipped with an Ouster OS1-64 spanning LiDAR and a M39 GPS/IMU device. Coordinate frames illustrate the direction of each sensor on the vehicle with the convention: X-axis (red), Y-axis (green), and Z-axis (blue). + +SGLoc [30] in this work, to get the global pose. Specifically, SCR regresses the corresponding point cloud in the world coordinate, and matches them inside a RANSAC loop for robust pose estimation. + +Loss function. The output of RALoc consists of canonicalized point clouds and poses. Therefore, the total loss function $\mathcal{L}$ should contain both canonicalization $\mathcal{L}_{pcc}$ and regression $\mathcal{L}_{reg}$ components. Following the definition of previous work [24, 38]. We choose the original LiDAR coordinate system in the training data as the canonical orientation, and utilize $\mathcal{L}_{pcc}$ to keep point cloud canonicalization output point cloud orientation as constant as possible. Specifically, we use Mean Squared Error loss to optimize the point cloud canonicalization: + +$$ +\mathcal {L} _ {p c c} = \frac {1}{n} \sum_ {i = 1} ^ {n} (R _ {c} - I) ^ {2}, R _ {c} \in S O _ {3}, \tag {7} +$$ + +where the rotation matrix $R_{c}$ is obtained by applying Schmidt orthogonalization to the equivariant features and $I$ denotes the identity matrix. + +For the pose estimation, we employ $L_{1}$ loss to guide SCR prediction correspondences to achieve localization. + +The overall loss function of the entire network is: + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {r e g} + \lambda_ {2} \mathcal {L} _ {p c c}, \tag {8} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are hyper-parameters to balance the losses. + +# 3.2. Dataset + +At first glance, SCR yields satisfactory results. However, accuracy degrades markedly with changes in vehicle orientation (see Fig. 2 (a)). In real-world settings, avoiding such + +
DatasetLiDARGTReverseUnobstructed viewLong Interval
KITTI [21]1×64GPS/IMUX✓X
Oxford [4]2×32GPS/IMUXXX
NCLT [11]1×32SLAMX✓✓
Boreas [10]1×128GPS/IMUX✓✓
MulRan [26]1×64SLAM✓XX
BiLiLo (Ours)1×64GPS/IMU✓✓✓
+ +Table 1. Related Datasets. GT: Ground truth poses. Reverse: Presence of trajectories in the opposite direction. Unobstructed View: Minimal obstructions in the LiDAR's field of view. Long Interval: Extended time intervals between data collections. + +directional shifts during lane changes and turns proves difficult (Fig. 2 (b)). Due to the lack of localization datasets with significant directional variation, previous work [18] simulates these situations by rotating point clouds. While these simulations can somewhat validate the model's robustness to rotation, there are still discrepancies with the real scanned data, such as missing data caused by changes in scanning direction, etc. This motivates us to create a dataset with significant rotation changes in real-world scenarios to better evaluate existing methods. + +Sensors setup. Our dataset mainly consists of two sensors, as shown in Fig. 5, including (1) Ouster OS1-64, a 64-beam spinning LiDAR to collect point cloud data, and (2) M39, a GPS/IMU device, to obtain the ego-motion pose. We set the FOV, resolution, and FPS as $[-22.5^{\circ}, 22.5^{\circ}]$ , $64*1024$ and 10, respectively. The FPS of M39 is set to 200, to provide dense and smoothing trajectories. We design a rigid support system for securing sensors to ensure reliably. + +Sensors calibration. Following KITTI [21]. We use ICP [5] to accurately register point clouds of a parking sequence, which provides a wide variety of orientations and translations essential for effectively conditioning the minimization problem. Next, we randomly sample 1,000 pairs of poses from this sequence and obtain the desired result using [17] to achieve LiDAR-to-GPS/IMU calibration. + +Target environments. Our dataset is collected in a campus environment featuring multiple distinguishable structures and a broad field of view. To assess the robustness of localization methods against translation and rotation, we deliberately incorporated multiple reverse revisits and revisits with lane-level translations. + +Furthermore, to enhance data diversity, the collection period spanned both high-traffic intervals between classes and low-traffic periods during class sessions on campus. Note that we include two reverse trajectories, which pose challenges for localization due to missing data resulting from changes in scanning direction. + +The dataset comprises 8 sequences for training and testing, each approximately 10 kilometers long, including 6 forward trajectories and 2 reverse trajectories. More details can be found in the supplementary. + +Comparison with other datasets. Tab. 1 summarizes the + +
MethodsSO(3) Rotation AugmentationWithout Augmentation
DiffLocSGLocLiSADiffLocSGLocLiSARALoc
15-13-06-374.87m, 1.62°3.79m, 5.58°3.78m, 5.83°4.09m, 1.58°45.82m, 27.77°52.81m, 21.47°3.17m, 4.08°
17-13-26-397.35m, 2.01°4.44m, 5.00°4.48m, 4.92°4.17m, 1.30°42.59m, 25.14°56.02m, 22.89°3.87m, 3.96°
17-14-03-006.52m, 2.14°3.66m, 4.98°3.93m, 5.22°4.61m, 1.26°41.05m, 25.32°50.65m, 22.54°3.32m, 3.83°
18-14-14-423.83m, 1.43°2.76m, 4.38°2.90m, 4.25°3.22m, 1.23°38.51m, 24.75°51.25m, 21.94°2.58m, 3.70°
Average5.64m, 1.80°3.66m, 4.99°3.77m, 5.06°4.02m, 1.34°41.99m, 25.75°52.68m, 22.21°3.24m, 3.89°
+ +Table 2. Quantitative results on the Oxford dataset with small rotation. Mean position error (m) and mean orientation error $(^{\circ})$ are reported, with best results in bold and second best results underlined. Here, small rotation refers to a random rotation of $[-10^{\circ}, 10^{\circ}]$ degrees applied to Roll, Pitch, and Yaw axes. SO(3) rotation augmentation involves random rotations of training data within $[-180^{\circ}, 180^{\circ}]$ . + +
MethodsSO(3) Rotation AugmentationWithout Augmentation
DiffLocSGLocLiSADiffLocSGLocLiSARALoc
15-13-06-374.90m, 1.61°4.18m, 5.66°3.66m, 5.43°122.70m, 43.57°341.79m, 94.04°332.43m, 79.40°3.19m, 4.07°
17-13-26-398.03m, 2.09°4.55m, 5.05°4.64m, 5.04°153.37m, 43.03°341.70m, 89.57°324.07m, 76.47°3.87m, 3.96°
17-14-03-007.05m, 2.26°3.70m, 5.11°4.00m, 5.37°138.43m, 43.16°334.34m, 88.43°323.45m, 76.94°3.33m, 3.88°
18-14-14-424.11m, 1.52°2.82m, 4.45°2.94m, 4.38°129.65m, 41.08°343.08m, 91.18°336.76m, 78.59°2.58m, 3.73°
Average6.02m, 1.87°3.81m, 5.07°3.81m, 5.06°136.04m, 42.71°340.23m, 90.81°329.18m, 77.85°3.24m, 3.91°
+ +Table 3. Quantitative results on the Oxford dataset with large rotation. Mean position error (m) and mean orientation error $(^{\circ})$ are reported, with best results in bold and second best results underlined. Here, large rotation refers to a random rotation of $[-180^{\circ}, 180^{\circ}]$ degrees applied to the Yaw axis. SO(3) rotation augmentation involves random rotations of training data within $[-180^{\circ}, 180^{\circ}]$ . + +comparison of our dataset with the existing widely used datasets. Our objective is to verify the method's robustness to rotation by collecting datasets with rotation changes. Compared to Oxford Radar Robotcar [4] and NCLT [11], which are widely used for localization evaluation, we provide multiple reverse trajectories. MulRan [26] offers reverse direction data but suffers from Radar occlusion that restricts visibility and challenges localization consistency. In contrast, our dataset emphasizes the following features: (1) sensors are vertically mounted to reduce inter-sensor occlusions, ensuring near-complete 360-degree scene coverage. (2) data collection spans three months, offering diverse conditions for a robust evaluation of localization methods. See the supplementary for detailed comparisons. + +# 4. Experiment + +# 4.1. Setup + +Datasets and metrics. We evaluate RALoc for LiDAR localization on two large-scale outdoor datasets: Oxford Radar RobotCar [4] and BiLiLo datasets. During the evaluation, we select the mean position and orientation error as the evaluation metrics. + +Oxford Radar RobotCar. The Oxford Radar RobotCar (Oxford) dataset is a critical benchmark for urban scene localization. It covers a $2\mathrm{km}^2$ area, with each trajectory spanning nearly $10\mathrm{km}$ . The dataset collects data from multiple sensors, such as LiDAR, cameras, radar, and GPS/INS. Our method only utilizes the information from LiDAR. The Oxford dataset captures a wide range of traffic conditions, + +making it an ideal resource for evaluating LiDAR-based localization models. Following previous works [30, 31, 55], we use data from 11-14-02-26, 14-12-05-52, 14-14-48-55, and 18-15-20-12 for training, and data from 15-13-06-37, 17-13-26-39, 17-14-03-00, and 18-14-14-42 for testing. + +BiLiLo. Our collected BiLiLo dataset is an outdoor campus scene localization benchmark, where each trajectory is nearly $10\mathrm{km}$ in length. It contains 8 trajectories obtained through LiDAR, with 2 trajectories specifically designed to have driving directions entirely opposite to the others. This makes the BiLiLo dataset a complementary resource to other localization datasets, enabling a more comprehensive evaluation of model performance across various aspects. We use the data from 11-28-11-20, 11-28-14-54, 11-28-15-26, and 12-06-13-57 for training, and data from 12-06-10-30, 12-06-10-59, 03-04-16-04, and 03-04-16-29 for testing, where the last two trajectories in each set are reverse data. Note the naming convention for the trajectories is based on the data collection start time, formatted as month-date-hour-minute. + +Implementation. RALoc is implemented by Pytorch [41] and Spconv [15]. We conduct our experiments on a server equipped with an Intel Xeon(R) Gold 6342 CPU, and 4 NVIDIA RTX 4090 GPUs. For the point cloud canonicalization process, we randomly down sample 4,096 point clouds from the scene as input, and the subsequent LiDAR localization network is based on SGLoc. During training, we use the Adam [28] optimizer with a learning rate of 0.002 and a batch size of 120. On Oxford Radar RobotCar dataset, we set EkFA with $k = 2048$ , and set the loss + +
MethodsSO(3) Rotation AugmentationWithout Augmentation
DiffLocSGLocLiSADiffLocSGLocLiSARALoc
15-13-06-374.88m, 1.62°4.42m, 6.17°3.77m, 5.81°3.57m, 0.88°3.01m, 1.91°2.36m, 1.29°3.19m, 4.10°
17-13-26-397.89m, 2.01°4.62m, 5.60°4.58m, 5.47°3.65m, 0.68°4.07m, 2.07°3.47m, 1.43°3.87m, 3.96°
17-14-03-007.16m, 2.26°3.84m, 5.63°4.08m, 5.67°4.03m, 0.70°3.37m, 1.89°3.19m, 1.34°3.32m, 3.87°
18-14-14-424.22m, 1.40°2.89m, 4.83°3.05m, 4.67°2.86m, 0.60°2.12m, 1.66°1.95m, 1.23°2.59m, 3.71°
Average6.04m, 1.82°3.94m, 5.56°3.87m, 5.41°3.53m, 0.72°3.14m, 1.88°2.74m, 1.32°3.24m, 3.91°
+ +Table 4. Quantitative results on the Oxford dataset without rotation. Mean position error (m) and mean orientation error $(^{\circ})$ for various methods are reported, with best results in bold and second best results underlined. SO(3) rotation augmentation involves random rotations of training data along Roll, Pitch, and Yaw axes within $[-180^{\circ}, 180^{\circ}]$ . + +
MethodsSO(3) Rotation AugmentationWithout Augmentation
DiffLocSGLocLiSADiffLocSGLocLiSARALoc
12-06-10-306.93m, 10.26°3.35m, 2.93°2.87m, 1.96°2.78m, 0.77°2.14m, 1.81°1.92m, 1.02°1.91m, 2.90°
12-06-10-596.69m, 9.66°3.29m, 2.88°2.73m, 1.80°2.62m, 0.73°2.13m, 1.69°1.85m, 0.96°1.89m, 2.75°
03-04-16-0427.99m, 23.75°10.12m, 10.47°24.93m, 62.85°383.85m, 72.70°379.96m, 90.73°322.93m, 94.56°5.24m, 9.59°
03-04-16-2945.61m, 29.46°11.08m, 13.60°27.66m, 66.42°395.29m, 78.82°378.96m, 92.43°322.32m, 95.81°6.97m, 11.02°
Average21.81m, 18.28°6.96m, 7.47°14.55m, 33.26°196.14m, 38.26°190.80m, 46.67°162.26m, 48.09°4.00m, 6.57°
+ +Table 5. Quantitative results on the BiLiLo dataset. Mean position error (m) and mean orientation error $(^{\circ})$ for various methods are reported, with best results in bold and second best results underlined. SO(3) rotation augmentation involves random rotations of training data along Yaw, Pitch, and Roll axes within $[-180^{\circ}, 180^{\circ}]$ . The red-marked trajectories are aligned with the overall driving direction during training, while the blue ones are in the reverse direction. + +weights to $\lambda_1 = 1$ , $\lambda_2 = 1200$ . On BiLiLo dataset, we set $k = 512$ , $\lambda_1 = 1$ and $\lambda_2 = 1600$ . + +Baselines. To more effectively evaluate our model's performance, we select several representative and competitive baselines. For the APR method, we include the current state-of-the-art DiffLoc [31] for comprehensive comparison. For the SCR method, we choose SGLoc [30] and LiSA [55], where LiSA incorporates semantics into SGLoc through distillation while preserving a consistent framework. Note LiSA is the current state-of-the-art regression-based LiDAR localization method. + +# 4.2. Results + +Results on Oxford with Rotation. In this experiment, we introduce random rotations within 10 degrees in Roll, Pitch, and Yaw axes to the point cloud. Tab. 2 presents quantitative results on the Oxford dataset with small rotations (within $[-10^{\circ}, 10^{\circ}]$ ), comparing baseline performance without augmentation to results with random rotation augmentation applied during training. RALoc achieves an average error of $3.24\mathrm{m} / 3.89^{\circ}$ . It is clear that the results of the previous method experienced a significant decline without the introduction of rotation augmentation. Even after introducing data augmentation, our method still outperforms our baseline, SGLoc, with $3.24\mathrm{m} / 3.89^{\circ}$ vs $3.66\mathrm{m} / 4.99^{\circ}$ . + +Additionally, to account for the potential need for frequent direction changes during vehicle operation, we apply random rotations around the Yaw axis (within $[-180^{\circ}, 180^{\circ}]$ ) to the point clouds and adjust the ground-truth values accordingly. As shown in Tab. 3, RALoc achieves + +![](images/a6a9c7f4af6dce0720b64af42516fe29612d302be666172dc76a6aea78344e67.jpg) +(a) SGLoc (10.12m, 10.47°) + +![](images/e8cda00149f50fa264e59fed472ad4429dfe6727c0f5f433e76d3d593a962bc8.jpg) +(b) DiffLoc $(27.99\mathrm{m},23.75^{\circ})$ + +![](images/3210250f161680364e9e4b18d549bb9e4bd9115f6666373f88c12556a81a89a7.jpg) +(c) LiSA (24.93m,62.85°) + +![](images/4da7166aa6355faa191f2754723f91762c0a91df9a5b49c4e48b05249d4d1d4c.jpg) +(d) RALoc $(5.24\mathrm{m},9.59^{\circ})$ +Figure 6. Visualization of different methods on BiLiLo. We select a reverse trajectory for testing. The black and red points represent the ground-truth and estimation poses, respectively, and the star indicates the first frame. The caption of each subfigure shows the mean position error (m) and orientation error $(^{\circ})$ . + +$3.24 \mathrm{~m} / 3.91^{\circ}$ , reaching state-of-the-art performance in position, with no observable accuracy loss compared to previous evaluations. This result effectively demonstrates our model's resilience to significant rotational changes. + +Results on Oxford. We also evaluate RALoc on the Oxford dataset without rotation interference. Tab. 4 reports + +the results of RALoc and all other baselines with and without random rotation augmentation during training. RALoc achieves $3.24\mathrm{m} / 3.91^{\circ}$ average error. Although the proposed method shows a decrease in orientation accuracy compared to our baseline, SGLoc, it becomes more robust across various rotation scenarios. SGLoc has significantly lower accuracy than us after adding rotation augmentation. Meanwhile, our method is not significantly affected in terms of position accuracy and even outperforms our baseline on certain trajectories. + +Results on BiLiLo. As mentioned earlier, simulated experiments struggle to comprehensively evaluate model performance, so we collected a real dataset with significant rotational variations for validation. Tab. 5 presents the evaluation results on the BiLiLo dataset. In particular, trajectories 03-04-16-04 and 03-04-16-29 have driving directions entirely opposite to the training data, enabling a rigorous evaluation of each model's rotational robustness. Our method achieves $4.00\mathrm{m} / 6.57^{\circ}$ average error. Compared to previous methods, RALoc achieves a substantial improvement in position accuracy on the forward trajectories, reaching state-of-the-art performance. On the reverse trajectories, our model produces superior results, significantly outperforming other baselines. + +Fig. 6 illustrates the qualitative results of compartments with rotation augmentation on trajectory 03-04-16-04 with mean position error (m) and orientation error $(^{\circ})$ . It is clear that RALoc offers a clean trajectory, indicating its capacity to produce more robust results. + +Speed and parameter. Real-time performance is a critical metric for localization tasks. Given that commonly used LiDAR sensors operate at a frequency of $20\mathrm{Hz}$ , our inference time must remain below 50ms. We quantitatively evaluate the inference speed and parameter count of our proposed model, RALoc. The average inference time on the two datasets is 26ms. Additionally, in terms of total parameter count, we employ a relatively lightweight approach to achieve equivariance in the localization model. This requires only an additional 0.4M parameters compared to SGLoc, representing an increase of only $0.0043\%$ . + +# 4.3. Ablation study + +Study on Rotation Augmentation. As shown in Tab. 6 between Row 1 and Row 2, using data augmentation can, to some extent, mitigate catastrophic performance losses caused by rotation. However, as shown in Tab. 5, for the normal trajectories, it also brings a negative impact. By contrast, our method (Row 4, Tab. 6), outperforms it, achieving $4.00\mathrm{m} / 6.57^{\circ}$ vs. $6.96\mathrm{m} / 7.47^{\circ}$ , while avoiding significant localization accuracy losses on normal trajectories. These results highlight the superior robustness of our approach. + +Study on PCC. We further conduct ablation experiments to demonstrate the importance of PCC. The comparison be + +
AugmentationPCCEKFAAverage
1190.80m, 46.67°
2✓6.96m, 7.47°
3✓5.79m, 8.02°
4✓✓4.00m, 6.57°
+ +Table 6. Ablation study on the BiLiLo dataset. Augmentation: using data augmentation to increase the rotational diversity of training data. PCC: using Point Cloud Canonicalization process to perceptive rotation and align point cloud into canonicalization coordinate. EKFA: using Equivariant Key Feature Aggregation module to guide feature aggregation. + +tween Row 1 and Row 3 show PCC can solve the challenges brought by rotation changes. Moreover, from Row 2 and Row 3, when compared to data augmentation, PCC also takes an average improvement of $16.8\%$ on position accuracy. This shows that PCC can get a better result than rotation augmentation. + +Study on EKFA. We evaluate the EKFA module's impact on handling noise in real-world LiDAR point clouds. The EKFA module is designed to mitigate scene noise, aiming to enhance the equivariant network's robustness to rotation and reduce noise-induced localization errors. Compared to Row 3 and Row 4, it yields a process of $30.9\% /18.1\%$ on position and orientation accuracy, respectively. Compared to the data augmentation, it also get an improvement of $42.5\% /12.0\%$ . Moreover, as shown in Tab. 5, integrating EKFA into PCC improves position accuracy by an average of $11.0\%$ on normal trajectories compared to the vanilla method (SGLoc in this paper). These results show that the EKFA module effectively extracts scene-relevant information, reducing noise impact on the equivariant network and enhancing localization robustness. + +# 5. Conclusion + +In this paper, we reveal that current learning-based regression methods are sensitive to rotations, resulting in significant accuracy loss under rotational conditions. We propose RALoc, a novel rotation-robust scene coordinate regression framework for LiDAR-based localization. To more realistically evaluate the rotational robustness of the localization method, we establish equipment setups and collect a large-scale outdoor localization dataset, BiLiLo, with significant rotational variations. Extensive experiments demonstrate that the proposed RALoc significantly outperforms the state-of-the-art methods in scenarios with large rotations. In the future, we aim to enhance the training efficiency of RALoc and integrate multimodal information to further improve localization accuracy. + +Acknowledgements. This work is supported in part by the Fundamental Research Funds for the Central Universities (No. 20720230033) and Xiaomi Young Talents Program. + +# References + +[1] Sheng Ao, Qingyong Hu, Bo Yang, Andrew Markham, and Yulan Guo. Spinnet: Learning a general surface descriptor for 3d point cloud registration. In CVPR, pages 11753-11762, 2021. 2 +[2] Sheng Ao, Qingyong Hu, Hanyun Wang, Kai Xu, and Yulan Guo. Buffer: Balancing accuracy, efficiency, and generalizability in point cloud registration. In CVPR, pages 1255-1264, 2023. 2 +[3] Serge Assaad, Carlton Downey, Rami Al-Rfou, Nigamaa Nayakanti, and Ben Sapp. Vn-transformer: Rotation-equivariant attention for vector neurons. arXiv preprint arXiv:2206.04176, 2022. 3, 4 +[4] Dan Barnes, Matthew Gadd, Paul Murcutt, Paul Newman, and Ingmar Posner. The oxford radar robotcar dataset: A radar extension to the oxford robotcar dataset. In ICRA, pages 6433-6438. IEEE, 2020. 2, 5, 6 +[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 5 +[6] Georg Bokman, Johan Edstedt, Michael Felsberg, and Fredrik Kahl. Steerers: A framework for rotation equivariant keypoint descriptors. In CVPR, pages 4885-4895, 2024. 2, 3 +[7] Eric Brachmann and Carsten Rother. Visual camera re-localization from rgb and rgb-d images using dsac. IEEE TPAMI, 44(9):5847-5865, 2021. 2 +[8] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In CVPR, pages 6684-6692, 2017. +[9] Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. In CVPR, pages 5044-5053, 2023. 2 +[10] Keenan Burnett, David J Yoon, Yuchen Wu, Andrew Z Li, Haowei Zhang, Shichen Lu, Jingxing Qian, Wei-Kang Tseng, Andrew Lambert, Keith YK Leung, et al. Boreas: A multi-season autonomous driving dataset. The International Journal of Robotics Research, 42(1-2):33-42, 2023. 5 +[11] Nicholas Carlevaris-Bianco, Arash K Ushani, and Ryan M Eustice. University of michigan north campus long-term vision and lidar dataset. The International Journal of Robotics Research, 35(9):1023-1035, 2016. 2, 5, 6 +[12] Daniele Cattaneo, Matteo Vaghi, and Abhinav Valada. Lcd-net: Deep loop closure detection and point cloud registration for lidar slam. IEEE T-RO, 38(4):2074–2093, 2022. 1 +[13] Haiwei Chen, Shichen Liu, Weikai Chen, Hao Li, and Randall Hill. Equivariant point network for 3d point cloud analysis. In CVPR, pages 14514-14523, 2021. 3 +[14] Yunlu Chen, Basura Fernando, Hakan Bilen, Matthias Nießner, and Efstratios Gavves. 3d equivariant graph implicit functions. In ECCV, pages 485-502. Springer, 2022. 3 +[15] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022. 6 + +[16] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In ICCV, pages 12200-12209, 2021. 3, 4 +[17] Fadi Dornaika and Radu Horaud. Simultaneous robot-world and hand-eye calibration. IEEE transactions on Robotics and Automation, 14(4):617-622, 1998. 5 +[18] Zhaoxin Fan, Zhenbo Song, Wenping Zhang, Hongyan Liu, Jun He, and Xiaoyong Du. Rpr-net: A point cloud-based rotation-aware large scale place recognition network. In ECCV, pages 709-725. Springer, 2022. 5 +[19] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 15:381-395, 1981. 2 +[20] Fabian Fuchs, Daniel Worrall, Volker Fischer, and Max Welling. Se (3)-transformers: 3d roto-translation equivariant attention networks. NeurIPS, 33:1970–1981, 2020. 3 +[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 5 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 3 +[23] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In CVPR, pages 4267-4276, 2021. 2 +[24] Oren Katzir, Dani Lischinski, and Daniel Cohen-Or. Shape-pose disentanglement using se (3)-equivariant vector neurons. In ECCV, pages 468-484. Springer, 2022. 3, 5 +[25] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In ICCV, pages 2938-2946, 2015. 2, 3 +[26] Giseop Kim, Yeong Sang Park, Younghun Cho, Jinyong Jeong, and Ayoung Kim. Mulran: Multimodal range dataset for urban place recognition. In ICRA, pages 6246-6253. IEEE, 2020. 5, 6 +[27] Giseop Kim, Sunwook Choi, and Ayoung Kim. Scan context++: Structural place recognition robust to rotation and lateral variations in urban environments. IEEE T-RO, 38(3): 1856-1874, 2021. 1 +[28] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[29] Jacek Komorowski. Minkloc3d: Point cloud based large-scale place recognition. In WACV, pages 1790-1799, 2021. 2 +[30] Wen Li, Shangshu Yu, Cheng Wang, Guosheng Hu, Siqi Shen, and Chenglu Wen. Sgloc: Scene geometry encoding for outdoor lidar localization. In CVPR, pages 9286-9295, 2023. 2, 3, 4, 5, 6, 7 +[31] Wen Li, Yuyang Yang, Shangshu Yu, Guosheng Hu, Chenglu Wen, Ming Cheng, and Cheng Wang. Diffloc: Diffusion model for outdoor lidar localization. In CVPR, pages 15045-15054, 2024. 2, 3, 6, 7 +[32] Chien Erh Lin, Jingwei Song, Ray Zhang, Minghan Zhu, and Maani Ghaffari. Se (3)-equivariant point cloud-based place recognition. pages 1520-1530. PMLR, 2023. 3 + +[33] Quan Liu, Hongzi Zhu, Zhenxi Wang, Yunsong Zhou, Shan Chang, and Minyi Guo. Extend your own correspondences: Unsupervised distant point cloud registration by progressive distance extension. In CVPR, pages 20816-20826, 2024. 2 +[34] Weixin Lu, Yao Zhou, Guowei Wan, Shenhua Hou, and Shiyu Song. L3-net: Towards learning based lidar localization for autonomous driving. In CVPR, pages 6389-6398, 2019. 1 +[35] Lun Luo, Shuhang Zheng, Yixuan Li, Yongzhi Fan, Beinan Yu, Si-Yuan Cao, Junwei Li, and Hui-Liang Shen. Bevplace: Learning lidar-based place recognition using bird's eye view images. In ICCV, pages 8700-8709, 2023. 1 +[36] Shitong Luo, Jiahan Li, Jiaqi Guan, Yufeng Su, Chaoran Cheng, Jian Peng, and Jianzhu Ma. Equivariant point cloud analysis via learning orientations for message passing. In CVPR, pages 18932-18941, 2022. 3 +[37] Junyi Ma, Jun Zhang, Jintao Xu, Rui Ai, Weihao Gu, and Xieyuanli Chen. Overlaptransformer: An efficient and yaw-angle-invariant transformer network for lidar-based place recognition. 7(3):6958-6965, 2022. 2 +[38] Arnab Kumar Mondal, Siba Smarak Panigrahi, Oumar Kaba, Sai Rajeswar Mudumba, and Siamak Ravanbakhsh. Equivariant adaptation of large pretrained models. NeurIPS, 36: 50293-50309, 2023. 2, 3, 5 +[39] Mohamed Adel Musallam, Vincent Gaudilliere, Miguel Ortiz Del Castillo, Kassem Al Ismaeil, and Djamila Aouada. Leveraging equivariant features for absolute pose regression. In CVPR, pages 6876-6886, 2022. 2 +[40] Yue Pan, Xingguang Zhong, Louis Wiesmann, Thorbjörn Posewsky, Jens Behley, and Cyril Stachniss. Pin-slam: Li-dar slam using a point-based implicit neural representation for achieving global map consistency. IEEE T-RO, 2024. 1 +[41] 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. NeurIPS, 32, 2019. 6 +[42] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In CVPR, pages 11143-11152, 2022. 2 +[43] Victor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In ICML, pages 9323-9332. PMLR, 2021. 3 +[44] Torsten Sattler, Qunjie Zhou, Marc Pollefeys, and Laura Leal-Taixe. Understanding the limitations of cnn-based absolute camera pose regression. In CVPR, pages 3302-3312, 2019. 3 +[45] Chenghao Shi, Xieyuanli Chen, Junhao Xiao, Bin Dai, and Huimin Lu. Fast and accurate deep loop closing and relocalization for reliable lidar slam. IEEE T-RO, 2024. 1 +[46] Linus Svärn, Olof Enqvist, Fredrik Kahl, and Magnus Oskarsson. City-scale localization for cameras with known vertical direction. IEEE TPAMI, 39(7):1455-1461, 2016. 2 +[47] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field + +networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 2 +[48] Mikaela Angelina Uy and Gim Hee Lee. Pointnetvlad: Deep point cloud based retrieval for large-scale place recognition. In CVPR, pages 4470-4479, 2018. 2 +[49] Sijie Wang, Qiyu Kang, Rui She, Wei Wang, Kai Zhao, Yang Song, and Wee Peng Tay. Hypliloc: Towards effective lidar pose regression with hyperbolic fusion. In CVPR, pages 5176-5185, 2023. 3 +[50] Wei Wang, Bing Wang, Peijun Zhao, Changhao Chen, Ronald Clark, Bo Yang, Andrew Markham, and Niki Trigoni. Pointloc: Deep pose regressor for lidar point cloud localization. IEEE Sensors Journal, 22(1):959-968, 2021. 2, 3 +[51] Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. NeurIPS, 32, 2019. 3 +[52] Hai Wu, Chenglu Wen, Wei Li, Xin Li, Ruigang Yang, and Cheng Wang. Transformation-equivariant 3d object detection for autonomous driving. In AAAI, pages 2795-2802, 2023. 2, 3 +[53] Yan Xia, Mariia Gladkova, Rui Wang, Qianyun Li, Uwe Stilla, Joao F Henriques, and Daniel Cremers. Casspr: Cross attention single scan place recognition. In ICCV, pages 8461-8472, 2023. 2 +[54] Xuecheng Xu, Sha Lu, Jun Wu, Haojian Lu, Qiuguo Zhu, Yiyi Liao, Rong Xiong, and Yue Wang. Ring++: Rototranslation-invariant gram for global localization on a sparse scan map. IEEE T-RO, 2023. 1 +[55] Bochun Yang, Zijun Li, Wen Li, Zhipeng Cai, Chenglu Wen, Yu Zang, Matthias Muller, and Cheng Wang. Lisa: Lidar localization with semantic awareness. In CVPR, pages 15271-15280, 2024. 1, 2, 3, 4, 6, 7 +[56] Jingyun Yang, Congyue Deng, Jimmy Wu, Rika Antonova, Leonidas Guibas, and Jeannette Bohg. Equivact: Sim (3)-equivariant visuomotor policies beyond rigid object manipulation. In ICRA, pages 9249-9255. IEEE, 2024. 3 +[57] Huan Yin, Xuecheng Xu, Sha Lu, Xieyuanli Chen, Rong Xiong, Shaojie Shen, Cyril Stachniss, and Yue Wang. A survey on global lidar localization: Challenges, advances and open problems. IJCV, 132(8):3139-3171, 2024. 1 +[58] Shangshu Yu, Xiaotian Sun, Wen Li, Chenglu Wen, Yunuo Yang, Bailu Si, Guosheng Hu, and Cheng Wang. Nidaloc: Neurobiologically inspired deep lidar localization. IEEE TITS, 2023. 3 +[59] Chongjian Yuan, Jiarong Lin, Zuhao Zou, Xiaoping Hong, and Fu Zhang. Std: Stable triangle descriptor for 3d place recognition. In ICRA, pages 1897-1903, 2023. 2 +[60] Liyuan Zhu, Shengyu Huang, Konrad Schindler, and Iro Armeni. Living scenes: Multi-object relocalization and reconstruction in changing 3d environments. In CVPR, pages 28014-28024, 2024. 3 +[61] Minghan Zhu, Maani Ghaffari, William A Clark, and Huei Peng. E2pn: Efficient se (3)-equivariant point network. In CVPR, pages 1223-1232, 2023. 3 \ No newline at end of file diff --git a/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/images.zip b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7eaab3efee56087ee401a164bf8fbb29cd68a19e --- /dev/null +++ b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b93da8423a002abaffd4cd41fa7aa5a0ae35e402ea7805549e5f32f4542fe21d +size 638784 diff --git a/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/layout.json b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e9df5cbd03aca366a68f3145f04059628bfaf02f --- /dev/null +++ b/ICCV/2025/RALoc_ Enhancing Outdoor LiDAR Localization via Rotation Awareness/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd536b6f96253fdceda0712a9feb13f34ec2598cc56c014299f0524b7fc9cda4 +size 399337 diff --git a/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/36b702cc-9b8c-4901-b825-63e6ad466de8_content_list.json b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/36b702cc-9b8c-4901-b825-63e6ad466de8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7b203388b954780193a2a82ff70dfbf3b0c5f13d --- /dev/null +++ b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/36b702cc-9b8c-4901-b825-63e6ad466de8_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb13f6ae1ce026700d4fefa043185b582a68a463c863ce1b584f6fd1275d35e2 +size 74951 diff --git a/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/36b702cc-9b8c-4901-b825-63e6ad466de8_model.json b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/36b702cc-9b8c-4901-b825-63e6ad466de8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c0eaabb743e3ddc9bd0355af43328549507b07f2 --- /dev/null +++ b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/36b702cc-9b8c-4901-b825-63e6ad466de8_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b92496f26e9c6222be5af8b96eec73e8cb86939783b501df2d958f0d5ceab0c4 +size 94423 diff --git a/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/36b702cc-9b8c-4901-b825-63e6ad466de8_origin.pdf b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/36b702cc-9b8c-4901-b825-63e6ad466de8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dce3a1f5f1b13a38615b8b2ec732bcc12e35301e --- /dev/null +++ b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/36b702cc-9b8c-4901-b825-63e6ad466de8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f5c12148758b661e67df38c5259d16359e96c0e3f4eafd4fc8e41e01483d649 +size 8796446 diff --git a/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/full.md b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c277770c0244917f63e4895c41eb277287b9d114 --- /dev/null +++ b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/full.md @@ -0,0 +1,306 @@ +# RANKCLIP: Ranking-Consistent Language-Image Pretraining + +Yiming Zhang\*1,2, Zhuokai Zhao\*3, Zhaorun Chen\*3, Zhili Feng4, Zenghui Ding\*1, Yining Sun\*1 HFIPS, Chinese Academy of Sciences 2 University of Science and Technology of China 3 University of Chicago 4 Carnegie Mellon University + +yimingzhang@mail.ustc.edu.cn, zhuokai@uchicago.edu + +# Abstract + +Self-supervised contrastive learning models, such as CLIP, have set new benchmarks for vision-language models in many downstream tasks. However, their dependency on rigid one-to-one mappings overlooks the complex and latent semantic attributes relationships between and within texts and images. To this end, we introduce RANKCLIP, a novel pretraining method that extends beyond the rigid one-to-one matching framework of CLIP and its variants. By extending the traditional pair-wise contrastive loss to list-wise ranking consistency loss, and leveraging both in-modal and cross-modal ranking consistency, RANKCLIP improves the alignment process, enabling it to capture the nuanced many-to-many relationships between and within each modality. Through comprehensive experiments, we demonstrate the effectiveness of RANKCLIP in various downstream tasks, achieving significant gains over state-of-the-art methods, underscoring the importance of this enhanced learning process. Code is available at https://github.com/Jam1ezhang/RankCLIP + +# 1. Introduction + +In the realm of computer vision (CV) [55], natural language processing (NLP) [7], and multimodal deep learning [4, 25, 60], the alignment between visual and textual modalities [5, 49] has emerged as a cornerstone for downstream applications, ranging from image captioning [18] to zero-shot classification [42]. Contrastive Language-Image Pretraining (CLIP) [43] marks a significant advancement in this field, demonstrating incredible performance from training on large amounts of text-image pairs to create self-supervised models that understand [6, 22, 23] and generate [11, 44] descriptions of visual contents. Following the success of this contrastive learning paradigm, many recent works have been developed upon the original CLIP. More specifically, these enhancements focus on optimizing data efficiency through intrinsic supervision [28], as well + +![](images/6674a22815789f9455932c8adef998f927653f0d2b8f49d514c4e4c0952310b8.jpg) +Figure 1. Comparison of learning outcomes between CLIP and RANKCLIP using three text-image pairs: dog (red), cat (blue), and car (yellow). (a) Contrastive loss treats all unmatched relationships equally, failing to distinguish latent similar attributes between dog and cat versus airplane. RANKCLIP addresses this issue by leveraging the shared attributes in (c) during training, improving the final trained embedding distribution from (b) to (d). + +as improving downstream performance via cross-modal late interaction [59], hierarchical feature alignment [15], geometric consistency regularization [19], additional learning [36], adaptive loss [58], hierarchy-aware attentions [17], and softer cross-modal alignment [16]. + +Despite the improvements, these methods often have reliance on strict pairwise, cross-modal, and one-to-one mappings between images and texts, overlooking the actual many-to-many relationships that exist both cross-modal and + +in-modal in real-world data [8]. For example, as shown in Fig. 1, while pretrained models like CLIP can correctly classify dog, cat and airplane, they do not explicitly learn that dog and cat are more close to each other than dog and airplane, in terms of both in-modal (dog text is more similar to cat text than to airplane text) and cross-modal (dog text is more matched to cat image than to airplane image) similarities. Since it is rooted from the current contrastive loss that only the correct pairs are optimized while the rest of the unmatched pairs are treated the same, a large amount of information not used and unknown to the model during and after the training process. + +Recognizing the complex many-to-many relationships as well as the rich information contained within both in-modal and cross-modal data, we introduce **Ranking-Consistent Language-Image Pretraining, (RANKCLIP)**, which employs ranking consistency to learn and optimize similarity levels both between (cross-modal) and within (in-modal) the text-image pairs. The concept of ranking consistency stems from the simple observations that similar texts often correlate with similar images, as seen with the dog, cat and airplane example in Fig. 1. It effectively captures secondary similarity relationships among unmatched pairs, enabling the model to learn more efficiently for free compared to relying solely on matched pairs. Ranking consistency is conveniently modeled as an additional loss term to the traditional contrastive loss, requiring no extra external modules. It acts as a plug-and-play improvement for many existing methods, including those focusing on data-efficiency [28], potentially boosting performance in both efficiency and effectiveness. + +The main contributions of this paper are: 1) RANKCLIP, a novel contrastive pretraining method that uses ranking consistency to exploit the many-to-many relationships within data, thereby enhancing performance in downstream tasks such as zero-shot classification and retrieval accuracy; and 2) through comprehensive experiments conducted on multiple datasets, we demonstrate the superior effectiveness of RANKCLIP in improving pretraining model performance without requiring any additional data or computational resources. + +# 2. Related Work + +Vision-language pretraining has witnessed significant advancements over the past years [3, 14, 31]. Models such as CLIP [43], ALIGN [26] and FLAVA [49] have pioneered the contrastive learning paradigm applied with text-image pairs, showcasing remarkable performance and robustness in downstream tasks. Many follow-up works, mostly built upon CLIP, have been proposed since then. Li et al. [28] introduced DeCLIP, improving zero-shot performance through intrinsic supervision. FILIP [59] advances CLIP's alignment between image patches and text with a cross + +modal interaction mechanism. Gao et al. [15] developed PyramidCLIP, using hierarchical feature alignment to boost model efficiency and performance. Additionally, SLIP [36] merges self-supervised learning with CLIP pre-training for improved visual representation and accuracy. Goel et al. [19] introduced CyCLIP, augmenting CLIP with geometric consistency regularizers to enhance robustness and performance under varied conditions. + +Recently, Yang et al. [58] introduced ALIP, an adaptive pre-training model that enhances language-image alignment using raw text and synthetic captions with dynamic adjustments. HiCLIP [17] refines CLIP by adding hierarchy-aware attentions to uncover semantic hierarchies in images and texts. EqSim [57] incorporates equivariance loss into vision-language models, significantly improving sensitivity to semantic changes in image-text pairs. Additionally, SoftCLIP [16] softens CLIP's one-to-one constraint, enabling more flexible cross-modal alignment through fine-grained adjustments. + +Compared with existing approaches, RANKCLIP sets itself apart by fully leveraging the many-to-many relationships within each batch of text-image pairs, promoting learning from both matched and unmatched pairs with varying similarities by integrating in-modal and cross-modal list-wise ranking consistencies into the contrastive training objective. Crucially, RANKCLIP diverges from existing models' pair-wise training objective by adopting a global, list-wise optimization approach. In other words, it considers the rankings of all images and texts collectively within each batch, rather than focusing on pairwise similarities as seen in other methods. + +# 3. RANKCLIP + +RANKCLIP efficiently leverages the many-to-many relationships in real-world data by focusing on both matched and unmatched pairs. As in Fig. 2, it not only identifies if an image-text pair matches but also assesses their relative semantic similarities to other images and texts of both modalities in the dataset through self-supervised ranking consistency. Uniquely, RANKCLIP employs a list-wise loss for training batches, distinguishing it from other methods that solely rely on pair-wise relationships, as discussed in §2. + +# 3.1. Ranking Model Formulation + +RANKCLIP leverages the Plackett-Luce (PL) ranking model [20, 32, 41] to estimate the probability distribution over rankings for every image-text pair $(V_{i},T_{j})$ , so that the consistency in their relative ordering with respect to a reference ranking can be measured. Specifically, for a given data pair, whether it is in-modal (image-image, text-text), or cross-modal (image-text), we calculate its in- or cross-modal cosine similarity $S_{ij}$ to serve as the score when measuring the alignment of its ranking with respect to another + +![](images/01f6702ad53f0fb47b06dd4ba1bb8d103ed8cf83024f198174a0331d3cec5ee3.jpg) +Figure 2. Overview of RANKCLIP. Unlike conventional contrastive loss, which includes only the middle term, RANKCLIP introduces both cross-modal and in-modal consistency terms by minimizing a self-supervised, list-wise ranking loss. Paired images and texts are indicated by matching contour line colors. $V$ , $T$ , and $S$ represent image embeddings, text embeddings, and similarity scores, respectively. + +reference ranking $y_{\mathrm{ref}}$ + +Following [41], we first sort the reference ranking in a descending order to construct the optimal ranking $y^{*}$ , and assume that the ego ranking $y$ is sampled from $y^{*}$ . The probability that item $d$ with score $S_{ij}$ is ranked $k^{\text{th}}$ in the ego ranking $y$ from a set of items $\mathcal{D}$ is the score of $e^{S_{ij}}$ divided by the sum of scores for the items that have not been placed yet: + +$$ +\pi (d \mid y _ {1: k - 1}, y _ {\text {r e f}}, \mathcal {D}) = \frac {e ^ {S _ {i j}}}{\sum_ {d ^ {\prime} \in \mathcal {D} \backslash y _ {1 : k - 1}} e ^ {S _ {i j} ^ {\prime}}}, \tag {1} +$$ + +where $y_{1:k-1} = [y_1, y_2, \dots, y_{k-1}]$ denotes the set of items ranked before $d$ . Consequently, the probability of the entire ranking $y$ is the product of all individual placement probabilities: + +$$ +\mathcal {P} \left(y, y _ {\text {r e f}}\right) = \prod_ {k = 1} ^ {K} \pi \left(y _ {k} \mid y _ {1: k - 1}, \mathbf {y} _ {\text {r e f}}, \mathcal {D}\right). \tag {2} +$$ + +RANKCLIP's objective is to maximize the consistency log-likelihood of the list ranking in one modality towards the reference ranking (from the same/in-modal and different/cross-modal data), which conveniently aligns with minimizing the negative log-likelihood loss: + +$$ +\mathcal {L} _ {\mathrm {P L}} = - \log \mathcal {P} (y, y _ {\text {r e f}}) \tag {3} +$$ + +# 3.2. Cross-modal Consistency Ranking + +As illustrated by the green box in Fig. 2, RANKCLIP utilizes secondary relationships between unmatched visual and + +textual representations by constructing a list-wise rank loss. This approach ensures that the semantic similarity rankings between one image and multiple texts align with those between one corresponding text and multiple images. For example, as shown in Fig. 1, from the dog perspective, the semantic distance between dog image and cat text is closer compared to the plane text. This relationship should also apply between the dog text and the cat, plane images. Mathematically, Eq. (3) can be specified as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c r o s s - m o d a l}} = - \log \mathcal {P} (\mathbf {y} _ {\text {i m a g e - t e x}}, \mathbf {y} _ {\text {t e x t - i m a g e}}) (4) \\ = - \log \mathcal {P} (\hat {\mathbf {v}} \cdot \hat {\mathbf {t}} ^ {\mathbf {T}}, \hat {\mathbf {t}} \cdot \hat {\mathbf {v}} ^ {\mathbf {T}}) (5) \\ \end{array} +$$ + +By optimizing Eq. (4), RANKCLIP enhances its ability to bridge the semantic gap between modalities by leveraging nuanced unmatched correlations. This can also be viewed as learning a symmetric cosine-similarity matrix, further reinforcing semantic consistency across modalities. + +# 3.3. In-modal Consistency Ranking + +The pink box in Fig. 2 highlights the in-modal consistency component of the proposed rank loss. RANKCLIP ensures semantic consistency within each modality – image to image and text to text – enhancing the use of secondary unmatched relationships as an optimization objective. The underlying principle is that similar images should correspond to similar texts. For example, in Fig. 1, from the dog image perspective, the cat image is the most similar, followed by the plane image. This relationship should hold true + +
ImageNet1KMSCOCO
Image-to-Text RetrievalText-to-Image Retrieval
Top-1Top-3Top-5Recall@1Recall@5Recall@10Recall@1Recall@5Recall@10
CLIP [43]9.06%16.94%21.63%6.68%18.36%26.94%3.70%9.74%14.04%
CyCLIP [19]9.40%17.32%21.72%6.50%19.34%29.14%3.72%11.16%16.06%
ALIP [58]9.71%18.31%23.07%6.04%18.04%26.92%3.70%10.22%14.38%
RANKCLIP10.16%19.57%24.01%7.18%19.46%28.48%3.74%10.28%14.18%
+ +Table 1. Zero-shot top-1, top-3, and top-5 classification accuracy on ImageNet1K, along with retrieval performance on MS-COCO. The proposed RANKCLIP consistently outperforms all baselines across both tasks. All models are trained on CC3M with ViT-B/32 backbone. + +for their corresponding texts as well, where we utilize this to construct our $y$ and $y_{\mathrm{ref}}$ from Eq. (3). Mathematically, Eq. (3) can be specified as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {i n - m o d a l}} = - \log \mathcal {P} \left(\mathbf {y} _ {\text {t e x t - t e x t}}, \mathbf {y} _ {\text {i m a g e - i m a g e}}\right) (6) \\ = - \log \mathcal {P} (\hat {\mathbf {t}} \cdot \hat {\mathbf {t}} ^ {\mathbf {T}}, \hat {\mathbf {v}} \cdot \hat {\mathbf {v}} ^ {\mathbf {T}}) (7) \\ \end{array} +$$ + +where $\hat{\mathbf{t}}$ and $\hat{\mathbf{v}}$ are the text and image batch embedding matrix, respectively. Via Eq. (6), the model can efficiently leverage the nuanced in-modal relationships to learn a richer and more structured semantic representation. + +# 3.4. RANKCLIP Loss + +Combining both cross-modal and in-modal consistency with the traditional contrastive loss (more details in Appendix B), the complete rank loss is thus formulated as: + +$$ +\mathcal {L} _ {\text {R A N K C L I P}} = \mathcal {L} _ {\text {C L I P}} + \lambda_ {1} \mathcal {L} _ {\text {i n - m o d a l}} + \lambda_ {2} \mathcal {L} _ {\text {c r o s s - m o d a l}} \tag {8} +$$ + +which is also depicted in Fig. 2. By supplementing the pairwise contrastive loss with cross-modal and in-modality ranking consistency loss, RANKCLIP systematically organizes embeddings to fully exploit both global and fine-grained unmatched relationships, which enhances the learning of more informative and accurate representations, better supporting downstream multi-modal tasks. + +# 3.5. Training Recipe on Selecting $\lambda_{1}$ and $\lambda_{2}$ + +In the early stage of pre-training, rank consistency is highly unstable due to random initialization. Overemphasizing ranking consistency at this stage can impede the optimization of the embedding space. To address this, we gradually increase the weights $\lambda_{1}$ and $\lambda_{2}$ of the ranking loss as training progresses. Specifically, we have: + +$$ +\lambda_ {1} = \lambda_ {2} = \operatorname {c l i p} \left(\frac {3 i - 1}{n - 1}, 0, 2\right) +$$ + +where $i$ and $n$ denote the current training epoch and total number of epoch, respectively. The full RANKCLIP framework is outlined in Algorithm 1. + +# 4. Experiments + +# 4.1. Experimental Setup + +Baselines. The most direct baseline to RANKCLIP is the original CLIP [43]. To further demonstrate the supe + +rior performance of RANKCLIP, we include CyCLIP [19], which introduces cyclic consistency constraints to enforce more robust alignment between visual and textual representations, improving generalization and semantic coherence. We also include ALIP [58], which leverages synthetic captions to enhance vision-language representation learning. More specifically, it employs a unique architecture that dynamically adjusts sample and pair weights to mitigate the impact of noisy or irrelevant data, making its approach complementary to ours. The training procedures and parameters for all models are detailed in Appendix A. + +Data. All approaches are pretrained on the Conceptual Captions 3M (CC3M) dataset [48], which contains approximately 3.3 million text-image pairs. Although significantly smaller than CLIP's original dataset [24], CC3M remains a standard benchmark for vision-language pretraining, enabling strong zero-shot performance [2, 19, 28, 36, 53]. As discussed in §5.2, we also train CLIP and RANKLIP on a larger 15M-image subset of YFCC100M [54] (YFCC15M) to study the effect of scaling up the dataset size. + +# 4.2. Zero-shot Classification + +We evaluate the zero-shot classification performance of CLIP [43], CyCLIP [19], ALIP [58], and RANKCLIP on ImageNet1K [12, 46]. As shown in Table 1, RANKCLIP consistently outperforms CLIP, which highlights the effectiveness of ranking consistency in enhancing language-image alignment with the same training data. Compared to CyCLIP [19], which enforces cyclic consistency to improve semantic coherence, RANKCLIP achieves higher accuracy across all metrics, suggesting that ranking consistency provides a more direct and effective regularization for representation learning. Additionally, RANKCLIP surpasses ALIP [58], indicating that ranking consistency is a stronger alternative to synthetic caption-based supervision. Notably, RANKCLIP shows the most significant improvement in top-1 accuracy, reinforcing its practical advantages where the highest-ranked prediction is most critical. + +# 4.3. Zero-shot Cross-modal Retrieval + +We further evaluate RANKCLIP on zero-shot cross-modal retrieval tasks, including image-to-text and text-to-image retrieval, using the MSCOCO [30] dataset. As shown in Table 1, RANKCLIP outperforms all baselines, though the + +
ImageNetV2-MatchedImageNetV2-ThresholdImageNetV2-TopImageNet-R
Top1Top3Top5Top1Top3Top5Top1Top3Top5Top1Top3Top5
CLIP [43]7.53%14.99%19.61%8.89%17.22%21.86%10.76%19.80%24.87%9.36%10.56%19.76%
CyCLIP [19]7.68%15.07%19.11%9.10%17.42%21.94%11.20%20.18%25.34%9.23%16.72%21.64%
ALIP [58]7.82%15.56%19.81%9.65%18.31%22.85%11.43%20.88%26.10%10.92%20.27%26.24%
RANKCLIP9.01%16.95%21.12%10.32%19.31%24.13%12.31%22.11%27.17%11.34%20.88%26.94%
+ +Table 2. Zero-shot top-1, 3, and 5 accuracy on ImageNet1K variants with natural distribution shifts. Compared to baselines, RANKCLIP achieves higher accuracies. Notably, these gains are more pronounced than on standard ImageNet1K, highlighting improved robustness. + +
CIFAR-10CIFAR-100DTDFGVCAircraftFood101GTSRBOxfordPetsSST2STL10SVHNAverage
CLIP [43]77.6%56.2%43.2%22.6%39.7%60.0%40.4%51.0%79.0%50.5%52.0%
CyCLIP [19]76.8%54.3%45.8%19.2%37.5%58.6%44.2%51.5%82.3%41.3%51.2%
ALIP [58]71.1%49.1%47.1%17.4%36.1%51.5%41.9%53.3%81.0%38.3%48.7%
RANKCLIP78.4%56.6%42.4%23.4%40.2%60.6%40.6%53.4%79.6%47.7%52.3%
+ +![](images/31d7130f45b1d82da3a58469ac680f37ad1eeb69df1a93def1b4b015c0546910.jpg) +Table 3. Linear probing accuracy on 10 downstream datasets using a ViT backbone. +Figure 3. Effect of $\lambda_{1}$ and $\lambda_{2}$ on zero-shot classification (ImageNet1K) and retrieval (MSCOCO). + +improvements are less pronounced compared to zero-shot classification. This smaller margin of improvement may be attributed to the increased complexity of retrieval tasks, which require fine-grained image-text alignment across varying resolutions and object details - challenges distinct from the more direct pattern recognition in classification. + +# 4.4. Robustness to Distribution Shifts + +To evaluate the robustness of RANKCLIP under distribution shifts, we test all approaches on three variants of ImageNetV2 [45] and ImageNet-R [22], which assess resilience to different real-world deviations from ImageNet1K. As shown in Table 2, RANKCLIP consistently outperforms all baselines, achieving the highest accuracy across all datasets. These results indicate that RANKCLIP not only improves standard zero-shot classification but also enhances adaptation to real-world distribution shifts. + +# 4.5. Linear Probing + +We further assess whether the advantages of ranking consistency persist when supplemented with in-domain supervision. Specifically, we apply linear probing, where pretrained encoders remain fixed while a logistic regression + +classifier is trained on domain-specific datasets. We evaluate on 10 standard image classification benchmarks, including CIFAR-10, CIFAR-100, DTD [9], FGVG-Aircraft [33], Food101 [1], GTSDB [52], OxfordPets [40], SST2 [50], STL-10 [10], and SVHN [37]. As shown in Table 3, RANKCLIP achieves the highest average accuracy, demonstrating that ranking consistency enhances generalization even with additional in-domain supervision. + +# 5. Ablation Studies + +# 5.1. Different Weights of RANKCLIP Loss + +In Eq. (8), we define the RANKCLIP loss as a linear combination of the original contrastive loss and the in-modal and cross-modal ranking consistency losses, weighted by $\lambda_{1}$ and $\lambda_{2}$ , respectively. In §3.5, we introduce a training strategy that adaptively adjusts $\lambda_{1}$ and $\lambda_{2}$ at different stages of training. In this section, we analyze the impact of these weights and demonstrates that adaptive weighting further enhances RANKCLIP's performance. Notice that all model variants follow the same pretraining setup detailed in Appendix A. As shown in Fig. 3, RANKCLIP outperforms CLIP even with fixed $\lambda_{1}$ and $\lambda_{2}$ , highlighting the effectiveness of ranking consistency. And the adaptive weighting strategy further boosts accuracy by preventing the underutilization of ranking consistency at low weights and avoiding disruptions to contrastive learning at high weights. + +# 5.2. Different Data Sizes + +To assess the scalability of RANKCLIP, we trained both CLIP and RANKCLIP on 500k, 750k, 1M, and 3M text-image pairs from CC3M, as well as 15M pairs from YFCC15M, following the procedure in Appendix A. Fig. 4 compares their performance on zero-shot top-1 classifica + +![](images/a59e84a4cf5c6fc6742860b82a89dfd5bce782a26d1d87ab88745c44ffcbc592.jpg) +Figure 4. Ablation studies of CLIP and RANKCLIP trained with different data sizes. Left: zero-shot top-1 classification accuracy on ImageNet1K with various data sizes randomly sampled from CC3M. RANKCLIP consistently outperforms CLIP with significant margins. Right: zero-shot top-1 classification accuracy on ImageNet1K (horizontal axis) and ImageNet1K-R (vertical axis). RANKCLIP demonstrates better robustness as well as accuracy. + +![](images/2b36aa2075c9fe495932ca5c5fd21b998d81326c6c303daac59adcba2f33297c.jpg) + +
MethodVision BackboneImageNet1KMSCOCOLinear Probing Avg. Acc.
Image-to-Text RetrievalText-to-Image Retrieval
Top-1Top-3Top-5R@1R@5R@10R@1R@5R@10
CLIP [43]RN5021.6%36.9%44.9%15.6%36.4%48.4%6.7%15.2%20.1%64.2%
RANKCLIP30.9%49.4%57.6%19.5%42.6%54.8%7.5%16.2%21.6%68.9%
CLIP [43]ViT-B/3220.7%35.0%42.4%11.9%29.4%40.8%5.1%12.9%17.9%60.7%
RANKCLIP26.2%41.4%48.9%13.8%33.8%45.9%6.0%13.6%18.6%61.3%
+ +Table 4. Zero-shot evaluation of CLIP and RANKCLIP trained with different vision backbones (ResNet-50 (RN50) and ViT-B/32) on ImageNet1K classification, MSCOCO cross-modal retrievals, and linear probing. “R@k” denotes Recall@k. + +tion accuracy for ImageNet1K (left) and averaged linear probing results (middle), where RANKCLIP consistently outperforms CLIP. Full, non-averaged linear probing results are provided in Appendix 7. Notably, the performance gains of RANKCLIP become more pronounced as dataset size scales from $1\mathrm{m}$ to $15\mathrm{m}$ , highlighting its superior scalability, which is critical for language-image pretraining. + +Fig. 4 (right) further illustrates RANKCLIP's robustness across different dataset sizes. The horizontal axis represents top-1 accuracy on standard ImageNet1K, while the vertical axis shows accuracy on ImageNet1K-R. The black diagonal $(y = x)$ denotes ideal robustness, with deviations below it indicating degradation under distribution shifts. RANKCLIP remains well above the red baseline, which represents typical in-distribution to out-of-distribution generalization [35], and stays close to the ideal line, demonstrating strong robustness. + +# 5.3. Different Backbones: RN50 vs. ViT + +We compare RANKCLIP and CLIP across ResNet-50 (RN50) and ViT-B/32 backbones to assess its generalization performance and overall robustness. As shown in Table 4, RANKCLIP consistently outperforms CLIP across all evaluated tasks. Specifically, for zero-shot classification on ImageNet1K, RANKCLIP improves top-1 accuracy by $+9.3\%$ with RN50 and $+5.5\%$ with ViT-B/32, demonstrating its fundamentally stronger feature learning capabilities. + +In cross-modal retrieval, RANKCLIP achieves $+3.9\%$ in image-to-text and $+0.8\%$ in text-to-image with RN50, with smaller but consistent gains for ViT-B/32. RANKCLIP also improves linear probing accuracy by $+4.7\%$ with RN50 and $+0.6\%$ with ViT-B/32, confirming its advantage in representation learning. While both architectures benefit, RN50 sees the largest gains, a result suggesting that ranking consistency particularly enhances the hierarchical feature extraction process in CNNs. + +# 6. Analysis + +# 6.1. Modality Gap + +We analyze the modality gaps of CLIP and RANKCLIP by visualizing 250 text-image pair embeddings, reduced to two dimensions using UMAP [34], and presenting a histogram of the gaps. The modality gap [29] refers to the separation between text and image embeddings in multimodal models, hindering joint representation learning. This gap, inherent from initialization and reinforced by contrastive learning in CLIP, challenges effective language-image modeling. Recent studies [27, 38, 51] suggest that reducing this gap improves multimodal representations and downstream performance. As shown in Fig. 6, RANKCLIP exhibits a significantly smaller modality gap than CLIP, demonstrating that our ranking consistency approach effectively enhances text-image alignment. + +![](images/e001b595e7648a8a2d8bd15a564d13f68b6d1c226f65c23265035c56578fc121.jpg) +Caption: A painting of a table with fruit on top of it + +# 6.2. Alignment and Uniformity + +Beyond reducing modality gap, effective contrastive learning should ensure a broad and uniform distribution over a hypersphere [56]. These objectives – similarity and uniformity – are quantified by alignment and uniformity scores, respectively. Following Goel et al. [19] and the notations in §3, we compute the alignment score $S_{\mathrm{A}}$ and uniformity score $S_{\mathrm{U}}$ as: + +$$ +S _ {\mathrm {A}} = \frac {1}{N} \sum_ {j = 1} ^ {N} \hat {I} _ {j} ^ {T} \hat {T} _ {j}, \tag {9} +$$ + +$$ +S _ {\mathrm {U}} = \log \left(\frac {1}{N (N - 1)} \sum_ {j - 1} ^ {N} \sum_ {k = 1, j \neq k} ^ {N} \exp^ {- \hat {I} _ {j} ^ {T} \hat {T} _ {k}}\right) \tag {10} +$$ + +where $N$ is the number of text-image pairs. $S_{\mathrm{A}}$ captures the average cosine similarity between corresponding text and image embeddings, while $S_{\mathrm{U}}$ measures how evenly embeddings are spread across the space. High alignment indicates strong correlation between paired embeddings, whereas low uniformity suggests diverse and efficient embedding distribution—desirable for tasks like retrieval. As shown in Table 5, CLIP achieves stronger alignment but suffers from poor uniformity, leading to redundant representations. On the other hand, RANKCLIP along with two of its ablated version, $\mathrm{RANKCLIP}_I$ and $\mathrm{RANKCLIP}_C$ presents much better balance between alignment and uniformity. These results further suggest that optimizing solely for alignment or uniformity does not necessarily translate to better task performance. + +![](images/6029461dbf5afd29185a0c4409dcf981f34f0ec9939e18d6432fa43a4ab6e660.jpg) +Figure 5. For a given text query, we present the top ten most semantically relevant images (ordered from left to right) obtained through both CLIP and RANKCLIP. In comparison to CLIP, our approach consistently retrieves images that more comprehensively align with the textual description, maintaining this advantage even after the correct reference image appears in the ranked results. + +![](images/70c8ce0199c7bfd605095ae5a8a3ba529f5fea4ed4471374544fdaeeeacb6853.jpg) +Figure 6. Scatter and histograms plots illustrating modality gaps of (a) CLIP and (b) RANKCLIP. + +
CIFAR-10CIFAR-100ImageNet1K
SASUSASUSASU
CLIP0.28-0.190.28-0.180.33-0.19
RANKCLIP0.23-0.140.26-0.130.29-0.13
RANKCLIPc0.23-0.130.24-0.120.29-0.14
RANKCLIPI0.25-0.150.28-0.140.32-0.17
+ +Table 5. Alignment and uniformity scores of CLIP, RANKCLIP, RANKCLIP $_C$ and RANKCLIP $_I$ indicate the solely cross-modal and in-modal rank loss. + +# 6.3. Qualitative Examples + +Class activation maps. To further examine the effects of ranking consistency, we visualize class activation maps + +![](images/002172d1d24a3021c222dd4ae1c1ce63531e03ec7a87d832e312021962627e99.jpg) +Figure 7. Class activation maps for RANKCLIP and CLIP on different objects in the caption from MSCOCO. RANKCLIP has more precise responses to some nouns compared to CLIP and can accurately locate the region related to the noun. + +(CAMs) [47] for RANKCLIP and CLIP in Fig. 7. The results show that RANKCLIP consistently attends to more semantically relevant regions in the images. For example, when given the caption 'An airport filled with planes sitting on tarmac', CLIP mistakenly highlights surrounding areas, whereas RANKCLIP focuses precisely on the planes. Similar improvements are observed across other examples, demonstrating that RANKCLIP better aligns textual descriptions with visual concepts. This suggests that ranking consistency enhances fine-grained feature learning, leading to more localized and accurate visual grounding. + +Text-to-image retrieval. Fig. 5 compares text-to-image retrieval results from RANKCLIP and CLIP. Given a query, we display the top-ranked images retrieved by each model. RANKCLIP consistently retrieves more semantically aligned images, even beyond the correct reference image. Regarding the qualitative example in Fig. 5, CLIP retrieves a table with fruit but overlooks the intended painting of fruits. Moreover, RANKCLIP consistently retrieves more relevant results: in the top 10 results, CLIP includes clear mismatches such as a cat sitting on a table or a motorcycle instead of a sink. This demonstrates RANKCLIP's + +ability to capture fine-grained semantic distinctions, thereby reinforcing its distinct advantage in retrieval tasks that demand a precise and nuanced understanding of the queries. + +# 7. Conclusion + +In this paper, we introduce RANKCLIP, a novel language-image pretraining method that integrates ranking consistency into the contrastive learning paradigm. RANKCLIP aims to better understand the complex many-to-many relationships in diverse text-image pairs by optimizing a self-supervised, list-wise rank loss. Through extensive experiments, including zero-shot classification, robustness to distribution shifts, linear probing, and zero-shot image-text retrieval, RANKCLIP not only enhances performance but also improves model robustness and semantic comprehension, outperforming the baseline CLIP and another state-of-the-art model ALIP. Our ablation studies and analyses further demonstrate and interpret the significance of each component of RANKCLIP in boosting performance and understanding across modalities. We believe that the methodologies and principles of RANKCLIP will inspire further research and lead to the development of models with a deeper understanding of the intricate interactions between visual and textual data. + +# Acknowledgements + +# References + +[1] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5 +[2] Nicholas Carlini and Andreas Terzis. Poisoning and backdooring contrastive learning. arXiv preprint arXiv:2106.09667, 2021. 4 +[3] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023. 2 +[4] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Qinglan Huang, et al. Mj-bench: Is your multimodal reward model really a good judge for text-to-image generation? arXiv preprint arXiv:2407.04842, 2024. 1 +[5] Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. Halc: Object hallucination reduction via adaptive focal-contrast decoding. arXiv preprint arXiv:2403.00425, 2024. 1 +[6] Zhaorun Chen, Zhuokai Zhao, Zhihong Zhu, Ruiqi Zhang, Xiang Li, Bhiksha Raj, and Huaxiu Yao. Autoprm: Automating procedural supervision for multi-step reasoning via controllable question decomposition. arXiv preprint arXiv:2402.11452, 2024. 1 +[7] KR1442 Chowdhary and KR Chowdhary. Natural language processing. Fundamentals of artificial intelligence, pages 603-649, 2020. 1 +[8] Sanghyuk Chun. Improved probabilistic image-text representations. arXiv preprint arXiv:2305.18171, 2023. 2 +[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 5 +[10] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 5 +[11] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In European Conference on Computer Vision, pages 88–105. Springer, 2022. 1 + +This work was supported by the National Key Research and Development Program of China (No. 2024YFF0507603); the Anhui Provincial Clinical Medical Research Transformation Project (No. 202204295107020004); and the Anhui Provincial Major Science and Technology Project (Nos. 202303a07020006, 202304a05020071). +[12] 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. 4 +[13] 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. 12 +[14] Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A survey of vision-language pre-trained models. arXiv preprint arXiv:2202.10936, 2022. 2 +[15] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, Rongrong Ji, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. Advances in neural information processing systems, 35:35959-35970, 2022. 1, 2 +[16] Yuting Gao, Jinfeng Liu, Zihan Xu, Tong Wu, Enwei Zhang, Ke Li, Jie Yang, Wei Liu, and Xing Sun. Softclip: Softer cross-modal alignment makes clip stronger. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1860-1868, 2024. 1, 2 +[17] Shijie Geng, Jianbo Yuan, Yu Tian, Yuxiao Chen, and Yongfeng Zhang. Hiclip: Contrastive language-image pretraining with hierarchy-aware attention. arXiv preprint arXiv:2303.02995, 2023. 1, 2 +[18] Taraneh Ghandi, Hamidreza Pourreza, and Hamidreza Mahyar. Deep learning approaches on image captioning: A review. ACM Computing Surveys, 56(3):1-39, 2023. 1 +[19] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. Advances in Neural Information Processing Systems, 35:6704-6719, 2022. 1, 2, 4, 5, 7 +[20] John Guiver and Edward Snelson. Bayesian inference for plackett-luce ranking models. In proceedings of the 26th annual international conference on machine learning, pages 377-384, 2009. 2 +[21] 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. 12 +[22] 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 Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 1, 5 +[23] 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. 1 +[24] 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. If you use this software, please cite it as below. 4 + +[25] Summaira Jabeen, Xi Li, Muhammad Shoib Amin, Omar Bourahla, Songyuan Li, and Abdul Jabbar. A review on methods and applications in multimodal deep learning. ACM Transactions on Multimedia Computing, Communications and Applications, 19(2s):1-41, 2023. 1 +[26] 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 +[27] Yogesh Kumar and Pekka Marttinen. Improving medical multi-modal contrastive learning with expert annotations. arXiv preprint arXiv:2403.10153, 2024. 6 +[28] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 1, 2, 4 +[29] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 6 +[30] 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. 4 +[31] Siqu Long, Feiqi Cao, Soyeon Caren Han, and Haiqin Yang. Vision-and-language pretrained models: A survey. arXiv preprint arXiv:2204.07356, 2022. 2 +[32] R Duncan Luce. Individual choice behavior: A theoretical analysis. Courier Corporation, 2005. 2 +[33] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. Technical report, 2013. 5 +[34] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 6 +[35] John P Miller, Rohan Taori, Aditi Raghunathan, Shiori Sagawa, Pang Wei Koh, Vaishaal Shankar, Percy Liang, Yair Carmon, and Ludwig Schmidt. Accuracy on the line: on the strong correlation between out-of-distribution and indistribution generalization. In International conference on machine learning, pages 7721-7735. PMLR, 2021. 6 +[36] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In European conference on computer vision, pages 529-544. Springer, 2022. 1, 2, 4 +[37] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, page 7. Granada, Spain, 2011. 5 +[38] Changdaee Oh, Junhyuk So, Hoyoon Byun, YongTaek Lim, Minchul Shin, Jong-June Jeon, and Kyungwoo Song. + +Geodesic multi-modal mixup for robust fine-tuning. Advances in Neural Information Processing Systems, 36, 2024. 6 +[39] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 12 +[40] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5 +[41] Robin L Plackett. The analysis of permutations. Journal of the Royal Statistical Society Series C: Applied Statistics, 24 (2):193-202, 1975. 2, 3 +[42] Farhad Pourpanah, Moloud Abdar, Yuxuan Luo, Xinlei Zhou, Ran Wang, Chee Peng Lim, Xi-Zhao Wang, and QM Jonathan Wu. A review of generalized zero-shot learning methods. IEEE transactions on pattern analysis and machine intelligence, 45(4):4051-4070, 2022. 1 +[43] 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, 4, 5, 6, 12 +[44] 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. 1 +[45] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International conference on machine learning, pages 5389-5400. PMLR, 2019. 5 +[46] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 4 +[47] 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 international conference on computer vision, pages 618-626, 2017. 8 +[48] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 4 +[49] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 1, 2 +[50] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher + +Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642, 2013. 5 +[51] Siddharth Srivastava and Gaurav Sharma. Omnivec: Learning robust representations with cross modal sharing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1236-1248, 2024. 6 +[52] J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. Neural Networks, (0):-, 2012. 5 +[53] Ajinkya Tejankar, Maziar Sanjabi, Bichen Wu, Saining Xie, Madian Khabsa, Hamed Pirsiavash, and Hamed Firooz. A fistful of words: Learning transferable visual models from bag-of-words supervision. arXiv preprint arXiv:2112.13884, 2021. 4 +[54] 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. 4 +[55] Athanasios Voulodimos, Nikolaos Doulamis, Anastasios Doulamis, and Eftychios Protopapadakis. Deep learning for computer vision: A brief review. Computational intelligence and neuroscience, 2018, 2018. 1 +[56] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 7 +[57] Tan Wang, Kevin Lin, Linjie Li, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Equivariant similarity for vision-language foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11998-12008, 2023. 2 +[58] 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. 1, 2, 4, 5, 12 +[59] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 1, 2 +[60] Zhuokai Zhao, Harish Palani, Tianyi Liu, Lena Evans, and Ruth Toner. Multi-modality guidance network for missing modality inference. arXiv preprint arXiv:2309.03452, 2023.1 \ No newline at end of file diff --git a/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/images.zip b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a1556dffa57a6f09780d6b52134540e08d14bacc --- /dev/null +++ b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e70814835e30803cddd48443b91012d07cd6160da73189a4ac8b5970391baa34 +size 878797 diff --git a/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/layout.json b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9f2da04882f27c41edc6d99fcf9a51b1f24f54d1 --- /dev/null +++ b/ICCV/2025/RANKCLIP_ Ranking-Consistent Language-Image Pretraining/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1feb3d994d77722f611ad0a7ab8564c34d0b507aab27a591ad19dd2c0e3049df +size 348872 diff --git a/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/31084f33-ade9-4a22-b199-dbcd5f86db88_content_list.json b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/31084f33-ade9-4a22-b199-dbcd5f86db88_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0a2dd937fcfd2df1bf524a2e97d7dd64f8a28488 --- /dev/null +++ b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/31084f33-ade9-4a22-b199-dbcd5f86db88_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e564e91fdaa93d6a5b2ab35f2f46abf220a3c71946bd0871da85ae99f8efd1c7 +size 77800 diff --git a/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/31084f33-ade9-4a22-b199-dbcd5f86db88_model.json b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/31084f33-ade9-4a22-b199-dbcd5f86db88_model.json new file mode 100644 index 0000000000000000000000000000000000000000..93878239faf12889ebbcd4bd227b6bb1cb5b1761 --- /dev/null +++ b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/31084f33-ade9-4a22-b199-dbcd5f86db88_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1e264d504047da94b5dd53b5355528ddd9802285fbe886114ed38159cc993a9 +size 96501 diff --git a/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/31084f33-ade9-4a22-b199-dbcd5f86db88_origin.pdf b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/31084f33-ade9-4a22-b199-dbcd5f86db88_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..461a658ee34c4d3d5f54c7f348203554eb3b8f74 --- /dev/null +++ b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/31084f33-ade9-4a22-b199-dbcd5f86db88_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9312c68dccdcc49e80f4cf7d6dee382f2edd6c46d82d652185827c65045f9cbe +size 4551055 diff --git a/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/full.md b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a71529a9037385297b5cd29443b54be3757b88e8 --- /dev/null +++ b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/full.md @@ -0,0 +1,310 @@ +# RARE: Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning + +Chengyu Zheng $^{1,2*}$ Jin Huang $^{1*}$ Honghua Chen $^{3\dagger}$ Mingqiang Wei $^{1,2\dagger}$ + +$^{1}$ College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics $^{2}$ Shenzhen Research Institute, Nanjing University of Aeronautics and Astronautics $^{3}$ School of Data Science, Lingnan University, Hong Kong SAR, China + +zhengcy@nuaa.edu.cn; jinhuang.nuaa@gmail.com; chenhonghuacn@gmail.com; mqwei@nuaa.edu.cn + +![](images/d9d22b112d3542ed0bf483b5974effee5458abd30122b62210efc1147bccbd93.jpg) + +![](images/4c0ada7e457b8af2c557322ee289f1ccc304983f8e923bb73cd4da91f962cd51.jpg) +(a) Source point cloud +(e) Source view depth map +Figure 1. RARE refines arbitrary point cloud registration methods effectively in a zero-shot learning manner. (a) and (e) show the source point cloud and its projected depth map from the source view. (b) and (f) depict the source point cloud rotated to the reference view using the initial transformation $T_{init}$ and its corresponding projected depth map. (c) and (g) display the reference point cloud and its projected depth map from the reference view. From (f) and (g), it is evident that the initial transformation allows us to extract high-quality correspondences from the depth maps. (d) shows the registration result using the initial transformation by GeoTransformer [31], while (f) presents the refined registration result by diffusion features. + +![](images/3ee93ed3bd384dc76c8da8a0f60588d87b2968de668eae8e3d7abbca1afd8e44.jpg) +(b) Rotated Source Cloud + +![](images/651900c060fd75480665e8d82d109309519b484a238bc5a2ae0d32df6b19b6a7.jpg) +(f) Rotated source view depth map + +![](images/b234e572c9839d5d7398a2873c3afa30a0eb16927d41068a9decd75d8bc137ee.jpg) + +![](images/261d66229b0fd34a714b9ab7165848529b1cd0d35570ec518d04f11f5f51e65a.jpg) +(c) Reference point cloud +(g) Reference view depth map + +![](images/8533ffdc71f2a0cccf5fa65ce7694593d0813b4460c6349281780a5b2806e4a3.jpg) +(d) Initial registration result +RE: $4.99^{\circ}$ +TE: 15.23cm + +![](images/f013db72f14e7ec9592cb7ae2dfca8d93a98176cc334293e59d14ce1ccc64a6f.jpg) +(h)Refined registration result +RE: 1.37° +TE: 5.87cm + +![](images/5f1034906f8a8b64bc0dc663e9039f63dd0920decf5e577c59d9bc51a67e60a4.jpg) + +![](images/275f5e11c7ed28ae93471edb0f29f64f8ee3f9047d8813081afeabf2e036d56f.jpg) + +# Abstract + +Recent research leveraging large-scale pretrained diffusion models has demonstrated the potential of using diffusion features to establish semantic correspondences in images. Inspired by advancements in diffusion-based techniques, we propose a novel zero-shot method for refining point cloud registration algorithms. Our approach leverages correspondences derived from depth images to enhance point feature representations, eliminating the need for a dedicated training dataset. Specifically, we first project the point cloud into depth maps from multiple perspectives and extract implicit knowledge from a pretrained diffusion network as depth diffusion features. These features are then integrated with geometric features obtained from existing methods to establish more accurate corre + +spondences between point clouds. By leveraging these refined correspondences, our approach achieves significantly improved registration accuracy. Extensive experiments demonstrate that our method not only enhances the performance of existing point cloud registration techniques but also exhibits robust generalization capabilities across diverse datasets. Codes are available at https://github.com/zhengcy-lambo/RARE.git. + +# 1. Introduction + +Point cloud registration is a fundamental yet challenging task in the field of 3D computer vision, with broad applications in localization [13], object detection [15], and scene reconstruction [29]. The objective is to compute a rigid transformation that aligns two partial point clouds captured from different viewpoints of the same scene or object. + +Point cloud registration typically comprises three key + +![](images/c92096c4783cb6c9baaf1d2bdee7f5a1ae4af70bc8b290a87c24e378c1bfb68a.jpg) + +![](images/52817172711d66492320b3ead9064a4d08e0ece4fbcadbfb7086402a1e6a7201.jpg) +Figure 2. Comparison of correspondence matching using projected depth maps from different viewing angles. In the top pair of images, the default viewing angle introduces challenges in obtaining accurate correspondences from depth maps due to the reliance on diffusion features. In contrast, in the bottom pair of images, aligning both point clouds to a consistent viewing angle enables the projected depth maps to produce more precise correspondences. Incorrect correspondences are indicated by red lines, while correct correspondences are marked by green lines. + +steps: feature extraction, correspondence establishment, and robust geometric transformation estimation. Feature extraction is a critical component, as distinctive and unique features enable the establishment of more accurate correspondences, thereby improving registration accuracy. Traditional methods [22, 34] often rely on hand-crafted features, which tend to exhibit limited performance. Recently, rapid advancements in 3D representations have driven the development of learning-based 3D features for point cloud registration. Learning-based methods [10, 17, 31, 39, 45, 46] extract 3D features using feature extractors trained on large datasets of paired point clouds. While these approaches demonstrate robust performance under challenging conditions, such as low overlap, they require substantial amounts of paired data with correspondences for training, limiting their practical applicability. However, recent advancements in pre-trained generative models offer promising alternatives that reduce the dependency on paired data. For instance, Stable Diffusion (SD) [32] exhibits remarkable capabilities in synthesizing high-quality images from textual input, highlighting its powerful internal representation of images. Diffusion-based methods [38, 49] leverage SD features to establish semantic correspondences across images. Building on this, FreeReg [40] extends the approach by establishing robust correspondences between images and point clouds using pre-trained diffusion models. + +Inspired by diffusion-based approaches, we explore the + +use of diffusion model features to establish correspondences between two point clouds. A straightforward method for converting point clouds into an image modality involves projecting the points onto an image plane to create a depth map, which can then be transformed into an image using depth-to-image diffusion models, such as ControlNet [50]. However, as highlighted by FreeReg, a single depth map can correspond to multiple plausible images, leading to significant discrepancies between the generated images and the original input. To address this, FreeReg leverages intermediate feature maps from the depth-to-image ControlNet to directly establish correspondences between depth maps derived from point clouds. Nevertheless, existing methods like FreeReg face significant challenges in generating accurate correspondences on depth maps with large parallax, as illustrated in Figure 2. To mitigate this issue, we propose using an initial transformation derived from existing point cloud registration methods to convert matched depth maps from large parallax to small parallax. This approach is based on our observation that diffusion features on depth maps with smaller parallax exhibit higher distinctiveness. By leveraging these diffusion features from small parallax depth maps, we refine existing methods to achieve more accurate rigid transformations. + +In this paper, we propose a zero-shot approach to refine existing point cloud registration methods by leveraging correspondences derived from depth images, enhancing point feature representations without requiring a training dataset. First, we project the point cloud into depth maps using the initial transformation obtained from existing point cloud registration methods and extract implicit knowledge as depth diffusion features using pre-trained diffusion models. These features are then integrated with the 3D features extracted from existing methods to establish more accurate correspondences between point clouds. To evaluate our approach, we conduct experiments on three datasets: the scene-scale indoor datasets 3DMatch and 3DLoMatch [48], and the scene-scale outdoor dataset KITTI [14]. Our method not only improves the performance of existing point cloud registration techniques but also demonstrates robust generalization capabilities across diverse datasets. To summarize, our contributions are as follows: + +- We propose a zero-shot approach to refine point cloud registration using depth information, eliminating the need for training on point cloud datasets. +- We leverage pre-trained diffusion models to extract deep features from depth maps and integrate them with 3D features derived from point clouds through a series of aggregation operations, enabling the establishment of more accurate correspondences. +- We conduct extensive experiments on three benchmarks such as 3DMatch, 3DLoMatch, and KITTI, and achieve competitive results. + +# 2. Related Work + +# 2.1. Point Cloud Registration + +Point cloud registration aims to estimate a rigid transformation that aligns two partially overlapping 3D point clouds and is widely utilized in graphics, vision, and robotics. The traditional approach focuses on hand-crafted descriptors that can discriminatively characterize the local geometry. Rusu et al. [33] use persistent Point Feature Histograms (PFH) to estimate a robust set of 16D features that describe the local pointwise geometry. They [34] refine PFH and propose Fast Point Feature Histograms (FPFH), retaining most of PFH's discriminative power while improving computational efficiency. However, the traditional registration methods typically require complicated optimization strategies, and may fail in complex scenes. + +Recently, many learning-based works [1, 17, 25, 27, 45] about 3D descriptors have been proposed. FCGF [10] computes the features in a single pass through a fully convolutional neural network without relying on keypoint detection. ImLoveNet [5] uses an intermediate misaligned image and multi-modal features to enhance low-overlap registration performance. GeoTransformer [31] embeds the geometric features of point clouds into a Transformer to achieve robust superpoint matching. PSReg [18] mitigates outlier matches caused by ambiguous features within overlapping regions. However, supervised methods are constrained by their dependence on large paired datasets, limiting real-world applicability. In contrast, our framework enhances registration accuracy by utilizing learned geometric priors, offering a more scalable solution without extensive labeled data. + +Compared to the supervised methods, unsupervised registration methods [30, 41, 42] attract less attention. BYOC [3] learns visual and geometric features from RGB-D video without relying on ground-truth pose or correspondence. UDPReg [28] leverages Gaussian Mixture Models to achieve self-supervised. EYOC [26] trains a feature extractor progressively, iterating from near to far to achieve self-supervision of distant point clouds. Yuan et al. [47] propose that self-supervised frameworks have unverified suitability for registration tasks. Buffer-X [36] utilizes a zero-shot registration pipeline to improve robustness in diverse scenes. However, these methods still require a substantial amount of data for self-supervised training. + +# 2.2. Outlier Removal Methods + +Although our method is fundamentally based on feature extraction, it extends beyond this by refining the registration results achieved by existing approaches, effectively serving as a refinement mechanism. This capability aligns closely with the objectives of outlier removal techniques, which aim to improve data integrity by eliminating inconsistencies. Therefore, we provide a brief introduction to out- + +lier removal methods. Traditional outlier filtering methods, such as RANSAC [13] and its variants [4, 23, 24], use repeated sampling and verification for outlier rejection. $\mathrm{SC^2}$ - $\mathrm{PCR}++$ [9] proposes a second-order spatial compatibility metric to compute the similarity between correspondences. VBReg [19] develops a novel variational non-local network-based outlier rejection framework for robust alignment. MAC [43] loosens the previous maximum clique constraint and mines more local consensus information in a graph for accurate pose hypotheses generation. Fast-MAC [51] enhances MAC by introducing graph signal processing into the domain of correspondence graph, achieving real-time speed while leading to little performance drop. Since higher-quality correspondences result in more accurate transformations, enhancing the power of the features becomes essential to achieve superior performance. + +# 2.3. Diffusion Feature + +Diffusion models have recently demonstrated remarkable performance in image generation tasks [11, 32, 35], capable of producing more diverse images and proven to be unaffected by mode collapse. Building on this foundation, diffusion models have been extensively studied and applied to other vision tasks, such as object detection [6] and image segmentation [7, 21, 37]. Recent studies [38, 49] leverage diffusion features for representation learning and semantic matching across RGB images, effectively capturing object variations both within and across categories. Additionally, FreeReg [40] explores the integration of diffusion features into image-to-point cloud registration tasks, demonstrating their effectiveness in learning cross-modal representations for image-to-point cloud alignment. Morreale et al. [30] propose NSM to distill semantic matches from pre-trained vision models. Wimmer et al. [42] first propose to back-project features from large pre-trained 2D vision models onto 3D shapes and employ them for this task. ZeroMatch [20] uses the powerful image representations of SD to enhance geometric descriptors for robust registration. Inspired by these works, we leverage diffusion features in point cloud registration tasks to optimize existing registration methods, thereby enhancing their performance. + +# 3. Method + +# 3.1. Overview + +Given a reference point cloud $P = \{p_i \in \mathbb{R}^3 | i = 1,2,\dots,M\}$ and a source point cloud $Q = \{q_i \in \mathbb{R}^3 | i = 1,2,\dots,N\}$ , point cloud registration aims to compute a rigid transformation $T = \{R,t\}$ that aligns the two partial point clouds, captured from different viewpoints of the same scene or object. Here, $R \in SO(3)$ represents a 3D rotation, and $t \in \mathbb{R}^3$ represents a 3D translation. Let $p^k$ and $q^k$ denote the key points in $P$ and $Q$ , respectively, while $F_p$ + +![](images/f4909829980318212323d161c5a7b3d97db6eb0b91c38f4b54868c0fc1162538.jpg) +Figure 3. Pipeline of our framework. Our framework begins by employing GeoTransformer [31] as a 3D feature extractor to derive an initial transformation for aligning the point clouds. The initial transformation is applied to project the point clouds into both the reference and source views, resulting in two pairs of depth images. Subsequently, a pre-trained diffusion model is utilized to extract depth diffusion features. Next, we use the Feature Transfer Module to establish common points between the depth diffusion features and the previously extracted 3D features, and the features of these common points are generated through a feature concatenation operation. To leverage both common and non-common features effectively, an inlier aggregation module is introduced. It integrates point matching, weighted SVD, and transformation selection to consolidate correspondences. Finally, this module identifies a set of high-confidence correspondences by analyzing common and non-common points correspondences, then computes the final rigid transformation. + +and $F_{q}$ are the corresponding pointwise features generated by feature descriptors. An initial transformation $T_{init}$ is estimated using these pointwise 3D features. Our proposed method refines the rigid transformation between $P$ and $Q$ by leveraging $F_{p}, F_{q},$ and $T_{init}$ . Our method is technically straightforward, as demonstrated in Figure 3. + +# 3.2. Stable Diffusion and ControlNet + +Preliminary. Our depth map diffusion features are built on the ControlNet framework, which leverages diffusion models—parameterized Markov chains trained via variational inference. These models operate through two Markov chain processes: a forward process and a reverse process. In the forward process, Gaussian noise $(\epsilon)$ is incrementally added to the input image over multiple steps, gradually transforming the image into an unstructured noise representation. At any time step $\ell$ , the image during the forward diffusion process is defined as: + +$$ +x _ {\ell} = \sqrt {\alpha_ {\ell}} x _ {0} + \sqrt {1 - \alpha_ {\ell}} \epsilon \tag {1} +$$ + +Here, $\alpha_{\ell}$ is a coefficient controlling the noise schedule, as defined in [16], and $\ell \in \mathrm{range}(S)$ . In the reverse process, + +the diffusion model employs a UNet to iteratively denoise the unstructured noise image, ultimately reconstructing the output image. The UNet is trained using the noisy RGB image $x_{\ell}$ and the corresponding noise added during the forward process as input data. The overall objective of the reverse process can be expressed as: + +$$ +L _ {D M} = \mathbb {E} _ {x, \epsilon \sim \mathcal {N} (0, 1), \ell} \left[ \| \epsilon - \epsilon_ {\theta} (x _ {\ell}, \ell) \| _ {2} ^ {2} \right] \tag {2} +$$ + +This loss function $L_{DM}$ minimizes the difference between the predicted noise $\epsilon_{\theta}(x_{\ell},\ell)$ and the actual noise $\epsilon$ , enabling the UNet to effectively reconstruct images during the reverse diffusion process. Building upon this foundation, Stable Diffusion incorporates an encoder $E$ and a decoder $D$ , derived from VQGAN [12], to enable seamless conversion between pixel space and latent space. + +Recent advancements like ControlNet extend this framework by introducing an additional encoder to process depth maps. The extracted depth features are then used to guide the reverse process of SD, enabling it to generate images that align coherently with the input depth map, starting from pure Gaussian noise. Inspired by these developments, we + +leverage ControlNet to extract diffusion features from depth maps, which are then used for matching purposes. + +Depth Diffusion Feature Extraction. Given the reference and source point clouds, we first apply the initial transformation $T_{init}$ , calculated using existing point cloud registration methods, to project the reference and source point clouds into depth maps from the perspective of the reference point cloud. Conversely, we apply the inverse of the initial transformation to project the reference and source point clouds into depth maps from the perspective of the source point cloud. This process yields two pairs of depth maps: one from the reference's perspective and the other from the source's perspective. For depth maps, as discussed in FreeReg, we apply traditional erosion and dilation operations to densify them before extracting diffusion features using ControlNet. These depth maps are then input into ControlNet, where an additional encoder encodes the depth maps into the latent space to guide the reverse process of Stable Diffusion. During Stable Diffusion reverse process, the pure Gaussian noise is progressively denoised through a specified number of sampling steps. Finally, as mentioned in [49], we aggregate feature maps from specific layers in the decoder of the U-Net to obtain deep diffusion features. For feature extraction, we selected layers [0, 3, 6] based on previous work [38, 40, 49] and verified their effectiveness through ablation experiments. However, simple concatenation of these feature maps often results in a high-dimensional fusion feature, which not only contains redundant information but also leads to excessive memory usage and computational overhead. To address this, we employ principal component analysis (PCA) to reduce the dimensionality of the feature maps for each layer. The reduced feature maps are then upsampled to the same resolution and combined to form the final depth diffusion features. + +# 3.3. Refine Using Deep Diffusion Features + +Although we have obtained diffusion features, it is challenging to directly integrate them with 3D features. Therefore, we design a series of module aggregating these two types of features to refine the initial transformation. + +Pixel-to-point Feature Transfer. The pixel-wise coordinates associated with diffusion-derived feature maps exhibit a geometric misalignment with 3D point-wise coordinate systems. This discrepancy stems from the inherent disparity of camera projection parameters and the stochastic sampling mechanisms of diffusion models. To bridge this cross-modal representation mismatch and aggregate the depth diffusion features with the 3D features, we design a Feature Transfer module. + +For point cloud $P$ , the Feature Transfer module takes three inputs: the 3D feature $F_{p}$ , the reference view diffusion feature $F_{p}^{ref}$ , and the source view diffusion feature $F_{p}^{src}$ , with their associated coordinates sets $\text{coor}_{p}$ + +$coor_{p}^{ref}$ , and $coor_{p}^{src}$ , respectively. This module utilizes a nearest-neighbor matching to identify common points between $coor_{p}$ and $coor_{p}^{ref}$ , which are points with Euclidean Distances below a predefined threshold and are generally considered the same point, denoted as $coor_{p}^{com1}$ . Similarly, common points between $coor_{p}$ and $coor_{p}^{src}$ are identified and denoted as $coor_{p}^{com2}$ . The same process is applied to point cloud $Q$ to obtain $coor_{q}^{com1}$ and $coor_{q}^{com2}$ . Then, to fuse the correspondences of different views robustly, we implement a two-stage feature aggregation strategy. The core idea involves independently normalizing the two features to align their scales and distributions, followed by concatenating them. The fused feature is calculated as: + +$$ +F u s e d F _ {\{p, q \}} ^ {\{r e f, s r c \}} = C o n c a t \left(\| F _ {\{p, q \}} \| _ {2}, \| F _ {\{p, q \}} ^ {\{r e f, s r c \}} \| _ {2}\right) \tag {3} +$$ + +where $\| \cdot \| _2$ represents L2 norm, $\{\cdot ,\cdot \}$ represents choosing between two options, and $Concat(\cdot ,\cdot)$ denotes concatenation along the feature dimension. + +Inlier Aggregation Module. While the Feature Transfer module enhances correspondence accuracy through aggregated 3D-diffusion features, two limitations persist: First, exclusive reliance on common points discards potentially valid correspondences; Second, in specific cases where the number of common points is limited, the results may display considerable variability. In extreme situations, such as when fewer than three common points are available, estimating the rigid transformation may become infeasible. To address this, we introduce an Inlier Aggregation Module that reweights the confidence scores for each correspondence and obtains a robust registration result. + +The module operates on five input feature sets: the fused common point features from the reference view $FusedF_{p}^{ref}$ and $FusedF_{q}^{ref}$ , the fused common point features from the source view $FusedF_{p}^{src}$ and $FusedF_{q}^{src}$ , the reference-view diffusion features $F_{p}^{ref}$ and $F_{q}^{ref}$ , the source-view diffusion features $F_{p}^{src}$ and $F_{q}^{src}$ , and the 3D features $F_{p}$ and $F_{q}$ . Then, it builds five correspondence sets via feature similarity matching: $\mathcal{C}_{common}^{ref}$ (reference view common features), $\mathcal{C}_{common}^{src}$ (source view common features), $\mathcal{C}_{diffu}^{ref}$ (reference view diffusion features), $\mathcal{C}_{diffu}^{src}$ (source view diffusion features), $\mathcal{C}_{3d}$ (3D features). Finally, the transformations $T_{i} = \{R_{i}, t_{i}\}$ are computed for the five sets of correspondences. It can be determined in closed form using weighted singular value decomposition (SVD): + +$$ +R _ {i}, t _ {i} = \min _ {R, t} \sum_ {\left(p _ {x _ {j}}, q _ {y _ {j}}\right) \in \mathcal {C} _ {i}} w _ {j} ^ {i} \| R \cdot p _ {x _ {j}} + t - q _ {y _ {j}} \| _ {2} ^ {2} \tag {4} +$$ + +where $\mathcal{C}_i$ belong to five correspondence sets, $w_{j}^{i}$ is confidence score for each correspondence. Then, we select the transformation that allows the most inlier matches: + +$$ +R, t = \max _ {R _ {i}, t _ {i}} \sum_ {\left(p _ {x _ {j}}, q _ {y _ {j}}\right) \in \mathcal {C}} \llbracket \left\| R _ {i} \cdot p _ {x _ {j}} + t _ {i} - q _ {y _ {j}} \right\| _ {2} ^ {2} < \tau_ {a} \rrbracket \tag {5} +$$ + +where $\llbracket \cdot \rrbracket$ is the Iverson bracket and $\tau_{a}$ is the acceptance radius. We then iteratively re-weight the confidence scores based on the surviving inlier matches, refining the transformation $R$ , $t$ for $Nr$ iterations using weighted SVD. Finally, the $K$ highest-scoring correspondences are used to estimate the final transformation. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We consider three datasets, i.e., the scene-scale indoor datasets 3DMatch & 3DLoMatch [48], and the scene-scale outdoor dataset KITTI [14]. 3DMatch is a collection of 62 scenes, from which we use 8 for testing. Official 3DMatch dataset considers only scan pairs with $>30\%$ overlap. For the 3DLoMatch, It considers only scan pairs with overlaps between 10 and $30\%$ from 3Dmatch. KITTI contains 11 sequences of LiDAR-scanned outdoor driving scenarios. We follow [10] and use 8-10 for testing. + +Evaluation Criteria. Following [17, 44], we evaluate the performance with three metrics: (1) Registration Recall (RR), i.e., the fraction of point cloud pairs whose transformation error is smaller than a certain threshold; (2) Relative Rotation Error (RE), i.e., geodesic distance between estimated and GT rotation matrices; (3) Relative Translation Error (TE), i.e., Euclidean distance between the estimated and GT translations. + +Implementation Details. While our approach and traditional outlier removal methods commonly employed in point cloud registration share an optimization-based framework, our method is fundamentally distinct in its underlying principles. Nonetheless, the input-output characteristics of the two approaches exhibit significant similarities. Therefore, we compare our method with the baseline outlier removal methods. To ensure a fair comparison, we utilize Geotransformer [31] as a initial method—a widely applied method in point cloud registration—to generate correspondences, 3D features, and initial transformations. Regarding the number of sampling points, we have struck a balance between efficiency and effectiveness by sampling 500 points for optimization on the 3DMatch dataset and setting 1000 points for the KITTI dataset. Since the input requirements of our method differ from those of outlier removal methods, for a fair comparison, we compute the initial transformation on the same sampling points. This typically implies that the initial transformation results may be inferior to those obtained by the prior methods themselves. + +Comparison methods. We evaluate both traditional and deep-learning-based approaches, including RANSAC [13], $\mathrm{SC^2}$ -PCR [8], PointDSC [2], $\mathrm{SC^2}$ -PCR++ [9], VBReg [19], MAC [43] and FastMAC [51]. + +Table 1. Registration results on 3DMatch dataset. + +
Method(†:Learning-based)RR↑RE↓TE↓
RANSAC-50K [13]91.42.240.071
RANSAC-1M [13]91.62.240.070
SC2-PCR [8]93.11.860.052
PointDSC† [2]92.31.750.051
VBReg† [19]94.11.870.052
SC2-PCR++ [9]91.21.840.052
MAC [43]93.81.600.052
FastMAC [51]93.81.860.065
Ours94.51.560.050
+ +Table 2. Registration results on 3DLoMatch dataset. + +
Method (†:Learning-based)RR↑RE↓TE↓
RANSAC-50K [13]65.93.620.101
RANSAC-1M [13]66.43.710.105
SC2-PCR [8]76.72.740.077
PointDSC† [2]69.82.600.075
VBReg† [19]77.72.680.075
SC2-PCR++ [9]71.82.690.077
MAC [43]76.42.570.075
FastMAC [51]75.42.920.085
Ours76.72.550.074
+ +# 4.2. Indoor Benchmarks: 3DMatch & 3DLoMatch + +The visual results presented in Figures 4 and 5, combined with the quantitative results summarized in Tables 1 and 2, support the following conclusions: (1) Both the geometric-only and deep-learned methods are outperformed by our method on the 3DMatch and 3DLoMatch datasets, demonstrating its ability to effectively optimize point cloud registration for indoor scenes. This underscores its proficiency in enhancing the registration capabilities of point cloud methods. (2) Even when compared to deep learning approaches, our method achieves superior performance without any data training, thereby proving its robust generalization capabilities. (3) In addition to the RR metric, we also attain the best results in terms of RE and TE metrics. It is worth noting that in Table 1 and 2, we use Geotransformer to generate the initial transformation and 3D features. For other refinement methods in the tables, we also use Geotransformer to generate their initial correspondences. + +Boosting other supervised methods. To further evaluate the effectiveness and generalization ability of our method, we selected several classic learning-based point cloud registration methods as baseline approaches for comparison. The considered methods include Predator [17], CoFiNet [46], and GeoTransformer [31]. + +As shown in Table 3, we have achieved significant improvements in registration recall rate, rotation error, and + +![](images/b3474bc94573d0880c471700c40fdb70e3fae16eeafb6efcbb1433e97ce45147.jpg) + +![](images/ffd4840e82982a6fc02d249c43d1bdfd524d8ecb9cc4dbb1313b9d782571b3b6.jpg) + +![](images/f3b12fa91b62015a518ab37f77a7bcad516bfd472c67e65c28b600c986bbb1a0.jpg) + +![](images/70483c8205f41a7d735dc2680ae3014ded4e11053c201e3bc4d15430d63ab8a2.jpg) + +![](images/00274e6fcc7269c56032a37e78b8001abfb2d63f5c1482d79d836a74c75f3b48.jpg) + +![](images/2bbb2b03f63fd91e047608a11f5fc2fb6ef6ed15da7d0b5467a6e1c8295ea669.jpg) + +![](images/86101847361dcb0d7641165ef5eea4f3e74c6d97d3f19a3ef5c44fc899f89fda.jpg) +Figure 4. Registration visualization results on 3DMatch. + +![](images/b3cff328ab2fa9d293560a16153ba197f6a29ffc2ff029b62bbf8ddc73b944c4.jpg) + +![](images/adb9a3e80fa65ba8a0299bb67abfd7184ccdfb83f8b904581e9fb40d033af0f7.jpg) + +![](images/02d0accb613ddda29909c977bf814f6e3ebd6d8a086efc1afe302f6442e51e48.jpg) + +translational error across all tested methods on the 3DMatch dataset, demonstrating robustness and generalization capabilities. Notably, both Predator and CoFiNet outperformed GeoTransformer after being optimized by our method. Building upon GeoTransformer, we attain a registration recall rate of $94.5\%$ on the 3DMatch dataset. Furthermore, we achieve further enhancements in the rotation error and translational error metrics, demonstrating the ability to refine samples with accurate registrations. + +Table 3. Performance boosting for deep-learned methods when combined with our method. + +
3DMatch
RR↑RE↓TE↓
FCGF [10]67.92.200.078
Predator [17]89.02.020.064
CoFiNet [46]89.32.440.067
GeoTransformer [31]92.01.720.062
FCGF+Ours76.22.070.055
8.3↑0.13↓0.023↓
Predator+Ours92.71.630.051
4.7↑0.39↓0.013↓
CoFiNet+Ours93.01.600.051
3.7↑0.84↓0.016↓
GeoTransformer+Ours94.51.560.050
2.5↑0.16↓0.012↓
+ +# 4.3. Outdoor Benchmark: KITTI odometry + +We have also tested our method with other SOTA methods on the KITTI odometry dataset. As shown in Table 4, for the three metrics of RR, RE, and TE, the performance of certain outlier removal methods has declined to some extent due to GeoTransformer already approaching their performance limits. Although our method is also unable to further enhance the registration recall rate, it can still improve the rotation error and translational error metrics without compro + +Table 4. Registration results on KITTI dataset. + +
Method (†:Learning-based)RR↑RE ↓TE ↓
Geotransformer [31]99.80.240.068
RANSAC-50K [13]99.80.420.069
SC2-PCR [8]99.80.570.105
PointDSC† [2]99.30.870.176
VBReg† [19]99.80.570.115
MAC [43]99.60.390.128
FastMAC [51]99.50.380.120
Ours99.80.240.064
+ +mising the effectiveness of the original methods. Consistent with Tables 1 and 2, Table 4 also uses Geotransformer as the initialization method. + +# 4.4. Ablation Experiments + +We conduct ablation studies and analytical experiments on the 3DMatch dataset. We independently assess the contribution of each methodological component introduced in Section 3, with results detailed in Tables 5 and 6. More ablation studies are provided in the supplementary materials. + +Diffusion layer selection. In table 5, we report the size of output feature maps of 13 layers in the UNet of Stable Diffusion. The feature map size is divided into four levels, i.e. small group $(8 \times 11$ , layers 0-2), medium group $(16 \times 22$ , layers 3-5), large group $(32 \times 44$ , layers 6-8), and extra-large group $(64 \times 88$ , layers 9-12). To better assess the contribution of each layer to the depth map matching process, we conduct an experiment where Principal Component Analysis (PCA) was excluded, relying solely on diffusion features to estimate correspondences. The transformation errors obtained from the derived correspondences are summarized in the table. The experimental results demonstrate that integrating features from layers 0, 3, and 6 leads to a significant improvement in registration performance. Conversely, incorporating features from layer 9 adversely + +![](images/09ba753fe6a68f45af181e27a8a5021bbe1532c2f1356013e3a9e8884d378a81.jpg) +Figure 5. Registration visualization results on 3DLoMatch. + +affects the registration process, indicating a potential limitation of extra-large layer features in this context. + +Table 5. Layer selection in diffusion feature extraction. + +
LayerFeature Map3DMatch
RR ↑RE ↓TE ↓
08 × 1154.219.021.772
316 × 2270.716.450.554
632 × 4467.717.000.505
964 × 8828.240.841.128
[0,3]16 × 2269.815.600.582
[0,6]32 × 4474.616.030.476
[0,9]64 × 8853.426.360.735
[0,3,6]32 × 4476.014.670.442
[0,3,9]64 × 8860.921.900.624
[0,3,6,9]64 × 8863.320.330.600
+ +The effectiveness of components. We conduct an ablation study to evaluate the contribution of each component in our framework, as summarized in Table 6. The PCA module exhibited minimal impact on overall performance, which is consistent with its primary purpose of reducing the dimensionality of diffusion features to alleviate computational overhead. In contrast, the Inlier Aggregation module demonstrated the most significant contribution to the method's effectiveness. While diffusion features are successful in generating a large number of correct correspondences, they are prone to introducing outliers. The Inlier Aggregation module addresses this issue by refining these correspondences, ensuring greater accuracy. Additionally, the feature transfer and feature fusion operations + +were found to be crucial for the overall performance, underscoring their importance in enhancing the method's robustness and efficacy. + +Table 6. Ablation of Components. + +
3DMatch
RR ↑RE ↓TE ↓
Geotransformer [31]92.01.720.062
w/o PCA94.11.560.052
w/o Feature Transfer93.11.650.051
w/o Inlier Aggregation92.81.690.060
Ours94.51.560.050
+ +# 5. Conclusion + +In this paper, we introduce a novel zero-shot approach to enhance point cloud registration by incorporating diffusion-based features extracted from depth images. Inspired by recent advances in leveraging pre-trained diffusion models for semantic correspondence in images, our method refines traditional point cloud registration techniques by combining depth diffusion features with geometric features. This fusion improves the accuracy of correspondences between misaligned point clouds without requiring a training dataset. Extensive experiments on multiple datasets demonstrate that our approach significantly enhances the performance of existing registration methods and exhibits robust generalization capabilities across various scenarios. This work opens up new possibilities for more effective and efficient point cloud registration, particularly in real-world applications where labeled data may be scarce. + +# Acknowledgement + +This work was supported by the National Natural Science Foundation of China (No. T2322012, No. 62172218), the Shenzhen Science and Technology Program (No. JCYJ20220818103401003, No. JCYJ20220530172403007), and the Shenzhen Longhua Science and Technology Innovation Special Funding Project (Industrial Sci-Tech Innovation Center of Low-Altitude Intelligent Networking). + +# References + +[1] Sheng Ao, Qingyong Hu, Hanyun Wang, Kai Xu, and Yulan Guo. BUFFER: balancing accuracy, efficiency, and generalizability in point cloud registration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1255-1264, 2023. 3 +[2] Xuyang Bai, Zixin Luo, Lei Zhou, Hongkai Chen, Lei Li, Zeyu Hu, Hongbo Fu, and Chiew-Lan Tai. Pointdsc: Robust point cloud registration using deep spatial consistency. In IEEE Conference on Computer Vision and Pattern Recognition, pages 15859-15869, 2021. 6, 7 +[3] Mohamed El Banani and Justin Johnson. Bootstrap your own correspondences. In IEEE/CVF International Conference on Computer Vision, pages 6413-6422, 2021. 3 +[4] Daniel Barath and Jiri Matas. Graph-cut RANSAC. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6733-6741, 2018. 3 +[5] Honghua Chen, Zeyong Wei, Yabin Xu, Mingqiang Wei, and Jun Wang. Imlovenet: Misaligned image-supported registration network for low-overlap point cloud pairs. In ACM SIGGRAPH, pages 1-9, 2022. 3 +[6] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusional: Diffusion model for object detection. In IEEE/CVF International Conference on Computer Vision, pages 19773-19786, 2023. 3 +[7] Ting Chen, Lala Li, Saurabh Saxena, Geoffrey E. Hinton, and David J. Fleet. A generalist framework for panoptic segmentation of images and videos. In IEEE/CVF International Conference on Computer Vision, pages 909-919, 2023. 3 +[8] Zhi Chen, Kun Sun, Fan Yang, and Wenbing Tao. $\mathrm{Sc}^2$ -pcr: A second order spatial compatibility for efficient and robust point cloud registration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13211-13221, 2022. 6, 7 +[9] Zhi Chen, Kun Sun, Fan Yang, Lin Guo, and Wenbing Tao. $\mathrm{Sc}^2$ -pcr++: Rethinking the generation and selection for efficient and robust point cloud registration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 12358-12376, 2023. 3, 6 +[10] Christopher B. Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In IEEE/CVF International Conference on Computer Vision, pages 8957-8965, 2019. 2, 3, 6, 7 +[11] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794, 2021. 3 + +[12] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021. 4 +[13] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 1, 3, 6, 7 +[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 2, 6 +[15] Yulan Guo, Mohammed Bennamoun, Ferdous Ahmed Sohel, Min Lu, and Jianwei Wan. 3d object recognition in cluttered scenes with local surface features: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11): 2270-2287, 2014. 1 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, 2020. 4 +[17] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4267-4276, 2021. 2, 3, 6, 7 +[18] Xiaoshui Huang, Zhou Huang, Yifan Zuo, Yongshun Gong, Chengdong Zhang, Deyang Liu, and Yuming Fang. Psreg: Prior-guided sparse mixture of experts for point cloud registration. In AAAI, pages 3788-3796, 2025. 3 +[19] Haobo Jiang, Zheng Dang, Zhen Wei, Jin Xie, Jian Yang, and Mathieu Salzmann. Robust outlier rejection for 3d registration with variational bayes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1148-1157, 2023. 3, 6, 7 +[20] Haobo Jiang, Jin Xie, Jian Yang, Liang Yu, and Jianmin Zheng. Zero-shot rgb-d point cloud registration with pretrained large vision model. In IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 16943-16952, 2025. 3 +[21] Peng Jiang, Fanglin Gu, Yunhai Wang, Changhe Tu, and Baoquan Chen. Difnet: Semantic segmentation by diffusion networks. In Advances in Neural Information Processing Systems, pages 1637-1646, 2018. 3 +[22] Andrew E. Johnson and Martial Hebert. Using spin images for efficient object recognition in cluttered 3d scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(5):433-449, 1999. 2 +[23] Huu M. Le, Thanh-Toan Do, Tuan Hoang, and Ngai-Man Cheung. SDRSAC: semidefinite-based randomized approach for robust point cloud registration without correspondences. In IEEE Conference on Computer Vision and Pattern Recognition, pages 124-133, 2019. 3 +[24] Jiayuan Li, Qingwu Hu, and Mingyao Ai. Gesac: Robust graph enhanced sample consensus for point cloud registration. ISPRS Journal of Photogrammetry and Remote Sensing, 167:363-374, 2020. 3 +[25] Quan Liu, Hongzi Zhu, Yunsong Zhou, Hongyang Li, Shan Chang, and Minyi Guo. Density-invariant features for distant + +point cloud registration. In IEEE/CVF International Conference on Computer Vision, pages 18169-18179, 2023. 3 +[26] Quan Liu, Hongzi Zhu, Zhenxi Wang, Yunsong Zhou, Shan Chang, and Minyi Guo. Extend your own correspondences: Unsupervised distant point cloud registration by progressive distance extension. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20816-20826, 2024. 3 +[27] Guofeng Mei, Fabio Poiesi, Cristiano Saltori, Jian Zhang, Elisa Ricci, and Nicu Sebe. Overlap-guided gaussian mixture models for point cloud registration. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4500-4509, 2023. 3 +[28] Guofeng Mei, Hao Tang, Xiaoshui Huang, Weijie Wang, Juan Liu, Jian Zhang, Luc Van Gool, and Qiang Wu. Unsupervised deep probabilistic approach for partial point cloud registration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13611-13620, 2023. 3 +[29] Ajmal S. Mian, Mohammed Bennamoun, and Robyn A. Owens. Automatic correspondence for 3d modeling: an extensive review. Int. J. Shape Model., 11(2):253-291, 2005. 1 +[30] Luca Morreale, Noam Aigerman, Vladimir G. Kim, and Niloy J. Mitra. Neural semantic surface maps. Computer Graphics Forum, 43(2):i-iii, 2024. 3 +[31] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11133-11142, 2022. 1, 2, 3, 4, 6, 7, 8 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10674-10685, 2022. 2, 3 +[33] Radu Bogdan Rusu, Nico Blodow, Zoltan Csaba Marton, and Michael Beetz. Aligning point cloud views using persistent feature histograms. In International Conference on Intelligent Robots and Systems, pages 3384-3391, 2008. 3 +[34] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (FPFH) for 3d registration. In IEEE International Conference on Robotics and Automation, pages 3212-3217, 2009. 2, 3 +[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In SIGGRAPH, pages 15:1-15:10, 2022. 3 +[36] Minkyun Seo, Hyungtae Lim, Kanghee Lee, Luca Carlone, and Jaesik Park. Buffer-x: Towards zero-shot point cloud registration in diverse scenes. arXiv preprint arXiv:2503.07940, 2025. 3 +[37] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. In Advances in Neural Information Processing Systems, pages 8702-8716, 2022. 3 +[38] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. In Advances in Neural Information Processing Systems, 2023. 2, 3, 5 + +[39] Haiping Wang, Yuan Liu, Zhen Dong, and Wenping Wang. You only hypothesize once: Point cloud registration with rotation-equivariant descriptors. In ACM International Conference on Multimedia, pages 1630-1641, 2022. 2 +[40] 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. In International Conference on Learning Representations, 2024. 2, 3, 5 +[41] Weijie Wang, Guofeng Mei, Bin Ren, Xiaoshui Huang, Fabio Poiesi, Luc Van Gool, Nicu Sebe, and Bruno Lepri. Zero-shot point cloud registration. arXiv preprint arXiv:2312.03032, 2023. 3 +[42] Thomas Wimmer, Peter Wonka, and Maks Ovsjanikov. Back to 3d: Few-shot 3d keypoint detection with back-projected 2d features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4154-4164, 2024. 3 +[43] Jiaqi Yang, Xiyu Zhang, Peng Wang, Yulan Guo, Kun Sun, Qiao Wu, Shikun Zhang, and Yanning Zhang. MAC: maximal cliques for 3d registration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10645-10662, 2024. 3, 6, 7 +[44] Zi Jian Yew and Gim Hee Lee. RPM-net: Robust point matching using learned features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11821-11830, 2020. 6 +[45] Zi Jian Yew and Gim Hee Lee. REGTR: end-to-end point cloud correspondences with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6667-6676, 2022. 2, 3 +[46] Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobodan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. In Advances in Neural Information Processing Systems, pages 23872-23884, 2021. 2, 6, 7 +[47] Mingzhi Yuan, Qiao Huang, Ao Shen, Xiaoshui Huang, and Manning Wang. Exploring self-supervised learning for 3d point cloud registration. IEEE Robotics and Automation Letters, 10(1):25-31, 2025. 3 +[48] Andy Zeng, Shuran Song, Matthias Nießner, Matthew Fisher, Jianxiong Xiao, and Thomas A. Funkhouser. 3dmatch: Learning local geometric descriptors from RGB-D reconstructions. In IEEE Conference on Computer Vision and Pattern Recognition, pages 199-208, 2017. 2, 6 +[49] 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. In Advances in Neural Information Processing Systems, 2023. 2, 3, 5 +[50] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision, pages 3813-3824, 2023. 2 +[51] Yifei Zhang, Hao Zhao, Hongyang Li, and Siheng Chen. Fastmac: Stochastic spectral sampling of correspondence graph. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17857-17867, 2024. 3, 6, 7 \ No newline at end of file diff --git a/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/images.zip b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c363dca1f7db429e8bb61f3edce55c657ffd99d7 --- /dev/null +++ b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84abcc30a2eadfa61462d5cc7e7a6b06f5a413d5aa042be640d0c3e97db386fa +size 727530 diff --git a/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/layout.json b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fcd72e410a3ea367087a049beec22dd18d84eae0 --- /dev/null +++ b/ICCV/2025/RARE_ Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b60551e7bb9ac3fe290024b9762ee695b96e1bc256b7293d246b3c378fab7885 +size 397637 diff --git a/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/c3344904-7a3d-450c-9b49-0bfc0eef5237_content_list.json b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/c3344904-7a3d-450c-9b49-0bfc0eef5237_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0261b5300f2c6bee9f59315d760a5c8ac4a268b1 --- /dev/null +++ b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/c3344904-7a3d-450c-9b49-0bfc0eef5237_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8b9aa8ff33c8c96759d4aad07e7e7938c8ae01bd5bcbf733397b8e03dde956a +size 81353 diff --git a/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/c3344904-7a3d-450c-9b49-0bfc0eef5237_model.json b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/c3344904-7a3d-450c-9b49-0bfc0eef5237_model.json new file mode 100644 index 0000000000000000000000000000000000000000..35bb26f0834c236081e41f7666608d3f60298e2f --- /dev/null +++ b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/c3344904-7a3d-450c-9b49-0bfc0eef5237_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26a7b009e494f2c19c8d9f26bc68c55f540e47a25cc893ac54311013ee8fa05b +size 98807 diff --git a/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/c3344904-7a3d-450c-9b49-0bfc0eef5237_origin.pdf b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/c3344904-7a3d-450c-9b49-0bfc0eef5237_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d5dc77f32a8257bfd8e0725bf291929c7c14b146 --- /dev/null +++ b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/c3344904-7a3d-450c-9b49-0bfc0eef5237_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8251057b322bf62fd3af093c7fbbfc603bacb2b5e526aa68596adda345c500ea +size 12335443 diff --git a/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/full.md b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c8489e5dbc3a6458932c04e8e81b10af6bea58c9 --- /dev/null +++ b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/full.md @@ -0,0 +1,319 @@ +# RCTDistill: Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion + +Geonho Bang $^{1*}$ Daye Oh $^{3}$ + +Minjae Seong $^{2*}$ +Junhyung Kim $^{3}$ + +Jisong Kim $^{2*}$ +Junho Koh $^{3}$ + +Geunju Baek1 +Jun Won Choi1† + +$^{1}$ Seoul National University $^{2}$ Hanyang University $^{3}$ Hyundai Motor Company $\{\text{ghbang}, \text{gjbaek}\} @ \text{spa.snu.ac.kr} \quad \{\text{mjseong}, \text{jskim}\} @ \text{spa.hanyang.ac.kr} \quad \{\text{daye.oh}, \text{univjun}, \text{junhkoh}\} @ \text{hyundai.com} \quad \text{junwchoi@snu.ac.kr}$ + +# Abstract + +Radar-camera fusion methods have emerged as a cost-effective approach for 3D object detection but still lag behind LiDAR-based methods in performance. Recent works have focused on employing temporal fusion and Knowledge Distillation (KD) strategies to overcome these limitations. However, existing approaches have not sufficiently accounted for uncertainties arising from object motion or sensor-specific errors inherent in radar and camera modalities. In this work, we propose RCTDistill, a novel cross-modal KD method based on temporal fusion, comprising three key modules: Range-Azimuth Knowledge Distillation (RAKD), Temporal Knowledge Distillation (TKD), and Region-Decoupled Knowledge Distillation (RDKD). RAKD is designed to consider the inherent errors in the range and azimuth directions, enabling effective knowledge transfer from LiDAR features to refine inaccurate BEV representations. TKD mitigates temporal misalignment caused by dynamic objects by aligning historical radar-camera BEV features with current LiDAR representations. RDKD enhances feature discrimination by distilling relational knowledge from the teacher model, allowing the student to differentiate foreground and background features. RCTDistill achieves state-of-the-art radar-camera fusion performance on both the nuScenes and View-of-Delft (VoD) datasets, with the fastest inference speed of 26.2 FPS. + +# 1. Introduction + +Radar-camera-based 3D detection serves as a crucial component in autonomous driving systems. While various Radar-Camera fusion approaches exist, recent methods predominantly focus on feature-level fusion, with particular emphasis + +![](images/f4f4546322998f54cedf1aab43a3ba44ab5dab1d7e2a44918e3aaa54ddf358fa.jpg) +Figure 1. RCTDistill enhances the quality of BEV features for 3D object detection by aligning spatial features in the range-azimuth direction and temporal features in dynamic object trajectories. + +on Bird's Eye View (BEV) domain [13, 15, 22]. Although these methods effectively address the inherent limitations of individual sensors, there remains a notable performance gap when compared to LiDAR-based 3D detectors, as illustrated in Figure 1, where radar-camera fusion models show spatially and temporally misaligned BEV representations. + +In an effort to overcome these performance gaps, Knowledge Distillation (KD) [16, 40] and temporal fusion approaches [14, 15, 20] have recently emerged as promising solutions in 3D object detection. KD-based approaches employ detectors based on low-fidelity sensors (e.g., cameras, radar, or camera-radar fusion) as student models while lever + +aging detectors based on higher-fidelity sensor inputs (e.g., LiDAR or LiDAR-camera fusion) as teacher models to enhance the student model's performance. In parallel, temporal fusion methods aim to improve performance by incorporating additional information from past frames and merging it with current features to compensate for information that may be missed at a single time point. + +However, several significant challenges remain in these approaches. First, many existing KD methods do not fully account for sensor-specific characteristics [16, 40], suggesting that explicitly incorporating these characteristics could further improve KD effectiveness. Cameras suffer from depth ambiguity, introducing uncertainties in object distance estimation during BEV feature generation [26, 38], whereas radar provides reliable depth but has low angular resolution, leading to imprecise object localization in the horizontal plane [33, 43]. While these sensors appear complementary in nature, integrating both modalities remains challenging, especially when it comes to mitigating their inherent uncertainties. Second, despite recent advances [14], addressing error propagation caused by object movements remains a challenge for temporal fusion models. The independent motion of dynamic objects leads to temporal misalignment, requiring explicit motion estimation to align features across frames. This process increases latency, ultimately degrading real-time detection performance. + +In this paper, we propose RCTDistill, a novel cross-modal knowledge distillation method that transfers knowledge from a LiDAR model to a temporal radar-camera fusion model. RCTDistill performs knowledge distillation in three directions: Range-Azimuth Knowledge Distillation (RAKD), Temporal Knowledge Distillation (TKD), and Region-Decoupled Knowledge Distillation (RDKD). RAKD mitigates feature uncertainties in the fused BEV representation using an elliptical Gaussian mask, which accounts for varying uncertainties along the range and azimuth directions. By focusing on elliptical regions with Gaussian mask intensities above a threshold, it performs targeted feature distillation to effectively transfer knowledge from the teacher model. TKD addresses temporal feature misalignment in dynamic object detection by employing HA-Net to align historical BEV features. It generates trajectory-aware Gaussian regions for distillation, effectively capturing temporal dynamics while suppressing misaligned features. RDKD leverages an affinity map to distill relational knowledge between features, enabling the student model to learn discriminative feature relationships between foreground and background regions from the teacher model. RCTDistill achieves state-of-the-art performance and latency on both the nuScenes [3] and View-of-Delft [27] radar-camera 3D detection benchmarks, outperforming all previous methods. + +The main contributions of this paper are as follows: + +- Our RCTDistill is the first method to introduce a knowl + +edge distillation technique specifically designed for radar-camera temporal fusion. Previous methods [16, 40] focused solely on a single-frame setup, without explicitly utilizing temporal information. In this paper, we propose a novel cross-modal knowledge distillation approach that aligns BEV features over time, enhancing 3D object detection performance. + +- We propose three knowledge distillation techniques, RAKD, TKD, and RDKD designed to address misalignment issues in radar-camera temporal 3D object detection. RAKD incorporates sensor-specific characteristics into the distillation process, while TKD focuses on temporal fusion by compensating for dynamic object motion. RDKD enhances the model's discriminative power by distilling the relationships between foreground and background features from the teacher model. +- RCTDistill achieves improvements of a $4.7\%$ mean Average Precision (mAP) and $4.9\%$ NuScenes Detection Score (NDS) over the student model. Furthermore, it outperforms existing state-of-the-art radar-camera fusion methods on both the nuScenes and VoD datasets, while maintaining a real-time inference speed of 26.2 FPS. + +# 2. Related Work + +# 2.1. Radar-Camera 3D Object Detection + +Recent radar-camera fusion models have adopted view transformation-based methods, which can be categorized into two directions: implicitly mapping features through attention mechanisms [21, 23, 31], or explicitly using depth distributions obtained from camera features [19, 30]. RCM-Fusion [13], following the former approach, integrated radar and camera features into a unified bird's-eye-view (BEV) space using deformable attention. Similarly, SpaRC [33] enhances the positional information of sparse object queries and 2D camera features through an attention mechanism based on radar point features. Conversely, CRN [15] and CRT-Fusion [14] adopted the latter strategy to enhance camera BEV features via radar-guided view transformation before feature-level fusion. RCBEVDet [22] introduced a dualpath radar backbone combining point-based and transformer-based modules, leveraging deformable cross-attention for effective radar-camera feature fusion. To incorporate temporal modeling, some works [14, 15, 22, 33] have extended their methods by integrating multi-frame radar and camera features. Notably, CRT-Fusion [14] introduces dynamic object modeling to mitigate temporal misalignment. + +# 2.2. Knowledge Distillation in 3D Object Detection + +Cross-modal KD approaches have been widely adopted to facilitate knowledge transfer across heterogeneous sensor modalities to overcome sensing limitations [4, 5, 8, 10, 16, 18, 32, 41, 42]. BEVDistill [4] proposed feature- and + +![](images/cf2538a3c8e3c16a78cb9a38e5d8ffd73e43583d461130c0a7e7ac8eb58558b1.jpg) +Figure 2. Overall architecture of RCTDistill. The student model fuses radar and camera features using a gating module to generate low-level BEV features, which are stored in a memory bank. These historical features are then temporally aggregated with the current BEV features, followed by a decoder that produces high-level BEV representations. RAKD and TKD enhance the low-level features by accounting for sensor uncertainty and object motion, respectively, while RDKD learns discriminative feature relationships between foreground and background regions from the teacher detector. Note that the teacher model is utilized only during training and omitted during inference. + +instance-level distillation in the BEV space, and UniDistill [42] employed flexible KD pathways—such as LiDARTo-camera, camera-to-LiDAR, and fusion-to-single modality—by projecting both teacher and student features into a unified BEV representation. + +Growing interest in leveraging radar for 3D object detection has led to increased attention on cross-modal KD, where data sparsity remains a key challenge. RadarDistill [2] and CRKD [40] addressed this by transferring rich knowledge from LiDAR or LiDAR-camera models using feature- and relation-level supervision. LEROjD [28] and SCKD [35] further explored 4D radar-based KD; LEROjD employed a multi-stage training and distillation framework to leverage the spatial density of LiDAR, while SCKD adopted a semi-supervised approach using a LiDAR-camera teacher. + +In parallel, temporal KD approaches have emerged to address object misalignment across time, a critical issue in dynamic driving scenes. VCD [11] projected the historical positions of ground truth into the current frame and applied KD at these aligned positions. STXD [12] leveraged similarity maps to align past frames of the teacher model with both the current frame of the teacher and the student. + +# 3. Method + +The overall architecture of RCTDistill is illustrated in Figure 2. Our model leverages a cross-modal knowledge distillation framework to enhance the performance of radar-camera fusion models. In Section 3.1, we provide a comprehensive overview of our teacher and student model architectures. Sections 3.2-3.4 present the detailed descriptions of our proposed knowledge distillation methods, RAKD, TKD, and RDKD. + +# 3.1. Baseline Model + +Student model. We adopt BEVFusion-R as our student model, which extends the BEVFusion [25] framework for radar-camera fusion. The model extracts radar BEV features and camera features using separate backbones. These camera features are then transformed into BEV representation via a View Transformation (VT) module. Subsequently, the radar and camera BEV features are concatenated and fused through a $1 \times 1$ convolution, followed by 2D CNN blocks and a detection head to produce the final predictions. + +We improve the performance of the baseline by incorporating several modifications to BEVFusion-R. First, for multi-modal feature fusion, we replace $1 \times 1$ convolution layer with an adaptive gating network [40] that adaptively adjusts the influence of each modality. This gating network outputs low-level fused BEV features $B_{\mathrm{low}}^{\mathrm{RC}} \in \mathbb{R}^{D \times H \times W}$ . These BEV features are aggregated using a streaming-based temporal fusion, where a memory bank retains historical BEV features obtained from previous time steps. These historical features are transformed into the current coordinate and then concatenated channel-wise with the current BEV features. Finally, the temporally fused features pass through a decoder consisting of 2D CNN layers to produce high-level BEV features $B_{\mathrm{high}}^{\mathrm{RC}} \in \mathbb{R}^{D' \times H \times W}$ . + +Teacher model. We employed CenterPoint, a widely adopted LiDAR-based 3D object detector, as our teacher model. CenterPoint groups raw LiDAR point clouds into 3D voxels, then extracts low-level BEV features $B_{\mathrm{low}}^{L} \in \mathbb{R}^{C \times H \times W}$ through a 3D sparse convolutional backbone. These low-level features are further processed by a decoder consisting of 2D CNN layers to generate high-level BEV features, $B_{\mathrm{high}}^{L} \in \mathbb{R}^{C' \times H \times W}$ . + +![](images/6cdbb5a7d7f40da7c62849dcae554642c383dfe43d1a9dca49e09f996849b124.jpg) +Figure 3. Elliptical Gaussian Mask Regions for RAKD and TKD. Elliptical masks in the RAKD (center) and TKD (right) regions are used for knowledge distillation, targeting range-azimuth uncertainty and temporal misalignment, respectively. + +# 3.2. Range-Azimuth Knowledge Distillation + +Cross-modal knowledge distillation aims to transfer the desirable representations from the teacher modality to the student modality. Earlier methods [4, 39] have used ground truth (GT) box-centered Gaussian masks to determine the region for feature distillation, as shown in Figure 3 (left). However, these approaches fail to capture modality-specific characteristics, such as depth ambiguity in cameras and range-azimuth uncertainties in radar. Gaussian masks apply weights that decrease uniformly from the object center, which makes it difficult to adapt to the unique distribution of each modality. To address these challenges, we propose Range-Azimuth Knowledge Distillation (RAKD), which utilizes an elliptical Gaussian mask. The shape of the ellipse is determined by the inherent uncertainties along range and azimuth directions, as shown in Figure 3 (middle). + +For each $i$ -th GT 3D box projected to bird's eye view, we obtain a 2D box defined by center position $(p_x^i,p_y^i)$ , heading angle $\theta^i$ , and dimensions $(l^i,w^i)$ . Inspired by DORN [6], we determine the radius $(r_1^i,r_2^i)$ of an elliptical region in the major and minor axes based on the object size and distance from the ego-vehicle as + +$$ +r _ {1} ^ {i} = l ^ {i} \cdot \left(\frac {\alpha l}{l ^ {i}}\right) ^ {\beta}, \quad r _ {2} ^ {i} = w ^ {i} \cdot \left(\frac {\alpha_ {w}}{w ^ {i}}\right) ^ {\beta}, \tag {1} +$$ + +where $\alpha_{l}$ and $\alpha_{w}$ denote the hyperparameters that control the scale of $r_1^i$ and $r_2^i$ , and $\beta (< 1)$ denotes the normalized distance from the ego-vehicle to the center position of the object. Notice that $r_1^i$ and $r_2^i$ are proportional to the size $l^i$ and $w^i$ of the 2D box, respectively. Using these parameters, we generate an Elliptical Gaussian mask $E_{i,x,y}$ of the $i$ -th GT box at the position $(x,y)$ as + +$$ +E _ {i, x, y} = \exp \left(- \frac {1}{2} \left(\frac {\left(x ^ {\prime}\right) ^ {2}}{\left(r _ {1} ^ {i}\right) ^ {2}} + \frac {\left(y ^ {\prime}\right) ^ {2}}{\left(r _ {2} ^ {i}\right) ^ {2}}\right)\right), \tag {2} +$$ + +where + +$$ +\left[ \begin{array}{l} x ^ {\prime} \\ y ^ {\prime} \end{array} \right] = \left[ \begin{array}{c c} \cos \theta^ {i} & - \sin \theta^ {i} \\ \sin \theta^ {i} & \cos \theta^ {i} \end{array} \right] \left[ \begin{array}{l} x - p _ {x} ^ {i} \\ y - p _ {y} ^ {i} \end{array} \right]. \tag {3} +$$ + +We then merge the elliptical Gaussian masks $E \in \mathbb{R}^{N \times H \times W}$ , generated from the $N$ GT boxes, into $\bar{E} \in \mathbb{R}^{H \times W}$ . When elliptical masks overlap between objects, $\bar{E}_{x,y}$ is obtained by taking the maximum magnitude in those areas. Next, we obtain a Distillation Mask $W_{\mathrm{RA}} \in \mathbb{R}^{H \times W}$ by assigning the corresponding value from $\bar{E}_{x,y}$ to each spatial position $(x,y)$ if it surpasses the threshold $\tau$ , and zero otherwise. RAKD aims to align the low-level features of the student model with those of the teacher model within the Distillation Mask. To this end, we introduce a RAKD loss function + +$$ +L _ {\mathrm {R A}} = \frac {1}{\left| N _ {\mathrm {R A}} \right|} \sum_ {j = 1} ^ {H} \sum_ {k = 1} ^ {W} W _ {\mathrm {R A}, j, k} \left\| B _ {\text {l o w}, j, k} ^ {\mathrm {L}} - \bar {B} _ {\text {l o w}, j, k} ^ {\mathrm {R C}} \right\| _ {2}, \tag {4} +$$ + +where $N_{\mathrm{RA}}$ represents the number of non-zero elements in $W_{\mathrm{RA}}$ . The student feature map $\bar{B}_{\mathrm{low}}^{\mathrm{RC}}$ is obtained by applying a $1\times 1$ convolution to the original features $B_{\mathrm{low}}^{\mathrm{RC}}$ to match its channel dimension with that of $B_{\mathrm{low},i,j}^{\mathrm{L}}$ . + +# 3.3. Temporal Knowledge Distillation + +While previous temporal fusion methods [15, 22] in 3D object detection have achieved significant performance improvements through simple concatenation of historical BEV features followed by CNN layers, they often struggle to adequately capture the movement of dynamic objects. To address this issue, we propose Temporal Knowledge Distillation (TKD), a novel framework that effectively takes object motion information into account for knowledge distillation. + +As illustrated in Figure 2, we first concatenate historical BEV feature maps and process them through Historical + +![](images/ce4e0b20c972323eb65f34405b40aa371a77e6db52cf4bf5f6a9920239cf9844.jpg) +(a) RCTDistill w/ TKD + +![](images/c64af38bf4785ad1e874aba6acb94061993b6561d8330143ba15e0ed7ce73667.jpg) +(b) RCTDistill w/o TKD +Figure 4. Visualization of the BEV feature map. Green boxes denote static object Ground Truth (GT) boxes, Red boxes indicate dynamic object GT boxes. + +Alignment Network (HA-Net), a 2D CNN-based architecture. In the presence of rapid object motion, BEV features exhibit significant variations between consecutive frames. HA-Net aggregates these BEV features through large-scale kernel CNNs to alleviate inter-frame variations. The detailed architecture of HA-Net is provided in the Supplementary Material. + +We identify the spatial regions in temporal features for knowledge transfer. To achieve this, we introduce temporal elliptical Gaussian masks $T \in \mathbb{R}^{N \times H \times W}$ whose shapes are determined based on object trajectories as shown in Figure 3 (right). These masks are designed to transfer temporal representations within regions that encompass the trajectories of dynamic objects. For the $i$ -th object, we determine the elliptical Gaussian mask $T_{i}$ by revising the center position and major axis radius of the elliptical function $E_{i}$ in Equation 2. Given the velocity vector $\mathbf{v}^i = (v_x^i, v_y^i)$ and the center point $\mathbf{p}^i = (p_x^i, p_y^i)$ for the $i$ -th object, the ellipse center $\hat{\mathbf{p}}^i = (\hat{p}_x^i, \hat{p}_y^i)$ is computed as + +$$ +\hat {\mathbf {p}} ^ {i} = \left\{\mathbf {p} ^ {i} - \frac {t _ {s}}{2} \mathbf {v} ^ {i}, \quad \text {i f} \| \mathbf {v} ^ {i} \| _ {2} ^ {2} > \tau_ {v} \right. \tag {5} +$$ + +where $t_s$ is the temporal duration chosen to be less than the total duration of the historical frame. This implies that when the magnitude of the velocity vector exceeds a threshold $\tau_v$ , the center of the ellipse is obtained by shifting the object center in the direction opposite to its motion. Subsequently, the radius $\hat{r}_1^i$ and $\hat{r}_2^i$ in the major and minor axis is calculated as + +$$ +\hat {r} _ {1} ^ {i} = l ^ {i} + \sqrt {\left(\hat {p} _ {x} ^ {i} - p _ {x} ^ {i}\right) ^ {2} + \left(\hat {p} _ {y} ^ {i} - p _ {y} ^ {i}\right) ^ {2}} \tag {6} +$$ + +$$ +\hat {r} _ {2} ^ {i} = w ^ {i}, \tag {7} +$$ + +to allow the masks to cover a historical trajectory of the $i$ -th dynamic object. Using the computed $\hat{\mathbf{p}}^i$ , $\hat{r}_1^i$ and $\hat{r}_2^i$ , the temporal mask $T_{i,x,y}$ is given by + +$$ +T _ {i, x, y} = \exp \left(- \frac {1}{2} \left(\frac {\left(x ^ {\prime \prime}\right) ^ {2}}{\left(\hat {r} _ {1} ^ {i}\right) ^ {2}} + \frac {\left(y ^ {\prime \prime}\right) ^ {2}}{\left(\hat {r} _ {2} ^ {i}\right) ^ {2}}\right)\right), \tag {8} +$$ + +where + +$$ +\left[ \begin{array}{l} x ^ {\prime \prime} \\ y ^ {\prime \prime} \end{array} \right] = \left[ \begin{array}{c c} \cos \theta^ {i} & - \sin \theta^ {i} \\ \sin \theta^ {i} & \cos \theta^ {i} \end{array} \right] \left[ \begin{array}{l} x - \hat {p} _ {x} ^ {i} \\ y - \hat {p} _ {y} ^ {i} \end{array} \right]. \tag {9} +$$ + +When the elliptical masks overlap, we obtain $\bar{T} \in \mathbb{R}^{H \times W}$ by taking the maximum value of $T_{i,x,y}$ over $i$ . Similarly to RAKD, the Temporal Distillation Mask $W_{\mathrm{T}} \in \mathbb{R}^{H \times W}$ is computed by retaining $\bar{T}_{x,y}$ at position $(x,y)$ if it exceeds the threshold $\tau$ , or setting it to zero otherwise. To facilitate temporal knowledge transfer in these regions, we introduce a temporal distillation loss function + +$$ +L _ {\mathrm {T}} = \frac {1}{\left| N _ {\mathrm {T}} \right|} \sum_ {j = 1} ^ {H} \sum_ {k = 1} ^ {W} W _ {\mathrm {T}, j, k} \left\| B _ {\text {l o w}, j, k} ^ {L} - \hat {B} _ {\text {l o w}, j, k} ^ {R C} \right\| ^ {2}, \tag {10} +$$ + +where $N_{\mathrm{T}}$ denotes the number of non-zero elements in $W_{\mathrm{T}}$ , $\hat{B}_{\mathrm{low}}^{RC}$ represents the temporally aggregated feature maps obtained through the HA-Net, and $B_{\mathrm{low}}^{L}$ denotes the current low-level feature map from the teacher model. This distillation process aims to align the historical camera-radar features with the current LiDAR features within the regions defined by $W_{\mathrm{T}}$ . As shown in Figure 4, the BEV feature maps produced with our proposed TKD module exhibit clearer object boundaries and reduced temporal artifacts for dynamic objects. In contrast, the baseline displays noticeable trailing effects around moving objects. The enhanced temporal alignment achieved by TKD leads to more accurate localization of dynamic objects at the current timestamp. + +# 3.4. Region-Decoupled Knowledge Distillation + +High-level feature maps—especially those from LiDAR-based models—are effective at capturing rich semantic information that distinctly separates foreground objects from background regions. To transfer this discriminative capability to radar-camera models, we propose RDKD, which aligns the relational structure between foreground and background regions in the high-level feature maps. Guided by the teacher model, the student model is encouraged to maintain high similarity among foreground features while effectively distinguishing them from background features. + +The high-level feature map $B_{\mathrm{high}}^{\mathrm{RC}}$ obtained from the student detector is processed through the detection head to generate a classification score map $B_{\mathrm{cls}}^{\mathrm{RC}}$ . We obtain the confidence score map $\bar{B}_{\mathrm{cls}}^{RC} \in \mathbb{R}^{H \times W}$ by taking the maximum value of $B_{\mathrm{cls}}^{\mathrm{RC}}$ across all classes. From $\bar{B}_{\mathrm{cls}}^{\mathrm{RC}}$ , we identify all positions with confidence scores above a predefined threshold $\tau$ . At these $K$ selected positions, we extract the corresponding high-level features from $\bar{B}_{\mathrm{cls}}^{\mathrm{RC}}$ , resulting in a set of $K$ feature vectors, denoted as $\mathbf{f}_1^{\mathrm{RC}}, \mathbf{f}_2^{\mathrm{RC}}, \dots, \mathbf{f}_K^{\mathrm{RC}}$ . Using these extracted features, the $K \times K$ affinity map is computed as + +$$ +S _ {j k} ^ {\mathrm {R C}} = \frac {\mathbf {f} _ {j} ^ {\mathrm {R C}} \cdot \mathbf {f} _ {k} ^ {\mathrm {R C}}}{\| \mathbf {f} _ {j} ^ {\mathrm {R C}} \| _ {2} \| \mathbf {f} _ {k} ^ {\mathrm {R C}} \| _ {2}}, \quad \text {f o r} j, k = 1, 2, \dots , K, \tag {11} +$$ + +
DatasetMethodsInputKDBackboneImage SizeFramesNDSmAPmATEmASEmAOEmAVEmAAEFPS
Val.Teacher [37]L---167.160.30.2980.2510.2880.2810.190-
BEVDepth†[19]C-R50256×704247.535.10.6390.2670.4790.4280.19811.6
SOLOFusion†[29]C-R50256×70416+153.442.70.5670.2740.4110.2520.18811.4
X3KD†[16]C+R✓R50256×704153.842.30.4870.2770.5420.3440.197-
CRKD†[40]C+R✓R50256×704154.943.20.4500.2670.4420.3390.176-
RCTDistill-S†C+R✓R50256×704155.546.40.4880.2620.4620.3810.17628.0
CRN [15]C+R-R50256×704456.049.00.4870.2770.5420.3440.19720.4
RCBEVDet†[22]C+R-R50256×704256.845.30.4860.2850.4040.2200.19221.3
CRT-Fusion†[14]C+R-R50256×7046+159.750.80.4610.2640.4190.2340.18614.5
RCTTrans†[20]C+R-R50256×704458.650.90.5370.2690.4910.2030.18319.2
SpaRC [33]C+R-R50256×704862.054.50.4960.2690.4030.1770.18119.1
BaselineC+R-R50256×7048+157.350.50.4910.2710.5600.2800.18927.8
RCTDistillC+R✓R50256×7048+162.255.20.4340.2410.4320.2630.17426.2
Teacher [37]L---167.360.60.2860.2510.3150.2610.185-
BEVDepth†[19]C-R101512×1408253.541.20.5650.2660.3580.3310.1905.0
SOLOFusion†[29]C-R101512×140816+158.248.30.5030.2640.3810.2460.20711.4
CRN [15]C+R-R101512×1408459.252.50.4600.2730.4430.3520.1807.2
CRT-Fusion [14]C+R-R101512×14086+162.155.40.4250.2640.4330.2370.1934.9
SpaRC [33]C+R-R101512×1408864.457.10.4840.2640.3080.1750.1787.2
BaselineC+R-R101512×14088+160.954.20.4330.2720.4690.2520.1988.7
RCTDistillC+R✓R101512×14088+165.559.00.3780.2500.3460.2360.1868.4
TestBEVDepth [19]C-ConvX-B640×1600260.952.00.4450.2430.3520.3470.128-
SOLOFusion [29]C-ConvX-B640×160016+161.954.00.4530.2570.3760.2760.148-
CRN‡[15]C+R-ConvX-B640×1600462.457.50.4160.2640.4560.3650.130-
RCBEVDet [22]C+R-V2-99640×1600263.955.00.3900.2340.3620.2590.113-
CRTFusion [14]C+R-ConvX-B512×14086+164.958.30.3650.2610.4050.2620.1323.7
RCTTrans [20]C+R-V2-99640×1600464.757.80.4590.2450.3920.1980.121-
SpaRC [33]C+R-V2-99640×1600867.160.0------
RCTDistillC+R✓ConvX-B512×14088+167.460.30.3340.2450.3340.2510.1135.0
+ +where $\cdot$ denotes the inner product operation. From the same $K$ selected positions, we also extract $K$ high-level feature vectors from the high-level feature map $B_{\mathrm{high}}^{\mathrm{L}}$ obtained by the teacher model. Then, we compute the teacher affinity map $S^{\mathrm{L}}$ in the same manner as Equation 11. + +We define the RDKD loss $L_{RD}$ to guide the student model to learn the ability to discriminate between the foreground and background regions from the teacher model as + +$$ +L _ {\mathrm {R D}} = \frac {1}{| K ^ {2} |} \sum_ {j = 1} ^ {K} \sum_ {k = 1} ^ {K} \left| S _ {j k} ^ {L} - S _ {j k} ^ {R C} \right|. \tag {12} +$$ + +# 3.5. Loss Function + +The total loss function comprises a 3D detection loss $L_{det}$ and three KD loss terms. The total loss $L_{total}$ is obtained as + +$$ +L _ {\text {t o t a l}} = L _ {\text {d e t}} + \lambda_ {\mathrm {R A}} L _ {\mathrm {R A}} + \lambda_ {\mathrm {T}} L _ {\mathrm {T}} + \lambda_ {\mathrm {R D}} L _ {\mathrm {R D}}, \tag {13} +$$ + +where $\lambda_{\mathrm{RA}}$ , $\lambda_{\mathrm{T}}$ , and $\lambda_{\mathrm{RD}}$ are hyperparameters that control the relative importance of each loss term. + +Table 1. Performance comparisons with 3D object detectors on the nuScenes validation (Val.) and test set. Teacher denotes CenterPoint [37]. 'C', and 'R' represent camera and radar, respectively. RCTDistill-S indicates the use of single-frame input. †: trained with CBGS [44]. ‡: use Test Time Augmentation. + +
MethodsRAKDRDKDTKDNDS↑mAP↑
Baseline55.047.1
RCTDistill✓57.249.3
✓56.849.0
✓57.048.5
✓✓57.550.1
✓✓57.449.3
✓✓57.649.9
✓✓✓58.451.0
+ +Table 2. Ablation study for evaluating the main components. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets and metrics. We conducted the experiments on the nuScenes [3] and View-of-Delft (VoD) [27] datasets. + +The nuScenes dataset consists of 1,000 driving scenes, each scene is approximately 20 seconds long with 360-degree coverage from six cameras, five radars, and one LiDAR sensor. Keyframes are annotated at $2\mathrm{Hz}$ , covering 10 object classes. Our evaluation follows the official nuScenes benchmark metrics, including mean Average Precision (mAP) and nuScenes Detection Score (NDS). + +The VoD dataset comprises 8,693 frames, each capturing forward view using a single 4D radar, a stereo camera, and a LiDAR. Evaluation is performed using mAP measured over the Entire Annotated Area (EAA) and within the Driving Corridor (RoI) for three object classes (Car, Pedestrian, Cyclist). Since the VoD test server is not yet available, evaluation is conducted on the validation set. + +Implementation details. For the student model, we utilized ResNet [9] and ConvNeXt [24] as the image backbone networks and a modified Pillar Feature Network [14] building on PointPillars [17] as the radar backbone network. The teacher model employs the SECOND [36] backbone for processing LiDAR point clouds. Following the SOLOFusion [29], we incorporate a streaming-based parallel temporal fusion mechanism that leverages BEV features from the preceding eight frames. During the training, the teacher model weights are frozen while the student model is trained for 60 epochs. Detailed training configurations and hyperparameter settings are provided in the supplementary material. + +# 4.2. Comparison to the state-of-the-art + +Table 1 compares RCTDistill with existing camera-radar fusion and camera-only 3D object detectors on both the nuScenes validation and test sets. RCTDistill sets a new state-of-the-art performance across various backbone configurations, consistently surpassing previous methods. With ResNet-50, RCTDistill achieves $62.2\%$ NDS and $55.2\%$ mAP while maintaining the highest efficiency at 26.2 FPS. For the ResNet-101 backbone configuration, RCTDistill exhibits significant performance gains of $1.1\%$ in NDS and $1.9\%$ in mAP compared to SpaRC [33]. Moreover, on the test set, our method outperforms all existing radar-camera fusion models. Even when a single frame input is used (RCTDistill-S), the proposed method achieves superior results. Notably, RCTDistill achieves the highest inference speed (up to 28.0 FPS) across all evaluated configurations. + +# 4.3. Ablation Studies + +We conducted ablation studies on both the nuScenes and VoD validation sets. All experiments were conducted using a ResNet-50 image backbone and trained for 20 epochs. + +Components analysis. Table 2 presents an ablation study of the RCTDistill, focusing on the contributions of RAKD, RDKD, and TKD. Each component, when applied individually, improves performance by at least $1.4\%$ in mAP and $1.8\%$ in NDS. Pairwise combinations further boost results, + +
MethodsBEVDistill [4]CRKD [40]OursNDS↑mAP↑
Baseline55.047.1
RAKD✓56.448.3
✓56.748.6
✓57.249.3
+ +(a) Performance Comparison Based on Mask Types. + +
MethodsVCD [11]STXD [12]OursNDS↑mAP↑
Baseline55.047.1
TKD✓55.948.3
✓56.548.1
✓57.048.5
+ +(b) Performance Comparison Based on Temporal Knowledge Distillation Methods. + +
MethodsMonoDistill [5]CRKD [40]OursNDS↑mAP↑
Baseline55.047.1
RDKD✓56.047.8
✓55.948.1
✓56.849.0
+ +(c) Performance Comparison by Relation KD Application Scope. +Table 3. Ablation studies of proposed KD methods. + +
MethodsInputEAA AP↑RoI AP↑
BEVFusion [25]C+R49.2568.52
LXL [34]C+R56.3172.93
RCBEVDet [22]C+R49.9969.80
HGSFusion [7]C+R58.9679.46
SGDet3D [1]C+R59.7577.42
RCTDistillC+R62.3782.25
+ +Table 4. Performance Comparison on VoD [27] Dataset. RoI denotes the driving corridor area. EAA: Entire Annotated Area. + +while integrating all three components achieves the highest performance, reaching $58.4\%$ NDS and $51.0\%$ mAP. This demonstrates the effective synergy between spatial, temporal, and relational distillation techniques. + +Impact of KD region selection on model performance. Table 3 presents an ablation study evaluating the performance impact of applying KD to different regions for each module. In Table 3a, we compare our proposed elliptical Gaussian mask with previous methods that adopt different strategies for determining knowledge distillation regions. BEVDistill [4] employs a circular Gaussian distribution centered on the object. In contrast, CRKD [40] proposes MSFD, which expands the ground-truth bounding box region based on object velocity and range thresholds. We replaced RAKD with these methods in our RCTDistill framework for comparison. While the previous approaches outperform the baseline, our method achieves greater performance gains—specifically, a $2.2\%$ improvement in both NDS and mAP. This improvement may be attributed to the explicit modeling of sensor- + +![](images/288eca9666cd70ee380a7aed6ce6018295929c7c40edd4ac5922315837b8aff6.jpg) +(a) + +![](images/d3853ae7157a387c8ad57cebcb12f0cd372638c5bcfb39a50972bcfabdc2a0fd.jpg) +(b) + +![](images/b0fbb1507c24c920072a6473118343519a0ff132825d464296950b0a8c145b00.jpg) +(c) +Figure 5. Qualitative results comparing RCTDistill and Baseline. Blue, green, and red boxes denote predictions, static ground truth, and dynamic ground truth, respectively. Pink circles highlight areas where the baseline model fails to detect objects, but RCTDistill successfully identifies them. Orange and yellow circles indicate regions where false positives have been corrected. + +specific uncertainty. + +Table 3b demonstrates the effectiveness of our proposed TKD method compared to existing temporal KD approaches. While STXD [12] relies on constructing temporal similarity maps between past and current BEV features for KD, and VCD [11] applies distillation over circular Gaussian masks across all time steps, our method employs a simpler yet more effective approach. TKD leverages only current-frame ground truth boxes and velocity information to identify adaptive distillation regions that reflect potential vehicle movement patterns. Our method outperforms existing temporal KD approaches, achieving significant improvements of $2.0\%$ in NDS and $1.4\%$ in mAP over the baseline. Additional analysis is provided in the Supplementary Material. + +Table 3c highlights the effectiveness of our RDKD compared to conventional approaches that apply relation KD across the entire region. MonoDistill [5] proposes distillation between feature maps, and CRKD [40] presents a ReID strategy that utilizes distillation on affinity maps. In contrast, our method selectively targets specific regions. Our approach outperforms the existing methods, improving mAP by at least $0.9\%$ and NDS by $0.8\%$ . + +Performance comparison on VoD dataset. Table 4 presents a performance comparison between RCTDistill and existing models based on the VoD [27] validation set. RCTDistill achieves state-of-the-art performance by outperforming the latest SGDet3D method [1], achieving a $2.62\%$ increase in EAA AP and a $4.83\%$ improvement in RoI AP. These results demonstrate that RCTDistill consistently delivers superior + +performance across both the nuScenes [3] and VoD [27] datasets, underscoring its robust performance across diverse environments and sensor characteristics. + +Qualitative results. Figure 5 presents a qualitative comparison between the proposed RCTDistill and the baseline model. The highlighted circles illustrate RCTDistill's enhanced ability to detect missed objects (pink), reduce false positives caused by range and azimuth uncertainties (orange), and resolve false positives caused by dynamic object movement (yellow). These results demonstrate its effectiveness in addressing the limitations of previous methods. + +# 5. Conclusions + +In this paper, we presented RCTDistill, a novel cross-modal knowledge distillation framework designed to improve temporal radar-camera fusion for 3D object detection. RCTDistill effectively addresses key challenges, including sensor-specific uncertainty and temporal misalignment, leading to significant performance gains. The proposed RAKD module mitigates sensor-specific uncertainty by applying elliptical Gaussian regions, thereby enhancing the quality of BEV representations. TKD alleviates temporal misalignment caused by dynamic objects by aligning historical BEV features more accurately with current LiDAR features. Lastly, RDKD improves feature discrimination by transferring relational knowledge from the teacher model, enabling the student model to more effectively distinguish foreground from background features. Overall, RCTDistill establishes a new state-of-the-art in radar-camera-based 3D object detection. + +# 6. Acknowledgement + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) [NO. RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)], the National Research Foundation (NRF) funded by the Korean government (MSIT) (No. RS-2024-00421129), and the Technology Innovation Program (20018112, Development of autonomous manipulation and gripping technology using imitation learning based on visual tactile sensing) funded By the Ministry of Trade, Industry & Energy(MOTIE, Korea). + +# References + +[1] Xiaokai Bai, Zhu Yu, Lianqing Zheng, Xiaohan Zhang, Zili Zhou, Xue Zhang, Fang Wang, Jie Bai, and Hui-Liang Shen. Sgdet3d: Semantics and geometry fusion for 3d object detection using 4d radar and camera. IEEE Robotics and Automation Letters, 2024. 7, 8 +[2] Geonho Bang, Kwangjin Choi, Jisong Kim, Dongsuk Kum, and Jun Won Choi. Radardistill: Boosting radar-based object detection performance via knowledge distillation from lidar features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15491-15500, 2024. 3 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 2, 6, 8 +[4] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. In International Conference on Learning Representations, 2023. 2, 4, 7 +[5] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. In International Conference on Learning Representations, 2022. 2, 7, 8 +[6] 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/CVF Conference on Computer Vision and Pattern Recognition, pages 2002-2011, 2018. 4 +[7] Zijian Gu, Jianwei Ma, Yan Huang, Honghao Wei, Zhanye Chen, Hui Zhang, and Wei Hong. Hgsfusion: Radar-camera fusion with hybrid generation and synchronization for 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3185-3193, 2025. 7 +[8] Xiaoyang Guo, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Liga-stereo: Learning lidar geometry aware representations for stereo-based 3d detector. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3153-3163, 2021. 2 + +[9] 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, pages 770-778, 2016. 7 +[10] Yu Hong, Hang Dai, and Yong Ding. Cross-modality knowledge distillation network for monocular 3d object detection. In European Conference on Computer Vision, pages 87-104, 2022. 2 +[11] Linyan Huang, Zhiqi Li, Chonghao Sima, Wenhai Wang, Jingdong Wang, Yu Qiao, and Hongyang Li. Leveraging vision-centric multi-modal expertise for 3d object detection. Advances in Neural Information Processing Systems, 36:38504-38519, 2023. 3, 7, 8 +[12] Sujin Jang, Dae Ung Jo, Sung Ju Hwang, Dongwook Lee, and Daehyun Ji. Stxd: structural and temporal cross-modal distillation for multi-view 3d object detection. Advances in Neural Information Processing Systems, 36:29323-29342, 2023. 3, 7, 8 +[13] Jisong Kim, Minjae Seong, Geonho Bang, Dongsuk Kum, and Jun Won Choi. Rcm-fusion: Radar-camera multi-level fusion for 3d object detection. In IEEE International Conference on Robotics and Automation, pages 18236-18242. IEEE, 2024. 1, 2 +[14] Jisong Kim, Minjae Seong, and Jun Won Choi. CRT-Fusion: Camera, radar, temporal fusion using motion information for 3d object detection. Advances in Neural Information Processing Systems, 37:108625-108648, 2024. 1, 2, 6, 7 +[15] Youngseok Kim, Juyeb Shin, Sanmin Kim, In-Jae Lee, Jun Won Choi, and Dongsuk Kum. Crn: Camera radar net for accurate, robust, efficient 3d perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17615-17626, 2023. 1, 2, 4, 6 +[16] Marvin Klingner, Shubhankar Borse, Varun Ravi Kumar, Behnaz Rezaei, Venkatraman Narayanan, Senthil Yogamani, and Fatih Porikli. X3kd: Knowledge distillation across modalities, tasks and stages for multi-camera 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13343-13353, 2023. 1, 2, 6 +[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12697-12705, 2019. 7 +[18] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. Advances in Neural Information Processing Systems, 35:18442-18455, 2022. 2 +[19] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2, 6 +[20] Yiheng Li, Yang Yang, and Zhen Lei. Rctrans: Radar-camera transformer via radar densifier and sequential decoder for 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5048-5056, 2025. 1, 6 + +[21] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European Conference on Computer Vision, pages 1-18. Springer, 2022. 2 +[22] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14928-14937, 2024. 1, 2, 4, 6, 7 +[23] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 2 +[24] 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. 7 +[25] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In IEEE International Conference on Robotics and Automation, pages 2774-2781. IEEE, 2023. 3, 7 +[26] Shu-Wei Lu, Yi-Hsuan Tsai, and Yi-Ting Chen. Toward real-world bev perception: Depth uncertainty estimation via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17124–17133, 2025. 2 +[27] Andras Palffy, Ewoud Pool, Srimannarayana Baratam, Julian FP Kooij, and Dariu M Gavrila. Multi-class road user detection with 3+ 1d radar in the view-of-delft dataset. IEEE Robotics and Automation Letters, 7(2):4961-4968, 2022. 2, 6, 7, 8 +[28] Patrick Palmer, Martin Krüger, Stefan Schütte, Richard Al-terendorfer, Ganesh Adam, and Torsten Bertram. Lerojd: Lidar extended radar-only object detection. In European Conference on Computer Vision, pages 379–396. Springer, 2024. 3 +[29] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In International Conference on Learning Representations, 2023. 6, 7 +[30] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 2 +[31] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2 +[32] Zeyu Wang, Dingwen Li, Chenxu Luo, Cihang Xie, and Xiaodong Yang. Distillbev: Boosting multi-camera 3d object detection with cross-modal knowledge distillation. In + +Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8637-8646, 2023. 2 +[33] Philipp Wolters, Johannes Gilg, Torben Teepe, Fabian Herzog, Felix Fent, and Gerhard Rigoll. Sparc: Sparse radar-camera fusion for 3d object detection. arXiv preprint arXiv:2411.19860, 2024. 2, 6, 7 +[34] Weiyi Xiong, Jianan Liu, Tao Huang, Qing-Long Han, Yuxuan Xia, and Bing Zhu. Lxl: Lidar excluded lean 3d object detection with 4d imaging radar and camera fusion. IEEE Transactions on Intelligent Vehicles, 9(1):79–92, 2023. 7 +[35] Ruoyu Xu, Zhiyu Xiang, Chenwei Zhang, Hanzhi Zhong, Xijun Zhao, Ruina Dang, Peng Xu, Tianyu Pu, and Eryun Liu. Sckd: Semi-supervised cross-modality knowledge distillation for 4d radar object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8933-8941, 2025. 3 +[36] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. +[37] 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, pages 11784-11793, 2021. 6 +[38] Songsong Yu, Yifan Wang, Yunzhi Zhuge, Lijun Wang, and Huchuan Lu. Dme: unveiling the bias for better generalized monocular depth estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6817-6825, 2024. 2 +[39] Haimei Zhao, Qiming Zhang, Shanshan Zhao, Zhe Chen, Jing Zhang, and Dacheng Tao. Simdistill: Simulated multi-modal distillation for bev 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7460-7468, 2024. 4 +[40] Lingjun Zhao, Jingyu Song, and Katherine A Skinner. Crkd: Enhanced camera-radar object detection with cross-modality knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15470-15480, 2024. 1, 2, 3, 6, 7, 8 +[41] Wu Zheng, Mingxuan Hong, Li Jiang, and Chi-Wing Fu. Boosting 3d object detection by simulating multimodality on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13638-13647, 2022. 2 +[42] Shengchao Zhou, Weizhou Liu, Chen Hu, Shuchang Zhou, and Chao Ma. Unidistill: A universal cross-modality knowledge distillation framework for 3d object detection in bird's-eye view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5116-5125, 2023. 2, 3 +[43] Taohua Zhou, Junjie Chen, Yining Shi, Kun Jiang, Mengmeng Yang, and Dange Yang. Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection. IEEE Transactions on Intelligent Vehicles, 8(2): 1523-1535, 2023. 2 +[44] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019. 6 \ No newline at end of file diff --git a/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/images.zip b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bb96e292080f224235dcbea1c7c170614855520a --- /dev/null +++ b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ad4d467bb72737a672d7b4da80ed910075fe268df0140cd0fae51d777929d94 +size 790499 diff --git a/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/layout.json b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0b94c10ce50f36cd5c086fe238306751d6bdf746 --- /dev/null +++ b/ICCV/2025/RCTDistill_ Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ca2d6dfe6218e23a5bcbb19b63775d04e1b794d3a35a8cb11ea909bc48f9eb8 +size 396478 diff --git a/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/184a3ddf-9191-495c-ac71-0c98fabd0f16_content_list.json b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/184a3ddf-9191-495c-ac71-0c98fabd0f16_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2001915e6560a3887deb60d826c95a9ef74c209 --- /dev/null +++ b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/184a3ddf-9191-495c-ac71-0c98fabd0f16_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7eba0ed9004b0cebed3907ffc8419b902e982f7e929930f18a32136ce8e22400 +size 77621 diff --git a/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/184a3ddf-9191-495c-ac71-0c98fabd0f16_model.json b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/184a3ddf-9191-495c-ac71-0c98fabd0f16_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e74cf931753f9353c4100cdc0c7a507d6630f467 --- /dev/null +++ b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/184a3ddf-9191-495c-ac71-0c98fabd0f16_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5282359fe7bc1fd83ab9bfc7384edc166cb4b7643e15681a0d97416d819f42c +size 102681 diff --git a/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/184a3ddf-9191-495c-ac71-0c98fabd0f16_origin.pdf b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/184a3ddf-9191-495c-ac71-0c98fabd0f16_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..92298b2a94cb8556479c7c9f539f37ac3ad4e2b3 --- /dev/null +++ b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/184a3ddf-9191-495c-ac71-0c98fabd0f16_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:570b2727cecdf0591e386fb22685d9590c6f05b76f1af8c09192cbcab869b1dc +size 5267973 diff --git a/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/full.md b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d07fb1a6b3d32711104af9608e2bb301520afe95 --- /dev/null +++ b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/full.md @@ -0,0 +1,323 @@ +# REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents + +Rui Tian $^{1,2}$ Qi Dai $^{3*}$ Jianmin Bao $^{3}$ Kai Qiu $^{3}$ Yifan Yang $^{3}$ Chong Luo $^{3}$ Zuxuan Wu $^{1,2*}$ Yu-Gang Jiang $^{1,2}$ + +$^{1}$ Institute of Trustworthy Embodied AI, Fudan University + +$^{2}$ Shanghai Collaborative Innovation Center of Intelligent Visual Computing $^{3}$ Microsoft Research + +![](images/83516294a4f8733fa2c5556534d31120d39335f1688b4032c0bcb381d1ba9e48.jpg) + +![](images/ddc7ad11b9173ca6a2bbf846f71737b30e389bfd8f0888309b9391257ad384ec.jpg) +Impressionist style, a yellow rubber duck floating on the waves at sunset. + +![](images/b7c7a6f2d9fbc0f63d37ad5d6de7a40a5a14257afce36b4337b4fdfebfc606d3.jpg) +An aerial shot of a lighthouse standing tall on a rocky cliff, its beacon cutting through the early dawn, waves crashing against the rocks below. +A wizard waves his wand. +Figure 1. Examples of generated high-resolution videos. + +# Abstract + +Commercial video generation models have exhibited realistic, high-fidelity results but are still restricted to limited access. One crucial obstacle for large-scale applications is the expensive training and inference cost. In this paper, we argue that videos contain significantly more redundant information than images, allowing them to be encoded with very few motion latents. Towards this goal, we design an image-conditioned VAE that projects videos into extremely compressed latent space and decode them based on content images. This magic Reducio charm enables $64 \times$ reduction of latents compared to a common 2D VAE, without sacrificing the quality. Building upon Reducio-VAE, we + +can train diffusion models for high-resolution video generation efficiently. Specifically, we adopt a two-stage generation paradigm, first generating a condition image via text-to-image generation, followed by text-image-to-video generation with the proposed Reducio-DiT. Extensive experiments show that our model achieves strong performance in evaluation. More importantly, our method significantly boosts the training and inference efficiency of video LDMs. Reducio-DiT is trained in just 3.2K A100 GPU hours in total and can generate a 16-frame $1024 \times 1024$ video clip within 15.5 seconds on a single A100 GPU. Code is available at https://github.com/microsoft/Reducio-VAE. + +# 1. Introduction + +Recent advances in latent diffusion models (LDMs) [40] for video generation have presented inspiring results [1-5, 36, 48-51, 55, 61, 65, 68], showing great potential in various applications. Commercial models like Sora [31], Runway Gen-3 [41], Movie Gen [35], and Kling [19] can already generate photorealistic and high-resolution video clips. However, training and deploying such models are computationally demanding—thousands of GPUs and millions of GPU hours are required for training, and the inference for a one-second clip costs several minutes. Such a high cost is becoming a significant obstacle for research and large-scale real-world applications. + +Extensive effort has been made to relieve the computational burden by means of efficient computation modules in the backbone [47, 64], optimizing diffusion training strategies [13, 60], and adopting few-step sampling methods [44, 67]. However, in this paper, we argue that the very nature of the problem is long neglected. Concretely, most video LDMs stick to the paradigm of text-to-image diffusion, e.g., Stable Diffusion (SD) [40], and inherently employ the latent space of its pre-trained 2D variational autoencoder (VAE) [18, 38]. This scheme compresses the input videos by $8 \times$ in each spatial dimension, which is suitable for images while being much surplus for video. + +As videos naturally carry much more redundancy than images, we believe they can be projected into a much more compressed latent space with a VAE based on 3D convolutions. Surprisingly, for videos, a $16\times$ spatial downsampling factor does not sacrifice much of the reconstruction performance. While the resulting distribution of latent space shifts from those of the pre-trained image LDMs, we can still wield their well-learned spatial prior by factorizing text-to-video generation with a condition frame [12, 57, 63, 72]. In particular, we can obtain a content frame with off-the-shelf text-to-image LDMs and then generate samples with text and image as joint priors into video LDMs. + +More importantly, as the image prior delivers enriched spatial information of video contents, factorized video generation offers the potential for further compression. We argue that video can be encoded into very few latents that represent motion variables, i.e., motion latents, plus a content image. Hereby, we introduce the magical Reducio charm by building a VAE comprised of a 3D encoder that aggressively compresses input videos into a $4096 \times$ down-sampled compact space, and a 3D decoder fused with feature pyramids of the middle frame as the content condition. Notably, the resulting Reducio-VAE surpasses the common 2D VAE by 5db in PSNR while using a $64 \times$ smaller latent space. + +Subsequently, we establish the LDM with a diffusion transformer (DiT) [6, 32]. Besides using T5 features [37] as the text condition, we employ an image semantic encoder as well as a context encoder to provide additional + +image conditions, which inform the model of the spatial content. Thanks to the extremely compressed video latent, our diffusion model, namely Reducio-DiT, thus enjoys both the fast training (inference) speed and high generation quality. Specifically, Reducio-DiT achieves 318.5 FVD score on UCF-101, surpassing a bunch of previous work [1, 4, 57, 58, 73] while being much faster. It can also be easily upscaled to a larger resolution, e.g. $1024^2$ , with affordable cost. Our experiments show that Reducio-DiT can generate a 16-frame video clip of $1024^2$ resolution within 15.5 seconds on a single A100 GPU, and hence achieves a $16.6 \times$ speedup over Lavie [58]. + +In summary, our contributions are as follows: (1) We encode video into very few motion latents plus a content image, where the designed Reducio-VAE can compress video into a $64 \times$ smaller latent space compared to common 2D VAE. (2) Based on Reducio-VAE, a diffusion model called Reducio-DiT is devised by incorporating multimodal conditions to perform text-image to video generation. (3) Experiments show that our Reducio-DiT can significantly speed up the generation process while producing high-quality videos. + +# 2. Related Work + +Video diffusion models have become the latest craze in the field of text-to-video generation. Due to the huge success of text-to-image LDMs [14, 28, 29, 40, 43], attempts have been made to leverage the powerful UNet-based stable diffusion models pre-trained in 2D space by inserting temporal convolution and attention layers [3, 4, 11, 55, 64]. Recently, the strong performance of transformers has continued in image generation domain [6, 9, 20, 26, 54]. Consequently, recent works have focused on extending 2D diffusion transformer (DiT) [32] to spatiotemporal domain. + +While the mainstream exploration [21, 35, 78] follows the paradigm of image diffusion models by using text as the only condition, an emerging stream of works [12, 57, 72] decompose text-to-video diffusion into two steps, i.e., text-to-image and text-image-to-video generation. The factorized framework encourages models to focus on motion modeling instead of challenging spatiotemporal joint modeling. In this work, we build an image-to-video model upon DiT. Thanks to the delicate design of Reducio-VAE, our diffusion model preserves fine details of image priors in generated videos and hence can be used for more high-demanding real-world applications, e.g., advertisement generation. + +Efficient diffusion models have gained increasing attention due to the tremendous computational resources and time required for training and inferring samples, especially for high-resolution images and videos. On one hand, researchers have made attempts to build LDMs with more efficient architectures [10, 47, 62, 77]. For example, DIM [47] uses Mamba as its backbone to ease the computational bur + +![](images/bb611a5ba987e90c0568376bc41094be709efcc03a09eb8a40839bfdd944a17c.jpg) + +![](images/20a56a25930a0fee06cffa13aea17c0184211f186016abc568645990ef4da494.jpg) +(c) $H / 4\times W / 4$ + +![](images/92e68f5dccdd77492a6aac4fb88ecdf2a078fb4a8adb9b0e2f9c57414820b896.jpg) +Latent Dimension + +![](images/850ae5a0addd8f20312b516e75540b3a31065c114f18ba22b2d20cf86ccbf77a.jpg) +Reconstruction Performance + +![](images/85b8132888db9dc325912d4781d88d82ae4cea52e8eefcb87a01da057fdf41e7.jpg) + +![](images/2543478daf45dcb6675840b1c0e0a6283157132d2e6e31ec50e362e2b74ed29e.jpg) +Figure 2. Architecture and performance comparisons between vanilla 2D VAE in (a) SDXL, (b) 3D VAE and (c) the proposed Reducio-VAE. Reducio-VAE enables $64\times$ reduction of latent size while achieving much higher reconstruction performance. + +PSNR + +![](images/f5915d960c660969c5482167f21209cd5117735a9bcf28e436b043c590014548.jpg) + +![](images/19ba4ce2c0f63897aeda407e8bf5be5035433f97db517917f978e8bbfad5f5b3.jpg) + +![](images/12c072c21412769631cafc1610cbcd92a833dacb9766057ea15ed4136e85809f.jpg) +SSIM + +PSNR + +den compared to vanilla transformers. Another line of work leverages efficient diffusion techniques to perform fewer sampling steps during inference [23, 25, 44, 67, 74] or to accelerate convergence during training [13, 60]. For instance, CausVid [69] applies distribution matching distillation between a causal student video generation model and a bidirectional teacher model. + +Our work focuses on designing a more compact latent space [33], which straightforwardly and effectively reduces the overall training and inference costs of LDM. PVDM [70] and CMD [71] propose projecting 3D video samples into three separate yet compact 2D latent spaces. However, this approach requires training separate diffusion models for each space, which inevitably leads to discrepancies during inference. Our Reducio-VAE takes after LaMD [16] to adopt a factorized, compact motion latent space with the help of a content frame for reconstruction. In contrast, we emphasize aggressive compression of the spatial dimension and suggest that a moderate temporal down-sampling factor provides the best trade-off between performance and efficiency. Concurrent research in the image domain also indicates that using a high spatial down-sampling factor on redundant visual content [46, 62], such as high-resolution images, significantly improves efficiency. We propose that it is crucial to recognize the importance of spatial redundancy in the video latent space. + +# 3. Reducio + +# 3.1. Reducio Video Autoencoder + +Variational autoencoder (VAE) [18] projects input images or videos into a compressed latent space obeying a certain distribution and maps the latent sampled from the obtained distribution into RGB space with a decoder. Latent diffusion models [40] take advantage of the down-sampled latent space of VAE to support generation with improved efficiency. In this paper, we follow the common practice and build our video autoencoder upon VAE. + +We denote the temporal down-sampling factor of a VAE as $f_{t}$ and the spatial one as $f_{s}$ . Given an input video $\mathcal{V}$ with the shape of $3\times T\times H\times W$ , the compressed latent + +$z_{\mathcal{V}}$ as the input into diffusion models fall into the space of $|z|\times T / f_t\times H / f_s\times W / f_s$ , where $z$ denotes the latent channel. Thanks to the superior performance of open-source Stable Diffusion (SD) [40] models, it has been a common practice to leverage SD-VAE to process images and videos into compressed latent. Typically, as shown in (a) of Fig. 2, SD-VAE is built upon 2D convolution with $f_{s} = 8$ and $f_{t} = 1$ . + +Reducio charm on videos. As adjacent frames share large similarities in pixel space, video signals inherently contain more redundancy than images. Therefore, we apply a more aggressive down-sampling strategy for video autoencoder. Specifically, we employ 3D convolutions and increase the number of autoencoder blocks, increasing $f_{t}$ and $f_{s}$ to 4 and 16, as shown in (b) of Fig. 2, respectively. Though the overall down-sampling factor increases by $16\times$ , the compressed latent space suffices to achieve comparable pixel-space reconstruction quality, highlighting the potential for extremely shrinking video latent dimensions. + +As for image-to-video generation, we argue that video latent can be further compressed as the content image contains rich visual details and substantial overlapping information. In such a scenario, video latents can be further compressed to represent the motion information. We can find similar spirits in the video codec [39], where carefully designed modules use reference frames as inputs to perform high-fidelity reconstruction of a given compressed code. Specifically, we propose to cast the Reducio charm on video latent space, as shown in (c) of Fig. 2. Generally, we choose the middle frame $V_{T/2}$ as the content guidance, compress input videos by an extreme down-sampling factor with a 3D encoder, and reconstruct from the latent space with the aid of content frame features pyramids. Specifically, we infuse content features of sizes varying in $H/8 \times W/8$ and $H/4 \times W/4$ via cross-attention within the decoder, adopt an extreme compression ratio, i.e., $f_s = 32$ and $f_t = 4$ . While Reducio-VAE projects videos into $64 \times$ more compressed latent space than one of SDXL-VAE [34], it significantly outperforms the latter in reconstruction performance. + +Note that we train Reducio-VAE on videos of $256 \times 256$ resolution. To encode and decode high-resolution videos without exceeding the memory limit, we split the input + +![](images/ef55971d6ba6687137fb495643dfb4d731ebe5ad63b649a9a4a187b84d9cff64.jpg) +Figure 3. The overview of Reducio-DiT training framework. Compared to the vanilla DiT, Reducio-DiT adopts an additional image-condition module to inject the semantics and content of the keyframe via cross-attention. + +video into overlapping tiles along the spatial dimension, following the approach used in Movie Gen [35]. In practice, we use a tile size of 256 (8 in latent space) with an overlap of 64 pixels (2 latents). Overlapping latents and decoded video regions are blended using linear combinations. + +# 3.2. Reducio Diffusion Transformer + +Thanks to the magical compression capability of ReducioVAE, we project input videos into compressed latent space with a down-sampling factor of 4096, hence significantly speeding up the training and inference of diffusion models. Specifically, we adopt the DiT-XL [32] model and follow most of the design choices adopted by PixArt- $\alpha$ [6], i.e., AdaLN-single modules, cross attention layers with Flant5-XXL [37] text conditions. + +To adapt the image diffusion model to video, we consider two options: (1) Directly transforming the 2D attention to full 3D attention without adding additional parameters [7, 68]. (2) Adding temporal layers and transforming the model to perform 2D spatial attention plus 1D temporal attention [21, 72, 78]. We adopt option (1) by default. Studies on the evaluation of these two options are shown in Sec. 4.3.2. In addition, we introduce Reducio-DiT (illustrated in Fig. 3) with additional image-condition modules, as described below. + +Content frame modules consist of a semantic encoder built upon pretrained OpenCLIP ViT-H [8] and a content encoder initialized with the SD2.1-VAE [40]. The former encoder projects content frames into a high-level semantic space while the latter one mainly focuses on extracting the spatial information. We concatenate the obtained features with text tokens output by T5 to form the image-text-joint condition, and then cross-attend them with the noisy video latent. + +Scaling up to high-resolution videos consumes tremendous computational resources for vanilla video diffusion + +![](images/fbb5301dfa9874b5f5e48e6cbcc9980c448725ce0e70f3e9b38f65ff39684e18.jpg) +Figure 4. Comparison between the $1^{st}$ and $16^{th}$ frames of videos reconstructed by SDXL-VAE (top) and Reducio-VAE (bottom). + +models. Nevertheless, Reducio-DiT relieves this limit to a great extent. To support high-resolution (e.g., $1024^2$ ) video generation, we employ a progressive training strategy: 1) in the first stage of training, the model learns to align the video latent space with text-image prior by taking in a vast number of $256^2$ videos as input; 2) in the second stage, we fine-tune the model on videos with higher resolution, i.e., $512 \times 512$ . In consequence, the content encoder augments $4 \times$ more tokens with the noisy latent; 3) in the third stage, we conduct further fine-tuning on videos with spatial resolution around 1024 with multi-aspect augmentation. + +To better accommodate the model to videos with varying aspect ratios, we inject the aspect ratio and size of input videos as embeddings [6] into the DiT model. We then add the dynamic size embedding with timestep embedding before being fed into the MLP. More importantly, for models in the first and the second stages, the content encoder cross-attends $H / 16 \times W / 16$ tokens with the noisy latent. As computational costs grow rapidly when resolution increases, we introduce a strategy to ensure the efficiency of Reducio-DiT on high-resolution videos. Specifically, we aim to shrink the content embedding by $2 \times$ in the spatial dimension in each cross-attention layer. Inspired by Deepstack [27], we + +Table 1. Quantitative comparison between Reducio-VAE and open-source state-of-the-art VAE for generation. + +
ModelDownsample Factor|z|PexelsUCF-101 rFVD↓
PSNR↑SSIM↑LPIPS↓rFVD↓
SD2.1-VAE [40]1 × 8 × 8429.230.820.0925.9621.00
SDXL-VAE [34]1 × 8 × 8430.540.850.0819.8723.68
OmniTokenizer [53]4 × 8 × 8827.110.890.0723.8830.52
OpenSora-1.2 [78]4 × 8 × 81630.720.850.1160.8867.52
Cosmos-VAE [30]8 × 8 × 81630.840.740.1229.4422.06
8 × 16 × 161628.140.650.1877.87119.37
Reducio-VAE4 × 32 × 321635.880.940.0517.8865.17
+ +divide input tokens into 4 groups of grids and iteratively concatenate each group of $H / 32 \times W / 32$ tokens with the other condition tokens. + +# 4. Experiments + +# 4.1. Training and Evaluation Details + +Our Reducio-VAE is trained on a 400K video dataset collected from Pexels*. Pexels contains a large number of free and high-quality stock videos, where each video is accompanied by a short text description. We train Reducio-VAE from scratch on $256 \times 256$ videos with 16 FPS. To perform the VAE comparison, PSNR, SSIM [59], LPIPS [75] and reconstruction FVD (rFVD) [52] are employed as the evaluation metrics. + +Reducio-DiT is trained on the above Pexels dataset and an internal dataset with 5M videos containing high-resolution text-video pairs. As described in Sec. 3.2, we adopt a multi-stage training strategy to train the model from low resolution to high resolution. We first train Reducio-DiT-256 using a batch size of 512 on 4 Nvidia A100 80G GPUs for around 900 A100 hours and initialize model weights with PixArt- $\alpha$ -256 [6]. Then we fine-tune on $512^2$ videos shortly for 300 A100 hours to obtain Reducio-DiT-512. In the third stage, we randomly sample video batches from 40 aspect ratio buckets, which is the same as the setting in PixArt- $\alpha$ . We leverage 8 AMD MI300 GPUs to support a training batch size of 768 and fine-tune Reducio-DiT-1024 for 1000 GPU hours. + +We adopt DPM-Solver++ [24] as the sampling algorithm for efficient inference and set the sampling step to 20. For evaluation in Tab. 3, we use PixArt- $\alpha$ -256 and PixArt- $\alpha$ 1024 to generate the content frame for $256^{2}$ and $1024^{2}$ videos, respectively. We denote out-of-memory as OOM and calculate the speed of generating 16-frame video clips on a single A100 80G GPU, including both text-to-image and text-image to video cost. For Reducio-DiT, We report the FVD and IS scores of Reducio-DiT-512 under the zero-shot setting on UCF-101 [45] and MSR-VTT [66]. Since + +Table 2. Quantitative comparison between Reducio-DiT and state-of-the-art image-to-video LDMs. + +
ModelFVD↓
UCF-101MSR-VTT
VideoComposer [56]576.8377.3
I2VGen-XL [76]571.1-
DynamiCrafter [63]429.2234.7
Reducio-DiT318.5291.9
+ +our model performs image-to-video generation and requires a condition image prior for the text-to-video pipeline, we follow MicroCinema [57] to use SDXL [34] to generate the condition image and LLaVA-1.5 [22] to generate captions for FVD evaluations on UCF-101 and MSR-VTT. Additionally, we evaluate Reducio-DiT-512 on the recent video generation benchmark VBench [17]. + +# 4.2. Main Results + +Reducio-VAE preserves fine details of video inputs. Fig. 4 displays the first and the last frame of $256^{2}$ videos reconstructed by SDXL-VAE and Reducio-VAE, with the yellow bounding box highlighting the region with obvious differences in details. While SDXL-VAE causes corruptions and blurs, Reducio-VAE generally maintains the subtle texture within the middle frame. Tab. 1 also reflects the advantage of Reducio-VAE in quantitative reconstruction metrics, e.g., PSNR, SSIM, and LPIPS on a randomly sampled validation subset of 1K videos with the duration of 1s on Pexels, and test on both UCF-101 and Pexels for FVD. + +As shown in the table, Reducio-VAE strikes obvious advantages on overall reconstruction performance. Specifically, our model significantly outperforms state-of-the-art 2D VAE, e.g., SD2.1-VAE and SDXL-VAE, by more than 5 db in PSNR. Moreover, Reducio-VAE also performs better than VAEs designed for video in recent literature, e.g., OmniTokenizer and OpenSora-1.2. Our VAE also outperforms the concurrent work, Cosmos-VAE, which also adopts a further compressed video latent space, by 0.2 in SSIM and 5 db in PSNR with an $8 \times$ higher down-sampling factor. + +Table 3. Quantitative comparison between Reducio-DiT and state-of-the-art text-to-video LDMs. + +
ModelParamsVBatchFVD↓Throughput (#videos/sec)GPU Hours A100 Training
UCF-101MSR-VTT256 × 2561024 × 1024
Make-A-Video [42]--367.23----
CogVideo [15]7.8B67.01701.5912940.0023OOM>368K
Show-1 [73]1.7B78.93394.46538.00.0016OOM>56K
Lavie [58]1B77.08526.30-0.18380.0039-
MicroCinema [57]2B-342.86377.40.0234OOM>6K
Lumiere [2]>2B-332.49----
Reducio-DiT1.2B81.39318.50291.90.98240.06503.2K
+ +![](images/8ab430b71a6047a0178c46b3b3688400f06bad2d701fa5b3803dfae0a3f3ea8a.jpg) + +![](images/d7cbe6da8666e3481b06b56cd9cd38055e71eab11b5606db81656dbe8ea0c45b.jpg) +(a) +A child excitedly swings on a rusty swing set, laughter filling the air. +Figure 5. Comparison between frames generated given an identical frame and prompt, by (a) DynamicCrafter [63], (b) SVD-XT [3] and (c) Reducio-DiT, respectively. We resize the output frames from $1344 \times 768$ to $1024 \times 576$ to match with baselines. + +![](images/b711495fa7815c5abe50f71de176bf5d0daccbc9e02ffd56254800aaa2144c57.jpg) +Figure 6. The user study of human preference rate for Reducio-DiT, Dynami-Crafter [63] and SVD-XT [3]. + +We observe that Reducio-VAE has worse rFVD than others on UCF-101. By careful inspection, we find the selected prior content image is blurring in many cases due to the low visual quality of UCF-101 and hence results in a lower reconstruction quality. In contrast, Reducio-VAE places first on rFVD on pexels, which consists of videos with much improved visual quality and much higher resolution. + +Reducio-DiT is a powerful image-to-video model. As shown in Tab. 2, Reducio-DiT achieves a 90-point lower FVD in the zero-shot setting on UCF-101 compared to the state-of-the-art image-to-video model, DynamiCrafter. Additionally, Reducio-DiT surpasses VideoComposer in FVD metrics on MSR-VTT but falls short of DynamiCrafter in this aspect. For qualitative evaluation, we conduct a user study using identical images generated by PixArt- $\alpha$ -1024 as input. Participants evaluate 15 video groups based on three criteria: (1) visual quality, (2) motion quality, and (3) alignment of the condition image. Responses from 20 participants are presented in Fig. 6. + +Notably, Reducio-DiT significantly outperforms in visual quality (64%) and content alignment (57%) by overcoming temporal flickering and preserving fine details from the condition image. Moreover, thanks to its strong temporal consistency, Reducio-DiT achieves a preference rate comparable to SVD-XT in motion quality, yet needs much + +fewer training hours and data. Fig. 5 displays the comparison between Reducio-DiT and the other two stat-of-the-art image-to-video models. While DynamicCrafter and SVD-XT fail to generate stable frames consistent with the given content frame, Reducio-DiT yields high visual quality with maintained facial details. Overall, Reducio-DiT demonstrates strong performance in image-to-video generation. + +Reducio-DiT effectively balances efficiency and performance. By integrating Reducio-DiT with a text-to-image LDM for condition image generation, the factorized text-to-video task can be performed efficiently. Consequently, our model achieves FVD scores of 318.5 on UCF-101 and 291.0 on MSR-VTT and the VBench score of 81.39, beating a range of state-of-the-art video LDMs, as recorded in Tab. 3. More importantly, as most mainstream text-to-video models are built upon pre-trained text-to-image LDMs, we compare Reducio-DiT with baselines on training costs for video samples. Thanks to the extremely compressed latent space, we achieve competitive performance with a training cost of merely 3.2K A100 hours. + +Moreover, Reducio-DiT significantly boosts the throughput of video LDMs. The compressed latent space facilitates Reducio-DiT to output $256^{2}$ videos of 16 frames in almost 1 second, which realizes a $5.3 \times$ speed-up compared with Lavie. Moreover, 16-frame $1024^{2}$ video generation costs + +Figure 7. Examples of a 2s video (top) and a 3s video (bottom) generated by Reducio-DiT in a recursive manner. +![](images/17ac9e393964391303beb1c7fc50f9a302ababc64c078e43cf81b7bc485f458e.jpg) +top: pouring milk into a cup of coffee. +bottom: A storm trooper vacuuming the beach. + +Table 4. Ablation on the down-sampling factor of Reducio-VAE. + +
ftfsPSNR↑SSIM↑
161634.140.93
23236.790.96
43235.880.94
16435.510.93
+ +Reducio-DiT only for 15.4 seconds on average, leading to a striking speed-up of $16.6 \times$ over Lavie. We argue that the core philosophy of Reducio lies in the extremely compressed latent space, which accelerates both the training and inference speed of text-to-video generation. + +Although Reducio-DiT is trained exclusively on 16-frame videos sampled at 16 fps, we present that it can be adapted for longer video generation in a zero-shot manner. This is achieved by recursively using the last frame of a generated clip as the input condition for the subsequent clip. Fig. 7 presents the visualization of frames sampled uniformly from 32-frame and 48-frame videos, demonstrating that Reducio-DiT maintains consistent motion while preserving high visual quality. + +# 4.3. Ablation Studies + +# 4.3.1. Design of Reducio-VAE + +Spatial redundancy is the key for compression. As Tab. 4 displays, when adopting an identical overall down-sampling factor $\mathcal{F} = f_t \cdot f_s^2$ , Reducio-VAE achieves worse performance with a larger compression factor in the temporal dimension. The result suggests that when a reference frame is provided, videos contain heavier redundancy in the spatial dimension than in the temporal dimension. Similarly, when $f_s = 32$ is kept unchanged, increasing $f_t$ from 2 to 4 leads to a drop in PSNR by 0.91 db. To balance the trade-off between more compact latent space and better reconstruction performance, we set $f_t = 4$ and $f_s = 32$ in default. + +Scaling up latent channel is helpful for improving the reconstruction performance, as results shown in Tab. 5. When $z$ grows from 4 to 16, the PSNR of reconstructed videos increases by 2.71 db. However, increasing the latent channel has a ceiling effect, as setting $z$ to 32 hardly achieves fur + +![](images/4a98da468082453b3e2a28c9836f2d6118978ef1f9b5ed1a9dcff0ace567fc68.jpg) +Figure 8. Comparison between the $8^{\text{th}}$ to $11^{\text{th}}$ frames of videos reconstructed by Reducio-VAE with different latent space channel dimensions, i.e., (a) $|z| = 16$ , (b) $|z| = 8$ , (c) $|z| = 4$ . + +ther gains in both PSNR and SSIM. Therefore, we opt to use latent space with $|z| = 16$ for Reducio as the default setting. A visual qualitative comparison is presented in Fig. 8, where we compare the results when $|z| = 16, 8,$ and 4. + +Table 5. Ablation on latent channel dimensions of Reducio-VAE. + +
|z|PSNR↑SSIM↑
433.170.91
834.150.93
1635.880.94
3235.240.94
+ +Fusion with cross-attention achieves the most competitive performance in Tab. 6. As for add-based fusion, we feed the content frame features into a convolution and replicate the output on the temporal dimension. Then we add video features with the content feature, together with a temporal embedding. Concerning linear-based fusion, we view the content condition as a single-frame video feature and concatenate it with the video feature by the temporal dimension. A 3D convolution layer then refines the concatenated features. As observed in Tab. 6, the attention-based ReducioVAE outperforms other baselines by 1.71 in PSNR and 0.01 in SSIM, while demanding higher computational costs. By + +Table 6. Ablation on the fusion type in Reducio-VAE when infusing the content image. + +
FusionPSNR↑SSIM↑Params
w/o.27.910.80435M
linear30.860.87439M
add34.170.93450M
attention35.880.94437M
+ +default, we display results using the attention-based fusion. Note that all Reducio-VAE displayed in Tab. 6 share the same 3D encoder and hence an identical latent space, where users are free to choose any of the decoders with different resource constraints. + +# 4.3.2. Design of Reducio-DiT + +For ablation studies on DiT, we conduct training for 300 A100 hours for each experiment on Pexels and report FVD and IS scores on UCF-101. + +Using patch size of one yields the best performance. As shown in Tab. 7, keeping a fixed spatial down-sampling factor of 32, we compare the performance of using a $2 \times$ higher temporal down-sampling factor in latent space and a temporal patch size of 1 ( $f_{t} = 4$ , $p_{t} = 1$ ), with temporal patch size of 2 when patchify the input latents for obtaining patch embedding ( $f_{t} = 2$ , $p_{t} = 2$ ). We observe that compressing the temporal dimension with VAE outperforms. Similarly, we ablate using VAE with $2 \times$ lower spatial down-sampling factor yet setting the patch size of spatial dimension to 2 ( $f_{s} = 16$ , $p_{s} = 2$ ). The experimental results identify that compressing the video with Reducio-VAE to an extreme extent and using a small patch size, i.e., 1, in both the temporal and spatial dimensions achieves better performance. + +Table 7. Ablation on the temporal patch size $p_t$ and the spatial patch size $p_s$ in Reducio-DiT. + +
f_tf_sp_tp_sFVD↓IS↑
23221351.032.7
41612534.130.4
43211337.634.1
+ +Fusion with both semantic and content information contributes to the best quality. As shown in Tab. 8, using semantic features, i.e., OpenClip-based features alone may contribute to distorted visual detail, leading to a lower FVD. On the other hand, using content-based features helps to make video smooth and stable, achieving a higher FVD yet less versatile motion. A comparison example is illustrated in Fig. 9. The collaboration of content and semantic features helps Reducio strike the best balance between content consistency and motion richness. + +![](images/d4db6e9bfc82d4ee9d0b37be5a117f6c82ad8fa4cd10c5a947bc4e180c19c5e0.jpg) +(a) +A stunning sight as a steam train leaves the bridge, traveling over the arch-covered viaduct. The landscape is dotted with lush greenery and rocky mountains. +Figure 9. Comparison on conditioning type in Reducio-DiT. (a) using semant.+content features. (b) using semantic features only. + +Table 8. Ablation on the conditioning type in Reducio-DiT. + +
ConditionFVD↓IS↑
semantic472.134.5
content363.731.1
semantic + content337.634.1
+ +# 5. Conclusion + +Video generation has shown promising results with many potential applications while it is still trapped by the unaffordable computation cost. In this paper, we explored how to effectively cut the overhead by reducing the latent size. In particular, we found that video can be encoded into extremely compressed latents with the help of a content image prior, where the latent codes only need to represent the motion variables. With this observation, we designed the Reducio-VAE to compress videos to $4096 \times$ smaller latents. Using this powerful Reducio-VAE, we trained the Reducio-DiT that enables fast high-resolution video generation. Our method is also compatible with other accelerating techniques, e.g., efficient attention, rectified flow, diffusion distillation, and etc., allowing for further speedup, where we leave it for future exploration. + +Limitations and future work. While Reducio-DiT maintains strong consistency with the center frame, the number of training videos is still relatively limited for good instruction following performance. Also, 16-frame training samples somehow limit the magnitude of the motion in generated videos. As Reducio is much more compute-friendly with increased video frames and extended training data. Therefore, we believe in the potential of adapting our work for longer video generation on large-scale training and will explore it as future work. + +Acknowledgement This work was supported in part by the National Natural Science Foundation of China (Grant 62032006 and Grant 62472098). + +# References + +[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2 +[2] 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 space-time diffusion model for video generation. In SIGGRAPH Asia, 2024. 6 +[3] 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, 6 +[4] 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 +[5] 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 +[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2, 4, 5 +[7] Shoufa Chen, Chongjian Ge, Yuqi Zhang, Yida Zhang, Fengda Zhu, Hao Yang, Hongxiang Hao, Hui Wu, Zhichao Lai, Yifei Hu, et al. Goku: Flow based video generative foundation models. In CVPR, 2025. 4 +[8] 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. 4 +[9] 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. 2 +[10] Zhengcong Fei, Mingyuan Fan, Changqian Yu, Debang Li, Youqiang Zhang, and Junshi Huang. Dimba: Transformermamba diffusion models. arXiv preprint arXiv:2406.01159, 2024. 2 +[11] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Ccredit: Creative and controllable video editing via diffusion models. In CVPR, 2024. 2 +[12] 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. In ECCV, 2024. 2 + +[13] Tiankai Hang, Shuyang Gu, Chen Li, Jianmin Bao, Dong Chen, Han Hu, Xin Geng, and Baining Guo. Efficient diffusion training via min-snr weighting strategy. In ICCV, 2023. 2, 3 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2 +[15] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. *ICLR*, 2023. 6 +[16] Yaosi Hu, Zhenzhong Chen, and Chong Luo. Lamd: Latent motion diffusion for video generation. IJCV, 2025. 3 +[17] 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 CVPR, 2024. 5 +[18] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 2, 3 +[19] Kuaishou. Kling. https://kling.kuaishou.com/en, 2024. 2 +[20] Black Forest Labs. Flux.1. https://github.com/black-forest-labs/flux, 2014.2 +[21] Zongyu Lin, Wei Liu, Chen Chen, Jiasen Lu, Wenze Hu, Tsu-Jui Fu, Jesse Allardice, Zhengfeng Lai, Liangchen Song, Bowen Zhang, et al. Stiv: Scalable text and image conditioned video generation. arXiv preprint arXiv:2412.07730, 2024. 2, 4 +[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 5 +[23] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In ICLR, 2023. 3 +[24] 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. 5 +[25] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 3 +[26] Xu Ma, Peize Sun, Haoyu Ma, Hao Tang, Chih-Yao Ma, Jialiang Wang, Kunpeng Li, Xiaoliang Dai, Yujun Shi, Xuan Ju, et al. Token-shuffle: Towards high-resolution image generation with autoregressive models. arXiv preprint arXiv:2504.17789, 2025. 2 +[27] Lingchen Meng, Jianwei Yang, Rui Tian, Xiyang Dai, Zuxuan Wu, Jianfeng Gao, and Yu-Gang Jiang. Deepstack: Deeply stacking visual tokens is surprisingly simple and effective for Imms. In NeurIPS, 2024. 4 +[28] Alex 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. 2 +[29] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2 +[30] Nvidia. Cosmos tokenizer: A suite of image and video neural tokenizers. https://github.com/NVIDIA/Cosmos-Tokenizer, 2024.5 + +[31] OpenAI. Sora. https://openai.com/index/sora/, 2024.2 +[32] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 4 +[33] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In ICLR, 2024. 3 +[34] 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. 3, 5 +[35] 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, 4 +[36] Zhiwu Qing, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yujie Wei, Yingya Zhang, Changxin Gao, and Nong Sang. Hierarchical spatio-temporal decoupling for text-to-video generation. In CVPR, 2024. 2 +[37] 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. JMLR, 2020. 2, 4 +[38] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, 2014. 2 +[39] Iain E Richardson. The H. 264 advanced video compression standard. John Wiley & Sons, 2011. 3 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5 +[41] Runway. Gen-3. https://runwayml.com/research/introducing-gen-3-alpha, 2024.2 +[42] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. *ICLR*, 2023. 6 +[43] 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 +[44] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3 +[45] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5 +[46] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. In ICLR, 2025. 3 +[47] 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. 2 + +[48] Shuyuan Tu, Qi Dai, Zuxuan Wu, Zhi-Qi Cheng, Han Hu, and Yu-Gang Jiang. Implicit temporal modeling with learnable alignment for video recognition. In ICCV, 2023. 2 +[49] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. +[50] Shuyuan Tu, Qi Dai, Zihao Zhang, Sicheng Xie, Zhi-Qi Cheng, Chong Luo, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motionfollower: Editing video motion via lightweight score-guided diffusion. arXiv preprint arXiv:2405.20325, 2024. +[51] Shuyuan Tu, Zhen Xing, Xintong Han, Zhi-Qi Cheng, Qi Dai, Chong Luo, and Zuxuan Wu. Stableanimator: High-quality identity-preserving human image animation. In CVPR, 2025. 2 +[52] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5 +[53] Junke Wang, Yi Jiang, Zehuan Yuan, Binyue Peng, Zuxuan Wu, and Yu-Gang Jiang. Omnitokensizer: A joint image-video tokenizer for visual generation. In NeurIPS, 2024. 5 +[54] Junke Wang, Zhi Tian, Xun Wang, Xinyu Zhang, Weilin Huang, Zuxuan Wu, and Yu-Gang Jiang. Simplear: Pushing the frontier of autoregressive visual generation through pretraining, sft, and r1. arXiv preprint arXiv:2504.11455, 2025. 2 +[55] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Swap attention in spatiotemporal diffusions for text-to-video generation. IJCV, 2025. 2 +[56] 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. In NeurIPS, 2023. 5 +[57] Yanhui Wang, Jianmin Bao, Wenming Weng, Ruoyu Feng, Dacheng Yin, Tao Yang, Jingxu Zhang, Qi Dai, Zhiyuan Zhao, Chunyu Wang, et al. Microcinema: A divide-and-conquer approach for text-to-video generation. In CVPR, 2024. 2, 5, 6 +[58] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 2, 6 +[59] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 5 +[60] Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang Wang, Weizhu Chen, Mingyuan Zhou, et al. Patch diffusion: Faster and more data-efficient training of diffusion models. NeurIPS, 2024. 2, 3 +[61] Wenming Weng, Ruoyu Feng, Yanhui Wang, Qi Dai, Chunyu Wang, Dacheng Yin, Zhiyuan Zhao, Kai Qiu, Jian-min Bao, Yuhui Yuan, et al. Art-v: Auto-regressive text-to-video generation with diffusion models. In CVPRW, 2024. 2 + +[62] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. In ICLR, 2025. 2, 3 +[63] 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, 5, 6 +[64] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Simda: Simple diffusion adapter for efficient video generation. In CVPR, 2024. 2 +[65] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 2024. 2 +[66] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 5 +[67] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 2, 3 +[68] 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. In ICLR, 2025. 2, 4 +[69] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast autoregressive video diffusion models. In CVPR, 2025. 3 +[70] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In CVPR, 2023. 3 +[71] Sihyun Yu, Weili Nie, De-An Huang, Boyi Li, Jinwoo Shin, and Anima Anandkumar. Efficient video diffusion models via content-frame motion-latent decomposition. In ICLR, 2024. 3 +[72] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In CVPR, 2024. 2, 4 +[73] 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, 6 +[74] Hui Zhang, Zuxuan Wu, Zhen Xing, Jie Shao, and Yu-Gang Jiang. Adadiff: Adaptive step selection for fast diffusion. In AAAI, 2025. 3 +[75] 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 +[76] 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. 5 +[77] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. In ICLR, 2025. 2 + +[78] 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. arXiv preprint arXiv:2412.20404, 2024. 2, 4, 5 \ No newline at end of file diff --git a/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/images.zip b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bd5b14022e00670ef1f5814aacfde402245ebdd9 --- /dev/null +++ b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b88639c29fe55314bf7e0e13c30d0d0a1c84eb6858f4e2e387390726f81b605 +size 798470 diff --git a/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/layout.json b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a530a6670eb77e137e8c8de652fc498b8fb291a0 --- /dev/null +++ b/ICCV/2025/REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:104b2ff905bb0c6f016d20283e776f2b752cfa56d653964f90a8867583907277 +size 436992 diff --git a/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/fc895970-d957-4614-a966-4a644fd27e5a_content_list.json b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/fc895970-d957-4614-a966-4a644fd27e5a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..82f943312f1ee2ee76ba36355edb23caf8fa493d --- /dev/null +++ b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/fc895970-d957-4614-a966-4a644fd27e5a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dee1e43364b050aeba80955193f02e4b10e0499576b259e9a33d62a6daef4d5 +size 83498 diff --git a/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/fc895970-d957-4614-a966-4a644fd27e5a_model.json b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/fc895970-d957-4614-a966-4a644fd27e5a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c917103cafa68411445fc666204d2bb97800ec4 --- /dev/null +++ b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/fc895970-d957-4614-a966-4a644fd27e5a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e403b3d088e6a716b1470d7c8b1c00bec29696c5416cc3f382457839e5f9e4f +size 97473 diff --git a/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/fc895970-d957-4614-a966-4a644fd27e5a_origin.pdf b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/fc895970-d957-4614-a966-4a644fd27e5a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..761e595032734643e0d29dd41a46fd76a1a4dac7 --- /dev/null +++ b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/fc895970-d957-4614-a966-4a644fd27e5a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9329b4b9487151e37cb4d1d468a6c130655d996fa2d2f0aa348e6daa07013444 +size 2660704 diff --git a/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/full.md b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2afad9f218abd1fd857d6df7e8a447bd2a8ae1c3 --- /dev/null +++ b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/full.md @@ -0,0 +1,365 @@ +# REGEN: Learning Compact Video Embedding with (Re-)Generative Decoder + +Yitian Zhang $^{1,2\dagger}$ Long Mai $^{1}$ Aniruddha Mahapatra $^{1}$ David Bourgin $^{1}$ Yicong Hong $^{1}$ Jonah Casebeer $^{1}$ Feng Liu $^{1}$ Yun Fu $^{2}$ + +$^{1}$ Adobe Research $^{2}$ Northeastern University + +https://bespontaneous.github.io/REGEN/ + +![](images/b6409c5a0ce7d3e22c497c4ee16148c20f858ed5eeca2b32d2961778fd26fffd.jpg) + +![](images/e487dc0b1bf894493aa7fda42a01e5d9f3bfde523f69907a04ff1f1981d447f3.jpg) + +![](images/313436a7199787ce6b5384aa6a4d8eaebdc67152a9089c0943e9f4d092e5d187.jpg) + +![](images/946619b6db01e072c689400f9a4709e3a146ba33bbbe7ea0e735e7fe140c59c4.jpg) + +![](images/d947da09e43a845baae4ca4e4cf5974fcc23e9cbc72a4d3a16b4716532e8c4c4.jpg) + +![](images/79694412527bc31b62132bbcf14ccf503a5458cc431551a46222b5350b8d5fe3.jpg) + +![](images/bfa989a7dd0369f9d59f25f70484a0d2e12dc698b9f4315784fc88e945889be2.jpg) + +![](images/6f986c4ff8c5e049e7c9c58e656a654804f3334845cbcef785209e8b582ee55e.jpg) + +![](images/811a69a2b35199ae8a93e981f25e93ccb7d74664e77c300a3ad4b348332d46b7.jpg) + +![](images/d0d48387eb7f26b6382c38c7ef13308b419cb9457ed73c7b2100eadd6669771c.jpg) + +![](images/f36b476174c8005d15fa89c7f57322ad69f1ac462a14dfe4cf1903b7754d113c.jpg) + +![](images/267800550215c749c839c979bcfaeca174c67a0cba221cad316b22d7a7a6850e.jpg) + +![](images/50e77940160936693a10677940d3555a85b79281de04aabc5224c3ccecc5bb3e.jpg) +Ours + +![](images/b9e51a009a2c2460521628df5bd559ab99bb93e27b755d7df888499e5a2b501f.jpg) + +![](images/9d89e78e102636bbfb107a913c7adec28ed3151a614cced38884dad09487d01d.jpg) + +![](images/ce8aa954cbf895afadf0f3dc9675e785b64366f8745bb678306054dfc2651d54.jpg) +Prompt: A sheep behind a fence looking at the camera. +Figure 1. Reconstruction (top) and text-to-video (T2V) generations (bottom) results at $32 \times$ temporal compression. This figure contains video results (for T2V results), best viewed with Adobe Acrobat Reader. Colored bounding boxes in reconstruction results denote the regions with the most difference where MAGVIT-v2 leads to clear artifacts in the tuning fork. + +![](images/66307e87b4f269c8c290e2dd49b1c765d462d5e512d1a5b140d9fc25391fae5e.jpg) +Prompt: Young woman with red hair and leafy crown. + +![](images/5eb49ee9b486152491d5df0bbbce2d111cb214bec104d409f68e4c2f2c37c92e.jpg) +Prompt: Time lapse at the snow land with aurora in the sky. + +# Abstract + +We present a novel perspective on learning video embedders for generative modeling: rather than requiring an exact reproduction of an input video, an effective embedder should focus on synthesizing visually plausible reconstructions. This relaxed criterion enables substantial improvements in compression ratios without compromising the + +quality of downstream generative models. Specifically, we propose replacing the conventional encoder-decoder video embedder with an encoder-generator framework that employs a diffusion transformer (DiT) to synthesize missing details from a compact latent space. Therein, we develop a dedicated latent conditioning module to condition the DiT decoder on the encoded video latent embedding. Our experiments demonstrate that our approach enables superior encoding-decoding performance compared to state-of- + +the-art methods, particularly as the compression ratio increases. To demonstrate the efficacy of our approach, we report results from our video embedders achieving a temporal compression ratio of up to $32 \times (8 \times$ higher than leading video embedders) and validate the robustness of this ultracompact latent space for text-to-video generation, providing a significant efficiency boost in latent diffusion model training and inference. + +# 1. Introduction + +Diffusion models [11, 34] have recently emerged as a major paradigm in generative content creation, achieving breakthroughs across various domains such as images [3, 12, 13, 28, 30], videos [25, 27, 31, 36, 42, 45], and audio [6, 14, 18, 21]. For efficiency, most modern diffusion models compress the input media into a compact latent space before modeling, commonly referred to as latent diffusion models (LDMs) [30]. A core component of the LDM is the embedding model (which we refer to as **embedder** in this paper), which typically follows an encoder-decoder architecture. The encoder transforms the raw data into a compact latent representation, while the diffusion models operate within this latent space and subsequently transform the data back to the raw space using the decoder. + +For video generation, state-of-the-art (SOTA) video embedder, exemplified by the continuous-token variants of the MAGVIT-v2 model [43], leverage architectures that enable spatiotemporal processing, offering compression in both spatial and temporal dimensions. Despite these advances, SOTA video embedders typically offer $8 \times$ spatial compression but only $4 \times$ temporal compression. + +Increasing compression is inherently challenging due to a fundamental trade-off between representational efficiency and reconstruction fidelity. At high compression ratios, the model struggles to retain sufficient information within the latent space to reproduce high-fidelity details accurately. We verify that adapting existing video embedders to support temporal compression beyond $4 \times$ is highly challenging without significantly degrading reconstruction quality. + +This paper presents an alternative perspective on learning video tokenizers tailored for latent diffusion modeling. We propose a generation-oriented view of latent representation learning. "We argue that, in the context of latent diffusion, the key desired property of the latent space is the ability to generate visually plausible content rather than to faithfully recover an input video." We propose transforming the traditional encoder-decoder video embedder into an encoder-generator. This generation-oriented perspective enables a more flexible approach to allowing high-compression in the latent space. By leveraging generative capabilities in the decoder, the encoder can focus only on preserving the essential semantic and structural information. The decoder then synthesizes realistic finer details. + +To this end, we introduce REGEN, an approach for learning a high-compression video embedder via a diffusion decoder with a resynthesis objective. REGEN utilizes a diffusion Transformer (DiT) [28] as the video decoder due to its exceptional video modeling capability compared to diffusion U-Net, as consistently demonstrated by impactful works [19, 25, 27, 45]. Our DiT decoder treats the latent features encoded from an input video sequence as the conditioning signal during the diffusion-based generation process. Both the encoder and the DiT decoder are jointly trained within the diffusion-based training framework to learn to generate the original video content. + +To connect the encoded latent features with the DiT decoder, we design a dedicated latent conditioning module. This module introduces a novel DiT conditioning mechanism that transforms the latent maps into a content-aware positional embedding, enabling strong conditioning signals for the decoder's diffusion process. Our proposed latent conditioning mechanism also naturally enables continuous-time decoding that supports not only reconstruction but also interpolation and extrapolation capabilities. More importantly, it also addresses the challenge of encoding and decoding for arbitrary resolutions and aspect ratios, unseen during training, which is difficult to achieve with DiT. + +In this work, we aim to enhance the compactness of the latent space beyond the current limits and focus on the axis of increasing the temporal compression of our embedder, keeping the spatial compression fixed at $8 \times$ . The contributions of the current work are as follows: + +1. We present the idea of using a Diffusion Transformer as a decoder in video modeling instead of the conventional VAE [17]-based encoder-decoder. We demonstrate that this modeling strategy allows a re-design of the learning objective in video encoding, allowing us to bypass the fundamental compression-reconstruction trade-off. +2. We introduce a novel latent conditioning module with a content-aware positional encoding formulation. This enables us to effectively turn the encoded latent features into spatiotemporal control signals for the Diffusion Transformer (DiT) decoder. Our decoder not only enables DiT to encode and decode videos of arbitrary aspect ratios and resolutions, but also supports even one-step sampling without utilizing external distillation. +3. We show that REGEN enables superior reconstruction performance compared to extending the VAE-based embeddings at high compression ratios. We also show that our highest-compression $(32\times$ temporal) latent space is friendly for text-to-video generation. + +# 2. Related work + +Video Embedders. Early video diffusion models adapted image LDM, using the same image-based embedder for per-frame latent feature extraction [1, 8, 10, 23, 32, 41]. This + +![](images/f224ed40e9ccd0379f8996c72f1713ad3dd69870ed87537ae6697009b873c432.jpg) +Figure 2. Overall framework. Our spatiotemporal video encoder $E(\cdot)$ encodes the input video $x_{input}$ into two latent frames, content and motion $(z_c, z_m)$ . They are processed by the latent expansion module $C_e$ and serve as conditioning for the generative decoder. + +frame-wise approach neglects temporal relationships, causing inconsistencies and restricting the compactness of the video latent. Recent video diffusion frameworks utilize spatiotemporal embeddings for joint processing. The seminal work of MAGVIT-v2 [43] employs causal 3D convolutions for compression across both spatial and temporal dimensions. Originally designed for discrete tokenization, MAGVIT-v2 has been adapted for continuous-token video encoding, facilitating its integration into various prominent video diffusion models [19, 42, 45]. Despite their effectiveness, they typically offer limited temporal compression (e.g., $4 \times$ ). We aim to enhance the compactness of the latent space beyond the current limits, extending up to $32 \times$ temporal compression. To accomplish it, we offer a novel perspective of using encoder-generator with a (re-)generative diffusion decoding process to escape the reconstruction-compression trade-off. + +Diffusion Autoencoders. Our work draws inspiration from recent studies that demonstrate the potential of image diffusion models to generate content conditioned on abstract feature vectors extracted from conditioning images [2, 15, 29]. Specifically, [29] shows that an image diffusion model can be semantically guided, akin to style-code manipulation in StyleGAN [16]. Similarly, [2] validates the improvement brought by generative decoders, and [15] highlights the ability to reconstruct images from conditioning feature vectors. We adopt this latent-conditioning diffusion concept to the context of video embedder learning. Complementary to them, to the best of our knowledge, our work is the first to explore diffusion transformer autoencoders in the context of learning highly compact latent spaces for videos. + +# 3. Method + +Our model consists of two major components. First, the spatiotemporal video encoder that projects the input video sequence into a compact latent space (Section 3.1). The second is the DiT-based generative decoder with that converts the latents back to pixels by taking the latents as a conditioning signal (Section 3.2). The whole model is trained + +from scratch in an end-to-end fashion with the diffusion objective. Figure 2 illustrates our overall framework. + +# 3.1. Spatiotemporal Video Encoder + +The goal of our encoder is to encode a video into a compact latent space. Inspired by MAGVIT-v2 [43], we adopt the continuous version of their encoder design due to its ability to encode both images and videos in consistent latent space. + +The encoder is composed of multiple 3D convolution blocks, which are causal in the temporal dimension. For a video of $T + 1$ frames with dimensions $H \times W$ , at a spatial downsampling factor of $m$ , and a temporal downsampling rate of $k$ , due to the causal nature, the encoder produces a sequence of $1 + \frac{T}{k}$ latent feature maps each of dimension $\frac{H}{m} \times \frac{W}{m}$ . To provide flexibility in consistently encoding long videos without going out-of-memory, we follow a chunk-wise encoding scheme that encodes videos of fixed-length frames (a chunk). Thus, at a desired temporal compression ratio $k$ , we encode any input video chunk $x_{input}$ of length $k + 1$ into two latent frames: + +$$ +z _ {c}, z _ {m} = E \left(x _ {\text {i n p u t}}\right), \tag {1} +$$ + +where $z_{c}$ and $z_{m}$ denote the resulting two latent maps. As the encoder is causal in nature, $z_{c}$ only contains information from the first frame. We call this the content latent frame. $z_{m}$ , contains the compressed motion information of the rest of the frames. We refer to this as the motion latent frame. We use 8 latent channels for both $z_{c}$ and $z_{m}$ . + +# 3.2. Diffusion Transformer Decoder + +As opposed to the conventional convolution-based decoder, we use a transformer [38] and model the decoding task as a conditional diffusion process. Given the input video sequence $x_{input}$ and its corresponding content and motion latents $[z_c, z_m]$ , the diffusion decoder $G_d$ is trained to (re-)generate a target video output $x_{target}$ from a noise map sequence $N$ , conditioned on the latent maps $[z_c, z_m]$ . For the reconstruction task, $x_{target}$ is chosen to be the same as $x_{input}$ . It's worth noting that this formulation inherently + +![](images/aef3a829e4af399f3396edafa8d616bfbe97815d5d92a1352a4f61c810f669a7.jpg) +Figure 3. Latent conditioning module $C_e$ . The SIREN network $M_t$ maps the time coordinate $t_f$ to a feature vector modulated by the motion latent $z_m$ . The resulting feature is concatenated with the feature value of $z_c$ at the corresponding spatial coordinate $(x, y)$ . The concatenated feature is mapped into the DiT hidden dimension by the projector $M_s$ . We utilize the first frame prediction from SIREN to replace the first frame of expanded $z_c$ to ensure consistent representation for both image and video inputs. + +allows for a more flexible definition of $x_{target}$ to handle different tasks beyond pure reconstruction. For example, setting $x_{target}$ to be the temporally upsampled version of $x_{input}$ corresponds to modeling the temporal interpolation, while setting $x_{target}$ to be a time-shifted version of $x_{input}$ leads to the temporal extrapolation. + +During training, we corrupt the clean data $x_{target}$ (denoting as $x_{target}^{0}$ for diffusion noise timestep $t = 0$ ) with the forward diffusion process: + +$$ +q \left(x _ {\text {t a r g e t}} ^ {t} \mid x _ {\text {t a r g e t}} ^ {0}\right) = \mathcal {N} \left(x _ {\text {t a r g e t}} ^ {t}; \sqrt {\alpha^ {t}} x _ {\text {t a r g e t}} ^ {0}, (1 - \alpha^ {t}) I\right), \tag {2} +$$ + +where $t$ is the diffusion time step, and $\alpha^t$ is determined via the noise scheduling procedure [22]. The sample at time step $t$ is obtained with the reparameterization trick: + +$$ +x _ {t a r g e t} ^ {t} = \sqrt {\alpha^ {t}} x _ {t a r g e t} ^ {0} + \sqrt {1 - \alpha^ {t}} \epsilon , \tag {3} +$$ + +where $\epsilon \sim \mathcal{N}(0,I)$ . In the reverse denoising process, the generative decoder learns to invert the forward corruption under the condition $[z_c,z_m]$ . Specifically, a denoising model $\epsilon_{\theta}$ with parameters $\theta$ is trained to predict the noise $\epsilon$ at each time step $t$ given the corresponding noisy version $x_{target}^{t}$ , conditioning on $[z_c,z_m]$ with the simplified learning objective: + +$$ +\mathcal {L} (\theta) = \left\| \epsilon - \epsilon_ {\theta} (x _ {t a r g e t} ^ {t}, [ z _ {c}, z _ {m} ]) \right\| ^ {2}, \qquad (4) +$$ + +Under this formulation, we train our video encoder along with the generative decoder in an end-to-end fashion. Our diffusion decoder leverages a DiT architecture for the denoiser backbone due to its superior modeling and scaling capabilities compared to U-Net as pointed to many prior works [19, 25, 27, 45]. Our decoder operates in pixel space, at a patch size of $p$ . The patch size $p$ is determined according to the spatial down-sampling ratio in the encoder, which in our experiments is $8 \times$ . Thus, we keep $p = 8$ . + +Latent Conditioning via Content-Aware Positional Encoding. The transformer model for video data usually incorporates spatiotemporal positional encoding (PE) of the space-time coordinates as additional inputs. PE is critical not only for addressing the order-invariant nature of transformers but also for capturing the spatiotemporal structure of the generated content. Existing DiT models often adopt a fixed PE scheme where the coordinate-to-embedding mapping. Such a fixed PE scheme struggles with generalizing to input sizes unseen during training [42]. This is particularly problematic in our context, where we leverage DiT as the decoder since the model after training should be applied to decode the latents extracted from inputs of different resolutions and aspect ratios. + +We introduce a novel conditioning mechanism to address this challenge. The key idea is to generate the positional embedding from the conditioning latent $[z_c, z_m]$ instead of using a fixed spatial PE. Specifically, we develop a latent expansion module $C_e$ that takes the condition $[z_c, z_m]$ as inputs and expand it to the expanded latent $z_e$ same spatiotemporal dimension as the target output $x_{\text{target}}$ . $z_e$ is then added to the token embedding and timestep embedding and input into the DiT. In this way, $C_e$ operates as a mini-decoder that decodes the latent into a full spatiotemporal form, and $z_e$ operates as a content-aware positional embedding that participates in controlling the spatiotemporal aspects in the synthesized videos. + +As a positioning encoding module, $C_e$ is designed to map the coordinates $(x,y,t_f)$ of a token location (after patchification) in DiT input space into an embedding vector where the mapping is conditioned on $[z_c,z_m]$ . We defined such mapping as: + +$$ +C _ {e} \left(x, y, t _ {f} \mid \left[ z _ {c}, z _ {m} \right]\right) = M _ {s} \left(z _ {c} (x, y) \circledast M _ {t} \left(t _ {f} \mid z _ {m} (x, y)\right)\right) \tag {5} +$$ + +where $\star$ denotes the channel-wise concatenation operation. Note that the $(x,y)$ coordinates in the DiT input space match with those in the feature maps $[z_c,z_m]$ as we matched the patchification's patch size with the spatial downsampling factors. $M_{s}$ is a projector to align the channel dimension, consisting of a single linear layer and one RM-SNorm [44]. $M_{t}$ is a neural network sub-module that maps the time coordinate $t_f$ into a feature vector. Inspired by the success of implicit neural representation in modeling video data [5, 24], we implement $M_{t}$ with the SIREN [33]. We condition the mapping in $M_{t}$ on the motion feature $z_{m}$ by modulating the SIREN network with the feature values of $z_{m}$ at the queried $(x,y)$ location. Thus, the positional information is integrated into the expanded latent so that we can entirely remove the original spatial and temporal positional embedding in DiT, enabling our generative decoder to be generalize across arbitrary resolutions and aspect ratios. + +
MethodCompressionMCL-JCVDAVIS 2019
256×256512×512256×256512×512
PSNRSSIMrFVD↓PSNRSSIMrFVD↓PSNRSSIMrFVD↓PSNRSSIMrFVD↓
● MAGVIT-v28×8×826.610.718105.7229.140.77172.0722.820.602183.5224.750.660125.03
● REGEN28.820.78585.3732.740.84629.8826.000.711152.4629.340.77889.98
● MAGVIT-v28×8×1625.060.672205.7526.620.717185.6920.620.527441.2421.210.572417.43
● REGEN27.270.736174.2930.410.79892.4823.850.635328.8326.270.699235.13
● MAGVIT-v28×8×3222.970.573536.01†††18.230.4191080.15†††
● REGEN26.050.695265.9628.710.758224.5622.200.575488.8923.490.625522.20
+ +Table 1. Reconstruction comparison at high temporal compression. We compare our method, REGEN, with MAGVIT-v2 at different compression rates on MCL-JCV and DAVIS 2019 datasets. The best results are bold-faced. $\dagger$ MAGVIT-v2 $(32\times)$ faces out of memory issue at $512\times 512$ , due to the 3D convolution layers in decoder. + +![](images/1b5852542a045c2531dd211b267c80aa7fe494e51d416ec4b82e5d3dd829c88d.jpg) +Figure 4. Effectiveness of REGEN at high temporal compression. MAGVIT-v2 suffers from strong temporal artifacts in regions of high motion, such as the dog's face (left) and the toy (right). Areas enclosed in boxes show regions of maximum difference. + +# 4. Experiments + +We evaluate our encoder-generative decoder framework in three aspects: (1) effectiveness of generative decoder at various temporal compression ratios, and (2) the compatibility of our compressed latent space for video generation. + +# 4.1. Effectiveness of REGEN as Video Embedder + +Implementation Details. We configure our DiT-based decoder to match the model latency of MAGVIT-v2 [43]: it consists of 24 Transformer blocks, with 16 heads and a hidden dimension of 2048 using full satiotemporal self-attention. We set the patch size to be 8 to accommodate the costs. For each compression ratio, we train our model for $\sim$ 100K iterations under the scenarios of reconstruction, interpolation, and extrapolation objectives. Our encoder shares the same architecture with our implemented MAGVIT-v2. Please refer to the supplementary material for details. + +Evaluation. Following MAGVIT-v2, we evaluate the reconstruction quality on 2 benchmarks: MCL-JCV [39] and + +DAVIS 2019 (full resolution) [4]. In both datasets, videos are rescaled and center-cropped to $256 \times 256$ and $512 \times 512$ resolution for evaluation. We use standard quantitative metrics, PSNR, SSIM, and Fréchet video distance (rFVD) [9, 37], to examine the reconstruction quality. This covers both pixel-based and perceptual quality metrics. + +Baselines. At $8 \times 8 \times 4$ compression, we compare REGEN with SOTA 8-channel video embedders since the number of latent channels greatly affects reconstruction quality. We compare against OmniTokenizer [40], VidTok [35] and, WF-VAE [20] and MAGVIT-v2 [43]. Since the MAGVIT-v2 weights are not released, we implement it and train it on our dataset with over 200K iterations. For higher compression experiments, we adapt MAGVIT-v2's design to handle different temporal compression ratios. Note that we adjust the location of upsampling layers in MAGVIT-v2 decoder at $8 \times 8 \times 32$ to avoid out of GPU memory during training and we follow the original design in other experiments to ensure the optimal reconstruction quality. + +REGEN's effectiveness at high compression. We in- + +![](images/f3aef45880ee014ab40dc36d14a17d497a8e104d343e0ed7505e3f254dbecff8.jpg) + +![](images/efcbb5c55d449dee0e5a488b15435498b296adb1bcc68a858bc9746d4d9214fc.jpg) + +![](images/1ab85adc949c2ec26a77da1868682948ead0a6929a5f0046acabbcf73d22f904.jpg) + +![](images/1b66bb46af4ea03b1002e4c75e147b73cfb531f9e3e71644d2e0faed954b13da.jpg) + +![](images/937fd55f1fece9c68e2cbae6224c76ba8071648c576d24d6cfd6769ab35fbaba.jpg) + +![](images/b29cc01fa941486450901275dc468f527361b1efa8b3bde674a030da66c587d3.jpg) + +![](images/a13be0e32319fd02c5a17224a2cc62faff85d74f00162ad8acfe0392809f41e7.jpg) + +![](images/a27502f7c2670a725c2ef3c4d0ac03b3edbabbec9061d5cd99176a00771ce617.jpg) + +![](images/f30a149302170fd0c9d48483235d8ef7831ca891269c5ba57046604343925a28.jpg) + +![](images/d211ec9a41d17425187f3a4b6bbe4bbb88d0fef916ff0a201916d67ebbc8afc3.jpg) + +![](images/82877f54ebb932e2550140a474ccb594ce49ae14d58bee4660029ff923cc6d62.jpg) + +![](images/98215034806ef8994c9c355c2a1e91a66a0a5ba87898532e707854aba3c6eb50.jpg) + +![](images/e034a3435436546bbf56634ef6197f00c61b7b383e4fe3e726c5b4598750760e.jpg) +Figure 5. Effectiveness of REGEN at base $4 \times$ temporal compression. Current video embedders suffer from ghosting artifacts for videos with large motion, especially in faces (last row). REGEN performs well and retains plausible spatiotemporal structures from the input. + +![](images/34799ee64169b1c630449d9d0d2c3ae97fc3539d8fa03a6aabd830e67a0a3ca8.jpg) + +![](images/e8146270b4949ba4fc92c7b0b734d8a0ab51ba6a1ee90e733c1033c8385cfde4.jpg) + +![](images/bae9b0bf2e8eb51b1919fd95c80d65362fb4cb46200bfa820cd4551b833d9a8b.jpg) + +![](images/4c86305abae726ba4cc46a406adb3230de41fdf8255119834f121840b38e8c24.jpg) + +![](images/ed5f8c5e5e956dd058ae265d64eee8788f0bb2caba685696c0d25388346b8372.jpg) + +
MethodMCL-JCVDAVIS 2019
PSNRSSIMrFVD↓PSNRSSIMrFVD↓
• Omni24.630.71093.3523.390.628152.01
• WF-VAE31.000.80455.0127.950.737107.67
• VidTok32.060.83638.8528.670.76067.24
• MAGVIT-v231.490.82928.6328.160.75856.46
• REGEN32.940.85722.4030.250.80148.38
+ +Table 2. Reconstruction comparison at base $4 \times$ temporal compression. We compare REGEN with various SOTA 8-channel video embedders at $4 \times$ temporal compression on MCL-JCV and DAVIS 2019 datasets under $512 \times 512$ inputs. The best results are bold-faced and the second best results are underlined. + +vestigate our primary objective, whether a diffusion-based encoder-generator framework can learn a very compact latent space with a high degree of temporal compression that can faithfully reconstruct the input much more effectively than traditional video embeddings. As shown in Tab. 1, MAGVIT-v2 lags behind our method on all metrics at different temporal compression rates. Moreover, the advantage of REGEN over MAGVIT-v2 widens with the increasing temporal compression ratios, which suggests that naively extending conventional video embeddings to high compression rates cannot deliver satisfying results and highlights the generation ability of our decoder. Shown in Fig. 4, MAGVIT-v2 suffers from significant degradation in reconstruction quality and often exhibits visual artifacts at high temporal compressions. In contrast, our method maintains the reconstruction ability to a good extent. Moreover, our decoder is more efficient in terms of GPU memory when decoding long sequences compared to MAGVIT-v2, which + +is quite important in practical applications. + +Comparison with SOTA video embedders at base $4 \times$ temporal compression. Although our goal is to strive for higher temporal compression instead of SOTA performance at $8 \times 8 \times 4$ compression, we compare REGEN with existing 8-channel video embedders in Tab. 2 to validate the effectiveness of our method. We observe that our implemented MAGVIT-v2 achieves compelling performance compared to recent SOTA methods VidTok [35], WF-VAE [20], and OmniTokenizer [40], which proves the soundness of our experimental setup. Nevertheless, all the approaches lag behind our method in all metrics, and the qualitative comparison. In Fig. 5, we see that REGEN preserves better spatiotemporal structures from the input, especially in the human faces, which other convolution-based embedders, designed explicitly for $4 \times$ compression ratio, fail to reconstruct accurately. We further provide a holistic comparison with other video embedders that have different latent channels in supplementary material. + +# 4.2. Text-to-Video Generation + +The previous experiments demonstrate that the compact video embeddings from our methods can achieve promising performance for video reconstruction. We now verify whether our compact latent space is friendly to text-to-video generation, which is our end objective for learning a compact latent space. To validate it, we train a 5B DiT-based latent diffusion model on the ultra-compact latent with $32 \times$ temporal compression for text-to-video generation. The design of the text-to-video model is based on + +![](images/3d3b9d1c3a511ad56ffda739a079aac5e92a6c20ab4ae68b2b3120407692b193.jpg) +Prompt: A slow cinematic push in on an ostrich standing in a 1980s kitchen. + +![](images/f499251cecb5d37d86bfb70ec985744118a42a89294b0881d45a447954d135e4.jpg) + +![](images/beadd705bfc650ddb208cfffd5cc15d705d93870d7a0129c1cbb41463a3e9557.jpg) + +![](images/e67b75d39a21be596adbce0a46a7488fdc2a1c3cfb3e06091850d62e2c12a086.jpg) + +![](images/5a8e50d4413ea39d6a791c23af1effcedfcc0c617a9955c7269e7ecd48ac3b0b.jpg) + +![](images/75826a13f12cf8caed764ce402f7c877e58a5cb465ba107f4795d51ce7996fec.jpg) +Prompt: A stunning aerial drone footage time lapse of El Capitan in Yosemite National Park at sunset. + +![](images/383d5c4996dd49ab66f6178f8ec7072abd50fb1d0e3319a0f42009b27ca7ae94.jpg) + +![](images/5a0d326f50e6e07efb75787a18bbad536fe8b3e0d2d264b32c6b87ce14f7d5b1.jpg) + +![](images/ef785c0100137aec3736f8c7ac6ce8fa7d18916dbb272831d7dfe716e6fe7adc.jpg) + +![](images/cb5a66632a07d74d7e08b35d3147b05b4b2602bd76d0f03ac4037225b1340384.jpg) + +![](images/ebd3307eccae8469fdd98469a18381cdc5b9c891ff19cb9cfa9fc23134b0c88f.jpg) +Prompt: A cinematic documentary hand held close up of a woman standing in a busy Italian plaza smirking to herself. + +![](images/d7f5118611134293ec06e0e7ccfd39ed4145bd6934f269eca6bd373d69edf57c.jpg) + +![](images/2f5555bf247c3a3f0ce9d79dca0605229b24dc7ea01b78a7d2c3f38fb6f1491a.jpg) + +![](images/1e1260d1924e6b73dc94a48e895b677997607af48e46320b923edec5e3dfbe23.jpg) + +![](images/cd9226d977573ecfedd45d0a3cef3c5cbb44da23bc35f1aeb9a9c6a0fe9cdfb1.jpg) + +![](images/ab661580ab484f1f7c874115531159caca0bb35559c8cb4aba6cfb354d742f73.jpg) +1 Step + +![](images/c71835e0f62cf0545c364b30ce4289f9860afc0e18126bb4cdcf262a65a33fd7.jpg) +10 Step + +![](images/02a281845f7b515a89e9068e38bdf6aba3041feb8cde7d9a8fe7153c414aa863.jpg) +50 Step + +![](images/85fb92b5fc37c48eb129a95e5f1cf17c6e487c551461d7d86f4ff9475cf6ad6a.jpg) +100 Step + +![](images/5925081c815156cf5cf60e6ac021ccb51d1a00c045c19edc46af1b261ebd5473.jpg) +(a) Example reconstruction results at different sampling steps. +(b) Quantitative results on MCL-JCV. +Figure 7. Reconstruction results at different sampling steps. Dash lines in (b) and (c) denote the PSNR and rFVD of MAGVIT-v2. Our method exhibits strong performance across varying sampling steps and supports few-step and even single-step sampling. + +![](images/c070bf2b2727418a02f29976a7261eab40763c56280cf6ceeb2d99967e748a8d.jpg) +Figure 6. Text-to-video generation results on our ultra-compact latent space with $32 \times$ temporal compression. The latent diffusion model generates 132-frame videos with only 8 latent frames on our ultra-compact latent space, offering $\sim 5 \times$ reduction in the number of latent frames compared to current video embeddings at $4 \times$ temporal compression. +(c) Quantitative results on DAVIS 2019. + +MMDiT [7]. We first train the model with image inputs for 260K iterations, and the model is further trained with mixed image and video inputs for another 80K iterations. We show samples of the generated videos in Fig. 6. Note that due to constraints in computing resources, we trained the model at a relatively small scale with a limited number of iterations. Nevertheless, even with small-scale training, we validate that the diffusion model is capable of generating plausible video content on a much more compact latent space, promising the application of our method in video generation. Notably, the latent diffusion model generates 132-frame videos with only 8 latent frames on our ultracompact latent space $(32\times$ temporal compression), offering $\sim 5\times$ reduction in the number of latent frames compared to current video embedders that operate with $4\times$ temporal compression, thus reducing the training and inference costs remarkably. Please find more results in the supplementary. + +# 5. Discussion + +In this section, we address two of the main concerns with the practical usability of our transformer-based diffusion model for embedders: (1) the diffusion model inherently + +requires many denoising steps at inference, making it very costly, and (2) transformers are difficult to generalize to resolutions and aspect ratios unseen during training. We also provide a method to alleviate the chunking issue by leveraging the extrapolation ability of our decoder. + +Few-step and one-step sampling. One of the main bottlenecks with the diffusion model is that it requires a number of denoising steps, even at inference, to achieve good-quality results, which causes a lot of overhead for practical use. However, our decoder is in fact tasked with an easier generation problem compared to text-to-video generation as it is equipped with a very strong conditioning signal provided by our latent conditioning module. Thus, we perform inference on the same model with different numbers of steps to inspect how significantly the reconstruction quality degrades as the number of sampling steps reduces. Fig. 7a shows reconstruction examples obtained at different sampling steps using our $4 \times$ compression model. Interestingly, we observe that it is possible to reduce the number of sampling steps significantly (even down to 1-step sampling) without noticeable visual quality degradation. We further evaluate our model at different sampling steps on all test videos from DAVIS 2019 and MCL-JCV datasets. + +![](images/8e8d279809c2965b0a859c27ce8f9275efa420e9239783a8daccb35fbbe9f5e2.jpg) +Figure 8. Video frame reconstruction at different spatial resolutions. The models are trained under $192 \times 320$ input and evaluated at the resolution of $192 \times 320$ (pink bounding boxes) and $384 \times 640$ (blue bounding boxes). In-context conditioning results in gridding artifacts at a larger resolution, while our method exhibits strong generalization due to the proposed content-aware PE. + +![](images/a805328ea0b8126a417b5dca27d414871ab552aee8f3009517e2b8223075bda8.jpg) + +
Method192×320384×640
PSNRSSIMrFVD↓PSNRSSIMrFVD↓
In-context25.710.709135.8923.390.587441.98
Ours26.040.720128.8029.410.78557.01
+ +Table 3. Ablation of conditioning mechanism on DAVIS 2019 dataset. In-context conditioning exhibits inferior performance compared to ours at the training resolution and cannot generalize well to unseen resolutions. The best results are bold-faced. + +Fig. 7b and Fig. 7c suggest that more sampling steps lead to slightly worse PSNR and better rFVD scores, which can be attributed to the increase in sharpness in the reconstructed videos. To conclude, our diffusion decoder offers a flexible decoding scheme and can even act like a feed-forward model without utilizing external distillation, promising its application for practical purposes. + +Ablation of conditioning mechanism. A significant challenge with using a transformer as a decoder is that conventional transformers with fixed positional encoding (PE) can not generalize to unseen resolutions at inference. In contrast, our content-aware positional encoding makes REGEN inherently flexible to varying aspect ratios and resolutions at inference, unseen during training. We verify that with an experiment where we replace our condition module with the in-context conditioning design in DiT [28]. This conditioning scheme retains the fixed-PE scheme like other conventional conditioning designs and performs conditioning via concatenating the conditioning tokens (i.e. the encoded latents $[z_c, z_m]$ in our context) with the input tokens along the sequence dimension. Fig. 8 shows that our method generalizes well to different resolutions. In contrast, in-context conditioning leads to strong gridding artifacts when evaluating at higher resolutions not seen during training, severely impairing its functionality as a video decoder. The quantitative results in Tab. 3 further validate this. + +Alleviating the chunking issue. Like other conventional video embedders, REGEN sometimes exhibits slight jumps across the intersection of two chunks due to the chunk-wise encoding scheme, which is widely adapted to process arbitrary video lengths. To mitigate this issue, we leverage the extrapolation capability of the decoder to generate the last + +![](images/f34f141af891d8aa5bd42d473cef40b5b2fede473a67e404c4946f1e2f4716aa.jpg) +Figure 9. Alleviating the chunking issue with latent extension. The x-t slice is obtained by extracting a short segment (shown as the red line in the video frame) from 2 chunk frames and latent extension offers a smoother transition across the chunk boundary. + +chunk frames based on the latent of the current chunk. Following SDEdit [26], we utilize the predicted last frame from the previous chunk to guide the generation of the next chunk so that we can better align consecutive chunks and reduce the jumping in an autoregressive manner. Figure 9 demonstrates that our latent extension strategy mitigates the jumping and offers a smoother transition across the chunk boundary. Note the visible horizontal line in the middle time-slice visualization in the Vanilla decoding, which becomes much smoother when the latent extension is applied. + +# 6. Conclusion and Limitation + +In this work, we provide a new perspective on constructing the latent space for video generation. We propose that a good latent space needs to be able to generate visually plausible content that faithfully recovers an input video without needing to match the input to the pixel-perfect level of detail. To this end, we present an encoder-generator framework, instead of the commonly-used encoder-decoder framework trained with VAE objective, where we leverage a diffusion transformer as the decoder to ensure the latent space only captures essential semantic and structural features of the input video and the details are synthesized during decoding by the generative decoder. We show that this approach allows us to achieve a temporal compression ratio as high as $32 \times$ with reconstruction quality far surpassing existing state-of-the-art conventional video embedded trainers for such an aggressive level of compactness. + +Limitation. While our method enables highly compact video embedding, there still remain some challenges and future directions: (1) While one-step (or few-step) sampling is efficient, the training of diffusion-based decoders remains computationally expensive. Due to limited resources, we could not conduct extensive ablation on architecture configuration with the number of transformer blocks and other architectural details. Investigating this is a future direction. (2) Although our method of latent extension could mitigate the transition challenge between chunks to great extent, it cannot fully remove the jumping issue in such a training-free manner. (3) Our decoder operates in the pixel space, requiring a large patch size for efficiency (8 in our case), which can degrade generation quality [28] compared to the commonly used patch size of 1 or 2. + +# Acknowledgment + +We thank Richard Zhang, Zhifei Zhang, Jianming Zhang, Gunjan Aggarwal, Kevin Duarte, Yang Zhou, Zhan Xu, and Lakshya for several fruitful discussions. + +# References + +[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2 +[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 +[3] blackforestlabs. Flux. https://blackforestlabs.ai/, 2024. Accessed: 2024-XX-XX. 2 +[4] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737, 2019. 5 +[5] Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vedit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong Wang. Videoirr: Learning video implicit neural representation for continuous space-time super-resolution. In CVPR, 2022. 4 +[6] Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi, and Alexandre Défossez. Simple and controllable music generation. NeurIPS, 2024. 2 +[7] 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. 7 +[8] 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 +[9] Songwei Ge, Aniruddha Mahapatra, Gaurav Parmar, JunYan Zhu, and Jia-Bin Huang. On the content bias in fréchet video distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7277-7288, 2024. 5 +[10] 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 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2 +[12] Jonathan Ho, William Chan, Chitwan Saharia, 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 + +[13] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 2022. 2 +[14] Qingqing Huang, Daniel S Park, Tao Wang, Timo I Denk, Andy Ly, Nanxin Chen, Zhengdong Zhang, Zhishuai Zhang, Jiahui Yu, Christian Frank, et al. Noise2music: Text-conditioned music generation with diffusion models, march 2023a. arXiv preprint arXiv:2302.03917. 2 +[15] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthey, Felix Hill, and Alexander Lerchner. Soda: Bottleneck diffusion models for representation learning. In CVPR, 2024. 3 +[16] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 3 +[17] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 2 +[18] Felix Kreuk, Gabriel Synnaeve, Adam Polyak, Uriel Singer, Alexandre Défossez, Jade Copet, Devi Parikh, Yaniv Taigman, and Yossi Adi. Audiogen: Textually guided audio generation. arXiv preprint arXiv:2209.15352, 2022. 2 +[19] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2,3,4 +[20] Zongjian Li, Bin Lin, Yang Ye, Liuhan Chen, Xinhua Cheng, Shenghai Yuan, and Li Yuan. Wf-vae: Enhancing video vae by wavelet-driven energy flow for latent video diffusion model. arXiv preprint arXiv:2411.17459, 2024. 5, 6 +[21] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. arXiv preprint arXiv:2301.12503, 2023. 2 +[22] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In ICLR, 2023. 4 +[23] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. arXiv preprint arXiv:2303.08320, 2023. 2 +[24] Long Mai and Feng Liu. Motion-adjustable neural implicit video representation. In CVPR, 2022. 4 +[25] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. In CVPR, 2024. 2, 4 +[26] 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. arXiv preprint arXiv:2108.01073, 2021. 8 +[27] OpenAI. Sora. https://openai.com/index/video-generation-models-as-world-simulators/, 2024. Accessed: 2024-XX-XX. 2, 4 +[28] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 2, 8 + +[29] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 3 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[31] Runway. Runway gen-3: Next-generation video generation model. https://runwayml.com, 2023. Accessed: 2023-XX-XX. 2 +[32] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[33] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 4 +[34] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2 +[35] Anni Tang, Tianyu He, Junliang Guo, Xinle Cheng, Li Song, and Jiang Bian. Vidtok: A versatile and open-source video tokenizer. arXiv preprint arXiv:2412.13061, 2024. 5, 6 +[36] Kuaishou AI Team. Kling. https://kling.kuaishou.com/en, 2024. Accessed: 2024-XX-XX. 2 +[37] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5 +[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[39] Haiqiang Wang, Weihao Gan, Sudeng Hu, Joe Yuchieh Lin, Lina Jin, Longguang Song, Ping Wang, Ioannis Katsavounidis, Anne Aaron, and C-C Jay Kuo. Mcl-jcv: a jnd-based h. 264/avc video quality assessment dataset. In ICIP, 2016. 5 +[40] Junke Wang, Yi Jiang, Zehuan Yuan, Bingyue Peng, Zuxuan Wu, and Yu-Gang Jiang. Omnitokensizer: A joint image-video tokenizer for visual generation. NeurIPS, 2025. 5, 6 +[41] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. 2023. 2 +[42] 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, 3, 4 +[43] 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, 3, 5 +[44] Biao Zhang and Rico Sennrich. Root mean square layer normalization. NeurIPS, 2019. 4 + +[45] 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, 2024. 2, 3, 4 \ No newline at end of file diff --git a/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/images.zip b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ec045413000e42b199718fd18245b9cb8faaf1f0 --- /dev/null +++ b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa3bdb7cdb139857f8145b9012ea4d2db34016f5b25a77a11c0cae0ff935528f +size 892526 diff --git a/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/layout.json b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dff09216cf2fe28aaa3e31fab6e5c9ddb44e4f34 --- /dev/null +++ b/ICCV/2025/REGEN_ Learning Compact Video Embedding with (Re-)Generative Decoder/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51057eb29989a9f4fa3431e5ec70fdb8b76775b0bc7e978a416d0e509f279116 +size 478853 diff --git a/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/04874538-2450-4e14-8844-b1d1f738bbdb_content_list.json b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/04874538-2450-4e14-8844-b1d1f738bbdb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4e91c58fe7dc4e144acfb4abc03af61afb5141a5 --- /dev/null +++ b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/04874538-2450-4e14-8844-b1d1f738bbdb_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f11713a82084f65d568062cfef995ef133693ac2f01cc1844fb6005f958ba77 +size 83500 diff --git a/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/04874538-2450-4e14-8844-b1d1f738bbdb_model.json b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/04874538-2450-4e14-8844-b1d1f738bbdb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0e1f1afc3cd822c461ab4c1e0cce96df3a57f5a6 --- /dev/null +++ b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/04874538-2450-4e14-8844-b1d1f738bbdb_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27c432f3fa1c893ed6407db9f2ec717d846cca39c5390985802cd4ce3d968db0 +size 103377 diff --git a/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/04874538-2450-4e14-8844-b1d1f738bbdb_origin.pdf b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/04874538-2450-4e14-8844-b1d1f738bbdb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..72793e9bd85edcaae22c912ea56dd3e9715a42b0 --- /dev/null +++ b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/04874538-2450-4e14-8844-b1d1f738bbdb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98558f12db06f4f0a88fef663b7c5427147817947d02990d83ccea642dfdd136 +size 2932811 diff --git a/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/full.md b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..470b076601e8e44fadcc1cc0fb6513889cb4f965 --- /dev/null +++ b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/full.md @@ -0,0 +1,307 @@ +# REPA-E: Unlocking VAE for End-to-End Tuning with Latent Diffusion Transformers + +Xingjian Leng $^{\alpha \star}$ Jaskirat Singh $^{\alpha \star}$ Yunzhong Hou $^{\alpha}$ Zhenchang Xing $^{\beta}$ Saining Xie $^{\chi}$ Liang Zheng $^{\alpha}$ + +$^{\alpha}$ Australian National University $\beta$ Data61 CSIRO $\chi$ New York University $\{x$ ingjian.length*, jaskirat.singh*, yunzhong.hou, liang.zheng\}@anu.edu.au zhenchang.xing@data61.csiro.au saining.xie@nyu.edu + +![](images/1d8e324ca31d2b6bd5784df586a6aa8d8a02f27f80d38f1045fb8dcf329fc28d.jpg) +Figure 1. Can we unlock VAE for end-to-end tuning with latent-diffusion models? - Traditional deep learning wisdom dictates that end-to-end training is often preferable when possible. However, latent diffusion models usually only update the generator network while keeping the variational auto-encoder (VAE) fixed (a). This is because directly using the diffusion loss to update the VAE (b) causes the latent space to collapse. We show that while direct diffusion-loss is ineffective, end-to-end training can be unlocked through the representation-alignment (REPA) loss - allowing both encoder and diffusion model to be jointly tuned during the training process (c). Notably, this allows for significantly accelerated training; speeding up training by over $17\times$ and $45\times$ over REPA and vanilla training recipes, respectively (d). + +![](images/00cb03ff93682e09a30e873bce6640a775227d68934bbfe6177625059d88b133.jpg) + +![](images/b77fd795b276056047d93fc6d8c1c1d2c65db0de2a1fb6872a2d3bd0e526b60b.jpg) + +# Abstract + +In this paper we tackle a fundamental question: "Can we train latent diffusion models together with the variational auto-encoder (VAE) tokenizer in an end-to-end manner?" Traditional deep-learning wisdom dictates that end-to-end training is often preferable when possible. However, for latent diffusion transformers, it is observed that end-to-end training both VAE and diffusion-model using standard diffusion-loss is ineffective, even causing a degradation in final performance. We show that while diffusion loss is ineffective, end-to-end training can be unlocked through the representation-alignment (REPA) loss - allowing both VAE and diffusion model to be jointly tuned during the training process. Despite its simplicity, the proposed training recipe (REPA-E) shows remarkable performance; speeding up diffusion model training by over $17 \times$ and $45 \times$ over REPA and + +vanilla training recipes, respectively. Interestingly, we observe that end-to-end tuning with REPA-E also improves the VAE itself; leading to improved latent space structure and downstream generation performance. In terms of final performance, our approach sets a new state-of-the-art; achieving FID of 1.26 and 1.83 with and without classifier-free guidance on ImageNet $256 \times 256$ . Code is available at https://end2end-diffusion.github.io. + +# 1. Introduction + +End-to-end training has propelled the field forward for the past decade. It is understood that incorporating more components into end-to-end training can lead to increased performance, as evidenced by the evolution of the RCNN family [14, 15, 38]. With that said, training schemes of latent diffusion models (LDMs) [39] remain two-stage: first, the variational auto-encoder (VAE) [22] is trained with the re + +![](images/912e18b57fd33649ecf9c8a1b7e061430429d4c8b094dcc6635e5902074ca692.jpg) +(a) PCA Analysis on VAE Latent Space Structure + +![](images/36080a8ae8a0876c5c070de12a20150b73836f24ca8349b0cb6b71110f7afd5b.jpg) +(b) Performance Improvements with REPA-E (400K Steps) +Figure 2. End-to-End Training Automatically Improves VAE Latent-Space Structure. (a) Following [24], we visualize latent space structure from different VAEs before and after end-to-end training using principal component analysis (PCA) that projects them to three channels colored by RGB. We consider SD-VAE [39], and IN-VAE $^1$ , a $16 \times$ downsampling, 32-channel VAE trained on ImageNet [6]. For SD-VAE we find that latent representations have high-frequency noise. Applying end-to-end tuning helps learning a more smooth and less noisy latent representation. Interestingly to the contrast, the latent space for IN-VAE is over-smoothed (e.g., row-2). Applying end-to-end tuning automatically helps learn a more detailed latent space structure to best support final generation performance. (b) Jointly tuning both VAE and latent diffusion model (LDM) significantly improves final generation performance (gFID) across different VAE architectures. + +construction loss; then, the diffusion model is trained with the diffusion loss while keeping the VAE fixed (see Fig. 1a). + +The above two-stage division of the LDM training process, though popular, leads to a challenging optimization task: "How to best optimize the representation from first stage (VAE) for optimal performance while training the second stage (diffusion model)?" While recent works study the interplay between the performance of the two stages [24, 43], they are often limited to empirical analysis, which may vary depending on the architecture and training setting for both the VAE and the diffusion model. For instance, in a concurrent work [43] show that the latent space of popular autoencoders e.g., SD-VAE [39] suffer from high-frequency noise / components. However, as seen in Fig. 2 & 5, while the same holds for some VAEs (e.g. SD-VAE), it might not be true for other VAE architectures — which instead might suffer from an over-smoothed latent space (Fig. 2, 5). + +In this paper, we therefore ask a fundamental question: "Can we jointly tune both VAE and LDM in an end-to-end manner to best optimize final generation performance?" Technically, it is straightforward to do end-to-end LDM training by simply back-propagating the diffusion loss to the VAE tokenizer. However, experiments (§3) reveal that this naive approach for end-to-end training is ineffective. The diffusion loss encourages learning a simpler latent space structure which is easier for denoising objective (refer §3.1), but leads to reduced generation performance (Fig. 1d). + +To address this, we propose REPA-E; an end-to-end training recipe using representation alignment loss [52]. We show that while the diffusion loss is ineffective, end-to-end tuning can be unlocked through the recently proposed representation-alignment (REPA) loss - allowing both VAE + +and diffusion model to be jointly tuned during training process. Through extensive evaluations, we demonstrate that end-to-end tuning with REPA-E offers several advantages; End-to-End Training Leads to Accelerated Generation Performance; speeding up diffusion training by over $17 \times$ and $45 \times$ over REPA and vanilla training recipes (Fig. 1d). Furthermore, it also helps significantly improve the final generation performance. For instance as seen in Fig. 1d, we find that when using the popular SiT-XL [30] architecture, REPA-E reaches an FID of 4.07 within 400K steps, significantly boosting final performance over even REPA which only only reaches a final FID for 5.9 after 4M steps [52]. + +End-to-End Training improves VAE latent-space structure. As seen in Fig. 2 and §4.4, we find that jointly tuning the VAE and latent diffusion model during training, automatically improves the latent space structure across different VAE architectures. For instance, for SD-VAE [39], it is observed that the original latent space suffers from high-frequency noise (Fig. 2). Applying end-to-end tuning helps learn a more smooth latent space representation. In contrast, the latent space for IN-VAE1 is over-smoothed. Applying REPA-E automatically helps learn more detailed latent space structure to best support generation performance. End-to-End Tuning Improves VAE Performance. Finally, we find that once tuned using REPA-E, the end-to-end tuned VAE can be used as a drop-in replacement for their original counterparts (e.g. SD-VAE) showing improved generation performance across diverse training settings and model architectures (refer §4.4). + +To summarize, key contributions of this paper are: 1) We propose REPA-E; an end-to-end training recipe for jointly + +tuning both VAE and LDM using representation alignment loss (§3). 2) We find that despite its simplicity, REPA-E leads to accelerated generation performance; speeding up diffusion training by over $17 \times$ and $45 \times$ over REPA and vanilla training recipes, respectively (§4.2). 3) We show that end-to-end training is able to adaptively improve the latent space structure across diverse VAE architectures. 4) We demonstrate that once tuned using REPA-E, the end-to-end tuned VAE can be used as a drop-in replacement for their original counterparts (e.g., SD-VAE), exhibiting significantly better downstream generation performance (§4.4). + +# 2. Related Work + +Tokenizer or autoencoders (AE) [3] use either the variational objective [22] for continuous tokenization or a vector quantization objective [9, 47] for discrete tokenization [8-10, 16, 21, 22, 36, 39, 47, 51, 53]. However, current tokenizers are primarily trained for minimizing the reconstruction error, which maybe not provide the optimal latent space for generation [24]. We show that improved latent space structure is achieved by end-to-end training of LDMs. + +Latent diffusion models leverage pre-trained image tokenizers to compress images into a lower-dimensional latent space to simplify the generative task [5, 10, 10, 11, 26, 32, 36, 39, 42, 46]. Despite their effectiveness, existing tokenizers and diffusion models are trained separately [10, 36, 39]. In this paper, we explore jointly optimizing tokenizers and diffusion models to achieve faster convergence and improved generation performance (Sec. 4). + +Representation alignment for generative learning has recently shown huge promise for improving the training speed and performance of diffusion models [35, 48, 52]. We find that instead of applying the REPA loss separately over LDM [52] or VAE [48], significantly better performance and training speed can be achieved through E2E training. + +End-to-End Diffusion. LSGM [46] explores joint training with score-based generative models, which uses a variational lower bound objective with an entropy term for preventing latent space collapse while backpropagating the diffusion loss. We empirically find that while this helps prevent latent space collapse, REPA-E shows significantly faster convergence during E2E training (refer App. B). + +# 3. REPA-E: Unlocking VAE for Joint Training + +Overview. Given a variational autoencoder (VAE) and latent diffusion transformer (e.g., SiT [30]), we wish to jointly tune the VAE latent representation and diffusion model features in an end-to-end manner to best optimize the final generation performance. To this end, we first make three key insights in §3.1: 1) Naive end-to-end tuning - directly back-propagating the diffusion loss to the VAE is ineffective. The diffusion loss encourages learning a more simpler latent space structure (Fig. 3a) which is easier for min + +imizing the denoising objective [39], but degrades the final generation performance. We next analyze the recently proposed representation-alignment loss [52] showing that; 2) Higher representation-alignment score [52] correlates with improved generation performance (Fig. 3b). This offers an alternate path for improving final generation performance using representation-alignment score as a proxy. 3) The maximum achievable alignment score with vanilla-REPA is bottlenecked by the VAE latent space features. We further show that backpropagating the REPA loss to the VAE during training can help address this limitation, significantly improving final representation-alignment score (Fig. 3c). + +Given the above insights, we finally propose REPA-E (§3.2); an end-to-end tuning recipe for both VAE and LDM features. Our key idea is simple: instead of directly using diffusion loss for end-to-end tuning, we can use the representation alignment score as a proxy for the final generation performance. This motivates our final approach, where instead of the diffusion loss, we propose to perform end-to-end training using the representation-alignment loss. The end-to-end training with REPA loss helps better improve the final representation-alignment score (Fig. 3b), which in turn leads to improved final generation performance (§3.1). + +# 3.1. Motivating End-to-End Training with REPA + +Naive End-to-End Tuning is Ineffective. We first analyze the naive approach for end-to-end tuning; directly backpropagating the diffusion loss to the VAE tokenizer. As shown in Fig. 3a, we observe that directly backpropagating the diffusion loss encourages learning a more simpler latent space structure with lower variance along the spatial dimensions (Tab. 9). The simpler latent-space structure poses an easier problem for the denoising objective [39], but leads to reduced generation performance (Fig. 1). Consider an intermediate latent $z_{t} = \alpha_{t}z_{\mathrm{VAE}} + \sigma_{t}\epsilon_{orig}$ for any timestep $t$ . The denoising objective [34] mainly aims to predict $\epsilon_{pred}$ ; estimating the originally added noise $\epsilon_{orig}$ from VAE features $z_{\mathrm{VAE}}$ and timestep $t$ . As the variance along the spatial dimensions for VAE latent $z_{\mathrm{VAE}}$ goes down, the denoising objective effectively reduces to predicting a bias term for recovering back the originally added noise $\epsilon_{orig}$ . Thus, backpropagation the diffusion loss effectively hacks the latent space structure to create an easier denoising problem, but leads to a reduced generation performance (Fig. 1). + +Higher Representation Alignment Correlates with Better Generation Performance. Similar to the findings of [52], we also measure representation alignment using CKNNA scores [19] across different model sizes and training iterations. As seen in Fig. 3b, we observe that higher representation alignment during the training process leads to improved generation performance. This suggests an alternate path for improving generation performance by using the representation alignment objective instead of the diffusion loss for end-to-end training (refer §3.2). + +![](images/05b9f702935dc141adcc26d5d238550c17c209692ccdf2112f12f5048ca49900.jpg) +RGB +Image + +![](images/bf9b324ba4f0c293f481c0aedb4bcef8793a9edef400c252401ec845c9a542c0.jpg) +SDVAE +w/o E2E + +![](images/7fedd752eab54eb27f3e2d4f0c3268d8a71acbd6c3d9a291ca6f87e686230bd5.jpg) +E2E with +REPA Loss + +![](images/464299cb1a3ce59f45f6c1aea9806d17a3fbe79a65f18888dcf40788f029ec35.jpg) +E2E with +Diff, Loss + +![](images/87c33a980d6a2ef8f114377f442ca405b0538539dc0b1bf93c1f186a8c711039.jpg) + +![](images/a7f6da97577b21da144487b287331b69fe0b39774e345e1babcabed7b4e5a508.jpg) + +![](images/1e1087940c13bacc74f5c1f447614478b1ed3791d2d8842917179d95576cb8c5.jpg) + +![](images/2b696b9c613618a962d041c37f1527d11ceadf1b43401a3dde7df17b3f98036e.jpg) + +![](images/06736ab8e5a778deb6608503e27cc08cb1b10fdabdf2d48c5c23fec96457b5f4.jpg) +(a) PCA Visualization of Latent Spaces + +![](images/b5bc7b610ed92bc9896d201b899af4ae312a4eb6b9ea5f11eb437ab43070d1bb.jpg) + +![](images/466e108bd5f7a4e20612ce7fbe190eaa18625778830205199c9d649a96e5667e.jpg) + +![](images/aca45814db44dbc2b9a027cc7699549632877ae0aa529f3fe591767d1338a338.jpg) + +![](images/29cb10bc83c7a2590d61f611c4c31fac0067c4a445cd1e9f5abc3eac2bf2fe00.jpg) + +![](images/8ce92ff348b9c69930efd541887a95c7c1f4d66016cc89812559a1579bcc1040.jpg) +(b) Correlation: gFID & CKNNA Score +(c) E2E tuning with REPA improves CKNNA Score +Figure 3. Motivating End-to-End Tuning using Representation Alignment (REPA) Loss. We make three key insights: 1) Naive end-to-end (E2E) tuning using diffusion loss is ineffective. The diffusion encourages learning a more simpler latent space structure (a) which is easier for denoising objective (refer §3.1) but degrades final generation performance (Fig. 1). We next analyze the recently proposed representation alignment (REPA) loss [52] showing: 2) Higher representation alignment (CKNNA) leads to better generation performance. This suggests an alternate path for improving performance by using representation-alignment (CKNNA) as proxy for generation performance. 3) The maximum achievable CKNNA score with vanilla-REPA is bottlenecked by the VAE features (c) saturating around $\sim 0.42$ . Back-propagating the REPA-loss to the VAE helps address this limitation and improve the final CKNNA score. Given the above insights: we propose REPA-E ( $\S 3.2$ ) for end-to-end LDM training. The key idea is simple: instead of using the diffusion loss, we perform end-to-end training using the REPA loss. The end-to-end training with REPA loss helps improve the final representation-alignment (CKNNA), which in turn leads to improved generation performance ( $\S 4$ ). + +Representation Alignment is Bottlenecked by the VAE Features. Fig. 3c shows that while the naive application of REPA loss [52] leads to improved representation-alignment (CKNNA) score, the maximum achievable alignment score is still bottlenecked the VAE features saturating around a value of 0.4 (maximum value of 1). Furthermore, we find that backpropagating the representation-alignment loss to the VAE helps address this limitation; allowing end-to-end optimization of the VAE features to best support representation-alignment objective [52]. + +# 3.2. End-to-End Training with REPA + +Given the above insights, we next propose REPA-E (§3.2); an end-to-end tuning recipe for jointly training both VAE and LDM features. Instead of directly using diffusion loss, we propose to perform end-to-end training using the representation-alignment loss. The end-to-end training with REPA loss helps better improve the final representation-alignment score (Fig. 3c), which in turn leads to improved final generation performance (refer §4.2). We next discuss key details for implementation of REPA-E for training. + +Batch-Norm Layer for VAE Latent Normalization. To enable end-to-end training, we first introduce a batchnorm layer between the VAE and latent diffusion model (Fig. 1). Typical LDM training involves normalizing the VAE features using precomputed latent statistics (e.g., std $= 1 / 0.1825$ for SD-VAE [39]). This helps normalize the VAE latent outputs to zero mean and unit variance for more efficient training for the diffusion model. However, with end-to-end training the statistics need to be recomputed whenever the VAE model is updated - which is expensive. To address this, we propose the use of a batch + +norm layer [20] which uses the exponential moving average (EMA) mean and variance as a surrogate for dataset-level statistics. The batch-norm layer thus acts as a differentiable normalization operator without the need for recomputing dataset level statistics after each optimization step. + +End-to-End Representation-Alignment Loss. We next enable end-to-end training, by using the REPA loss [52] for updating the parameters for both VAE and LDM during training. Formally, let $\mathcal{V}_{\phi}$ represent the VAE, $\mathcal{D}_{\theta}$ be the diffusion model, $f$ be the fixed pretrained perceptual model (e.g., DINO-v2 [33]) for REPA [52] and $\mathbf{x}$ be a clean image. Also similar to REPA, consider $h_{\omega}(\mathbf{h}_t)$ be the projection of diffusion transformer output $\mathbf{h}_t$ through a trainable projection layer $h_{\omega}$ . We then perform end-to-end training by applying the REPA loss over both LDM and VAE as, + +$$ +\mathcal {L} _ {\mathrm {R E P A}} (\theta , \phi , \omega) = - \mathbb {E} _ {\mathbf {x}, \epsilon , t} \left[ \frac {1}{N} \sum_ {n = 1} ^ {N} \operatorname {s i m} \left(\mathbf {y} ^ {[ n ]}, h _ {\omega} \left(\mathbf {h} _ {t} ^ {[ n ]}\right)\right) \right], +$$ + +where $\mathbf{y} = f(\mathbf{x})$ is the output of the pretrained perceptual model (e.g., DINO-v2 [33]), $N$ is number of patches, $\mathrm{sim(< ., . >)}$ computes the patch-wise cosine similarities between pretrained representation $\mathbf{y}$ from perceptual model (e.g., DINO-v2) and diffusion transformer hidden state $\mathbf{h}_t$ . + +Diffusion Loss with Stop-Gradient. As discussed in Fig. 3a and §3.1, backpropagating the diffusion loss to the VAE causes a degradation of latent-space structure. To avoid this, we introduce a simple stopgrad operation which limits the application of diffusion loss $\mathcal{L}_{\mathrm{DIFF}}$ to only the parameters $\theta$ of the latent diffusion model $\mathcal{D}_{\theta}$ . + +VAE Regularization Losses. Finally, we introduce regularization losses $\mathcal{L}_{\mathrm{REG}}$ for VAE $\nu_{\phi}$ , to ensure that the end-to-end training process does not impact the reconstruction + +performance (rFID) of the original VAE. In particular, following [1], we use three losses, 1) Reconstruction Losses $(\mathcal{L}_{\mathrm{MSE}},\mathcal{L}_{\mathrm{LPIPS}})$ , 2) GAN Loss $(\mathcal{L}_{\mathrm{GAN}})$ , 3) KL divergence loss $(\mathcal{L}_{\mathrm{KL}})$ as regularization loss $\mathcal{L}_{\mathrm{REG}}$ for the VAE $\nu_{\phi}$ . + +Overall Training. The overall training is then performed in an end-to-end manner using the following loss, + +$$ +\mathcal {L} (\theta , \phi , \omega) = \mathcal {L} _ {\mathrm {D I F F}} (\theta) + \lambda \mathcal {L} _ {\mathrm {R E P A}} (\theta , \phi , \omega) + \eta \mathcal {L} _ {\mathrm {R E G}} (\phi), +$$ + +where $\theta, \phi, \omega$ refer to the parameters for the LDM, VAE and trainable REPA projection layer [52], respectively. Further implementation details are provided in §4.1 and Appendix. + +# 4. Experiments + +We next validate the performance of REPA-E and the effect of proposed components through extensive evaluation. In particular, we investigate three key research questions: + +1. Can REPA-E significantly improve generation performance and training speed? (Sec. 4.2, Tab. 1, Fig. 1, 4) +2. Does REPA-E generalize across variations in training settings including model-scale, architecture, encoder model for REPA etc.? (Sec. 4.3, Tab. 2, 3, 4, 5, 6, 7) +3. Analyze the impact of end-to-end tuning (REPA-E) on VAE latent-space structure and downstream generation performance. (please refer Sec. 4.4, Fig. 5, Tab. 8, 13) + +# 4.1. Setup + +Implementation Details. We follow the same setup as in SiT [30] and REPA [52] unless otherwise specified. All training is conducted on the ImageNet [6] training split. We adopt the same data preprocessing protocol as in ADM [7], where original images are center-cropped and resized to $256 \times 256$ resolution. We experiment with publicly available VAEs, including SD-VAE (f8d4) [39], VA-VAE (f16d32) [39], and our own f16d32 VAE trained on ImageNet, referred to as IN-VAE. Depending on the VAE downsampling rate, we adopt SiT-XL/1 and SiT-XL/2 for $4 \times$ and $16 \times$ downsampling rates, respectively, where 1 and 2 denote the patch sizes in the transformer embedding layer. We disable affine transformations in the BN [20] layer between the VAE and SiT, relying solely on the running mean and standard deviation. The VAE regularization loss combines multiple objectives and is defined as: $\mathcal{L}_{\mathrm{REG}} = \mathcal{L}_{\mathrm{KL}} + \mathcal{L}_{\mathrm{MSE}} + \mathcal{L}_{\mathrm{LPIPS}} + \mathcal{L}_{\mathrm{GAN}}$ . For alignment loss, we use DINOv2 [33] as external visual features and apply alignment to the eighth layer of the SiT model. Empirically, we set the alignment loss coefficient to $\lambda_{\mathrm{REPA}_g} = 0.5$ for updating SiT and $\lambda_{\mathrm{REPA}_v} = 1.5$ for VAE. For optimization, we use AdamW [23, 29] with a constant learning rate of $1 \times 10^{-4}$ , and a global batch size of 256. During training, we apply gradient clipping and exponential moving average (EMA) to the generative model for stable optimization. All experiments are conducted on 8 NVIDIA H100 GPUs. + +Evaluation. For image generation evaluation, we strictly follow the ADM setup [7]. We report generation quality using Fréchet inception distance (gFID) [17], structural FID (sFID) [31], inception score (IS) [41], precision (Prec.) and recall (Rec.) [25], measured on 50K generated images. For sampling, we follow the approach in SiT [30] and REPA [52], using the SDE Euler-Maruyama sampler with 250 steps. In terms of VAE benchmark, we measure the reconstruction FID (rFID) on 50K images from the ImageNet [6] validation set at a resolution of $256 \times 256$ . + +# 4.2. Impact on Training Performance and Speed + +We first analyze the impact of end-to-end tuning using REPA-E (Sec. 3.2) for improving generation performance and speed when training latent-diffusion transformers. + +Quantitative Evaluation. We compare REPA-E against various latent diffusion model (LDM) baselines in Tab. 1. We evaluate models of similar sizes ( $\sim$ 675M parameters) on ImageNet $256 \times 256$ generation task. All results are reported without classifier-free guidance [18] using popular SiT-XL [30] model for training. We make two observations; 1) End-to-End tuning leads to faster training: consistently improving generation FID (gFID) from $19.40 \rightarrow 12.83$ (20 epochs), $11.10 \rightarrow 7.17$ (40 epochs), and $7.90 \rightarrow 4.07$ (80 epochs), even when comparing with REPA [52]. 2) End-to-End training leads to better final performance: REPA-E at 80 epochs surpasses FasterDiT [49] (gFID=7.91) trained for 400 epochs and even MaskDiT [54], DiT [34], and SiT [30] which are trained over 1400 epochs. For instance, REPA-E reaches an FID of 4.07 within 400K steps, significantly boosting final performance over even REPA which only reaches a final FID for 5.9 after 4M steps [52]. + +Qualitative Evaluation. We provide qualitative comparisons between REPA [52] and REPA-E in Fig. 4. We generate images from the same noise and label using checkpoints at $50\mathrm{K}$ , $100\mathrm{K}$ , and $400\mathrm{K}$ training iterations, respectively. As seen in Fig. 4, we observe that REPA-E demonstrates superior image generation quality compared to the REPA baseline, while also generating more structurally meaningful images during early stages of training process. + +# 4.3. Generalization and Scalability of REPA-E + +We next analyze the generalization of the proposed approach to variation in training settings including model-size, tokenizer architecture, representation encoder, alignment depth [52] etc. Unless otherwise specified, all analysis and ablations use SiT-L [30] as the generative model, SD-VAE as the VAE, and DINOv2-B [33] as the pretrained vision model for REPA loss [52]. Default REPA alignment-depth of 8 is used. We train each variant for 100K iterations and report results without classifier-free guidance [18]. All baseline numbers are reported using vanilla REPA and compared with end-to-end training using REPA-E. + +![](images/f31586e0f57411e1372077ad844a09a2817c61bce236304b09387a98634d6752.jpg) +Figure 4. End-to-End Tuning (REPA-E) Improves Visual Scaling. We observe that REPA-E produces higher-quality images at $400\mathrm{K}$ steps compared with the vanilla-REPA and generates more structurally meaningful images even in the early stages of training. Results for both methods are sampled using the same seed, noise and class label. We use a classifier-free guidance scale of 4.0 during sampling. + +
MethodTokenizerEpochsgFID↓sFID↓IS↑
Without End-to-End Tuning
MaskDiT [54]16005.6910.34177.9
DiT [34]SD-VAE14009.626.85121.5
SiT [30]14008.616.32131.7
FasterDiT [49]4007.915.45131.3
REPA [52]SD-VAE2019.406.0667.4
4011.106.0667.4
807.905.06122.6
8005.905.73157.8
With End-to-End Tuning (Ours)
REPA-ESD-VAE*2012.835.0488.8
407.174.39123.7
804.074.60161.8
+ +Impact of Model Size. Tab. 2 compares SiT-B, SiT-L, and SiT-XL to evaluate the effect of model size. We make two key observations. First, across all configurations, REPA-E consistently improves performance over the REPA baseline. Specifically, it reduces gFID from $49.5 \rightarrow 34.8$ for SiT-B, $24.1 \rightarrow 16.3$ for SiT-L, and $19.4 \rightarrow 12.8$ for SiT-XL, demonstrating the effectiveness. Second, surprisingly the percentage gains in gFID achieved with REPA-E (over REPA) improve with increasing model size. For instance, for SiT-B model REPA-E leads to a $29.6\%$ improvement in gFID over REPA. Surprisingly even more gains + +Table 1. REPA-E for Accelerated Generation Performance. End-to-End training with REPA-E achieves significantly better performance (lower gFID) while using fewer epochs. Notably, REPA-E with only 80 epochs surpasses vanilla REPA using $10 \times$ epochs. * indicates that VAE is updated during end-to-end training. All results are w/o classifier-free guidance on ImageNet 256 × 256. Additional system-level comparisons with classifier-free guidance and state-of-the-art results are provided in Tab. 13. + +
Diff. ModelgFID↓sFID↓IS↑Prec.↑Rec.↑
SiT-B (130M)49.57.0027.50.460.59
+REPA-E (Ours)34.86.3139.10.570.59
SiT-L (458M)24.16.2555.70.620.60
+REPA-E (Ours)16.35.6975.00.680.60
SiT-XL (675M)19.46.0667.40.640.61
+REPA-E (Ours)12.85.0488.80.710.58
+ +Table 2. Variation in Model-Scale. We find that REPA-E brings substantial performance improvements across all model-scales. All baselines are reported using vanilla-REPA [52] for training. + +are achieved for bigger models improving gFID by $32.3\%$ and $34.0\%$ for SiT-L and SiT-XL models respectively. This trend highlights the scalability of REPA-E; larger models achieve better percentage gains over vanilla-REPA. + +Variation in Representation Encoder. We report results across different perception model encoders (CLIP-L, I-JEPA-H, DINOv2-B, and DINOv2-L) Tab. 3. We observe that REPA-E gives consistent performance improvements over REPA, across different choices of the perceptual encoder model. In particular, with DINOv2-B and DINOv2-L, REPA-E significantly reduces gFID from $24.1 \rightarrow 16.3$ and from $23.3 \rightarrow 16.0$ , respectively. + +Variation in VAE. Tab. 4 evaluates the impact of different VAEs on REPA-E performance. In particular, we report results using three different VAEs 1) SD-VAE [1], 2) VA-VAE [48] and 3) IN-VAE (a $16 \times$ downsampling, 32-channel VAE trained on ImageNet [6] using official training code from [39]). Across all variations, REPA-E consistently improves performance over the REPA baseline. REPA-E reduces gFID from $24.1 \rightarrow 16.3$ , from $22.7 \rightarrow 12.7$ , and $12.8 \rightarrow 11.1$ , for SD-VAE, IN-VAE and VA-VAE, respectively. The results demonstrate that REPA-E robustly im + +![](images/8468ce81fd22679b0ed4c8eaf60da1840a9932da9a153062187a05c4ba7e517f.jpg) +(a) PCA Visualization of Latent Space Structure [24] + +![](images/eea707e67d1951bd1d3d18842143d659bf01ce8e17c81d28e9da944af3faeb0e.jpg) +(b) Impact of End-to-End Tuning for Automatically Improving Latent Space Structure +Figure 5. End-to-End Training Improves Latent Space Structure. (a) We observe that the latent space of pretrained VAEs can suffer either high noise components (e.g., SDXL-VAE, SD-VAE [39]), or, be over-smoothed and lack details (e.g., VA-VAE [48]). (b) The use of end-to-end tuning ( $\S 3.2$ ) automatically helps improve the latent space structure in a model-agnostic manner across different VAE architectures. For instance, similar to findings of concurrent work [43], we observe that SD-VAE suffers from high noise components in the latent space. Applying end-to-end training automatically helps adjust the latent space to reduce noise. In contrast, other VAEs such as recently proposed VA-VAE [48] suffer from an over-smoothed latent space. The use of end-to-end tuning with REPA-E automatically helps learn a more detailed latent-space structure to best support generation performance. + +
Target Repr.gFID↓sFID↓IS↑Prec.↑Rec.↑
I-JEPA-H [2]23.05.8160.30.620.60
+REPA-E (Ours)16.55.1873.60.680.60
CLIP-L [37]29.25.9846.40.590.61
+REPA-E (Ours)23.46.4457.10.620.60
DINOv2-B [33]24.16.2555.70.620.60
+REPA-E (Ours)16.35.6975.00.680.60
DINOv2-L [33]23.35.8959.90.610.60
+REPA-E (Ours)16.05.5977.70.680.58
+ +Table 3. Variation in Representation Encoder. REPA-E yields consistent performance improvements across different choices for the representation-encoder used for representation-alignment [52]. All baselines are reported using vanilla-REPA [52] for training. + +
AutoencodergFID↓sFID↓IS↑Prec.↑Rec.↑
SD-VAE [39]24.16.2555.70.620.60
+REPA-E (Ours)16.35.6975.00.680.60
IN-VAE (f16d32)22.75.4756.00.620.62
+REPA-E (Ours)12.75.5784.00.690.62
VA-VAE [48]12.86.4783.80.710.58
+REPA-E (Ours)11.15.3188.80.720.61
+ +proves generative quality across diverse variations in architecture, pretraining dataset and training setting of the VAE. + +Variation in Alignment Depth. Tab. 5 investigates the effect of applying the alignment loss at different layers the diffusion model. We observe that REPA-E consistently enhances generation quality over the REPA baseline across variation in choice of alignment depth; with gFID improving from $23.0 \rightarrow 16.4$ (6th layer), $24.1 \rightarrow 16.3$ (8th layer), and $23.7 \rightarrow 16.2$ (10th layer). + +Table 4. Variation in VAE Architecture. All baselines are reported using vanilla-REPA [52] for training. + +
Aln. DepthgFID↓sFID↓IS↑Prec.↑Rec.↑
6th layer23.05.7259.20.620.60
+REPA-E (Ours)16.46.6474.30.670.59
8th layer24.16.2555.70.620.60
+REPA-E (Ours)16.35.6975.00.680.60
10th layer23.75.9156.90.620.60
+REPA-E (Ours)16.25.2274.70.680.58
+ +Table 5. Variation in Alignment Depth. End-to-End tuning (REPA-E) gives consistent performance imrpoements over original REPA [52] across varying alignment-depths. + +
ComponentgFID↓sFID↓IS↑Prec.↑Rec.↑
w/o stopgrad444.1460.31.490.000.00
w/o batch-norm18.15.3272.40.670.59
w/o LGAN19.26.4768.20.640.58
REPA-E (Ours)16.35.6975.00.680.60
+ +Table 6. Ablation Study on Role of Different Components. + +Ablation on Design Components. We also perform ablation studies analyzing the importance of each component discussed in Sec. 3.2. Results are shown in Tab. 6. We observe that each component plays a key role in the final performance for REPA-E. In particular, we observe that the stop-grad operation on the diffusion loss helps prevent degradation of the latent-space structure. Similarly, the use of batch norm is useful adaptively normalizing the latent-statistics and helps improve the gFID from $18.09 \rightarrow 16.3$ . Similarly, the regularization losses play a key role in maintaining the reconstruction performance of the finetuned VAE, thereby improving the gFID from $19.07 \rightarrow 16.3$ . + +End-to-End Training from Scratch. We next analyze the impact of VAE initialization on end-to-end training. As shown in Tab. 7, we find that while initializing the + +
MethodgFID↓sFID↓IS↑Prec.↑Rec.↑
100K Iterations (20 Epochs)
REPA [52]19.406.0667.40.640.61
REPA-E (scratch)14.127.8783.50.700.59
REPA-E (VAE init.)12.835.0488.80.710.58
200K Iterations (40 Epochs)
REPA [52]11.105.05100.40.690.64
REPA-E (scratch)7.546.17120.40.740.61
REPA-E (VAE init.)7.174.39123.70.740.62
400K Iterations (80 Epochs)
REPA [52]7.905.06122.60.700.65
REPA-E (scratch)4.344.44154.30.750.63
REPA-E (VAE init.)4.074.60161.80.760.62
+ +Table 7. End-to-End Training from Scratch. We find that while initializing the VAE with pretrained weights (SD-VAE [39]) helps slightly improve performance, REPA-E can be used to train both VAE and LDM from scratch in an end-to-end manner; still achieving significantly superior performance over REPA which requires a separate stage for training VAE in addition to LDM training. + +VAE from pretrained weights helps slightly improve performance, REPA-E can be used to train both VAE and LDM from scratch still achieving superior performance over REPA, which technically requires a separate stage for VAE training in addition to LDM training. For instance, while REPA achieves a FID of 5.90 after 4M iterations, REPA-E while training entirely from scratch (for both VAE and LDM) achieves much faster and better generation FID of 4.34 within just 400K iterations. + +# 4.4. Impact of End-to-End Tuning on VAE + +We next analyze the impact of end-to-end tuning on the VAE. In particular, we first show that end-to-end tuning improves the latent-space structure (Fig. 5). We next show that once tuned using REPA-E, the finetuned VAEs can be used as a drop-in replacement for their original counterparts offering significantly improved generation performance. + +End-to-End Training improves Latent Space Structure. Results are shown in Fig. 5. Following [24], we visualize latent space structure using principal component analysis (PCA) that projects them to three channels colored by RGB. We consider three different VAEs: 1) SD-VAE [39], 2) IN-VAE (a $16\times$ downsampling, 32-channel VAE trained on ImageNet [6]). 3) VA-VAE from recent work from [48]. We observe that end-to-end tuning using REPA-E automatically improves the latent space structure of the original VAE. For instance, similar to findings of concurrent work [43], we observe that SD-VAE suffers from high noise components in the latent space. Applying end-to-end training automatically helps adjust the latent space to learn reduce noise. In contrast, other VAEs such as recently proposed VA-VAE [48] suffer from over-smother latent space. Application of E2E tuning automatically helps learn a more detailed latent-space to best support generation performance. + +End-to-End Training Improves VAE Performance. + +
VAEDiffusion modelREPAgFID-50K
SD-VAE [39]DiT-XL [34]✘19.82
VA-VAE [48]DiT-XL [34]✘6.74
E2E-VAE (Ours)DiT-XL [34]✘6.75
SD-VAE [39]SiT-XL [30]✘17.20
VA-VAE [48]SiT-XL [30]✘5.93
E2E-VAE (Ours)SiT-XL [30]✘5.26
SD-VAE [39]DiT-XL [34]✓12.29
VA-VAE [48]DiT-XL [34]✓4.71
E2E-VAE (Ours)DiT-XL [34]✓4.20
SD-VAE [39]SiT-XL [30]✓7.90
VA-VAE [48]SiT-XL [30]✓4.88
E2E-VAE (Ours)SiT-XL [30]✓3.46
+ +Table 8. Impact of End-to-End Tuning on VAE Performance. We find that once tuned using REPA-E, the finetuned VAEs can be used as a drop-in replacement for their original counterparts offering significantly accelerated generation performance. We fix all the VAEs and only train the diffusion models (with and w/o REPA). E2E-VAE is obtained from REPA-E fine-tuning (VA-VAE + SiT-XL). All results are reported at 80 epochs (400K iterations). + +We next evaluate the impact of end-to-end tuning on downstream generation performance of the VAE. To this end, we first use end-to-end tuning for finetuning the recently proposed VA-VAE [48]. We then use the resulting end-to-end finetuned-VAE (named E2E-VAE), and compare its downstream generation performance with current state-of-the-art VAEs; including SDVAE [39] and VA-VAE [48]. To do this, we conduct traditional latent diffusion model training (w/o REPA-E), where only the generator network is updated while keeping the VAE frozen. Tab. 8 shows the comparison of VAE downstream generation across diverse training settings. We observe that end-to-end tuned VAEs consistently outperform their original counterparts for downstream generation tasks across variations in LDM architecture and training settings. Interestingly, we observe that a VAE tuned using SiT-XL yields performance improvements even when using a different LDM architecture such as DiT-XL; thereby demonstrating the robustness of our approach. + +# 5. Conclusion + +"Can we unlock VAE's for performing end-to-end training with latent diffusion transformers?" Directly backpropagating diffusion loss to the VAE is ineffective and even degrges final performance. We show that while diffusion loss is ineffective, end-to-end training can be unlocked using REPA loss. Our end-to-end training recipe (REPA-E), significantly improves latent-space structure, shows remarkable performance; speeding up diffusion model training by over $17 \times$ and $45 \times$ over REPA and vanilla training recipes. Overall, our approach achieves a new state-of-the-art results with generation FID of 1.26 and 1.83 with and without use of classifier-free guidance. We hope that our work can help foster further research for enabling end-to-end training with latent diffusion transformers. + +# Acknowledgments + +We would like to extend our deepest appreciation to Zeyu Zhang, Qinyu Zhao, and Zhanhao Liang for insightful discussions. We would also like to thank all reviewers for their constructive feedback. This work was supported in part by the Australian Research Council under Discovery Project DP210102801 and Future Fellowship FT240100820. SX acknowledges support from the OpenPath AI Foundation, IITP grant funded by the Korean Government (MSIT) (No. RS-2024-00457882) and NSF Award IIS-2443404. + +# References + +[1] Stability AI. Improved autoencoders ... https://huggingface.co/stabilityyai/sd-vae-ft-mse, n.d. Accessed: April 11, 2025. 5, 6 +[2] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023. 7 +[3] Dana H Ballard. Modular learning in neural networks. In Proceedings of the sixth National conference on Artificial intelligence-Volume 1, pages 279-284, 1987. 3 +[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 13 +[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3 +[6] 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. 2, 5, 6, 8, 12 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3 +[9] 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 +[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 *Forty-first International Conference on Machine Learning*, 2024. 3, 12 +[11] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv preprint arXiv:2405.05945, 2024. 3 +[12] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23164-23173, 2023. 13 +[13] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Mdtv2: Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2023. 13 +[14] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 1 +[15] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 1 +[16] 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 +[17] 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. 5 +[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5 +[19] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. The platonic representation hypothesis. In International Conference on Machine Learning, 2024. 3 +[20] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 4, 5 +[21] Dongwon Kim, Ju He, Qihang Yu, Chenglin Yang, Xiaohui Shen, Suha Kwak, and Liang-Chieh Chen. Democratizing text-to-image masked generative models with compact text-aware one-dimensional tokens. arXiv preprint arXiv:2501.07730, 2025. 3 +[22] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 3 +[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[24] Theodoros Kouzelis, Ioannis Kakogeorgiou, Spyros Gidaris, and Nikos Komodakis. Eq-vae: Equivalence regularized latent space for improved generative image modeling. arXiv preprint arXiv:2502.09509, 2025. 2, 3, 7, 8 + +[25] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 5 +[26] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.3 +[27] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems, 37:56424-56445, 2025. 13 +[28] 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 European conference on computer vision, pages 740-755. Springer, 2014. 12 +[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[30] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In European Conference on Computer Vision, pages 23-40. Springer, 2024. 2, 3, 5, 6, 8, 12, 13 +[31] Charlie Nash, Jacob Menick, Sander Dieleman, and Peter Battaglia. Generating images with sparse representations. In International Conference on Machine Learning, pages 7958-7968. PMLR, 2021. 5 +[32] OpenAI. Sora. https://openai.com/sora, 2024.3 +[33] Maxime Oquab, Timothee Darct, Theo 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. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 4, 5, 7 +[34] 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. 3, 5, 6, 8, 13 +[35] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 3 +[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. 3 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 +[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 1 + +[39] 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, 4, 5, 6, 7, 8, 12, 13 +[40] Leonid I. Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: Nonlinear Phenomena, 60(1):259-268, 1992. 12 +[41] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5 +[42] Jaskirat Singh, Stephen Gould, and Liang Zheng. High-fidelity guided image synthesis with latent diffusion models. arXiv preprint arXiv:2211.17084, 2022. 3 +[43] Ivan Skorokhodov, Sharath Girish, Benran Hu, Willi Menapace, Yanyu Li, Rameen Abdal, Sergey Tulyakov, and Aliaksandr Siarohin. Improving the diffusability of autoencoders. arXiv preprint arXiv:2502.14831, 2025. 2, 7, 8 +[44] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 13 +[45] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing systems, 37:84839-84865, 2025. 13 +[46] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In Advances in Neural Information Processing Systems, pages 11287-11302. Curran Associates, Inc., 2021. 3, 12 +[47] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3 +[48] Jingfeng Yao and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. arXiv preprint arXiv:2501.01423, 2025. 3, 6, 7, 8, 13 +[49] Jingfeng Yao, Wang Cheng, Wenyu Liu, and Xinggang Wang. Fasteredit: Towards faster diffusion transformers training without architecture modification. arXiv preprint arXiv:2410.10356, 2024. 5, 6, 13 +[50] 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. 13 +[51] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. Advances in Neural Information Processing Systems, 37:128940-128966, 2025. 3 +[52] 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. 2, 3, 4, 5, 6, 7, 8, 12, 13 + +[53] Kaiwen Zha, Lijun Yu, Alireza Fathi, David A Ross, Cordelia Schmid, Dina Katabi, and Xiuye Gu. Language-guided image tokenization for generation. arXiv preprint arXiv:2412.05796, 2024. 3 +[54] Hongkai Zheng, Weili Nie, Arash Vahdat, and Anima Anandkumar. Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305, 2023. 5, 6, 13 \ No newline at end of file diff --git a/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/images.zip b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f9abb08d96e9b782a8555ca9b2a7e70f53709098 --- /dev/null +++ b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67f2ede1c1ba7dfda47e73a549a99178185617980b056dc3a1b772b7c3e03693 +size 752110 diff --git a/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/layout.json b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7a187ff9ca0dfda7fd3d53225a09d8665adb6743 --- /dev/null +++ b/ICCV/2025/REPA-E_ Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7f5c626f0958f5c448a7896a1c634fc3716d54b3bd0332ed0285584ec47dda9 +size 417576 diff --git a/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/e303d607-a2aa-4ba0-8faf-b6ee8aef2213_content_list.json b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/e303d607-a2aa-4ba0-8faf-b6ee8aef2213_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..be36e7b3436a7be96632ff5f66f4c0f1bd327d01 --- /dev/null +++ b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/e303d607-a2aa-4ba0-8faf-b6ee8aef2213_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3c46062fd0ddf86b6424a8bd997d79bee7a8b2b505272aacc61750db015ff19 +size 80812 diff --git a/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/e303d607-a2aa-4ba0-8faf-b6ee8aef2213_model.json b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/e303d607-a2aa-4ba0-8faf-b6ee8aef2213_model.json new file mode 100644 index 0000000000000000000000000000000000000000..741304738ed14ac2878c781680132f188a3b0bb3 --- /dev/null +++ b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/e303d607-a2aa-4ba0-8faf-b6ee8aef2213_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:daa1027da64b9364edb8298bc48475187ed5300e3403183875a9404e6e123495 +size 105722 diff --git a/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/e303d607-a2aa-4ba0-8faf-b6ee8aef2213_origin.pdf b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/e303d607-a2aa-4ba0-8faf-b6ee8aef2213_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..99132691ae072f543e464c1441e6ae5154975248 --- /dev/null +++ b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/e303d607-a2aa-4ba0-8faf-b6ee8aef2213_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36bd1c30d9ddced721931cb774f03183d2591ae53216b0b1ffe9cefe9f20e1a0 +size 4774229 diff --git a/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/full.md b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6b64cd9786f63984256bf3df1bfc0f124783c962 --- /dev/null +++ b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/full.md @@ -0,0 +1,343 @@ +# REPARO: Compositional 3D Assets Generation with Differentiable 3D Layout Alignment + +Haonan Han $^{1*}$ Rui Yang $^{2*}$ Huan Liao $^{1*}$ Jiankai Xing $^{1}$ Zunnan Xu $^{1}$ Xiaoming Yu $^{3}$ Junwei Zha $^{3}$ Xiu Li $^{1\dagger}$ Wanhua Li $^{4\dagger}$ + +$^{1}$ Tsinghua University $^{2}$ The University of Hong Kong $^{3}$ Tencent $^{4}$ Harvard University + +![](images/d4fde1c2331a9d424f840906513161373bf851262f042c1c1a72e156e06e070d.jpg) +(a) GT +(b) Individually generated asset 1 +(c) Individually generated asset 2 +(d) Compositional generation +(e) Ours +Figure 1. Qualitative comparison of generated 3D assets of the single object and multiple objects. (b) and (c) are generated 3D assets of the single object assets using DreamGaussian [51]; (d) is a 3D asset of multiple objects generated by DreamGaussian [51]; (e) is our result. + +# Abstract + +Traditional image-to-3D models often struggle with scenes containing multiple objects due to biases and occlusion complexities. To address this challenge, we present REPARO, a novel approach for compositional 3D asset generation from single images. REPARO employs a two-step process: first, it extracts individual objects from the scene and reconstructs their 3D meshes using image-to-3D models; then, it optimizes the layout of these meshes through differentiable rendering techniques, ensuring coherent scene composition. By integrating optimal transport-based long-range appearance loss term and high-level semantic loss term in the differentiable rendering, REPARO can effectively recover the layout of 3D assets. The proposed method can significantly enhance object independence, detail accuracy, and overall scene coherence. Extensive evaluation of multi-object scenes demonstrates that our REPARO offers a comprehensive approach to address the complexities of multi-object 3D scene generation from single images. The demo have been available at https://reparo-3d.github.io/. + +# 1. Introduction + +The field of 3D content creation has advanced rapidly due to progress in reconstruction methods [16, 35] and generative frameworks [40]. Automated 3D content generation enables creators in augmented reality (AR), virtual reality (VR), gaming, and filmmaking [20, 25, 50] by significantly reducing manual labor requirements. + +Recent research in 3D generation primarily focuses on optimization-based 2D approaches, particularly text-to-3D [26, 40, 60] and image-to-3D [27, 31, 34, 52] models. For example, Dreamfusion [40] overcomes 3D data scarcity by distilling geometry and appearance from 2D diffusion models; and DreamGaussian [51] integrates 3D Gaussian Splatting into generative pipelines with mesh extraction and texture refinement. However, despite their effectiveness in single-object scenes, these methods face challenges in reconstructing multi-object scenes. This limitation stems from dataset biases: most 3D training samples feature centrally aligned single objects, and preprocessing pipelines often re-center inputs, thereby introducing inherent positional bias. Consequently, occluded objects are frequently misrepresented as fused entities, leading to erroneously merged assets in rendered outputs, as shown in Figure 1 (d). Fur + +thermore, typical 3D generation models output assets as monolithic mesh representations. This forces users to rely on error-prone post-processing steps to segment individual object mesh, which does not ensure the quality of the segmented assets. + +To address this challenge, we propose REPARO, a compositional 3D generation pipeline capable of generating multiple objects from a single input image. Our approach decomposes the scene into discrete assets, then systematically reassembles and globally optimizes their spatial arrangement. This methodology not only leverages the strengths of off-the-shelf models (the ability to generate single objects with high fidelity) but also addresses inherent challenges in the multi-object scene with complex occlusions and interactions. + +Specifically, our pipeline comprises two core stages. First, individual objects are isolated from the input image through cropping and inpainting to generate context-complete single-object images, which are processed using off-the-shelf image-to-3D models [33, 41, 54] to produce high-fidelity 3D assets. Second, the 3D assets are assembled into a cohesive scene and geometrically aligned with the input image. After placing all objects within a unified coordinate system, we employ differentiable rendering to optimize the spatial arrangement of each asset. During rendering, the pose parameters (translation, rotation, and scale) of each 3D asset are treated as estimizable variables, which are iteratively refined via gradient descent using a loss function. Unlike conventional pixel-wise loss function [3], we introduce an optimal transport (OT)-based long-range appearance loss that integrates RGB color, depth, and positional information to constrain global structural relationships between objects. Therefore, this loss can match the rendered image and the reference image. Additionally, a high-level semantic loss is introduced to ensure semantic alignment and coherence between the synthesized scene and the original image context using high-level visual features. + +Through these two stages, individual-object reconstruction and combined-scene optimization, our REPARO enhances object independence and geometric detail accuracy, particularly for occluded regions. Simultaneously, it preserves both spatial arrangement fidelity and semantic-geometric consistency across the synthesized scene. Critically, our method obviates the need for post hoc mesh segmentation or manual joining operations by generating instance-level meshes directly within the scene context, streamlining the 3D asset creation workflow and improving end-user usability. + +To evaluate our method, we conducted experiments on 20 multi-object scenes from the Google Scanned Objects (GSO) dataset [9] (containing 2-6 objects per scene) and 20 additional in-the-wild images for qualitative analysis. Quantitative and qualitative comparisons demonstrate that + +REPARO offers significant improvements over previous techniques in ensuring the quality of assets, managing multiple assets, and processing occlusion. Our contributions are summarized as follows: + +1. A two-stage compositional generation Pipeline: We propose REPARO, a framework to decouple multi-object 3D generation into two stages: (1) per-object high-fidelity reconstruction via inpainting-aware cropping and off-the-shelf models, and (2) global scene optimization using novel differentiable rendering. This structure ensures object-level geometric independence while preserving scene-level contextual integrity. +2. Optimal transport-driven layout loss: We design a long-range appearance loss based on optimal transport (OT) theory, which enforces multi-modal constraints (RGB, depth, positional channels) to align 3D asset layouts with the reference image. By solving for global pixel correspondences, OT mitigates local minima issues inherent in the pixel-wise loss. +3. Semantic loss: We further introduce a semantic alignment loss that leverages high-level visual features to refine object placements. This loss ensures synthesized scenes adhere to the original image's semantic context. + +# 2. Related work + +Text-to-3D generation. Early exploration in 3D generation based on text [15, 35–37, 44, 56, 63] relied on a largescale pre-trained model of text and image, CLIP [44], to optimize representation as the prior. In light of the promising capability of diffusion models, some works [1, 21, 26, 40, 45, 46, 57, 60] introduce Score Distillation Sampling (SDS) to narrow the gap between novel view images rendered from 3D representation and diffusion prior. Recent works [4, 24, 51, 65] have used 3D Gaussian Splatting [16] instead of NeRF [35] for 3D generation, due to the fact that this efficient 3D representation shows excellent quality [23, 43, 66] in reconstruction tasks. + +Image-to-3D generation. Differing from the ambiguity and diversity associated with textual description, 3D generation based on a single image predominantly concentrates on maintaining geometric and textural consistency between the generated assets and the input images. Some early works [34, 52] used 2D diffusion models as prior knowledge to guide the training of 3D representation of the object in input images. However, with the enrichment of 3D data [6, 7], a series of adapting fine-tuned diffusion models [27, 29, 30, 42, 47, 61, 68] represented by Zero-1-to-3 [31] have been continuously developed. Subsequently, further advancements [32, 33, 48, 49] have been proposed, which offer improved quality and multi-view consistency. + +Compositional 3D generation. Unlike single-object reconstruction, multi-object 3D generation requires attention to both local details and global spatial relation + +![](images/76f93e59f785cc48331768ab4e63d792cfe6023f7dd94833d8d67cf2a5845a3b.jpg) +Figure 2. The diagram of the proposed REPARO. (a) is the pipeline to reconstruct the 3D asset of each object in the reference image. $B$ and $O$ denote bounding boxes and occlusion information of each object, respectively. If an object is occluded, the preprocessing module will complement it using the inpainting model. (b) is the process of layout alignment based on differentiable rendering. The parameters of reconstructed meshes are optimized by gradient descent. The loss function $L$ (Eq. 9) consists of the long-range appearance loss $L_{a}$ and the high-level semantic loss $L_{s}$ . + +ships. Existing approaches address this through two main paradigms: (1) Single-object construction using geometric surface models [10] and editable generative models [53]; (2) Scene composition employing bounding boxes [39], text prompts [59], Gaussian radiance fields [55], and scene graphs [11]. Emerging approaches integrate LLMs for layout interpretation [28], leverage 2D semantic maps [64], or combine Gaussian Splatting with LLMs [67]. Distinct from text/layout-based methods, our approach directly generates multiple high-quality 3D assets from a single compositional image while preserving input-aligned spatial arrangements. + +# 3. Method + +The field of 3D generation is advancing rapidly, with numerous image-to-3D models [51, 54] capable of generating 3D assets from a single image. However, these models typically focus on individual objects, making it challenging to apply them to scenes containing multiple objects. To address this limitation, we propose REPARO for compositional 3D object generation from a single image. Our approach consists of two steps: (1) Extracting each target object from the given image and reconstructing their 3D meshes using off-the-shelf 3D reconstruction models. (2) Optimizing the parameters of each mesh through differentiable rendering to align their layout. In the following section, we provide a detailed explanation of these two steps. + +# 3.1. Reconstruction of Each Single Object + +Since most data samples in 3D datasets only have one object located centrally within the image, and most image-to-3D models recenter the object in the preprocessing step, there is an inherent center bias in these models. As a result, the generated 3D assets for single objects are better + +than for multiple objects. As shown in Figure 1, Dream-Gaussian [51] can generate the sound 3D asset for a single object, but it fails in generation for multiple objects. Instead of reconstructing 3D assets of multiple objects at the same time, our REPARO first focuses on extracting individual objects from images containing several objects and then generating their 3D assets using image-to-3D models. + +Object extraction. As illustrated in Figure 2 (a), given an image $I^{ref}$ with the bounding boxes $B$ and occlusion information $O$ , we use the foundation model SAM [18] to segment each object and obtain their binary masks $M$ . For occluded objects, we utilize a Stable-Diffusion-based inpainting model to reconstruct the occluded parts in the preprocessing module, thereby obtaining complete RGB priors for each object. Plus, we resize and crop the images to centralize the involved objects, which adapts to the center bias of existing image-to-3D models. + +Single object reconstruction. After that, we employ off-the-shelf image-to-3D models, such as Dreamgaussian [51] and TripoSR [54], to generate the corresponding 3D assets for each obtained object. + +# 3.2. Layout Alignment by Differentiable Rendering + +After obtaining every single object, we need a layout. To achieve the goal, first, we put every object together in one coordinate system as demonstrated in Figure 2 (b), then we use differentiable rendering techniques [19, 22, 62] to optimize the layout of each object. Differentiable rendering aims to recover the scene parameters $\theta$ from reference image(s) $I^{ref}$ through analysis by synthesis process. Given an initial estimation of the scene, differentiable rendering can produce a rendered image $I$ together with the gradients with respect to arbitrary scene parameters $\theta$ . Briefly, the gradient + +can be represented as: + +$$ +\frac {\partial I}{\partial \theta} = \left[ \frac {\partial I \left(p _ {1}\right)}{\partial \theta}, \dots , \frac {\partial I \left(p _ {N}\right)}{\partial \theta} \right] = \left[ \frac {\partial I _ {1}}{\partial \theta}, \dots , \frac {\partial I _ {N}}{\partial \theta} \right], \tag {1} +$$ + +where $p_i$ refers to the position of $i$ -th pixel, $I(p_i)$ is the RGB color of $i$ -th pixel, and $N$ is the total number of pixels. We denote $I(p_i) = I_i$ for simplicity in the following. Combining with the loss function between rendered and reference image, we can leverage a gradient-decent-based method like Adam [17] to optimize the scene parameters of interest, like geometry, material, or lights. According to the chain rule, the derivative of the loss function $L(I, I^{ref})$ with respect to scene parameters $\theta$ is: + +$$ +\frac {\partial L}{\partial \theta} = \frac {\partial L}{\partial I} \cdot \frac {\partial I}{\partial \theta}. \tag {2} +$$ + +Specifically, for our task, since we have put every generated object in the same coordinate system, the layout of objects in the initial rendered image and reference image may have quite large differences. Such a case is not suitable for the commonly used pixel-wised $L_{2}$ loss function and differentiable rendering methods that only compute color derivatives. Imagine that when an object in the reference image has no overlapped area with it in the rendered image, the $L_{2}$ loss will not change when we slightly move the object, the gradients of the loss function could get stuck in an undesired local minima and not indicate the right way for optimization. To mitigate this issue, we propose to introduce a loss function that could find the global correspondences. For one thing, inspired by previous work [12, 62], we use the optimal transport algorithm to find a match between the rendered image and the reference image, thus obtaining long-range correspondences. For another thing, we utilize the feature embedding from the visual backbone to get semantic correspondences. + +# 3.2.1. Long-range Appearance Loss Term + +Optimal transport theory describes the following problem: suppose there are $N$ suppliers and $M$ demanders within a region. The $i$ -th supplier holds $s_i$ units of goods, and the $j$ -th demander needs $d_j$ units of goods. The transportation cost per unit of goods from $i$ -th supplier to $j$ -th demander is $c_{ij}$ . The goal of the optimal transport algorithm is to find a transportation matrix $T = \{T_{ij} > 0 \mid i = 1, \dots, N; j = 1, \dots, M\}$ that minimizes the total transportation cost: + +$$ +\min _ {T} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {M} T _ {i j} c _ {i j}, \quad T _ {i j} > 0, \tag {3} +$$ + +$$ +\text {s . t .} \sum_ {i = 1} ^ {N} T _ {i j} = d _ {j}, \quad \sum_ {j = 1} ^ {M} T _ {i j} = s _ {i}, \quad \sum_ {i = 1} ^ {N} s _ {i} = \sum_ {j = 1} ^ {M} d _ {j}. +$$ + +In our layout alignment, we consider all the pixels in the rendered image $I$ as suppliers and all the pixels in the + +reference image $I^{ref}$ as demanders, where $N = M$ and $s_i = d_j = 1$ . For the cost, instead of only considering RGB color distance between $i$ -th pixel and $j$ -th pixel, we define it based on RGB color, depth value, and position distance at the same time: + +$$ +c _ {i j} = \alpha \cdot \| I _ {i} - I _ {j} ^ {r e f} \| _ {2} + \beta \cdot \| D _ {i} - D _ {j} ^ {r e f} \| _ {2} + \gamma \cdot \| p _ {i} - p _ {j} \| _ {2}, \tag {4} +$$ + +where $p_i$ and $p_j$ refer to the screen space position of $i$ -th pixel and $j$ -th pixel, respectively; $D$ is the depth map predicted from the rendered image $I$ using a frozen depth foundation model $F_D(\cdot)$ ; $D_i$ denotes the depth value of $i$ -th pixel; and $\alpha, \beta,$ and $\gamma$ are hyper-parameters. + +Based on Eq. 3 and Eq. 4, we can obtain a transport matrix $T$ using Sinkhorn divergences [5]. $T$ recording a one-to-one mapping between the rendered image $I$ and the reference image $I^{ref}$ . In other words, we can find a target pixel for each pixel in the rendered image. Formally, we define $I_{\sigma(i)}^{ref}$ is the target pixel of $I_i$ , where $\sigma(\cdot)$ is the one-to-one mapping function from the transport matrix $T$ . In addition, rather than taking only RGB color distance as the differentiable rendering loss, we also consider RGB color, depth value, and position distance together as Eq. 4. Accordingly, our appearance loss function can be expressed as: + +$$ +L _ {a} (I, I ^ {r e f}) = \frac {1}{N} \sum_ {i} ^ {N} c _ {i \sigma (i)}. \tag {5} +$$ + +Here, $N$ denotes the number of pixels of the rendered image, and the settings of all hyper-parameters are consistent with Eq. 4. To propagate the gradient of the pixel position to the scene parameter, we can define the pixel position as the projection of the shading point used for this pixel in the rasterization process as DROT [62]. According to Eq. 2 and the chain rule, the derivative of $L_{a}$ with respective to scene parameters $\theta$ is: + +$$ +\frac {\partial L _ {a}}{\partial \theta} = \frac {\partial L _ {a}}{\partial I} \cdot \frac {\partial I}{\partial \theta} + \frac {\partial L _ {a}}{\partial D} \cdot \frac {\partial F _ {D}}{\partial I} \cdot \frac {\partial I}{\partial \theta} + \frac {\partial L _ {a}}{\partial p} \cdot \frac {\partial p}{\partial \theta} \tag {6} +$$ + +Note that since we freeze the depth model $F_{D}(\cdot)$ , the gradients will totally contribute to scene parameters. As the mapping function $\sigma (\cdot)$ considers the long-range correspondence, our appearance loss term can also leverage long-range information during differentiable rendering. + +# 3.2.2. High-level Semantic Loss Term + +To enhance the semantic information during the alignment process, we further propose incorporating high-level features into the loss function of differentiable rendering. Our semantic loss term is defined as: + +$$ +L _ {s} (I, I ^ {r e f}) = \frac {1}{K} \sum_ {i} ^ {K} \| f _ {i} - f _ {i} ^ {r e f} \| _ {2}, \tag {7} +$$ + +Figure 3. GPU memory usage and approximate elapsed time for optimization of compositional assets layout under different parameter settings. +![](images/6fe4fe21a97aa23b3e403967ff1033975ad687b05bb0d9365fac0332cde047b8.jpg) +$\theta = \{\mathrm{trans\_x,trans\_y,trans\_z}\}$ ,GPU Memory: 4 GB + +where $f_{i}$ and $f_{i}^{ref}$ are the $i$ -th embedding in the feature map $f$ and $f^{ref}$ , respectively; $f$ and $f^{ref}$ are extracted using a frozen DINO-v2 [38] backbone $F(\cdot)$ ; and $K$ is the total number of embedding from the last hidden state. Based on this term, we can align the semantic relation between the rendered image and reference image from part to part. According to Eq. 2, the derivative of $L_{s}$ with respective to layout parameters $\theta$ is: + +$$ +\frac {\partial L _ {s}}{\partial \theta} = \frac {\partial L _ {s}}{\partial F} \cdot \frac {\partial F}{\partial I} \cdot \frac {\partial I}{\partial \theta} \tag {8} +$$ + +So far, we have integrated the proposed loss terms together to align the layout of multiple reconstructed 3D objects: + +$$ +L (I, I ^ {r e f}) = \lambda L _ {a} (I, I ^ {r e f}) + (1 - \lambda) L _ {s} (I, I ^ {r e f}), \tag {9} +$$ + +where $\lambda$ is a hyper-parameter to adjust the weight between appearance and semantic loss term. In practice, we align the layout by optimizing the object's translation parameter $t = \{\mathrm{trans\_x},\mathrm{trans\_y},\mathrm{trans\_z}\}$ and scale parameter $s = \{\mathrm{scale\_x},\mathrm{scale\_y},\mathrm{scale\_z}\}$ . We exclude the rotation parameter $r$ of assets from the optimization process, as assets generated via image-to-3D models are loaded into the scene with the same orientation as in the input image, eliminating the need for rotation optimization. Thanks to the proposed loss function, we can get the layout of the compositional 3D assets through differentiable rendering. + +# 4. Experiments + +# 4.1. Implementation details + +In the reconstruction of 3D assets from a single object, we utilize SAM-ViT-H [18] as the segmentation model to obtain masks corresponding to each bounding box. In the preprocessing module, we extract images of the objects using these masks and subsequently resize and crop the images to center the objects. We then apply the Stable Diffusion-based inpainting model [14] to address occlusions in the objects. During the generation of 3D assets from a single object, we employ the DreamGaussian [51] and TripoSR [54]. + +Table 1. Performance of different 3D generation models for compositional 3D assets generation. REPARO\* and REPARO\* separately denote utilizing DreamGaussian and TripoSR as the image-to-3d model based on our framework. + +
MethodCLIP ↑PSNR ↑SSIM ↑LPIPS ↓
DreamGaussian [51]0.80713.2800.8020.240
TripoSR [54]0.79517.2480.8630.218
Wonder3D [33]0.80113.6890.8070.238
LRM [13]0.81213.6640.8060.237
REPARO♣ (ours)0.83317.2790.8260.234
REPARO♣ (ours)0.82217.7510.8650.216
+ +Table 2. Runtime and memory usage of each stage. +Table 3. Comparison with representative methods on perceptual quality and resource usage. + +
StageVRAM Timequality and resource usage.
6 GB < 1sMethodCLIP ↑GPU Time
SAM Segmentation8 GB 20 sDreamGaussian0.8078 GB 120 s
InpaintingTripoSR0.7956 GB < 1s
Individual Generation6 GB < 1sWonder3D0.80116 GB 180 s
Individual Generation8 GB 120 sLRM0.81214 GB 20 s
Layout Alignment6 GB 90 s
Total≤8GB 120 sREPARO (ours)0.833≤8GB240 s
Total≤8GB 240 sREPARO (ours)0.822≤8GB120 s
+ +For layout alignment, we use Nvdiffrast [19] as our differentiable rendering framework. During the differentiable rendering process, we optimize the translation $x, y, z$ , and rotation $r$ parameters of each 3D asset using gradient descent. Specifically, we employ the Adam optimizer [17] with a learning rate of 0.02 and a weight decay of 0.999. The total number of iterations is set to 500. All experiments were conducted on RTX 3090 GPUs, and as shown in Figure 3, optimizing parameter $t$ alone takes approximately 20 seconds and utilizes 4GB of memory, while optimizing both parameters $t$ and $s$ to their optimal states requires about 90 seconds and utilizes 6GB of memory. We report per-stage runtime and memory in Table 2. All modules run on $\leq 8\mathrm{GB}$ VRAM, covering SAM, inpainting, 3D generation, and layout alignment. Table 3 compares our method with baselines in runtime and memory. REPARO achieves better performance under similar or lower resource budgets. + +To validate the effectiveness of our method, we selected 20 samples containing multiple objects from the Google Scanned Objects (GSO) dataset [8] as our test set. We used images from 18 views to compute quantitative metrics, including CLIP score [44], Peak signal-to-noise ratio (PSNR), the structural similarity (SSIM) [58], and Perceptual Similarity (LPIPS) [69]. Moreover, we also conducted a subjective evaluation of the generation quality for different models. We collected opinions from 40 participants. Participants selected the visually best subjective quality option among the generation results of 4 models for the same input image. We computed the preference score for each model by dividing the selection count of a specific model by the total selections, reflecting the human preference distribution for each model. + +![](images/e7dd96afd57782150021626984b7446e15faa573f656fb63987f05d41f296bc0.jpg) +Figure 4. Qualitative comparison with different image-to-3D generation models. Given an input image, previous methods produce inaccurate textures and geometry with noticeable artifacts. Our method generates high-quality, high-fidelity compositional assets with the correct spatial layout. + +Table 4. Comparison of different loss functions. + +Table 5. Comparison of optimization parameters. + +Table 6. Comparison of information in appearance loss $L_{a}$ + +(a) DreamGaussian-based REPARO. + +
LaLsCLIP ↑PSNR ↑SSIM ↑LPIPS ↓
✓0.83317.2960.8260.233
✓0.80917.8200.8490.210
✓✓0.83317.2790.8260.234
+ +(b) TripoSR-based REPARO. + +
LaLsCLIP ↑PSNR ↑SSIM ↑LPIPS ↓
✓0.82217.7650.8650.216
✓0.81317.9060.8670.212
✓✓0.82217.7510.8650.216
+ +(a) DreamGaussian-based REPARO. + +
TRSCLIP ↑PSNR ↑SSIM ↑LPIPS ↓
✓0.83217.3690.8260.232
✓✓0.82616.9060.8200.243
✓✓0.83317.2790.8260.234
✓✓✓0.83117.1220.8230.241
+ +(b) TripoSR-based REPARO. + +
TRSCLIP ↑PSNR ↑SSIM ↑LPIPS ↓
✓0.81917.7970.8650.216
✓✓0.82217.7680.8640.219
✓✓0.82217.7510.8650.216
✓✓✓0.82617.7520.8640.219
+ +(a) DreamGaussian-based REPARO. + +
Info in LaOTCLIP ↑PSNR ↑SSIM ↑LPIPS ↓
RGB×0.80018.0540.8530.206
RGB✓0.82817.0090.8210.245
RGBXY✓0.82817.0690.8210.243
RGBDXY✓0.83317.2790.8260.234
+ +(b) TripoSR-based REPARO. + +
Info in LaOTCLIP ↑PSNR ↑SSIM ↑LPIPS ↓
RGB×0.77417.9180.8680.216
RGB✓0.81917.8370.8650.215
RGBXY✓0.82317.8470.8650.213
RGBDXY✓0.82217.7510.8650.216
+ +# 4.2. Main results + +Quantitative experiment. Table 7 compares the performance (CLIP score, PSNR, SSIM, and LPIPS) of various methods for reconstructing 3D assets of multiple objects using one reference image. DreamGaussian-based REPARO and Triposr-based REPARO, our proposed methods, show significant improvements over baseline methods (DreamGaussian [51], Wonder3D [33], and LRM [13]). DreamGaussian-based REPARO achieves $83.3\%$ CLIP score, 17.279 PSNR, 0.826 SSIM, and 0.234 LPIPS, indicating better alignment between reconstructed 3D assets and ground-truth 3D assets. Triposr-based REPARO exhibits superior performance in PSNR (17.751) and SSIM (0.865) and the lowest LPIPS (0.216), highlighting its exceptional quality in image reconstruction and perceptual similarity. These results show that better reconstruction quality on single objects is beneficial for structural similarity and perceptual similarity during differentiable rendering. Qualitative experiment. Some qualitative comparison results are shown in Figure 4, where our Triposr-based REPARO can produce high-quality 3D assets with multiple objects. By contrast, DreamGaussian [51], Wonder3D [33], and LRM [13] have problems in layout and completeness, since they possess preferences on an individual object. + +User study. In addition to numerical metrics, we conducted a user study to compare our method with others, gathering responses from participants who assessed the realism of different 3D assets. Our approach received $61\%$ approval, while Dreamgaussian received $19\%$ approval, LRM received $12\%$ approval, and Wonder 3D received $8\%$ approval. The results consistently showed that our approach was favored over previous methods. For more detailed results, please refer to the supplementary material. + +# 4.3. Ablation study + +Ablation for different loss functions in differentiable rendering. We conducted ablation experiments on different + +![](images/201e4c9fb7869704cbae003ba8a257dc291cfd9e9bfa7bd5bd9286f536782b77.jpg) +Input + +![](images/09e71c4cff820d36236dbbe07b305fc73d8091086467c1d90afa85ff5a251c6f.jpg) +$\theta = \{\mathbf{t}\}$ + +![](images/38b302201e765db3e3bce85ebab0da2ae3d0cf0d8a09b277c7734698e3d6cd09.jpg) +$\theta = \{\mathrm{t},\mathrm{r}\}$ + +![](images/71b6af0c4160f734f621658dd69f968e45113992135fb26ae73b2b6c5e1cf9a5.jpg) +$\theta = \{\mathrm{t},\mathrm{s}\}$ + +![](images/1e41f00fca25aea1ea8777f78a2bb789e94dfabba4e1b98969f7910e179f7f83.jpg) +$\theta = \{\mathrm{t},\mathrm{r},\mathrm{s}\}$ + +![](images/15f4b6b99ad228cbd5964457308f798365fdec77c80b9033ce5025990042264b.jpg) +Figure 5. Qualitative comparison for different options of optimization parameters $\theta$ . +Input +Figure 6. Qualitative comparison for different terms in the appearance loss $L_{a}$ (Eq. 5). + +![](images/1c579a75f9ba4ff4a5179c7ab2f34cc950ce00fb2f7683f96f1c2ebe0c7573da.jpg) +RGB-3D + +![](images/b11a575be0bb703193d01b5420d1b27d0d36baddb8adbc0291d158ed6c795101.jpg) +RGBXY-5D + +![](images/d2ab931279d0944a00f08c5928a14d1dcbe001c8df4afe14af86bbb743fb1b4f.jpg) +RGBDXY-6D + +![](images/c168a6fa12c528da4eb291b99766b6503cfe2bda421fc9599f7c35369980ca5d.jpg) +GT + +loss components in Eq. 9, with the results presented in Table 4. The results indicate that incorporating the high-level semantic loss function $L_{s}$ can bring improvements in both SSIM and LPIPS. However, the enhancement in CLIP score is not as pronounced. Additionally, when only $L_{s}$ is used, there is a notable decrease in CLIP score. For instance, in the DreamGaussian-based model, the CLIP score decreased by $2.1\%$ . This suggests that relying solely on feature supervision can reduce the effectiveness of layout alignment. + +Ablation for different options of scene parameters $\theta$ . Table 5 provides an ablation study on the impact of various parameter optimization settings within $\theta$ across both DreamGaussian-based and TripoSR-based REPARO models. Specifically, we compare combinations of translation $(t)$ , rotation $(r)$ , and scale $(s)$ optimizations. For DreamGaussian-based REPARO, the combination of parameters $t$ and $s$ achieves the highest CLIP score (0.833), competitive PSNR (17.279), and the lowest LPIPS (0.234), indicating a high level of perceptual quality and fidelity. Similarly, for TripoSR-based REPARO, optimizing $t$ and $s$ yields a balance between perceptual similarity (CLIP 0.822) and structural fidelity, with the lowest LPIPS (0.216) and strong SSIM (0.865), aligning closely with the input image. + +![](images/f7434435d271fc6c3d865677205e952e06208e02b1563aae346a97dc5b996f4e.jpg) +Figure 7. 3D scene editing based on the implementation of 2D image editing. Based on our REPARO framework, it is possible to implement four types of 3D editing guided by 2D image editing, namely adding, removing, position exchanging and instance replacement. + +Table 7. DreamGaussian-based (♣), TripoSR-based (♠), and GT-based (♦) results. + +
MethodCLIPs↑PSNR ↑SSIM ↑LPIPS ↓
REPARO♣0.83317.2790.8260.234
REPARO♣0.82217.7510.8650.216
REPARO♣0.90615.7080.7890.321
+ +Notably, as shown in Figure 5, including the rotation parameter $r$ in the optimization tends to degrade alignment with the input image orientation, as shown by the rotation-inclusive configurations, which generally exhibit slight reductions in CLIP scores and increased LPIPS. As a result, we select the combination of $t$ and $s$ for the final configuration, as it achieves optimal perceptual alignment without sacrificing the fidelity to the original image orientation. + +Ablation for long-range appearance loss term. As shown in Eq. 5, long-range appearance loss $L_{a}$ incorporates RGB color distance, position distance, and depth value distance. More importantly, we use the optimal transport to compute pixel matching from a global perspective, thereby obtaining long-range correspondences. Quantitative results in Table 6 demonstrate that in the DreamGaussian-based model, introducing long-range correspondence results in a $2.8\%$ improvement in CLIP score; in the TripoSR-based model, it results in a $4.5\%$ improvement in CLIP score. Additionally, when position distance and depth distance are introduced separately, there is no significant change in quantitative results, but there is a noticeable improvement in qualitative results, as illustrated in Figure 6. + +The impact of 3D asset quality. As shown in Tab. 7, the experimental results indicate that the process of generating individual assets in the first stage introduces significant errors. Therefore, we still need a model capable of generating high-quality 3D assets. + +# 4.4. Extension: Using 2D editing to guide 3D editing + +We further explored potential applications of our combinatorial generation framework. As depicted in the Figure 7, the entire framework relies on a single input image as guidance, enabling both the reconstruction of individual objects and layout alignment through differentiable rendering. Leveraging this capability, we integrated pre-existing image editing techniques to facilitate editing of the generated 3D scenes. Specifically, we demonstrate four types of editing operations: adding, removing, position exchange, and instance replacement. The 2D editing process is implemented using the AnyDoor [2] model. By comparing the 3D scenes before and after editing, it is evident that the REPARO framework exhibits robust spatial arrangement capabilities and maintains consistency for each assets. + +# 5. Conclusion + +In conclusion, we introduce REPARO, a comprehensive approach for generating compositional 3D scenes from single images. Our REPARO addresses the multi-object issue by decomposing the scene into individual objects, reconstructing their 3D meshes, and optimizing their layout through differentiable rendering. During differentiable rendering, we incorporate an optimal transport-based long-range appearance loss which considers RGBDXY information, and a high-level semantic loss which aligns the feature correspondence. Consequently, REPARO is able to obtain multi-object 3D assets with visual-spatial arrangement and contextual consistency. One limitation of our method is that severe occlusion may lead to hallucinated parts inconsistent with the image, and in semantically implausible scenes, layout optimization tends to favor common-sense interpretations over the actual image. We hope our method can pave the way for image to 3D generation in multi-object scenes. + +# References + +[1] 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), 2023. 2 +[2] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization, 2024. 8 +[3] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets creation using spatially-aware diffusion guidance. arXiv preprint arXiv:2403.12409, 2024. 2 +[4] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting, 2024. 2 +[5] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 4 +[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects, 2022. 2 +[7] 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. arXiv preprint arXiv:2307.05663, 2023. 2 +[8] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5 +[9] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 2 +[10] Thomas Funkhouser, Michael Kazhdan, Philip Shilane, Patrick Min, William Kiefer, Ayellet Tal, Szymon Rusinkiewicz, and David Dobkin. Modeling by example. 23(3):652-663, 2004. 3 +[11] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[12] Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 303-312, 2021. 4 +[13] 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. 5, 7 + +[14] Stable Diffusion Infinity. Stable diffusion infinity. https://github.com/lkwq007/stablediffusion-infinity, 2023. Accessed: 2023-10-01. 5 +[15] 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 +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2 +[17] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, ICLR, 2015. 4, 5 +[18] 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, 5 +[19] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 3, 5 +[20] Chenghao Li, Chaoning Zhang, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung-Ho Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 1 +[21] Ming Li, Pan Zhou, Jia-Wei Liu, Jussi Keppo, Min Lin, Shuicheng Yan, and Xiangyu Xu. Instant3d: Instant text-to-3d generation. International Journal of Computer Vision, pages 1-17, 2024. 2 +[22] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 3 +[23] Wanhua Li, Renping Zhou, Jiawei Zhou, Yingwei Song, Johannes Herter, Minghan Qin, Gao Huang, and Hanspeter Pfister. 4d langsplat: 4d language gaussian splatting via multimodal large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 22001-22011, 2025. 2 +[24] Zhiqi Li, Yiming Chen, Lingzhe Zhao, and Peidong Liu. Controllable text-to-3d generation via surface-aligned gaussian splatting, 2024. 2 +[25] Jingbo Zhang3 Zhihao Liang4 Jing Liao, Yan-Pei Cao, and Ying Shan. Advances in 3d generation: A survey. arXiv preprint arXiv:2401.17807, 2024. 1 +[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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 300–309, 2023. 1, 2 +[27] Yukang Lin, Haonan Han, Chaoqun Gong, Zunnan Xu, Yachao Zhang, and Xiu Li. Consistent123: One image to + +highly consistent 3d asset using case-aware diffusion priors. arXiv preprint arXiv:2309.17261, 2023. 1, 2 +[28] Yiqi Lin, Hao Wu, Ruichen Wang, Haonan Lu, Xiaodong Lin, Hui Xiong, and Lin Wang. Towards language-guided interactive 3d generation: Llms as layout interpreter with generative feedback. arXiv preprint arXiv:2305.15808, 2023. 3 +[29] 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. 2 +[30] 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 +[31] 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. 1, 2 +[32] 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. 2 +[33] 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. arXiv preprint arXiv:2310.15008, 2023. 2, 5, 7 +[34] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: 360 reconstruction of any object from a single image. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2 +[35] 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 Proceedings of the European conference on computer vision (ECCV), pages 405-421, Cham, 2020. Springer International Publishing. 1, 2 +[36] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, 2022. +[37] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2 +[38] Maxime Oquab, Timothee Darcet, Theo 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. 5 +[39] Ryan Po and Gordon Wetzstein. Compositional 3d scene generation using locally conditioned diffusion. arXiv preprint arXiv:2303.12218, 2023. 3 + +[40] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2 +[41] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations (ICLR), 2023. 2 +[42] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2 +[43] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 2 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5 +[45] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2349–2359, 2023. 2 +[46] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Hyeonsu Kim, Jaehoon Ko, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation, 2024. 2 +[47] 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. arXiv preprint arXiv:2310.15110, 2023. 2 +[48] Yukai Shi, Jianan Wang, He Cao, Boshi Tang, Xianbiao Qi, Tianyu Yang, Yukun Huang, Shilong Liu, Lei Zhang, and Heung-Yeung Shum. Toss: High-quality text-guided novel view synthesis from a single image. arXiv preprint arXiv:2310.10644, 2023. 2 +[49] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2 +[50] Ryo Suzuki, Adnan Karim, Tian Xia, Hooman Hedayati, and Nicolai Marquardt. Augmented reality and robotics: A survey and taxonomy for ar-enhanced human-robot interaction and robotic interfaces. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1–33, 2022. 1 +[51] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 1, 2, 3, 5, 7 + +[52] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22819-22829, 2023. 1, 2 +[53] Konstantinos Tertikas, Despoina Paschalidou, Boxiao Pan, Jeong Joon Park, Mikaela Angelina Uy, Ioannis Emiris, Yannis Avrithis, and Leonidas Guibas. Generating part-aware editable 3d shapes without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4466-4478, 2023. 3 +[54] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 2, 3, 5 +[55] Alexander Vilesov, Pradyumna Chari, and Achuta Kadambi. Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907, 2023. 3 +[56] Can Wang, Mengelei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3825–3834, 2022. 2 +[57] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation, 2022. 2 +[58] 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. 5 +[59] Zhaoning Wang, Ming Li, and Chen Chen. Luciddreaming: Controllable object-centric 3d generation. arXiv preprint arXiv:2312.00588, 2023. 3 +[60] 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. 1, 2 +[61] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, CL Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3d object synthesis. arXiv preprint arXiv:2310.08092, 2023. 2 +[62] Jiankai Xing, Fujun Luan, Ling-Qi Yan, Xuejun Hu, Houde Qian, and Kun Xu. Differentiable rendering using rgbxy derivatives and optimal transport. ACM Transactions on Graphics (TOG), 41(6):1-13, 2022. 3, 4 +[63] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20908–20918, 2023. 2 +[64] Han Yan, Yang Li, Zhennan Wu, Shenzhen Chen, Weixuan Sun, Taizhang Shang, Weizhe Liu, Tian Chen, Xiaqiang Dai, Chao Ma, et al. Frankenstein: Generating semantic-compositional 3d scenes in one tri-plane. arXiv preprint arXiv:2403.16210, 2024. 3 + +[65] 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. In CVPR, 2024. 2 +[66] 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 +[67] Xueling Yuan, Hongyu Yang, Yueming Zhao, and Di Huang. Dreamscape: 3d scene creation via gaussian splatting joint correlation modeling. arXiv preprint arXiv:2404.09227, 2024.3 +[68] Junwu Zhang, Zhenyu Tang, Yatian Pang, Xinhua Cheng, Peng Jin, Yida Wei, Wangbo Yu, Munan Ning, and Li Yuan. Repaint123: Fast and high-quality one image to 3d generation with progressive controllable 2d painting. arXiv preprint arXiv:2312.13271, 2023. 2 +[69] 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. 5 \ No newline at end of file diff --git a/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/images.zip b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0ff10b65054cd961dc5f7771801a7265b68d93a9 --- /dev/null +++ b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60bcb21f2c218c81c2a0fcd42ec6dd687abf580b6d22b1a97345c53cc0cbe0da +size 707513 diff --git a/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/layout.json b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..832abe36a99eb228e7cb8a03b5826060534b683d --- /dev/null +++ b/ICCV/2025/REPARO_ Compositional 3D Assets Generation with Differentiable 3D Layout Alignment/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:186d567af60780100ad2a207b4f2129396a34bdf308e7f62baad35a15de53137 +size 462796 diff --git a/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/38368614-6c60-4c3b-89b7-3b5161f0d296_content_list.json b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/38368614-6c60-4c3b-89b7-3b5161f0d296_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a106d9e2af1357b04e93f3319708439a9d31e3a3 --- /dev/null +++ b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/38368614-6c60-4c3b-89b7-3b5161f0d296_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7251a8aa52d174090e365d47679217d02cd95f056cd4e0978dddcc93bd83a0a4 +size 70744 diff --git a/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/38368614-6c60-4c3b-89b7-3b5161f0d296_model.json b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/38368614-6c60-4c3b-89b7-3b5161f0d296_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f633aa553f8fba6319a1490391acc15a4c681fa0 --- /dev/null +++ b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/38368614-6c60-4c3b-89b7-3b5161f0d296_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c3ac6b911ad5ee8de8e78f304421f5b61d217bc604a9a9185e7790462e0b73d +size 89778 diff --git a/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/38368614-6c60-4c3b-89b7-3b5161f0d296_origin.pdf b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/38368614-6c60-4c3b-89b7-3b5161f0d296_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..478f1e386f5694aa06515a738543feb5f64b93a9 --- /dev/null +++ b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/38368614-6c60-4c3b-89b7-3b5161f0d296_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a14a63fc158b991807ad116609bfde8e5e407f36a17cb5e60576c6d40e7c3ac +size 23968655 diff --git a/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/full.md b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d498eeba8025b4fe1612dfcde6d1dc874cb43474 --- /dev/null +++ b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/full.md @@ -0,0 +1,293 @@ +# RESCUE: Crowd Evacuation Simulation via Controlling SDM-United Characters + +Xiaolin Liu $^{1,\dagger}$ , Tianyi Zhou $^{1,\dagger}$ , Hongbo Kang $^{1}$ , Jian Ma $^{1}$ , Ziwen Wang $^{1}$ , Jing Huang $^{1}$ , Wenguo Weng $^{2}$ , Yu-Kun Lai $^{3}$ , Kun Li $^{1,*}$ + +$^{1}$ Tianjin University, $^{2}$ Tsinghua University, $^{3}$ Cardiff University + +{liuxiaolin, tianyizhou, hbkang, jianma, 3021244417, hj00, lik}@tju.edu.cn, + +wgweng@tsinghua.edu.cn, Yukun.Lai@cs.cardiff.ac.uk + +$^{\dagger}$ Equal contribution *Corresponding author + +![](images/e70bc7a1c059ca882309e9893db1ca05b6ecde3fff64ea2806d1a59a97fa3ce9.jpg) +Figure 1. Our SDM-United Framework enables online simulation of personalized, physically realistic, and 3D-adaptive multi-agent evacuation scenarios. + +# Abstract + +Crowd evacuation simulation is critical for enhancing public safety, and demanded for realistic virtual environments. Current mainstream evacuation models overlook the complex human behaviors that occur during evacuation, such as pedestrian collisions, interpersonal interactions, and variations in behavior influenced by terrain types or individual body shapes. This results in the failure to accurately simulate the escape of people in the real world. In this paper, aligned with the sensory-decision-motor (SDM) flow of the human brain, we propose a real-time 3D crowd evacuation simulation framework that integrates a 3D-adaptive SFM (Social Force Model) Decision Mechanism and a Personalized Gait Control Motor. This framework allows multiple agents to move in parallel and is suitable for various scenarios, with dynamic crowd awareness. Additionally, we introduce Part-level Force Visualization to assist in evacuation analysis. Experimental results demonstrate that our framework supports dynamic trajectory planning and personalized behavior for each agent throughout the evacuation process, and is compatible with uneven terrain. Visually, our method generates evacuation results that are more realistic and plausible, providing enhanced insights for crowd simulation. The code is available at http://cic.tju.edu.cn/faculty/likun/projects/RESCUE. + +# 1. Introduction + +Evacuation simulations serve as valuable tools for as- + +sessing the likelihood of crowding and trampling incidents, estimating evacuation times, and supporting virtual reality escape training. However, no existing methods can simulate personalized and proxemics-aware 3D movements of hundreds of people online. This paper aims to develop an adaptive multi-agent 3D evacuation simulation framework. As shown in Figure 1, our framework, called RESCUE (cRowd Evacuation Simulation via Controlling SDM-United characters), is capable of achieving online physically realistic simulations performing avoidance and personalized gait. + +Traditional crowd simulation methods [3, 6, 22] have explored various strategies in different scenarios. However, due to their simplified representations, these methods fail to integrate 3D motion with realistic behavior, resulting in physically implausible actions. Most crowd simulation methods cannot make decisions based on scene changes and are not designed for crowd evacuation simulation. Based on physics engines [18], character control [4, 26, 29, 36, 37] achieves autonomous 3D motion. However, these methods face the lack of personalization motions and may result in falls and collisions in densely populated scenarios. Some motion generation methods [10, 31, 32, 43] based on diffusion models can generate diverse 3D motions, but they still face challenges in controllability and physical realism. Neither of these methods can directly produce realistic, personalized, and online evacuation motions. The detailed capabilities of the methods are summarized in Table 1. + +The limitations of the existing methods are mainly due to + +Table 1. Comparison with Related Methods + +
Related WorkTraditional Crowd Simulation [3, 6, 7, 22]Path-Guided Motion Generation [27, 39, 43]Path-Guided Character Control [29, 37, 41]Our Method
Online Path PlanningPartialXX✓
3D Human ModelsX✓✓✓
Personalized GaitX✓X✓
Collision VisualizationXXX✓
Uneven TerrainsXX✓✓
+ +the significant behavioral and environmental complexities inherent in evacuation scenarios: 1) 3D Proxemics-aware Ability: Crowded pathways and frequent physical interactions require the use of online 3D movement decision to enable dynamic strategy adjustments, such as avoidance of collisions and maintaining stability; 2) Personalized Gait: Individuals with different attributes, such as older or disabled persons, exhibit varying behaviors in identical situations, adapting to the scene with different speed and performance. + +Neuroscientific studies [17, 20, 35] demonstrate that humans employ a Sensory-Decision-Motor (SDM) loop [28], integrating environmental cues through distributed neural networks to evaluate and dynamically adjust behavior in complex environments. In this paper, we propose an online SDM-United 3D Evacuation Simulation Framework that integrates personalized decision-making with character control in a physics engine. This allows each agent to perceive its surroundings and make online adjustments. To handle congestion, we propose a 3D-Adaptive Social Force Model for decision-making and obstacle avoidance in 3D environments. In particular, we introduce an evasive force that adapts to dense scenarios to avoid congestion. + +To achieve behavioral diversity, we introduce a Proxemics-aware Personalization Method with a personalized gait controller and an optimized SFM to generate individualized behaviors for agents based on attributes such as ages and physical conditions. Additionally, we introduce Part-level Force Visualization, offering unprecedented insights into contact forces. Our experiments demonstrate that the proposed framework effectively adapts to uneven terrain, maintains balance during collisions and crowd congestion, and generates personalized behaviors. Although our experiments have focused on evacuation scenarios, the framework readily generalizes to a wide variety of multi-agent environments, such as autonomous driving simulations in which dynamic vehicles and diverse pedestrians coexist. Our contributions can be summarized as follows: + +- We propose the first online SDM-United 3D Evacuation Simulation Framework, with online decision-making and physics-based motion generation. This proxemics aware framework can be seamlessly extended to arbitrary dynamic environments and supports real time parallel simulation of hundreds of agents. +- We propose a 3D-adaptive Social Force Model, designing optimized forces and personalized coefficients to enable accurate decision-making in 3D environments. This approach ensures correct guidance for the speed of agents. + +- We propose a Personalized Gait Controller, enabling agents to generate personalized motions tailored to attributes such as age and physical condition. Additionally, a Part-level Force Visualization is designed for analysis. +- We validate the framework in various evacuation scenarios, demonstrating its ability to simulate diverse evacuation events, including the effects of crowd density, corridor width and terrain influence trampling incidents, as well as crowd behaviors like collisions and falls. + +# 2. Related Work + +# 2.1. Crowd Simulation + +Crowd simulation is a tool for analyzing and optimizing crowd dynamics. The aim is to capture key phenomena such as bottlenecks and panic-induced behaviors. Macroscopic models describe collective crowd movement using fluid dynamics [5, 8, 11, 30], while providing detailed insights through approaches. Social Force Models (SFM) [14, 15, 34] are among the most widely used methods, treating pedestrians as Newtonian particles influenced by the forces driving acceleration. Additional forces such as panic, sliding friction, and body compression have been introduced [6, 48]. However, these refinements often target specific scenarios. With advances in computational power, Agent-Based Models (ABM) have gained popularity [3, 7], simulating individuals as autonomous agents with unique decision-making capabilities. Although the above-mentioned methods attempt to simulate real-world escape, they rely on simplified 2D representations, where force feedback can only use unrealistic surrogate computations, leading to distorted speed outputs. Although several attempts at physically plausible 3D simulations have emerged [1, 9, 13, 46], these approaches still fall short in integrating personalized motion control and decision-making mechanisms. Most methods rely on generic movement patterns that overlook individual variations and fail to construct comprehensive strategies for handling crowd dynamics and escape scenarios. + +# 2.2. Character Control + +The practicable controllers for physics-based character simulation remain a challenge in animation. Initial methods predominantly rely on carefully designed control structures, such as finite state machines and trajectory optimization, to achieve specific motions. These approaches demonstrate the feasibility of simulating a wide range of behaviors, but their reliance on task-specific engineering makes them inflexi- + +![](images/ba111a2c4b3071ea663a28b8c52baddc1da55499dc8563403e828a75af7728c2.jpg) +Figure 2. The detailed architecture of RESCUE. Our framework simulates the neuroscience-validated paradigm (Sensory-Decision-Motor) to achieve a realistic and personalized crowd evacuation simulation. Specifically, our framework includes: (a) Sensory: each agent senses observations, including the fully observable self-state, the partially observable other-state, and the environment state; (b) Decision: each agent uses Social Force Model with personalized SFM coefficients to decide its speed in the next timestep, which is then used to compute path condition; and (c) Motor: the path condition and personal attribute of each agent are conditioned to simulate locomotion with a personalized gait controller. The simulation then updates the states of all agents, allowing them to be sensed by the sensory module. + +ble and time consuming to adapt to new tasks [4]. Recent advances in deep reinforcement learning (DRL) greatly expand the capabilities of physics-based characters, enabling them to perform diverse tasks, such as walking and running [24, 44], and recovering from disturbances [16]. Designing effective reward functions can produce natural motions, but remains challenging. To address this, motion imitation methods [25, 38] gain popularity, where controllers learn from reference data, either through direct tracking or adversarial learning, to produce lifelike behaviors. Furthermore, hierarchical control frameworks are proposed to combine low-level imitation with high-level motion planning, enabling characters to execute complex sequences of actions [29, 37]. However, existing methods struggle to generate realistic behaviors in densely populated scenes, lack online decision-making capabilities, and are insufficient in generating diverse and personalized motions. + +# 2.3. Motion Generation + +Motion generation aims to synthesize realistic and varied human motions. Recent advances in diffusion models have significantly improved the diversity and quality of motion synthesis [31, 32]. Key research directions include text-driven motion generation, stylized motion synthesis [2, 50], scene-aware motion generation [23, 40, 49], and human-object interaction modeling [45]. Recent efforts [33, 47] have also integrated physical constraints to enhance the realism of generated motions. Trajectory-controlled [27, 39, 43] and multi-person motion synthesis [42] have been explored. However, challenges remain in generating long-duration, physically plausible motions, particularly in multiperson, contact-rich scenarios. In addition, these methods are usually time consuming and not suitable for online applications. + +In conclusion, all existing methods fail to accurately simulate 3D multi-agent with personalized behaviors in densely populated escape scenarios. Therefore, we propose the first SDM-unified evacuation simulation framework to generate + +realistic, personalized, and online 3D evacuation motions. + +# 3. Method + +# 3.1. Overview + +The structure of our framework is shown in Figure 2. Our goal is to simulate realistic evacuation scenarios with diverse individuals. Given a scene mesh, an exit location and initial positions, the framework generates a personalized evacuation animation. Notably, it is suitable for evacuation tasks by incorporating human-like paradigm, while ensuring personalized decisions and motions. + +In Section 3.2, we introduce our SDM-United Framework, comprising the Sensory, Decision, and Motor modules. The Sensory module provides self-awareness, partial awareness of others, and exit perception. Our 3D-adaptive SFM Mechanism (Section 3.3) enhances decision-making and obstacle avoidance in 3D environments. Proxemics-aware Personalization is achieved through both the Decision and Motor modules: the Decision module, with Personalized Optimization for SDM Coefficients (Section 3.3), computes speeds tailored to individual attributes (Section 3.3), while the Motor module, with the Personalized Gait Controller (Section 3.4), generates locomotion with personalized gaits. We also introduce the Visualization of Part-Level Forces (Section 3.5) to aid analysis. + +# 3.2. SDM-United Framework + +The human brain follows a workflow of sensory, decision, and motor processes [28], and hence we develop our framework. Each person in the evacuation process is abstracted as an agent that operates in an online loop, where the sensory module gathers observations, the decision module computes speeds, and the motor module executes motions, with updated states feeding back into the sensory module for continuous processing and making them available for next cycle. + +Sensory. Each agent perceives its self-state $s^{self}$ , the partially observable other-state $s^{other}$ , and the environment state $s^{envir}$ . At time $t$ , $s_{i,t}^{self}$ includes the agent's position, speed, and its humanoid state $h_{i,t}$ . The other-state $s_{i,t}^{other}$ contains other agents' root positions relative to the agent, and the environment state $s_{i,t}^{envir}$ . The form of $h_{i,t}$ and $s_{i,t}^{envir}$ follows [29]. + +Decision. The Social Force Model (SFM) [14, 15] models pedestrian behavior, including panic and crowd dynamics. We propose a 3D-adaptive SFM Decision Mechanism to compute the desired speed $\tilde{\mathbf{v}}_{i,t + 1}$ . The key points of the escape path are found via A* search [12], and we use the SFM to drive each agent toward the key points while avoiding collisions, resulting in the computation of the agent's speed $\tilde{\mathbf{v}}_{i,t + 1}$ . SFM coefficients are optimized based on individual attributes to enable personalized decisions. + +Motor. We use Pacer [29] for path-following, which computes the motion $\pmb{a}_{i,t}$ based on the speed $\tilde{\mathbf{v}}_{i,t+1}$ . This motion is non-personalized, so a Personality Gait Controller assigns a personalized gait based on attributes. The resulting motion is simulated in the physics engine and then sensed in by each agent for continuous processing. + +# 3.3. 3D-adaptive SFM Decision Mechanism + +In 3D simulations, interaction dynamics differs from 2D point-based models due to physics engine integration. While 3D collision handling prevents interpenetration, human-to-human interactions require substantial adaptation—agents in 3D may stumble or fall under congestion that would merely slow movement in 2D models. Effective navigation requires detouring behavior rather than waiting or collision. To address these challenges, we introduce personalized SFM coefficient optimization and evasive forces specifically for 3D environments. Our decision process transforms agent $i$ 's observations (position, speed, and nearby agents) into desired speed $\tilde{v}_{i,t+1} \in \mathbb{R}^3$ for the next timestep. + +Base Forces. Our implementation enhances the basic social force model [14] with two essential components. The driving force guides agents efficiently toward destinations by first employing A* pathfinding [12] to generate optimal routes, then calculating acceleration based on the discrepancy between desired and current speeds, smoothed by a relaxation parameter. The complementary repulsive force maintains safe inter-agent distances through exponential repulsion that intensifies as proximity increases, with magnitude controlled by repulsion coefficient, spatial decay, and interaction radius parameters. + +Evasive Force. We propose the evasive force to evade the front agents stopping or falling due to collisions. By computing perpendicular directions to the desired path, the agent can detour efficiently. The force is calculated by considering a $45^{\circ}$ sector in front of the agent, representing the perceived area. The evasive force $\pmb{F}_{\mathrm{evasive}}$ is defined as: + +$$ +\boldsymbol {F} _ {\text {e v a s i v e}} = A \operatorname {s g n} \left(\boldsymbol {o} _ {i} \cdot \boldsymbol {p} _ {i}\right) \boldsymbol {p} _ {i}, \tag {1} +$$ + +where $A$ is a binary mask that equals 1 when there is an obstacle in front of the agent and there is available space to the side, and 0 otherwise. $\mathbf{p}_i$ is the perpendicular vector to the desired direction, and $\mathbf{o}_i$ represents the average position of nearby agents. This improvement aligns the SFM with the requirements of 3D simulations. + +Personalized Optimization for SDM Coefficients. In crowd evacuation simulations, it seems not reasonable to use uniform social force coefficients, as individuals have varying escape abilities based on their attributes. Personalized coefficients better reflect these differences. Moreover, factors, for example friction in the physics engine, can cause individuals to move more slowly than the intended speed. To compensate for these, we adjust the target speed setting to achieve the desired speed in simulation. Specifically, $v_{\mathrm{real}}$ is the empirically obtained real-world escape speed from literature (see Supplementary), $v_{\mathrm{sim}}$ is the simulated speed, and $v_{\mathrm{setting}}$ is the speed input to the engine. We run agents of different attributes on a straight path in the physics engine and optimize $v_{\mathrm{setting}}$ so that $v_{\mathrm{sim}}$ approximates $v_{\mathrm{real}}$ . We categorize agents into five groups, i.e. the young, middle-aged, elderly, patients, and disabled1. Through this process, we determine personalized coefficients with an accuracy of 0.005 m/s for use in simulation. The algorithm details are presented in the Supplementary Materials. + +Final Decision. The final desired speed $\tilde{v}_{i,t + 1}$ is determined by combining the forces: + +$$ +\tilde {\boldsymbol {v}} _ {i, t + 1} = \boldsymbol {v} _ {i} + \Delta t \left(\boldsymbol {F} _ {\text {d r i v e}} + \boldsymbol {F} _ {\text {r e p u l s i v e}} + \boldsymbol {F} _ {\text {e v a s i v e}}\right), \tag {2} +$$ + +where $\Delta t$ is the timestep size, and $\pmb{v}_i$ is the current speed of agent $i$ . This framework ensures a smooth, collision-free, and efficient movement toward the destination. + +# 3.4. Personalized Gait Control + +We propose a Personalized Gait Controller that generates personalized escape gaits by transforming unpersonalized actions to personalized actions based on agent attributes. This section corresponds to the Motor process in Section 3.2. Pacer [29], a robust physics-based controller, serves as our backbone for path-following. Its policy $\pi_{\mathrm{PACER}}$ tracks 2D trajectories $\tau_{i,t}$ , sampled based on the desired speed. Given the state $S$ (agent's position, humanoid state, and environment state), the action $a_{i,t}$ is computed by $\pi_{\mathrm{PACER}}$ . + +People with different attributes have different gaits. We correspond these gaits to a finite number of personalization styles in the 100style dataset [19]. This converter is a diffusion-based generative model, taking personalized gait labels, characterless action frames, and randomly sampled Gaussian noise as inputs, and outputs personalized action + +frames given to the characterless action frames corresponding to the personalized gait labels. This requires the characterless action frames as inputs and the personalized action frames as ground-truths during training should be matched. + +Gait Frame Matching. Inspired by [21], a gait cycle can be fundamentally divided into 8 events [21]. We only use 4 events which are Initial Contact, Mid Stance, Opposite Initial Contact and Feet Adjacent. By calculating the distance between the two ankles, the image can be obtained similar to the sine-cosine function. The four gait events align with two consecutive cycles of the ankle distance waveform, corresponding to the peak, trough, peak, and trough, respectively. To ensure gait connectivity for gait frame matching, we assign gait values for all gait frames. We identify the keyframes of four events based on the above pattern and assign gait values of 0, 0.3, 0.5, and 0.75, while linearly interpolating the non-keyframes between them. We match personalized gait frames (excluding neutral style data in 100STYLE [19]) and non-personalized gait frames (neutral style data in 100STYLE [19]) which have the same gait values as data pairs. The algorithm details are presented in the Supplementary Materials. When forming multiple candidate data frames, we select the two frames with the closest joint angles to match. During training, we also perform data augmentation by randomly assigning the same rotation to the root joints of the data pair. + +![](images/5c727cd66399f94d7e2741601c5050c9f29ad646701fd76952f1129222e6f892.jpg) +Figure 3. Personalized Gait Converter. + +Network of Personalized Gait Converter. We use the CAMDM network [2] as the backbone to implement the nonpersonalized action frame $\pmb{a}_{i,t}$ to personalized action frame $\tilde{\pmb{a}}_{i,t}$ , see Figure 3. This network is a probability diffusion model. At each denoising step, the model takes as input a noisy motion sample $\pmb{a}_{i,t}^{t}$ , diffusion step $\mathbf{t}$ , along with personalized gait label $c$ , then learns to predict the original clean $\tilde{\pmb{a}}_{i,t}^{0}$ (see Figure 3): + +$$ +\boldsymbol {a} ^ {0} = G (\boldsymbol {a} ^ {t}, \mathbf {t}; c). \tag {3} +$$ + +The gait converter only involves single frame action conversion, and we use MSE loss $\mathcal{L}_{\mathrm{samp}}$ and 3D joint position loss + +$\mathcal{L}_{\mathrm{pos}}$ . $\lambda_{\mathrm{samp}}$ and $\lambda_{\mathrm{pos}}$ are hyperparameters that control the weights of $\mathcal{L}_{\mathrm{samp}}$ and $\mathcal{L}_{\mathrm{pos}}$ in the total loss function: + +$$ +\mathcal {L} = \lambda_ {\text {s a m p}} \mathcal {L} _ {\text {s a m p}} + \lambda_ {\text {p o s}} \mathcal {L} _ {\text {p o s}}. \tag {4} +$$ + +Lower body movement is affected by terrain, and upper body movement is more flexible [41]. Hence, in the simulation of the physics engine, we replace the upper body actions of non-personalized action frames $\mathbf{a}_{i,t}$ with the upper body actions of personalized action frames $\tilde{\mathbf{a}}_{i,t}$ . The simulation of actions in the physics engine can be sensed by the sensory process for the decision and motor in the next timestep. + +# 3.5. Visualization of Part-Level Forces + +Each physical agent consists of 24 distinct body parts. Each body part contributes to the overall motion of the physical character. The interactions between these characters, as well as their contact with the environment, generate forces that drive their movements. In evacuation scenarios, assessing the forces acting on each individual is crucial for mitigating potential hazards in crowded situations. We integrate force sensors into each body part of the characters, which are capable of recording the magnitude of contact forces exerted at every timestep. To visualize the forces acting on part-level, we design a color-based mapping approach. Specifically, a gradient is employed, with lighter colors indicating lower forces and darker colors representing higher forces. + +To validate the effectiveness of our force visualization method, we conduct part-level collision simulations under various scenarios. Some examples are shown in Figure 4. + +![](images/e19678fc23d64f77e0927c3bb730245e916da5bfe6c660c9a9b2b22ab7bcf80e.jpg) +Figure 4. Visualization of forces under multiple collisions. + +# 4. Experimental Results + +# 4.1. Comparison Methods + +Since there is no personalized 3D evacuation simulation method, we compare our approach with three related methods: Social Force Model [14, 15], OmniControl [43] and MaskedMimic [37]. Among the methods considered, only the Social Force Model [14, 15] is originally designed for crowd simulation, while OmniControl [43] is a motion generation method and MaskedMimic [37] is a character control method. We adapt OmniControl [43] and MaskedMimic [37] to accomplish the crowd evacuation task as well for a fair comparison. The implementation details are as follows. For all compared methods, we first use A* [12] to find waypoints to the exit, and then: (1) Social Force Model [14, 15] - The individual agents of the Social Force Model utilize strategies + +![](images/12871005cedec3b5e88fdb011091b8b7f9903e31b779ed8134d719352357ccda.jpg) +Figure 5. Qualitative comparison. + +to reach each waypoint step by step. (2) OmniControl [43] - We calculate the path and interpolate the path points into dense trajectories, which serve as root joint conditions for generating escape motions. (3) MaskedMimic [37] - We compute the escape trajectory, and the path-following task is employed to control the character's escape actions. + +Table 2. Quantitative comparison. + +
MethodAverage Success RateAverage Fallen Count
OmniControl0.48—
MaskedMimic0.6018.55
Ours0.8412.26
+ +# 4.2. Quantitative Comparison + +To validate the performance of our framework in evacuation tasks, we compare it with OmniControl and MaskedMimic across four scenes. We evaluate the average escape success rate and the average number of falls for 50 agents over 1000 frames per scene, with each scene run 10 times. Due to the OmniControl not having physical realism, it does not produce falls, so the fallen Count is not calculated. As shown in Table 2, our method outperforms OmniControl and MaskedMimic in all scenes, with higher success rates and fewer falls. The OmniControl has insufficient controllability for long-distance movements, resulting in agents deviating from trajectories and stopping prematurely. Com + +pared to MaskMinic, in our method, agents demonstrate better pathfinding, mutual avoidance, and balance abilities. + +# 4.3. Qualitative Comparison + +To validate the superiority of our proposed framework, we use the above-mentioned three methods for comparison. Figure 5 shows the evacuation processes and evacuation motions executed by the three comparison models and our framework on two classic evacuation scenarios and one large-scale scenario. The Social Force Model is a 2D approach that can only depict the positional changes of people during evacuation, and it cannot represent 3D human movement. It assumes that individuals will always reach their desired position when making decisions, which make the simulation distorted. OmniControl fails to correctly generate long-distance trajectory-constrained actions, leading to motion distortion, trajectory confusion, and premature stopping. MaskedMimic, due to the lack of a collision avoidance mechanism, may cause collisions during linear movement, leading to crowd accumulation. Furthermore, the actions of all agents tend to be similar, failing to reflect individual variations. There is also no noticeable distinction in speed personalization among any of the three methods. Our approach achieves more rational evacuation processes and simulates personalized evacuation motions tailored to individuals with different attributes. More details can be found + +![](images/e4ec51b15d5374d4b8b6dcfe86831a148b03340d7c1a791057bbb09184e4f25d.jpg) +Figure 6. Ablation study results. + +in demo video and supplementary document. + +We also conduct a user study to evaluate our method, along with OmniControl and MaskedMimic, in terms of the plausibility of the evacuation process, the realism and personalization of evacuation motions, and the effectiveness of visualization part-level forces. For more details, please refer to the supplementary document. + +# 4.4. Ablation Study + +Qualitative comparison. In qualitative comparison, we compare the pipeline of our framework after removing the 3D-adaptive SFM Decision Mechanism, the Personalizing Gait Control Motor, and both components together, which degrades into the PACER[29] baseline. For all methods, we first find waypoints. Figure 6 shows that, when it comes to excluding 3D adaptive SFM while keeping personalized gait control, the generated the agents fail to avoid each other, leading to collisions and crowd accumulation. As a result, some agents are unable to escape, and their escape speeds are nearly identical. This indicates that the 3DA-SFM Decision Mechanism effectively prevents potential collisions as much as possible and offers more rational escape processes tailored to individual attributes. Moreover, although removing the Personalized Gait Control Module allows most agents to fully escape, their gaits and movements become similar. + +Personalized Gait Control Motor helps to generate personalized motions for agents based on their attributes. In contrast, the complete pipeline can dynamically adjust escape routes based on evacuation speed tailored to each agent's human attributes, minimizing collisions and ensuring personalized escape gaits that reflect individual characteristics. More details can be found in the demo video and supplementary document. + +![](images/dc291cdfa666dd6cb42f939c94deb4ac0e34097dc0249664d54e67f68e75a954.jpg) +Figure 7. Boxplot of speed distributions across categories. + +Speed Diversity. We conducted experiments across six scenarios, including four small-scale scenarios with 50 agents each and two large-scale scenarios with 100 agents. Agents were divided into five personalized groups—the young, the middle-aged, the old, patients, and the disabled, + +as well as a non-personalized group. Each category reflects distinct attributes such as age and mobility style. + +Figure 7 illustrates the speed distributions across all categories. The results reveal distinct patterns aligned with mobility attributes. For instance, the young achieve the highest median speeds and greater variability, reflecting their agility. In contrast, the old and patients exhibit slower speeds with narrower ranges, corresponding to their reduced mobility. Similarly, the disabled show limited speeds due to physical constraints, while the non-personalized group displays average speed distributions. + +![](images/3bc10610e5b7cacdee9d379965742216fce85e8140d1cc6f8bd82b909f8acde6.jpg) +(a) + +![](images/8410596ef0ccf9e17e6ba360c707a0fef5ed483d8d7913f5f70927b1fdda37fb.jpg) +(b) + +# 4.5. Illustrative Experiment + +Our framework can reflect many real-life evacuation in various scenarios. This is more helpful and meaningful for evacuation analysis. + +Analysis of Stampede Incidents. Stampede incidents often occur in bottleneck areas. We analyze the impact of evacuation density and corridor width on stampede occurrences at bottleneck locations. In a 2-meter-wide corridor, we randomly place 120, 80, and 40 agents to simulate evacuation scenarios, and the results are shown in Figure 8 (a). Under the same corridor width, the greater the number of evacuees, the more severe the stampede. With 120 agents in the same scenario, we test corridor widths of 2 meters, 3 meters, and 4 meters, with the results displayed in Figure 8 (b). Our results indicate that, for the same evacuation density, narrower corridors lead to more severe trampling. + +Analysis of stampede incidents shows they occur when overcrowding leads to falls—from unstable movement—with victims unable to recover, then being stepped on or crushed. Bottlenecks worsen this: large crowds pushing through narrow spaces heighten trampling risks, posing severe hazards. See the demo video for full simulation details. + +Analysis of the impact of terrain on evacuation. Our framework can simulate the impact of various terrains on + +crowd evacuation. In the same scene, we incorporate normal ground, uneven ground, ground with discrete obstacles, and slippery ground (with reduced friction), and test 50 individuals starting from the same initial position to escape from the same room. Results are presented in Figure 9. + +Figure 9 illustrates the different escape processes of individuals within the same scene, using the same number of participants and identical initial positions, but navigating various types of terrain. This demonstrates that terrain factors significantly influence motor control and further impact perception and decision-making in subsequent steps. Additionally, uneven and slippery surfaces are more likely to cause individuals to fall. + +![](images/65636eeb3dae24bca3724b3a9e050f9f113b983b2a7f39ae6cfb3fbabc8a32af.jpg) +(a) Normal Ground + +![](images/45bb1030a0f890a77c8ae6ce5837ddc16ca9901ac78463d9fd8b5079f0eeca31.jpg) +(b)Uneven Ground + +![](images/b3f666cf23a09d1a7b0be04a9b7d1decf879d63bbf7e3d83663b9d0682390dd4.jpg) +Figure 8. Evacuation simulation results: (a) with the same corridor width but different evacuation densities; (b) with the same evacuation density but different corridor widths. +(c) Ground with Discrete Obstacles + +![](images/0035b3f6c89c20cf4fcd46f2e9518628009bfaac354064de7ada57125746840e.jpg) +(d) Slippery Ground +Figure 9. Evacuation simulations in various terrains. + +# 5. Conclusions + +We propose a crowd evacuation simulation framework that mimics the sensory-decision-motor flow of a brain-like intelligence, enabling personalized simulations at the individual level. Additionally, we design part-level force visualization, which enhances evacuation analysis. Our framework is capable of simulating the personalized and diverse evacuation dynamics of individuals with varying attributes throughout the evacuation process. The escape motions produced by our method outperform existing methods. Through this framework, we can also validate many common phenomena observed during evacuation scenarios, offering new insights into crowd evacuation and public safety. + +# Acknowledgements + +This work was supported in part by the National Key R&D Program of China (2023YFC3082100), the National Natural Science Foundation of China (62171317), the Science Fund for Distinguished Young Scholars of Tianjin (22JCJJQJC00040), and the Engineering and Physical Sciences Research Council (No. EP/Y028805/1). + +# References + +[1] Panayiotis Charalambous, Julien Pettre, Vassilis Vassiliades, Yiorgos Chrysanthou, and Nuria Pelechano. Greil-crowds: Crowd simulation with deep reinforcement learning and examples. ACM Transactions on Graphics (TOG), 42(4):1-15, 2023. 2 +[2] Rui Chen, Mingyi Shi, Shaoli Huang, Ping Tan, Taku Komura, and Xuelin Chen. Taming diffusion probabilistic models for character control. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 3, 5 +[3] X Chen and F B Zhan. Agent-based modelling and simulation of urban evacuation: relative effectiveness of simultaneous and staged evacuation strategies. Journal of the Operational Research Society, 59(1):25-33, 2008. 1, 2 +[4] Stelian Coros, Philippe Beaudoin, and Michiel van de Panne. Generalized biped walking control. ACM Trans. Graph., 29 (4), 2010. 1, 3 +[5] Emiliano Cristiani, Benedetto Piccoli, and Andrea Tosin. Multiscale modeling of pedestrian dynamics. Springer, 2014. 2 +[6] Ning Ding, Yu Zhu, Xinyan Liu, Dapeng Dong, and Yang Wang. A modified social force model for crowd evacuation considering collision predicting behaviors. Applied Mathematics and Computation, 466:128448, 2024. 1, 2 +[7] J. Fang, S. El-Tawil, and B. Aguirre. New agent-based egress model allowing for social relationships. Journal of Computing in Civil Engineering, 30(4):04015066, 2016. 2 +[8] I.G. Georgoudas, P. Kyriakos, G.Ch. Sirakoulis, and I.Th. Andreadis. An FPGA implemented cellular automaton crowd evacuation model inspired by the electrostatic-induced potential fields. Microprocessors and Microsystems, 34(7):285-300, 2010. 2 +[9] Gonzalo Gomez-Nogales, Melania Prieto-Martin, Cristian Romero, Marc Comino-Trinidad, Pablo Ramon-Prieto, Anne-Hélène Olivier, Ludovic Hoyet, Miguel Otaduy, Julien Pette, and Dan Casas. Resolving collisions in dense 3d crowd animations. ACM Transactions on Graphics, 43(5):1-14, 2024. 2 +[10] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, 2022. 1 +[11] Horst W Hamacher and Stevanus A Tjandra. Mathematical modelling of evacuation problems: A state of art. 2001. 2 +[12] Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE Transactions on Systems Science and Cybernetics, 4(2): 100-107, 1968. 4, 5 +[13] Brandon Haworth, Glen Berseth, Seonghyeon Moon, Petros Faloutsos, and Mubbasir Kapadia. Deep integration of physical humanoid control and crowd navigation. In Proceedings of the 13th ACM SIGGRAPH Conference on Motion, Interaction and Games, New York, NY, USA, 2020. Association for Computing Machinery. 2 +[14] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E, 51(5):4282, 1995. 2, 4, 5 + +[15] Dirk Helbing, Illés Farkas, and Tamas Vicsek. Simulating dynamical features of escape panic. Nature, 407(6803):487-490, 2000. 2, 4, 5 +[16] Alexis Jensen, Thomas Chatagnon, Niloofar Khoshsiyar, Daniele Reda, Michiel Van De Panne, Charles Pontonnier, and Julien Pettré. Physical simulation of balance recovery after a push. In Proceedings of the 16th ACM SIGGRAPH Conference on Motion, Interaction and Games, New York, NY, USA, 2023. Association for Computing Machinery. 3 +[17] Andrei Khilkevich, Michael Lohse, Ryan Low, Ivana Orsovic, Tadej Bozic, Paige Windmill, and Thomas D Mrsic-Flogel. Brain-wide dynamics linking sensation to action during decision-making. Nature, pages 1-11, 2024. 2 +[18] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac Gym: High performance GPU-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 1 +[19] Ian Mason, Sebastian Starke, and Taku Komura. Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 5(1):1-18, 2022. 4, 5 +[20] Earl K Miller and Jonathan D Cohen. An integrative theory of prefrontal cortex function. Annual review of neuroscience, 24(1):167-202, 2001. 2 +[21] Donald A. Neumann. Kinesiology of the musculoskeletal system: foundations for physical rehabilitation. 2002. 5 +[22] Manh Hung Nguyen, Tuong Vinh Ho, and Jean-Daniel Zucker. Integration of smoke effect and blind evacuation strategy (sebes) within fire evacuation simulation. Simulation Modelling Practice and Theory, 36:44-59, 2013. 1, 2 +[23] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 2024 International Conference on 3D Vision (3DV), pages 1498-1507. IEEE, 2024. 3 +[24] Xue Bin Peng, Glen Berseth, Kangkang Yin, and Michiel Van De Panne. DeepLoco: dynamic locomotion skills using hierarchical deep reinforcement learning. ACM Trans. Graph., 36(4), 2017. 3 +[25] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3 +[26] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Trans. Graph., 41(4), 2022. 1 +[27] Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Korrawe Karunratanakul, Pu Wang, Hongfei Xue, Chen Chen, Chuan Guo, Junli Cao, Jian Ren, and Sergey Tulyakov. ControlMM: Controllable masked motion generation. arXiv preprint arXiv:2410.10780, 2024. 2, 3 +[28] Michael L Platt. Neural correlates of decisions. Current Opinion in Neurobiology, 12(2):141-148, 2002. 2, 3 +[29] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13756-13766, 2023. 1, 2, 3, 4, 7 +[30] Siamak Sarmady, Fazilah Haron, and Abdullah Zawawi Talib. Simulating crowd movements using fine grid cellular automata. In 2010 12th International Conference on Computer Modelling and Simulation, pages 428-433, 2010. 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, 3 +[32] Yi Shi, Jingbo Wang, Xuekun Jiang, and Bo Dai. Controllable motion diffusion model. arXiv preprint arXiv:2306.00416, 2023. 1, 3 +[33] Yi Shi, Jingbo Wang, Xuekun Jiang, Bingkun Lin, Bo Dai, and Xue Bin Peng. Interactive character control with autoregressive motion diffusion models, 2024. 3 +[34] Mohammed Shuaib and Zarita Zainuddin. Incorporating intelligence into exit choice model for typical evacuation. Sains Malaysia, 46(10):1997-2005, 2017. 2 +[35] Markus Siegel, Timothy J Buschman, and Earl K Miller. Cortical information flow during flexible sensorimotor decisions. Science, 348(6241):1352-1355, 2015. 2 +[36] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. CALM: Conditional adversarial latent models for directable virtual characters. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 1 +[37] Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, and Xue Bin Peng. Maskedmimic: Unified physics-based character control through masked motion inpainting. ACM Transactions on Graphics (TOG), 43(6):1-21, 2024. 1, 2, 3, 5, 6 +[38] Takara Everest Truong, Michael Piseno, Zhaoming Xie, and Karen Liu. PDP: Physics-based character animation via diffusion policy. In SIGGRAPH Asia 2024 Conference Papers, page 1-10. ACM, 2024. 3 +[39] 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, 3 +[40] Jingbo Wang, Sijie Yan, Bo Dai, and Dahua Lin. Scene-aware generative network for human motion synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12206-12215, 2021. 3 +[41] Jingbo Wang, Zhengyi Luo, Ye Yuan, Yixuan Li, and Bo Dai. Pacer+: On-demand pedestrian animation controller in driving scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 718-728, 2024. 2, 5 +[42] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. InterControl: Generate human motion interactions by controlling every joint. CoRR, abs/2311.15864, 2023. 3 +[43] 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. 1, 2, 3, 5, 6 + +[44] Zhaoming Xie, Hung Yu Ling, Nam Hee Kim, and Michiel van de Panne. ALLSTEPS: Curriculum-driven learning of stepping stone skills, 2020. 3 +[45] 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. 3 +[46] Jianming Ye, Zhen Liu, Tingting Liu, Yanhui Wu, and Yuanyi Wang. Crowd evacuation simulation based on hierarchical agent model and physics-based character control. Computer Animation and Virtual Worlds, 35(3):e2263, 2024. 2 +[47] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. PhysDiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16010-16021, 2023. 3 +[48] Zarita Zainuddin and Mohammed Shuaib. Modification of the decision-making capability in the social force model for the evacuation process. Transport Theory and Statistical Physics, 39(1):47-70, 2010. 2 +[49] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14738-14749, 2023. 3 +[50] Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, and Huaizu Jiang. Smoodi: Stylized motion diffusion model. In European Conference on Computer Vision, pages 405-421. Springer, 2025. 3 \ No newline at end of file diff --git a/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/images.zip b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ec5a42067027759f1b16b8dc177f726002dda1d3 --- /dev/null +++ b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e69fe6f64b8a3d058a0bc1527729bcbee7de9faeb874d852f16c67be6f4b530 +size 674981 diff --git a/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/layout.json b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c5cb957d971e36be96c9ea503c991e653fc9c1e0 --- /dev/null +++ b/ICCV/2025/RESCUE_ Crowd Evacuation Simulation via Controlling SDM-United Characters/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7745a908461fe1f8af3ec7dd6d8ac3e32bde3d8d9582e36ce054854352add65 +size 345287 diff --git a/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/f5c0c5e2-fe9c-4ffa-a82c-0d11d9a5d550_content_list.json b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/f5c0c5e2-fe9c-4ffa-a82c-0d11d9a5d550_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ac00222b17cc75af33aa28d4ce0f3084c4f0ee12 --- /dev/null +++ b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/f5c0c5e2-fe9c-4ffa-a82c-0d11d9a5d550_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b00659c47699ce8d52327d6551916bd6996da4cc52bd75ad99096a8a086e36d0 +size 67275 diff --git a/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/f5c0c5e2-fe9c-4ffa-a82c-0d11d9a5d550_model.json b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/f5c0c5e2-fe9c-4ffa-a82c-0d11d9a5d550_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e037954bf208cd5c36365c01e530ad22eda7ff3d --- /dev/null +++ b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/f5c0c5e2-fe9c-4ffa-a82c-0d11d9a5d550_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2237ba708553574d1398411107cae91a56cca2f36839beeeeb09f6f72f6f046 +size 83659 diff --git a/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/f5c0c5e2-fe9c-4ffa-a82c-0d11d9a5d550_origin.pdf b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/f5c0c5e2-fe9c-4ffa-a82c-0d11d9a5d550_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..42273d2c1a5591115351cfef745382bfc314b829 --- /dev/null +++ b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/f5c0c5e2-fe9c-4ffa-a82c-0d11d9a5d550_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a050ed28026c9468047a901c2d9486ef59ac3c0bc165404db41702635fd60028 +size 6058806 diff --git a/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/full.md b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..76a09d4852c3a1557b5d747330cbaa2dafc0f75e --- /dev/null +++ b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/full.md @@ -0,0 +1,289 @@ +# RGE-GS: Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors + +Sicong Du $^{1,\dagger}$ Jiarun Liu $^{1,2,\dagger}$ Qifeng Chen $^{1}$ Hao-Xiang Chen $^{3}$ Tai-Jiang Mu $^{3}$ Sheng Yang $^{1,\boxtimes}$ $^{1}$ Unmanned Vehicle Dept., CaiNiao Inc., Alibaba Group $^{2}$ Zhejiang University $^{3}$ BNRist, Tsinghua University +{liujiarun.ljr, shengyang}@cainiao.com, iedusicong@126.com + +# Abstract + +A single-pass driving clip frequently results in incomplete scanning of the road structure, making reconstructed scene expanding a critical requirement for sensor simulators to effectively regress driving actions. Although contemporary 3D Gaussian Splatting (3DGS) techniques achieve remarkable reconstruction quality, their direct extension through the integration of diffusion priors often introduces cumulative physical inconsistencies and compromises training efficiency. To address these limitations, we present RGE-GS, a novel expansive reconstruction framework that synergizes diffusion-based generation with reward-guided Gaussian integration. The RGE-GS framework incorporates two key innovations: First, we propose a reward network that learns to identify and prioritize consistently generated patterns prior to reconstruction phases, thereby enabling selective retention of diffusion outputs for spatial stability. Second, during the reconstruction process, we devise a differentiated training strategy that automatically adjust Gaussian optimization progress according to scene converge metrics, which achieving better convergence than baseline methods. Extensive evaluations of publicly available datasets demonstrate that RGE-GS achieves state-of-the-art performance in reconstruction quality. Our source-code will be made publicly available at https://github.com/CN-ADLab/RGE-GS. + +# 1. Introduction + +Autonomous driving simulators necessitate sophisticated street scene modeling that can effectively synthesize multiview sensor data while adhering to realistic physical principles - a fundamental capability critical for scalable data augmentation [35] and closed-loop simulation [16]. + +Expansive scene modeling - which involves performing physically plausible and perceptually realistic reconstruction + +on under-observed regions from single-pass real-world scans – has become a pivotal capability for autonomous driving simulators to support reactive regressions. While recent breakthroughs in controllable diffusion models [7, 28] have achieved significant improvements of visual effects, these approaches remain limited in their adaptability for scene-specific training and tuning, as well as their computational efficiency for real-time Novel View Synthesis (NVS) along predefined trajectories. + +Therefore, many state-of-the-art diffusion-aided reconstruction frameworks [5, 20, 29] incorporate generated scans into an explicit reconstruction pipeline to improve physical consistency, facilitate real-time photorealistic rendering, and enable interactive editing of reconstructed driving scenes. Nevertheless, integrating diffused scans with physical inconsistency into Gaussian reconstruction process may lead to unpredictable artifacts or compromised training efficiency. To date, existing frameworks have yet to adequately resolve this challenge, as they predominantly rely on reconstruction methods [11, 13, 19] originally designed for real-world scans without necessary adaptations. Fine-tuning diffusion models on domain-specific data can help narrow the gap between generated images and real scans. However, this approach may still fall short in addressing inherent biases in the generated outputs. Without a robust mechanism to identify and reject these biases, the generated images may lack physical consistency, leading to potentially misleading results. + +To more effectively leverage diffusion priors during expansive modeling, we propose RGE-GS, a reconstruction method that is compatible with recent advancements in controllable diffusion models [2] and represents scenes using Gaussians. For generated scans, we quantify two pixel-level metrics revealing the reconstruction progress: (1) the coherence between newly generated diffusion priors and existing Gaussians and (2) the current reconstruction progress of under-observed and ongoing regions. As for the first metric, it can be used for quality inspection and data filtering for input diffusion priors, preserving mostly correct information for reconstruction. While fine-tuning diffusion + +models on domain-specific data or incorporating additional control information can help narrow the gap between generated images and real scans, these approaches alone may not fully address the challenges of continuous integration during reconstruction. To stress this issue, we propose a reward network which predicts the pixel-wise confidence of each generated image and assists in training the GS scene. Our experiments demonstrate that adversarially verifying the generated results as a complementary strategy is an effective means to reject biased outputs and enhance physical consistency. The second metric identifies the under-reconstructed areas to focus on incomplete regions while preserving the already reconstructed parts. We design a differentiated training strategy that automatically applies different optimization techniques based on the convergence level of the Gaussians. This strategy not only enhances the overall learning performance of the Gaussian scene but also reduces training time by applying targeted optimization techniques for different regions. + +In conclusion, RGE-GS makes the following key contributions: + +- A novel reward-guided GS reconstruction method for expansive scene modeling via diffusion priors, achieving state-of-the-art quality and efficiency compared to previous diffusion-prior aided reconstruction methods. +- A reward network to evaluate the pixel-level confidence of input diffusion priors considering the compatibility and progress of existing Gaussians. +- A differentiated training strategy is proposed to adjust the optimization process of GS primitives based on different convergence, enhancing both efficiency and stability. + +Extensive evaluations on publicly available datasets [23] have demonstrated the effectiveness of our approach in leveraging these two metrics for adaptively handling diffusion priors, thereby improving both quality and efficiency during progressive reconstruction. + +# 2. Related Work + +# 2.1. Controllable Diffusion Models + +Controllable diffusion models for autonomous driving [9, 15] generate high-fidelity videos based on text prompts, reference sensor frames, and road structural information, including maps and obstacles. DriveDreamer [29] aligns the above conditions to concatenate with noise images for diffusion, where trained temporal attention layers are leveraged to ensure the consistency of generated video frames. DriveDreamer2 [39] further utilizes a fine-tuned large language model (LLM) to translate user prompts into agent trajectories and High Definition (HD) maps for controlling generation. MagicDrive [7] also encodes these conditions through tailored encoding strategies and further supports multi-sensor consistency. FreeVS [28] leverages projected + +colored LiDAR point cloud as pseudo-images for controlling the generation. Streetscapes [4] incorporates street and height maps for a layout-conditioned long-term generation. + +In summary, these methods mainly focus on leveraging auxiliary information together with existing diffusion models. Since they use a unified big model for different driving clips to maintain temporal consistency, they also require a sufficient dataset and training process for this structural information. + +# 2.2. Diffusion-aided Reconstruction + +While controllable diffusion models propose to directly generate temporally coherent videos, their subsequent insights on leveraging reconstruction methods provide an applicable way of reducing both structural information and heavy models: DriveDreamer4D [38] leverages diffusion priors via constructing temporally aligned cousin pairs between diffusion generated frames and Gaussian rendered frames. MagicDrive3D [8] uses consistent monocular depth prior, aligns exposure between cameras, and regards Gaussians as deformable for local dynamics to enhance the scene representation. FreeSim [5] proposes a progressive reconstruction strategy to add generated images from under-observed perspectives. + +Although these methods with 3DGS (also see [10, 22, 32, 34]) have successfully built appropriate loss functions leading reconstructed Gaussians compliant with diffusion priors, 3DGS – as an explicit scene representation – necessitates high cross-view geometry and appearance consistency, and unavoidably magnifies the challenge of convergence or results in conspicuous artifacts. Based on enhanced loss functions, our RGE-GS further guards the expansive reconstruction by examining the input diffusion priors based on the current state of Gaussians. Our proposed reward-guidance network does not require heavy parameters and training processes, and can be compatible with these reconstruction methods as an independent access-in module. + +# 2.3. Indicating Gaussian Reconstruction Progress + +GS-Planner [12] represents the completeness of partially reconstructed Gaussians for planning the active scanning trajectory w.r.t. the information gain. 3DGS-Enhancer [18] proposes to compute image and pixel levels of confidence through perspective parallax and Gaussians' covariance, respectively. ReconX [17] focuses on a global consistency on a series of generated frames to simultaneously obtain multiple confidence maps through the training process of DUSt3R. InstantSplat [6] introduces a confidence-aware perpoint optimizer designed to facilitate convergence during self-supervised optimization. GaussianPro [3] proposes a progressive propagation strategy using geometrical values from PatchMatch [1] to indicate under-reconstruction regions. Among methods using Neural Radiance Field (NeRF) + +![](images/1863e377575ca2d27597ba1cc044410ca8f9a2f3ae76e542807393783f957937.jpg) +Figure 1. The workflow of our proposed RGE-GS for expansive scene reconstruction. We use a single trajectory clip to reconstruct gaussians and feed rendered images to a diffusion model for novel view synthesis. Before sending diffused images back to the reconstruction, we design and train a reward guidance network to evaluate the consistency of generation and progress of reconstruction and infer a pixel-wise confidence during reconstruction. + +Representation [21], NeRF On-the-go [24] uses uncertainty from MLP to train models. Our method combines progress and consistency metrics through a unified model and proposes strategic improvements to better reconstruct scene regions at different completeness levels. + +# 3. Methodology of RGE-GS + +# 3.1. Preliminaries + +3D Gaussian Splatting. 3DGS [13] characterizes the scene through a collection of anisotropic Gaussians defined within the 3D space. Each Gaussian $G$ is assigned several parameters: opacity $\boldsymbol{o} \in \mathbb{R}$ , spherical harmonics (SH) coefficients $z \in \mathbb{R}^k$ , a position vector $\boldsymbol{\mu} \in \mathbb{R}^3$ , a rotation quaternion $\boldsymbol{q} \in \mathbb{R}^4$ , and a scale factor $s \in \mathbb{R}^3$ . The Gaussian kernel distribution is expressed as: + +$$ +\boldsymbol {G} (\mathbf {x}) = \exp \left(- \frac {1}{2} (\mathbf {x} - \boldsymbol {\mu}) ^ {\top} \boldsymbol {\Sigma} ^ {- 1} (\mathbf {x} - \boldsymbol {\mu})\right), \tag {1} +$$ + +where the covariance matrix is defined as $\boldsymbol{\Sigma} = \mathbf{RSS}^{\top}\mathbf{R}^{\top}$ , with $\mathbf{S}$ representing the scaling matrix dictated by $s$ and $\mathbf{R}$ being the rotation matrix derived from $q$ . Given the camera's extrinsic parameters $\mathbf{W}$ , the 2D covariance matrix in screen space is computed as follows: + +$$ +\boldsymbol {\Sigma} ^ {*} = \mathbf {J} \mathbf {W} \boldsymbol {\Sigma} \mathbf {W} ^ {\top} \mathbf {J} ^ {\top}, \tag {2} +$$ + +Latent Diffusion Models. Latent diffusion models (LDMs) [25] represent a category of generative models that effectively learn the distribution of the latent variable $\mathbf{X}$ derived from the dataset $\mathbf{D}$ . These models work in conjunction with an autoencoder (or tokenizer) that transforms the data into the latent space defined as $\mathbf{X} = \epsilon (\mathbf{D})$ . A diffusion + +model initiates the process with a noise vector $\mathcal{N}(0,\mathbf{I})$ and progressively denoises it to produce a sample $\mathbf{X}$ , which can be influenced by a condition $\mathbf{C}$ . Subsequently, the decoder is employed on the denoised latent representation to create the output data $\hat{\mathbf{D}}$ . LDMs have demonstrated their effectiveness and efficiency in capturing various data modalities, such as images [25] and videos [2]. + +# 3.2. Framework Overview + +The overall framework of our RGE-GS is shown in Fig.1. The first round, we use the ground truth images $\bigcup I_o$ from the original viewpoints to reconstruct the GS primitives, denoted as $G_{o}$ . To prepare for the expansion of GS, we shift to extrapolated free viewpoints along the original sensor trajectory to render degraded images $\bigcup I_d$ , and feed these into a diffusion model to generate pseudo images $\bigcup I_e$ . The Second round, $\bigcup I_e$ and $\bigcup I_o$ are used as input to retrain $G_{o}$ . Especially, we introduce a reward network $\mathcal{F}_c$ (Sec. 3.3) to identify high-demand and high-confidence regions corresponding to each pixel of generated image. It is jointly optimized with $G_{o}$ , and the output of $\mathcal{F}_c$ , reward map $\bigcup C_e$ , served as an adversarial verification to assist the consistency of scene expansion. Meanwhile, we enhance the expansive reconstruction process with such auxiliary info by differentiating the initialization and training strategies and obtain the final scene representation $G$ (Sec. 3.4). + +# 3.3. Reward Network + +Our framework systematically integrates the prior images generated by the diffusion model into GS training through a lightweight reward network. This auxiliary module is trained in an adversarial strategy to produce pixel-wise reward scores $C_e$ that dynamically modulate the contribution + +![](images/c699b3e6c44aee849ad4f4477cef6aaa55f6433de1f402213b817b04701286c8.jpg) +Figure 2. The illustration of reward network. + +of generated priors. + +Network structure. As shown in Fig. 2, the reward network $\mathcal{F}_c: I_e \in \mathbb{R}^{\mathrm{H} \times \mathrm{W} \times 3} \to C_e \in \mathbb{R}^{\mathrm{H} \times \mathrm{W}}$ adopts a lightweight U-Net architecture with encoder-decoder symmetry and residual skip connections [27]. Comprising three downsampling and upsampling stages, the network processes input images through successive convolution blocks to extract multi-scale features, followed by transposed convolutions for spatial recovery. Skip connections bridge encoder and decoder features to preserve structural details, culminating in a prediction head with sigmoid activation that outputs pixel-wise confidence scores between 0 and 1. + +Training strategy. Given the absence of ground truth for confidence values, the reward network is constructed as an adversarial verification module, facilitating unsupervised learning during joint training with the GS reconstruction framework. This design allows for the auxiliary generation of confidence maps. Specifically, for each scene, we first integrate the reward network into the GS pipeline and train it for multiple iterations. During this process, the reward network is guided by the following supervisions: (1) a reprojection error loss $\mathcal{L}_{\mathrm{proj}}$ that enforces consistency between point cloud projections and diffusion priors, and (2) a binarization regularization term $\mathcal{L}_{\mathrm{reg}}$ that prevents forming indistinguishable reward maps. + +The reprojection error loss is defined as Eq. 3: + +$$ +\mathcal {L} _ {\text {r e p r o j}} = \frac {1}{\mathrm {H} \cdot \mathrm {W}} \| C _ {e} \odot (\pi (\mathbf {P}, \mathbf {T} _ {e}, \mathbf {K}) - I _ {e}) \| _ {2}, \tag {3} +$$ + +where $\mathbf{P}$ is the colored LiDAR point cloud generated by recolor each LiDAR points with their positions in the world reference frame and their visible colors. $\pi (\mathbf{P},\mathbf{T},\mathbf{K})$ represents the projection function to transpose 3D points to 2D image with camera pose $\mathbf{T}$ and intrinsic $\mathbf{K}$ . The reprojection of colored point cloud provides coarse explicit spervision to the reward map, which facilitates the gradient optimization of the reward network during adversarial training. + +The binarization regularization term in Eq. 4 is designed to encourage the confidence distribution to concentrate around 0 or 1, thereby avoiding ambiguous predictions. + +$$ +\mathcal {L} _ {\text {r e g}} = \frac {C _ {e} \cdot \left(1 - C _ {e}\right)}{\mathrm {H} \cdot \mathrm {W}}. \tag {4} +$$ + +The overall training loss of reward network $\mathcal{L}_{\mathrm{reward}}$ is defined as Eq. 5: + +$$ +\mathcal {L} _ {\text {r e w a r d}} = \lambda_ {\text {r e p r o j}} \mathcal {L} _ {\text {r e p r o j}} + \lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}} + \mathcal {L} _ {\text {G S}}, \tag {5} +$$ + +where $\lambda_{\mathrm{proj}} = 0.5$ and $\lambda_{\mathrm{reg}} = 0.3$ . $\mathcal{L}_{\mathrm{GS}}$ represents the total loss of GS reconstruction, which will be discussed further in Sec. 3.4. The output of reward network $C_e$ is applied to corresponding $I_e$ during training. + +# 3.4. Differentiated Training Strategy + +After $G_{o}$ , $\bigcup I_{e}$ (Sec. 3.1) and $\bigcup C_{e}$ (Sec. 3.3) are obtained, we categorize the GS primitives into three groups and apply different training strategies based on the convergence status. + +1. Missing Gaussians. For regions visible in the novel views $\bigcup I_e$ but occluded in $\bigcup I_o$ , we adopt the same initialization strategy used in the first reconstruction pass and employ the following two methods: (1) for areas covered by stitched LiDAR point clouds, we initialize the positions of the Gaussians based on the geometric distribution of the LiDAR points, and (2) for regions inadequately scanned by LiDAR sensors, we utilize depth estimation from DepthAnything-V2 [33] and calibrate the scale by minimizing the Euclidean distance relative to existing LiDAR measurements. This latter strategy enables us to successfully initialize a significant number of unobserved regions, thereby alleviating the time-consuming and unstable aspects of multi-view geometry during Gaussian training. + +2. Immature Gaussians. Incomplete scans would result in Gaussians with large view-space positional gradients when observed on novel view images $I_{e}$ since they cannot be sufficiently trained due to limited multi-view parallax. Intuitively, this phenomenon is because some regions are not yet fully reconstructed, and the optimization process tries to address this by moving Gaussians. Consequently, we identify these immature primitives using a predefined gradient threshold of $\lambda = 5 \times 10^{-4}$ for further optimization until convergence with extra supervision from $I_{e}$ . + +3. Mature Gaussians. For the rest of the existing Gaussians regarded as matured, they no longer need to be affected by generative scans, so we fix their attributes to ensure the stability of reconstruction, which also accelerates the training procedure. + +The reward maps $\bigcup C_e$ are then applied to the corresponding diffusion prior for GS reconstruction. The reconstruction loss function is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {I _ {e}} = \lambda_ {\text {R G B}} \mathcal {L} _ {R G B} \left(I _ {d}, I _ {e}, C _ {e}\right) \\ + \left(1 - \lambda_ {\mathrm {R G B}}\right) \mathcal {L} _ {\mathrm {S S I M}} \left(\boldsymbol {I} _ {d}, \boldsymbol {I} _ {e}, \boldsymbol {C} _ {e}\right) + \lambda_ {\boldsymbol {I} _ {e}} \mathcal {L} _ {\mathrm {L P I P S}} \left(\boldsymbol {I} _ {d}, \boldsymbol {I} _ {e}\right), \tag {6} \\ \end{array} +$$ + +where $\mathcal{L}_{\mathrm{RGB}} = 0.8$ is the L1 norm with the weighted reward map $\bigcup C_e$ multiplied on the gaussian rendering $I_d$ and generated image $\bigcup I_e$ respectively. $\mathcal{L}_{\mathrm{ssim}}$ is the structural similarity index measure [30]. $\mathcal{L}_{\mathrm{LPIPS}}$ is the perceptual similarity metric[37] and $\lambda_{I_e}$ is set to 0.01. + +The definition of loss function during the GS reconstruction process on differentiated training strategy is as follows, + +$$ +\mathcal {L} _ {\mathrm {G S}} = \mathcal {L} _ {I _ {o}} + \mathcal {L} _ {I _ {e}}, \tag {7} +$$ + +$$ +\mathcal {L} _ {\boldsymbol {I} _ {o}} = \lambda_ {\boldsymbol {I} _ {o}} \mathcal {L} _ {1} (\boldsymbol {I}, \boldsymbol {I} _ {o}) + (1 - \lambda_ {\boldsymbol {I} _ {o}}) \mathcal {L} _ {\text {S S I M}} (\boldsymbol {I}, \boldsymbol {I} _ {o}), +$$ + +where $\lambda_{I_o} = 0.8$ , $I$ represents the ground truth image of the original lane. + +# 4. Evaluation + +# 4.1. Implementation and Experimental Settings + +Dataset. We conduct experiments on Para-Lane [23] dataset and Waymo Open Datset. On the Para-Lane dataset, we select four scenes containing three lanes (left, middle, right), comprising 250 front-view images of each lane scanned at $1280 \times 720 \times 10\mathrm{Hz}$ to evaluate the aforementioned method with publicly available baselines + +Implementation details of the diffusion model. Our generative model is initialized using the pre-trained weights from SD v1.5 [25]. We train the parameters of U-Net [26] and ControlNet [36] for 50,000 iterations across 4 A100 GPUs with a batch size of 8 per GPU while keeping the VAE [14] parameters frozen. The training dataset is constructed from the Para-Lane dataset using method in [28], containing about 60,000 pairs. During training, the learning rate is set to 5e-5 and the resolution of generated images remains the same with input real-scans. + +Implementation details of the reward network. The reward network is initialized with random parameters, except for a bias constraint applied before the prediction head to keep its output away from all-zero. The network employs the AdamW optimizer with standard parameter settings. The learning rate is initialized at 5e-4, undergoes linear decay during the first 1,000 iterations, and then follows a cosine annealing schedule. The network is jointly trained with Gaussian scenarios while synchronously updating parameters, undergoing a total of 5,000 iterations. After that, the reward network is frozen and we only use it to generate reward output $\bigcup C_e$ for the following training process. + +Implementation details of the Gaussian reconstruction. Our proposed method is compatible for most publicly available Gaussian reconstruction frameworks. In our ablation (also see Tab. 1), we select 3DGS [13] and Scaffold-GS [19] as our baselines to demonstrate our adapt ability. The Gaussian densification process begins and ends at 500 and 15,000 iterations, respectively, with a gradient threshold for mature primitive selection set at 4e-4. The number of initialization points is configured to 50,000. + +Repeat of publicly available methods. We replicate the current state-of-the-art cross-lane view synthesis methods FreeSim [5] and ReconDreamer [22] since both are not + +completely open-sourced as of now1. Specifically, our training framework is built upon a dynamic-static decoupled GS reconstruction pipeline, where the generation results from our proposed diffusion model serve as cross-lane novel view priors. To simplify the training process and directly get the upper-bound performance of the baseline method, we directly use generated images of cross-lane views for training, bypassing the progressive view generation approach that transitions from near to far. + +Table 1. Fidelity of novel view synthesis with different methods on Para-Lane dataset. Asterisks stand for our reimplementation. + +
Method MetricsOne-Lane ExpansionTwo-Lane Expansion
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS [13]16.120.5190.46114.500.4620.526
2DGS [11]15.040.5270.47613.040.4830.529
Street Gaussians [31]16.760.5170.45016.260.4850.506
Scaffold-GS [19]17.250.5190.45316.700.4710.511
FreeSim* [5]19.070.5370.42417.960.5140.466
RGE on 3DGS18.730.5560.46118.010.5420.483
RGE on Scaffold-GS19.720.5860.41518.110.5190.468
+ +# 4.2. Main Results + +Quantitative comparison. On the Para-Lane Dataset, we present the performance of novel view synthesis for scene expansion across one lane and two lanes in Table 1, with lateral trajectory shifts of approximately $3.5\mathrm{m}$ and $7\mathrm{m}$ , respectively. In both comparisons, our proposed RGE strategy significantly outperforms previous reconstruction-based methods. This improvement is primarily due to the incorporation of a confidence-aware diffusion prior during the reconstruction process for the novel views. Additionally, our method exhibits a slight performance drop in two-lane expansion compared to one-lane expansion, which can be attributed to the less effective guidance information provided to the diffusion model from the rendering results when larger lateral shifts are involved. Furthermore, RGE, when applied to both 3DGS and Scaffold-GS, demonstrates a consistent trend, highlighting the adaptability of our method. + +On the Waymo dataset, we select 10 scenes that offer relatively spacious movement areas, ensuring that camera movements do not intrude into the obstacle regions to accommodate the trajectory shifting. Due to the cross-lane ground truth images being unavailable in the Waymo dataset, we adopt the FID score based on InceptionV3 for off-trajectory evaluation following [22], with different lateral trajectory shifts from $1\mathrm{m}$ to $5\mathrm{m}$ . As a result shown in Tab. 2, our proposed method outperforms the previous methods, especially when the lateral shift distance is relatively large. This is be + +![](images/54b82ac5114ed8d3f4a30d97b7387375685bebed046995de5bd972e0c12c443d.jpg) +Figure 3. Comparisons for cross-lane view synthesis with publicly available methods, demonstrating our effectiveness of using a reward network and differentiated training strategy for expansive reconstruction. Please refer to our supplementary video for more examples. + +cause the degradation of GS rendering quality due to large lateral movements has a cascading effect on the diffusion output, making the reward network mechanism increasingly critical for reconstruction in such scenarios. + +Table 2. Fidelity of novel view synthesis with different methods on Waymo dataset. Asterisks stand for our reimplementation. + +
Method Metrics@1mLane Shift views (FID ↓)
@2m@3m@4m@5m
Scaffold-GS62.9384.32124.57148.32189.25
FreeSim*56.8779.23111.77138.07174.12
ReconDreamer*59.7676.44117.46142.67167.97
RGE on Scaffold-GS52.4171.56102.37117.08129.12
+ +Qualitative comparison. Fig. 3 shows the visual comparisons between RGE on Scaffold-GS and other methods in novel views on Para-Lane dataset. We can find that images rendered by the baseline algorithms exhibit are filled with speckles and ghosting. Our method, however, significantly improves the rendering quality. + +# 4.3. Ablation Results + +Ablations on the reward network. We use Scaffold-GS as our baseline method in ablations. Tab. 3 and Fig. 4 demonstrates the significance of the reward inference module outputting $C_e$ . The experimental results demonstrate that our reward net effectively constrains the diffusion outputs through adaptive regulation, thus mitigating potential distortions in diffusion priors caused by implausible generation artifacts while preserving critical anatomical authenticity. Notice that although classifying Gaussian primitives can improve the effectiveness of training under pseudo views, the supervision of training comes directly from the diffusion output, which risks propagating error information. The + +reward network can limit these error priors at the source, ensuring stable training. We further validate the correctness of the network output by overlaying the reward maps on diffusion outputs and measuring MAE (Mean Absolute Error) of valid pixels to ground truth images on Para-Lane dataset. As shown in Fig. 5, the reward map effectively limits the adverse effects of erroneous outputs, confirming our training strategy's effectiveness. + +Table 3. Ablation study on novel view synthesis with different pipeline settings on Para-Lane dataset with one-lane expansion. + +
MetricsPSNR↑SSIM↑LPIPS↓Training (min) ↓
w/o RewardNet19.240.5440.42791
w/o Diff-Train19.550.5520.419112
RGE on Scaffold-GS19.720.5860.41589
+ +Table 4. Ablation study on novel view synthesis with different diffusion prior models on ParaLane dataset. + +
Method MetricsOne-Lane ExpansionTwo-Lane Expansion
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
RGE_SD on Scaffold-GS19.720.5860.41518.110.5190.468
SVD on Scaffold-GS19.520.5520.42218.080.4990.474
RGE_SVD on Scaffold-GS19.950.6010.40418.630.5200.432
+ +Ablations on the differentiated training strategy. Tab. 3 also showcase the quantitative results of the ablation study on differentiated training strategy. The results indicate the effectiveness of our proposed graduated training method. Specifically, the training time is reduced with differentiated training strategy because the number of Gaussians requiring gradient computation is reduced. + +![](images/c63d9fc5b1976e49e8f698a261a549f0f80d4c97f706dff974d7fc086aa30f7f.jpg) + +![](images/eedf0b6bb2fcbd776b53cbc6a50a7cb39c6d2224172548e6bf627ad2963297f6.jpg) +Generated Image + +![](images/6ed4c1c8b593d56ca29ea882137ca0a0b2627ee5406405cb19fd8ae2450a370e.jpg) + +![](images/b0591e7e564e801922e9aa6b74c591c53c731f7647235f6b3f70ae168132738c.jpg) +GT +Ours w/o Differentiated Training + +![](images/4407eb6c81d62f54bacce99edcee9c1d7ce08ff3e9721c7a79cd68cd81802d11.jpg) +Ours w/o Reward Guidance + +![](images/e5c16242a5b0daaa4f08b4c83dd6afaf1a4811b932f1d7314f808d490ba31fe6.jpg) +Reward Map +Ours + +![](images/297cafc76e71a39a1a5bca5332f287399730d697ee7d9abe8ffa92ec6b2e0270.jpg) +Figure 4. Ablations of cross-lane view synthesis results with different experiment settings. Notice the restoration of vegetation color and the expression of high-frequency details. +GT +Figure 5. Visualization of reward map and layered diffusion output. The darker the areas in the reward map, the lower the confidence. + +![](images/15f4db9d75aab272d1a87d9c0861b18117652758e3b91a47becffc3b691812d5.jpg) +Reward Map +# + +![](images/bc3e2fba6f8c62a70209bfd68fa5379074f49577762c1674fa9569c5cc8240ec.jpg) + +![](images/13998830d8fef2a8478aa15a6621478cfa4734780b367e0bce21bc0af29cec1c.jpg) +Masked Image + +Additional comparisons on different diffusion prior models. We conduct an ablation study with two fine-tuned diffusion models on the ParaLane dataset: stable diffusion v1.5 (SD) [25] and stable video diffusion (SVD) [2]. As shown in Tab. 4, RGE_SVD results in light improvements compared to RGE_SD due to its enhanced temporal consistency, and the proposed RGE strategy can also enhance the performance of SVD. The ablation study proves that the reward map along with the differentiable training strategy can reduce the impact of hallucinations generated by different diffusion prior source. + +# 5. Conclusion and Future Work + +In this paper, we introduced RGE-GS, a novel method for expansive scene reconstruction that synergizes diffusion-based generation with reward-guided Gaussian integration. Our approach addresses the challenges of incomplete scene reconstruction from single-pass real scans by leveraging diffusion priors in a controlled and efficient manner. The key contributions of RGE-GS include a reward network that evaluates the pixel-level confidence of input diffusion priors, and a differentiated training strategy that categorizes Gaussians into missing, immature, and mature for optimized re + +construction. Extensive evaluations on publicly available datasets demonstrate that RGE-GS achieves state-of-the-art performance in both reconstruction quality and training efficiency. + +Our future work includes adopting video generation models to enhance the temporal consistency in diffusion model outputs, and exploring the integration of the reward network in a feed-forward Gaussian Splitting reconstruction pipeline for real-time reconstruction. + +# Acknowledgements + +We thank the reviewers for the valuable discussions. This research was supported by the Zhejiang Provincial Natural Science Foundation of China under Grant No. LD24F030001. + +# References + +[1] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM TOG, 28 (3):24, 2009. 2 +[2] 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:2311.15127, 2023. 1, 3, 7 +[3] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In ICML, pages 1-18, 2024. 2 +[4] Boyang Deng, Richard Tucker, Zhengqi Li, Leonidas Guibas, Noah Snavely, and Gordon Wetzstein. Streetscapes: Large-scale consistent street view generation using autoregressive video diffusion. In ACM SIGGRAPH, pages 27:1-27:11, 2024. 2 +[5] Lue Fan, Hao Zhang, Qitai Wang, Hongsheng Li, and Zhaoxiang Zhang. FreeSim: Toward free-viewpoint camera simulation in driving scenes. In CVPR, pages 12004-12014, 2025. 1, 2, 5 +[6] 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. arXiv:2403.20309, 2024. 2 +[7] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. MagicDrive: Street view generation with diverse 3D geometry control. In ICLR, pages 1-20, 2024. 1, 2 +[8] Ruiyuan Gao, Kai Chen, Zhihao Li, Lanqing Hong, Zhenguo Li, and Qiang Xu. Magicdrive3d: Controllable 3d generation for any-view rendering in street scenes. In ICLR, pages 1-17, 2025. 2 +[9] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. In NeurIPS, 2025. 2 +[10] Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, and Chunxia Xiao. Ggs: Generalizable gaussian splatting for lane switching in autonomous driving. In AAAI, pages 3329-3337, 2025. 2 +[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH, pages 32:1-32:11, 2024. 1, 5 +[12] Rui Jin, Yuman Gao, Yingjian Wang, Yuze Wu, Haojian Lu, Chao Xu, and Fei Gao. Gs-planner: A gaussian-splatting-based planning framework for active high-fidelity reconstruction. In IEEE IROS, pages 11202-11209, 2024. 2 +[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):139:1-139:14, 2023. 1, 3, 5 +[14] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, pages 1–14, 2014. 5 +[15] Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. DrivingDiffusion: Layout-guided multi-view driving scene video generation with latent diffusion model. In ECCV, pages 469-485, 2024. 2 +[16] Yueyuan Li, Wei Yuan, Songan Zhang, Weihao Yan, Qiyuan Shen, Chunxiang Wang, and Ming Yang. Choose your simulator wisely: A review on open-source simulators for autonomous driving. IEEE TIV, 9(5):4861-4876, 2024. 1 + +[17] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. ReconX: Reconstruct any scene from sparse views with video diffusion model. In ICLR, pages 1-22, 2025. 2 +[18] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3DGS-Enhancer: Enhancing unbounded 3D gaussian splatting with view-consistent 2d diffusion priors. In NeurIPS, pages 133305–133327, 2024. 2 +[19] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-GS: Structured 3D gaussians for view-adaptive rendering. In CVPR, pages 20654-20664, 2024. 1, 5 +[20] Yifan Lu, Xuanchi Ren, Jiawei Yang, Tianchang Shen, Zhangjie Wu, Jun Gao, Yue Wang, Siheng Chen, Mike Chen, Sanja Fidler, and Jiahui Huang. InfiniCube: Unbounded and controllable dynamic 3D driving scene generation with world-guided video models. arXiv:2412.03934, 2024. 1 +[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. Commun. ACM, 65(1):99-106, 2021. 3 +[22] Chaojun Ni, Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Wenkang Qin, Guan Huang, Chen Liu, Yuyin Chen, Yida Wang, Xueyang Zhang, Yifei Zhan, Kun Zhan, Peng Jia, Xianpeng Lang, Xingang Wang, and Wenjun Mei. ReconDreamer: Crafting world models for driving scene reconstruction via online restoration. In CVPR, pages 1559-1569, 2025. 2, 5 +[23] Ziqian Ni, Sicong Du, Zhenghua Hou, Chenming Wu, and Sheng Yang. Para-Lane: Multi-lane dataset registering parallel scans for benchmarking novel view synthesis. In 3DV, pages 1-10, 2025. 2, 5 +[24] 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 CVPR, pages 8931-8940, 2024. 3 +[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3, 5, 7 +[26] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241, 2015. 5 +[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 4 +[28] Qitai Wang, Lue Fan, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. FreeVS: Generative view synthesis on free driving trajectory. In ICLR, pages 1-18, 2025. 1, 2, 5 +[29] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jia-gang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-drive world models for autonomous driving. In ECCV, pages 55-72, 2024. 1, 2 +[30] 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. 4 + +[31] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians: Modeling dynamic urban scenes with gaussian splatting. In ECCV, pages 156-173, 2024. 5 +[32] Yunzhi Yan, Zhen Xu, Haotong Lin, Haian Jin, Haoyu Guo, Yida Wang, Kun Zhan, Xianpeng Lang, Hujun Bao, Xiaowei Zhou, and Sida Peng. StreetCrafter: Street view synthesis with controllable video diffusion models. In CVPR, pages 822-832, 2025. 2 +[33] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. NeurIPS, 37:21875-21911, 2024. 4 +[34] Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. In WACV, pages 3812-3822, 2025. 2 +[35] Hengyuan Zhang, David Paz, Yuliang Guo, Xinyu Huang, Henrik I Christensen, and Liu Ren. MapGS: Generalizable pretraining and data augmentation for online mapping via novel view synthesis. arXiv:2501.06660, 2025. 1 +[36] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 5 +[37] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 4 +[38] Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Xueyang Zhang, Yida Wang, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, Wenjun Mei, and Xingang Wang. DriveDreamer4D: World models are effective data machines for 4D driving scene representation. In CVPR, pages 12015-12026, 2025. 2 +[39] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. In AAAI, pages 10412–10420, 2025. 2 \ No newline at end of file diff --git a/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/images.zip b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bc137349714490676feeb42b861636ba3b90b3eb --- /dev/null +++ b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22f46044d082c9e10a0e023a11ca976ac5759883c4b992ee57a84198058dab8f +size 553964 diff --git a/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/layout.json b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..67194248e4c8f19d2cfd753bc31d7ebfd2ae8ca5 --- /dev/null +++ b/ICCV/2025/RGE-GS_ Reward-Guided Expansive Driving Scene Reconstruction via Diffusion Priors/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1816bc656c218a3829fb1acc8ff1602991c15fce2b4c5d5adf0251accc51d0ac +size 342673 diff --git a/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/f87eaa72-b153-4f71-96eb-cd933a44db5d_content_list.json b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/f87eaa72-b153-4f71-96eb-cd933a44db5d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e266d29c34dd695170d89183ddd222d9c7f81ab7 --- /dev/null +++ b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/f87eaa72-b153-4f71-96eb-cd933a44db5d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e209b9248a67c273a32a9cfae506c809fa86d158151f134ac584cf0a3990670 +size 90069 diff --git a/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/f87eaa72-b153-4f71-96eb-cd933a44db5d_model.json b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/f87eaa72-b153-4f71-96eb-cd933a44db5d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..242812dc69e5af7e2d1a1236bc0e87f1fc31f0be --- /dev/null +++ b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/f87eaa72-b153-4f71-96eb-cd933a44db5d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fc1557361ea61551068005e5ac4bf5c8385e7267e0569dc05c0f95b2411fa46 +size 110304 diff --git a/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/f87eaa72-b153-4f71-96eb-cd933a44db5d_origin.pdf b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/f87eaa72-b153-4f71-96eb-cd933a44db5d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f95508634171cfb36d868405acc5c4c3cbb43d9a --- /dev/null +++ b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/f87eaa72-b153-4f71-96eb-cd933a44db5d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a17e9de5bf9b98ba29bc8a9431ca7552c65705f3318c413d5ce83d5631d0fbf9 +size 12064486 diff --git a/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/full.md b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..08a61d79df901fd9b2a7a9a8fc4b11ddcdda8a9d --- /dev/null +++ b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/full.md @@ -0,0 +1,393 @@ +# RI3D: Few-Shot Gaussian Splitting With Repair and Inpainting Diffusion Priors + +Avinash Paliwal1 + +Xilong Zhou $^{1,3}$ + +Wei Ye2 + +Jinhui Xiong² + +Rakesh Ranjan + +Nima Khademi Kalantari + +$^{1}$ Texas A&M University + +Meta Reality Labs + +3Max Planck Institute for Informatics + +![](images/abb5387a396450cc610c6a90df5f60c67a67ba05e1b029a72cd3b46677b3b77a.jpg) + +![](images/fa2f27cfa6bf3e994f1e581a1c36fb8eccb120d4f252e60b210d6c94b1d2e923.jpg) +Inputs + +![](images/2c8444449af19fe586c90312322abbc0624f86116c07d1b3320a18a937a4dcb3.jpg) +Figure 1. We introduce a novel sparse view synthesis method that employs two diffusion models, "repair" and " inpainting", which are responsible for aiding in the reconstruction of visible regions and hallucinating missing regions, respectively. Our approach involves a two-stage optimization process. In the first stage, we use the repair model to constrain the 3DGS optimization and reconstruct the regions covered by the input images. As shown, the output of the first stage properly reconstructs the visible areas, but contains missing regions, which are marked in white. In the second stage, we utilize the inpainting model to fill in these missing areas and continue optimization using the repair model to seamlessly integrate the hallucinated regions with the rest of the scene. Here, we compare our method ("Stage 2") against several state-of-the-art techniques on a $360^{\circ}$ scene using only three input images. + +![](images/8ff416abe0bb11b7a75516b1a294493ea2f7f6bb12b2f7d6e1f1bef9974e48c8.jpg) + +![](images/59b245cbaeb897ce48b887c673341c7728b4b497d0935a8d651c6c3104d50cd1.jpg) + +![](images/ffadcc7e5869c9244f022153763b94dc38f97d3808c8a011780665ac3d3eba18.jpg) + +![](images/405085813b64e8c5184fc4a2940f8c37f9c15f134fe1b3db52f921694c6ab478.jpg) + +![](images/e63f7048fdef790164c16b7cf2ea1e37cc3f089c19daf2e4dbe5a6fab21747fb.jpg) + +![](images/7a72ff76dd495dc6751a92a02f9c9187df16938919d4bc4eb30c90b053bc441d.jpg) + +# Abstract + +In this paper, we propose R13D, a novel 3DGS-based approach that harnesses the power of diffusion models to reconstruct high-quality novel views given a sparse set of input images. Our key contribution is separating the view synthesis process into two tasks of reconstructing visible regions and hallucinating missing regions, and introducing two personalized diffusion models, each tailored to one of these tasks. Specifically, one model ('repair') takes a rendered image as input and predicts the corresponding high-quality image, which in turn is used as a pseudo ground truth image to constrain the optimization. The other model (' inpainting') primarily focuses on hallucinating details in unobserved areas. To integrate these models effectively, we introduce a two-stage optimization strategy: the first stage reconstructs visible areas using the repair model, and the second stage reconstructs missing regions with the inpainting model while ensuring coherence through further optimization. Moreover, we augment the optimization with a novel Gaussian initialization method that obtains per-image + +depth by combining 3D-consistent and smooth depth with highly detailed relative depth. We demonstrate that by separating the process into two tasks and addressing them with the repair and inpainting models, we produce results with detailed textures in both visible and missing regions that outperform state-of-the-art approaches on a diverse set of scenes with extremely sparse inputs. + +# 1. Introduction + +The introduction of novel 3D representations, such as neural radiance fields (NeRF) [18] and 3D Gaussian splatting (3DGS) [14], has revolutionized the field of novel view synthesis. While these techniques excel at reconstructing 3D scenes from a large number of images, view synthesis from a sparse set of images remains a challenging problem. + +Most state-of-the-art sparse novel view synthesis approaches [16, 32, 48] introduce a series of regularizations to constrain the optimization process and avoid overfitting to the input images. While these approaches produce re + +sults with reasonable texture in areas visible in a few input images, they often struggle to hallucinate details in the occluded regions. This issue is especially pronounced in $360^{\circ}$ scene reconstruction with extremely sparse input images where there are larger missing regions and areas covered by only a single image (see Fig. 1). + +Recently, a couple of techniques [8, 38] propose to address this issue using a diffusion model as a prior during the optimization process. Specifically, they first train a diffusion model [25] on a large multiview image dataset to synthesize novel views given a set of input images. They then perform the optimization using a combination of the input images and novel views, synthesized using this view synthesis diffusion model. Although these methods produce significantly better results than the previously discussed techniques, particularly in $360^{\circ}$ scenes, they often overblur details, especially in missing areas (see supplementary video). This is because although their view synthesis diffusion models produce visually pleasing novel views, the reconstructed images, particularly in occluded regions, are not 3D-consistent. Using such images during optimization thus leads to overblurring of details. Additionally, since these methods use NeRF as their 3D representation, they suffer from slow rendering times. + +To address this issue, we propose to separate the view synthesis process into two tasks of reconstructing the visible and missing areas and introduce two diffusion models, each specialized to help with one of these two tasks. Specifically, inspired by Yang et al.'s approach [41], one diffusion model (repair model) is responsible for taking a rendered image as the input and producing a clean image. This model effectively suppresses the artifacts in the reconstructed images during the optimization process. The second diffusion model ( inpainting model) is solely responsible for hallucinating details in the missing regions. To ensure these two diffusion models produce results consistent with the scene at hand, we personalize them by tuning the models on the input images. + +To ensure fast inference speed, instead of using NeRF as done in the previous approaches [8, 38], we utilize 3DGS as our 3D representation. To initialize the Gaussians, we propose to obtain per image depth by combining the 3D-consistent, but smooth, depth estimates from DUSt3R [37] with the highly detailed relative depth from a monocular depth estimation approach through Poisson blending [22]. Using the estimated depth, we then assign one Gaussian to each pixel of every input image and project them into 3D space. Moreover, to effectively utilize the two diffusion models, we propose a two-stage optimization process. The goal of the first stage is to reconstruct the scene with detailed texture in the visible areas (see Fig. 1 "Stage 1") by utilizing the repair model to enhance the renderings and using them as pseudo ground truth. During the second stage, + +we hallucinate details in the missing areas using the inpainting model and continue the optimization by utilizing the repair model to seamlessly integrate the hallucinated details with the rest of the scene. + +We show that our approach, dubbed RI3D, produces high-quality textures particularly in occluded areas, for challenging scenarios. We further demonstrate that our results outperform the state of the art, both numerically and visually. In summary, we make the following contributions: + +- To improve the view synthesis process, we propose to utilize two personalized diffusion models: one for enhancing the rendered images and using them as pseudo ground truth during optimization and another for hallucinating details in the missing areas. +- We propose a novel approach to initialize Gaussians by combining 3D-consistent, smooth depth maps with highly detailed relative depth from monocular approaches. +- We introduce a two-stage optimization strategy that seamlessly incorporates the two diffusion models. + +# 2. Related Work + +# 2.1. Radiance Field and 3D Gaussian Splitting + +Neural Radiance Field (NeRF) [18] is an optimization-based technique for reconstructing 3D scenes from dense input images. The key idea of NeRF is to encode a scene into an implicit neural network that takes a 3D position and view direction as inputs and outputs opacity and view-dependent color. By minimizing the loss between input images and renderings, the implicit neural network is optimized to represent real-world scenes. NeRF has gained considerable attention, inspiring a significant amount of follow-up work to improve rendering quality [1, 2, 4, 35] and efficiency [5, 9, 19, 23, 26]. + +The key limitation of NeRF, however, is its slow training and inference, as rendering requires evaluating the network multiple times along a ray. Kerbl et al. [14] address this problem by proposing 3D Gaussian Splatting (3DGS), which explicitly models the scene using a set of Gaussian primitives. We build our approach on 3DGS, as it produces results comparable to NeRF at a lower computational cost. + +# 2.2. Sparse Novel View Synthesis + +While both NeRF and 3DGS demonstrate high-quality rendering given dense input image sampling, they struggle with sparse input views. Several techniques have been proposed to address the problem of sparse-input novel view synthesis using NeRF [6, 13, 20, 30, 33, 42, 44] and 3DGS [21, 40, 48] representations. + +More specifically, for NeRF-based methods, PixelNeRF [44] learns an image-based 3D feature extractor as a prior to optimize NeRF from sparse input views. DietNeRF [13] introduces an auxiliary semantic consistency + +loss to encourage realistic renderings from novel views. RegNeRF [20] proposes geometry and appearance regularization from novel viewpoints. DS-NeRF [6] incorporates depth supervision provided by structure-from-motion into the NeRF pipeline. SparseNeRF [40] introduces a local depth ranking method and spatial continuity constraints to regularize optimization. FreeNeRF [42] regularizes the optimization by reducing the frequency of positional encoding, while SimpleNeRF [30] provides additional supervision through point augmentation. + +Among 3DGS-based methods, FSGS [48] introduces monocular depth supervision and proposes a specially designed densification strategy. SparseGS [40] presents a novel explicit operator for 3D representations to prune floating Gaussians. CoherentGS [21] enhances coherence in the Gaussian representation by constraining the movement of Gaussians and introducing single- and multi-view constraints. DNGaussian [16] employs both hard and soft depth regularization to improve sparse-view reconstruction by enforcing surface completeness. + +While these methods significantly reduce the number of input images required for high-quality view synthesis, they still struggle to achieve robust results with extremely sparse inputs (e.g., three images). Additionally, they do not provide a reliable approach for reconstructing missing regions and typically fill these areas with overly smooth content. + +# 2.3. Diffusion-based Methods + +In recent years, Diffusion Models (DM) [12, 25] have stood out in image generation tasks due to stable training and high-quality results. Because of their success, these models have been used extensively as a prior for various tasks including view synthesis [7, 8, 10, 11, 17, 27, 28, 34, 38, 41]. + +The majority of these methods utilize diffusion models for synthesizing novel views of objects. Specifically, GaussianObject [41] leverages ControlNet [46] to repair the artifacts introduced by 3DGS optimization and synthesize 3D objects only from four input images. We follow a similar strategy to enhance the quality of rendered images, but focus on reconstructing scenes. + +For general scenes, ReconFusion [38] and CAT3D [8] train a view synthesis diffusion model, which is further used to regularize NeRF optimization. The main issue with these approaches is that the images reconstructed using the diffusion model are not 3D-consistent, causing the optimized NeRF model to produce blurry results, particularly in the missing areas. In contrast, we introduce two diffusion models to enhance renderings and inpaint missing areas, and we propose a two-stage optimization process that enables the reconstruction of detailed textures in both visible and missing regions. + +# 3. Background + +In this section, we review two fundamental techniques used in our algorithm: 3D Gaussian splatting and diffusion models. Specifically, we use 3DGS as our 3D representation due to its fast rendering pipeline and diffusion models are used as priors to regularize the optimization process for spare input views. + +# 3.1. 3D Gaussian Splatting + +3DGS [14] is a point-based rendering technique representing a scene with a dense set of 3D Gaussians, which can achieve high-quality, fast and differentiable scene rendering. In 3DGS, each 3D Gaussian is defined using a set of estimizable parameters: 3D position $\mathbf{x}$ , opacity $\sigma$ , anisotropic covariance matrix $\Sigma$ , and color $\mathbf{c}$ represented with spherical harmonics (SH) coefficients. Given the representation, point-based $\alpha$ -blending is used to render the color $\mathbf{c}$ at each pixel $\mathbf{p}$ as follows: + +$$ +\mathbf {c} (\mathbf {p}) = \sum_ {i \in N} \mathbf {c} _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {1} +$$ + +where $N$ represents the number of Gaussians overlapping with pixel $\mathbf{p}$ , $\mathbf{c}_i$ is view-dependent color computed from SH coefficients of the $i^{\mathrm{th}}$ Gaussian, and $\alpha_{i}$ is the effective opacity, obtained by evaluating the Gaussian and multiplying it with the per-Gaussian opacity $\sigma_{i}$ . During optimization, given a set of input images, the parameters of the Gaussian particles are optimized by minimizing the loss between the input and rendered images. + +# 3.2. Diffusion Models + +Diffusion models [25] are a class of generative models that create data matching the target distribution $q(\mathbf{x}_0)$ by progressively denoising Gaussian noise $\varepsilon$ . Specifically, in the forward process, noise is added to clean data $\mathbf{x}_0$ over $T$ steps, producing a sequence of increasingly noisy data $\mathbf{x}_0, \ldots, \mathbf{x}_T$ . The reverse process then uses the diffusion model to invert this sequence, iteratively denoising from $\mathbf{x}_T$ back to reconstruct $\mathbf{x}_0$ . Generating high-resolution images with diffusion models is computationally intensive and memory-demanding. To address this issue, latent diffusion models (LDMs) [25] perform the diffusion process in the latent space of a variational autoencoder (VAE) [15], reducing memory and computational requirements. Our repair and inpainting models are based on LDMs. + +# 4. Algorithm + +Given a sparse set of $N$ images $\mathbf{I}_1^{\mathrm{ref}},\dots ,\mathbf{I}_N^{\mathrm{ref}}$ with their corresponding camera poses, our goal is to reconstruct the scene using a 3D Gaussian representation. To do this, we start by initializing the Gaussians using a detailed and 3D-consistent depth estimate at each input image. We then + +![](images/f158bbb2547a035d5fecbf1819f1a0dd4b8ebdb0c1da74f8f117634ba738734c.jpg) +Figure 2. We provide an overview of the different stages of our approach. First, we initialize the Gaussians by generating high-quality per-view depth maps (Sec.4.1). Next, we fine-tune the repair and inpainting diffusion models on the scene at hand (Sec.4.2). Finally, we use these models to optimize the 3DGS representation in two stages (Sec. 4.3). In the first stage, we reconstruct the areas covered by the input images (blue), using the repair model to generate pseudo ground truth images at $M$ novel views (orange) to constrain the optimization. In the second stage, we first select a subset of novel views (green) to inpaint the missing regions (left) and continue the optimization using the repair model (right). This process of inpainting and optimization is repeated multiple times until all missing areas are reconstructed. + +perform a two-stage optimization by utilizing two diffusion models ("repair" and " inpainting"), personalized for the scene at hand, to enhance the rendered images and inpaint the missing areas. An overview of our approach is provided in Fig. 2. In the following sections, we discuss our initialization process, the two diffusion models, and the two-stage optimization process. Additional implementation details, such as hyperparameters and learning schedules, can be found in the supplementary materials. + +# 4.1.3D Gaussian Initialization + +Initialization is one of the key factors in effectiveness of 3DGS optimization, especially in sparse input settings. Ideally, we would like to start with dense and 3D-consistent Gaussians in areas covered by the input images. A potential solution is to use the point cloud estimated by the multi-view stereo (MVS) technique by Wang et al. [37] (DUSt3R), which has been shown to perform significantly better than other MVS techniques, particularly with a small number of images. However, DUSt3R's point cloud is usually only accurate in the high-confidence regions where there is overlapping content from multiple input images. For example, for a $360^{\circ}$ scene with extremely sparse inputs (e.g., 3 images), the high confidence point cloud generally covers only a small foreground region and is usually sparse. The point cloud in the background areas, which are covered by only a single image, is often highly inaccurate. + +To obtain a dense initialization, we instead propose generating per-image depth maps. We then assign a Gaussian to each pixel of every input image and project them into 3D space according to their depth. The key challenge here is obtaining 3D-consistent and detailed depth maps, even in regions covered by only a single image. Our main observation is that the depth estimated by DUSt3R (one of the byproducts of this approach) and monocular depth estimation methods, such as Depth Anything V2 [43], have complementary properties (see Fig. 3). Specifically, DUSt3R depth in high-confidence regions is 3D-consistent and accurate but smooth, as the estimation is performed on low-resolution images. Additionally, its depth is inaccurate + +![](images/70e2930b6d0491dd935d0ace92c991c59249495a8212897fded7c406a9523265.jpg) +Reference Image + +![](images/0de839e1b57997ebbc4a99c110a7491a95565cdf057eb6ec7e4957b4509921c9.jpg) +DUSt3R + +![](images/41490c07ec27e0bc68c81875cce74ffa1f534a15a42d67eb13290378cae1ef86.jpg) +Confidence Mask + +![](images/13951c14db7000d2670db8ddd8dfaa7ada3d0f765efb69993385fee6f0df610e.jpg) +Monocular Depth +Figure 3. The depth estimated by DUSt3R is geometrically consistent in the high confidence regions (marked in yellow), but of poor quality in the remaining areas. Monocular depth is highly detailed, but is not 3D consistent. Our proposed method combines the two depth maps into a detailed and geometrically consistent depth. Applying bilateral filtering, further sharpens the boundaries. + +![](images/3b873cfecd4023364c4e3ce22edbeebc3538281c15685dd70897894d16e30626.jpg) +Combined + +![](images/e509e357e4625c4af1c65490d0775d2ba518962cdd60aa83e7ee303f05934caa.jpg) +Combined + Bilateral Filtering + +in low-confidence regions. On the other hand, monocular depth is detailed but relative and not 3D-consistent, as the depth maps of each image are obtained independently. + +We therefore propose combining the two depth maps to leverage the strengths of both approaches. Specifically, we aim for the combined depth to have two properties: 1) it should preserve the details (edges) from the monocular depth, and 2) it should adhere to the absolute depth values from DUSt3R in high-confidence regions. Combining these properties results in the following objective with two terms: + +$$ +\mathbf {d} ^ {*} = \arg \min _ {\mathbf {d}} \left[ \mathbf {M} \odot \| \mathbf {d} - \mathbf {d} ^ {D} \| _ {2} + \lambda \| \nabla \mathbf {d} - \nabla \mathbf {d} ^ {\mathrm {M}} \| _ {2} \right], \tag {2} +$$ + +where $\mathbf{d}^D$ and $\mathbf{d}^{\mathrm{M}}$ are the DUSt3R and monocular depth estimates, respectively, and $M$ is a mask indicating the regions where DUSt3R has high confidence. The first term enforces similarity between the combined and DUSt3R depth maps in high-confidence regions, while the second term ensures that the gradient of the combined depth matches that of the monocular depth everywhere. The weight $\lambda$ controls the contribution of the gradient term, which we set to 10 in our implementation. + +Notably, this objective is similar to the Poisson blend + +ing equation [22], but with a key difference: we enforce the gradient loss everywhere, not just in the low-confidence areas $(1 - M)$ , since DUSt3R depth is smooth, and we aim to incorporate details from monocular depth even in high-confidence regions. As the objective is quadratic, we obtain the solution by solving a linear system of equations (derived by setting the gradient of the objective to zero), which can be efficiently computed using sparse matrix solvers. + +Directly using the monocular depth as $\mathbf{d}^M$ in our objective is problematic, as DUSt3R and monocular depth can have widely different ranges. To address this, we globally align the monocular depth map to DUSt3R depth using a piecewise linear function. Since DUSt3R depth is inaccurate in low-confidence regions, we perform the alignment only in high-confidence regions and linearly extrapolate the function for other values. Once we obtain this piecewise linear function, we apply it to the monocular depth to align it with DUSt3R depth. We then use this aligned monocular depth as $\mathbf{d}^M$ in Eq. 2 to compute the combined depth. Finally, we apply bilateral filtering, as in previous approaches [29, 36], to the combined depth to produce per-image depth maps with sharp edges (see Fig. 3). + +Given the per-image depth maps, we assign a Gaussian to each pixel and project it into 3D space along the ray connecting the camera and pixel center [21]. We initialize each Gaussian's color using the corresponding pixel color and set the rotation matrix to identity. Additionally, we use an isometric scale, setting it so the projected Gaussians cover 1.4 times the pixel size, and assign an initial opacity of 0.1 to all Gaussians. Figure 4 shows our output after initialization. + +# 4.2. Repair and Inpainting Diffusion Models + +In extremely sparse input settings, 3DGS optimization, even with highly accurate initialization, is brittle and unable to properly reconstruct the scene due to two issues (see Fig. 5). First, the optimization can quickly overfit to the input views, resulting in a representation that produces poor-quality results when rendered from novel camera angles. Second, due to the lack of supervision in occluded regions, these areas are often not reconstructed properly, resulting in blurry or dark appearances in the rendered images. + +To address these issues, we propose using two diffusion models. Specifically, the first diffusion model, referred to as the "repair" model, takes as input a rendered image from a view not covered by the input cameras and produces a clean version of the image. The key idea is that the repair model can generate clean images for a set of novel views, which can then be used as pseudo ground truth images during optimization to address the first problem. The second diffusion model, referred to as the " inpainting" model, tackles the second issue by filling in the missing areas. + +Following Yang et al.'s approach [41], we use a pretrained ControlNet [46] as our base repair model and fine + +![](images/3ba110302497d288ba491ee67f88caa0b899974842d34d9a8e033830fde06d07.jpg) +Initialization + +![](images/22c3bca9a1b61638f1dce48614085da567f666ac8f4999c8b3d771f5a1a75d43.jpg) +Optimization Stage 1 + +![](images/5b8ae3c4b39b19ff992753c231f97b7a15b6b41c33ec1d5b62709e437e972405.jpg) +Optimization Stage 2 + +![](images/d50d81d3b2b6ef8c3184ec46ed40128fa443ddf7cb7bc71df07640fa113ac87f.jpg) +Rendered Novel +Figure 5. We show the result of 3DGS optimization using our initialization. In the absence of any constraints, 3DGS optimization quickly overfits to the input images (compare rendered and input images) but produces distracting artifacts in the novel view image. Additionally, unobserved areas will not be reconstructed during optimization, resulting in a dark and blurry appearance. We address these issues using our repair and inpainting models to constrain the optimization and hallucinate missing areas. + +![](images/f3426f54d5a47ea008f34b440433031a41d6992317b0370c63946fe46791b6fb.jpg) +Rendered Input + +![](images/039d4605cafb0ab43dfe7370adf901b66850ab5e8e00106f0c27c4fafd0d492f.jpg) +Figure 4. We show the output at different stages of our approach. Our initialization strategy ensures that Gaussians from different input images are roughly aligned and cover the visible areas of the scene. During the first stage of optimization, we use the repair model to constrain the problem, which in turn helps reconstruct the visible regions with detailed texture. The missing areas are then hallucinated and seamlessly incorporated into the scene during the second and final stage of optimization. +Input + +tune it on the target scene by generating pairs of corrupted and clean images through a leave-one-out strategy. Specifically, we create $N$ subsets of images, each containing $N - 1$ images by excluding one input image. We then optimize $N$ separate 3DGS representations on these subsets. After a set number of iterations, we reintroduce the excluded image to its subset and continue optimization for an additional fixed number of iterations. During this process, we render the left-out view at different stages of optimization, pairing these intermediate renderings (corrupted images) with the original left-out image (clean image) to form the training data. Using the corrupted image as the condition for ControlNet and the clean image as ground truth, we train the diffusion model using the standard loss [41]. Note that by initially excluding an image and later reintroducing it, we generate progressively refined corrupted images, improving the repair model's ability to handle the final 3DGS optimization (discussed in Sec. 4.3). + +For inpainting, we use the Stable Diffusion inpainting model [25]. Although this model is powerful, the missing regions in our problem are often large, providing limited context for hallucinating consistent details. To address this, we follow Tang et al. [31]'s approach and fine-tune the diffusion model on the input images by simulating a large set of input-output pairs through random masking. + +Since both diffusion models operate on $512 \times 512$ images, we resize the input images during fine-tuning so that their smallest dimension (typically height) is 512 pixels, fol + +lowed by random $512 \times 512$ cropping to create training data. Resizing is applied first to preserve scene context, which is essential for both the repair and inpainting models, as directly cropping high-resolution images could remove important contextual information. + +# 4.3. Optimization + +We now have all the elements to begin discussing our proposed optimization strategy, consisting of two stages. The goal of the first stage of the optimization is to obtain a reasonable 3D representation covering only the regions that are visible in the input images. This is a critical step for properly identifying the missing areas. In the second stage, the missing regions are inpainted and the representation is further optimized to obtain a complete scene (see Fig. 4). Below, we discuss the two stages in detail. + +Stage 1: Optimizing solely by minimizing the loss between the rendered and input images can lead to overfitting. To address this, we introduce a set of $M$ novel cameras along an elliptical path aligned with the input cameras. To leverage these novel cameras during optimization, we generate pseudo ground truth images, $\mathbf{G}_j^{\mathrm{nov}} = R(\hat{\mathbf{I}}_j^{\mathrm{nov}})$ , by applying our repair diffusion model $R$ to the rendered images $\hat{\mathbf{I}}_j^{\mathrm{nov}}$ . In summary, we optimize the following objective: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {s t a g e 1}} = \sum_ {i = 1} ^ {N} \mathcal {L} _ {\text {r e c}} \left(\hat {\mathbf {I}} _ {i} ^ {\text {r e f}}, \mathbf {I} _ {i} ^ {\text {r e f}}\right) + \sum_ {j = 1} ^ {M} \lambda_ {j} \mathbf {M} _ {j} ^ {\alpha} \mathcal {L} _ {\text {r e c}} \left(\hat {\mathbf {I}} _ {j} ^ {\text {n o v}}, \mathbf {G} _ {j} ^ {\text {n o v}}\right) \\ + \sum_ {j = 1} ^ {M} \| \mathbf {A} _ {j} \odot \left(1 - \mathbf {M} _ {j} ^ {\alpha}\right) \odot \mathbf {M} _ {j} ^ {b} \| _ {1}. \tag {3} \\ \end{array} +$$ + +The first term enforces similarity between the rendered $\hat{\mathbf{I}}_i^{\mathrm{ref}}$ and the input images. The reconstruction loss comprises the original 3DGS losses ( $L_{1}$ and SSIM), as well as LPIPS and depth correlation [48] losses. For the depth correlation loss, we compute the Pearson correlation between the rendered depth from 3DGS and the monocular depth estimated from the input image. The second term encourages similarity between the rendered images and their repaired counterparts from the $M$ novel cameras. The factor $\lambda_{j}$ represents the camera distance weight [41], giving higher weight to novel cameras closer to the input cameras. Additionally, $\mathbf{M}_j^\alpha$ is a binary mask obtained by thresholding the rendered opacity, ensuring that the loss is enforced only in regions covered by the input images. Finally, the third term encourages the rendered opacity $\mathbf{A}_j$ to be low in the missing regions ( $1 - \mathbf{M}_j^\alpha$ ), as we do not want 3DGS to place any visible Gaussians in these areas. Note that $\mathbf{M}_j^b$ is a mask identifying the background regions, which we use here because the occluded regions in the foreground are typically small and can be easily inpainted with our repair model. We obtain this background mask by applying agglomerata + +tive clustering [36] on the monocular depth estimated for repaired images. + +To summarize, this objective provides additional supervision using the repair model for $M$ novel cameras to constrain the optimization. Moreover, we focus solely on reconstructing the visible regions by ensuring that the opacity of Gaussians in the missing areas is minimized through the third term. See our output at this stage in Fig. 4. + +Stage 2: At the end of stage 1, the 3DGS representation can reconstruct the visible areas from the input images but lacks information in the missing regions, $(1 - \mathbf{M}_j^\alpha) \odot \mathbf{M}_j^b$ . To address this, we first use our personalized inpainting model to fill in the missing areas in novel camera views and then integrate the hallucinated details into the scene through optimization. Specifically, we select a subset of $K < M$ novel view images with non-overlapping content and inpaint the missing areas using our inpainting diffusion model. We inpaint only a subset of images to prevent independently inpainting overlapping content, which could lead to inconsistent results. We then project the inpainted areas into the scene using monocular depth estimated from the inpainted images. Since monocular depth is relative and has a different range than the reconstructed 3D scene, it cannot be used directly. We address this by combining the monocular depth in the inpainted regions with the rendered depth using Eq. 2. + +This process results in $K$ inpainted images, $\hat{\mathbf{L}}_j^{\mathrm{nov}}$ . We then apply our repair model to these images, as well as the remaining $M - K$ novel view images, to obtain the pseudo ground truth images, $\mathbf{G}_j^{\mathrm{nov}}$ . These images are then used to optimize the 3DGS representation by minimizing the following loss: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {s t a g e 2}} = \sum_ {i = 1} ^ {N} \mathcal {L} _ {\text {r e c}} \left(\hat {\mathbf {I}} _ {i} ^ {\text {r e f}}, \mathbf {I} _ {i} ^ {\text {r e f}}\right) + \sum_ {j = 1} ^ {M} \lambda_ {j} \mathcal {L} _ {\text {r e c}} \left(\hat {\mathbf {I}} _ {j} ^ {\text {n o v}}, \mathbf {G} _ {j} ^ {\text {n o v}}\right) \\ + \sum_ {k = 1} ^ {K} \left(1 - \mathbf {M} _ {k} ^ {\alpha}\right) \odot \mathbf {M} _ {k} ^ {b} \odot L _ {p} \left(\hat {\mathbf {I}} _ {k} ^ {\text {n o v}}, \hat {\mathbf {L}} _ {k} ^ {\text {n o v}}\right). \tag {4} \\ \end{array} +$$ + +The first and second terms are similar to those in Eq. 3. The main difference is that, since most of the scene has now been inpainted, we enforce the loss in the second term across the entire image by removing the visibility mask $\mathbf{M}_j^\alpha$ . The third term ensures that the rendered images $\hat{\mathbf{I}}_j^{\mathrm{nov}}$ remain close to the inpainted images $\hat{\mathbf{L}}_j^{\mathrm{nov}}$ in the inpainted areas, $(1 - \mathbf{M}_j^\alpha) \odot \mathbf{M}_j^b$ . We perform this optimization for a fixed number of iterations, then repeat the inpainting step with a different subset of $K$ images before optimizing again. This iterative process continues until the remaining missing areas in the 3D scene are progressively filled. + +At the end of this stage, we obtain a complete 3D representation of the scene, which can be rendered from novel camera views, as shown in Fig. 4. + +![](images/fe02c38502a4a098c565c494d3757c316fdaa77f73571be20c5ea8ace222d710.jpg) + +![](images/14bfbb5c11ffcb02be807980f5f4b9928ade312f2fd0b0bba8f232aead5e51f6.jpg) + +![](images/4520f640a4f34e6a6e30c1db201c5a1123066795ab77f19089764e8abb1af092.jpg) +DNGaussian + +![](images/66efd99b6390a0782aa1b709168d6ffbc14178a051efe3f48e094221b4990de7.jpg) + +![](images/ba01647f59470c9774ab27390be4c2838a7d203f16c3effa8182359e44a29d62.jpg) + +![](images/9275373cb268fa91d980a1224962645727a61a56370f3a6ddcfe11a7c5c1466e.jpg) +CoR-GS + +![](images/7f3455521256c19e52c736eddd9f8304aa015c4d7295cc0e87d158e00268cdd5.jpg) + +![](images/5cf9a96429ae7bf25838c57e431c2499205327fdd4ccded1105fa3214ed0e55d.jpg) + +![](images/bc6567bb09d2f5648ab0028b9b1efad32ba9e21ef20d91a1b6f5b8a176a59490.jpg) +FSGS + +![](images/915b2f598ad74af0aaf7f299bc5548d1da9ff8c0ec989d2e2d72c5398b7ed741.jpg) + +![](images/3fb604149c9539bdb718822f323635a0e7843f690c0fb9d69b4f8e2c191686e6.jpg) + +![](images/67b3e9869aabf1b4f2df3023d72576985e8ac7bbe5778bea4fff00526e97c1b3.jpg) +Ours + +![](images/56a12b288cdd5ded9521127fa127c59a13ec69d2b0aa8ae335fe1b3128c5b27a.jpg) + +![](images/b659de679083795dc1e246d9525a74c76d0a4da14eaea077debc56505b62226f.jpg) + +![](images/60037d4d756f2057d2cd109c5bf020fe01bc9f6eb304e3886d00736ddbbbacaf.jpg) +Ground Truth + +![](images/885824f2227cccc6af895bc31581bb82e4fc9078882ce1375e8a407a0b75e12a.jpg) +Figure 6. We compare our approach against the other state-of-the-art methods. Existing methods are not able to properly handle both the visible (green arrows) and missing (red arrows) regions. Through the use of repair and inpainting models, our approach is able to produce high-quality textures in all areas. +Rendered Novel +Figure 7. We show the rendered novel view image of an inset of the Garden scene and the enhanced version using our repaired model. The repaired image is used as the pseudo ground truth during optimization to help with reconstructing detailed textures. + +![](images/a1dbf5e0054d8208b0ace72e594cea489140c3ebff936973eb1559a3c5f25fd7.jpg) +Repaired Novel + +# 5. Results + +We compare our approach against several state-of-the-art methods both visually and numerically and evaluate the effect of various aspects of our proposed method. Here, we show the results on the Mip-NeRF 360 [3] dataset but provide additional results on CO3D [24] and video comparisons in the supplementary materials. + +# 5.1. Visual Results + +In Fig. 6, we compare our approach against several 3DGS-based approaches: DNGaussian[16], CoR-GS [45], and FSGS [48]. We use the authors' publicly available source code for these methods. As the code for Reconfusion [38] and CAT3D [8] is unavailable, we cannot include their results in this figure. However, we show a comparison against their publicly available video results in our supplementary video. As shown in Fig. 6, existing approaches fill the missing areas (red arrows) using either a few large Gaussians or many small Gaussians, leading to a blurry or noisy appearance (the latter in the second and third rows of DNGaussian). In contrast, by utilizing our powerful personalized inpainting model, we are able to reconstruct these areas with detailed textures. + +Moreover, existing approaches struggle with properly reconstructing the visible areas (indicated by the green arrows) in these extremely sparse input settings. On the other hand, by providing additional novel view supervision through the repair model, our method is able to reconstruct the visible areas reasonably well. For example, as shown in Fig. 7, our repair model is able to produce high-quality texture on the table from the ghosted rendering. Using these repaired images as pseudo ground truth during optimization enhances texture detail in visible areas. + +# 5.2. Numerical Results + +In Table 1, we numerically evaluate our approach by reporting the average error in PSNR, SSIM, and LPIPS [47] across all scenes in the Mip-NeRF 360 [3] dataset. Our method outperforms all other approaches in PSNR and SSIM, except for ReconFusion and CAT3D, where our results are slightly worse. However, this is primarily because ReconFusion and CAT3D generate smooth, blurry textures in the missing regions (see supplementary video), which are favored by PSNR and SSIM. In contrast, our method achieves the best performance in LPIPS, highlighting its ability to synthesize detailed textures. + +# 5.3. Ablations + +We evaluate the effect of the core components of our method both visually and numerically in Fig. 8 and Table 2, respectively. Directly using the DUSt3R depth (w/o depth enhancement) results in floating artifacts, as indicated by the red arrow, since DUSt3R produces inaccurate depth in low-confidence areas. By combining DUSt3R with monocular depth, our approach largely mitigates this issue. + +To evaluate the impact of two-stage optimization, we compare our approach against an alternative method where + +
Method3-view6-view9-view
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
DiffusioNeRF* [39]11.050.1890.73512.550.2550.69213.370.2670.680
FreeNeRF* [42]12.870.2600.71513.350.2830.71714.590.3190.695
SimpleNeRF* [30]13.270.2830.74113.670.3120.72115.150.3540.676
Zip-NeRF* [4]12.770.2710.70513.610.2840.66314.300.3120.633
ZeroNVS* [27]14.440.3160.68015.510.3370.66315.990.3500.655
DNGaussian† [16]12.020.2260.66512.710.2480.64612.970.2580.637
CoR-GS† [45]13.510.3140.63314.530.3280.59615.480.3560.574
FSGS† [48]13.140.2880.57814.640.3480.51516.000.3970.470
ReconFusion* [38]15.500.3580.58516.930.4010.54418.190.4320.511
CAT3D16.620.3770.51517.720.4250.48218.670.4600.460
RI3D (Ours)15.740.3420.50516.510.3920.46217.480.4390.415
+ +![](images/0977080f3826777c580dc728c5178a84c1000be3e9904c89bad13592d3e73a90.jpg) +w/o depth enhancement +Figure 8. We show the effect of various components of our system. The differences can be best seen in the supplementary video. + +![](images/78ae01d1b0d24046fbbdbd898f8cea140f411c3c5a457a45131ee7d059c73f70.jpg) +Ours + +![](images/fa3899d2c634d34b923a8ae2a934b1409d6b8fdbdcc53e52aa8f0c30e697591b.jpg) +single stage + +![](images/1e105e38362ed1d3a65121ea6d29e20b4fa1a5700570ec627cfbd760caef3e15.jpg) +Ours + +![](images/57d506a762126926b14f63765d92f2d6a4aaa5a0ced8882385f800dd4e02bd66.jpg) +w/o ft inpainting + +![](images/eba97ad4689a93ce34f94e32cad6871c01703adf196f43c9da04d2ecb1667d26.jpg) +Ours + +Table 1. We compare our approach against other sparse view synthesis methods. PSNR and SSIM measure pixel-wise error and as such ReconFusion and CAT3D score higher in most cases, since these metrics favor blurrier results. Our LPIPS scores, however, demonstrate that our approach is better at preserving texture details. The numbers for approaches marked by * are directly grabbed from ReconFusion. CAT3D results are obtained from the original paper. The methods marked by † are initialized using DUSt3R to improve their results. + +
PSNR ↑SSIM ↑LPIPS ↓
w/o depth enhancement15.610.3360.513
single stage14.920.3090.527
w/o ft inpainting15.640.3410.508
RI3D (Ours)15.740.3420.505
+ +Table 2. We evaluate the effect of different components of our method on the 3-input Mip-NeRF 360 [3] dataset. + +optimization is performed in a single stage using only the repair model. Note that in this case, the repair model is responsible for both enhancing the rendered images and hallucinating details in the missing areas. As seen, the single-stage approach produces floating artifacts due to insufficient constraints to place the inpainted content in the correct 3D location. Additionally, since the repair model is not specifically trained for inpainting, it sometimes hallucinates low-quality textures, as indicated by the blue arrow. + +Finally, to assess the impact of personalizing the inpainting diffusion model, we compare our approach against a variant that directly uses our base inpainting model without fine-tuning (w/o ft inpainting). Although the baseline inpainting model generates visually pleasing textures, they sometimes do not perfectly match the surrounding textures. For example, the synthesized textures include a grayish plant and leaf-like texture on the ground (red arrows). + +# 6. Limitations + +Our dense initialization, and consequently, the final reconstruction quality, depend on the DUSt3R depth. When DUSt3R produces a highly inaccurate depth map, our op + +timization cannot fully correct the misaligned initialization, resulting in ghosting artifacts (see supplementary). However, the modular nature of our approach allows us to replace DUSt3R with potentially more accurate models in the future as better methods become available, mitigating this issue. Moreover, unlike approaches like CAT3D [8], our method is not able to handle a single input image, mainly because of the use of leave-one-out strategy for training our repair model. In the future, it would be interesting to avoid this issue by training the repair model on a large scale training data in an offline manner. + +# 7. Conclusion + +In this paper, we have presented a novel technique based on 3DGS that utilizes diffusion models to reconstruct novel views from a sparse set of input images. Specifically, we initialize 3D Gaussians by proposing a strategy to obtain per-view depth maps by combining multiview stereo depth with monocular depth. We then perform 3DGS optimization using two personalized diffusion models through two stages. We demonstrate that our approach produces better results than the state of the art on challenging scenes. + +# Acknowledgements + +The project was funded in part by a generous gift from Meta. Portions of this research were conducted with the advanced computing resources provided by Texas A&M High Performance Research Computing. + +# 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] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 7, 8 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2, 8 +[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2 +[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2, 3 +[7] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9935-9946, 2023. 3 +[8] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 3, 7, 8 +[9] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14346-14355, 2021. 2 +[10] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. Advances in Neural Information Processing Systems, 35:14715-14728, 2022. 3 +[11] Alper Güngör, Salman UH Dar, Şaban Öztürk, Yilmaz Korkmaz, Hasan A Bedel, Gokberk Elmas, Muzaffer Ozbey, and Tolga Çukur. Adaptive diffusion priors for accelerated mri reconstruction. Medical image analysis, 88:102872, 2023. 3 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[13] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. + +In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2 +[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2, 3 +[15] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[16] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 1, 3, 7, 8 +[17] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffir: Toward blind image restoration with generative diffusion prior. In European Conference on Computer Vision, pages 430-448. Springer, 2024. 3 +[18] 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, 2 +[19] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2 +[20] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 2, 3 +[21] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. Coherentgs: Sparse novel view synthesis with coherent 3d gaussians. arXiv preprint arXiv:2403.19495, 2024. 2, 3, 5 +[22] Patrick Pérez, Michel Gangnet, and Andrew Blake. Poisson image editing. In ACM SIGGRAPH 2003 Papers, page 313-318, New York, NY, USA, 2003. Association for Computing Machinery. 2, 5 +[23] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14335-14345, 2021. 2 +[24] 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. 7 +[25] 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 +[26] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2 +[27] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 3, 8 +[28] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 3 +[29] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5 +[30] Nagabhushan Somraj, Adithyan Karanayil, and Rajiv Soundararajan. Simplenerf: Regularizing sparse input neural radiance fields with simpler solutions. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2, 3, 8 +[31] Luming Tang, Nataniel Ruiz, Qinghao Chu, Yuanzhen Li, Aleksander Holynski, David E. Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, and Michael Rubinstein. Realfill: Reference-driven generation for authentic image completion. ACM Trans. Graph., 43(4), 2024. 5 +[32] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1 +[33] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 2 +[34] 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, 132(12):5929-5949, 2024. 3 +[35] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4150–4159, 2023. 2 +[36] Qianqian Wang, Zhengqi Li, David Salesin, Noah Snavely, Brian Curless, and Janne Kontkanen. 3d moments from near-duplicate photos. In CVPR, 2022. 5, 6 +[37] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 4 +[38] 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. arXiv preprint arXiv:2312.02981, 2023. 2, 3, 7, 8 + +[39] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4180-4189, 2023. 8 +[40] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 $\{\backslash \mathrm{deg}\}$ sparse view synthesis using gaussian splatt-ing. arXiv preprint arXiv:2312.00206, 2023. 2, 3 +[41] Chen Yang, Sikuang Li, Jiemin Fang, Ruofan Liang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Gaussianobject: Just taking four images to get a high-quality 3d object with gaussian splatting. arXiv preprint arXiv:2402.10259, 2024. 2, 3, 5, 6 +[42] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 2, 3, 8 +[43] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 4 +[44] 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 +[45] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: sparse-view 3d gaussian splatting via co-regularization. In European Conference on Computer Vision, pages 335–352. Springer, 2025. 7, 8 +[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. 3, 5 +[47] 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 (CVPR), 2018. 7 +[48] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00451, 2023. 1, 2, 3, 6, 7, 8 \ No newline at end of file diff --git a/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/images.zip b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5bdc7e00419968125dbd408e16a7f05684cc98de --- /dev/null +++ b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb489a36264c5a1a75e49acd39d3da2a98fc19e9abf9e7585e579aab9d055faa +size 647039 diff --git a/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/layout.json b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..940f99d2284bd47d968f70b38eb54fcfee87b823 --- /dev/null +++ b/ICCV/2025/RI3D_ Few-Shot Gaussian Splatting With Repair and Inpainting Diffusion Priors/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e43c78efc7a56c976543226344eb56de7b8eb4dc0feacd2164e291fa95787c08 +size 447587 diff --git a/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/3e6e9afd-08c2-45b5-bccd-cd0e0a6e9794_content_list.json b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/3e6e9afd-08c2-45b5-bccd-cd0e0a6e9794_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9e0b028c738800071f944543d0005f92887508c0 --- /dev/null +++ b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/3e6e9afd-08c2-45b5-bccd-cd0e0a6e9794_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5be5f4c3cba37a7da94fdc4f698979486a37cfd4943e66144fbe31a31644d40a +size 90554 diff --git a/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/3e6e9afd-08c2-45b5-bccd-cd0e0a6e9794_model.json b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/3e6e9afd-08c2-45b5-bccd-cd0e0a6e9794_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b124517eca8f0985eaaad17ca19663a85a5a00cf --- /dev/null +++ b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/3e6e9afd-08c2-45b5-bccd-cd0e0a6e9794_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2df651f2628fbceaed288b06c50a38dd7962b73defb2c2a6e55312f537b68144 +size 110528 diff --git a/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/3e6e9afd-08c2-45b5-bccd-cd0e0a6e9794_origin.pdf b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/3e6e9afd-08c2-45b5-bccd-cd0e0a6e9794_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b8af034e70cba4b91c9d8876a8f0314dbd6df0f1 --- /dev/null +++ b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/3e6e9afd-08c2-45b5-bccd-cd0e0a6e9794_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6945d94c4fe59ab97056e5c07b4d62d76729f8ea478bfc2192f183de26f1ae5a +size 2244697 diff --git a/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/full.md b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..38bb331ce3660f7711a0d51408320f95bfae6b4f --- /dev/null +++ b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/full.md @@ -0,0 +1,362 @@ +# RIOcc: Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction + +Baojie Fan $^{1*}$ Xiaotian Li $^{1}$ Yuhan Zhou $^{1}$ Yuyu Jiang $^{1}$ Jiandong Tian $^{2}$ Huijie Fan $^{2}$ $^{1}$ Nanjing University of Posts and Telecommunications + $^{2}$ Shenyang Institute of Automation, Chinese Academy of Sciences + +{jobfbj, xiaotianli981, yuhanzhou521, jyy01731}@gmail.com {tianjd, fanhuijie}@sia.cn + +# Abstract + +The multi-modal 3D semantic occupancy task provides a comprehensive understanding of the scene and has received considerable attention in the field of autonomous driving. However, existing methods mainly focus on processing large-scale voxels, which bring high computational costs and degrade details. Additionally, they struggle to accurately capture occluded targets and distant information. In this paper, we propose a novel LiDAR-Camera 3D semantic occupancy prediction framework called RIOcc, with collaborative feature refinement and multi-scale cross-modal fusion transformer. Specifically, RIOcc encodes multi-modal data into a unified Bird's Eye View (BEV) space, which reduces computational complexity and enhances the efficiency of feature alignment. Then, multi-scale feature processing substantially expands the receptive fields. Meanwhile, in the LiDAR branch, we design the Dual-branch Pooling (DBP) to adaptively enhance geometric features across both the Channel and Grid dimensions. In the camera branch, the Wavelet and Semantic Encoders are developed to extract high-level semantic features with abundant edge and structural information. Finally, to facilitate effective cross-modal complementarity, we develop the Deformable Dual-Attention (DDA) module. Extensive experiments demonstrate that RIOcc achieves state-of-the-art performance, with $54.2\mathrm{mIoU}$ and $25.9\mathrm{mIoU}$ on the Occ3D-nuScenes and nuScenes-Occupancy datasets, respectively. + +# 1. Introduction + +3D semantic occupancy prediction is a complex and vital task that aims to jointly estimate the geometric structure and semantic categories of voxels within a scene. Compared to traditional tasks [13, 25, 37, 42] like 3D object detection, 3D semantic occupancy prediction focuses more on understanding the overall scene, which includes not only objects + +![](images/4d8b857bfecdac5397d55449c8735c3ee5b71389e52288987509ae0e6f4efb72.jpg) + +![](images/81093d99d7a282c87c15ac1d2b6dbed8e1884cedaaed72c6b6cfc9e26da08559.jpg) +Figure 1. Comparison between OpenOccupancy and the proposed RIOcc. Instead of processing voxel features like OpenOccupancy, we choose BEV features to achieve higher computational efficiency. Additionally, we extracted refined multi-modal BEV features and fed them into the efficient fusion module (DDA). The fused BEV features are then transformed into 3D features for occupancy prediction. "VT" is view transform. + +but also details in the background and environment. Consequently, it provides richer and more reliable environmental perception, which is essential for applications such as high-precision map construction and path planning in autonomous driving [5, 50]. + +In contrast to traditional methods that rely solely on a single modality [19, 22, 23, 51], multi-modal fusion [1, 6, 15, 41, 59] provides higher accuracy and robustness for 3D perception. In various 3D perception tasks, effectively combining data from cameras and LiDAR presents a crucial challenge for achieving high-precision predictions. Cameras provide rich semantic and texture information, while LiDAR captures accurate geometric and depth information. The complementary of the two has driven research on fusion strategies to improve the performance of 3D semantic + +occupancy prediction. Nevertheless, due to the heterogeneity of camera and LiDAR data, along with the complexity of feature integration, interactions between different modalities often lead to inconsistencies in occupancy prediction. It can degrade semantic information and introduce significant uncertainty into the predictions. + +However, the task of semantic occupancy prediction [2, 9, 10, 12, 39, 40, 49] also faces significant computational challenges, especially when it involves real-time processing of large-scale voxel data, which demands high computational resources. To manage this computational burden, previous single-modal approaches [7, 17, 18, 21, 24, 55, 57] processed data from a Bird's Eye View (BEV) perspective and achieved considerable results. While multi-modal occupancy methods can significantly improve performance compared to previous approaches, they also increase computational complexity. To find a better balance between computational efficiency and perception accuracy in multimodal methods, reasonable and effective design is essential. Additionally, data imbalance in complex environments hinders models from learning rare but important spatial features, which is critical issue that needs to be addressed. + +To address the aforementioned issues, we propose RI-Occ, a novel multi-modal 3D semantic occupancy prediction method. The comparison with the pioneering work OpenOccupancy is shown in Figure 1. RIOcc avoids reliance on 3D convolutions, enabling the extraction of finer features from the BEV perspective and enhancing information complementarity between different modalities. Specifically, we first extract features from LiDAR and camera data and project them into a unified BEV space. To reduce the impact of redundant features on overall performance while preserving rich geometric depth information, we design Dual-branch Pooling (DBP). In addition, the Wavelet Encoder can effectively capture important edge and structural features in images. It compresses data in the channel dimension, which helps reduce computational complexity. The Semantic Encoder is employed to extract high-level features with rich semantic information from the input data. These features are then combined with geometric features in multi-modal fusion, enhancing the comprehensive understanding of the scene. + +In the feature fusion stage, RIOcc employs Deformable Dual-Attention (DDA) to facilitate interaction and dynamic adjustment of features from LiDAR and camera. This approach significantly bridges the differences between modalities and enhances information exchange and prediction consistency. Through the DDA module, our model can leverage data from multiple modalities to achieve a comprehensive understanding of both the geometric and semantic aspects of the scene. Finally, the integrated features are fed into the occupancy prediction module. Similar to FlashOcc, we convert the fused BEV features from the channel dimen + +sion to the height dimension. This operation provides semantic labels for each voxel and generates a 3D scene representation rich in semantic information. + +Our contributions are summarized as follows: + +- We propose a novel multi-modal 3D semantic occupancy prediction framework, RIOcc. The LiDAR and camera branches respectively extract refined structural information and semantic features, with balanced computational load and performance. +- The proposed Deformable Dual-Attention facilitates multi-modal feature fusion, bridging the gaps caused by modality differences, which enhances semantic consistency and reduces modality heterogeneity. +- We conduct extensive experiments on the nuScenes-Occapacity and nuScenes-Occ3d dataset, which demonstrate the effectiveness of our method and achieve state-of-the-art performance. + +# 2. Related Work + +BEV-based 3D Perception. In recent years, BEV perception has become crucial for autonomous driving due to its global view and efficiency. With declining camera costs, vision-based 3D object detection has rapidly advanced. A key work is LSS [34], which uses geometric methods to create depth maps. This approach was first integrated into a full detection pipeline in BEVDet [13]. Later, BEVDepth [22] added explicit depth supervision, and Far3D [16] introduced sparse queries and adaptive 3D query generation. However, single-modal methods still struggle with accurate depth and spatial understanding, leading to a growing focus on multi-modal BEV perception. BEVFusion [27] was a pioneering multi-modal approach, projecting different features into BEV for unified fusion. MaskBEV [59] further added task-agnostic masks to the BEV space. While effective, these methods still face challenges in detecting nontarget objects. + +Camera-based 3D Semantic Occupancy Prediction. To overcome the intrinsic limitations of traditional object detection approaches, 3D semantic occupancy prediction has gained prominence. MonoScene [5] leverages optically inspired 2D-3D feature projection to infer dense geometric and semantic information. TPVFormer [14] employs a three-view representation combined with Transformers to project 2D image features into 3D space for prediction. Recently, OccupancyM3D [33] directly learns occupancy states in 3D space, and integrates occupancy information into the feature representation of monocular 3D detection. PanoOcc [48] employs voxel queries and a coarse-to-fine approach, effectively processing semantic information for all objects in complex scenes. However, due to inaccuracies in depth estimation and insufficient geometric information in pure visual occupancy, incorporating multi-modal perception presents a more reliable solution. + +![](images/945946babf32bb9a3accc1d1785632261c233543c1677d3c8e318d9470fe7bba.jpg) +Figure 2. The overall framework of RIOcc. This framework includes three main branches: LiDAR, Camera, and Interaction Branch. The LiDAR Branch processes LiDAR points through an encoder and DBP modules to produce multi-scale BEV features. The Camera Branch encodes multi-view images, refining the output with Wavelet and Semantic Encoders to capture spatial and semantic details. In the Interaction Branch, Deformable Dual-Attention integrates features from both branches. These multi-scale features are then adaptively fused and passed to an Occupancy Head for the final occupancy prediction. + +Multi-modal 3D Semantic Occupancy Prediction. To further improve the precision of 3D semantic occupancy prediction, researchers have increasingly focused on exploring multi-modal perception approaches. OpenOccupancy [47] introduces the first benchmark model for LiDAR-camera semantic occupancy. Building upon this foundation, subsequent work OccGen [46] proposes a "noise-to-occupancy" generative paradigm, which iteratively removes random Gaussian noise to generate and refine 3D semantic occupancy maps. The Co-Occ [31] method presents a multi-modal 3D semantic occupancy prediction framework that combines explicit features fusion with volume rendering regularization. OccFusion [30] improves the utilization of information from different sensors through a dynamic fusion module. OccMamba [20] introduces the Mamba architecture and a 3D-to-1D reordering strategy to enhance the efficiency of processing large-scale 3D voxels. Despite significant advancements in multimodal perception, challenges like limited interaction, high computational load, and coarse feature representations persist, requiring further optimization. + +# 3. Method + +# 3.1. Overall Architecture + +The overall architecture of RIOcc is illustrated in Figure 2. The framework takes images and LiDAR point clouds as inputs, extracting consistent BEV features for subsequent fusion (Sec 3.2). The designed Dual-branch Pooling (DBP) + +removes redundant features while retaining geometric depth information from the LiDAR branch (Sec 3.3). Simultaneously, to extract more effective camera BEV features, a Wavelet Encoder is employed to suppress noise and enhance hierarchical information. The Semantic Encoder is used to enrich semantic information, improving the understanding of the scene (Sec 3.4). Then, we design the Deformable Dual-Attention (DDA) to strengthen the interaction of BEV features at different scales between modalities (Sec 3.5). Finally, the fused multi-scale features are input into the occupancy prediction module (Sec 3.6). + +# 3.2. Features Extraction + +During the feature extraction stage, we design LiDAR and camera branches to encode multi-modal input, following the BEVFusion [25] setup. The LiDAR branch voxelizes point clouds and uses a 3D sparse convolutional network to generate BEV features $F_{L}^{BEV}$ . The camera branch extracts multi-scale image features using ResNet50 and maps them to the BEV space using a view transformation, resulting in feature $F_{C}^{BEV}$ . Finally, the features from the two branches are fused in the BEV perspective, providing basic feature representations for subsequent occupancy prediction. + +# 3.3. Dual-branch Pooling + +To enhance the feature representation of the LiDAR branch, we design Dual-branch Pooling (DBP), which primarily focuses on adaptively refining important geometric features and removing redundant features. Firstly, in order to en + +![](images/35528b80bb497c4d5a0d22d93f5515b871a71229455042ca5a583df71c97d3fc.jpg) +Figure 3. The schema of Dual-branch Pooling (DBP). LiDAR feature representation is improved by adaptively highlighting important semantic channels and significant geometric regions. + +hence the ability to capture long-range semantics and multiscale spatial information of global BEV features, the input feature $F_{L}^{\prime}$ passes through the Windowed Attention module followed by the Bottleneck ASPP module. Then, the features are passed through the Channel-wise Attention and Grid-wise Attention modules, optimizing information representation across different dimensions. The detailed structures of DBP are shown in Figure 3. + +To adaptively highlight channels that are crucial for semantic information, the channel-wised attention utilizes max pooling and average pooling to aggregate the spatial dimensions of the input BEV features $F_{L}^{\prime}$ , generating two descriptors $F_{Avg}$ and $F_{Max}$ . They are processed through a shared MLP to produce a channel attention weight map, which is then activated by a sigmoid function to emphasize the important channel information within the features. The output from the Channel-wise Attention are given by: + +$$ +F _ {\text {c h a n n e l}} = \sigma \left(M L P \left(F _ {\text {A v g}}\right) + M L P \left(F _ {\text {M a x}}\right)\right) \tag {1} +$$ + +To focus on regions that are particularly important for the overall geometric description, the spatial attention module also utilizes max pooling and average pooling operations to the input BEV features to compress them along the channel dimension to obtain $F_{Avg}^{\prime}$ and $F_{Max}^{\prime}$ . Subsequently, these compressed features are processed through convolutional layers and ReLU to generate a spatial attention weight map, which identifies positions that are significant in terms of geometric structure. The features outputted from the Grid-wise Attention can be represented as: + +$$ +F _ {g r i d} = \sigma \left(R e L U \left(f ^ {7 \times 7} \left(F _ {A v g} ^ {\prime}; F _ {M a x} ^ {\prime}\right)\right) \right. \tag {2} +$$ + +The final output $F_{DBP}$ can be represented as: + +$$ +F _ {D B P} = \left(F _ {\text {c h a n n e l}} \odot F _ {L} ^ {\prime}\right) \otimes \left(F _ {\text {g r i d}} \odot F _ {L} ^ {\prime}\right) \tag {3} +$$ + +where $\sigma$ denotes the sigmoid activation operation, and $f^{7\times 7}$ represents a convolution with a kernel size of $7\times 7$ . + +# 3.4. Efficiency Camera BEV Features + +# 3.4.1. Wavelet Encoder + +Due to differences in viewpoints and hardware characteristics among the different cameras, various noises and inconsistencies are often introduced. To reduce redundancy, + +![](images/335cc14cbca3c781bc7e1fed3a1e4df6ee0f94757cc5a259cef744a71323cc06.jpg) +Figure 4. Detailed structure diagram of the wavelet encoder. The input BEV features undergo DWT and IWT to obtain richer structure and details. + +noise impact, and decrease computational burden, we design the Wavelet Encoder, as shown in Figure 4. This encoder retains the primary structure and details of the images by decomposing features, which enhances the efficiency and accuracy of the fusion stage. Specifically, the Wavelet Encoder first applies a wavelet transform to the input camera BEV features, decomposing them into low-frequency and high-frequency components. The low-frequency components typically contain the main structural information and global features of the images, while the high-frequency components capture details such as edges and textures. This approach allows for a more refined multi-scale decomposition of features, leading to rich semantic and geometric details. We apply Discrete Wavelet Transform (DWT) to the input features $F_{C}^{BEV}$ to obtain low-frequency features $F_{low}$ and high-frequency features $F_{high}$ : + +$$ +F _ {C} ^ {B E V} \xrightarrow {D W T} \left\{F _ {\text {l o w}}, F _ {\text {h i g h}} \right\} \tag {4} +$$ + +Next, the low-frequency features $F_{low}$ are retained for subsequent feature fusion, while the high-frequency features $F_{high}$ undergo further encoding through a series of convolutional layers and nonlinear activation to extract high-contrast information and to eliminate noise. Finally, we concatenate the processed high-frequency features with the low-frequency features and then passed through the Inverse Discrete Wavelet Transform (IWT) to form a multiscale feature representation $F_{wavelet}$ . This process can be roughly expressed as: + +$$ +F _ {\text {w a v e l e t}} \xrightarrow {I W T} \operatorname {C o n c a t} \left(F _ {\text {l o w}}, \operatorname {C o n v} \left(F _ {\text {h i g h}}\right)\right) \oplus F _ {H} \tag {5} +$$ + +# 3.4.2. Semantic Encoder + +To enhance the semantic expressiveness of the BEV features, we propose a lightweight 2D Semantic Encoder for efficiently extracting rich semantic information. This encoder employs a lightweight 2D U-Net [36] architecture, utilizing multi-scale feature downsampling and upsampling to extract and fuse both global and local information. The Semantic Encoder first downsamples the input BEV features to capture global contextual information. Subsequently, it upsamples to restore the spatial dimensions of the + +![](images/f5d5ccd48bb99385396c5fb3cf77c072823adafc791561d0bec53c8a5e2926ad.jpg) +Figure 5. Overview of Deformable Dual-Attention (DDA), which reduces the disparity between LiDAR and Camera BEV features and enhances scene understanding. + +features and utilizes residual connections to fuse features of different scales. Additionally, we introduce an Auxiliary Semantic Loss at the output stage to enhance the semantic consistency of the features and improve the model's understanding of complex scenes. + +# 3.5. Deformable Dual-Attention + +Due to the inherent sparsity of point clouds, certain regions in the LiDAR BEV representation lack sufficient geometric details, particularly at longer distances. Conversely, although image BEV features provide extensive semantic and texture information, they exhibit limitations in depth estimation accuracy. To enhance distant perception and aggregate local information, we propose a multi-modal fusion transformer that combines the geometric depth information from LiDAR with the semantic information from cameras, which enhance the perception of distant objects for more accurate 3D scene understanding. + +As shown in Figure 5, the LiDAR BEV features $F_{LiDAR} \in \mathbb{R}^{H \times W \times C}$ and Camera BEV features $F_{Camera} \in \mathbb{R}^{H \times W \times C}$ are used as inputs. For the LiDAR stream, we specially compute the distance $d_r$ from each LiDAR point to the sensor and encode it using a Gaussian function to generate a distance vector $e_d$ : + +$$ +e _ {d} = \text {G a u s s i a n E n c o d i n g} \left(d _ {r}\right) \tag {6} +$$ + +Next, we apply a linear transformation using the learned weight vector $w_{linear}$ and the bias $b$ obtained through backpropagation to compute a distance-weighted value $w_{i,j}$ , + +where $(i,j)$ corresponds to the index of the BEV grid. + +$$ +w _ {i, j} = e _ {d} \cdot w _ {\text {l i n e a r}} + b \tag {7} +$$ + +$$ +F _ {L i D A R} ^ {\text {e n h a n c e d}} (i, j) = w _ {i, j} \cdot F _ {L i D A R} (i, j) \tag {8} +$$ + +It allows us to enhance the LiDAR BEV features, compensating for distant features. To enhance the correlation of adjacent spatial features, the features from the $k$ local regions $region_{k}$ within $F_{LiDAR}^{enhanced}$ are aggregated and to extract a single feature vector. Finally, we extract a set of representative tokens from the enhanced features $T_{LiDAR}^{(k)}$ . The specific formula is as follows: + +$$ +t _ {L i D A R} ^ {(k)} = \text {A g g r e g a t e} \left(F _ {L i D A R} ^ {\text {e n h a n c e d}}, \text {r e g i o n} _ {k}\right) \tag {9} +$$ + +$$ +T _ {L i D A R} = \left\{t _ {L i D A R} ^ {1}, t _ {L i D A R} ^ {2}, \dots , t _ {L i D A R} ^ {N} \right\} \tag {10} +$$ + +where Aggregate combines multiple features from local regions into a single feature vector. Subsequently, these enhanced tokens are processed through a linear layer to obtain $Q_{l}$ , $K_{l}$ , and $V_{l}$ for the attention mechanism. + +For the Camera stream, the input features are tokenized and undergo self-attention, generating updated $Q_{c}$ , $K_{c}$ and $V_{c}$ , which allow for a comprehensive understanding of different regions. Then, $Q_{L}$ interacts with $K_{c}$ , and $V_{c}$ through a deformable cross-modal attention, and vice versa. The deformable cross-attention dynamically captures intermodal relationships by flexibly selecting feature from the target modality. It effectively focuses on the complementary information between the LiDAR BEV features and the dense camera semantic features. After cross-modal interaction, the LiDAR and Camera BEV features are each combined with their respective linear weights and then concatenated. The DDA module effectively enhances feature interaction between the LiDAR and camera modalities, providing abundant global features for subsequent predictions. + +# 3.6. Occupancy Prediction Module + +In our framework, the BEV features obtain from the multiscale fusion stage are input into the occupancy prediction module. These fused features contain rich semantic information and fine-grained geometric structures, which allow for a more comprehensive description of the scene information. Similar to FlashOcc [53], the fused BEV features are processed through the occupancy prediction head before being passed to the Channel-to-Height module. This module rearranges the features from the shape $C \times W \times H$ to $C^* \times Z \times W \times H$ , where $C, C^*, W, H, Z$ represent the number of channels, the number of categories, and the dimensions in the three-dimensional space $(x, y, z)$ . This transformation enables intuitive semantic classification and occupancy prediction of the features in 3D space, significantly enhancing the model's ability to express occupancy situations in the scene. + +
MethodModalityResolutionImage BackbonemIoUothersbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmannmadevegetation
LangOcc [3]C256×704R5011.840.003.1090.006.3014.200.4010.806.209.003.8010.7043.702.239.5026.4019.6026.40
FB-Occ [24]C256×704R5023.120.0437.1516.8134.1738.2213.4116.9719.6918.9411.6521.9455.9426.9829.6526.9210.2414.33
UniVision* [11]C256×704R5037.5011.0044.7023.1043.0050.5021.6024.9026.9025.7030.7035.8079.8041.4049.1053.8040.3034.70
GEOcc* [43]C256×704R5043.6414.2951.2731.1146.1355.0929.1230.4630.9935.4735.2041.8284.0047.0055.5259.5050.0344.82
TEOcc* [26]C&R256×704R5042.9010.8250.3324.2848.9957.3229.3824.4130.1428.4636.4643.0183.9643.0956.0059.3454.1849.16
EFFOcc* [38]C&L256×704R1849.2910.5756.1621.7358.6863.1631.9837.7155.4036.1545.8750.8181.0239.0753.0857.1570.4168.90
OccFormer [58]C900×1600R10121.935.9430.2912.3234.4019.1714.4416.4517.229.2713.9026.3450.9930.9634.6622.736.766.97
RenderOcc [32]C900×1600R10126.114.8431.7210.7227.6736.4513.8718.2017.6717.8421.1923.2563.2036.4246.2144.2619.5820.72
TPVFormer [14]C900×1600R10128.346.6739.2014.2441.5446.9819.2122.6417.8714.5430.2035.5156.1833.6535.6931.6119.9716.12
CTF-Occ [45]C928×1600R101-DCN28.538.0939.3320.5638.2942.2416.9324.5222.7221.0522.9831.1153.3333.8437.9833.2320.7918.00
PanoOcc* [48]C640×1600R101-DCN42.1311.6750.4829.6449.4455.5223.2933.2630.5530.9934.4342.5783.3144.2354.4056.0445.9440.40
OctreeOcc* [28]C900×1600R101-DCN44.0211.9651.7029.9353.5256.7730.8333.1730.6529.9937.7643.8783.1744.5255.4558.8649.5246.33
OccFusion* [30]C&L900×1600R10146.7911.6547.8132.0757.2757.5131.8040.1147.3533.7445.8150.3578.7937.1744.3653.3661.1863.20
OccFusion* [56]C&L900×1600R10148.7412.3551.7733.0154.5657.6533.9943.0348.3535.5441.2248.5583.0044.6557.1360.0162.4661.25
RadOcc* [54]C&L512×1408Swin-B49.3810.9358.2325.0157.8962.8534.0433.4550.0732.0548.8752.1182.9042.7355.2758.3468.6466.01
RIOcc* (Ours)C&L256×704R5054.2111.8259.7336.9862.2168.7236.4547.4548.2544.2049.9254.3985.1044.6059.6761.7770.5169.56
+ +Table 1. 3D Occupancy prediction performance on the Occ3D-nuScenes dataset. * means the performance using the camera mask during training. C, L, and R represent camera, LiDAR, and radar, respectively. + +
MethodModalityResolutionImage BackboneLiDAR BackboneIoUmIoUbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmannmadevegetation
MonoScene [5]C900×1600R101-DCN-18.46.97.13.99.37.25.63.05.94.44.94.214.96.37.97.410.07.6
C-CONet [47]C900×1600R50-20.112.813.28.115.417.26.311.210.08.34.712.131.418.818.716.34.88.2
SparseOcc [44]C704×256R50-21.814.116.19.315.118.67.39.411.29.47.213.031.821.720.718.86.110.6
LMSCNet [35]L--VoxelNet27.311.512.44.212.812.16.24.76.26.38.87.224.212.316.614.113.922.2
JS3C-Net [50]L--VoxelNet30.212.514.23.413.612.07.24.37.36.89.29.127.915.314.916.214.024.9
L-CONet [47]L--VoxelNet30.915.817.55.213.318.17.85.49.65.613.213.634.921.522.421.719.223.5
PointOcc [60]L--VoxelNet34.123.924.919.020.925.713.425.630.617.916.721.236.525.625.724.924.829.0
M-CONet [47]C&L900×1600R50VoxelNet29.520.123.313.321.224.315.315.918.013.315.320.733.221.022.521.519.623.2
Co-Occ [31]C&L900×1600R101VoxelNet30.621.926.516.822.327.010.120.920.714.516.421.636.923.525.523.720.523.5
OccGen [46]C&L896×1600R50VoxelNet30.322.024.916.422.526.114.020.121.614.617.421.935.824.524.724.020.523.5
OccFusion [56]C&L900×1600R101VoxelNet32.422.425.317.022.525.916.522.424.016.116.022.135.622.124.023.921.324.0
EFFOcc [38]C&L256×704R18VoxelNet30.822.928.116.722.127.313.024.836.222.616.821.629.413.918.220.626.528.8
OccMamba [20]C&L900×1600R50VoxelNet33.725.129.620.225.728.516.725.023.219.920.324.536.125.325.124.827.728.9
RIOcc (Ours)C&L256×704R50VoxelNet35.425.930.219.825.828.718.324.831.821.820.524.937.224.525.524.927.028.8
+ +Table 2. 3D Occupancy prediction performance on nuScenes-Occupancy validation set. C represents camera and L represents LiDAR. + +# 3.7. Loss + +To effectively train the proposed semantic occupancy prediction framework, we combine various types of losses to comprehensively optimize network performance. The cross-entropy loss $L_{ce}$ and Lovasz-Softmax loss $L_{ls}$ are used to optimize the overall framework. Affinity loss $L_{geo}$ and $L_{sem}$ are applied to optimize scene-wise and class-wise metrics, while $L_{d}$ provides feedback for the depth-aware view transform module. Additionally, we introduce an Auxiliary Semantic Loss $L_{aux}$ to optimize the refined semantic features extracted by the semantic encoder. Therefore, the overall loss function can be expressed as: + +$$ +L _ {\text {t o t a l}} = L _ {c e} + L _ {l s} + L _ {g e o} + L _ {s e m} + L _ {d} + L _ {a u x} \tag {11} +$$ + +# 4. Experiment + +# 4.1. Dataset and Metrics + +Dataset. Similar to previous works [29, 31, 47, 52], we conducted extensive experiments based on the nuScenes dataset. The dataset is divided into nuScenes-occupancy [4] and Occ3D-nuScenes [45] according to the source of the annotated data. Both datasets inherit the data format of nuScenes, containing 700 training scenes and 150 validation scenes, with annotations for 17 categories. The evaluation range for OpenOccupancy is $[-51.2\mathrm{m}, 51.2\mathrm{m}]$ in the X and Y directions, and $[-3\mathrm{m}, 5\mathrm{m}]$ in the Z direction, using a voxel resolution of 0.2 meters. In comparison, the data coverage for Occ3D-nuScenes is $[-40\mathrm{m}, 40\mathrm{m}]$ in the X and Y directions, and $[-1\mathrm{m}, 5.4\mathrm{m}]$ in the Z direction, with + +![](images/61c3c39d9706005e0e7284bb698c935c29caa811b8daa3034d9bde8c3a904834.jpg) + +![](images/18932a4938acd263d201cb5e30ddcd639fc4c30877095df1f025b2e7c31945d7.jpg) + +![](images/8c592e40e08a33d59c29aaa96017264006b05487fb6e69fe6b6d7707ed876e4e.jpg) + +![](images/d73347f15d72e7151111704fd5589af400903d49d4ab9340e70469789db1315b.jpg) +Images + +![](images/bb7823fac23d0390b3cd81a42966a633a0403bc527ec7da6dd648784f4b4a8c9.jpg) + +![](images/37abce9a7c9ad066dc205f2e3d8e40337d27696f325b468082d57ef0718b1521.jpg) + +![](images/6c6a255411dec0cd96f2847760ce0a8bc2e66e893de7516958d07314878aa6de.jpg) +LiDAR + +![](images/30922fbf06bce7e142ae4e8ae48a2ac07c9f92a5339fa209e444ce522f118c3f.jpg) + +![](images/71f37404986fbfd09d5b3cf5389ddd23126cb531ef49600217f5642e281c9f1c.jpg) + +![](images/edd2cec75b96172681c9f2e7b0bb9340c1e41880ee9c77c08aeb622e78c1dea4.jpg) +M-CONet +Figure 6. The additional qualitative comparisons results between RIOcc and M-CONet. The red box highlights the effectiveness in dealing with distant and occluded targets. + +![](images/c6776fa380ef9d3853c54ac066ca74e0d5b7c17ec8cd53d1ae68d68e0332365b.jpg) + +![](images/12a269afeae5770a848fafabe552f4a1f71a174eced979e10957fefe42dc50d0.jpg) + +![](images/6e8a76fe9d42ee469e15cb3f7386862706a697622b6f05446a5f586e75d3cedd.jpg) +RIOcc + +![](images/4b5d1fb22a751323cd055a9852531e352524128215488f710cb6443037fb0e17.jpg) + +![](images/96719c7610adc1b888fb9d219fbaaf460dba668bd77a027f4806bdaa2bc57fc4.jpg) + +![](images/3d3bca8e2cb876763a75c85a63938c9b909306b02dabefde8c1bce096d29239d.jpg) +GT + +discretization using voxels of size $[0.4\mathrm{m}, 0.4\mathrm{m}, 0.4\mathrm{m}]$ . + +Metrics. We adopt the official evaluation metrics, including IoU and mIoU. + +# 4.2. Implementation Details + +Our method is implemented based on MMDetection3D [8]. For the camera branch, we use ResNet50 pretrained on ImageNet as the image backbone, and the input image size is cropped to $256 \times 704$ . For the LiDAR branch, we voxelize 10 LiDAR sweeps and employ a voxel encoder for the nuScenes dataset. During training, we use the AdamW optimizer, set the weight decay to 0.01, and an initial learning rate of $1e^{-4}$ , with a multi-step learning rate scheduler for optimization. Training is conducted on four NVIDIA 3090 GPUs with a batch size of 4, for a total of 24 epochs. + +# 4.3. Comparison with State-of-the-Art Methods + +To ensure fair comparison, all results are either provided by the original authors or reproduced using the official code. We first compare RIOcc with some classical and advanced methods on Occ3D-nuScenes. The results are shown in Table 1, RIOcc achieves the latest state-of-the-art performance with an impressive $54.20\mathrm{mIoU}$ , outperforming RadOcc by $4.82\mathrm{mIoU}$ . To further evaluate the effectiveness of our proposed framework, we conduct additional performance tests on the nuScenes-Occupancy validation set. As shown in Table 2, RIOcc again achieves the best performance, with 35.4 IoU and $25.9\mathrm{mIoU}$ . + +# 4.4. Ablation Studies + +We conduct ablation experiments using the camera mask on the Occ3D-nuScenes dataset to evaluate the contribution of + +each module to the overall framework. + +Effects of Downsampling Layers. We conduct extensive experiments to evaluate the number of downsampling layers. Notably, all experiments are performed without the Wavelet Encoder and Semantic Encoder. As shown in the Table 3, when the number of downsampling layers is set to 0 (no downsampling), the model achieves an mIoU of 48.21, with a memory usage of 5.02 GB. As the number of downsampling layers increases, the mIoU continues to grow by 3.82 at 2 layers, with memory usage increasing slightly to 5.58 GB. However, when further downsampling to 3 layers, the performance improvement tends to plateau. Therefore, we choose to use 2 downsampling layers, which strikes a reasonable balance between computational load and significant performance improvement. + +Influences of Wavelet and Semantic Encoders. To validate the effectiveness of the Wavelet Encoder and Semantic Encoder, we present an ablation study shown in Table 4. The results indicate that adding either the Wavelet Encoder or the Semantic Encoder alone can improve model performance. When both are incorporated, the mIoU increases from the baseline of 51.96 to 54.21. It demonstrates that introducing the Wavelet Encoder and Semantic Encoder effectively enhances the model's ability to capture refined features and understand semantic information. + +Effectiveness of DBP. As illustrated in the Table 5, we explore the contributions of the submodules within the DBP module. It obtains $0.63\mathrm{mIOU}$ with channel-wise alone and $0.71\mathrm{mIOU}$ with grid-wise alone. When both branches are combined to form the complete DBP module, the mIoU increases by 1.14. It indicates that the synergy between the + +![](images/36f299b7a5f81aa29a23e49d7edeab057c2b08a8c4bf74d97a86db0bec42abac.jpg) +Figure 7. Features heatmaps after adaptive fusion. + +![](images/cfe73518eae94d76029cced90ef90af9702a7f89dbdb0f9675c2df7f38634ba7.jpg) + +
Downsampling Layer0123
mIoU48.2151.1252.0352.15
Memory(GB)5.025.325.585.82
+ +Table 3. Ablation study of downsampling layer. + +
#SemanticmIoU
w/ Wavelet✓53.32
54.21
w/o Wavelet✓51.96
53.37
+ +Table 4. Ablation Study of Wavelet and Semantic Encoders. It is worth noting that the default downsampling operation is used in the absence of wavelet or semantic encoding. + +Channel-wise and Grid-wise branches has a greater advantage in capturing multi-level feature information. + +Impacts of Aggregation Region Size. We illustrate the results to demonstrate that different region sizes in the DDA module help identify the optimal local aggregation size for enhancing feature representation and fusion. To ensure consistency in the number of final tokens, we uniformly sampling of the same number of anchors and test 4 different region sizes. $1 \times 1$ means no feature aggregation, with each token corresponding to a single feature point. As indicated in Table 6, when the region size is set to $5 \times 5$ , the model achieves the highest mIoU at 48.21. This setting proves an ability to preserve sufficient specificity while enhancing detection accuracy for large areas and distant objects. In contrast, when the region size is further increased to $7 \times 7$ , detection accuracy slightly decreases. This result may be attributed to an overly large aggregation region, which smoothes the feature maps and blurs local details. + +Comparison of Interaction Methods. In Table 7, we compare the effectiveness of conventional BEV feature fusion methods with our proposed DDA module. The conventional addition and concatenation fusion strategies achieve mIoU of 46.58 and 46.69, respectively. In contrast, the DDA module results in a significant improvement, with the mIoU gaining about 1.57. The DDA module demonstrates a stronger impact on feature fusion, effectively enhancing + +
#Channel WisedGrid WisedmIoU
153.07
2✓53.70
3✓53.78
4✓✓54.21
+ +Table 5. Ablation Study on Channel-Wised and Grid-Wised Operations. To reduce training time, this experiment uses only a single-scale feature for training. + +
region1 x 13 x 35 x 57 x 7
mIoU47.4347.9248.2148.16
+ +Table 6. Ablation Study of Aggregation Region Size. + +
#StrategymIoU
1Addition46.58
2Concatenation46.69
3DDA48.21
+ +Table 7. Ablation study of the Dual-BEV fusion strategy. + +representation and improving scene understanding. + +Feature Alignment on Heatmaps. To demonstrate that our model effectively enhances feature alignment in the LiDAR-camera fusion process, we present the feature heatmaps after adaptive fusion for both M-CONet and our method, as shown in the Figure 7. Note that we only select the first channel of the sample for visualization. + +# 4.5. Visualization + +As shown in Figure 6, we present the visualization of RI-Occ on Occ3D-nuScenes without the camera mask. The results indicate that our RIOcc provides a more comprehensive prediction of the scene. It demonstrates improved detection performance for distant and occluded objects, along with a finer-grained occupancy prediction of the scene. + +# 5. Conclusion + +We propose RIOcc, a novel multi-modal 3D semantic occupancy prediction method, which is equipped with advanced BEV features refinement and interaction mechanisms. RIOcc integrates multi-modal data in a unified BEV space to achieve less computational burden. The LiDAR and camera branches extract refined BEV features independently, while the interaction branch effectively mitigates the disparity between multi-modal BEV features. Finally, the fused BEV features are transformed into voxel representation for occupancy prediction. Extensive experiments on the Occ3D-nuScenes and nuScenes-Occupancy datasets demonstrate the superiority of RIOcc over existing approaches. We believe our model will inspire new insights in the field. + +# Acknowledgements + +# References + +[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1090-1099, 2022. 1 +[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9297-9307, 2019. 2 +[3] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Langocc: Self-supervised open vocabulary occupancy estimation via volume rendering. arXiv preprint arXiv:2407.17310, 2024. 6 +[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6 +[5] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2, 6 +[6] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1 +[7] Xiaowei Chi, Jiaming Liu, Ming Lu, Rongyu Zhang, Zhaoqing Wang, Yandong Guo, and Shanghang Zhang. Bev-san: Accurate bev 3d object detection via slice attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17461-17470, 2023. 2 +[8] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.7 +[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2 +[10] Michael Firman, Oisin Mac Aodha, Simon Julier, and Gabriel J Brostow. Structured prediction of unobserved voxels from a single depth image. In Proceedings of the IEEE + +This work is supported in part by the National Natural Science Foundation of China under Grant 62473205, Natural Science Foundation of Jiangsu Province No. BK20241893, and sponsored by Yong Academic Leader of Qing Lan Project in Jiangsu Province. +Conference on Computer Vision and Pattern Recognition, pages 5431-5440, 2016. 2 +[11] Yu Hong, Qian Liu, Huayuan Cheng, Danjiao Ma, Hang Dai, Yu Wang, Guangzhi Cao, and Yong Ding. Univision: A unified framework for vision-centric 3d perception. arXiv preprint arXiv:2401.06994, 2024. 6 +[12] Binh-Son Hua, Quang-Hieu Pham, Duc Thanh Nguyen, Minh-Khoi Tran, Lap-Fai Yu, and Sai-Kit Yeung. Scenenn: A scene meshes dataset with annotations. In 2016 fourth international conference on 3D vision (3DV), pages 92–101. IEEE, 2016. 2 +[13] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 2 +[14] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 2, 6 +[15] Qi Jiang, Hao Sun, and Xi Zhang. Semanticbefvusion: Rethink lidar-camera fusion in unified bird's-eye view representation for 3d object detection. arXiv preprint arXiv:2212.04675, 2022. 1 +[16] Xiaohui Jiang, Shuai Lin Li, Yingfei Liu, Shihao Wang, Fan Jia, Tiancai Wang, Lijin Han, and Xiangyu Zhang. Far3d: Expanding the horizon for surround-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2561-2569, 2024. 2 +[17] Zheng Jiang, Jinqing Zhang, Yanan Zhang, Qingjie Liu, Zhenghui Hu, Baohui Wang, and Yunhong Wang. Fsd-bev: Foreground self-distillation for multi-view 3d object detection. arXiv preprint arXiv:2407.10135, 2024. 2 +[18] Yang Jiao, Zequn Jie, Shaoxiang Chen, Lechao Cheng, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. Instance-aware multi-camera 3d object detection with structural priors mining and self-boosting learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2598–2606, 2024. 2 +[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 1 +[20] Heng Li, Yuenan Hou, Xiaohan Xing, Xiao Sun, and Yanyong Zhang. Occmamba: Semantic occupancy prediction with state space models. arXiv preprint arXiv:2408.09859, 2024.3,6 +[21] Peidong Li, Wancheng Shen, Qihao Huang, and Dixiao Cui. Dualbev: Cnn is all you need in view transformation. arXiv preprint arXiv:2403.05402, 2024. 2 +[22] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1, 2 +[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: + +Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1 +[24] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 6 +[25] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 1, 3 +[26] Zhiwei Lin, Hongbo Jin, Yongtao Wang, Yufei Wei, and Nan Dong. Teocc: Radar-camera multi-modal occupancy prediction via temporal enhancement. In ECAI 2024, pages 129-136. IOS Press, 2024. 6 +[27] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2, 1 +[28] Yuhang Lu, Xinge Zhu, Tai Wang, and Yuexin Ma. Octreeocc: Efficient and multi-granularity occupancy prediction using octree queries. arXiv preprint arXiv:2312.03774, 2023.6 +[29] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19936-19945, 2024. 6 +[30] Zhenxing Ming, Julie Stephany Berrio, Mao Shan, and Stewart Worrall. Occfusion: Multi-sensor fusion framework for 3d semantic occupancy prediction. IEEE Transactions on Intelligent Vehicles, 2024. 3, 6 +[31] Jingyi Pan, Zipeng Wang, and Lin Wang. Co-occ: Coupling explicit feature fusion with volume rendering regularization for multi-modal 3d semantic occupancy prediction. IEEE Robotics and Automation Letters, 2024. 3, 6 +[32] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaqi Li, Hongwei Xie, Bing Wang, Li Liu, and Shang-hang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 12404-12411. IEEE, 2024. 6 +[33] Liang Peng, Junkai Xu, Haoran Cheng, Zheng Yang, Xiaopei Wu, Wei Qian, Wenxiao Wang, Boxi Wu, and Deng Cai. Learning occupancy for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10281-10292, 2024. 2 +[34] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2 +[35] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. + +In 2020 International Conference on 3D Vision (3DV), pages 111-119. IEEE, 2020. 6 +[36] 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 +[37] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincarc: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 1 +[38] Yining Shi, Kun Jiang, Ke Wang, Kangan Qian, Yunlong Wang, Jiusi Li, Tuopu Wen, Mengmeng Yang, Yiliang Xu, and Diange Yang. Effocc: A minimal baseline for efficient fusion-based 3d occupancy network. arXiv preprint arXiv:2406.07042, 2024. 6 +[39] 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. 2 +[40] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1746-1754, 2017. 2 +[41] Yang Song and Lin Wang. Bico-fusion: Bidirectional complementary lidar-camera fusion for semantic-and spatial-aware 3d object detection. arXiv preprint arXiv:2406.19048, 2024. 1 +[42] Ziying Song, Guoxing Zhang, Lin Liu, Lei Yang, Shaoqing Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Robofusion: Towards robust multi-modal 3d object detection via sam. arXiv preprint arXiv:2401.03907, 2024. 1 +[43] Xin Tan, Wenbin Wu, Zhiwei Zhang, Chaojie Fan, Yong Peng, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Geocc: Geometrically enhanced 3d occupancy network with implicit-explicit depth fusion and contextual self-supervision. arXiv preprint arXiv:2405.10591, 2024. 6 +[44] Pin Tang, Zhongdao Wang, Guoqing Wang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15035-15044, 2024. 6 +[45] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 6 +[46] Guoqing Wang, Zhongdao Wang, Pin Tang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Occgen: Generative multi-modal 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2404.15014, 2024. 3, 6 + +[47] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17850-17859, 2023. 3, 6, 2 +[48] Yuqi Wang, Yuntao Chen, Xingyu Liao, Lue Fan, and Zhaoxiang Zhang. Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17158-17168, 2024. 2, 6 +[49] Jianxiong Xiao, Andrew Owens, and Antonio Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. In Proceedings of the IEEE international conference on computer vision, pages 1625-1632, 2013. 2 +[50] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 1, 6 +[51] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 1 +[52] Zhen Yang, Yanpeng Dong, and Heng Wang. Daocc: 3d object detection assisted multi-sensor fusion for 3d occupancy prediction. arXiv preprint arXiv:2409.19972, 2024. 6 +[53] Zichen Yu, Changyong Shu, Jiajun Deng, Kangjie Lu, Zong-dai Liu, Jiangyong Yu, Dawei Yang, Hui Li, and Yan Chen. Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058, 2023. 5 +[54] Haiming Zhang, Xu Yan, Dongfeng Bai, Jiantao Gao, Pan Wang, Bingbing Liu, Shuguang Cui, and Zhen Li. Radocc: Learning cross-modality occupancy knowledge through rendering assisted distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7060-7068, 2024. 6 +[55] Jinqing Zhang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Sa-bev: Generating semantic-aware bird's-eye-view feature for multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3348-3357, 2023. 2 +[56] Ji Zhang, Yiran Ding, and Zixin Liu. Occfusion: Depth estimation free multi-sensor fusion for 3d occupancy prediction. arXiv preprint arXiv:2403.05329, 2024. 6 +[57] Jinqing Zhang, Yanan Zhang, Yunlong Qi, Zehua Fu, Qingjie Liu, and Yunhong Wang. Geobev: Learning geometric bev representation for multi-view 3d object detection. arXiv preprint arXiv:2409.01816, 2024. 2 +[58] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 6 +[59] Xiao Zhao, Xukun Zhang, Dingkang Yang, Mingyang Sun, Mingcheng Li, Shunli Wang, and Lihua Zhang. Maskbev: Towards a unified framework for bev detection and map segmentation. arXiv preprint arXiv:2408.09122, 2024. 1, 2 + +[60] Sicheng Zuo, Wenzhao Zheng, Yuanhui Huang, Jie Zhou, and Jiwen Lu. Pointocc: Cylindrical tri-perspective view for point-based 3d semantic occupancy prediction. arXiv preprint arXiv:2308.16896, 2023. 6 \ No newline at end of file diff --git a/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/images.zip b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0587edbfed670a6ccf688286dadb7643f0192a20 --- /dev/null +++ b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d17a9bc2c288a150047be621f6bfb7f6679d5cb9fb080a610f3bd5791c3c37a9 +size 889736 diff --git a/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/layout.json b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..15dfc0be38d4761c14a4959da4ed616c29bb51d6 --- /dev/null +++ b/ICCV/2025/RIOcc_ Efficient Cross-Modal Fusion Transformer with Collaborative Feature Refinement for 3D Semantic Occupancy Prediction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:993f281963631f1bf999df591af5a43e4a8ccb9c2b644c806058c2280d1e228d +size 420737 diff --git a/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/bb17e26a-dee4-4a8f-a256-4affea9ba7e5_content_list.json b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/bb17e26a-dee4-4a8f-a256-4affea9ba7e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5393d2c9da9f4594f56e7ec1bbdc81a2dcc85d8d --- /dev/null +++ b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/bb17e26a-dee4-4a8f-a256-4affea9ba7e5_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:975466c2884f1cfdfbb41075721d3c8943feca74bec4dc299b649bf850f71091 +size 80960 diff --git a/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/bb17e26a-dee4-4a8f-a256-4affea9ba7e5_model.json b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/bb17e26a-dee4-4a8f-a256-4affea9ba7e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..207b34677d777d6d0da9c31da49b87ecc0712a0c --- /dev/null +++ b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/bb17e26a-dee4-4a8f-a256-4affea9ba7e5_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8d548982b62175696c96e29d892fa3dfb3470d306679fd918e7ae75f51bd7c5 +size 98360 diff --git a/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/bb17e26a-dee4-4a8f-a256-4affea9ba7e5_origin.pdf b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/bb17e26a-dee4-4a8f-a256-4affea9ba7e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1ff391622f68012d39be789892a8911285ca63ff --- /dev/null +++ b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/bb17e26a-dee4-4a8f-a256-4affea9ba7e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5865fe0b0ab2957805c08c3742dc81be7d95c8bc085d6e1b667b4fab2227a695 +size 1600826 diff --git a/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/full.md b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a5542cb0742f76acc9ce6dfdf2429bb25d9caa65 --- /dev/null +++ b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/full.md @@ -0,0 +1,321 @@ +# RIPE: Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction + +Johannes Kunzel1,2 Anna Hilsmann1 Peter Eisert1,2 + +1Fraunhofer Heinrich-Hertz-Institut, HHI, Germany, 2Humboldt University Berlin, Germany, + +# Abstract + +We introduce RIPE, an innovative reinforcement learning-based framework for weakly-supervised training of a keypoint extractor that excels in both detection and description tasks. In contrast to conventional training regimes that depend heavily on artificial transformations, pre-generated models, or 3D data, RIPE requires only a binary label indicating whether paired images represent the same scene. This minimal supervision significantly expands the pool of training data, enabling the creation of a highly generalized and robust keypoint extractor. + +RIPE utilizes the encoder's intermediate layers for the description of the keypoints with a hyper-column approach to integrate information from different scales. Additionally, we propose an auxiliary loss to enhance the discriminative capability of the learned descriptors. + +Comprehensive evaluations on standard benchmarks demonstrate that RIPE simplifies data preparation while achieving competitive performance compared to state-of-the-art techniques, marking a significant advancement in robust keypoint extraction and description. To support further research, we have made our code publicly available at https://github.com/fraunhoferhhi/RIPE. + +# 1. Introduction + +Given two images, how can we determine whether they depict the same scene and precisely identify matching keypoints? This task, as shown in Fig. 1, is intuitive for humans: we can identify distinctive keypoints in one image and look for their counterparts in the other, ignoring distractions such as moving cars, changing foliage, all while remaining unaffected by noise or lighting variations. This natural human capability raises an intriguing question: can neural networks learn robust keypoint extraction exclusively from binary labels indicating whether two images depict the same scene? + +Traditional keypoint detection methods such as SIFT [21], ORB [29], and SURF [3] rely on handcrafted feature detectors and descriptors that struggle significantly with long-term registration tasks, particularly when images are taken hours, + +![](images/6c7d47b180505dec56d8867d63370e937589d595d089629452d73f827bcec4dc.jpg) +Figure 1. Can you tell if these images depict the same scene? While humans naturally ignore noise and lighting variations to solve this effortlessly, teaching a neural network to do the same using only image pairs poses significant challenges. (Images: MegaDepth [18]) + +![](images/11ddd2c0c616f2766a6ed48cbe5d1a1fa1280cf8373ef652ea5ce78889ab4e43.jpg) + +days or months apart. Variations in weather, lighting conditions and appearance (most notably from vegetation) pose significant problems for traditional detectors. To overcome these limitations, deep-learning-based methods were introduced to directly learn feature representations from images. Current state of the art (SOTA) approaches, including DeDoDe [9], DISK [38] and ALIKED [47] rely on training datasets like MegaDepth [18], which provide relative pose and depth information. MegaDepth, for example, consists of touristic imagery along with 3D models generated with Structure from Motion using COLMAP [34], itself reliant on classical SIFT keypoints. Other methods, like SuperPoint [6] or SiLK [13] rely on self-supervision through applying artificial augmentations, but face limitations from the available training data and the domain-gap between real and simulated scenarios, which inadequately replicate long-term changes of the real world. + +In this paper we propose RIPE (Reinforcement Learning on Image Pairs for Keypoint Extraction), a novel approach that trains keypoint detectors using only image pairs labeled as either showing the same image scene or not, eliminating the need for depth or pose information. This significantly expands the pool of usable datasets to include diverse real-world scenarios, such as large-scale autonomous driving data (ACDC [30]) or place recognition data (Tokyo 24/7 [37]), including challenging weather and illumination changes. + +Training on this broader set of datasets helps keypoint detectors become more robust to real-world conditions, especially for challenging long-term and dynamic localization tasks, as we show in our evaluation (Sec. 4.6 and Sec. 6.1). + +However, using only binary labels substantially weakens training supervision. To overcome this and address the non-differentiable nature of the keypoint selection process, we introduce a probabilistic formulation for keypoint selection via Reinforcement Learning (RL). Unlike previous RL-based methods [4, 26, 38] that still require depth or pose information, we propose to derive the reward exclusively from labeled image pairs. Our key insight is to leverage the epipolar constraint, a fundamental concept in computer vision, that all true keypoint matches in a positive pair must satisfy. This eliminates the need for pre-generated 3D models and further broadens the range of suitable training datasets. + +To efficiently associate every keypoint location with its descriptor, we incorporate intermediate multi-scale information by leveraging hyper-column features from intermediate layers of the encoder, rather than relying solely on the final low-resolution output. We further strengthen the descriptiveness of the descriptors by introducing a robust loss function explicitly designed for binary-labeled labeled image pairs. + +We evaluate RIPE on MegaDepth 1500 and HPatches, achieving competitive results compared to the state-of-the-art. On Aachen Day-Night and Boreas we demonstrate its increased robustness to adverse weather and illumination changes thanks to our minimal supervision strategy. + +In summary, the key contributions of our work include: + +- Innovative Weakly-Supervised Training Framework: Introduction of RIPE, a novel method based on Reinforcement Learning that trains keypoint detectors using only labeled image pairs, effectively removing the dependency on depth or pose information. +- Improved Generalizability: Ability to utilize diverse training datasets, improving keypoint detector performance in varying real-world conditions. +- Epipolar Geometry-based reward: Utilization of the epipolar constraint to derive rewards from labeled image pairs, ensuring that the optimization process adheres to fundamental principles of computer vision. +- Multi-Scale Feature Representation: Integration of multi-scale hyper-column features to enhance the association between keypoint locations and descriptors, leading to more informative and discriminative representations. +- Robust Descriptor Loss: Development of a robust loss function based on labeled image pairs, further strengthening the descriptiveness and reliability of the keypoint descriptors. + +# 2. Related Work + +Evolving from classical hand-crafted approaches, current training-based methods for learning keypoint detection and + +description either rely on artificial augmentations or the availability of 3D information such as pose or depth, often derived from pre-generated 3D data. We systematically categorize current state-of-the-art methods based on their underlying training principles. + +Artificial augmentations Methods in this category generate training data by applying artificial augmentations to existing images, creating image pairs with known transformations. This can be achieved by using photometric and/or homographic data augmentation (SuperPoint [6], DomainFeat [45], SiLK [13]). As these are limited in their representation of illumination changes, style-transfer techniques were introduced to synthesize night images (DomainFeat [45], Melekhov et al. [22]). + +Additionally, different losses are used to ensure that feature maps are robust against domain changes. For instance, the triplet loss (DomainFeat [45]) enhances descriptor descriptiveness, while the minimization of margins for corresponding patches (HardNet [24]) or hard-negative sampling of local features (Melekhov et al. [22]) is also utilized. + +Pose or depth information Methods of the second category depend on known pose or depth information, typically pre-generated by Structure-from-Motion (SfM) techniques. Many methods [7, 9, 12, 40, 41] utilize the work of Li and Snavely [18], leveraging the MegaDepth dataset (see Fig. 1) for two example images), which contains photos of landmark collections, automatically annotated with pixel-wise depth information. This dataset is constructed using COLMAP [34, 35] based on SIFT [21], to generate a 3D model along with pixel-wise depth information for each image. + +Known depth information is used to calculate keypoint correspondences to train detection and description (Dusmanu et al. [7]) or to calculate reward values (DISK [38]). DeDoDe [9] introduced the direct use of 3D tracks originating from reconstructed points in the 3D model as a supervision signal. ALIKE [40] proposed to ground the learning of descriptors on relative pose information by introducing a differentiable matching layer and translating the relative poses into epipolar constraints. ALIKED [39] additionally introduced attention-weighted local descriptors to include image-level spatial awareness into the descriptor, training on pixel-wise correspondences enriched with random homography transformations. Recently, the dense (RoMa [10], Mast3r [16], Dust3r [42]) and semi-dense (S2DNet [12], LoFTR [36], Efficient LoFTR [43]) matching methods moved into focus. These methods also use the available depth information from the pre-generated 3D models. + +Tyszkiewicz et al. [38] (DISK) and Bhowmik et al. [4] (Reinforced Feature Points) independently introduced Reinforcement Learning for learning keypoint detection and description to overcome the non-differentiability of keypoint detection. Bhowmik et al. trained with a complete com + +puter vision pipeline, treating the matching and pose estimation stages as non-differentiable black boxes, requiring known poses to calculate the reward value. In contrast, DISK computes rewards based on the number of correct feature matches (determined by the known relative position), allowing for precise calculation of matching probabilities. Potje et al. [26] (DEAL) extended the DISK approach by introducing an additional Warp Module, increasing robustness against non-rigid image deformations. + +SOTA Limitations Ultimately, existing methods rely on known pixel-wise correspondences derived from artificial augmentations, depth information (measured or estimated with SfM) or relative pose. Consequently, these methods remain dependent on limited datasets or augmentation techniques. RL also remains underutilized, as depth (DISK [38]), pose (Reinforced Feature Points [4]) or artificial augmentations (DEAL [26]) are still required. + +To this end, we introduce a more radical Reinforcement Learning approach, which allows us to train without known poses, without depth information, and without pixel-wise correspondences – just using paired images. This is comparable to semantic keypoint matching works like Rocco et al. [28], but differs in the usage of RL, the inclusion of negative pairs and the enforcement of a valid epipolar geometry. + +# 3. Method + +In the following, we present our approach to learning keypoint detection and description solely from unlabeled image pairs by using Reinforcement Learning (RL). A visual overview of our method is provided in Fig. 2. For each image in a given pair, a neural network generates a heatmap from which keypoint positions are sampled (Sec. 3.1). Each keypoint is then associated with a descriptor, extracted from the decoder using hyper-column features (Sec. 3.2). We process pairs of images and the resulting keypoints are matched and filtered by estimating the fundamental matrix. The final number of successfully matched keypoints is used as the reward signal, based on the label of the input pair. Using Reinforcement Learning (Eq. (4)), the reward encourages the network to generate a greater number of matchable keypoints - consistent with the epipolar constraint - and to produce fewer keypoints for negative pairs. This formulation effectively leverages the feedback from both geometric consistency and image similarity to guide the learning process. + +Consequently, the only required label is binary, indicating whether the paired images depict the same scene (i.e. a sufficient number of 2D image keypoints correspond to projections of the same 3D point from the depicted scene) or not. Hence, we assume a dataset $\mathcal{D} = \{(I_{\kappa}, I_{\kappa}', \lambda_{\kappa}) | \kappa = 1, \dots, N\}$ where each tuple contains two images and a + +![](images/ddbd68bf1dec0d5c214a89ce80eb234e76349e5c9919fb6d27675a12875e6626.jpg) +Figure 2. Overview of RIPE, our approach for learning keypoint detection and description from unlabeled image pairs using reinforcement learning. For an image pair, heatmaps are generated for probabilistic keypoint sampling, with descriptors derived from hyper-column features. These keypoints are matched and filtered via the fundamental matrix, with the number of matchable keypoints serving as the reward signal. This encourages the network to produce a large number of keypoints fulfilling the epipolar constraint for positive pairs and fewer keypoints for negative pairs. + +binary label $\lambda_{\kappa}$ with + +$$ +\lambda_ {\kappa} = \left\{ \begin{array}{l l} 1, & \text {i f} I _ {\kappa} \text {a n d} I _ {\kappa} ^ {\prime} \text {s h o w t h e s a m e s c e n e} \\ - 1, & \text {o t h e r w i s e} \end{array} \right. \tag {1} +$$ + +to indicate positive and negative pairs. + +# 3.1.Keypoint detection + +Our proposed method RIPE starts with detecting possible keypoint positions in each image, employing an hourglass network composed of an encoder $\mathbf{e}_{\theta}(\cdot)$ and a decoder $\mathbf{d}_{\theta}(\cdot)$ connected via skip connections with learnable parameters $\theta$ . A graphical overview is given in the upper part of Fig. 3. For an input image $\mathbf{I} \in \mathbb{R}^{h \times w \times c}$ , the network generates a detection heatmap $\mathbf{H} \in \mathbb{R}^{h \times w}$ with $\mathbf{H} = \mathbf{d}_{\theta}(\mathbf{e}_{\theta}(\mathbf{I}))$ , indicating potential keypoint locations. The heatmap $\mathbf{H}$ is divided into a regular grid, with each cell $\mathbf{c}_i$ having a size $m \times m$ . The total number of cells is denoted by $C = \left\lfloor \frac{h}{m} \right\rfloor \times \left\lfloor \frac{w}{m} \right\rfloor$ , with $\lfloor \cdot \rfloor$ denoting the floor function. Each cell $\mathbf{c}_i \in \mathbb{R}^{m \times m}$ holds the corresponding logit values. The logit values in each cell constitute a categorical probability distribution from which exactly one keypoint location per cell is sampled, resulting in a keypoint position $s_i$ , with an initial probability $\hat{p}_i$ and logit $l_i$ . As a result, the network learns to define the keypoint locations by shaping the logit values accordingly. Acknowledging that some image regions (such as overexposed sections, sky, etc.) may not be optimal for reliable keypoint detection, a sigmoid function is applied to the keypoint logit forming an acceptance indicator $a_i = \sigma(l_i)$ for each cell. This enables the network to probabilistically discard unsuitable keypoints. Consequently, the final probability $p_i = \sigma(l_i) \cdot \hat{p}_i$ captures both the initial sampling chance and the likelihood of keypoint retention. + +![](images/c760eb9642fa165c742b7226e19601286145a2041bf7358fd431e81d9976a2a3.jpg) +Figure 3. Keypoint detection (top) and description (bottom) for a single input image $\mathbf{I}$ . The network outputs a logit heatmap $\mathbf{H}$ for each input image. Potential keypoint locations $\mathbf{S}$ are sampled from equally sized patches based on logit values, with rejected keypoints marked in orange and accepted ones in green. At the keypoint locations, descriptors $\mathbf{D}$ are generated from the intermediate encoder layers, linking each location to a descriptor integrating context at different scales from the intermediate layers. + +The keypoint detection results in a list $\mathbf{S}_{\mathbf{I}} \in \mathbb{R}_{2 \times C}$ of keypoint locations, acceptance indicators $\mathbf{a}_{\mathbf{I}} \in \mathbb{R}_{1 \times C}$ , and associated probability values $\mathbf{p}_{\mathbf{I}} \in \mathbb{R}_{1 \times C}$ gathered from the individual cells of each keypoint in the image $\mathbf{I}$ . The total number of keypoints is equivalent to the number of cells $C$ . + +# 3.2. Keypoint description + +Once the keypoints are detected, the next step is to assign a descriptor to each location in order to enable the matching between images. This process is visualized in the lower part of Fig. 3. Potje et al. [26] demonstrated that features derived from the encoder and bilinearly upsampled to the input resolution retain a high distinctiveness under small to moderate photometric and geometric changes. However, this approach proved insufficient for our requirements, as it does not facilitate a reliable matching, as discussed in Sec. 6.2. To enhance (with minimal computational overhead) the quality of the features created simply by upsampling the final feature layer, we employ hyper-column features [14], a concept also proposed by Germain et al. [11] and tested in diverse settings (e.g. Li et al. [17], Min et al. [23]) to combine multi-scale information for increased descriptiveness. + +In each layer $l$ of the encoder, feature vectors $\mathbf{e}_i^l$ at the detected keypoint locations $\mathbf{s}_i\in \mathbf{S}$ are bilinearly interpolated (please refer to Fig. 3 for a visualization). This results in a list + +$(\mathbf{e}_i^1,\mathbf{e}_i^2,\dots \mathbf{e}_i^L)$ of feature vectors that encode information at different image scales. These feature vectors are then concatenated to form the intermediate hyper-column features $\mathbf{D}_{\mathrm{hyper}}\in \mathbb{R}_{\hat{d}\times C}$ , with $\hat{d}$ being the overall sum of channels for the encoder layers. As the final dimensionality $\hat{d}$ is usually large (960 for a VGG-19), a final 1x1 convolution is applied to reduce the dimensionality to $d$ , resulting in compact descriptors $\mathbf{D}\in \mathbb{R}_{d\times C}$ . + +For a single image $\mathbf{I}$ , the detection and description results in: keypoint locations $\mathbf{S} \in \mathbb{R}_{2 \times C}$ (in image coordinates), acceptance indicators $\mathbf{a} \in \mathbb{R}_{1 \times C}$ , selection probabilities $\mathbf{p} \in \mathbb{R}_{1 \times C}$ , and the descriptor map $\mathbf{D} \in \mathbb{R}_{d \times C}$ . During training, we repeat the entire process to obtain $\mathbf{S}'$ , $\mathbf{a}'$ , $\mathbf{p}'$ , and $\mathbf{D}'$ for the second image $\mathbf{I}'$ of a pair. + +# 3.3. Reinforcement of matchable keypoints + +To address the discrete nature of the keypoint detection process and to allow training from unlabeled image pairs only, we formulate it as a reinforcement problem. Starting from a brief general introduction, we develop the methodical foundation of RIPE. + +The policy is defined as a probability distribution over actions $\mathcal{A}$ , conditioned on the current state $S$ and parameterized by $\theta$ with + +$$ +\pi_ {\theta} (\mathcal {S}) = \mathbb {P} [ \mathcal {A} | \mathcal {S}, \theta ]. \tag {2} +$$ + +This constitutes a probability distribution, from which an action $\mathcal{A}$ is sampled. Based on the sampled action, the agent receives a reward signal that indicates a good or bad action. The learning objective is then formulated as maximizing the expected cumulative reward over a trajectory $\tau$ (a sequence of state, action, reward tuples) scaled by the reward $R$ : + +$$ +\max _ {\theta} J (\theta) = \mathbb {E} _ {x \sim \pi_ {\theta}} [ R (\tau) ], \tag {3} +$$ + +However, directly calculating the derivative of $J$ is not feasible, as it would require differentiating through all possible trajectories and the state distribution. Nonetheless, REINFORCE [44] provides an approximation for the derivative: + +$$ +\nabla_ {\theta} J (\theta) \approx \hat {g} = \sum_ {t = 0} \nabla_ {\theta} \log \pi_ {\theta} \left(a _ {t} \mid s _ {t}\right) R (\tau). \tag {4} +$$ + +In RIPE, the encoder-decoder network acts as a trainable policy, with the input image $\mathbf{I}$ representing the state. This leads to the following formulation for the policy: + +$$ +\begin{array}{l} \pi_ {\theta} (\mathbf {s}) = d _ {\theta} \left(e _ {\theta} (\mathbf {I})\right) = \mathbf {p} \\ = \left[ \mathbb {P} _ {1} [ a _ {1} | I, \theta ], \mathbb {P} _ {2} [ a _ {2} | I, \theta ], \dots , \mathbb {P} _ {c} [ a _ {c} | I, \theta ]) \right], \tag {5} \\ \end{array} +$$ + +where $\mathbf{p}$ is a list of distributions for each cell $\mathbf{c}$ in the heatmap $\mathbf{H}$ . As described in Sec. 3.1, keypoint locations are sampled from these distributions (i.e. the keypoint localization corresponds to an action). + +![](images/88b734154e8e5f367504fa0bf52f092b1007acdb89d4aee2bda3f3804325b7f0.jpg) +Figure 4. Overview of the proposed Reinforcement Learning formulation for learning keypoint detection and description from unlabeled image pairs. A network/ agent generates probability distributions over potential keypoint locations. Actions (keypoint locations S) are sampled from these distributions and associated with their respective probability $\mathbf{p}$ . The agent receives rewards based on the number of mutual nearest neighbors further filtered by estimating the fundamental matrix. This encourages the detection of matching keypoints in positive pairs ( $\lambda = 1$ ) while penalizing incorrect detections in negative pairs ( $\lambda = -1$ ). Using REINFORCE [44], gradients are derived and utilized to update the network parameters accordingly. + +Working with image pairs, we introduce a second list of probabilities $\mathbf{p}'$ and compute the joint probability across all combinations of cells. This allows us to approximate the gradient, as given by: + +$$ +\hat {g} = \sum_ {c = 0} ^ {C} \nabla_ {\theta} \left(\log p _ {c} \oplus \log p _ {c} ^ {\prime}\right) R, \tag {6} +$$ + +with $\oplus$ denoting the outer sum + +The only ingredient missing is the reward $R$ . As motivated in our introductory example, the goal is to produce many matching keypoints for positive pairs and few for a negative pairs. Consequently, the network should be rewarded for detecting keypoints in positive image pairs that are both matchable and pass the geometric filtering based on the epipolar constraint. Conversely, a negative reward (penalty) should be applied when the network incorrectly detects keypoints that appear matchable and filterable, yet originate from different scenes. The reward is computed using mutual nearest-neighbor estimation and RANSAC filtering, which, despite being non-differentiable, are used here solely for the reward calculation, and thus do not require gradient computation. + +RIPE identifies mutual nearest neighbors between the keypoint descriptors in $\mathbf{D}$ and $\mathbf{D}'$ by computing their L2-distance, yielding a list of index pairs $\mathbf{C}$ . This list is again filtered by estimating the fundamental matrix $\mathbf{F}$ using the 8-point algorithm in combination with RANSAC. The resulting mask $\mathbf{M}$ indicates all mutual nearest neighbors that can be explained by a common epipolar geometry. + +The reward matrix $\mathbf{R} \in \mathbb{R}^{C \times C}$ with the individual reward values $r_{i,j} \in \mathbf{R}$ is created by + +$$ +r _ {p, q} = \left\{ \begin{array}{l l} \operatorname {s i g n} \left(\lambda_ {\kappa}\right) \rho , & (p, q) \in \mathbf {C} \text {a n d} \mathbf {M} _ {p, q} \text {i s T r u e} \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {7} +$$ + +with the reward $\rho$ and the signum function $\mathrm{sign}(\cdot)$ . Simply inverting the reward for negative pairs produces in a symmetric training signal, which we found to be beneficial during training. + +The resulting gradients of Eq. (6) adjust the network parameters $\theta$ to increase the likelihood of selecting keypoints (i.e., actions taken) that result in a positive reward. Since the expectation is approximated by the mean, a substantial number of samples is required for a reliable estimate. In our approach, the sampling of $C$ keypoint locations for every image acts as an equivalent of sampling multiple trajectories (as in classical reinforcement learning), and allows a sufficient approximation of the expectation. + +# 3.4. Descriptor loss + +To enhance the descriptiveness and robustness of the descriptors, we integrate a second loss + +$$ +L _ {\text {d e s c}} = \left\{ \begin{array}{l l} \frac {1}{N} \sum_ {n = 1} ^ {N} \max \left(0, \mu + \delta_ {+} ^ {n} - \delta_ {h} ^ {n}\right), & \lambda_ {\kappa} = 1 \\ \frac {1}{N} \sum_ {n = 1} ^ {N} \max \left(0, \mu - \delta_ {+} ^ {n}\right), & \text {o t h e r w i s e} \end{array} \right. \tag {8} +$$ + +where $N$ is the number of inliers after the RANSAC filtering, $\mu$ is a positive threshold, and $\delta$ the L2-distance between + +![](images/b3a99914e9b7b9df1fabcdf6d4149e9f188481d15df2b7d72e6ac2ff6ce54afa.jpg) +Figure 5. Example results for RIPE (ours), ALIKED [47], and DeDoDe [9] from the MegaDepth 1500 benchmark. RIPE demonstrates its ability to effectively utilize the unlabeled paired images to disregard irrelevant regions and identify highly discriminative keypoints. DeDoDe tends to cluster its keypoints in specific areas, whereas the keypoints of ALIKED and RIPE are more evenly distributed across the image. + +two descriptors. This loss behaves differently depending on the type of image pair, with $\delta_{+}^{n}$ being the L2-distance between two keypoint descriptors, which are matched and were validated (i.e., identified as inliers) by geometric filtering, while $\delta_h^n$ is the distance to the second closest neighbor. In the positive case, this loss forces the descriptors of putative matches (putative, as we have no ground truth verification for belonging to the same 3D point) to become more similar while repelling all other descriptors. In contrast, in the negative case, where no matches should exist, the loss drives the descriptors of incorrectly matched keypoints further apart, thus increasing robustness by minimizing false positives. + +# 3.5. Final loss + +We define the final loss for the descriptor as $\mathcal{L}_{\mathrm{desc}} = -\mathbb{E}_{\mathbf{K}}[\mathbf{R}]$ . Inspired by the work of Potje et al. [26], we add a regularization term + +$$ +\mathcal {L} _ {\text {l o w}} = - \sum_ {p} \sum_ {q} \log p _ {i} \cdot \epsilon , \tag {9} +$$ + +with $\epsilon$ being a small negative constant to prevent the network from generating low probability keypoints. The final loss is calculated by + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d e c t}} + \mathcal {L} _ {\text {l o w}} + \psi \mathcal {L} _ {\text {d e s c}}, \tag {10} +$$ + +where $\psi$ is a small positive constant, to balance detection and description loss. + +By framing keypoint detection as an RL problem, RIPE can backpropagate the final loss to optimize both keypoint locations and descriptors, effectively tackling the challenge of learning without relying on extensive labeled datasets. The use of RL and formulating the detection process as a probabilistic decision also overcomes the inherent discreteness of the keypoint selection, enabling direct optimization. With this approach, we demonstrate how keypoints can be learned effectively using only unlabeled image pairs – mirroring the way humans would perform the task. + +# 4. Experiments + +# 4.1. Implementation details + +We used the TORCHVISION implementation of VGG-19 with strides of $[1,2,4,8]$ and pretrained on ImageNet + +
VariableValuePurpose
h, w560image size
m8cell size
ρ1.0reward Eq. (7)
ε-7e-8constant regularization Eq. (9)
ψ5.0descriptor loss weight Eq. (10)
+ +Table 1. Overview of our hyperparameter configuration. as our encoder. The feature maps on each scale have [64, 128, 256, 512] channels, respectively, resulting in $d' = 960$ (see Sec. 3.2), reduced to $d = 256$ , by the 1x1 convolutional layer. For the decoder, we follow DeDoDe [9] and use the depthwise convolutional refiners proposed by Edstedt et al. in DKM [8], with 8 blocks per scale and internal dimension [64, 128, 256, 512]. We use the GPU-accelerated mutual nearest neighbor implementation from the Kornia library [27] and the fundamental matrix estimation from PoseLib [15]. The network is trained with AdamW [20] with a learning rate starting from 0.001 and linearly decaying to 1e-6. We train for 80,000 steps with a batch size of 6 and a gradient accumulation over 4 batches, taking three days on a single A100 GPU. Input images are normalized, resized and padded to meet the desired input resolution, without altering the aspect ratio. In addition, we set the hyperparameters according to Tab. 1. To facilitate the initial stages of training, we linearly increase the influence of $\epsilon$ during the first third of the training process. + +# 4.2. Training data + +We use the same Megadepth subset as [38], reducing the dataset to image pairs which show a sufficient number of covisible 3D points. However, our innovative training regime enables the use of an extended data basis. In Sec. 4.6 and Sec. 6.1 we therefore demonstrate how this can improve the robustness of our method. + +# 4.3. Inference + +During inference, we pass a single image through the network and sample the top $K$ most likely keypoints and their positions from $\mathbf{H}$ , accompanied with non-maximum suppression in a 3x3 window. Each keypoint is associated with its hyper-column descriptor, as described in Sec. 3.2. The mean inference time (measure for the setting of Sec. 4.4 was + +
CategoryMethodMegaDepth 1500LabelHPatches
AUC@5°AUC@10°AUC@20°AUC@1pxAUC@3pxAUC@5px
(Semi-)DenseEff. LoFTR[43]CVPR'2464.6378.0387.21D42.3863.2672.75
Mast3r[16]ECCV'2422.8435.2947.62D29.6354.0466.86
RoMa[10]CVPR'2468.2880.5388.77D38.3863.7573.37
SparseSIFT[21]UCV'0437.2950.6161.89—29.9853.9365.01
SuperPoint[6]CVPRW'1849.562.5172.12H36.6156.9267.28
DISK[38]Neurlps'2051.8364.4974.3P/D37.5457.6868.08
ALIKED[47]TIM'2356.7169.8679.54P+H31.9557.6569.35
SiLK[13]ICCV'2329.8340.5949.98H39.0557.9667.64
DeDoDe-B[9]3DV'2455.0267.9977.31P39.8860.5970.36
RIPE55.1168.3478.03Pair37.9358.9369.20
+ +Table 2. Results for the evaluation of relative pose estimation on MegaDepth-1500 (left) and homography estimation on HPatches (right). We specify the training data for each approach, with D indicating depth, P position and H homography/ artificial augmentations. Crucially, our method is the only one that requires only unlabeled image pairs, yet achieves competitive performance to the sparse SotA methods. The best and second-best performances for the sparse methods are highlighted. + +0.47 s for RIPE (for comparison: DeDoDe 0.45 s, ALIKED 0.08 s, DISK 0.17 s). + +# 4.4. Relative pose estimation + +Dataset We evaluated the relative pose estimation performance using the MegaDepth-1500 subset. It contains two (Brandenburger Tor and St. Peters Square) out of the 196 scenes from the original MegaDepth dataset and was introduced in LoFTR [36]. The main challenges are large viewpoint, illumination changes and repetitive patterns. Following recent evaluation protocols ([43], [19]), we resized the longer side of the input images to 1600 for the dense methods and to 1200 for the (semi-)sparse methods. To assess the quality of the extracted keypoints, we use mutual-nearest-neighbor matching for all sparse methods. + +Baselines We compare RIPE against SotA methods for sparse keypoint detection and description. Due to their recent success, we also integrate dense (RoMA [10], Mast3r [16]) and semi-dense matching methods for comparison (Efficient LoFTR [43]). For the sparse methods, we use mutual-nearest-neighbor matching to establish correspondences. + +Metrics Building on preceding approaches, the accuracy of matches is assessed by evaluating the relative poses they yield. The pose error is characterized as the greatest of the angular discrepancies in both rotation and translation. We provide the Area Under the Curve (AUC) of the pose error at the thresholds of $5^{\circ}$ , $10^{\circ}$ , and $20^{\circ}$ . We used the glue-factory library, kindly provided by the authors of GlueStick [25] and LightGlue [19]. We use the poselib [15] for the robust pose estimation and select the top 2048 keypoints. + +Results The left half of Tab. 2 and demonstrate that our method, RIPE, ranks a close second to the current state-of-the-art sparse method, ALIKED, while not requiring images with known poses during training and leveraging a significantly weaker training signal. Remarkably, RIPE even out + +performs DeDoDe, despite the latter employing two distinct networks for detection and description, which nearly doubles the number of required parameters. Furthermore, our method is the first to achieve this performance without relying on artificial homographies (as seen in methods like SuperPoint and SiLK) or requiring pose or depth information from a pre-registered 3D model. Qualitative results can be found in Fig. 5 and in the supplementary (Fig. 6). + +# 4.5. Homography Estimation + +Dataset We evaluate on the HPatches dataset [2], which contains sequences of planar scenes, taken with viewpoint or illumination changes. We resized the input smaller side of the input images to 480 pixels. + +Metrics To assess the quality of the homography estimation we calculate the mean reprojection error of the corner points and report the AUC for the thresholds of 1, 3 and 5 pixels. For all methods, we use mutual-nearest-neighbor matching and the implementation of poselib [15] for the robust homography estimation and restricted the number of keypoint to 1024. + +Results As illustrated on the right side of Tab. 2, RIPE once again performs on par with the state-of-the-art methods SiLK and DeDoDe, despite not utilizing artificial homographies like SiLK or pose information as employed by DeDoDe. The strong performance of SiLK highlights the advantages of its training regimen, which relies solely on artificial augmentations and therefore closely resembles the testing data. Conversely, the impressive performance of ALIKED diminishes on the HPatches benchmark, where it only narrowly surpasses SIFT. + +# 4.6. Outdoor localization day-night + +Dataset To evaluate our approach in the context of visual localization (i.e. the estimation of the 6-DoF pose for + +
MethodDayNight
.25m/2°.5m/5°5m/10°.25m/2°.5m/5°5m/10°
ALIKED87.393.997.373.388.096.9
DeDoDe82.289.092.647.156.564.4
SIFT82.588.591.930.938.246.1
RIPE81.689.293.152.967.579.1
↓+Tokyo+0.0-0.2-0.7+7.3+5.3+4.7
RIPE81.689.092.460.272.883.8
+ +a query image relative to a 3D scene model), we use the Aachen v1.1 [32, 33, 46] dataset. This dataset is especially challenging because of its large viewpoint and illumination (day-night) changes. We used the HLoc localization framework [31] for the evaluation and first triangulate a 3D model from the 6,697 reference images. For each of the 1015 (824 daytime, 191 nighttime) queries we retrieve 50 images using NetVLAD [1] and match them. To evaluate the influence of additional training data, we replaced $20\%$ of the training samples with images from the Tokyo 24/7 et al. [37] query dataset, originally intended for place recognition. This dataset contains images from 125 distinct locations. Images were captured at each location from three different viewing directions, across day, dusk, and night, resulting in a total of 1,125 images. The images are paired based on their geoposition only. No 3D positions or depth maps are available. + +Metrics The camera pose is estimated with a Perspective-n-Point solver in conjunction with RANSAC and the AUC is reported for thresholds $0.25\mathrm{m} / 2^{\circ}$ , $0.5\mathrm{m} / 5^{\circ}$ and $1.0\mathrm{m} / 10^{\circ}$ . + +Results Our results in Tab. 3 highlight the importance of diverse training data for the robustness of trained keypoint extractors. As ALIKED was not only trained on MegaDepth, but also on synthetic and training images from the Aachen[33] dataset, it shows strong performance for the localization of night-time queries. The results from DeDeDo demonstrate the influence of the limited training data (MegaDepth only). RIPE clearly outperforms DeDoDe, even if only trained on MegaDepth. When day-to-nighttime images are added to training data, the margin further increases. This improvement is made possible by our innovative training regime, which facilitates the addition of diverse training data. + +Additional experiments presented in the supplementary material (Sec. 6.1) further support these findings, with results obtained from the Boreas [5] dataset, which encompasses challenging weather conditions. + +# 4.7. Evaluation dataset Composition + +Table 3. Outdoor visual localization on the Aachen Day-Night v1.1. The results emphasize the significance of available training data: ALIKED outperforms DeDoDe by incorporating images from the Aachen dataset. Additionally, RIPE demonstrates substantial improvements by including day-night training pairs from Tokyo 24/7. The best and second-best performances are highlighted. + +
MegaDepth TokyoDayNight
0.25m/20.5m/55m/100.25m/20.5m/55m/10
1.00.081.689.293.152.967.579.1
0.90.181.387.593.257.168.182.7
0.80.281.689.092.460.272.883.8
0.70.380.688.19356.571.782.2
0.60.479.486.092.158.170.785.3
0.50.578.084.289.456.568.683.2
0.40.671.881.28754.570.283.8
+ +Table 4. Evaluation on how training data from the Tokyo 24/7 dataset improves the ability of RIPE to handle day to night illumination changes. The best performances are highlighted. + +We further investigated which data mix between MegaDepth and Tokyo 24/7 yields the best improvement on the Aachen Day-Night benchmark. We trained RIPE with different compositions, as presented in Tab. 4, and found that $80\%$ MegaDepth data and $20\%$ Tokyo data were the most advantageous. A further increase in the proportion of Tokyo data begins to degrade the results, likely due to the lack of viewpoint variability in this dataset, as the images primarily differ in illumination. Consequently, RIPE struggles to learn to cope with the strong viewpoint variations that accompany the day-to-night changes in the Aachen benchmark. + +Additional experiments regarding the influence of our hyperparameters can be found in the supplementary (Sec. 6.2). + +# 5. Conclusion + +We present a fundamentally new approach to learning keypoint detection and description. By integrating Reinforcement Learning with essential computer vision principles, we successfully train RIPE using only unlabeled image pairs, expanding the pool of available training data. This eliminates the constraints imposed by traditional approaches that depend on precise geometric annotations, making our method scalable and adaptable to diverse real-world scenarios. + +Despite leveraging a significantly weaker training signal, RIPE achieves performance on par with state-of-the-art sparse keypoint extractors. Furthermore, our results demonstrate that RIPE effectively benefits from the inclusion of diverse training data, improving its generalization capabilities and robustness to challenging conditions. This highlights the potential of our approach to redefine keypoint learning, enabling broader applicability across various domains. + +Acknowledgments We thank our colleague, Wieland Morgenstern, for his valuable feedback on the manuscript. This work was partly funded by the Federal Ministry of Education and Research (REFrame, grant no. 01IS2407A) and the German Federal Ministry for Economic Affairs and Climate Action (DeepTrain, grant no. 19S23005D). + +# References + +[1] Relja Arandjelović, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN Architecture for Weakly Supervised Place Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5297-5307, 2016. 8 +[2] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. HPatches: A Benchmark and Evaluation of Handcrafted and Learned Local Descriptors. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3852-3861, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 7 +[3] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. SURF: Speeded Up Robust Features. ECCV 2006, 9th European Conference on Computer Vision, pages 404-417, 2006. 1 +[4] Aritra Bhowmik, Stefan Gumhold, Carsten Rother, and Eric Brachmann. Reinforced Feature Points: Optimizing Feature Detection and Description for a High-Level Task. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 00:4947-4956, 2020. 2, 3 +[5] Keenan Burnett, David J Yoon, Yuchen Wu, Andrew Z Li, Haowei Zhang, Shichen Lu, Jingxing Qian, Wei-Kang Tseng, Andrew Lambert, Keith YK Leung, Angela P Schoellig, and Timothy D Barfoot. Boreas: A multi-season autonomous driving dataset. The International Journal of Robotics Research, 42(1-2):33-42, 2023. 8 +[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-Supervised Interest Point Detection and Description. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 337-349, 2018. 1, 2 +[7] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A Trainable CNN for Joint Description and Detection of Local Features. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8084–8093, 2019. 2 +[8] Johan Edstedt, Ioannis Athanasiadis, Marten Wadenback, and Michael Felsberg. DKM: Dense kernelized feature matching for geometry estimation. arXiv, 2022. 6 +[9] Johan Edstedt, Georg Bokman, Mårten Wadenback, and Michael Felsberg. DeDoDe: Detect, Don't Describe — Describe, Don't Detect for Local Feature Matching. 2024 International Conference on 3D Vision (3DV), pages 148–157, 2024. 1, 2, 6 +[10] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 00:19790-19800, 2024. 2, 7 +[11] Hugo Germain, Guillaume Bourmaud, and Vincent Lepetit. Sparse-to-Dense Hypercolumn Matching for Long-Term Visual Localization. 2019 International Conference on 3D Vision (3DV), 00:513-523, 2019. 4 +[12] Hugo Germain, Guillaume Bourmaud, and Vincent Lepetit. S2dnet: Learning image features for accurate sparse-to-dense matching. In Computer Vision - ECCV 2020, pages 626-643, Cham, 2020. Springer International Publishing. 2 + +[13] Pierre Gleize, Weiyao Wang, and Matt Feiszli. SiLK: Simple Learned Keypoints. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22442-22451, 2023. 1, 2 +[14] Bharath Hariharan, Pablo Arbelaez, Ross Girshick, and Jitendra Malik. Hypercolumns for Object Segmentation and Fine-Grained Localization. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 447–456, 2015. 4 +[15] Viktor Larsson and contributors. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 6, 7 +[16] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding Image Matching in 3D with MASTR. In European Conference on Computer Vision, pages 71-91, 2024. 2, 7 +[17] Wanhua Li, Yueqi Duan, Jiwen Lu, Jianjiang Feng, and Jie Zhou. Learning toCompose Hypercolumns for Visual Correspondence. Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV, pages 346-363, 2020. 4 +[18] Zhengqi Li and Noah Snavely. MegaDepth: Learning Single-View Depth Prediction from Internet Photos. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2041–2050, 2018. 1, 2 +[19] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In International Conference on Computer Vision (ICCV), 2023. 7 +[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 6 +[21] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2 +[22] Iaroslav Melekhov, Zakaria Laskar, Xiaotian Li, Shuzhe Wang, and Juho Kannala. Digging Into Self-Supervised Learning of Feature Descriptors. In 2021 International Conference on 3D Vision (3DV), pages 1144-1155, Los Alamitos, CA, USA, 2021. IEEE Computer Society. 2 +[23] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Hyperpixel Flow: Semantic Correspondence with Multi-layer Neural Features. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 00:3394-3403, 2019. 4 +[24] Anastasiya Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor's margins: Local descriptor learning loss. In Neural Information Processing Systems, 2017. 2 +[25] Rémi Pautrat*, Iago Suárez*, Yifan Yu, Marc Pollefeys, and Viktor Larsson. GlueStick: Robust Image Matching by Sticking Points and Lines Together. In International Conference on Computer Vision (ICCV), 2023. 7 +[26] Guilherme Potje, Felipe Cedar, Andre Araujo, Renato Martins, and Erickson R. Nascimento. Enhancing Deformable Local Features by Jointly Learning to Detect and Describe Keypoints. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1306-1315, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 2, 3, 4, 6 + +[27] E. Riba, D. Mishkin, D. Ponsa, E. Rublee, and G. Bradski. Körnia: an open source differentiable computer vision library for pytorch. In Winter Conference on Applications of Computer Vision, 2020. 6 +[28] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. End-to-end weakly-supervised semantic alignment. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6917-6925, 2018. 3 +[29] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: an efficient alternative to SIFT or SURF. 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 1 +[30] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. ACDC: The Adverse Conditions Dataset with Correspondences for Semantic Driving Scene Understanding. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 00: 10745-10755, 2021. 1 +[31] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From Coarse to Fine: Robust Hierarchical Localization at Large Scale. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12708–12717, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 8 +[32] Torsten Sattler, Tobias Weyand, Bastian Leibe, and Leif Kobbelt. Image Retrieval for Image-Based Localization Revisited. In *British Machine Vision Conference (BMCV)*, 2012. 8 +[33] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, Fredrik Kahl, and Tomas Pajdla. Benchmarking 6DOF Outdoor Visual Localization in Changing Conditions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8 +[34] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2 +[35] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 2 +[36] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-Free Local Feature Matching with Transformers. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 00:8918-8927, 2021. 2, 7 +[37] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 Place Recognition by View Synthesis. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1808-1817, 2015. 1, 8 +[38] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. DISK: Learning local features with policy gradient. In Advances in Neural Information Processing Systems, pages 14254-14265. Curran Associates, Inc., 2020. 1, 2, 3, 6 +[39] Changwei Wang, Rongtao Xu, Ke Lu, Shibiao Xu, Weiliang Meng, Yuyang Zhang, Bin Fan, and Xiaopeng Zhang. Attention Weighted Local Descriptors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10632-10649, 2023. 2 + +[40] Qianqian Wang, Zhou Xiaowei, Bharath Hariharan, and Noah Snavely. Learning Feature Descriptors Using Camera Pose Supervision. In Computer Vision - ECCV 2020, pages 757-774. Springer International Publishing, 2020. 2 +[41] Qianqian Wang, Xiaowei Zhou, Bharath Hariharan, and Noah Snavely. Learning feature descriptors using camera pose supervision. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I, page 757–774, Berlin, Heidelberg, 2020. Springer-Verlag. 2 +[42] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D Vision Made Easy. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20697-20709, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2 +[43] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 00:21666-21675, 2024. 2, 7 +[44] Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3-4):229-256, 1992. 4, 5 +[45] Rongtao Xu, Changwei Wang, Shibiao Xu, Weiliang Meng, Yuyang Zhang, Bin Fan, and Xiaopeng Zhang. DomainFeat: Learning Local Features With Domain Adaptation. IEEE Transactions on Circuits and Systems for Video Technology, 34(1):46-59, 2024. 2 +[46] Zichao Zhang, Torsten Sattler, and Davide Scaramuzza. Reference pose generation for long-term visual localization via learned features and view synthesis. Int. J. Comput. Vision, 129(4):821-844, 2021. 8 +[47] Xiaoming Zhao, Xingming Wu, Weihai Chen, Peter C. Y. Chen, Qingsong Xu, and Zhengguo Li. Aliked: A lighter keypoint and descriptor extraction network via deformable transformation. IEEE Transactions on Instrumentation and Measurement, 72:1-16, 2023. 1, 6 \ No newline at end of file diff --git a/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/images.zip b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bfa7227f281e045a08f1303db9a25b16ab537c56 --- /dev/null +++ b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f309b15bdc04f002932ec5574c5c84be4a056d35ed2497e6a67fdf456a42f6e +size 448698 diff --git a/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/layout.json b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d3307d5b1b048d5987a9bc4a45eb7bbd4d182ab3 --- /dev/null +++ b/ICCV/2025/RIPE_ Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e53bf91b1534b20afd90f327a0f9575b83cc4756f2e1cdc5f5699a403d0af96 +size 385527 diff --git a/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/21de1f49-8d33-4331-9b98-fe11d063adec_content_list.json b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/21de1f49-8d33-4331-9b98-fe11d063adec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0e2b127842c7fed04e4a5c2a4da439c57bc89acf --- /dev/null +++ b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/21de1f49-8d33-4331-9b98-fe11d063adec_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2cdc6aa4ca94f86589494fc4beef4674fec0fbaeca2fce20c45eca5a7b39a43 +size 86363 diff --git a/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/21de1f49-8d33-4331-9b98-fe11d063adec_model.json b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/21de1f49-8d33-4331-9b98-fe11d063adec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b71da828b6794677caa2f1fa8c6ec1e09ba88721 --- /dev/null +++ b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/21de1f49-8d33-4331-9b98-fe11d063adec_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d43f4dd495001a03ae1804041680eeca2bd3e8a580c6da7d9598b8bd09ba0dae +size 110245 diff --git a/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/21de1f49-8d33-4331-9b98-fe11d063adec_origin.pdf b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/21de1f49-8d33-4331-9b98-fe11d063adec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..90e59334e443e326431044102ebf852e793012ed --- /dev/null +++ b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/21de1f49-8d33-4331-9b98-fe11d063adec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:751f74f635d0b759ca1946ab4b93e5dbff7b365138d266f3f03386cd9bac6a16 +size 2104268 diff --git a/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/full.md b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aee1455eb2cafe65bd979d8742ef0f5ef49b145e --- /dev/null +++ b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/full.md @@ -0,0 +1,320 @@ +# RMultiplex200K: Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications + +Sijia Chen +Hong Kong University of Science and Technology (Guangzhou) +sijiachen@hkust-gz.edu.cn + +Bin Song† +Xidian University +bsong@mail.xidian.edu.cn + +# Abstract + +Visual Language Models (VLMs) have achieved remarkable success in many domains due to their ability to perform step-by-step reasoning. However, progress in the telecommunication (Telecom) domain remains limited, primarily due to the lack of high-quality datasets and domain-specific insights. In this paper, we introduce RMultiplex200K, a multimodal dataset designed to present step-wise reasoning rationales and correctness scores for real-world Telecom questions. This enables VLMs to engage in step-level reasoning and verification using multimodal information, thereby facilitating reliable problem-solving. RMultiplex200K is highly scalable as it is constructed without human annotations, relying instead on our automatic plan-based annotation (ApPA) method, which automatically synthesizes reasoning steps labeled with reward scores. With this dataset, we introduce TC-NAVIGATOR, a new mechanism for training multimodal process reward models to serve as reliable reasoning verifiers for VLMs. For instance, the Qwen-2-VL-72B and Llama-3.2-90B models, which initially achieve only $21.3\%$ and $19.8\%$ respectively on practice Telecom questions, reached $48.5\%$ and $46.1\%$ accuracy, respectively, after training with RMultiplex200K and verifying with TC-NAVIGATOR. + +# 1. Introduction + +Visual Language Models (VLMs) can perform step-by-step reasoning to solve complex problems across a wide range of tasks. By leveraging datasets that capture this step-wise reasoning, VLMs have achieved markedly improved performance in fields such as physics [19], medicine [36], and particularly mathematics [44, 46, 58]. However, relatively little work has been done to develop these datasets and validate the effectiveness of VLMs in the critical domain of telecommunications (Telecom) [65], which is characterized by domain-specific terminologies, concepts, and reasoning + +![](images/92147a2e205df683dcc8a191ee21eb2cbcdeda88e24bf9c85c37201e44b7f845.jpg) +Figure 1. Illustration of a solution sample of the RMultiplex200K dataset. It contains a question, an image, and a step-by-step reasoning process in italics, along with the steps and rewards generated by ApPA highlighted in a shadow box. + +# patterns. + +Existing literature [5, 34, 42] typically employs only textual datasets [20, 33, 37]. However, omitting visual information limits the applicability of VLMs, particularly because problem-solving in Telecom often relies on interpreting images such as circuit diagrams and waveform plots. Although more recent works [8, 51, 59] incorporate visual data into their datasets, they fail to provide answers with the necessary step-wise rationales and human-level correctness scores, which are critical to the success of large language foundation models (LLMs), as demonstrated in mathematics [24, 46]. + +This paper introduces RMultiplex200K, the first multimodal dataset built for reliable process supervision, which support the training of VLMs toward complex reasoning ability on real-world Telecom problems. Compared with existing Telecom datasets, mathematics datasets (e.g., MATHSHEPHERD [46] and PRM800K [24]), and even the mul + +![](images/b037d6d8bde4fe18c0bbc2bf9711b8038e8743d4ef4ca89a46264447cba46ae1.jpg) +Figure 2. Performance of advanced LLMs and VLMs using Llama-3.1-70B (LLM) and Qwen2-VL-72B (VLM) as reward models for verification. With the verifier TC-NAVIGATOR trained as a PRM (i.e., TC), the problem-solving rates of these models are significantly enhanced. The "Fine-tuned" refers to directly fine-tuning the large models on the RMultiplex200K dataset. "ORM" denotes the outcome reward model. The self-consistency (SC) method gets a solution by majority voting across 100 solutions. + +timodal, multidisciplinary dataset MMMU [59], $RMultiplex200K$ provides step-level reasoning rationales and reliable correctness scores, thereby significantly broadening the applications of VLMs. We systematically examine the primary components of our contributions and demonstrate that: + +- Proposing automatic plan-based process annotation (ApPA) mechanism, designed on Monte Carlo Tree Search (MCTS), which automatically synthesizes answers with detailed step-by-step reasoning descriptions and corresponding correctness scores. +- Organizing a manually-free and scalable RMultiplex200K dataset composed of high-quality samples derived from Telecom problems. As shown in Fig. 1, each sample contains a step-level reasoning process as the answer along with human-level correctness scores. +- Introducing a novel mechanism, TC-NAVIGATOR, to support the training of multimodal process reward models (PRMs) that dynamically adjust visual attention for reliable step-wise verification. + +By combining these outcomes, we enable a series of VLMs that originally fail on Telecom problems to achieve a new state-of-the-art (SOTA) benchmark on the RMultiplex200K testbed, as illustrated in Fig. 2. Specifically, the accuracy of Llama-3.2-90B, initially only $19.8\%$ increases to $41.8\%$ with supervised fine-tuning on our dataset. With PRMs obtained via TC-NAVIGATOR using Proximal Policy Optimization (PPO), its accuracy further improves to $46.1\%$ . More importantly, we observe that the trained PRMs are able to adaptively adjust their attention across image regions at each reasoning step to enable reliable ver + +ification, thereby paving the way for further research on the interpretability and modeling of such step-wise attention adjustments. + +# 2. Related Work + +Visual language models (VLMs) [2, 4, 27, 28, 45] significantly extend the capabilities of large language models (LLMs) by incorporating visual information — typically images — into the step-by-step reasoning process for problem solving. Datasets for evaluating [22, 31] and finetuning [39, 63] VLMs typically consist of a textual question and an accompanying image as input, paired with a golden multi-step reasoning answer. In mathematics, fine-tuning VLMs on datasets like GeoGPT4V [7] and MathV360K [41] has enabled advanced multimodal reasoning. Likewise, datasets curated for physics [16, 19], medicine [26, 36], and biology [29, 62] have been used to adapt VLMs to broader and more specialized domains. + +These Apart from fine-tuning models on step-wise reasoning datasets, recent advancements in enhancing model reliability primarily focus on LLMs. These advancements include the Outcome Reward Model (ORM) [11, 57], which is trained to assign a correctness score to the final outcome of model responses, and the Process Reward Model (PRM) [23, 24, 30, 46], which is trained to assign correctness scores to each reasoning step of model responses. However, as demonstrated in [24, 30, 46], preparing datasets with step-wise scores is both challenging and resource-intensive. Moreover, advanced automatic labeling methods [30, 46], such as MATH-SHEPHERD [46], do not perform well on Telecom problems due to the scarcity of useful synthetic samples. + +Some efforts [5, 42] have discussed the application of LLMs in the frontier science domain of Telecom, and several works [25, 34, 54, 66] have proposed training telecom-specific LLMs using small-scale, collected datasets. Yet, these datasets — including benchmark datasets such as Spec5G [20], NetEval [37], and TeleQnA [33] — contain only textual samples. By incorporating visual information, VLMs can be trained on multimodal datasets to enhance performance in solving engineering [59], 6G [8] and networking [51] problems. However, unlike in mathematics, where significant innovations have been achieved, no existing work in Telecom emphasizes process supervision in fine-tuning VLMs to develop reliable step-by-step reasoning abilities for practical problem-solving. + +# 3. RMultiplex200K without Human Annotation + +# 3.1. Dataset Overview + +RMultiplex200K consists of 200000 samples with step-level correctness labels from multimodal problems, categorized into five domains: Wireless Communication (WC), Satel- + +![](images/cd70ad23061f95573d989e307b0391f4be7a94cd24db89c51d11aea36592620e.jpg) +(a) Reasoning Steps + +![](images/6c5c65c94a8f6393dd937cd17957dda5a736a48765d8355c950a28591046b2e9.jpg) +(b) $ApPA$ + +![](images/a077a967c8beefb165ec7ee276c12536221735d808717991a27e154145dbf5aa.jpg) +(c) TC-SHEPHERD +Figure 3. Illustration of details of the $RMultiplex200K$ dataset. (a) presents the number of reasoning steps required for a solution across the samples of five categories, while (b)(c) illustrate of the reward label distribution comparison between our $RMultiplex200K$ dataset and TC-SHEPHERD, which is generated by applying MATH-SHEPHERD [46] to our Telecom problems. + +lite Communication (SC), Network Theory and Optimization (NTO), Information Security and Encryption (ISE), and Communication Signal Processing (CSP), with category abbreviations provided in parentheses. + +Each problem in the dataset includes a textual question, presented as a sequence of words denoted by Q, along with a visual image denoted by V. The gold answer for each problem is represented by $z_{1\dots n}$ , which comprises $n$ reasoning steps, each represented as a sequence of words $z_{i}$ for $i \in [1,n)$ , leading to the final solution $z_{n}$ . The ground-truth solution contained in $z_{n}$ is denoted by $y$ , which presents a textual answer, such as "36 dB/K". To support the process supervision, each reasoning step $z_{i}$ is associated with a correctness score $c_{i} \in [0,1]$ , where $c_{i} = 1$ indicates a correct reasoning step, and $c_{i} = 0$ indicates an incorrect one. Without losing generality, we follow the terminology in PRM800K [24] to denote a sample as $s = (Q,V,A,z,c,y)$ while a solution sample as $\mathbb{S} = \left(Q,V,A,\left\{z_{1\dots n}^{k},c_{1\dots n}^{k},y^{k}\right\}_{k = 1}^{K}\right)$ , where A is the answer, $c_{1\dots n}$ represents the scores for the $n$ reasoning steps and $K$ is the number of reasoning paths generated by our annotation method, ApPA Sec. 3.2. + +Detailed data information is presented in Fig. 3. For instance, the five categories' average number of reasoning steps is 9, 8, 11, 9, and 12, respectively. + +# 3.2. Automatic Plan-based Process Annotation + +Telecom problems involve numerous complex, domain-specific concepts and reasoning mechanisms. Thus, unlike in mathematics, understanding the answer requires a high level of expertise, making it nearly impossible for human annotation. Besides, the current SOTA annotation method, MATH-SHEPHERD [46], fails to generate a useful Telecom dataset because it yields few positive samples (correctness scores greater than 0.5), as shown in Fig. 3(c). Thus, we propose an automatic plan-based process annotation (ApPA) mechanism comprising three stages for creating multimodal samples with process supervision, as illustrated in Fig. 4, thereby yielding balanced scores, as shown in Fig. 3(b). + +Data collection. We use Mathpix [35] to extract paired + +questions, answers, and visual images from various open sources, including Telecom textbooks, research papers, online resources, and quizzes, into LaTeX format, resulting in raw problems consisting of Q, A, and V. + +Answer decomposition. To decompose the answer A, which contains the problem-solving process, into sequential reasoning steps, we apply an advanced VLM denoted as $p_{\theta}$ , such GPT-4o, based on the idea of zero-shot prompting [21]. Specifically, we design the prompt $\mathbb{I}_1(\mathrm{Q},\mathrm{A},E)$ where $E$ is the human-made instruction. Then, we prompt the VLM to generate the reasoning steps, formulated as $z_{1\dots n}\sim p_{\theta}(z_{1\dots n}|\mathbb{I}_1(\mathrm{Q},\mathrm{A},E))$ . Both the data creation and evaluation phases rely on this approach to decompose answers into individual reasoning steps. The detailed prompt design is provided in the Appendix for reference + +Annotation assignment. Existing automatic annotation methods for process supervision (for example, MATHSHEPHERD [46] and AutoPSV [30]) are not well suited for Telecom problems, as LLMs frequently generate incorrect reasoning paths that result in sparse positive samples. To address this issue without introducing additional finetuning, our annotation process is based on the principle that, like humans, VLMs potentially follow a plan - a high-level, generalized logic - as an inherent guide for generating each reasoning step. Thus, we introduce ApPA, a plan-based MCTS mechanism, as outlined below. + +- plan generation. For each reasoning step $\mathbf{z}_i \in \mathbf{z}_{1\dots n}$ , we prompt the VLM with the instruction $\mathbb{I}_2(\mathrm{Q},\mathrm{V},\mathbf{z}_{1\dots i},S)$ to summarize the plan $\psi_i$ from it, where $S$ is the summarization guidance. This is formulated as $\psi_i \sim p_\theta (\psi_i|\mathbb{I}_2)$ . Thus, we obtain $\psi_{1\dots n}$ , which presents a chain of plans toward guiding the VLM to generate $\mathbf{z}_{1\dots n}$ . +- reasoning creation. For $(\mathrm{Q},\mathrm{V},z_{1\dots n})$ , we use the Monte Carlo Tree Search (MCTS) to create various reasoning paths as a tree structure, where each node represents a reasoning step. Specifically, Selection stage begins at the root node $(\mathrm{Q},\mathrm{V})$ and iteratively selects child nodes based on UCT (Upper Confidence bound applied to Trees) score until it reaches a leaf node. The Expansion stage expands the tree from this node (Step denoted as $z_i^{\prime}$ ) by prompting the VLM with the generated plan $\psi_{i + 1}$ to deduce one next reasoning step $z_{i + 1}^{\prime}$ , which is formulated as $z_{i + 1}^{\prime}\sim p_{\theta}\left(z_{i + 1}^{\prime}|\mathbb{I}_{3}\left(\mathrm{Q},\mathrm{V},z_{1\dots i}^{\prime},\psi_{i + 1}\right)\right)$ . Thus, the Rollout stage completes the reasoning to obtain the solution $y^\prime$ by using the VLM to decode subsequent reasoning steps $z_{i + 1\dots n}^{\prime}$ , where $z_{j + 1}^{\prime}\sim p_{\theta}\left(z_{j + 1}^{\prime}|\mathbb{I}_{3}\left(\mathrm{Q},\mathrm{V},z_{1\dots j}^{\prime},\psi_{j + 1}\right)\right)$ for $j\in i + 1\dots n$ +- label assignment. The Backpropagation stage computes $\mathbf{1}(y', y)$ to yield the reasoning result, which is used to update the score of each node in the reasoning path $z_{1\dots n}^{\prime}$ . For each question, we set the tree width to $M$ and perform $T$ iterations of MCTS, assigning the final score of each node as its label $c \in [0, 1]$ . + +![](images/e80e1b566bf726f71ac617868b06a7346d3e992b09a6aa252254a34ee270ba0c.jpg) +Figure 4. Illustration of the automatic plan-based process annotation (ApPA) mechanism. Given the extracted question, answer, and image as raw data from the SC category, ApPA first decomposes the answer into step-by-step reasoning and summarizes the plans. It then performs MCTS to explore step-wise reasoning samples and assigns rewards to them. During this reasoning creation process, an VLM, such as GPT-4o, is prompted by the summarized plans to generate the reasoning steps, shown in the small rectangular boxes along the lines, such as the one from Q, V to $z_{1,2}$ . The orange box indicates that the reasoning is complete and the result has been obtained. + +![](images/a103e00fa02aa4e548d7005542d8d854e3597089f35130ae7469d7e65bc472fc.jpg) +Figure 5. Illustration of the highly abstract structure of TC-NAVIGATOR, built upon QA-ViT [13]. It comprises RA1, shown in the upper part, which generates $F_{RV}^{\prime}$ , and RA2, depicted in the lower part, which generates $F_{RV}$ . + +- sample selection. Performing $T$ iterations of MCTS for each question results in a tree with $M^n$ paths, each corresponding to a reasoning path containing $z_{1\dots n}^{\prime}$ . We then obtain samples by filtering out (1) False Positives: paths where each step has a high score but leads to an incorrect solution, and (2) False Negatives: paths where any intermediate step has a low score but results in a correct solution. After filtering, the number of samples for this question is denoted as $K$ . + +Therefore, eventually, $ApPA$ generates multimodal samples with process supervision for each question, which can be formulated as a solution sample $\mathbb{S}$ , in which $y^{k}$ is the solution extracted from $z^{k}$ and $K$ is not a constant value but varies across questions. + +# 4. TC-NAVIGATOR with Step-level Attention + +With RMultiplex200K, we introduce TC-NAVIGATOR, a mechanism to train multimodal process reward model to en + +sure reliable step-by-step reasoning with VLMs in Telecom. It is based on the intuition that an expert focuses on specific regions of the science image for in-depth analysis, guided by existing reasoning steps, before verifying the correctness of each step. The model implements this human-like behavior by including two reasoning-aware parts including RA1 and RA2, as shown in Fig. 5. + +RA1. The design largely draws from the proven QA-ViT framework [13], with modifications for step-level reasoning. RA1 uses step-level features by first projecting Q and $z_{1\dots i}$ into embeddings, denoted as $f_{\mathrm{Q}} \in R^{d}$ and $f_{z_{1\dots i}} \in R^{i\times d}$ where $d$ is the dimension size. This reduces computational cost when there are many steps and allows the model to focus on the most relevant ones. For language features and the visual feature $f_{V} \in R^{A\times d}$ for $A$ patches, we employ the fusion mechanism of QA-ViT [13] to derive the fused visual feature $F_{RV}'$ , which is a step-aware visual feature formulated as $F_{RV}' = \zeta(f_{RV}') + \zeta_g(f_{RV}') \cdot \tanh(\beta)$ . Here, $\beta$ is a learnable parameter initialized to zero, $\zeta$ is the projection, $\zeta_g$ is the gated projection, and $f_{RV}' = Self\_Attn((f_{\mathrm{Q}}, f_{z_{1\dots i}}), f_{V}))_{[0:A]}$ . + +RA2. We encode the Q and $\boldsymbol{z}_i$ with a pre-trained model, such as a LLM, to obtain token-level features, denoted as $F_{\mathrm{Q}} \in R^{|Q| \times d}$ and $F_{\boldsymbol{z}_i} \in R^{|z_i| \times d}$ , respectively. We then apply the QA-ViT [13] to derive $F_{RV}$ , following the formulation in RA1. Subsequently, as a high-level demonstration, the combined terms are processed by an LLM to generate the hidden state of the final token, which is then projected through a linear layer to produce the score $\widetilde{c}_i$ . + +Training. Using RMultiplex200K and TC-NAVIGATOR, we can train a verifier as either an outcome reward model (ORM) or a process reward model (PRM). The ORM is typically trained with a cross-entropy loss defined as $\mathcal{L}_O = c_n\log c_n' + (1 - c_n)\log (1 - c_n')$ , where $c_{n}^{\prime}$ is the predicted sigmoid score of the final step $n$ . The loss of PRM is $\sum_{i = 1}^{n}c_{i}\log c_{i}^{\prime} + (1 - c_{i})\log (1 - c_{i}^{\prime})$ , where $i\in [1,n]$ . More generally, TC-NAVIGATOR can also reinforce VLMs + +Table 1. Comparison of the performance of various VLMs and LLMs on the RMultiplex200K testbed. The models are evaluated under different settings: (1) Prompting methods, such as ZeroshotCoT [21] with "let's think step by step" and CoT prompting [49] with 5 shots, (2) Fine-Tune, and (3) Fine-Tune + Verification, where the reward models for the verification are Llama-3.1-70B (LLM) and Qwen2-VL72B (VLM). For self-consistency (SC) [47], the accuracy is reported as the majority voting of 100 generations. For our TC-NAVIGATOR (TC) and ORM, accuracy is computed on a best-of-100 basis, meaning the answer with the highest predicted score among 100 candidates is selected. The unit is percent (\%) and the best results are highlighted in bold. Note that the GPT-\* series of models are not fine-tuned but are instead prompted directly, even though they may appear in the third column. + +
PromptingFine-TuneFine-Tune + Verification
ZeroshotCoTCoT_5SCZeroshotCoTCoT_5SCORMORM+SCTCTC+SC
MultiModal Large Language Models (VLMs): Text and Image as the input
OpenFlamingo2-9B11.512.112.321.223.324.525.827.12627.5
LLaMA-Adapter2-7B11.912.712.720.121.722.624.525.92526.6
MiniGPT4-Vicuna-13B12.412.612.925.227.129.331.333.532.833.9
InstructBLIP-T5-XXL13.314.514.624.225.927.228.930.429.631
Qwen-VL-7B-Chat13.714.214.528.930.131.832.933.93334.7
LlaVA-NEXT-34B15.217.317.929.832.234.236.13837.338.5
LlaVA-OneVision-7B15.715.915.929.431.832.533.635.534.336.5
LlaVA-NEXT-72B16.018.519.53031.934.536.939.138.540
Llama-3.2-11B16.116.417.331.432.834.235.737.236.437.8
LlaVA-1.6-34B*16.417.718.131.333.535.937.539.638.940.9
Qwen-VL-MAX*17.719.420.933.435.837.239.140.739.941.2
LlaVA-OneVision-72B18.320.321.835.137.739.941.844.243.745.6
InternVL-Chat-V1.2*18.418.919.332.934.136.237.939.63940.4
Llama-3.2-90B19.821.123.734.336.839.241.844.943.746.1
Qwen-2-VL-72B21.323.525.136.839.442.144.547.446.848.5
GPT-4o mini22.425.227.9---40.345.645.146.2
Large Language Models (LLMs): Only Text as Input
Llama-2-7B8.18.48.411.512.212.913.914.11414.7
Qwen-14B10.311.611.916.317.818.42020.720.321.4
Vicuna-13B7.48.58.511.612.313.414.815.11515.8
Llama-3-8B5.76.36.79.49.710.211.111.511.411.9
Llama-3-70B11.212.112.815.917.519.721.622.822.123.7
Llama-3.1-8B11.81212.416.817.618.319.220.119.620.9
Llama-3.1-70B14.616.117.118.220.123.425.827.22728.4
GPT-410.213.915.2---19.721.821.722.2
+ +through step-by-step Proximal Policy Optimization (PPO). See more details in the Appendix. + +# 5. Experiments + +Dataset. We organize $RMultiplex200K$ by applying $ApPA$ to 5000 multimodal problems from WC (1300), SC (600), NTO (1400), ISE (650), and CSP (1050). On average, each problem involves 10 reasoning steps. For every problem, we extract $K = 15$ reasoning paths from the MCTS tree and, as detailed in Sec. 3.2, select the best paths while eliminating duplicates, yielding roughly 200,000 labeled samples. We have a testbed of 2000 problems for evaluation. To ensure reliability, the GPT-4o [1] model is employed within $ApPA$ . For comparison, we generate labeled samples for Telecom problems $(Q, A, V)$ using the state-of-the-art mathematics annotation method, MATH-SHEPHERD [46] (referred to as TC-SHEPHERD in Fig. 7b). + +Models. For verification, TC-NAVIGATOR is used to train verifiers by fine-tuning on Llama-3.2-3B, Llama3.1-8B, and Llama-3.1-70B under the text-only condition, + +while Qwen-VL-7B, Llama-3.2-11B, and Qwen2-VL-72B are used under the text-visual condition. We primarily use Llama-3.1-70B and Qwen2-VL-72B as verifiers for advanced LLMs and VLMs, respectively. The remaining models are used for comparison purposes, as shown in Fig. 7a. For broader applicability, TC-NAVIGATOR is also used to reinforce LLMs and VLMs through step-by-step Proximal Policy Optimization (PPO). Due to resource constraints, we conduct this reinforcement learning only for training Llama-3.1-8B and Llama-3.2-11B, as presented in Tab. 2. Throughout the experiments, when TC-NAVIGATOR is applied to train the PRM, we abbreviate it as TC; otherwise, we refer to it as ORM. + +Training. To fine-tune LLMs toward reward models for verification, we follow the MATH-SHEPHERD framework [46]. Specifically, reward models are optimized over one training epoch, using a learning rate of $3 \times 10^{-4}$ for 3B, 8B, and 70B LLMs. For VLMs, we use a learning rate of $2 \times 10^{-4}$ for 11B models and $1 \times 10^{-6}$ for the 72B model. Toward reinforce VLMs with reinforcement learn- + +Table 2. Comparison of the performance of various reinforcement learning and verification strategies on the RMultiplex200K testbed. The verifiers in "with Verifier" are Llama-3.1-70B for LLM and Qwen2-VL-72B for VLM. Accuracy is measured in percent (\%) and, for SC, is reported as the majority vote from 100 generated solutions. For example, "+"ORM+SC" refers to a reinforced model paired with the verifier TC-NAVIGATOR (TC) Llama-3.1-70B to generate the solution by voting among 100 verified solutions. + +
ModelsCategoryOverall 2000/10WC 520/9SC 240/8NTO 560/11ISE 260/9CSP 420/12
3.2-11Bwith PPOReinforce LLMs with step-by-step PPO
ORM3435.836.732.534.631.9
TC3737.938.935.837.536.2
TC w/o RA136.337.237.735.336.435.8
TC w/o RA235.936.538.534.237.234.6
with VerifierCombine step-by-step PPO and the verifier (+*)
+ORM40.341.343.338.842.338.1
+ORM+SC42.142.744.241.143.140.7
+TC4141.943.839.542.739
+TC+SC43.343.844.642.544.242.4
Large Language Models (LLMs): Only Text as Input
3.1-8BPPO - ORM12.513.614.211.813.110.7
PPO - TC13.814.4151314.612.9
+ +ing, we train the PRMs with PPO for one epoch at a learning rate of $4 \times 10^{-7}$ . For a series of VLMs and LLMs as presented in Tab. 1, we fine-tune these models directly on the partial problem Q, $z_{1\dots n}$ , V from the training set of RMultiplex200K to enable learning from our dataset. We implement a cosine learning rate scheduler, with a minimum learning rate of $1 \times 10^{-8}$ . All models are trained with a maximum sequence length of 512 on the "unsloth" platform *, using QLoRA under four NVIDIA A100 GPUs, each equipped with 80 GB of VRAM. + +Baselines and Metrics. Following the MATHSHEPHERD framework [46], we evaluate the trained verifiers by comparing them against the Self-consistency (SC, majority voting) [47] and other baseline methods, such as Zeroshot [21] and CoT [49]. With SC, we generate many solutions and use voting among the best-of-100 solutions to obtain the final answer. Otherwise, we directly select the solution with the correct scores for reasoning steps. We measure accuracy, defined as the problem-solving rate, as the evaluation metric. + +# 5.1. Main Results + +Table 3. Cross-data Evaluation. After fine-tuning on $RMultiplex200K$ , DeepSeek-R1 (LLM) is evaluated on the TeleQnA, MMLU, and MATH datasets, while Qwen-2-VL-72B (VLM) is tested on TeleQnA, MMMU, and MATH-Vision. The 'Base' setting uses the original, unfine-tuned model, whereas the 'PR' setting employs the same model with zero-shot prompting. + +
MethodsTeleQnAMMLU / MMMUMATH / MATH-Vision
BasePRORMTCBasePRORMTCBasePRORMTC
DeepSeek-R189.891.792.194.490.891.292.894.197.297.497.898.7
Qwen-2-VL-72B66.371.774.476.864.564.865.466.422.323.725.427.1
+ +We present a comprehensive evaluation of various LLMs + +and VLMs on the RMultiplex200K testbed under different settings, as shown in Tab. 1, where the verifier is trained with TC-NAVIGATOR. Using zero-shot and CoT prompting methods, we find that even the latest large models struggle to solve scientific Telecom problems. Most models achieve around $20\%$ accuracy, which is impractical for real-world applications; for example, Llama-3.2-11B attains only $16.1\%$ accuracy, and GPT-4o reaches $22.4\%$ in zero-shot prompting. Even with CoT prompting (using five demonstration question-answer pairs) and self-consistency [47], VLM accuracy improves by only about $2\%$ , while text-only LLMs remain below $12\%$ . These results underscore that current VLMs are far from practical for Telecom, highlighting the importance of our released dataset. + +Subsequently, fine-tuning large models on specific scientific domains, such as Telecom, is necessary to enhance their practicality. For example, fine-tuning on the RMultiplex200K dataset significantly improves model performance. The average accuracy improvements for these models from prompting to fine-tuning in ZeroshotCoT, CoT_5, and SC are $14.06\%$ , $14.79\%$ , and $15.82\%$ , respectively. Additionally, we observe that large models fine-tuned on RMultiplex200K show greater improvements when using CoT and SC, with Qwen-2-VL-72B achieving a $5.3\%$ increase using SC. This improvement is attributed to the detailed step-by-step reasoning process provided by the answers in the our dataset, which enables the models to learn how to perform or follow reasoning steps toward an accurate solution. + +Tab. 1's third column shows that integrating large models with reward models, particularly those employing process supervision, significantly enhances Telecom problem-solving reliability. With verification, average accuracy increases from the best Fine-tune with SC to ORM, SC+ORM, and TC-NAVIGATOR by $1.85\%$ , $3.88\%$ , and $2.87\%$ , respectively. In particular, TC-NAVIGATOR improves accuracy by $4.58\%$ , enabling advanced models such as Llama-3.2-90B and Qwen-2-VL-72B to achieve $46.1\%$ and $48.5\%$ accuracy, with similar gains for GPT-4V(ison), GPT-4o mini, and other text-only LLMs. Notably, our findings extend previous conclusions [24, 46] by showing that, even for VLMs, process supervision outperforms outcome supervision, making it essential for reliable Telecom problem-solving. + +Fig. 6 shows the relationship between the number of reasoning steps needed for a solution and the performance of different verifiers: the outcome reward model (ORM), the process reward model (PRM), and ORM combined with self-consistency [47] (ORM+SC). As the number of steps increases, PRM performance remains stable while ORM's performance drops sharply. Thus, for more challenging problems that require more reasoning steps, PRM becomes essential and outperforms ORM, underscoring the need for + +![](images/49e609ed484aff4c2d23fd4292fa178903fde48c56cca0a6bdaae2b24a67a4bd.jpg) + +![](images/700748acc052ec5ed9c71762b6ac7c736b440b43ca805e354bb1bbbe9525a791.jpg) +Figure 6. Illustration of how the performance of different types of verifiers varies with the number of steps involved in solving the problem. In this analysis, Qwen-2-VL-72B is used as the verifier, while Llama-3.2-90B is used to generate reasoning steps to address the given problem. +(a) +Figure 7. Illustration of the ablation study: (a) shows how the number of generated reasoning solutions influences the performance of TC across different LLMs and VLMs. (b) presents the performance gap of TC with different VLMs trained on our $RMultiplex200K$ and TC-SHEPHERD. For both results, we use large models with zero-shot CoT as the baseline, indicated by the gray lines. + +![](images/8eb33d3346b6c3662cd1994579eaecb7dab9d81b032400464475fc297aea905b.jpg) +(b) + +process supervision. + +Tab. 3 presents cross-data evaluation results, showing the performance of LLMs and VLMs on external datasets when verified by verifiers trained on our dataset in ORM and TC modes. DeepSeek-R1 and Qwen-2-VL-72B exhibit significant accuracy improvements, demonstrating that the verification capabilities learned from our dataset generalize well to other datasets. + +# 5.2. Reinforce Large Models + +We present the performance of Llama-3.1-8B (LLM) and Llama-3.2-11B (VLM) generators when TC-NAVIGATOR (TC) is employed as the reward model for supervision, reinforcing them with step-by-step PPO. We also combine the reinforced models with the verifier to further enhance the reliability of problem-solving, as summarized in Tab. 2. + +Compared to Llama-3.2-11B with zero-shot CoT [21], which achieves $31.4\%$ accuracy, reinforcing it with ORM + +![](images/12e02bb687f47cddf5c81b062b8a71764e2dd21d15bf442a01ae2157ec62b959.jpg) +(a) + +![](images/9795eeec8517b188cd02d96e996e9e831464510eb65f92f3e679e3b2d0216cc9.jpg) +(b) +Figure 8. Illustration of the ablation study for RA1 and RA2 of the verifier: (a) and (b) depict the performance drop of Qwen-VL-7B-Chat and Qwen-2-VL-72B when switching from the full TCNAVIGATOR with Llama-3.2-11B to the Llama-3.2-11B under different conditions: without RA1 (w/o RA1), without RA2 (w/o RA2), and without both RA1 and RA2 (w/o RA12). + +yields a $2.6\%$ improvement. Using TC for step-wise supervision further increases accuracy to $37\%$ , compared to the $34.2\%$ achieved by the fine-tuned Llama-3.2-11B with SC, as shown in Tab. 1. This demonstrates that reinforcing VLMs with TC-NAVIGATOR significantly enhances their reasoning ability. A similar trend is observed for Llama-3.1-8B. + +Combining the reinforced VLMs with the verifier sets a new SOTA for Llama-3.2-11B. With ORM, accuracy increases to $40.3\%$ , and $\mathrm{ORM + SC}$ further boosts it to $42.1\%$ . Using TC and $\mathrm{TC + SC}$ as the verifier, accuracy reaches $41\%$ and $43.3\%$ , respectively, approaching the $43.7\%$ achieved by the advanced Qwen2-VL-72B with Qwen2-VL-72B as the verifier. These results extend the conclusions from the LLM work [24, 46] to VLMs, demonstrating that combining reinforcement learning with verification can make VLMs more reliable for problem-solving in Telecom. + +Finally, compared to ORM, TC and $\mathrm{TC + SC}$ , which perform process supervision, achieve higher accuracy in the Telecom category, where problems involve more reasoning steps to solve. For example, in the CSP category, which has an average of 12 steps, accuracy increases by $4.3\%$ when switching from ORM to TC. In contrast, the increase in the SC category, which requires an average of 8 steps, is only $2.2\%$ . This extends the insight from [24], further demonstrating that process supervision not only surpasses outcome supervision but significantly outperforms it when problems require more reasoning steps. + +# 5.3. More Analysis + +First, TC verifiers trained with larger VLMs more effectively boost generator accuracy as the number of generated solutions increases. As shown in Fig. 7, Qwen2-VL-72B's accuracy nearly reaches its optimum (achieved with 100 solutions) when using 64 solutions, whereas Llama-3.2-11B and Qwen-VL-7B-Chat remain suboptimal even with 64. A similar trend is observed for TC verifiers trained on these models. These results confirm that using larger VLMs for + +![](images/99bfeac6ce7b1d05816ef83e865e72e8a660a39c7455646404b727914dda51d2.jpg) + +![](images/ea1c53c33b0ed1fff34159ef880668e312522267d4e4b5c8b3a43c4a2983bf56.jpg) +0.79 + +![](images/56a42ba5086ad9e3c73a6bcfc0deeaf2b68123091f559f4c30be804dbef6128d.jpg) + +![](images/97e1c0d12b646355931610504f474c9093ae58547f7637186dc7d1639c8a1ea7.jpg) + +![](images/4709ff1d81cc78d6a6e104b6d455bbe528cd2b349003a48b27ec3778a0625c98.jpg) +0.92 + +![](images/de6ea5e5df1ededa3552a3fa0b23dc41f8ae0983f096011b846e0f22b80dacc4.jpg) + +![](images/2f4bc9ce8601efcf946af07c1c1c98263b92a6d4f0dcf6c99bd84abbbc231d65.jpg) +0.99 + +![](images/80a134e2376f6585ced1628fd5b3e66d6fb375149f3b2e3fc42d3361c97bea75.jpg) + +![](images/f2daf3f88d3ac13dce3457a21f2ca98534b90130ddc44622e35a748fb3e178bc.jpg) + +![](images/0053665418f0586ac9ee800fe2db1dd9b00e3f54d698ea310bbbca20ec61015d.jpg) + +![](images/50ad58744ca52f4b7f4fce716fd88ccb23317f8ec363865c3b77f58e2bf44ece.jpg) +Figure 9. Illustration of the reasoning process generated by Llama-3.2-90B for answering a CSP question, with Qwen-2-VL-72B (QW) serving as the verifier (bold numbers represent predicted rewards). $Q$ : Compute the power spectral density of a random binary wave shown in Figure 4.7, with symbols $+A$ and $0$ occurring with equal probability. Assume that the symbols are independent. + +![](images/b4ef5b21cb3f1f2556c926291db0515fc5e9488a1f8837385d624133db919994.jpg) + +![](images/24dd3b42a79bb42d9f80bd128035521d326feb8f387974727b417424daf136a1.jpg) +0.91 + +![](images/a8e485858f121c933c248ed5094dd38073816ed9a0ced8df47bd8ba22cb98180.jpg) +Figure 10. Illustration following Fig. 9. $Q$ : Consider the system in Figure. The input $w_{n}$ is a real-valued, zero-mean white noise process with autocorrelation defined as: $R_{ww,m} = E[w_n w_{n - m}] = 2\delta_K(m)$ where $\delta_K(\cdot)$ is the Kronecker delta function. The impulse response of the filter is given by: $h_n = \delta_K(n) + 2\delta_K(n - 1) + 3\delta_K(n - 2)$ Using the Levinson-Durbin algorithm, compute the coefficients of the optimum $3^{rd}$ -order. + +![](images/8a4f0e7088cf7ae0864b91bc5fa79ae2bf59f84d92e1f6f774351cbb7732f9a3.jpg) +0.39 +0.27 + +![](images/9f62b8ba2aa7aa21b30434ba81b64c688c99ae7208519b7feb33c44016e9a2cb.jpg) +0.63 + +training TC verifiers is beneficial and that selecting the best-of-100 solutions is sufficient for Telecom problems. + +Second, we train the ORM/TC verifier using Qwen2-VL-72B as the base model on a new dataset, TC-SHEPHERD, which leverages annotations generated by the SOTA mathematics annotation method [46]. However, compared with the results in Tab. 1, the performance of models such as Qwen-VL-7B-Chat, Qwen-2-VL-72B, and Llama-3.2-11B declines significantly when the verifier is trained on TC-SHEPHERD. For instance, Qwen-2-VL-72B's accuracy decreases by over $6\%$ across all four verifier cases, and Qwen-VL-7B-Chat performs even worse than its zero-shot CoT reasoning on ORM, TC, and $\mathrm{TC + SC}$ . Moreover, larger models suffer an even greater drop. These results indicate that the annotation mechanism effective in mathematics does not transfer well to Telecom, underscoring the necessity of our domain-specific ApPA. + +Third, our ablation study (Tab. 2 and Fig. 8) shows that: (1) a weaker VLM (e.g., Llama-3.2-11B) fails to effectively verify solutions from a weaker generator (Qwen-VL-7B-Chat) when either RA1 or RA2 is removed; and (2) RA1 is more critical for ORM while RA2 has a greater influence on PRM. Specifically, Fig. 8 indicates that the accuracy drop for ORM under "w/o RA1" is particularly high, and under "w/o RA2" the drop in TC-related terms generally increases. In Tab. 2, removing RA2 from PRM lowers accuracy to 35.9, which is $0.4\%$ lower than when RA1 is removed. Moreover, a stronger generator such as Qwen-2-VL-72B is largely unaffected by verifier changes (Fig. 8b). + +Fourth, Fig. 9 and Fig. 9 illustrate two examples of how the verifier trained with TC-NAVIGATOR assigns correctness scores to each reasoning step and adjusts the corre + +sponding attention maps during the reasoning process. In Fig. 9, Qwen-2-VL-72B correctly verifies a solution, as indicated by high scores and a check mark. However, in another example, the model mistakenly assigns a high score to an incorrect solution, even though it identifies errors in Step 2 and Step 4 (shown by red, low scores). More importantly, to verify different reasoning steps, the verifier assigns varying attention to different image regions, mimicking human-level intelligence by adaptively adjusting its focus for reliable judgment ("see closer to see better"). Although these attention maps improve the interpretability of verification, one issue remains: understanding the underlying logic or rules that VLMs follow to adjust their attention. + +# 6. Concluding Remarks + +In this paper, we introduced RMultiplex200K, the first multimodal dataset created to support step-wise supervision for reliable problem-solving by VLMs in the Telecom domain. To enable data creation without human annotations, we proposed an automatic plan-based process annotation method, ApPA, which involves a Monte Carlo Tree Search to assign a correctness score to each reasoning step. Based on RMultiplex200K, we introduce TC-NAVIGATOR, a novel mechanism for training multimodal process reward models that either verify VLMs' solutions or directly reinforce them via step-by-step Proximal Policy Optimization (PPO). Extensive experiments demonstrate the necessity of our dataset and the effectiveness of our method. + +Acknowledgements. This research was supported by the Guangzhou-HKUST(GZ) Joint Funding Program (No. 2024A03J0630) and National Natural Science Foundation of China (No.62372357). Sijia Chen was supported by the + +Guangzhou Municipal Key Laboratory of Financial Technology Cutting-Edge Research, and Bin Song was supported by the State Key Laboratory of Integrated Services Networks. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5, 13 +[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] 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. 13 +[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2, 13 +[5] Lina Bariah, Qiyang Zhao, Hang Zou, Yu Tian, Faouzi Bader, and Merouane Debbah. Large generative ai models for telecom: The next big thing? IEEE Communications Magazine, 2024. 1, 2, 13 +[6] Ali Behrouz, Parsa Delavari, and Farnoosh Hashemi. Unsupervised representation learning of brain activity via bridging voxel activity and functional connectivity. In *Forty-first International Conference on Machine Learning*, 2024. 13 +[7] Shihao Cai, Keqin Bao, Hangyu Guo, Jizhi Zhang, Jun Song, and Bo Zheng. Geogpt4v: Towards geometric multi-modal large language models with geometric image generation. In The Association for Computational Linguistics, 2024. 2 +[8] Daipeng Cao, Jun Wu, and Ali Kashif Bashir. Multimodal large language models driven privacy-preserving wireless semantic communication in 6g. In Proc IEEE International Conference on Communications Workshops, pages 171-176. IEEE, 2024. 1, 2 +[9] Wenhu Chen, Ming Yin, Max Ku, Pan Lu, Yixin Wan, Xueguang Ma, Jianyu Xu, Xinyi Wang, and Tony Xia. Theoremq: A theorem-driven question answering dataset. In Proc. Conference on Empirical Methods in Natural Language Processing, 2023. 14 +[10] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 14 +[11] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Train- + +ing verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 2 +[12] Anurag Dutta, Madhurima Panja, Uttam Kumar, Chittaranjan Hens, and Tanujit Chakraborty. Van der pol-informed neural networks for multi-step-ahead forecasting of extreme climatic events. In NeurIPS 2023 AI for science workshop, 2023. 13 +[13] Roy Ganz, Yair Kittenplon, Aviad Aberdam, Elad Ben Avraham, Oren Nuriel, Shai Mazor, and Ron Litman. Question aware vision transformer for multimodal reasoning. In Proc IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13861-13871, 2024. 4, 18 +[14] Timin Gao, Peixian Chen, Mengdan Zhang, Chaoyou Fu, Yunhang Shen, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Xing Sun, Liujuan Cao, et al. Cantor: Inspiring multimodal chain-of-thought of mllm. arXiv preprint arXiv:2404.16033, 2024. 13 +[15] Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Minlie Huang, Nan Duan, and Weizhu Chen. Tora: A tool-integrated reasoning agent for mathematical problem solving. In Proc International Conference on Learning Representations, 2024. 13 +[16] Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Leng Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, et al. Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems. In The Association for Computational Linguistics, 2024. 2 +[17] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874, 2021. 14 +[18] Anton A Huurdeman. The worldwide history of telecommunications. John Wiley & Sons, 2003. 13 +[19] Serwan Jassim, Mario Holubar, Annika Richter, Cornelius Wolff, Xenia Ohmer, and Elia Bruni. Grasp: A novel benchmark for evaluating language grounding and situated physics understanding in multimodal language models. In The International Joint Conference on Artificial Intelligence, 2024. 1, 2 +[20] Imtiaz Karim, Kazi Samin Mubasshir, Mirza Masfiqur Rahman, and Elisa Bertino. Spec5g: A dataset for 5g cellular network protocol analysis. In Proc International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics, pages 20-38, 2023. 1, 2, 13 +[21] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 3, 5, 6, 7, 13 +[22] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proc IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 2 +[23] Yifei Li, Zeqi Lin, Shizhuo Zhang, Qiang Fu, Bei Chen, JianGuang Lou, and Weizhu Chen. Making language models + +better reasoners with step-aware verifier. In Proc the Association for Computational Linguistics, pages 5315-5333, 2023. 2, 13 +[24] 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. In Proc International Conference on Learning Representations, 2023. 1, 2, 3, 6, 7, 13, 14, 19, 20 +[25] Boxun Liu, Xuanyu Liu, Shijian Gao, Xiang Cheng, and Liuqing Yang. Llm4cp: Adapting large language models for channel prediction. arXiv preprint arXiv:2406.14440, 2024. 2, 13 +[26] Fenglin Liu, Tingting Zhu, Xian Wu, Bang Yang, Chenyu You, Chenyang Wang, Lei Lu, Zhangdaihong Liu, Yefeng Zheng, Xu Sun, et al. A medical multimodal large language model for future pandemics. NPJ Digital Medicine, 6(1): 226, 2023. 2 +[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proc IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 13 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 13 +[29] Xiang Liu, Zhaoxiang Liu, Huan Hu, Zezhou Chen, Kohou Wang, Kai Wang, and Shiguo Lian. A multimodal benchmark dataset and model for crop disease diagnosis. In Proc The European Conference on Computer Vision, 2024. 2 +[30] Jianqiao Lu, Zhiyang Dou, Hongru Wang, Zeyu Cao, Jianbo Dai, Yingjia Wan, and Zhijiang Guo. Autopsy: Automated process-supervised verifier. In Advances in Neural Information Processing Systems, 2024. 2, 3, 13 +[31] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In The International Conference on Learning Representations, 2024. 2, 14 +[32] Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583, 2023. 14 +[33] Ali Maatouk, Fadhel Ayed, Nicola Piovesan, Antonio De Domenico, Merouane Debbah, and Zhi-Quan Luo. Teleqna: A benchmark dataset to assess large language models telecommunications knowledge. arXiv preprint arXiv:2310.15051, 2023. 1, 2, 13, 14 +[34] Ali Maatouk, Nicola Piovesan, Fadhel Ayed, Antonio De Domenico, and Merouane Debbah. Large language models for telecom: Forthcoming impact on the industry. IEEE Communications Magazine, 2024. 1, 2, 13 +[35] Mathpix. Mathpix Snip, 2024. Available at: https://mathpix.com/.3,15 + +[36] Bertalan Mesko. The impact of multimodal large language models on health care's future. Journal of medical Internet research, 25:e52865, 2023. 1, 2 +[37] Yukai Miao, Yu Bai, Li Chen, Dan Li, Haifeng Sun, Xizheng Wang, Ziqiu Luo, Yanyu Ren, Dapeng Sun, Xiuting Xu, et al. An empirical study of netops capability of pre-trained large language models. arXiv preprint arXiv:2309.05557, 2023. 1, 2, 13, 14 +[38] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proc IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 13 +[39] Debjyoti Mondal, Suraj Modi, Subhadarshi Panda, Rituraj Singh, and Godawari Sudhakar Rao. Kam-cot: Knowledge augmented multimodal chain-of-thoughts reasoning. In Proc AAAI Conference on Artificial Intelligence, pages 18798-18806, 2024. 2, 13 +[40] Michael Plainer, Hannes Stärk, Charlotte Bunne, and Stephan Gunnemann. Transition path sampling with boltzmann generator-based mcmc moves. arXiv preprint arXiv:2312.05340, 2023. 13 +[41] Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models. In Empirical Methods in Natural Language Processing, 2024. 2 +[42] Sumit Soman and Ranjani HG. Observations on llms for telecom domain: capabilities and limitations. In Proc International Conference on AI-ML Systems, pages 1–5, 2023. 1, 2, 13 +[43] 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. 13 +[44] Trieu H Trinh, Yuhuai Wu, Quoc V Le, He He, and Thang Luong. Solving olympiad geometry without human demonstrations. Nature, 625(7995):476-482, 2024. 1 +[45] 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. 2, 13 +[46] Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. In Proc Annual Meeting of the Association for Computational Linguistics, pages 9426-9439, 2024. 1, 2, 3, 5, 6, 7, 8, 13, 14, 19 +[47] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. In International Conference on Learning Representations, 2022. 5, 6, 13 +[48] Yanzheng Wang, Boyue Wang, Tianyu Shi, Jie Fu, Yi Zhou, and Zhizhuo Zhang. Sample-efficient antibody de + +sign through protein language model for risk-aware batch bayesian optimization. bioRxiv, pages 2023-11, 2023. 13 +[49] 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. 5, 6, 13 +[50] Daniel S Wigh, Joe Arrowsmith, Alexander Pomberger, Kobi C Felton, and Alexei A Lapkin. Orderly: Data sets and benchmarks for chemical reaction data. Journal of Chemical Information and Modeling, 64(9):3790-3798, 2024. 13 +[51] Duo Wu, Xianda Wang, Yaqi Qiao, Zhi Wang, Junchen Jiang, Shuguang Cui, and Fangxin Wang. Netllm: Adapting large language models for networking. In Proc ACM SIGCOMM 2024 Conference, pages 661-678, 2024. 1, 2, 14 +[52] Shuang Wu, Zhen Qin, Zhen Wang, Siwei Rao, Qiang Ye, Xingyue Quan, and Guangjian Tian. Flexible o&m for telecom networks at huawei: A language model-based approach. In Proc ACM International Conference on Information & Knowledge Management, pages 4224-4233, 2021. 13 +[53] Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. Fine-grained human feedback gives better rewards for language model training. Advances in Neural Information Processing Systems, 36:59008-59033, 2023. 13 +[54] Minrui Xu, Dusit Niyato, Jiawen Kang, Zehui Xiong, Shiwen Mao, Zhu Han, Dong In Kim, and Khaled B Letaief. When large language model agents meet 6g networks: Perception, grounding, and alignment. IEEE Wireless Communications, 2024. 2, 13 +[55] Zelin Xu, Tingsong Xiao, Wenchong He, Yu Wang, and Zhe Jiang. Infusing spatial knowledge into deep learning for earth science: A hydrological application. In NeurIPS 2023 AI for Science Workshop, 2023. 13 +[56] Kaiyu Yang, Aidan Swope, Alex Gu, Rahul Chalamala, Peiyang Song, Shixing Yu, Saad Godil, Ryan J Prenger, and Animashree Anandkumar. Leandojo: Theorem proving with retrieval-augmented language models. Advances in Neural Information Processing Systems, 36, 2024. 13 +[57] Fei Yu, Anningzhe Gao, and Benyou Wang. Outcomesupervised verifiers for planning in mathematical reasoning. In Findings of North American Chapter of the Association for Computational Linguistics, 2023. 2 +[58] Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathematical questions for large language models. In Proc International Conference on Learning Representations, 2024. 1, 13, 14 +[59] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proc IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 1, 2, 13, 14 + +[60] Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mammoth: Building math generalist models through hybrid instruction tuning. In Proc International Conference on Learning Representations, 2024. 13, 14 +[61] Shuo Zhang and Lei Xie. Protein language model-powered 3d ligand binding site prediction from protein sequence. In NeurIPS 2023 AI for Science Workshop, 2023. 13 +[62] Shanghang Zhang, Gaole Dai, Tiejun Huang, and Jianxu Chen. Multimodal large language models for bioimage analysis. Nature Methods, 21(8):1390-1393, 2024. 2 +[63] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2023. 2, 13 +[64] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 13 +[65] Hao Zhou, Chengming Hu, Ye Yuan, Yufei Cui, Yili Jin, Can Chen, Haolun Wu, Dun Yuan, Li Jiang, Di Wu, et al. Large language model (llm) for telecommunications: A comprehensive survey on principles, key techniques, and opportunities. arXiv preprint arXiv:2405.10825, 2024. 1, 13 +[66] Hang Zou, Qiyang Zhao, Yu Tian, Lina Bariah, Faouzi Bader, Thierry Lestable, and Merouane Debbah. Telecomgpt: A framework to build telecom-specific large language models. arXiv preprint arXiv:2407.09424, 2024. 2, 13 \ No newline at end of file diff --git a/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/images.zip b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9abb8b97600c8f102aada1dfebab0500c05d25e1 --- /dev/null +++ b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bded78865efc717c295a759c88b5efbb29743e4a0d39cc6e8b8cc05e1b3a534c +size 611359 diff --git a/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/layout.json b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9048068977b6d07688a2265102157628173ed591 --- /dev/null +++ b/ICCV/2025/RMultiplex200K_ Toward Reliable Multimodal Process Supervision for Visual Language Models on Telecommunications/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1a459794dee96db1e183da8f8a36167cda5e13a119d7e9fae66ce7dc83d78a8 +size 487390 diff --git a/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/a282fced-716e-43b8-bc5e-05fba052ae84_content_list.json b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/a282fced-716e-43b8-bc5e-05fba052ae84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..29eeb4187d6ee2a78c0d8122701b5f7340661e1c --- /dev/null +++ b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/a282fced-716e-43b8-bc5e-05fba052ae84_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0ec523a2f664b5948c2a6508c79d462d665ea76aea943dfad09a229305029a5 +size 90779 diff --git a/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/a282fced-716e-43b8-bc5e-05fba052ae84_model.json b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/a282fced-716e-43b8-bc5e-05fba052ae84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b799252d749ecf8dc53132deeb1c57281247e2dd --- /dev/null +++ b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/a282fced-716e-43b8-bc5e-05fba052ae84_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bb7b3b635f63da11305c836cdcb205bb2035b1ac8eaaaeeb0f3efa3af7697d5 +size 116902 diff --git a/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/a282fced-716e-43b8-bc5e-05fba052ae84_origin.pdf b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/a282fced-716e-43b8-bc5e-05fba052ae84_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fb8794167365b75d500c88d1c604f3bf6f501e5d --- /dev/null +++ b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/a282fced-716e-43b8-bc5e-05fba052ae84_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cebf9ace78721c0f3d7a71692ba5fd57dec72a5385c050d26ec31add58b3a3f +size 20964255 diff --git a/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/full.md b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/full.md new file mode 100644 index 0000000000000000000000000000000000000000..74cb9396a5904f3cfaacbd3d7c0d51832c3428c8 --- /dev/null +++ b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/full.md @@ -0,0 +1,316 @@ +# ROADWork: A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones + +Anurag Ghosh Shen Zheng Robert Tamburo Khiem Vuong Juan Alvarez-Padilla +Hailiang Zhu* Michael Cardei* Nicholas Dunn* Christoph Mertz Srinivasa G. Narasimhan +Carnegie Mellon University +https://www.cs.cmu.edu/~roadwork/ + +# Abstract + +Perceiving and autonomously navigating through work zones is a challenging and underexplored problem. Open datasets for this long-tailed scenario are scarce. We propose the ROADWork dataset to learn to recognize, observe, analyze, and drive through work zones. State-of-the-art foundation models fail when applied to work zones. Finetuning models on our dataset significantly improves perception and navigation in work zones. With ROADWork, we discover new work zone images with higher precision $(+32.5\%)$ at a much higher rate $(12.8\times)$ around the world. Open-vocabulary methods fail too, whereas fine-tuned detectors improve performance $(+32.2AP)$ . Vision-Language Models (VLMs) struggle to describe work zones, but finetuning substantially improves performance $(+36.7SPICE)$ . + +Beyond fine-tuning, we show the value of simple techniques. Video label propagation provides additional gains $(+2.6\text{AP})$ for instance segmentation. While reading work zone signs, composing a detector and text spotter via crop-scaling improves performance $(+14.2\%)$ 1-NED). Composing work zone detections to provide context further reduces hallucinations $(+3.9\text{SPICE})$ in VLMs. We predict navigational goals and compute drivable paths from work zone videos. Incorporating road work semantics ensures $53.6\%$ goals have angular error $(AE) < 0.5^{\circ}$ $(+9.9\%)$ and $75.3\%$ pathways have $AE < 0.5^{\circ}$ $(+8.1\%)$ . + +# 1. Introduction + +Navigating safely through work zones is a significant challenge for both autonomous and human-driven vehicles. In recent years, several reports suggest that self-driving cars are stumped by the simplest work zones [10, 11, 29, 41, 42, 44, 47, 75, 76]. In fact, even humans find driving through work zones challenging [46, 62, 64]. Since 2010, more than 700 people have died every year in the United States in accidents involving work zones [48, 63]. Although road work + +![](images/55b3f10d04a33e7dc861933976b0270136c5010cae874234fc617eabfa589185.jpg) + +![](images/6812cce506fedb9c69e73fa97416ed7d88c6d2a2155481a77d53c6e4e1f2a0bb.jpg) + +![](images/394662fa682bce3c84fa544ce2d7cf59c1903f6247b6b39b3ee3c9f43b9c861b.jpg) +Figure 1. Autonomous Driving In Work Zones. Work zone objects block the road (in red), making some road regions unsafe (red arrows), necessitating the development of forecasting traversable paths from driven paths (green dots). Prior datasets do not comprehensively address this long-tailed challenge. Foundation models struggle with recognizing and interpreting work zone objects and signs, discovering new work zones, and analyzing work zones. We additionally formulate work zone navigation as a learnable task. The ROADWork dataset and benchmark highlight key challenges in work zone perception and navigation, and we demonstrate simple techniques that improve performance on these tasks. + +zones are relatively rare, learning to navigate them is essential for full vehicular autonomy. + +Despite the research potential, few studies focus on this domain. Work zones fall into the long-tail of scene distributions for vehicular autonomy. Mining data for such long-tail situations is expensive, difficult, and labor-intensive [35]. Even large data sources contain few informative and diverse long-tailed samples. Thus, such data and methods are often trade secrets [18, 66]. For perspective, in this work, we semi-automatically filtered only 5078 keyframes (less than $0.01\%$ ) from nearly 45 million frames in one data source. + +Both common [8, 49, 69] and long-tailed datasets [33, 70, 71] contain few work zones. Among common datasets, our analysis estimates that BDD100K [69] and Mapil- + +
DatasetsTypeDataset SizeWork Zone Related ObjectsNumber of CitiesPixel-Level AnnotationsObject-Level AnnotationsFine-grained AnnotationsScene DescriptionsDriven Pathway
BDD100k [69]C10K ImagesNone4 US Cities✓
100K ImagesNone✓
NuScenes [1]C1000 Scenes (40K Frames)5 Object CategoriesBoston and Singapore✓✓
Waymo [60]C1150 ScenesNone3 US Cities✓✓
Mapillary [49]C25K Images2 Object CategoriesAround The World✓✓
Cityscapes [8]C5000 Images1 Object Category27 Cities in Germany✓✓
WildDash [70]LT1800 ImagesNoneAround The World✓
WildDash 2 [71]LT5000 ImagesNoneAround The World✓
CODA [33]LT1500 Scenes (5937 Images)6 Object CategoriesNuScenes, Karlsruhe and few cities in China✓
SegmentMeIfYouCan [2]AN327 Images (Eval)NoneUnknown✓
BDD-Anomaly [20]AN810 Images (Eval)NoneData from BDD100K✓
ROADWork (Ours)LT4375 Videos, 9650 Images, 129K Images (Path.)15 Object Categories 360 Unique Signs18 US Cities In-The-Wild Images from Around The World✓✓✓✓✓
+ +Table 1. Comparing ROADWork With Other Driving Datasets/Benchmarks. Compared to common (C) datasets [1, 8, 49, 69], ROADWork has a wider variety of annotations for a rare driving scenario while being similar in size. Compared to long-tailed (LT) and anomalous (AN) datasets [2, 20, 33, 70, 71], ROADWork is larger with rich pixel, object and scene-level annotations across a broader range of tasks. Others focus on one or two aspects of self-driving but ROADWork holistically covers work zone perception and navigation. + +lary [49] collectively contain fewer than 1000 work zone images. Although long-tailed datasets such as CODA [33] contain bounding boxes for conventional work zone objects such as cones and work vehicles, they do not cover all work zone objects and do not provide any instance segmentations, fine-grained object attributes, scene descriptions, or drivable paths. Others such as WildDash [70, 71] contain images with heavy weather, lighting and geographic variations but largely annotate common Cityscapes [8] categories. Previous works contain little research on work zones, focusing mainly on individual issues and edge cases [13, 14, 16, 17, 27, 43, 50, 57, 67]. + +Because work zones are neglected and under-represented in existing datasets, even after training with massive amounts of data and compute, foundation models such as open vocabulary detectors [38, 65, 74], promptable segmentation models [28, 53], scene text models [54] and vision language models [30, 32, 36], do not perform well in this scenario. The key to improving self-driving in work zones lies in obtaining work zone driving data - even simple baselines such as Mask R-CNN [19] outperform foundation models [65, 74] when trained on a well-curated dataset. + +Thus, a concerted and structured approach is needed to address this challenge. We present a novel ROADWork dataset and benchmark for addressing self-driving in work zones. In total, our dataset contains 4375 thirty-second videos, 9650 richly annotated keyframes, and 129017 images with automatically computed pathways from more than 5000 work zones. We cover 18 US cities and in-the-wild images from around the world. Our dataset has 15 types of objects, 360 types of Temporary Traffic Control (TTC) signs and boards, scene descriptions and 2D/3D traversable paths (Figure 2). Table 1 provides a comparison with other datasets. ROADWork contains richer annotations for a greater number of tasks, is larger than other long-tailed datasets, and covers many aspects of perception and navigation in work zones. To the best of our knowledge, no large-scale public dataset targets work zones. For detailed related work, see Supp. A. + +What kind of data is needed for driving in work zones? No two work zones are truly alike, with unique arrangements of barriers, cones, and signs. Navigating these work zones requires context-specific understanding of atypical object configurations that standard driving datasets fail to capture. We propose that understanding and navigating work zones mirrors human cognition, where seeing differs from observing, and knowing differs from understanding [59]. We address: Recognizing work zones and their constituent objects, Observing fine-grained details (e.g. sign text) that provide navigation instructions, Analyzing global scene context to capture spatial relationships, and Driving through work zones via pathway prediction. + +# Our contributions are summarized below: + +- We mine thousands of hours of driving videos to curate the ROADWork dataset containing 4375 videos, 9650 richly annotated images, and 129017 images with automatically computed pathways from over 5000 work zones across 18 US cities and around the world. We formulate major challenges as problems reflecting human cognition: recognizing work zones and their objects, observing fine-grained details such as sign text, analyzing global scene context, and driving safely through these zones. +- Foundation models perform poorly on work zone tasks. For example, open vocabulary segmentation models achieve only 2.9-4.2 AP. We dramatically improve perception and navigation with our data (39.0 AP). +- Beyond fine-tuning, we show the value of simple techniques: (a) video label propagation improves detection $(+2.8\text{AP})$ , (b) crop-rescaling improves reading sign texts $(+14.2\%$ 1-NED), (c) composing work zone detections with VLMs reduces hallucinations in scene descriptions $(+3.9\text{SPICE})$ , and (d) incorporating work zone semantics improves pathway prediction $(+9.9\%$ in goals with angular error $< 0.5^{\circ}$ ). Fine-tuning foundation models on the ROADWork dataset does not hurt performance on common driving situations allowing label unification. The ROADWork dataset enables studying other long-tailed scenarios such as geographic domain adaptation. + +![](images/3dc4a1c43231df9f415f28ccfa7f459d2113b822c2ef4b7e10fc9fa17b6203d5.jpg) +Work on corner next to road. Workers off road. Line of vertical panels on both sides of road lane. Work vehicle partially blocking road. + +![](images/62677d2a91ca89cfa7f3d0febe772dec99a33d2e494d581fd0d1bb06896d4723.jpg) +Work on both sides of road. Line of drums on both sides of the road. Lane shift by drums on right side of road. Barriers on left side of road in front of work zone. Work vehicles behind barriers which fully block sidewalk. TTC sign partially blocking sidewalk on right. + +![](images/59c9c584c7f065ba0862cafe29a8760e159b6afbb8ad8c5c9e7a4c15f7db886c.jpg) +Tubular markers and cones next to work vehicles on right side of road. + +# 2. ROADWork Dataset + +The ROADWork dataset contains road work navigation videos and images with pixel-level annotations, object-level annotations, fine-grained object annotations, scene level annotations, and 2D/3D driven pathways. The annotations encompass multiple levels mirroring human cognition: (a) Recognizing work zones and detection of constituent objects using bounding boxes and segmentation labels, (b) Observing fine-grained details about the work zones using arrow board states, labeled text, arrows, and graphics on traffic signs, (c) Analyzing the work zones enabled by detailed and consistent human description about type of work along with spatial understanding about blockage and activity, and (d) the traversable Driving path computed automatically from driving videos. Figure 2 shows examples of work zones and annotations from our dataset. + +ROADWork consists of 4375 videos, 9650 richly annotated images, and 129017 images with passable trajectories. The videos were used to propagate labels [53] to 11959 additional keyframes to train our models. Like other recent papers for scaling data [28], we use bootstrapping to collect long-tailed work zone data. Our first bootstrapping stage involved manually capturing 2338 work zone images by driving in Pittsburgh to train our initial models. In our second bootstrapping stage, we manually curated 5078 keyframes after automatically mining nearly 100K potential images from Michelin Mobility Intelligence (MMI) Open Dataset [25] containing 45 million frames. From the selected keyframes, we obtained 4375 thirty-second videos of driving through work zones. All data is captured using smartphones with known intrinsic's mounted inside the vehicle. In total, we estimate that our data covers around 5000 work zones (see Supp. B) from the first two data curation stages. Most of the results presented in the paper correspond to training using data from the first two stages, unless otherwise stated. In our third bootstrapping stage, we expanded to in-the-wild data sources. We mined, discovered and annotated 969 work zone images from BDD100K [69] and Mapillary [49] for in-the-wild evaluation. We also + +![](images/e1131f0ae5a397c85dbdaf51f77376a59ed8fbd41a768a9d81e6481d04b98e32.jpg) +Figure 2. The ROADWork Dataset consists of work zone videos and images. We have segmented 15 object categories such as workers, vehicles and barriers. We provide object attributes for signs and arrow boards to enable fine-grained understanding. Our work zone scene descriptions analyze the scene globally and one passable trajectory automatically estimated from the associated video sequences learns how to drive through work zones. See Supp. B for more details. + +![](images/ab8395e575443fe70a12fda40d38d0dcc840c751010132f164a3255e95ff6912.jpg) + +![](images/392e45f08c74e7544ae7ce88bdf975d601fd0433203dae4fa5899c78128feb36.jpg) + +![](images/497d5e44e1bd167e4ff030f3831c3936c363488085fdff7eeacc61c50f543b5a.jpg) + +![](images/058aae6bafc18ff189a4f1a963afd8764aa58c3723e4413806caf374352d896c.jpg) + +![](images/c0a71e59be106244bfeadaeba686da1219b92266b8c02746e193a305aba91ea2.jpg) +Figure 3. Generating Driving Trajectory. Using driving video frames (a-c) as input, we estimated camera poses using COLMAP [56] (d-f), where camera poses are depicted as viewing frustums shown in red. These poses were projected onto the estimated ground plane to form the driven trajectory, then backprojected onto the frames as the future driving trajectory (g). + +![](images/0f12a8fa692d9734d7df5b3f21dcbd2a922100dd9dbf97cfb458aedbb7ed6461.jpg) + +semi-automatically discovered images in bad weather and night conditions. We discovered 465 images by driving in Pittsburgh and other rural US areas, along with 800 images from a commuter bus that followed a fixed route in Pittsburgh over the course of two years. + +From the videos, we automatically extracted the actual path trajectory taken by the ego-vehicle. For this, we implemented a 3D reconstruction pipeline using off-the-shelf-modules (described in Supp. B.3). We then obtained driven pathways by projecting the camera poses first onto the ground plane and then back onto the image. To ensure correctness, we manually verified all associated keyframes. Figure 3 visualizes some of these steps. This process resulted in 1936 unique sequences with passable trajectories for 129017 frames. A complete description of the acquisition, annotations, data cleaning protocols, and key statistics about ROADWork are in Supp. B. + +![](images/05838dfa2a278d4962133ae00b3a599696ab68b915418c99124d64c363c6c474.jpg) +Figure 4. Challenges And Variations In Work Zone Objects. There exists a significant geographical variation in work zone objects and work vehicles in the ROADWork dataset. For instance, the appearance of barriers, vertical panels, and tabular markers (first row) varies across cities. Similarly, work vehicles (second row) demonstrate large variations, as they are specialized for specific tasks within the work zones. + +
APAP50AP75APsAPmAPl
Foundation Models (Zero-Shot)
OpenSeeD (COCO+O365) [72]2.95.82.51.32.84.0
Detic (LVIS+121K) [74]4.26.34.62.15.56.0
OMG-LLaVA (Many Datasets) [73]0.20.30.200.10.8
Supervised with ROADWork Dataset
Mask R-CNN [19]29.948.332.716.032.943.8
Mask DINO [31]36.455.240.720.237.452.2
Mask DINO [31] w/ Video Prop. [53]39.058.843.422.140.055.0
+ +# 3. Recognizing Work Zones + +"Recognizing" work zones is underexplored, with little research in this domain. We define "recognition" as being able to identify a work zone and objects present in it - no matter the geography or conditions. Figure 4 shows some of the associated challenges. The ROADWork Dataset enables us to understand and improve these recognition challenges in the era of foundation models. + +Detecting And Segmenting Work Zone Objects. Foundation models for open vocabulary scene understanding [68, 72-74] have shown impressive zero-shot generalization on many datasets. No closed vocabulary detector exists for recognizing work zones objects, leading us to consider these open vocabulary foundation models. Although earlier datasets considered few work zone objects such as cones [1, 49], they did not consider all types of work zone objects, such as "Arrow Board". + +We investigate whether popular open vocabulary instance segmentation methods, Detic [74] and OpenSeeD [72] generalize to our data. We also investigate whether OMG-LLaVA [73], a foundation model trained for pixel level tasks, can reason about and segment work zone objects. Table 2 shows our results. Although few of the object categories are common concepts, such as "Fence" and "Police Officer," open vocabulary methods [72, 74] fail to detect and segment most objects, such as "Temporary Traffic Control (TTC) Message Board" and "Arrow Board". This is irrespective of the large-scale training source. Note + +Table 2. Segmenting Work Zone Objects. We train instance segmentation [19, 31] on ROADWork using a coarse vocabulary. Open vocabulary methods [38, 72, 74] fail to recognize work zone objects. In contrast, supervised models perform significantly better $(+32.2\text{AP})$ , and label propagation through video [53] further improves performance $(+2.6\text{AP})$ . + +
SupervisionAPAP50AP75APAP50AP75
SAM [28]SAM2 [53]
Bbox24.747.927.326.447.726.4
Bbox + 5 pts25.446.525.326.247.426.5
Bbox + 10 pts25.847.225.925.046.824.6
Ground Truth29.948.332.729.948.332.7
+ +Table 3. Are Manual Segmentations Still Necessary? We train instance segmentation models [19] with varying levels of weak supervision from the ROADWork dataset. We use SAM [28] and SAM2 [53] to generate pseudo-ground truth masks from ground truth boxes and points. Manual ground truth masks (last row) outperform SAM at tighter IoU thresholds (+5.4 AP75) and for rare categories such as "Arrow Board" (+26.9 AP). SAM2 pseudoground truth masks, while better than SAM pseudo-ground truth overall (+3.5 AP), perform worse at tighter thresholds (+6.3 AP75) when compared to manual ground-truth masks. + +that all methods additionally leverage CLIP [51], which is trained on a 400 million image-text paired dataset, which likely under-represents work zones. For comparison, we train coarse vocabulary work zone object detectors by fine-tuning Mask R-CNN [19] and Mask DINO [31] on our dataset, producing significantly improved detection (+25.7 AP and +32.2 AP, respectively). However, many work zone categories are rare, e.g. the ROADWork annotations itself has fewer than 1000 instances of "Arrow Board". To increase the number of instances, we use the Segment Anything 2 (SAM2) [53] model to propagate segmentations through ROADWork videos - we prompt all ground truth masks as input for an annotated keyframe to SAM2. Then we propagate masks for the next 60 frames and uniformly add every 10th frame to our augmented dataset. We obtain 11959 more labeled frames using 1947 training videos from ROADWork, in addition to 5318 images in the training set. Training on this augmented dataset yields further improvements (+2.6 AP). + +In summary, work zone objects are woefully underrepresented in existing foundation models and datasets, highlighting the crucial role of ROADWork dataset. We perform analysis of open-vocabulary detectors [38] and state-of-the-art object detectors in Supp. C.1, which follow similar trends. In Supp. C.1, we also show that these foundation models can be fine-tuned on our data with no or minimal loss in performance on common driving datasets. + +Are Manual Segmentations Still Necessary? Given foundation models for promptable segmentation [28, 53] have shown competitive zero-shot generalization on many datasets, are manual segmentations even needed anymore? + +We use Segment Anything Model (SAM) [28] and Segment Anything Model 2 (SAM2) [53] to perform zero-shot instance segmentation. We use ground truth boxes as input to generate pseudo-ground truth masks for the ROAD-Work dataset. We also simulate $K = \{5,10\}$ positive and negative points from manual ground-truth masks as input prompts [4]. We then train detectors (in this case, Mask R-CNN [19]) using manual and pseudo-ground truth masks. + +
MethodSupervisionBDD100K [69]Mapillary [49]
#DiscoveredPrecision#DiscoveredPrecision
Detic [74]I21K + LVIS3252.4 %12542.9 %
Mask R-CNN [19]COCO + ROADWork41184.9 %55877.0 %
+ +Table 4. Characterizing and Discovering Work Zones. The ROADWork dataset improves discovery of new work zones compared to open-vocabulary foundation models. Using the same classification rule, we employ a coarse-vocabulary detector [19] trained on ROADWork dataset. We discovered $12.8 \times$ and $4.5 \times$ more work zone images than Detic [74], improving precision by $+32.5\%$ and $+34.0\%$ on BDD100K [69] and Mapillary [49]. + +![](images/c08b446c6db6c08b53da6c644dfa24dc83e698c9296db6645b464b4c6b106275.jpg) +Figure 5. Work Zones Discovered Around The World. Mapillary [49] dataset contains driving images from around the world. Despite the ROADWork training dataset images being restricted to the U.S., we discovered work zones captured in Europe and Asia. + +From Table 3, manual ground truth masks improve performance by $+5.2$ AP over pseudo-ground truth masks from SAM. This gap is more pronounced for rare objects, example, "Arrow Board" $(+26.9$ AP) and "TTC Message Board" $(+8.7$ AP). Pseudo-ground truth masks are comparable for common classes such as "Worker" $(+1.2$ AP) and "Police Officer" $(-2.1$ AP). Performance gap at tighter thresholds is higher $(+5.4$ AP75). Using additional annotated points as weak supervision reduced the overall gap to $+4.1$ AP. Performance of SAM2 is not much better than SAM. Ground truth masks are still better than SAM2 $(+3.5$ AP) but the gap is smaller than SAM. SAM2 is worse than SAM at tighter IOU thresholds $(+6.7$ AP). In this case, additional point supervision affects performance at tighter thresholds, increasing the gap to $+8.1$ AP with ground truth masks. Thus, manual ground truth masks are still necessary. Characterizing And Discovering Work Zones. During dataset curation, we manually filtered work zone images (See Section 2) — a process too cumbersome to scale to millions of images at lower precision levels. Thus, automatically discovering new work zones from vast amounts of unlabeled data is important. It is also useful to understand whether models trained on the ROADWork dataset generalize to new data sources. However, the existence of a work zone object (e.g., a work vehicle) alone does not make a scene a work zone. We classify a work zone as any activity on the road, defined by specific objects and human actors, that redefines the road network and traffic rules. A proxy for work zone activity is its "intensity," measured by the + +![](images/a758a3aa457c8780bd4777fd69ebdb4619466cf09cb9c945db486d6706746162.jpg) +Figure 6. Challenges In Fine-Grained Observations Like Signs. TTC signs contain "text", "graphics", or both. We annotated 62 types of graphics and 360 text types across 8242 sign instances. These signs convey specific information (e.g., for pedestrians or delivery vehicles), require interpretation with other signs, and are may be vandalized. They also impact long-term navigation (e.g., driving towards Burger King and reading "Burger King Closed During Construction"). Thus, interpreting signs is not only challenging but also requires understanding the broader scene. + +number and relative area of work zone objects in the scene. + +We discover work zones in existing self-driving datasets devising a simple classification rule: an image is a work zone iff at least three instances of work zone objects from two unique categories that occupy at least $10\%$ of the image area. We chose Mask R-CNN for discovery due to its good accuracy-latency trade-offs [23, 34] for processing millions of images. Compared to foundation model Detic [74], Mask R-CNN trained on ROADWork discovers work zone images with higher precision in BDD100K [69] and Mapillary [49] datasets. We found fewer than 1000 roadwork images, representing only $0.4\%$ and $2.3\%$ of these datasets respectively. Surprisingly, while the ROADWork dataset contains images only from US cities, the model discovered work zones worldwide (Figure 5). Furthermore, detectors trained on the ROADWork dataset perform significantly better than open-vocabulary methods on discovered images in a variety of other conditions along with BDD100K and Mapillary; see Supp. C.1. + +Summary. The ROADWork dataset significantly helps to recognize and discover work zones. Further, In Supp. C.1, open-vocabulary detectors [38] fine-tuned on the ROAD-Work data still generalize to common driving scenes, while not degrading performance on Cityscapes [8] $(+0.2 AP)$ compared to a pretrained model. ROADWork enables study of other interesting long-tailed recognition problems. In Supp. C.1.1, we show that there exist large domain gaps $(+7.4 AP)$ between cities even with SOTA adaptation methods [26] $(-13.8 AP_{50}$ for "barricade" compared to upper-bound). We also show results for label unification. Detector performance degrades $(-5.0 AP)$ when models are jointly trained on datasets with common and rare object labels. The state-of-the-art label-unification algorithm [65] marginally improves performance $(+2.0 AP)$ . + +
APAP50AP75APsAPmAPl
Foundation Models (Zero-Shot)
OpenSEED (COCO+O365) [72]1.02.00.90.51.12.1
Detic (LVIS+I21K) [74]1.11.51.30.31.82.2
OMG-LLaVA (Many Datasets) [73]00.10.1000.2
Supervised with ROADWork Dataset
Mask R-CNN [19]17.626.219.511.321.236.0
Mask R-CNN w/ Copy Paste [15]25.034.627.811.026.946.8
Mask DINO [31]30.040.933.420.634.450.1
Mask DINO [31] w/ Video Prop. [53]32.843.436.418.732.760.2
+ +Table 5. Segmenting Fine-Grained Signs. We expand the "TTC sign" category to classify different kinds of TTC signs into separate categories. Open vocabulary methods [68, 74] fail to detect most of these objects with this expanded vocabulary. Supervised Mask DINO [31] improves considerably $(+23.9\text{AP})$ . Additionally, simple copy-paste [72] ensures simpler Mask R-CNN [19] baseline compares favorably with SOTA transformers. + +# 4. Observing Work Zones + +As noted in Section 2, recognizing objects alone is insufficient; fine-grained local observations and understanding is essential, such as sign interpretation and human intent recognition. We define this challenge as "observing". For example, consider temporary traffic control (TTC) signs in Figure 6. Sign interpretation is valuable for both short-term navigation and long-term planning. For the purposes of our discussion, we break down sign interpretation into two subproblems: recognizing the graphical elements and reading the text. We discuss challenges in current methods. + +Segmenting Fine-Grained Signs. "Graphics" refers to the pictorial diagrams that are frequently present in signs, such as arrows or stick figures that depict humans. ROAD-Work contains 62 types of signs board graphics. Frequency distribution of these graphics is long-tailed and we show some examples in Figure 6. In the coarse vocabulary considered in Section 3, all TTC signs were binned into the general "TTC sign" category. Here, we consider a fine-grained vocabulary, where each sign-board graphics is a separate category, expanding the label vocabulary to 49 classes. However, scarce categories (sign graphics with less than 5 instances) were binned into "Other" category. + +As discussed in Section 3 we employ methods like Detric [74], OpenSeeD [68] and OMG-LLaVA [73]. We fine-tune instance segmentation models [19, 31] on ROADWork with fine-grained vocabulary for comparison. Our observations are in Table 5. All the foundation models [72-74] are dismal in segmenting these signs while Mask DINO [31] performs considerably better $(+28.9\text{AP})$ . Following Section 3, we adopt a similar strategy to exploit ROADWork videos and to propagate ground-truth labels using SAM2 [53] and obtain an additional improvement of $+2.8\text{AP}$ . However, overall performance of Mask R-CNN was still low (17.6 AP) compared to Mask DINO. We use simple copy-paste augmentation [15] which improves rare category segmentation. It further improved the accuracy of our Mask R-CNN [19] model by $+7.1\text{AP}$ . + +
Edit. (1 - NED)Word Acc.Ch. RecallCh. PrecisionCh. F1
Foundation Models (Zero-Shot)
Gemini 2.5 Flash [7]22.74.594.114.625.2
GPT-4o [24]40.311.984.727.341.3
Specialized Text Spotting Model (Zero-Shot)
Glass [54]65.456.065.898.478.9
Glass (w/ 1x crops)76.856.378.984.281.4
Glass (w/ 3x crops)79.659.680.686.183.2
Small Signs Subset (Area less than 32 × 32)
Glass [54]19.613.218.193.130.3
Glass (w/ 1x crops)76.720.942.270.452.8
Glass (w/ 3x crops)81.332.348.778.460.0
+ +Table 6. Reading Sign Text. The pretrained text spotting method Glass [54] performs well on nearby signs, but it fails to detect and read distant text (See Figure 7). Closed foundation models like GPT-4o [24] and Gemini 2.5 [7] are worse. A simple crop-rescale strategy composing a work zone detector with Glass [54] improves text spotting and recognition by $+14.2\%$ (1 - NED). + +![](images/ab5f6ceadabe0cb7901c0300161a3c41cb005f0af51086aeb273ba826e294425.jpg) + +![](images/4f14cbf65377405184011bd20e26bbd0c90700348b9dbb72209f59fd3bccc7ad.jpg) + +![](images/08a6ba73f108f42b7e5490e1418528ed5659a7f05141bd9982aa11c536380975.jpg) +Figure 7. Reading Sign Text. Glass [54] reads nearby signs but fails for distant signs (a-c). Dirt, occlusions and vandalism are issues for close-by signs too. In (d), prediction is "Keep Bight". + +![](images/07a7958c5d173f78ec4a1453076eb7eca1a0ac780ee7e2e87fcd35b377cb873a.jpg) + +Reading Sign Text. ROADWork contains 360 different TTC sign texts. Sign texts are unconstrained and provide navigational instructions. Spotting and reading them accurately is important. We consider foundation models like GPT-4o [24] and Gemini 2.5 Flash [7] and a popular scene text spotting method, Glass [54]. We evaluate the spotter on TTC signs, and ignore other text in the scene. Foundation Models are unable to localize the text regions accurately, which is consistent with their underperformance on standard tasks [52]. We then try Glass [54], trained on the TextOCR [58] dataset. Glass accurately spots and reads most TTC signs, but fails on small signs far away (see Figure 7). Our annotations do not include character-level annotations required by Glass [54] so we did not fine-tune it on our data. + +We report sentence, word, and character level metrics for text spotting [5, 54]. From Table 6, we observe that while the precision of Glass [54] is high, the recall is low. Thus, we adopted a simple cropping strategy, composing Glass [54] with a coarse vocabulary work zone object detector [19] (Section 3). We infer the text spotter only on TTC sign crops. Our strategy significantly increased recall $(+13.1\%)$ at the expense of precision. Increasing the scale of the crops improved all metrics, such as word accuracy $(+4.5\%)$ . There is a large performance gap for small signs - around $20\%$ to $30\%$ lower. The crop rescale strategy sig- + +![](images/1b97eb688ee2cbab38566db17fb1dbcf5aedeec7a837883c3f26f050c8f5413d.jpg) +5 +Road partially blocked by cones and police vehicle. Workers on road and sidewalk. Right sidewalk partially blocked. +Prrnneaae +In the image, there is a work zone with multiple vehicles, including a police car, a truck, and a car. The police car is located on the left side of the scene, while the truck is located on the right side of the scene. The vehicle is situated on the right side of the scene. There are also several traffic lights in the area, with one near the center of the scene, another on the right side, and the third one further to the right. +A person is standing on the right side of the scene, possibly a construction worker or a pedestrian. Additionally, there is a handbag placed near the center of the scene, possibly belonging to one of the individuals in the area. +00000000000000000000000000000000000000 +Worker on right side of road. Worker holding TTC sign. Worker holding TTC sign on right sidewalk. Worker holding TTC sign on left sidewalk. Line of cones next to work vehicle on left side of road. Worker on left side of road. +w/ +Police vehicle driving on road. Workers on right side of road. Worker on right sidewalk. Barriers and fence on right side of road. Cones on left side of road. +Figure 8. Generating Work Zone Descriptions. We present examples of work zone descriptions generated by a pretrained LLaVA [37] and a LoRA [21] fine-tuned LLaVA (with and without additional object context). The generated descriptions can be classified as factual, incorrect or uninformative. As we can observe, while pretrained LLaVA correctly identifies certain scene elements, it also generates many uninformative or inaccurate sentences. Fine-tuning LLaVA without object context reduces uninformative outputs but introduces hallucinated scene elements that do not exist. Incorporating LLaVA with our workzone detector to provide object context significantly improves faithfulness; however, it still misses some scene elements and spatial interactions (See San Antonio example's ground truth). + +![](images/3a4774203f486f410e4c5569641d363d41c64a058408ae2e813bc1d1f172dd0c.jpg) +Line of cones between lanes on right side of road next to work vehicles containing barricades. +The work zone in the scene features a construction area with orange cones placed around the site. There are multiple cars and trucks parked or driving nearby, as well as a few pedestrians. The orange cones are placed to mark the boundaries of the construction zone, ensuring that vehicles and pedestrians maintain a safe distance from the ongoing work. +Cones on right side of road. Work vehicles on right side of road. Fence around work zone on right side of road. Workers and work vehicle in work zone on right side of road. +Barricades and TTC message board on right side of road. Cones on right sidewalk. + +![](images/ac2c5c892e717966e6ab2beee367bc70da0de605b25c420f32085b6b88c617a8.jpg) +Line of cones on right side of road next to work vehicle. Worker on right intersecting road. Line of cones, workers, and work vehicle on left intersecting road. +In the scene, there is a work zone with several orange traffic cones placed on the street. The cones are positioned in a way that they create a barrier, marking the boundaries of the work zone. There are also multiple cars and trucks in the area, indicating that the work zone is active and vehicles need to navigate around it. The presence of traffic lights suggests that the area is regulated for safety. +Line of cones next to work vehicles on left side of road. Workers on left sidewalk. Workers on right sidewalk. Worker on right side of road. Worker on right sidewalk. Worker on right side of road. +Line of cones on right side of road. Workers and work vehicles on right side of road. Workers on left sidewalk. + +![](images/7505968c47eaf93ee6974aa5284f75eced91b382fa3b37e937c8fd1d1a6c9866.jpg) +Line of vertical panels on right side of right lane of road. Barricade on left side of road. Arrow board on right side of road. Work vehicles on overpass. +In the image, there is a work zone with multiple vehicles, including cars and trucks, driving on a highway. There are several traffic lights and road signs present in the scene, indicating that the area is under construction or has ongoing maintenance. The vehicles are navigating through the work zone, which is characterized by a mix of traffic and construction equipment. +The work zone is located near a hill, and the highway is surrounded by trees. The presence of construction equipment, such as a crane, suggests that the area is undergoing infrastructure development or repairs. The vehicles are driving on a divided highway, with some cars passing through the center of the road. The traffic light is located in the center of the direction. The traffic lights and road signs help regulate the flow of traffic and ensure the safety of the drivers and construction workers in the area. +Barriers on left side of road. Vertical panels on left side of road. Barriers on right side of road. Vertical panels on right side of road. Work vehicles on right side of road. +Vertical panels on right side of road. Line of vertical panels between lanes on right side of road. Arrow board on right side of road. Work vehicles on right side of road. + +nificantly improves word accuracy $(+21.3\%)$ for such signs. Summary. The ROADWork dataset enables studying foundation models to interpret sign graphics and spot sign text. Addressing failures, simple strategies such as copy-paste [15], video label propagation [53] and crop-rescaling image regions all improve downstream performance. + +# 5. Analyzing Work Zones + +Sections 3 and 4 focused on object recognition and fine-grained understanding and observations. However, global scene analysis is equally crucial for navigating work zones. In this section, we study how well foundation models generate global work zone descriptions and use "analyzing" to refer to generating scene descriptions that may guide downstream planning. + +Generating Work Zone Descriptions. Vision-Language Models (VLM) have shown strong performance in scene understanding in various scenarios [32, 36, 37, 51]. However, work zones are underrepresented in their training data. Our experiments show that while pre-trained BLIP [32] and LLaVA [30, 37] fail to generate informative work zone descriptions, fine-tuning on ROADWork enables them to produce accurate and contextually relevant descriptions. + +Table 7 shows improvements with fine-tuning. LLaVA-7B [37] outperforms BLIP-361M [32] on all metrics by a large margin (e.g. +32.8 SPICE) - effectively utilizing the learned priors of a bigger model. Apart from fine-tuning, we added work zone object context by appending detec + +
MethodBLEU@4METEORROUGECIDErSPICE
Pretrained Foundation Models
BLIP-361M [32]0.24.38.61.63.9
LLaVA-1.5-7B [37]0.411.09.409.9
GPT-4o [24]0.916.414.40-
Gemini 2.5 Flash [7]0.613.211.70-
Fine-tuned Foundation Models
BLIP-361M [32]20.721.543.783.541.3
LLaVA-1.5-7B [37]27.024.748.0112.142.7
Fine-tuned w/ Object Context
[37] + Pred. Obj31.127.750.9140.346.6
[37] + GT Obj32.528.552.5166.049.9
+ +Table 7. Generating Work Zone Descriptions. Pretrained VLMs, BLIP [32] and LLaVA-1.5-7B [37], do not generate informative descriptions about work zones. Finetuning with LoRA [21] on ROADWork dramatically improves performance. LLaVA-1.5-7B [37] significantly outperforms BLIP [32], though larger LLaVA models perform similarly (See Supp. C.2). Using object context enhances description faithfulness, but a performance gap remains between ground truth and uncertain object predictions. + +tion predictions from a trained work zone object detector, which further improves performance (+3.9 SPICE). Using ground truth objects as context establishes an upper bound, and a gap exists compared to providing uncertain object predictions (-3.3 SPICE). Figure 8 shows some generated descriptions from pre-trained LLaVA [37], with untrue and irrelevant sentences. In comparison, training on ROAD-Work dramatically improves description quality, and providing object context reduces hallucinations. + +Summary. VLMs fail to describe work zones without the ROADWork dataset. Beyond fine-tuning, providing ob- + +
AE% < θGoalPathway
All Curv.High Curv.All Curv.High Curv.
CRCRCRCR
θ = 0.5°43.753.631.238.467.275.367.175.4
θ = 1°63.573.853.860.582.887.282.787.0
θ = 2°79.085.973.883.090.593.090.493.0
θ = 5°89.693.291.496.896.097.396.097.3
θ = 10°94.597.095.798.998.599.398.599.3
+ +Table 8. Pathway Prediction In Work Zones. YNet [40] models are trained on ROADWork dataset pathway data using pretrained segmentation models from Cityscapes (C) and ROADWork (R). Our metric measures the percentage of predictions where the angular error (AE) between the predicted and ground-truth pathway falls below a given threshold, considering the image's field of view $(\sim 50^{\circ})$ . The model trained with ROADWork segmentations consistently outperforms the Cityscapes-trained model at all thresholds. We also report results on a subset of high-curvature paths, hypothesizing that navigating irregular work zone paths is more challenging. As expected, $AE\% < \theta$ of predicted goal is lower $(-15.4\%)$ for this subset at tight thresholds $(0.5^{\circ}$ to $2^{\circ})$ . + +ject context significantly improves performance. Supp. C.2 shows that larger foundation models, e.g. LLaVA-NEXT [30], also perform poorly, highlighting the need for the ROADWork dataset. Fine-tuning VLMs on our dataset preserves the overall performance and even improves COCO-Captions [3] $(4.7\rightarrow 12.6$ BLEU@4). + +# 6. Driving Through Work Zones + +In Sections 3, 4 and 5, we studied challenges in recognizing, observing, and analyzing work zones. In this section, we discuss the challenges in driving through work zones and show how to predict goals and pathways in the image space to address these challenges. + +Pathway Prediction In Work Zone Images. Work zones demand long-term planning beyond bird's eye view (BEV) [6] representations, which are common for driving scenarios [9, 22, 45, 61] such as highways. Rare work zone objects and their spatial interactions are difficult to capture using BEV alone. Additionally, navigational goals in work zones are inherently multi-modal and cannot be effectively constrained within a fixed temporal window. Thus, we represent goals and pathways as spatial probability distributions in the image space, similar to long-horizon human path planning [40]. Perception cost maps for BEV planners [12, 39, 55] can be derived from these distributions. + +We employ YNet [40]. We pre-train the segmentation model that YNet employs using the Cityscapes [8] dataset which contains no work zone objects and with the ROAD-Work dataset. Both models are trained on pathways from our dataset. Pixel-level metrics [40] do not account for the field of view of images (See Supp. D). Thus, we employ Percentage of Paths within Angular Error threshold $\theta$ ( $AE\% < \theta$ ) as our metric. Table 8 presents our results. Our model, employing work zone semantic segmenta + +![](images/ef36d3dae56ab1ca72ac8ed272254984050edcc02e6c8cf672616b6350c0ef9b.jpg) +Figure 9. Pathway Prediction in Work Zones. We show goal and trajectory heatmaps predicted by YNet [40]. The model inputs are the image and the observed pathway. We display the future pathway (both computed from driving videos). The top row shows the predicted goal heatmaps, and the bottom row shows the predicted pathway heatmaps conditioned on a sampled goal. The predicted goal heatmap (highlighted with an orange box) closely aligns with the ground-truth, and the pathway appears plausible. + +tions is better at all $\theta$ thresholds, than the model employing Cityscapes semantic segmentations. For example, at threshold $0.5^{\circ}$ , our model improves AE% for goals by $+9.9\%$ and AE% for pathways by $+8.1\%$ (see Figure 9 for examples). We hypothesize navigating irregular work zone paths that deviate from a straight line is more difficult. Thus we also report results on subsets of paths thresholded at different curvatures. AE% at $0.5^{\circ}$ for goals on paths with high curvature is lower by $-15.4\%$ than all paths. Our dataset demonstrates that curved paths are indeed more difficult. + +Summary. The ROADWork dataset enables the prediction of long-term navigation goals and paths in work zones. We introduced a novel and important task for navigating work zones, while providing evaluation metrics and baselines. More analysis and details can be found in Supp. C.3. + +# 7. Conclusion + +We introduced the ROADWork dataset to address scene understanding and navigation in the challenging yet critical problem of self-driving in work zones. By mining thousands of hours of driving data, we curated nearly 5000 annotated work zones. Foundation models struggle on many work zone tasks. We highlighted several challenges, from work zone object recognition, to fine-grained scene understanding, to work zone description, to long-horizon pathway prediction. Our findings demonstrate the effectiveness of simple strategies such as video label propagation, copy-pasting, crop-rescaling, and providing object context as well as emphasize the need for better data-scarce algorithms. Our dataset, primarily sourced from smartphones, lacks multiple calibrated cameras seen on self-driving platforms. Not all traversable paths are captured, and 3D perception in work zones remains an open challenge. Despite these limitations, we hope our dataset inspires further research into this overlooked challenge. + +Acknowledgements: This work was supported by a research contract from General Motors Research-Israel, NSF Grant CNS-2038612, a US DOT grant 69A3551747111 through the Mobility21 UTC and grants 69A3552344811 and 69A3552348316 through the Safety21 UTC. We thank N. Dinesh Reddy, Shefali Srivastava, Neha Boloor, Tiffany Ma for insightful discussions. + +# 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. In CVPR, 2020. 2, 4 +[2] Robin Kien-Wei Chan, Krzysztof Lis, Svenja Uhlemeyer, Hermann Blum, Sina Honari, Roland Siegwart, Pascal Fua, Mathieu Salzmann, and Matthias Rottmann. Segmentmeifyoucan: A benchmark for anomaly segmentation. NeurIPS Track on Datasets and Benchmarks, 2021. 2 +[3] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 8 +[4] Bowen Cheng, Omkar Parkhi, and Alexander Kirillov. Pointly-supervised instance segmentation. In CVPR, 2022. 4 +[5] Chee Kheng Chng, Yuliang Liu, Yipeng Sun, Chun Chet Ng, Canjie Luo, Zihan Ni, ChuanMing Fang, Shuaiqiao Zhang, Junyu Han, Errui Ding, et al. Icdar2019 robust reading challenge on arbitrary-shaped text-rrc-art. In ICDAR, 2019. 6 +[6] Laurene Claussmann, Marc Revilloud, Dominique Gruyer, and Sébastien Glaser. A review of motion planning for highway autonomous driving. T-ITS, 2019. 8 +[7] Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025. 6, 7 +[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 CVPR, 2016. 1, 2, 5, 8 +[9] Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with misconceptions about learning-based vehicle motion planning. In CoRL, 2023. 8 +[10] Vinayak Dixit, Sai Chand, and Divya Nair. Autonomous vehicles: Disengagements, accidents and reaction times. PLOS ONE, 2016. 1 +[11] Joe Eskenazi. Waymo rolls toward san francisco airport. a showdown is brewing. https://missionlocal.org/2024/12/waymo-rolls-toward-san-francisco-airport-showdown-brewing/, 2024. 1 + +[12] Dave Ferguson, Thomas M Howard, and Maxim Likhachev. Motion planning in urban environments. Journal of Field Robotics, 2008. 8 +[13] David I Ferguson and Donald Jason Burnette. Mapping active and inactive construction zones for autonomous driving, 2015. US Patent 9,141,107. 2 +[14] David Ian Ferguson, Dirk Haehnel, and Ian Mahon. Construction zone object detection using light detection and ranging, 2015. US Patent 9,199,641. 2 +[15] 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. 6, 7 +[16] Regine Graf, Andreas Wimmer, and Klaus CJ Dietmayer. Probabilistic estimation of temporary lanes at road work zones. In ITSC, 2012. 2 +[17] Thomas Gumpp, Dennis Nienhuser, Rebecca Liebig, and J Marius Zollner. Recognition and tracking of temporary lanes in motorway construction sites. In Intelligent Vehicles Symposium, 2009. 2 +[18] Sean Harris. Cruise's continuous learning machine predicts the unpredictable on san francisco roads, 2020. 1 +[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2, 4, 5, 6 +[20] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joe Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 2 +[21] 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 ICLR, 2022. 7 +[22] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In CVPR, 2023. 8 +[23] Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. In CVPR, 2017. 5 +[24] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 6, 7 +[25] Michelin Mobility Intelligence. Roadbotics open data set. https://www.roadbotics.com/2021/03/15/roadbotics-open-data-set/, 2021.3 +[26] Mikhail Kennerley, Jian-Gang Wang, Bharadwaj Veeravalli, and Robby T Tan. 2pcnet: Two-phase consistency training for day-to-night unsupervised domain adaptive object detection. In CVPR, 2023. 5 +[27] Jinwoo Kim, Kyounghwan An, and Donghwan Lee. Rosa dataset: Road construct zone segmentation for autonomous driving. In ECCV 2024 Workshop on Multimodal Perception and Comprehension of Corner Cases in Autonomous Driving, 2024. 2 + +[28] 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. 2, 3, 4 +[29] Michael Levenson. Driverless car gets stuck in wet concrete in san francisco. https://www.nytimes.com/2023/08/17/us/driverless-car-accident-sf.html, 2023.1 +[30] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llava-last: Stronger llms supercharge multimodal capabilities in the wild, 2024. 2, 7, 8 +[31] 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 CVPR, 2023. 4, 6 +[32] 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. 2, 7 +[33] Kaican Li, Kai Chen, Haoyu Wang, Lanqing Hong, Chaoqiang Ye, Jianhua Han, Yukuai Chen, Wei Zhang, Chunjing Xu, Dit-Yan Yeung, et al. Coda: A real-world road corner case dataset for object detection in autonomous driving. In ECCV, 2022. 1, 2 +[34] Mengtian Li, Yu-Xiong Wang, and Deva Ramanan. Towards streaming perception. In ECCV, 2020. 5 +[35] Mingfu Liang, Jong-Chyi Su, Samuel Schulter, Sparsh Garg, Shiyu Zhao, Ying Wu, and Manmohan Chandraker. Aide: An automatic data engine for object detection in autonomous driving. In CVPR, 2024. 1 +[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 7 +[37] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 7 +[38] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In ECCV, 2024. 2, 4, 5 +[39] David V Lu, Dave Hershberger, and William D Smart. Layered costmaps for context-sensitive navigation. In IROS, 2014. 8 +[40] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In ICCV, 2021. 8 +[41] Aarian Marshall. Why self-driving cars *can't even* with construction zones. https://www.wired.com/2017/02/self-driving-cars-cant-even-construction-zones/, 2017.1 +[42] Aarian Marshall. An autonomous car blocked a fire truck responding to an emergency. https://www.wired.com/story/cruise-fire-truck-block-san-francisco-autonomous-vehicles/, 2022.1 + +[43] Bonolo Mathibela, Ingmar Posner, and Paul Newman. A roadwork scene signature based on the opponent colour model. In IROS, 2013. 2 +[44] Matt McFarland. Traffic cones confused a waymo self-driving car. then things got worse. https://www.cnn.com/2021/05/17/tech/waymo-arizona-confused/, 2021.1 +[45] Matthew McNaughton, Chris Urmson, John M Dolan, and Jin-Woo Lee. Motion planning for autonomous driving with a conformal spatiotemporal lattice. In ICRA, 2011. 8 +[46] J.F. Morgan, A.R. Duley, and P.A. Hancock. Driver responses to differing urban work zone configurations. Accident Analysis & Prevention, 2010. 1 +[47] Power Nation TV. Robotaxi causes traffic jam after attempting to drive through construction site; elon Musk responds. https://www.powernationtv.com/post/robotaxi-traffic-elon-musk-responds,2023.1 +[48] National Safety Council Injury Facts. Motor Vehicle Safety Issues, Visited on March 1, 2024. 1 +[49] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 1, 2, 3, 4, 5 +[50] David Pannen, Martin Liebner, Wolfgang Hempel, and Wolfram Burgard. How to keep hd maps for automated driving up to date. In ICRA, 2020. 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 ICML, 2021. 4, 7 +[52] Rahul Ramachandran, Ali Garjani, Roman Bachmann, Andrei Atanov, Oğuzhan Fatih Kar, and Amir Zamir. How well does gpt-4o understand vision? evaluating multimodal foundation models on standard computer vision tasks. arXiv preprint arXiv:2507.01955, 2025. 6 +[53] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dolar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 3, 4, 6, 7 +[54] Roi Ronen, Shahar Tsiper, Oron Anschel, Inbal Lavi, Amir Markovitz, and R Manmatha. Glass: Global to local attention for scene-text spotting. In ECCV, 2022. 2, 6 +[55] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In ECCV, 2020. +[56] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 3 +[57] Weijing Shi and Ragunathan Raj Rajkumar. Work zone detection for autonomous vehicles. In ITSC, 2021. 2 +[58] 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 CVPR, 2021. 6 + +[59] Li Su, Howard Bowman, and Philip Barnard. Glancing and then looking: on the role of body, affect, and meaning in cognitive control. Frontiers in psychology, 2011. 2 +[60] 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 CVPR, 2020. 2 +[61] Siyu Teng, Xuemin Hu, Peng Deng, Bai Li, Yuchen Li, Yunfeng Ai, Dongsheng Yang, Lingxi Li, Zhe Xuanyuan, Fenghua Zhu, et al. Motion planning for autonomous driving: The state of the art and future perspectives. T-IV, 2023. 8 +[62] Diwas Thapa, Sabyasachee Mishra, Asad Khattak, and Muhammad Adeel. Assessing driver behavior in work zones: A discretized duration approach to predict speeding. Accident Analysis & Prevention, 2024. 1 +[63] U.S. Department of Transportation, National Highway Traffic Safety Administration. FARS Data: People Killed in Construction or Maintenance Zones, Visited on March 1, 2024. 1 +[64] Jun Wang, Warren E. Hughes, Forrest M. Council, and Jeffrey F. Paniati. Investigation of highway work zone crashes: What we know and what we don't know. Transportation Research Record, 1996. 1 +[65] Zhenyu Wang, Yali Li, Xi Chen, Ser-Nam Lim, Antonio Torralba, Hengshuang Zhao, and Shengjin Wang. Detecting everything in the open world: Towards universal object detection. In CVPR, 2023. 2, 5 +[66] Matthew Wilson, Tim Zaman, and Long Tran. Clip search with multimodal queries, 2025. Tesla, US Patent US20240419724A1. 1 +[67] Andreas Wimmer, Thorsten Weiss, Francesco Flogel, and Klaus Dietmayer. Automatic detection and classification of safety barriers in road construction sites using a laser scanner. In Intelligent Vehicles Symposium, 2009. 2 +[68] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 4, 6 +[69] 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 CVPR, 2020. 1, 2, 3, 5 +[70] Oliver Zendel, Katrin Honauer, Markus Murschitz, Daniel Steininger, and Gustavo Fernandez Dominguez. Wilddash-creating hazard-aware benchmarks. In ECCV, 2018. 1, 2 +[71] Oliver Zendel, Matthias Schörghuber, Bernhard Rainer, Markus Murschitz, and Csaba Beleznai. Unifying panoptic segmentation for autonomous driving. In CVPR, 2022. 1, 2 +[72] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In ICCV, 2023. 4, 6 + +[73] 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. NeurIPS, 2024. 4, 6 +[74] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 2, 4, 5, 6 +[75] Julie Zigoris. Driverless waymo car almost digs itself into hole—literally. https://sfstandard.com/2023/01/15/driverless-waymo-car-digs-itself-into-hole-literally/, 2023. 1 +[76] David Zipper. San francisco has a problem with robotaxis. https://www.theatlantic.com/ideas/archive/2023/08/robotaxis-san-francisco-self-driving-car/674956/, 2023.1 \ No newline at end of file diff --git a/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/images.zip b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..03acabd9cd04be32fc61b79092b28ce88cb95813 --- /dev/null +++ b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4c6ae55b3acb96d651aa3556aa65ce27d17c3f31fc6004e3adca134a5895d70 +size 808531 diff --git a/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/layout.json b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..395d868b4fb73023ab365a2874fc324d943b8dff --- /dev/null +++ b/ICCV/2025/ROADWork_ A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ff9d5b2c0cdd10f8e3232a8bdf8a9d47d6b5d774b42f522633c2694ea0de27f +size 436399 diff --git a/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/967e3628-761e-4c34-b1f5-4f310b06c46e_content_list.json b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/967e3628-761e-4c34-b1f5-4f310b06c46e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4f0c81d26ee7ee403db581f97f0fbb32b8b30bac --- /dev/null +++ b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/967e3628-761e-4c34-b1f5-4f310b06c46e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c0ba239cec22eaa0530be24bd82737f31846d7d87ccbb2b49fafdd1fc95321c +size 80588 diff --git a/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/967e3628-761e-4c34-b1f5-4f310b06c46e_model.json b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/967e3628-761e-4c34-b1f5-4f310b06c46e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..47e10f7b3a41e89fc035cca28bf24b3985eb0f12 --- /dev/null +++ b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/967e3628-761e-4c34-b1f5-4f310b06c46e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a96f81f841d591c07ff8a6034be9f488882c00abab126b281481505a8d2cf468 +size 95886 diff --git a/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/967e3628-761e-4c34-b1f5-4f310b06c46e_origin.pdf b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/967e3628-761e-4c34-b1f5-4f310b06c46e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..46f963b288789005ac2d60effe997f441a8e6440 --- /dev/null +++ b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/967e3628-761e-4c34-b1f5-4f310b06c46e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:900205a4361d826638a555b4c44d52325b470b6426d6a744c91b3c038f840a63 +size 1162802 diff --git a/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/full.md b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9d7d7400927fbfef4ff149e198a8960bbdc8c025 --- /dev/null +++ b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/full.md @@ -0,0 +1,326 @@ +# ROAR: Reducing Inversion Error in Generative Image Watermarking + +Hanyi Wang $^{1}$ Han Fang $^{2,*}$ Shi-Lin Wang $^{1}$ Ee-Chien Chang $^{2}$ $^{1}$ Shanghai Jiao Tong University $^{2}$ National University of Singapore + +why_820@sjtu.edu.cn, fanghan@nus.edu.sg, wsl@sjtu.edu.cn, changec@comp.nus.edu.sg + +# Abstract + +Generative image watermarking enables the proactive detection and traceability of generated images. Among existing methods, inversion-based frameworks achieve highly concealed watermark embedding by injecting watermarks into the latent representation before the diffusion process. The robustness of this approach hinges on both the embedding mechanism and inversion accuracy. However, prior works have predominantly focused on optimizing the embedding process while overlooking inversion errors, which significantly affect extraction fidelity. In this paper, we address the challenge of inversion errors and propose ROAR, a dual-domain optimization-based framework designed to mitigate errors arising from two key sources: 1) Latent-domain errors, which accumulate across inversion steps due to inherent approximation assumptions. 2) Pixel-domain errors, which result from channel distortions such as JPEG compression. To tackle these issues, we introduce two novel components: A Regeneration-based Optimization (RO) mechanism, which incorporates an estimable starting latent to minimize latent-domain errors; A Mixture of Experts (MoE)-based distortion-adaptive restoration (AR) network, which effectively recovers watermarked distributions from pixel-level distortions. Extensive experiments demonstrate that ROAR significantly reduces inversion errors and enhances watermark extraction robustness, thereby improving the reliability of generative image watermarking. + +# 1. Introduction + +Recent advancements in diffusion models [7, 15] have significantly improved the quality of generated images. Modern text-to-image latent diffusion models, such as Stable Diffusion [27], have further streamlined the process, allowing non-experts to create high-quality images directly from text prompts. However, these powerful capabilities also raise concerns about false content fabrication. Conse + +![](images/32c55b13b11b72a0afe81e76cdd899d3cd4216694acc901739231376be925758.jpg) +Figure 1. The framework of inversion-based generative image watermarking method and proposed ROAR. + +sequently, ensuring the detectability and traceability of generated images or their underlying models, as well as identifying responsible users, has become an urgent necessity. + +Generative image watermarking [5, 10, 18, 26, 32, 33] offers a promising solution to these issues. By injecting watermarks into generated images, detection and source attribution are achieved through extracted watermark. Among all these solutions, the inversion-based framework has emerged as a particularly effective approach. The general structure of this framework is depicted in Fig. 1. In the embedding phase, the watermark signal is first injected into the initial starting latent of the diffusion model through a robust embedding mechanism. Then, the forward sampling process begins with the watermarked latent and iteratively generates images with imperceptible watermarks. For watermark extraction, an inversion technique is employed to reverse the generated image back to its origin latent, enabling the extraction of the watermark. + +The robustness of inversion-based watermarking frameworks is inherently dependent on both the embedding scheme and inversion accuracy. Existing methods [5, 32, 33] primarily focus on designing robust embedding mechanisms while relying on standard Denoising Diffusion Implicit Models (DDIM) [30] for inversion. However, we argue that while DDIM inversion establishes a baseline for watermark robustness, significant improvements in precision and reliability remain unaddressed. + +Through our investigation, we identify two primary sources of inversion errors: (1) Latent-domain errors: DDIM inversion assumes that noise predictions at consecutive time steps are approximately equal, i.e., $\epsilon_{\theta}(x_t,t)\approx$ + +$\epsilon_{\theta}(x_{t - 1},t)$ . While this assumption theoretically minimizes discrepancies, even small errors per step accumulate across multiple iterations, leading to significant deviations that degrade inversion precision. (2) Pixel-domain errors: Transmission distortions introduce pixel-level artifacts, altering the starting point for inversion and further compromising accuracy. + +To address these issues, we propose ROAR, a dual-domain optimization mechanism designed to mitigate both latent- and pixel-domain errors. Specifically, we introduce: + +Regeneration-based Optimization (RO) strategy, which aims to optimize the starting latent such that its generated image closely matches the watermarked image. By enforcing similarity between the regenerated and target extraction images, we can approximately reconstruct every internal latent $\tilde{x}_t$ at step $t$ that strikingly similar to the original watermarked latent $x_t$ , ensuring $\epsilon_{\theta}(x_t,t)\simeq \epsilon_{\theta}(\tilde{x}_t,t)$ . Such a optimization process mitigates errors introduced by inversion approximations and improves inversion accuracy from a forward generation perspective. + +Distortion-Adaptive Restoration (AR) network based on Mixture of Experts (MoE), designed to restore distorted images before inversion. Trained on diverse noise layers, this network dynamically selects specialized experts to adaptively correct different types of distortions. By improving image restoration, it provides a more reliable starting point for inversion, thereby enhancing watermark extraction accuracy. + +In summary, we make the following contributions: + +- We proposed ROAR, a dual-domain optimization-based approach that effectively reduces inversion errors, thereby improving the robustness of inversion-based generative watermarking. Besides, ROAR is fully compatible with existing inversion-based watermarking frameworks. +- To mitigate latent-domain errors, we introduce a regeneration-based optimization (RO) mechanism, which enhances inversion accuracy by refining the forward generation process. To address pixel-domain errors, we design a MoE-based distortion-adaptive restoration (AR) network, which provides a more reliable starting point for inversion. +- Extensive experimental results demonstrate that ROAR significantly reduces inversion errors and enhances the robustness of inversion-based generative image watermarking. Our approach achieves an average extraction accuracy of $99.71\%$ , outperforming state-of-the-art methods by $10.66\%$ . + +# 2. Related Work + +# 2.1. Digital watermarking + +Digital watermarking is an effective technique for addressing copyright protection and content authentication by em + +bedding copyright or traceable identification information directly into carrier data. The functionality of a watermark typically depends on its capacity. A single-bit watermark can indicate whether an image was generated by a specific diffusion model, ensuring detectability, while a multibit watermark can encode additional information to identify users responsible for generation, enabling traceability. Traditional image watermarking often embeds watermark into the transform domain[1, 12, 19] to achieve a balance between fidelity and robustness. In recent years, deep learning-based image watermarking techniques[3, 9, 16, 39] have demonstrated significant potential due to their powerful learning capabilities. + +# 2.2. Diffusion models + +Diffusion models [15] have become the leading architecture for generative tasks such as image generation [7], editing [17], and inpainting [22]. By iteratively denoising, they generate high-quality images. To enhance computational efficiency, latent diffusion models (LDMs), such as Stable Diffusion [27], perform generation in the latent space using a pre-trained Variational Autoencoder (VAE). To further reduce sampling steps, DDIM [30] employs an ODE solver for deterministic sampling. This deterministic nature enables the inversion of latent noise from the generated image, which is a crucial property for applications such as generative image watermarking. + +# 2.3. Generative image watermarking + +Generative image watermarking techniques for diffusion models, which embed watermarks into the generated images, can be broadly classified into three main categories: 1). Post-hoc Methods: These methods[6, 8, 34] involve applying traditional image watermarking techniques directly to diffusion-generated images. 2). Fine-tuning-based Methods: These methods mainly focus on fine-tuning the VAE decoder in LDMs to embed watermarks. For example, Stable Signature [10] uses a pre-trained watermark extractor to guide fine-tuning, while LaWa [26] integrates coarse-to-fine watermarking modules into frozen intermediate layers of the VAE decoder. 3). Inversion-based Methods: These methods embed watermarks into the initial latent representation before the forward sampling process generates watermarked images. Tree-Ring [32] encodes a 1-bit watermark in the Fourier domain of a random latent, while RingID [5] employs a multi-channel heterogeneous watermarking framework for enhanced identification. Additionally, [33] proposes a distribution-preserving sampling approach to map multi-bit watermarks into pseudo-random latents. For watermark extraction, existing inversion-based methods rely on DDIM inversion [30] to recover the initial latent representation, enabling the retrieval of the embedded watermark. + +# 3. Preliminary + +# 3.1. DDIM + +Denoising diffusion models [15] are trained using a denoising objective. During the forward diffusion process, images (or autoencoded latents) $x_0 \in X$ are progressively noised over $T$ steps by adding Gaussian noise $\epsilon \sim \mathcal{N}(0, 1)$ according to a monotonically increasing noise schedule $\{\alpha_t\}_{t=0}^T$ . At each step $t$ , the forward process is defined as: + +$$ +x _ {t} = \sqrt {\alpha_ {t}} x _ {0} + \sqrt {1 - \alpha_ {t}} \epsilon . \tag {1} +$$ + +The diffusion model $\epsilon_{\theta}$ is trained to predict the added noise $\epsilon$ at each step by minimizing the $l_{2}$ -norm $\| \epsilon_{\theta}(x_t,t) - \epsilon \| _2^2$ . Once trained, the model employs an iterative reverse diffusion process over $T$ steps, starting from random noise, to generate images. A commonly used sampling scheme is DDIM [30], which calculates intermediate steps as follows: + +$$ +x _ {t - 1} = \sqrt {\frac {\alpha_ {t - 1}}{\alpha_ {t}}} x _ {t} - (\sqrt {\frac {\alpha_ {t - 1} \left(1 - \alpha_ {t}\right)}{\alpha_ {t}}} + \sqrt {1 - \alpha_ {t - 1}}) \epsilon_ {\theta} (x _ {t}, t). \tag {2} +$$ + +For simplicity, this expression can be reformulated as: + +$$ +x _ {t - 1} = \gamma_ {t} x _ {t} + \varphi_ {t} \epsilon_ {\theta} \left(x _ {t}, t\right), \tag {3} +$$ + +where $\gamma_{t} = \sqrt{\alpha_{t - 1} / \alpha_{t}}$ and $\varphi_t = -\sqrt{\alpha_{t - 1}(1 - \alpha_t) / \alpha_t} +$ $\sqrt{1 - \alpha_{t - 1}}$ + +In the classifier-free guidance mechanism [14], the image generation process is conditioned on text input $C$ and controlled by a guidance scale $w$ (typically $w > 1$ ), which modulates the influence of the conditioning. Under this mechanism, the forward sampling process is given by: + +$$ +x _ {t - 1} = \gamma_ {t} x _ {t} + \varphi_ {t} \left\{\epsilon_ {\theta} \left(x _ {t}, t, \emptyset\right) + w \left[ \epsilon_ {\theta} \left(x _ {t}, t, C\right) - \epsilon_ {\theta} \left(x _ {t}, t, \emptyset\right) \right] \right\}. \tag {4} +$$ + +# 3.2. DDIM Inversion + +As outlined in DDIM [30], the denoising process is approximately invertible. Typically, during the inversion process, the specific prompt used in the generation process is generally unknown. As a result, inversion is performed under a "Null" condition $(C = \emptyset)$ with no guidance scale applied $(w = 1)$ . Under these assumptions, $x_{t}$ can be recovered from $x_{t-1}$ using the following equation: + +$$ +x _ {t} = \frac {1}{\gamma_ {t}} x _ {t - 1} - \frac {\varphi_ {t}}{\gamma_ {t}} \epsilon_ {\theta} (x _ {t}, t, \emptyset). \tag {5} +$$ + +To simplify the inversion process, a key approximation is introduced: $\epsilon_{\theta}(x_t,t,\emptyset)\approx \epsilon_{\theta}(x_{t - 1},t,\emptyset)$ [31]. This assumption suggests that the noise predictions at consecutive steps are approximately equal. Applying this assumption yields + +![](images/16531806f333f51058c2dfcae3b9b4be526d407a40d7d6fe002ad125dbd0ce28.jpg) +Figure 2. Schematic illustration of generation, DDIM inversion, and our optimization mechanism. $X_{T}, \ldots, X_{0}$ denote forward sampling steps under classifier-free guidance. $X_{1}', \ldots, X_{T}'$ represent DDIM inversion steps. $X_{T}^{*}, \ldots, X_{1}^{*}$ indicate regenerated steps from our optimization strategy. + +the following approximation for the reverse representation of $x_{t}$ , denoted as $x_{t}^{\prime}$ : + +$$ +x _ {t} ^ {\prime} \approx \frac {1}{\gamma_ {t}} x _ {t - 1} - \frac {\varphi_ {t}}{\gamma_ {t}} \epsilon_ {\theta} (x _ {t - 1}, t, \emptyset). \tag {6} +$$ + +Notably, existing inversion-based generative watermarking methods rely entirely on this process to recover the watermarked initial latent representations $x_{T}$ , from which the embedded watermark is then extracted. + +# 4. Method + +# 4.1. Analysis and solutions + +The primary objective of our method is to minimize inversion errors. To achieve this, we first analyze the fundamental causes of both latent and pixel-domain errors in the inversion process, and provide our intuitive solutions. + +Latent-Domain Errors. During the generation process, the image synthesis is guided by a text prompt $C$ , unconditional prompt $\emptyset$ and the guidance scale $w$ , as formulated in Eq. 4. However, in the inversion process, due to the absence of $C$ and $w$ , inversion can only be performed under $\emptyset$ with $w = 1$ , as shown in Eq. 6. This discrepancy leads to two types of inversion errors, as illustrated in Eq. 7. + +$$ +\begin{array}{l} x _ {t} ^ {\prime} - x _ {t} = \frac {\varphi_ {t}}{\gamma_ {t}} \left\{\underbrace {w \left[ \epsilon_ {\theta} \left(x _ {t} , t , C\right) - \epsilon_ {\theta} \left(x _ {t} , t , \emptyset\right) \right]} _ {\text {C o n d i t i o n D r i f t}} \right. \tag {7} \\ + \underbrace {\epsilon_ {\theta} (x _ {t} , t , \emptyset) - \epsilon_ {\theta} (x _ {t -}} _ {\text {E s t i m a t i o n E r r o r}} \\ \end{array} +$$ + +Condition Drift, arising from the mismatch between the conditions used during generation and inversion; Estimation Error, resulting from the approximation of $\epsilon_{\theta}(x_t,t,\emptyset)$ across consecutive time steps. To facilitate a clearer understanding, we provide a schematic representation in Fig. 2, where unconditional and conditional generations are depicted as horizontal and vertical vectors, respectively. Given that accurately recovering $C$ and $w$ from a single image in real-world scenarios is exceedingly difficult, our + +![](images/3c4e7bffdd8d74bcaba7c012f4db52712d6b809febea77712a6d5dcd641f03e7.jpg) +Figure 3. The framework of the proposed ROAR, which consists of two key components: a distortion-adaptive restoration network $(\Theta^{\mathrm{AR}})$ and an regeneration-based optimization module (RO). During the training stage, only $\Theta^{\mathrm{AR}}$ is trained. In testing stage, watermark extraction is performed using the frozen $\Theta^{\mathrm{AR}}$ along with the regeneration-based optimization module. + +Regeneration-based Optimization (RO) approach shifts the focus to mitigating estimation errors. Specifically, RO aims to reduce the accumulated errors introduced by $\varnothing$ -condition inversion. + +To achieve this, we optimize the initial latent such that the image regenerated under the unconditional setting aligns closely with the watermarked image. By enforcing pixel-level similarity constraints, the optimized latent is effectively guided toward the truly starting point along the unconditional inversion trajectory. We provide a mathematical analysis in the supplementary material to offer deeper theoretical insights and further validate the effectiveness of our approach. + +Pixel-domain Errors. Distortions in the image domain inevitably induce distribution shifts, causing significant deviations in the latent domain via direct inversion. Therefore, addressing image domain distortions before the inversion process is crucial. To this end, we propose employing a distortion-adaptive restoration network to align distorted images with their original watermarked distribution. + +Roughly, we categorize distortions into two types: entropy-increasing distortions (e.g., noise) and entropy-decreasing distortions (e.g., blurring, JPEG compression). To effectively address both, we design a MoE-based distortion-adaptive restoration network, which dynamically assigns specialized experts to handle various types of distortion. This adaptive restoration process helps ensure that the restored images are more closely aligned with the expected distribution before latent-domain optimization, providing a more favorable initialization for subsequent optimization. + +# 4.2. Framework overview + +By integrating the aforementioned solutions, we construct the proposed framework, as illustrated in Fig. 3. The framework comprises five key components: a robust embedding mechanism $\mathcal{F}$ , a diffusion model $\mathcal{G}$ coupled with a pretrained VAE $\{\mathcal{E}_{\mathrm{VAE}}, \mathcal{D}_{\mathrm{VAE}}\}$ , a distortion-adaptive restoration network $\Theta^{\mathrm{AR}}$ , and a regeneration-based optimization module, denoted as RO. Here, we define $\overrightarrow{\mathcal{G}}$ as the forward sampling process and $\overleftarrow{\mathcal{G}}$ as the DDIM inversion. + +The workflow of the proposed framework is as follows: Given a watermark $m$ , the embedding mechanism $\mathcal{F}$ is first applied to generate the watermarked initial latent $z_T^m = \mathcal{F}(m)$ . Next, the diffused latent, conditioned on $C$ and guided by scale $w$ , is iteratively refined over $T$ steps, producing $z_0^m = \overrightarrow{\mathcal{G}}(z_T^m, T, C, w)$ . Finally, $z_0^m$ is processed by $\mathcal{D}_{\mathrm{VAE}}$ to generate the watermarked image $X^m = \mathcal{D}_{\mathrm{VAE}}(z_0^m)$ . In the pixel domain, $X^m$ undergoes various types of distortions, producing the distorted image $X_{Dis}^m$ . This distorted image is then passed through our proposed distortion-adaptive restoration network, yielding the restored image $\tilde{X}^m = \Theta^{\mathrm{AR}}(X_{Dis}^m)$ . Subsequently, $\tilde{X}^m$ is fed into $\mathcal{E}_{\mathrm{VAE}}$ to obtain the restored latent $\tilde{z}_0^m = \mathcal{E}_{\mathrm{VAE}}(\tilde{X}^m)$ . Next, DDIM inversion is performed on $\tilde{z}_0^m$ for $V$ steps with an empty condition $\emptyset$ , generating the inverted latent $\tilde{z}_V^m = \overleftarrow{\mathcal{G}}(\tilde{z}_0^m, V, \emptyset)$ . To reduce accumulated inversion errors and enhance reconstruction fidelity, the result of DDIM inversion serves as the initialization point $z_{start}^*$ for subsequent optimization process. This optimization is conducted over $K$ iterations using the proposed regeneration-based optimization module, $\mathcal{R}$ , to obtain $z_{final}^* = \mathcal{R}(z_{start}^*, K)$ . Fi + +nally, the optimized latent representation is used to extract the recovered watermark $m^{*} = \mathcal{F}^{-1}(z_{final}^{*})$ . + +# 4.3. Training of $\Theta^{\mathrm{AR}}$ + +The distortion-adaptive restoration network $\Theta^{\mathrm{AR}}$ is designed based on the MoE [29] architecture, which dynamically assigns specialized experts to handle both entropy-increasing and entropy-decreasing distortions. Each expert network follows a hierarchical structure to extract multiscale features. It begins with a convolutional layer as the initial feature extractor, followed by a series of Conv-ReLU blocks for non-linear feature transformation. To enhance feature representation, channel attention blocks [36] are incorporated, allowing the network to emphasize informative channels while suppressing redundant ones. The backbone of the network is built upon a standard U-Net [28], which combines an encoder-decoder structure for both global context understanding and precise spatial feature recovery. Instead of using transposed convolutions for upsampling, we employ bilinear interpolation followed by a convolutional layer in the decoder to mitigate checkerboard artifacts [24], resulting in smoother feature interpolation and more stable reconstructions. + +To train $\Theta^{\mathrm{AR}}$ , we minimized the difference between the restored image $X_{Res}$ and the clean image $X_{Cln}$ . The objective function is formulated as: + +$$ +\mathcal {L} _ {\mathrm {A R}} = | | X _ {R e s} - X _ {C l n} | | _ {2} ^ {2} = | | \Theta^ {\mathrm {A R}} (X _ {D i s}) - X _ {C l n} | | _ {2} ^ {2}, +$$ + +where $X_{Dis}$ is the distorted image. + +# 4.4. Regeneration-based Optimization + +In the optimization process, we employ the latent representation obtained from DDIM inversion as the starting point for iterative optimization. As previously analyzed, at each step, the latent representation is forward-sampled under an empty condition $\varnothing$ to generate an image. To optimize the reversed latent representation, we minimize the discrepancy between the regenerated image and the watermarked image. This optimization objective is formulated as: + +$$ +\mathcal {L} _ {\mathrm {R O}} = | | \mathcal {D} _ {\mathrm {V A E}} (\overrightarrow {\mathcal {G}} (z _ {V} ^ {*}, T, \emptyset)) - \tilde {X} ^ {m} | | _ {2} ^ {2}, +$$ + +where $z_{V}^{*}$ represents the optimized latent variable, and $\tilde{X}^{m}$ denotes the restored watermarked image obtained from the distortion-adaptive restoration network. By iteratively updating $z_{V}^{*}$ using gradient-based optimization over $K$ steps, we progressively reduce the accumulated inversion errors. + +# 5. Experiments + +# 5.1. Experimental setup + +Implementation details. We conduct experiments using Stable Diffusion (SD) [27], provided by Hugging Face, + +![](images/cf4a3b21ccd0c66a8c3cd88d98b7d8f9c30e78c8194ebc09b8c8f6c4e5bb4fae.jpg) +Figure 4. Histogram between $\| x_{T} - x_{T}^{*}\|$ and $\| x_{T} - x_{T}^{\prime}\|$ + +evaluating two widely used versions: SD-v1.4 and SD-v2.1. The generated images have a resolution of $512 \times 512 \times 3$ , with corresponding latent features of size $4 \times 64 \times 64$ . During generation, we use prompts from the SD-Prompt [11]. To train $\Theta^{\mathrm{AR}}$ , we introduce entropy-increasing distortions, including Gaussian Noise ( $\mu = 0$ , $\sigma = 0.05$ ), and Salt-and-Pepper Noise ( $p = 0.05$ ); as well as entropy-decreasing distortions, including Median Filtering ( $k = 7$ ), and JPEG Compression (Quality Factor $QF = 25$ ). During optimization, we perform DDIM inversion and reconstruction under the $\emptyset$ -condition with 50 steps. The iteration count $K$ is set to 20. We use the Adam optimizer with an initial learning rate of $5 \times 10^{-3}$ , gradually reduced to $1 \times 10^{-6}$ via a cosine annealing learning rate scheduler. For robust embedding $\mathcal{F}$ , we employ Tree-Ring [32] and RingID [5] for embedding 1-bit message, while Gaussian Shading [33] is used for embedding 256-bit binary sequences. + +Evaluation metrics. To assess the robustness of ROAR, we follow [33], considering both detection and traceability scenarios. For detection, we measure the true positive rate (TPR) at a fixed false positive rate (FPR) of $10^{-10}$ . For traceability, we use extraction bit accuracy. Additionally, we evaluate watermarked image quality using Fréchet Inception Distance (FID) [13] and CLIP-Score [25], comparing them to watermark-free images. All evaluations are performed on 1000 watermarked images generated from randomly sampled prompts in the SD-Prompt dataset. + +Baselines. We compare ROAR with 8 state-of-the-art generative model watermarking frameworks, categorized into three groups: (1) Image watermarking-based methods, including three officially used by SD, namely DwtDct[6], DwtDctSvd [6] and RivaGAN [34]; (2) Fine-tuning-based methods, comprising Stable Signature [10] and LaWa [26]; and (3) Inversion-based methods, including Tree-Ring [32], RingID [5], and Gaussian Shading (GS) [33]. + +# 5.2. Effectiveness of the Optimization Mechanism + +We begin by numerically evaluating whether the optimization strategy effectively reduces the absolute error between the optimized latent representation, $x_{T}^{*}$ , and the initial latent representation, $x_{T}$ , used for image generation. Specifically, we compare the post-optimization distance, denoted as $\| x_{T} - x_{T}^{*}\|$ , with the distance obtained from DDIM inversion, represented as $\| x_{T} - x_{T}^{\prime}\|$ , where $x_{T}^{\prime}$ is the latent + +
MethodsMetrics
TPR (Cln.)TPR (Dis.)Bit Acc. (Cln.)Bit Acc. (Dis.)FID ↓CLIP-Score ↓
Stable Diffusion----25.23±.180.3629±0.0006
DwtDct [6]0.800/0.7500.111/0.1080.839/0.7990.541/0.54024.97±.190.3419±0.0630/0.3590±0.0420
DwtDctSvd [6]1.000/1.0000.325/0.3350.999/0.9990.636/0.63424.45±.220.3381±0.0622/0.3550±0.0427
RivaGAN [34]0.950/0.9600.344/0.3670.973/0.9910.771/0.78324.24±.160.3389±0.0728/0.3574±0.0431
Stable Signature [10]0.940/1.0000.279/0.2740.964/0.9940.646/0.65225.45±.180.3410±0.0018/0.3622±0.0023
LaWa [26]1.000/1.0000.258/0.2641.000/1.0000.775/0.77725.42±.160.3503±0.0454/0.3659±0.0401
Tree-Ring [32]1.000/1.0000.838/0.878--25.43±.130.3634±0.0006/0.3653±0.0002
RingID [5]1.000/1.0000.894/0.899--25.43±.170.3526±0.0028/0.3663±0.0012
GS [33]1.000/1.0000.912/0.9241.000/1.0000.885/0.89425.20±.220.3597±0.0009/0.3638±0.0005
Tree-Ring-ROAR1.000/1.0001.000/1.000--25.43±.130.3634±0.0006/0.3653±0.0002
RingID-ROAR1.000/1.0001.000/1.000--25.43±.170.3526±0.0028/0.3663±0.0012
GS-ROAR1.000/1.0001.000/1.0001.000/1.0000.988/0.99025.20±.220.3597±0.0009/0.3638±0.0005
+ +Table 1. Comparison results. We control the FPR at $10^{-10}$ and evaluate the TPR and bit accuracy for SD V1.4/2.1. "Cln" indicates the results without distortions, while "Dis" represents the average results across the 9 tested distortions. + +recovered via DDIM inversion. The histogram presented in Figure 4 illustrates that the optimized distance is generally smaller than the inversion-induced distance. Moreover, the distribution of optimized distances is more concentrated in the lower range. This indicates that the optimization strategy effectively reduces the inversion error. + +# 5.3. Comparison to baselines + +In this section, we access the performance of ROAR. Specifically, we follow [33] considering nine representative types of distortions, as illustrated in Fig. 5. The results are presented in Tab. 1. Tree-Ring-ROAR, RingID-ROAR, and GS-ROAR represent the integration of our proposed mechanism into the inversion-based methods Tree-Ring, RingID, and GS, respectively. + +Experimental results demonstrate that our approach significantly enhances robustness while fully preserving the inherent characteristics of inversion-based methods, measured by CLIP-Score and FID for generation quality and diversity. Compared to all baseline methods, our method achieves notable improvements in both bit accuracy and detection TPR. Specifically, it increases the overall average TPR of distorted images by $17\%$ compared to Tree-Ring and RingID. Additionally, it improves TPR and bit accuracy by $13\%$ and $11\%$ , respectively, relative to GS. Notably, our method is designed as a plug-and-play solution, ensuring seamless compatibility with existing inversion-based watermarking frameworks without modifying their original embedding mechanisms or other intrinsic characteristics. + +To further analyze the improvements, we evaluate the robustness of Tree-Ring, RingID, and GS against five challenging distortion types: Gaussian Noise (with $\sigma$ ranging from 0.05 to 0.2), JPEG Compression (with $QF$ between 5 and 30), Median Filtering (with $w$ from 11 to 21), Salt & Pepper Noise (with $p$ from 0.05 to 0.3), and Gaussian Blurring (with $r$ from 4 to 9). The results are presented in Fig. 6. ROAR demonstrates significant robustness improvements across all three methods, particularly under se + +![](images/cbcc43f6bd55202ca0f467aa27511df839248953b34c02f599683d094313f48a.jpg) +Figure 5. (a) Watermarked image; (b) JPEG, $QF = 25$ ; (c) $60\%$ area Random Crop; (d) $80\%$ area Random Drop; (e) Gaussian Blur $r = 4$ ; (f) Median Filter, $k = 7$ ; (g) Gaussian Noise, $\mu = 0$ , $\sigma = 0.05$ ; (h) Salt and Pepper Noise, $p = 0.05$ ; (i) $25\%$ Resize and restore; (j) Brightness, factor $= 6$ . + +vere distortions. For instance, when the standard deviation of Gaussian Noise is 0.2, our method increases TPR by $82\%$ for Tree-Ring and $29\%$ for RingID. Similarly, for GS, ROAR improves TPR by $42\%$ and bit accuracy by $28\%$ . This trend is consistent across other distortion types, where ROAR outperforms baseline methods, further validating its effectiveness in challenging conditions. + +# 5.4. Adaptive attacks + +In this section, we examine two types of adaptive attacks to assess the robustness of our method against adversarial manipulations: (1) Reconstruction attack, proposed by [38], leverages an autoencoder or a diffusion model to encode and subsequently reconstruct watermarked images, aiming to eliminate watermark features during the reconstruction process. (2) Purification attack, introduced by [23], involves adding random noise to the watermarked image, followed by a diffusion-based denoising process designed to remove both the noise and the watermark. For reconstruction attacks, we utilize four widely adopted autoencoders: Cheng [4], Bmshj [2], VQ-VAE and KL-VAE [27], along with SDv2.1 to conduct our experiments. For purification attacks, we introduce noise with strengths ranging from 0.05 to 0.7, followed by a diffusion-based denoising process with 100 steps to generate the purified images. + +The results presented in Tab. 2 demonstrate the robust- + +![](images/4e2b7a12597dcba842978de5904b9b9a77c867c83b42b8212b8975916725b887.jpg) + +![](images/00495c5368eacba5f3cdd6934a57d784c571e97d8a8938cc7a2511a4c939dc57.jpg) + +![](images/0cf5c7b17fcdc0da72f02bb53ce68d7f85f76dc522cd9a2879e7ded2c92d4fde.jpg) + +![](images/58757414233ae1b999d3dddcb139786349afcdba30cf8a1008545e63fd2b9cc4.jpg) + +![](images/8b618d941faa205be4c48e8b0efe9cb0813624ae380cc6639a005b7f8b4a0da5.jpg) + +![](images/bbf43f8629baaf30976f861221a674979fd83e46033fe4efa49985855fccff5a.jpg) + +![](images/caa3b7cacb67ac606f05cc0c50e3506c6ae7bb71598351c63768779554609a76.jpg) + +![](images/ccfa2e90685d6e5e969f4a0d5f5c51bf6f2a9039b36c04c6e69f78d8fc27c079.jpg) + +![](images/530b9a02be7fa265d4a6467dfb3891d3818a8fba6ecc05f6b0e79e34d1f3da39.jpg) +(a) Gaussian Noise. + +![](images/83dfe039a1a8c99a0bb6b2f69ab3fde76f66d30bbd9ace812034889b7e6a3d4f.jpg) +(b) JPEG Compression. + +![](images/1c5bed41c043952484581ace634e0477ad196896b6fd3053d40bb66232380d82.jpg) +(c) Median Filter. +Figure 6. Robustness enhancement of ROAR against four different types of distortions, evaluated on Tree-Ring, RingID, and GS. For Tree-Ring and RingID, we evaluate TPR, while for GS, we evaluate both TPR and bit accuracy. + +![](images/046737a9a2cb2d1dc3e9de057800512db0e145cd736d98f6c98e362a07b4adc0.jpg) +(d) Salt & Pepper Noise. + +
AttacksReconstruction Attack
ChengBmshjVQ-VAEKL-VAESD-v1.4
PSNR (dB)32.0930.6929.6829.2125.64
TPR1.0001.0001.0001.0001.000
Bit Acc.1.0000.9991.0000.9981.000
AttacksPurification Attack
s = 0.050.10.30.50.7
PSNR (dB)25.8223.7920.1518.7516.65
TPR1.0001.0001.0000.9800.663
Bit Acc.1.0000.9870.9630.8550.776
+ +ness against reconstruction attacks. Specifically, the detection TPR consistently achieves $100\%$ across all tested scenarios, and the bit accuracy remains above $99\%$ . This strong resilience can be attributed to the watermark being embedded in the latent domain of diffusion models, making it difficult for reconstruction-based attacks to disrupt the watermark signal, as these attacks primarily alter pixel arrangements. For purification attacks, our method also demonstrates notable robustness. When the attack strength is below 0.3, both TPR and bit accuracy remain at high levels. However, as the attack intensity increases, both metrics gradually decline. Higher attack strengths also result in noticeable visual degradation. When the attack strength reaches 0.5, the peak signal-to-noise ratio (PSNR) drops to 18.75dB, indicating a significant loss of visual fidelity. + +# 5.5. Generalization across generation conditions + +To evaluate the generalization capability of our method, we conduct watermark extraction experiments on SD V2.1 by + +Table 2. Robustness against adaptive attacks, evaluated using GS in terms of TPR and bit accuracy. + +
SettingsGuidance ScaleSampling step
2510102550
w/o Distortion1.0001.0001.0001.0001.0001.000
w Distortion0.9920.9860.9830.9870.9880.990
SettingsSampling methodsAverage
DDIMUniPCPNDMDEISDPMSolver
w/o Distortion1.0001.0001.0001.0001.0001.000
w Distortion0.9900.9870.9920.9960.9910.991
+ +Table 3. Generalization across various generation conditions, evaluated using GS in terms of bit accuracy. + +varying key parameters in the generation process, including the guidance scale, sampling steps, and sampling methods. Specifically, we consider five widely used continuous-time samplers based on ODE solvers [30]: DDIM [30], UniPC [37], PNDM [20], DEIS [35], and DPMSolver [21]. Sampling steps range from 10 to 100, and guidance scale is tested at 2, 5, and 10. By default, we use a guidance scale of 7.5, 50 sampling steps, and DPMSolver. In each experiment, only the parameter under investigation is modified, keeping other settings fixed. Extraction accuracy results are summarized in Tab. 3. The results demonstrate that our method generalizes well across diverse generation conditions, achieving approximately $99\%$ bit accuracy even in the presence of noise. + +# 5.6. Ablation studies + +Effectiveness of RO and AR. In ROAR, we introduce two key mechanisms to mitigate inversion errors. To evaluate their individual contributions, we conduct an ablation study + +
MethodsMechanismsSeenUnseen
Gaussian Noiseσ = 0.1Median Filteringw = 13JPEG Compression21QF = 1510Gaussian Blurr = 59Brightnessfactor = 26
Tree-RingBase0.4420.1010.9630.5110.7850.3380.9830.6870.981
RO0.7530.5251.0000.5820.9520.9310.9970.9561.000
AR0.9900.8571.0001.0001.0000.9890.9890.7310.983
ROAR1.0001.0001.0001.0001.0001.0001.0001.0001.000
RingIDBase0.9500.7180.9610.5120.8620.7411.0000.7250.989
RO0.9770.8930.9800.7480.9050.8911.0000.9411.000
AR0.9890.9800.9910.9740.9770.9621.0000.7420.989
ROAR1.0001.0001.0001.0001.0001.0001.0001.0001.000
GSBase0.914/0.8470.492/0.6971.000/0.9530.894/0.7870.991/0.9530.962/0.8911.000/0.9550.927/0.7731.000/0.999
RO0.926/0.8580.583/0.8021.000/0.9570.917/0.7881.000/0.9630.978/0.9101.000/0.9780.988/0.9681.000/1.000
AR1.000/0.9870.991/0.8531.000/0.9931.000/0.9311.000/0.9790.996/0.9421.000/0.9600.934/0.8521.000/0.999
ROAR1.000/0.9961.000/0.9941.000/0.9991.000/0.9871.000/0.9991.000/0.9751.000/0.9991.000/0.9741.000/1.000
+ +Table 4. Ablation study on the effectiveness of the distortion-adaptive network and the regeneration-based optimization module. For TreeRing and RingID, we evaluate TPR, while for GS, we evaluate both TPR and bit accuracy. + +
Inference StepInversion Step
102550100
101.0000.9981.0000.999
250.9981.0001.0001.000
500.9950.9991.0001.000
1000.9940.9981.0001.000
+ +Table 5. Ablation study on the impact of the inversion step, evaluated using GS in terms of bit accuracy. + +assessing their effectiveness under three seen types of distortions (Gaussian noise, Median filtering, and JPEG compression) and two unseen types of distortions (Gaussian blur and Brightness). Watermarked images are generated using different embedding methods, and the watermark is subsequently extracted using four distinct extraction settings: (1) without RO and AR ("Base"), (2) with RO but without AR ("RO"), (3) with AR but without RO ("AR"), and (4) with both RO and AR ("ROAR"). As shown in Tab. 4, both RO and AR significantly enhance robustness. + +For seen distortions, compared to the "Base" setting, the average TPR of Tree-Ring, RingID, and GS improves by $27\%$ , $11\%$ , and $4\%$ under "RO," while "AR" yields further gains of $35\%$ , $19\%$ , and $12\%$ , respectively. Additionally, GS's bit accuracy increases by $9\%$ . The most significant robustness boost occurs when both RO and AR are applied, achieving TPR improvements of approximately $48\%$ , $21\%$ , and $13\%$ for Tree-Ring, RingID, and GS. This demonstrates that distortion restoration before inversion provides a more well-aligned starting point with the original distribution, facilitating subsequent optimization. Notably, AR generalizes well to seen distortions (e.g., Gaussian Blur, an entropy-increasing distortion). For completely unseen distortions (e.g., Brightness), RO demonstrates superior performance in improving inversion accuracy, compensating for AR's limitations in such cases. These results underscore the distinct yet complementary roles of RO and AR in reducing inversion errors and enhancing extraction performance. + +Impact of the inversion step. In practice, the inference step is often unknown, potentially causing a mismatch with the inversion step. However, as shown in Tab. 5, this mismatch results in minimal accuracy loss. Given the efficiency + +![](images/309ee46f59158321d1c19bf3102552b4a156c7c72dab8a43a977c59909150835.jpg) +Figure 7. Impact of the optimization iterations. We evaluate TPR for Tree-Ring, RingID; and both TPR and bit accuracy for GS. + +of modern samplers, inference steps rarely exceed 50. To ensure consistency and optimize extraction, we set the inversion step to 50 in the RO process. + +Impact of the optimization iterations. In the RO process, iteration count plays a crucial role in balancing convergence speed and extraction accuracy. To assess its impact, we conduct an ablation study by varying iterations while keeping other parameters fixed. As shown in Fig. 7, increasing iterations generally improves watermark extraction accuracy. However, beyond a certain threshold, the improvement becomes marginal, while computational cost continues to rise. Based on these observations, we set the iteration count to 20 for an optimal trade-off between efficiency and accuracy. + +# 6. Conclusion + +In this paper, we propose ROAR, a dual-domain optimization-based framework to enhance the robustness of inversion-based generative image watermarking frameworks from the inversion error reduction aspect. Specifically, we introduce a regeneration-based optimization strategy to mitigate latent-domain errors and a MoE-based distortion-adaptive restoration network to reduce the pixel-domain errors. Extensive experimental results demonstrate that ROAR significantly improves the robustness of existing inversion-based watermark extraction. + +# 7. Acknowledgment + +The work was supported by the National Natural Science Foundation of China under Grant 62271307 and 61771310. + +# References + +[1] Ali Al-Haj. Combined dwt-dct digital image watermarking. Journal of computer science, 3(9):740-746, 2007. 2 +[2] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 6 +[3] Tu Bui, Shruti Agarwal, Ning Yu, and John Collomosse. Rosteals: Robust steganography using autoencoder latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 933-942, 2023. 2 +[4] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 6 +[5] Hai Ci, Pei Yang, Yiren Song, and Mike Zheng Shou. Ringid: Rethinking tree-ring watermarking for enhanced multi-key identification. In European Conference on Computer Vision, pages 338-354. Springer, 2025. 1, 2, 5, 6 +[6] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 2, 5, 6 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2 +[8] Han Fang, Yupeng Qiu, Kejiang Chen, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. Flow-based robust watermarking with invertible noise layer for black-box distortions. In Proceedings of the AAAI conference on artificial intelligence, pages 5054-5061, 2023. 2 +[9] Pierre Fernandez, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3054-3058. IEEE, 2022. 2 +[10] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 2, 5, 6 +[11] Gustavosta. Stable-diffusion-prompts dataset. Hugging Face, 2024. Accessed: 18-Feb-2025. 5 +[12] Mohamed Hamidi, Mohamed El Haziti, Hocine Cherifi, and Mohammed El Hassouni. Hybrid blind robust image watermarking technique based on dft-dct and arnold transform. Multimedia Tools and Applications, 77:27181-27214, 2018. 2 +[13] 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. 5 +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3 + +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3 +[16] Zhaoyang Jia, Han Fang, and Weiming Zhang. Mbrs: Enhancing robustness of cnn-based watermarking by minibatch of real and simulatedJPEG compression. In Proceedings of the 29th ACM international conference on multimedia, pages 41-49, 2021. 2 +[17] 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. 2 +[18] Changhoon Kim, Kyle Min, Maitreya Patel, Sheng Cheng, and Yezhou Yang. Wouaf: Weight modulation for user attribution and fingerprinting in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8974-8983, 2024. 1 +[19] Deepa Kundur and Dimitrios Hatzinakos. A robust digital image watermarking method using wavelet-based fusion. In Proceedings of International Conference on Image Processing, pages 544-547. IEEE, 1997. 2 +[20] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 7 +[21] 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. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 7 +[22] 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. 2 +[23] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460, 2022. 6 +[24] Augustus Odena, Vincent Dumoulin, and Chris Olah. Deconvolution and checkerboard artifacts. Distill, 1(10):e3, 2016. 5 +[25] 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 +[26] Ahmad Rezaei, Mohammad Akbari, Saeed Ranjbar Alvar, Arezou Fatemi, and Yong Zhang. Lawa: Using latent space for in-generation image watermarking. In European Conference on Computer Vision, pages 118-136. Springer, 2025. 1, 2, 5, 6 +[27] 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, 5, 6 + +[28] 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 +[29] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 5 +[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2, 3, 7 +[31] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22532-22541, 2023. 3 +[32] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 1, 2, 5, 6 +[33] Zijin Yang, Kai Zeng, Kejiang Chen, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12162-12171, 2024. 1, 2, 5, 6 +[34] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 2, 5, 6 +[35] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 7 +[36] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 5 +[37] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 7 +[38] Xuandong Zhao, Kexun Zhang, Yu-Xiang Wang, and Lei Li. Generative autoencoders as watermark attackers: Analyses of vulnerabilities and threats. 2023. 6 +[39] J Zhu. Hidden: hiding data with deep networks. arXiv preprint arXiv:1807.09937, 2018. 2 \ No newline at end of file diff --git a/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/images.zip b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8fa76b9f801ea05d5e320644dd1dad3f132b7594 --- /dev/null +++ b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ede4c2b4afe7b9bd6a6dcfcbc1c2e8e2176234b12f2e49ff4e2529a01ef2ab60 +size 684728 diff --git a/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/layout.json b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..86f627272bcbee3fab785ffd8a9feb2932671d98 --- /dev/null +++ b/ICCV/2025/ROAR_ Reducing Inversion Error in Generative Image Watermarking/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8489dd6a81b12b5faab7123f8eae27c83d3672acb96216e33e4b8151d6a74da +size 437655 diff --git a/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/44a1ed72-9329-4443-a25d-f09953e202c6_content_list.json b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/44a1ed72-9329-4443-a25d-f09953e202c6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7058ad5b4f1170a0b3f713cc6a04080e3cd19f5c --- /dev/null +++ b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/44a1ed72-9329-4443-a25d-f09953e202c6_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8143a9261b4db258d98c7ec38069f5792c127f1879962fc5c2455131db1f547 +size 88839 diff --git a/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/44a1ed72-9329-4443-a25d-f09953e202c6_model.json b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/44a1ed72-9329-4443-a25d-f09953e202c6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f91336b085e60748aa680375c57fe7138eefb1f0 --- /dev/null +++ b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/44a1ed72-9329-4443-a25d-f09953e202c6_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fa92c108dfc4aa3fe275824d448e780f168580e9000bb83bb77ba37e951b8b8 +size 126034 diff --git a/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/44a1ed72-9329-4443-a25d-f09953e202c6_origin.pdf b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/44a1ed72-9329-4443-a25d-f09953e202c6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cc81fd7124f3875ae459c811d24f2801a04ebc0f --- /dev/null +++ b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/44a1ed72-9329-4443-a25d-f09953e202c6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44f87781b4ca210658ec7e8712ebaabd661d082277a85d12e426c99638be46eb +size 6284060 diff --git a/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/full.md b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8256550459a5a05bf028765d48ca6272ba8463e7 --- /dev/null +++ b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/full.md @@ -0,0 +1,452 @@ +# ROVI: A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation + +Cihang Peng + +Qiming Hou + +Zhong Ren + +Kun Zhou* + +State Key Lab of CAD&CG, Zhejiang University + +cihangpeng@zju.edu.cn, hqm03ster@gmail.com, renzhong@zju.edu.cn, kunzhou@acm.org + +![](images/3d7db0b141a1b41b4b6e4bb4374b59dd3713ffb8f4a608a3be1b7cd458e1da1c.jpg) +(a) Detected by Web Cap. +Figure 1. Effect of our pre-detection strategy. Image (a) shows a web image with original caption where OVDs only detect cockatoo, jungle, and wallpaper mural. Our method (b) discovers additional elements: bedding, pillows, and blanket, with detailed composition descriptions. The base generator [40] (c) captures composition but lacks instance grounding. The instance-grounded model [25] (d) fine-tuned on sparse web captions only generates a bird-like shadow, omitting the blanket and other elements. The same algorithm trained on our re-captioned dataset (e) correctly produces all details. + +![](images/6016eea5b242e1d0b3d9b1dccbc0aef92b8613fb1267e800075aac43bc2cd354.jpg) +Web Image +(b) Detected after Re-Captioning + +![](images/75a3c8c741a0b98d108b42bf12b32482afe2532ff47103a8de47242b10dfda22.jpg) +(c) Stable Diffusion + +![](images/b10e913747a1d23a8ddb38595208ed555699f78f90ee9b8eecde8fc841d2ff91.jpg) +Generated Images +(d) Trained on Others + +![](images/9acb997110ad273325ed0c2b63fe92b3340b8cbb6f3115a6afc6060149f56eb4.jpg) +(e) Trained on ROVI + +Original Web Caption: Nostalgic Cockatoo Jungle Wallpaper Mural + +A photograph of a bedroom with a mural-style wallpaper featuring tropical plants and birds. In the center, there is a wooden bed with white bedding, pillows and a grey throw blanket. The wallpaper includes illustrations of two pink cockatoos and a grey cockatoo. + +# Abstract + +We present ROVI, a high-quality synthetic dataset for instance-grounded text-to-image generation, created by labeling 1M curated web images. Our key innovation is a strategy called re-captioning, focusing on the pre-detection stage, where a VLM (Vision-Language Model) generates comprehensive visual descriptions that are then processed by an LLM (Large Language Model) to extract a flat list of potential categories for OVDs (Open-Vocabulary Detectors) to detect. This approach yields a global prompt inherently linked to instance annotations while capturing secondary visual elements humans typically overlook. Evaluations show that ROVI exceeds existing detection datasets in image quality and resolution while containing two orders of magnitude more categories with an open-vocabulary nature. For demonstrative purposes, a text-to-image model GLIGEN [25] trained on ROVI significantly outperforms state-of-the-art alternatives in instance grounding accuracy, prompt fidelity, and aesthetic quality. Our dataset and reproducible pipeline are available at https://github.com/CihangPeng/ROVI. + +# 1. Introduction + +Text-to-image generation has rapidly matured in recent years with huge industrial impact. While in basic form it solely takes a text prompt as input, many practical applications would benefit from instance grounding, where users specify precisely where and what kind of object should appear using box-label pairs to enable finer control. Many state-of-the-art research [2, 19, 25, 51, 52, 60, 61] can provide this capability, giving significant insight on algorithm design. Unfortunately, their performance is restricted by the poor availability of dedicated datasets. + +Typically, the training of instance-grounded generation models consumes images with individually labeled object instances with bounding boxes and captions. Existing public datasets like COCO [27] and LVIS [17] provide annotations that superficially meet these requirements but suffer from critical limitations. First, these datasets are primarily designed for detection purposes, thus cannot guarantee high-quality text-image pairs. The image descriptions are either missing or of poor quality with insufficient detail, while many images feature poor lighting conditions, + +
DatasetImagesCategoriesAvg. CategoryAvg. BoxAvg. ResolutionAvg. Aesthetic Score
COCO [27]123K802.937.04578 × 4844.97
LVIS [17]123K12033.6212.73578 × 4844.97
Objects365 [43]1936K3656.1214.58663 × 5364.92
Open Images [20]1893K6002.628.38976 × 7914.90
V3det [48]245K132044.017.16789 × 6604.99
ROVI1012K144336012.5124.212102 × 14886.00
+ +heavy occlusion, and cluttered backgrounds, all of which adversely affect text-to-image generation quality. Second, their label vocabulary remain restricted to a limited set of predefined categories, severely constraining the expressive potential of instance grounding tasks. Recent algorithm research on instance grounding [12, 25, 51, 52, 60, 61] tend to generate ad hoc synthetic instance labels to enrich the training data. Such dataset synthesis is enabled by the recent advancement of OVDs (Open-Vocabulary Detectors) [8, 18, 29, 30, 47, 55, 59]. While OVDs can detect arbitrary objects mentioned in the accompanying text, captions rarely include all visual elements, and the detection often remains incomplete. Figure 1 illustrates how such incompleteness eventually translates to a low-quality generated image: a web image of a cockatoo mural in (a) has a caption mentioning only the cockatoo, omitting visible elements such as bed, pillows, and blanket. Using a model trained on the resulting annotations, the generated images fail to properly integrate previously "unseen" objects, as shown in (d). + +To address this limitation, we introduce ROVI (Recaptaioned Open-Vocabulary Instances), a high-quality synthetic dataset of 1M images created through novel methodology. Our core innovation is a pre-detection re-captioning strategy with two main components. First, we use a VLM (Vision-Language Model) to create a detailed description of recognizable objects for each image along with compositional information such as semantic relationships between instances. Then we use an LLM (Large Language Model) to summarize a flat list of categories from the VLM description. The end result is illustrated in Fig. 1 (b) and the accompanying text. Various components of the bed are detected and the birds are explicitly stated as a part of the wallpaper in the caption. Such improved dataset eventually leads to a network that can correctly interpret the instance layout in Fig. 1 (e). + +Table 1 compares the overall statistics with existing alternatives. Our ROVI dataset is able to discover two orders of magnitude more distinct object categories, producing significantly more grounded instances for each image. We also curated the dataset with a considerably higher standard in terms of resolution and aesthetic scores. + +Our dataset and codes are publicly available, including image URLs, detailed annotations, and intermediate results from VLM/LLM/OVD steps, along with the code to reproduce each step. We rely solely on open source models and all third-party code can be run offline with downloaded weights. Table 2 summarizes the open source models used in our pipeline and their parameter sizes. + +# Contributions We make the following contributions: + +- A systematic instance annotation pipeline, featuring a novel methodology of pre-detection VLM-LLM re-cognition that enables multi-OVD usage and captures compositional information between instances; + +- A large-scale, high-quality open dataset synthesized with this pipeline. + +Table 1. Dataset statistics computed on combined training and validation sets. "Categories" counts the total number of distinct categories in the entire dataset. "Avg. Category" counts the average number of distinct categories in each image. + +
Model TypeModel NameParam Size
VLMInternVL1.526B
Qwen2VL8B
LLMLlama 38B
OVDGrounding-DINO172M
YOLO-World110M
OWLv2408M
OV-DINO166M
+ +Table 2. Open source models involved in the creation of ROVI. + +# 2. Related work + +# 2.1. Text-to-Image Generation + +Text-to-image generation started with GANs (Generative Adversarial Networks) [15]. Recently diffusion [5, 11, 21, 26, 33, 35, 40] and auto-regressive [24, 38, 45, 56] methods have achieved a lot of success. By definition, all aforementioned methods take a natural-language text prompt as input and generate an image. Our dataset is specifically designed to support advanced text-to-image generation training, featuring high-quality images paired with comprehensive textual descriptions and others. + +# 2.2. Language Models and Vision + +With the advent of LLMs [1, 4, 10, 14, 16, 46, 54], language models have received significant attention recently, with numerous approaches integrating them with computer vision tasks. The VLMs [7, 22, 28, 32, 49, 50] extend language + +![](images/2fc95644b867c72473528882aafc38936ec4ac8c3a8ca85d42aade43075a24d7.jpg) +Figure 2. Pipeline overview: stages of our systematic instance annotation framework. + +models with visual capabilities, enabling them to interpret and reason about image content. + +In our framework, we use VLMs primarily to generate image descriptions that guide detection and secondarily for cross-checking on detected box-label pairs. While some VLMs can produce bounding boxes directly, they lack the precision and reliability of dedicated detectors. Therefore, we opted to use specialized OVDs for detection purpose. + +# 2.3. Open-Vocabulary Detectors + +Object detection has been a long-standing topic in computer vision. The open-vocabulary variant is flexible in what objects can be detected, although a limited set of potential objects still has to be specified somehow. Among the OVDs (Open-Vocabulary Detectors) [8, 18, 23, 29, 30, 47, 55, 57, 59] we surveyed, phrase-grounding models [23, 29, 55, 57] such as Grounding-DINO[29] take raw text as input and select relevant phrases to associate with each detected bounding box. Other OVDs require a list of potential object categories in text form, with each category potentially containing multiple words. + +# 2.4. Datasets for Detection and Image Description + +Traditional vision datasets such as COCO [27], LVIS [17], Object 365 [43], and Open Images [20] have limited categories. V3Det [48] expands the category scope by building a large taxonomy of 13,204 categories, improving the diversity of instance descriptions. However, when repurposed for text-to-image generation, such datasets pose shared challenges including low image quality, terse image descriptions, and the lack of global contextual connections. Their + +instance labels also exhibit limited attributes, which is essential for capturing nuanced object characteristics. + +Recent efforts have explored detailed image descriptions with compositional information, also an important component for enhancing text-to-image generation, though they lack the explicit instance grounding our work addresses. DOCCI [31] and ImageInWords [13] datasets emphasize detailed, human-annotated descriptions that capture richer object attributes and improved spatial relationships. From Pixels to Prose [44] leverages open-source VLMs to generate dense captions, effectively expanding the dataset scale. + +# 2.5. Instance Grounding + +As text alone has limited capability to specify spatial constraints such as position and layout, many instance grounding methods [6, 25, 34, 51-53, 60, 61] have been developed to support additional forms of input for image generation, in which bounding boxes and per-instance captions are used to provide precise spatial and semantic guidance. These methods typically extend pretrained text-to-image generators by incorporating such spatial information, while still relying on the original text prompt for global context. + +A key challenge in this paradigm lies in preparing high-quality training data that integrates bounding boxes, object-level tags, and complementary text prompts in a cohesive manner. Several strategies have been proposed in recent algorithm research. GLIGEN [25] directly apply phrase-grounding OVDs [23, 29, 55, 57] on sourced image captions to establish instance-level grounding. MIGC [61] and MIGC++ [60] refined this process by extracting instance descriptions from source captions using a natural lan + +guage processing tool [36], thereby improving the attribute binding of OVD-detected tags. InstanceDiffusion [51] and IFAdapter [52] adopt a more structured approach, using RAM [58] to generate instance annotations from the image instead of its caption, which are then passed through OVDs. A VLM is finally applied to produce captions for independently detected instances. + +To the best of our knowledge, our method is the first to focus on the pre-detection stage, using a meticulous combination of VLMs and LLMs to improve the OVD input and supplement the results with compositional information. We also employed non-phrase-grounding OVDs [8, 30, 47] for improve detection recall, which would have been impossible in previous methods where the detector input is unstructured natural text. + +# 3. Method + +Figure 2 is an overview of our pipeline. We start from a set of curated images fetched from public datasets. We then describe each image with a VLM and prepare a flat list of categories with an LLM. Categories are detected by OVDs. Finally, we merge and resample the detection results and cross-check detected bounding boxes with a different VLM. + +Compared to ad hoc dataset synthesis, the main innovation we make is the pre-detection re-captioning. By invoking a novel VLM-LLM strategy before detection, we are able to create an accurate and comprehensive list of potential objects, which allows us to detect a broader set of instances using the same OVDs. The re-captioning also mitigates any bias present in the text part of source datasets, significantly increasing the overall consistency. + +The complete pipeline processing 1.012M images required approximately 1,960 A800 GPU hours, averaging 6.97 seconds per image or 0.29 seconds per instance. The computational cost breakdown by stage is: VLM description (820h), LLM summarization (250h), object detection (440h), and VLM cross-checking (450h). + +# 3.1. Image Curation + +Since the ultimate purpose of our dataset is image generation, we need some basic guarantee on image quality. We start with the large COYO-700M [3], LAION2B-enaesthetic [41, 42] and LAION POP [9] datasets, and filter for images with aesthetic score [42] larger or equal to 5.75 and resolution above $1024 \times 1024$ . We also filter out oversized images with more than 6144 pixels along the longer edge or 4096 pixels along the shorter edge. After pHash dedduplication with a Hamming distance of 10, the curation process yielded a total of 1M image-text pairs out of the sourced billions. Only $4.7\%$ images can pass our resolution standards, in which only $2.1\%$ can pass our aesthetic score criteria. The dedduplication eliminates $31\%$ of redundant images. + +![](images/9a24e88bc406f7e7b80167cea6bb265a365c5c71e709f421e3f9aca7fecea549.jpg) +(a) Unwanted Texts & Abstract Concepts + +including one marked "CELIC." ... A warning sign reads "HOT DO NOT THROW RUBBISH IN HOPPER," Text + +There's a sense of disrepair and desolation, emphasizing the passage of time and the decline of industrial activity in this once bustling facility. Abstract Concept + +![](images/5b30052ad36db50e7f2fde0b3f64a6df1776db03ed89bc5f5d1f5ec754d0509b.jpg) + +There are no visible human activities or figures in the scene, emphasizing the natural and untouched beauty of the landscape ... + +There is no visible text within the image ... + +![](images/26df1669bcd2c71ba9a5150ab55eb708aad51460cc8f4d9bf1d0d8dcb57fbd21.jpg) +(c) Potential OV. Categories in Web Caption + +VLM Description: A photograph of a traditional red and yellow horse-drawn caravan with intricate patterns and designs ... + +Web Caption: Gypsy caravan weekend breaks rhyader mid wales + +a two-layered cake with a chocolate ganache drip on the sides and a smooth chocolate frosting on top ... + +adorned with six swirls of a lighter-colored frosting, possiblypeanut butter or caramel … VLM Description + +![](images/bb4a34cdc6068f12e137d74b8bf230e4d1118e582c0649545d442ad4b184a300.jpg) +(b) Negative Forms to Remove +(d) Two-Step LLM Decomposition +Figure 3. Challenges identified in VLM/LLM inference trials. (a)-(b) show possible problematic elements in raw VLM outputs, (c) demonstrates potential open-vocabulary categories in web captions, and (d) illustrates our two-step LLM approach converting detailed VLM descriptions into structured object categories. + +LLM Step 1: two-layered cake, chocolate ganache drip, smooth chocolate frosting, lighter-colored frosting, peanut butter or caramel frosting ... + +LLM Step 2: cake, chocolate, ganache, drip, frosting, chocolate frosting, peanut butter, caramel frosting + +# 3.2. VLM Description + +As introduced in Sec. 1, text in source datasets are often terse webpage captions, while we prefer a comprehensive inventory of visual contents. As a solution, we prompt InternVL1.5 [7], a recent VLM, to generate detailed visual descriptions from images. For clarity, we refer to the original dataset text as the web caption, and to our generated description as the VLM description. + +To accommodate token limits of downstream models, we begin our descriptions with "This is" followed by the overall image style like "black and white photograph", preserving such essential information in naïve token clamping. We remove "This is" as a post-process. + +As illustrated in Fig. 3 (a), VLMs might spend tokens discussing non-visual concepts, such as expressions like "a sense of disrepair and desolation." Similarly, VLMs often transcribe text within images, which can introduce misleading categories for detectors (e.g., "RUBBISH"). Since such content is difficult to filter otherwise, we dedicate approximately half of our input tokens to explicitly instruct the VLM to avoid such speculation and text transcription. + +Additionally, we post-process the VLM output to remove negative statements. As illustrated in Fig. 3 (b), sentences like "There are no visible human activities" provide little value to detectors and waste resources in subsequent steps. + +# 3.3. LLM Summarization + +Before object detection, we choose to generate a flat list of categories suitable for less flexible OVDs. This text summarization task requires no image processing, so we employ an LLM, Llama3 [10], for its advanced linguistic capabilities. We delay the introduction of the web caption to this step. + +As illustrated in Fig. 3 (c), the web caption can contain vital information like "Gypsy caravan" which cannot be easily derived from visual features alone. However, providing them prematurely in the VLM step risks suppressing the factual "red and yellow horse-drawn" qualifiers of the same caravan which are equally important. Deferring the web caption to the image-oblivious LLM summarization ensures both versions have a chance to reach downstream OVDs, where anything incorrectly summarized linguistically will likely be ignored. + +As shown in Fig. 3 (d), we first instruct the LLM to extract a list of concise categories from the combined text. In this pass, our LLM has a strong tendency to preserve compound phrases in the VLM description, sometimes refusing to break conjunctions like "peanut butter or caramel frosting". While such behavior enhances attribute binding, the lack of simple words can challenge downstream detectors and limit our coverage of simple constituent objects. + +As mitigation, we implement a second decomposition step to extract basic forms, like standalone nouns, from the output phrases. Specifically, we prompt the LLM to break complex concepts into their constituent parts (e.g., extracting cake, chocolate, ganache, and drip). We then combine the results from both steps as OVD input. The redundancy improves recall by providing core components as backup when a complex phrase fails detection. + +# 3.4. Object Detection + +For the actual detection, we employ Grounding-DINO [29], OV-DINO [47], YOLO-World [8] and OWLv2 [30], sending the category list to all four OVDs. For each OVD, we set its confidence score threshold slightly lower than conventional usage to retain more candidate boxes while balancing their average detection counts. We concatenate the boxes from all OVDs and perform IoU (Intersection over Union)-based NMS (Non-Maximum Suppression) with a threshold of 0.4 between boxes of the same category. + +After concatenation, we apply a resampling process to address excessive box coverage resulted from using multiple OVDs. This process balances OVDs, penalizes overlapping boxes, duplicate captions, distance from image center, and small box sizes, effectively removing approximately + +$70\%$ of instances while preserving detection quality. Further details on this resampling methodology are provided in the supplementary materials. + +Table 3 presents the distribution of detection boxes and category coverage across our four OVDs on 5K randomly sampled images. Each OVD demonstrates distinct strengths in different categories, and a considerable portion of categories can only be detected by one of them. Such multidetector fusion is enabled by our fine-grained VLM description and LLM summarization, which produces the flat category list required by the less flexible OVDs [8, 30, 47]. + +
OVD TypeGD [29]YW [8]OW [30]OD [47]
Box Contribution26.7%21.7%28.0%23.6%
Cat. Coverage49.5%42.9%49.1%31.2%
Unique Cat.17.0%11.9%15.8%9.4%
+ +Table 3. Per-OVD statistics on 5K images. GD: Grounding-DINO, YW:YOLO-World,OW:OWLv2,OD:OV-DINO. + +# 3.5. VLM Cross-Checking + +At the end of our pipeline, we crop out the bounding box of each resampled instance. We subsequently prompt a second VLM, Qwen2VL [49], to check whether the image content matches its expected caption. This process can be expensive due to running a VLM or each box. We manage the cost by asking a yes/no question and restricting the output token count to 1. We only keep an instance when the total probability of all capitalization variants of "yes" is significantly more than that of "no". + +# 4. Experiments + +# 4.1. Study on Data Generation Alternatives + +Our primary contribution is the pre-detection VLM-LLM re-captioning. In this section, we assess its impact by comparing it with alternatives that served similar roles in previous instance grounding works. + +The first group relies on recent phrase-grounding OVDs, such as GLIP, GLIPv2, and Grounding-DINO, to extract categories as token subsets from natural language text. Among those OVDs, Grounding-DINO offers state-of-the-art performance, which we choose as a comparison target following established practice [51, 52, 60, 61]. We tested two variants of this straightforward approach: + +- Web Cap. + OVD uses the web captions directly as input to phrase-grounding OVD, similar to the approach adopted by GLIGEN [25]. +- VLM Desc. + OVD follows Sec. 3.2 to generate a VLM description of the image, which is sent to the OVD. This is a straightforward extension of the GLIGEN approach to VLMs. + +The second group follows methods first introduced in InstanceDiffusion [51], which operates in two stages: first + +ROVI: VLM Description and Guided Box Annotations +Figure 4. Effect of RAM++ tagging plus per-box VLM captioning. RAM++ tags: dirt field, floor, hill, hillside, land, mound, terrain, squat, man, sit, plaid, shirt, beard. Note the high redundancy within this list. For clarity, detected boxes are de-duplicated in both approaches. +![](images/adf8beb949ba0d91031029aaa335fae90d343269aa1fbd8abe3abaa11b3869f6.jpg) +Web Caption: 10 Hour Package + +A photograph featuring a man with a beard and long hair, seated on a sandy hill. He is wearing a red and black plaid shirt, blue jeans, and hiking boots, and is equipped with a large green backpack. The man appears contemplative, gazing into the distance. The landscape around him is natural, with sparse vegetation and a few trees scattered in the background under a clear sky. + +RAM++ -> Grounding-DINO -> Per-Box VLM Captioning +![](images/cb8d04a053a442699fb7ca4e70e84583f1fbe653b8d51d7ec43d3d3c388029c5.jpg) +Dedup. +Per-Box VLM Caption + +dirt A person kneeling on the ground (wrong) in a field natural outdoor setting. +squat A man sitting on the ground with a backpack, wrong) wearing a plaid shirt and a beard. + +plaid A person wearing a plaid shirt and a backpack, sitting outdoors. + +beard A close-up image of a person's beard. + +performing recognition-based tagging on images and inputting these tags to OVD for detection, then generating VLM captions individually for each detected box, producing isolated instance-level descriptions. For the tagging component, we tested both RAM [58] and $\mathrm{RAM} + +$ [58]. + +Table 4 compares the number of categories and bounding boxes extracted using each approach on 5K randomly sampled images. RAM tags allowed the detectors to discover more boxes but are limited by predefined categories. The OVD-extracted categories are more diverse, while detecting less boxes. Our dataset outperforms both significantly. + +Figure 4 uses a concrete example to illustrate the difference between our approach and RAM-based alternatives. While RAM tagging addresses the web caption bias, many less prominent objects remain undetected. They also tend to produce boxes for non-object concepts such as "squat" or "plaid," which can be counterproductive for image generation. + +The right figure also illustrates the unexpected effect of the per-box VLM captioning strategy. While non-object boxes receive more reasonable captions, all boxes overlapping the person described him to some extent. Such content duplication across differently shaped boxes can mislead the trained image generator. Furthermore, the VLM hallucinated details based on incomplete crops, such as interpreting a cropped human body showing only legs as "kneeling on the ground." In contrast, the pre-detection re-captioning allows ROVI to provide higher object detection coverage, better context integration, and improved category precision. + +
Data Gen. MethodCat.Avg. Cat.Avg. Box
Web Cap.→OVD30010.720.86
VLM Desc.→OVD144127.679.20
RAM→OVD18366.4411.54
RAM++→OVD18466.4911.58
ROVI's2537412.4524.01
+ +Table 4. Quantitative comparisons between data generation methods. RAM/RAM++ settings follow the InstanceDiffusion [51] implementation. We only count box-caption pairs that passed our VLM cross-check in Sec. 3.5. Cat.: number of distinct categories discovered over the entire sampled set. Avg.: per-image averages. + +
ModelGen Inst.↑FID↓Aes.↑CLIP Sim.↑
SD v1.40.530/0.39817.6/38.35.66/5.170.283/0.257
Offi. G.0.750/0.81219.3/25.45.33/4.780.273/0.233
MIGC0.821/0.83327.6/32.25.47/5.230.228/0.213
Instdiff0.854/0.87121.1/28.45.47/5.220.262/0.229
Our G.0.880/0.87215.7/16.75.83/5.400.286/0.245
+ +Table 5. Apples-to-apples quantitative comparisons on ROVI validation set (left values) and Open Images validation set (right values in brown). Gen Inst.: per-box cross-checking pass rate using VLM-based evaluation. Our G.: GLIGEN trained on ROVI. SD v1.4: Stable Diffusion baseline [39]. Offi. G.: Official GLIGEN [25]. InstDiff: InstanceDiffusion [51]. Aes.: Aesthetic score [41]. CLIP Sim.: CLIP-based image-text similarity [37]. + +# 4.2. Demonstrations on Generated Images + +To evaluate the end-to-end effect of our dataset, we trained a GLIGEN network on the training portion of our dataset, us + +ing the official Stable Diffusion v1.4 checkpoint as the base model. This configuration is deliberately chosen to better isolate improvements induced by our dataset from algorithmic improvements in newer methods. During comparison we generate images using each method from the same text prompt and the same set of box-label pairs. + +Quantitative Results. Table 5 compares qualitative metrics between our GLIGEN, existing methods, and a Stable Diffusion baseline. We stochastically sampled 5000 prompt-image pairs from both the ROVI validation set (black numbers) and the Open Images validation set [20] (brown numbers) and compared instance-level and image-level metrics on generated images. + +At the instance level, we focus on instance grounding accuracy, evaluated using the VLM-based cross-checking method in Sec. 3.5. Specifically, we cropped each expected instance from its specified bounding box on a generated image and used Qwen2VL [49] to compute the likelihood of the cropped window matching its assigned label. Our GLI-GEN has achieved pass rates comparable with more recent methods on both datasets. + +At the image level, our GLIGEN consistently outperforms existing alternatives quality-wise (FID, Aes.) while also remaining more faithful to the global prompt (CLIP Sim.). Notably, we also outperformed the baseline quality-wise, even on the Open Images validation set. Such image-level metrics are known to disadvantage fine-tuned models as noted in the MIGC [61] paper, and the results indicate that training GLIGEN on ROVI improved image quality sufficiently to offset this disadvantage. We attribute this to our improved grounding and prompt quality. + +Comparing the two validation datasets, the Open Images set results in lower image quality for every method. Our manual inspection observed visually objectionable artifacts in a large portion of generated images regardless of model. We hypothesize that the degraded performance stems from three properties critical for grounded text-to-image evaluation: (1) comprehensive open-vocabulary labels, (2) high-quality captions suitable as prompts, and (3) aesthetic quality matching generation standards. For example, the Open Images dataset we used only contains 600 categories. Its image captions are seldom consistent with image generation prompts. Many images in the set are also aesthetically questionable for evaluation purposes. We have explored commonly used instance-annotated datasets and found they all have similar problems, as they were mostly designed for detection tasks. In contrast, our ROVI validation set was specifically curated to address these limitations, which can provide more reliable evaluation for real-world grounded text-to-image generation scenarios. + +Qualitative Results. We also present qualitative results to evaluate attribute bindings and object composition. To + +make the results more interpretable for visualization and user study, we hand-picked box-label pairs that provide a semantically meaningful description of the global scene. + +Figure 5 (a) compares the reproduction accuracy of instance attributes. All methods successfully produced the correct object types such as "lamp" or "chair" at the requested locations. However, when both the prompt and instance grounding specified a "lamp with black shade," only our model trained on ROVI correctly rendered the black shade, while other algorithms defaulted to white shades. Similar accuracy improvements were observed with the "slat back chair". + +We attribute this improved attribute fidelity to the category diversity enabled by our VLM-LLM re-captioning strategy. While visually evident, attributes like lamp shade colors and chair back types are rarely mentioned in conventional image-text pairs. We direct the VLM to comprehensively catalog these self-evident attributes, which are then fully revealed by the LLM, allowing subsequent OVDs to detect and ultimately enabling their accurate generation. + +Figure 5 (b) illustrates how the compositional information in VLM descriptions improves text-image alignment. When prompted to generate a black and white photograph of an elderly man kneeling in front of a wall and holding a paintbrush, networks trained on other datasets produced correct overall layouts but neglected compositional details, suggesting over-reliance on instance grounding. In contrast, our GLIGEN faithfully reproduced most details, a capability directly enabled by our re-captioning methodology. + +User Study. We have also conducted a user study to analyze how well our trained GLIGEN network respects instance grounding, the prompt, and its aesthetic quality when compared to alternatives. We ran all candidate implementations to synthesize images from 200 box-grounded inputs. Each survey session presented two images from the same input alongside their shared instance grounding and text prompt, and asked the user to select which image they prefer aesthetically / better aligns with the prompt. We also asked the users to specify which image better respects the grounding constraint of each box-label pair. The image pairs had their placement randomized so that users could not infer their sources. The users are volunteered researchers with ample expertise in image generation. + +
Our GLIGEN vsWin Rate
Official G.MIGCInstDiff
Instance Alignment82.0%94.8%85.0%
Prompt Alignment83.3%90.5%82.3%
Aesthetic Quality85.3%98.3%83.3%
+ +Table 6. Ratio of users preferring GLIGEN [25] trained on ROVI compared to each alternative on each aspect. Official $G$ : official implementation of GLIGEN. InstDiff: InstanceDiffusion [51]. + +Table 6 shows the result of the user study. GLIGEN + +A photograph of a neatly arranged interior space featuring a black wooden desk. On the desk, there's a framed picture, and a lamp with a black shade. A slat back chair with a gray cushion is placed in front of the desk. To the left of the desk, there's a large terracotta pot with a green plant. + +![](images/a801ba00711d1c959fad55d437381281eba2a430c8afb366ca9db4e1c39c4af5.jpg) +Input + +![](images/c9fc2e600f2ae4d50e1fbbf60e62c26834f41e8a022a8fbf1008892cc528fe91.jpg) +Our GLIGEN +lamp with black shade +gray cushion + +![](images/f39facb7a59e9a29459e34e3c36660078966b99feea9cfec5c14caa28f03dc47.jpg) +(a) Handling Less Common Attributes +slat back chair +large terracotta pot +black wooden desk +amed p + +![](images/799d2ca7fac43a72bb6d70d9620913e2b068812e1bc4030ddbe81302b9ad2fa7.jpg) +Official GLIGEN +lamp with black shade +gray cushion +slat back chair +large terracotta pot +black wooden desk + +![](images/b3791b36ba879e90d9b9f5694637588dbb10f7873b14c6418379736bc2ffcf3e.jpg) +MIGC +lamp with black shade +gray cushion +slat back chair +large terracotta pot +black wooden desk + +![](images/54be6b08bdd64c9521f15f1d9b8eaa56e21e2ace4d4d4e78cb2a7b7f24cda15f.jpg) +InstanceDiffusion +lamp with black shade +gray cushion +slat back chair +large terracotta pot +black wooden desk + +A black and white photograph depicting an elderly man engaged in an artistic activity. He is kneeling in front of a wall that has remnants of paint and sketches on it. The man is holding a paintbrush and painting the wall. He is wearing a dark jacket and has a full beard and mustache. There is a wooden cabinet in the background. + +![](images/64f652da6c18f6831611765280433fb52c72c97c64a1df5eaf8ba5884abbee1f.jpg) +Input +(b) Handling Compositional Information + +![](images/b5da9d6bf759699c50a10a6c2844200c51028506d5866d6c13356be61939b84b.jpg) +Our GLIGEN +black and white photo +remnants on wall +holding a paintbrush +keeling in front of wall +dark jacket +wooden cabinet + +![](images/891dc43fc840a1e3cb594b8ac4866f10721b2bba4b71024d4488a5c03c30ff3b.jpg) +framed picture +Official GLIGEN +black and white photo +remnants on wall +holding a paintbrush +keeling in front of wal +dark jacket +wooden cabinet + +![](images/f858a4c9b3944871a47ff95fd6ee66e8b94a06a939ff8c66281369753f61fae5.jpg) +framed picture +MIGC +black and white photo +remnants on wall +holding a paintbrush +keeling in front of wa +dark jacket +wooden cabinet + +![](images/c1aa3237f73791375eaf1632402daced3f5659cf6936d3a0c4d2973440ae2271.jpg) +framed picture +InstanceDiffusion +black and white photo + +![](images/2ad3ca09ed38a94e981f40cf4c1c68fc9faf99abd1895a49058dc598e29471b6.jpg) +remnants on wall +holding a paintbrush +keeling in front of wall +dark jacket +wooden cabinet +√ +X +# +X +Figure 5. Comparing model performance on compositional prompts with instance grounding. Besides, note the attribute binding issues such as gray cushion. Grayish-blue: instance attributes. Brown: compositional elements. Red: mismatch in generation. + +trained on ROVI is preferred in all aspects by a decisive margin. ROVI allows the network to learn many subtly different attributes of each core object, thanks to our more diverse categories and more detected boxes. The improved alignment appears obvious to human users and significantly contributes to the overall quality of the final images. + +# 5. Limitations and Discussion + +We have presented ROVI, a synthetic dataset for instance-grounded text-to-image generation that addresses key limitations in existing approaches. By employing VLMs for comprehensive image description followed by LLM-based open-vocabulary category summarization, our pre-detection strategy enables more complete instance grounding and superior text-to-image generation. + +We manually inspected 24k bounding boxes from 1,000 randomly selected images, finding that $3.3\%$ of the detected boxes contain errors. These errors stem from OVD false-positives that bypassed VLM cross-verification. The two main failure modes are: (1) detecting a partially occluded + +object as several small visible parts, and (2) misinterpreting phrases, such as labeling a rock as a "book about rocks". + +A few minor language model artifacts persist through our pipeline, including inconsistent singular/plural handling, left/right confusion when models interpret human figures as mirrored, and awkward phrasing such as "blue dressed woman" or "stethoscope-wearing woman". Bounding box grounding is also inherently inaccurate for visually occluded objects and non-contiguous elements. Our resampling strategy partially mitigates these inaccuracies. + +While our dataset features high-quality, high-resolution images, current GLIGEN based on Stable Diffusion v1 experiments cannot fully utilize these improvements. Newer architectures would better showcase ROVI's strengths in high-resolution rendering, aesthetic quality, and complex grounding. We aim to implement our dataset with advanced models like Flux-dev [21] as future work. + +We refer readers to our supplementary materials for more detailed examples and limitation discussions. + +# Acknowledgments + +This work is supported by NSF China (No. 62421003) and the XPLORER PRIZE. + +# 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] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation, 2023. 1 +[3] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.4 +[4] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2 +[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2 +[6] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343-5353, 2024. 3 +[7] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei 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. arXiv preprint arXiv:2404.16821, 2024. 2, 4 +[8] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 2, 3, 4, 5 +[9] Peter Bevan Christoph Schuhmann. Laion pop: 600,000 high-resolution images with detailed descriptions. https://laion.ai/blog/laion-pop, 2023.4 +[10] 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, 5 +[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. 2 +[12] Yutong Feng, Biao Gong, Di Chen, Yujun Shen, Yu Liu, and Jingren Zhou. Ranni: Taming text-to-image diffu + +sion for accurate instruction following. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4744-4753, 2024. 2 +[13] Roopal Garg, Andrea Burns, Burcu Karagol Ayan, Yonatan Bitton, Ceslee Montgomery, Yasumasa Onoe, Andrew Bunner, Ranjay Krishna, Jason Baldridge, and Radu Soricut. Imageinwords: Unlocking hyper-detailed image descriptions. arXiv preprint arXiv:2405.02793, 2024. 3 +[14] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 2 +[15] 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. 2 +[16] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 2 +[17] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2, 3 +[18] Jooyeon Kim, Eulrang Cho, Sehyung Kim, and Hyunwoo J Kim. Retrieval-augmented open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17427-17436, 2024. 2, 3 +[19] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In ICCV, 2023. 1 +[20] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 2,3,7 +[21] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.2,8 +[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 International conference on machine learning, pages 19730–19742. PMLR, 2023. 2 +[23] Lianian 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 3 +[24] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 2 + +[25] 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. 1, 2, 3, 5, 6, 7 +[26] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained Chinese understanding. arXiv preprint arXiv:2405.08748, 2024. 2 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 1, 2, 3 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[29] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 5 +[30] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 5 +[31] Yasumasa Onoe, Sunayana Rane, Zachary Berger, Yonatan Bitton, Jaemin Cho, Roopal Garg, Alexander Ku, Zarana Parekh, Jordi Pont-Tuset, Garrett Tanzer, et al. Docci: Descriptions of connected and contrasting images, 2024. URL https://arxiv.org/abs/2404.19753. 3 +[32] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.2 +[33] 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. 2 +[34] 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 +[35] 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 +[36] Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D Manning. Stanza: A python natural language processing toolkit for many human languages. arXiv preprint arXiv:2003.07082, 2020. 4 +[37] 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. 6 + +[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] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. Stable diffusion v1.4 checkpoint. https://github.com/CompVis/stable-diffusion, 2022.6 +[40] 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 +[41] Christoph Schuhmann. Laion-aesthetics. https:// laion.ai/blog/laion-aesthetics, 2022. 4, 6 +[42] 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. 4 +[43] 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 Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 2, 3 +[44] Vasu Singla, Kaiyu Yue, Sukriti Paul, Reza Shirkavand, Mayuka Jayawardhana, Alireza Ganjdanesh, Heng Huang, Abhinav Bhatele, Gowthami Somepalli, and Tom Goldstein. From pixels to prose: A large dataset of dense image captions. arXiv preprint arXiv:2406.10328, 2024. 3 +[45] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2 +[46] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2 +[47] Hao Wang, Pengzhen Ren, Zequn Jie, Xiao Dong, Chengjian Feng, Yinlong Qian, Lin Ma, Dongmei Jiang, Yaowei Wang, Xiangyuan Lan, et al. Ov-dino: Unified open-vocabulary detection with language-aware selective fusion. arXiv preprint arXiv:2407.07844, 2024. 2, 3, 4, 5 +[48] Jiaqi Wang, Pan Zhang, Tao Chu, Yuhang Cao, Yujie Zhou, Tong Wu, Bin Wang, Conghui He, and Dahua Lin. V3det: Vast vocabulary visual detection dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19844-19854, 2023. 2, 3 +[49] 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. 2, 5, 7 +[50] 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. 2 +[51] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation, 2024. 1, 2, 3, 4, 5, 6, 7 +[52] Yinwei Wu, Xianpan Zhou, Bing Ma, Xuefeng Su, Kai Ma, and Xinchao Wang. Ifadapter: Instance feature control for grounded text-to-image generation. arXiv preprint arXiv:2409.08240, 2024. 1, 2, 4, 5 +[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] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2 +[55] Lewei Yao, Renjie Pi, Jianhua Han, Xiaodan Liang, Hang Xu, Wei Zhang, Zhenguo Li, and Dan Xu. Detclipv3: Towards versatile generative open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27391-27401, 2024. 2, 3 +[56] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2 +[57] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems, 35:36067-36080, 2022. 3 +[58] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 4, 6 +[59] Shiyu Zhao, Samuel Schulter, Long Zhao, Zhixing Zhang, Yumin Suh, Manmohan Chandraker, Dimitris N Metaxas, et al. Improving pseudo labels for open-vocabulary object detection. arXiv preprint arXiv:2308.06412, 2023. 2, 3 +[60] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Migc++: Advanced multi-instance generation controller for image synthesis, 2024. 1, 2, 3, 5 +[61] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis, 2024. 1, 2, 3, 5, 7 \ No newline at end of file diff --git a/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/images.zip b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d1e75b4d49db10434fb1700966d45679a2932167 --- /dev/null +++ b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0b237f8108d77a7684d43cd8bac3fd931d3b404fe606e727aa0cd4f3af9ba81 +size 715180 diff --git a/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/layout.json b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d279729df7b4e44b67e70f9b216a13505d2b430d --- /dev/null +++ b/ICCV/2025/ROVI_ A VLM-LLM Re-Captioned Dataset for Open-Vocabulary Instance-Grounded Text-to-Image Generation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ade4510c7c5d765a481b0d85725e54436ad9e158bd08ee949c3eed1c6759498 +size 479161 diff --git a/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/94cd09f3-cf1f-46f6-98bc-304743ee47c4_content_list.json b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/94cd09f3-cf1f-46f6-98bc-304743ee47c4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1321bb317a233efcb65465958a6be3c5790b0ae1 --- /dev/null +++ b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/94cd09f3-cf1f-46f6-98bc-304743ee47c4_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f57636f5c6b65a82167b51f217a693ef2cbfa9c73f15b2e427329189a268c9be +size 87513 diff --git a/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/94cd09f3-cf1f-46f6-98bc-304743ee47c4_model.json b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/94cd09f3-cf1f-46f6-98bc-304743ee47c4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c60d40f94ce228c83707c410433b20bafa393a9d --- /dev/null +++ b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/94cd09f3-cf1f-46f6-98bc-304743ee47c4_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dfe83516a3d88b26a94eaa5cfe9c79591ae3d990136af7a2f520dc19b6d31e4 +size 112620 diff --git a/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/94cd09f3-cf1f-46f6-98bc-304743ee47c4_origin.pdf b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/94cd09f3-cf1f-46f6-98bc-304743ee47c4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cba73c5e7f31e206edcee3e3cf1e84d389c077f0 --- /dev/null +++ b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/94cd09f3-cf1f-46f6-98bc-304743ee47c4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ef3a0c0ba5871df69b858bd3f817cdcac598e95ab3c50159e263904ff1cc69d +size 5705794 diff --git a/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/full.md b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5781df94027b6dd8fd5895e36e3ea5260467c946 --- /dev/null +++ b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/full.md @@ -0,0 +1,324 @@ +# RS-vHeat: Heat Conduction Guided Efficient Remote Sensing Foundation Model + +Huiyang Hu $^{1,2,3,4}$ Peijin Wang $^{1,2,3,4}$ * Hanbo Bi $^{1,2,3,4}$ Boyuan Tong $^{1,2,3,4}$ Zhaozhi Wang $^{2,3,5}$ + +Wenhui Diao $^{1,2,3,4}$ Hao Chang $^{1,2,3,4}$ Yingchao Feng $^{1,2,3,4}$ Ziqi Zhang $^{2}$ Yaowei Wang $^{5}$ Qixiang Ye $^{2,3,5}$ Kun Fu $^{1,2,3,4}$ Xian Sun $^{1,2,3,4}$ + +$^{1}$ Aerospace Information Research Institute, Chinese Academy of Sciences $^{2}$ University of Chinese Academy of Sciences (UCAS) +$^{3}$ School of Electronic, Electrical and Communication Engineering, UCAS $^{4}$ Key Laboratory of Target Cognition and Application Technology (TCAT) $^{5}$ Peng Cheng Lab + +{huhuiyang22, wangpeijin17, bihanbo21, tongboyuan22, wangzhaozhi22}@mails.ucas.ac.cn + +# Abstract + +Remote sensing foundation models largely break away from the traditional paradigm of designing task-specific models, offering greater scalability across multiple tasks. However, they face challenges such as low computational efficiency and limited interpretability, especially when dealing with large-scale remote sensing images. To overcome these, we draw inspiration from heat conduction, a physical process modeling local heat diffusion. Building on this idea, we are the first to explore the potential of using the parallel computing model of heat conduction to simulate the local region correlations in high-resolution remote sensing images, and introduce RS-vHeat, an efficient multi-modal remote sensing foundation model. Specifically, RS-vHeat 1) applies the Heat Conduction Operator (HCO) with a complexity of $O(N^{1.5})$ and a global receptive field, reducing computational overhead while capturing remote sensing object structure information to guide heat diffusion; 2) learns the frequency distribution representations of various scenes through a self-supervised strategy based on frequency domain hierarchical masking and multi-domain reconstruction; 3) significantly improves efficiency and performance over state-of-the-art techniques across 4 tasks and 10 datasets. Compared to attention-based remote sensing foundation models, we reduce memory usage by $84\%$ , FLOPs by $24\%$ and improves throughput by 2.7 times. The code will be made publicly available. + +# 1. Introduction + +Recently, remote sensing (RS) technology has become a vital data source for scientific research, resource management, and environmental monitoring [30, 47, 52, 76] + +![](images/f1569c29bcc98ee9ef8fcce55611633c5321084e9713900ffbe29c35f38ccc6d.jpg) +Figure 1. Calculation of RS images in the heat conduction system. In heat conduction theory, different materials exhibit varying diffusion rates. Inspired by this, the heat-conduction-based encoder maps optical (OPT) and synthetic aperture radar (SAR) into a unified thermal space, enhancing sensitivity to shape features by simulating heat flow and accumulation within the object region. + +by capturing surface information via satellites. Traditional models, designed as single-task networks for specific RS tasks [36, 50], struggle with multi-payload, multi-resolution, multi-temporal, and multi-feature RS data [20]. However, the emergence of remote sensing foundation models (RSFMs) has overcome these limitations, enabling unified handling of diverse tasks scenarios, significantly enhancing the scalability and versatility of the models [2, 3, 15, 31, 39-41, 44, 53, 56, 59, 60, 70]. RSFMs can automatically extract features from remote sensing imagery (RSI), providing a robust foundation for real-world tasks. + +Prior RSFMs typically use visual encoders and decoders, leveraging large-scale RS datasets for pre-training while + +incorporating self-supervised learning strategies. Specifically, RSFMs can be categorized based on the type of backbone used: CNN-based methods often use ResNet18/50 [25] to extract features [2, 31, 39, 40], while attention-based methods mainly use ViT [17] or Swin transformer [35] with attention mechanisms [55] to capture global dependencies [3, 15, 41, 44, 53, 56]. Both approaches use strategies such as masked reconstruction, knowledge distillation, or contrastive learning during pre-training to enhance model robustness (a detailed comparison is available in the supplementary materials). Despite the significant advancements achieved by these RSFMs, they still face two limitations: + +Balancing Efficiency and Receptive Field. To accurately capture information about the large objects in RSI, model outputs must be responsive to sufficiently large regions [37]. However, this need significantly increases computational complexity [14, 38]. CNN-based networks lack a global receptive field due to their reliance on sliding computations with fixed-size convolutional kernels. While attention-based models achieve global modeling, their attention mechanisms incur quadratic computational complexity. Therefore, existing RSFMs struggle to deliver both fast and high-accuracy inference in practical applications. + +Weak Physical Interpretability. RS objects often exhibit irregular polygonal shapes [29], and current RSFMs struggle to integrate physical principles to explain how object features propagate. This deficiency makes it challenging for researchers to adjust learning strategies effectively [43, 46, 54]. In the long term, RSFMs need to possess a certain degree of information interpretability. + +To address these, this paper introduces RS-vHeat, a heat-conduction-based RSFM that supports multi-modal inputs inspired by the idea of vHeat [63]. First, heat conduction represents the natural process of energy diffusion from high- to low-temperature regions, as shown in the conceptual physical model in Fig. 1. This process transitions from an unsteady to a steady state based on the material. Since its computational process resembles feature extraction in neural networks, it can be applied to RS image processing. Second, we hypothesize that object types correspond to special feature distributions, with the model predicting diffusion rates based on RS-specific properties, simulating parameter computation through heat flow. This approach projects all modalities into a common thermal space, following the constraint that complex areas containing RS objects are high-temperature regions where heat accumulates, while sparse regions are low-temperature areas where heat diffuses easily, as shown in the lower part of Fig. 1. Third, building on this theory, we further design a RSFM with physical interpretability using 3 million optical and SAR data for pre-training, as shown in Fig. 2. The heat conduction network simulates the diffusion process of heat across large-scale multi-modal RS data, facilitating the alignment + +of feature propagation with the structural characteristics of the objects. Moreover, its computational approach provides guidance for the efficient operation of the network. + +To summarize, our contributions are as follows: + +1. We introduce RS-vHeat, a RSFM designed based on the heat conduction differential equation to process RS data. It conceptualizes the semantic relationships between pixels in RS images as the propagation of heat. +2. We propose a self-supervised strategy based on frequency domain hierarchical masking and multi-domain reconstruction that preserves small objects, driving the model to reconstruct fine and coarse frequency signals. +3. We design spatial correction embeddings, which operate directly on the global features to capture local details, assisting in simulating the rate of thermal diffusion. +4. We evaluate RS-vHeat on 10 datasets, showing it outperforms advanced RSFMs in accuracy while maintaining lower computational complexity. When processing large-scale images, RS-vHeat reduces memory usage by $84\%$ , decreases FLOPs by $24\%$ , and improves throughput by 2.7 times compared to attention-based RSFMs. + +# 2. Related Work + +# 2.1. Mainstream Network Architectures in RS + +Convolutional Neural Networks. CNNs have been adapted for RS to tackle challenges in satellite imagery [28, 74]. However, local receptive fields, mainly limited by the kernel size, restrict the capture of long-range dependencies [17]. This drawback is especially critical in RS, where large-scale patterns and complex spatial relationships are prevalent, making the balance between local feature extraction and global context a key research focus [11, 16]. + +Transformers. Self-attention mechanisms [55] empower networks to capture long-range dependencies [17, 35]. Research studies have shown that plain ViTs performs better than traditional CNN models in RS applications [56, 72]. However, the limitation stems from the ViTs' restricted ability to effectively handle long sequences. When processing large-scale RS images, the computational load of networks grows quadratically, leading to significant overhead. + +Mamba-Based Models. Mamba, an efficient implementation of State Space Models (SSMs) [23], leverages selective scan mechanisms for long-sequence processing [34, 78], balancing computational efficiency and high accuracy. [22, 71]. It has been rapidly adopted in RS [8, 9, 79], with its ability to process long sequences, although Mamba, as an innovative architecture, suffers from limited interpretability. + +# 2.2. Self-Supervised Learning Strategies for RSFMs + +Contrastive Learning. By establishing rules to distinguish between positive and negative samples, contrastive learning aims to fully understand the relationships between these + +![](images/dd96b30e13c09dfbc5ec3c2749b4a9a2e93d9345863a12e23615616afa0dcf30.jpg) +Figure 2. The pre-training process of RS-vHeat. It performs frequency domain hierarchical masking based on randomly generated sector regions to separate each image into high- and low-frequency components. These component images are fed into the network and projected into the thermal space via embedding layers. The heat diffusion is computed within the heat-conduction-based visual encoder to simulate complex RS objects. The deep features undergo multi-domain reconstruction loss and contrastive loss computation via decoders. + +samples. GASSL [2] treats RS images of the same scene captured at different times as positive pairs, using them as self-supervised signals. Based on seasonal contrast method, SeCo [40] and CACo [39] effectively leverage temporal information within the networks by comparing scene variations across different years or seasons as perceptual signals. SkySense [24] utilizes multi-modal RS images as input and implements a multi-granularity contrastive learning framework. Motivated by the aforementioned methods, we apply contrastive constraints in the thermal space to optical and SAR data, encouraging the model to focus on deep, fine-grained semantic relationships within the RS images. + +Masked Image Modeling. By involving masking parts of the image and predicting the missing information, networks can understand the details of RSI. RingMo [51] attempts to solve the issue that directly masking image patches can easily lead to the loss of small objects by designing an incomplete masking strategy that is implemented proportionally within the patches. SpectralGPT [26] models multispectral data as three-dimensional data and applies masking in three-dimensional space. Scale-MAE [44] masks pixels in the spatial domain and reconstructs high- and low-frequency images to learn image representations at different scales. Inspired by these, to preserve and learn the information of complex RS scenes, we employ dual reconstruction in both the spatial and frequency domains as constraints. + +# 3. Proposed RS-vHeat + +We introduce RS-vHeat, a multi-modal remote sensing foundation model with three key components: a frequency domain hierarchical masking strategy for multi-modal RS data, a visual encoder for modeling internal heat flow within RS images, and decoders for multi-domain reconstruction. As shown in Fig. 2, RS-vHeat enables shared representation learning of multi-modal data and projection into a thermal space to simulate heat propagation. Our architecture is inspired by vHeat [63], but differs in its emphasis on leveraging heat propagation within the self-supervised process, where we introduce a novel internal design. Details of vHeat are provided in the supplementary materials. + +Frequency Domain Hierarchical Masking Strategy. Traditional spatial masking [69] on RS images often obscures smaller objects, hindering reconstruction. Inspired by the frequency-aware dynamic network [68], we implement a frequency domain hierarchical masking strategy to preserve object structure, enabling more accurate heat propagation by focusing on frequency domain signals. + +Specifically, given the input (optical and SAR), denoted as $I(x,y,c) = \{I_o,I_s\} ,I_o\in \mathbb{R}^{H\times W\times 3},I_s\in \mathbb{R}^{H\times W\times 1}$ , two parallel streams process paired and unpaired data. The Discrete Cosine Transform (DCT) is applied along each image dimension, converting $I(x,y)$ to its frequency + +![](images/f55a7fa3df473a0c4c6f8137f6d537bff62441959191f5605d33c0854dabac0f.jpg) +Figure 3. The overall structure of the heat-conduction-based visual encoder, simulates the general solution of heat conduction. SCEs first dynamically perform spatial domain correction, adjusting heat distribution based on the RS scene. The enhanced image undergoes a 2D DCT in the HCO block, interacting with the heat diffusion rate predicted by the FVEs to complete a heat conduction calculation, which is then transformed back into a visual representation via 2D IDCT. + +representation $\tilde{I}(u,v)$ , where low-frequency components are concentrated in the top-left of the spectrum. A sector mask separates high-frequency $\tilde{I}^{high}(u,v)$ and low-frequency $\tilde{I}^{low}(u,v)$ regions, using a random masking rate of $20\% - 30\%$ . Next, the Inverse Discrete Cosine Transform (IDCT) is applied along the image dimensions to revert the data back to the spatial domain, obtaining $I^{low}(x,y)$ and $I^{high}(x,y)$ . Notably, the DCT and IDCT operations are differentiable, enabling efficient computation and seamless integration into network training on both CPUs and GPUs. Details can be found in the supplementary material. + +Heat-Conduction-Based Visual Encoder. High- and low-frequency information from multi-modal data is mapped to a shared thermal space and then fed into the heat-conduction-based visual encoder for thermal simulation. In most RS downstream tasks, areas containing RS objects tend to appear as high-temperature zones, while sparse or empty regions display as low-temperature zones. + +The Heat Conduction Operator (HCO) simulates the process of visual information transmission as thermal conduction. The two-dimensional temperature distribution at time $t$ , $u(x,y,t)$ is extended to a multi-dimensional feature distribution $U(x,y,c,t)$ . The HCO block specifically models the general solution of physical heat conduction: + +$$ +U _ {m} ^ {t} = \mathcal {F} ^ {- 1} \left(\mathcal {F} \left(U _ {m} ^ {0}\right) e ^ {- k \left(\omega_ {x} ^ {2} + \omega_ {y} ^ {2}\right) t}\right) \tag {1} +$$ + +where $U^0$ and $U^t$ represents the input $U(x,y,c,0)$ and output $U(x,y,c,t)$ . We denote the Discrete Fourier Transform (DFT) and its inverse (IDFT) as $\mathcal{F}$ and $\mathcal{F}^{-1}$ . $m\in \{o,s\}$ refers to the modality (optical or SAR). Based on the Neumann boundary condition [12], we replace the 2D DFT and IDFT with the 2D DCT $(\mathrm{DCT}_{2\mathrm{D}})$ and IDCT $(\mathrm{IDCT}_{2\mathrm{D}})$ [48]: + +$$ +U _ {m} ^ {t} = I D C T _ {2 D} \left(D C T _ {2 D} \left(U _ {m} ^ {0}\right) e ^ {- k \left(\omega_ {x} ^ {2} + \omega_ {y} ^ {2}\right) t}\right) \tag {2} +$$ + +where $e^{-k(\omega_x^2 +\omega_y^2)t}$ functions as an adaptive filter in the frequency domain, executing heat conduction. + +A set of learnable frequency value embeddings (FVEs) is predicted to estimate the heat diffusion coefficient $k$ ( $k \coloneqq k(\omega_x, \omega_y)$ ), facilitating adaptive heat transfer, as shown in Fig. 3. The weights are denoted as $W_{FVEs} \in \mathbb{R}^{M \times N \times C}$ : + +$$ +U _ {m} ^ {\prime p} = H C O \left(U _ {m} ^ {p}, W _ {F V E s}\right) \tag {3} +$$ + +where $U$ and $U'$ represent the temperature states of the features from different modalities before and after passing through the HCO block, respectively, with $p \in \{high, low\}$ indicating the frequency level. + +To enhance the spatial representation of different structural components, we introduce a lightweight correction learner. By predicting a set of spatial correction embeddings (SCEs) from large-scale pre-trained images, it interacts with the existing temperature field and perform activation, as shown in Fig. 3. The weight $W_{SCEs} \in \mathbb{R}^{M \times N \times C}$ adaptively adjusts the object boundaries in the spatial domain based on content by adding them to the original image $I(x,y)$ . Given the upper-layer temperature feature $Z_m^p$ , the correction learner can be expressed as $CL$ : + +$$ +Z _ {m} ^ {\prime p} = C L \left(Z _ {m} ^ {p}, W _ {S C E s}\right) \tag {4} +$$ + +Where $Z_{m}^{\prime p}$ represents the temperature features after adaptive adjustment. + +Multi-Domain Reconstruction Decoders. Considering that the restoration of heat distribution is modality-specific, the optical and SAR decoders (Fig. 2 (2)), $D_{o}^{\delta}$ and $D_{s}^{\delta}$ concurrently operate on the encoded outputs $(F_{o}^{\prime p}, F_{s}^{\prime p} \in \mathbb{R}^{M \times N \times C})$ . By employing convolution and pixel shuffling, decoders effectively upsample feature maps. The reconstruction outputs $I'$ for both modalities— $I_{o}^{\prime high}$ , $I_{o}^{\prime low}$ for optical, and $I_{s}^{\prime high}$ , $I_{s}^{\prime low}$ for SAR—are as follows: + +$$ +I _ {o} ^ {\prime p} = D _ {o} ^ {\delta} \left(F _ {o} ^ {\prime p}\right), I _ {s} ^ {\prime p} = D _ {o} ^ {\delta} \left(F _ {s} ^ {\prime p}\right) \tag {5} +$$ + +![](images/53448d94475e0b985e2dd945c94adf7105942c9f8353895b482366f7ee6e1d5e.jpg) +Figure 4. Comparison of throughput (left), GPU memory (middle), and FLOPs (right) across image sizes for RSFMs. RS-vHeat shows significantly higher throughput, lower memory usage and lower FLOPs than current attention-based RSFMs, especially for large-scale images. All tests were performed in a consistent environment using a single A100 (80G) GPU with a batch size of 32. + +![](images/56cbce0e355f8ac94446ce3680ad047dbe4b37c39b6c208fbe88c7ed6cb26b63.jpg) + +![](images/460e3938043f42efb7fa1036c8cfbc7c9b79f689032dd3aef2dc0199e5da66bd.jpg) + +The reconstruction is guided by the $\mathcal{L}_1$ loss in the frequency domain. After applying DCT, the loss $\mathcal{L}_{Fre}$ is computed by measuring the difference between transformed features $\tilde{I}_m^{p}(u,v)$ and $\tilde{I}_m^p (u,v)$ . + +To learn more robust higher-level features, spatial decoders (Fig. 2 (3)) are responsible for fusion and reconstruction. Specifically, outputs, $F_{m}^{\prime p}$ and $F_{m}^{p}$ , from the third and fourth stages extract modality-specific frequency features. The fusion layer integrates the frequency components into higher-level image features, $F_{m}^{\prime}$ and $F_{m}$ : + +$$ +F _ {m} ^ {\prime} = \operatorname {C o n c a t} \left[ C O N V \left(F _ {m} ^ {\prime h i g h}\right), C O N V \left(F _ {m} ^ {\prime l o w}\right) \right] \tag {6} +$$ + +$$ +F _ {m} = C o n c a t [ C O N V (F _ {m} ^ {h i g h}), C O N V (F _ {m} ^ {l o w}) ] +$$ + +Furthermore, the integration of multi-stage features is achieved through the function $P_{m} = g(F_{m},F_{m}^{\prime})$ , where $g$ combines convolution, ReLU activation and element-wise summation operations to collectively produce the final high-level output $P_{m} = \{P_{o},P_{s}\}$ . Subsequently, these features are fed into modality-specific spatial domain decoders $D_{o}^{\phi}$ and $D_{s}^{\phi}$ , they are upsampled to the original image size to restore the high-level semantics of the original image: + +$$ +I _ {o} ^ {\prime} = D _ {o} ^ {\phi} \left(P _ {o}\right), I _ {s} ^ {\prime} = D _ {s} ^ {\phi} \left(P _ {s}\right) \tag {7} +$$ + +Where $I_o'$ and $I_s'$ represent the reconstructed high-level optical and SAR data. The reconstruction loss in the spatial domain $\mathcal{L}_{Spa}$ is then computed using the $\mathcal{L}_1$ loss. + +To explore the network's understanding of fine-grained semantic information, we employ contrastive loss to compute the semantic discrepancy between different modalities. The loss operates on embeddings $F_{m}^{p}$ obtained from two different preprocessing methods applied to the same image, representing low- and high-frequency features. The contrastive loss $\mathcal{L}_{Con}$ is calculated leveraging cosine similarity, following the computational methodology outlined in [1]. + +The overall loss function is shown in Eq. (8). The combined approach encourages the model to fully capture both global structures and fine-grained details: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {C o n}} + \mathcal {L} _ {\text {S p a}} + \mathcal {L} _ {\text {F r e}} \tag {8} +$$ + +# 4. Experiments + +# 4.1. Training Implementation + +The training of RS-vHeat consists of self-supervised pretraining and downstream fine-tuning. The visual encoder follows the Swin-B [35] configuration with four stages of 2, 2, 18, and 2 blocks. During pre-training, we use a largescale multi-modal dataset with $450\mathrm{k}$ matched optical and SAR image pairs, totaling over 3 million entries, following the methodology of RingMo [51]. The model is trained on eight A100 (80G) GPUs for 200 epochs with images of size 224. The training starts with a learning rate of 1e-6 for 10 warm-up epochs, gradually increasing to 2e-4, with a cosine annealing schedule and a minimum learning rate of 1e-5. For fine-tuning, we transfer the pre-trained embedding layers and visual encoder structure, adjusting the fixed-size FVEs and SCEs to match the image dimensions. + +# 4.2. Performance + +We comprehensively evaluate the performance of RS-vHeat compared to other representative RSFMs. Primarily, as shown in Fig. 4, we analyze throughput, memory usage, and FLOPs across various image sizes. RS-vHeat (depicted by the red line) demonstrates significantly higher throughput, lower memory consumption and FLOPs, with these advantages becoming more pronounced as image sizes increase. For instance, when processing $1024 \times 1024$ resolution images, RS-vHeat achieves 2.7 times the throughput of SwinB-based models such as SatLas [3], GFM [41], RingMo [51], and CMID [42], while reducing memory usage by $84\%$ . Similarly, when compared to the improved ViT-B-based SpectralGPT [26], RS-vHeat achieves 2.5 times the throughput while reducing memory usage by $77\%$ . + +Next, we evaluate the accuracy and computational efficiency of fine-tuning RS-vHeat on four tasks—semantic segmentation, object detection, classification and change detection—to demonstrate the effectiveness, along with extensive ablation studies. Many RSFMs primarily focus on optical data and do not support SAR input, so we compare RS-vHeat with other RSFMs using optical datasets, while SAR-specific tasks are compared against specialized mod- + +Table 1. Comparison of mF1, mIoU, parameters, and FLOPs on iSAID and Potsdam with other RSFMs. + +
Backbone TypeMethodBackbonePotsdam mF1 ↑iSAID mIoU ↑Image SizeParamsFLOPs ↓
CNN-BasedGASSL (ICCV'2021) [2]ResNet-5091.2765.95896264M722G
SeCo (ICCV'2021) [40]ResNet-5089.0357.20896264M722G
SSL4EO (GRSM'2023) [58]ResNet-5091.5464.01896264M722G
CACo (CVPR'2023) [39]ResNet-1891.3564.32896241M671G
TOV (JSTARS'2023) [53]ResNet-5092.0366.24896264M722G
SAMRS (NeurIPS'2023) [57]ResNet-5091.4366.26896264M722G
RS-vHeat (Ours)vHeat-B + SCEs92.8268.728962148M921G
ViT-BasedRVSA (TGRS'2022) [56]ViT-B + RVSA-64.498962128M1043G
SatMAE (NeurIPS'2022)[15]ViT-L90.6362.978962341M1536G
Scale-MAE (ICCV'2023) [44]ViT-L91.5465.778962341M1536G
RS-vHeat (Ours)vHeat-B + SCEs92.8268.728962148M921G
Swin-BasedRingMo (TGRS'2022) [51]Swin-B91.2767.208962121M968G
SatLas (ICCV'2023) [3]Swin-B91.2868.718962121M968G
GFM (ICCV'2023) [41]Swin-B91.8566.628962121M968G
CMID (TGRS'2023) [42]Swin-B91.8666.218962121M968G
SkySense (CVPR'2024) [24]Swin-H93.9970.918962>702M>2708G
RS-vHeat (Ours)vHeat-B + SCEs92.8268.728962148M921G
+ +Table 2. Comparison of OA and user's accuracy on WHU-OPT-SAR with other specialized models. + +
MethodPublicationUser's Accuracy ↑OA ↑
FarmlandCityVillageWaterForestRoadOthers
SegFormer [67]NeurIPS'202179.172.938.064.788.10.30.475.5
Segmenter [49]ICCV'202182.375.251.774.489.416.012.079.9
MCANet [32]JAG'202274.362.253.165.795.531.09.882.9
VMamba [34]NeurIPS'202482.174.157.881.489.141.518.481.4
MMOKD [33]TGRS'202470.058.150.069.780.140.125.082.5
RS-vHeat (Ours)-81.167.367.579.090.254.955.383.9
+ +Table 3. Comparison of mIoU, OA, and AA on AIR-PolSAR-Seg with other specialized models. + +
MethodPublicationmIoU ↑OA ↑AA ↑
DeepLab V3+ [10]ECCV'201848.2176.8163.55
EncNet [75]CVPR'201847.7575.6757.51
PSANet [77]ECCV'201847.1476.2162.92
CCNet [27]ICCV'201946.4675.5355.83
DANet [18]CVPR'201951.9376.9162.79
GCNet [5]ICCV'201947.5676.7557.10
RS-vHeat (Ours)-57.4681.4665.92
+ +els. RSFMs are categorized into three main groups based on the backbone: CNN-based, ViT-based, and Swin-based. All datasets follow official partitioning methods for training and testing. More details on datasets and visualizations are provided in the supplementary materials. + +Single- and Multi-Modal Semantic Segmentation. Evaluation is conducted on two optical datasets (Potsdam [47], iSAID [64]), one SAR dataset (Air-PolSAR-Seg [61]), and one multi-modal dataset (WHU-OPT-SAR [32]), using UPerNet [66] as the output head. Tab. 1 compares RS-vHeat with 14 other RSFMs on iSAID and Potsdam. RS-vHeat surpasses CNN-based models in accuracy and + +achieves lower FLOPs than ViT-based models, improving over Scale-MAE [44] by $1.28\%$ and $2.95\%$ . It also exhibits the lowest FLOPs among Swin-based models, and substantially reduces parameters compared to SkySense [24], with only $1.17\%$ and $2.19\%$ decreases in accuracy. Besides, on AIR-PolSAR-Seg (Tab. 3) and WHU-OPT-SAR (Tab. 2), RS-vHeat outperforms other specialized segmentation models, demonstrating the applicability of the heat conduction model in SAR and multi-modal segmentation tasks. + +Object Detection. We conduct coarse- and fine-grained experiments on two optical datasets (FAIR1M [52], DIOR [30]) and one SAR dataset (SAR-AIRCraft-1.0 [62]), usingYOLOX [21] as the output head. On the DIOR dataset, Tab. 4 shows that RS-vHeat demonstrates not only lower FLOPs but also surpasses the results of SkySense [24] by $3.57\%$ . As shown in Tab. 6 and Tab. 5, RS-vHeat also achieves strong performance and high computational efficiency on two additional fine-grained datasets. + +Change Detection. We train and test on the LEVIR-CD dataset [6], employing the BIT architecture [7] with cross-entropy loss. RS-vHeat demonstrates superior adaptability, achieving an F1 score of $93.48\%$ outperforming existing methods and surpassing SkySense [24] by 0.9 points, as + +Table 4. Comparison of ${\mathrm{{mAP}}}_{50}$ ,parameters,and FLOPs on DIOR with other RSFMs. + +
Backbone TypeMethodBackboneImage SizemAP50↑ParamsFLOPs ↓
CNN-BasedGASSL (ICCV'2021) [2]ResNet-50800267.4041M134G
CACo (CVPR'2023) [39]ResNet-18800266.9128M101G
TOV (JSTARS'2023) [53]ResNet-50800270.1641M134G
SSL4EO (GRSM'2023) [58]ResNet-50800264.8241M134G
RS-vHeat (Ours)vHeat-B + SCEs800282.30128M266G
ViT-BasedRVSA (TGRS'2022) [56]ViT-B + RVSA800273.22113M378G
SatMAE (NeurIPS'2022)[15]ViT-L800270.89324M1094G
Scale-MAE (ICCV'2023) [44]ViT-L800273.81324M1094G
RS-vHeat (Ours)vHeat-B + SCEs800282.30128M266G
Swin-BasedRingMo (TGRS'2022) [51]Swin-B800275.90105M322G
SatLas (ICCV'2023) [3]Swin-B800274.10105M322G
GFM (ICCV'2023) [41]Swin-B800272.84105M322G
CMID (TGRS'2023) [42]Swin-B800275.11105M322G
SkySense (CVPR'2024) [24]Swin-H800278.73>674M>1679G
RS-vHeat (Ours)vHeat-B + SCEs800282.30128M266G
+ +Table 5. Comparison of mAP, parameters, and FLOPs on FAIR1M-2.0 with other RSFMs. + +
MethodBackbonemAP ↑ParamsFLOPs ↓
CACo [39]ResNet-1847.8342M64G
GASSL [2]ResNet-5048.1555M77G
TOV [53]ResNet-5049.6255M77G
SSL4EO [58]ResNet-5049.3755M77G
RVSA [56]ViT-B + RVSA47.04126M160G
SatMAE [15]ViT-L46.55336M241G
Scale-MAE [44]ViT-L48.31336M241G
SatLas [3]Swin-B46.19119M167G
GFM [41]Swin-B49.69119M167G
RingMo [51]Swin-B46.21119M167G
CMID [42]Swin-B50.58119M167G
SkySense [24]Swin-H54.57>688M>900G
RS-vHeatvHeat-B + SCEs48.29130M137G
+ +Table 6. Comparison of ${\mathrm{{mAP}}}_{50}$ and ${\mathrm{{mAP}}}_{75}$ on SAR-AIRcraft-1.0 with other specialized models. + +
MethodPublication\( \mathrm{mAP}_{50} \uparrow \)\( \mathrm{mAP}_{75} \uparrow \)
Faster R-CNN [45]TPAMI'201676.162.2
Cascade R-CNN [4]CVPR'201875.758.9
RepPoints [73]ICCV'201972.653.3
SKG-Net [19]JSTARS'202170.746.4
SA-Net [62]RADARS'202377.762.8
RS-vHeat (Ours)-87.167.4
+ +shown in Tab. 7. Despite a 256-pixel input size, RS-vHeat maintains FLOPs comparable to the Swin-B-based baseline while delivering higher precision. + +Image Classification. We validate our model on two benchmark datasets (AID [65], NWPU-RESISC45 [13]) by attaching a classification head designed, and employ cross-entropy loss for computation. As shown in Tab. 8, RS-vHeat surpasses CNN-based models like CACo [39] in classification accuracy. Compared to ViT- and Swin-based models, + +Table 7. Comparison of F1 on LEVIR-CD with other RSFMs. + +
MethodBackboneF1 ↑ParamsFLOPs ↓
CACo [39]ResNet-1881.0412M11G
GASSL [2]ResNet-5078.1927M25G
SeCo [40]ResNet-5090.1427M25G
SSL4EO [58]ResNet-5089.0527M25G
RVSA [56]ViT-B + RVSA90.8694M57G
SatMAE [15]ViT-L87.65304M162G
Scale-MAE [44]ViT-L92.07304M162G
SatLas [3]Swin-B90.6288M45G
GFM [41]Swin-B91.7388M45G
RingMo [51]Swin-B91.8688M45G
CMID [42]Swin-B91.7288M45G
SkySense [24]Swin-H92.58>656M>307G
RS-vHeat (Ours)vHeat-B + SCEs93.4893M46G
+ +RS-vHeat demonstrates superior computational efficiency while also achieving competitive accuracy. + +# 4.3. Ablations + +To validate the effectiveness of the features learned through pre-training based on heat conduction theory, we conduct ablation studies on key components. + +Effectiveness of Masked Training. We visualize the accuracy curves of the visual encoder before and after introducing new structures across downstream tasks, as shown in Fig. 5. All three networks are pre-trained on the same multi-modal RS data. In (a), we use vHeat-B with the SimMIM [69] pixel mask training scheme; in (b), we add frequency domain masking and multi-domain reconstruction; and in (c), we integrate the innovative RS-vHeat structure, which includes SCEs. For optical tasks, RS-vHeat shows faster accuracy improvement, stabilizing at a higher ceiling. It surpasses $80\%$ $\mathrm{mAP}_{50}$ on the DIOR dataset by the 120th epoch. For SAR tasks, while there are fluctuations, RS-vHeat demonstrates better adaptability and prediction accuracy. Overall, the improvements in pre-training and SCEs + +Table 8. Comparison of OA, parameters, and FLOPs on AID and NWPU-RESISC45 with other RSFMs. + +
Backbone TypeMethodBackboneAIDNWPU-RESISC45Image SizeParamsFLOPs ↓
TR=20%TR=50%TR=10%TR=20%
CNN-BasedGASSL (ICCV'2021) [2]ResNet-5093.5595.9290.8693.061024224M87G
SeCo (ICCV'2021) [40]ResNet-5093.4795.9989.6492.911024224M87G
CACo (CVPR'2023) [39]ResNet-1890.8895.0588.2891.941024211M38G
TOV (JSTARS'2023) [53]ResNet-5095.1697.0990.9793.791024224M87G
SSL4EO (GRSM'2023) [58]ResNet-5091.0694.7487.6091.271024224M87G
RS-vHeat (Ours)vHeat-B + SCEs96.8197.5892.0195.6610242150M340G
ViT-BasedRVSA (TGRS'2022) [56]ViT-B + RVSA97.0398.5093.9395.691024289M460G
SatMAE (NeurIPS'2022)[15]ViT-L95.0296.9491.7294.1010242310M2070G
Scale-MAE (ICCV'2023) [44]ViT-L96.4497.5892.6395.0410242310M2070G
RS-vHeat (Ours)vHeat-B + SCEs96.8197.5892.0195.6610242150M340G
Swin-BasedRingMo (TGRS'2022) [51]Swin-B96.9098.3494.2595.671024287M450G
SatLas (ICCV'2023) [3]Swin-B94.9697.3892.1694.701024287M450G
GFM (ICCV'2023) [41]Swin-B95.4797.0992.7394.641024287M450G
CMID (TGRS'2023) [42]Swin-B96.1197.7994.0595.531024287M450G
SkySense (CVPR'2024) [24]Swin-H97.6898.6094.8596.3210242>660M>2760G
RS-vHeat (Ours)vHeat-B + SCEs96.8197.5892.0195.6610242150M340G
+ +Table 9. Results of RS-vHeat under various constraints of self-supervised learning strategies. + +
#LossObject DetectionSemantic Segmentation
DIOR (Optical)SAR-AIRcraft-1.0 (SAR)iSAID (Optical)AIR-PolSAR-Seg (SAR)WHU-OPT-SAR (Optical+SAR)
SDRFDRCLmAP50↑mAP50↑mIoU↑mIoU↑OA↑
(a)✓××78.284.666.154.780.5
(b)✓✓×79.586.467.255.182.3
(c)✓✓✓82.387.168.757.583.9
+ +![](images/b3926a7ddd392db341a22e2222cb79a01f9f54c12ed3eb81b1ec21f9fda3b0e2.jpg) +Figure 5. Accuracy curves of various methods across modalities. (a) optical object detection on DIOR, (b) SAR object detection on SAR-AIRcraft-1.0, (c) optical semantic segmentation on iSAID, and (d) SAR semantic segmentation on AIR-PoSAR-Seg. + +lead to advanced feature extraction performance. + +Reconstruction Learning Strategy. Tab. 9 demonstrates the effectiveness of different loss components—spatial domain reconstruction loss (SDR), frequency domain reconstruction loss (FDR), and contrastive loss (CL). When only SDR is applied (row a), performance drops slightly on both optical and SAR datasets. Adding FDR (row b) improves performance by leveraging frequency domain information. Introducing CL (row c) further enhances performance by enforcing similarity constraints in the multi-modal feature space, improving feature learning. This leads to an $83.9\%$ OA on WHU-OPT-SAR, surpassing rows (a) and (b) by $3.4\%$ and $1.6\%$ , respectively. These results highlight the importance of combining these three losses for improved predictions across tasks and modalities. + +# 5. Conclusion + +In this work, we introduce the concept of heat conduction into RS tasks for the first time and establish a multi-modal R SFM, RS-vHeat. By employing a self-supervised learning strategy that integrates frequency domain masking and multi-domain reconstruction, along with a heat conduction operator incorporating spatial correction embeddings, we propose an approach that balances computational complexity and global receptive field coverage in RS. Our method captures global details from both spatial and frequency domains, significantly reducing the issue of small object omission, and achieves consistent multi-modal feature representation by mapping images into a high-dimensional thermal space. In future work, we plan to propose new solutions that offer fresh insights and methods for tackling visual modeling challenges across various industries. + +# 6. Acknowledgment + +# References + +[1] Sara Atito, Muhammad Awais, and Josef Kittler. Sit: Self-supervised vision transformer. arXiv preprint arXiv:2104.03602, 2021. 5 +[2] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10190, 2021. 1, 2, 3, 6, 7, 8 +[3] Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinando, and Aniruddha Kembhavi. Satlaspretrain: A large-scale dataset for remote sensing image understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16772-16782, 2023. 1, 2, 5, 6, 7, 8 +[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7 +[5] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 6 +[6] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10), 2020. 6 +[7] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 6 +[8] Hongruixuan Chen, Jian Song, Chengxi Han, Junshi Xia, and Naoto Yokoya. Changemamba: Remote sensing change detection with spatio-temporal state space model. arXiv preprint arXiv:2404.03425, 2024. 2 +[9] Keyan Chen, Bowen Chen, Chenyang Liu, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsmamba: Remote sensing image classification with state space model. IEEE Geoscience and Remote Sensing Letters, 2024. 2 +[10] 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. 6 +[11] Xi Chen, Zhiqiang Li, Jie Jiang, Zhen Han, Shiyi Deng, Zhihong Li, Tao Fang, Hong Huo, Qingli Li, and Min Liu. Adaptive effective receptive field convolution for semantic segmentation of vhr remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 59(4):3532-3546, 2020. 2 + +This work is supported by the National Natural Science Foundation of China under Grant 62331027, and the Key Laboratory of Target Cognition and Application Technology under Grant 2023-CXPT-LC-005. +[12] Alexander H-D Cheng and Daisy T Cheng. Heritage and early history of the boundary element method. Engineering analysis with boundary elements, 29(3):268-302, 2005. 4 +[13] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 7 +[14] Emmanuel Christophe, Julien Michel, and Jordi Inglada. Remote sensing processing: From multicore togpu. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 4(3):643-652, 2011. 2 +[15] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 1, 2, 6, 7, 8 +[16] Xiaohu Dong, Ruigang Fu, Yinghui Gao, Yao Qin, Yuanxin Ye, and Biao Li. Remote sensing object detection based on receptive field expansion block. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2021. 2 +[17] 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. 2 +[18] 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. 6 +[19] Kun Fu, Jiamei Fu, Zhirui Wang, and Xian Sun. Scatteringkeypoint-guided network for oriented ship detection in high-resolution and large-scale sar images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:11162-11178, 2021. 7 +[20] Kun Fu, Xian Sun, Xiaolan Qiu, Wenhui Diao, Zhiyuan Yan, Lijia Huang, and Hongfeng Yu. Multi-satellite integrated processing and analysis method under remote sensing big data. National Remote Sensing Bulletin, 25(3):691-707, 2021. 1 +[21] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6 +[22] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2 +[23] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2 +[24] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27672-27683, 2024. 3, 6, 7, 8 + +[25] 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. 2 +[26] 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. 3, 5 +[27] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 603-612, 2019. 6 +[28] Haifeng Li, Haikuo Huang, Li Chen, Jian Peng, Haozhe Huang, Zhenqi Cui, Xiaoming Mei, and Guohua Wu. Adversarial examples for cnn-based sar image classification: An experience study. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:1333-1347, 2020. 2 +[29] Jiang Li and Ram M Narayanan. A shape-based approach to change detection of lakes using time series remote sensing images. IEEE transactions on geoscience and remote sensing, 41(11):2466-2477, 2003. 2 +[30] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 1, 6 +[31] Wenyuan Li, Keyan Chen, Hao Chen, and Zhenwei Shi. Geographical knowledge-driven representation learning for remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 1, 2 +[32] Xue Li, Guo Zhang, Hao Cui, Shasha Hou, Shunyao Wang, Xin Li, Yujia Chen, Zhijiang Li, and Li Zhang. Mcanet: A joint semantic segmentation framework of optical and sar images for land use classification. International Journal of Applied Earth Observation and Geoinformation, 106: 102638, 2022. 6 +[33] Xiao Liu, Fei Jin, Shuxiang Wang, Jie Rui, Xibing Zuo, Xiaobing Yang, and Chuanxiang Cheng. Multimodal online knowledge distillation framework for land use/cover classification using full or missing modalities. IEEE Transactions on Geoscience and Remote Sensing, 2024. 6 +[34] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model, 2024. 2, 6 +[35] 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. 2, 5 +[36] Xiaonan Lu, Xian Sun, Wenhui Dao, Yingchao Feng, Peijin Wang, and Kun Fu. Lil: Lightweight incremental learning approach through feature transfer for remote sensing image scene classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-20, 2021. 1 +[37] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolu + +tional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 2 +[38] Yan Ma, Haiping Wu, Lizhe Wang, Bormin Huang, Rajiv Ranjan, Albert Zomaya, and Wei Jie. Remote sensing big data computing: Challenges and opportunities. Future Generation Computer Systems, 51:47-60, 2015. Special Section: A Note on New Trends in Data-Aware Scheduling and Resource Provisioning in Modern HPC Systems. 2 +[39] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change-aware sampling and contrastive learning for satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5261-5270, 2023. 1, 2, 3, 6, 7, 8 +[40] Oscar Manas, Alexandre Lacoste, Xavier Giroi-Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2, 3, 6, 7, 8 +[41] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, and Chen Chen. Towards geospatial foundation models via continual pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 1, 2, 5, 6, 7, 8 +[42] Dilxat Muhtar, Xueliang Zhang, Pengfeng Xiao, Zhenshi Li, and Feng Gu. Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing, 61:1-17, 2023. 5, 6, 7, 8 +[43] Adrián Pérez-Suay, Jose E. Aduvara, María Piles, Laura Martínez-Ferrer, Emiliano Díaz, Alvaro Moreno-Martínez, and Gustau Camps-Valls. Interpretability of recurrent neural networks in remote sensing. In IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium, pages 3991-3994, 2020. 2 +[44] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 1, 2, 3, 6, 7, 8 +[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 7 +[46] Huanfeng Shen, Menghui Jiang, Jie Li, Chenxia Zhou, Qiangqiang Yuan, and Liangpei Zhang. Coupling model- and data-driven methods for remote sensing image restoration and fusion: Improving physical interpretability. IEEE Geoscience and Remote Sensing Magazine, 10(2):231-249, 2022. 2 +[47] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 1, 6 +[48] Gilbert Strang. The discrete cosine transform. SIAM review, 41(1):135-147, 1999. 4 +[49] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmenta + +tion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7262-7272, 2021. 6 +[50] Xian Sun, Peijin Wang, Cheng Wang, Yingfei Liu, and Kun Fu. Pbnet: Part-based convolutional neural network for complex composite object detection in remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 173:50-65, 2021. 1 +[51] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 2022. 3, 5, 6, 7, 8 +[52] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 1, 6 +[53] Chao Tao, Ji Qi, Guo Zhang, Qing Zhu, Weipeng Lu, and Haifeng Li. Tov: The original vision model for optical remote sensing image understanding via self-supervised learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16:4916-4930, 2023. 1, 2, 6, 7, 8 +[54] Anastasios Temenos, Nikos Temenos, Maria Kaselimi, Anastasios Doulamis, and Nikolaos Doulamis. Interpretable deep learning framework for land use and land cover classification in remote sensing using shap. IEEE Geoscience and Remote Sensing Letters, 20:1-5, 2023. 2 +[55] 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. 2 +[56] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2022. 1, 2, 6, 7, 8 +[57] 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. Advances in Neural Information Processing Systems, 36, 2024. 6 +[58] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eos12: A large-scale multimodal, multitemporal dataset for self-supervised learning in earth observation [software and data sets]. IEEE Geoscience and Remote Sensing Magazine, 11(3):98-106, 2023. 6, 7, 8 +[59] Yi Wang, Conrad M Albrecht, and Xiao Xiang Zhu. Multi-label guided soft contrastive learning for efficient earth observation pretraining. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1 +[60] Yi Wang, Conrad M. Albrecht, Nassim Ait Ali Graham, Chenying Liu, Zhitong Xiong, and Xiao Xiang Zhu. Decoupling common and unique representations for multimodal self-supervised learning. In Computer Vision - ECCV 2024, + +pages 286-303, Cham, 2025. Springer Nature Switzerland. +[61] Zhirui Wang, Xuan Zeng, Zhiyuan Yan, Jian Kang, and Xian Sun. Air-pulsar-seg: A large-scale data set for terrain segmentation in complex-scene pulsar images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15:3830-3841, 2022. 6 +[62] Zhirui Wang, Yuzhuo Kang, Xuan Zeng, Yuelei Wang, Ting Zhang, and Xian Sun. Sar-aircraft-1.0: High-resolution sar aircraft detection and recognition dataset. Journal of Radars, 12(4):906–922, 2023. 6, 7 +[63] Zhaozhi Wang, Yue Liu, Yunfan Liu, Hongtian Yu, Yaowei Wang, Qixiang Ye, and Yunjie Tian. vheat: Building vision models upon heat conduction. arXiv preprint arXiv:2405.16555, 2024. 2, 3 +[64] 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/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 6 +[65] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 7 +[66] 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), 2018. 6 +[67] 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. 6 +[68] Wenbin Xie, Dehua Song, Chang Xu, Chunjing Xu, Hui Zhang, and Yunhe Wang. Learning frequency-aware dynamic network for efficient super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4308-4317, 2021. 3 +[69] 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, 7 +[70] Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J. Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural plasticity-inspired multimodal foundation model for earth observation, 2024. 1 +[71] Rui Xu, Shu Yang, Yihui Wang, Bo Du, and Hao Chen. A survey on vision mamba: Models, applications and challenges. arXiv preprint arXiv:2404.18861, 2024. 2 +[72] Tianyu Yan, Zifu Wan, and Pingping Zhang. Fully transformer network for change detection of remote sensing images. In Proceedings of the Asian Conference on Computer Vision, pages 1691-1708, 2022. 2 + +[73] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Reppoints: Point set representation for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9657-9666, 2019. 7 +[74] Chi Zhang, Shiqing Wei, Shunping Ji, and Meng Lu. Detecting large-scale urban land cover changes from very high resolution remote sensing images using cnn-based classification. ISPRS International Journal of Geo-Information, 8(4): 189, 2019. 2 +[75] Hang Zhang, Kristin Dana, Jianping Shi, Zhongyue Zhang, Xiaogang Wang, Ambrish Tyagi, and Amit Agrawal. Context encoding for semantic segmentation. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7151-7160, 2018. 6 +[76] Yuxiang Zhang, Wei Li, Ran Tao, Jiangtao Peng, Qian Du, and Zhaoquan Cai. Cross-scene hyperspectral image classification with discriminative cooperative alignment. IEEE Transactions on Geoscience and Remote Sensing, 59(11): 9646-9660, 2021. 1 +[77] Hengshuang Zhao, Yi Zhang, Shu Liu, Jianping Shi, Chen Change Loy, Dahua Lin, and Jiaya Jia. Pisanet: Pointwise spatial attention network for scene parsing. In Proceedings of the European conference on computer vision (ECCV), pages 267-283, 2018. 6 +[78] 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. 2 +[79] Qinfeng Zhu, Yuanzhi Cai, Yuan Fang, Yihan Yang, Cheng Chen, Lei Fan, and Anh Nguyen. Samba: Semantic segmentation of remotely sensed images with state space model. arXiv preprint arXiv:2404.01705, 2024. 2 \ No newline at end of file diff --git a/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/images.zip b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7f4fec5714e936249fc08a6d0b20ef8b0dd020f3 --- /dev/null +++ b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1359d720f753d3e0b5d04bcf58328e8636eecd1c1a8615acfebecc18155d1a53 +size 1108922 diff --git a/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/layout.json b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dfc70c2b7898f2d0cfb26cc82350bfc954fa5745 --- /dev/null +++ b/ICCV/2025/RS-vHeat_ Heat Conduction Guided Efficient Remote Sensing Foundation Model/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f8a4250bb3e28186adaa536bf0119fd0a593ce9a58e790958ba1c02408c7e56 +size 432995 diff --git a/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/c637999d-0f0d-4030-97b8-2a89b2be133d_content_list.json b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/c637999d-0f0d-4030-97b8-2a89b2be133d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f48bf5d826e146af129a5ec26d834f6999d872fc --- /dev/null +++ b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/c637999d-0f0d-4030-97b8-2a89b2be133d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ae3b1b2792d42f570879226ec01f9039a6491fa956cd352fefd23f16dc9d33b +size 75781 diff --git a/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/c637999d-0f0d-4030-97b8-2a89b2be133d_model.json b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/c637999d-0f0d-4030-97b8-2a89b2be133d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e098789d42f30feb0a4792613529cfe1a57d6d9 --- /dev/null +++ b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/c637999d-0f0d-4030-97b8-2a89b2be133d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f92cf42e452cc336a2a6e6b2d5686875065d114bebab52383c962015f317be6 +size 93601 diff --git a/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/c637999d-0f0d-4030-97b8-2a89b2be133d_origin.pdf b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/c637999d-0f0d-4030-97b8-2a89b2be133d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..80dde2e079e7454df3f93e349d2a307a3a2d1e3b --- /dev/null +++ b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/c637999d-0f0d-4030-97b8-2a89b2be133d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b11afab2b9499eb7f905682c892fc58057b1cf75a9c4cd1c0a5f60a56788849 +size 9659798 diff --git a/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/full.md b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..958abd411153b35cea14486a01fe4e2d196a6e8d --- /dev/null +++ b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/full.md @@ -0,0 +1,296 @@ +# RTMap: Real-Time Recursive Mapping with Change Detection and Localization + +Yuheng Du $^{1,\dagger}$ Sheng Yang $^{1,\dagger,\boxtimes}$ Lingxuan Wang $^{1}$ Zhenghua Hou $^{1}$ Chengying Cai $^{1}$ Zhitao Tan $^{1}$ Mingxia Chen $^{1}$ Shi-Sheng Huang $^{2}$ Qiang Li $^{1}$ $^{1}$ Unmanned Vehicle Dept., CaiNiao Inc., Alibaba Group $^{2}$ Beijing Normal University {guofan.dyh, shengyang}@cainiao.com + +![](images/393fe8e65e5945059141a93b9d51537157124082c656716322cf2ae5bba133b2.jpg) +Figure 1. RTMap performs real-time online HD mapping on onboard driving agents, simultaneously solving map-based localization, map change detection, and online map fusion tasks. Meanwhile, a cloud service asynchronously stitches crowdsourced online HD maps into a global prior-map for subsequent traverses. + +![](images/e884d441308924c7fb95ad2426f2ed06ec9a8f9cf6f9243fc16b6c19bea034b6.jpg) + +# Abstract + +While recent online HD mapping methods relieve burdened offline pipelines and solve map freshness, they remain limited by perceptual inaccuracies, occlusion in dense traffic, and an inability to fuse multi-agent observations. We propose RTMap to enhance these single-traversal methods by persistently crowdsourcing a multi-traversal HD map as a self-evolutional memory. On onboard agents, RTMap simultaneously addresses three core challenges in an end-to-end fashion: (1) Uncertainty-aware positional modeling for HD map elements, (2) probabilistic-aware localization w.r.t. the crowdsourced prior-map, and (3) real-time detection for possible road structural changes. Experiments on several public autonomous driving datasets demonstrate our solid performance on both the prior-aided map quality and the localization accuracy, demonstrating our effectiveness of robustly serving downstream prediction and planning modules while gradually improving the accuracy and freshness of the crowdsourced prior-map asynchronously. Our source-code will be made publicly available at https://github.com/CN-ADLab/RTMap. + +# 1. Introduction + +High-Definition (HD) maps enriched with geometric, semantic, and behavioral annotations, are foundational to autonomous driving systems – enabling robust obstacle avoidance, centimeter-level parking precision, and customized driving policies. As highlighted in a recent survey [35], the field has shifted from reliance on costly offline mapping pipelines to online HD mapping [24–26], which generates maps in real-time using consumer-grade sensors and embedded computing resources. These methods empower vehicles to simultaneously perceive and construct HD maps onboard, unlocking a dual-mode Operational Design Domain (ODD): (1) First-Traversal Autonomy: In unexplored environments, vehicles operate without a prior-map, leveraging real-time perception for basic autonomy, e.g., Lane Centering Control (LCC) and Autonomous Emergency Braking (AEB). (2) Crowdsourced Enhancement: Once aggregated from multi-agent traversals, such fused prior-maps can provide previous perceptual knowledge to extend sensing range, resolve occlusions, and elevate autonomy to L4 standards, e.g., dense urban navigation and unprotected left turns. + +While the community rapidly evolves methods and strategies for the first-traversal autonomy [35], crowd-sourced enhancement remains under-explored following + +such a shifted architecture. Revisiting online HD mapping frameworks [24-26], most of them treat mapping as a single-pass process, failing to effectively utilize the rich contextual knowledge available from multiple traversals of the same scene. To fully harness this multi-traversal knowledge, two fundamental capabilities are essential: (1) precise localization within the prior-map to accurately align the current surroundings with existing spatial data, and (2) robust detection and adaptation to structural changes in road networks – such as lane modifications or construction zones – to maintain map freshness by eliminating outdated information. Although numerous studies have proposed effective methods to solve these problems independently [15, 32], these two sub-tasks and the overall prior-aided online HD mapping task essentially solve the retrieval, correspondence, and differentiation problems between multi-traversal map elements. To address these challenges, we propose RTMap with novel comprehensive query and matching, to solve the element identification and map association in a unified end-to-end framework for addressing all these tasks. Moreover, we further introduce an explicit modeling of the geometric uncertainty of extracted map elements, to facilitate both a probabilistic-aware state estimation of the 6 degrees of freedom (DOF) vehicle poses and noise-aware multi-traversal prior-map fusion. Thus, we lead to highly accurate real-time HD mapping, especially suitable for autonomous driving scenarios. + +In summary, our RTMap offers three key contributions: + +- The first end-to-end framework supports multi-traversal online HD mapping while simultaneously addressing map-based localization and change detection to robustly and stably serve downstream autonomous driving modules. +- By quantitatively inherencing the probabilistic density of perceived vectorized HD map elements, RTMap further improves localization accuracy through end-to-end learning and explicit state estimation. +- RTMap proposes a crowdsourcing mechanism to recursively update the offline prior-map asynchronously. The inferred probabilistic density and change information further improve the accuracy of the prior-aided online HD map. + +Experiments on localization, mapping, and map change detection tasks on multiple publicly available datasets [3, 18] have demonstrated our effectiveness in simultaneously addressing them through an end-to-end framework. + +# 2. Related Work + +In decades, offline mapping pipelines [34, 40] have become the mainstream for mass-producing HD maps [11]. Once these maps are constructed, centimeter-level map-based localization deployed onboard can accurately fetch and transform these map elements for downstream prediction and + +planning tasks. However, recent advances [35] in onboard real-time HD mapping are deeply reforming the solution of mapping (Sec. 2.1) and localization (Sec. 2.2). + +# 2.1. Online HD Mapping + +Single-traversal single moment methods. HDMapNet [20] and MapTR [24, 25] use the encoder-decoder paradigm to transform sensor input into a unified Bird's Eye View (BEV) representation and output an explicit local HD map. VectorMapNet [27] aggregates features generated from different modalities into a common BEV feature space for multi-sensor mapping. GKT [8] leverages calibrated camera parameters to determine the correspondence between 2D positions and BEV grids for constructing full correlations. LaneSegNet [22] proposes a heads-to-regions mechanism for capturing long-range attention and an identical initialization strategy for enhancing the learning of positional priors for lane attention. + +Single-traversal temporal fusion methods. To facilitate the stability of online HD mapping between consecutive frames, many subsequent methods [21, 41] leverage BEVFormer [23] for temporal self-attention to fuse temporally adjacent BEV information recurrently. Other explicit methods [7] use odometry to transform perceived results and perform voxel mapping for multi-frame fusion. MapTracker [6] leverages BEV and vector memory fusion to better exploit temporal information. PrevPredMap [30] use previous predictions as priors. + +Prior-aided methods. Previous works [1, 33] demonstrate that prior-informed online mapping can further improve completeness and stability. Unlike the single-traversal methods mentioned above, prior-aided methods require both localization and change detection to establish correct correspondences between maps. We mainly found two categories of prior-aided methods for online HD mapping: The first category [37, 39] uses neural representation as priormaps, which is storage-intensive and not suitable for largescale deployment or manual inspection. The second category uses Standard-Definition (SD) or HD maps as prior knowledge: SMERF [28] queries the SD map at the ego vehicle's location for lane-topology reasoning, and U-BEV [4] extracts BEV features from both sensors and SD map to reach meters-level accuracy through template matching. P-MapNet [17] leverages SD priors for coarse localization and HD priors for final map refinement. PriorDrive [42] integrates diverse prior maps through the Unified Vector Encoder. Our method belongs to the latter category, stepping further to reach real-time centimeter-level localization accuracy beyond previous methods through an end-to-end model. Meanwhile, RTMap manages to self-enclose the maintenance of prior HD maps, inferencing the probabilistic density of online HD map elements for noise-aware HD crowdsourcing. + +![](images/9bd7cd5e063458b710ca23a5bc284b56832f5a97b7cb8bfb88b2539cbfaf1a5e.jpg) +Backbone +2D to BEV + +![](images/b526c5f6f66208fd0eeea6d08b5d53cb06c18cfadb6d0d3114aa0998e5dfba6a.jpg) +CrowdSrcMap +Fusion + +![](images/3b56c1fe69df4bdea071162470433c58dd95031ea99fd80647437fb1d81479a6.jpg) + +![](images/6294beffed3e085c04f45711c51e4c90c50772ba7e337f40c8d439a002670b28.jpg) +Offline Crowdsourcing +Figure 2. Online and offline modules of RTMap. We encode sensors and the crowdsourced HD map to perform hybrid queries and existence-aware matching to obtain matched, outdated, and newly observed HD map elements online. Matched queries can be used for either a pose head or a maximum a posteriori (MAP) state estimator for obtaining the 6-DOF vehicle poses. During offline, we gather multi-traversal local HD maps to fuse them to improve the accuracy and completeness of the crowdsourced map. + +Change detection. Considering possible road structural changes, recent methods such as ExelMap [36] add additional heads for detecting element-wise insertion and deletion events. Also, recent advances for offline pipelines [38] leverage prior-map encoding and unify the feature space for the predicted and historical instances using distinct multilayer perceptron (MLP) networks to associate and pick up changed instances. Compared to these methods, our method operates onboard and thus instantly serves downstream tasks without relying on engineering practices to ensure the refreshed HD map is efficiently deployed. + +# 2.2. Map-based Localization + +Online HD mapping reforms localization separately into odometry [19, 45] for temporal reasoning and map-based localization [9, 29] for aggregating prior knowledge if such prior exists. We found several learning-based methods contain similar architectures above but are individually designed for map positioning tasks: BEV-Locator [44] adopts the transformer structure in cross-modal feature association to address cross-modality matching between semantic map elements and camera images. EgoVM [15] designs learnable embeddings and a transformer decoder to bridge the representation gap between vectorized HD maps and sensor BEV features to reach centimeter-level localization accuracy. However, these methods treat localization as an isolated task and still require a standalone HD mapping progress for referring to. Considering the main objective of a precise map-based localization - better aggregating prior knowledge for map servicing, we choose to union the lo + +calibration task with all related tasks that may influence its accuracy – online map prediction accuracy, road structural changes, erroneous map elements matching, and possible occlusions – and thus lean toward a multi-task end-to-end manner. + +Meanwhile, inspired by Gu et al. [13], who additionally output uncertainty estimates for downstream prediction tasks, we contribute to introduce uncertainty as thorough information, which can either explicitly or implicitly improve the quality of localization and prior-aided mapping, enabling end-to-end architecture to optionally perform optimization-based state estimation as Simultaneous Localization and Mapping (SLAM) approaches [2, 5, 10, 16, 46] for interpretability while maintaining accuracy. + +# 3. Methodology + +# 3.1. Method Overview + +Problem formulation. In the current traverse $t$ given a single moment of multi-sensor frames $\mathbf{I}_t$ (typically contains calibrated surround view camera frames) and a prior-map $\mathcal{M}_{t-1}$ formed by previous multi-agent traverses, our end-to-end framework simultaneously addresses the following tasks: + +$$ +\begin{array}{l} \left\{\mathbf {M} _ {t}, \mathbf {U} _ {t}, \mathbf {D} _ {t}, \mathbf {T} _ {t} ^ {\mathbb {E}} \right\} \leftarrow \operatorname {R T M a p M o d e l} \left(\mathbf {I} _ {t}, \mathcal {M} _ {t - 1}\right), \\ \left\{\mathbf {T} _ {t} ^ {\mathbb {R}} \right\} \gets \operatorname {L o c a l i z e} \left(\mathbf {M} _ {t}, \mathbf {U} _ {t}, \mathbf {D} _ {t}, \mathcal {M} _ {t - 1}\right), \\ \left\{\mathcal {M} _ {t} \right\} \leftarrow \operatorname {C S r c} \left(\mathbf {M} _ {t}, \mathbf {U} _ {t}, \mathbf {D} _ {t}, \mathbf {T} _ {t}, \mathcal {M} _ {t - 1}\right). \tag {1} \\ \end{array} +$$ + +As illustrated in Fig. 2, we use a multi-task onboard model RTMapModel $(\cdot)$ with hybrid queries (Sec. 3.2) and + +an existence-aware matching scheme (Sec. 3.3) to detect three classes of map elements: matched $\mathbf{M}_t^{\mathbb{M}}$ , outdated $\mathbf{M}_t^{\mathbb{F}}$ , and newly observed $\mathbf{M}_t^{\mathbb{N}}$ . Therefore, we can separately infer possible map change events $\Delta_t \triangleq \{\mathbf{M}_t^{\mathbb{F}}, \mathbf{M}_t^{\mathbb{N}}\}$ besides matched elements $\mathbf{M}_t^{\mathbb{M}}$ to reduce their impact on localization and update the crowdsourced map afterward. We design our loss function (Sec. 3.4) for multiple additional tasks, jointly inferring (1) the per-vertex probabilistic density $\mathbf{U}_t$ of these map elements, (2) point-wise correspondences $\mathbf{D}_t$ between re-observed and prior-map elements $\langle \mathbf{M}_t^{\mathbb{M}}, \mathcal{M}_{t-1} \rangle$ , and (3) an end-to-end pose w.r.t. the prior-map $\mathbf{T}_t^{\mathbb{E}} \in \mathbb{S}\mathbb{E}^3$ . For explicitly addressing the 6-DOF localization if required, we also implement a state optimizer Localize(\cdot) leveraging deducted correspondences $\mathbf{D}_t$ for an optimization-based pose solution $\mathbf{T}_t^{\mathbb{R}} \in \mathbb{S}\mathbb{E}^3$ as an option (Sec. 3.5). As $\mathbf{T}_t^{\mathbb{E}}$ and $\mathbf{T}_t^{\mathbb{R}}$ share the same purpose, we compare their performance (Tab. 4) and retain the estimator of $\mathbf{T}_t^{\mathbb{R}}$ for tightly coupling inertial measurements if necessary. Finally, CSrc(\cdot) operates on-the-cloud to fuse multi-traversal inferences and update the crowdsourced map (Sec. 3.5). We use GPS measurements to provide a meters-level initial pose $\mathbf{T}_t^0$ for fetching a truncated local crowdsourced HD map $\mathcal{M}_{t-1}$ for these onboard operations. + +# 3.2. Hybrid Queries + +To tackle the challenges of localization, mapping, and change detection concurrently, we propose hybrid queries enabling the network to process various query types and obtain matched, outdated, and newly observed elements to address all three tasks effectively. In this context, the decoder initially receives two types of queries $\mathbf{Q}_{\mathrm{prior}}$ and $\mathbf{Q}_{\mathrm{new}}$ from prior-map embedding and sensor encoding, respectively. After training with query-based matching (Sec. 3.3), we further differentiate $\mathbf{Q}_{\mathrm{prior}} \rightarrow \{\mathbf{Q}_{\mathrm{map}}, \mathbf{Q}_{\mathrm{fake}}\}$ , with the first part $\mathbf{Q}_{\mathrm{map}}$ telling reliable correspondences for localization and the latter part $\mathbf{Q}_{\mathrm{fake}}$ specifying outdated elements. + +Inspired by MapTR [24] and MapEX [33], we use unified representation to encode each $\mathbf{Q}_{\mathrm{prior}}$ as an instance represented by fixed points, where the first two dimensions of each point are encoded as XOY coordinates, and the remaining N dimensions represent category information using one-hot encoding. Due to the characteristics of the transformer, the number of queries is fixed. Therefore, the remaining queries are designated as $\mathbf{Q}_{\mathrm{new}}$ padded with zeros. Finally, after passing through a linear layer, these queries are combined with the hierarchical query embeddings $\mathbf{Q}_{\mathrm{hie}}$ [24] to obtain the hybrid queries $\mathbf{Q}_{\mathrm{hybrid}}$ formulated as follows: + +$$ +\mathbf {Q} _ {\text {h y b r i d}} = \left\{\mathbf {Q} _ {\text {m a p}}, \mathbf {Q} _ {\text {f a k e}}, \mathbf {Q} _ {\text {n e w}} \right\} + \mathbf {Q} _ {\text {h i e}}. \tag {2} +$$ + +![](images/e166403fc9a78236535669e19340b9cc1a88dc671393f2735875ad4223567c0e.jpg) +Prior-Map + +![](images/c09fdf0d634ab071461d136d7fbc41834f92cadf5676bbdc9c75d7e78d0a89bc.jpg) +Dec. Layer 2 + +![](images/b67dac1ea77eac5ee6deea63bc933bba7b466d55754975bcf7b18a60f95eb41c.jpg) +Dec. Layer 4 + +![](images/b0b93595c3436d1623b9ded110c735f06f3ec3133e678e8e27ad6858fea164d2.jpg) +Dec. Layer 6 + +![](images/4b4f2d93e9e49b2b693e7a568da985ef0e51813d6cb2298dac1b990254cbc136.jpg) +GT + +![](images/5de6ef8c70bf0ea43cd01ebe6a8907959c8dfd5e58ed599faf643d4f3b34f257.jpg) +Figure 3. Trend of queries in different decoder layers. $\mathbf{Q}_{\mathrm{map}}$ , $\mathbf{Q}_{\mathrm{fake}}$ , $\mathbf{Q}_{\mathrm{new}}$ are colorized in blue, red, and green, respectively. The ellipsoid reflects their uncertainty. Hence, outdated map elements are gradually filtered from matched map elements. + +![](images/26e72c846f3e7daa7fb09ddcbcb67e0beb7c593e5d4d852d6b6b38c483d596b5.jpg) + +![](images/83050705d6e3b8315e7be8ffe9a8e5cc5c76fba1f73277cd1444c9b79d76133e.jpg) + +![](images/77960b7ef73073e0cd60292587c3616cac70bbd12a491965b723a3da768cf6c7.jpg) + +![](images/152a9207d2385008e3fa1cd46733d54d45df64894e68cb05030dc2686ae3fc4e.jpg) + +# 3.3. Existence-aware Matching + +Matching in training. In contrast to MapEX [33], which leverages pre-attribute from ground truth for training, we need to tell $\mathbf{Q}_{\mathrm{fake}}$ apart from $\mathbf{Q}_{\mathrm{map}}$ for change detection and improving multi-task performances. Therefore, with the help of ground truth map-changing events, we choose to implement only a pre-attribute of $\mathbf{Q}_{\mathrm{map}}$ predictions to their corresponding existing map elements. Besides, the remaining map elements are matched with $\mathbf{Q}_{\mathrm{new}}$ predictions using conventional Hungarian matching. + +As an observable consequence of the training process (Fig. 3), different types of queries exhibit distinct behaviors under this matching strategy. Due to pose perturbations, Map queries $\mathbf{Q}_{\mathrm{map}}$ are projected near their actual locations. After that, their corresponding reference points gradually move to the correct positions. In contrast, since fake queries $\mathbf{Q}_{\mathrm{fake}}$ do not exist in the prior-map, they lead to unstable behavior in their reference points. The rest of the new queries $\mathbf{Q}_{\mathrm{new}}$ detect newly occurring elements from the current traverse. + +Matching in inference. During the inference stage, without ground truth annotations, $\mathbf{Q}_{\mathrm{map}}$ and $\mathbf{Q}_{\mathrm{fake}}$ are mixed within the prior queries $\mathbf{Q}_{\mathrm{prior}}$ . Since $\mathbf{Q}_{\mathrm{fake}}$ were not pre-attributed during the training phase, the category confidence associated with $\mathbf{Q}_{\mathrm{fake}}$ will be significantly lower than those for $\mathbf{Q}_{\mathrm{map}}$ , providing strong evidence to differentiate between the two types of queries. Therefore, we believe that using the confidence score of the predictions corresponding to $\mathbf{Q}_{\mathrm{prior}}$ can effectively distinguish between them. + +# 3.4. Loss Function for Map Elements and Poses + +Composite Geometric Loss for Element Vertices. Following Gu et al. [13], we choose univariate Laplace distri + +butions to model map elements, allowing us to additionally output the uncertainty of the location $\mathbf{M}_t$ of map element vertices. For inferencing such additional information, we augment a negative log-likelihood (NLL) loss $\mathbf{L}_{\mathrm{nll}}$ for each vertex, as: + +$$ +\begin{array}{l} \mathbf {L} _ {\mathrm {p t s}} = \lambda_ {1} \cdot \mathbf {L} _ {\mathrm {n l l}} + \lambda_ {2} \cdot \mathbf {L} _ {\mathrm {m h t}}, \\ \mathbf {L} _ {\mathrm {n l l}} = \sum_ {v = 1} ^ {\mathrm {V}} \sum_ {k = 1} ^ {2} \left(\log \left(2 \sigma_ {v} ^ {k}\right) + \frac {\left| \mathbf {m} _ {v} ^ {k} - \mu_ {v} ^ {k} \right|}{\sigma_ {v} ^ {k}}\right), \tag {3} \\ \end{array} +$$ + +We keep their Z-value representation unchanged due to insufficient observability. The parameters $\mu_v^k,\sigma_v^k\in \mathbb{R}$ denote the location and scale parameters of the $k^{\mathrm{th}}$ dimensional Laplace distributions for a vertex $\mathbf{m}_v$ , and we use a fixed total size $\mathrm{V} = 20$ for each map element. We remain the original Manhattan distance loss $\mathbf{L}_{\mathrm{mht}}$ [25] with hyperparameters $\lambda_{1}$ and $\lambda_{2}$ to balance the convergence ability and the element accuracy. + +Pose Auxiliary Loss. Given a coarse initial pose $\mathbf{T}^0$ , the $\mathbf{Q}_{\mathrm{map}}$ of the input prior-map may be incorrectly positioned. Hence, the pose auxiliary loss $\mathbf{L}_{\mathrm{pose}}$ is designed to provide additional supervision for the transformer decoder, helping it learn to correct such misalignment more effectively. Our end-to-end design utilizes the output features from the decoder to precisely align perceived map elements with the prior-map. Regarding map alignment, information can be retrieved with in the map queries $\mathbf{Q}_{\mathrm{map}}$ , which are then added to the output of each decoder layer. Subsequently, we utilize a shared MLP and a max-pooling layer to obtain the predicted delta pose $\partial \hat{\mathbf{T}}_{\mathrm{aux}}$ . Finally, the predicted delta pose is supervised by ground truth delta pose $\partial \mathbf{T}$ with a smooth L1 Loss. + +# 3.5. MAP for Localization and Crowdsourcing + +Optimization-based Localization. For the Localize( $\cdot$ ) operation, we can explicitly remove map-changing events $\Delta_t \triangleq \{\mathbf{M}_t^\mathbb{F}, \mathbf{M}_t^\mathbb{N}\}$ from the matched associations $\mathbf{D}_t: \langle \mathbf{M}_t^\mathbb{M}, \mathcal{M}_{t-1} \rangle$ for outlier rejection, and employ the additional probabilistic density of vertices $\mathbf{U}_t$ . Specifically, we launch the following maximum a posteriori (MAP) state estimation for solving the vehicle pose: + +$$ +\begin{array}{l} \min _ {\mathbf {T} ^ {\mathbb {R}}} \mathbf {E} ^ {1} (\mathbf {D} _ {t}), \\ \mathbf {f} ^ {1} \propto \exp \left(- \frac {1}{2} \left\| \mathbf {T} ^ {\mathbb {R}} \cdot \mathbf {m} _ {t} ^ {i} - m _ {t - 1} ^ {i} \right\| _ {\mathbf {g} _ {t} ^ {i}} ^ {2}\right), \tag {4} \\ \end{array} +$$ + +where $\mathbf{E}^1 (\mathbf{D}_t) = \sum_{\mathbf{D}_t} - \log (\mathbf{f}^1)$ is the sum of the negative log-likelihood of point-to-point residuals $\mathbf{f}^1$ , and we use the squared Mahalanobis distance whose covariance is assigned as the Gaussian mixture $\mathbf{g}_t^i\triangleq \mathbf{u}_t^i\oplus u_{t - 1}^i$ of both inferred and crowdsourced probabilistic densities [5]. Each pair of associations denoted as $d_t^i:\langle \mathbf{m}_t^i,m_{t - 1}^i\rangle \in \mathbf{D}_t$ constructs a residual between corresponded vertices across + +two maps. One can combine more residuals like IMU pre-integration [12] or odometry [31, 45] for a tightly coupled system, and in this paper, we mainly test the difference between the single residual form $\mathbf{E}^1$ and the end-to-end regressed pose $\mathbf{T}^{\mathbb{E}}$ . Specifically, we use the Levenberg-Marquardt algorithm to minimize $\mathbf{E}^1(\mathbf{D}_t)$ . + +Probabilistic-aware Crowdsourcing. For the $\mathrm{CSrc}(\cdot)$ operated on-the-cloud, we gather multi-traversal and multiframe observations $\langle \mathbf{M}_t^j,\mathbf{U}_t^j,\mathbf{D}_t^j\rangle$ with their relative pose w.r.t. a base frame as $\mathbf{T}_t^j$ , to fuse the crowdsourced map again in a MAP manner. For each tracked map vertex with multiple observations, we use the union-find algorithm to construct the following position solver for its latest position $m_t\in \mathcal{M}_t$ : + +$$ +\min _ {m _ {t}} \sum_ {\hat {\mathbf {m}} _ {t} ^ {j} \in \mathbf {M}} \frac {1}{2} \| m _ {t} - \mathbf {T} _ {t} ^ {j} \cdot \hat {\mathbf {m}} _ {t} ^ {j} \| _ {\mathbf {u} _ {t} ^ {j}} ^ {2} + \frac {1}{2} \| m _ {t} - m _ {t - 1} \| _ {u _ {t - 1}} ^ {2}, (5) +$$ + +where $\hat{\mathbf{m}}_t^j$ represents an associated observation $\mathbf{M}_t^j$ , and the union $\mathbf{M}$ gathers all matched observations, and we again use the Gaussian mixture model to continuously refine the probabilistic density $u_t$ for each vertex. Meanwhile, we perform Hungarian voting to determine the predecessor and successor vertices of each vertex, to preserve the topology of HD map elements. For the first traverse, the second term leveraging the previous crowdsourced version is omitted for an initial crowdsourcing. + +# 4. Implementation, Results, and Evaluation + +# 4.1. Implementation details + +Training Sample Generation. To better align with real-world scenarios and enhance network robustness, we first apply random perturbations to the ground truth pose $\mathbf{T}^{\mathrm{gt}}$ to generate a noisy initial pose estimate $\mathbf{T}^{\mathrm{init}}$ . Using $\mathbf{T}^{\mathrm{init}}$ , we then crop the ground truth map. Occasionally, when the local map obtained from the noisy pose is the same size as the perceptual range, some map elements may fall outside the perceptible area. To ensure all ground truth map elements remain accessible, we pad the perceptual range as necessary. Next, we randomly remove certain elements from the prior-map and generate synthetic fake map elements. We do not perturb the ground truth map elements to simulate specific changes (such as shifting lane markings), as such modifications could be interpreted as deleting the original map elements and adding new ones. + +Implementation and Training Details. We choose ResNet50 [14] as the image backbone. Our RTMap model is trained using 8 NVIDIA GeForce RTX 3090 GPUs with a total batch size of 32 for 36 epochs. We utilize the AdamW optimizer with a learning rate of $6 \times 10^{-4}$ . To ensure fair comparisons, we adopt the same training settings for MapTRv2 [25] and ensure its convergence. For $\mathbf{L}_{\mathrm{pts}}$ in Eq. 3, we + +set $\lambda_{1} = 0.03$ and $\lambda_{2} = 5.0$ for balancing their influence. For $\mathbf{L}_{\mathrm{pose}}$ , the translation weight is set to 0.04, corresponding to the scale in radians used when predicting the heading angle. We set the longitudinal range to $[-36,36]$ and the lateral range to $[-18,18]$ meters for online HD mapping. In contrast to prior-aided method such as HRMapNet [43], we do not use crowdsourced prior-maps for training. Instead, we use artificially perturbed prior-maps to ensure we have incorporated sufficient map-changing events regarding the current scale and maturity of publicly available datasets. + +# 4.2. Experiment Setup + +Datasets. We evaluate the performance of the proposed network using two datasets: TbV [18] and nuScenes [3]. The TbV dataset [18] provides over 200 scenarios involving real-world changes. These map-changing events are primarily on lane topology, road boundaries, and pedestrian crossings. It is specifically designed to detect discrepancies between sensor data and HD maps caused by these changes, with the val set containing scenarios that feature actual map alterations. For training, the dataset also incorporates synthetically modified ground truth maps. The nuScenes dataset [3] is a large-scale dataset for autonomous driving, consisting of 1,000 driving scenes in urban environments, each sampled at $2\mathrm{Hz}$ . It includes sensor data such as RGB images, LiDAR point clouds, inertial measurements, and labeled HD maps for evaluation. + +Tasks and Metrics. Regarding the insufficiency of multitraversal scanning and map-changing events, we tested the main task - crowdsourcing - as well as two sub-tasks, change detection and localization, on the TbV dataset. Meanwhile, we use the nuScenes dataset to test the performance of our map-based localization w.r.t. its labeled HD map. + +Unlike conventional online HD mapping tasks, our crowdsourcing task accounts for localization errors. Therefore, we introduce localization perturbations on top of the evaluation method used in online HD mapping tasks. Specifically, for a more realistic simulation of localization noise, we sample perturbations from Gaussian distributions: a lateral perturbation of $\mathcal{N}(0,0.75^2)$ , a longitudinal perturbation of $\mathcal{N}(0,1.5^2)$ in meters, and a yaw perturbation of $\mathcal{N}(0,0.85^2)$ in degrees. To obtain ground truth for evaluation, we augment the map data by aligning the previous validation set with sensor data for ensuring 3D consistency. Specifically, we use average precision to evaluate the quality of the generated map and Chamfer distance to determine the alignment between the ground truth and predicted maps. We compute the AP at the following thresholds $\{0.5,1.0,1.5\}$ in meters keeping the same with experiments in other methods [25], and then calculate the mean to obtain the final mean average precision (mAP). + +For the change detection task, we follow the evaluation + +Table 1. Quantitative comparison of different online mapping approaches and strategies on TbV [18]. We additionally perform an ablation study on whether or not to use the vertex-level probabilistic density for crowdsourcing. + +
MethodCyclemAP(%) ↑
Ped.Div.Bou.Avg.
Straight
MapTRv2Ave.31.742.037.337.1
HRMapNetAve.34.243.739.839.2
MapTrackerAve.35.744.639.639.9
Ours (w/o U)228.660.531.740.2
Ours232.768.635.545.6
Ours (w/o U)335.774.442.050.7
Ours340.984.347.657.6
Turning
MapTRv2Ave.28.231.618.326.0
HRMapNetAve.31.032.819.627.8
MapTrackerAve.31.433.219.328.0
Ours (w/o U)230.265.530.141.9
Ours233.171.032.345.5
Ours (w/o U)335.971.232.646.5
Ours342.385.238.855.4
+ +implementation of the officially released code in TbV [18], to treat the change detection task as a binary classification problem (Acc), with classes indicating whether a change has occurred Acc or not Acc, and the mean accuracy (mAcc) is also calculated to assess overall performance. + +For the localization task, we measure the mean and $90^{\mathrm{th}}$ absolute error of ego-poses relative to the ground truth HD map in three main dimensions: lateral, longitudinal, and yaw. + +# 4.3. Performance on Crowdsourcing + +We cluster multi-traversal spatially overlapped driving clips in TbV for performing crowdsourcing, group 15 found clips into two scenarios – straight (6) and turning (9), and listed our results in comparison with state-of-the-art online HD mapping methods, as shown in Tab. 1. Due to the introduction of localization noise, the accuracy of existing online HD mapping methods is significantly reduced. Our approach is essentially the same as existing methods in the first cycle, when neither localization nor change detection is required. In later cycles, by utilizing the prior-map, our method gradually decreases localization errors and improves mapping accuracy. We also compare to a direct version, which does not infer and use probabilistic densities $\mathbf{U}$ during crowdsourcing, highlighting the effectiveness of leveraging such uncertainty in mapping tasks. We also refer readers to Fig. 5, which presents the trend of quality im + +provement by solving a precise trajectory and experiencing multi-traversals. + +# 4.4. Performance on Change Detection + +To evaluate the effectiveness of the proposed RTMap under dynamic driving conditions, we performed change detection tests on the TbV dataset, with modifications including 'insert crosswalk' and 'delete crosswalk'. + +The results of our experiments, presented in Tab. 2, indicate a clear distinction in performance across the different scenarios, with our model demonstrating a particularly high sensitivity to the 'changed' category. This suggests that RTMap is well-suited to recalling possible map-changing events. Practically, a higher recall enables crowd-mining on these dubious events and ensures safety. Nevertheless, our method achieves higher overall accuracy than the previous method. + +From the results, we can conclude that the RTMap's performance is robust in dynamic contexts. This capability not only enhances the model's effectiveness but also indicates its potential applicability in real-world scenarios, where rapid adaptation to environmental changes is essential. + +Table 2. Quantitative comparison of change detection between ours and the original approach proposed for TbV [18]. + +
MethodAccc(%) ↑Accr(%) ↑mAcc(%) ↑
TbV [18]40.068.254.1
RTMap48.966.057.4
+ +# 4.5. Performance on Localization + +Our end-to-end framework also supports localizing the vehicle w.r.t. the prior-map. In the first experiment, we perform an ablation study on whether or not to incorporate the change detection task - differentiating $\mathbf{Q}_{\mathrm{map}}$ from $\mathbf{Q}_{\mathrm{prior}}$ , and show the results in Tab. 3 and Fig. 4-(b), which proves the advantage of tightly coupling the change detection task to reject mismatched map elements for localization. + +Table 3. An ablation on whether or not using hybrid queries to distinguish $\mathbf{Q}_{\mathrm{fake}}$ from $\mathbf{Q}_{\mathrm{map}}$ on TbV [18]. + +
MethodLat (m)↓Lon (m)↓Yaw (°)↓
Mean90thMean90thMean90th
RTMap (Qprior)0.1630.3180.6861.6160.3320.766
RTMap (Qmap)0.1250.2560.6331.5300.3170.713
+ +We also evaluate localization on the nuScenes dataset as listed in Tab. 4, which includes 100 different scenarios + +![](images/b8e932e04f1b4247a89e0cf02762111ad5b825e86db7699b01f3819a9dbf696f.jpg) +Figure 4. Qualitative comparison of the localization performance between two configurations on TbV [18], we show deviations of our estimated trajectory w.r.t. the ground truth, to illustrate the effectiveness of leveraging state optimization $(\mathbf{T}^{\mathbb{R}})$ and map changing events $(\mathbf{Q}_{\mathrm{map}})$ for further improving the accuracy. + +from the validation set. We find the following conclusions through experimental results: (1) The explicit optimization-based pose estimation $\mathbf{T}^{\mathbb{R}}$ still outperforms the end-to-end manner $\mathbf{T}^{\mathbb{E}}$ , (2) the proposed loss functions in Sec. 3.4, including the localization loss $\mathbf{L}_{\mathrm{pose}}$ and the probabilistic density aware vertex regression loss $\mathbf{L}_{\mathrm{pts}}$ during training, can effectively contribute to the localization task. + +In the current framework, we no longer require a standalone map-based localization task [15, 44] to serve downstream planning and control modules. Instead, we concentrate on tightly coupling change detection, localization, and the prior-map together, to benefit the completeness and robustness of the local map for serving downstream planning and control modules. + +Table 4. Ablations of accuracy between end-to-end pose estimation $\mathbf{T}^{\mathbb{E}}$ , and optimization-based pose estimation $\mathbf{T}^{\mathbb{R}}$ under different settings on nuScenes [3]. + +
MethodLat (m)↓Lon (m)↓Yaw (°)↓
Mean90thMean90thMean90th
RTMap (TΕ)0.1420.3060.5891.4470.5211.091
RTMap (TΦ)0.1210.2570.5861.4290.3680.799
TΦ: Base0.1220.2540.6181.5100.3900.840
TΦ: +Lpose0.1220.2610.5901.4360.3710.800
TΦ: +Lpts0.1180.2420.6091.5070.3760.841
+ +# 5. Conclusion + +We presented RTMap, an end-to-end framework for real-time crowdsourcing mapping. Integrating prior-aided localization, change detection, and multi-agent crowdsourcing. Its functionality includes: (1) a multi-task architecture deployed onboard for concurrently performing localization, map change detection, vectorized HD mapping, reaching real-time performance on robust HD map servic + +![](images/972874aa086c06e5eb879eb4384acdb944a4c410f58007f57b99fdf485438435.jpg) +Figure 5. Qualitative comparison trending the quality improvement through solving a precise pose for alignment, and a multi-traversal probabilistic-aware crowdsourced HD mapping. We zoom-in several map regions for further comparisons. + +ing, (2) uncertainty-aware modeling, which probabilistically represents element uncertainties to improve localization accuracy and map quality, and (3) crowdsourced mapping, enabling continuous global map refinement via collaborative multi-agent multi-traversal data. Experiments on TbV and nuScenes datasets have demonstrated RTMap's performance for all these tasks, with dynamic updates ensuring temporal consistency. By bridging onboard perception and asynchronous cloud computing, RTMap advances safe, adaptive autonomous driving solutions to maintain and leverage a self-evolutional memory. + +Future work. Our future directions include incorporating more sensors (e.g., LiDAR) to form a multi-modal sensor fusion and uploading auxiliary but highly compact data for better probabilistic modeling on the cloud. Expanding validation to unstructured terrains and more complex urban environments will better establish the framework's gener + +alizability. Besides, crowdsourcing enables a maximum-a-posteriori policy to vote out temporary results gradually in our current version. In the future, our backbone supports adding occlusion detection head as an auxiliary task to distinguish between permanent structural changes and transient obstructions. + +# Acknowledgements + +We thank the reviewers for the valuable discussions. We respect our past and present colleagues for their unwavering dedication to pursuing innovative light-weight map solutions, whose contributions have profoundly initiated this work. This research was supported by the Zhejiang Provincial Natural Science Foundation of China under Grant No. LD24F030001. + +# References + +[1] Samuel M Bateman, Ning Xu, H Charles Zhao, Yael Ben Shalom, Vince Gong, Greg Long, and Will Maddern. Exploring real world map change generalization of prior-informed hd map prediction models. In CVPR, pages 4568-4578, 2024. 2 +[2] Sean L. Bowman, Nikolay Atanasov, Kostas Daniilidis, and George J. Pappas. Probabilistic data association for semantic SLAM. In ICRA, pages 1722-1729, 2017. 3 +[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, pages 11621-11631, 2020. 2, 6, 7 +[4] Andrea Boscolo Camiletto, Alfredo Bochicchio, Alexander Liniger, Dengxin Dai, and Abel Gawel. U-bev: Height-aware bird's-eye-view segmentation and neural map-based relocalization. In IROS, pages 5597-5604, 2024. 2 +[5] Yan-Pei Cao, Leif Kobbelt, and Shi-Min Hu. Real-time high-accuracy three-dimensional reconstruction with consumer rgb-d cameras. ACM TOG, 37(5):171:1-171:16, 2018. 3, 5 +[6] Jiacheng Chen, Yuefan Wu, Jiaqi Tan, Hang Ma, and Yasutaka Furukawa. Maptracker: Tracking with strided memory fusion for consistent vector hd mapping. In ECCV, pages 90-107, 2024. 2 +[7] Jiagang Chen, Liangliang Pan, Shunping Ji, Ji Zhao, and Zichao Zhang. Online temporal fusion for vectorized map construction in mapless autonomous driving. IEEE Robotics and Automation Letters, 10(4):3948-3955, 2025. 2 +[8] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv:2206.04584, 2022. 2 +[9] Wentao Cheng, Sheng Yang, Maomin Zhou, Ziyuan Liu, Yiming Chen, and Mingyang Li. Road mapping and localization using sparse semantic visual features. IEEE Robotics and Automation Letters, 6(4):8118-8125, 2021. 3 +[10] Zheng-Jun Du, Shi-Sheng Huang, Tai-Jiang Mu, Qunhe Zhao, Ralph R Martin, and Kun Xu. Accurate dynamic slam using crf-based long-term consistency. IEEE TVCG, 28(4): 1745-1757, 2020. 3 +[11] Gamal Elghazaly, Raphaël Frank, Scott Harvey, and Stefan Safko. High-definition maps: Comprehensive survey, challenges and future perspectives. IEEE Open Journal of Intelligent Transportation Systems, 4:527-550, 2023. 2 +[12] Christian Forster, Luca Carlone, Frank Dellaert, and Davide Scaramuzza. On-manifold preintegration for real-time visual-inertial odometry. IEEE Transactions on Robotics, 33 (1):1-21, 2016. 5 +[13] Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, and Boris Ivanovic. Producing and leveraging online map uncertainty in trajectory prediction. In CVPR, pages 14521-14530, 2024. 3, 4 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5 + +[15] Yuzhe He, Shuang Liang, Xiaofei Rui, Chengying Cai, and Guowei Wan. Egovm: Achieving precise ego-localization using lightweight vectorized maps. In IROS, pages 12248-12255, 2024. 2, 3, 7 +[16] Shi-Sheng Huang, Ze-Yu Ma, Tai-Jiang Mu, Hongbo Fu, and Shi-Min Hu. Lidar-monocular visual odometry using point and line features. In ICRA, pages 1091-1097, 2020. 3 +[17] Zhou Jiang, Zhenxin Zhu, Pengfei Li, Huan-ang Gao, Tianyuan Yuan, Yongliang Shi, Hang Zhao, and Hao Zhao. P-MapNet: Far-seeing map generator enhanced by both sdmap and hdmap priors. In ICLR, pages 1–12, 2024. 2 +[18] John Lambert and James Hays. Trust, but verify: Cross-modality fusion for hd map change detection. In NeurIPS, pages 1-14, 2021. 2, 6, 7 +[19] Dongjae Lee, Minwoo Jung, Wooseong Yang, and Ayoung Kim. Lidar odometry survey: recent advancements and remaining challenges. Intelligent Service Robotics, 17(2):95-118, 2024. 3 +[20] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework. In ICRA, pages 4628-4634, 2022. 2 +[21] Siyu Li, Jiacheng Lin, Hao Shi, Jiaming Zhang, Song Wang, You Yao, Zhiyong Li, and Kailun Yang. DTCLMapper: Dual temporal consistent learning for vectorized hd map construction. IEEE Transactions on Intelligent Transportation Systems, 25(12):21672-21686, 2024. 2 +[22] Tianyu Li, Peijin Jia, Bangjun Wang, Li Chen, Kun Jiang, Junchi Yan, and Hongyang Li. LaneSegNet: Map learning with lane segment perception for autonomous driving. In ICLR, 2024. 2 +[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. BEVFormer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, pages 1-18, 2022. 2 +[24] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. MapTR: Structured modeling and learning for online vectorized hd map construction. In ICLR, pages 1-18, 2023. 1, 2, 4 +[25] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. MapTRv2: An end-to-end framework for online vectorized hd map construction. International Journal of Computer Vision, 133(3):1352-1374, 2025. 2, 5, 6 +[26] Xiaolu Liu, Song Wang, Wentong Li, Ruizi Yang, Junbo Chen, and Jianke Zhu. MGMap: Mask-guided learning for online vectorized hd map construction. In CVPR, pages 14812-14821, 2024. 1, 2 +[27] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. VectorMapNet: End-to-end vectorized hd map learning. In ICML, pages 22352-22369, 2023. 2 +[28] Katie Z Luo, Xinshuo Weng, Yan Wang, Shuang Wu, Jie Li, Kilian Q Weinberger, Yue Wang, and Marco Pavone. Aug-mentation lane perception and topology understanding with standard definition navigation maps. In ICRA, pages 4029-4035, 2024. 2 +[29] Jinyu Miao, Kun Jiang, Tuopu Wen, Yunlong Wang, Peijin Jia, Benny Wijaya, Xuhe Zhao, Qian Cheng, Zhongyang + +Xiao, Jin Huang, et al. A survey on monocular relocalization: From the perspective of scene map representation. IEEE TIV, pages 1-33, 2024. 3 +[30] Nan Peng, Xun Zhou, Mingming Wang, Xiaojun Yang, Songming Chen, and Guisong Chen. Prevpredmap: Exploring temporal modeling with previous predictions for online vectorized hd map construction. In WACV, pages 8134-8143, 2025. 2 +[31] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 5 +[32] Tong Qin, Changze Li, Haoyang Ye, Shaowei Wan, Minzhen Li, Hongwei Liu, and Ming Yang. Crowd-sourced nerf: Collecting data from production vehicles for 3d street view reconstruction. IEEE Transactions on Intelligent Transportation Systems, 25(11):16145-16156, 2024. 2 +[33] Rémy Sun, Li Yang, Diane Lingrand, and Frédéric Precioso. Mind the map! accounting for existing map information when estimating online hdmaps from sensor data. In WACV, pages 1-11, 2025. 2, 4 +[34] Kun Tang, Xu Cao, Zhipeng Cao, Tong Zhou, Erlong Li, Ao Liu, Shengtao Zou, Chang Liu, Shuqi Mei, Elena Sizikova, et al. THMA: Tencent hd map ai system for creating hd map annotations. In AAAI, pages 15585-15593, 2023. 2 +[35] Benny Wijaya, Kun Jiang, Mengmeng Yang, Tuopu Wen, Yunlong Wang, Xuewei Tang, Zheng Fu, Taohua Zhou, and Dange Yang. High definition map mapping and update: A general overview and future directions. arXiv:2409.09726, 2024.1,2 +[36] Lena Wild, Ludvig Ericson, Rafael Valencia, and Patric Jensfelt. Exelmap: Explainable element-based hd-map change detection and update. arXiv:2409.10178, 2024. 3 +[37] Chenming Wu, Jiadai Sun, Zhelun Shen, and Liangjun Zhang. MapNeRF: Incorporating map priors into neural radiance fields for driving view simulation. In IROS, pages 7082-7088, 2023. 2 +[38] Deguo Xia, Weiming Zhang, Xiyan Liu, Wei Zhang, Chenting Gong, Xiao Tan, Jizhou Huang, Mengmeng Yang, and Diange Yang. LDMapNet-U: An end-to-end system for city-scale lane-level map updating. In ACM SIGKDD, page 2693-2702, 2025. 3 +[39] Xuan Xiong, Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Neural map prior for autonomous driving. In CVPR, pages 17535-17544, 2023. 2 +[40] Sheng Yang, Xiaoling Zhu, Xing Nian, Lu Feng, Xiaozhi Qu, and Teng Ma. A robust pose graph approach for city scale LiDAR mapping. In IROS, pages 1175-1182, 2018. 2 +[41] Tianyuan Yuan, Yicheng Liu, Yue Wang, Yilun Wang, and Hang Zhao. StreamMapNet: Streaming mapping network for vectorized online hd map construction. In WACV, pages 7341-7350, 2024. 2 +[42] Shuang Zeng, Xinyuan Chang, Xinran Liu, Zheng Pan, and Xing Wei. Driving with prior maps: Unified vector prior encoding for autonomous vehicle mapping. arXiv preprint arXiv:2409.05352, 2024. 2 +[43] Xiaoyu Zhang, Guangwei Liu, Zihao Liu, Ningyi Xu, Yunhui Liu, and Ji Zhao. Enhancing vectorized map perception + +with historical rasterized maps. In European Conference on Computer Vision, pages 422-439. Springer, 2024. 6 +[44] Zhihuang Zhang, Meng Xu, Wenqiang Zhou, Tao Peng, Liang Li, and Stefan Poslad. Bev-locator: An end-to-end visual semantic localization network using multi-view images. Science China Information Sciences, 68(2):122106, 2025. 3, 7 +[45] Bin Zhao, Yunzhou Zhang, Junjie Huang, Xichen Zhang, Zeyu Long, and Yulong Li. L-VIWO: Visual-inertial-wheel odometry based on lane lines. In ICRA, pages 18079-18085, 2024. 3, 5 +[46] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, Hongbo Fu, and Song-Hai Zhang. GP-Recon: Online monocular neural 3d reconstruction with geometric prior. IEEE TVCG, 2024. 3 \ No newline at end of file diff --git a/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/images.zip b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2ca15583d96719e7b392914bf1de83296120ae28 --- /dev/null +++ b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55f564ebbdad21844de93918e02a73ddc94efe3271ae866d961f4fbc3d54774a +size 610300 diff --git a/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/layout.json b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0579362d1ec3cdbc0a1815cbf69fb9040cdc98f7 --- /dev/null +++ b/ICCV/2025/RTMap_ Real-Time Recursive Mapping with Change Detection and Localization/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59c059eb1dbde497c7e3d1be27aa06f048771b54ca13d7925643f1bb9ec0d3fe +size 406709 diff --git a/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/8d30b116-8006-4dfe-9551-2c5c1f0a9398_content_list.json b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/8d30b116-8006-4dfe-9551-2c5c1f0a9398_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c6e402f0473018f755c593819c2c8ab0433c9bc6 --- /dev/null +++ b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/8d30b116-8006-4dfe-9551-2c5c1f0a9398_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f46ac4585b51b8511cf3b7df9a93ff3c343f898cfc2369321a065726a6109621 +size 95426 diff --git a/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/8d30b116-8006-4dfe-9551-2c5c1f0a9398_model.json b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/8d30b116-8006-4dfe-9551-2c5c1f0a9398_model.json new file mode 100644 index 0000000000000000000000000000000000000000..710f2820ccc6d2afe1dd72b6bb381e21aa10fa91 --- /dev/null +++ b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/8d30b116-8006-4dfe-9551-2c5c1f0a9398_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c218d68b30ea5c05f710187396343e9aaafe26845ff7c7faf84223144e494e8 +size 118722 diff --git a/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/8d30b116-8006-4dfe-9551-2c5c1f0a9398_origin.pdf b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/8d30b116-8006-4dfe-9551-2c5c1f0a9398_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5c6a0d3109ca1ce0629b1119fd948c1db761c58 --- /dev/null +++ b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/8d30b116-8006-4dfe-9551-2c5c1f0a9398_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b3d7e7d2c083aefbe5128d33bd635e0b7e06e658c6c9b927e41299ab499e7b0 +size 9719601 diff --git a/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/full.md b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dd1423d95861611880433a906cfb0af31c8558f4 --- /dev/null +++ b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/full.md @@ -0,0 +1,377 @@ +# RadGPT: Constructing 3D Image-Text Tumor Datasets + +Pedro R. A. S. Bassi $^{1,2,3}$ Mehmet Can Yavuz $^{4}$ Ibrahim Ethem Hamamci $^{6,7}$ Sezgin Er $^{5}$ + +Xiaoxi Chen8 Wenxuan Li1 Bjoern Menze6 Sergio Decherchi3 + +Andrea Cavalli $^{2,3,9}$ Kang Wang $^{4}$ Yang Yang $^{4}$ Alan Yuille $^{1}$ Zongwei Zhou $^{1,*}$ + +$^{1}$ Johns Hopkins University $^{2}$ University of Bologna $^{3}$ Italian Institute of Technology + +$^{4}$ University of California, San Francisco $^{5}$ Istanbul Medipol University $^{6}$ University of Zurich + +$^{7}$ ETH AI Center $^{8}$ University of Illinois Urbana-Champaign + +$^{9}$ École Polytechnique Fédérale de Lausanne + +Code, dataset, and models: https://github.com/MrGiovanni/RadGPT + +# Structured Reports (Section 3.1) + +![](images/bbef0773c8f91d9899a99e2c274b1e433426a8d16819445ca55925beb64539ad.jpg) + +Normal size (volume: $1249.9\mathrm{cm}^3$ + +Mean HU value: 131.4 +/- 23.2 + +![](images/671be995cc2564fc9eee18e42209c21b860e78dbec17b0db673ced3c608cfa39.jpg) + +# Liver tumor 1: + +Location: hepatic segment 7/8. + +Size: $3.1 \times 2.7 \, \text{cm}$ (slice 387). Volume: $10.9 \, \text{cm}^3$ . + +Enhancement relative to liver: Hypoattenuating (HU + +value is 58.7+/-29.7 + +![](images/0574dba93d8791b6672d2926ee8810626eb057ac9138ff710b3496711503130e.jpg) + +Pancreas is enlarged (volume: $84.6\mathrm{cm}^3$ + +Mean HU value: 105.7 +/-33.1 + +![](images/3d0e956d1090886ec1bcdf07a714542157d117137491e590b10ccbdd00878bc3.jpg) + +# Pancreas tumor 1: + +Location: pancreas head/body. + +Size: $2.9 \times 2.2 \mathrm{~cm}$ (slice 298). Volume: $8.2 \mathrm{~cm}^{3}$ . + +Tumor Stage (T stage): T2. + +Enhancement relative to pancreas: Hypoattenuating (HU value is 52.6+/-26.8). + +# Narrative Reports (Section 3.2) + +The patient has a hypoattenuating liver mass located in hepatic segments 7 and 8, measuring $3.1 \times 2.7 \, \text{cm}$ (slice 387) with a volume of $10.9 \, \text{cm}^3$ . The liver itself is of normal size, with a volume of $1249.9 \, \text{cm}^3$ and a mean HU value of $131.4 \div 23.2$ . The pancreas is enlarged, ... + +# IMPRESSION: + +1. A hypoattenuating liver mass located in hepatic segments 7 and 8, measuring $3.1 \times 2.7 \mathrm{~cm}$ . +2. A hypoattenuating pancreas mass located in the head and body, measuring $2.9 \times 2.2 \mathrm{~cm}$ , classified as T2. +3. Multiple small hypoattenuating kidney masses, with the largest one measuring $0.6 \times 0.3 \mathrm{~cm}$ . + +# Enhanced Human Reports (Section 3.3) + +# Kidney tumor 1: + +Location: left kidney. + +Size: $5.1 \times 4.1 \, \text{cm}$ (slice 90). Volume: $45.7 \, \text{cm}^3$ . + +Enhancement relative to kidney: Hypoattenuating (HU + +value is $-8.9 + / - 20.6)$ + +# Human-Made Report: + +Left lower lobe of the lung sees strip-shaped high-density shadow, considering as discoid atelectasis the lung. A low-density, round-shaped mass in the + +kidney, suggestive of a renal tumor, possibly malignant... + +![](images/6ab73341fb6d1d75bf3012a5f3a96c37ce80b0437c57b5a80009b44b47582b44.jpg) +Figure 1. AbdomenAtlas 3.0 is a large-scale, image-text tumor dataset of 9,262 3D CT scans. Each CT scan has per-voxel tumor annotations and reports, including 5,582 liver tumors, 368 pancreatic tumors and 4,424 kidney tumor, 7,003 of which are small tumors $(\leq 2\mathrm{cm})$ . In addition, AbdomenAtlas 3.0 provides detailed annotations for pancreatic cancer staging (T1-T4), as well as per-voxel segmentation of liver sub-segments (1-8) and pancreatic sub-segments (head, body, and tail). Structured, narrative, and enhanced reports were created by a team of 12 board-certified radiologists assisted by our proposed Radiology Generative Pretrained Transformer (RadGPT). + +# Kidney tumor 1: + +Location: left kidney. + +Size: $5.1 \times 4.1 \mathrm{~cm}$ (slice 90). Volume: + +${45.7}{\mathrm{\;{cm}}}^{3}$ + +Enhancement relative to kidney: + +Hypoattenuating (HU value is -8.9+/- + +20.6). + +Possibly malignant renal tumor. + +Lung: Left lower lobe sees strip-shaped + +high-density shadow, suggestive of + +discoid atelectasis... + +# Kidney tumor 2: + +Location: left kidney. + +Size: $0.5 \times 0.4 \, \text{cm}$ (slice 321). Volume: $0.1 \, \text{cm}^3$ . + +Enhancement relative to kidney: Hypoattenuating + +(HU value is 97.8+/-54.5). + +# Kidney tumor 3: + +Location: left kidney. + +Size: $0.6 \times 0.4 \, \text{cm}$ (slice 283). Volume: $0.1 \, \text{cm}^3$ . + +Enhancement relative to kidney: Hypoattenuating + +(HU value is $49.5\% - 48.4$ ). + +9,262 structured | 9,262 narrative reports | 240 fusion reports +9,262 CT scans from 138 institutions +10,374 tumor instances +3 pancreatic sub-segments +per-voxel annotated tumors, organs, vessels + +1.8 million tokens +7,003 small tumor $(< 2\mathrm{cm})$ +8 hepatic sub-segments +4 pancreatic tumor stages + +![](images/d327ebc57694afaaa43040f15c4861b0fdbdb0f792e4c795f44258ff34864924.jpg) + +![](images/ff92b43f3c773893a768b02a456ef0a0119add294470055d6f22d255091d292e.jpg) + +# Abstract + +Cancers identified in CT scans are usually accompanied by detailed radiology reports, but publicly available CT datasets often lack these essential reports. This absence limits their usefulness for developing accurate report generation AI. To address this gap, we present AbdomenAtlas 3.0, the first public, high-quality abdominal CT dataset with detailed, expert-reviewed radiology reports. All reports are paired with per-voxel masks and they describe liver, kidney and pancreatic tumors. AbdomenAtlas 3.0 has 9,262 triplets of CT, mask and report—3,955 with tumors. These CT scans come from 17 public datasets. Besides creating the reports for these datasets, we expanded their number of tumor masks by $4.2\times$ , identifying 3,011 new tumor cases. Notably, the reports in AbdomenAtlas 3.0 are more standardized, and generated faster than traditional human-made reports. They provide details like tumor size, location, attenuation and surgical resectability. These reports were created by 12 board-certified radiologists using our proposed RadGPT, a novel framework that converted radiologist-revised tumor segmentation masks into structured and narrative reports. Besides being a dataset creation tool, RadGPT can also become a fully-automatic, segmentation-assisted report generation method. We benchmarked this method and 5 state-of-the-art report generation vision-language models. Our results show that segmentation strongly improves tumor detection in AI-made reports. + +# 1. Introduction + +Each year, over 85 million CT scans are performed in the United States [44, 52], growing $6\%$ per year, and significantly outpacing the $0.7\%$ annual growth rate of the medical imaging workforce [13]. This disparity puts radiologists under significant time pressure, making it challenging to generate detailed, accurate radiology reports. AI may support report generation, but it requires data. To address this gap, we present AbdomenAtlas 3.0 (summarized in Figure 1 and Table 1), the first high-quality abdominal CT dataset with reports. It has 9,262 3D CTs in NIfTI format (2,789,975 CT slices) sourced from 17 public datasets (Table 1), which originally had no radiology report. 12 board-certified radiologists, assisted by RadGPT (introduced below), generated reports for all CTs—totaling 1,843,262 tokens. For each CT, we document tumor size, location, attenuation (HU), and volume for each identified tumor. Reports also include T-stage for pancreatic cancer (PDAC), derived from tumor size and vessel involvement, critical for surgery. Each CT has both structured (template-based) and narrative (free-text) reports, and precise voxel-level annotations. Reports cover tumors in the liver, pancreas, and kidneys, including 3,011 tumors newly identified by the radiologists. + +Our reports also describe organ abnormalities (e.g., fatty liver, enlarged spleen), patient demographics, and contrast phase. They locate tumors in liver segments (1-8) and pancreas segments (head, body, tail)—all annotated per-voxel. This is the largest liver sub-segment dataset, and the first public pancreas sub-segment dataset. Also, we enhanced 240 existing human-made reports, covering 66 distinct diagnoses, with more detailed tumor analyses. + +To create AbdomenAtlas 3.0, we developed Radiology Generative Pre-trained Transformers (RadGPT), an anatomy-aware vision-language AI agent that assists radiologists in creating CT-report datasets. We started with our previous AbdomenAtlas 1.1 [31], composed of 17 public datasets and their organ segmentation masks, but no tumor segmentation nor report. First, RadGPT segments liver, kidney, and pancreas tumors, along with liver/pancreas sub-segments, surrounding organs, and blood vessels1. Then, radiologists revise the segmented tumors, annotating missed ones and removing false positives. We call the dataset with CT scans and tumor segmentation masks AbdomenAtlas 2.0, and we also release it here. From the revised segmentations, RadGPT extracts attributes (e.g., tumor size, volume, attenuation, stage) via deterministic, rule-based algorithms. These attributes are used to fill a radiologist-designed template, producing structured reports. RadGPT's deterministic algorithms ensure that the structured reports are fully explainable and fully coherent with the radiologist-revised segmentations. Next, RadGPT converts the structured reports into free-text narrative reports, using large language models (LLMs) that emulate the style (word choice and organization) of radiologists at a major US hospital—through in-context learning with special example selection (\$3.2). Last, RadGPT fuses per-voxel segmentations with human-made reports/clinical notes to produce enhanced human reports (\$3.3), combining precise and detailed tumor analysis from segmentation with broader diagnostic range (66 diagnoses) from human-made reports. Reports were verified by radiologists (Appendix C). We call the final triplet dataset—CT scans, tumor masks, reports—AbdomenAtlas 3.0. + +We evaluated six CT report generation models on AbdomenAtlas 3.0 (internal validation) and a private dataset (external validation): CT2Rep [21], M3D [4], CT-CHAT [20], Merlin [8], RadFM [54] and RadGPT. Besides a dataset creation tool, RadGPT can also become a fully-automatic, segmentation-assisted report generation model, by converting the outputs of a segmentation model into reports, without radiologist revision. We expect AbdomenAtlas 3.0 to foster segmentation-assisted report generation, as the dataset has CTs, per-voxel annotations and reports. We evaluated all report generation models with a new diagnostic metric (§B.3). It first uses an LLM to extract labels (tumor presence) from AI- and human-made reports. Then, it compares the labels from AI- and human-made re + +
datasetCTsinstitutionscountriesannotated liver tumorsannotated pancreatic tumorsannotated kidney tumors
FLARE'23 [2022] [link]4,1003510 → 5640 → 380 → 941
KiTS'23 [2020] [link]489110 → 10452
LiTS [2019] [link]131755000
TCIA-Pancreas-CT [2015] [link]4211000
CT-ORG [2020] [link]140860 → 4400 → 21
Trauma Det. [2023] [link]4,71423130 → 1130 → 320 → 38
BTCV [2015] [link]4711000
CHAOS [2018] [link]201100 → 10
AbdomenCT-1K [2021] [link]1,0501270 → 1170 → 940 → 181
MSD CT Tasks (6) [2021] [link]94511251 → 4621910 → 388
WORD [2021] [link]120110 → 470 → 10 → 45
AMOS [2022] [link]200210 → 740 → 40 → 56
AbdomenAtlas 3.0 (ours)9,26213819301 → 1,472191 → 361452 → 2,122
datasetliver sub-segmentspancreas sub-segmentsperipancreatic blood vessels1tumor stageradiology reportstext tokens
FLARE'23 [2022] [link]XXXX00
KiTS'23 [2020] [link]XXXX00
LiTS [2019] [link]✓XXX00
TCIA-Pancreas-CT [2015] [link]XXXX00
CT-ORG [2020] [link]XXXX00
Trauma Det. [2023] [link]XXXX00
BTCV [2015] [link]XXXX00
CHAOS [2018] [link]XXXX00
AbdomenCT-1K [2021] [link]XXXX00
MSD CT Tasks (6) [2021] [link]XXXX00
WORD [2021] [link]XXXX00
AMOS [2022] [link]XXXX00
AbdomenAtlas 3.0 (ours)✓✓✓✓18,5241,843,262
+ +$\rightarrow$ represents the number of CT scans with tumor annotations in the original dataset, followed $(\rightarrow)$ by our updated number of CT scans with tumor annotations, including the additional annotations AbdomenAtlas 3.0 provided with radiologist support. + +Table 1. Besides being the only public abdominal CT dataset with paired radiology reports, AbdomenAtlas 3.0 offers $4.2 \times$ more annotated tumors than the combined total of its constituent datasets. The table highlights how AbdomenAtlas 3.0 enhances public datasets with reports and tumor annotations. It includes 1,472 CT scans with liver tumors, 361 with pancreatic tumors, and 2,122 with kidney tumors, most newly annotated with radiologist support. Each sample includes per-voxel annotations and reports. AbdomenAtlas 3.0 is also the first dataset to provide per-voxel segmentations of pancreas sub-segments and peripancreatic blood vessels. AbdomenAtlas 1.1 [31] has the same CTs as AbdomenAtlas 3.0, but it has only organ segmentation masks—no tumor masks, reports, organ sub-segments, nor peripancreatic blood vessels. AbdomenAtlas 2.0 has the same CTs and masks as 3.0, no report—it is our intermediate step before 3.0. + +ports to evaluate AI's diagnostic sensitivity and specificity (§3.4). To validate this new metric, radiologists manually evaluated LLM labeling—it achieved 96% zero-shot accuracy (Figure 4). Our contributions are: + +1. AbdomenAtlas 3.0 is the first public dataset with high-quality abdominal CT scans (9,262), radiology reports (structured, narrative, and enhanced), and tumor masks. +2. With 12 radiologists, we annotated 3,011 new tumors in the 17 public datasets inside AbdomenAtlas 3.0—expanding their number of tumor masks by $4.2 \times$ +3. Our reports locate liver and pancreas tumors within sub-segments of the organs. They also measure contact between tumors and blood vessels for pancreatic tumor staging. Staging and sub-segments are key for surgery. +4. We developed Rad-GPT to assist dataset creation: unlike current VLMs, it uses deterministic algorithms to convert radiologist-revised tumor masks into reports, improving reports' trustworthiness and interpretability. Also, RadGPT can generate fully-automated reports. +5. We benchmarked 5 SOTA VLMs for report generation and showed segmentation improves report generation. + +# 2. Related Work + +Per-voxel tumor annotations are scarce. Most public abdominal CT datasets concentrate on a single tumor type (e.g., liver [7], pancreas [3], or kidney [22]) and contain only a few hundred tumor annotations (Table 1). This small volume of annotations hinders effective AI training and evaluation. To address this, our radiologists have quadrupled the number of per-voxel tumor annotations in the 17 public datasets included in AbdomenAtlas 3.0 (Table 1). + +Real-world radiology reports are even rarer than per-voxel tumor annotations. At the time of writing, no publicly available abdominal CT dataset contains authentic clinical reports. Only one dataset, M3D-Cap [4], provides textual captions (sourced from Radiopaedia [19]), but its scans are 2D JPG/PNG image series rather than standard 3D NIfTI or DICOM volumes. Consequently, crucial information such as inter-slice spacing and Hounsfield units (HU) is missing [59]. In contrast, CT scans in AbdomenAtlas 3.0 were collected in standard formats from 138 medical institutions, retaining clinically important metadata. As an + +![](images/7bce6517f12fc41ad5fcddeb103b7d0f5c0b1326d3c137cdc8ee8a106e503c0d.jpg) +Figure 2. The RadGPT 3-stage pipeline for report generation. Blue arrows denote human revision used to create AbdomenAtlas 3.0. By skipping revision, RadGPT can also become a fully-automatic, segmentation-assisted report generation AI. Stage I. Segmentation. DiffTumor [11] and nnU-Net [25] segment 26 anatomical structures important for cancer detection and staging1. Radiologists corrected wrong tumor segmentations in AbdomenAtlas 3.0, and ground-truths from public datasets were used when available. Stage II. Structured Report Generation. Deterministic algorithms (§3.1.1-3.1.3) extract radiologist-selected attributes—important for cancer detection, staging and treatment—from CTs and segmentations. Attributes fill a radiologist-designed template, generating structured reports detailing liver, kidney, and pancreatic tumors. The rule-based deterministic algorithms ensure the reports are fully coherent with segmentations and explainable. Stage III. Style Adaptation. LLM adapts structured reports into a target hospital's narrative style, leveraging example reports from the hospital—in-context learning prioritizing examples of similar diagnoses (§3.2). LLM is asked to preserve medical information and double checks for consistency. Radiologists revised reports in AbdomenAtlas 3.0. Also, LLM can fuse structured and human-made reports, creating enhanced human reports combining segmentation-based precision with humans' broad diagnostic range (§3.3). + +other unique quality, our reports are paired with tumor and organ masks—fostering the development of segmentation-assisted report generation AI. + +Due to the scarcity of reports in public datasets, only two models specifically target abdominal CT report generation: M3D [4] (publicly released) and Merlin [8] (partially released). Text-similarity metrics were used to evaluate both models (e.g., BLEU and ROUGE [35]; Merlin was also evaluated with RadGraph-F1), but these metrics can be skewed by style variations even when the underlying diagnoses remain unchanged (§B.3). In contrast, we propose the evaluation of AI-generated reports using diagnostic sensitivity and specificity (Table 2)—clinically meaningful and acceptable metrics [9, 55]. Lastly, although many report-generation models exist for 2D X-ray [12, 33, 34, 45, 51, 57], adapting them to 3D CT may require profound re-design, which may unfairly represent the originals. Why? First, tumors in CTs can occupy $\leq 0.0001\%$ of the full volume, vs. $5 - 10\%$ in X-rays. Second, many X-ray models rely on 2D pre-trained models, but CT data is 3D. Processing CT slices individually is computationally prohibitive, and it is difficult to align slices with findings in reports. Thus, all models we evaluated in AbdomenAtlas 3.0 [4, 8, 20, 21, 54] are designed for CT. + +# 3. AbdomenAtlas 3.0 & RadGPT + +Table 1 shows advantages of AbdomenAtlas 3.0 over its 17 source datasets—providing reports, organ sub-segments and blood vessels annotated per-voxel, and $4 \times$ more tumor + +annotations. Sections §3.1–§3.3 explain RadGPT (summarized in Figure 2), and how it empowered 12 radiologists to generate reports for the 9,262 CTs in AbdomenAtlas 3.0. + +# 3.1. Creating Structured Reports + +Structured reports use a radiologist-designed template, enhancing clarity and aiding medical decisions [1] (Figure 1). To fill the template, RadGPT uses segmentation and deterministic algorithms to: (1) sub-segment organs to locate tumors ( $\S 3.1.1$ ); (2) measure tumor size, volume, and attenuation ( $\S 3.1.2$ ); (3) perform cancer staging from tumor and blood vessel segmentations ( $\S 3.1.3$ ). + +# 3.1.1. Sub-segment Organs to Locate Tumors + +Human-made reports use organ sub-segments to locate tumors. Location is key for prognosis, tracking tumor progression, and treatment planning. E.g., the possibility of tumor surgical removal depends on its location [47]. To locate liver and pancreas tumors in structured reports, RadGPT sub-segments the organs and checks which sub-segments intersect with the tumor. RadGPT segments tumors with DiffTumor [11], a public segmentation model, and radiologists revise the segmentations (Appendix C). + +For liver sub-segmentation, we leverage whole-liver ground-truth per-voxel annotations to help the AI find liver sub-segments. First, we offset the liver intensity (by 200 HU), following its ground-truth per-voxel annotation. Second, using these CT scans with offsets as input, we trained an nnU-Net [25] for liver sub-segmentation. The sub-segments follow the Couinaud standard [15], which divides + +the liver into eight sub-segments that can be independently removed in surgeries. Couinaud sub-segment annotations are publicly available for 131 LiTS CT scans [7, 58], which we used for training. Given the small size of this dataset, we fine-tuned an nnU-Net pre-trained on 9,262 CT scans in AbdomenAtlas 1.1 [31]. After fine-tuning, we inferred the nnU-Net on AbdomenAtlas 3.0. The HU value offsetting ensured the precise alignment between the generated sub-segments and the existing liver ground-truth pervoxel annotations. AbdomenAtlas 3.0 is the second [58] but largest public dataset with liver sub-segments. + +For pancreas sub-segmentation, there is no public dataset with per-voxel annotations of pancreas head, body, and tail—ours is the first. Thus, to subsegment the pancreas, we used the superior mesenteric artery (SMA) as a landmark. We trained an nnU-Net to segment the SMA (using private data) and developed a deterministic algorithm that uses the SMA segmentation to sub-segment the pancreas (Sup. Alg. 1). First, it uses the SMA to find the pancreatic neck, since it curves around the SMA. The neck locates the head-body boundary. Then, the body-tail boundary is set at the midpoint along their length. Our landmark-based deterministic algorithm closely mimics how radiologists use mesenteric vessels to subsegment the pancreas [48]. AbdomenAtlas 3.0 is the first public dataset with pancreas sub-segments. + +# 3.1.2. Measure Tumors Like Radiologists + +Radiologists commonly measure tumors using use the World Health Organization (WHO) standard, which provides two diameters: the longest tumor diameter in any axial plane $(D)$ , and its perpendicular diameter in the same plane $(d)$ [40]. Standardization of measurements is key for accurate cancer prognosis and treatment planning [32, 40]. Thus, RadGPT also uses the WHO standard, measuring tumors like radiologists. AbdomenAtlas 3.0 presents radiologist-revised segmentations of liver, kidney and pancreas tumors. From segmentations, RadGPT extracts tumor measurements using a deterministic algorithm that implements the WHO standard (Sup. Alg. 2). Besides diameters, our structured reports present tumor & organ volume and attenuation (HU values), also extracted from segmentation1. Using volumes, our reports diagnose enlarged organs, and attenuation diagnoses fatty liver (average HU $< 40$ [28]) and pancreas (pancreas-to-spleen attenuation $< 0.7$ [18])—a condition related to diabetes and pancreatic cancer [18]. Meanwhile, tumor attenuation helps identify tumor type. + +![](images/65ac0c1f98c5eae5a742ccf070a513c9b2bb2aabb4777586504d1949fa8d600a.jpg) +Figure 3. Automated T staging. Our RadGPT first segments the tumor and key vascular structures from CT scans, then measures tumor size and blood vessel contact angle to automatically assign T stage and resectability. If the tumor-vessel contact angle surpasses 180 degrees, the tumor becomes unresectable (T stage 4). + +![](images/aa24ae534b377c324697cd2e8eeb8806f636e4714dc0f5a60610b8653b040ebe.jpg) + +![](images/0f34d29d7d7f5f8173e4ff3ff1711649c8e241228192c248d4a1aeb0cb519b59.jpg) + +# 3.1.3. Stage Pancreatic Cancer using Segmentation + +Tumor T-stage summarizes tumor size and relationship to nearby structures. It is key for surgical planning and survival, especially for pancreatic adenocarcinoma (PDAC), an aggressive cancer [1]. However, staging is time-consuming. As shown in Figure 3, for PDAC staging, radiologists must measure tumors (§3.1.2) and analyze its interaction with blood vessels (SMA, CHA, CA, SA) [1]. Accordingly, RadGPT first segments vessels and tumors (using nnU-Net and DiffTumor1) and radiologists revise segmentations (Appendix C). Then, a deterministic algorithm uses the revised segmentations to measure the tumor-vessel contact angle (Sup. Alg. 3). Large angles ( $>180^{\circ}$ ) make surgery difficult, increasing stage. For interpretability, reports justify stages with tumor size and tumor-vessel degree of contact, and our deterministic algorithm faithfully implements the guidelines radiologists use to stage PDAC [1]. AbdomenAtlas 3.0 is first public dataset with PDAC T stage labels. + +# 3.2. Creating Narrative Reports + +Structured reports use rigid templates to improve clarity and clinical decision-making [1]. However, rigid templates may conflict with the reporting style of an institution. Thus, RadGPT can create narrative reports that mimic the style of a target institution. In AbdomenAtlas 3.0, they mimic human-made reports at a major US hospital (Figure 1). The narrative reports are created through style adaptation with in-context learning: we provide a pre-trained LLM (Llama3.1 70B, AWQ quantization [17]) with a structured report and 10 human-made reports from the target institution, and the LLM adapts the structured report to the style of the human-made reports. We ask the LLM not to change diagnoses or details. Thus, narrative reports contain all the detailed information from structured reports (§3.1). + +However, style of human-made reports varies with diagnoses. E.g., pancreatic tumor differ from liver tumor reports [41]. Thus, we verify diagnoses to give the LLM example reports with the correct style. First, another LLM categorizes human-made reports according to tumors (liver, pancreas, kidney, none). Then, when adapting a structured report to narrative, the first LLM receives example human-made reports with the same tumor as the structured report. + +Internal validation on the test set of AbdomenAtlas 3.0 (IID) + +
Modelpancreatic tumor (%)kidney tumor (%)liver tumor (%)
Sen. (≤2 cm)Sen. (>2 cm)Spec.Sen. (≤2 cm)Sen. (>2 cm)Spec.Sen. (≤2 cm)Sen. (>2 cm)Spec.
CT-CHAT [20]66.751.961.231.132.874.25.73.294.7
CT2Rep [21]0.00.092.536.539.370.435.849.270.4
M3D [4]0.07.497.28.116.484.19.412.786.0
Merlin [8]33.351.971.828.445.986.630.241.395.9
RadFM [54]0.00.099.93.76.395.63.35.793.9
RadGPT (ours)66.781.593.254.893.351.839.696.864.4
+ +External validation on unseen hospital—UCSF (OOD) + +
ModelSen. (≤2 cm)Sen. (>2 cm)Spec.Sen. (≤2 cm)Sen. (>2 cm)Spec.Sen. (≤2 cm)Sen. (>2 cm)Spec.
CT-CHAT [20]27.5N/A73.124.329.774.65.24.294.0
CT2Rep [21]2.1N/A96.74.010.098.00.00.0100.0
M3D [4]3.3N/A97.914.813.186.310.717.387.3
Merlin [8]7.5N/A100.08.19.2100.09.119.2100.0
RadFM [54]0.0N/A100.07.56.890.910.911.185.0
RadGPT (ours)76.9N/A76.692.097.378.379.689.473.4
+ +Table 2. In tumor detection, fully-automated reports by RadGPT surpass reports created by end-to-end report generation models. We use RadGPT as a fully-automated segmentation-assisted report generation model (Figure 2). The results indicate that per-voxel segmentation (step 1 in the RadGPT pipeline) may strongly improve report generation. We tested out-of-distribution (OOD) at UCSF, a hospital not seen in training, and in-distribution (IID). In the IID set, the ground-truth contained 9 small and 27 large pancreatic tumors, 74 small and 61 large kidney tumors, and 53 small and 63 large liver tumors, with 890, 791, and 810 negatives, respectively. In OOD, we have 385 (small) and 0 (large) for pancreas, 50 (small) and 219 (large) for kidney, and 142 (small) and 301 (large) for liver, with 244 negatives for each organ. Decision thresholds are analyzed in 15. While other methods were evaluated zero-shot, CT-CHAT, CT2Rep and Merlin were trained in AbdomenAtlas 3.0, giving them an advantage in the IID dataset. To compute sensitivity and specificity, we used our proposed diagnostic evaluation (\$3.4): an LLM extracted binary tumor presence labels per-organ, and we compared the labels for AI-made reports and ground-truth human-made reports. LLM label extraction accuracy is $96\%$ (Figure 4). Table 5 provides additional metrics (BLEU, ROUGE, BERT, RadGraph-F1), showing they are usually sensible to variations in report style, unlike our diagnostic evaluation. + +After adapting a structured report into a narrative report, the LLM performed a quality check. It extracted diagnoses and quantitative information (e.g., tumor size and stage) from both reports and checked for consistency. We prompted the LLM to correct in the narrative report any information diverging from the structured report, and to remove any diagnosis not present in the structured report. + +# 3.3. Creating Enhanced Human Reports + +Like most abdominal CT datasets, AbdomenAtlas 3.0 focuses on tumors—as cancer is a major cause of death. Our reports can precisely measure and analyze multiple tumors in a CT, while human-made reports usually measure the largest tumors only (§4.3). However, human-made reports cover multiple diagnoses unrelated to tumors. To combine their strengths, RadGPT prompts the zero-shot LLM (Llama 3.1 70B AWQ) to fuse the details in structured reports with the many diagnoses in human-made reports/clinical notes (Figures 12 and 1), generating enhanced human reports. AbdomenAtlas 3.0 has 240 of them: 209 used clinical notes for TotalSegmentator CT scans [4], and 31 used notes from our radiologists. They span 66 diagnoses. + +# 3.4. Evaluating Diagnoses in AI-Made Reports + +We propose a new strategy to evaluate the clinical utility of AI-made reports: a straightforward, LLM-based diagnostic + +evaluation. First, we prompt a zero-shot LLM (Llama 3.1 70B AWQ, prompts in §B.4) to identify in which organ the report mentions tumors. Then, we convert the LLM answer into categorical labels. We compare labels for AI-made and human-made reports (ground-truth) to calculate tumor detection sensitivity and specificity. This evaluation strategy is scalable and practical: with zero-shot inference, it does not need fine-tuning and is easily adaptable to multiple hospitals. Importantly, our strategy produces clinically relevant metrics (detection sensitivity / specificity), which are easy to interpret by clinicians. Here, we limit our evaluation strategy to tumor detection. However, it can be expanded to evaluate other relevant clinical information and diseases beyond tumors—with simple prompt modifications. + +# 4. Experiment & Result + +We randomly selected $10\%$ of AbdomenAtlas 3.0 as a test set, where we evaluated 6 CT report generation models: CT2Rep [21], M3D [4], CT-CHAT [20], Merlin [8], RadFM [54] and RadGPT—as baselines for future work. In AbdomenAtlas 3.0, RadGPT transforms radiologist-revised tumor masks into reports. In this section, we evaluate RadGPT as a fully-automated, segmentation-assisted method, without radiologist revision (Figure 2). We have both AI trained on AbdomenAtlas 3.0 (CT2Rep and CTCHAT, see Appendix B.1 for training details) and those + +![](images/ce585dd621dda7ac57775069f00f359d5a4b1da5fba8435bcf180cb9501bbe09.jpg) +Figure 4. Confusion matrices. (a) A zero-shot LLM (Llama 3.1 70B AWQ) has $96\%$ accuracy, 0.953 F1-score in determining if radiology reports show tumors. Thus, the LLM can accurately calculate tumor detection sensitivity and specificity for AI-made reports ( $\S 3.4$ ). LLM's accuracy rivals established labelers, as those in CheXpert [24] and CheX-ray14 [50]. Results were manually evaluated by radiologists on 447 reports with kidney, pancreas, and liver tumors. (b) PDAC staging confusion matrix for RadGPT, the first public AI for staging abdominal CT tumors. Results on a private dataset with ground-truth tumor stage annotations ( $N = 42$ ). + +![](images/5f9a3115ecf964157c00c7b02c074dd06b6f9ac2120e11b85ff235c4b7d638fa.jpg) + +trained on abdominal CTs in other works (M3D, Merlin, DiffTumor inside RadGPT). To ensure realistic evaluation [5, 6], we evaluate on the AbdomenAtlas 3.0 test set and on a private out-of-distribution (OOD) dataset, from the University of California San Francisco hospital (UCSF, California, USA) never seen by any AI in training. + +Zero-shot LLMs can accurately evaluate report generation. For automated evaluation on a large test dataset, we will use an LLM (Llama-3.1) to assess the reports generated by 6 AI models ( $\S 3.4$ ). Before that, radiologists verified the LLM's ability to determine whether a report indicates tumors or not. They read the zero-shot LLM answers for 447 different reports, verifying that it achieved $96\%$ accuracy (Figure 4). Results demonstrate the LLM reliability in evaluating tumor detection, per-organ. + +Segmentation can assist report generation models. The LLM-based evaluation, (Table 2) showed that the reports generated by RadGPT strongly surpassed the other abdominal CT report generation models, especially in the OOD test set (unseen hospital) $^{2}$ . End-to-end trained methods had difficulty detecting tumors in the OOD dataset (low sensitivity), and RadGPT strongly outperformed them for small and large tumors in the liver, pancreas, and kidneys. This performance difference shows the benefits of using segmentation to improve report generation: DiffTumor produces accurate tumor segmentations, which RadGPT translates into reports. By releasing AbdomenAtlas 3.0, the first abdominal CT dataset with triplets of CT scans, reports, and per-voxel annotations, our objective is to catalyze further research on segmentation-assisted report generation. + +RadGPT is the first public AI model to perform cancer staging on abdominal CT. Figure 4 shows the perfor + +
liver tumorpancreatic tumorkidney tumor
Detection Precision (%)92.3 (12/13)50.0 (8/16)91.7 (11/12)
Size Accuracy (%)100.0 (12/12)75.0 (6/8)100.0 (11/11)
+ +Table 3. RadGPT has $75.6\%$ tumor detection precision and $93.5\%$ tumor measurement accuracy. A radiologist manually evaluated reports RadGPT created for 23 external test CTs (UCSF). A tumor measurement was considered correct if it deviated by $\leq 10\%$ from the radiologist's measurement (both use the WHO measuring standard [40]). As evaluation is time-consuming, the radiologist evaluated 23 reports. Using an LLM for automatically evaluating tumor measurements is challenging: it requires pairing tumors in AI-made reports and ground-truth reports. + +mance of RadGPT for staging of pancreatic adenocarcinoma. RadGPT fully-automated reports achieved accuracy of $71.43\%$ in determining tumor T stages 1 to 3. The results show that AI is a promising tool for assisting cancer staging, a key but time-consuming task for radiologists. Still, these fully-automatic results show radiologist revision is essential to ensure staging accuracy in AbdomenAtlas 3.0. + +# 4.1. RadGPT Accurately Measures Tumor Size + +An expert radiologist manually evaluated structured reports generated by RadGPT. He analyzed each reported tumor, evaluating its measurement and checking if the tumor is a false-positive (tumor not present in the CT volume) or a true-positive (present). The radiologist deemed $75.6\%$ of the tumors reported by RadGPT true-positives, and $93.5\%$ of them were correctly measured (Table 3). RadGPT only made measuring mistakes for pancreatic tumors (PDAC), but even the radiologist could not measure 3 PDACs. + +# 4.2. RadGPT Locates Tumors in Organs + +RadGPT uses use organ sub-segments to locate tumors. It achieved a Dice similarity coefficient (DSC) of 0.85 in segmenting eight liver sub-segments, according to the test set from Zhang et al. [58]. For pancreas sub-segmentation, we do not have a ground-truth or dataset for testing, because AbdomenAtlas 3.0 is the first public dataset to present pancreas sub-segments (head, body, and tail). However, our algorithm to sub-segment the pancreas closely follows radiologist-accepted standards (see Figure 5), and we asked radiologists to qualitatively evaluate our annotations. + +# 4.3. RadGPT Enhances Human-made Reports + +Human-made radiology reports often omit critical quantitative details—such as tumor volumes and attenuation (HU) values—compromising clinical decision-making. In our evaluation of 90 human reports from a UCSF, none reported organ or tumor volumes and only $63\%$ measured all detected tumors. In contrast, our structured reports (RadGPT) consistently provide full, quantitative data. As shown in Table 4, while human-made reports measure volume and + +![](images/9d96436911d97be6a2b4e1029f29d29e733c9f780c154d587de36d496e83ff48.jpg) +Figure 5. Pancreas and liver sub-segments. (a) RadGPT segments the pancreas based on radiology standards [48]. (I-II) the SMA separates the pancreas head (H) from the body (B), and (III) the remaining pancreas is divided at its midpoint into the body and tail (T). (b) Our liver sub-segmentation model achieved a DSC of 0.85 in segmenting eight liver sub-segments on a public test set [58]. Sub-segments are in different colors and tumors in brown. Sub-segments are essential for RadGPT to locate tumors. + +HU in $0\%$ of cases and capture all tumors in only $63\%$ of cases, RadGPT achieves $100\%$ for all metrics. This level of consistency streamlines clinical assessments and improves prognostic accuracy by ensuring every tumor is precisely measured [10, 39, 53]. + +In AbdomenAtlas 3.0, 240 CT scans include clinical notes, which lack quantitative tumor measurements; for instance, among 63 TotalSegmentator notes mentioning tumors, none provide such data—even though they report other findings like calcified arterial plaques. By merging these notes with our structured and narrative reports using an LLM, RadGPT generates 240 enhanced reports that integrate the notes' comprehensive clinical findings (covering 66 diagnoses) with precise tumor sizes (see Figure 12). + +# 4.4. Discussion: End-to-End, Segmentation-based? + +Table 2 shows RadGPT outperformed end-to-end VLMs. The unique design of RadGPT offers several advantages over end-to-end training. (i) Interpretable: RadGPT generates reports from tumor segmentation, allowing clinicians and developers to visualize and verify tumor locations and sizes in the CT. In contrast, errors in end-to-end methods are harder to diagnose and debug. (ii) Interactive: Our goal is creating a high-quality dataset of CTs and reports to drive innovation in report generation. As all algorithms make mistakes, a human-in-the-loop approach is key. Segmentation is a safeguard, allowing radiologists to ensure reports are correct, and easily transforming radiologist-revised masks into reports. (iii) Strong supervision: tumor segmentation AI has been a long-term focus of the research community, achieving high accuracy by leveraging precise per-voxel masks. Our benchmark shows that segmentation-assisted models can transfer this high accuracy to reports. + +
VolumeHUDiameters
Human-made0%0%63%
RadGPT (ours)100%100%100%
+ +Table 4. Comparison of human-made reports vs. RadGPT reports for 90 UCSF CTs. Values indicate the percentage of reports containing tumor volume, HU and diameter measurements for all detected tumors. RadGPT reports provide more clinically relevant [10, 39, 53] quantitative details about tumors. + +# 5. Conclusion & Future Work + +Dataset curation and report-generation are inter-dependent. Developing image-report-mask methods requires image-report-mask datasets, but creating these datasets requires reliable methods and human-in-the-loop involvement. Our focus is creating high-quality image-report-mask datasets to support further methodological advancements. + +AbdomenAtlas 3.0 is the first public dataset providing high-quality abdominal CT scans with reports and pervoxel tumor annotations, encompassing 9,262 CT scans from 138 institutions. It uniquely includes pancreas sub-segments, peripancreatic blood vessels, and pancreatic cancer stages—absent in existing public datasets. RadGPT transforms per-voxel annotations into structured reports using deterministic algorithms. These reports align with the accuracy of segmentations revised by radiologists in AbdomenAtlas 3.0. Additionally, RadGPT enables fully-automated report generation, surpassing existing approaches in detecting tumors. Together, AbdomenAtlas 3.0 and RadGPT bridge the gap between tumor segmentation and report generation, offering valuable resources and tools to advance AI in abdominal CT interpretation. + +We are committed to expanding AbdomenAtlas 3.0 to include reports for more types of tumors. Additionally, we plan to host benchmarks using AbdomenAtlas 3.0 with two train/test splits. IID Split: Randomly sets aside $10\%$ of the dataset for testing, where training and testing data come from the same institutions, following standard AI evaluation practices. Used in Table 2. OOD Split: Uses data from 23 unseen institutions (4,500 CT scans) for testing, providing a large test set to evaluate AI generalization to new environments. This benchmark will assess report generation models using standard text similarity metrics but will prioritize tumor detection sensitivity and specificity, enabled by our proposed LLM-based diagnostic evaluation. + +Although 66 out of 240 fusion reports present diverse diagnoses, AbdomenAtlas 3.0 is cancer-centric. Cancer is a leading cause of death, and over $40\%$ of medical imaging reports focus on cancer detection. Thus, AI-assisted report generation has the potential for significant impact. We hope our release of the cancer-centric AbdomenAtlas 3.0 will stimulate further AI advancements in the field. + +Acknowledgments. This work was supported by the Lustgarten Foundation for Pancreatic Cancer Research and the National Institutes of Health (NIH) under Award Number R01EB037669. We would like to thank the Johns Hopkins Research IT team in IT@JH for their support and infrastructure resources where some of these analyses were conducted; especially DISCOVERY HPC. We thank the funding of Italian Institute of Technology and the HPC infrastructure at Italian Institute of Technology. P.R.A.S.B. thanks the funding from the Center for Biomolecular Nanotechnologies, Istituto Italiano di Tecnologia (73010, Arnesano, LE, Italy). + +# References + +[1] Mahmoud M Al-Hawary, Isaac R Francis, Suresh T Chari, Elliot K Fishman, David M Hough, David S Lu, Michael Macari, Alec J Megibow, Frank H Miller, Koenraad J Mortele, et al. Pancreatic ductal adenocarcinoma radiology reporting template: consensus statement of the society of abdominal radiology and the american pancreatic association. Radiology, 270(1):248-260, 2014. 4, 5 +[2] Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M Summers, Bram van Ginneken, et al. The medical segmentation decathlon. arXiv preprint arXiv:2106.05735, 2021. 3 +[3] 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):1-13, 2022. 3 +[4] Fan Bai, Yuxin Du, Tiejun Huang, Max Q-H Meng, and Bo Zhao. M3d: Advancing 3d medical image analysis with multi-modal large language models. arXiv preprint arXiv:2404.00578, 2024.2,3,4,6 +[5] Pedro RAS Bassi, Sergio SJ Dertkigil, and Andrea Cavalli. Improving deep neural network generalization and robustness to background bias via layer-wise relevance propagation optimization. Nature Communications, 15(1):291, 2024. 7 +[6] Pedro RAS Bassi, Wenxuan Li, Yucheng Tang, Fabian Isensee, Zifu Wang, Jieneng Chen, Yu-Cheng Chou, Yannick Kirchhoff, Maximilian Rokuss, Ziyan Huang, Jin Ye, Junjun He, Tassilo Wald, Constantin Ulrich, Michael Baumgartner, Saikat Roy, Klaus H. Maier-Hein, Paul Jaeger, Yiwen Ye, Yutong Xie, Jianpeng Zhang, Ziyang Chen, Yong Xia, Zhaohu Xing, Lei Zhu, Yousef Sadegheih, Afshin Bozorgpour, Pratibha Kumari, Reza Azad, Dorit Merhof, Pengcheng Shi, Ting Ma, Yuxin Du, Fan Bai, Tiejun Huang, Bo Zhao, Haonan Wang, Xiaomeng Li, Hanxue Gu, Haoyu Dong, Jichen Yang, Maciej A. Mazurowski, Saumya Gupta, Linshan Wu, Jiaxin Zhuang, Hao Chen, Holger Roth, Daguang Xu, Matthew B. Blaschko, Sergio Decherchi, Andrea Cavalli, Alan L. Yuille, and Zongwei Zhou. Touchstone benchmark: Are we on the right way for evaluating ai algorithms for medical segmentation? Conference on Neural Information Processing Systems, 2024. 7 + +[7] Patrick Bilic, Patrick Ferdinand Christ, Eugene Vorontsov, Grzegorz Chlebus, Hao Chen, Qi Dou, Chi-Wing Fu, Xiao Han, Pheng-Ann Heng, Jirgen Hesser, et al. The liver tumor segmentation benchmark (lits). arXiv preprint arXiv:1901.04056, 2019. 3, 5 +[8] Louis Blankemeier, Joseph Paul Cohen, Ashwin Kumar, Dave Van Veen, Syed Jamal Safdar Gardezi, Magdalini Paschali, Zhihong Chen, Jean-Benoit Delbrouck, Eduardo Reis, Cesar Truyts, et al. Merlin: A vision language foundation model for 3d computed tomography. arXiv preprint arXiv:2406.06512, 2024. 2, 4, 6, 12 +[9] Kai Cao, Yingda Xia, Jiawen Yao, Xu Han, Lukas Lambert, Tingting Zhang, Wei Tang, Gang Jin, Hui Jiang, Xu Fang, et al. Large-scale pancreatic cancer detection via non-contrast ct and deep learning. Nature medicine, 29(12): 3033-3043, 2023. 4 +[10] Z. Cao and et al. Volumetric versus linear measurements in lung cancer prognosis. Annals of Surgical Oncology, 26(12): 3757-3763, 2019. 8 +[11] Qi Chen, Xiaoxi Chen, Haorui Song, Zhiwei Xiong, Alan Yuille, Chen Wei, and Zongwei Zhou. Towards generalizable tumor synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4 +[12] Zeyuan Chen, Lingxi Xie, Xin Huang, Yu Rong, Shuguang Cui, and Bo Dai. Generating radiology reports via memory-driven transformer. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 231-240, 2020. 4 +[13] Noel CF Codella, Ying Jin, Shrey Jain, Yu Gu, Ho Hin Lee, Asma Ben Abacha, Alberto Santamaria-Pang, Will Guyman, Naiteek Sangani, Sheng Zhang, et al. Medimageinsight: An open-source embedding model for general domain medical imaging. arXiv preprint arXiv:2410.06542, 2024. 2 +[14] Errol Colak, Hui-Ming Lin, Robyn Ball, Melissa Davis, Adam Flanders, Sabeena Jalal, Kirti Magudia, Brett Marinelli, Savvas Nicolaou, Luciano Prevedello, Jeff Rudie, George Shih, Maryam Vazirabad, and John Mongan. Rsna 2023 abdominal trauma detection. https://kaggle.com/competitions/rsna-2023-abdominal-trauma-detection, 2023. Kaggle. 3 +[15] Claude Couinaud. Le foie: études anatomiques et chirurgicales. Masson, 1957. 4 +[16] Jean-Benoit Delbrouck, Pierre Chambon, Zhihong Chen, Maya Varma, Andrew Johnston, Louis Blankemeier, Dave Van Veen, Tan Bui, Steven Truong, and Curtis Langlotz. Radgraph-xl: A large-scale expert-annotated dataset for entity and relation extraction from radiology reports. In Findings of the Association for Computational Linguistics: ACL 2024, pages 12902–12915, 2024. 13 +[17] 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, 2024.5 +[18] Yasunari Fukuda, Daisaku Yamada, Hidetoshi Eguchi, Tomoki Hata, Yoshifumi Iwagami, Takehiro Noda, Tadafumi Asaoka, Koichi Kawamoto, Kunihito Gotoh, Shogo + +Kobayashi, et al. Ct density in the pancreas is a promising imaging predictor for pancreatic ductal adenocarcinoma. Annals of surgical oncology, 24:2762-2769, 2017. 5 +[19] F Gaillard et al. Radiopaedia: building an online radiology resource. European Congress of Radiology-RANZCR ASM 2011, 2011. 3 +[20] Ibrahim Ethem Hamamci, Sezgin Er, Furkan Almas, Ayse Gulnihan Simsek, Sevval Nil Esirgun, Irem Dogan, Muhammed Furkan Dasdelen, Bastian Wittmann, Enis Simsar, Mehmet Simsar, et al. A foundation model utilizing chest volumes and radiology reports for supervised-level zero-shot detection of abnormalities. CoRR, 2024. 2, 4, 6, 12 +[21] Ibrahim Ethem Hamamci, Sezgin Er, and Bjoern Menze. Ct2rep: Automated radiology report generation for 3d medical imaging. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 476-486. Springer, 2024. 2, 4, 6, 12 +[22] 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. 3 +[23] Nicholas Heller, Sean McSweeney, Matthew Thomas Peterson, Sarah Peterson, Jack Rickman, Bethany Stai, Resha Tej-paul, Makinna Oestreich, Paul Blake, Joel Rosenberg, et al. An international challenge to use artificial intelligence to define the state-of-the-art in kidney and kidney tumor segmentation in ct imaging., 2020. 3 +[24] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 590-597, 2019. 7 +[25] 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. 4 +[26] Yuanfeng Ji, Hongliang Bai, Le Gu, Fucang Liu, Yingda Xia, Haofeng Lu, Zhennan Chen, Yunlong Gao, Hongyang Pan, Weilin Wang, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 111–120. Springer, 2022. 3 +[27] Erin S. Kilpatrick and colleagues. Thresholds for low and high kidney volume in adult men and women. National Center for Biotechnology Information, 2018. 16 +[28] Yoshihisa Kodama, Chaan S Ng, Tsung T Wu, Gregory D Ayers, Steven A Curley, Eddie K Abdalla, Jean Nicolas Vauthy, and Chusilp Charnsangavej. Comparison of ct methods for determining the fat content of the liver. American Journal of Roentgenology, 188(5):1307-1312, 2007. 5 +[29] Shigeru Kondoh and colleagues. Ct imaging characteristics of the pancreas: Normal and abnormal findings. PubMed, 2018. 16 + +[30] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, T 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. 3 +[31] Wenxuan Li, Chongyu Qu, Xiaoxi Chen, Pedro RAS Bassi, Yijia Shi, Yuxiang Lai, Qian Yu, Huimin Xue, Yixiong Chen, Xiaorui Lin, et al. Abdomenatlas: A large-scale, detailed-annotated, & multi-center dataset for efficient transfer learning and open algorithmic benchmarking. Medical Image Analysis, page 103285, 2024. 2, 3, 5, 17 +[32] Wenxuan Li, Pedro RAS Bassi, Tianyu Lin, Yu-Cheng Chou, Xinze Zhou, Yucheng Tang, Fabian Isensee, Kang Wang, Qi Chen, Xiaowei Xu, et al. Scalemai: Accelerating the development of trusted datasets and ai models. arXiv preprint arXiv:2501.03410, 2025. 5 +[33] Yuanhan Li, Jingyuan Chen, and Dinggang Shen. Knowledge-driven encode, knowledge-based decode: A novel pipeline for automatic chest x-ray report generation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 395-403, 2019. 4 +[34] Y. Li, X. Wang, and Z. Zhang. Exploring cross-modal transformers for chest x-ray report generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages —, 2022. 4 +[35] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out. Association for Computational Linguistics, 2004. 4, 13 +[36] Xiangde Luo, Wenjun Liao, Jianghong Xiao, Tao Song, Xiaofan Zhang, Kang Li, Guotai Wang, and Shaoting Zhang. Word: Revisiting organs segmentation in the whole abdominal region. arXiv preprint arXiv:2111.02403, 2021. 3 +[37] Jun Ma, Yao Zhang, Song Gu, Cheng Zhu, Cheng Ge, Yichi Zhang, Xingle An, Congcong Wang, Qiyuan Wang, Xin Liu, et al. Abdomenct-1k: Is abdominal organ segmentation a solved problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 3 +[38] Jun Ma, Yuxin Zhang, Yuankai Gu, Haozhe Zhang, Zong-hao Han, Jiancheng Dong, Yong Xie, Yefeng Zheng, and Yaowei Chen. Fast and low-gpu-memory abdominal organ segmentation from 3d ct images with nformer and swimunetr. In International Challenge on Abdominal Multi-Organ Segmentation from CT Images, FLARE 2022, pages 83–94. Springer, 2022. 3 +[39] A. McErlean and et al. Evaluation of tumor volume as a response metric in clinical trials and patient management. Journal of Surgical Oncology, 121(4):523-531, 2015. 8 +[40] A.B. Miller, B. Hoogstraten, M. Staquet, and A. Winkler. Reporting results of cancer treatment. Cancer, 47(1):207-214, 1981. 5, 7 +[41] Radiological Society of North America (RSNA). Radreport: Reporting templates. https://www.radreport.org, 2025. Accessed: 2025-01-06. 5 +[42] Blaine Rister, Darwin Yi, Kaushik Shivakumar, Tomomi Nobashi, and Daniel L Rubin. Ct-org, a new dataset for multiple organ segmentation in computed tomography. Scientific Data, 7(1):1-9, 2020. 3 + +[43] Holger R Roth, Le Lu, Amal Farag, Hoo-Chang Shin, Jiamin Liu, Evrim B Turkbey, and Ronald M Summers. Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation. In International conference on medical image computing and computer-assisted intervention, pages 556-564. Springer, 2015. 3 +[44] Laura Schöckel, Gregor Jost, Peter Seidensticker, Philipp Lengsfeld, Petra Palkowitsch, and Hubertus Pietsch. Developments in x-ray contrast media and the potential impact on computed tomography. Investigative radiology, 55(9):592-597, 2020. 2 +[45] J. Tang, P. Hu, and R. Qian. Multi-granular vision-language modeling for automated radiology reporting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages —, 2023. 4 +[46] Andrew Taylor, William Dodds, Sandra Erickson, and Edward Stewart. Ct of acquired abnormalities of the spleen. AJR American Journal of Roentgenology, 157(6):1213-1219, 1991. 16 +[47] Gianluca Tomasello, Michele Ghidini, Antonio Costanzo, Antonio Ghidini, Alessandro Russo, Sandro Barni, Rodolfo Passalacqua, and Fausto Petrelli. Outcome of head compared to body and tail pancreatic cancer: a systematic review and meta-analysis of 93 studies. Journal of gastrointestinal oncology, 10(2):259, 2019. 4 +[48] Alexandre Triay Bagur, Paul Aljabar, Gerard R Ridgway, Michael Brady, and Daniel P Bulte. Pancreas mri segmentation into head, body, and tail enables regional quantitative analysis of heterogeneous disease. Journal of Magnetic Resonance Imaging, 56(4):997-1008, 2022. 5, 8 +[49] Vanya V Valindria, Nick Pawlowski, Martin Rajchl, Ioannis Lavdas, Eric O Aboagye, Andrea G Rockall, Daniel Rueckert, and Ben Glocker. Multi-modal learning from unpaired images: Application to multi-organ segmentation in ct and mri. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 547-556. IEEE, 2018. 3 +[50] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 7 +[51] Xiaosong Wang, Yifan Peng, Lu Li, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. Tienet: Text-image embedding network for common thorax disease classification and reporting in chest x-rays. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9049-9058, 2018. 4 +[52] Mateusz Winder, Aleksander Jerzy Owczarek, Jerzy Chudek, Joanna Pilch-Kowalczyk, and Jan Baron. Are we overdoing it? changes in diagnostic imaging workload during the years 2010-2020 including the impact of the sarcoma-2 pandemic. In Healthcare, page 1557. MDPI, 2021. 2 +[53] H.C. Woodruff and et al. The role of tumor volume in assessing treatment response in cancer imaging. Cancer Imaging, 15(1):5-12, 2015. 8 + +[54] Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards generalist foundation model for radiology. arXiv preprint arXiv:2308.02463, 2023. 2, 4, 6 +[55] Yingda Xia, Qihang Yu, Linda Chu, Satomi Kawamoto, Seyoun Park, Fengze Liu, Jieneng Chen, Zhuotun Zhu, Bowen Li, Zongwei Zhou, et al. The felix project: Deep networks to detect pancreatic neoplasms. medRxiv, 2022. 4 +[56] Feiyang Yu, Mark Endo, Rayan Krishnan, Ian Pan, Andy Tsai, Eduardo Pontes Reis, Eduardo Kaiser Uurahy Nunes Fonseca, Henrique Min Ho Lee, Zahra Shakeri Hossein Abad, Andrew Y. Ng, Curtis P. Langlotz, Vasantha Kumar Venugopal, and Pranav Rajpurkar. Evaluating progress in automatic chest x-ray radiology report generation. medRxiv, 2022. 13 +[57] A. Zhang, B. Chen, and C. Li. Hierarchical perception for chest x-ray image captioning and radiology report generation. In European Conference on Computer Vision (ECCV), pages —, 2022. 4 +[58] Xukun Zhang, Sharib Ali, Tao Liu, Xiao Zhao, Zhiming Cui, Minghao Han, Shuwei Ma, Jingyi Zhu, Yanlan Kang, Le Wang, et al. Robust and smooth couinaud segmentation via anatomical structure-guided point-voxel network. Computers in Biology and Medicine, 182:109202, 2024. 5, 7, 8 +[59] Zongwei Zhou, Michael B Gotway, and Jianming Liang. Interpreting medical images. In Intelligent Systems in Medicine and Health, pages 343-371. Springer, 2022. 3 \ No newline at end of file diff --git a/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/images.zip b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..716a5bb122773bf2adeff2bef75b7701b5d1f0a6 --- /dev/null +++ b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:939f9ece544ec98866fa242ad03a64ed96158ac790c430a8e7cb53a73067fd9f +size 387007 diff --git a/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/layout.json b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3a0f5a44d27c10f400f559096dfbadd094882adf --- /dev/null +++ b/ICCV/2025/RadGPT_ Constructing 3D Image-Text Tumor Datasets/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:594d6ab544467cb1fb36b6753354e3987e7ccb7afb85a62840ae4c292a14ae44 +size 442883 diff --git a/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/0e2a74b3-6b72-46b1-9a54-5ca1e9e4663e_content_list.json b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/0e2a74b3-6b72-46b1-9a54-5ca1e9e4663e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..330b31f0651db078a2e394a750a053813462a667 --- /dev/null +++ b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/0e2a74b3-6b72-46b1-9a54-5ca1e9e4663e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79273a073c7e61b66483896207e867dbfe4c647da9f3675d99d343ecbbcbc6d5 +size 78739 diff --git a/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/0e2a74b3-6b72-46b1-9a54-5ca1e9e4663e_model.json b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/0e2a74b3-6b72-46b1-9a54-5ca1e9e4663e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9cfd621a294784f79601ccd53e3798349927d11f --- /dev/null +++ b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/0e2a74b3-6b72-46b1-9a54-5ca1e9e4663e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c18d644c572acdb5ef6d6d277c4e362e4fd24f447e97d465e19ef1b86e924ef1 +size 100944 diff --git a/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/0e2a74b3-6b72-46b1-9a54-5ca1e9e4663e_origin.pdf b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/0e2a74b3-6b72-46b1-9a54-5ca1e9e4663e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ae21f22f218238cd9d2d3edaa021bf8c91bbc288 --- /dev/null +++ b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/0e2a74b3-6b72-46b1-9a54-5ca1e9e4663e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:742f1758567836b535635d6c720a10cd432eca0a6f3c8b79a7c37dd87dcbf2e6 +size 23760313 diff --git a/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/full.md b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..564ca3b14ed46ab21953ba7c4a146ed562b63a0f --- /dev/null +++ b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/full.md @@ -0,0 +1,363 @@ +# RadarSplat: Radar Gaussian Splating for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes + +Pou-Chun Kung Skanda Harisha Ram Vasudevan Aline Eid Katherine A. Skinner University of Michigan + +{pckung, skandah, ramv, alineeid, kskin}@umich.edu + +![](images/01d671acc9eb3e679476391218002cd19a453b2e4dbce95c71333fa72c3a9218.jpg) +Radar 3D Reconstruction + +![](images/31d3a635f6d993a174ffa76c360d1022f81bd22826ffb2d68011262bf6a21162.jpg) +Radar Image Rendering +Figure 1. RadarSplat enables radar 2D-to-3D scene reconstruction, image synthesis, and occupancy estimation. RadarSplat outperforms the state-of-the-art neural rendering method [5] both qualitatively and quantitatively while also enabling additional radar inverse rendering. + +![](images/3e9a46425f90c6d47a7f3786e2dd4279781544715f33e98de4f664f0ded2f3ff.jpg) +Radar Inverse Rendering + +# Abstract + +High-fidelity 3D scene reconstruction plays a crucial role in autonomous driving by enabling novel data generation from existing datasets. This allows simulating safety-critical scenarios and augmenting training datasets without incurring further data collection costs. While recent advances in radiance fields have demonstrated promising results in 3D reconstruction and sensor data synthesis using cameras and LiDAR, their potential for radar remains largely unexplored. Radar is crucial for autonomous driving due to its robustness in adverse weather conditions like rain, fog, and snow, where optical sensors often struggle. Although the state-of-the-art radar-based neural representation shows promise for 3D driving scene reconstruction, it performs poorly in scenarios with significant radar noise, including receiver saturation and multipath reflection. Moreover, it is limited to synthesizing preprocessed, noise-excluded radar images, failing to address realistic radar data synthesis. To address these limitations, this paper proposes Radar-Splat, which integrates Gaussian Splatting with novel radar noise modeling to enable realistic radar data synthesis and enhanced 3D reconstruction. Compared to the state-of- + +the-art, RadarSplat achieves superior radar image synthesis $(+3.4\text{PSNR} / 2.6 \times \text{SSIM})$ and improved geometric reconstruction $(-40\% \text{RMSE} / 1.5 \times \text{Accuracy})$ , demonstrating its effectiveness in generating high-fidelity radar data and scene reconstruction. A project page is available at https://umautobots.github.io/radarsplat. + +# 1. Introduction + +Data-driven, learning-based methods have significantly advanced autonomous driving; however, acquiring suitable training data remains a substantial challenge. Real-world data collection to train models is time-consuming and prohibitively expensive, while developing realistic sensor simulations during real-world driving scenarios is hindered by the persistent simulation-to-reality gap. + +Recent advances in 3D scene reconstruction using neural radiance field (NeRF) and Gaussian Splatting (GS) enabled closed-loop simulation [57] and synthetic data generation [27, 33, 49], which are critical for enhancing data-driven autonomous driving systems. While radar is an essential sensor for modern autonomous vehicles, most existing 3D scene reconstruction and data synthesis methods + +focus exclusively on cameras and LiDAR [37, 49, 61], leaving radar's potential largely unexplored. Radar can enable weather-resilient SLAM [6, 17, 20], localization [1, 30, 39] and perception [31, 38, 46], and can boost performance for multimodal perception tasks [16, 23, 47, 50, 59]. + +Recently, DART [22] achieved realistic data synthesis of range-Doppler radar signals using a NeRF. However, dense range-Doppler images are uncommon in radar systems for autonomous driving. In contrast, Radar Fields [5] designed a NeRF-based approach for 3D reconstruction and novel view synthesis using scanning radar images, which are widely used in autonomous driving research [3, 7, 28, 29, 46]. While Radar Fields demonstrates encouraging results, due to the lack of noise modeling, it can only synthesize preprocessed, noise-excluded radar images, making realistic radar data synthesis still a challenging problem. + +This work proposes a GS-based framework for radar with noise detection and modeling. The proposed radar rendering formulation integrates azimuth and elevation antenna gain and accounts for spectral leakage to model the blurred radar range measurement. The radar noises are handled by view-dependent noise probability and periodic modeling. This enables more realistic radar image synthesis and improved 3D geometry estimation compared to [5], as shown in Figure 1. The key contributions of this paper are: + +- The first 3D Gaussian Splatting formulation for radar in autonomous driving. +- High-fidelity radar data synthesis that incorporates modeling of multipath effects and noises. +- Efficient radar noise detection and a denoising method that improves signal clarity and occupancy estimation. +- Radar inverse rendering that disentangles the real target, noise, and multipath effects from noisy radar images. + +# 2. Related Works + +# 2.1. Radiance Fields for Data Synthesis + +Radiance field methods have opened a new era of high-fidelity reconstruction and novel view synthesis using cameras. NeRF [36] first introduced an implicit neural scene representation and image rendering approach. This work was soon extended to autonomous driving scenes for data synthesis and closed-loop simulation [37, 53, 56, 57]. + +More recently, Gaussian Splitting (GS) [26] replaced the implicit neural network with explicit 3D Gaussians for scene representation, offering faster and more accurate reconstruction and data synthesis. GS has also been adapted for autonomous driving applications [10, 55, 61], facilitating the generation of high-quality autonomous driving data. + +# 2.2. LiDAR-Integrated Radiance Fields + +While cameras are the primary sensors used by radiance field methods, other sensors have also been incorporated + +to improve the synthesis of sensor data. Several studies have explored radiance fields with LiDAR/depth measurement [21, 24, 32, 48, 52, 54, 60]. Recent research on LiDAR-camera integration [8, 43, 49] has shown that using LiDAR measurements helps improve 3D scene reconstruction and leads to better image rendering. LiDAR-integrated GS methods have also been introduced [9, 12, 19, 25, 33, 58], combining the strengths of GS for more accurate scene reconstruction and data synthesis. + +# 2.3. Radar and Sonar-Integrated Data Synthesis + +In addition to cameras and LiDAR, researchers have explored integrating non-optical sensors such as radar and sonar into NeRF/GS. Unlike optical sensors that provide images or point clouds, radar and sonar produce range-power images, requiring redesigned sensor models for rendering. Sonar NeRF/GS methods [40, 42, 45] have been developed for 3D reconstruction in underwater environments. More recently, Z-Splat [41] introduced a camera-sonar fusion GS approach to enhance indoor scene reconstruction. + +On the other hand, most radar-based NeRF methods [13, 15, 34, 35] focus on 3D reconstruction using synthetic aperture radar (SAR). These methods require a specific sensor trajectory (usually linear or circle) to simulate a much larger antenna for high-resolution imaging. This limits their applicability to satellites or aircraft, making them unsuitable for automotive use. Recently, [51] proposed weather radar forecasting with a Gaussian representation, but their method approximates the radar model through camera image projection, and is not designed for scene reconstruction and data synthesis. DART [22] introduced a NeRF-based method using a TI MIMO radar, showing promising results for scene reconstruction and novel view synthesis in indoor, room-scale environments. However, it requires a range-Doppler image, which is not available from radars commonly used in autonomous driving. + +Radar Fields [5], the work closest to ours, proposes a NeRF-based method for scanning radar, achieving promising scene reconstruction and radar image rendering. However, it does not model the impact of radar noise, making the method unreliable in high-noise conditions, which are common in real-world autonomous driving. Additionally, its training data is denoised using a dynamic threshold, preventing the model from accurately simulating noise in synthetic radar images. This limitation reduces the accuracy of radar data synthesis. In contrast, we introduce noise decomposition and modeling to improve occupancy estimation and enable noise synthesis from novel views. + +# 2.4. Radar Data Simulation and Synthesis + +Simulating radar measurements has long been of interest. The most accurate method uses electromagnetic full-wave solvers, but iteratively solving Maxwell's equations is com + +![](images/d65a7c69394f209ae3e7cc1ca63243d67520af5c58bcbf7a7969ca3ab858217d.jpg) +Figure 2. System Overview. RadarSplat takes radar images and poses as input. The preprocessing step includes noise detection and initial occupancy mapping. The multipath source map and Gaussian splat reconstruct the 3D scene and model multipath effects for novel view synthesis. + +putationally expensive. To save time and resources, researchers have developed model-based simulations using physics-driven environmental models and ray tracing [2, 11, 18]. While these simulators capture complex radar phenomena such as occlusion, path loss, and multipath effects, they rely on manually defined scenes, limiting realism. Additionally, data-driven methods reconstruct real-world scenes from sensor data [14], but typically produce sparse radar points based on constant false alarm rate (CFAR) detection [44]. In contrast, the proposed method synthesizes dense, high-fidelity radar images without information loss. + +# 3. Methods + +This section first introduces the radar sensing equation and common noise types in radar images (Sec. 3.1). Figure 2 provides an overview of our pipeline. To account for radar noise, we propose a noise detection method (Sec. 3.2) and model multipath effects by identifying their sources, enabling recovery in novel view rendering (Sec. 3.3). Based on noise detection, we develop a radar denoising approach for noise-free occupancy mapping, serving as a supervisory signal for RadarSplat (Sec. 3.4). For scene reconstruction, we present a radar model that renders radar images from 3D Gaussians based on radar physics (Sec. 3.5). Finally, we define the RadarSplat training loss (Sec. 3.6). + +# 3.1. Radar Sensing Foundation + +# 3.1.1. Radar Equation + +A radar image is comprised of multiple range-power signals captured from different azimuth angles. In a range-power signal, the received power $P_{r}(n)$ at bin $n$ with range $R_{n}$ is determined by the radar equation [4]: + +$$ +P _ {r} (n) = \frac {P _ {t} G ^ {2} \lambda^ {2} \sigma}{(4 \pi) ^ {3} R _ {n} ^ {4} L}, \tag {1} +$$ + +![](images/9eaf4b8a2f06547beee9e1e5d4eb2b0c710120d351203852442d37ff6527a757.jpg) +Figure 3. Three types of radar noise of scanning radar highlighted in a raw radar image in polar space (bottom) and Cartesian space (top). The source and ghost reflections of multipath effects are highlighted with red and orange boxes. + +where $P_{t}$ is the radar peak transmit power, $G$ is the antenna gain, $\lambda$ is the wavelength, $L$ represents system and propagation loss, and $\sigma$ is the total radar cross-section of objects within the radar wave at range $R_{n}$ . + +# 3.1.2. Radar Noise Types + +Figure 3 illustrates three common radar noises. Multipath effects (red) occur when radar signals reflect off multiple surfaces before reaching the receiver, causing multiple echoes to arrive at different times. This results in periodic ghost targets appearing behind real targets. Receiver saturation (green) occurs when a strong signal return overwhelms the radar receiver, causing signal distortion and resulting in + +![](images/712528aa17a9aac19171761f40716dc6b3da2cf2e79191f143bc7a0a6164958e.jpg) +Figure 4. Range-power signal and its FFT of a radar azimuth beam with multipath effects. The constant and peak magnitude in the FFT results are used for noise detection and denoising. + +a uniform power offset in radar images. Speckle noise (yellow) is a granular interference pattern that appears in radar images due to the coherent nature of radar signals, appearing as weak background noise. + +# 3.2. Multipath and Saturation Noise Detection + +Unlike [5], which removes noise using a dynamic threshold that does not adapt to the radar equation, we instead detect and model radar noise to synthesize more realistic radar images. We adopt a noise detection method based on the Fast Fourier Transform (FFT), as in Figure 4. First, we apply FFT to all azimuth beams: + +$$ +X [ k ] = \mathcal {F} \{x [ n ] \} = \sum_ {n = 0} ^ {N - 1} x [ n ] e ^ {- j \frac {2 \pi}{N} k n}, \tag {2} +$$ + +where $x[n]$ represents the range-power signal from a row in a radar image, and $n$ is the range bin index with a total of $N$ bins. $X[k]$ is the frequency domain output from the discrete Fourier transform, with $k$ as the frequency index. + +Receiver Saturation. A beam experiencing receiver saturation typically exhibits a uniform radar power offset across all ranges. This uniform power distribution results in a constant term in the frequency domain. Based on this observation, we define a constant ratio value, $\mathcal{C}$ : + +$$ +\mathcal {C} = \frac {| X [ 0 ] |}{\sum_ {k = 1} ^ {N - 1} | X [ k ] |} \tag {3} +$$ + +Beams with $\mathcal{C} > \mathcal{C}_{th}$ are classified as saturation beams, $\Theta_{sat}$ , where $C_{th}$ is an experimentally determined magnitude threshold. + +Multipath Effects. Multipath noise appears as a fixed periodic pattern in a radar beam. This periodic power pattern contributes to a significant amplitude peak in the frequency domain (Figure 4). Based on this observation, we identify the significant frequency with index $k_{m}$ that has maximum magnitude. Beams with magnitude $|X[k_m]| >$ + +$A_{th}$ and a constant value ratio $\mathcal{C} > \mathcal{C}'_{th}$ are classified as multipath beams, $\Theta_{multi}$ . $A_{th}$ and $C_{th}'$ are experimentally determined magnitude thresholds. Note that since multipath typically occurs alongside saturation beams, we use a relaxed criteria $C'_th$ to reduce false positive multipath detection. The azimuth angles of detected noisy beams, $\Theta_{sat}$ and $\Theta_{multi}$ , are saved for later use in multipath modeling and occupancy mapping. + +# 3.3. Modeling Multipath Effects + +After detecting the multipath beams, we aim to model the noise source by estimating its position, reflectivity, and power attenuation, allowing us to simulate multipath effects from novel views. Figure 5 illustrates our proposed pipeline for noise source modeling. First, the source distance $d_{m}$ is determined from the peak frequency index, $k_{m}$ : + +$$ +f _ {m} = \frac {k _ {m}}{N \Delta r} \quad (1 / \mathrm {m}) \tag {4} +$$ + +$$ +d _ {m} = 1 / f _ {m} \quad (\mathrm {m}) \tag {5} +$$ + +where $\Delta r$ is range resolution. With known azimuth angle $\theta_{m}\in \Theta_{multi}$ , distance $d_m$ , and radar pose $\mathcal{T}$ , we can get the multipath source location, $\mu$ , activated view angle, $\theta_{view}$ , and range, $r_{view}$ . Therefore, a rough multipath signal reconstruction can be obtained by performing an inverse Fourier Transform with magnitude $|X[k_m]|$ and phase $\angle X[k_m]$ : + +$$ +x _ {\mathrm {m}} [ n ] = \frac {1}{N} | X [ k _ {m} ] | \cos \left(\frac {2 \pi}{N} k _ {m} n + \angle X [ k _ {m} ]\right) \tag {6} +$$ + +To further reconstruct multipath effects from novel views, we then define the power reflection, $A_{m}$ , and exponential attenuation rate, $\gamma_{m}$ , of the source, so that the multipath effect is represented as: + +$$ +x _ {m} ^ {\prime} [ n ] = A _ {m} e ^ {- \gamma_ {m} n} x _ {m} [ n ] \tag {7} +$$ + +$A_{m}$ and $\gamma_{m}$ can be estimated by fitting raw data, $x[n]$ , with least squares. Finally, the map of multipath sources is stored as: + +$$ +\mathcal {M} = \left\{\mathcal {M} _ {i}: \left(\mu^ {i}, \theta_ {\text {v i e w}} ^ {i}, r _ {\text {v i e w}} ^ {i}, A _ {m} ^ {i}, \gamma_ {m} ^ {i}\right) \mid i = 1, \dots , l \right\}. \tag {8} +$$ + +When rendering from novel viewpoints, $\mathcal{T}_{\text{novel}}$ , we compare the new view angle and distance differences, $\Delta r_{\text{view}}$ and $\Delta \theta_{\text{view}}$ , to determine if multipath effects are present. If $\Delta r_{\text{view}} < r_{\text{th}}$ and $\Delta \theta_{\text{view}} < \theta_{\text{th}}$ , we reconstruct multipath effects $x_m'[n]$ from the novel view using the updated $d_m$ and Eq. 4-7, where $r_{\text{th}}$ and $\theta_{\text{th}}$ are experimental thresholds. The multipath rendering is denoted as $\Psi$ , where the multipath image from a given view $\mathcal{T}$ is expressed as $I_{\mathcal{M}} = \Psi(\mathcal{M}, \mathcal{T})$ . + +![](images/4bec72dd1b1152ac65c71c00e924a2f8d6ae12139794c6ff3cccb24f1fad03bf.jpg) +Figure 5. Multipath modeling and denoising. The multipath effect is modeled by peak frequency and source power reflection and attenuation. The denoising method removes ghost detection of saturation and multipath effects, and the denoised image is used to construct an occupancy map to guide RadarSplat training. Note that the denoising is apply to both detected saturated and multipath beams. + +![](images/7a25ed5a1d8539ffddf617fa2fd886129d92ab654d576e7daa00bfafe802cdbb.jpg) +Figure 6. Our proposed radar image denoising method preserves rich information while remaining robust to multipath effects. In contrast, the dynamic threshold approach used in [5] struggles with multipath scenarios that include strong power returns. The remaining low-power speckle noise in the denoised image is later removed by binary occupancy mapping. + +# 3.4. Denoising and Occupancy Map Pre-processing + +Following Radar Fields [5], we use an initial estimated occupancy map to guide training, which helps to disentangle noise and real objects in radar images. We propose a denoising algorithm that removes noise across detected noisy azimuth angles, $\theta_{noise} \in \Theta_{sat} \cup \Theta_{multi}$ , identified in Sec. 3.2. The process involves Gaussian smoothing of the raw signal, identifying the distance with the maximum magnitude, and searching the decay region to generate a noise-free mask. The denoising process is illustrated in Figure 5, with the pseudo-code provided in the supplementary material. Figure 6 shows qualitative results of our proposed denoising method. Our method produces a clear denoised image, whereas Radar Fields struggles with multipath effects. The denoised image is then used to generate an initial occupancy map following [5]. The denoised image and occupancy map are shown in Figure 6. + +# 3.5. Radar Gaussian Splitting + +We represent the scene as a set of 3D Gaussians following [26]. The model $\mathcal{G}$ has $N$ Gaussians, and each Gaussian is composed of the mean, $\mu$ , rotation quaternion, $q$ , scaling vector, $S$ , and radar power return ratio, $\sigma$ , which is the radar cross-section (RCS) value when a Gaussian represents a real object rather than noise: + +$$ +\mathcal {G} = \left\{G _ {i}: \left(\mu_ {i}, q _ {i}, S _ {i}, \sigma_ {i}\right) | i = 1, \dots , N \right\}. \tag {9} +$$ + +Prior work [5] strictly defines the power return ratio as the RCS value, $\sigma_{i} = \alpha_{i}\cdot \rho_{i}$ , which fails to account for radar noise, where $\alpha_{i}$ represents occupancy probability and $\rho_{i}$ represents reflectance. Instead, we compute the power return ratio, $\sigma_{i}$ , by considering an additional noise probability term, $\eta_{i}$ : + +$$ +\sigma_ {i} = \rho_ {i} \cdot \min \left(\alpha_ {i} + \eta_ {i}, 1\right) \tag {10} +$$ + +To account for view dependency, the reflectivity is defined by spherical harmonics (SH) dependent on the view angle: $\rho_{i} = SH(\theta_{view})$ . Additionally, a regularization loss term is designed to make $\alpha_{i}$ and $\eta_{i}$ sum to one. + +Next, we introduce our rendering pipeline, which incorporates elevation and azimuth projection along with spectral leakage modeling. An overview illustration is provided in Figure 6 of the supplementary material. + +# 3.5.1. Rendering with Elevation Projection + +To render radar images in polar space, we first convert all Gaussians into spherical coordinates $(r_i,\theta_i,\phi_i) = F(x_i,y_i,z_i)$ , following [33]. To obtain the radar image, a 3D-to-2D Gaussian projection is applied to the $\phi$ -axis. The rasterization step used in [26] is simplified by accumulating the weighted power return ratio, $\sigma$ , for all Gaussians having the same $r$ and $\theta$ . $\sigma$ is weighted by the radar elevation antenna gain profile, $G_{\phi}$ . The rendering equation is defined as: + +$$ +P _ {r} (\theta , n) = \sum_ {\phi_ {i}} ^ {i} \frac {P _ {t} \cdot G \left(\phi_ {i}\right) ^ {2} \cdot \sigma_ {i}}{\left(4 \pi\right) ^ {3} R _ {n} ^ {4}} \tag {11} +$$ + +Note that the wavelength, $\lambda$ , and the general loss factor, $L$ , in Eq. 1 are ignored here since they are both constants. After the projection, we obtain an intermediate rendering, $I_{Elev}$ . + +# 3.5.2. Rendering with Azimuth Projection + +Radar typically has a wide beamwidth; in our experiments, the scanning radar covers $1.8^{\circ}$ in azimuth with a $0.9^{\circ}$ scanning resolution. To address this, we define a scaling factor, $Q$ , along the image height ( $\theta$ -axis). Given a ground truth image of size $H \times W$ , we render polar images with an + +![](images/912ef946149a3acb7e186ea8a0340b752d891a5158d8ada4926879f96e8691b8.jpg) +Figure 7. Modeling spectral leakage in the radar-power signal. (a) Ideal range FFT. (b) Practical range FFT with spectral leakage. (c) Practical range FFT sharpened by a Hamming window. (d) Proposed approximation for Hamming-window-sharpened FFT. + +intermediate azimuth resolution, $HQ$ , using elevation projection, ensuring $I_{Elev} \in \mathbb{R}^{HQ \times W}$ . Azimuth projection is then applied via a 1D convolution along the azimuth axis with kernel size $2Q$ and stride size $Q$ , and a kernel weighted by the azimuth antenna profile, $G_{\theta}(\theta)$ , producing the output image $I_{Azi} \in \mathbb{R}^{H \times W}$ . + +# 3.5.3. Spectral Leakage Modeling + +Radar range measurements are usually blurry due to spectral leakage. The spectral leakage is due to finite-time sampling of received signals. We use a Gaussian distribution to approximate the blurred effect, as shown in Figure 7. More details about the spectral leakage modeling are provided in supplemental materials. + +# 3.5.4. Final Radar Rendering + +Details Finally, the complete radar rendering process is denoted as $\Pi$ . The radar image is obtained by $\sigma$ -rendering, given by $I_{\sigma} = \Pi_{\sigma}(\mathcal{G},\mathcal{T})$ . The occupancy state is derived via $\alpha$ -rendering by replacing $\sigma$ in Eq. 11 with $\alpha$ , expressed as $I_{\alpha} = \Pi_{\alpha}(\mathcal{G},\mathcal{T})$ . The radar inverse rendering process decomposes the noise-free image $I_{target}$ and the noisy image $I_{noise}$ through $\rho \alpha$ -rendering, $\Pi_{\rho \alpha}$ , and $\rho \eta$ -rendering, $\Pi_{\rho \eta}$ . + +# 3.6. Training Losses + +The total loss to train the 3D Gaussian scene representation is as follows: + +$$ +\begin{array}{l} \mathcal {L} = \lambda_ {1} \mathcal {L} _ {l 1} + \lambda_ {2} \mathcal {L} _ {s s i m} + \lambda_ {3} \mathcal {L} _ {o c c} \tag {12} \\ + \lambda_ {4} \mathcal {L} _ {s i z e} + \lambda_ {5} \mathcal {L} _ {r e g} \\ \end{array} +$$ + +The $\mathcal{L}_{l1}$ and $\mathcal{L}_{ssim}$ components represent the L1 error and SSIM score between the rendered image $I_{render}$ and ground truth radar image $I_{gt}$ . Additionally, $\mathcal{L}_{occ}$ corresponds to the L1 error between the rendered occupancy state $I_{\alpha}$ output by RadarSplat and the initial occupancy map $I_{occ}$ estimated in the preprocessing step to aid in training. $\lambda_{i}$ are weights for different loss terms. + +To refine the model, we introduce two regularization losses. $\mathcal{L}_{size}$ constrains Gaussians to have a maximum size, $s_{max}$ , to prevent distortions from the first-order Taylor approximation in the Cartesian-to-Spherical transformation. + +In addition, $\mathcal{L}_{reg}$ is defined as $\mathcal{L}_{reg} = \mathrm{ReLU}(\alpha +\eta -1)$ to ensure that the sum of the occupancy and noise probability is smaller than one. + +# 4. Experiments + +# 4.1. Experimental Setup + +We train and test our proposed RadarSplat on the public Boreas Dataset [7], which includes a Navtech $360^{\circ}$ CIR304-H scanning radar, a 128-beam LiDAR, and ground truth poses from a GNSS/IMU sensor. Following [5], we select every 5 frames as the test frame to create a train-test split. We selected 13 driving sequences, comprising 7 sunny, 2 night, 2 rainy, and 2 snowy scenes. Each selected sequence has $>10$ seconds duration, which contains more than 40 radar frames. + +# 4.2. Novel Radar View Rendering + +We evaluate the synthetic radar image quality with common image rendering metrics, including PSNR, LPIPS, and SSIM. The quantitative evaluation is shown in Table 1. With the correct noise modeling and rendering, our proposed method outperforms state-of-the-art, Radar Fields, by $+3.4$ PSNR and achieves more than $2.6 \times$ better in SSIM score. Figure 8 shows the qualitative radar rendering results. Our method accurately renders noise and multipath effects, producing more realistic synthetic radar data. In contrast, Radar Fields fails to model the noise, resulting in noticeable performance degradation. We further demonstrate radar inverse rendering in Figure 9. RadarSplat can decompose radar images into different sources. The radar image rendering speed reaches 4.5 FPS on an NVIDIA A6000 GPU. + +# 4.3. Occupancy State Estimation + +While reconstructing noise in the scene, RadarSplat still supports accurate occupancy state estimation by decoupling the occupancy and noise probability from the power return ratio of the Gaussians. To assess the quality of occupancy estimation, we report the RMSE, Relative Chamfer Distance (R-CD), and Accuracy. Accuracy is computed using a $0.5\mathrm{m}$ threshold. Ground truth geometry is obtained from the LiDAR point cloud map, which is cropped according to + +
MethodImage SynthesisScene Reconstruction
PSNR↑SSIM↑LPIPS↓RMSE↓R-CD↓Acc.↑
Radar Fields22.660.200.603.030.290.59
Ours26.060.510.371.810.040.91
+ +Table 1. Image synthesis and geometry reconstruction evaluation on Boreas dataset [7]. Image synthesis is evaluated from unseen views, and geometry reconstruction is evaluated against the LiDAR map. Two snowy scenes are excluded from geometry evaluation due to LiDAR inaccuracy. Bold indicates the best result. + +![](images/1d2f713a312ff8e0142cfd256832d4c65ef335243ab5f02ba22689affa806588.jpg) +Figure 8. Radar image rendering. Our method better synthesizes multipath and noise effects compared to the baseline. + +![](images/edc2448779d9a244cfd4987b549930512fbb3f77b337601c54254218e92b77ff.jpg) +Figure 9. Radar inverse rendering for radar image decomposition. Our method can disentangle real target render, multipath effect, and other noises (mainly speckle noise and receiver saturation) from radar images. + +the radar field of view. The point cloud is then projected onto a 2D BEV point cloud for evaluation, following [5]. For the Radar Fields occupancy estimation, we follow the original implementation, integrating occupancy along the elevation axis. For RadarSplat, we use the rendered $I_{occ}$ as the final occupancy estimation. A 0.5 probability threshold is then applied to the output of both RadarSplat and Radar Fields to generate the final occupancy map for evaluation. + +The quantitative evaluation is presented in Table 1, alongside the qualitative comparison shown in Figure 10. The proposed method outperforms Radar Fields across all metrics, achieving better reconstruction by reducing RMSE by $1.22\mathrm{m}$ and improving accuracy more than $1.5\times$ compared to Radar Fields. We found that Radar Fields can be significantly affected by severe multipath effects, leading to ghost artifacts behind the wall. Also, it provides blurry occupancy estimations due to saturation and speckle noise. While Radar Fields mitigates most noise using a dynamic threshold during preprocessing and applies Bayesian grid mapping to generate the occupancy grid map, the simple dynamic threshold appears insufficient to effectively handle + +![](images/c24b53a5da5d7a8d5dce42878590e1c3b445fb5f7bf6bef15751e24a1cac40d9.jpg) +Figure 10. Occupancy estimation. Our method provides clear and noise-free occupancy estimation compared to the baseline. + +radar noise in challenging conditions. + +In addition, we provide qualitative 3D scene reconstruction results in Figure 11. The results indicate that Radar-Splat achieves accurate 3D reconstruction similar to Li-DAR, by taking only 2D noisy radar images as input. + +# 4.4. Ablation Studies + +Table 2 presents the results of our ablation study on image synthesis. We disable multipath modeling (Sec. 3.3) and noise probability (Sec. 3.5), which are key features we propose for more realistic rendering. We found that disentangling occupancy and noise probability has a significant impact on radar image synthesis, as speckle noise and receiver saturation frequently occur in radar images. While multipath modeling also improves rendering quality, its quantitative impact is less pronounced compared to noise probability due to the lower frequency of multipath effects. However, in frames with severe multipath interference, it provides substantial qualitative improvements (Figure 12). + +Table 3 presents the results of our ablation study on scene + +
Image SynthesisPSNR↑SSIM↑LPIPS↓
RadarSplatw/o Noise Prob.23.520.230.59
w/o Multipath25.950.500.38
Full Method26.060.510.37
+ +Table 2. Ablation studies on image synthesis. + +
Scene ReconstructionRMSE↓R-CD↓Acc.↑
Init Occ. MapRadar Fields3.400.120.90
Proposed1.810.040.90
RadarSplatw/o Occ. Map1.860.230.30
w/o Spectral Leakage2.050.050.90
Full Method1.810.040.91
+ +Table 3. Ablation studies on scene reconstruction. + +![](images/76f77d5536fa3dd8f15cd8d9bcd4310cf0e29f3d793a9241db2107e328ebde54.jpg) +Figure 11. Comparison of 3D reconstructions from our proposed method RadarSplat and Radar Fields [5]. Our method exhibits greater robustness to noise by leveraging multipath modeling and noise probability estimation, which Radar Fields struggles with. + +![](images/47c28eaeb208649355d77e0b68e5090577db2ec31dfba4fba255e4038cd24929.jpg) +Figure 12. Ablation studies on image synthesis. RadarSplat fails to model multipath effects when disabling the proposed multipath modeling. RadarSplat also fails to model other noises when disabling the proposed noise probability. + +reconstruction. In this study, we compare with results with and without spectral leakage modeling (Sec. 3.5.3) and occupancy map supervision. we also compare the quality of proposed and existing occupancy map. Our findings indicate that the proposed occupancy map significantly reduces RMSE by nearly $2\times$ , as occupancy mapping with a dynamic threshold and a Bayesian grid map alone is insufficient to handle severe radar noise. Also, the occupancy map supervision improves $3\times$ RadarSplat reconstruction accuracy. On the other hand, spectral leakage modeling also reduce about 0.2 RMSE. This is expected since the Gaussian variance $\sigma_w$ derived from the radar we use is about $0.17\mathrm{m}$ . + +# 4.5. Adverse Weather and Lighting Conditions + +To further demonstrate the robustness of RadarSplat under various weather conditions, Figure 13 illustrates novel view rendering and occupancy maps in snow, rain, and nighttime scenarios. In the snow scene, the LiDAR point cloud exhibits significant artifacts caused by snowfall. In the rain and night scenes, the camera is either blurred due to raindrops or has limited visibility due to low illumination. However, the radar data remains unaffected by these extreme conditions, and the proposed RadarSplat continues to provide robust occupancy map estimations. + +![](images/c6fb2c7860be804b05174d512170eca64863a2dca34af9c29ed591b91187d774.jpg) +Figure 13. Robust radar occupancy estimation in extreme conditions, where either the LiDAR or camera sensor is degraded. + +# 5. Conclusion + +We propose RadarSplat, a Gaussian Splitting method for realistic radar image rendering and accurate occupancy prediction. We take advantage of GS for better rendering quality and present a rendering process that takes radar's unique physical properties into account. Furthermore, we propose a novel noise detection and removal method. In addition, the different types of noise are modeled in a Gaussian primitive and multipath source maps. This enables radar inverse rendering for radar signal decomposition, high-fidelity radar data synthesis, and robust noise-free occupancy prediction. The proposed RadarSplat outperforms the state-of-the-art in both image synthesis and scene reconstruction in both qualitative and quantitative evaluation. We also demonstrate high-quality radar 3D reconstruction with similar performance to LiDAR. + +Limitations. The current method only works in static scenes. We plan to extend this work to dynamic scenes, similar to [55, 61], in the future. + +Acknowledgements. This work is supported by ChiaLun Lo Fellowship at University of Michigan, Ann Arbor. + +# References + +[1] Daniel Adolfsson, Martin Magnusson, Anas Alhashimi, Achim J Lilienthal, and Henrik Andreasson. Lidar-level localization with radar? the cfer approach to accurate, fast, and robust large-scale radar odometry in diverse environments. IEEE Transactions on robotics, 39(2):1476-1495, 2022. 2 +[2] Stefan Auer, Richard Bamler, and Peter Reinartz. Raysar-3d sar simulator: Now open source. In 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pages 6730–6733. IEEE, 2016. 3 +[3] Dan Barnes, Matthew Gadd, Paul Murcutt, Paul Newman, and Ingmar Posner. The oxford radar robotcar dataset: A radar extension to the oxford robotcar dataset. In 2020 IEEE international conference on robotics and automation (ICRA), pages 6433-6438. IEEE, 2020. 2 +[4] David Knox Barton. Radar equations for modern radar. Artech House, 2013. 3 +[5] David Borts, Erich Liang, Tim Broedermann, Andrea Ramazzina, Stefanie Walz, Edoardo Palladin, Jipeng Sun, David Brueggemann, Christos Sakaridis, Luc Van Gool, et al. Radar fields: Frequency-space neural scene representations for fmcw radar. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 1, 2, 4, 5, 6, 7, 8 +[6] Keenan Burnett, Yuchen Wu, David J Yoon, Angela P Schoellig, and Timothy D Barfoot. Are we ready for radar to replace lidar in all-weather mapping and localization? IEEE Robotics and Automation Letters, 7(4):10328-10335, 2022. 2 +[7] Keenan Burnett, David J Yoon, Yuchen Wu, Andrew Z Li, Haowei Zhang, Shichen Lu, Jingxing Qian, Wei-Kang Tseng, Andrew Lambert, Keith YK Leung, et al. Boreas: A multi-season autonomous driving dataset. The International Journal of Robotics Research, 42(1-2):33-42, 2023. 2, 6 +[8] Alexandra Carlson, Manikandasriram S Ramanagopal, Nathan Tseng, Matthew Johnson-Roberson, Ram Vasudevan, and Katherine A Skinner. Cloner: Camera-lidar fusion for occupancy grid-aided neural representations. IEEE Robotics and Automation Letters, 8(5):2812-2819, 2023. 2 +[9] Qifeng Chen, Sheng Yang, Sicong Du, Tao Tang, Peng Chen, and Yuchi Huo. Lidar-gs: Real-time lidar re-simulation using gaussian splatting. arXiv preprint arXiv:2410.05111, 2024. 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 preprint arXiv:2311.18561, 2023. 2 +[11] CJ Coleman. A ray tracing formulation and its application to some problems in over-the-horizon radar. Radio Science, 33 (4):1187-1197, 1998. 3 +[12] Jiadi Cui, Junming Cao, Yuhui Zhong, Liao Wang, Fuqiang Zhao, Penghao Wang, Yifan Chen, Zhipeng He, Lan Xu, Yu-jiao Shi, et al. Letsgo: Large-scale garage modeling and rendering via lidar-assisted gaussian primitives. arXiv preprint arXiv:2404.09748, 2024. 2 +[13] Junyuan Deng, Pengfei Xie, Lei Zhang, and Yulai Cong. Isar-nerf: Neural radiance fields for 3-d imaging of space + +target from multiview isar images. IEEE Sensors Journal, 24(7):11705-11722, 2024. 2 +[14] Christopher Doer and Gert F Trommer. Yaw aided radar inertial odometry using manhattan world assumptions. In 2021 28th Saint Petersburg International Conference on Integrated Navigation Systems (ICINS), pages 1-9. IEEE, 2021. 3 +[15] Thibaud Ehret, Roger Mari, Dawa Derksen, Nicolas Gasnier, and Gabriele Facciolo. Radar fields: An extension of radiance fields to sar. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 564-574, 2024. 2 +[16] Adam W Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. Simple-bev: What really matters for multi-sensor bev perception? In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2759-2765. IEEE, 2023. 2 +[17] Kyle Harlow, Hyesu Jang, Timothy D Barfoot, Ayoung Kim, and Christoffer Heckman. A new wave in robotics: Survey on recent mmwave radar applications in robotics. IEEE Transactions on Robotics, 2024. 2 +[18] Nils Hirsenkorn, Paul Subkowski, Timo Hanke, Alexander Schaermann, Andreas Rauch, Ralph Rasshofer, and Erwin Biebl. A ray launching approach for modeling an fmcw radar system. In 2017 18th International Radar Symposium (IRS), pages 1-10. IEEE, 2017. 3 +[19] Sheng Hong, Junjie He, Xinhu Zheng, Chunran Zheng, and Shaojie Shen. Liv-gaussmap: Lidar-inertial-visual fusion for real-time 3d radiance field map rendering. IEEE Robotics and Automation Letters, 2024. 2 +[20] Ziyang Hong, Yvan Petillot, Andrew Wallace, and Sen Wang. Radar slam: A robust slam system for all weather conditions. arXiv preprint arXiv:2104.05347, 2021. 2 +[21] Shengyu Huang, Zan Gojcic, Zian Wang, Francis Williams, Yoni Kasten, Sanja Fidler, Konrad Schindler, and Or Litany. Neural lidar fields for novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18236-18246, 2023. 2 +[22] Tianshu Huang, John Miller, Akarsh Prabhakara, Tao Jin, Tarana Laroia, Zico Kolter, and Anthony Rowe. Dart: Implicit doppler tomography for radar novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24118-24129, 2024. 2 +[23] Jyh-Jing Hwang, Henrik Kretzschmar, Joshua Manela, Sean Rafferty, Nicholas Armstrong-Crews, Tiffany Chen, and Dragomir Anguelov. Cramnet: Camera-radar fusion with ray-constrained cross-attention for robust 3d object detection. In European conference on computer vision, pages 388-405. Springer, 2022. 2 +[24] Seth Isaacson, Pou-Chun Kung, Mani Ramanagopal, Ram Vasudevan, and Katherine A Skinner. Loner: Lidar only neural representations for real-time slam. IEEE Robotics and Automation Letters, 2023. 2 +[25] Changjian Jiang, Ruilan Gao, Kele Shao, Yue Wang, Rong Xiong, and Yu Zhang. Li-gs: Gaussian splattering with lidar incorporated for accurate large-scale reconstruction. arXiv preprint arXiv:2409.12899, 2024. 2 + +[26] 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. 2, 5 +[27] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2407.02598, 2024. 1 +[28] Giseop Kim, Yeong Sang Park, Younghun Cho, Jinyong Jeong, and Ayoung Kim. Mulran: Multimodal range dataset for urban place recognition. 2 +[29] Hanjun Kim, Minwoo Jung, Chiyun Noh, Sangwoo Jung, Hyunho Song, Wooseong Yang, Hyesu Jang, and Ayoung Kim. Hercules: Heterogeneous radar dataset in complex urban environment for multi-session radar slam. arXiv preprint arXiv:2502.01946, 2025. 2 +[30] Pou-Chun Kung, Chieh-Chih Wang, and Wen-Chieh Lin. A normal distribution transform-based radar odometry designed for scanning and automotive radars. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 14417-14423. IEEE, 2021. 2 +[31] Pou-Chun Kung, Chieh-Chih Wang, and Wen-Chieh Lin. Radar occupancy prediction with lidar supervision while preserving long-range sensing and penetrating capabilities. IEEE Robotics and Automation Letters, 7(2):2637-2643, 2022. 2 +[32] Pou-Chun Kung, Seth Isaacson, Ram Vasudevan, and Katherine A Skinner. Sad-gs: Shape-aligned depth-supervised gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2842-2851, 2024. 2 +[33] Pou-Chun Kung, Xianling Zhang, Katherine A Skinner, and Nikita Jaipuria. Lihi-gs: Lidar-supervised gaussian splatt- ing for highway driving scene reconstruction. arXiv preprint arXiv:2412.15447, 2024. 1, 2, 5 +[34] Zhengxin Lei, Feng Xu, Jiangtao Wei, Feng Cai, Feng Wang, and Ya-Qiu Jin. Sar-nerf: Neural radiance fields for synthetic aperture radar multi-view representation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2 +[35] Afei Liu, Shuanghui Zhang, Chi Zhang, Shuaifeng Zhi, and Xiang Li. Ranerf: Neural 3-d reconstruction of space targets from isar image sequences. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 2 +[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. 2 +[37] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 2 +[38] Dong-Hee Paek, Seung-Hyun Kong, and Kevin Tirta Wijaya. K-radar: 4d radar object detection for autonomous driving in various weather conditions. Advances in Neural Information Processing Systems, 35:3819-3829, 2022. 2 +[39] Guohao Peng, Heshan Li, Yangyang Zhao, Jun Zhang, Zhenyu Wu, Pengyu Zheng, and Danwei Wang. Transloc4d: + +Transformer-based 4d radar place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17595-17605, 2024. 2 +[40] Mohamad Qadri, Michael Kaess, and Ioannis Gkioulekas. Neural implicit surface reconstruction using imaging sonar. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1040-1047. IEEE, 2023. 2 +[41] Ziyuan Qu, Omkar Vengurlekar, Mohamad Qadri, Kevin Zhang, Michael Kaess, Christopher Metzler, Suren Jayasuriya, and Adithya Pediredla. Z-splat: Z-axis gaussian splatting for camera-sonar fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[42] Albert Reed, Juhyeon Kim, Thomas Blanford, Adithya Pediredla, Daniel Brown, and Suren Jayasuriya. Neural volumetric reconstruction for coherent synthetic aperture sonar. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2 +[43] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12932-12942, 2022. 2 +[44] Hermann Rohling. Radar cfar thresholding in clutter and multiple target situations. IEEE transactions on aerospace and electronic systems, (4):608-621, 1983. 3 +[45] Advaith V Sethuraman, Max Rucker, Onur Bagoren, PouChun Kung, Nibarkavi NB Amutha, and Katherine A Skinner. Sonarsplat: Novel view synthesis of imaging sonar via gaussian splatting. arXiv preprint arXiv:2504.00159, 2025. 2 +[46] Marcel Sheeny, Emanuele De Pellegrin, Saptarshi Mukherjee, Alireza Ahrabian, Sen Wang, and Andrew Wallace. Radiate: A radar dataset for automotive perception in bad weather. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 1-7. IEEE, 2021. 2 +[47] Jingyu Song, Lingjun Zhao, and Katherine A Skinner. Lirafusion: Deep adaptive lidar-radar fusion for 3d object detection. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 18250-18257. IEEE, 2024. 2 +[48] Tang Tao, Longfei Gao, Guangrun Wang, Yixing Lao, Peng Chen, Hengshuang Zhao, Dayang Hao, Xiaodan Liang, Mathieu Salzmann, and Kaicheng Yu. Lidar-nerf: Novel lidar view synthesis via neural radiance fields. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 390–398, 2024. 2 +[49] Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895-14904, 2024. 1, 2 +[50] Yingjie Wang, Jiajun Deng, Yao Li, Jinshui Hu, Cong Liu, Yu Zhang, Jianmin Ji, Wanli Ouyang, and Yanyong Zhang. Bi-lrfusion: Bi-directional lidar-radar fusion for 3d dynamic object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13394-13403, 2023. 2 +[51] Ziye Wang, Yiran Qin, Lin Zeng, and Ruimao Zhang. High-dynamic radar sequence prediction for weather nowcasting + +using spatiotemporal coherent gaussian representation. arXiv preprint arXiv:2502.14895, 2025. 2 +[52] Hanfeng Wu, Xingxing Zuo, Stefan Leutenegger, Or Litany, Konrad Schindler, and Shengyu Huang. Dynamic lidar resimulation using compositional neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1998-19998, 2024. 2 +[53] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 2 +[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. 2 +[55] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2, 8 +[56] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 2 +[57] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 1, 2 +[58] Cheng Zhao, Su Sun, Ruoyu Wang, Yuliang Guo, Jun-Jun Wan, Zhou Huang, Xinyu Huang, Yingjie Victor Chen, and Liu Ren. TcIc-gs: Tightly coupled lidar-camera gaussian splatting for surrounding autonomous driving scenes. arXiv preprint arXiv:2404.02410, 2024. 2 +[59] Lingjun Zhao, Jingyu Song, and Katherine A Skinner. Crkd: Enhanced camera-radar object detection with cross-modality knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15470-15480, 2024. 2 +[60] Zehan Zheng, Fan Lu, Weiyi Xue, Guang Chen, and Changjun Jiang. Lidar4d: Dynamic neural fields for novel space-time view lidar synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5145-5154, 2024. 2 +[61] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 2, 8 \ No newline at end of file diff --git a/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/images.zip b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bf73f8dfaa7e940beb22dd04ae8d8360ea87532b --- /dev/null +++ b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:524a43fce3316879c06e4f9614809bd0c4b70e9f76f6eafb576de146b9d7482a +size 839637 diff --git a/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/layout.json b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..54b92fdf77d258daef6659bf4bb21c1f587b8c3f --- /dev/null +++ b/ICCV/2025/RadarSplat_ Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22eca7647f563a269b1d5f9f7074f42ddf29d348180387968a492966e01481af +size 453727 diff --git a/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/b4ba8099-5baa-4d40-b7b9-e018713360cf_content_list.json b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/b4ba8099-5baa-4d40-b7b9-e018713360cf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8e17662fe7f2f1742098cfba06da6153a98ac2f4 --- /dev/null +++ b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/b4ba8099-5baa-4d40-b7b9-e018713360cf_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc2f65c957aeb01ddcf2da7b7d1c005715da6eb1b9664950a24bb4cb929c4681 +size 81091 diff --git a/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/b4ba8099-5baa-4d40-b7b9-e018713360cf_model.json b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/b4ba8099-5baa-4d40-b7b9-e018713360cf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f8c834b7490007efa6a23d8d88b6dbda012c47f1 --- /dev/null +++ b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/b4ba8099-5baa-4d40-b7b9-e018713360cf_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48a2478e86c59d69c6d89fe72ac53d5ee01724e6b46dc58ae59e25f0c8f02d90 +size 99254 diff --git a/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/b4ba8099-5baa-4d40-b7b9-e018713360cf_origin.pdf b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/b4ba8099-5baa-4d40-b7b9-e018713360cf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fa3dd19b2f6ad0637801e712b21a81b0aa52a7e4 --- /dev/null +++ b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/b4ba8099-5baa-4d40-b7b9-e018713360cf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d3d800f40a24a8d8364b00af67e7ff65098be6d9ea0b82cfd55a9088208720e +size 17676520 diff --git a/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/full.md b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..025c3e8ac4d3943f0d636de8903dcaaa6b761a9e --- /dev/null +++ b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/full.md @@ -0,0 +1,314 @@ +# Radiant Foam: Real-Time Differentiable Ray Tracing + +Shrisudhan Govindarajan\*, Daniel Rebain\*, Kwang Moo Yi\*, Andrea Tagliasacchi\*1,3,4 + +1Simon Fraser University, 2University of British Columbia, 3University of Toronto, 4Google DeepMind, *equal contributions + +radfoam.github.io + +![](images/7cd4696300918639e30882b732e7413fb7afffee18d2bee8a5e6a024f8dc504b.jpg) +Figure 1. Teaser - We introduce the Radiant Foam differentiable 3D representation, which can be used to learn accurate radiance fields for any novel view synthesis applications (left). If we slice our foam along the plane highlighted by the red "laser", we expose (right) the internal structure of our representation: a polyhedral mesh that provides an injective parameterization of the 3D domain. Our representation is a foam, as the polyhedral cell structure is analogous to a closed-cell foam which partitions space into regions physically separated by thin, flat walls. It is radiant, as each foam bubble emits a view-dependent radiance that can be used to model the plenoptic function. + +![](images/a97b0b80d74b9c4c509cb6e4a9997adcbae98c4125f83d00805603485b009024.jpg) + +# Abstract + +Research on differentiable scene representations is consistently moving towards more efficient, real-time models. Recently, this has led to the popularization of splatting methods, which eschew the traditional ray-based rendering of radiance fields in favor of rasterization. This has yielded a significant improvement in rendering speeds due to the efficiency of rasterization algorithms and hardware, but has come at a cost: the approximations that make rasterization efficient also make implementation of light transport phenomena like reflection and refraction much more difficult. We propose a novel scene representation which avoids these approximations, but keeps the efficiency and reconstruction quality of splatting by leveraging a decades-old efficient volumetric mesh ray tracing algorithm which has been largely overlooked in recent computer vision research. The resulting model, which we name Radiant Foam, achieves rendering speed and quality comparable to Gaussian Splatting, without the constraints of rasterization. Unlike ray traced Gaussian models that use hardware ray tracing acceleration, our method requires no special hardware or APIs beyond the standard features of a programmable GPU. + +![](images/8ace5de3dec28e0a60c11f3d019edc4f6015ab36c1af6199b7e8c347f32820eb.jpg) +Figure 2. Radiant Foam - (left) In a stable foam, the pressure inside each bubble is roughly equal. The interfaces between bubbles settle into shapes that balance forces, leading to polygonal cells resembling Voronoi patterns (right). Our representation is nothing but a dense Voronoi tessellation of 3D space, where each point belongs to exactly one Voronoi cell. The position of Voronoi sites is differentiable, making it amenable to gradient-based optimization. + +![](images/4da65853dc29aac5a0cdbef645abfe713a76882fdc20c835d8ee4a207ae4a615.jpg) + +![](images/34e461b84da799f2f66aae573be2d6dc855b9e93035b910b79d738e6e555f90a.jpg) +Figure 3. Ray-based effects - Ray tracing simplifies the implementation of many effects which are difficult to approximate with rasterization. To motivate our work, we show here examples of integrating reflections, refractions, and non-linear camera models into our rendering pipeline. Each would be complicated to achieve with rasterization, but requires only minor modification to our rendering code. S + +![](images/d40dc05a18c1290bc945413cb7551ac2df0d6d9867d69c22d8532eb218d410b5.jpg) + +![](images/655e11bfa7f7b15fe879fc31769fa08a92542a77005817396f0d52afcac75ee8.jpg) + +# 1. Introduction + +Neural radiance fields (NeRF) have revolutionized 3D computer vision by allowing the extraction of dense 3D representations from collections of 2D images [29]. In their early days, radiance fields suffered from low training and testing/ rendering speed. Since then, researchers have developed more efficient models [31], as well as techniques for distilling models which enable highly efficient rendering/testing [5]. In particular, the real-time rendering performance was made possible by leveraging the rasterization pipeline readily available on GPU hardware. Soon thereafter, researchers proposed to also incorporate rasterization into the training process, leading to the development of the now wildly popular 3D Gaussian Splatting (3DGS) [19]. + +By comparing the development of radiance fields to that of classical graphics, one may argue that history is repeating itself: rasterization became ubiquitous in computer graphics because it is amenable to real-time rendering workloads. However, light effects that are trivial to implement in ray tracing (e.g. reflections, refractions, and transparency; see Fig. 3), are rather difficult to implement with rasterization. Consider the "Graphics Gems" series of books as a practical example, given how much of their content describes clever implementation tricks to express advanced light effects with rasterization engines. Nonetheless, since the introduction of dedicated hardware support (NVIDIA RTX) in 2018, real-time rendering engines have, at least partially, returned to simulate light transport via ray tracing. + +Recently, research on 3DGS representations has flourished, and much of this research is seemingly trying to fix issues caused by rasterization, such as the removal of 3DGS popping artifacts [38], or the introduction of more complex camera projection models [16]. Others integrated ray tracing with 3DGS, so to accelerate rendering and enable more complex light behavior [30]. But these developments have not discouraged researchers from investigating new, better, 3D representations. The community craves a return to polyg- + +onal meshes, as meshes are the unquestionable workhorse of modern computer graphics. And while we can find very interesting attempts at employing meshes for the modeling of radiance fields [11, 23, 49], as we will later discuss, none of these has realistically been able to oust 3DGS as the representation for learning radiance fields. + +Rather than revisiting history, and proposing clever engineering tricks that enable rasterization to work slightly better, we take a drastically different approach. In particular, we highlight how, over two decades ago, Weiler et al. [57] showed that fields represented by volumetric meshes admit a very efficient ray-tracing algorithm which requires no special hardware. This approach has been subsequently overlooked in the resurgence of volume rendering methods, and we hope to re-introduce volumetric mesh models which can benefit from it to the differentiable rendering community.3 In this paper, we make this representation differentiable, and carefully design refinement techniques for the underlying mesh. These refinements allow us to accurately represent the surface of objects, and yet render efficiently by skipping empty portions of volume. + +In a nutshell, our solution, which we name Radiant Foam (Fig. 2) provides 3DGS-like rendering speed and quality, but has a training modality based on rays that resembles the one from NeRF [29]. This implies that many NeRF techniques can now be seamlessly adapted to our method, with the significant advantage that the underlying geometry is explicitly represented by a volumetric mesh. We parameterize this volumetric mesh as a 3D Voronoi diagram (Fig. 1), which enables training of a mesh structure with dynamic connectivity through gradient descent by avoiding the discontinuities associated with discrete changes in other representations. We also propose a coarse-to-fine training approach, which enables rapid construction of mesh models with adaptive resolution. + +![](images/55954f9115df0bb6445f830e12aa78ac2fdcc81fc04c04ca775eba83341098ab.jpg) +Figure 4. Ray tracing foams - When a ray (red) enters a cell, we iterate through all the planar cell faces to identify the face through which the ray exits. This exit intersection is the first intersection along the ray with a normal vector less than 90 degrees from the ray direction (green); other intersections are considered back-facing (blue) and ignored. As the faces each correspond to a neighboring cell, the tracing then proceeds by stepping into the cell associated with the exit intersection and repeating the process. + +# 2. Related Work + +Neural Radiance Fields (NeRFs) [29] represent 3D scenes as volumetric radiance fields encoded within coordinate-based neural networks. This representation allows querying the network at any spatial location to retrieve volumetric density and view-dependent color, facilitating the generation of photo-realistic novel views. The success of NeRFs has led to numerous follow-up works. For instance, significant effort has been devoted to enhancing training speed [31, 41], quality [1, 2], and to extract surfaces from the representation [54, 59]. Finally, several works investigated ways to speed up the inference by baking the neural fields to more performant representations [5, 8, 42, 48, 55]. While achieving high quality and fast rendering speeds, these methods often employ multi-stage training procedures. + +Point-based rasterization. Point-based rendering with primitives such as circles, spheres, or ellipsoids [36, 44, 61] laid the foundations for point-based rasterization. Differentiable rendering using depth-based blending has also been extended to volumetric particles in Pulsar [24], which employs sphere-based differentiable rasterization to render scenes with millions of spheres in real time. While Pulsar provides a differentiable representation, Kerbl et al. [19] provides more expressive primitives in the form of soft, anistropic Gaussians, which can be differentiably optimized to fit the scene content. This approach has inspired several follow-up works aimed at reducing its reliance on strong initialization [21], reducing rendering time and memory footprints [9, 32, 35], enabling the reconstruction of surfaces [12, 15], and improving their ability to be trained for large spatial extents [20, 43]. + +Extensions of 3DGS. While significant progress has been + +made, these approaches still inherit the inherent limitations of rasterization. They struggle to handle camera distortions, model secondary lighting effects, or simulate sensor-specific properties like rolling shutter or motion blur. Recent efforts have aimed to address these challenges. For example Niemeyer et al. [33] distilled a Zip-NeRF [3] into a 3DGS so to model distorted cameras and rolling shutter effects. To capture secondary lighting effects, recent works have explored incorporating occlusion information into spherical harmonics for each Gaussian [10, 25], or leveraging shading models and environment maps [17]. These methods either rely on rasterization during inference [10], or during training [17, 25] hence inheriting the limitations of rasterization. For complex lens effects, Seiskari et al. [47] modeled motion blur and rolling shutter by approximating them in screen space through rasterization and pixel velocities. Our approach introduces a principled framework for efficient ray tracing of volumetric primitives, overcoming the aforementioned limitations and enabling the simulation of effects like reflection, refraction, and camera distortion. + +# 2.1. Ray tracing of volumetric primitives + +Ray tracing has been a cornerstone of photo-realistic rendering since its introduction [58]. It enables the accurate simulation of light interactions with scene geometry, making it indispensable for applications requiring effects such as shadows, reflections, and refractions. Modern advancements in hardware have further accelerated ray tracing, allowing for real-time applications [18]. Recently, Moenne-Loccoz et al. [30] and Condor et al. [6] integrated the NVIDIA OptiX ray tracer with 3DGS [19] for fast ray tracing, while Mai et al. [26] employed it with constant ellipsoids. While these methods perform real-time ray tracing, they suffer from the tendency of 3DGS [19] to produce overlapping primitives which degrade the quality of the hierarchical acceleration structure and increase the number of intersections per ray. Our method represents the scene using a non-overlapping polyhedral mesh without the need for a secondary acceleration structure, thereby efficiently avoiding these limitations. + +# 2.2. Delaunay triangulation and Voronoi diagrams + +Voronoi diagrams have been used for representing information in various fields. In computer graphics, Delaunay triangulations [7] and Voronoi diagrams [53] have been extensively studied for meshing (surfaces and volumes), spatial partitioning [13, 50, 56], and differentiable simulation [34]. Recently, they have also found novel applications in the realm of differentiable rendering. DMTet [49] introduced a deformable tetrahedral grid, generated using Delaunay triangulation, as an underlying 3D representation, while Tet-Splatting [11] implements a differentiable rasterizer for this representation. Similarly, DMesh [51] proposes a differential representation to model Delaunay triangulations + +![](images/e0fb0d78781616c3f514f7cf752571e98c5848c7b827e37e2d17965a096eea82.jpg) +Figure 5. Delaunay and its dual Voronoi - (left) Given a set of points (red) in $\mathbb{R}^N$ , we can find circumspheres (blue) which each pass through $N + 1$ points. (center) The set of all circumspheres which contain no points on their interior defines the Delaunay triangulation, where the $N + 1$ points tangent to each circumsphere form a simplex. (right) These circumspheres also describe the Delaunay triangulation's dual, the Voronoi diagram: the centers of circumspheres tangent to a point become the vertices of the Voronoi cell containing that point. + +![](images/50994da4ea18462111e1920598bc30e19294deedcf30061411f2a8b68092d988.jpg) + +![](images/54e1df381877be75058c3f612cf8d496722598587a39bf342b154edc7d7dc988.jpg) + +to optimize for mesh connectivity. DeRF [39] leverages Voronoi diagrams to spatially decompose a scene, resulting in faster training and inference, while Tetra-NeRF [23] employs a tetrahedral Delaunay mesh to model scenes more effectively. While these methods inherit the ray-marching approach of [29] for volume rendering, they share a common limitation: slow rendering speeds caused by many MLP evaluations required per ray. In contrast, with our method we propose to optimize the mesh topology generated by the Voronoi diagram as a differentiable polygonal mesh. Our efficient ray tracing of volumetric particles also significantly accelerates the rendering speed. + +# 3. Method + +Our method addresses the now-familiar problem of constructing representations of scenes from image collections. As with NeRF [29] and its numerous successors, we achieve this reconstruction by gradient-based optimization of a differentiable scene representation. In the following sections we propose a volumetric mesh-based differentiable representation, and explain how we are able to effectively optimize it from image supervision. + +# 3.1. Volume rendering + +Volume rendering has become the workhorse of modern differentiable scene reconstruction methods. Volume rendering allows all points in space to make a continuously varying contribution to the observed color of viewing rays which pass through those points. The effect of this contribution is controlled by a density field, which creates occlusions, and a radiance field, which determines the brightness of light observed. Unlike many alternative rendering formulations, volume rendering of continuous fields is fully continuous with respect to all degrees of freedom, including both the + +viewpoint, as well as the density and radiance field values. This property makes it very amenable to gradient-based optimization. Volume rendering is defined by an integral over the segment $(t_{\mathrm{min}}, t_{\mathrm{max}})$ of a viewing ray. Specifically the observed color $\mathbf{c}_{\mathbf{r}}$ for a ray $\mathbf{r}$ is: + +$$ +\mathbf {c} _ {\mathbf {r}} = \int_ {t _ {\min }} ^ {t _ {\max }} T (t) \cdot \sigma (\mathbf {r} (t)) \cdot \mathbf {c} (\mathbf {r} (t)) d t, \tag {1} +$$ + +$$ +T (t) = \exp \left(- \int_ {t _ {\min }} ^ {t} \sigma (\mathbf {r} (u)) d u\right), \tag {2} +$$ + +where $\mathbf{r}(t)$ denotes the point in 3D space at distance $t$ along ray $\mathbf{r}$ , and $\sigma(\cdot)$ and $\mathbf{c}(\cdot)$ denote the density and radiance fields respectively; see [52] for more details. + +Piecewise constant volumes. In the case of piecewise constant $\sigma(\cdot)$ and $\mathbf{c}(\cdot)$ , the integral can be expressed as a sum over all $N$ ray segments with constant field values: + +$$ +\mathbf {c} _ {\mathbf {r}} = \sum_ {n = 1} ^ {N} T _ {n} \cdot \left(1 - \exp \left(- \sigma_ {n} \delta_ {n}\right)\right) \cdot \mathbf {c} _ {n} d t, \tag {3} +$$ + +$$ +T _ {n} = \prod_ {j = 1} ^ {n} \exp (- \sigma_ {j} \delta_ {j}), \tag {4} +$$ + +where $\delta_{n}$ is the width of segment $n$ . This formulation admits a simple implementation as a loop over the segments in order of depth. NeRF-based methods typically use Eq. (3) as an approximation to Eq. (1), with segments sampled according to some importance sampling scheme, e.g. see [2]. + +Conversely, in our model these forms are exactly equivalent. We propose to leverage the algorithm by Weiler et al. [57] along with a model of constant field values within the cells of a volumetric mesh (see Fig. 4), arriving at a representation for which Eq. (3) gives the exact volume rendering + +![](images/c604c285817bf1861ba6a759f0d499a1541072f3f3f225fc82789a45cfe813d7.jpg) +Figure 6. Edge flips - The connectivity of the Delaunay graph (green) is sensitive to small positional perturbations, leading to "edge flips" in the triangulation. These discrete changes occur at configurations where the circumspheres (blue) of two neighbouring simplices become identical. In the dual Voronoi diagram, this configuration also corresponds to a discrete change, but the cell boundary which changes has an area of zero at the moment of the flip (center). Consequently, while there are still discrete changes in the adjacency structure of the Voronoi diagram, the shapes of the cells vary continuously with the positions of the points (red), which enables gradient-based optimization. + +![](images/3ad1552dc122eef38ba40a38bc876ac731eaa5e8d837263d43cf917b4fbe57ff.jpg) + +![](images/8b940f0818ccdadd587322de1680d3ad86d488ccb560ffaf984ab9ec572c2b2e.jpg) + +result. In the volume rendering literature, this is also known as regular tracking [37, Chapter 15.2.2]. While this choice is highly advantageous in avoiding any complicated or expensive sampling schemes, we must take great care to not interfere with the continuity of the representation, which is critical for gradient-based optimizers. In particular, the volumetric mesh itself receives gradients only through the segment widths $\delta_{n}$ , which are determined by the locations of ray intersections with cell boundaries. It is therefore critical that these boundary intersections vary continuously with the estimizable parameters of the model. + +# 3.2. Differentiable mesh representation + +Constructing mesh representations of volumes or surfaces through gradient-based optimization is hardly a new problem, and many methods tackle it using a variety of strategies; see Shen et al. [49] and citations therein. Our method, however, encounters and must solve a challenge that most mesh optimization methods either avoid, or solve with discrete optimization techniques. Generally, meshes are determined by degrees of freedom in two groups: vertex locations $\mathcal{V}$ and cells $\mathcal{C}$ (i.e. connectivity, or mesh topology). Assuming all other parameters are fixed, optimizing for vertex locations is typically straightforward, as the intersections between rays and cell boundaries vary smoothly with vertex locations. The true challenge arises if one desires to also optimize the connectivity $\mathcal{C}$ of the representation. + +The Delaunay triangulation. Because mesh connectivity is inherently discrete, we can not optimize it directly with gradient-based methods. To avoid this problem, we can define connectivity in terms of the vertex locations, such that each possible configuration of vertices corresponds to a unique set of cells. The most obvious choice for this mapping is the Delaunay triangulation [7], which in 3D + +comprises the set of all tetrahedral cells $\mathbf{c}_i\in \mathcal{C}$ which may be formed from four vertices such that their circumspheres contain no additional vertices (the so-called “Delaunay criterion”). This construction is unique for point sets in general position (those lacking groups of 5 or more co-spherical points), and is easily computed using well-known and efficient algorithms; see Section 3.4. This strategy encounters two significant problems if we wish to use it as a basis for differentiable volume rendering: (1) the discrete nature of the mesh connectivity is not entirely avoided, as the boundaries of Delaunay cells undergo discrete “flips” whenever a vertex enters the circumsphere of another cell. These flips introduce discontinuities into the optimization landscape, which interfere with the convergence of gradient descent; (2) the number of tetrahedra in this model is not fixed in this model, and therefore it is not straightforward to associate each cell with estimizable $\sigma$ and $\mathbf{c}$ values as required for volume rendering. The latter issue could be solved by associating field values with vertices (vs. tets), and interpolating those values within the cell, but this complicates volume rendering. + +The Voronoi diagram. Rather than utilize the Delaunay mesh directly and suffer its limitations, we instead look to the dual graph of the Delaunay triangulation: the Voronoi diagram [53]. As shown in Figure 5, this is constructed by placing a vertex at the circumcenter of each Delaunay tetrahedron; it partitions space into convex polyhedral cells $\mathbf{c}_i \in \mathcal{C}$ consisting of points which share a nearest neighbour among the primal vertices $^4 \mathbf{p}_i$ of the Delaunay triangulation: + +$$ +\mathbf {c} _ {i} = \left\{x \in \mathbb {R} ^ {3}: \underset {j} {\arg \min } | | x - \mathbf {p} _ {j} | | = i \right\}. \tag {5} +$$ + +This may at first seem like a strange choice... after all, if the + +Delaunay Triangulation is unavoidably discontinuous in optimization, will not its dual also suffer this issue? The answer to this question lies in the fact that any discrete change in the connectivity of the Voronoi Diagram exactly coincides with the point that the affected cell faces attain zero surface area. As shown in Figure 6, the discrete flips are effectively hidden within these zero-volume regions of space, and the resulting field representation remains completely continuous with respect to the primal vertex locations. For a rigorous treatment of this property of the Voronoi diagram, see the proof by Reem [40] that Voronoi cells in $\mathbb{R}^d$ are stable with respect to primal vertex positions. The number of cells in the Voronoi diagram is also constant regardless of configuration, which makes the association of $\sigma$ and $\mathbf{c}$ values to cells... trivial. The resulting model is then effectively a learnable point cloud, not dissimilar to the formulation of 3D Gaussian Splatting [19], though lacking the per-point covariance matrix. The only remaining issue is that our ray tracing algorithm [27] expects tetrahedral cells. We therefore modify the algorithm to handle the more general case of convex cells. This modified tracing process pre-fetched the neighboring vertices of each vertex, thereby allowing an efficient iteration over the cell faces to find ray intersections. For more detail on this algorithm, see the supplementary material. + +# 3.3. Optimization + +Similarly to Kerbl et al. [19], the (mostly) local nature of Voronoi cells renders the optimization landscape more prone to local minima. We follow a similar strategy, by first carefully initializing the optimization, and then adaptively densifying and pruning Voronoi sites. Additionally, to promote the formation of surface-like densities, we also employ a regularization objective similar to the distortion loss [2] commonly used by NeRF methods. + +Densification. Similarly to Kerbl et al. [19], to initialize training we start with a sparse point cloud obtained from COLMAP [46]. Over training, we perform densification and pruning operations to control the number of Voronoi sites and their density, allowing the model to adaptively re-allocate representational capacity to areas of space with more geometric and/or photometric detail. We observe that gradients of the reconstruction loss with respect to Voronoi site locations can be used to identify cells which are underfitting the training signal. We therefore use the norm of this gradient multiplied by the approximate radius of the cell as a measure of which cells require further densification. Inspired by Kheradmand et al. [21], we select the candidates for densification by sampling a Multinomial distribution with probability mass function proportional to this measure. + +Pruning. Towards building a parsimonious representation, we remove cells from the Voronoi diagram that do not con + +tribute to rendering. However, it is not sufficient to simply delete "empty" cells (i.e. zero density), as the geometry of Voronoi cells is determined by the positions of adjacent sites, even when the density in those cells is zero. To accurately represent object boundaries, it is therefore essential to retain cells with near-zero density that define the boundary. For this reason, our pruning strategy removes Voronoi sites that have very low density, and are surrounded by neighbors with very low density. This pruning ensures that we eliminate sites that neither contribute to nor define the surfaces, thereby maintaining the accuracy of the object boundaries. + +Training objectives. Similarly to the distortion loss of MipNeRF 360 [2], we apply a regularization on the distribution of contribution to the volume rendering integral along the ray. This additional loss function encourages the density to concentrate at surfaces and reduces visible "floater" artifacts. This objective is computed as + +$$ +\mathcal {L} _ {\text {q u a n t i l e}} = \mathbb {E} _ {t _ {1}, t _ {2} \sim \mathcal {U} [ 0, 1 ]} \left[ \left| W ^ {- 1} \left(t _ {1}\right) - W ^ {- 1} \left(t _ {2}\right) \right| \right], \tag {6} +$$ + +where $W^{-1}(\cdot)$ denotes the quantile function (inverse CDF) of the volume rendering weight distribution along the ray. This form has the same effect as the distortion loss [2], but avoids the need for a quadratic nested sum which would increase the computational cost and memory footprint of training. Denoting with $\mathcal{L}_{\mathrm{rgba}}$ the typical L2 photometric reconstruction loss (including ground truth opacity equal to one for all rays), our overall training objective is + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {r g b a}} + \lambda \mathcal {L} _ {\text {q u a n t i l e}} \tag {7} +$$ + +# 3.4. Implementation details + +We implemented our method in Python and $\mathrm{C + + }$ within the PyTorch framework, with custom CUDA kernels for ray tracing, Delaunay triangulation, and other operations requiring high efficiency. This implementation includes an interactive viewer and a (very) low overhead renderer, which we used to measure frame rates. Importantly, nothing in our implementation is dependent on dedicated ray tracing hardware, or the OptiX library, which is required by methods like [30]. Therefore, with some engineering effort, our entire rendering loop could easily be implemented in a portable rendering framework like WebGL. For more details on training and Voronoi optimization please refer to Section 7 in the supplementary material. + +# 4. Experiments + +We evaluated our algorithm on a total of 9 real-world scenes sourced from two publicly available datasets. Specifically, we utilized the complete set of scenes from the Mip-NeRF 360 dataset [2] except for two private scenes (flowers and treehill) and two scenes from the Deep Blending dataset [14]. These datasets contain scenes with a diverse range of capture styles, including bounded indoor scenes and expansive, + +
MipNeRF 360 [2]Deep Blending [14]
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
Rasterization
3DGS* [19]28.690.870.2229329.410.900.32319
Mip-Splating [60]29.390.880.2024129.470.900.32260
3DGS-MCMC [21]29.720.890.1930229.710.900.32662
Ray Tracing / Marching
Plenoxels [45]23.630.670.44<3023.060.800.51<30
iNGP-Big [31]26.750.750.30<3024.960.820.39<30
MipNeRF360 [2]29.230.840.21<129.400.900.25<1
3DGRT** [30]28.710.850.257829.230.900.32119
RadiantFoam (v1)28.470.830.2120028.950.890.26301
+ +Table 1. Novel view synthesis - We evaluate our method's accuracy in reconstructing held-out views for two standard datasets. Our method has similar performance to 3DGS and 3DGSRT, while providing significantly higher frame rates than the latter. FPS is measured as an average on the test set for each scene at their native resolutions. *For 3DGS, we report corrected LPIPS scores provided by Bulò et al. [4] **Note that 3DGRT FPS results were measured with an RTX 6000 Ada GPU. We report the results from the original publication, as the code is not open source. + +unbounded outdoor environments. For Mip-NeRF 360, to make our results compatible with [19, 30], we downsample images for the indoor scenes by a factor of two, and the outdoor scenes by four. For Deep Blending scenes, we use the original image resolutions. All frame rates were measured on a consumer-grade RTX 4090 GPU. + +Metrics. We assess each method using three widely recognized image quality metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). In the supplementary web page we include rendered video paths for selected scenes, showcasing views significantly different from the input images. + +Quantitative results. We report our quantitative results for the Mip-NeRF 360 [2] and Deep Blending [14] in Table 1. In terms of quality, our method achieves results comparable to, or slightly below, those of 3DGS [19] and 3DGRT [30] (the state of the art differentiable ray tracing method). However, our method excels in rendering speed. As shown in Table 1, our efficient ray-tracing implementation achieves in some cases over 300 FPS, more than twice as fast as 3DGRT (119 FPS), while maintaining a similar rendering speed to rasterization methods. + +Qualitative results. We include in the supplementary material additional qualitative comparison showing the differences between our reconstruction and that of 3D Gaussian splatting. These results highlight how 3DGS tends to over-smooth some details and miss some view dependence, but also more gracefully extrapolates to under-observed regions. + +
SfMDensifyPruneQuantilePoint optBonsaiGardenPlayroomMean
X✓✓✓✓29.6525.8326.3427.00
XX✓✓✓20.2318.8819.5519.36
✓✓X✓✓32.2526.5829.4629.15
✓✓✓X✓29.6225.3529.5927.90
✓✓✓✓X30.5125.0328.3527.96
✓✓✓✓✓32.1526.5829.5929.15
+ +Table 2. Ablation table - We evaluate the impact of various components in our method by systematically excluding them and analyzing the reconstruction quality (PSNR $\uparrow$ ) on the Bonsai and Garden scenes from MipNeRF 360 [2], as well as the Playroom scene from Deep Blending [14]. + +# 4.1. Ablation + +We isolated the different contributions and choices we made and constructed a set of experiments to measure their effect. Specifically, we test the following aspects of our algorithm: initialization from SfM, our densification and pruning strategies, and regularization loss. The quantitative effect of each choice is summarized in Table 2. + +Initialization from SFM. We assess the importance of initializing our Voronoi sites from the SfM point cloud. For this, we initialize our representation with $2^{17}$ points from a normal distribution with a standard deviation of 25. We observe that our method performs relatively well even without the SfM points. Instead, it degrades mainly in the background and tends to have more floaters in regions that are sparsely covered in the training views, see Figure 7. + +Densification and pruning. Due to numerical approximations, our triangulation algorithm can fail when processing very close (or identical) points. This limitation prevents us from initializing our representation directly with the final set of points obtained by duplicating and perturbing the Structure from Motion (SfM) points, as SfM typically produces numerous closely spaced points. Consequently, for the ablation study on densification, we utilize the random initialization strategy described earlier to initialize our representation with the intended number of points. We observe that our method significantly underperforms without densification, resulting in an under-represented scene where resources are not adequately allocated to regions with complex geometry or texture. In the ablation study of our pruning strategy, we cease pruning those Voronoi sites that are neither surface nor boundary points. We notice that this pruning approach has minimal impact on the quality of the renderings because the number of pruneable points is very low. By starting with a sparse point set and progressively densifying only in under-represented areas, we ensure that points are allocated exclusively to necessary regions, resulting in few pruneable points. + +![](images/90dd4b3bf46478698f830a5d45626b63faa16532639909833126aba5cc7fcd5e.jpg) + +![](images/4c9766ae44c718ea66f915f056c1d7a358fc633f52f6182b8a837f3142057753.jpg) + +![](images/d020d141cd118a1e9e768871d7ff94ce22ff8d83215222ebe451850a52823ff6.jpg) + +![](images/f7ae37bc7e1dbb7877fe20ff481aa252047df3a74019a84697bb5b757fbed15d.jpg) + +![](images/be25665342cc317096bad21732e672832c59e7340a018b712fcf8215f325703a.jpg) + +![](images/24d66f5ba99426e81f071f1c9ccd81cfc3f6465196c6c8563c5bdf3fe07e4f52.jpg) + +![](images/31e0d8e3d1275b94c8748630bc9994fe9dbb2023e9bec9b0149d453a2b0b90ff.jpg) + +![](images/dbac236e8fce713eb709c1b47bfe1ea49cd7cf631af8b299cc6c6a116d58afe6.jpg) + +![](images/f2c2c4ae4827506c5cada37f4c649960c65f085e4a6560c8f2434af17f61f359.jpg) + +![](images/c29cf259dc399f4e21c8c4f7d608aa1e18e99beada4465bb0889c3e5e396b224.jpg) + +![](images/263e2cd1983b5504534002a9e08142a1825db15238b408789288217b476298e2.jpg) + +![](images/dd9b81554546e0a59c3c28f9ada44ab3bbaa728fc3143c920471c1f9018a825c.jpg) + +![](images/394d63fedf81c999657f2aba44f0bc78ef040b4beb027c3bb9d743389c099a2d.jpg) +No SFM + +![](images/c7991a72311237cf1e0ca1df425693ccc4f0206cf699849ab04ab8a29f2fe7e9.jpg) +No Densify+ SFM + +![](images/975ba03b8c9ee43843b026f52e3a92d89d4ca8016ccf6892c8446d48f1c9f862.jpg) +No Quantile + +![](images/9beb0afd7c44dd2ee878b6284365758eaf13506a9d601fdd8ce95db8eaaaf0b9.jpg) +No Prune +Figure 7. Ablations - We qualitatively analyze the impact of excluding various components from our method. For better visualization, we zoom in on specific image regions and present their corresponding error maps. No SFM: While the method performs relatively well without initialization, it degrades in sparsely covered regions of the training views. No Densify+SFM: Without densification, the method significantly underperforms, leading to under-represented scenes. No Quantile: Disabling the quantile loss introduces floaters, degrading rendering quality. No Pruning: Ceasing point pruning has minimal impact due to the low number of prunable points. Points freeze: Fixing points position leads to sub-optimal mesh topology, resulting in a lack of fine details. + +![](images/40962d979d2c55510a7ad41b3aac5ae18393022fd0752c5d8324a35ae440b741.jpg) +Points freeze + +![](images/6f63d340546fed2524ef3bfef3216abf7dde90c289f2fadceff3d4fe6dcd9667.jpg) +Full + +Quantile loss. We disable the quantile regularization while training, which results in floater artifacts that degrade rendering quality for novel views. Some scenes are more affected by this than others, as the floaters result from data-dependent ambiguities, but on average we find that the quantile regularization improves reconstruction metrics. + +Point optimization. For this ablation, we keep the positions of the points fixed during training, leading to a suboptimal mesh topology that lacks adaptability to scene geometry. This rigidity prevents the mesh from dynamically optimizing its structure, resulting in degraded novel view quality, as the fixed topology fails to accurately represent fine details. + +# 5. Conclusions + +We have introduced Radiant Foam, a novel representation that allows real-time differentiable ray tracing. The core of our method is a foam structure of polyhedral cells, which allows efficient volumetric mesh ray tracing algorithms to be applied without relying on dedicated hardware such as NVIDIA RT cores. We allow these cells to be continuously optimized by parameterizing them as a Voronoi diagram, which we show to be differentiable under volume rendering. By doing so, we have achieved similar modeling quality as + +3D Gaussian Splating without sacrificing either the benefits of a true ray tracing-based volume renderer, or the fast rendering speed of rasterization-based renderers. + +Limitations and future work. While the Voronoi-based representation we have proposed is very effective at constructing foam models through continuous optimization, the space of possible foam models which could be used in our rendering pipeline is much larger than what is parameterized by Voronoi. Most notably, our current model always requires that cell boundaries be equidistant between neighbouring points, which leads to many small, empty cells being needed to define a surface. Future work could potentially relax this requirement by generalizing beyond Voronoi diagrams. Other open research questions include how to compose multiple foam models together efficiently and accounting for varying illumination, how to model dynamic content instead of static scenes, how to enable editing of scenes, and how to integrate generative modeling with our representation. Progress in these directions could make foam models relevant in real-time ray tracing applications currently dominated by triangle meshes, as we have already found that foam-based ray tracing can exceed the performance of dedicated ray tracing hardware. + +Acknowledgements. This work was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant [2023-05617], NSERC Collaborative Research and Development Grant, the SFU Visual Computing Research Chair, Google Research, Digital Research Alliance of Canada, and Advanced Research Computing at the University of British Columbia. We would like to thank George Kopanas, Lily Goli, Sherwin Bahmani, Alex Evans, Thomas Muller, Bernhard Kerbl, Vincent Sitzmann, Forrester Cole, Or Litany, David Fleet, and Delio Vicini for their feedback and/or early research discussions. + +# 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. ICCV, 2021. 3 +[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 3, 4, 6, 7, 12, 13 +[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 3 +[4] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting, 2024. 7, 12 +[5] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In CVPR, 2023. 2, 3 +[6] Jorge Condor, Sebastien Speierer, Lukas Bode, Aljaz Bozic, Simon Green, Piotr Didyk, and Adrian Jarabo. Don't splat your gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media. ACM Trans. Graph., 44(1), 2025. 3 +[7] Boris Delaunay. Sur la sphère vide. Bulletin of the Academy of Sciences of the USSR Classe des Sciences Mathematiques et Naturelles, 1934. 3, 5 +[8] Daniel Duckworth, Peter Hedman, Christian Reiser, Peter Zhizhin, Jean-François Thibert, Mario Lučić, Richard Szeliski, and Jonathan T Barron. Smerf: Streamable memory efficient radiance fields for real-time large-scene exploration. ACM Transactions on Graphics (TOG), 43(4):1-13, 2024. 3 +[9] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. 3 +[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043, 2023. 3 +[11] Chun Gu, Zeyu Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Tetrahedron splattering for 3d generation. In NeurlPS, 2024. 2, 3 +[12] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and + +high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 3 +[13] Leonidas Guibas and Jorge Stolfi. Primitives for the manipulation of general subdivisions and the computation of voronoi. ACM Trans. Graph., 4(2):74-123, 1985. 3 +[14] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. 37(6):257:1-257:15, 2018. 6, 7, 12, 13 +[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splating for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 3 +[16] Letian Huang, Jiayang Bai, Jie Guo, Yuanqi Li, and Yanwen Guo. On the error analysis of 3d gaussian splattering and an optimal projection strategy. arXiv preprint arXiv:2402.00752, 2024. 2 +[17] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 3 +[18] Pushkar Joshi, Mark Meyer, Tony DeRose, Brian Green, and Tom Sanocki. Harmonic coordinates for character articulation. ACM transactions on graphics (TOG), 26(3):71-es, 2007. 3 +[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 3, 6, 7, 12, 13 +[20] 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, 43(4), 2024. 3 +[21] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Jeff Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. NeurIPS, 2024. 3, 6, 7, 12, 13 +[22] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 12 +[23] Jonas Kulhanek and Torsten Sattler. Tetra-nerf: Representing neural radiance fields using tetrahedra. In ICCV, 2023. 2, 4 +[24] 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. 3 +[25] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splattering for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21644–21653, 2024. 3 +[26] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T. Barron, and Yinda Zhang. Ever: Exact volumetric ellipsoid rendering for real-time view synthesis, 2024. 3 + +[27] Gerd Marmitt and Philipp Slusallek. Fast ray traversal of tetrahedral and hexahedral meshes for direct volume rendering. In Proceedings of the Joint Eurographics/IEEE VGTC conference on Visualization, pages 235-242, 2006. 2, 6 +[28] Nelson Max and Min Chen. Local and global illumination in the volume rendering integral. Technical report, Lawrence Livermore National Lab. (LLNL), Livermore, CA (United States), 2005. 2 +[29] 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, 3, 4 +[30] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Riccardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics and SIGGRAPH Asia, 2024. 2, 3, 6, 7, 12 +[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. SIGGRAPH, 2022. 2, 3, 7, 13 +[32] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3d gaussian splattering for accelerated novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10349-10358, 2024. 3 +[33] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radslat: Radiance field-informed gaussian splatt-ting for robust real-time rendering with $900+$ fps. arXiv.org, 2024. 3 +[34] Logan Numerow, Yue Li, Stelian Coros, and Bernhard Thomaszewski. Differentiable voronoi diagrams for simulation of cell-based mechanical systems. ACM Trans. Graph., 43(4), 2024. 3 +[35] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7 (1):1-17, 2024. 3 +[36] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3 +[37] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. Third edition, 2023. 5 +[38] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. Stophe-pop: Sorted gaussian splatting for view-consistent real-time rendering. ACM Transactions on Graphics (TOG), 2024. 2 +[39] Daniel Rebain, Wei Jiang, Soroosh Yazdani, Ke Li, Kwang Moo Yi, and Andrea Tagliasacchi. Derf: Decomposed radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14153-14161, 2021. 4 + +[40] Daniel Reem. The geometric stability of voronoi diagrams with respect to small changes of the sites. In Proceedings of the twenty-seventh annual symposium on Computational geometry, pages 254-263, 2011. 6 +[41] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 3 +[42] Christian Reiser, Stephan Garbin, Pratul Srinivasan, Dor Verbin, Richard Szeliski, Ben Mildenhall, Jonathan Barron, Peter Hedman, and Andreas Geiger. Binary opacity grids: Capturing fine geometric detail for mesh-based view synthesis. ACM Transactions on Graphics (TOG), 43(4):1-14, 2024. 3 +[43] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898, 2024. 3 +[44] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. In Computer Graphics Forum, pages 461-470. Wiley Online Library, 2002. 3 +[45] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 7, 13 +[46] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6 +[47] Otto Seiskari, Jerry Yilammi, Valtteri Kaatrasalo, Pekka Rantalankila, Matias Turkulainen, Juho Kannala, Esa Rahtu, and Arno Solin. Gaussian splattering on the move: Blur and rolling shutter compensation for natural camera motion. In European Conference on Computer Vision, pages 160-177. Springer, 2025. 3 +[48] Gopal Sharma, Daniel Rebain, Kwang Moo Yi, and Andrea Tagliasacchi. Volumetric rendering with baked quadrature fields. ECCV, 2024. 3 +[49] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. NeurIPS, 2021. 2, 3, 5 +[50] Jonathan Richard Shewchuk. Delaunay refinement algorithms for triangular mesh generation. Computational Geometry, 22 (1):21-74, 2002. 16th ACM Symposium on Computational Geometry. 3 +[51] Sanghyun Son, Matheus Gadelha, Yang Zhou, Zexiang Xu, Ming C. Lin, and Yi Zhou. Dmesh: A differentiable representation for general meshes, 2024. 3 +[52] Andrea Tagliasacchi and Ben Mildenhall. Volume rendering digest (for nerf), 2022. 4 +[53] Georges Voronoi. Nouvelles applications des paramètres continus à la théorie des formes quadratiques. Journal für die reine und angewandte Mathematik (Crelles Journal), 1908, 3, 5 +[54] 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. 3 +[55] Zian Wang, Tianchang Shen, Merlin Nimier-David, Nicholas Sharp, Jun Gao, Alexander Keller, Sanja Fidler, Thomas Müller, and Zan Gojcic. Adaptive shells for efficient neural radiance field rendering. arXiv preprint arXiv:2311.10091, 2023. 3 +[56] D. F. Watson. Computing the n-dimensional delaunay tessellation with application to voronoi polytopes. The Computer Journal, 24(2):167-172, 1981. 3 +[57] M. Weiler, M. Kraus, M. Merz, and T. Ertl. Hardware-based view-independent cell projection. IEEE Transactions on Visualization and Computer Graphics, 9(2):163-175, 2003. 2, 4, 12 +[58] Turner Whitted. An improved illumination model for shaded display. In ACM Siggraph 2005 Courses, pages 4-es. 2005. 3 +[59] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3 +[60] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 7, 13 +[61] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, page 371-378, New York, NY, USA, 2001. Association for Computing Machinery. 3 \ No newline at end of file diff --git a/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/images.zip b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..758ceeffb1295d5090209e88e20cd09416739ca4 --- /dev/null +++ b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00d28d40808c211e60e0e315fca2b45fc25631fe4303210786d02def3df16c63 +size 689683 diff --git a/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/layout.json b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..db4be3788c851054fe5c339de0278bc1550a577c --- /dev/null +++ b/ICCV/2025/Radiant Foam_ Real-Time Differentiable Ray Tracing/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f069b133b8ae419522e7c4d02caed13d39fb3d94da223230b364d0da896bc77a +size 360332 diff --git a/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/a0c0b5bc-92d8-4fb7-af17-7e01c3190672_content_list.json b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/a0c0b5bc-92d8-4fb7-af17-7e01c3190672_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ee4ef4962b1f26cbe0863196f12fcdeacf0947ac --- /dev/null +++ b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/a0c0b5bc-92d8-4fb7-af17-7e01c3190672_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14424740447e271e6361ed546679dcd5d732f4093320b5cf7d385b1ee499dd7c +size 79729 diff --git a/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/a0c0b5bc-92d8-4fb7-af17-7e01c3190672_model.json b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/a0c0b5bc-92d8-4fb7-af17-7e01c3190672_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dd63c215b6faa2599fe965376535a0f00ea5afe4 --- /dev/null +++ b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/a0c0b5bc-92d8-4fb7-af17-7e01c3190672_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21a47d2cf578f0c35f55176fa8fec715bd4b32fbc93590aa7eee7e75d3116fbc +size 97834 diff --git a/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/a0c0b5bc-92d8-4fb7-af17-7e01c3190672_origin.pdf b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/a0c0b5bc-92d8-4fb7-af17-7e01c3190672_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4a0fce9503e8882c437fbf8449e2ae29e3d42251 --- /dev/null +++ b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/a0c0b5bc-92d8-4fb7-af17-7e01c3190672_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b8e0827e662e7dd57fa06fcb52bf1d96e522383e79d95ddd66c86fe22e4e0a3 +size 1296694 diff --git a/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/full.md b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2baa6a2c97204c10b2081ad3f5ba78f9d947d381 --- /dev/null +++ b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/full.md @@ -0,0 +1,303 @@ +# RainbowPrompt: Diversity-Enhanced Prompt-Evolving for Continual Learning + +Kiseong Hong +Department of AI +Chung-Ang University +ghdrltjd@cau.ac.kr + +Gyeong-hyeon Kim +School of CSE +Chung-Ang University +leonardkkh@cau.ac.kr + +Eunwoo Kim* +School of CSE +Chung-Ang University +eunwoo@cau.ac.kr + +# Abstract + +Prompt-based continual learning provides a rehearsal-free solution by tuning small sets of parameters while keeping pre-trained models frozen. To meet the complex demands of sequential tasks, it is crucial to integrate task-specific knowledge within prompts effectively. However, existing works rely on either fixed learned prompts (i.e., prompts whose representations remain unchanged during new task learning) or on prompts generated from an entangled task-shared space, limiting the representational diversity of the integrated prompt. To address this issue, we propose a novel prompt-evolving mechanism to adaptively aggregate base prompts (i.e., task-specific prompts) into a unified prompt while ensuring diversity. By transforming and aligning base prompts, both previously learned and newly introduced, our approach continuously evolves accumulated knowledge to facilitate learning new tasks. We further introduce a learnable probabilistic gate that adaptively determines which layers to activate during the evolution process. We validate our method on image classification and video action recognition tasks in class-incremental learning, achieving average gains of $9.07\%$ and $7.40\%$ over existing methods across all scenarios. + +# 1. Introduction + +Continual learning (CL) aims to enable a model to learn sequential tasks while retaining knowledge from previous tasks. However, sequential learning faces a major challenge, catastrophic forgetting [9], where learning new tasks deteriorates performance on previously learned tasks. To address this, various CL approaches [20, 22, 36] have been proposed. A recent paradigm in CL, known as prompt-based CL (PCL) [33, 45], has shown promise in mitigating catastrophic forgetting with its rehearsal-free nature, which applied to a pre-trained vision transformer (ViT) [8]. These methods leverage prompt-tuning [25], a transfer learning + +![](images/89f739a55d3c3860b2e2e563e9b8c66718fe97c1b3f72fef39cbcb6d5c1d3eb9.jpg) +(a) Existing prompt integration + +![](images/3dd1d7c61ee12a70fdc8806212bdc673225a582fcfcb9137845076011fe6688d.jpg) +(b) Prompt-evolution (ours) +Figure 1. Conceptual illustration of (a) existing prompt integration approaches and (b) the proposed prompt-evolving approach. (a) Existing approaches integrate prompts using input-conditioned weights $w$ , with fixed old task prompts (blue box) or those generated between task-shared and task-specific spaces (green box). (b) Our approach progressively transforms and aligns prompts to make their representations more adaptable to new tasks. + +strategy initially developed for natural language processing (NLP). Instead of modifying the weights of ViT, prompting keeps the model frozen and fine-tunes additional small sets of parameters called prompts. These prompts act as task-specific instructions, guiding the model without the need for direct weight updates. + +Existing PCL works either select task-relevant prompts from a prompt pool [45] or extend this approach by incorporating a task-invariant prompt shared across all tasks [44]. Recent studies attempt to integrate prompts into a unified prompt [35, 38]. They produce a prompt by combining prompts via an input-conditioned weighted sum [38] or by generating task-specific prompts and merging them [35] (see Fig. 1a). However, they often yield prompts with limited representational diversity. This limitation arises from relying on learned representations that remain unchanged without adapting to new tasks [38] or deriving prompts in a task-shared space vulnerable to task interference and dominance [35]. This reduces adaptability and generalization, highlighting the need for more effective strategies to balance integration with diversity. + +Furthermore, our findings in Fig. 2 reveal that this limitation hinders achieving high accuracy while minimizing forgetting during sequential task learning. We observe that existing methods fail to achieve optimal knowledge integra + +![](images/f39ad4c07b416e5ef14ff519a4dd47bc060b72c4b740e7cba9d1610f49d3efa1.jpg) +(a) Number of seen classes vs. Diversity + +![](images/3d29e69f03d9a6ed54a348ddea0f3e565945360c4fbdb69fe3110a5a8cb0eeb1.jpg) +(b) Diversity vs. Accuracy +Figure 2. Results of the representation diversity in a class-incremental learning scenario with a 10-task setting on CIFAR-100. We compare ours with two baselines using fixed [38] and generated [35] prompts. In (a), we measure the average nuclear norm of the prompt at each time step as new classes are introduced, using test samples from all seen classes. We insert prompts into all layers and calculate diversity using the prompt in the last layer. In (b) and (c), we compare the representation diversity (after all classes are learned) against accuracy and forgetting, respectively, with each dot denoting a distinct random trial. + +![](images/f3c89e90e4f115896c8b60aa5e97a4e357697dabc96f0846c7abb1df27b7d2ed.jpg) +(c) Diversity vs. Forgetting + +tion with enhanced diversity1. + +Merging prompts via a weighted sum without adequate adaptation to new tasks yields low representational diversity (blue line in Fig. 2a). This limitation restricts the ability to capture distinctive features introduced by new classes. Generating prompts for new classes improves representational diversity, as it adapts to class-specific traits (green line in Fig. 2a). However, the diversity of prompts generated from the increasingly entangled task-shared space remains insufficient, as overlapping knowledge across tasks leads to less distinctive representations. Figures 2b and 2c demonstrate that higher representation diversity improves accuracy and reduces forgetting, respectively. These analyses underscore the importance of optimal knowledge integration to ensure diversity and preserve previously learned knowledge in sequential task learning. + +In this work, we propose a prompt-evolving approach to integrate task-specific knowledge into a diversity-enhanced representation. Prompts undergo transformation and alignment, gradually adapting their representations to new tasks while preserving previously learned information. Our method progressively reconfigures the prompts to seamlessly integrate complementary knowledge from both previous and current tasks. It consists of two steps: attention-based transformation and task-guided alignment. Our attention-based transformation assesses the relevance of multiple prompts to new tasks. Unlike existing works that derive prompt weights via cosine similarity between input-conditioned queries and prompt keys [38], our method dynamically reweights the contribution of each prompt based on its relevance to new tasks at multi-level granularity. The task-guided alignment further refines these representations by progressively tailoring them to a new task. It incor + +porates non-linear transformations to align the representations of the new task while preserving the intrinsic properties of each prompt. Finally, we introduce RainbowPrompt, a novel prompt that integrates enhanced representations of task-specific prompts. We further optimize the proposed method by introducing a learnable probabilistic gate that adaptively regulates layer activation during evolution, leveraging task-specific differences. + +To demonstrate the effectiveness of RainbowPrompt, we compare it with existing prompt-based CL approaches [11, 33, 35, 38, 44, 45] in class-incremental learning scenarios. The evaluation includes image classification tasks using ImageNet-R [14], CIFAR-100 [24], and CUBS [42]. We also evaluate video action recognition tasks on UCF-101 [39] and ActivityNet [2]. Extensive experiments show that the proposed method significantly outperforms the compared methods across all benchmarks, demonstrating its effectiveness and versatility in image and video tasks. The main contributions of this work are as follows: + +- We propose a novel prompt-evolving approach for CL, enabling effective knowledge integration while ensuring diversity across tasks. +- Our method employs attention-based transformation and task-guided alignment to integrate knowledge from previous tasks while adapting to new tasks. +- Experimental results on a wide range of tasks demonstrate that ours outperforms its competitors with an overall average margin of $8.23\%$ across all scenarios. + +# 2. Related Work + +Continual Learning. Continual learning (CL) aims to progressively acquire new knowledge while retaining previously learned information [7, 15]. A major challenge in CL is catastrophic forgetting [9], where learning new tasks degrades performance on earlier ones. Solutions to this challenge can be broadly divided into three categories: + +Regularization-based methods [1, 22] mitigate catastrophic forgetting by adding penalty terms that consolidate knowledge from earlier tasks. Rehearsal-based methods tackle catastrophic forgetting by retaining a subset of the data from previous tasks [4, 29, 34] or generating data for previous tasks using adversarial techniques [32, 37, 46]. Architecture-based methods [17, 19, 47, 48] dynamically adjust the network structure to accommodate new tasks. Despite the advancements of these approaches, the emergence of foundation models trained on large-scale data has driven a paradigm shift in CL [45], moving towards replay-free and parameter-efficient approaches [21, 27]. + +Prompt-based Continual Learning. Prompting [28] initially involved manually designed task-specific instructions to induce desired responses from pre-trained models. In the context of CL, prompt-based methods [33, 45] retain task-specific knowledge without rehearsal buffers. They use prompts as keys to retrieve stored knowledge, eliminating the need for model parameter updates. L2P [45] introduces prompt-tuning with key-query matching to retrieve task-relevant knowledge. DualPrompt [44] further introduces task-invariant prompts along with task-specific prompts to distinctly encode task-relevant and task-agnostic knowledge. CODA-Prompt [38] introduces a decomposed attention-based prompting method, while ConvPrompt [35] leverages convolution for prompt generation. PGP constrains prompt updates to be orthogonal to prior prompt directions via a projection mechanism. CPrompt [11] introduces consistency prompting to address inconsistencies between the classifier and the prompt. + +Limitations. Among the competitive approaches, CODA-Prompt [38] and ConvPrompt [35] focus on integrating multiple base prompts but often overlook diversity within the resulting unified prompt. Fixed representations hinder the capture of diverse information [3, 5], while generated representations shaped by a task-shared space often suffer from task interference and dominance, resulting in overfitting and limited generalization [23, 30]. In contrast, we address these limitations by progressively evolving the representations of all accumulated base prompts. We accumulate independent base prompts for each task and integrate them into a unified prompt for optimal knowledge integration. + +# 3. Methodology + +# 3.1. Prerequisites + +Continual learning (CL) enables a model to progressively learn knowledge from a sequence of tasks $\{\mathcal{T}^1,\mathcal{T}^2,\ldots ,\mathcal{T}^T\}$ . The $t$ -th task $\mathcal{T}^t$ consists of $N_{t}$ input-label pairs $\{(x_i,y_i)\}_{i = 1}^{N_t}$ , where $x_{i}\in \mathcal{X}^{t}$ represents the input data and $y_{i}\in \mathcal{Y}^{t}$ denotes the corresponding label. In this work, we focus on class-incremental learning, where task identities are unknown during testing [16]. We also + +adopt a practical rehearsal-free setting [44], which prohibits the storage of data from previous tasks. + +In prompt-based continual learning (PCL), the model $f_{\theta}(\cdot)$ is typically chosen to be a pre-trained vision transformer (ViT) [8]. Rather than adjusting the weights of ViT directly, we use small learnable parameters, i.e., prompts $\pmb{p} \in \mathbb{R}^{L_p \times D}$ , to provide task-specific instructions [25], where $L_{p}$ represents the prompt length and $D$ is the embedding dimension. Specifically, we employ prefix tuning (P-T) [26], where the prompts comprising $\pmb{p}_K \in \mathbb{R}^{L_p / 2 \times D}$ and $\pmb{p}_V \in \mathbb{R}^{L_p / 2 \times D}$ are prepended to the key and value representations in the multi-head self-attention layers of ViT. + +# 3.2. Prompt-Evolving Mechanism + +Our goal is to consolidate task-specific prompts into a cohesive prompt through a prompt-evolving mechanism, which dynamically reconfigures their representations to adapt to new tasks while enhancing representational diversity. At each time step $t$ , a new task is introduced, with its corresponding base prompts (i.e., task-specific prompts) $\pmb{p}^t = \{\pmb{p}_l^t\}_{l=1}^L$ , where $\pmb{p}_l^t \in \mathbb{R}^{L_p \times D}$ is the base prompt at layer $l$ , and $L$ is the number of layers. To facilitate sequential learning, we maintain a set of accumulated prompts denoted as $\pmb{\mathcal{P}} = \{\pmb{p}^1, \pmb{p}^2, \dots, \pmb{p}^t\}$ , which includes all base prompts up to time step $t$ . Our approach seeks to integrate these prompts into unified prompts, i.e., $\pmb{p}^{\mathrm{rainbow}(t)} = \{\pmb{p}_l^{\mathrm{rainbow}(t)}\}_{l=1}^L$ , where $\pmb{p}_l^{\mathrm{rainbow}(t)}$ is the unified prompt at layer $l$ , without hurting the base prompts. + +The accumulated prompts $\mathcal{P}$ present a challenge when combining them in continual learning. Direct integration of prompts risks losing knowledge from previously learned tasks, while freezing them restricts adaptation and hinders knowledge transfer to the $t$ -th task. To address this, we update the base prompts for the current task at each time step while keeping the base prompts from previous tasks frozen to preserve their knowledge. To facilitate knowledge transfer to new tasks, we transform and align the representations of all base prompts by introducing learnable components. Specifically, we apply self-attention to dynamically reweight the contributions of accumulated prompts, facilitating their transformation by emphasizing salient information while suppressing less relevant details. This adaptive reweighting mitigates knowledge dilution and enhances representational diversity. By aligning the prompts to their transformed representations for the new task, we encapsulate the diverse knowledge in $\mathcal{P}$ into a set of unified prompts. The overall framework is illustrated in Fig. 3. + +# 3.2.1. Attention-based Transformation + +We first apply an attention-based transformation that enables accumulated prompts to self-identify their influence on the new task via two-level interactions: task-level and feature-level. Interactions among prompt vectors capture + +![](images/085dc04c2a4c70d529c33491e9a074b908c08c711d1c582b3f51d5e90ba1fecb.jpg) +Figure 3. Illustration of the proposed framework. When a new task $t$ arrives, we evolve the base prompt representations, comprising all task-specific prompts up to the $t$ -th time step, through attention-based transformation and task-guided alignment. Finally, RainbowPrompt, $\pmb{p}_l^{\mathrm{rainbow}(t)}$ , is constructed at each layer $l$ by integrating diverse knowledge from the accumulated base prompts. We further introduce a learnable probabilistic gate $g_l^t$ to selectively insert $\pmb{p}_l^{\mathrm{rainbow}(t)}$ into intermediate layers. During testing, the proposed method requires the RainbowPrompts $\{\pmb{p}_l^{\mathrm{rainbow}(i)}\}_{l=1}^{L_i}$ for each task $i$ , which are appended to $L_i$ selected layers determined by $\{\pmb{g}_l^i\}_{l=1}^L$ . + +overall contributions at the task level, while interactions among the individual features (elements) of each vector measure fine-grained contributions. Before the transformation process, we precede a task-conditioning step to inject task-relevant information into $\mathcal{P}_l^2$ . + +We use a learnable task embedding vector $e^t \in \mathbb{R}^D$ to compute attention weights that emphasize the task-relevant components of $\mathcal{P}_l$ at the $t$ -th time step. We condition the set of base prompts by $\mathcal{P}_l \gets \text{softmax}\left(\sigma(e^t)\mathcal{P}_l^T / \sqrt{d_p}\right)\mathcal{P}_l$ , where $\sigma(\cdot)$ and $d_p$ denote the broadcasting function and the dimensionality of the $\mathcal{P}_l$ , respectively. + +We treat $\pmb{p}_l^{\mathrm{new}} = \pmb{p}_l^t$ as the query and use the concatenated task-conditioned base prompts set $\mathcal{P}_l$ as the key and value inputs to the transformation process. First, we project each prompt into a lower-dimensional space to align their dimensions for the transformation, reducing the complexity of subsequent operations: + +$$ +Q = W _ {l} ^ {Q} \boldsymbol {p} _ {l} ^ {\text {n e w}}, K = W _ {l} ^ {K} \boldsymbol {\mathcal {P}} _ {l}, V = W _ {l} ^ {V} \boldsymbol {\mathcal {P}} _ {l}, \tag {1} +$$ + +where $W_{l}^{Q}\in \mathbb{R}^{D\times D_{p}}$ , $W_{l}^{K}\in \mathbb{R}^{D\times D_{p}}$ , and $W_{l}^{V}\in \mathbb{R}^{D\times D_{p}}$ are learnable projection matrices for the respective query, key, and value at the $l$ -th layer, and $D_{p}$ represents the projected dimension ( $D_{p}\ll D$ ). + +For the task-level transformation, we define the attention-based transformation function, $\mathrm{AT}(\cdot)$ , which computes the inter-task affinity matrix $\mathcal{G}$ from the attention between $Q$ and $K$ , and then uses $\mathcal{G}$ to weight the value representations: + +$$ +\operatorname {A T} (Q, K, V) = \mathcal {G} \cdot V \triangleq \tilde {V} \in \mathbb {R} ^ {t \times L _ {p} \times D _ {p}}, \tag {2} +$$ + +where $\mathcal{G} = \mathrm{softmax}\left(QK^T / \sqrt{d_k}\right)$ , $d_k$ denotes the dimensionality of the key, $\cdot$ represents the matrix multiplication operation, and $\tilde{V}$ is the transformed representation. $\mathcal{G}$ quantifies how much information from existing tasks should contribute to the new task, allowing us to adaptively weigh the influence of each task. By integrating $\mathcal{G}$ with $V$ , we transform the representations associated with the prompts of each task based on their contribution to the new task $t$ . + +The feature-level transformation captures cross-feature influences at a finer granularity, complementing the task-level transformation. It leverages transposed query $Q^T$ and key $K$ to enable comparison of feature dimensions. This approach is inspired by bilinear pooling [10], which enjoys multiplicative interactions between feature dimensions. We update the output of the task-level transformation $\tilde{V}$ as + +$$ +\operatorname {A T} (Q, K, \tilde {V}) = \mathcal {F} \cdot \tilde {V} ^ {T} \triangleq \hat {V} \in \mathbb {R} ^ {t \times D _ {p} \times L _ {p}}, \tag {3} +$$ + +where the inter-feature affinity matrix $\mathcal{F}$ is computed as $\mathcal{F} = \mathrm{softmax}\left(Q^T K / \sqrt{d_k}\right)$ . It quantifies the contributions between individual features across tasks, capturing fine-grained dependencies. By applying $\mathcal{F}$ to $\tilde{V}$ , the feature-level transformation integrates cross-feature influences, refining $\tilde{V}$ to incorporate contributions distinct from those of the task-level transformation. + +To integrate the transformed representations, we redefine the output of the feature-level transformation as $\hat{V} \gets \mathrm{LN}(\mathcal{P}_l + (\hat{V}^T W_l^O)) \in \mathbb{R}^{t \times L_p \times D}$ . $\mathrm{LN}(\cdot)$ represents layer normalization, and $W_l^O \in \mathbb{R}^{D_p \times D}$ is a learnable projection matrix. This integrates $\mathcal{P}_l$ with the transformed representations, preserving the original information in $\mathcal{P}_l$ while enriching the resulting representations with newly adapted features. + +# 3.2.2. Task-Guided Alignment + +Following the attention-based transformation, the task-guided alignment refines the transformed representations $\hat{V}$ . This decodes the representations to meet the traits of the new task, preserving the attributes of each prompt: + +$$ +\tilde {\mathcal {P}} _ {l} = \mathrm {L N} (\hat {V} + \mathrm {L T} (\hat {V})) \in \mathbb {R} ^ {t \times L _ {p} \times D}, \tag {4} +$$ + +where $\mathrm{LT}(x) = \max (0,xW_l^1)W_l^2$ aligns the transformed representations to identify task-relevant patterns. It refines their distribution across tasks within a reduced space $D_{n}$ $(D_{n}\ll D)$ [12, 31], with $W_{l}^{1}\in \mathbb{R}^{D\times D_{n}}$ and $W_{l}^{2}\in$ $\mathbb{R}^{D_n\times D}$ as learnable weight matrices (bias terms omitted for simplicity). + +Finally, we derive a RainbowPrompt, $\pmb{p}_l^{\mathrm{rainbow}(t)}$ , by combining the evolved representations of prompts from all tasks learned up to time step $t$ into a unified prompt: + +$$ +\boldsymbol {p} _ {l} ^ {\text {r a i n b o w} (t)} = \frac {1}{t} \sum_ {i = 1} ^ {t} \tilde {\boldsymbol {\mathcal {P}}} _ {l} [ i ] \in \mathbb {R} ^ {L _ {p} \times D}, \tag {5} +$$ + +where $\tilde{\mathcal{P}}_l[i]$ denotes the evolved representation of the $i$ -th base prompt. $\pmb{p}_l^{\mathrm{rainbow}(t)}$ is a distinct type of prompt that differs from task-specific and task-shared prompts by evolving and integrating accumulated knowledge without predefined roles. Note that the proposed prompt-evolving mechanism is performed exclusively during training, where task-specific RainbowPrompts are produced and stored. At test time, they are directly used for prediction without requiring the evolution components. + +# 3.3. Adaptive Prompting + +Determining which layers to insert RainbowPrompts poses an additional challenge due to varying learning complexities across tasks. Manually selecting these layers is impractical, as it fails to account for task-specific differences [15]. To this end, we introduce a task-specific learnable probabilistic gate $G^{t} = \{g_{l}^{t}\}_{l = 1}^{L}$ to learn where to insert RainbowPrompts in the model. $g_{l}^{t}$ is a Bernoulli random variable deciding whether to insert a RainbowPrompt at layer $l$ . Since the discrete nature of $g_{l}^{t}$ prevents gradient-based optimization, we employ the Gumbel-Softmax trick [18] to relax it into a differentiable form: + +$$ +\hat {g} _ {l} ^ {t} (p) = \frac {\exp \left(\log \delta_ {l} ^ {t} (p) + Z _ {l} ^ {t} (p) / \tau\right)}{\sum_ {i \in \{0 , 1 \}} \exp \left(\log \delta_ {l} ^ {t} (i) + Z _ {l} ^ {t} (i) / \tau\right)}, \tag {6} +$$ + +where $p \in \{0,1\}$ , $Z_{l}^{t} = -\log (-\log U_{l}^{t})$ is the Gumbel noise generated from $U_{l}^{t} \sim \mathcal{U}[0,1]$ , and $\tau$ controls the temperature for continuous relaxation. We sample discrete task-specific decisions from the learned distribution $\delta_{l}^{t} = [\alpha_{l}^{t}, 1 - \alpha_{l}^{t}]$ , where $\alpha_{l}^{t}$ is the probability of prompt insertion for layer $l$ . This flexible activation aligns more effectively with adaptive prompting, compared to the manual + +prompting [35, 38]. Section 4.4 provides a detailed discussion on this aspect. + +# 3.4. Optimization + +We jointly optimize $\delta_l^t$ , the new task prompt $p_l^t$ , and the learnable parameters of the prompt-evolving mechanism, defined as $W^{\mathrm{evolution}} = \{(W_{l}^{Q}, W_{l}^{K}, W_{l}^{V}, W_{l}^{O}, W_{l}^{1}, W_{l}^{2})\}_{l=1}^{L}$ using the cross-entropy loss. To encourage a sparse yet effective RainbowPrompt insertion pattern, we impose a regularization term to suppress insertion probabilities $\alpha_l^t$ , i.e., $\mathcal{L}_{\mathrm{sparse}} = \sum_{l \leq L} \log \alpha_l^t$ . In addition, we optimize the task embedding vector $e^t$ using a matching loss [44], $\mathcal{L}_{\mathrm{match}} = \gamma(q(\boldsymbol{x}), e^t)$ , where $\boldsymbol{x} \in \mathcal{T}^t$ , $\gamma$ denotes cosine similarity, and $q(\cdot)$ is a query function [44]. Finally, the proposed method minimizes the following total loss: + +$$ +\min _ {\Theta^ {t}} \sum_ {i = 1} \mathrm {C E} \left(\boldsymbol {z} _ {i}, y _ {i}\right) + \lambda_ {\mathrm {s}} \mathcal {L} _ {\text {s p a r s e}} + \lambda_ {\mathrm {m}} \mathcal {L} _ {\text {m a t c h}}, \tag {7} +$$ + +where $\Theta^t = \{\pmb{p}^t, \pmb{e}^t, \pmb{G}^t, W^{\text{evolution}}, \phi\}$ , where $\phi$ denotes a classifier. CE is the cross-entropy loss, $z_i$ represents the final output, and $\lambda_{\mathrm{s}}$ and $\lambda_{\mathrm{m}}$ are balancing parameters, both set to 0.01 in our experiments. + +# 4. Experiments + +# 4.1. Setup + +Scenarios and Datasets. To evaluate RainbowPrompt $^3$ , we demonstrated it on image classification and video action recognition tasks under class-incremental learning scenarios. For image classification, we used the ImageNet-R [14], CIFAR-100 [24], and CUBS [42] datasets. ImageNet-R is a challenging benchmark due to its classes featuring distinct styles (e.g., cartoon, graffiti, origami) and significant intra-class diversity [44]. We divided its 200 classes into disjoint subsets, forming 10-task (20 classes per task) and 20-task (10 classes per task) settings. CIFAR-100, a widely used benchmark in prompt-based continual learning (PCL), was split into 10-task (10 classes per task) and 20-task (5 classes per task) settings. CUBS, a fine-grained dataset featuring bird species, poses another challenge due to subtle inter-class differences. We applied the same task split configuration as used for ImageNet-R. When splitting each dataset, we ensure no class overlaps occur between them. + +For video action recognition, we conducted experiments using the vCLIMB benchmark [40] on trimmed versions of the UCF-101 [39] and ActivityNet [2] datasets. We evaluated our approach on UCF-101 (101 classes) and ActivityNet (200 classes) under both 10-task and 20-task settings. Following [40], we partitioned each dataset into disjoint subsets of randomly selected classes, with each subset treated as an independent task. + +Table 1. Results on ImageNet-R and CIFAR-100. $A_N$ and $F_N$ represent the average accuracy and forgetting for the $N$ -task setting, respectively. The best performance is highlighted in bold. + +
MethodImageNet-RCIFAR-100
A10(↑)F10(↓)A20(↑)F20(↓)A10(↑)F10(↓)A20(↑)F20(↓)
Joint training79.60 ± 0.87-79.60 ± 0.87-93.22 ± 0.16-93.22 ± 0.16-
L2P [45]63.49 ± 0.406.85 ± 0.4259.38 ± 0.505.89 ± 0.3682.76 ± 1.177.86 ± 0.3977.95 ± 0.729.88 ± 0.30
DualPrompt [44]68.50 ± 0.525.14 ± 0.1863.21 ± 0.495.28 ± 0.4585.07 ± 0.495.57 ± 0.2080.49 ± 0.318.84 ± 0.68
DualPrompt-PGP [33]69.34 ± 0.054.53 ± 0.0464.75 ± 0.386.04 ± 0.1586.92 ± 0.055.35 ± 0.1983.74 ± 0.017.91 ± 0.15
CODA-Prompt [38]74.24 ± 0.564.92 ± 0.2170.86 ± 0.426.87 ± 0.2587.00 ± 0.384.78 ± 0.2482.15 ± 0.176.33 ± 0.23
CPrompt [11]76.71 ± 0.614.66 ± 0.9374.45 ± 0.254.98 ± 0.2187.83 ± 0.374.88 ± 0.7984.66 ± 0.135.69 ± 0.06
ConvPrompt [35]77.86 ± 0.254.33 ± 0.2475.10 ± 0.394.10 ± 0.2988.87 ± 0.334.75 ± 0.1587.37 ± 0.135.16 ± 0.01
RainbowPrompt (ours)79.09 ± 0.133.90 ± 0.2378.36 ± 0.473.44 ± 0.2989.86 ± 0.113.44 ± 0.2690.15 ± 0.053.75 ± 0.23
+ +Compared Methods. We evaluated RainbowPrompt against existing PCL approaches: L2P [45], DualPrompt [44], DualPrompt-PGP [33], CODA-Prompt [38], CPrompt [11], and ConvPrompt [35]. We also reported the results of joint training, where the ViT model [8] was trained on a combined dataset encompassing the training data from all tasks. We evaluated the methods using two metrics: average accuracy and average forgetting [45]. + +Implementation Details. We followed the training details from [38], including the optimizer and input resolution, while using a learning rate of 0.03. We used a single base prompt with a prompt length of $L_{p} = 20$ per task across all scenarios in RainbowPrompt. We set the dimensionality of the projection matrices $(D_{p})$ to 96 for the 10-task setting and 56 for the 20-task setting. We set the dimensionality of the weight matrices $(D_{n})$ as 56 for the 10-task setting and 28 for the 20-task setting. $\mathrm{LT}(\cdot)$ consists of two linear layers stacked with ReLU activations. For adaptive prompting, we used soft decisions from the distribution $\delta_l^t$ during training for a certain number of epochs. After that, we sampled task-specific decisions from the learned distribution. We implemented all methods, including RainbowPrompt, using the pre-trained ViT-B/16 as in [45]. We conducted each experiment three times with random trials and reported the results as the mean and standard deviation of the runs. + +# 4.2. Image Classification + +We first evaluated the proposed approach on image classification using the ImageNet-R dataset. Tab. 1 (left) presents the results for both 10-task and 20-task settings. In the 10-task setting, RainbowPrompt surpasses all the competitors, L2P, DualPrompt, DualPrompt-PGP, CODA-Prompt, CPrompt, and ConvPrompt, in average accuracy by gaps of $15.60\%$ , $10.59\%$ , $9.75\%$ , $4.85\%$ , $2.38\%$ , and $1.23\%$ , respectively. Compared to ConvPrompt, which achieves the lowest average forgetting among the compared methods, RainbowPrompt further reduces forgetting by $0.43\%$ , highlighting its superior capability to retain knowledge while accommodating new tasks. The results remain consistent in the more challenging 20-task setting. Notably, Rainbow- + +Prompt achieves an accuracy improvement of $3.26\%$ and a forgetting reduction of $0.66\%$ compared to the strong competitor, ConvPrompt. These results highlight the effectiveness of ours in both the 10-task and 20-task settings. + +We further evaluated RainbowPrompt on CIFAR-100. The results for the 10-task and 20-task settings are summarized in Tab. 1 (right). Notably, in the 20-task setting, L2P, DualPrompt, DualPrompt-PGP, CODA-Prompt, CPrompt, and ConvPrompt exhibit accuracy drops of $4.81\%$ , $4.58\%$ , $3.18\%$ , $4.85\%$ , $3.17\%$ , and $1.50\%$ , respectively, compared to their performance in the 10-task setting. In contrast, RainbowPrompt not only avoids performance degradation but achieves a $0.29\%$ increase in accuracy on average. It demonstrates consistently competitive performance as the number of tasks increases, while others struggle to adapt under extended task sequences. + +We evaluated RainbowPrompt on the fine-grained CUBS dataset. Fig. 4 illustrates the average accuracy and forgetting across all seen classes. RainbowPrompt outperforms all baseline methods in terms of the metrics. In the 10-task setting, RainbowPrompt and the strongest baseline, ConvPrompt, exhibit stable performance as the number of seen classes increases. RainbowPrompt achieves a $3.62\%$ higher accuracy and $2.28\%$ lower forgetting than ConvPrompt. Notably, in the 20-task setting, RainbowPrompt significantly outperforms ConvPrompt, achieving a notable $10.44\%$ improvement in average accuracy and a $5.44\%$ reduction in forgetting. It demonstrates the ability to produce diversity-enhanced prompts while accommodating new tasks. + +# 4.3. Video Action Recognition + +To extend the evaluation of the proposed prompt-evolving mechanism beyond image classification, we applied it to video action recognition tasks. We utilized trimmed versions of the UCF-101 and ActivityNet datasets [41] to focus on well-defined actions and facilitate analysis by excluding irrelevant background segments. Each video was divided into three equal segments, and one frame was randomly sampled from each segment [43]. The datasets + +![](images/e267b36728d822dad2fdf13b0eb4577ec978a984209b1acd7d84d9be22950d4c.jpg) + +![](images/d7fa3a5a65af2aa1e2b447531afd79692def04937ce34eb8e4ef91d5d1332c05.jpg) + +![](images/687743c0d77ae372c7a48c99383343760e26cf630bc9bb062060abb51952980a.jpg) +Figure 4. Results on CUBS. The average accuracy for all seen classes at each time step is shown with its standard deviation indicated by the shaded area. + +Table 2. Results on the video action recognition tasks using UCF-101 and ActivityNet. + +
MethodUCF-101ActivityNet
A10(↑)F10(↓)A20(↑)F20(↓)A10(↑)F10(↓)A20(↑)F20(↓)
L2P [45]78.35 ± 1.545.46 ± 0.2970.29 ± 0.198.31 ± 0.4063.46 ± 0.476.47 ± 1.1356.47 ± 1.129.96 ± 1.33
DualPrompt [44]83.15 ± 0.825.41 ± 0.3874.96 ± 1.237.97 ± 0.4163.76 ± 0.885.78 ± 0.5056.59 ± 0.819.89 ± 0.05
DualPrompt-PGP [33]84.39 ± 0.565.92 ± 0.1675.37 ± 0.167.50 ± 0.5964.65 ± 0.366.20 ± 0.6457.46 ± 0.648.56 ± 0.25
CODA-Prompt [38]84.77 ± 0.755.81 ± 0.2775.35 ± 0.905.66 ± 0.5066.13 ± 0.656.23 ± 0.1558.62 ± 0.588.27 ± 0.34
CPrompt [11]87.16 ± 0.595.30 ± 0.2581.78 ± 0.044.42 ± 0.1966.81 ± 0.487.25 ± 0.0862.17 ± 0.667.49 ± 0.57
ConvPrompt [35]85.58 ± 0.505.34 ± 0.2578.83 ± 0.294.03 ± 0.7967.32 ± 0.305.04 ± 0.2960.01 ± 0.346.02 ± 0.52
RainbowPrompt (ours)89.03 ± 0.914.91 ± 0.3084.05 ± 1.273.59 ± 0.1769.87 ± 0.243.96 ± 0.2870.55 ± 1.575.47 ± 0.47
+ +were prepared using the temporal segment sampling strategy outlined in [43]. We evaluated the proposed approach on the 10-task and 20-task settings for each dataset, with the results presented in Tab. 2. + +In the 10-task setting on UCF-101, RainbowPrompt outperforms L2P, DualPrompt, DualPrompt-PGP, CODA-Prompt, CPrompt, and ConvPrompt in average accuracy by $10.68\%$ , $5.88\%$ , $4.64\%$ , $4.26\%$ , $1.87\%$ , and $3.45\%$ , respectively. Also, it reduces forgetting by $0.55\%$ , $0.50\%$ , $1.01\%$ , $0.90\%$ , $0.39\%$ , and $0.43\%$ compared to these methods. In the 20-task setting, it outperforms all competing methods by a larger gap than in the 10-task setting. Specifically, it achieves an average improvement of $7.95\%$ in accuracy and a $3.92\%$ reduction in forgetting compared to other methods. This is because a larger number of tasks enhances diversity of accumulated knowledge, allowing for a more effective evolution process. + +In the 10-task setting on ActivityNet, RainbowPrompt outperforms CPrompt and ConvPrompt by achieving $3.06\%$ higher accuracy and $3.29\%$ lower forgetting than CPrompt and an additional $2.55\%$ accuracy gain and $1.08\%$ lower forgetting over ConvPrompt. In the 20-task setting, it achieves a significant performance gain over ConvPrompt, with a $10.54\%$ improvement in accuracy and a $0.55\%$ reduction in forgetting. Notably, while ConvPrompt experiences a $7.31\%$ drop in accuracy from its 10-task setting, RainbowPrompt not only prevents degradation but instead improves + +Table 3. Ablation study of RainbowPrompt on ImageNet-R. + +
TCTLTFLTTGAAPA10(↑)F10(↓)
✓✓✓✓✓79.09 ± 0.133.90 ± 0.23
-✓✓✓✓78.92 ± 0.064.19 ± 0.14
✓-✓✓✓78.70 ± 0.414.14 ± 0.00
✓✓-✓✓78.57 ± 0.114.29 ± 0.36
✓✓✓-✓66.31 ± 0.484.84 ± 0.26
✓✓✓✓-78.13 ± 0.224.07 ± 0.19
+ +by $0.68\%$ . The diversity-enhancing integration of RainbowPrompt promotes predictive discriminability [6], enabling it to handle complex action classes more effectively. + +# 4.4. Analysis + +Ablation Study. We conducted an ablation study on ImageNet-R in the 10-task setting to evaluate the impacts of the key components in RainbowPrompt. Tab. 3 presents the effect of excluding each component, including the task-conditioning step (TC), task-level transformation (TLT), feature-level transformation (FLT), task-guided alignment (TGA), and adaptive prompting (AP). Excluding the task-conditioning step results in small declines in average accuracy and forgetting, suggesting that it provides supplementary task-specific information. Compared to the task-level transformation, which captures overall task dependencies, the feature-level transformation plays a more critical + +![](images/901a082b98a2b593af9991033e0454d4c2f1cb3a40ec99d90c5b018a4bd469e8.jpg) +Figure 5. Visualization of prompt insertion patterns across layers: manual prompting (red box) [35] vs. the proposed adaptive prompting (green boxes). Gray represents layers with inserted prompts, whereas white denotes layers without prompts. Accuracy and forgetting gaps denote the average differences in performance across all seen classes, measured as the difference between ours and manual prompting. + +role by capturing fine-grained feature interactions, leading to a larger performance drop when removed. Excluding the task-guided alignment leads to the most significant decline in performance, underscoring its essential role in Rainbow-Prompt. It aligns prompts based on their transformed representations, creating a coherent prompt that facilitates new task learning and integrates diverse knowledge. We also observe that replacing manual prompting with an adaptive approach yields performance gains through task-specific tuning strategies. + +Adaptive Prompting. We conducted an empirical analysis to demonstrate the advantages of the proposed adaptive prompting, as shown in Fig. 5. We visualize the task-specific prompt insertion layers of manual [35] and adaptive prompting in the 10-task setting using ImageNet-R and CUBS. Manual prompting applies predetermined prompt-insertion layers uniformly across all tasks [11, 38], regardless of the dataset. In contrast, our adaptive prompting dynamically determines the optimal prompt-insertion layers for each dataset and task. For ImageNet-R, adaptive prompting outperforms manual prompting, achieving a $1.98\%$ accuracy gain and $0.26\%$ less forgetting across all tasks. We observe a drop in accuracy in the first three tasks, while forgetting remains consistently lower or comparable to manual prompting. Note here that integrating historical knowledge is crucial for transferring knowledge to new tasks, but its absence in early sequential learning limits accuracy as shown in the figure. For CUBS, adaptive prompting achieves a $3.61\%$ higher average accuracy and $2.34\%$ lower forgetting than manual prompting. Our approach demonstrates exceptional performance across most tasks, with even larger performance gaps than those observed on ImageNet-R. + +Sensitivity. To assess the robustness of our method to hyper-parameter variations, we performed sensitivity analysis on the prompt length $(L_{p})$ and the dimensions of the projection and weight matrices $(D_{p}, D_{n})$ . As illustrated in Fig. 6, RainbowPrompt consistently outperforms the strong baselines, CODA-Prompt and ConvPrompt, across various + +![](images/962652ce4f9b1227d974da16a2bc22c630b8823db967e440522644ddc046ca96.jpg) +Figure 6. Sensitivity analysis on ImageNet-R. + +![](images/8abd048d8274f452bc3ab57a61f6eec4df826bc8243f93718747791651030a41.jpg) + +![](images/ffb3d5b1bf2bf97238d2451b2e349c4e5b65456e01f55aa97bcfd74fb2c2f64e.jpg) + +$L_{p}$ values (see first figure). Additionally, varying $D_{p}$ or $D_{n}$ shows low sensitivity to changes in dimensionality, as shown in the second and third figures. These results indicate that our method is robust to hyper-parameter choices and maintains stable performance. + +Efficiency. To evaluate the computational efficiency of our approach, we assessed performance on CIFAR-100 (20-task). CODA-Prompt and ConvPrompt incur 33.7B and 17.1B MACs with 4.8M and 5.7M trainable parameters. Our method trains with 8.2M parameters, including prompt-evolving components, but discards $76.5\%$ (6.2M) at inference, requiring only 18.5B MACs. This offers a favorable balance between accuracy and efficiency. + +# 5. Conclusion + +We have proposed RainbowPrompt, a novel prompt-evolving mechanism to address a fundamental challenge in sequential task learning, which effectively integrates accumulated historical knowledge without compromising previously learned information. Unlike existing methods, RainbowPrompt progressively improves the representations of task-specific prompts for new tasks to promote integration that enhances diversity. By dynamically regulating layer activation with a learnable probabilistic gate, our method optimizes the evolution process based on task-specific differences. Extensive experiments on image classification and video action recognition benchmarks demonstrate that RainbowPrompt consistently outperforms state-of-the-art methods, achieving significant improvements in accuracy and forgetting across diverse scenarios. + +Acknowledgements. This work was supported in part by the National Research Foundation of Korea (NRF) grant funded by the Korea government(MSIT) (RS-2023-00279019) and in part by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) [RS-2021-II211341, Artificial Intelligence Graduate School Program (Chung-Ang University)]. + +# References + +[1] 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 +[2] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 2, 5 +[3] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In Proceedings of the IEEE/CVF International conference on computer vision, pages 9516-9525, 2021. 3 +[4] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420, 2018. 3 +[5] Zhiyuan Chen and Bing Liu. Lifelong machine learning. Morgan & Claypool Publishers, 2018. 3 +[6] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Liang Li, Qingming Huang, and Qi Tian. Towards discriminability and diversity: Batch nuclear-norm maximization under label insufficient situations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3941-3950, 2020. 2, 7 +[7] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales 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. 2 +[8] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 6 +[9] Robert M French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 3(4): 128-135, 1999. 1, 2 +[10] Yang Gao, Oscar Beijbom, Ning Zhang, and Trevor Darrell. Compact bilinear pooling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 317-326, 2016. 4 + +[11] Zhanxin Gao, Jun Cen, and Xiaobin Chang. Consistent prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28463-28473, 2024. 2, 3, 6, 7, 8 +[12] Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are key-value memories. arXiv preprint arXiv:2012.14913, 2020. 5 +[13] Dongyoon Han, Sangdoo Yun, Byeongho Heo, and YoungJoon Yoo. Rethinking channel dimensions for efficient model design. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 732-741, 2021. 2 +[14] 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 Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 2, 5 +[15] Kiseong Hong, Hyundong Jin, Sungho Suh, and Eunwoo Kim. Exploration and exploitation in continual learning. Neural Networks, 188:107444, 2025. 2, 5 +[16] Y Hsu. Re-evaluating continual learning scenarios: A categorization and case for strong baselines. arXiv preprint arXiv:1810.12488, 2018.3 +[17] 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 +[18] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5 +[19] Hyundong Jin and Eunwoo Kim. Helpful or harmful: Inter-task association in continual learning. In European Conference on Computer Vision, pages 519-535. Springer, 2022. 3 +[20]Hyundong Jin, Gyeong-hyeon Kim, Chanho Ahn, and Eunwoo Kim. Growing a brain with sparsity-inducing generation for continual learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 18961-18970, 2023. 1 +[21] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11847-11857, 2023. 3 +[22] Sangwon Jung, Hongjoon Ahn, Sungmin Cha, and Taesup Moon. Continual learning with node importance based adaptive group sparse regularization. Advances in Neural Information Processing Systems, 33:3647-3658, 2020. 1, 3 + +[23] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 3 +[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 5 +[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 3 +[26] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3 +[27] 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 +[28] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35, 2023. 3 +[29] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in Neural Information Processing Systems, 30, 2017. 3 +[30] Chengcheng Ma, Yang Liu, Jiankang Deng, Lingxi Xie, Weiming Dong, and Changsheng Xu. Understanding and mitigating overfitting in prompt tuning for vision-language models. IEEE Transactions on Circuits and Systems for Video Technology, 33(9): 4616-4629, 2023. 3 +[31] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359-17372, 2022. 5 +[32] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11321-11329, 2019. 3 +[33] Jingyang Qiao, Xin Tan, Chengwei Chen, Yanyun Qu, Yong Peng, Yuan Xie, et al. Prompt gradient projection for continual learning. In The Twelfth International Conference on Learning Representations, 2023. 1, 2, 3, 6, 7 +[34] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Pro + +ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 3 +[35] 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. 1, 2, 3, 5, 6, 7, 8 +[36] Gobinda Saha, Isha Garg, and Kaushik Roy. Gradient projection memory for continual learning. arXiv preprint arXiv:2103.09762, 2021. 1 +[37] 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 +[38] James Seale Smith, Leonid Karlinsky, Vysshnavi 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, 7, 8 +[39] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 5 +[40] Andrés Villa, Kumail Alhamoud, Victor Escorcia, Fabian Caba, Juan León Alcazar, and Bernard Ghanem. vclimb: A novel video class incremental learning benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19035-19044, 2022. 5 +[41] Andrés Villa, Juan León Alcazar, Motasem Alfarra, Kumail Alhamoud, Julio Hurtado, Fabian Caba Heilbron, Alvaro Soto, and Bernard Ghanem. Pivot: Prompting for video continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24214-24223, 2023. 6 +[42] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 5 +[43] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 6, 7 +[44] 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, 5, 6, 7 + +[45] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi 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, 7 +[46] Chenshen Wu, Luis Herranz, Xialei Liu, Joost Van De Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. Advances in Neural Information Processing Systems, 31, 2018. 3 +[47] 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. 3 +[48] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 3 \ No newline at end of file diff --git a/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/images.zip b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e2c9527984149816d847a3d90da755c168af3676 --- /dev/null +++ b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b66dc45741a85824fecccb97b46d764d3df21a61db5ed59d3bd2618ebcaa3c34 +size 577058 diff --git a/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/layout.json b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..07ad18835a03874d4072f539f380986961dff6e8 --- /dev/null +++ b/ICCV/2025/RainbowPrompt_ Diversity-Enhanced Prompt-Evolving for Continual Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70e78b59cbac95abee4cf53475ec43c10178b2f2502c246d33c2cccab2b4a1bd +size 458029 diff --git a/ICCV/2025/Randomized Autoregressive Visual Generation/7a271cf3-67f5-4d9e-8304-b922e236ce90_content_list.json b/ICCV/2025/Randomized Autoregressive Visual Generation/7a271cf3-67f5-4d9e-8304-b922e236ce90_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..43d573c35b2b81f3013a37907940494e13bfe18f --- /dev/null +++ b/ICCV/2025/Randomized Autoregressive Visual Generation/7a271cf3-67f5-4d9e-8304-b922e236ce90_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec0eca5e15f6e10b3316f570e99f58a2cfa64e0659d2f52b5f7320ce4f169bc9 +size 89479 diff --git a/ICCV/2025/Randomized Autoregressive Visual Generation/7a271cf3-67f5-4d9e-8304-b922e236ce90_model.json b/ICCV/2025/Randomized Autoregressive Visual Generation/7a271cf3-67f5-4d9e-8304-b922e236ce90_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5c66cea962cadf477788775ff317901ded588782 --- /dev/null +++ b/ICCV/2025/Randomized Autoregressive Visual Generation/7a271cf3-67f5-4d9e-8304-b922e236ce90_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df86e958ee30847568fd1a8a2b0ef642deaa94b038c229a2ba4114703c9fb6ff +size 112222 diff --git a/ICCV/2025/Randomized Autoregressive Visual Generation/7a271cf3-67f5-4d9e-8304-b922e236ce90_origin.pdf b/ICCV/2025/Randomized Autoregressive Visual Generation/7a271cf3-67f5-4d9e-8304-b922e236ce90_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bb61c79cc8865dddbc7089b7501f203507ab581a --- /dev/null +++ b/ICCV/2025/Randomized Autoregressive Visual Generation/7a271cf3-67f5-4d9e-8304-b922e236ce90_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be1f3672cce1102e418569435964f8377d4ce52e2d301ba972777f5d0d995599 +size 1704573 diff --git a/ICCV/2025/Randomized Autoregressive Visual Generation/full.md b/ICCV/2025/Randomized Autoregressive Visual Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..706660804ac46edbf17a31f4257feef74390eb80 --- /dev/null +++ b/ICCV/2025/Randomized Autoregressive Visual Generation/full.md @@ -0,0 +1,348 @@ +# Randomized Autoregressive Visual Generation + +Qihang Yu Ju He Xueqing Deng Xiaohui Shen Liang-Chieh Chen ByteDance + +# Abstract + +This paper presents Randomized AutoRegressive modeling (RAR) for visual generation, which sets a new state-of-the-art performance on the image generation task while maintaining full compatibility with language modeling frameworks. The proposed RAR is simple: during a standard autoregressive training process with a next-token prediction objective, the input sequence—typically ordered in raster form—is randomly permuted into different factorization orders with a probability $r$ , where $r$ starts at 1 and linearly decays to 0 over the course of training. This annealing training strategy enables the model to learn to maximize the expected likelihood over all factorization orders and thus effectively improve the model's capability of modeling bidirectional contexts. Importantly, RAR preserves the integrity of the autoregressive modeling framework, ensuring full compatibility with language modeling while significantly improving performance in image generation. On the ImageNet-256 benchmark, RAR achieves an FID score of 1.48, not only surpassing prior state-of-the-art autoregressive image generators but also outperforming leading diffusion-based and masked transformer-based methods. Code and models are available at https://github.com/bytedance/1d-tokenizer. + +# 1. Introduction + +AutoRegressive (AR) models have driven remarkable advancements across both natural language processing and computer vision tasks in recent years. In language modeling, they serve as the fundamental framework for Large Language Models (LLMs) such as GPT [45], Llama [64, 65], and Gemini [62], along with other state-of-the-art models [1, 72]. In the realm of computer vision, autoregressive models1 have also shown substantial potential, delivering competitive performance in image generation tasks [22, 37, 41, 52, 53, 57, + +![](images/3be0da59bca749074c0640dd70410304dcb011677d5c7bf04a31892bd0e24144.jpg) +Figure 1. Comparison among different language modeling compatible autoregressive (AR) image generators. The proposed RAR demonstrates significant improvements over previous AR methods. RAR-B, with only $261M$ parameters, achieves an FID score of 1.95, outperforming both LlamaGen-XXL (1.4B parameters) and Open-MAGVIT2-XL (1.5B parameters). + +75, 76] to diffusion models [6, 18, 29, 38, 47, 55] or non-autoregressive transformers [10, 34, 70, 77-79]. More importantly, autoregressive modeling is emerging as a promising pathway toward unified models across multiple modalities and tasks [5, 9, 14, 60, 61, 71]. + +Despite the dominance of autoregressive models in language modeling, they often yield suboptimal performance in comparison to diffusion models or non-autoregressive transformers in visual generation tasks [41, 57]. This discrepancy can be attributed to the inherent differences between text and visual signals. Text is highly compact and semantically meaningful, while visual data tends to be more low-level and redundant [30, 79], making bidirectional context modeling more critical. For instance, several studies [7, 21, 38] have demonstrated that causal attention applied to image tokens leads to inferior performance compared to bidirectional attention in vision tasks. + +To address this, recent works [38, 63] have attempted to reintroduce bidirectional attention by redesigning the autoregressive formulation, achieving state-of-the-art results in image generation. However, these approaches often deviate from the traditional autoregressive paradigm. For example, + +![](images/f41813e4148761275fa6141eb4cd4c3e37f6a37e33b5a4af9a89126009db3bc2.jpg) +Figure 2. Overview of the proposed Randomized AutoRegressive (RAR) model, which is fully compatible with language modeling frameworks. Left: RAR introduces a randomness annealing training strategy to enhance the model's ability to learn bidirectional contexts. During training, the input sequence is randomly permuted with a probability $r$ , which starts at 1 (fully random permutations) and linearly decreases to 0, transitioning the model to a fixed scan order, such as raster scan, by the end of training. Right: Randomly selected images generated by RAR, trained on ImageNet. + +![](images/4eb648e6734b454ea321a8164a12621231c0a4ebacc89a1329d3065735102d11.jpg) + +![](images/9e782cacf0ce9a8d01f21e6c5151ca53e52a331919c66a532b0a5ad9dd9d1fb2.jpg) + +![](images/ebb6792280826fcaeff3129ac7b27911449d3c6cfca3d857dae10a19f6ccce3a.jpg) + +![](images/888d5cfa30075e25b2f97192a28faa503cc3d325d3e7b0ca5e0f91cf8130edeb.jpg) + +![](images/04fe53e993f2264e2697cbcc27dc0ab4d8fbea6ff5bd9478e63e0c9e552dc0de.jpg) + +![](images/12017a4ae40e4a116dbb45f0fbd2fbf95c178de4dd0d2ceb7f99688874d25183.jpg) + +![](images/ae436e9ef9079e469bda47b79235dfc3ffac280e18cfc8affd1dccca444fa90e.jpg) + +![](images/8dd2307319831bfb8b92b5a1d4b6b377ec7fe54cfd86a286a177ec26db87c1c2.jpg) + +![](images/f26fa60c5b7f0db5ca079a2dad5c90f89c320ebedd4c373d07534939a80e4cfd.jpg) + +![](images/f4a00fa878edb195f51c9e88f9c5139ea2e82d50a8a74d9f7c8ff9240db8e582.jpg) + +![](images/0b39574914676b95f3a03909501d2daa6df1eb0038d40b6a546d19a04dd44864.jpg) + +![](images/ba40cba81fb0fb7ab03c6618ca3d47f9d2e5f7e2de1d026cfd44d028bd797884.jpg) + +![](images/11f65831f6f0da3ee01aef7e59bc82867e40c19e3ae50068641464b6f855052a.jpg) + +![](images/320104c1117243176b97422d09471eb7522dc7ca631f5d644a231537401fadd1.jpg) + +![](images/ca32ffb4b13179952b1da677e2222802482d8a8fa8a0acc79e9b8cdc4159205f.jpg) + +![](images/79823f7cc47611ec8a2e3378b4f1d4651a8547d65eb112e0844036fd89a0fe05.jpg) + +![](images/c5513fe0566b5bfcf0c436f93bef5fadec3343d4032f5cbaca593b9d1eddfb21.jpg) + +![](images/22de1fff3dfd803ef209a55ccd4299ca9e07c33c42b2d062392d505e189260d4.jpg) + +VAR [63] shifts from next-token prediction to next-scale prediction, enabling bidirectional attention within each scale, and MAR [38] generalizes MaskGIT-style framework [10] to the autoregressive definition, which naturally introduces back the bidirectional attention. While effective, these modifications complicate their integration into universal transformer architectures that aim to unify different modalities, which proves to work well with conventional autoregressive modeling [60, 61]. + +In this paper, we aim to enhance the generation quality of autoregressive image models while preserving the core autoregressive structure, maintaining compatibility with language modeling frameworks. Specifically, we enable bidirectional context learning within an autoregressive transformer by maximizing the expected likelihood over all possible factorization order. In this way, all tokens will be trained and predicted under all possible contexts, facilitating learning bidirectional representation. Moreover, we introduce a permutation probability $r$ , which controls the ratio of training data between a random factorization order and the standard raster order. Initially, $r$ is set to 1 (fully random factorization) and it linearly decays to 0 over the course of training, gradually reverting the model to the raster order commonly used by other autoregressive image generators. + +To this end, we present a simple, effective, and scalable autoregressive model training paradigm named Randomized AutoRegressive modeling (RAR). RAR retains the original autoregressive model architecture and formulation, ensuring full compatibility with language modeling. At the same time, it significantly improves the generation quality of autoregressive models at no additional cost. On the + +ImageNet-256 benchmark [16], RAR achieves an FID score of 1.48, substantially outperforming previous state-of-the-art autoregressive image generators, as illustrated in Fig. 1. By addressing the limitations of unidirectional context modeling, RAR represents a critical step towards autoregressive visual generation and opens up new possibilities for further advancements in the field. + +# 2. Related Work + +Autoregressive Language Modeling. The advent of autoregressive language models [1-4, 9, 13, 20, 45, 48, 49, 62, 64, 65, 72] has paved a promising path toward general-purpose AI systems. At the core of these models is a simple yet powerful next-token prediction paradigm, where the objective is to predict the next word or token in a sequence based on preceding inputs. This approach has demonstrated both scalability, as evidenced by scaling laws, and versatility through zero-shot generalization, enabling explorations beyond traditional language tasks to diverse modalities. + +Autoregressive Visual Modeling. Pioneering research [12, 27, 46, 67, 68] in autoregressive visual modeling has focused on representing images as sequences of pixels. Nevertheless, inspired by advancements in autoregressive language modeling, a subsequent wave of studies has transitioned to modeling images as sequences of discrete-valued tokens [22, 50, 51, 69, 75], resulting in notable improvements in performance. This direction has been further explored through efforts [41, 57] aimed at enhancing tokenization quality and leveraging modern autoregressive architectures initially developed for language tasks. However, all of these + +works strictly adhere to a raster-scan order for processing pixels or tokens, resulting in a unidirectional information flow that is sub-optimal for visual modeling. In this work, we instead explore learning across all possible factorization orders to enhance bidirectional context learning while retaining the core autoregressive framework. + +Other Visual Generation Models. In addition to autoregressive visual modeling, there have been numerous efforts in exploring other formats of visual generation models, including generative adversarial networks (GANs) [8, 26, 33], diffusion models [18, 23, 32, 39, 47, 54, 73], masked transformers [10, 11, 70, 77, 79], scale-wise autoregressive modeling (VAR) [43, 59, 63, 81], and masked autoregressive modeling with diffusion loss (MAR) [24, 38]. It is worth noting that MAR [38] also experimented a random order based AR framework similar to the proposed RAR. However, as indicated in our experiments (see Sec. 4.2), simply replacing the raster order with random order only brings marginal improvement, coinciding the observation in [38]. This further demonstrates the importance on the randomness annealing strategy in RAR, leading to a substantial improvement for the AR image generators. + +# 3. Method + +In this section, we first provide an overview of autoregressive modeling in Sec. 3.1, followed by our proposed Randomized AutoRegressive modeling (RAR) in Sec. 3.2. + +# 3.1. Background + +We provide a brief overview of autoregressive modeling with a next-token prediction objective. Given a discrete token sequence $\mathbf{x} = [x_1, x_2, \dots, x_T]$ , the goal of autoregressive modeling is to maximize the likelihood of the sequence under a forward autoregressive factorization. Specifically, the objective is to maximize the joint probability of predicting the current token $x_t$ based on all preceding tokens $[x_1, x_2, \dots, x_{t-1}]$ , $\forall t = 1, \dots, T$ : + +$$ +\max _ {\theta} p _ {\theta} (\mathbf {x}) = \prod_ {t = 1} ^ {T} p _ {\theta} \left(x _ {t} \mid x _ {1}, x _ {2}, \dots , x _ {t - 1}\right), \tag {1} +$$ + +where $p_{\theta}$ denotes a token distribution predictor with a model parameterized by $\theta$ + +As shown in the equation, each token $x_{t}$ at position $t$ is conditioned solely on the preceding tokens, which limits context modeling to a unidirectional manner. This contrasts with methods such as masked transformer [10, 70, 77, 78] and diffusion models [32, 39, 47, 54], which can leverage bidirectional context at the training time. Additionally, while natural language has an inherent sequential order (left-to-right in most languages), image data lacks a clear, predefined order for processing tokens. Among the possible orders for image generation, the row-major order (i.e., raster scan) + +is the most widely adopted and has demonstrated superior performance compared to other alternatives [22]. + +# 3.2. RAR: Randomized AutoRegressive Modeling + +Visual signals inherently exhibit bidirectional correlations, making effective global context modeling essential. However, conventional autoregressive models rely on causal attention masking, which enforces a unidirectional dependency on the token sequence, contradicting the nature of visual data, as noted in prior works [7, 21, 38], where bidirectional attention works significantly better than causal attention for visual modality. Furthermore, there is no universally "correct" way to arrange image tokens into a causal sequence. While the widely adopted raster order has achieved some success, it introduces biases in the autoregressive training process. For instance, each token is conditioned solely on the preceding tokens in the scanning order, restricting the model's ability to learn dependencies from other directions. + +To address these challenges, we propose a randomized autoregressive modeling approach that incorporates optimization objective with bidirectional context: + +$$ +\max _ {\theta} p _ {\theta} (\mathbf {x}) = \prod_ {t = 1} ^ {T} p _ {\theta} \left(x _ {t} \mid x _ {1}, \dots , x _ {t - 1}, x _ {t + 1}, \dots , x _ {T}\right). \tag {2} +$$ + +Unlike BERT-style [17] or MaskGIT-style [10] methods, our method follows the permuted objective approach [66, 74], where the model is trained in an autoregressive manner across all possible factorization orders. This enables the model to gather bidirectional context while preserving the autoregressive framework in expectation. Formally, we have: + +$$ +\max _ {\theta} p _ {\theta} (\mathbf {x}) = \mathbb {E} _ {\tau \sim \mathcal {S} _ {T}} \left[ \prod_ {t = 1} ^ {T} p _ {\theta} \left(x _ {\tau_ {t}} \mid x _ {\tau_ {< t}}\right) \right], \tag {3} +$$ + +where $S_T$ denotes the set of all possible permutations of the index sequence $[1, 2, \dots, T]$ , and $\tau$ represents a randomly sampled permutation from $S_T$ . The notation $\tau_t$ refers to the $t$ -th element in the permuted sequence, and $\tau_{ 10^{506}$ , which can overwhelm the model and reduce training efficiency. (2) Although images can be processed in arbitrary orders, certain scan orders tend to outperform others. For instance, [22] evaluated six different scan orders (row-major, spiral in, spiral out, z-curve, subsample, and alternate) and found that row-major (i.e., raster order) consistently performed the best, a result that has made it the most widely used order for visual generation. + +To address these issues, we propose Randomness Annealing, a strategy designed to balance the randomness of permutations with the known effectiveness of the raster order. This method introduces a single parameter, $r$ , which controls the probability of using a random permutation versus the raster order. At the start of training, $r = 1$ , meaning that the model exclusively uses random permutations. Over the course of training, $r$ linearly decays to 0, transitioning the model to the raster order by the end of training. Specifically, we define a training schedule for $r$ , controlled by two hyper-parameters start and end indicating the training epoch when $r$ starts to anneal and when the annealing ends. Formally, we have: + +$$ +r = \left\{ \begin{array}{l l} 1. 0, & \text {i f e p o c h < s t a r t}, \\ 0. 0, & \text {i f e p o c h > e n d}, \\ 1. 0 - \frac {\text {e p o c h - s t a r t}}{\text {e n d - s t a r t}}, & \text {o t h e r w i s e}, \end{array} \right. \tag {5} +$$ + +where epoch is the current training epoch. We will ablate the hyper-parameters start and end in the experiments. + +The schedule allows the model to initially explore the diverse random permutations for better bidirectional representation learning, and ultimately converge to the more effective row-major scan order for better visual generation quality, as is used by other typical autoregressive methods [22]. It is worth noting that this strategy not only improves generation performance but also maintains compatibility with the optimization techniques for standard scan orders used in previous works. + +# 4. Experimental Results + +In this section, we outline the implementation details of our method in Sec. 4.1. Next, we present ablation studies on key + +
modeldepthwidthmlpheads#params
RAR-B24768307216261M
RAR-L241024409616461M
RAR-XL321280512016955M
RAR-XXL4014086144161499M
+ +Table 1. Architecture configurations of RAR. We follow prior works scaling up ViT [19, 80] for different configurations. + +design choices in Sec. 4.2. The main results are discussed in Sec. 4.3, followed by scaling study and visualizations. + +# 4.1. Implementation Details + +We implement the RAR on top of language modeling autoregressive framework with minimal changes. + +VQ Tokenizer. Following prior works [10, 22] which use a VQ tokenizer to tokenize the input images into discrete token sequences, we use the MaskGIT-VQGAN [10] with the official weight trained on ImageNet. This tokenizer is a purely CNN-based tokenizer which tokenizes a $256 \times 256$ image into 256 discrete tokens (i.e., downsampling factor 16) with a codebook size (i.e., vocabulary size) 1024. + +Autoregressive Transformer. We use vision transformers [19] of different model configurations [80] including RAR-S (133M), RAR-B (261M), RAR-L (461M), RAR-XL (955M), and RAR-XXL (1499M). For all of these model variants, we apply causal attention masking in the self-attention module and QK LayerNorm [15] to stabilize the large-scale model training. We use plain ViT for all ablation studies to speed up the experiments, and we enhance the model with adaLN [47] for final models. The detailed architecture configuration and model size are available at Tab. 1. + +Positional Embedding. We use learnable embeddings for both original positional embedding in ViT and target-aware positional embedding. Notably, as our model anneals to raster order-based autoregressive image generation after the training is finished, the two positional embeddings can be combined into one, making it identical to a conventional autoregressive image generator. + +Dataset. We train our model on ImageNet-1K [16] training set, which contains 1, 281, 167 training images across 1000 object classes. We pre-tokenize the whole training set with MaskGIT-VQGAN tokenizer [10] to speed up the training. For ablation studies, we pre-tokenize the dataset with only center crop and horizontal flipping augmentation, while we further enhance the diversity in pretokenized datasets with ten-crop transformation [57, 58] for final models. + +Training Protocols. We use the same training hyperparameters for all model variants. The model is trained with batch size 2048 for 400 epochs (250k steps). The learning rate will be linearly increased from 0 to $4 \times 10^{-4}$ at the first 100 epochs (warm-up), then it will be gradually decayed to $1 \times 10^{-5}$ following a cosine decay schedule. + +
start epochend epochFID↓IS↑Pre.↑Rec.↑
00†3.08245.30.850.52
01002.68237.30.840.54
02002.41251.50.840.54
03002.40258.40.840.54
04002.43265.30.840.53
1001002.48247.50.840.54
1002002.28253.10.830.55
1003002.33258.40.830.54
1004002.39266.50.840.54
2002002.39259.70.840.54
2003002.18269.70.830.55
2004002.55241.60.840.54
3003002.41269.10.840.53
3004002.74236.40.830.54
400400‡3.01305.60.840.52
+ +We use AdamW [35, 40] optimizer with beta1 0.9, beta2 0.96, and weight decay 0.03. We perform gradient clipping with maximum gradient norm 1.0. During training, the class condition will be dropped at a probability 0.1. The training setting remains the same for both ablation studies and main results across all RAR model variants. + +Sampling Protocols. We sample 50000 images for FID computation using the evaluation code from [18]. We do not use any top-k or top-p based filtering techniques. We also follow prior arts [11, 25, 79] to use classifier-free guidance [31]. In the ablation study, we use a simpler linear guidance schedule [11], and for final models we use the improved power-cosine guidance schedule [25]. The final detailed hyper-parameters for each model variant can be found in the appendix. + +# 4.2. Ablation Studies + +We study different configurations for RAR, including the randomness annealing strategy and scan orders that RAR converges to. + +Randomness Annealing Strategy. In Tab. 2 we compare different randomness annealing strategies. We adopt a linear decaying schedule and focus on when should the randomization annealing starts and ends by changing two hyperparameters start and end, as defined in Eq. (5). For a training lasting for 400 epochs, we enumerate all possible combinations for every 100 epochs. For example, when $start = 200$ and end $= 300$ , the model is trained with random permu + +Table 2. Different start and end epochs for randomness annealing, with a total of 400 training epochs and model size RAR-L. The final setting is labeled in gray. $\dagger$ : When start epoch and end epoch are both 0 (1st row), the training reverts to a standard raster order training. $\ddagger$ : When start epoch and end epoch are both 400 (last row), the training becomes a purely random order training. After training is finished, all results are obtained with raster order sampling, except for the purely random order training (i.e., last row), where we also randomly sample the scan order following [38], which otherwise could not produce a reasonable result. + +
scan orderFID↓IS↑Precision↑Recall↑
row-major2.18269.70.830.55
spiral in2.50256.10.840.54
spiral out2.46256.60.840.54
z-curve2.29262.70.830.55
subsample2.39258.00.840.54
alternate2.48270.90.840.53
+ +Table 3. Effect of different scan orders RAR-L converges to. We mainly consider 6 different scan orders (row major, spiral in, spiral out, z-curve, subsample, alternate) as studied in [22]. Our default setting is marked in gray. A visual illustration of different scan orders are available in the appendix. + +tations from 0 to 200 epochs and raster order from 300 to 400 epochs. During 200 to 300 epoch, the model is trained via random permutation with probability $r$ and raster order with probability $1 - r$ , where $r$ is computed as in Eq. (5). It is noteworthy that when start = end = 0, the model is trained with purely raster order, i.e., the standard autoregressive training. When start = end = 400, the model is always trained with randomly permuted input sequence. Both cases are important baselines of the proposed randomness annealing, and they achieve FID scores of 3.08 and 3.01, respectively. Notably, pure random order training does not bring notable performance improvements. Interestingly, we observe all other variants with randomness annealing achieve substantial improvement over these two baselines. For example, even simply replacing the first 100 epochs of raster order with random permutation, it (i.e., start = 100 and end = 100) improves the FID to 2.48 by 0.6. Besides, we also note that the model prefers to keep some beginning epochs for pure random permutation training and some last epochs for better adapting to raster scan order, which usually leads to a better performance compared to other variants. All the results demonstrate that adding randomized autoregressive training with a permuted objective is beneficial to the autoregressive visual generator and leads to a boosted FID score, thanks to the improved bidirectional representation learning process. + +Additionally, among all variants, we found that the case, where start = 200 and end = 300, works the best, which improves the baseline (purely raster order) FID from 3.08 to 2.18. This strategy allocates slightly more computes on the training with random permutation order, and focuses on the purely raster order for the last 100 epochs. Therefore, we default to adopt this annealing strategy for all RAR models. + +Different Scan Orders Besides Raster. Although row-major order (i.e., raster scan) has been the de facto scan order in the visual generation, there lacks a systematic study on how good it is compared to other scan orders. We note that the work [22] conducted a similar study 4 years ago. However, it is worth re-examining the conclusion considering the significant progress generative models have achieved + +in recent years. Specifically, we consider 6 different scan orders (row-major, spiral in, spiral out, z-curve, subsample, and alternative) following [22] that RAR may converge to. Instead of reporting the training loss and validation loss as the comparison metric [22], we directly evaluate their generation performance. The results are summarized in Tab. 3. Interestingly, we observe that all variants achieve a reasonably good score, which indicates that RAR is capable of handling different scan orders. Considering that the row-major (raster scan) still demonstrates advantages over the other scan orders, we thus use the raster scan order for all final RAR models. + +We provide more ablation experiments regarding training epochs, different tokenizers etc. are available in the supplementary material. + +# 4.3. Main Results + +We report RAR results against state-of-the-art image generators on ImageNet-1K $256 \times 256$ benchmark [16]. + +As shown in Tab. 4, RAR achieves significantly better performance compared to previous AR image generators. Specifically, the most compact RAR-B with 261M parameters only, achieves an FID score 1.95, already significantly outperforming current state-of-the-art AR image generators LlamaGen-3B-384 (3.1B, FID 2.18, crop size 384) [57] and Open-MAGVIT2-XL (1.5B, FID 2.33) [41], while using $91\%$ and $81\%$ fewer model parameters respectively. It also surpasses the widely used diffusion models such as DiT-XL/2 (FID 1.95 vs. 2.27) and SiT-XL (FID 1.95 vs. 2.06) while only using $39\%$ model parameters compared to them. + +In Tab. 4, we further explore RAR at different model sizes (from 261M to 1.5B), where we observe strong scalability behavior with consistent performance improvement as model size scales up. Notably, the largest variant RAR-XXL sets a new state-of-the-art result on ImageNet benchmark, with an FID score 1.48. When compared to the other two recent methods VAR [63] and MAR [38], both of which attempt to amend AR formulation for better visual generation quality, RAR not only demonstrates a superior performance (FID 1.48 from RAR vs. 1.73 from VAR and 1.55 from MAR), but also keeps the whole framework compatible with language modeling and thus is more friendly for adapting the mature optimization and speed-up techniques for large language models to visual generation [57]. + +Moreover, RAR demonstrates superior performance to state-of-the-art visual generators in different frameworks. It performs better against the leading autoregressive models, diffusion models and masked transformer models, surpassing LlamaGen-3B-384 [57], MDTv2-XL/2 [25] and MaskBit [70] respectively (FID 1.48 from RAR vs. 2.18 from LlamaGen, 1.58 from MDTv2, and 1.52 from MaskBit). To the best of our knowledge, this is the first time that the language modeling style autoregressive visual + +
tokenizertypegenerator#paramsFID↓IS↑Pre.↑Rec.↑
VQ [54]Diff.LDM-8 [54]258M7.76209.50.840.35
VAE [54]Diff.LDM-4 [54]400M3.60247.70.870.48
UViT-L/2 [6]287M3.40219.90.830.52
UViT-H/2 [6]501M2.29263.90.820.57
DiT-L/2 [47]458M5.02167.20.750.57
VAE [56]Diff.DiT-XL/2 [47]675M2.27278.20.830.57
SiT-XL [42]675M2.06270.30.820.59
DiMR-XL/2R [39]505M1.70289.00.790.63
MDTv2-XL/2 [25]676M1.58314.70.790.65
VQ [10]Mask.MaskGIT [10]177M6.18182.1--
VQ [79]Mask.TiTok-S-128 [79]287M1.97281.8--
VQ [78]Mask.MAGVIT-v2 [78]307M1.78319.4--
VQ [70]Mask.MaskBit [70]305M1.52328.6--
VAE [38]MARMAR-B [38]208M2.31281.70.820.57
MAR-L [38]479M1.78296.00.810.60
MAR-H [38]943M1.55303.70.810.62
VQ [63]VARVAR-d30 [63]2.0B1.92323.10.820.59
VAR-d30-re [63]2.0B1.73350.20.820.60
VQ [22]ARGPT2 [22]1.4B15.7874.3--
GPT2-re [22]1.4B5.20280.3--
VQ [75]ARVIM-L [75]1.7B4.17175.1--
VIM-L-re [75]1.7B3.04227.4--
Open-MAGVIT2-B [41]343M3.08258.30.850.51
VQ [41]AROpen-MAGVIT2-L [41]804M2.51271.70.840.54
Open-MAGVIT2-XL [41]1.5B2.33271.80.840.54
LlamaGen-L [57]343M3.80248.30.830.51
LlamaGen-XL [57]775M3.39227.10.810.54
LlamaGen-XXL [57]1.4B3.09253.60.830.53
LlamaGen-3B [57]3.1B3.05222.30.800.58
LlamaGen-L-384 [57]343M3.07256.10.830.52
LlamaGen-XL-384 [57]775M2.62244.10.800.57
LlamaGen-XXL-384 [57]1.4B2.34253.90.800.59
LlamaGen-3B-384 [57]3.1B2.18263.30.810.58
VQ [57]ARRAR-B (ours)261M1.95290.50.820.58
RAR-L (ours)461M1.70299.50.810.60
RAR-XL (ours)955M1.50306.90.800.62
RAR-XXL (ours)1.5B1.48326.00.800.63
+ +Table 4. ImageNet-1K $256 \times 256$ generation results evaluated with ADM [18]. "type" refers to the type of the generative model, where "Diff." and "Mask." stand for diffusion models and masked transformer models, respectively. "VQ" denotes discrete tokenizers and "VAE" stands for continuous tokenizers. "-re" stands for rejection sampling. "-384" denotes for generating images at resolution 384 and resize back to 256 for evaluation, as is used in [57]. + +generators outperform state-of-the-art diffusion models and masked transformer models. + +Sampling Speed. One key advantage of AR methods is their ability to leverage established optimization techniques from LLMs, such as KV-caching. In Tab. 5, we compare the sampling speed (measured as images/sec) of RAR against other types of generative models, such diffusion models [47], masked transformers [70, 79], VAR [63], and MAR [38]. Among them, AR models (RAR) and VAR models (VARd30) are compatible with the KV-cache optimization, providing a significant advantage in generation speed over other methods. As shown in Tab. 5, RAR achieves a state-of- + +![](images/61e3d24f8c0913cca4b00b2d53008536c1afc589435269b544d1ae5d8a44b7da.jpg) +(a) training losses + +![](images/02c090cfd618f0ad4e561807b96076649f95caaffbddd566566ede027fa51a1f.jpg) +(b) FID scores w/o classifier-free guidance + +![](images/7a2ca73e82ff2a38ecfab743767e1f39eb7ea9b4db95b3c3ee2ae3f2ab3a1048.jpg) +(c) FID scores w/ classifier-free guidance + +
methodtype#paramsFID↓stepsimages/sec
DiT-XL/2 [47]Diff.675M2.272500.6
TiTok-S-128 [79]Mask.287M1.97647.8
VAR-d30 [63]VAR2.0B1.921017.3
MAR-B [38]MAR208M2.312560.8
RAR-B (ours)AR261M1.9525617.0
MAR-L [38]MAR479M1.782560.5
RAR-L (ours)AR461M1.7025615.0
MaskBit [70]Mask.305M1.522560.7
MAR-H [38]MAR943M1.552560.3
RAR-XL (ours)AR955M1.502568.3
RAR-XXL (ours)AR1.5B1.482566.4
+ +Table 5. Sampling throughput comparison (including tokenization process) categorized by methods with similar FID scores. Throughputs are measured as samples generated per second on a single A100 using float32 precision and a batch size of 128, based on their official codebases. For VAR [63] and our RAR, KV-cache is applied. "Diff." and "Mask." refer to diffusion models and masked transformer models, respectively. + +the-art FID score while also significantly surpassing other methods in generation speed. For instance, at an FID score around 1.5, MaskBit [70] and MAR-H [38] generate image samples at 0.7 and 0.3 images per second, respectively. In comparison, RAR-XL not only achieves a better FID score but can generate 8.3 high-quality visual samples per second— $11.9 \times$ faster than MaskBit and $27.7 \times$ faster than MAR-H. The largest RAR variant, RAR-XXL, further improves the FID score while maintaining a notable speed advantage, being $9.1 \times$ faster than MaskBit and $21.3 \times$ faster than MAR-H. Additionally, RAR may benefit further from LLM optimization techniques such as vLLM [36], as seen with other AR methods [57]. + +Scaling Behavior. We study the scaling behavior of RAR. Specifically, we plot the training loss curves and FID score curves (with and without classifier-free guidance [31]) in Fig. 4. As shown in the figure, we observe that RAR scales well at different model sizes, where larger model size leads to a consistently lower training loss and better FID score, regardless of using the enhancement of classifier-free + +![](images/94e40e0d32de89e484382976d423c071ccadfe9939f1aa706262f6388284146e.jpg) +Figure 4. Scaling behavior of RAR models. The scaled-up RAR models demonstrate (a) reduced training losses, and improved FID scores both (b) without and (c) with classifier-free guidance. +Figure 5. Visualization of samples generated by RAR across various model sizes. RAR generates high-quality visual samples across all model sizes. As model size increases, fidelity and diversity improve, especially in challenging classes (e.g., dogsled). + +guidance or not. We note that as RAR keeps the AR formulation and framework intact, it also inherits the scalability from AR methods. + +Visualization. We visualize generated samples by different RAR variants in Fig. 5, which shows that RAR is capable of generating high-quality samples with great fidelity and diversity. More visualizations are provided in the appendix. + +# 5. Conclusion + +In this paper, we introduced a simple yet effective strategy to enhance the visual generation quality of language modeling-compatible autoregressive image generators. By employing a randomized permutation objective, our approach enables improved bidirectional context learning while preserving the autoregressive structure. Consequently, the proposed RAR model not only surpasses previous state-of-the-art autoregressive image generation methods but also outperforms leading non-autoregressive transformer and diffusion models. We hope this research contributes to advancing autoregressive transformers toward a more powerful unified framework for visual understanding and generation. + +# References + +[1] 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. 1, 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. +[3] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. +[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. 2 +[5] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In CVPR, 2024. 1 +[6] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 1, 7 +[7] 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. 1, 3 +[8] Andrew Brock. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018.3 +[9] 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. *NeurIPS*, 2020. 1, 2, 3 +[10] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1, 2, 3, 4, 5, 7 +[11] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In ICML, 2023. 3, 6 +[12] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 2 +[13] 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, 24(240):1-113, 2023. 2 +[14] 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. JMLR, 25(70):1-53, 2024. 1 +[15] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdul-mohsin, et al. Scaling vision transformers to 22 billion parameters. In ICML, pages 7480-7512. PMLR, 2023. 5 +[16] 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, 5, 7 +[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2018. 3 +[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 1, 3, 6, 7 +[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. In ICLR, 2021. 4, 5 +[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 +[21] Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai, Miguel Angel Bautista, Alexander Toshev, Vaishaal Shankar, Joshua M Susskind, and Armand Joulin. Scalable pre-training of large autoregressive image models. ICML, 2024. 1, 3 +[22] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 1, 2, 3, 4, 5, 6, 7 +[23] 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 +[24] Lijie Fan, Tianhong Li, Siyang Qin, Yuanzhen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 3 +[25] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. In ICCV, 2023. 6, 7 +[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 3 +[27] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In ICML, 2014. 2 +[28] Jiatao Gu, Changhan Wang, and Junbo Zhao. Levenshtein transformer. NeurIPS, 32, 2019. 4 +[29] Ju He, Qihang Yu, Qihao Liu, and Liang-Chieh Chen. Flowtok: Flowing seamlessly across text and image tokens. In ICCV, 2025. 1 + +[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1 +[31] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6, 8 +[32] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3 +[33] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 3 +[34] Dongwon Kim, Ju He, Qihang Yu, Chenglin Yang, Xiaohui Shen, Suha Kwak, and Liang-Chieh Chen. Democratizing text-to-image masked generative models with compact text-aware one-dimensional tokens. In ICCV, 2025. 1 +[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6 +[36] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles, pages 611-626, 2023. 8 +[37] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 1 +[38] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. NeurIPS, 2024. 1, 2, 3, 6, 7, 8 +[39] Qihao Liu, Zhanpeng Zeng, Ju He, Qihang Yu, Xiaohui Shen, and Liang-Chieh Chen. Alleviating distortion in image generation via multi-resolution diffusion models. NeurIPS, 2024. 3, 7 +[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 6 +[41] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magtvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 1, 2, 7 +[42] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In ECCV, 2024. 7 +[43] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 3 +[44] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. NeurIPS, 36, 2023. 4 +[45] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.1, 2 +[46] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In ICML, 2018. 2 + +[47] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 1, 3, 5, 7, 8 +[48] Alec Radford. Improving language understanding by generative pre-training. OpenAI, 2018. 2 +[49] 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 ICML, 2021. 2 +[50] 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 +[51] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NeurIPS, 2019. 2 +[52] Sucheng Ren, Qihang Yu, Ju He, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Beyond next-token: Next-x prediction for autoregressive visual generation. In ICCV, 2025. 1 +[53] Sucheng Ren, Qihang Yu, Ju He, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Flowar: Scale-wise autoregressive image generation meets flow matching. In ICML, 2025. 1 +[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 7 +[55] Inkyu Shin, Chenglin Yang, and Liang-Chieh Chen. Deeply supervised flow-based generative models. In ICCV, 2025. 1 +[56] stabilityyai, 2023. 7 +[57] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 4, 5, 7, 8 +[58] 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, 2015. 5 +[59] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024.3 +[60] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2 +[61] Emu3 Team. Emu3: Next-token prediction is all you need. Tech Report, 2024. 1, 2 +[62] 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. 1, 2 +[63] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NeurIPS, 2024. 1, 2, 3, 7, 8 +[64] 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. 1, 2 +[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2 +[66] Benigno Uria, Marc-Alexandre Côté, Karol Gregor, Iain Murray, and Hugo Larochelle. Neural autoregressive distribution estimation. JMLR, 17(205):1-37, 2016. 3 +[67] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NeurIPS, 2016. 2 +[68] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. 2 +[69] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 2 +[70] Mark Weber, Lijun Yu, Qihang Yu, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. Maskbit: Embedding-free image generation via bit tokens. arXiv preprint arXiv:2409.16211, 2024. 1, 3, 7, 8 +[71] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. In ICLR, 2022. 1 +[72] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1, 2 +[73] Chenglin Yang, Celong Liu, Xueqing Deng, Dongwon Kim, Xing Mei, Xiaohui Shen, and Liang-Chieh Chen. 1.58-bit flux. arXiv preprint arXiv:2412.18653, 2024. 3 +[74] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. Xlnet: Generalized autoregressive pretraining for language understanding. NeurIPS, 2019. 3 +[75] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. In ICLR, 2022. 1, 2, 7 +[76] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. TMLR, 2022. 1 +[77] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In CVPR, 2023. 1, 3 +[78] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. In ICLR, 2024. 3, 7 +[79] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth + +32 tokens for reconstruction and generation. NeurIPS, 2024. 1, 3, 6, 7, 8 +[80] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, pages 12104-12113, 2022. 5 +[81] Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. Var-clip: Text-to-image generator with visual auto-regressive modeling. arXiv preprint arXiv:2408.01181, 2024.3 \ No newline at end of file diff --git a/ICCV/2025/Randomized Autoregressive Visual Generation/images.zip b/ICCV/2025/Randomized Autoregressive Visual Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0978ee5787b75c46b70e3ea39bdfbd0672c68e9b --- /dev/null +++ b/ICCV/2025/Randomized Autoregressive Visual Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e91bfbfb48176c94cace2c2b820fda05c026cd45901e37f11da71b9873f54fe7 +size 614436 diff --git a/ICCV/2025/Randomized Autoregressive Visual Generation/layout.json b/ICCV/2025/Randomized Autoregressive Visual Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bb7296a6c56adf36c1eb9fdd4af2c0d0e479be8e --- /dev/null +++ b/ICCV/2025/Randomized Autoregressive Visual Generation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90aa03677e2612ba5d1313ef25d97a862fb1a55de8223b99c6544ba2dba4b97e +size 424524 diff --git a/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/83029d55-f117-48b0-b2a1-d5c53fa1a6f9_content_list.json b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/83029d55-f117-48b0-b2a1-d5c53fa1a6f9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eed58bc8065fe0f78479bb40e71f1052c1e5020e --- /dev/null +++ b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/83029d55-f117-48b0-b2a1-d5c53fa1a6f9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b835cd91ecb57b4b45675529f66eebfd6e101ab52be8aa0e8d9cfee9feec5e01 +size 81087 diff --git a/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/83029d55-f117-48b0-b2a1-d5c53fa1a6f9_model.json b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/83029d55-f117-48b0-b2a1-d5c53fa1a6f9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..32dafb3218a0c86bc4a012271e89d651a370b747 --- /dev/null +++ b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/83029d55-f117-48b0-b2a1-d5c53fa1a6f9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12dba93b6fe206ef506270e4573c606e4faef73218e4d972d94ad64c1506acb5 +size 103517 diff --git a/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/83029d55-f117-48b0-b2a1-d5c53fa1a6f9_origin.pdf b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/83029d55-f117-48b0-b2a1-d5c53fa1a6f9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e2c7b74006e4f716086c9590fa376dff1342d7ef --- /dev/null +++ b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/83029d55-f117-48b0-b2a1-d5c53fa1a6f9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6155e4c8ff50bcd22b1c72549d67ad3c8913b15a412009c7c2f3a6a6dfdaae8f +size 1804969 diff --git a/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/full.md b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7a80ef5bd3e4415d07ecb769214a959d522abdd8 --- /dev/null +++ b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/full.md @@ -0,0 +1,310 @@ +# RapVerse: Coherent Vocals and Whole-Body Motion Generation from Text + +![](images/dabb9065a7b65c160dff63896cd4cc2177abf1778226736737b6cbce6d5f090b.jpg) + +![](images/2cb0504eaf083709254004cd2fffe1f7227387c19fc1e72a8942f2ad4abeb149.jpg) + +Generate a holistic body motion and the corresponding singing vocals from the following lyrics: The way I sing on these tracks now I been back now digging but broke for a minute I been up since the evening + +![](images/e450d4f965c4990e3f8cb6f26fb3461ff7ec5fbd7836b67861925ec92468bd66.jpg) +Figure 1. RapVerse. We present a unified text-conditioned multi-modality generation framework, for jointly generating holistic body motions and singing vocals from textual lyrics inputs only. Note that the corresponding video frames are just shown for reference. + +# Abstract + +In this work, we introduce a challenging task for simultaneously generating 3D holistic body motions and singing vocals directly from textual lyrics inputs, advancing beyond existing works that typically address these two modalities in isolation. To facilitate this, we first collect the RapVerse dataset, a large dataset containing synchronous rapping vocals, lyrics, and high-quality 3D holistic body meshes. With the RapVerse dataset, we investigate the extent to which scaling autoregressive multimodal transformers across language, audio, and motion can enhance the coherent and realistic generation of vocals and whole-body human motions. For modality unification, a vector-quantized variational autoencoder is employed to encode whole-body motion sequences into discrete motion tokens, while a vocal-to-unit model is leveraged to obtain quantized audio tokens preserving content, prosodic information and singer identity. By jointly performing transformer modeling on + +these three modalities in a unified way, our framework ensures a seamless and realistic blend of vocals and human motions. Extensive experiments demonstrate that our unified generation framework not only produces coherent and realistic singing vocals alongside human motions directly from textual inputs, but also rivals the performance of specialized single-modality generation systems, establishing new benchmarks for joint vocal-motion generation. Video demonstration and more information can be found on the project page1. + +# 1. Introduction + +In the evolving landscape of multi-modal content generation in terms of sound and motion, significant strides have been made in individual modalities, including text-to-music [1, 24, 51], text-to-vocal [34], text-to-motion [7, 16, 26, 35, + +[66], and audio-to-motion [18, 33, 65] generation. These developments have paved the way for creating more dynamic and interactive digital content. Despite these advancements, existing works predominantly operate in silos, addressing each modality in isolation. However, there's strong psychological evidence that for human beings, the generation of sound and motion are highly related and coupled [30]. A unified system for joint generation allows for a more expressive and nuanced communication of emotions, intentions, and context, where the generation of one modality could guide and assist the other in a coherent and efficient way. + +In this paper, we tackle a crucial problem: can a machine not only sing with emotional depth but also perform with human-like expressions and motions? We propose a novel task for generating coherent singing vocals and whole-body human motions (including body motions, hand gestures, and facial expressions) simultaneously, see Fig. 1. This endeavor holds practical significance in fostering more immersive and naturalistic digital interactions, thereby elevating virtual performances, interactive gaming, and the realism of virtual avatars. + +An important question naturally arises: what constitutes a good model for unified generation of sound and motion? Firstly, we consider textual lyrics as the proper form of inputs for the unified system, since text provides a highly expressive, interpretable flexible means of conveying information by human beings, and could serve as a bridge between various modalities. Previous efforts explore scores [34], action commands [7, 26, 66], or audio signals [65] as inputs, which are inferior to textual inputs in terms of semantic richness, expressiveness and flexible integration of different modalities. + +Secondly, we reckon that a joint generation system that could produce multi-modal outputs simultaneously is better than a cascaded system that executes the single-modal generation sequentially. A cascaded system, combining a text-to-vocal module with a vocal-to-motion module, risks accumulating errors across each stage of generation. For instance, a misinterpretation in the text-to-vocal phase can lead to inaccurate motion generation, thereby diluting the intended coherence of the output. Furthermore, cascaded architectures necessitate multiple training and inference phases across different models, substantially increasing computational demands. + +To build such a joint generation system, the primary challenges include: 1) the scarcity of datasets that provide lyrics, vocals, and 3D whole-body motion annotations simultaneously; and 2) the need for a unified architecture capable of coherently synthesizing vocals and motions from text. In response to these challenges, we have curated RapVerse, a large-scale dataset featuring a comprehensive collection of lyrics, singing vocals, and 3D whole-body motions. Despite the existence of datasets available for text-to + +vocal [11, 25, 34, 52], text-to-motion [16, 32, 37, 43], and audio-to-motion [5, 8, 12, 15, 18, 62], the landscape lacks a unified dataset that encapsulates singing vocals, whole-body motion, and lyrics simultaneously. Most notably, large text-to-vocal datasets [25, 67] are predominantly in Chinese, limiting their applicability for English language research and lacking any motion data. And text-to-motion datasets [16, 32, 43] typically focus on text descriptions of specific actions paired with corresponding motions without audio data, often not covering whole body movements. Moreover, audio-to-motion datasets [34, 36] focus primarily on speech rather than singing. A comparison of existing related datasets is demonstrated in Table. 1. The RapVerse dataset is divided into two distinctive parts to cater to a broad range of research needs: 1) a Rap-Vocal subset containing a large number of pairs of vocals and lyrics, and 2) a Rap-Motion subset encompassing vocals, lyrics, and human motions. The Rap-Vocal subset contains 108.44 hours of high-quality English singing voice in the rap genre without background music. Paired lyrics and vocals are crawled from the Internet from 32 singers, with careful cleaning and post-processing. On the other hand, the Rap-Motion subset contains 26.8 hours of rap performance videos with 3D holistic body mesh annotations in SMPL-X parameters [41] using the annotation pipeline of Motion-X [32], synchronous singing vocals and corresponding lyrics. + +With the RapVerse dataset, we explore how far we can push by simply scaling autoregressive multimodal transformers with language, audio, and motion for a coherent and realistic generation of vocals and whole-body human motions. To this end, we unify different modalities as token representations. Specifically, three VQVAEs [60] are utilized to compress whole-body motion sequences into three-level discrete tokens for head, body, and hand, respectively. For vocal generation, previous works [10, 34, 38, 38] share a common paradigm, producing mel-spectrograms of audio signals from input textual features and additional music score information, following with a vocoder [40, 59, 64] to reconstruct the phase. We draw inspiration from the speech resynthesis domain [44], and learn a self-supervised discrete representation to quantize raw audio signal into discrete tokens while preserving the vocal content and prosodic information. Then, with all the inputs in discrete representations, we leverage a transformer to predict the discrete codes of audio and motion in an autoregressive fashion. Extensive experiments demonstrate that this straightforward unified generation framework not only produces realistic singing vocals alongside human motions directly from textual inputs but also rivals the performance of specialized single-modality generation systems. + +To sum up, this paper makes the following contributions: + +- We collect RapVerse, a large dataset featuring synchronous singing vocals, lyrics, and high-quality 3D + +holistic SMPL-X parameters. + +- We design a simple but effective unified framework for the joint generation of singing vocals and human motions from text with a multi-modal transformer in an autoregressive fashion. +- To unify representations of different modalities, we employ a vocal-to-unit model to obtain quantized audio tokens and utilize compositional VQVAEs to get discrete motion tokens. +- Experimental results show that our framework rivals the performance of specialized single-modality generation systems, setting new benchmarks for joint generation of vocals and motion. + +# 2. Related Work + +# 2.1. Text to Vocal Generation + +Text-to-audio Dataset. There exists several singing vocal datasets, yet they each has constraints. For instance, PopCS [34] and OpenSinger [25] are limited to Chinese, while NUS-48E [11] and NHSS [52] have only a few hours of songs. Nevertheless, JVS-MuSiC [55] and NUS-48E [11] offer a few hours of songs from dozens of singers, whereas OpenSinger [25] provides a more extensive collection with tens of hours from a single singer. Notably, our dataset represents the first to be specifically curated for rap songs of multiple singers with 108 hours. + +Text-to-vocal Models. Recent advancements in text-to-speech (TTS) models, including WavNet [40], FastSpeech 1 and 2 [48, 49], and EATS [10], have significantly improved the quality of synthesized speech. However, singing voice synthesis (SVS) presents a greater challenge due to its reliance on additional musical scores and lyrics. Recent generation models [19, 19, 29, 34, 71] perform excellently in generating singing voices. However, their performance tends to deteriorate when encountering out-of-distribution data. To handle this, StyleSinger [70] introduces a Residual Style Adaptor and an Uncertainty Modeling Layer Normalization to handle this. + +# 2.2. Text to Motion Generation + +Text-to-motion Dataset. Current text-motion datasets, such as KIT [43], AMASS [37], and HumanML3D [16], are constrained by limited data coverage, typically spanning only tens of hours and lacking whole-body motion representation. To overcome these shortcomings, Motion-X [32] was developed, providing an extensive dataset that encompasses whole-body motions. + +Text-to-motion Models. Text-to-motion models [7, 16, 17, 35, 42, 68] have gained popularity for their user convenience. Recent advancements have focused on diffusion models [28, 58, 68], which, unlike deterministic generation models [2, 14], enable finer-grained and diverse output gen + +eration. Chen et al. [7] proposed a motion latent-based diffusion model to enhance generative quality and reduce computational costs. However, the misalignment between natural language and human motions presents challenges. Lu et al. [35] introduced the Text-aligned Whole-body Motion generation framework to address these issues by generating high-quality, diverse, and coherent facial expressions, hand gestures, and body motions simultaneously. + +# 2.3. Audio to Motion Generation + +Audio-to-motion Dataset. Co-Speech Datasets can be classified into 1: pseudo-labeled (PGT) and 2: motion-captured (mocap). PGT datasets [5, 15, 18, 65] utilize disconnected 2D or 3D key points to represent the body, allowing extracting at a lower cost but with limited accuracy. On the other hand, mocap datasets provide annotations for limited body parts, with some focusing solely on the head [8, 12, 62] or the body [13, 54]. Differing from them, BEATX [33] contains mesh data of both head and body. However, these datasets typically focus on speech to motion generation. Our RapVerse, in contrast, contains paired text-vocal-motion data, enabling simultaneous motion and vocal generation. + +Audio-to-motion Models. Audio-to-motion models [3, 4, 6, 33, 53, 56, 65, 69] aim to produce human motion from audio inputs. Recognizing the intricate relationship between audio and human face, TalkSHOW [65] separately generates face parts and other body parts. However, this approach exhibits limitations, such as the absence of generating lower body motion. Motioncraft [4] proposes a DiT structure for motion generation. EMAGE [33] utilizes masked gestures together with vocals for a generation. Our model, however, goes a step further by being the first to generate paired audio-motion data directly from text. + +# 3. RapVerse Dataset + +In this section, we introduce RapVerse, a large rap music motion dataset containing synchronized singing vocals, textual lyrics and whole-body human motions. A comparison of our dataset with existing datasets is shown at Table. 1. The RapVerse dataset is divided into two subsets to cater to a broad range of research needs: a Rap-Vocal subset and a Rap-Motion subset. The overall collection pipeline of Rap-Verse is shown at Fig. 2. + +# 3.1. Rap-Vocal Subset + +The Rap-Vocal subset contains 108.44 hours of high-quality English singing voice in the rap genre with paired lyrics. We will introduce each step in detail. + +Data Crawling. In a bid to obtain a large number of rap songs and corresponding lyrics from the Internet, we utilize Spotdl and Spotify to collect songs, lyrics, and metadata of different rap singers. To ensure the quality of the dataset, we + +
DatasetLanguage#SingersHours
NUS-48E [11]English121.91
NHSS [52]English107
JVS-MuSiC [55]Japanese1002.28
Tohoku Kiritan [39]Japanese11
PopCS [34]Chinese15.89
OpenSinger [25]Chinese6650
Opencpop [61]Chinese15.25
M4Singer [67]Chinese2029.77
OursEnglish32108.44
+ +(a) Audio Datasets + +
DatasetModalityWhole†Hours
KIT-ML [43]Text5511.2
AMASS [37]Text5540.0
HumanML3D [16]Text5528.6
Motion-X [32]Text51127.1
HA2G [34]Audio5533
Yoon et.al [36]Audio5530
Speech2Gesture [15]Audio55144
Talkshow [65]Audio5127.0
OursText-Audio5126.8
+ +(b) Motion Datasets + +Table 1. Comparison of Audio and Motion Datasets. (a) compares our Rap-Vocal Subset with existing singing vocal datasets. (b) compares our Rap-Motion Subset with existing motion datasets. †: “Whole” means whole-body motion including face, body, and hand. + +![](images/c125ff3c38b1f81a0e5d42178b84bb46195b22cf42e68a07a63f3db2573f3162.jpg) +Figure 2. RapVerse dataset collection pipeline. There are two pathways for rap-vocal and rap-motion subsets, respectively. + +perform cleaning on the crawled songs by removing songs with misaligned lyrics and filtering out songs that are too long or too short. + +Vocal and Background Music Separation. Since the crawled songs are mixed with rapping vocals and background music, and we aim to synthesize singing vocals from separated clean data, we utilize Spleeter [21], the state-of-the-art open-source vocal-music source separation tool to separate and extract rap vocal voices and accompanying background music from the collected songs. Following [50], we normalize the loudness of the vocal voices to a fixed loudness level. + +Vocal Data Processing. The raw crawled lyrics from the Internet are in inconsistent formats, we conduct data cleaning on the lyrics by removing meta information (singer, composer, song name, bridging words, and special symbols). To ensure that the lyrics are aligned with the singing vocals, we collect lyrics only with the correct timestamps of each sentence, and we separate each song into around 10-second to 20-second segments for model training. + +# 3.2. Rap-Motion Subset + +The Rap-Motion subset contains 26.8 hours of rap performance videos with 3D holistic body mesh annotations in SMPL-X parameters [41], synchronous singing vocals, and corresponding lyrics. We introduce the collection pipeline of this subset as follows. + +Data Crawling. We crawled over 1000 studio performance + +videos from YouTube under the Common Creative License. We filter out low-quality videos manually to ensure the videos meet the following criteria: stable camera work, performers centered in the frame, clear visibility of the performer's entire body to capture detailed motion data, and high-quality audio for accurate vocal analysis. + +Audio Data Processing. Similar to the Rap-Vocal subset, we leverage Spleeter [22] to isolate singing vocals from accompanying music. Given that YouTube videos typically lack paired lyrics, we employ an ASR model, Whisper [45], to accurately transcribe vocals into corresponding text. + +Video Data Processing. To ensure the collection of high-quality video clips for motion annotation, we implemented a semi-automatic process to filter out undesirable content, such as advertisements, transition frames, changes in shots, and flashing lights. Initially, we applied YOLO [47] for human detection to discard frames where no humans were detected. Subsequently, we utilized RAFT [57] to assess the motion magnitude, employing a threshold to eliminate frames affected by camera instability. We then perform meticulous manual curation on the extracted clips, retaining only those of the highest quality. Finally, we follow the pipeline of the optimized-based method Motion-X [32] to extract 3D whole-body meshes from monocular videos. Specifically, we adopt the SMPL-X [41] for motion representations, given a T-frame video clip, the corresponding pose states $\mathcal{M}$ are represented as: + +$$ +\mathcal {M} = \left\{\mathcal {M} _ {f}, \mathcal {M} _ {b}, \mathcal {M} _ {h}, \zeta , \epsilon \right\} \tag {1} +$$ + +where $\mathcal{M}_f\in \mathbb{R}^{T\times 3},\mathcal{M}_b\in \mathbb{R}^{T\times 63}$ and $\mathcal{M}_h\in \mathbb{R}^{T\times 90}$ denotes jaw poses, body poses and hand poses, respectively. $\zeta \in \mathbb{R}^{T\times 100}$ and $\epsilon \in \mathbb{R}^{T\times 3}$ are the facial expression and global translation. + +# 4. Method + +Given a piece of lyrics text, our goal is to generate rap-style vocals and whole-body motions, including body movements, hand gestures, and facial expressions that resonate with the lyrics. With the help of our RapVerse dataset, we propose a novel framework that not only represents texts, vocals, and motions as unified token forms but also integrates token modeling in a unified model. As illustrated in Fig. 3, our model consists of multiple tokenizers for motion (Sec. 4.2) and vocal (Sec. 4.3) token conversions, as well as a general Large Text-Motion-Audio Foundation Model (Sec. 4.4) that targets for audio token synthesize and motion token creation, based on rap lyrics. + +# 4.1. Problem Formulation + +Let $\mathcal{L}$ be the rap lyric, such as "I going to make it a flashback set a picture from way back.", our model will compose the text-related vocal $\mathcal{V} \in \mathbb{R}^{t \times d_v}$ and whole-body motion $\mathcal{M} \in \mathbb{R}^{t \times d_m}$ , where $t$ , $d_v$ , $d_m$ denote the temporal length and the feature dimensions for vocal and motion in every time unit, respectively. Modality tokenizer encoders $\phi^{\mathcal{X}}(\mathcal{T} \mid \mathcal{X})$ and decoders $\varphi^{\mathcal{X}}(\mathcal{X} \mid \mathcal{T})$ will bridge each modality with the token format $\mathcal{T}$ utilizing different tokenizers, where $\mathcal{X} \in \{\mathcal{L}, \mathcal{V}, \mathcal{M}\}$ . The large foundation model $\psi^{\mathcal{X}}$ will incorporate tokens in unified modeling. The whole problem of coherent vocal and model generation via lyric text can be formulated as follows: + +$$ +\Theta^ {*} = \underset {\Theta} {\arg \max } \mathbb {P} _ {\Theta} (\mathcal {V}, \mathcal {M} \mid \mathcal {L}), \tag {2} +$$ + +where $\Theta$ denotes the whole model parameters and $\mathbb{P}_{\Theta}(\cdot)$ denotes the model distributions. + +# 4.2. Motion VQ-VAE Tokenizer + +Similar to prior arts, a motion VQ-VAE is adopted to convert 3D human motions into discrete tokens, where the tokenizer encoder $\phi$ generates motion tokens enriched with high-frequency information, while the decoder $\varphi$ seamlessly reconstructs discrete tokens into continuous motion sequences. Specifically, the tokenizer will maintain a learnable codebook $\mathcal{C}$ , which firstly encodes the input motion features $\mathcal{M}$ to vectors $z^{\mathcal{M}}$ , and then the motion quantizer $\mathcal{Q}^{\mathcal{M}}$ will look up the nearest neighbor in codebook $\mathcal{C}^{\mathcal{M}} = \{c_k^{\mathcal{M}}\}_{k=1}^{d_c}$ for the motion representation. The quantized code can be calculated as + +$$ +Q \left(z ^ {\mathcal {M}}; \mathcal {C} ^ {\mathcal {M}}\right) = \underset {k} {\arg \min } \| z ^ {\mathcal {M}} - c _ {k} ^ {\mathcal {M}} \| _ {2}. \tag {3} +$$ + +The tokenizer decoder $\varphi$ will project the quantized code back into the human motion after quantization. The reconstruction objective in the tokenizer training stage can be represented as + +$$ +\mathbb {E} _ {\mathcal {T} \sim \phi_ {\mathcal {M}} (\mathcal {T} | \mathcal {M})} \left[ \log \varphi_ {\mathcal {M}} (\mathcal {M} \mid \mathcal {T}) \right]. \tag {4} +$$ + +As the whole-body motion $\mathcal{M}$ can be decomposed into face $\mathcal{M}_f$ , body $\mathcal{M}_b$ , and hand $\mathcal{M}_h$ , where $\mathcal{M} \in \mathbb{R}^{t \times d}$ , $d_f + d_b + d_h = d_m$ , we build three tokenizers to obtain tokens for these three parts. + +# 4.3. Vocal2unit Audio Tokenizer + +Overall, we leverage the self-supervised framework [44] in speech resynthesis domain to learn vocal representations from the audio sequences. Specifically, we train a Vocal2unit audio tokenizer to build a discrete tokenized representation for the human singing voice. The vocal tokenizer consists of three encoders and a vocoder. The encoders include three different parts: (1) the semantic encoder; (2) the F0 encoder; and (3) the singer encoder. We will introduce each component of the model separately. + +Semantic Encoder. In order to extract the semantic information of the audio into discrete tokens, we use a pretrained Hubert encoder [23] as the main component of our semantic encoder network $E^{h}$ . Given a raw audio sequence $v$ , the output of the Hubert encoder is a continuous representation. Then we apply a K-means algorithm on the representations to finally generate discrete units $z^{S} = \{z_{i}^{S}\}_{i=1}^{L_{s}}$ where $z_{i}^{S} \in \{1, 2, \dots, K\}$ . $L_{s}$ is the length of hubert token sequence, and $K$ represents the number of clusters in K-means algorithm. + +F0 Encoder. We use the YAAPT algorithm [27] to extract the F0 from the input audio signal $v$ , which is then fed into an encoder $E_{p}$ to generate low-frequency F0 pitch representation. The encoder $E^{p}$ is trained using a VQ-VAE framework with Exponential Moving Average updates similar to [9], maintaining a learnable codebook $\mathcal{C}^p = \{c_k^p\}_{k=1}^{d_c}$ , where $d_c$ is the number of codes in the codebook. The raw F0 sequence is first preprocessed according to the singer, converted into a sequence of latent vectors, then mapped to its nearest neighbors in the codebook $c^p$ , which could be represented using integer number $k \in \{1, 2, \dots, d_c\}$ as its index in the codebook. The index sequence will be further formatted as the token $j$ pitch $k_{c}$ to produce pitch token sequence $z^{\mathcal{P}} = \{z_i^{\mathcal{P}}\}_{i=1}^{L_p}$ , where $L_p$ is the length of the pitch token sequence. Finally, the decoder will reconstruct the F0 representation leveraging the pitch token sequence. + +Singer Encoder. As the last component of our encoders, the singer encoder $E_{si}$ from [20, 44] is used to extract the Mel-spectrogram from the raw audio sequence and to output a singer representation $z^{\mathcal{I}} \in \mathbb{R}^{256}$ . The singer embedding only depends on the singer and is trained globally. + +![](images/8b118e5d1e9e6aa7b00631a95fd09c529ccc56ad089623217f1b467f369014ec.jpg) +Figure 3. Pipeline overview. We first pre-train all tokenizers on vocal-only and motion-only data. After we have pretrained the modality tokenizers, we can unify text, vocal, and motion in the same token space. We adopt a mixing organizing algorithm for input tokens to align via the temporal axis. These mixed input tokens are fed into the large Text-Motion-Audio foundation model to train on token prediction tasks, guided by the encoded features from textual input. + +Vocoder. In order to decode the vocal signal from the discrete tokens, we adopt a modified version of the HiFiGAN neural vocoder similar to the one in [44], consisting of a generator and multiple discriminators. The generator takes the encoded discrete representation $z^S$ and $z^P$ , and the singer embedding vector $z^T$ as inputs. After a lookup table and a series of blocks composed of transposed convolution and a residual block with dilated layers, it finally outputs waveform audio. The discriminators include multiple sub-discriminators, analyzing on different scales and different periodic structures. + +# 4.4. General Auto-regressive Modeling + +Model Architectures. The whole model of our general modeling pipeline consists of several pre-trained vocal and motion tokenizers and a large Text-Motion-Audio foundation model. The vocal and motion tokenizers are based on VQ-VAE architectures, enabling us to represent these two modalities as discrete tokens. We also adopt a T5-Tokenizer [46] to convert lyric text into tokens, so that all three modalities $\mathcal{X} \in \{\mathcal{L}, \mathcal{V}, \mathcal{M}\}$ are united in the token space $\mathcal{T}^{\mathcal{X}} = \{t_k^\mathcal{X}\}_{k=1}^{\mathcal{N}^\mathcal{X}}$ , where $\mathcal{N}^{\mathcal{X}}$ is the token sequence length for each modality. The large foundation model, a temporal transformer based on the decoder-only architecture, performs the next-token-prediction task based on textual features encoded by the T5-Encoder. + +Specifically, each token is generated after the start token by the probability distribution from the foundation model + +$p_{\delta}(t^{\mathcal{X}};h^{\mathcal{L}}) = \prod_{i}p_{\delta}(t_{i}^{\mathcal{X}}\mid t_{< i}^{\mathcal{X}};h_{\mathcal{L}})$ . During the training period, we calculate the cross-entropy loss to measure the likelihood between ground truth tokens and predicted token probabilities as + +$$ +- \sum_ {i = 1} ^ {\mathcal {N} ^ {\mathcal {X}}} \log p _ {\delta} \left(t _ {i} ^ {\mathcal {X}}; h ^ {\mathcal {L}}\right). \tag {5} +$$ + +After optimizing via this training objective, our model learns to predict the next token, which can be decoded into different modality features. This process is similar to text word generation in language models, while the "word" in our method such as $\text{face\_0123}$ , does not have explicit semantic information, but can be decoded into continuous modality features. + +Multimodality Tokens Organization. Since we have designed a coherent vocal and motion generation pipeline with multiple token types in our method, including hubert and pitch tokens in rap vocals, and face, body, hand in whole-body motions, the organization method of all these tokens in one sentence is important for training. Similar to previous works [26, 35], we adopt an interleaved style within each modality. Specifically, with the input whole-body motion tokens $\mathcal{T}^{\mathcal{M}_s} = \{t_k^{\mathcal{M}_s}\}_{k=1}^{\mathcal{N}^{\mathcal{M}_s}}$ , where $s = f, b, h$ which denotes face, body and hand respectively, and $\mathcal{N}^{\mathcal{M}_s}$ denotes the token length for each part. Therefore the motion token sequence can be formulated as $\mathcal{T}^{\mathcal{M}} = \{t_1^{\mathcal{M}_f}, t_1^{\mathcal{M}_b}, t_1^{\mathcal{M}_h}, t_2^{\mathcal{M}_f}, \ldots\}$ . Similarly, the vocal to + +kens consist of the hubert tokens $t^{\nu_h}$ and pitch tokens $t^{\nu_p}$ , which can be formulated as $\mathcal{T}^{\nu} = \{t_1^{\nu_h}, t_1^{\nu_p}, t_2^{\nu_h}, t_2^{\nu_p}, \ldots\}$ . This interleaved way of organization allows us to align the timesteps of each part. When performing coherent vocal and motion generation, we mix these two modality sequences sequentially with modality-wise start token as $\mathcal{T} = \{t_{start}^{V}, \mathcal{T}^{V}, t_{start}^{M}, \mathcal{T}^{M}\}$ . Note that we put vocal sequences before motion sequences because 1) the vocals are more related to the lyrics, and 2) motion tokens have a relationship with previously generated vocals, e.g. face motions with lips movement are directly related to vocals. Joint training of mixed modalities allows the transformer to capture the temporal dependencies and dynamics inherent in vocal and motion data. + +Inference and Decoupling. In the inference stage, we use different start tokens to specify which modality to generate. The textual input is encoded as features to guide token inference. We adopt a top-k algorithm to control the diversity of the generated content by adjusting the temperature, as generating vocals and motions based on lyrics is a creation process with multiple possible answers. After token prediction, a decoupling algorithm is used to process output tokens to make sure tokens from different modalities are separated and temporally aligned. These discrete tokens will be further decoded into text-aligned vocals and motions. + +# 5. Experiments + +In this section, we evaluate our proposed model on our proposed benchmark designed for joint vocal and whole-body motion generation from textual inputs. + +# 5.1. Experimental Setup + +Metrics. To evaluate the generation quality of singing vocals, we utilize the Mean Opinion Score (MOS) to gauge the naturalness of the synthesized vocal. For motion synthesis, we evaluate the generation quality of the body hand gestures and the realism of the face, respectively. Specifically, for gesture generation, we use Frechet Inception Distance (FID) based on a feature extractor from [16] to evaluate the distance of feature distributions between the generated and real motions, and Diversity (DIV) metric to assess the motions diversity. For face generation, we compare the vertex MSE [63] and the vertex L1 difference LVD [65]. Finally, we adopt Beat Constancy (BC) [31] to measure the synchrony of generated motion and singing vocals. + +Baselines. We compare the vocal generation quality with the state-of-the-art vocal generation method DiffSinger [34]. And we also adapt the text-to-speech model FastSpeech2 [49] for vocal generation. For motion generation, we compare our method with both text-to-motion methods and audio-to-motion methods. For text-to-motion methods, since there is no existing open-sourced work for text to whole-body motion generation, we compare with + +transformer-based T2M-GPT [66] and MLD [7] for body generation. For the audio-to-motion generation, we compare with Habibie et al. [18], Talkshow [65], Motioncraft [4] and EMAGE [33]. We report all the results on RapVerse with an $85\% / 7.5\% / 7.5\%$ train/val/test split. + +# 5.2. Main Results Analysis + +Evaluations on joint vocal and whole-body motion generations. We compared both text-driven and audio-driven motion generation baselines in Table 2 (a). To be noted, our setting is different from all existing methods in the following ways. First, we use rap lyrics as our textual input instead of motion textual descriptions, which contain direct action prompt words, such as walk and jump; Second we use text to jointly generate both audio and motion, instead of using audio to generate motion as audio-driven methods did. As is demonstrated, our model rivals with both text-to-motion and audio-to-motion methods in terms of metrics measuring body motion quality and face motion accuracy. + +![](images/2842d21e2a1c58bc46490d416082d7e04a30130504e3e0946d1e2e390e84bfdc.jpg) +Figure 4. User study on generated motion. + +Furthermore, the cornerstone of our approach lies in the simultaneous generation of vocals and motion, aiming to achieve temporal alignment between them. This objective is substantiated by our competitive results on the BC metric, which assesses the synchrony between singing vocals and corresponding motions, underscoring our success in closely synchronizing the generation of these two modalities. For the cascaded system, we integrate the text-to-vocal model DiffSinger with the audio-to-motion model Talkshow. Compared with the cascaded system, our joint-generation pipeline demonstrates superior outcomes while also reducing computational demands during both training and inference phases. In the cascaded architectures, errors tend to accumulate through each stage. Specifically, if the text-to-vocal module produces unclear vocals, it subsequently hampers the audio-to-motion model's ability to generate accurate facial expressions that align with the vocal content. + +To better evaluate the generation quality, we conducted an additional user study (see Fig. 4) using ten 10-second-long videos from the RapVerse dataset, sampled from the test set. We recruited 20 participants to evaluate four aspects: the holistic realism of human motion, expression-vocal synchronism, gesture-vocal synchronism, and holistic diversity, using real data as a reference. Participants + +
MethodFID↓DIV↑BC↑MSE↓LVD↓
Text-to-Motion
T2M-GPT [66]23.4511.75---
MLD [7]26.3412.15---
Audio-to-Motion
Habibie et al. [18]32.1410.080.4762.139.54
Talkshow [65]18.2313.140.4822.059.20
EMAGE [33]21.1812.650.4881.968.45
MotionCraft [4]17.7513.760.4822.069.23
Cascaded Result23.4212.870.4792.099.38
Text-to-Audio+Motion
Ours17.5814.080.4852.037.23
+ +(a) Motion Generation + +
MethodMOS↑†
GT4.45 ± 0.06
Reconstruction4.02 ± 0.08
FastSpeech2 [49]3.41 ± 0.18
DiffSinger [34]3.72 ± 0.12
Ours3.64 ± 0.15
+ +(b) Vocal Generation +Table 2. Quantitative results of generated motion and vocal. (a) compares with different motion generation baselines, best result is shown in bold and the second-best result is shown in underline. (b) compares with text-to-vocal generation baselines. †: The MOS is calculated with 95% confidence intervals of song samples. + +
MethodFID↓DIV↑BC↑MSE↓LVD↓
GT014.530.49900
Pretrained LLM48.2512.650.4623.2812.15
Single motion token19.1512.750.4772.1810.12
Ours17.5814.080.4852.037.23
+ +Table 3. Ablation study. We compare with common designs in motion generation frameworks. + +were asked to sort shuffled videos generated by different methods based on these criteria. As shown in the results, our model achieved the highest scores in human motion realism, gesture-vocal synchronism, and holistic diversity when compared with T2MGPT (text-to-motion) and Talkshow (audio-to-motion). + +Evaluations on vocal generations. We compare our method against other state-of-the-art text-to-vocal generation baselines in Table 2 (b). While our unified model is trained to simultaneously generate vocals and motion, a task considerably more complex than generating vocals alone, its vocal generation component still manages to achieve results comparable to systems designed solely for vocal generations. + +# 5.3. Ablation Study + +We present the outcomes of our ablation study in Table. 3. Initially, we explored the integration of a pre-trained large language model [46] for multi-modality generation, akin to the approach in [26]. However, the efficacy of utilizing pre-trained language models significantly lags behind our tailored design, underscoring that pre-training primarily on linguistic tokens does not facilitate effective prediction across multiple modalities like vocal and motion. Additionally, we study the impact of our compositional VQ-VAEs + +on motion generation. In contrast, a baseline employing a single VQVAE for the joint quantization of facial, body, and hand movements was implemented. This approach led to a noticeable degradation in performance, particularly marked by a -2.89 decrease in LVD. This decline can be attributed to the preponderance of facial movements in a singer's performance. Using a single VQ-VAE model for full-body dynamics compromises the detailed representation of facial expressions, which are crucial for realistic and coherent motion synthesis. + +# 6. Conclusion + +In this work, we present a new framework for the simultaneous generation of 3D whole-body motions and singing vocals, directly from textual lyrics. To address this challenging task, we first collect RapVerse, a large dataset containing synchronous rap vocals, alongside lyrics and 3D whole-body motions. Utilizing RapVerse, we demonstrate that simply scaling autoregressive transformers across language, audio, and motion yields a coherent generation of singing vocals and 3D holistic human motions. We anticipate that this work will inspire novel avenues in the joint modeling of text, audio, and motion. + +# References + +[1] Andrea Agostinelli, Timo I Denk, Zalán Borsos, Jesse Engel, Mauro Verzetti, Antoine Caillon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, et al. Musicl: Generating music from text. arXiv preprint arXiv:2301.11325, 2023. 1 +[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 3 +[3] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 3 +[4] Yuxuan Bian, Ailing Zeng, Xuan Ju, Xian Liu, Zhaoyang Zhang, Wei Liu, and Qiang Xu. Adding multi-modal controls to whole-body human motion generation. arXiv preprint arXiv:2407.21136, 2024. 3, 7, 8 +[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 2, 3 +[6] Junming Chen, Yunfei Liu, Jianan Wang, Ailing Zeng, Yu Li, and Qifeng Chen. Diffsheg: A diffusion-based approach for real-time speech-driven holistic 3d expression and gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7352-7361, 2024. 3 +[7] 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 Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 1, 2, 3, 7, 8 +[8] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 2, 3 +[9] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020. 5 +[10] Jeff Donahue, Sander Dieleman, Mikołaj Binkowski, Erich Elsen, and Karen Simonyan. End-to-end adversarial text-to-speech. arXiv preprint arXiv:2006.03575, 2020. 2, 3 +[11] Zhiyan Duan, Haotian Fang, Bo Li, Khe Chai Sim, and Ye Wang. The nus sung and spoken lyrics corpus: A quantitative comparison of singing and speech. In 2013 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, pages 1-9. IEEE, 2013. 2, 3, 4 +[12] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia, 12(6):591-598, 2010. 2, 3 +[13] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In + +Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 93-98, 2018. 3 +[14] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1396-1406, 2021. 3 +[15] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 2, 3, 4 +[16] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 1, 2, 3, 4, 7 +[17] 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 European Conference on Computer Vision, pages 580-597. Springer, 2022. 3 +[18] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 2, 3, 7, 8 +[19] Jinzheng He, Jinglin Liu, Zhenhui Ye, Rongjie Huang, Chenye Cui, Huadai Liu, and Zhou Zhao. Rmssinger: Realistic-music-score based singing voice synthesis. arXiv preprint arXiv:2305.10686, 2023. 3 +[20] Georg Heigold, Ignacio Moreno, Samy Bengio, and Noam Shazeer. End-to-end text-dependent speaker verification. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5115-5119. IEEE, 2016. 5 +[21] Romain Hennequin, Anis Khlif, Felix Voituret, and Manuel Moussallam. Spleeter: a fast and efficient music source separation tool with pre-trained models. Journal of Open Source Software, 5(50):2154, 2020. 4 +[22] Romain Hennequin, Anis Khlif, Felix Voituret, and Manuel Moussallam. Spleeter: a fast and efficient music source separation tool with pre-trained models. Journal of Open Source Software, 5(50):2154, 2020. Deezer Research. 4 +[23] Wei-Ning Hsu, Yao-Hung Hubert Tsai, Benjamin Bolte, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: How much can a bad teacher benefit asr pre-training? In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6533-6537. IEEE, 2021. 5 +[24] Qingqing Huang, Daniel S Park, Tao Wang, Timo I Denk, Andy Ly, Nanxin Chen, Zhengdong Zhang, Zhishuai Zhang, Jiahui Yu, Christian Frank, et al. Noise2music: Text-conditioned music generation with diffusion models. arXiv preprint arXiv:2302.03917, 2023. 1 +[25] Rongjie Huang, Feiyang Chen, Yi Ren, Jinglin Liu, Chenye Cui, and Zhou Zhao. Multi-singer: Fast multi-singer singing + +voice vocoder with a large-scale corpus. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3945-3954, 2021. 2, 3, 4 +[26] 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, 2024. 1, 2, 6, 8 +[27] Kavita Kasi and Stephen A Zahorian. Yet another algorithm for pitch tracking. In 2002 IEEE international conference on acoustics, speech, and signal processing, pages I-361. IEEE, 2002. 5 +[28] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 3 +[29] Sungjae Kim, Yewon Kim, Jewoo Jun, and Injung Kim. Muse-svs: Multi-singer emotional singing voice synthesizer that controls emotional intensity. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023. 3 +[30] George Lakoff and Mark Johnson. Metaphors We Live By. University of Chicago Press, 2003 edition, 1980. First published January 1, 1980. 2 +[31] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 7 +[32] 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. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4 +[33] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via masked audio gesture modeling. arXiv preprint arXiv:2401.00374, 2023. 2, 3, 7, 8 +[34] Jinglin Liu, Chengxi Li, Yi Ren, Feiyang Chen, and Zhou Zhao. Diffsinger: Singing voice synthesis via shallow diffusion mechanism. In Proceedings of the AAAI conference on artificial intelligence, pages 11020-11028, 2022. 1, 2, 3, 4, 7, 8 +[35] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978, 2023. 1, 3, 6 +[36] Shuhong Lu, Youngwoo Yoon, and Andrew Feng. Co-speech gesture synthesis using discrete gesture token learning. arXiv preprint arXiv:2303.12822, 2023. 2, 4 +[37] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 2, 3, 4 +[38] Dongchan Min, Dong Bok Lee, Eunho Yang, and Sung Ju Hwang. Meta-stylespeech: Multi-speaker adaptive text-to-speech generation. In International Conference on Machine Learning, pages 7748–7759. PMLR, 2021. 2 + +[39] Itsuki Ogawa and Masanori Morise. Tohoku kritan singing database: A singing database for statistical parametric singing synthesis using japanese pop songs. Acoustical Science and Technology, 42(3):140-145, 2021. 4 +[40] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016. 2, 3 +[41] 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 the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2, 4 +[42] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 3 +[43] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236–252, 2016. 2, 3, 4 +[44] Adam Polyak, Yossi Adi, Jade Copet, Eugene Kharitonov, Kushal Lakhotia, Wei-Ning Hsu, Abdelrahman Mohamed, and Emmanuel Dupoux. Speech resynthesis from discrete disentangled self-supervised representations, 2021. 2, 5, 6 +[45] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, pages 28492-28518. PMLR, 2023. 4 +[46] 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, 8 +[47] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 4 +[48] Yi Ren, Yangjun Ruan, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech: Fast, robust and controllable text to speech. Advances in neural information processing systems, 32, 2019. 3 +[49] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech. arXiv preprint arXiv:2006.04558, 2020.3,7,8 +[50] Yi Ren, Xu Tan, Tao Qin, Jian Luan, Zhou Zhao, and Tie-Yan Liu. Deepsinger: Singing voice synthesis with data mined from the web. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1979-1989, 2020. 4 +[51] Flavio Schneider, Ojasv Kamal, Zhijing Jin, and Bernhard Scholkopf. Mo\^ usai: Text-to-music generation with longcontext latent diffusion. arXiv preprint arXiv:2301.11757, 2023. 1 +[52] Bidisha Sharma, Xiaoxue Gao, Karthika Vijayan, Xiaohai Tian, and Haizhou Li. Nhss: A speech and singing parallel database. Speech Communication, 133:9-22, 2021. 2, 3, 4 + +[53] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized talking head synthesis. arXiv preprint arXiv:2301.03786, 2023. 3 +[54] Kenta Takeuchi, Souichirou Kubota, Keisuke Suzuki, Dai Hasegawa, and Hiroshi Sakuta. Creating a gesture-speech dataset for speech-based automatic gesture generation. In HCI International 2017-Posters' Extended Abstracts: 19th International Conference, HCI International 2017, Vancouver, BC, Canada, July 9–14, 2017, Proceedings, Part I 19, pages 198–202. Springer, 2017. 3 +[55] Hiroki Tamaru, Shinnosuke Takamichi, Naoko Tanji, and Hiroshi Saruwatari. Jvs-music: Japanese multispeaker singing-voice corpus. arXiv preprint arXiv:2001.07044, 2020. 3, 4 +[56] Sarah Taylor, Taehwan Kim, Yisong Yue, Moshe Mahler, James Krahe, Anastasio Garcia Rodriguez, Jessica Hodgins, and Iain Matthews. A deep learning approach for generalized speech animation. ACM Transactions On Graphics (TOG), 36(4):1-11, 2017. 3 +[57] 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. 4 +[58] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3 +[59] Jean-Marc Valin and Jan Skoglund. Lpcnet: Improving neural speech synthesis through linear prediction. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5891-5895. IEEE, 2019. 2 +[60] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2 +[61] Yu Wang, Xinsheng Wang, Pengcheng Zhu, Jie Wu, Hanzhao Li, Heyang Xue, Yongmao Zhang, Lei Xie, and Mengxiao Bi. Opencpop: A high-quality open source chinese popular song corpus for singing voice synthesis. arXiv preprint arXiv:2201.07429, 2022. 4 +[62] Cheng-hsin Wuu, Ningyuan Zheng, Scott Ardisson, Rohan Bali, Danielle Belko, Eric Brockmeyer, Lucas Evans, Timothy Godisart, Hyowon Ha, Xuhua Huang, et al. Multiface: A dataset for neural face rendering. arXiv preprint arXiv:2207.11243, 2022. 2, 3 +[63] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 7 +[64] Geng Yang, Shan Yang, Kai Liu, Peng Fang, Wei Chen, and Lei Xie. Multi-band melgan: Faster waveform generation for high-quality text-to-speech. In 2021 IEEE Spoken Language Technology Workshop (SLT), pages 492-498. IEEE, 2021. 2 +[65] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. + +Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 2, 3, 4, 7, 8 +[66] 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. arXiv preprint arXiv:2301.06052, 2023. 2, 7, 8 +[67] Lichao Zhang, Ruiqi Li, Shoutong Wang, Liquin Deng, Jinglin Liu, Yi Ren, Jinzheng He, Rongjie Huang, Jieming Zhu, Xiao Chen, et al. M4singer: A multi-style, multi-singer and musical score provided mandarin singing corpus. Advances in Neural Information Processing Systems, 35:6914-6926, 2022. 2, 4 +[68] 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. 3 +[69] 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 audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 3 +[70] Yu Zhang, Rongjie Huang, Ruiqi Li, JinZheng He, Yan Xia, Feiyang Chen, Xinyu Duan, Baoxing Huai, and Zhou Zhao. Stylesinger: Style transfer for out-of-domain singing voice synthesis. arXiv preprint arXiv:2312.10741, 2023. 3 +[71] Zewang Zhang, Yibin Zheng, Xinhui Li, and Li Lu. Wesinger: Data-augmented singing voice synthesis with auxiliary losses. arXiv preprint arXiv:2203.10750, 2022. 3 \ No newline at end of file diff --git a/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/images.zip b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b5c9ccb8963313e086117345968c648c76fa0e88 --- /dev/null +++ b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba88a6bd78b7ff301f47cead5c54bd77ed9cbfc8c643d4c6e38c3015f6f74659 +size 349026 diff --git a/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/layout.json b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..76bc19a57960fddaf64316054d32c99b39772c02 --- /dev/null +++ b/ICCV/2025/RapVerse_ Coherent Vocals and Whole-Body Motion Generation from Text/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15d010858a5f3b2f7e08f61719d217e72e9e7c0c23033d244095c85df8a43206 +size 378709 diff --git a/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/c89271c8-200b-4b69-9134-2b2e40fd2420_content_list.json b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/c89271c8-200b-4b69-9134-2b2e40fd2420_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fc567c5c6f158705352b246c6e470118cfc40326 --- /dev/null +++ b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/c89271c8-200b-4b69-9134-2b2e40fd2420_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38ca8d2eb821eb6c181ff3a1135814d8f78ec0e5723f6195cb8b0baa3f1561d5 +size 88892 diff --git a/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/c89271c8-200b-4b69-9134-2b2e40fd2420_model.json b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/c89271c8-200b-4b69-9134-2b2e40fd2420_model.json new file mode 100644 index 0000000000000000000000000000000000000000..70bb2b84ca107fde688c5a9b3caa1e9ae323152f --- /dev/null +++ b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/c89271c8-200b-4b69-9134-2b2e40fd2420_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffba495b9230241f1ce43b44a8a3049c5681f9a809ebb2acc65d4d6112729d4a +size 106367 diff --git a/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/c89271c8-200b-4b69-9134-2b2e40fd2420_origin.pdf b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/c89271c8-200b-4b69-9134-2b2e40fd2420_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..612dbcb2d18a6425ec998899d004e4def7737ea7 --- /dev/null +++ b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/c89271c8-200b-4b69-9134-2b2e40fd2420_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:558f1d129c5808c12316d834a3faf01584b061ede3bf551fd8efb4b753c4969c +size 1143995 diff --git a/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/full.md b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cae13714d50f8dfab2e033fe744265d855e30e11 --- /dev/null +++ b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/full.md @@ -0,0 +1,396 @@ +# RareCLIP: Rarity-aware Online Zero-shot Industrial Anomaly Detection + +Jianfang He1,2 Min Cao3,* Silong Peng1 Qiong Xie1,* + +$^{1}$ Institute of Automation, Chinese Academy of Sciences, Beijing, China + $^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences + $^{3}$ Soochow University, Suzhou, China + +{hejianfang2022, silong.peng, qiong.xie}@ia.ac.cn, mcao@suda.edu.cn + +# Abstract + +Large vision-language models such as CLIP have made significant strides in zero-shot anomaly detection through prompt engineering. However, most existing methods typically process each test image individually, ignoring the practical rarity of abnormal patches in real-world scenarios. Although some batch-based approaches exploit the rarity by processing multiple samples concurrently, they generally introduce unacceptable latency for real-time applications. To mitigate these limitations, we propose RareCLIP, a novel online zero-shot anomaly detection framework that enables sequential image processing in real-time without requiring prior knowledge of the target domain. RareCLIP capitalizes on the zero-shot capabilities of CLIP and integrates a dynamic test-time rarity estimation mechanism. A key innovation of our framework is the introduction of a prototype patch feature memory bank, which aggregates representative features from historical observations and continuously updates their corresponding rarity measures. For each incoming image patch, RareCLIP computes a rarity score by aggregating the rarity measures of its nearest neighbors within the memory bank. Moreover, we introduce a prototype sampling strategy based on dissimilarity to enhance computational efficiency, as well as a similarity calibration strategy to enhance the robustness of rarity estimation. Extensive experiments demonstrate that RareCLIP attains state-of-the-art performance with $98.2\%$ image-level AUROC on MVTec AD and $94.4\%$ on VisA, while achieving a latency of 59.4 ms. Code is available at https://github.com/hjf02/RareCLIP. + +# 1. Introduction + +Industrial anomaly detection (AD) plays a critical role in intelligent manufacturing by enabling robust quality control and efficient production. Traditional unsupervised AD + +methods (also known as full-shot methods) [7, 27, 31, 35, 38, 40] have achieved strong performance by training on large datasets of normal images. However, recent research has shifted toward reducing data dependency. Few-shot AD approaches [13, 14, 23, 33, 37] tackle this challenge by requiring only a limited number of normal samples. Extending this trend, zero-shot AD methods [3, 6, 16, 20, 41] eliminate the need for any prior domain-specific training, enabling direct deployment on previously unseen products. + +Most zero-shot AD methods use large vision-language models [17, 26, 30] and compare test images against predefined or learned textual prompts. However, by processing each test image in isolation, they overlook inter-image relationships. We refer to these as vanilla or offline zero-shot AD (Figure. 1a). They suffer from two major limitations: (1) They do not exploit the inherent rarity of anomalies in industrial environments, where defects are sparse compared to normal patterns; (2) Their isolated processing neglects the temporal continuity inherent in real-world production lines, where images are captured sequentially. + +To overcome these limitations, recent work has explored batch zero-shot AD [18, 22], which models the distribution of an entire test dataset (Figure. 1b). However, these methods often rely on computationally expensive pairwise comparisons across all images to achieve high performance, resulting in prohibitive latency for time-sensitive applications. + +In this paper, we introduce a novel paradigm: online zero-shot AD, where test images are processed sequentially in real-time (Figure. 1c). This setting introduces three key challenges: (1) Online Modeling: Rapid adaptation to data streams without prior domain knowledge. (2) Single-Pass Efficiency: Maximizing information extraction from images that are observed only once. (3) Real-Time Performance: Achieving low-latency and memory-efficient inference without compromising accuracy. + +To address these challenges, we propose RareCLIP, the first online zero-shot AD framework that integrates the zero-shot capabilities of CLIP [30] with dynamic rarity modeling. Our key insight is that normal patches main + +![](images/0ed0c44b9e006a99ed8d1de13ba4977a67285aa59853ce91554e7489bc540654.jpg) +(a) Vanilla/offline zero-shot AD + +![](images/e9b7da224712ff19f7564280125491cf8917ff875c19bcb4837d20ab291e5e0f.jpg) +(b) Batch zero-shot AD +Figure 1. Illustration of zero-shot AD paradigms during testing. (a) Vanilla/Offline zero-shot AD processes each test image individually. (b) Batch zero-shot AD requires simultaneous input of test images. (c) Our proposed online zero-shot AD processes test images sequentially. + +![](images/8ec55725558d6713effbd3f62f759b5103f90e64a04a0b445ca64ec6dd536e0f.jpg) +(c) Online zero-shot AD (Ours) + +![](images/5953c3c3982d98a17f0952e4327ceb1a157e8837fb772f9c8bd2dd29e248bc73.jpg) + +tain consistent appearances over time, whereas abnormal patches are rare and variable. We formalize this observation through the notion of patch-image similarity—defined as the maximum similarity between a given patch and all patches within an image. A patch is deemed normal if it exhibits high similarity to a sufficient fraction of historical images; otherwise, it is flagged as abnormal. + +A direct implementation, RareCLIP-d, computes similarities against all historical images, which incurs substantial memory and computational overhead. To alleviate this, our proposed RareCLIP framework incorporates a prototype patch feature memory bank that continuously accumulates representative patch features while dynamically updating their rarity measures via a patch-image similarity memory bank. For each test patch, RareCLIP estimates its patch-image similarity by aggregating the patch-image similarities of its nearest neighbors within the patch feature memory bank. To efficiently identify representative patch features, we introduce a dissimilarity-based sampling strategy—Sequential Coreset Sampling—which effectively prunes redundant pairs from the memory bank. In addition, we propose a similarity calibration strategy, Loose Similarity, to mitigate estimation errors and enhance the robustness of rarity scoring. + +Our main contributions are summarized as follows: + +- We introduce online zero-shot AD, a novel and practical setting that addresses the limitations of offline and batch zero-shot paradigms in industrial anomaly detection. +- We propose RareCLIP, a pioneering framework that combines CLIP's zero-shot capability with dynamic rarity modeling. Key innovations include Sequential Coreset Sampling for efficient memory management and Loose Similarity for robust anomaly scoring. +- Extensive experiments on the MVTec AD and VisA datasets demonstrate that RareCLIP significantly outperforms existing methods. + +# 2. Related Work + +Industrial Anomaly Detection. Recent advances in anomaly detection have predominantly focused on unsu + +pervised methods, also known as full-shot methods [1, 5, 7, 8, 11, 32, 38, 39], which rely on large collections of normal images during training. For example, PatchCore [31] builds a highly representative memory bank of normal patch features using a greedy coreset sampling strategy to minimize redundancy, thus reducing both storage requirements and inference time. In contrast, few-shot AD methods [13, 14, 23, 33, 37] address the challenge of limited normal data. GraphCore [37], for example, leverages graph neural networks to capture isometric invariant features, enabling the construction of a compact feature memory bank that performs effectively with only a few normal images. + +Zero-shot Industrial Anomaly Detection. Leveraging the strong generalization capabilities of large pre-trained vision-language models [17, 26, 30], zero-shot AD methods have emerged that bypass the need for any normal reference images. Most zero-shot AD methods [4, 6, 12, 16, 41] use text prompts to compare against test images. WinCLIP [16] pioneered this direction by introducing a prompt ensemble strategy combined with multiple window-based forward passes across image patches. Subsequent works [4, 6, 12, 41] have primarily focused on improving the alignment between patch features and text prompts. We refer to these as vanilla or offline zero-shot AD methods since they typically process each test image individually. On the other hand, approaches such as ACR [18] and MuSc [22] explore batch zero-shot AD by jointly evaluating a collection of test images, although the high performance comes at the cost of increased latency that may hinder real-time application. In this paper, we propose OnlineAD, a framework for online zero-shot AD that leverages visual information from historical images in real-time. + +Online Industrial Anomaly Detection. Online AD has attracted attention due to its closer alignment with industrial realities. Unlike incremental AD methods [25, 34], which iteratively refine models across multiple categories to enhance multi-class adaptability, online AD focuses on dynamically updating intra-class representations as new samples are acquired. For example, LeMO [10] supports online training using normal samples, making it well-suited + +![](images/94e3b39600799a0e98771057178cb656dbab710d77b8afb4283af863ea19cc36.jpg) +Figure 2. Overview of the proposed RareCLIP framework. RareCLIP mainly comprises two branches: the Text Prompt Branch (TPB, Section 3.1) and the Patch-level Parity Branch. The latter includes the Direct Patch-level Parity Branch (DPRB, Section 3.2.1) and the Indirect Patch-level Parity Branch (IPRB, Section 3.2.2), where DPRB is an intermediate variant of IPRB. The right portion illustrates the IPRB update process, where $\mathbf{S}_{t2H}$ and $\mathbf{S}_{\mathcal{M}2t}$ denote the patch-image similarities between test patches and historical images, and between memorized patches and the test image, respectively. + +![](images/2b39bd728b9b71add13fed6e6705b01138ae7b782e9ccf29ae51440bd3c5b18f.jpg) + +for streaming industrial data. FOADS [36] advances online few-shot AD by constructing a normal feature bank from a limited number of normal samples and boost performance during online testing, while O-InReach [28] mitigates reliance on normal samples by delaying predictions until multiple samples have been observed. In contrast, our proposed online zero-shot AD framework can make predictions even in the absence of historical images. + +# 3. Method + +Problem Setting. Figure. 1(c) illustrates the formal setting for our proposed online zero-shot AD. Given an unlabeled stream of test images $D_{u} = \{I_{1}, I_{2}, \dots \}$ from a single category without any prior domain-specific knowledge, the model is required to output the detection result for the $t$ -th test image $I_{t}$ before the next image $I_{t+1}$ arrives. + +Overview. Figure. 2 provides an overview of the proposed RareCLIP framework, which consists of two main branches: the Text Prompt Branch (TPB) and the Patch-level Rarity Branch. The Patch-level Rarity Branch is further divided into Direct Patch-level Rarity Branch (DPRB, corresponding to RareCLIP-d) and Indirect Patch-level Rarity Branch (IPRB, corresponding to RareCLIP), where DPRB serves as an ablation to highlight IPRB's reduced computational overhead while maintaining comparable performance. Each input test image is first encoded to generate patch features $\mathbf{F}_{test} = \{f_{test}^{m} \mid m \in [1, M]\} \in \mathbb{R}^{M \times C}$ , where $M$ is the number of patches and $C$ is the feature dimension, and a global image feature $F_{test}^{image}$ . + +# 3.1. Text Prompt Branch + +The Text Prompt Branch (TPB) forms the foundational component of RareCLIP, endowing it with initial zero-shot anomaly detection capabilities. In TPB, lightweight learnable projection layers adapt local patch features, denoted as $\widetilde{\mathbf{F}}_{test}$ , to align effectively with text features $\mathbf{F}_{text}$ that encode normal and abnormal semantic information from fixed text prompt ensemble. The patch-level anomaly map is computed as: + +$$ +\mathcal {A} _ {t e x t} = \operatorname {s o f t m a x} (\langle \widetilde {\mathbf {F}} _ {t e s t}, \mathbf {F} _ {t e x t} \rangle), \tag {1} +$$ + +where $\langle \cdot, \cdot \rangle$ denotes cosine similarity. Similarly, the global image feature $F_{test}^{image}$ is compared with $\mathbf{F}_{text}$ to yield an image-level anomaly score: + +$$ +c _ {t e x t} = \operatorname {s o f t m a x} \left(\left\langle F _ {t e s t} ^ {i m a g e}, \mathbf {F} _ {t e x t} \right\rangle\right). \tag {2} +$$ + +We observe that TPB tends to assign low anomaly scores in regions with missing components since background areas are generally learned as normal. To mitigate potential false negatives in such cases, we compute the cumulative average of $\mathcal{A}_{text}$ in the online setting, denoted as $\bar{\mathcal{A}}_{text}$ , and then refine the result by taking the element-wise maximum: + +$$ +\hat {\mathcal {A}} _ {t e x t} = \max \left(\mathcal {A} _ {t e x t}, \bar {\mathcal {A}} _ {t e x t}\right). \tag {3} +$$ + +Training Loss of TPB. To align local patch features with the semantics of normal and abnormal text prompts, we employ lightweight learnable projection layers. During training, the adapted patch features $\widetilde{\mathbf{F}}$ are optimized using a combination of focal loss [24] and Dice loss [21] against the + +corresponding text features. To further mitigate overfitting, we also enforce alignment between $\widetilde{\mathbf{F}}$ and the global image feature—which is already well-aligned with text. The final loss is formulated as: + +$$ +\begin{array}{l} \operatorname {L o s s} = \operatorname {F o c a l} \left(\mathcal {A} _ {\text {t e x t}}, \mathcal {A} _ {\text {t r u t h}}\right) + \operatorname {D i c e} \left(\mathcal {A} _ {\text {t e x t}}, \mathcal {A} _ {\text {t r u t h}}\right) \\ + \operatorname {M S E} (\langle \widetilde {\mathbf {F}}, F ^ {i m a g e} \rangle , \mathbf {1}), \\ \end{array} +$$ + +where $\mathcal{A}_{\text{truth}}$ denotes the ground-truth anomaly mask and MSE represents the Mean Squared Error. + +# 3.2. Patch-level Rarity Branch + +The TPB provides semantic-aware anomaly priors, which are further refined by the Patch-level Rarity Branch through cross-sample pattern analysis. + +# 3.2.1. Direct Patch-level Rarity Branch (RareCLIP-d) + +To capture the rarity of each test patch, we introduce the Direct Patch-level Rarity Branch (DPRB), the direct implementation of rarity mechanism. In DPRB, patch features from $N_{I}$ historical images are stored along with information about their source images to compute patch-image similarities. To manage memory usage, only patch features from the latest $N_{I,\max}$ historical images are retained. + +Patch-image Similarity Computation. For each test patch $f_{test}^{m}$ , we compute a patch-image similarity vector: + +$$ +\mathbf {s} _ {m} = \left\{\max _ {f _ {h i s} \in \mathbf {F} _ {h i s} ^ {i}} \langle f _ {t e s t} ^ {m}, f _ {h i s} \rangle \Big | i \in [ 1, N _ {I} ] \right\}, \tag {5} +$$ + +where $\mathbf{F}_{his}^{i}$ denotes the patch features from the $i$ -th historical image. + +Rarity Score Computation. A patch is considered non-rare if a sufficient proportion (exceeding rarity threshold $X\%$ ) of its patch-image similarities yield high value. Given the patch-image similarity vector $\mathbf{s} \in \mathbb{R}^{N_I}$ , we select the top $X\%$ of values, denoted as $\overline{\mathbf{s}}$ , and define the rarity (anomaly) score as: + +$$ +\operatorname {R a r i t y} (\mathbf {s}) = 1 - \operatorname {m e a n} (\overline {{\mathbf {s}}}). \tag {6} +$$ + +Accordingly, the patch-level anomaly map is given by: + +$$ +\mathcal {A} _ {\text {r a r e}} = \left\{\operatorname {R a r i t y} \left(\mathbf {s} _ {m}\right) \mid m \in [ 1, M ] \right\}. \tag {7} +$$ + +Sequential Coreset Sampling. To reduce both memory and computational costs, we propose Sequential Coreset Sampling (SCS), which eliminates redundant features in the memory bank. The key idea is that if two features are highly similar, the latter can be removed with minimal loss of representativeness. This approach ensures that the remaining features are dissimilar from each other and cover the feature space with minimal redundancy. + +SCS takes as input a feature set $\mathcal{F}$ and a target sample count $N_{\text{sample}}$ . For each feature $\mathcal{F}_i$ , we compute: + +$$ +s _ {i} = \left\{ \begin{array}{l l} 0, & i = 1, \\ \max _ {j < i} \langle \mathcal {F} _ {i}, \mathcal {F} _ {j} \rangle , & i > 1, \end{array} \right. \tag {8} +$$ + +![](images/cb0620aa96b792bbe4571baedd7cc1ade3cddb723c816a6bded89125a58a61b0.jpg) +Figure 3. A simple example of Sequential Coreset Sampling. Numbers represent the sequence of features in the memory. + +and then select the features with the smallest $s_i$ values: + +$$ +\mathcal {F} _ {\text {o u t p u t}} = \left\{\mathcal {F} _ {j} \mid s _ {j} \leq s _ {(N _ {\text {s a m p l e}})} \right\} \in \mathbb {R} ^ {N _ {\text {s a m p l e}} \times C}, \tag {9} +$$ + +where $s_{(N_{\text{sample}})}$ is the $N_{\text{sample}}$ -th smallest value. A simple example of SCS is shown in Figure. 3. + +Within DPRB, we apply SCS to the patch features of each historical image, updating Eq. 5 as: + +$$ +\mathbf {s} _ {m} = \left\{\max _ {f _ {h i s} \in \operatorname {S C S} \left(\mathbf {F} _ {h i s} ^ {i}, \alpha M\right)} \left\langle f _ {\text {t e s t}} ^ {m}, f _ {h i s} \right\rangle \mid i \in [ 1, N _ {I} ] \right\}, \tag {10} +$$ + +where $\alpha$ is the sampling ratio. + +# 3.2.2. Indirect Patch-level rarity Branch (RareCLIP) + +Although SCS reduces the number of features stored in DPRB, redundancy may still exist across different historical images. To address this, we propose the Indirect Patch-level Rarity Branch (IPRB) that employs dual memory banks: + +- Patch Feature Memory Bank $\mathcal{M}_{\mathcal{F}} \in \mathbb{R}^{N_F \times C}$ , which stores diversity-preserving patch features via SCS. +- Patch-image Similarity Memory Bank $\mathcal{M}_S \in \mathbb{R}^{N_F \times N_I}$ , which records the corresponding patch-image similarities of these memorized patch features. + +The patch-image similarity for a test patch is then indirectly estimated by aggregating the similarities of its nearest neighbors in $\mathcal{M}_{\mathcal{F}}$ with their corresponding values in $\mathcal{M}_{\mathcal{S}}$ . + +Update of the Patch Feature Memory Bank. After processing a test image, its patch features $\mathbf{F}_{test}$ are appended to $\mathcal{M}_{\mathcal{F}}$ . To maintain a fixed bank size, we apply SCS on the concatenated features: + +$$ +\hat {\mathcal {M}} _ {\mathcal {F}} = \operatorname {S C S} \left(\operatorname {C o n c a t} \left(\mathcal {M} _ {\mathcal {F}}, \mathbf {F} _ {\text {t e s t}}\right), N _ {F}\right), \tag {11} +$$ + +where $\hat{\mathcal{M}}_{\mathcal{F}}$ is the updated memory bank and $N_F$ controls its size. Notably, SCS tends to preserve earlier features, ensuring minimal changes in $\mathcal{M}_{\mathcal{F}}$ and, consequently, reducing the frequency of estimated updates needed for the associated similarity bank $\mathcal{M}_S$ . This stability enhances the robustness of the patch-image similarity estimation. + +Update of the Patch-image Similarity Memory Bank. The patch-image similarity memory bank $\mathcal{M}_S\in \mathbb{R}^{N_F\times N_I}$ stores the similarity between each feature in $\mathcal{M}_{\mathcal{F}}$ and the $N_{I}$ historical images. Its update can be divided into two components: + +- Memorized-to-Test Similarity $\mathbf{S}_{\mathcal{M}2t} \in \mathbb{R}^{N_F}$ : For each patch in $\mathcal{M}_{\mathcal{F}}$ , we compute its patch-image similarity with the test image: + +$$ +\mathbf {S} _ {\mathcal {M} 2 t} = \left\{\max _ {f _ {t e s t} \in \mathbf {F} _ {t e s t}} \left\langle f _ {\mathcal {M}} ^ {n}, f _ {t e s t} \right\rangle \mid f _ {\mathcal {M}} ^ {n} \in \mathcal {M} _ {\mathcal {F}} \right\}. \tag {12} +$$ + +- Test-to-Historical Similarity $\mathbf{S}_{t2H} \in \mathbb{R}^{M \times N_I}$ : Since full historical images are unavailable in IPRB, we estimate $\mathbf{S}_{t2H}$ indirectly. For each test patch $f_{test}^m$ , we first retrieve its $K$ nearest neighbors $\{\phi_k^m\}_{k=1}^K$ from $\mathcal{M}_{\mathcal{F}}$ and compute their similarities: + +$$ +\mathbf {S} _ {t 2 \mathcal {M}} = \left\{\left. \langle f _ {\text {t e s t}} ^ {m}, \phi_ {k} ^ {m} \rangle \right| m \in [ 1, M ], k \in [ 1, K ] \right\} \in \mathbb {R} ^ {M \times K}. \tag {13} +$$ + +Then, for each test patch, we estimate its patch-image similarities to historical images by aggregating: + +$$ +\mathbf {S} _ {t 2 H} = \left\{\frac {1}{K} \sum_ {k = 1} ^ {K} \mathbf {S} _ {t 2 \mathcal {M}} ^ {m, k} \cdot \mathcal {M} _ {\mathcal {S}} \left(\phi_ {k} ^ {m}\right) \mid m \in [ 1, M ] \right\}, \tag {14} +$$ + +where $\mathbf{S}_{t2\mathcal{M}}^{m,k}$ is the similarity between $f_{test}^{m}$ and its $k$ -th neighbor $\phi_k^m$ , and $\mathcal{M}_S(\phi_k^m) \in \mathbb{R}^{N_I}$ is the corresponding patch-image similarity vector of $\phi_k^m$ . + +Using $\mathbf{S}_{t2H}$ , the patch-level anomaly map can be obtained: + +$$ +\mathcal {A} _ {\text {r a r e}} = \left\{\operatorname {R a r i t y} \left(\mathbf {s} _ {m}\right) \mid \mathbf {s} _ {m} \in \mathbf {S} _ {t 2 H} \right\}. \tag {15} +$$ + +After computing $\mathbf{S}_{\mathcal{M}2t}$ and $\mathbf{S}_{t2H}$ , the similarity memory bank is preliminarily updated as: + +$$ +\mathcal {M} _ {\mathcal {S}} ^ {*} = \left[ \begin{array}{c c} \mathcal {M} _ {\mathcal {S}} & \mathbf {S} _ {\mathcal {M} 2 t} \\ \mathbf {S} _ {t 2 H} & \mathbf {1} \end{array} \right] \in \mathbb {R} ^ {\left(N _ {F} + M\right) \times \left(N _ {I} + 1\right)}, \tag {16} +$$ + +where $\mathbf{1}$ is a vector of ones. When SCS is applied to $\mathcal{M}_{\mathcal{F}}$ , the corresponding rows in $\mathcal{M}_{\mathcal{S}}^{*}$ associated with removed features are also deleted, resulting in an updated bank $\hat{\mathcal{M}}_{\mathcal{S}} \in \mathbb{R}^{N_F \times (N_I + 1)}$ . As before, we only retain patch-image similarities for the latest $N_{I,\max}$ historical images. + +**Loose Similarity.** Multiplying similarity measures to estimate patch-image similarity in Eq. 14 can introduce error. Assuming that the test patch and its neighbor are both similar to the same patch in a historical image, this can be simplified to estimating $\cos \theta_{AC}$ as $\cos \theta_{AB} \cdot \cos \theta_{BC}$ , where $\theta_{AC}$ is the angle between vectors $A$ and $C$ . The error $\cos \theta_{AC} - \cos \theta_{AB} \cdot \cos \theta_{BC}$ is bounded by + +$$ +\pm \sqrt {(1 - \cos^ {2} \theta_ {A B}) (1 - \cos^ {2} \theta_ {B C})}, +$$ + +which decreases as either $\cos \theta_{AB}$ or $\cos \theta_{BC}$ approaches 1. However, subtle differences among different images often prevent normal patches from achieving a similarity of 1, thus gradually reducing their estimated similarity after multiple multiplicative estimations. + +To alleviate this issue and enhance the robustness of rarity estimation, we propose a Loose Similarity (LS) strategy that loosens the requirement for similarities to reach 1. Given a set of similarities $\mathbf{s} = \{s_1,s_2,\dots ,s_n\}$ , we identify the $Y\%$ -th largest value $s_{[p_Y]}$ , where $p_{Y} = \lceil n\times Y\% \rceil$ and $Y$ controls the loose degree. We then divide all values by $s_{[p_Y]}$ and cap the results at 1: + +$$ +\operatorname {L S} (\mathbf {s}) = \min \left(\frac {\mathbf {s}}{s _ {[ p _ {Y} ]}}, 1\right). \tag {17} +$$ + +We apply LS to both $\mathbf{S}_{\mathcal{M}2t}$ (Eq. 12) and $\mathbf{S}_{t2\mathcal{M}}$ (Eq. 13), ensuring symmetric error compensation: + +$$ +\hat {\mathbf {S}} _ {\mathcal {M} 2 t} = \mathrm {L S} (\mathbf {S} _ {\mathcal {M} 2 t}), \quad \hat {\mathbf {S}} _ {t 2 \mathcal {M}} = \mathrm {L S} (\mathbf {S} _ {t 2 \mathcal {M}}). \tag {18} +$$ + +# 3.3. Anomaly Detection + +Image-level Rarity Branch. To better detect large abnormal regions, we introduce an Image-level Rarity Branch (IRB). In IRB, a local-aggregated image level feature $F^{laif}$ is calculated to assess image-level rarity. Since TPB tends to yield low anomaly scores in monotonous areas, we use $\bar{A}_{text}$ to roughly identify the most noticeable regions. Specifically, $F^{laif}$ is obtained by averaging the top half of the patch features corresponding to the highest values in $\bar{A}_{text}$ . The similarity between the test image and each historical image is then computed as: + +$$ +\mathbf {s} _ {t 2 H} = \left\{\left. \langle F _ {t e s t} ^ {l a i f}, F _ {i} ^ {l a i f} \rangle \right| i \in [ 1, N _ {I} ] \right\}, \tag {19} +$$ + +where $F_{test}^{laif}$ and $F_{i}^{laif}$ are the local-aggregated features for the test image and the $i$ -th historical image, respectively. The image-level rarity (anomaly) score is then computed as $c_{rare} = \mathrm{Rarity}(\mathbf{s}_{t2H})$ , and combined with the patch-level anomaly map: + +$$ +\hat {\mathcal {A}} _ {r a r e} = \mathcal {A} _ {r a r e} + c _ {r a r e} \cdot \left(1 + \text {n o r m a l i z e 0 1} (\bar {\mathcal {A}} _ {t e x t})\right). \tag {20} +$$ + +Pixel-level Anomaly Detection. We fuse the refined anomaly maps from TPB and the rarity branch to obtain the final anomaly map: + +$$ +\mathcal {A} _ {\text {t e s t}} = \lambda_ {\text {t e x t}} \hat {\mathcal {A}} _ {\text {t e x t}} + \lambda_ {\text {r a r e}} \hat {\mathcal {A}} _ {\text {r a r e}}, \tag {21} +$$ + +where $\lambda_{text}$ and $\lambda_{\mathrm{rare}}$ are weighting factors. The anomaly map $\mathcal{A}_{test} \in \mathbb{R}^{M}$ is reshaped to a $\sqrt{M} \times \sqrt{M}$ grid and upsampled to the original image resolution. A Gaussian filter is then applied to smooth the final pixel-level anomaly detection result. + +Image-level Anomaly Detection. For image-level detection, we first compute a preliminary anomaly score $c_{test}$ by taking the maximum value from the pixel-level anomaly + +map and adding $c_{text}$ . Inspired by MuSc [22], we further refine this score via an Image-level Re-scoring strategy. The underlying assumption is that visually similar images should exhibit consistent anomaly statuses. We identify the $B$ most similar historical images based on the average similarity of their image-level features (including both $F^{image}$ and $F^{laif}$ ). Let $s_b$ denotes the similarity between the test image and the $b$ -th similar historical image, and $c_b$ its preliminary anomaly score, for $b = 1, \ldots, B$ . The refined image-level anomaly score is computed as: + +$$ +\hat {c} _ {\text {t e s t}} = \frac {e ^ {\frac {1}{\tau}} \cdot c _ {\text {t e s t}} + \sum_ {b = 1} ^ {B} e ^ {\frac {s _ {b}}{\tau}} \cdot c _ {b}}{e ^ {\frac {1}{\tau}} + \sum_ {b = 1} ^ {B} e ^ {\frac {s _ {b}}{\tau}}}, \tag {22} +$$ + +where $\tau$ is a temperature hyper-parameter. + +# 4. Experiments + +Dataset. We conduct experiments on the widely-used industrial image datasets, MVTec AD [2] and VisA [42]. Both datasets contain multiple subsets, each with only one category per subset. MVTec AD includes 10 object categories and 5 texture categories with high-resolution images (from $700 \times 700$ to $1024 \times 1024$ ). VisA consists of 12 object categories with high-resolution images ( $1000 \times 1500$ ). We use the official test splits of both datasets, which include normal and abnormal images. + +Evaluation Metrics. In line with common practice, we report three metrics for image-level anomaly detection: image-level Area Under Receiver Operator Characteristic curve (I-AUC), image-level F1-score at the optimal threshold (I-F1-max), and image-level Average Precision (I-AP). Additionally, we report four metrics for pixel-level anomaly detection: pixel-level Area Under Receiver Operator Characteristic curve (P-AUC), pixel-level F1-score at the optimal threshold (P-F1-max), pixel-level Average Precision (P-AP), and Per-Region Overlap (PRO). + +Implementation Details. In accordance with prior studies [6, 41], we adopt CLIP [30] with the ViT-L-14-336 [9] backbone implemented via OpenCLIP [15] and resize all input images to $518 \times 518$ . For the text prompt branch, features are extracted from the 12th, 16th, 20th, and 24th layers of ViT which contain rich semantic information, and the associated linear layers are trained on one dataset when testing on another. In the rarity branch, features from the 6th, 12th, 18th, and 24th layers are used, with multi-scale patch features obtained via $1 \times 1$ and $3 \times 3$ neighborhood average pooling to enhance visual representation. The rarity threshold $X\%$ is set to $30\%$ , and $N_{I,\max}$ is fixed at 200 by default. In the Direct Patch-level rarity Branch (DPRB), the sampling ratio $\alpha$ is set to $\frac{1}{3}$ . In the Indirect Patch-level rarity Branch (IPRB), the patch feature memory bank size $N_F$ is set to 4107, $K$ is fixed at 3 for K-NN, and the loose degree $Y\%$ is set to $1\%$ . All experiments are performed on + +a single NVIDIA GeForce RTX 3090. Further implementation details are provided in the supplemental material. + +Offline and Online Modes. RareCLIP can operate in both offline and online modes. In offline mode, all memory banks are frozen and cumulative updates are halted, ensuring that detection results are independent of the test sequence. In online mode, the model dynamically updates its memory banks, and performance may be influenced by the sequence of test images. To ensure reproducibility of online mode, we use the official test split as fixed initial sequence, then report the average performance over five different random shuffles (using seeds 0, 1, 2, 3, and 4). + +Baselines. We compare the proposed RareCLIP with other state-of-the-art zero-shot AD methods, including: (1) offline zero-shot AD methods WinCLIP [16], April-GAN [6], and AnomalyCLIP [41], FiLo [12], AdaCLIP [4] and VCP-CLIP [29]; (2) batch zero-shot AD methods ACR [18] and MuSc [22]; (3) pseudo online (online-) zero-shot AD methods O-InReach [28] and MuSc*, which only make prediction starting from the second image, and MuSc* is a reproduced online version of MuSc [22] that only utilizes historical information from previous test images. + +# 4.1. Comparison with Zero-shot Methods + +Table. 1 presents the performance comparison between RareCLIP and other zero-shot AD methods. Our key observations are: (1) In offline mode, RareCLIP achieves competitive results compared to existing offline zero-shot methods; (2) In online mode, RareCLIP exhibits significant improvements across all metrics, even outperforming batch zero-shot AD methods; (3) RareCLIP attains higher image-level performance than RareCLIP-d (which incurs higher memory and computation costs) while only incurring a slight decrease in pixel-level performance. These results underscore the effectiveness and practicality of the proposed RareCLIP. + +# 4.2. Comparison with Few-/Full-shot Methods + +Table. 2 compares RareCLIP against state-of-the-art few-shot and full-shot AD methods [6, 16, 19, 23, 31, 35, 36, 40]. We also extend RareCLIP to a few-shot AD setting (details are provided in the supplemental material). Compared to state-of-the-art few-shot methods, RareCLIP in offline mode achieves more than a $1\%$ improvement in both I-AUC and P-AUC on both datasets. Moreover, RareCLIP in online mode attains competitive performance with full-shot AD methods. + +# 4.3. Offline Performance vs. Online Adaptation + +To analyze the evolution of the model's performance as it processes more images, we update the model using the first $N$ images of the test set and then evaluate its performance on the entire test set in offline mode. This sim + +
DatasetMethodPublicModeImage-levelPixel-level
I-AUCI-F1-maxI-APP-AUCP-F1-maxP-APPRO
MVTec ADWinCLIP [16]CVPR23offline91.892.996.585.131.7-64.6
April-GAN [6]CVPR23 WSoffline86.190.593.587.643.340.844.0
AnomalyCLIP [41]ICLR24offline91.692.796.491.139.134.581.4
FiLo [12]ACM MM24offline91.2--92.3---
AdaCLIP [4]ECCV24offline90.092.395.789.943.941.644.1
VCP-CLIP [29]ECCV24offline92.191.896.992.049.349.487.3
RareCLIP-offline91.592.996.691.547.546.186.2
O-InReach [28]ECCV24online-87.1±0.991.1±0.494.0±0.593.5±0.143.9±0.437.7±0.383.0±0.4
MuSc* [22]ICLR24online-96.0±0.496.2±0.398.0±0.397.0±0.257.8±0.555.9±0.993.3±0.1
RareCLIP-d-online97.9±0.297.4±0.299.1±0.297.8±0.162.9±0.564.2±1.193.4±0.1
RareCLIP-online98.2±0.297.6±0.199.3±0.197.7±0.264.1±0.766.1±1.093.5±0.1
ACR [18]NIPS23batch85.891.392.992.544.238.972.7
MuSc [22]ICLR24batch97.797.599.197.162.262.393.4
VisAWinCLIP [16]CVPR23offline78.179.081.279.614.8-56.8
April-GAN [6]CVPR23 WSoffline78.078.781.494.232.325.786.8
AnomalyCLIP [41]ICLR24offline82.080.485.395.528.321.386.7
FiLo [12]ACM MM24offline83.9--95.9---
AdaCLIP [4]ECCV24offline85.883.188.595.537.731.551.3
VCP-CLIP [29]ECCV24offline83.881.487.695.729.830.190.7
RareCLIP-offline86.183.189.095.733.527.090.2
O-InReach [28]ECCV24online-78.0±0.279.4±0.382.3±0.395.7±0.131.4±0.925.5±1.175.7±0.6
MuSc* [22]ICLR24online-90.0±0.587.1±0.790.5±0.398.6±0.048.5±0.244.9±0.392.4±0.1
RareCLIP-d-online93.5±0.290.2±0.394.1±0.598.9±0.048.1±0.444.2±0.593.2±0.1
RareCLIP-online94.4±0.390.8±0.495.3±0.298.8±0.050.9±0.347.5±0.393.5±0.1
MuSc [22]ICLR24batch92.689.193.398.748.945.492.4
+ +Table 1. Comparison of image-level and pixel-level zero-shot anomaly detection on the MVTec AD and VisA datasets. We compare the proposed RareCLIP with other state-of-the-art zero-shot methods. MuSc* denotes a reproduced online version of MuSc [22], and the evaluations of methods in "online-" mode exclude the first image since they predict result starting from the second image. Bold indicates the best performance, while underline denotes the second-best result. Methods under the batch setting are highlighted in gray as they concurrently utilize all test images for anomaly detection. All metrics are in %. + +
MethodSettingModeMVTec ADVisA
I-AUCP-AUCI-AUCP-AUC
WinCLIP+ [16]4-shotoffline95.296.287.397.2
April-GAN [6]4-shotoffline92.895.992.696.2
PromptAD [23]4-shotoffline96.696.589.197.4
FOADS [36]10-shotonline87.395.1--
RareCLIP0-shotonline98.297.794.498.8
RareCLIP4-shotoffline97.798.194.698.8
RareCLIP4-shotonline98.798.295.598.8
CutPaste [19]full-shotoffline96.196.0--
PatchCore [31]full-shotoffline99.198.194.898.5
RD++ [35]full-shotoffline99.498.395.998.7
RealNet [40]full-shotoffline99.699.097.898.8
+ +Table 2. Comparison with state-of-the-art few-shot and full-shot methods in image-level and pixel-level AUC on the MVtec AD and VisA datasets. + +ulates placing each test image at time step $N + 1$ and allows us to observe how performance evolves with the amount of historical data. As shown in Figure. 3, we observe that: (1) Performance steadily improves as $N$ increases, stabilizing after approximately $N = 24$ images; (2) RareCLIP achieves slightly higher image-level performance + +but marginally lower pixel-level performance compared to RareCLIP-d; (3) RareCLIP outperforms MuSc* especially when $N$ is small, demonstrating its stronger adaptability in online setting. + +# 4.4. Ablation Study + +Sampling Strategy. We compared SCS with other sampling strategies, including Random Sampling (RS), K-means Clustering Sampling (KCS), and Greedy Coreset Sampling (GCS). We use the implementation of GCS from Patchcore [31]. The results in Table. 3 indicate that: (1) KCS yields performance similar to RS but requires significantly more memory and time; (2) GCS achieves higher image-level performance than RS, though at the cost of increased computational time; (3) SCS offers optimal performance with minimal additional memory and time overhead by maintaining minimal changes in memory banks. + +Impact of $N_{I,\max}$ on Memory and Time Cost. Table. 4 illustrates the effect of varying $N_{I,\max}$ on GPU memory and computation time. We observe that: (1) Both MuSc* and RareCLIP-d exhibit a significant increase in time and GPU memory usage as $N_{I,\max}$ increases due to their reliance on + +![](images/c711492e913a61cba633694c1b4292c56b8901fb984a9532e02eb1d4b1282f1c.jpg) + +![](images/85361c175ddfcdbb1ecc69d0a1783d83ebdfbb9829b43acbbe3d5468d3eeb9a0.jpg) + +![](images/767470b55b7e064587188217d0c175518cb9eec098aa8167f57f393f339f8c20.jpg) +Figure 4. Offline performance as a function of the number of processed test images $(N)$ on the MVtec AD and VisA datasets. + +![](images/8313dbc53788bd1b078f011d5435a3741d2e3528595d0229ba7c0208293692ba.jpg) + +
SamplingMVTec ADVisAGPU (MB)Time (ms)
I-AUCP-AUCI-AUCP-AUC
RS88.1393.2175.3893.47429657.0
KCS88.0593.1075.3893.3548704280.8
GCS97.1893.1787.8088.4644682197.6
SCS98.1997.7094.4098.80435259.4
+ +Table 3. Ablation study of different sampling strategies on the MVTec AD and VisA datasets. RS, KCS, GCS and SCS respectively denote Random Sampling, K-means Clustering Sampling, Greedy Coreset Sampling and Sequential Coreset Sampling. + +
NI,maxMuSc* [22]RareCLIP-dRareCLIP
GPU (MB)Time (ms)GPU (MB)Time (ms)GPU (MB)Time (ms)
504697455.0469671.6429859.0
2007293983.76018106.2435259.4
1000198574054.719586310.8452061.2
+ +storing extensive patch features from historical images; (2) RareCLIP maintains low time and memory requirements across varying values of $N_{I,\max}$ , highlighting its computational efficiency. + +Impact of Loose Degree $Y$ . Table. 5 presents an ablation study on the hyper-parameter $Y$ , which controls the loose degree in LS. Our observations include: (1) Without LS or with $Y = 0$ (i.e., using the maximum similarity), per + +Table 4. Comparison of time and GPU memory consumption for $\mathbf{MuSc^{*}}$ [22], and RareCLIP(-d) under different values of $N_{I,\max}$ . + +
Y(%)MVTec ADVisA
I-AUCP-AUCI-AUCP-AUC
w/o97.0096.7593.2498.32
097.4097.0293.7398.49
0.598.1497.6394.4598.78
198.1997.7094.4098.80
298.1597.7594.2598.79
398.1297.7894.1298.78
598.0097.7993.9298.76
+ +Table 5. Ablation study of loose degree $Y$ on the MVTec AD and VisA datasets. "w/o" indicates that LS is not applied. + +
X(%)MVTec ADVisA
I-AUCP-AUCI-AUCP-AUC
597.2496.6893.5798.82
1097.6096.8894.0898.83
2098.1597.3194.3698.81
3098.1997.7094.4098.80
4097.9897.8094.1298.76
5097.6397.7893.9098.74
6097.2697.7693.6298.71
+ +Table 6. Ablation study of rarity threshold $X$ on the MVTec AD and VisA datasets. + +formance is lower compared to using other $Y$ values; (2) There is little difference in performance between $Y = 0.5$ and $Y = 2$ ; (3) As $Y$ increases further, I-AUC performance tends to decrease due to more weak abnormal regions being misclassified as normal. + +Impact of Rarity Threshold $X$ . Table. 6 shows how varying the rarity threshold $X$ affects performance. The results indicate that both excessively high and low values of $X$ lead to degraded performance, while a moderate value yields the best results. + +# 5. Conclusion + +To tackle the novel online zero-shot AD task, we proposed RareCLIP, a rarity-aware method that leverages the zero-shot capabilities of CLIP and integrates a dynamic test-time rarity estimation mechanism. RareCLIP contains a prototype patch feature memory bank, which aggregates representative features from historical observations and continuously updates their rarity measures. For each test patch, RareCLIP computes a rarity score by aggregating the rarity measures of its nearest neighbors within the memory bank. Furthermore, we introduced a dissimilarity-based prototype sampling strategy—Sequential Coreset Sampling—to improve computational efficiency, along with a similarity calibration mechanism, Loose Similarity, to enhance the robustness of rarity estimation. Experimental results on benchmark datasets demonstrate that RareCLIP achieves state-of-the-art performance with low latency. + +# Acknowledgments + +This work is supported in part by the National Natural Science Foundation of China under Grant No 62476188; in part by the Natural Science Foundation of the JiangsuHigher Education Institutions of China. + +# References + +[1] Jaehyeok Bae, Jae-Han Lee, and Seyun Kim. Pni: Industrial anomaly detection using position and neighborhood information. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6373-6383, 2023. 2 +[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9592-9600, 2019. 6 +[3] Yunkang Cao, Xiaohao Xu, Chen Sun, Yuqi Cheng, Zongwei Du, Liang Gao, and Weiming Shen. Segment any anomaly without training via hybrid prompt regularization. arXiv preprint arXiv:2305.10724, 2023. 1 +[4] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, pages 55-72. Springer, 2024. 2, 6, 7 +[5] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. arXiv preprint arXiv:2407.09359, 2024. 2 +[6] Xuhai Chen, Yue Han, and Jiangning Zhang. A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad. arXiv preprint arXiv:2305.17382, 2023. 1, 2, 6, 7 +[7] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 1, 2 +[8] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 2 +[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 +[10] Han Gao, Huiyuan Luo, Fei Shen, and Zhengtao Zhang. Towards total online unsupervised anomaly detection and localization in industrial vision. arXiv preprint arXiv:2305.15652, 2023. 2 +[11] Zhihao Gu, Liang Liu, Xu Chen, Ran Yi, Jiangning Zhang, Yabiao Wang, Chengjie Wang, Annan Shu, Guannan Jiang, and Lizhuang Ma. Remembering normality: Memory-guided knowledge distillation for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16401-16409, 2023. 2 + +[12] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high-quality localization. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2041–2049, 2024. 2, 6, 7 +[13] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 2024. 1, 2 +[14] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 2 +[15] 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. 6 +[16] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 2, 6, 7 +[17] 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, 2 +[18] Aodong Li, Chen Qiu, Marius Kloft, Padhraic Smyth, Maja Rudolph, and Stephan Mandt. Zero-shot anomaly detection via batch normalization. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 6, 7 +[19] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9664-9674, 2021. 6, 7 +[20] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 1 +[21] Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855, 2019. 3 +[22] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. Musc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. In International Conference on Learning Representations, 2024. 1, 2, 6, 7, 8 +[23] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompt: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838-16848, 2024. 1, 2, 6, 7 + +[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 3 +[25] Jiaqi Liu, Kai Wu, Qiang Nie, Ying Chen, Bin-Bin Gao, Yong Liu, Jinbao Wang, Chengjie Wang, and Feng Zheng. Unsupervised continual anomaly detection with contrastively-learned prompt. In Proceedings of the AAAI conference on artificial intelligence, pages 3639-3647, 2024. 2 +[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1, 2 +[27] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 1 +[28] Declan McIntosh and Alexandra Branzan Albu. Unsupervised, online and on-the-fly anomaly detection for nonstationary image distributions. In European Conference on Computer Vision, pages 428-445. Springer, 2024. 3, 6, 7 +[29] Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. Vcp-clip: A visual context prompting model for zero-shot anomaly segmentation. arXiv preprint arXiv:2407.12276, 2024. 6, 7 +[30] 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, 6 +[31] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 1, 2, 6, 7 +[32] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14902-14912, 2021. 2 +[33] Shelly Sheynin, Sagie Benaim, and Lior Wolf. A hierarchical transformation-discriminating generative model for few shot anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8495-8504, 2021. 1, 2 +[34] Jiaqi Tang, Hao Lu, Xiaogang Xu, Ruizheng Wu, Sixing Hu, Tong Zhang, Tsz Wa Cheng, Ming Ge, Ying-Cong Chen, and Fugee Tsung. An incremental unified framework for small defect inspection. In European conference on computer vision, pages 307-324. Springer, 2024. 2 +[35] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and + +Steven QH Truong. Revisiting reverse distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24511-24520, 2023. 1, 6, 7 +[36] Shenxing Wei, Xing Wei, Zhiheng Ma, Songlin Dong, Shaochen Zhang, and Yihong Gong. Few-shot online anomaly detection and segmentation. Knowledge-Based Systems, 300:112168, 2024. 3, 6, 7 +[37] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Feng Zheng, and Yaochu Jin. Pushing the limits of fewshot anomaly detection in industry vision: Graphcore. In International Conference on Learning Representations, 2023. 1, 2 +[38] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8330-8339, 2021. 1, 2 +[39] Xinyi Zhang, Naiqi Li, Jiawei Li, Tao Dai, Yong Jiang, and Shu-Tao Xia. Unsupervised surface anomaly detection with diffusion probabilistic model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6782-6791, 2023. 2 +[40] Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16699-16708, 2024. 1, 6, 7 +[41] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. In International Conference on Learning Representations, 2023. 1, 2, 6, 7 +[42] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 6 \ No newline at end of file diff --git a/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/images.zip b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..187da5c1e57dc5a1d2e3e00869d09d11430d866f --- /dev/null +++ b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b08670bacc848d3cbd431b6539b3a1f2ef6b730c0c9debb9d31d50cf1e9ab0da +size 657941 diff --git a/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/layout.json b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a0fe46472609d14ab0befb84b911709a078f9767 --- /dev/null +++ b/ICCV/2025/RareCLIP_ Rarity-aware Online Zero-shot Industrial Anomaly Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4318842309e89b5525cc5c766cc292665cb0eb9b7b348b93ec570ceba9bd564 +size 466248 diff --git a/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/f1d600ef-cd99-42cb-9996-11692aa89243_content_list.json b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/f1d600ef-cd99-42cb-9996-11692aa89243_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7480ea5003f5fe9e0f624caf7a6956886a2ff5da --- /dev/null +++ b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/f1d600ef-cd99-42cb-9996-11692aa89243_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8967e897dc6ad530d41f11ee19154a0a6ca12a22f8ebd0e047a8229e8628a88 +size 78614 diff --git a/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/f1d600ef-cd99-42cb-9996-11692aa89243_model.json b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/f1d600ef-cd99-42cb-9996-11692aa89243_model.json new file mode 100644 index 0000000000000000000000000000000000000000..65684674b6c2881a8a5c406f607a6d0e5da5981a --- /dev/null +++ b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/f1d600ef-cd99-42cb-9996-11692aa89243_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e00cb80f2b5480dd79c66b412432a4310c0e48fd92afaacaf2d15f33bfb852e0 +size 96080 diff --git a/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/f1d600ef-cd99-42cb-9996-11692aa89243_origin.pdf b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/f1d600ef-cd99-42cb-9996-11692aa89243_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7a06d00eadab5c42a9b9a24dd98649ed776c6548 --- /dev/null +++ b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/f1d600ef-cd99-42cb-9996-11692aa89243_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:063140d9e5a8017982a83211ad5063b7751d7c312893c3fbace73e9a6deb5774 +size 13556766 diff --git a/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/full.md b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..378e29f88c6b5770991a0fbc5f94dba904071d67 --- /dev/null +++ b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/full.md @@ -0,0 +1,325 @@ +# RayGaussX: Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis + +Hugo Blanc, Jean-Emmanuel Deschaud, Alexis Paljic Mines Paris, PSL University, Centre for robotics, 75006 Paris, France {hugo.blanc, jean-emmanuel.deschaud, alexis.paljic}@minesparis.psl.eu + +![](images/1206dfdd3cbfba3051142eebe2bb5c69900e978736ab98d402a93475a8fefca5.jpg) +Ground Truth + +![](images/abd54b604818c025bd794ca0ac5535fce32a87a44695dcb03c0f0d9b71f9cac0.jpg) +RayGaussX (ours) + +![](images/02fe8783b576ce7c3327fbcb7e55a6d38fba272b71a20385f53838406410aa81.jpg) +RayGauss +Figure 1. Comparison of RayGaussX with RayGauss, Zip-NeRF and 3D-GS. Test image from the garden scene of Mip-NeRF360. + +![](images/07c6eb210d5b744ec75b84ae5a11ee249518fc2559b60d9a95c4a3f35bf03291.jpg) +Zip-NeRF + +![](images/b36b23339d8bb4e4d491c3322d0ab09043487ee06e61941e0afd0bb93b6bc33e.jpg) +3D-GS + +# Abstract + +RayGauss has achieved state-of-the-art rendering quality for novel-view synthesis on synthetic and indoor scenes by representing radiance and density fields with irregularly distributed elliptical basis functions, rendered via volume ray casting using a Bounding Volume Hierarchy (BVH). However, its computational cost prevents real-time rendering on real-world scenes. Our approach, RayGaussX, builds on RayGauss by introducing key contributions that accelerate both training and inference. Specifically, we incorporate volumetric rendering acceleration strategies such as empty-space skipping and adaptive sampling, enhance ray coherence, and introduce scale regularization to reduce false-positive intersections. Additionally, we propose a new densification criterion that improves density distribution in distant regions, leading to enhanced graphical quality on larger scenes. As a result, RayGaussX achieves $5 \times$ to $12 \times$ faster training and $50 \times$ to $80 \times$ higher rendering speeds (FPS) on real-world datasets while improving visual quality by up to $+0.56$ dB in PSNR. The associated code is available at: github.com/hugobl1/raygaussx. + +# 1. Introduction + +Novel view synthesis (NVS) methods enable the generation of new images of a scene from a set of known viewpoints. This field has made significant progress with the introduction of Neural Radiance Fields (NeRF) [23]. This groundbreaking approach introduced a differentiable volumetric rendering equation coupled with a neural network to learn the scene's radiance fields. Subsequent research has further refined this method by enhancing both rendering quality and speed. One promising direction involves representing radiance and density fields using irregularly distributed local parametric functions. These approaches replace the computationally expensive neural networks used by NeRF with more localized structures that are less demanding and better suited to reconstruct the scene's geometry and appearance. Among these approaches, two main categories can be identified according to their rendering algorithms. The first relies on a differentiable rasterization algorithm for local primitives, such as Gaussians, as in the 3D Gaussian Splatting approach [16]. This rasterization algorithm enables high rendering speeds. However, it can introduce artifacts such as flickering, which is inherent to + +the rasterization process. The second category employs the volumetric ray-marching rendering algorithm introduced by NeRF, providing a more exhaustive sampling of the space and being less prone to rendering artifacts. The recent Ray-Gauss approach [5] leverages this type of algorithm to render scenes represented with irregularly distributed elliptical basis functions, achieving state-of-the-art graphical results on synthetic and indoor scenes. However, its performance slightly degrades in outdoor environments, and its training and inference times remain significant for complex scenes. Thus, this paper aims to address these limitations and enhance the method's applicability to a broader range of scenarios. To achieve this, we focus on key aspects to improve computational efficiency. First, we draw inspiration from acceleration strategies in classic volumetric ray marching, such as empty-space skipping and adaptive sampling, to speed up the rendering equation computation. Then, by analyzing the implementation specifics, we introduce a reordering of primitives and rays to better adapt the algorithm for parallel computation on GPUs, enhancing ray coherence and memory-access efficiency. Furthermore, we propose a scale-regularization function to minimize false-positive intersections. Finally, to improve rendering quality in largescale scenes, we present a new densification method that evenly increases the density of Gaussians in the scene, regardless of their distance from the camera. Thus, our contributions can be summarized as follows: + +- Integrating empty-space skipping and adaptive sampling strategies to accelerate the evaluation of the rendering equation. +- Enhancing ray coherence by reordering rays and primitives for efficient GPU parallelization. +- Introducing a scale-regularization function to mitigate false-positive intersections. +- Formulating a novel criterion for improving densification in distant regions. + +# 2. Related Work + +Novel view synthesis has been studied in the computer vision community for decades, with early approaches known as Image-Based Rendering (IBR), including [19], which interpolate new images from the closest known views. Another strategy involves using a proxy, such as a mesh [11], to enhance interpolation or even leveraging neural networks [14]. Point clouds can also serve as proxies, as seen in NPBG [1], NPBG++ [30], and ADOP [32]. However, these image-space interpolation methods suffer from artifacts and flickering. + +In 2020, NeRF [23] showed that scene density and radiance fields can be encoded within a Multi-Layer Perceptron (MLP), with its weights learned thanks to differentiable volumetric rendering. Numerous improvements have made NeRF faster, such as using grids of small + +MLPs in KiloNeRF [31], caching in FastNeRF [13], dense voxel grids [33], sparse voxel structures [15, 20], octree-based representations in PlenOctrees [39], or hash tables in Instant-NGP [27]. Other approaches have improved rendering quality by addressing aliasing, as in Mip-NeRF [2] and Mip-NeRF360 [3]. More recently, Zip-NeRF [4], which combines Mip-NeRF360 and Instant-NGP, has set the state of the art in rendering quality but suffers from slow training and does not allow interactive rendering. Furthermore, Plenoxels [12] demonstrated that NeRF's key aspect is the differentiability of the rendering equation. Instead of using a network to encode scene density and color parameters, they employ a voxel grid with spherical harmonics. Since then, other data structures have been explored, including point clouds in Point-NeRF [36], PointNeRF++ [34], and tetrahedra in Tetra-NeRF [18]. + +DSS [38], Point-Based Radiance [41], and the recent, widely adopted 3D Gaussian Splatting (3D-GS) [16] also use point clouds to encode scenes. However, instead of volumetric ray marching, they use a rasterization-based approximation: splat-based differentiable rendering enables training and rendering speeds several orders of magnitude faster than NeRF-like methods. Many approaches have since built upon 3D-GS, including Mip-Splatting [40] for anti-aliasing, Scaffold-GS [21] using anchors to better regularize Gaussian parameters, and Spec-Gaussian [37], which incorporates a MLP to better model specular and anisotropic colors. However, rasterization-based radiance-field rendering suffers from two key limitations [7]: flickering arising from image-space primitive ordering, which induces popping artifacts [29], and inherent approximations in the rendering equation (each primitive contributes only once per pixel, irrespective of its size). Furthermore, rasterization algorithms struggle to accommodate complex camera models and to simulate a broader range of physical phenomena, such as volumetric scattering. + +A final category relies on rendering primitives defined by radial basis functions (RBFs) via volumetric ray casting. NeuRBF [9] uses a hybrid representation with adaptive RBFs for fine details, but is limited to a small number of primitives per sample along a ray. More recently, 3DGRT [24], RayGauss [5] and EVER [22] have leveraged a Bounding Volume Hierarchy (BVH) structure and the OptiX library for fast ray casting in a scene using Gaussian-based primitives. RayGauss [5] showed state-of-the-art rendering quality on synthetic datasets like the NeRF dataset [23], but suffers from slow training and rendering on real-world datasets. + +Our approach, RayGaussX, extends RayGauss [5] by significantly reducing training and rendering times through key contributions that exploit irregularly distributed elliptical primitives rendered via volume ray marching. + +![](images/d583940bde6f13f460bdb2e4e754bbd9abe97378f3de3e026c7b7ddd54ea0fb8.jpg) +Figure 2. Main contributions of RayGaussX approach to improve training and rendering speed. + +![](images/03c86844506b08f4aef785613a4c82c3bd74ad31b942e06e19d61c50987db99e.jpg) + +# 3. RayGaussX approach + +# 3.1. Preliminary: gaussian-based ray-marching + +The RayGauss method offers a physically consistent formulation of the emitted radiance $c$ and density $\sigma$ , expressed through a decomposition using Gaussian functions combined with Spherical Gaussians/Harmonics (SG/SH) for an all-frequency colorimetric representation. Specifically, it represents the density $\sigma$ and radiance $c$ as: + +$$ +\sigma (\mathbf {x}) = \sum_ {l = 1} ^ {N} \tilde {\sigma} _ {l} \mathcal {G} (\mathbf {x}; \mu_ {l}, \mathbf {q} _ {l}, \mathbf {s} _ {l}) \tag {1} +$$ + +$$ +c (\mathbf {x}, \mathbf {d}) = \frac {\sum_ {l = 1} ^ {N} c _ {l} (\mathbf {d}) \tilde {\sigma} _ {l} \mathcal {G} (\mathbf {x} ; \mu_ {l} , \mathbf {q} _ {l} , \mathbf {s} _ {l})}{\sum_ {m = 1} ^ {N} \tilde {\sigma} _ {m} \mathcal {G} (\mathbf {x} ; \mu_ {m} , \mathbf {q} _ {m} , \mathbf {s} _ {m})} +$$ + +where $c_{l}$ is the radiance emitted by the $l$ -th primitive, parameterized with SH/SG, $\tilde{\sigma}_{l}$ is a density parameter of the $l$ -th primitive, and $\mathcal{G}$ is an unnormalized Gaussian. To render these fields, RayGauss relies on volumetric rendering theory, expressing the interaction of light with the medium through the classical volumetric rendering equation: + +$$ +C (\mathbf {r}) = \int_ {t _ {n}} ^ {t _ {f}} c (\mathbf {r} (t), \mathbf {d}) \sigma (\mathbf {r} (t)) T (t) d t, \tag {2} +$$ + +$$ +T (t) = e ^ {- \int_ {t _ {n}} ^ {t} \sigma (\mathbf {r} (s)) d s} +$$ + +To numerically compute this integral, the usual approach subdivides it into uniform segments and computes an approximation for each segment, yielding the Volume Ray Marching equation used in recent works [5, 23]: + +$$ +\begin{array}{l} C (\mathbf {r}) = \sum_ {i = 0} ^ {N} (1 - \exp (- \sigma_ {i} \Delta t)) c _ {i} T _ {i} \tag {3} \\ T _ {i} = \exp \left(- \sum_ {j = 0} ^ {i - 1} \sigma_ {j} \Delta t\right) \\ \end{array} +$$ + +where the step size $\Delta t$ is constant. In the RayGauss approach, this equation is efficiently computed using an algorithm that leverages a Bounding Volume Hierarchy (BVH) to rapidly compute ray-primitive intersections. + +In this section, we present our main contributions aimed at improving this algorithm: empty-space skipping, adaptive sampling, enhanced ray coherence through ray and primitive reordering, scale regularization to reduce false positive intersections, and finally, a novel densification criterion improving rendering quality in distant regions. + +# 3.2. Efficient Ray Sampling: Empty Space Skipping and Adaptive Sampling + +As illustrated by Eq. (3), the computational cost of evaluating the color per ray is directly related to the number of samples $N$ . Thus, to reduce computation time, a key approach is to minimize the samples needed for ray color computation. Two main types of approaches can be distinguished: the first, empty space skipping, aims to avoid sampling completely transparent regions with null density. Since these regions do not contribute to the integration, such methods do not affect rendering quality. A second type of approach consists in adjusting the sampling step based on the integrated space region, focusing samples on areas that contribute the most to color computation. These two types of approaches are complementary, and in the following, we introduce advancements in both areas that accelerate the computation times of the RayGauss approach. + +Empty Space Skipping: As previously discussed, the objective is to design a method that avoids sampling in regions with zero density. A key advantage of RayGauss [5] is its representation based on sparsely distributed truncated Gaussians, which explicitly encode regions of non-zero density. In contrast, other representations, such as those based on neural networks, do not define these regions ex + +plicity. Additionally, RayGauss stores these primitives in a Bounding Volume Hierarchy (BVH) for efficient intersection computations. Ray integration is performed sequentially: at step $i$ , a segment $S_{i} = [t_{i}, t_{i} + N_{s} \cdot \Delta t]$ containing a fixed number $N_{s}$ of samples is considered. The BVH is used to compute the intersection of the scene's Gaussians with $S_{i}$ . If no intersection is found, the next segment is evaluated; otherwise, the contribution of this segment to the color is computed. Thus, at the segment level, RayGauss can detect whether the region is empty. Our approach enhances this empty space skipping by leveraging the precomputed BVH, which is already required by the algorithm. + +Algorithm 1 Volume Ray Marching +Input: $(o,d)$ : ray origin/direction, $\Delta t$ : step size, $N_{s}$ : buffer size, $T_{\epsilon}$ : transmittance threshold, $(t_n,t_f)$ : integration bounds, $\mathcal{H}_b$ : hit primitive index buffer, $\mathcal{G}_p$ : gaussians parameters +Output: $C_R$ +1: $\Delta S\gets \Delta t\times N_s$ $\triangleright$ Segment size +2: $T\gets 1.0$ $\triangleright$ Ray transmittance +3: $C_R\gets (0.0,0.0,0.0)$ $\triangleright$ Ray color +4: $t_{min}\gets \mathrm{BVHClosestHit}(o,d,t_n,t_f)$ +5: $t_S\gets t_{min}$ $\triangleright$ Segment distance along the ray +6: while $t_S < t_f$ and $T > T_{\epsilon}$ do +7: $n_p\gets 0$ $\triangleright$ Number of primitives +8: $\triangleright$ Collect the intersected primitives +9: $(\mathcal{H}_b,n_p)\gets \mathrm{BVHTraversal}(o,d,t_S,t_S + \Delta S)$ +10: if $n_p == 0$ then +11: $t_S\gets t_S + \Delta S$ +12: $\triangleright$ Recompute closest hit +13: $t_S\gets \mathrm{BVHClosestHit}(o,d,t_S,t_f)$ +14: continue +15: end if +16: $\triangleright$ Update ray color and transmittance +17: $(C_R,T)\gets \mathrm{RayUpdate}(n_p,\Delta t,t_S,o,d,\mathcal{G}_p,\mathcal{H}_b,C_R,T)$ +18: $t_S\gets t_S + \Delta S$ +19: end while + +Our method follows the procedure described in Algorithm 1. Ray tracing libraries, such as the OptiX API used by RayGauss, are designed to efficiently compute the closest intersection point for a given ray, as this is a fundamental operation in classical ray tracing algorithms. Our idea is to leverage this by alternating between two types of processes during scene traversal. On the one hand, the ClosestHit shader computes the nearest intersection point with primitives for a ray constrained to a given region, e.g., $[t_n, t_f]$ . On the other hand, the AnyHit shader (referred to as BVHTraversal in line 9 of Alg. 1), already implemented in RayGauss [5], collects all primitives intersected along the segment. Thus, starting from the initial integration bound $t_n$ , we first compute the closest hit with the primitives at $t_{\mathrm{min}}$ which allows us to avoid sampling empty regions between + +the camera and the closest primitive. Indeed: + +$$ +\begin{array}{l} \sigma (\mathbf {r} (t)) = 0, \quad \forall t \in [ t _ {n}, t _ {\min } ], \\ \int_ {t _ {n}} ^ {t _ {\min }} c (\mathbf {r} (t), \mathbf {d}) \sigma (\mathbf {r} (t)) T (t) d t = 0. \tag {4} \\ \end{array} +$$ + +Once the closest intersection point is found, we begin the integration process as described in [5]. However, other large empty regions may still exist along $[t_{\mathrm{min}}, t_f]$ . Moreover, the AnyHit shader not only collects primitives but also determines their number. If no primitives are detected within the current segment $[t_S, t_S + \Delta S]$ , it indicates that the region has zero density. In this case, the ClosestHit program is invoked again to locate the next point of interest. This approach effectively avoids exhaustive sampling of empty regions while adapting to the specific requirements of the Volume Ray Casting algorithm developed in [5]. + +Adaptive Sampling: A second approach to accelerating the computation of the volumetric rendering equation is to adapt the spatial subdivision of the integral to reduce the number of samples. Several studies have explored this problem. Danskin et al. [10] leveraged importance sampling principles to construct a pyramidal structure enabling adaptive sampling. More recently, Morrical et al. [25] employed a kd-tree to adjust the sampling step based on the variance within the subpartitions. Campagnolo et al. [6] proposed an adaptive Simpson integration scheme to tackle this problem. Thus, adaptive sampling strategies are designed to align with the specific constraints of the scene representation and rendering algorithm to determine an adaptive step size. In our case, two criteria are important. During training, the structure is updated at each optimization step, meaning that an approach storing the sampling criterion in an auxiliary structure, as in Morrical et al. [25], would require frequent reconstruction, increasing computational cost. Furthermore, the RayGauss rendering algorithm sequentially accumulates properties segment by segment, with each segment containing multiple samples. To retain the algorithm's advantages, we adopt adaptive subdivision at the segment level while keeping a fixed number of samples $N_{S}$ per segment for improved ray coherence. Since our primary goal is to accelerate rendering, we introduce a criterion with minimal computational overhead. Danskin et al. [10] suggest that adaptive sampling should depend on transmittance. Here, we define a criterion using the transmittance at the end of $(i - 1)$ -th segment to determine the size of the $i$ -th segment. The idea is that for a low resulting transmittance, the subsequent segments will contribute less to the rendering integral, allowing for a larger step size. Furthermore, we also make the sampling criterion dependent on the distance to the camera, assuming that distant regions require less precision than closer ones. + +Consequently, the size of the $i$ -th segment, $\Delta S_{i}$ , can be dynamically adapted using a criterion based on the current distance from the origin, $d_{i} = \| s_{i} - o\|$ , and the transmittance, $T_{i} = T(s_{i})$ , where $s_i$ is the position at the beginning of the segment and $o$ is the camera origin: + +$$ +\Delta S _ {i} = N _ {S} \cdot \min \left(\max \left(\frac {d _ {i}}{\beta}, \Delta t _ {\min }\right) \cdot T _ {i} ^ {- \frac {1}{3}}, \Delta t _ {\max }\right) \tag {5} +$$ + +where $\beta$ quantifies the variation of the step size with respect to the distance from the origin, $\Delta t_{\mathrm{min}}$ is the minimum step size and $\Delta t_{\mathrm{max}}$ is the maximum step size to prevent excessively large variations in the sampling step. + +# 3.3. Optimizing Ray Coherence and Memory Access Efficiency + +This section explores improvements enabled by the specifics of parallel GPU computation, particularly through enhanced ray coherency, ensuring more consistent data access patterns and execution paths. + +Enhancing Memory Access Efficiency through Spatial Reordering of Gaussians: The Volume Ray Marching algorithm from [5] efficiently renders irregularly distributed Gaussians using the OptiX API. Scene parameter optimization and inference are performed per image, which is significant because rays from the same image tend to traverse spatially adjacent regions, a property not guaranteed if originating from different viewpoints. Consequently, during rendering, these rays pass through contiguous spatial areas and intersect nearby Gaussians. In this context, it is important to note that global memory access, where Gaussian parameters are stored, benefits from spatial locality in modern frameworks like CUDA, used by the OptiX API. Memory accesses occur via the transfer of contiguous blocks, and spatial locality in memory enables more efficient coalescing and higher data throughput. Moreover, spatially adjacent Gaussians are accessed simultaneously, suggesting that grouping their parameters in memory and reordering them based on their spatial distribution could enhance performance by improving memory coalescing. + +The objective is to efficiently reorder data, so that spatially proximate Gaussians are stored in contiguous memory regions. To achieve this, we employ Z-order curve theory, which maps multidimensional integer data point to an integer index while preserving data locality. This means that points that are close together in the multidimensional space are more likely to have nearby indices [26]. In practice, the Gaussian mean serves as an indicator of its spatial position, which is normalized, quantized into an integer, and mapped to a Z-order index. The Gaussians are then reordered in ascending Z-index order, allowing spatially close Gaussians to be stored in neighboring memory regions (illustrated in Fig. 2 and applied to the bonsai scene in Fig. 3). + +![](images/87601a5bd5e7cfad0c81b90f0880d80a9f1b2903a883262c83d62d521a14eff2.jpg) +(a) Default indexing +Figure 3. Comparison of the bonsai scene from the Mip-NeRF360 dataset: the left image shows the scene with default indexing, while the right image uses indexing obtained via a Z-order curve. Note that the color scale ranges from yellow to red, and that colorimetric similarity corresponds to a proximity of indices, hence contiguous memory allocation. + +![](images/821780a106649262388d7b6e2b169c9e189239614f3eedf6809571b4b95bea5f.jpg) +(b) Z-order curve indexing + +Improving Ray Coherence: In the classical parallel computing model supported by CUDA [28], computations are performed in parallel by threads. In the CUDA architecture, these threads are organized into groups of 32, called warps. Within a warp, all threads execute the same instruction simultaneously, although each may process different data. Maximum efficiency is achieved when all 32 threads follow the same execution path. However, if divergence occurs, such as through a data-dependent conditional branch, the warp must sequentially execute each path taken, temporarily disabling inactive threads. While necessary for handling different execution scenarios, this mechanism significantly degrades performance by reducing parallelism efficiency. This phenomenon is known as warp divergence. + +If each ray is processed in parallel, minimizing warp divergence requires maximizing coherence among rays within a warp, ensuring that as many rays as possible execute the same operations on similar data. However, achieving this coherence depends on how rays are organized. Ray-Gauss [5] does not perform any specific ray ordering, leading to an initial grouping in scanline order. Without prior scene information, dividing the image into tiles and grouping rays accordingly generally reduces variance among rays compared to a scanline-order approach. Rays within the same tile typically sample similar regions (and thus similar Gaussians), whereas scanline ordering may group rays observing significantly different parts of the scene, increasing divergence. In practice, we leverage the OptiX API, which transparently manages this initial distribution by launching rays on a 2D grid. Although threads may later be dynamically reordered, making the process more opaque, this approach provides an effective initial grouping. + +# 3.4. Limiting Highly Anisotropic Gaussian + +We observe that differentiable rendering algorithms relying on Gaussian kernels often produce highly anisotropic Gaussians, with substantial variations in their scale values (e.g., $s_{\min} \ll s_{\max}$ ). This phenomenon negatively impacts the + +computational efficiency of RayGauss [5]. Indeed, the algorithm leverages a BVH for efficient intersection computations. However, for custom primitives like ellipsoids, an Axis-Aligned Bounding Box (AABB) is first constructed around each ellipsoid, and a BVH is then built around these AABBs. Subsequently, during intersection computations, the optimized OptiX API first computes the intersection with the AABB and then calls the user-code intersection program for the ellipsoid. Here, the smallest enclosing AABB is used for each ellipsoid. If this AABB's volume is significantly larger than that of the ellipsoid, more false-positive calls to the user-code intersection program occur. This slows down BVH traversal and is particularly problematic in the context of hardware-accelerated ray intersection [35]. Thus, to accelerate rendering, reducing false-positive hits is beneficial. Also, their occurrence is correlated with the following volume ratio: + +$$ +r = \frac {\operatorname {V o l} (\mathrm {A A B B} (\mathcal {E}))}{\operatorname {V o l} (\mathcal {E})} \tag {6} +$$ + +with Vol the volume function, $\mathcal{E}$ a given ellipsoid, and $\mathrm{AABB}(\mathcal{E})$ its associated axis-aligned bounding box. When $r$ is high, many false-positive intersections occur. Conversely, when $r$ is low, false-positive intersections decrease. To ensure a rotation-invariant criterion and avoid bias toward any specific Gaussian orientation, we constrain the worst-case scenario with respect to rotation by computing an upper bound. (see supplementary material for details): + +$$ +\frac {\operatorname {V o l} (\mathrm {A A B B} (\mathcal {E}))}{\operatorname {V o l} (\mathcal {E})} \leq \frac {2}{\pi \sqrt {3}} \frac {\left(s _ {1} ^ {2} + s _ {2} ^ {2} + s _ {3} ^ {2}\right) ^ {\frac {3}{2}}}{s _ {1} s _ {2} s _ {3}} \tag {7} +$$ + +Thus, to minimize false-positive intersections, we add a loss term (hereafter referred to as the isotropic loss) that constrains this ratio for all Gaussians: + +$$ +L _ {s} = \frac {1}{N} \sum_ {l} \left(\max \left(r _ {\max , l}, r _ {0}\right) - r _ {0}\right) \tag {8} +$$ + +with: + +$$ +r _ {\max , l} = \frac {2}{\pi \sqrt {3}} \frac {\left(s _ {l , 1} ^ {2} + s _ {l , 2} ^ {2} + s _ {l , 3} ^ {2}\right) ^ {\frac {3}{2}}}{s _ {l , 1} s _ {l , 2} s _ {l , 3}} \tag {9} +$$ + +and $r_0$ is a threshold ratio beyond which a penalty is applied. Then, the total loss function is: $L = (1 - \lambda) \cdot L_1 + \lambda \cdot \mathrm{DSSIM} + \lambda_s L_s$ where $\lambda_s$ is a fixed hyperparameter. + +# 3.5. A Novel Densification Criterion + +RayGauss densification process [5] is inspired by the adaptive density control heuristic introduced in 3D Gaussian Splatting [16]. This heuristic enables cloning or splitting the Gaussians in the scene by evaluating a criterion every $I$ iterations. For a Gaussian $l$ with mean position $\mu_l$ , let $\mu_l^P$ + +be its projection in normalized device coordinates (NDC). If this Gaussian has been observed in $I_{l} < I$ images during recent iterations, the heuristic determines whether to densify it by comparing the average norm of the loss gradient with respect to $\mu_l^P$ across these $I_{l}$ images to a given threshold $\tau$ : + +$$ +\frac {1}{I _ {l}} \sum_ {i = 1} ^ {I _ {l}} \left\| \nabla_ {\mu_ {l} ^ {P}} L _ {i} \right\| > \tau . \tag {10} +$$ + +where $L_{i}$ is the loss computed for the $i$ -th image in which Gaussian $l$ appears. RayGauss does not compute $\mu_l^P$ , since rendering uses volumetric ray casting rather than projection and splatting, and replaces the densification criterion with: + +$$ +\frac {1}{I _ {l}} \sum_ {i = 1} ^ {I _ {l}} \| \nabla_ {\mu_ {l}} L _ {i} \| > \tau . \tag {11} +$$ + +where the gradient with respect to $\mu_l$ is computed during optimization. However, a limitation of the RayGauss formulation (Eq. 11) is that it introduces a dependency on the distance between the Gaussian and the camera. Specifically, distant Gaussians are more likely to exhibit low gradients compared to closer ones, even if their contributions to the image are similar. This occurs because a slight deviation of the mean in world space results in a smaller variation in the image contribution of a distant Gaussian compared to a closer one, due to perspective projection. As a result, Gaussians farther from the camera exhibit lower gradients and are less likely to be densified. This may explain the lower performance of RayGauss on outdoor scenes compared to indoor environments. In contrast, the original densification criterion from [16] does not suffer from this issue, as it relies on the view-space position gradient. This prevents the criterion from being biased by the Gaussian's distance to the camera. Indeed, a slight deviation in view space does not favor closer Gaussians over distant ones. For this reason, we modify the criterion from the RayGauss approach [5] by introducing a corrective factor that weights the gradients in 3D space, ensuring densification of distant Gaussians: + +$$ +\frac {1}{I _ {l}} \sum_ {i = 1} ^ {I _ {l}} \alpha_ {i} \| \nabla_ {\mu_ {l}} L _ {i} \| > \tau . \tag {12} +$$ + +where $\alpha_{i} = \frac{\|\mu_{l} - o_{i}\|}{f}$ , with $o_{i}$ being the center of camera $i$ , and $f$ the focal distance. Empirically, we observe that $\nabla_{\mu_l}L_i\cdot \frac{\mu_l - o_i}{\|\mu_l - o_i\|}$ is generally negligible compared to its tangential component. Under this approximation, our criterion corresponds to averaging the norms of the gradients with respect to the mean projected onto the sphere of radius $f$ (see supplementary for details). This criterion therefore remains closer to the original criterion of 3D Gaussian Splitting, accounts for the scaling factor, and is efficiently computable. + +# 4. Experiments and Results + +# 4.1. Experiments + +The experiments are conducted on two synthetic datasets, the NeRF Synthetic dataset [23] and the NSVF Synthetic dataset [20], as well as on three real-world datasets: Mip-NeRF360 [3], Tanks&Temples [17], and Deep Blending [14]. We evaluate our approach using standard image synthesis metrics, including rendering quality (PSNR, SSIM, and LPIPS computed with a VGG network), training time (Train), and rendering speed (FPS, Frames Per Second). Our RayGaussX approach is compared to the RayGauss method [5] as well as to state-of-the-art novel view synthesis techniques. Specifically, we include Zip-NeRF [4], which achieves the highest visual quality but requires a long training time and does not support real-time rendering, and 3D-GS [16], a rasterization-based Gaussian rendering method that offers the fastest performance. Additionally, we compare against more recent variants of 3D-GS: Mip-Splatting [40], which addresses aliasing, and Spec-Gaussian [37], which incorporates an MLP network to better model high-frequency color details. All methods were retrained from their official repositories on a single NVIDIA RTX 4090 to ensure a fair comparison of training and rendering performance (see supplementary). + +# 4.2. Implementation details + +The training parameters for the RayGaussX method are identical to those used for the RayGauss method [5]. The only differences concern the following contributions: + +- Sampling strategy: We use uniform sampling with $\Delta t = 0.0025$ for synthetic datasets (same as Ray-Gauss), whereas for real-world datasets, we adopt adaptive sampling with $\Delta t_{\mathrm{min}} = 0.005$ , $\Delta t_{\mathrm{max}} = 4 \times \Delta t_{\mathrm{min}}$ , $\beta = 1024$ . In contrast, to benchmark Ray-Gauss, we kept a fixed $\Delta t = 0.005$ for real-world datasets. +- Loss for isotropic Gaussians: This loss is applied to real-world datasets with the following hyperparameters $\lambda_{s} = 0.00025$ , $r_0 = 10$ . +- New densification: We use the parameter $\tau = 0.00015$ . in all experiments. + +No major hyperparameters are introduced by the Empty Space Skipping, Ray Coherence, or Memory Reordering of Gaussians techniques. + +# 4.3. Results on synthetic datasets + +Tab. 1 and Tab. 2 present the results of our RayGaussX approach in comparison with RayGauss [5] and other novel view synthesis methods on the synthetic NeRF [23] and NSVF [20] datasets. All methods were trained with a white background, a standard practice for comparisons on synthetic data. + +It can be observed that for both datasets, RayGaussX achieves the best rendering quality (slightly superior to RayGauss) while offering a rendering speed three times faster in terms of FPS. For synthetic datasets, we applied empty space skipping, z-curve indexing, ray reordering, and uniform sampling (to maximize rendering quality), which already provide significant improvements in rendering speed. + +NeRF Synthetic + +
Method \ MetricPSNR↑SSIM↑LPIPS↓Train↓FPS↑
Instant-NGP [27]33.180.9630.045--
Mip-NeRF360 [3]33.240.9610.042--
Point-NeRF [36]33.300.9620.049--
3D-GS [16]*33.390.9680.0314min681
Zip-NeRF [4]*33.690.9740.028445min0.3
Mip-Splatting [40]*33.740.9710.0295min468
Spec-Gaussian [37]*33.800.9700.02711min142
3DGRT [24]*33.860.9700.0379min304
NeuRBF [9]34.470.9740.035-≈1
RayGauss [5]*34.530.9740.02433min25
RayGaussX (ours)34.540.9740.02422min85
+ +Table 1. Results on NeRF Synthetic dataset [23]. Methods marked * were retrained on an NVIDIA RTX4090 for fair comparison of training and rendering (FPS) times. +NSVF Synthetic + +
Method \ MetricPSNR↑SSIM↑LPIPS↓Train↓FPS↑
TensoRF [8]36.530.9820.026--
3D-GS [16]*37.070.9870.0146min498
Mip-Splatting [40]*37.280.9880.0146min445
NeuRBF [9]37.800.9880.019-≈1
Spec-Gaussian [37]*37.970.9890.01215min102
RayGauss [5]*38.720.9900.01142min22
RayGaussX (ours)38.750.9910.01026min71
+ +Table 2. Results on NSVF Synthetic dataset [20]. Methods marked * were retrained on an NVIDIA RTX4090 for fair comparison of training and rendering (FPS) times. + +# 4.4. Results on real-world datasets + +Tab. 3 reports results on three real-world datasets: MipNeRF360, Tanks&Temples, and Deep Blending [3, 14, 17]. + +First, we observe that compared to RayGauss, Ray-GaussX achieves training times that are 5 to 12 times faster and rendering speeds that are 50 to 80 times faster in FPS, while also improving quality by up to $+0.56$ dB in PSNR on the Tanks&Temples dataset. Also, RayGaussX is 100x to 200x faster in FPS than the state-of-the-art method Zip-NeRF [4], while being only -0.13 dB lower in PSNR on Mip-NeRF360. Furthermore, RayGaussX outperforms Spec-Gaussian [37], the best Gaussian splatting variant, by $+0.38$ dB in PSNR on Mip-NeRF360 dataset, while being only twice as slow in rendering speed. Overall, RayGaussX achieves better rendering quality than Gaussian splatting-based methods (3D-GS [16], Mip-Splatting [40], and Spec-Gaussian [37]) while approaching the rendering speed of + +
DatasetMip-NeRF360Tanks&TemplesDeep Blending
Method\MetricPSNR↑SSIM↑LPIPS↓Train↓FPS↑PSNR↑SSIM↑LPIPS↓Train↓FPS↑PSNR↑SSIM↑LPIPS↓Train↓FPS↑
Instant-NGP [27]25.590.6990.331--21.920.7450.305--24.970.8170.390--
3DGRT [24]*26.890.8070.25960min5922.740.8440.19728min11129.740.9050.31549min65
EVER [22]27.510.8250.194-36----------
Mip-NeRF360 [3]27.690.7920.237--22.220.7590.257--29.400.9010.245--
3D-GS [16]*27.800.8250.20821min16123.720.8480.17814min19729.920.9050.24422min176
Mip-Splatting [40]*27.930.8380.17627min11423.720.8600.15715min17029.510.9030.24324min139
Spec-Gaussian [37]*28.050.8350.17741min4223.860.8560.16621min8529.670.9060.24526min75
Zip-NeRF [4]*28.560.8280.176425min0.110.850.3260.661422min0.330.760.9140.209416min0.2
RayGauss [5]*28.230.8340.193456min0.523.200.8490.167565min0.530.300.9140.242553min0.8
RayGaussX (ours)28.430.8420.17785min24.123.760.8650.15047min41.130.320.9150.24256min39.6
+ +the most recent variant, Spec-Gaussian [37]. As shown in Tab. 3, the addition of features to improve the rendering quality of 3D-GS, such as antialIASing in Mip-Splatting [40] and the MLP for high-frequency color prediction in Spec-Gaussian [37], has progressively slowed down rasterization-based methods, decreasing from 161 FPS for 3D-GS to 42 FPS for Spec-Gaussian on Mip-NeRF360 dataset. + +Secondly, we observe that the most notable improvement in rendering quality between RayGauss and RayGaussX occurs in outdoor scenes (from the Mip-NeRF360 [3] and Tanks&Temples [17] datasets). This improvement is due to the new densification strategy, which allows densifying Gaussians far from the camera and better reconstructing distant areas. + +Qualitative results on the Mip-NeRF360 [3] dataset are shown in Figure 1 and in the supplementary material. + +# 4.5. Ablation study on the main contributions + +Tab. 4 presents an ablation study on the different contributions of RayGaussX that significantly accelerate rendering. This study was conducted on the garden scene from the Mip-NeRF360 [3] dataset. We decomposed RayGaussX into four main contributions: the new densification (D), empty space skipping with adaptive sampling $(\mathrm{E} + \mathrm{A})$ , reordering of Gaussians with ray coherence $(\mathrm{Z} + \mathrm{R})$ , and finally, the isotropic loss (L). The first row (1) of Tab. 4 represents the original RayGauss method, row $(1^{*})$ is our optimized RayGauss baseline with only implementation-level optimizations and the last row (8) corresponds to our complete RayGaussX approach. + +It can be observed that all four contributions improve the training time and rendering speed of RayGauss. The new densification alone (row (2)) significantly improves rendering quality (28.38 dB) while also slightly enhancing training and rendering speed. This is due to the novel densification strategy, which, compared to RayGauss, adds more distant Gaussians while reducing the number of nearby ones, thereby slightly speeding up the ray marching process. The most impactful contributions for accelerating ray marching with Gaussian primitives are empty space skipping with + +adaptive sampling (row (3), achieving a $3.9 \times$ speedup in FPS compared to $(1^{*})$ ) and the isotropic loss (row (5), yielding a $6.7 \times$ increase in FPS). Additionally, the isotropic loss alone (row (5)) results in a -0.2 dB drop in PSNR compared to $(1^{*})$ , but it provides significant gains in training and rendering speed. For this reason, the isotropic loss is preferred for real-world datasets but is not retained for synthetic datasets. Finally, as more contributions are added, rendering speed progressively improves, reaching 27 FPS for interactive rendering, compared to 0.5 FPS for Ray-Gauss at a test image resolution of $1297 \times 840$ pixels. + +Table 3. Results on Mip-NeRF360 [3], Tanks&Temples [17] and Deep Blending [14] datasets. FPS are calculated on the resolutions of the test set images. Methods marked * were retrained on an NVIDIA RTX4090 for fair comparison of training and rendering (FPS) times. + +
DE+AZ+RLPSNR↑SSIM↑LPIPS↓Train↓FPS↑
(1)28.060.8790.103585min0.5
(1*)28.140.8850.090297min1.5
(2)✓28.380.8870.090294min1.8
(3)✓28.140.8840.090180min5.9
(4)✓28.150.8850.089200min3.6
(5)✓28.120.8850.090133min10.1
(6)✓✓28.350.8870.089168min6.3
(7)✓✓✓28.360.8880.089116min10.5
(8)✓✓✓✓28.350.8870.09084min27.4
+ +Table 4. Ablation study of key contributions on garden scene from Mip-NeRF360 dataset (D: new Densification criterion, E+A: Empty space Skipping + Adaptive sampling, Z+R: Z-curve re-indexing of gaussians + Ray coherence, L: Loss for isotropic gaussians). Row (1) represents the original RayGauss method [5], row $(1^{*})$ is our optimized RayGauss baseline with only implementation-level optimizations (excluding D, E, A, Z, R, L), and row (8) corresponds to our full method, RayGaussX. + +# 5. Conclusion + +Building on the RayGauss algorithm [5], which has achieved great graphical results on synthetic and indoor scenes, we propose a new approach for real-time novel view synthesis. Our approach introduces innovative contributions: empty-space skipping, adaptive sampling, reordering of rays and primitives for improved ray coherence, scale regularization, and a novel densification criterion. These enhancements enable the real-time synthesis of novel views with state-of-the-art graphical quality across indoor and outdoor scenes from multiple datasets. + +# References + +[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision (ECCV), 2020. 2 +[2] 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 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 7, 8 +[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7, 8 +[5] Hugo Blanc, Jean-Emmanuel Deschaud, and Alexis Paljic. Raygauss: Volumetric gaussian-based ray casting for photorealistic novel view synthesis, 2024. 2, 3, 4, 5, 6, 7, 8 +[6] Leonardo Quatron Campagnolo, Waldemar Celes, and Luiz Henrique de Figueiredo. Accurate volume rendering based on adaptive numerical integration. In 2015 28th SIB-GRAPI Conference on Graphics, Patterns and Images, pages 17–24, 2015. 4 +[7] Adam Celarek, Georgios Kopanas, George Drettakis, Michael Wimmer, and Kerbl Bernhard. Does 3d gaussian splatting need accurate volumetric rendering? Computer Graphics Forum, 44(2):e15256, 2025. 2 +[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 7 +[9] Zhang Chen, Zhong Li, Liangchen Song, Lele Chen, Jingyi Yu, Junsong Yuan, and Yi Xu. Neurbf: A neural fields representation with adaptive radial basis functions. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7 +[10] John Danskin and Pat Hanrahan. Fast algorithms for volume ray tracing. In Proceedings of the 1992 Workshop on Volume Visualization, page 91-98, New York, NY, USA, 1992. Association for Computing Machinery. 4 +[11] Paul Debevec, Yizhou Yu, and George Borshukov. Efficient view-dependent image-based rendering with projective texture-mapping. In Eurographics Rendering Workshop, 1998. 2 +[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[13] Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 + +[14] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Trans. Graph., 37(6), 2018. 2, 7, 8 +[15] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4), 2023. 1, 2, 6, 7, 8 +[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 7, 8 +[18] J. Kulhanek and T. Sattler. Tetra-nerf: Representing neural radiance fields using tetrahedra. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[19] Marc Levoy and Pat Hanrahan. Light field rendering. In Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 1996. 2 +[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 2, 7 +[21] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[22] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T. Barron, and Yinda Zhang. Ever: Exact volumetric ellipsoid rendering for real-time view synthesis, 2024. 2, 8 +[23] 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 (ECCV), 2020. 1, 2, 3, 7 +[24] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics and SIGGRAPH Asia, 2024. 2, 7, 8 +[25] Nate Morrical, Will Usher, Ingo Wald, and Valerio Pascucci. Efficient space skipping and adaptive sampling of unstructured volumes using hardware accelerated ray tracing. In 2019 IEEE Visualization Conference (VIS), pages 256-260, 2019. 4 +[26] G.M. Morton. A computer oriented geodetic data base; and a new technique in file sequencing. Technical report, IBM Ltd., Ottawa, Canada, 1966. 5 +[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires + +olution hash encoding. ACM Trans. Graph., 41(4), 2022. 2, 7, 8 +[28] John Nickolls, Ian Buck, Michael Garland, and Kevin Skadron. Scalable parallel programming with CUDA. In ACM SIGGRAPH 2008 Classes, New York, NY, USA, 2008. Association for Computing Machinery. 5 +[29] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering. ACM Transactions on Graphics, 4 (43), 2024. 2 +[30] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. NPBG++: Accelerating neural point-based graphics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[31] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[32] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: approximate differentiable one-pixel point rendering. ACM Trans. Graph., 41(4), 2022. 2 +[33] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[34] Weiwei Sun, Eduard Trulls, Yang-Che Tseng, Sneha Sambandam, Gopal Sharma, Andrea Tagliasacchi, and Kwang Moo Yi. PointNeRF++: A multi-scale, point-based Neural Radiance Field. arXiv e-prints, art. arXiv:2312.02362, 2023. 2 +[35] Ingo Wald, Nate Morrical, Stefan Zellmann, Lei Ma, Will Usher, Tiejun Huang, and Valerio Pascucci. Using Hardware Ray Transforms to Accelerate Ray/Primitive Intersections for Long, Thin Primitive Types. Proceedings of the ACM on Computer Graphics and Interactive Techniques (Proceedings of High Performance Graphics), 2020. 6 +[36] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5428-5438, 2022. 2, 7 +[37] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 2, 7, 8 +[38] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Trans. Graph., 38 (6), 2019. 2 +[39] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[40] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splat + +ting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 2, 7, 8 +[41] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia Conference, 2022. 2 \ No newline at end of file diff --git a/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/images.zip b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9e4764be1c9dbcfd675ad03c0335546f65ab9713 --- /dev/null +++ b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5e63d701d1b3aabfaf2250473db2e3d20da68b433b0e3b1c71a1a471c19e135 +size 527460 diff --git a/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/layout.json b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..33d2c7f015fe78fbdadbf15bc6822e1d15966c0d --- /dev/null +++ b/ICCV/2025/RayGaussX_ Accelerating Gaussian-Based Ray Marching for Real-Time and High-Quality Novel View Synthesis/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c63979149166c4e483fcd6f5d88bf5b454187f8c56b78e8c3f4bd9bb2036df34 +size 390805 diff --git a/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/8e133c3d-e963-4ff1-a4d4-46823ed648d2_content_list.json b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/8e133c3d-e963-4ff1-a4d4-46823ed648d2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6855f2ed21bc91c238d09474c6f823e1d7c9e2ee --- /dev/null +++ b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/8e133c3d-e963-4ff1-a4d4-46823ed648d2_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc97a1cf14ed8759d879bb5fb481fce2a6ac5a3ac64d2a2c958dc571b065bb97 +size 87374 diff --git a/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/8e133c3d-e963-4ff1-a4d4-46823ed648d2_model.json b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/8e133c3d-e963-4ff1-a4d4-46823ed648d2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..54ac112e1dc466efbb4660a47b1956e9c7a118c7 --- /dev/null +++ b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/8e133c3d-e963-4ff1-a4d4-46823ed648d2_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5cb0308004aea30ce3db0309ca6c73d025a9277d6fcd6bd654939a6fac00edb +size 108696 diff --git a/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/8e133c3d-e963-4ff1-a4d4-46823ed648d2_origin.pdf b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/8e133c3d-e963-4ff1-a4d4-46823ed648d2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6897edc64159fc6f17c5607231c925a45cbdc1c2 --- /dev/null +++ b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/8e133c3d-e963-4ff1-a4d4-46823ed648d2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00c455617115403a2ce633dd6ee935934c3c755c6a7da6e333466fad9faa2e34 +size 1412874 diff --git a/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/full.md b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..29b0604448a91b832891d09c268a727007c5ea87 --- /dev/null +++ b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/full.md @@ -0,0 +1,351 @@ +# RayPose: Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation + +Junwen Huang $^{1,2}$ + +Shishir Reddy Vutukur1,2 + +Peter KT Yu3 + +Slobodan Ilic + +Benjamin Busam1,2 + +Nassir Navab1,2 + +$^{1}$ Technical University of Munich $^{2}$ Munich Center for Machine Learning $^{3}$ XYZ Robotics + +# Abstract + +Typical template-based object pose pipelines estimate the pose by retrieving the closest matching template and aligning it with the observed image. However, failure to retrieve the correct template often leads to inaccurate pose predictions. To address this, we reformulate template-based object pose estimation as a ray alignment problem, where the viewing directions from multiple posed template images are learned to align with a non-posed query image. Inspired by recent progress in diffusion-based camera pose estimation, we embed this formulation into a diffusion transformer architecture that aligns a query image with a set of posed templates. We reparameterize object rotation using object-centered camera rays and model object translation by extending scale-invariant translation estimation to dense translation offsets. Our model leverages geometric priors from the templates to guide accurate query pose inference. A coarse-to-fine training strategy based on narrowed template sampling improves performance without modifying the network architecture. Extensive experiments across multiple benchmark datasets show competitive results of our method compared to state-of-the-art approaches in unseen object pose estimation. + +# 1. Introduction + +Multi-view vision is a core element for 3D perception [12]. Spatial understanding and measurements often depends on multiple cameras or temporally-varied perspectives over time to reason about the surrounding in 3D. Also for the task of object pose estimation – the prediction of rotation and translation of objects in space, multi-view constraints can be beneficial [25]. In many computer vision applications, like robotic bin picking, augmented reality, and autonomous driving, multiple cameras or acquisitions are not + +![](images/54ef0293c0bde25614a4527e79310da6ef4ca27e978d801528cc07d5366be4c9.jpg) +Figure 1. Given a novel object query image, our method accurately predicts the object's 6D pose using a multiview diffusion model conditioned on a set of template images with known poses. Leveraging our proposed structured 2D pose maps, represented as bundles of rays, the diffusion model recovers the query object's pose by progressively denoising these ray bundles. + +available and the system needs to function even with a single monocular RGB image. + +In object pose estimation literature, much effort has been put into learning other constraints, such as object appearance from visual data during training. Instance-based approaches [48, 53] therefore get their constraint from access to model appearance during training while category-level approaches [7, 22, 28, 54] use object shape and semantic priors. Despite the excellent results that benefit from deep learning, these approaches require training for every new object or object category from scratch and creating synthetic training data from a CAD model is also computationally expensive. To overcome per-object training, researchers have been working on unseen object pose estimation with access to textured CAD models during inference [6, 19, 26, 40, 42, 56, 60]. These advancements promise to overcome the scalability and flexibility hurdles of object-specific approaches. + +These approaches are unable to access multiple views by input design and template approaches typically solve a + +classification task first: which is the best template given an image query? Consecutive steps after template matching involve correspondence estimation, pose prediction, and optionally refinement [40, 46, 56]. Instead of finding the best possible posed template and then building pairwise correspondences, we think of the problem as an implicit bundle agreement among multiple views, using multiple template-query tuples to reason about 3D, with the advantage of having the template already posed. + +Learning to reason about 3D from multiview inputs has been extensively studied in prior work [14, 25, 31, 45]. More recently, diffusion models have emerged as powerful tools for 3D reasoning, demonstrating remarkable generalization capabilities [2, 35, 36, 51, 52, 55, 58, 59]. Among them, PoseDiffusion [55] addresses the inverse problem of structure-from-motion by directly diffusing camera poses within a probabilistic diffusion framework, modeling the conditional distribution of poses given input images. Building upon this, recent work [59] introduces an overparameterization of camera poses using Plücker coordinates [44], representing a pose as 2D maps of ray direction and ray moment. This formulation is shown to be more compatible with diffusion processes and leads to improved accuracy in relative pose estimation. These approaches exhibit strong generalization and can infer relative camera poses even in novel scenes composed of entirely unseen images. Motivated by this capability, we propose to leverage a set of posed template images and a single query image to estimate the 6D pose of an object in the query by building on the strengths of multiview diffusion-based backbones. + +Although diffusion models have shown success in relative camera pose estimation [55, 59], they are suboptimal for object pose estimation due to scale differences: camera poses are defined in a large world coordinate system, while object poses reside in a compact, object-centric space. To bridge this gap, we propose novel object-centric pose representations tailored for 6D object pose estimation. For rotation, we replace camera-centric Plücker coordinates with an object-centered formulation where rays are structured as a 2D image-aligned grid. For translation, we extend the Scale-Invariant Translation Estimation (SITE) framework [29] to generate a dense translation map. This object-centric parameterization enables more precise and disentangled reasoning about object-level 6D pose within the diffusion framework. Our structured pose diffusion framework takes a query image of an unseen object cropped from the scene and a set of posed images as templates, obtained by synthetic rendering from a CAD model, and generates precise 6D object pose predictions. We also propose a coarse-to-fine object pose estimation strategy by sampling the template with a narrower distribution based on the inputs. We evaluated our method on standard benchmark datasets from the pose estimation benchmark [50] and compared it to re + +cent methods for unseen object pose estimation. The performance of our method surpassed the results of the related works, and a detailed ablation study verified our design choices. This paper makes the following contributions: + +- we formulate unseen object pose estimation as ray bundling problem between multiview templates and RGB query, which helps the network to capture the correlation between query and templates in 3D space. +- we introduce object-centric orientation and translation over-parameterization suitable for learning within diffusion framework. +- we propose a flexible diffusion-based 6D object pose framework for unseen object pose estimation that can be extended to a coarse-to-fine prediction by using different template sampling + +# 2. Related Work + +The benchmark for object pose estimation (BOP) [16] has long been dominated by traditional handcrafted feature matching methods based on point pair features (PPF). In recent years, learning-based approaches such as GDR-Net [53], ZebraPose [48], and SurfEmb [13] have surpassed traditional methods in performance. However, these methods are instance-specific and require training on each target object. More recently, the community has placed increasing emphasis on unseen object pose estimation, which focuses on estimating the pose of novel objects not encountered during training. Below, we describe different pose estimation pipelines used in this setting. + +Model-free approaches. Without 3D model of target objects, Gen6D [34], OnePose [49], and OnePose++ [14] estimate its pose by flipping the structure from motion (SfM) at its head and matching features to align a posed object image to a test view. MFOS [27] uses posed template images as a model representation and establishes correspondences between the input query image patch and the rendered 3D bounding box of the object associated with each template image. While attractive, this leads to lower pose accuracy caused by this rough bounding box approximation of the object shape. + +Template-based approaches. OSOP [47] and OVE6D [4] utilize a template object representation for 2D segmentation and coarse to fine matching. MegaPose [26] proposes a generic render-and-compare refinement strategy. GigaPose [40] performs a template-matching approach in two stages: 1) estimates out-of-plane rotation (2 DoF) by finding discriminative synthetic templates rendered from a CAD model and then 2) establishes correspondences to estimate the four remaining 4 DoF of the object pose. + +Foundation models with CAD model prior. The idea of foundation models is a recent way to incorporate generic prior knowledge into pose estimation pipelines. Due to the need for an abundance of labeled data, all approaches are + +![](images/cd6f8cf1c5453789df50808011db7877a8c30f3bb408d3b0a707a174e011433b.jpg) +Figure 2. Pipeline overview of our method. We represent the 6D object pose using structured rotation and translation maps and employ a diffusion model to estimate the pose from random inputs. Given a query image of an unseen object and multiple template images with known object poses, our method first extracts query embeddings from a query encoder and multiview posed template embeddings from a template encoder. These embeddings serve as conditioning inputs for a diffusion transformer decoder, which is trained to denoise the object pose from random inputs. The model predicts the relative pose between the query and templates, from which the absolute 6D pose of the query object is reconstructed based on the known poses for the templates. + +![](images/18b02c0c1e1ad09b2a7dc44267abaf2558bba67f3c2411c07e7bdafec87eb8d0.jpg) + +trained on synthetic data. Several methods learn generic 3D descriptors such as Zeropose [6] and GCPose [60]. Zeropose predicts poses utilizing the foundation models of ImageBind [11] and SAM [24] together with 3D-3D feature matching. GCPose [60] uses explicit knowledge of object symmetries. FoundPose [42] combines features from the foundation model DINOv2 [41] and bag-of-words retrieval for coarse matching and then uses feature metric alignment for pose refinement. MatchU [18] and SAM6D [32] build discriminative descriptors by fusing RGB and depth information using transformers. + +Diffusion in Pose Estimation Diffusion models reconstruct a target distribution from noise over multiple time steps, inherently capturing multimodal distributions. They are, by design, capable of capturing multimodal distributions as different noisy initializations can lead to different predictions during inference in the case of a multimodal distribution. RayDiffusion [59] denoises camera poses using ray parameterization for multiview estimation, avoiding COLMAP [45] in NeRF training but is unsuitable for object pose estimation. Object pose diffusion [17] diffuses poses in SE(3) space, excelling in synthetic data but struggling with unseen objects and real datasets. PoseDiffusion [55] addresses the SfM problem by diffusing camera poses across multiple images, implicitly performing bundle adjustment. Other methods include DiffusionNOCS [20], an RGB-D approach that diffuses NOCS maps for pose estimation, and Diff9D [33], which estimates 9D pose by diffusing + +scale, translation, and rotation based on image conditioning. + +# 3. Method + +# 3.1. Method Overview + +In this paper, we represent the 6D object pose using pose maps $\mathcal{M}$ , which encode both orientation and translation. As illustrated in Fig. 2, we adopt a multiview diffusion transformer framework that learns to estimate object pose by denoising noisy pose maps conditioned on an input query object image and a set of reference images with known object poses (termed posed templates). We extract a query embedding $\mathcal{F}_Q$ and the multiview template embedding $\mathcal{F}_{MV}$ using the query and template encoders, respectively. Each encoder consists of an image encoder $\mathcal{E}_I$ that extracts 2D image features, and a view encoder $\mathcal{E}_V$ that encodes 6D object pose and/or 2D object location. Specifically, the multiview template features are fused using a Multiview Fuser to form the embedding $\mathcal{F}_{MV}$ . A Diffusion Transformer Decoder is then trained to reconstruct the clean pose maps $\mathcal{M}_0$ from noisy inputs $\mathcal{M}_t$ , conditioned on both $\mathcal{F}_Q$ and $\mathcal{F}_{MV}$ . We train our model with two different template sampling strategies to obtain both coarse and fine pose predictors. For the coarse predictor, template viewpoints are randomly sampled independently of the query pose. For the fine predictor, the same model is trained with templates sampled from a narrower distribution centered around the query pose. This strategy enables coarse-to-fine pose inference during testing without any changes to the network architecture. + +# 3.2. Object Pose Parameterization + +The 6D object pose is defined by its rotation $\mathbf{R} \in SO(3)$ and translation $t \in \mathbb{R}^3$ , representing the transformation from the object's local coordinate frame to the camera coordinate system. While compact pose regression is desirable, it remains challenging for neural networks, especially in generic or cluttered scenes. Recent work [59] overparameterizes camera poses using ray directions and ray moments based on Plücker coordinates [44], which has proven effective for scene-level camera pose estimation. However, this formulation entangles camera intrinsics, rotation, and translation, limiting its effectiveness for object-level pose tasks. Specifically, inaccuracies in the predicted direction map can propagate to the translation component, hindering the centimeter-level precision required in object pose estimation. To overcome this, we propose a novel object-centric representation that maps the 6D object pose into separate 2D rotation and translation maps, enabling more accurate and disentangled learning. + +Rotation Parameterization Camera pose estimation or novel view synthesis methods often model camera-centered rays, where rays originate from the camera center and pass through pixel coordinates in the image plane. In contrast, we introduce an object-centered ray representation, where the object center is treated as a virtual pinhole camera, emitting rays toward the camera coordinate system. Given the camera intrinsic matrix $\mathbf{K} \in \mathbb{R}^{3 \times 3}$ and extrinsic parameters—rotation $\mathbf{R} \in \mathrm{SO}(3)$ and translation $t \in \mathbb{R}^3$ , a 3D object point $\mathbf{x}$ is projected onto the image plane as $\mathbf{u} = \mathbf{K}[\mathbf{R} \mid t]\mathbf{x}$ . Instead of relying on this conventional image-based projection, we define a structured representation in which object-centered rays are mapped onto a normalized 2D square grid using a uniform intrinsic matrix, denoted as $\mathbf{K} = \mathbf{K}_I$ . The set of direction vectors originating from the object center is represented as + +$$ +\mathcal {M} _ {R} = \left\{\boldsymbol {d} _ {1}, \dots , \boldsymbol {d} _ {n} \right\} \tag {1} +$$ + +where each direction vector $\pmb{d}_i$ is normalized to unit length. This formulation enables us to map arbitrary rotation matrices $\pmb{R}$ onto a unique structured grid on the unit sphere surface. To construct the ray map, we uniformly select $\{d_i\}_{i=1}^n$ on the projected grid of the sphere surface, ensuring that each vector passes through the center of its corresponding grid cell. Consequently, we obtain a 2D grid map with the shape of $(p \times p \times 3)$ as our rotation representation in the diffusion process. The illustration is given in the supplementary. Given the object-centered ray representation, we recover the rotation matrix $\pmb{R}$ by aligning the predicted ray directions with a predefined canonical frame. Let $\mathcal{M}_R = \{d_1, \dots, d_n\}$ be the predicted ray set and $\mathcal{M}_R^* = \{d_1^*, \dots, d_n^*\}$ the reference rays corresponding to an identity rotation $\pmb{R} = \pmb{I}$ . The optimal rotation matrix $\pmb{R}^*$ is obtained by solving: + +![](images/5162b602dc63760031842c4f1447246eebe3e9491ded2cc4154340b004c24171.jpg) +Object centric ray bundles + +![](images/47f831937078c13a41d20fc10089a1a714cd08941f0fecbef36fc9586ac44986.jpg) +2D Rotation map +Figure 3. Visual illustration of the object-centric ray representation used for rotation prediction in our diffusion model. The rotation map $\mathcal{M}_R$ is defined as a bundle of rays originating from the object center $\mathcal{O}_{obj}$ , encoded as a 3-channel 2D map. + +$$ +\boldsymbol {R} ^ {*} = \underset {\boldsymbol {R} \in \mathrm {S O} (3)} {\arg \min } \sum_ {i = 1} ^ {n} \left\| \boldsymbol {R} \boldsymbol {d} _ {i} ^ {*} - \boldsymbol {d} _ {i} \right\| ^ {2} \tag {2} +$$ + +where $\pmb{R}$ is the relative rotation of the object with respect to the canonical frame. This problem can be solved using the Singular Value Decomposition (SVD) differentially, ensuring a valid rotation by enforcing $\pmb{R}^T\pmb{R} = \pmb{I}$ . This formulation allows for robust recovery of the object's orientation and enables the diffusion process on 3D rotations from a structured 2D ray representation. Translation Parameterization A major challenge in estimating an object's 6D pose from a single RGB image is minimizing translation error, particularly for previously unseen objects and scenes. Earlier work, SSD6D [23], estimates translation by locating the object centroid in 2D coordinates and comparing the bounding box scale with a pre-rendered template of the same rotation to determine object distance. However, this approach assumes the object center aligns with the bounding box center, making it sensitive to occlusion. Instance-level regression-based methods [29, 53] improve robustness by employing Scale-Invariant Translation Estimation (SITE), which predicts translation by computing the offset between the bounding box center and the object center. More recently, generalizable RGB-based methods [40, 42] estimate translation by establishing 2D correspondences between query and template images using a pre-trained feature matcher. While template depth can be rendered, these methods rely solely on one RGB image pair for correspondence extraction. In this paper, we extend SITE to a patch-level dense translation map. Given the object translation $\pmb{t} = [t_x,t_y,t_z]$ and the camera intrinsic matrix $\pmb{K}$ , the projected object centroid $[o_x,o_y,1]^T$ in image coordinates is computed as: + +$$ +\left[ \boldsymbol {o} _ {x}, \boldsymbol {o} _ {y}, 1 \right] ^ {T} = \boldsymbol {K} t. \tag {3} +$$ + +We estimate the offset from each pixel $(u, v)$ in the detected bounding box to the object centroid $(\boldsymbol{o}_x, \boldsymbol{o}_y)$ , forming a dense normalized translation offset map: + +$$ +\mathcal {M} _ {T} = \left(\frac {u - \boldsymbol {o} _ {x}}{w}, \frac {v - \boldsymbol {o} _ {y}}{h}, \frac {\boldsymbol {t} _ {z}}{r _ {z}}\right), \tag {4} +$$ + +where $w$ and $h$ denote the bounding box width and height, and $r_z$ is the zoom-in ratio of the bounding box. Similar + +to the rotation map, we uniformly sample the pixels in the bounding box with the same shape of $(p\times p\times 3)$ as the 2D translation map in the diffusion process. The 3D object translation is then recovered by back-projecting the estimated centroid offset using the camera intrinsics: + +$$ +\boldsymbol {t} ^ {*} = r _ {z} \cdot \boldsymbol {K} ^ {- 1} \left[ w \cdot \Delta \boldsymbol {o} _ {x} + \boldsymbol {o} _ {x}, h \cdot \Delta \boldsymbol {o} _ {y} + \boldsymbol {o} _ {y}, \Delta \boldsymbol {o} _ {z} \right] ^ {T}. \tag {5} +$$ + +To this end, we represent object pose as 2D pose maps $\mathcal{M} = (\mathcal{M}_R,\mathcal{M}_T)$ . This pose representation decouples rotation and translation as well as the camera intrinsics, enabling the model to predict rotation and translation independently and enabling the use of a diffusion model to denoise the pose on two dense 2D maps. + +# 3.3. Multiview Template Conditioned Diffusion + +In our framework, we employ a multiview diffusion model to estimate object pose by conditioning it with the input query image and the posed templates. This network formulates the learning as a denoising process that gradually refines noisy inputs into the proposed structured pose maps. + +# 3.3.1. Diffusion Preliminaries + +Diffusion process. The diffusion process consists of a forward (noising) and a reverse (denoising) process. Given a clean pose representation $\mathcal{M}_0$ (either the rotation map $\mathcal{M}_R$ or the translation map $\mathcal{M}_T$ ), the forward process adds Gaussian noise over a fixed number of timesteps $T$ . At each timestep $t \in \{1, \dots, T\}$ , the pose map is perturbed as: + +$$ +\mathcal {M} _ {t} = \sqrt {\alpha_ {t}} \mathcal {M} _ {0} + \sqrt {1 - \alpha_ {t}} \epsilon , \quad \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I}), \tag {6} +$$ + +where $\alpha_{t}$ is a noise schedule controlling the variance at timestep $t$ . + +Denoising process. The reverse process aims to recover the clean pose representation by learning to predict and remove the noise. A neural network $\epsilon_{\theta}(\mathcal{M}_t,t,\mathcal{F}_C)$ is trained to estimate the noise $\epsilon$ conditioned on an embedding $\mathcal{F}_c$ that encodes the query and template information. The predicted pose is obtained by iteratively refining $\mathcal{M}_t$ using the learned noise estimator: + +$$ +\begin{array}{l} \mathcal {M} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathcal {M} _ {t} - \sqrt {1 - \alpha_ {t}} \epsilon_ {\theta} \left(\mathcal {M} _ {t}, t, \mathcal {F} _ {c}\right)\right) \tag {7} \\ + \sigma_ {t} z, \quad z \sim \mathcal {N} (\mathbf {0}, \mathbf {I}). \\ \end{array} +$$ + +where $\sigma_{t}$ controls the stochasticity of the denoising step. This iterative process gradually refines the noisy pose representation into a structured output. + +# 3.3.2. Network Architecture + +Our diffusion-based framework for 6D object pose estimation consists of three main blocks: (1) the Query Encoder, which extract the query image features, (2) the Template Encoder, which encodes and fuse the multiple posed template information, and (3) the Diffusion Transformer Decoder, which attends the query and templates and predicts + +the denoised rotation and translation maps. The overall architecture is illustrated in Figure 2. + +Template Encoder. The template encoder consists of three components: the Image Encoder $\mathcal{E}_I$ , the View Encoder $\mathcal{E}_V$ , and the Multiview Fuser $\mathcal{E}_F$ . Given $N$ object templates rendered from different viewpoints, we first employ a frozen DINOv2 [41] backbone to extract the image feature maps $\mathcal{F}_T$ . Inspired by prior works [21, 36, 37, 61], which have demonstrated the effectiveness of Fourier encoding for camera rays in multiview scene understanding and reconstruction, we extend this idea to embed our structured pose information. + +View Encoder. In our framework, the View Encoder uses three Fourier encoders to process the structured rotation map, translation map, and the normalized bounding box coordinates in the 2D image. This view embedding $\mathcal{F}_v$ explicitly locates the objects by incorporating both their viewpoint in 3D space and their scale in 2D image coordinates, enforcing the network to learn implicit relationships across different views. The Fourier feature embedding for a scalar input $x$ is computed as $\gamma (x) = (x,\sin (2\pi Bx),\cos (2\pi Bx))$ where $B$ is a fixed frequency band that controls the resolution of the encoded features. Given the bounds $K_{r}$ , $K_{t}$ , and $K_{c}$ for rotation, translation, and object 2D coordinate maps, the total dimensionality of the view embedding is $D_V = (2(K_r + K_t + K_c) + 1)d$ , where $d$ is the number of frequency bands used for each Fourier encoding. This formulation ensures that each structured feature is transformed into a high-dimensional space, improving the model's ability to capture fine-grained variations in object pose. + +Multiview Fuser. The view embeddings $\mathcal{F}_v$ and the image embeddings $\mathcal{F}_I$ are concatenated and conditioned with view-level and patch-level positional encodings. This combined representation is then passed to the Multiview Fuser, which consists of $\mathcal{N}_F$ self-attention layers to extract the fused multiview template embedding $\mathcal{F}_{MV}$ . Following [59], we use DiT [43] blocks for cross-view information exchange and ensure that the network effectively aggregates object appearance and viewpoint-dependent geometric cues across all template images. The multiview embedding encodes the 2D and 3D priors of the object implicitly, which will be used as conditioning information to help reason the pose of the query image. + +Query Encoder. The Query Encoder shares the same image and view encoders as the Template Encoder but processes only a single-view input. Since the pose of the query is not known, only the object's 2D location in image coordinates is conditioned, instead of the full pose. Patch-level positional encoding is also incorporated to capture the spatial position of patches in the query image, facilitating implicit fusion with template patches. The resulting query embedding is denoted as $\mathcal{F}_Q$ . + +Diffusion Transformer Decoder. The fused multiview + +template embedding $\mathcal{F}_{MV}$ and the query embedding $\mathcal{F}_Q$ serve as conditioning information in our diffusion transformer decoder, which denoises the 2D rotation and translation maps over a series of timesteps. At each step $t$ , the noisy pose maps $\mathcal{M}_t$ are processed alongside their conditional embeddings through a sequence of $\mathcal{N}_D$ transformer-based diffusion blocks based on DiT [43]. Unlike DiT, which primarily uses only self-attention layers, we follow [30] to also incorporate cross-attention layers, allowing the query embedding to attend to the fused template embedding, enabling the feature exchange between different latent spaces that are constructed from a single query and multiple posed template views respectively. During training, for each time step $t$ , the decoder learns to predict the noise component $\epsilon_{\theta}(\mathcal{M}_t,t,\mathcal{F}_C)$ , where $\mathcal{F}_C = (\mathcal{F}_Q|\mathcal{F}_{MV})$ . During inference, the pose maps are randomly initialized and iteratively denoised to obtain the final 2D rotation and translation maps. To enhance generalization, instead of directly predicting the absolute pose of the query, we estimate the relative pose between the query and templates. Specifically, the rotation maps represent the relative rotation from the template to the query, while the translation maps encode a relative depth scale between them: + +$$ +\boldsymbol {t} _ {z} ^ {r e l} = \frac {\boldsymbol {t} _ {z} ^ {Q} r _ {z} ^ {T}}{\boldsymbol {t} _ {z} ^ {T} r _ {z} ^ {Q}} \tag {8} +$$ + +where $t_{z}^{rel}$ is the predicted relative depth scale, $t_{z}^{Q}$ and $t_{z}^{T}$ are the ground-truth depths of the object center for the query and templates respectively, and $r_{z}^{Q}$ and $r_{z}^{T}$ are the zoom-in scales of the query and template. Finally, the query pose $H_{Q} = [R_{Q}|t_{Q}]$ is recovered from the predicted relative pose $H_{rel}$ and the ground-truth pose $H_{gt}$ of the templates: + +$$ +\boldsymbol {H} _ {Q} = \boldsymbol {H} _ {r e l} \boldsymbol {H} _ {g t}. \tag {9} +$$ + +Our multiview DiT-based Diffusion Transformer Decoder estimates the relative pose between a query and each template by conditioning on implicit geometric priors encoded in the templates and the query features. Each template view produces an independent pose hypothesis, supervised by its specific relative pose to the query. This design enables efficient, batch-wise probabilistic sampling from randomly selected templates, yielding diverse pose hypotheses in parallel. + +# 3.4. Loss Functions + +In our diffusion model, instead of training the network to estimate the noise, we follow prior work [59] and train the denoising network $\epsilon_{\theta}(\mathcal{M}_t,t,\mathcal{F}_c)$ to learn the reverse diffusion process by predicting the original clean pose map $\mathcal{M}_0$ conditioned on the noisy input $\mathcal{M}_t$ at timestep $t$ . The loss function is defined as: + +$$ +\mathcal {L} _ {\text {d i f f}} = \mathbb {E} _ {t, \epsilon} \left[ \| \mathcal {M} _ {0} - \epsilon_ {\theta} \left(\mathcal {M} _ {t}, t, \mathcal {F} _ {c}\right) \| _ {2} ^ {2} \right], \tag {10} +$$ + +where $t$ is uniformly sampled from $[1, T]$ during training, and $\mathcal{M}_t$ is the noisy version of the original pose map $\mathcal{M}_0$ corrupted with Gaussian noise at timestep $t$ . This training strategy naturally integrates task-specific constraints on the original rotation and translation maps. + +Rotation losses. For the rotation map, we apply a pixel-level reconstruction loss to the target map $\mathcal{M}_R^*$ : + +$$ +\mathcal {L} _ {\text {r e c o n}} ^ {R} = \frac {1}{p ^ {2}} \| \mathcal {M} _ {R} - \mathcal {M} _ {R} ^ {*} \| _ {2} ^ {2}, \tag {11} +$$ + +where $p$ denotes the spatial resolution of the rotation ray maps. Since each element in the rotation map represents a directional vector, we also employ a cosine similarity loss $\mathcal{L}_{\cos}^{R}$ to supervise ray directions. Given that the rotation map consists of structured ray bundles, we introduce an angle-consistency loss to enforce geometric coherence across adjacent rays in the predicted ray map. Given the predicted ray directions $\{\pmb {d}_i\}_{i = 1}^n$ and the canonical ray set $\{\pmb {d}_i^*\}_{i = 1}^n$ we ensure that the predicted rays maintain consistent relative angles that reflect the intrinsic ray map geometry. For each pair of adjacent rays indexed by $(i,j)$ , the relative angle is computed as $\alpha_{ij} = \arccos (d_i^\top d_j)$ . Similarly, we precompute the reference angles $\alpha_{ij}^{*}$ from the canonical rays corresponding to an identity rotation. The ray-consistency loss is then defined as: + +$$ +\mathcal {L} _ {\text {r e g}} ^ {R} = \frac {1}{| \mathcal {N} _ {r} |} \sum_ {(i, j) \in \mathcal {N} _ {r}} \left(\alpha_ {i j} - \alpha_ {i j} ^ {*}\right) ^ {2}, \tag {12} +$$ + +where $\mathcal{N}_r$ is the set of adjacent ray index pairs, and $|\mathcal{N}_r|$ is its cardinality. This loss term encourages the network to respect intrinsic geometric constraints imposed by the projection, ensuring stable rotation estimation. The overall rotation loss is then formulated as a weighted combination of the reconstruction loss, cosine similarity loss, and ray-consistency loss: + +$$ +\mathcal {L} ^ {R} = \lambda_ {\text {r e c o n}} \mathcal {L} _ {\text {r e c o n}} ^ {R} + \lambda_ {\cos} \mathcal {L} _ {\cos} ^ {R} + \lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}} ^ {R}, \tag {13} +$$ + +where $\lambda_{\mathrm{recon}}, \lambda_{\mathrm{cos}}$ , and $\lambda_{\mathrm{reg}}$ are hyperparameters that balance the contributions of each term. This formulation ensures that the predicted rotation map is both accurate at the pixel level and geometrically consistent with the camera's intrinsic structure. + +Translation losses. For the translation map, similar to the rotation ray reconstruction, we apply a pixel-level L2 loss to ensure accurate reconstruction of the dense translation offset maps: + +$$ +\mathcal {L} _ {\text {r e c o n}} ^ {T} = \frac {1}{p ^ {2}} \| \mathcal {M} _ {T} - \mathcal {M} _ {T} ^ {*} \| _ {2} ^ {2}. \tag {14} +$$ + +Additionally, to explicitly supervise the final 3D translation prediction, we compute the object translation $\pmb{t}$ from the predicted translation map using the decoding formulation described in Eq. 5. We then impose an L1 loss on the + +predicted translation components along each axis: + +$$ +\mathcal {L} _ {\mathrm {x y z}} ^ {T} = \lambda_ {x} \left| \boldsymbol {t} _ {x} - \boldsymbol {t} _ {x} ^ {*} \right| + \lambda_ {y} \left| \boldsymbol {t} _ {y} - \boldsymbol {t} _ {y} ^ {*} \right| + \lambda_ {z} \left| \boldsymbol {t} _ {z} - \boldsymbol {t} _ {z} ^ {*} \right|, \tag {15} +$$ + +where $\lambda_x, \lambda_y, \lambda_z$ are weighting the supervision strength. The overall translation loss is defined as the weighted sum of these two terms: + +$$ +\mathcal {L} ^ {T} = \lambda_ {\text {r e c o n}} \mathcal {L} _ {\text {r e c o n}} ^ {T} + \lambda_ {t} \mathcal {L} _ {\text {x y z}}, \tag {16} +$$ + +where $\lambda_{\mathrm{recon}}$ and $\lambda_{t}$ are the hyperparameters to balance dense map-level supervision and explicit object-level translation regression. + +Overall loss function. The final training loss is defined as a weighted sum of the rotation and translation losses: + +$$ +\mathcal {L} = \lambda_ {\text {r o t}} \mathcal {L} ^ {R} + \lambda_ {\text {t r a n s}} \mathcal {L} ^ {T}, \tag {17} +$$ + +where $\lambda_{\mathrm{rot}}$ and $\lambda_{\mathrm{trans}}$ are hyperparameters controlling the relative importance of rotation and translation losses. + +# 3.5. Coarse to Fine Predictor Training Strategy. + +We train our network with different template distributions to obtain both a coarse and fine predictor. For the coarse predictor, we use eight templates with randomly sampled poses from pre-processed scene-cropped images, ensuring diverse viewpoint coverage. The fine predictor employs online template sampling, augmenting templates for $\pm 30^{\circ}$ , $\pm 5\mathrm{cm}$ based on the query pose to enforce closer template-query alignment. Both predictors share the same network architecture but differ in template distributions during training. During inference, the coarse predictor provides an initial pose estimate, which is then refined by the fine predictor using a more localized template distribution. + +# 4. Experiments + +# 4.1. Experimental Setup + +Evaluation Metrics. We adopt the metric Average Recall (AR) proposed by the Benchmark of Pose Estimation (BOP) [50]. The AR score is calculated with 3 pose-error functions: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). A pose is considered correct if the pose errors are within a predefined error threshold. The mean recall on the each error functions is computed over multiple error thresholds. The overall accuracy of a method is given by the Average Recall $AR = (AR_{\mathrm{VSD}} + AR_{\mathrm{MSSD}} + AR_{\mathrm{MSPD}}) / 3$ . + +Training and evaluation datasets. We train our model on realistic synthetic datasets generated by Megapose [26], comprising approximately 2 million images rendered with BlenderProc [8] using objects from Google Scanned Objects [10] and ShapeNet [5]. For novel object pose estimation, we evaluate our method on five benchmark datasets: + +LM-O [3], T-LESS [15], YCB-V [57], TUD-L [16], and IC-BIN [9]. Our evaluation is structured as follows: in Section 4.2, we compare our method with baselines on novel object pose estimation; in Section 4.3, we conduct ablation studies where we analyze design components, where we train and evaluate our method on LM-O dataset. + +# 4.2. Compare to Baselines + +We evaluate our method on five benchmark datasets: LM-O, T-LESS, TUD-L, IC-BIN, and YCB-V, which are unseen during training, and compare it with recent state-of-the-art methods that use only RGB images as input. All the methods use the same detection and segmentation results generated from CNOS [39] by default, except for OSOP [46]. As shown in Table 1, we analyze different setups, considering whether refinement and multi-hypothesis predictions are used. Our method achieves the highest average AR across all settings. In the single-prediction setting without refinement, it improves over the previous best method by $3.4\%$ on average, with notable gains of $6.3\%$ on LM-O and $9.2\%$ on T-LESS. With refinement, our method continues to outperform the baselines, particularly excelling on TUD-L and LM-O. In the multi-hypothesis setting, it achieves the best performance on most datasets, especially with T-LESS dataset being improved by $3.7\%$ . These results highlight the effectiveness of our approach in enhancing pose estimation by leveraging robust pose representations and a diffusion-based pipeline while ensuring strong generalization across diverse datasets. + +# 4.3. Ablation Study + +We conduct an ablation study on four key components of our approach: template ground-truth (GT) view embedding, the multiview setup, the fine-level predictor, and relative pose prediction. Each component is either removed or replaced with an alternative setup, and the results are summarized in Table 2. + +Fine predictor. In the refinement stage, we apply both our fine predictor and MegaPose refinement. As shown in (1) - (3) of Table 2, our fine predictor improves performance by $6.3\%$ compared to the coarse prediction. Notably, the fine predictor does not modify the network itself but instead utilizes a different template sampling strategy. Further improvements are achieved when incorporating an external refiner during the refinement stage. + +Relative pose prediction. To enhance generalization across different scenes, camera intrinsics, and viewing conditions, we predict the relative pose between posed templates and the query, using the ground-truth template pose to infer the absolute query pose. In this ablation, we modify the network to directly predict the absolute poses of all template and query frames using a sequence of DiT blocks. As shown in (4) of Table 2, this modification results in a per + +
MethodRefinementMulti-hypoLM-OT-LESSTUD-LIC-BINYCB-VAverage
OSOP [47]XX31.2---33.232.2
ZS6D [1]XX29.821.0--32.427.7
MegaPose [26]XX22.917.725.815.228.121.9
GenFlow [38]XX25.021.530.016.827.724.2
GigaPose [40]XX29.927.330.223.129.027.9
FoundPose [56]XX39.633.846.723.945.237.8
OursXX42.136.948.321.846.239.1
MegaPose [26]✓X49.947.765.336.760.151.9
GigaPose [40]✓X55.654.657.844.363.455.1
FoundPose [56]✓X55.751.063.343.366.155.9
Ours✓X56.253.866.541.662.856.2
GenFlow [38]✓✓56.352.368.445.363.357.1
MegaPose [26]✓✓56.050.768.441.462.155.7
GigaPose [40]✓✓59.957.064.546.766.358.9
FoundPose [56]✓✓61.057.069.447.969.060.9
Ours✓✓62.259.170.245.568.961.2
+ +Table 1. Comparison of our method with the baselines with only RGB image inputs. + +
MethodGT Temp. PoseMultiviewRelative PoseFine PredictorAR
(1) Fine+MegaPose✓✓✓✓56.2
(2) Fine✓✓✓✓42.1
(3) Coarse✓✓✓✗39.6
(4) Absolute pose✓✓✗✗35.4
(5) Single view✓✗✗✗32.5
(6) w/o GT Template Pose✗✓✓✗27.8
+ +Table 2. Ablation study for the key component of our method. + +
MethodTemplate DistributionAR
CoarseRandom65.58
Coarsefixed60.28
Fine±90°, ±10cm65.81
Ours-Fine±30°, ±5cm67.29
Fine±15°, ±3cm61.04
+ +Table 3. Comparison of different template selection strategies for the fine predictor on LM-O dataset. The bold is our default setup. + +formance drop of up to $10\%$ , highlighting the effectiveness of relative pose prediction. + +Multi-view prediction. We modify our pose map prediction head to perform single-view prediction, meaning the query pose is estimated directly without leveraging multiple templates. As shown in (5) of Table 2, this change leads to an approximately $18\%$ performance drop, confirming that multi-view prediction enables the model to learn stronger implicit correspondences between the query and templates. + +Ground-truth template pose. In our setup, we explicitly input template ground-truth pose maps as conditional information to help the network learn inherent correlations with the input query image. To evaluate its impact, we remove the ground-truth pose map from the input while retaining only the 2D position information, which is essential for predicting the relative distance from the camera. The ground-truth template pose is only used during inference to recover the absolute query pose. As shown in (6) of Table 2, removing the template ground-truth pose map results in a significant performance drop compared to the single-view setting. This finding underscores the importance of leveraging template pose priors in multi-view prediction, and also indicates the effectiveness of the view encoders. + +# 4.4. Effects of pose distribution of templates. + +We evaluate our method's sensitivity to template distribution by training with different template sampling strategies on the LM-O dataset, as shown in Table 3. For the coarse predictor, using randomly sampled templates yields better performance than the fixed-template setting. For the fine predictor, we examine different template distributions with varying pose and translation constraints. The best performance is achieved with a template distribution constrained to $\pm 30^{\circ}$ in rotation and $\pm 5\mathrm{cm}$ in translation, which aligns with the mean error of the coarse predictor. A wider distribution $(\pm 90^{\circ}, \pm 10\mathrm{cm})$ performs similarly to the randomly sampled distribution used in the coarse predictor, while overly narrow constraints $(\pm 15^{\circ}, \pm 3\mathrm{cm})$ lead to a slight performance drop. These results underscore the importance of selecting appropriate template distributions for both coarse and fine predictors to balance generalization and fine-level accuracy. + +# 5. Conclusion + +In this paper, we introduced a structured representation for object pose that enables effective deployment of diffusion models for object 6D pose estimation. Instead of pairwise matching, we propose aligning object-centered rays across multiple posed templates. Our multiview diffusion model is conditioned on embeddings extracted from both the query and multiple posed template images using dedicated encoders. A coarse-to-fine strategy further improves pose accuracy without altering the model architecture. By integrating multiview diffusion models into the 6D pose estimation pipeline, our approach enables probabilistic reasoning from multiview inputs instead of explicitly reconstructing 3D. While our model achieves competitive performance, it relies on posed templates and accurate object detections. Future work would be of interest in relaxing these constraints to further generalize multiview-based pose estimation. + +# References + +[1] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 463-469. IEEE, 2024. 8 +[2] Emmanuelle Bourigault and Pauline Bourigault. Mvdiff: Scalable and flexible multi-view diffusion for 3d object reconstruction from single-view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 7579-7586, 2024. 2 +[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In ECCV, 2014. 7 +[4] Dingding Cai, Janne Heikkilä, and ESA Rahtu. Ove6d: Object viewpoint encoding for depth-based 6d object pose estimation. In CVPR, 2022. 2 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 7 +[6] Jianqiu Chen, Mingshan Sun, Tianpeng Bao, Rui Zhao, Liwei Wu, and Zhenyu He. Zeropose: Cad-model-based zero-shot pose estimation, 2023. 1, 3 +[7] Yamei Chen, Yan Di, Guangyao Zhai, Fabian Manhardt, Chenyangguang Zhang, Ruida Zhang, Federico Tombari, Nassir Navab, and Benjamin Busam. Secondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9959-9969, 2024. 1 +[8] Maximilian Denninger, Martin Sundermeyer, Dominik Winkelbauer, Youssef Zidan, Dmitry Olefir, Mohamad Elbadrawy, Ahsan Lodhi, and Harinandan Katam. Blenderproc. arXiv preprint arXiv:1911.01911, 2019. 7 +[9] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In CVPR, 2016. 7 +[10] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, 2022. 7 +[11] 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 CVPR, 2023. 3 +[12] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1 +[13] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In CVPR, 2022. 2 + +[14] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without CAD models. In NeurIPS, 2022. 2 +[15] Tomáš Hodař, Pavel Haluza, Štepan Obdržálek, Jiří Matas, Manolis Lourakis, and Xenophon Zabulis. T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. WACV, 2017. 7 +[16] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, Caner Sahin, Fabian Manhardt, Federico Tombari, Tae-Kyun Kim, Jiri Matas, and Carsten Rother. Bop: Benchmark for 6d object pose estimation. In ECCV, 2018. 2, 7 +[17] Tsu-Ching Hsiao, Hao-Wei Chen, Hsuan-Kung Yang, and Chun-Yi Lee. Confronting ambiguity in 6d object pose estimation via score-based diffusion on se(3). In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3456–3465, 2023. 3 +[18] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10095-10105, 2024. 3 +[19] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1 +[20] Takuya Ikeda, Sergey Zakharov, Tianyi Ko, Muhammad Zubair Irshad, Robert Lee, Katherine Liu, Rares Ambrus, and Koichi Nishiwaki. Diffusionnocs: Managing symmetry and uncertainty in sim2real multi-modal category-level pose estimation. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7406-7413, 2024. 3 +[21] Andrew Jaegle, Felix Gimeno, Andrew Brock, Andrew Zisserman, Oriol Vinyals, and João Carreira. Perceiver io: A general architecture for structured inputs & outputs. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5 +[22] HyunJun Jung, Shun-Cheng Wu, Patrick Ruhkamp, Hannah Schieber, Pengyuan Wang, Giulia Rizzoli, Hongcheng Zhao, Sven Damian Meier, Daniel Roth, Nassir Navab, et al. Housecat6d-a large-scale multi-modal category level 6d object pose dataset with household objects in realistic scenarios. arXiv preprint arXiv:2212.10428, 2022. 1 +[23] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. SSD-6D: Making RGB-based 3d detection and 6d pose estimation great again. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1521–1529, 2017. 4 +[24] 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. 3 +[25] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Cosypose: Consistent multi-view multi-object 6d pose + +estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16, pages 574-591. Springer, 2020. 1, 2 +[26] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapore: 6d pose estimation of novel objects via render & compare. In CoRL, 2022. 1, 2, 7, 8 +[27] JongMin Lee, Yohann Cabon, Romain Brégier, Sungjoo Yoo, and Jerome Revaud. Mfos: Model-free & one-shot object pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2911-2919, 2024. 2 +[28] Weihang Li, Hongli Xu, Junwen Huang, Hyunjun Jung, Peter KT Yu, Nassir Navab, and Benjamin Busam. Gce-posede: Global context enhancement for category-level object pose estimation. arXiv preprint arXiv:2502.04293, 2025. 1 +[29] Yisheng Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. CDPN: Coordinates-based disentangled pose network for real-time RGB six-degrees-of-freedom object pose estimation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 7678-7687, 2019, 2, 4 +[30] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding. arXiv preprint arXiv:2405.08748, 2024. 6 +[31] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 2024 International Conference on 3D Vision (3DV), 2024. 2 +[32] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. arXiv preprint arXiv:2311.15707, 2023. 3 +[33] Jian Liu, Wei Sun, Hui Yang, Pengchao Deng, Chongpei Liu, Nicu Sebe, Hossein Rahmani, and Ajmal Mian. Diff9d: Diffusion-based domain-generalized category-level 9-dof object pose estimation. arXiv preprint arXiv:2502.02525, 2025.3 +[34] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 2 +[35] Yuanxun Lu, Jingyang Zhang, Tian Fang, Jean-Daniel Nahmias, Yanghai Tsin, Long Quan, Xun Cao, Yao Yao, and Shiwei Li. Matrix3d: Large photogrammetry model all-in-one. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 11250-11263, 2025. 2 +[36] Zhengzhe Luo, Yan Huang, Liang Wang, and Tieniu Tan. Zero-shot novel view and depth synthesis with multi-view geometric diffusion. arXiv preprint arXiv:2303.17598, 2023. 2, 5 +[37] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn + +thesis. Communications of the ACM, 65(1):99-106, 2021. 5 +[38] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. Genflow: Generalizable recurrent flow for 6d pose refinement of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10039-10049, 2024. 8 +[39] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. In ICCV, 2023. 7 +[40] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. GigaPose: Fast and robust novel object pose estimation via one correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 4, 8 +[41] Maxime Oquab, Timothee Darcet, Theo 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. 3, 5 +[42] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomáš Hodan. Found-Pose: Unseen object pose estimation with foundation features. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 3, 4 +[43] 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. 5, 6 +[44] Julius Plücker. Analytisch-geometrische Entwicklungen. GD Baedeker, 1828. 2, 4 +[45] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3 +[46] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In CVPR, 2022. 2, 7 +[47] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6835-6844, 2022. 2, 8 +[48] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6dof object pose estimation. In CVPR, 2022. 1, 2 +[49] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In CVPR, 2022. 2 +[50] Martin Sundermeyer, Tomás Hodań, Yann Labbe, Gu Wang, Eric Brachmann, Bertram Drost, Carsten Rother, and Jíří Matas. Bop challenge 2022 on detection, segmentation and pose estimation of specific rigid objects. In CVPR, 2023. 2, 7 +[51] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multi + +view image generation with correspondence-aware diffusion. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2 +[52] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. In European Conference on Computer Vision (ECCV), 2024. 2 +[53] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. GDR-Net: Geometry-guided direct regression network for monocular 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16611-16621, 2021. 1, 2, 4 +[54] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1 +[55] Jianyuan Wang, Christian Rupprecht, and David Novotny. PoseDiffusion: Solving pose estimation via diffusion-aided bundle adjustment. 2023. 2, 3 +[56] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17868-17879, 2024. 1, 2, 8 +[57] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. Robotics: Science and Systems, 2018. 7 +[58] Fuyang Zhang, Shitao Tang, Jiacheng Chen, and Yasutaka Furukawa. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2302.01329, 2023. 2 +[59] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 4, 5, 6 +[60] Heng Zhao, Shenxing Wei, Dahu Shi, Wenming Tan, Zheyang Li, Ye Ren, Xing Wei, Yi Yang, and Shiliang Pu. Learning symmetry-aware geometry correspondences for 6d object pose estimation. In ICCV, 2023. 1, 3 +[61] Wenhao Zhao, Yuchao Dai, Hongdong Li, and Zhibo Rao. Towards zero-shot scale-aware monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 5 \ No newline at end of file diff --git a/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/images.zip b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..411b4d28c9efb0579641669a681cf117e932072a --- /dev/null +++ b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d3ee28c0bc4bb1d5bc4a57c6c95a841826a4610e0173704ff375f456d9ee961 +size 303264 diff --git a/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/layout.json b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dc4e8ac64b40beb547c931d07859a7bda2135a93 --- /dev/null +++ b/ICCV/2025/RayPose_ Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4012558933d06ea7df099df39df5c3356b6e58d2a614d953c818d75aafd2d9fa +size 441324 diff --git a/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/6c54b274-5971-48c3-9462-c7f7ddb43594_content_list.json b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/6c54b274-5971-48c3-9462-c7f7ddb43594_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..37c35e481a1f0973caa827a55c82c034d3e38653 --- /dev/null +++ b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/6c54b274-5971-48c3-9462-c7f7ddb43594_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0aafde6ba351cec9e1276b0bf56fe5194f1c8d868290b86776ca38834d25b69 +size 97142 diff --git a/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/6c54b274-5971-48c3-9462-c7f7ddb43594_model.json b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/6c54b274-5971-48c3-9462-c7f7ddb43594_model.json new file mode 100644 index 0000000000000000000000000000000000000000..21ce6a39ca4b96b1c1cce4eb5dafd2ca54fcbffa --- /dev/null +++ b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/6c54b274-5971-48c3-9462-c7f7ddb43594_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27e86aa7b319c1128bebb75edae11b5889c9b6a6a7ac48f382feee063dab6fb4 +size 123109 diff --git a/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/6c54b274-5971-48c3-9462-c7f7ddb43594_origin.pdf b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/6c54b274-5971-48c3-9462-c7f7ddb43594_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a5f36fc2e510b6060a53b437696068aa86b7d40 --- /dev/null +++ b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/6c54b274-5971-48c3-9462-c7f7ddb43594_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddba5b8b7420424a5f8753960137886957afa41ff250c416a9ed2029c44c6593 +size 2519025 diff --git a/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/full.md b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..26b28be1c20c3230d842f27c352e8afd86d578a8 --- /dev/null +++ b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/full.md @@ -0,0 +1,388 @@ +# RayZer: A Self-supervised Large View Synthesis Model + +Hanwen Jiang $^{1}$ Hao Tan $^{2}$ Peng Wang $^{2}$ Haian Jin $^{3}$ Yue Zhao $^{1}$ Sai Bi $^{2}$ +Kai Zhang $^{2}$ Fujun Luan $^{2}$ Kalyan Sunkavalli $^{2}$ Qixing Huang $^{1}$ Georgios Pavlakos $^{1}$ $^{1}$ The University of Texas at Austin $^{2}$ Adobe Research $^{3}$ Cornell University + +![](images/b2266fe67978046ec63c1987c3af98e94a81ac1787ce8b829750eaffcb9813b6.jpg) + +![](images/1ff9909f06335fe804ddd9375dbe4da0e429a229fe81090fcbc8a80fce846b7d.jpg) +Figure 1. We propose RayZer, a self-supervised multi-view 3D Vision model trained on unlabeled data without any annotations, e.g., camera pose labels. At inference, RayZer supports feed-forward novel view synthesis from unposed & uncalibrated images. RayZer achieves novel view synthesis performance comparable to that of supervised "oracle" methods (GS-LRM and LVSM), which require camera labels in both training and inference, and even outperforms them when they rely on (potentially noisy) COLMAP camera annotations. We show two examples on the right, where COLMAP camera annotations lead to consistent failures of GS-LRM and LVSM during inference. + +![](images/51f19f53d41014698910990965f6130238d9ecebf0192d74ec58e6de45a018c6.jpg) + +![](images/d6bfd205d38fc154ddf50be4791cb1d5a722cf0d0d895f994a43681ab115935b.jpg) + +# Abstract + +We present RayZer, a self-supervised multi-view 3D Vision model trained without any 3D supervision, i.e., camera poses and scene geometry, while exhibiting emerging 3D awareness. Concretely, RayZer takes unposed and uncalibrated images as input, recovers camera parameters, reconstructs a scene representation, and synthesizes novel views. During training, RayZer relies solely on its self-predicted camera poses to render target views, eliminating the need for any ground-truth camera annotations and allowing RayZer to be trained with 2D image supervision. The emerging 3D awareness of RayZer is attributed to two key factors. First, we design a self-supervised framework, which achieves 3D-aware auto-encoding of input images by disentangling camera and scene representations. Second, we design a transformer-based model in which the only 3D prior is the ray structure, connecting camera, pixel, and scene simultaneously. RayZer demonstrates comparable or even superior novel view synthesis performance than "oracle" methods that rely on pose annotations in both training and testing. Project: https://hwjiang1510.github.io/RayZer/ + +# 1. Introduction + +Self-supervised learning has driven the rise of foundation models, enabling training on vast amounts of unlabeled data + +and fueled by the scaling law [35]. This paradigm has proven highly effective for LLMs [55], VLMs [2], and visual generation [51]. In contrast, 3D Vision models still rely heavily on ground-truth 3D geometry and camera pose labels [25, 72], which are usually estimated from time-consuming optimization methods, e.g., COLMAP [60], and are not always perfect. This reliance limits both learning scalability and effectiveness. To break free from this constraint, we move beyond the supervised paradigm and ask: how far can we push a 3D Vision model without any 3D supervision? + +In this paper, we present RayZer, a large multi-view 3D model trained with self-supervision and exhibiting emerging 3D awareness. The input of RayZer is unposed and uncalibrated multi-view images, sampled from continuous video frames or unordered multi-view captures. RayZer first recovers the camera parameters, then reconstructs the scene representation, and finally renders novel views. The key insight of our self-supervised training is to use the camera poses predicted by RayZer itself to render views that provide photometric supervision, rather than following the standard protocol of using ground truth poses for rendering [26, 69, 87]. Thus, RayZer can be trained with zero 3D supervision, i.e., no 3D geometry or camera pose supervision. During inference, RayZer predicts camera and scene representations in a feed-forward manner, without requiring per-scene optimization. We show inference results in Fig. 1. + +As RayZer uses the camera poses predicted by itself for training, this self-supervised task can be interpreted as 3D-aware image auto-encoding [38, 56, 90]. Initially, RayZer disentangles input images into camera parameters and scene representations (reconstruction). It then re-entangles these predicted representations back into images (rendering). To facilitate this disentanglement, we control the information flow. As shown in Fig. 2, we divide all images into two parts: one set predicts the scene representation (input views), while the other offers photometric self-supervision (target views). This is achieved by using estimated poses of the second set to render the scene representation predicted from the first set, thereby preventing trivial solutions that are not 3D-aware. + +To facilitate self-supervised learning, RayZer is built only with transformers – no 3D representation, hand-crafted rendering equation, or 3D-informed architectures. This design is motivated by self-supervised large models in other modalities [2, 6, 51], enabling RayZer to flexibly and effectively learn domain-specific knowledge. The only 3D prior incorporated in RayZer is the ray structure, which simultaneously models the relationship between camera, pixels (image), and scene. Concretely, RayZer first predicts camera poses, which are then converted into pixel-aligned Plücker ray maps [52] to guide the scene reconstruction that follows. This ray-based representation serves as a strong prior for addressing the chicken-and-egg problem of structure and motion [63], effectively allowing the camera and scene representations to regularize each other during training. + +We evaluate RayZer on three datasets, including both scene-level and object-level data with different camera configurations. We observe that RayZer demonstrates comparable or even better novel view synthesis performance than "oracle" methods [30, 86] that use pose labels in both training and testing. Interestingly, we identify that potentially noisy pose annotations from COLMAP can limit the performance of "oracle" models. The results not only demonstrate the effectiveness of RayZer, but also show the potential of 3D Vision models to break free from supervised learning. + +# 2. Related Work + +Large-scale 3D Vision Models. 3D Vision models learn 3D representations and priors from data [14, 21, 36, 53, 54, 66, 67, 88, 89]. Recently, researchers have developed large-scale models to acquire general 3D knowledge. One research direction focuses on designing improved model architectures that incorporate the inductive biases of multiview stereo [10, 13, 70, 81] and epipolar geometry [9, 12, 17, 23]. Another line of work leverages full transformer models that intentionally omit architectural 3D inductive biases [26, 49, 57]. For example, large reconstruction models (LRMs) [25, 69, 73, 86, 93], LEAP [26], and DUSt3R [18, 39, 71, 72, 78] employ transformers to convert 2D image input into 3D representations. SRT [57] and LVSM [30] + +![](images/b1116b7e80c0c3febebe0a75fea930aa175ad58f66d5a15768246256869e5f92.jpg) +Figure 2. Our proposed self-supervised training framework. This is an abstract design that we later operationalize with our RayZer model (illustrated in Fig. 3 and Sec. 4). We divide the input images into two sets $\mathcal{I}_A$ and $\mathcal{I}_{\mathcal{B}}$ . We predict the scene representation from $\mathcal{I}_A$ , and use the predicted cameras of $\mathcal{I}_{\mathcal{B}}$ (shown in orange) to render the scene. We leverage photometric loss between raw input $\mathcal{I}_{\mathcal{B}}$ and its prediction $\hat{\mathcal{I}}_{\mathcal{B}}$ for training. + +further replace 3D representations and physical rendering equations with latent representations and learned rendering functions, improving performance and scalability. However, they still require ground-truth camera poses for supervised training and/or accurate camera annotations during inference. To achieve scalable supervised learning, MegaSynth [29] and Stereo4D [31] leverage synthetic data and stereo videos to expand the data scale, however, curating data for different tasks can be laborious. In contrast, RayZer explores self-supervised training to break free from supervised learning. + +Self-supervised 3D Representation Learning. Learning 3D-aware representations from unlabeled image data is a long-standing problem in 3D Vision. One line of work leverages single-view images. However, they either only work for a specific category [7, 34, 43, 47, 48, 77] or can only recover partial observations [8, 59, 76]. Some works explore semi-supervised learning and achieve better scalability [27, 79], but performance is still highly restricted to the model weights, which are initialized by fully supervised training [80]. The most relevant work is self-supervised learning from multi-view images [64, 74, 75]. For example, Zhou et al. [90], Lai et al. [38], and their following works [19, 82] use camera motion as 2D or 3D warping operations to regularize learning. However, this strong inductive bias limits the learning effectiveness. RUST [58] is a pioneering work in learning latent scene representations from unposed imagery. RayZer is different in three aspects. First, RayZer initially estimates camera poses and uses poses to condition the following latent reconstruction. In contrast, RUST operates in an inverse pipeline – it first reconstructs the scene and then estimates the camera poses. Second, RayZer employs different explicit pose representations to improve information disentanglement and 3D awareness, en + +abling novel view synthesis by geometrically interpolating predicted poses. Instead, RUST uses a latent pose representation, which makes scene-posedisentanglement challenging and they are not explicitly 3D-aware. Third, RayZer estimates camera poses for all views, while RUST estimate latent poses only for target views. + +Optimization-based Unsupervised SfM, SLAM, and NVs. Although these methods are not directly comparable to RayZer, we discuss them due to the similar input-output formulations. In detail, these methods optimize target predictions on a per-scene basis [60, 62], while RayZer is a feedforward parametrized model, learning priors by training on large data. The traditional SfM, SLAM, and NVS methods are unsupervised [24, 60]. Although generally performing well, they are restricted by the complicated hand-crafted workflow, leading to requirements of dense-view inputs [84], slow speed [60], and sensitivity to hyper-parameters [65]. Recent optimization-based NeRF and 3DGS works can also perform NVs from unposed images [4, 20, 44, 61]. However, they do not have learnable model parameters to encode priors, thus requiring off-the-shelf models trained with 3D supervision as regularization or providing initialization. + +# 3. Preliminaries + +We introduce two important building blocks of RayZer, i.e., the latent set scene representation and how to render it. + +Latent set scene representation. Compressing data into tokens in latent space is a common practice in text, image, video, etc. Recently, this representation has also been extended to 3D research [30, 57, 58, 83]. In contrast to classical explicit (e.g., meshes and point clouds), implicit (e.g., NeRF [46] and SDF [50]), and hybrid (e.g., triplane [8] and 3DGS [37]) representations that are 3D-aware, the latent set representation is not explicitly 3D-aware. It serves as a compression of scene information, where the 3D-awareness properties are fully learned. The latent set scene representation can be denoted as $\mathbf{z} \in \mathbb{R}^{n \times d}$ , where $n$ is the number of tokens in the set and $d$ is the latent feature dimension. + +Rendering latent set scene representation requires a network, say $R^{\theta}$ , as introduced by SRT [57] and LVSM [30]. We formulate it as $v = R^{\theta}(\mathbf{z}, r)$ , where $r$ is a ray and $v$ is the rendered property, e.g., RGB values, of the corresponding pixel1. This formulation is the same as traditional Graphics rendering techniques [1, 33], as $v = R(\text{SCENE}, \text{RAY})$ , where $R$ is a pre-defined and handcrafted rendering equation, e.g., alpha-blending ray marching in NeRF. Differently, our "rendering equation" is a learned model parameterized with weights $\theta$ , and our scene representation is a latent token set as discussed previously. We omit the model parametrization, e.g., weight $\theta$ , in the following description for clarity. + +# 4. RayZer + +In this section, we first introduce RayZer's self-supervised learning framework (Sec. 4.1). Then, we present the details of the RayZer model architecture (Sec. 4.2). + +# 4.1. RayZer's Self-supervised Learning + +We first formulate the input and output of RayZer. We then introduce the self-supervised learning framework. + +We focus on the standard setting of modeling static scenes [60]. The input of RayZer is a set of unposed and uncalibrated multi-view images $\mathcal{I} = \{I_i\in \mathbb{R}^{H\times W\times 3}|i = 1,\dots,K\}$ , which can come from unlabeled video frames or image sets. The output is a parametrization of the inputs, i.e., camera intrinsics, per-view camera poses, and scene representation, enabling novel view synthesis. To predict these representations, we build the RayZer model and train it with self-supervised learning - no 3D supervision, i.e., 3D geometry, and camera pose annotations during training. + +To train RayZer with self-supervision, we control the data information flow. We split the input images $\mathcal{I}$ into two non-overlapping subsets $\mathcal{I}_A$ and $\mathcal{I}_B$ , where $\mathcal{I}_A \cup \mathcal{I}_B = \mathcal{I}$ and $\mathcal{I}_A \cap \mathcal{I}_B = \emptyset$ . RayZer uses $\mathcal{I}_A$ to predict the scene representation, and use $\mathcal{I}_B$ for providing supervision. Thus, RayZer renders images that correspond to $\mathcal{I}_B$ , denoted as $\hat{\mathcal{I}}_B$ , and we apply photometric losses: + +$$ +\mathcal {L} = \frac {1}{K _ {\mathcal {B}}} \sum_ {\hat {I} \in \hat {\mathcal {I}} _ {\mathcal {B}}} (\operatorname {M S E} (I, \hat {I}) + \lambda \cdot \operatorname {P e r c e p} (I, \hat {I})), \quad (1) +$$ + +where $K_{\mathcal{B}} = |\mathcal{I}_{\mathcal{B}}|$ is the size (number of images) of $\mathcal{I}_{\mathcal{B}}$ , $I \in \mathcal{I}_{\mathcal{B}}$ is the image that corresponds to a predicted image $\hat{I}$ , and $\lambda$ is the weight for perceptual loss [32, 42]. The two sets are randomly sampled during training. + +# 4.2. RayZer Model + +Overview. As introduced in Sec. 4.1, RayZer recovers both camera parameters and the scene representation from unposed, uncalibrated input images. A key design element of RayZer is its cascaded prediction of camera and scene representations. This is motivated by the fact that even noisy cameras can be a strong condition for better scene reconstruction [28, 60, 87], which is analogous to traditional structure-from-motion methods [60] and is in contrast with recent reconstruction-first methods [58, 69, 72]. This design can provide mutual regularization of predicting pose and scene during training, facilitating self-supervised learning. + +RayZer builds a pure transformer-based model, benefiting from its scalability and flexibility. As shown in Fig. 3, RayZer first tokenizes input images and uses a transformer-based encoder to predict camera parameters of all views. In this step, the cameras are represented by their intrinsics and SE (3) camera poses. This low-dimensional, geometri + +![](images/6848d1f591fa19de1137e03201eb6c57475cb49336b542dc330189a6e4384115.jpg) +Figure 3. RayZer self-supervised learning framework. RayZer takes in unposed and uncalibrated multi-view images $\mathcal{I}$ and predicts per-view camera parameters and a scene representation, which supports novel view rendering. (Left) RayZer first estimates camera parameters, where one view is selected as the canonical reference view (in blue box). RayZer predicts the intrinsics and the relative camera poses $\mathcal{P}$ of all views. The predicted cameras are then converted into pixel-aligned Plücker ray maps $\mathcal{R}$ . (Middle) RayZer uses a subset of input images, $\mathcal{I}_{\mathcal{A}}$ , as well as their previously predicted camera Plücker ray maps, $\mathcal{R}_{\mathcal{A}}$ , to predict a latent scene representation. Here, the Plücker ray maps, $\mathcal{R}_{\mathcal{A}}$ , serve as an effective condition for scene reconstruction. (Right) RayZer can render a target image given the scene representation $\mathbf{z}^{*}$ and a target camera. During training, we use $\mathcal{R}_{\mathcal{B}}$ , which is the previously predicted cameras Plücker ray maps of $\mathcal{I}_{\mathcal{B}}$ , to render $\hat{\mathcal{I}}_{\mathcal{B}}$ . This allows training RayZer end-to-end with self-supervised photometric losses between inputs $\mathcal{I}_{\mathcal{B}}$ and their renderings $\hat{\mathcal{I}}_{\mathcal{B}}$ . + +cally well-defined parametrization helps disentangle image information from the camera representation. + +RayZer then transforms the SE (3) camera poses and intrinsics into Plucker ray maps [52], representing the predicted cameras as pixel-aligned rays. This ray-based representation captures both the 2D ray-pixel alignment and the 3D ray geometry, providing fine-grained, ray-level details that encapsulate the physical properties of the camera model. The ray maps serve as a condition for improving the reconstruction stage that follows. + +From the image and predicted Plücker rays of $\mathcal{I}_A$ , RayZer uses another transformer-based encoder to predict the latent set scene representation (introduced in Sec. 3 and detailed later). Then, RayZer uses the previously estimated cameras of $\mathcal{I}_B$ to predict $\hat{\mathcal{I}}_B$ , providing photometric self-supervision (Eq. 1). We now formally introduce the RayZer model. + +Image Tokenization. For all $K$ input images $\mathcal{I} = \{I_i\in \mathbb{R}^{H\times W\times 3}|i = 1,\dots ,K\}$ , we patchify them into nonoverlapping patches following ViT [16]. Each patch is in $\mathbb{R}^{s\times s\times 3}$ , where $s$ is the patch size. We use a linear layer to encode each patch into a token in $\mathbb{R}^d$ , leading to a patch-aligned token map $f_{i}\in \mathbb{R}^{h\times w\times d}$ for each image, where $h = H / s$ , $w = W / s$ , and $d$ is the latent dimension. + +We then add positional embeddings (p.e.) to the tokens, enabling the following model to be aware of the spatial location and the corresponding image index of each token. Specifically, we combine the sinusoidal spatial p.e. [16] and the sinusoidal image index p.e. [3] using a linear layer; note that the image index p.e. is shared among all tokens from the same image. When training on continuous video frames, these image index embeddings also encode sequential priors, + +which benefits pose estimation. Finally, we reshape the token maps of all images into a set, denoted as $\mathbf{f} \in \mathbb{R}^{Khw \times d}$ (recall that the transformer is invariant to the permutation of tokens). For brevity, we will use this notation for latent token sets throughout the rest of the paper. + +Camera Estimator. The camera estimator $\mathcal{E}_{cam}$ predicts camera parameters, i.e., camera poses and intrinsics, for all input images. We use a learnable camera token in $\mathbb{R}^{1\times d}$ as the initial feature for this prediction for all views. We repeat the token $K$ times and add them with image index p.e. such that they correspond to the $K$ images. We denote this camera feature initialization as $\mathbf{p}\in \mathbb{R}^{K\times d}$ . We then use the camera estimator composed of full self-attention transformer layers to update the camera tokens, as: + +$$ +\left\{\mathbf {f} ^ {*}, \mathbf {p} ^ {*} \right\} = \mathcal {E} _ {c a m} (\left\{\mathbf {f}, \mathbf {p} \right\}), \tag {2} +$$ + +where $\{\cdot, \cdot\}$ denotes concatenation along the token dimension (the union set of two token sets), and $\mathbf{f}^*$ and $\mathbf{p}^*$ are the updated tokens. We note that $\mathbf{f}^*$ is not used for the following computation – it is only used as context to update $\mathbf{p}$ in the transformer layers. For clarity, we formulate the transformer layers as follows: + +$$ +\mathbf {y} ^ {0} = \{\mathbf {f}, \mathbf {p} \}, \tag {3} +$$ + +$$ +\mathbf {y} ^ {l} = \text {T r a n s f o r m e r L a y e r} ^ {l} \left(\mathbf {y} ^ {l - 1}\right), l = 1, \dots , l _ {T} \tag {4} +$$ + +$$ +\left\{\mathbf {f} ^ {*}, \mathbf {p} ^ {*} \right\} = \operatorname {s p l i t} \left(\mathbf {y} ^ {l _ {T}}\right), \tag {5} +$$ + +where $l_{T}$ is the number of layers, and the split operation recovers the two token sets, inverting Eq. 3. This notation remains consistent throughout the rest of the paper. + +We then predict the camera parameters for each image independently. For camera pose prediction, we follow prior + +works of using relative camera poses to resolve ambiguity [28, 84]. We select one view as the canonical reference (e.g., with identity rotation and zero translation), while for every non-canonical view, we predict its relative pose with respect to the canonical view. We parametrize the SO (3) rotation using a continuous 6D representation [92], and we predict the relative pose with a two-layer MLP as follows: + +$$ +p _ {i} = \mathrm {M L P} _ {p o s e} ([ \mathbf {p} _ {i} ^ {*}, \mathbf {p} _ {c} ^ {*} ]), \tag {6} +$$ + +where $[\cdot, \cdot]$ denotes concatenation along the feature dimension, $\mathbf{p}_i^*$ and $\mathbf{p}_c^*$ (all in $\mathbb{R}^d$ ) are the camera tokens for image $I_i$ and the canonical view, respectively. The output $p_i \in \mathbb{R}^9$ represents the predicted pose parameters, which are then transformed into an SE (3) pose $\mathbf{P}_i$ for image $I_i$ . + +For intrinsics prediction, following prior works [22, 38], we parameterize intrinsics using a single focal length value, under the assumptions that i) the focal lengths along the x and y axes are identical, ii) all views share the same intrinsics, and iii) the principal point is at the image center. We predict the focal length using a two-layer MLP: + +$$ +\operatorname {f o c a l} = \operatorname {M L P} _ {\text {f o c a l}} \left(\mathbf {p} _ {c} ^ {*}\right). \tag {7} +$$ + +The predicted focal length is then converted into the intrinsic matrix $\mathbf{K} \in \mathbb{R}^{3 \times 3}$ . + +Scene Reconstructor. As discussed in Sec. 4.1, we predict the scene representation from image set $\mathcal{I}_A$ and additionally condition it on the previously predicted camera parameters $\mathcal{P}_A = \{(\mathbf{P}_i,\mathbf{K})|I_i\in \mathcal{I}_A\}$ . We first convert $\mathcal{P}_A$ to pixel-aligned Plücker rays [52] for each image, denoted as $\mathcal{R}\in \mathbb{R}^{K\times H\times W\times 6}$ . Similar to image inputs, we also tokenize the Plücker rays into patch-level tokens using a linear layer, yielding $\mathbf{r}\in \mathbb{R}^{Khw\times d}$ . We index the image and Plücker rays tokens corresponding to the image set $\mathcal{I}_A$ , denoted as $\mathbf{f}_A$ and $\mathbf{r}_A$ (each in $\mathbb{R}^{K_Ahw\times d}$ , respectively). We fuse these tokens along the feature dimension with a two-layer MLP: + +$$ +\mathbf {x} _ {\mathcal {A}} = \operatorname {M L P} _ {f u s e} ([ \mathbf {f} _ {\mathcal {A}}, \mathbf {r} _ {\mathcal {A}} ]), \tag {8} +$$ + +where $\mathbf{x}_{\mathcal{A}} \in \mathbb{R}^{K_{\mathcal{A}}hw \times d}$ represents the fused tokens. Importantly, we use the raw image tokens $\mathbf{f}$ rather than the pose transformer output $\mathbf{f}^*$ for this fusion. This design choice prevents leakage of information from the image set $\mathcal{I}_{\mathcal{B}}$ , since the camera estimator transformer producing $\mathbf{f}^*$ has access to a global context that includes tokens from $\mathcal{I}_{\mathcal{B}}$ . + +We then employ a scene reconstructor $\mathcal{E}_{\text{scene}}$ consisting of full self-attention transformer layers to predict the latent scene representation. To initialize this representation, we use a set of learnable tokens $\mathbf{z} \in \mathbb{R}^{L \times d}$ , where $L$ denotes the number of tokens. We formulate the process as follows: + +$$ +\left\{\mathbf {z} ^ {*}, \mathbf {x} _ {\mathcal {A}} ^ {*} \right\} = \mathcal {E} _ {\text {s c e n e}} \left(\left\{\mathbf {z}, \mathbf {x} _ {\mathcal {A}} \right\}\right). \tag {9} +$$ + +The update rule is identical to the transformer layers in the camera estimator $\mathcal{E}_{cam}$ . Here, $\mathbf{z}^*$ represents the final latent + +scene representation predicted from $\mathcal{I}_A$ . Meanwhile, $\mathbf{x}_{A}^{*}$ is discarded. + +Rendering Decoder. We first define the rendering decoder and then describe its training usage. + +We use a transformer-based decoder with full self-attention for rendering, following LVSM [30]. For a target image, we begin by representing it as pixel-aligned Plücker rays and tokenize these rays using a linear layer to obtain target tokens $\mathbf{r} \in \mathbb{R}^{hw \times d}$ . Next, we fuse the scene information by updating the tokens with a decoder $\mathcal{D}_{\text{render}}$ comprising transformer layers: + +$$ +\left\{\mathbf {r} ^ {*}, \mathbf {z} ^ {\prime} \right\} = \mathcal {D} _ {\text {r e n d e r}} \left(\left\{\mathbf {r}, \mathbf {z} ^ {*} \right\}\right), \tag {10} +$$ + +where $\mathbf{z}'$ is subsequently discarded, while the update rule of $\mathcal{D}_{\text {render }}$ is the same as previously introduced modules. Finally, we decode the RGB values at the patch level with an MLP: + +$$ +\hat {I} = \mathbf {M L P} _ {r g b} \left(\mathbf {r} ^ {*}\right), \tag {11} +$$ + +where $\hat{I} \in \mathbb{R}^{hw \times (3s^2)}$ . We reshape $\hat{I}$ to recover the 2D spatial structure, yielding a final rendered image in $\mathbb{R}^{H \times W \times 3}$ . + +During training, we use the predicted Plücker ray maps $\mathcal{R}_{\mathcal{B}}$ , which correspond to $\hat{\mathcal{I}}_B$ , to render images of $\hat{\mathcal{I}}_B$ and then compute the self-supervised loss as defined in Eq. 1. + +# 5. Experiments + +In this section, we introduce the experimental setting and present the evaluation results. For the implementation, RayZer employs 24 transformer layers, with 8 layers for each of the camera estimator, scene encoder, and rendering decoder. We train RayZer with a learning rate of $4 \times 10^{-4}$ with a cosine scheduler for 50,000 iterations and a batch size of 256. The weight of perceptual loss is $\lambda = 0.2$ . For all experiments, we used a resolution of 256 with a patch size of 16. More details are in the Appendix. + +# 5.1. Experimental Setup + +We introduce our experimental setup, including datasets, evaluation protocol and metrics, as well as baseline methods. + +Datasets. We use three datasets to evaluate RayZer, including two scene-level datasets, DL3DV [45] and RealEstate [91], and an object-level dataset Objaverse [15] (rendered as videos). We train and test on each dataset separately. The numbers of input views $(\mathcal{I}_A)$ and target views $(\mathcal{I}_B)$ are set to 16 and 8 for DL3DV, 5 and 5 for RealEstate, and 12 and 8 for Objaverse, respectively. We sample input images with the index ranges of 64-96, 128-192, and 50-65 on DL3DV, RealEstate, and Objaverse, respectively. These values are chosen based on data difficulty, especially camera baseline, following prior works [9, 86, 93]. We use the official DL3DV train-test split, and split RealEstate following [9]. More details can be found in the Appendix. + +![](images/bf0438fc3de59a45df9564c78b7d47b3ec001220ea2f2768cf017961acde8243.jpg) +Figure 4. Visualization results on RealEstate and DL3DV. We compare RayZer with "oracle" methods GS-LRM and LVSM, which use COLMAP pose annotations in both training and testing. Our self-supervised RayZer model does not use any pose annotations. Generally, RayZer performs on par with "oracle" methods (first row), and can outperform them on cases that COLMAP usually struggles to handle, e.g., glasses and white walls (highlighted with red boxes). The results verify our analysis on the problems of using COLMAP in Sec. 5.2. + +![](images/cc1f69d2c7c38ad6aba582510c9141144708eafe634d4ab3b665d280ee09a26b.jpg) + +Evaluation Protocol and Metrics. We evaluate novel view synthesis quality. Specifically, the evaluation protocol of RayZer is different from the "oracle" and supervised methods, which use ground-truth poses to render images. Instead, we use predicted poses to render novel views, thereby assessing the compatibility between the predicted poses and the scene representation. Since the model is trained without explicit pose annotations, the learned poses exist in a different space, and their direct correspondence to standard pose annotations is unknown. This evaluation protocol follows RUST [58]. We note that the target views are used only for pose estimation and not for scene representation prediction, ensuring that no information leakage occurs. + +Baselines. We compare RayZer with two types of methods, including 1) "oracle" methods, i.e., GS-LRM [86] and LVSM [30] (encoder-decoder version), that use ground-truth camera poses during both training (as supervision) and inference (as pre-requisite). LVSM also uses latent set scene representation. Thus, it serves as the main comparison for the "oracle" methods; 2) supervised method, i.e., PF-LRM [69], which requires camera supervision to learn pose estimation and reconstruction; thus, it is pose-free during inference. For fair comparisons, we use 16 transformer layers in total for GS-LRM and LVSM. Thus, their number of parameters is the same as RayZer, except that RayZer has another camera estimator to handle unposed images. We use 24 transformer layers for PF-LRM. We also consider the self-supervised + +
Training SupervisionInference w. COLMAP Cam.Even SampleRandom Sample
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
“Oracle” methods (assume inputs are posed & use pose annotations during training)
GS-LRM2D + CameraYes23.490.7120.25223.020.7050.266
LVSM2D + CameraYes23.690.7230.24223.100.7030.257
Unsupervised methods (inputs are un-posed & no pose annotations used during training)
RayZer2DNo24.360.7570.20923.720.7330.222
+ +Table 1. Evaluation results on DL3DV. The camera annotations used by the "oracle" models come from COLMAP. The results are reported with continuous video frames (ordered) as the input. The results for the unordered image set input are in Table. 4. The input and target views can be evenly or randomly sampled from video frames. We bold our result if it is better than the "oracle" models. + +method RUST [58], but since it does not have an official public implementation, we ablate the key design differences between RUST and RayZer in Table 7 instead. + +# 5.2. Results + +Main results. Table 1-3 summarizes the results on the three datasets. Remarkably, without any 3D labels (e.g., camera pose annotations) during training, RayZer achieves performance comparable to the best "oracle" model, LVSM. In fact, RayZer even outperforms LVSM on DL3DV and RealEstate10k while performing slightly worse on Objverse. We conjecture that this is because the camera poses in DL3DV and RealEstate are annotated by COLMAP, which can be imperfect and set an upper bound for "oracle" methods that are supervised by COLMAP annotations. In contrast, our self-supervised approach enables the model to learn a + +![](images/bb3f9f6ffe69baba085f5acb2e719cdc7dc898996232777c4b9a53108c9570aa.jpg) +Figure 5. Visualization results on Objaverse. RayZer performs on par with LVSM and outperforms the supervised method PF-LRM. + +
Training SupervisionInference w. COLMAP Cam.Even SampleRandom Sample
PSNR†SSIM†LPIPS‡PSNR†SSIM†LPIPS‡
“Oracle” methods (assume inputs are posed & use pose annotations during training)
GS-LRM2D + CameraYes24.250.7700.22723.210.7480.251
LVSM2D + CameraYes27.000.8510.15725.880.8280.175
Unsupervised methods (inputs are un-posed & no pose annotations used during training)
RayZer2DNo27.480.8610.14626.320.8350.164
+ +pose space that optimally benefits latent reconstruction and novel view synthesis. This hypothesis is further supported by the results on Objaverse – a synthetic dataset with perfect pose annotations from the rendering tool – where LVSM, acting as a true oracle, outperforms RayZer. Nonetheless, the small performance gap showcases the effectiveness of our self-supervised training. Visualizations in Fig. 4 further support our conjecture regarding COLMAP's noisy poses, as both LVSM and GS-LRM consistently underperform on challenging cases that COLMAP usually fails. These results not only validate our self-supervised learning approach but also demonstrate its potential to break free from the limitations of supervised learning. + +Using unordered image sets for training. RayZer can be trained on continuous video frames (Table 1-3) or unordered image sets (Table 4). Note that these two training settings are applied separately. As shown in Table 4, we observe that the model trained with unordered image sets performs worse than the one trained with continuous video frames. We notice that the difference is at the pose estimation stage – specifically, the image index positional embedding encourages local pose smoothness that benefits the learning of pose estimation on continuous frames. This finding suggests that scaling training data using video resources, which are plentiful online, could be more advantageous than relying + +Table 2. Evaluation results on RealEstate with continuous video frames inputs. The camera annotations come from COLMAP. + +
Training SupervisionInference w. GT Cam.Even SampleRandom Sample
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
“Oracle” methods (assume inputs are posed & use pose annotations during training)
LVSM2D + GT Cam.Yes32.340.9500.05032.340.9490.051
Supervised methods (inputs are un-posed & use pose annotations during training)
PF-LRM2D + GT Cam.Yes (render)25.480.8820.11025.430.8810.111
Unsupervised methods (inputs are un-posed & no pose annotations used during training)
RayZer2DNo31.520.9450.05231.420.9430.053
+ +Table 3. Evaluation results on Objaverse with continuous video frames inputs. The camera annotations are Blender ground-truth. PF-LRM uses ground-truth poses to render novel views, same with oracle methods, and we evaluate its predicted pose in Table 5. + +
Training SupervisionInf. w. GT PoseContinuous InputsEven SampleRandom Sample
PSNR†SSIM†LPIPS‡PSNR†SSIM†LPIPS‡
(1)2DNo✓24.360.7570.20923.720.7330.222
(2)2DNoX20.560.5760.33420.020.5660.356
+ +Table 4. Evaluating RayZer performance when using continuous or unordered images for training on DL3DV. In evaluations, the input frames are sampled from continuous video frames. (1) keeps their temporal continuity (encoded by the image index p.e.) during training. (2) randomly shuffles the images during training. + +on unordered image sets that are often limited in scale and contain noisy content [41, 68]. + +# 5.3. Analysis of Camera Poses + +RayZer's learned camera pose space. We visualize some camera poses predicted by RayZer in Fig. 6. Although RayZer predicts SE (3) camera poses, we observe that these poses do not exactly match the real-world pose space. This result indicates that the SE (3) poses, which are later converted into Plücker ray maps, offer a degree of flexibility. Since both the rendering decoder and the scene representation operate in latent space, RayZer remains robust to any warping between the learned pose space and the actual real-world poses, as long as the poses are compatible with the scene representation and the decoder. + +3D Awareness of predicted camera poses. We further investigate whether the pose space learned by RayZer is 3D aware. To this end, we interpolate the predicted poses of input views to synthesize more novel views, where the camera pose of a novel view is interpolated from two neighboring input views. We use ground-truth camera poses to calculate the interpolation coefficients, checking whether predicted poses follow the same geometric interpolation rules. We include the details of the interpolation method in Appendix. As shown in Table 5, RayZer demonstrates significantly better performance than PF-LRM and the naive baseline of copying the nearest rendered input view. These results verify that poses predicted by RayZer are interpolatable and 3D-aware. + +Probing the learned camera pose space. To probe how much actual pose information is learned by RayZer, we follow RUST [58] to fit a lightweight 2-layer MLP head on the pose features. We freeze the camera estimator's transformers and train the MLP under camera supervision. As shown in Table 6, our probing outperforms the supervised baseline (which has the same model architecture and uses + +
Training SupervisionInference w. GT PoseEven SampleRandom Sample
PSNR†SSIM†LPIPS‡PSNR†SSIM†LPIPS‡
Supervised methods (inputs are un-posed & use pose annotations during training)
PF-LRM2D + GT PoseNo20.630.8190.16021.270.8270.154
Unsupervised methods (inputs are un-posed & no pose annotations used during training)
RayZer-copy2DNo19.560.8120.15920.170.8200.150
RayZer2DNo27.010.9000.07526.870.8960.078
+ +transformers trained from scratch), indicating that RayZer's novel view synthesis self-supervision facilitates a better latent pose space. In contrast, supervised learning struggles due to the challenges of low-dimensional pose representation [5, 11, 40, 85, 92]. + +# 5.4. Ablation Study + +We ablate the main design choices of RayZer from three aspects, including scene representation, 3D prior, and the overall model paradigm. As shown in Table 7 (1), when using the 3DGS representation rather than the latent set representation, the training does not converge. This verifies the optimization difficulty of explicit 3D representation [37, 86] and demonstrates the flexibility of the latent representation with its learned rendering decoder. + +Table 7 (2) and (3) ablate the prior of camera representation. Without Plücker ray maps, we observe a degraded performance in (2), showing the effectiveness of using Plücker ray maps to regularize the solution of structure-and-motion problem. Besides, we observe a slightly better performance of (3), which directly uses camera tokens $\mathbf{p}^*$ , compared to (2). The reason is that the camera tokens $\mathbf{p}^* \in \mathbb{R}^d$ can leak target image information, while SE (3) poses used in (2) serve as an information bottleneck to enforce this disentanglement. Moreover, SE (3) poses are geometrically well-defined, allowing us to interpolate them and generate novel views along the interpolated camera trajectory, while the latent camera representation is not directly interpolable. + +Table 7 (4) ablates the overall paradigm. When the model first predicts the latent scene and then estimates poses, we observe a degraded performance. In detail, the pose estimator takes the scene representation and target image feature tokens as inputs. The result verifies our insight that pose estimation can be a strong condition for scene reconstruction, championing traditional pose-first methods in the context of self-supervised learning. Note that combining (3) and (4) will be a model that is similar to RUST conceptually. + +# 6. Conclusion + +We introduce RayZer, a self-supervised large multi-view 3D Vision model trained with zero 3D supervision, i.e., no + +![](images/3fe3c88ab8ad674b380d47d03722c937c46a08edc6bf2f34463663240987c7b3.jpg) +Rendered Views (3/5) + +![](images/e93588bf8cddb8168151be556f5226da69bb61369ffe0be176e441b44451e014.jpg) + +![](images/eeee064022149830eadcfd42af36642c3336fb168ea5ef060df714505c8e031d.jpg) + +![](images/dc34271b703e100dcc255b2ac8c30d28c7d24e7824a6ffa270fb060696e60726.jpg) +Cameras + +![](images/033da0ddd8c4eab98a5a7d404c303b90ce4b5b844300faf77a391e9125593a40.jpg) + +![](images/43791b5e09b802a8ca3d0ef26b60a6448dd3f0fe10c8c7b345ec2ff8db140074.jpg) + +![](images/accb2c98b3ef0a83d65860af1840b142403128ac540ec2c54b6e345002aed6cd.jpg) + +![](images/a739dd3b2989e9fcada7aa3a68d3a69a256cf934bca7e0ebbc250167e94d8d85.jpg) + +![](images/50750d0d6f120f2f8fc3149fc8076070360d5abbab6c8e7645b1570492623565.jpg) +Figure 6. Visualization of RayZer predicted cameras learned with self-supervision. We visualize 3 out of 5 rendered views due to space limit, where the image index is highlighted by its color. + +![](images/3b4fa8f67f00900f588652bf9a379643827f070c7ec61fc7306909be454b2ca3.jpg) + +![](images/4d99acac4f26da5176277c4fb2b49e016c52be577675fae2ef1f90b44ff1c189.jpg) + +![](images/981981df2fee3a09e5b9dc93026bc8ae2f1a91be92473fa976b126e6c99d13f0.jpg) + +Table 5. Evaluating 3D awareness of predicted camera poses on Objaverse. Unlike Table 3, here we render novel views by interpolating predicted poses of input views, where the interpolation coefficients are calculated from GT poses. This experiment tests whether the learned SE (3) poses are geometrically well-defined and 3D-aware. We also compare against a naive baseline "RayZer-copy" that simply copies the nearest rendered input view. + +
Pose Encoder (Epose)Rotation Acc.↑ (%)Translation Acc.↑ (%)
R@10°R@20°R@30°t@0.1t@0.2t@0.3
DL3DVsupervised39.363.077.815.733.144.4
self-supervised47.672.584.020.844.060.5
RealEstatesupervised87.096.499.644.659.382.5
self-supervised99.699.910061.284.292.8
Objaversesupervised19.846.766.015.137.253.8
self-supervised33.669.286.820.152.775.5
+ +Table 6. Effectiveness of self-supervised pre-training for pose estimation. We train a two-layer MLP (with supervised learning) to read out latent camera tokens $\mathbf{p}^*$ predicted by the pose encoder $\mathcal{E}_{\text {pose }}$ , where the backbone is frozen. At the same time, we also compare with the baseline where both encoder $\mathcal{E}_{\text {pose }}$ and the pose prediction MLP are trained with supervised learning from scratch. + +
Even SampleRandom Sample
PSNRSSIMLPIPSPSNRSSIMLPIPS
(0)RayZer24.360.7570.20923.720.7330.222
(1)Representation - 3DGS + rasterization--failed---
(2)Prior - no Plücker ray, use SE (3) pose22.730.6870.24921.880.6470.274
(3)Prior - no explicit pose, use latent camera23.130.7000.25122.360.6680.272
(4)Paradigm - scene first, not pose first13.310.3380.73213.120.3370.729
+ +Table 7. Ablation study of RayZer designs on DL3DV with continuous inputs. (1) is a variant uses the 3D Gaussian representation rather than latent scene representation with its learned rendering decoder used by RayZer; (2) does not use Plücker ray maps $\mathcal{R}_{\mathcal{A}}$ for conditioning latent reconstruction. Instead, it encodes the SE (3) poses $\mathbf{P}_{\mathcal{A}}$ and intrinsics $\mathbf{K}$ into tokens as condition; (3) directly uses the latent camera tokens $\mathbf{p}^*$ , rather than converting it to any explicit forms of cameras, to condition the latent scene reconstruction; (4) first reconstructs latent scene and then estimates pose as Plücker ray maps, contrasting our pose-first paradigm. + +3D geometry and camera annotations. RayZer achieves comparable or even better novel view synthesis performance than prior works that use pose labels in both training and inference, verifying the feasibility of breaking free from supervised learning. + +Acknowledgment. We acknowledge the support of NSF-2047677, NSF-2413161, NSF-2504906, NSF-2515626, and Gifts from Adobe and Google. + +# References + +[1] Arthur Appel. Some techniques for shading machine renderings of solids. In Proceedings of the April 30-May 2, 1968, spring joint computer conference, pages 37-45, 1968. +[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In CVPR, pages 22861-22872, 2024. +[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. +[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 CVPR, pages 4160–4169, 2023. +[5] Romain Brégier. Deep regression on manifolds: a 3d rotation case study. In 2021 International Conference on 3D Vision (3DV), pages 166-174. IEEE, 2021. +[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. *NeurIPS*, 33:1877–1901, 2020. +[7] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, pages 5799-5809, 2021. +[8] 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 CVPR, pages 16123-16133, 2022. +[9] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, pages 19457-19467, 2024. +[10] 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 ICCV, pages 14124-14133, 2021. +[11] Jiayi Chen, Yingda Yin, Tolga Birdal, Baoquan Chen, Leonidas J Guibas, and He Wang. Projective manifold gradient layer for deep rotation regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6646-6655, 2022. +[12] Yuedong Chen, Haofei Xu, Qianyi Wu, Chuanxia Zheng, Tat Jen Cham, and Jianfei Cai. Explicit correspondence matching for generalizable neural radiance fields. arXiv preprint arXiv:2304.12294, 2023. +[13] 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. In ECCV, pages 370-386. Springer, 2024. +[14] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for + +single and multi-view 3d object reconstruction. In ECCV, pages 628-644. Springer, 2016. +[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. +[16] 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. +[17] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In CVPR, pages 4970-4980, 2023. +[18] Bardienus Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. arXiv preprint arXiv:2409.19152, 2024. +[19] Yang Fu, Ishan Misra, and Xiaolong Wang. Mononerf: Learning generalizable nerfs from monocular videos without camera poses. In International Conference on Machine Learning, pages 10392-10404. PMLR, 2023. +[20] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In CVPR, pages 20796-20805, 2024. +[21] Rohit Girdhar, David F Fouhey, Mikel Rodriguez, and Abhinav Gupta. Learning a predictable and generative vector representation for objects. In ECCV, pages 484-499. Springer, 2016. +[22] 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. +[23] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shoou-I Yu. Epipolar transformers. In CVPR, pages 7779-7788, 2020. +[24] Benno Heigl, Reinhard Koch, Marc Pollefeys, Joachim Denzler, and Luc Van Gool. Plenoptic modeling and rendering from image sequences taken by a hand-held camera. In *Mustererkennung* 1999: 21. DAGM-Symposium Bonn, 15–17. September 1999, pages 94–101. Springer, 1999. +[25] 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. +[26] Hanwen Jiang, Zhenyu Jiang, Yue Zhao, and Qixing Huang. Leap: Liberate sparse-view 3d modeling from camera poses. arXiv preprint arXiv:2310.01410, 2023. +[27] Hanwen Jiang, Qixing Huang, and Georgios Pavlakos. Real3d: Scaling up large reconstruction models with real-world images. arXiv preprint arXiv:2406.08479, 2024. +[28] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown categories and camera poses. In 2024 International Conference on 3D Vision (3DV), pages 31-41. IEEE, 2024. + +[29] Hanwen Jiang, Zexiang Xu, Desai Xie, Ziwen Chen, Haian Jin, Fujun Luan, Zhixin Shu, Kai Zhang, Sai Bi, Xin Sun, et al. Megasynth: Scaling up 3d scene reconstruction with synthesized data. arXiv preprint arXiv:2412.14166, 2024. +[30] 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. arXiv preprint arXiv:2410.17242, 2024. +[31] Linyi Jin, Richard Tucker, Zhengqi Li, David Fouhey, Noah Snavely, and Aleksander Holynski. Stereo4d: Learning how things move in 3d from internet stereo videos. arXiv preprint arXiv:2412.09621, 2024. +[32] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016. +[33] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. +[34] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, pages 371-386, 2018. +[35] 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. +[36] Abhishek Kar, Christian Hane, and Jitendra Malik. Learning a multi-view stereo machine. NIPS, 30, 2017. +[37] 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. +[38] Zihang Lai, Sifei Liu, Alexei A Efros, and Xiaolong Wang. Video autoencoder: self-supervised disentanglement of static 3d structure and motion. In ICCV, pages 9730-9740, 2021. +[39] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In ECCV, pages 71-91. Springer, 2024. +[40] Jake Levinson, Carlos Esteves, Kefan Chen, Noah Snavely, Angjoo Kanazawa, Afshin Rostamizadeh, and Ameesh Makadia. An analysis of svd for deep rotation estimation. Advances in Neural Information Processing Systems, 33:22554-22565, 2020. +[41] 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. +[42] Zhengqi Li, Wenqi Xian, Abe Davis, and Noah Snavely. Crowdsampling the plenoptic function. In ECCV, pages 178-196. Springer, 2020. +[43] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. Sdf-srn: Learning signed distance 3d object reconstruction from static images. NeurIPS, 33:11453-11464, 2020. +[44] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, pages 5741-5751, 2021. +[45] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. + +D13dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In CVPR, pages 22160-22169, 2024. +[46] 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. +[47] Siva Karthik Mustikovela, Varun Jampani, Shalini De Mello, Sifei Liu, Umar Iqbal, Carsten Rother, and Jan Kautz. Self-supervised viewpoint learning from image collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3981, 2020. +[48] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In ICCV, pages 7588-7597, 2019. +[49] 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. +[50] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019. +[51] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. +[52] Julius Plucker. XVii. on a new geometry of space. Philosophical Transactions of the Royal Society of London, (155): 725-791, 1865. +[53] Charles R Qi, Hao Su, Matthias Nießner, Angela Dai, Mengyuan Yan, and Leonidas J Guibas. Volumetric and multi-view cnns for object classification on 3d data. In CVPR, pages 5648-5656, 2016. +[54] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, pages 918-927, 2018. +[55] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. +[56] David E Rumelhart, Geoffrey E Hinton, Ronald J Williams, et al. Learning internal representations by error propagation, 1985. +[57] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In CVPR, pages 6229-6238, 2022. +[58] Mehdi SM Sajjadi, Aravindh Mahendran, Thomas Kipf, Etienne Pot, Daniel Duckworth, Mario Lučić, and Klaus Greff. Rust: Latent neural scene representations from unposed imagery. In CVPR, pages 17297-17306, 2023. +[59] Kyle Sargent, Jing Yu Koh, Han Zhang, Huiwen Chang, Charles Herrmann, Pratul Srinivasan, Jiajun Wu, and Deqing Sun. Vq3d: Learning a 3d-aware generative model onImagenet. In ICCV, pages 4240-4250, 2023. +[60] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. + +[61] Cameron Smith, David Charatan, Ayush Tewari, and Vincent Sitzmann. Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent. arXiv preprint arXiv:2404.15259, 2024. +[62] Randall C Smith and Peter Cheeseman. On the representation and estimation of spatial uncertainty. The International Journal of Robotics Research, 5(4):56-68, 1986. +[63] Noah Snavely. Lecture 15: Structure from motion. https://www.cs.cornell.edu/courses/cs5670/2017sp/lectures/lec15_sfm.pdf, 2017. +[64] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In ECCV, pages 776-794. Springer, 2020. +[65] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In CVPR, pages 4190-4200, 2023. +[66] Shubham Tulsiani, Saurabh Gupta, David F Fouhey, Alexei A Efros, and Jitendra Malik. Factoring shape, pose, and layout from the 2d image of a 3d scene. In CVPR, pages 302-310, 2018. +[67] Hsiao-Yu Fish Tung, Ricson Cheng, and Katerina Fragkiadaki. Learning spatial common sense with geometry-aware recurrent networks. In CVPR, pages 2595-2603, 2019. +[68] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. In European Conference on Computer Vision, pages 197-214. Springer, 2024. +[69] 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. +[70] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In $CVPR$ , pages 4690-4699, 2021. +[71] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. arXiv preprint arXiv:2501.12387, 2025. +[72] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, pages 20697-20709, 2024. +[73] 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 meshes. arXiv preprint arXiv:2404.12385, 2024. +[74] 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. +[75] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. In CVPR, pages 9065-9075, 2023. + +[76] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3d-aware image generation using 2d diffusion models. In ICCV, pages 2383-2393, 2023. +[77] Xinchen Yan, Jimei Yang, Ersin Yumer, Yijie Guo, and Honglak Lee. Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. NIPS, 29, 2016. +[78] Jianing Yang, Alexander Sax, Kevin J Liang, Mikael Henaff, Hao Tang, Ang Cao, Joyce Chai, Franziska Meier, and Matt Feiszli. Fast3r: Towards 3d reconstruction of $1000+$ images in one forward pass. arXiv preprint arXiv:2501.13928, 2025. +[79] 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, pages 10371-10381, 2024. +[80] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. NeurIPS, 37:21875-21911, 2025. +[81] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. +[82] Yanjie Ze, Nicklas Hansen, Yinbo Chen, Mohit Jain, and Xiaolong Wang. Visual reinforcement learning with self-supervised 3d representations. IEEE Robotics and Automation Letters, 8(5):2890-2897, 2023. +[83] 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. +[84] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. +[85] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. arXiv preprint arXiv:2402.14817, 2024. +[86] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. arXiv preprint arXiv:2404.19702, 2024. +[87] Shangzhan Zhang, Jianyuan Wang, Yinghao Xu, Nan Xue, Christian Rupprecht, Xiaowei Zhou, Yujun Shen, and Gordon Wetzstein. Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. arXiv preprint arXiv:2502.12138, 2025. +[88] Tianyuan Zhang, Zhengfei Kuang, Haian Jin, Zexiang Xu, Sai Bi, Hao Tan, He Zhang, Yiwei Hu, Milos Hasan, William T. Freeman, Kai Zhang, and Fujun Luan. Ritolrm: Generative relightable radiance for large reconstruction models, 2024. +[89] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In ECCV, pages 20-37. Springer, 2022. +[90] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851-1858, 2017. + +[91] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. +[92] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. +[93] 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 2410.12781, 2024. \ No newline at end of file diff --git a/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/images.zip b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..51141d13f5a71e6a1d0e8ace0fa798593ecbb274 --- /dev/null +++ b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b2bc4d8d951b46aee252efb55b9cceae777887b7889f22ba9739b929f0785d1 +size 722466 diff --git a/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/layout.json b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9fd5300a12c9df5a6e1427fe4e286c628b4c269f --- /dev/null +++ b/ICCV/2025/RayZer_ A Self-supervised Large View Synthesis Model/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f99363419859d7ec6f1a6c094a1b96eb51fe5c21456c4f549c4d8da3f978092 +size 521156 diff --git a/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/6772646b-a84b-4214-8cc8-8d4a062ae1b7_content_list.json b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/6772646b-a84b-4214-8cc8-8d4a062ae1b7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9c8016d92ff87f8cd23d0a327082a6f6306f62cd --- /dev/null +++ b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/6772646b-a84b-4214-8cc8-8d4a062ae1b7_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c2181447230dd759e1c24d8fac892d45217193b9860ff4fa343453721f3ac2e +size 86572 diff --git a/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/6772646b-a84b-4214-8cc8-8d4a062ae1b7_model.json b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/6772646b-a84b-4214-8cc8-8d4a062ae1b7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..904d4fb6021bb37c91e76adffc00eb6034d378e6 --- /dev/null +++ b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/6772646b-a84b-4214-8cc8-8d4a062ae1b7_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d62c118e5970bceda497d1eb23d2d3b077d11ad978dbd7d59d6470489ac0513f +size 113621 diff --git a/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/6772646b-a84b-4214-8cc8-8d4a062ae1b7_origin.pdf b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/6772646b-a84b-4214-8cc8-8d4a062ae1b7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cad6745f06cf84b00f8a1436131c9a9475047e74 --- /dev/null +++ b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/6772646b-a84b-4214-8cc8-8d4a062ae1b7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca2ddfc81f9721a1327e2df06ffd98a4f22f99b9472bf0fd305e167162d6a04b +size 1461895 diff --git a/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/full.md b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/full.md new file mode 100644 index 0000000000000000000000000000000000000000..88cf03a79759ff170c9d65eb4639e2a15ebde2bd --- /dev/null +++ b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/full.md @@ -0,0 +1,344 @@ +# RayletDF: Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians + +Shenxing Wei † Jinxi Li † Yafei Yang Siyuan Zhou Bo Yang * vLAR Group, The Hong Kong Polytechnic University + +{shenxing.wei, jinxi.li, ya-fei.yang, siyuan.zhou}@connect.polyu.hk, bo.yang@polyu.edu.hk + +![](images/c51ae3ff45f60980d0d45a5e1c3c5a2304482c5dc39674539c0853b908702a18.jpg) +Input Point Cloud / Gaussians + +![](images/4158f3b34f03fcf874e38da578a1ae73ac7abdb322afdb1b15e77d4e2ddb7eb7.jpg) +RayDF +Figure 1. We train all methods on the large-scale ARKitScene dataset and then directly test them on the unseen ScanNet++ dataset. Our RayletDF shows superior generalizability compared to baselines, achieving significantly more accurate 3D scene surface reconstructions. + +![](images/9a5a04abe85fbd1194188b588fcd6d4a355426c92c8777e89139025ae6ce9a14.jpg) +PFGS + +![](images/f6b24298d65b90b8b65d073d9ad8e349c67699c05f976a339a34de4c1016cb25.jpg) +RayletDF (Ours) + +![](images/91e4084b5bc1757e4d7139467886cc73175ebb760c0623df1e3620c7ef60f56e.jpg) +Ground Truth + +# Abstract + +In this paper, we present a generalizable method for 3D surface reconstruction from raw point clouds or pre-estimated 3D Gaussians by 3DGS from RGB images. Unlike existing coordinate-based methods which are often computationally intensive when rendering explicit surfaces, our proposed method, named RayletDF, introduces a new technique called raylet distance field, which aims to directly predict surface points from query rays. Our pipeline consists of three key modules: a raylet feature extractor, a raylet distance field predictor, and a multi-raylet blender. These components work together to extract fine-grained local geometric features, predict raylet distances, and aggregate multiple predictions to reconstruct precise surface points. We extensively evaluate our method on multiple public real-world datasets, demonstrating superior performance in surface reconstruction from point clouds or 3D Gaussians. Most notably, our method achieves exceptional generalization ability, successfully recovering 3D surfaces in a single-forward pass across unseen datasets in testing. Our code and datasets are available at https://github.com/vLAR-group/RayletDF. + +# 1. Introduction + +Learning efficient, accurate, and generalizable 3D surface representations is crucial for many applications in mixed reality, embodied AI, and graphics. Given RGB/D images and/or point clouds, a series of 3D representations has + +been developed to recover 3D geometry, including occupancy fields (OF) [38], un/signed distance fields (U/SDF) [13, 44], radiance fields (NeRF) [39], and vector fields (VF) [70]. While achieving excellent results in 3D reconstruction, these coordinate-based methods and their variants typically demand dense network evaluations to obtain explicit surfaces, thus being computationally heavy. + +Recently, the point-based method 3D Gaussian Splating (3DGS) [30] has emerged as an appealing alternative to those coordinate-based methods, thanks to its impressive real-time performance in synthesizing high-fidelity RGB images at high resolutions. However, it still falls short in rendering high-quality depth views, due to its failure in capturing fine-grained surface geometry, though various constraints such as depth priors [33, 62], local smoothness or planarity regularization [24, 63] can be added to enhance the quality of surfaces recovered. + +On the other hand, a line of recent ray-based methods such as DRDF [32], PRIF [20], DDFs [1], Pointintersect [8], and RayDF [36] have demonstrated excellent performance in representing intricate surfaces based on light rays. Nevertheless, due to the limitation of existing ray parametrizations such as Plucker and spherical coordinates, they are often limited to recovering object-level surfaces and require per-scene training, lacking the desired generalizability to infer diverse 3D scene structures in a single forward pass. + +In this paper, we present a generalizable 3D surface representation pipeline to accurately recover 3D geometry. The core module of our pipeline is a new technique called raylet distance field, with inspiration from ray-based methods. + +The term raylet is defined as a unit segment of a light ray + +![](images/1713e6bd04f19f2fd2a1b49fc13561096fc5365ed62299f90e30551b60aa3ee3.jpg) +Figure 2. An illustration of raylets and raylet distances. + +and its starting point is sampled (or located) near the surface of a shape. The raylet distance is defined as the signed distance between the surface hit point and the raylet starting point. As illustrated in Figure 2, given three light rays $\{r_1, r_2, r_3\}$ emitting from light sources, we sample three raylets $\{l_1, l_2, l_3\}$ near the surface $S$ , where raylet distance $d_1$ is assigned positive as its surface hit point $p$ locates further away from the raylet starting point $p_{l_1}$ , whereas $\{d_2, d_3\}$ are assigned negative accordingly. Note that, given a single light ray, we can sample numerous raylets on both sides of the surface hit point. Intuitively, such definitions of raylets and raylet distances primarily focus on nuanced local patterns of surfaces, and these local geometric patterns are usually generalizable across diverse shapes. + +With this merit of raylets, we simply formulate the problem of generalizable 3D surface reconstruction into learning raylet distance fields from visual observations. In particular, our pipeline comprises three modules: 1) a raylet feature extractor to extract geometry features from an input 3D scene for a query raylet; 2) a raylet distance field to predict the signed raylet-surface distance value for the input query raylet; and 3) a multi-raylet blender to aggregate multiple predicted raylet distances along every single ray, ultimately recovering the accurate surface point. + +These three modules together allow us to learn generalizable ray-based surface representations from large-scale training data with ground truth depth scans as supervision signals. Once the pipeline is well trained, it can be used to infer high-quality 3D scene surfaces on unseen datasets from arbitrary query viewing angles in a single forward pass. Notably, our pipeline can not only train and evaluate on point clouds, but is also amenable to any scene represented by 3D Gaussians recovered by the popular 3DGS [30] from RGB images. Our pipeline is called RayletDF and Figure 1 shows qualitative results on diverse indoor scenes. Our contributions are: + +- We propose a generic pipeline for explicit 3D surface reconstruction from either point clouds or 3D Gaussians. +- We introduce a new raylet distance field followed by a blender to learn generalizable surface patterns. +- We show superior accuracy in surface reconstruction across multiple datasets, clearly outperforming baselines, especially in terms of generalizability to new datasets. +- We benchmark generalizable surface reconstruction on 3D Gaussians and will release 7770 3D scene Gaussians trained on ScanNet/++, ARKit, and MultiScan datasets. + +# 2. Related Works + +Classical approaches to recovering 3D surfaces from images mainly include SfM [43] and SLAM [6] systems to obtain sparse point clouds, optionally followed by various global or local smoothness priors to recover continuous surfaces, but the reconstructed shapes usually lack fine details. Early learning methods to model explicit 3D structures mainly include voxel grids [14, 67, 68], point clouds [17], octree [54], meshes [29] based pipelines, but the fidelity of these discrete shape representations is often limited by the spatial resolutions and memory footprint. A comprehensive survey of these methods can be found in [4, 25]. + +Coordinate-based Methods for Surface Reconstruction: To avoid the discretization issue of classical explicit 3D representations, a series of implicit representations has been developed to use simple MLPs to recover continuous 3D shapes [21]. These representations, including OF [12, 38], SDF [44], UDF [13, 58], VF [47, 70], and NeRF [39, 56], typically take 3D coordinates as input and predict various properties of input coordinates. While demonstrating exceptional accuracy in recovering continuous 3D surfaces and/or rendering 2D views, these methods and their variants [55, 61, 65] demand dense 3D point sampling and network evaluations to regress explicit surface points given any query view, thus being computationally heavy. In this paper, our RayletDF learns efficient raylet fields and does not require dense sampling to regress surface points. + +Point-based Methods for Surface Reconstruction: Point-based methods aim to recover surfaces by rendering disconnected geometry samples, including the simplest point sample [23] and circular/ elliptic discs/ ellipsoids/ surfels sample rendering methods [45, 48, 77]. Recently, a new point-based method 3DGS [30] has been introduced to represent the scene by a set of 3D Gaussians with various properties such as position, covariance, and color. While achieving real-time rendering of 2D views thanks to the tile-based rasterizer, 3DGS falls short in synthesizing highly accurate depth views as it is hard to capture fine-grained surfaces, though advanced techniques [9, 11, 15, 18, 24, 26, 27, 33, 42, 50, 63, 64, 66, 73-76] can mitigate this issue but at the expense of generalization across various scenes. Meanwhile, researchers also introduce fast winding number to balance the efficiency and accuracy [2, 10]. These approaches achieve better geometry fidelity and convergence speed but still need per-scene optimization and point normals. In this paper, our introduced raylet distance field and the blender aim to learn nuanced local surface features, thus being accurate and generalizable. + +Ray-based Methods for Surface Reconstruction: A line of works formulates 3D shapes as ray-based neural functions. Existing ray-based methods simply take individual rays as input and estimate either RGB values such as LFN [49], NeuLF [34] and others [7, 19, 41, 51], or surface + +![](images/65f05fc5909da9a4c0d60991b119555cd6bd59cd6cad6a5460e839d8eb765504.jpg) +Figure 3. Overview of our proposed pipeline. The leftmost block shows the raylet feature extractor module, the middle block shows the raylet distance field module, and the rightmost block shows our multi-raylet blender module. + +![](images/b274e860ad472283c2776c2420d750fcdb739a42f835087818411c443e857520.jpg) + +![](images/03018a1c71d70561ecf973eb385c0ecc745a68c0be64995813ed928516bb1495.jpg) + +points including PRIF [20], DDFs [1], RayDF [36], and others [28, 32, 52, 53, 72]. Although these methods have shown excellent efficiency and accuracy in rendering surfaces, they can typically represent object-level shapes and lack generality due to the limitation of light ray representations. In this paper, instead of using global light rays to represent surfaces, we use raylets to recover geometry. + +Image-based Depth Estimation: With the advancement of techniques for novel view RGB synthesis, existing methods [5, 46, 69] for image-based depth estimation can also be used to recover per-view depth scans, followed by depth fusion for 3D surface reconstruction. However, such a pipeline inherently lacks consistent depth alignment and geometry consistency priors. Therefore, the recovered 3D structures are inferior in quality. + +# 3. RayletDF + +# 3.1. Overview + +Our RayletDF models 3D surface as a neural network $f$ . It takes a scene volumetric data $P$ and a query raylet $l$ as input, directly predicting the corresponding raylet distance $d_{l}$ , i.e., the distance between surface hit point and raylet starting point as shown in Figure 2. The scene volumetric data can be a sparse point cloud or on-the-shelf 3D Gaussians recovered by the popular 3DGS [30] from RGBs, both denoted by $P \in \mathcal{R}^{N \times C_0}$ for simplicity, where $N$ represents the total number of surface points or Gaussian centers and $C_0$ represents the number of per-point features such as point/Gaussian center $xyz$ optionally with other properties like colors or opacity. The query raylet $l \in \mathcal{R}^{1 \times 6}$ is parameterized by its starting point $xyz$ and a unit vector representing ray direction. Formally, it is defined as below: + +$$ +d _ {l} = f (\boldsymbol {P}, \boldsymbol {l}) \quad \boldsymbol {P} \in \mathcal {R} ^ {N \times C _ {0}}, \quad \boldsymbol {l} \in \mathcal {R} ^ {1 \times 6} \tag {1} +$$ + +Since our method aims at representing accurate and generalizable 3D surface representations, the pipeline is designed to be trained on a large number of 3D scenes with depth images as supervision signals. Once well-trained, it is expected to be applied to various unseen datasets, directly + +inferring high-quality depth maps at arbitrary query viewing angles. As shown in Figure 3, our pipeline has three components clarified in the following Sections 3.2&3.3&3.4. + +# 3.2. Raylet Feature Extractor + +This module is designed to learn local features from an input scene volumetric data point $\pmb{P}$ for any specific query raylet $l$ . As shown in the leftmost block of Figure 3, given an input scene $\pmb{P}$ , we simply adopt the powerful SparseConv architecture [22] without any pretraining step as the backbone network, obtaining per-point features $\pmb{F} \in \mathcal{R}^{N \times C}$ where the embedding length $C$ is predefined as 32. Implementation details are in Appendix 5.5. + +Having the input scene $\pmb{P}$ and its per-point features $\pmb{F}$ , we need to extract the local features for a specific input query raylet $l$ whose starting point is denoted by $p_l$ and ray direction by a unit vector $\pmb{u}_l$ . In particular, as shown in Figure 3, for the query raylet $l$ , we identify the top $K$ nearest points $\{\pmb{p}_l^1 \dots \pmb{p}_l^k \dots \pmb{p}_l^K\}$ in the input scene $\pmb{P}$ with regard to the raylet starting point $p_l$ . The simple K-nearest neighbors (KNN) algorithm is adopted based on the point-wise Euclidean distances. After that, we gather the information, denoted by $\hat{\pmb{f}}_l^k$ , for the $k^{th}$ neighboring point as follows: + +$$ +\hat {\boldsymbol {f}} _ {l} ^ {k} = \left(\boldsymbol {p} _ {l} ^ {k} \oplus \frac {\left(\boldsymbol {p} _ {l} ^ {k} - \boldsymbol {p} _ {l}\right)}{\left| \left| \boldsymbol {p} _ {l} ^ {k} - \boldsymbol {p} _ {l} \right| \right|} \oplus \left\| \boldsymbol {p} _ {l} ^ {k} - \boldsymbol {p} _ {l} \right\|\right) \oplus \boldsymbol {f} _ {l} ^ {k} \tag {2} +$$ + +where $\pmb{p}_l$ and $\pmb{p}_l^k$ are the xyz positions of points, $\oplus$ is the concatenation operation, $||\cdot||$ calculates the Euclidean distance between two points, and $\pmb{f}_l^k$ is the feature vector extracted from per-point features $\pmb{F}$ . + +Lastly, we stack all information of the total $K$ neighboring points into a local feature vector $\pmb{f}_l$ for the query raylet $\pmb{l}$ for simplicity as follows, though other methods such as max/mean and attentional pooling are also applicable. + +$$ +\boldsymbol {f} _ {l} = \hat {\boldsymbol {f}} _ {l} ^ {1} \oplus \dots \hat {\boldsymbol {f}} _ {l} ^ {k} \dots \oplus \hat {\boldsymbol {f}} _ {l} ^ {K} \tag {3} +$$ + +Overall, our key insight of this module is that the features extracted for a query raylet should retain local geometrical patterns near surfaces, allowing the learned raylet distance representations to be generalizable across diverse scenes. + +# 3.3. Raylet Distance Field + +This module aims to learn accurate raylet surface distance value for any specific query raylet $l$ . As illustrated in the middle block of Figure 3, we feed the query raylet, i.e., the starting point $\pmb{p}_l$ and direction $\pmb{u}_l$ , and its feature vector $\pmb{f}_l$ (ref to Section 3.2) into MLPs, directly predicting the corresponding raylet surface distance value $d_l$ . + +To improve the generalizability and robustness of raylet distance predictions, we further predict a confidence score $s_l$ alongside the distance value $d_l$ . This score is basically designed to weight the prediction and blend multiple raylets sampled along a single ray which will be detailed in Section 3.4. Formally, the raylet distance field is defined as: + +$$ +\left(d _ {l}, s _ {l}\right) = M L P s \left(\boldsymbol {p} _ {l} \oplus \boldsymbol {u} _ {l} \oplus \boldsymbol {f} _ {l}\right) \tag {4} +$$ + +For fast computation, the MLPs consist of 8 layers and each layer has 256 hidden neurons, though more sophisticated neural layers can be applied as well. Since the predicted distance value could be positive or negative, there is no non-linear activation function added at the output layer. More details are in Appendix 5.6. + +Our design of this raylet distance field is particularly simple in concept, but has a clear advantage over the prevailing coordinate-based surface representations such as OF, SDF, and UDF: our method takes raylets as input, without requiring sampling dense coordinates along a single ray, thus being efficient to regress surface points. + +# 3.4. Multi-raylet Blender + +The ultimate goal of our pipeline is to infer the surface point for any given light ray $r$ , i.e., the distance $D$ between the query camera center $p_{cam}$ and the surface hit point along the ray. Since our raylet distance field learns representations for raylets instead of entire rays, this allows us to sample multiple raylets along each single ray followed by blending multiple estimations, potentially improving the generalizability and robustness of our method. + +To this end, as shown in the rightmost block of Figure 3, we opt for sampling $T$ raylets $\{l_1 \cdots l_t \cdots l_T\}$ near the surface along a single ray $r$ . These $T$ raylets have the same unit vector of ray direction, but with different starting point positions, denoted by $\{\pmb{p}_{l_1} \cdots \pmb{p}_{l_t} \cdots \pmb{p}_{l_T}\}$ . Naturally, we feed these $T$ raylets into our raylet distance field in parallel, obtaining the corresponding distances and scores: $\{(d_{l_1}, s_{l_1}) \cdots (d_{l_t}, s_{l_t}) \cdots (d_{l_T}, s_{l_T})\}$ . After that, we blend all these predictions and estimate the distance between camera center $\pmb{p}_{cam}$ and surface hit point as follows: + +$$ +D = \sum_ {t = 1} ^ {T} \hat {s} _ {l _ {t}} \left(\left\| \boldsymbol {p} _ {c a m} - \boldsymbol {p} _ {l _ {t}} \right\| + d _ {l _ {T}}\right), \hat {s} _ {l _ {t}} = \frac {e ^ {s _ {l t}}}{\sum_ {t = 1} ^ {T} e ^ {s _ {l t}}} (5) +$$ + +where $\hat{s}_{l_t}$ is a normalized score. $\pmb{p}_{cam}$ and $\pmb{p}_{l_t}$ are always known during sampling in training and test. + +![](images/8a9ab5759e9a0af37ffb72c2686598cb52dc6afa017dbae7971b909f6fc5ad35.jpg) +Figure 4. Raylet sampling on virtual balls of point clouds. + +![](images/9460d1a5e86e26a20574c3a99f741545cc307b50ee88e89dcd5d7fb6ae9cff5d.jpg) + +The prediction $D$ is fully supervised using $\ell_1$ loss, and the ground truth distance values are converted from the available depth images in training set. + +# 3.5. Sampling Raylets for Training and Test + +Given a specific 3D scene $P$ as input, if it is a raw point cloud, for a specific query ray $r$ , we sample multiple raylets for both training or test in the following steps: + +- Step 1: For every point $\pmb{p}$ in the input scene point cloud $\pmb{P}$ , we search its nearest point $\pmb{p'}$ and identify the Euclidean distance $||\pmb{p} - \pmb{p'}||$ as a radius of the ball centering at point $\pmb{p}$ . In this way, every point in $\pmb{P}$ is assigned a virtual ball. Intuitively, the underlying continuous 3D scene surface, which is unknown and yet to be estimated, should be bounded by the union space of all virtual spheres, as illustrated in the left block of Figure 4. +- Step 2: as illustrated in the right block of Figure 4, given the query ray $r$ with a camera center $p_{cam}$ , it shoots through multiple virtual balls. We then project the center point of each intersected ball onto the ray, obtaining a perpendicular distance and the foot of the perpendicular. After that, we select $T$ foot points, belonging to intersected virtual balls with top $T$ shortest perpendicular distances, as the starting points of $T$ raylets. If the number of intersected balls is fewer than $T$ , we just keep the actual number of raylets. Note that, if there is no ball intersected, meaning that the ray shoots outside the target 3D surface, the ray is discarded in both training and test. + +If the input 3D scene $P$ is a set of 3D Gaussians recovered by 3DGS [30] from RGBs, we follow the technique [31, 74] to calculate an intersection point between the query ray and each Gaussian, and then we select $T$ intersection points, belonging to the Gaussians with top $T$ contributions based on alpha blending, as the starting points of $T$ raylets. + +For efficiency, we follow the tile-based splitting strategy in 3DGS [30] to calculate intersection points on both point clouds and Gaussians. More details are in Appendix 5.7. + +# 3.6. Surface Normal Derivation + +Similar to RayDF, our method also has a nice property of deriving a closed-form expression for a normal vector, thanks to the nature of our ray-based formulation. In particular, given an input raylet $\boldsymbol{l} = (\boldsymbol{p}_l, \boldsymbol{u}_l)$ and its estimated raylet distance value $d_l$ from the network, the corresponding normal vector $\boldsymbol{n}_l$ is provided in Appendix 5.8. + +This normal vector can be used in an additional loss to regularize surface points, or acts as a threshold to remove outlier predictions, which are left for future exploration. + +Table 1. Quantitative results of all methods for ray surface distance estimation. ${}^{ \dagger }$ indicates the scale is aligned with ground truth depth. + +
test on → ARKitScenestest on → ScanNet/ScanNet++test on → MultiScan
ADE↓RMSE↓Abs-Rel↓Sq-Rel↓δ↑ADE↓RMSE↓Abs-Rel↓Sq-Rel↓δ↑ADE↓RMSE↓Abs-Rel↓Sq-Rel↓δ↑
3DGS [30]per-scene train / test0.2680.3790.1730.0980.7070.3210.4610.1520.1490.7610.4310.5320.3220.2250.402
GOF [74]0.3180.4690.2130.1590.6590.3200.5070.1550.2140.7750.5710.7000.4040.3730.363
PGSR [9]0.2190.3760.1680.1830.7930.2020.5360.0990.3790.8770.3150.4300.2850.5160.567
DepthAnythingV2† [69]0.2060.2940.1440.0930.8600.1680.2480.1180.0660.8870.2280.3030.2050.1170.775
Depth-Pro† [5]0.2940.4030.2020.2420.7500.2200.3190.1200.1060.8610.2800.3630.2340.1570.659
MVSGaussian [35]train on ARKitScenes0.5840.7550.3920.4160.4250.5920.7970.3950.3760.5380.5430.6760.4440.4340.363
Pointersect [8]0.2860.3970.2090.1350.7490.3660.5130.1900.1510.7330.2660.3510.2170.1170.671
RayDF [36]0.1830.3030.1390.0880.8460.2270.3630.1310.1230.8380.3260.4250.2870.2020.583
PFGS [59]0.2640.3860.1710.0980.7190.3030.4550.1450.1450.7790.5360.6530.4420.3750.316
RayletDF (Ours)0.1150.2180.0820.0470.9280.1750.3200.0850.0980.8940.2160.3110.1820.1070.739
MVSGaussian [35]train on ScanNet/+0.6000.7740.4040.4480.4130.5910.7870.2950.3670.5280.5480.6810.4480.4450.363
Pointersect [8]0.3280.4620.1960.1420.6650.4330.6040.1980.1990.6580.4040.5040.2960.1980.449
RayDF [36]0.5870.7040.3480.3020.3170.2020.3370.1140.8180.8710.6040.6900.4470.4670.233
PFGS [59]0.2740.3990.1770.1040.7040.2860.4330.1370.1420.7980.5340.6690.4482.4490.327
RayletDF (Ours)0.1320.2410.0940.0530.9060.1450.2760.0720.0790.9220.2590.3530.2360.1480.662
+ +# 4. Experiments + +Datasets: Our method is evaluated on four real-world datasets based on the available train/test splits: 1) ScanNet [16] consisting of 1201 and 100 scenes for training and test; 2) ScanNet++ [71] comprising 865 and 50 scenes for training and test; 3) ARKitScenes [3] with 4498 and 549 scenes for training and test, which is the largest real-world indoor scene dataset captured by mobile RGBD sensor form Apple LiDAR scanner; 4) MultiScan [37] only consisting of about 200 indoor scenes in total captured by smartphones and tablets, where all scenes are used as a test set in our experiments. Due to the domain overlap, we merge ScanNet and ScanNet++, getting a joint train set and a joint test set. + +Baselines: We choose 5 representative groups of methods as our baselines: 1) the state-of-the-art per-scene optimization based 3D Gaussians splatting methods GOF [74] and PGSR [9] particularly designed for high-fidelity surface reconstruction; 2) the latest feed-forward generalizable 3D Gaussian splatting method MVSGaussian [35]; 3) the state-of-the-art point cloud rendering method PFGS [59]; 4) the state-of-the-art ray-based methods Pointersect [8] and RayDF [36] which is adapted as a generalizable version where the ray feature extractor is exactly the same as ours; 5) the latest image-based depth estimation methods DepthAnythingV2 [69] and DepthPro [5]. + +Metrics: Primarily, we evaluate the surface reconstruction by measuring the error between predicted ray-surface distance and ground truth distance. We report the per ray-surface absolute distance error (ADE) in meters across all test images, and other four commonly used metrics [57] including Root Mean Squared Error (RMSE), the Absolute Relative Distance (Abs Rel), the Squared Relative Distance (Sq Rel), and the Threshold accuracy ( $\delta \leq 1.25$ ) which measures the percentage of predicted value within a certain threshold $\delta$ of ground truth. In addition, we also evaluate the reconstructed 3D meshes, reporting Ac + +curacy, Completion, Precision, Recall, Chamfer-L1 distance, Normal Consistency, and F-scores with a threshold of 5cm. All details are in Appendix 5.1&5.2&5.3&5.4. + +# 4.1. Evaluation on 3D Gaussians + +In this setting, our method takes 3D Gaussians as input and predicts raylet distances for query rays. To prepare for this experiment, for every 3D scene in ScanNet/ScanNet++, ARKitScenes, and MultiScan, we first use the vanilla 3DGS to learn per-scene 3D Gaussians from RGB images in an offline fashion, and then evaluate our method in the following two groups of experiments. + +- Group 1: We train our method from scratch on the joint train set of ScanNet/ScanNet++. After that, we evaluate the trained model on the test sets of ScanNet/ScanNet++, ARKitScenes, and MultiScan. +- Group 2: We train our method from scratch on the train set of ARKitScenes, and then test it on the test sets of ScanNet/ScanNet++, ARKitScenes, and MultiScan. + +For a fair comparison, the four feed-forward generalizable baselines MVSGaussian [35], Pointintersect [8], RayDF [36], and PFGS [59] are all trained with the same amount of training data and supervision signals as our method. For a reference, the per-scene methods 3DGS [30], GOF [74] and PGSR [9] directly render depth images from their optimized 3D Gaussians on test set respectively. For the image-based depth estimation methods, we first render test view RGBs from 3D Gaussians and then feed them into a depth estimator, obtaining depth results. Lastly, we use ground truth depth values to align the scale of estimated depths, which is strongly in favor of these baselines. More details of experiments are in Appendix 5.4. + +Results & Analysis: Table 1 compares the quantitative results of all methods for estimating distance values of all query rays from test views. Table 2 compares the quantitative results of 3D meshes reconstructed by TSDF or Marching Tetrahedra (MT) from the rendered depth views for the + +Table 2. Quantitative results of 3D meshes reconstructed from estimated ray surface distances. Our model is trained on Scannet/++ dataset. + +
test on → ARKitScenestest on → ScanNet/ScanNet++
Acc.↓Comp.↓Pre.↑Rec.↑C-L1↓NC↑F1↑Acc.↓Comp.↓Pre.↑Rec.↑C-L1↓NC↑F1↑
3DGS [30]TSDF0.1110.2190.3050.2470.1650.6220.2680.1750.1600.2740.2900.1680.6500.270
GOF [74]MT0.1350.5750.3350.2940.3550.5710.3140.1650.1730.3720.3850.1690.6320.368
PGSR [9]TSDF0.1530.1880.4370.3440.1710.6910.3780.1130.0880.5270.5430.1000.7460.532
RayletDF (Ours)TSDF0.0510.0830.6710.6090.0670.6890.6330.1510.0660.5270.6330.1090.7050.566
+ +Table 3. Quantitative results of estimated ray-surface distances from 3D point clouds. + +
test on → ARKitScenestest on → ScanNet/ScanNet++test on → MultiScan
ADE↓RMSE↓Abs-Rel↓Sq-Rel↓δ↑ADE↓RMSE↓Abs-Rel↓Sq-Rel↓δ↑ADE↓RMSE↓Abs-Rel↓Sq-Rel↓δ↑
Point intersect [8]train on ARKitScenes0.3350.4560.1980.1320.6430.3890.5530.1770.1480.6800.2540.3300.2010.0930.637
RayDF [36]0.1660.2810.1120.0740.8760.1860.3210.0940.0550.8830.1540.2340.1340.0620.822
RayletDF (Ours)0.0880.2050.0660.0600.9490.1070.2690.0510.0490.9430.0670.1490.0640.0370.943
Point intersect [8]train on ScanNet/+0.2990.3980.2090.1410.6910.3440.4770.1810.1330.7080.2330.2890.2250.1030.656
RayDF [36]0.2890.4070.1890.1180.6990.1610.2910.0860.0490.8930.3490.4290.3130.2590.561
RayletDF (Ours)0.0960.2170.0740.0770.9460.0930.2340.0470.0400.9490.1300.2290.1310.2160.914
+ +three excellent per-scene Gaussian based methods and our method. Note that, in the original paper, GOF uses MT for higher quality surface reconstruction. Figure 5 shows qualitative results. From the results, we can see that: + +- When training/testing on ARKitScenes, ScanNet/ ScanNet++ datasets in domain, our method achieves the best accuracy, outperforming the second best method RayDF by large margins, with 0.115 vs 0.183 meters on ARKitScenes and 0.145 vs 0.202 meters on ScanNet/ScanNet++ on the key metric ADE. This means that our newly introduced raylet distance field has its clear advantage over existing ray-based representations. +- When evaluating all methods across new datasets, our method demonstrates superior generalizability on unseen datasets, clearly surpassing all other baselines usually by more than 0.10 meters in accuracy. This performance gap is much greater than the in-domain reconstruction results. Essentially, this is because our learned raylet distance representations capture the local surface geometric patterns which tend to be generalizable at various scenes. +- The per-scene optimization based 3D Gaussian methods and the depth estimation methods can achieve satisfactory results, but still being inferior than our method. + +# 4.2. Evaluation on Point Clouds + +In this setting, our method takes 3D point clouds as input and predicts raylet distances for query rays. To prepare for this experiment, for every 3D scene in ScanNet/ScanNet++, ARKitScenes, and MultiScan, we uniformly sample 10k points from the provided 3D scene mesh. Same as Section 4.1, we also conduct two groups of experiments. For a fair comparison, we also evaluate the top-performing ray-based methods Pointersect and RayDF with exactly the same training data and supervision signals as ours. + +Results & Analysis: Table 3 compares the quantitative results and Figure 6 shows the qualitative results. We can see that, similar to the analysis in Section 4.1, our method demonstrates clearly better reconstruction accuracy on the in-domain datasets. Most notably, we achieve an outstanding generalization ability when evaluating on unseen datasets, clearly surpassing two strong baselines. For example, when trained on ARKitScenes or ScanNet/ScanNet++, our method achieves \{0.067, 0.130\} meters in ADE on the novel MultiScan dataset respectively, while the baselines often have more than 0.20 meters errors. Basically, our strong generalization ability on point clouds is also attributed to the effectiveness of our virtual ball based raylet sampling strategy designed in Section 3.5. + +# 4.3. Evaluation on Raylet Sampling in Testing + +Regarding our designed multi-raylet blender in Section 3.4, the single hyperparameter $T$ of this module can be different in training and test phase, allowing flexibility during evaluation on various datasets. Intuitively, once our model is well-trained on a dataset, in testing, the more number of raylet samples along every single ray, we may obtain more accurate surface estimations. + +We further conduct experiments to validate the flexibility of raylet sampling in test phase. In particular, we use our model well-trained on ScanNet/ScanNet++ where $T$ is chosen as 5 during training, and then directly evaluate on the test splits of ScanNet/ScanNet++, ARKitScenes, and the whole MultiScan dataset, where $T$ is chosen as $\{1, 5, 10, 20\}$ respectively. + +Tables 4&5&6 show the quantitative results. It can be seen that, the more raylets sampled along every single ray during testing, we generally obtain more accurate surface reconstruction. This is particularly obvious when + +the trained model is evaluated on unseen datasets as shown in Tables 5&6. Primarily, the is because if more raylets are sampled along every single light ray, our multi-raylet blender tends to vote a more accurate surface point, instead of being affected by outlier predictions among all estimations. This validates the generalizability and robustness of our simple design. More results are in Appendix 5.10. + +Table 4. Qualitative results for different number of raylet samples per ray when evaluating on the test split of ScanNet/ScanNet++. + +
(train samples: T = 5)ADE↓RMSE↓Abs-Rel↓Sq-Rel↓δ↑
test samples: T = 10.1770.3370.0870.1900.896
test samples: T = 50.1450.2760.0720.0790.922
test samples: T = 100.1400.2680.0700.0700.927
test samples: T = 200.1410.2710.0710.0730.925
+ +Table 5. Qualitative results for different number of raylet samples per ray when evaluating on the test split of ARKitScenes. + +
(train samples: T = 5)ADE↓RMSE↓Abs-Rel↓Sq-Rel↓δ↑
test samples: T = 10.1610.2870.1130.0760.876
test samples: T = 50.1320.2410.0940.0530.906
test samples: T = 100.1270.2340.0920.0520.912
test samples: T = 200.1240.2300.0900.0520.917
+ +Table 6. Qualitative results for different number of raylet samples per ray when evaluating on the whole MultiScan dataset. + +
(train samples: T = 5)ADE↓RMSE↓Abs-Rel↓Sq-Rel↓δ↑
test samples: T = 10.2910.3930.2610.1770.619
test samples: T = 50.2590.3530.2360.1480.662
test samples: T = 100.2520.3460.2340.1480.674
test samples: T = 200.2470.3470.2340.1530.687
+ +# 4.4. Ablations + +Our pipeline has three major modules, 1) Raylet Feature Extractor, 2) Raylet Distance Field, and 3) Multi-raylet Blender. To evaluate the effectiveness of each module and the sensitivity of hyperparameters, we conduct the following ablations on the merged ScanNet/ScanNet++ dataset, and the input to our method is choose as 3D Gaussians. + +(1)~(3) Including different information of neighboring points. Regarding the raylet feature extractor, in Equation 2, for each neighboring point, we choose to include: (1) only the $xyz$ (i.e., $p_l^k$ ) of a neighboring point, (2) only the relative position information (i.e., $\left(\frac{(p_l^k - p_l)}{||p_l^k - p_l||} \right) \oplus ||p_l^k - p_l||$ ), (3) both $xyz$ and the relative position information. +(4)~(7) Choosing different number of neighboring points $K$ . For the raylet feature extractor, in Equation 3, the hyperparameter $K$ is chosen as $\{1,5,10,20\}$ . This aims to evaluate the influence of local neighborhood on raylet features. In all main experiments, we choose $K$ as 5. +(8)~(11) Sampling different number of raylets. For the multi-raylet blender, along each light ray, a total of $T$ raylets are sampled near the surface. In this ablation, the hyperparameter $T$ is chosen as $\{1, 5, 10, 20\}$ . In all main experiments, we choose $T$ as 5. + +(12) Removing the prediction of confidence score $s_l$ . Regarding the raylet distance field in Equation 4, we remove the confidence score, and the subsequent multi-raylet blender turns to simply average out multiple raylet distance predictions, i.e., $\hat{s}_{l_t}$ is $1 / T$ in Equation 5. +(13) Using alpha blending to aggregate multi-raylet predictions. For the multi-raylet blender, we choose to regard the confidence score $s_l$ in Equation 4 as opacity, and then use alpha blending to get the final prediction. +(14) Using sigmoid to normalize the confidence score $s_l$ . For the multi-raylet blender in Equation 5, we turn to use sigmoid function to independently normalize confidence scores, meaning that multiple raylets along a single ray will be trained independently. + +Table 7 shows the ablation results. We can see that: 1) The greatest impact is caused by the decrease of raylet sample $T$ to be just 1 in training. Basically, this means that our raylet distance field degenerates to the existing ray-based method such as RayDF, which is ineffective to learn local surface patterns. 2) The second greatest impact is caused by the removal of confidence score $s_l$ in Equation 4. Basically, this is closely related to our design of multi-raylet blender as ablated in (12)~(14). Without a suitable multiraylet blending strategy like ours, the accuracy of estimated surface drops. 3) Different number of neighboring points $K$ and choices of their features included could be helpful as shown in (1)~(7), but they are less important overall. + +Table 7. Results of all ablated models on ScanNet/ScanNet++. The main settings are bolded. + +
ADE↓RMSE↓Abs-Rel↓Sq-Rel↓δ↑
(1)Only xyz0.1480.2780.0730.0740.921
(2)Only (p^k_i - p^l_i) / ||p^k_i - p^l_i|| ⊕ ||p^k_i - p^l_i||0.1460.2750.0720.0780.921
(3)Both0.1450.2760.0720.0790.922
(4)K = 10.1510.2790.0770.0770.919
(5)K = 50.1450.2760.0720.0790.922
(6)K = 100.1500.2800.0740.0740.920
(7)K = 200.1520.2850.0750.0730.915
(8)T = 10.1740.3260.0850.1440.897
(9)T = 50.1450.2760.0720.0790.922
(10)T = 100.1410.2740.0700.0660.925
(11)T = 200.1410.2710.0700.0700.924
(12)Removing score sl0.1700.3020.0840.0970.907
(13)Using alpha blending0.1560.2900.0750.0730.914
(14)Using sigmoid0.1640.2870.0750.0780.915
(15)The Full Pipeline (RayletDF)0.1450.2760.0720.0790.922
+ +# 5. Conclusion + +In this paper, we present a generalizable 3D surface reconstruction method. By introducing a new technique of raylet distance field, our pipeline accurately captures intricate local surface patterns from both point clouds and 3D Gaussians, demonstrating superior performance across diverse real-world 3D scene datasets. Remarkably, thanks to the learned local raylet features, it exhibits excellent generalizability to new and unseen scenes in testing, while all baselines fail to do so. + +![](images/7fb622990dd6131c1ecd8a8f1a069f76ded6f80aa5656a443b8188faa649f7c2.jpg) +Figure 5. Qualitative results of our method and baselines for 3D surface reconstruction on multiple datasets. All methods are trained on pre-estimated 3D Gaussians. + +![](images/a7324d4f02215804926808ae91b464392e3ed0a745e0752330ee016d695d94fa.jpg) + +![](images/5b2b3c8376646dd336fd2c93a8d923972cd2a854ed8e060afb5b43424548c0be.jpg) +Figure 6. Qualitative results of our method and baselines for 3D surface reconstruction on multiple datasets. All methods are trained on point clouds. + +![](images/1dbaa521b2bd567c543e5275f0fe389e3ce2ee863025a374863c1efcf55a57c5.jpg) + +Acknowledgment: This work was supported in part by Research Grants Council of Hong Kong under Grants 15225522 & 25207822 & 15219125, in part by Otto Poon Charitable Foundation Smart Cities Research Institute (8-CDCQ), in part by Research Centre for Unmanned Autonomous Systems (1-CE3D), The Hong Kong Polytechnic University. + +# References + +[1] Tristan Aumentado-Armstrong, Stavros Tsogkas, Sven Dickinson, and Allan Jepson. Representing 3D Shapes with Probabilistic Directed Distance Fields. CVPR, 2022. 1, 3 +[2] Gavin Barill, Neil G Dickson, Ryan Schmidt, David IW Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. ACM Transactions on Graphics (TOG), 2018. 2 +[3] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARKitScenes: A Diverse Real-World Dataset For 3D Indoor Scene Understanding Using Mobile RGB-D Data. NeurIPS, 2021. 5 +[4] M Berger, A Tagliasacchi, L M Seversky, P Alliez, J A Levine, A Sharf, and C T Silva. A Survey of Surface Reconstruction from Point Clouds. Computer Graphics Forum, 2017. 2 +[5] 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. ICLR, 2025. 3, 5 +[6] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, Jose Neira, Ian D. Reid, and John J. Leonard. Past, Present, and Future of Simultaneous Localization and Mapping: Towards the Robust-Perception Age. *TRO*, 2016. 2 +[7] Junli Cao, Huan Wang, Pavlo Chemerys, Vladislav Shakhrai, Ju Hu, Yun Fu, Denys Makoviichuk, Sergey Tulyakov, and Jian Ren. Real-Time Neural Light Field on Mobile Devices. CVPR, 2023. 2 +[8] Jen-Hao Rick Chang, Wei-Yu Chen, Anurag Ranjan, Kwang Moo Yi, and Oncel Tuzel. Point intersect: Neural Rendering with Cloud-Ray Intersection. CVPR, 2023. 1, 5, 6, 4, 7 +[9] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction. TVCG, 2024. 2, 5, 6 +[10] Hanyu Chen, Bailey Miller, and Ioannis Gkioulekas. 3d reconstruction with fast dipole sums. ACM Transactions on Graphics (TOG), 2024. 2 +[11] Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yun-song Wang, and Gim Hee Lee. VCR-GauS: View Consistent Depth-Normal Regularizer for Gaussian Surface Reconstruction. NeurIPS, 2024. 2 +[12] Zhiqin Chen and Hao Zhang. Learning Implicit Fields for Generative Shape Modeling. CVPR, 2019. 2 + +[13] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural Unsigned Distance Fields for Implicit Function Learning. NeurIPS, 2020. 1, 2 +[14] Christopher B. Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction. ECCV, 2016. 2 +[15] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-Regularized Optimization for 3D Gaussian Splatting in Few-Shot Images. CVPRW, 2024. 2 +[16] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D Reconstructed of Indoor Scenes. CVPR, 2017. 5 +[17] Haoqiang Fan, Hao Su, and Leonidas Guibas. A Point Set Generation Network for 3D Object Reconstruction from a Single Image. CVPR, 2017. 2 +[18] Lue Fan, Yuxue Yang, Minxing Li, Hongsheng Li, and Zhaoxiang Zhang. Trim 3D Gaussian Splitting for Accurate Geometry Representation. arXiv:2406.07499, 2024. 2 +[19] Brandon Yushan Feng and Amitabh Varshney. SIGNET: Efficient Neural Representation for Light Fields. ICCV, 2021. 2 +[20] Brandon Y Feng, Yinda Zhang, Danhang Tang, and Ruofei Du. PRIF: Primary Ray-based Implicit Function. ECCV, 2022. 1, 3 +[21] A Tewari O Fried, J Thies V Sitzmann, S Lombardi K Sunkavalli, and R Martin-brualla T Simon J Saragih. State of the Art on Neural Rendering. Computer Graphics Forum, 2020. 2 +[22] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. CVPR, 2018. 3 +[23] Jeff P. Grossman and William J. Dally. Point Sample Rendering. Rendering Techniques, 1998. 2 +[24] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. CVPR, 2024. 1, 2 +[25] Xian-feng Han, Hamid Laga, and Mohammed Bennamoun. Image-based 3D Object Reconstruction: State-of-the-Art and Trends in the Deep Learning Era. TPAMI, 2019. 2 +[26] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. SIGGRAPH, 2024. 2 +[27] Han Huang, Yulun Wu, Chao Deng, Ge Gao, Ming Gu, and Yu-Shen Liu. FatesGS: Fast and Accurate Sparse-View Surface Reconstruction using Gaussian Splatting with Depth-Feature Consistency. AAAI, 2025. 2 +[28] Linyi Jin, Nilesh Kulkarni, and David Fouhey. 3DFIREs: Few Image 3D REconstruction for Scenes with Hidden Surface. CVPR, 2024. 3 +[29] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3D Mesh Renderer. CVPR, 2018. 2 +[30] Bernhard Kerbl, Université Côte, Georgios Kopanas, Université Côte, Thomas Leimkuhler, Max-planck-institut Informatik, and G R Aug. 3D Gaussian Splitting for Real-Time + +Radiance Field Rendering. SIGGRAPH, 2023. 1, 2, 3, 4, 5, 6 +[31] Leonid Keselman and Martial Hebert. Approximate Differentiable Rendering with Algebraic Surfaces. ECCV, 2022. 4, 3 +[32] Nilesh Kulkarni, Justin Johnson, and David F Fouhey. Directed Ray Distance Functions for 3D Scene Reconstruction. ECCV, 2022. 1, 3 +[33] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing Sparse-View 3D Gaussian Radiance Fields with Global-Local Depth Normalization. CVPR, 2024. 1, 2 +[34] Celong Liu, Zhong Li, Junsong Yuan, and Yi Xu. NeuLF: Efficient Novel View Synthesis with Neural 4D Light Field. Eurographics Symposium on Rendering, 2022. 2 +[35] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. MVSGaussian: Fast Generalizable Gaussian Splitting Reconstruction from Multi-View Stereo. ECCV, 2024. 5 +[36] Zhuoman Liu, Bo Yang, Yan Luximon, Ajay Kumar, and Jinxi Li. RayDF: Neural Ray-surface Distance Fields with Multi-view Consistency. NeurIPS, 2023. 1, 3, 5, 6, 4, 7 +[37] Yongsen Mao, Yiming Zhang, Hanxiao Jiang, Angel X Chang, and Manolis Savva. MultiScan: Scalable RGBD scanning for 3D environments with articulated objects. NeurIPS, 2022. 5 +[38] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy Networks: Learning 3D Reconstruction in Function Space. CVPR, 2019. 1, 2 +[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. ECCV, 2020. 1, 2, 6 +[40] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 2022. 6 +[41] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural Point Light Fields. CVPR, 2022. 2 +[42] Wanli Ouyang, Guofeng Zhang, Yu Qiao, and Tong He. GigaGS: Scaling up Planar-Based 3D Gaussians for Large Scene Surface Reconstruction. arXiv:2409.06685, 2024. 2 +[43] Onur Ozyesil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A Survey of Structure from Motion. Acta Numerica, 2017. 2 +[44] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation. CVPR, 2019. 1, 2 +[45] Hanspeter Pfister, Matthias Zwicker, Jeroen van Baar, and Markus H. Gross. Surfels: surface elements as rendering primitives. SIGGRAPH, 2000. 2 +[46] Rene Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. TPAMI, 2022. 3 + +[47] Edoardo Mello Rella, Ajad Chhatkuli, Ender Konukoglu, and Luc Van Gool. Neural Vector Fields for Surface Representation and Inference. IJCV, 2024. 2 +[48] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object Space EWA Surface Splating: A Hardware Accelerated Approach to High Quality Point Rendering. Computer Graphics Forum, 2003. 2 +[49] Vincent Sitzmann, Semon Rezchikov, William T Freeman, Joshua B. Tenenbaum, and Fredo Durand. Light Field Networks: Neural Scene Representations with Single-Evaluation Rendering. NeurIPS, 2021. 2 +[50] Gaochao Song, Chong Cheng, and Hao Wang. GVKF: Gaussian Voxel Kernel Functions for Highly Efficient Surface Reconstruction in Open Scenes. NeurIPS, 2024. 2 +[51] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Light Field Neural Rendering. CVPR, 2022. 2 +[52] Peder Bergebakken Sundt and Theoharis Theoharis. MARF: The Medial Atom Ray Field object representation. Computers & Graphics, 2023. 3 +[53] Peder Bergebakken Sundt and Theoharis Theoharis. Towards Multi-view Consistency in Neural Ray Fields using Parametric Medial Surfaces. Computers & Graphics, 2024. 3 +[54] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree Generating Networks: Efficient Convolutional Architectures for High-resolution 3D Outputs. ICCV, 2017. 2 +[55] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, Tomas Simon, Christian Theobalt, Matthias Niessner, Jonathan T. Barron, Gordon Wetzstein, Michael Zollhoefer, and Vladislav Golyanik. Advances in Neural Rendering. SIGGRAPH, 2021. 2 +[56] Alex Trevithick and Bo Yang. GRF: Learning a General Radiance Field for 3D Representation and Rendering. ICCV, 2021. 2 +[57] Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. DN-Splatter: Depth and Normal Priors for Gaussian Splatting and Meshing. WACV, 2025. 5 +[58] Bing Wang, Zhengdi Yu, Bo Yang, Jie Qin, Toby Breckon, Ling Shao, Niki Trigoni, and Andrew Markham. RangeUDF: Semantic Surface Reconstruction from 3D Point Clouds. arXiv:2204.09138, 2022. 2 +[59] Jiaxu Wang, Ziyi Zhang, Junhao He, and Renjing Xu. PFGS: High Fidelity Point Cloud Rendering via Feature Splatting. ECCV, 2024. 5 +[60] 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:2106.10689, 2021. 6 +[61] Zhengren Wang. 3D Representation Methods: A Survey. arXiv:2410.06475, 2024. 2 +[62] Yaniv Wolf, Amit Bracha, and Ron Kimmel. GS2Mesh: Surface Reconstruction from Gaussian Splitting via Novel Stereo Views. ECCV, 2024. 1 + +[63] Qianyi Wu, Jianmin Zheng, and Jianfei Cai. Surface Reconstruction from 3D Gaussian Splatting via Local Structural Hints. ECCV, 2024. 1, 2 +[64] Haodong Xiang, Xinghui Li, Xiansong Lai, Wanting Zhang, Zhichao Liao, Kai Cheng, and Xueping Liu. Gaussian-Room: Improving 3D Gaussian Splatting with SDF Guidance and Monocular Cues for Indoor Scene Reconstruction. arXiv:2405.19671, 2024. 2 +[65] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural Fields in Visual Computing and Beyond. Computer Graphics Forum, 2022. 2 +[66] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. DepthSplat: Connecting Gaussian Splatting and Depth. CVPR, 2025. 2 +[67] Bo Yang, Stefano Rosa, Andrew Markham, Niki Trigoni, and Hongkai Wen. Dense 3D Object Reconstruction from a Single Depth View. TPAMI, 2019. 2 +[68] Bo Yang, Sen Wang, Andrew Markham, and Niki Trigoni. Robust Attentional Aggregation of Deep Feature Sets for Multi-view 3D Reconstruction. *IJCV*, 2020. 2 +[69] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth Anything V2. NeurIPS, 2024. 3, 5 +[70] Xianghui Yang, Guosheng Lin, Zhenghao Chen, and Luping Zhou. Neural Vector Fields: Implicit Representation by Explicit Learning. CVPR, 2023. 1, 2 +[71] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes. ICCV, 2023. 5 +[72] Ruihong Yin, Yunlu Chen, Sezer Karaoglu, and Theo Gevers. Ray-Distance Volume Rendering for Neural Scene Reconstruction. ECCV, 2024. 3 +[73] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. GSDF: 3DGS Meets SDF for Improved Rendering and Reconstruction. NeurIPS, 2024. 2 +[74] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian Opacity Fields: Efficient and Compact Surface Reconstruction in Unbounded Scenes. SIGGRAPH Asia, 2024. 4, 5, 6, 2 +[75] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. RaDe-GS: Rasterizing Depth in Gaussian Splatting. arXiv:2406.01467, 2024. +[76] Ziyu Zhang, Binbin Huang, Hanqing Jiang, Liyang Zhou, Xiaojun Xiang, and Shunhan Shen. Quadratic Gaussian Splitting for Efficient and Detailed Surface Reconstruction. arXiv:2411.16392, 2024. 2 +[77] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus Gross. Surface splatting. SIGGRAPH, 2001. 2 \ No newline at end of file diff --git a/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/images.zip b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..86ecea7060f4a9617b9074f9ce5dad89b7614ba4 --- /dev/null +++ b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a358265bbc7de0fa966d6fbb2b7801f3ee70265d1421989410c9ae575d9aad1 +size 740825 diff --git a/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/layout.json b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1328b451484a66ebf516e8f0d25d85071d8e71bd --- /dev/null +++ b/ICCV/2025/RayletDF_ Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3440adc76caa5962083f5f649a846163cc3af3bfb94756161d61ab9e5b8ca565 +size 474004 diff --git a/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/0ac3ee88-2827-4a77-b75c-935c9749c9d5_content_list.json b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/0ac3ee88-2827-4a77-b75c-935c9749c9d5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fd5133eed8d8056c74f29259d30e8aa1b4a53a49 --- /dev/null +++ b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/0ac3ee88-2827-4a77-b75c-935c9749c9d5_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a9363fcc6615c083445e7a1573c7ce28cfb6e39a6a1d26eed9c5dfaf288c98f +size 89543 diff --git a/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/0ac3ee88-2827-4a77-b75c-935c9749c9d5_model.json b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/0ac3ee88-2827-4a77-b75c-935c9749c9d5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3e4c095d1993efcd60b2998135e716ffaa1f1f5 --- /dev/null +++ b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/0ac3ee88-2827-4a77-b75c-935c9749c9d5_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbcf1a77b4627e4e83ffdffc1b24c7edfb2b2941ec3b2cee33386ce63b3dfbe9 +size 110364 diff --git a/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/0ac3ee88-2827-4a77-b75c-935c9749c9d5_origin.pdf b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/0ac3ee88-2827-4a77-b75c-935c9749c9d5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..80d9a5f9f0ed5fd099386526b76e7e80dfaf05e7 --- /dev/null +++ b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/0ac3ee88-2827-4a77-b75c-935c9749c9d5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b53f0c1260b5bee952e80f395e49133af9dbb55f5dc2e5c107c175bd1b073d39 +size 4132302 diff --git a/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/full.md b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..837889919b4b1515627f9a97638fec5e0192b8b2 --- /dev/null +++ b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/full.md @@ -0,0 +1,415 @@ +# ReAL-AD: Towards Human-Like Reasoning in End-to-End Autonomous Driving + +Yuhang Lu $^{1}$ , Jiadong Tu $^{1}$ , Yuexin Ma $^{1,\dagger}$ , Xinge Zhu $^{2,\dagger}$ $^{1}$ ShanghaiTech University + $^{2}$ The Chinese University of Hong Kong + +{luyh2, tujd2023, mayuexin}@shanghaiitech.edu.cn, zhuxingel23@gmail.com + +# Abstract + +End-to-end autonomous driving has emerged as a promising approach to unify perception, prediction, and planning within a single framework, reducing information loss and improving adaptability. However, existing methods often rely on fixed and sparse trajectory supervision, limiting their ability to capture the hierarchical reasoning process that human drivers naturally employ. To bridge this gap, we propose Real-AD, a Reasoning-Augmented Learning framework that structures decision-making in autonomous driving based on the three-tier human cognitive model: Driving Strategy, Driving Decision, and Driving Operation, where Vision-Language Models (VLMs) are incorporated to enhance situational awareness and structured reasoning across these levels. Specifically, we introduce: (1) the Strategic Reasoning Injector, which formulates high-level driving strategies by interpreting complex traffic contexts from VLM-generated insights; (2) the Tactical Reasoning Integrator, which refines strategic intent into interpretable tactical choices such as lane changes, overtaking, and speed adjustments; and (3) the Hierarchical Trajectory Decoder, which progressively translates tactical decisions into precise control actions for smooth and human-like trajectory execution. Extensive evaluations show that integrating our framework improves planning accuracy and safety by over $30\%$ , making end-to-end autonomous driving more interpretable and aligned with human-like hierarchical reasoning. The project page can be found at: 4dvlab.github.io/project_page/realad + +# 1. Introduction + +End-to-end autonomous driving [6, 14, 15, 20, 39, 48, 57] has gained significant attention due to its advantages, such as minimizing information loss and simplifying system architecture. However, existing methods face fundamental limitations. Most notably, they rely on fixed and sparse tra + +![](images/24959ab55af9480d67377eb431e17f6f4c60f062489ec9b1ab9feb5d7f1637ef.jpg) +(a) Traditional end-to-end planning method + +![](images/c09ebdc8dd1cb6b7c36684f9cf81c5a86be951930f3461015191a410e2f9dfb4.jpg) +(b) The decision-making process of a human driver +Figure 1. Comparison between end-to-end networks and human driving logic. (a) illustrates the workflow and limitations of the end-to-end autonomous driving system, while (b) depicts the structured decision-making process of a human driver. + +jectory supervision, which fails to replicate the structured cognitive reasoning inherent to human drivers - a process involving hierarchical understanding of contextual relationships (e.g., traffic rules, interactive intentions) and adaptive decision-making hierarchy. + +Recent efforts [32, 33, 42, 46, 51, 52] to incorporate Vision-Language Models (VLMs) [1, 2, 8, 28, 45] aim to address these challenges by leveraging their semantic reasoning capabilities and cross-modal alignment. Most studies treat VLMs as auxiliary modules, providing additional semantic cues to enhance perception and planning. These approaches treat semantic reasoning as an isolated preprocessing step rather than embedding it organically into the decision hierarchy, where strategic planning, tactical decision-making, and operational control work in synergy. This lack of structured multi-stage reasoning hinders generalization and limits the effectiveness of current VLM-driven solutions in diverse real-world scenarios. + +So how do Human Drivers perform the decision-making? As illustrated in Fig. 1, human drivers make decisions in a structured, hierarchical manner, progressively refining their actions from high-level reasoning to precise execution. At the Driving Strategy level, drivers analyze + +the overall driving scene to establish a strategic understanding. This involves identifying key traffic elements, prioritizing relevant objects, and incorporating contextual factors such as traffic rules and road conditions. This strategic awareness provides the foundation for subsequent decision-making. At the Driving Decision level, drivers translate their strategic understanding into tactical choices, such as determining whether to maintain their lane, overtake, or adjust speed. These decisions are made dynamically based on traffic flow, surrounding vehicles, and predicted interactions, ensuring safe and efficient maneuvering. At the Driving Operation level, drivers refine tactical decisions into precise vehicle control actions. This includes adjusting steering, throttle, and braking to execute the intended maneuvers smoothly, conducting lateral control actions including lane changes to maintain stability and safety. However, most existing end-to-end autonomous driving systems fail to explicitly model this hierarchical cognitive process. Instead, they often rely on direct trajectory prediction without structured reasoning, leading to suboptimal planning and a lack of human-like decision-making granularity. + +Building upon this insight, we propose ReAL-AD, a novel Reasoning-Augmented Learning framework for Autonomous Driving, which leverages the reasoning capabilities of Vision-Language Models (VLMs) to embed human-like hierarchical decision-making into end-to-end autonomous driving systems. Inspired by the structured cognitive process of human drivers, ReAL-AD introduces a multi-level architecture that explicitly models decision-making from strategic reasoning to tactical execution, thereby enhancing interpretability and adaptability in dynamic environments. + +Specifically, the Strategic Reasoning Injector extracts high-level situational awareness from VLM-generated insights, encoding them into ego-centric queries to guide downstream planning. The Tactical Reasoning Integrator refines these high-level decisions into structured, interpretable tactical control commands, ensuring both feasibility and consistency with real-world driving constraints. Finally, the Hierarchical Trajectory Decoder employs a hierarchical variational decoder to iteratively refine trajectory planning in a two-stage process, ensuring coherence between high-level strategic intent and low-level tactical execution. By integrating VLM-driven reasoning across all levels of decision-making, ReAL-AD significantly improves interpretability, adaptability, and overall driving performance in complex and dynamic traffic scenarios. + +Our extensive experiments on the NuScenes [3] and Bench2Drive [19] datasets demonstrate the superior performance of our approach over existing methods. Compared to the baseline, our method achieves a $33\%$ reduction in L2 error and a $32\%$ decrease in collision rate, significantly improving trajectory accuracy and driving safety across di + +verse dynamic scenarios. Additionally, a comprehensive ablation study systematically evaluates each component's contribution, further validating our framework's effectiveness. Our contributions can be summarized as follows: + +- We propose RealAD, a novel reasoning-augmented end-to-end autonomous driving framework that explicitly incorporates hierarchical decision-making and aligns trajectory planning with human cognitive processes. +- We introduce the Strategic Reasoning Injector for VLM-driven decision integration, the Tactical Reasoning Integrator for structured control, and the Hierarchical Trajectory Decoder for hierarchical trajectory refinement, ensuring consistency from reasoning to execution. +- Our framework, when integrated into baseline approaches, demonstrates over $30\%$ improvement in both trajectory planning accuracy and safety metrics across NuScenes and Bench2Drive datasets, while ablation studies confirm the contribution of each component. + +# 2. Related Work + +# 2.1. End-to-End Autonomous Driving + +End-to-end autonomous driving planning methods [6, 7, 9, 11, 14, 15, 20, 24, 25, 27, 39, 48, 57] achieve their ultimate goal by simultaneously training multiple modules, reducing information loss during the pipeline process, and making them a popular area of research. ST-P3 [14] introduces a design that integrates multiple auxiliary information or tasks to enhance planning performance. UniAD [15] demonstrates impressive performance in open-loop evaluations. Subsequently, VAD [20] introduces compact vectorized scene representations, which not only improve planning efficiency but also reduce computational costs. PARADrive [48] investigates the impact of the design order of auxiliary tasks within an end-to-end framework. GenAD [57] models autonomous driving as a future generation problem, conducting motion prediction and ego planning simultaneously within a structured latent trajectory space. Following this, works such as VADv2 [6] and DiffusionDrive [27] explore the integration of probabilistic modeling into planning to enhance accuracy. + +However, these methods rely on a fixed set of future ego-vehicle trajectories as ground truth, leading to sparse supervision. This is problematic since human drivers use complex decision-making, which ego-vehicle trajectories alone fail to capture. This limits an autonomous driving system's learning and generalization. In contrast, our approach leverages Vision-Language Models (VLMs) for a deeper understanding of driving behavior. By integrating high-dimensional strategies and tactical commands while predicting trajectories with finer granularity through a hierarchical decoder, the network mimics human reasoning, enhancing planning decisions and overall autonomous driv + +ing performance. + +# 2.2. VLMs for Autonomous Driving + +Recently, Vision-Language Models (VLMs) have demonstrated remarkable performance across a wide range of general tasks [1, 2, 8, 23, 28, 40, 45, 54], and their integration into autonomous driving systems [4, 10, 26, 30, 41, 53] has become a prominent area of research. A number of studies [5, 16-18, 29, 31, 32, 35-38, 44, 46, 47, 50, 52] treat VLMs as agents that receive driving scene images and text prompts as inputs, generating driving decisions as outputs. For instance, Drive-with-LLMs [5] encode the perception information into a latent space, which is then fed into a large language model (LLM) to predict future planning trajectories. DriveGPT4 [52] takes front-camera video inputs, employing VLMs to predict control signals for planning and provide decision explanations. LanguageMPC [36] converts historical ground truth perception data and HD maps into language format, using chain-of-thought reasoning to analyze driving scenes and generate planning actions. AgentDriver [32] converts driving situations into textual descriptions with human-like intelligence, then uses an LLM to reason and plan. Additionally, DriveMLM [46] validates the effectiveness of VLM-based planning models within a closed-loop simulation environment. However, while VLMs can capture complex visual and language cues, they often lack a full grasp of the underlying physics and constraints of driving behavior and limited 3D spatial understanding, resulting in less accurate or unsafe trajectory predictions compared to end-to-end systems that consider more comprehensive environmental feedback. + +An alternative approach leverages the decision-making capabilities of VLMs as extra inputs to end-to-end autonomous driving systems. This approach primarily focuses on utilizing VLM outputs to refine or guide the learning process of these systems. DriveVLM [42] employs VLMs as a slower system to generate driving trajectories, with complementary networks acting as references when needed. VLMAD [51] uses the VLM as a teacher to generate free-form reasoning and construct action annotations, thereby assisting in the learning of the end-to-end network. VLP [33] enhances autonomous driving systems by strengthening their understanding of vehicle contexts and environments. Senna [21] decouples high-level planning from low-level trajectory prediction, creating a more modular and interpretable planning framework. Existing methods incorporate decision features via distillation and contrastive learning but treat semantic reasoning as a separate preprocessing step, limiting generalization in real-world scenarios. To address this, we leverage VLM to generate driving strategy and tactical commands and introduce a hierarchical trajectory decoder to translate these decisions into precise control actions, simulating the human driving thought process. + +# 3. Methods + +# 3.1. Overview + +In this section, we present a VLM-assisted framework for human-like reasoning-augmented learning, integrating a three-tier human cognitive model—Driving Strategy, Driving Decision, and Driving Operation—into end-to-end autonomous driving systems. First, we revisit the operation paradigm of conventional end-to-end planning systems (Sec. 3.2). To inject human-like reasoning patterns, we develop: (1) Strategic Reasoning Injector (Sec. 3.3) formulates high-level driving strategies by interpreting complex traffic contexts from VLM-generated insights; (2) Tactical Reasoning Integrator (Sec. 3.4) refines strategic intent into interpretable tactical choices; (3) Hierarchical Trajectory Decoder (Sec. 3.5) mimics human intuition-refinement patterns via hierarchical planning — first establishing rough motion patterns, then refining detailed trajectories. Finally, in Sec. 3.6, we introduce how to supervise the network. + +# 3.2. Preliminary + +In a vision-based end-to-end autonomous driving system, multi-view camera image streams are first processed by an image backbone (e.g., ResNet [12]) to extract 2D visual features. These features are then transformed into 3D or Bird's Eye View (BEV) scene representations via view transformation modules, capturing scene geometry such as road layouts, agent positions, and lane topology. Transformer-based task-specific decoders enable task queries (e.g., agent queries, map queries) to interact with scene features, modeling relevant representations for planning. The planning module initializes an ego-query embedding, which attends to scene features and task queries through cross-attention layers to construct ego-features, enabling the ego vehicle to contextualize its state. Finally, the trajectory decoder, typically a Multi-Layer Perceptron (MLP), predicts future waypoints by regressing from the ego-features. + +# 3.3. Strategic Reasoning Injector + +When navigating a driving scenario, human drivers first make driving strategies by identifying critical traffic participants or relevant regulations—cognitive processes that we replicate using VLM-generated reasoning. These structured insights are then encoded to guide the ego-query, which serves as the starting point of the planning module. + +The VLM first generates textual driving strategy through prompt-guided visual reasoning: + +$$ +T _ {\text {s t r a t e g y}} = \underbrace {M (V , P _ {\text {s t r a t e g y}})} _ {\text {V L M - b a s e d d r i v i n g s t r a t e g y g e n e r a t i o n}} \tag {1} +$$ + +where $M(\cdot)$ denotes the VLM processor, $V$ the visual input, and $P_{strategy}$ denotes the driving strategy prompt template. + +![](images/9200625dd0b8c65862136a819b0e5becf9d4f48442269f023e9e3b20ec966482.jpg) +Figure 2. Overall pipeline of ReAL-AD. Multi-view images are processed by the Scene Encoder to extract environmental features. The Strategic Reasoning Injector generates driving decisions using structured prompts and utilizes them to enhance ego-query. The Tactical Reasoning Integrator outputs reactive- and regulatory-level command features, which are then fed into the Hierarchical Trajectory Decoder. This decoder progressively refines the latent trajectory space to generate the final planning trajectory. + +The generated strategy text $T_{strategy}$ is then encoded into semantic space through a pretrained language encoder $E_{text}(\cdot)$ . To bridge the modality gap between linguistic strategy and visual perceptions, we choose a lightweight adapter module implemented as: + +$$ +F _ {\text {s t r a t e g y}} = M L P \left(E _ {\text {t e x t}} \left(T _ {\text {s t r a t e g y}}\right)\right) \tag {2} +$$ + +Concurrently, we encode the ground-truth trajectory $g_{traj}$ into planning features $F_{gt}$ using trajectory encoder $E_{gt}(\cdot)$ . A cosine similarity loss enforces consistency between strategy semantics and planning dynamics: + +$$ +\mathcal {L} _ {s i m} = 1 - \cos \left(F _ {s t r a t e g y}, F _ {g t}\right) \tag {3} +$$ + +The refined strategy features are then integrated into the ego-query through residual updating: + +$$ +Q _ {e g o} \leftarrow Q _ {e g o} + F _ {\text {s t r a t e g y}} \tag {4} +$$ + +This approach injects VLM-derived reasoning capabilities into ego-query, thereby guiding the acquisition of key planning-related features when interacting with environmental features. More details can be found in Appendix + +# 3.4. Tactical Reasoning Integrator + +While driving strategies provide semantic context, their abstract nature (e.g., "yield to approaching vehicles") may lack actionable controls for trajectory planning. To bridge this gap, we introduce tactical command that translates semantic strategies into executable choices, establishing dual-level reasoning from strategic planning to tactical driving decision and operation. + +The VLM generates structured commands through categorical-constrained visual reasoning: + +$$ +T _ {\text {c o m m a n d}} = \underbrace {M \left(V , P _ {\text {c o m m a n d}}\right)} _ {\text {V L M - b a s e d s t r u c t u r e d c o m m a n d g e n e r a t i o n}} \tag {5} +$$ + +where $M(\cdot)$ denotes the VLM processor, $V$ the visual input, and $P_{\text{command}}$ denotes the command prompt template and enforces four-category output format (Direction/Emergency/Lane/Speed). + +The raw text output $T_{command}$ undergoes deterministic parsing to extract executable commands: + +$$ +\left\{c _ {k} \right\} _ {k = 1} ^ {4} = \operatorname {R e g e x P a r s e} \left(T _ {\text {c o m m a n d}}, \left[ \begin{array}{l} \mathcal {P} _ {\text {d i r}} \\ \mathcal {P} _ {\text {l a n e}} \\ \mathcal {P} _ {\text {s p e e d}} \\ \mathcal {P} _ {\text {e m e r g}} \end{array} \right]\right) \tag {6} +$$ + +where $\mathcal{P}_*$ are pre-defined regular expression patterns for each category. Each text command $c_{k}$ is then mapped to its category-specific index: + +$$ +i _ {k} = \operatorname {L o o k u p} \left(\mathcal {D} _ {k}, c _ {k}\right), \quad \mathbf {c} _ {k} = \operatorname {o n e h o t} \left(i _ {k}\right) \in \{0, 1 \} ^ {n _ {k}} \tag {7} +$$ + +with $\mathcal{D}_k$ being the command dictionary for category $k$ containing all predefined options. This two-step conversion guarantees machine-readability while preserving VLM's semantic understanding. + +The discrete commands undergo category-specific encoding and strategic fusion: + +$$ +\left\{ \begin{array}{l} \mathbf {e} _ {d i r} = E _ {d i r} (\mathbf {c} _ {1}), \quad \mathbf {e} _ {e m e r g} = E _ {e m e r g} (\mathbf {c} _ {4}) \\ \mathbf {e} _ {l a n e} = E _ {l a n e} (\mathbf {c} _ {2}), \quad \mathbf {e} _ {s p e e d} = E _ {s p e e d} (\mathbf {c} _ {3}) \end{array} \right. \tag {8} +$$ + +where $\{E_k\}$ are learnable embedding matrices. The encoded features are then partitioned by vehicle control level: + +![](images/a3feeccd7e0ca674e0d5d6bdf456f93735181e45cf40cc441a104e85c367973a.jpg) + +![](images/d8eba021e72bca2af21a30b6566a0b4005cb07539c9e0a24f05fcd3325b35e20.jpg) + +# VLM-generated driving strategy + +Solid yellow lines prohibit lane changes; white arrows indicate the direction of traffic flow. + +A football is detected on the road. Caution advised—potential presence of children retrieving the object + +Maintain a safe following distance from the vehicle ahead. + +Drive at a moderate speed, keep a safe distance, and yield to pedestrians and cyclists when necessary. + +(Driving Strategy) + +![](images/1b059af0ffd791cb8164f663b5a8e18a3af526481ebd5ca058681fe8ab2fd3a8.jpg) +Figure 3. Example VLM output. Given a front-view image and prompt, VLM generates corresponding driving strategy, reactive-level and regulatory-level commands. + +VLM-generated reactive-level commands + +Direction Control: CONTINUE STRAIGHT + +Emergency Control: NO_ACTION + +(Driving Decision) + +![](images/d71c3bce0ed5d3ecb992ec569affb8dc53f3583c28648937a826f269c989a77b.jpg) + +VLM-generated regulatory-level commands + +Lane Management: KEEP LANE + +Speed Control: MAINTAIN_SPEED + +(Driving Operation) + +$$ +\mathbf {g} _ {\text {r e a c t i v e}} = \mathbf {e} _ {\text {d i r}} \oplus \mathbf {e} _ {\text {e m e r g}} \tag {9} +$$ + +$$ +\mathbf {g} _ {\text {r e g u l a t o r y}} = \mathbf {e} _ {\text {l a n e}} \oplus \mathbf {e} _ {\text {s p e e d}} +$$ + +$\mathbf{g}_{\text {reactive }}$ encodes the driving decision at the instantaneous reaction level, while $\mathbf{g}_{\text {regulatory }}$ encodes the driving operation command after careful consideration. These two commands function at different layers within the subsequent hierarchical trajectory decoder, offering detailed and precise guidance to facilitate the generation of human-like trajectories from coarse to fine. + +# 3.5. Hierarchical Trajectory Decoder + +Inspired by the hierarchical nature of human driving cognition—where instantaneous reactions occur before deliberate adjustments, we propose a two-layer variational decoder with latent trajectory spaces conditioned on control at both the driving decision and the driving operation level. + +The first-layer decoder models rough motion patterns using ego-vehicle features and reactive-level tactical com + +mands (directional intent and emergency indicators) via conditional variational inference: + +1. Distribution Parameterization: Map inputs to latent coarse trajectory space. + +$$ +\mu^ {c}, \log \sigma^ {c} = \operatorname {D i s t r i b u t i o n E n c o d e r} _ {\theta} ([ f _ {e g o}; \mathbf {g} _ {\text {r e a c t i v e}} ]) \tag {10} +$$ + +2. Latent Sampling: Extract global motion patterns. + +$$ +z ^ {c} \sim \mathcal {N} \left(\mu^ {c}, \sigma^ {c}\right) \tag {11} +$$ + +where $f_{ego} \in \mathbb{R}^d$ is ego-vehicle features, $\mathbf{g}_{\text{reactive}} \in \mathbb{R}^{16}$ denotes reactive-level command features, and coarse latent code $z^c \in \mathbb{R}^k$ captures rough motion pattern. + +The second layer refines the coarse motion pattern from the first layer by incorporating multi-source conditioning and mapping it to a fine-grained trajectory representation. + +1. Hierarchical Refinement: Condition latent fine trajectory space on coarse motion patterns and regulatory-level command. + +$$ +\mu^ {f}, \log \sigma^ {f} = \operatorname {D i s t r i b u t i o n E n c o d e r} _ {\phi} \left(\left[ f _ {e g o}; z ^ {c}; \mathbf {g} _ {\text {r e g u l a t o r y}} \right]\right) \tag {12} +$$ + +2. Latent Sampling: Sample fine-grained trajectory representation. + +$$ +z ^ {f} \sim \mathcal {N} \left(\mu^ {f}, \sigma^ {f}\right) \tag {13} +$$ + +where $\mathbf{g}_{\text{regulatory}} \in \mathbb{R}^{16}$ denotes regulatory-level command features which encode lane management and speed control choices. + +After obtaining latent representations $z^c$ and $z^f$ from the hierarchical variational process, the next step is to decode these latent variables into actual trajectory sequences. Our trajectory decoding extends the temporal modeling in [57] with dual latent-stream processing. The process is: + +Coarse Stream: $h^c = \mathrm{GRU}(\mathrm{repeat}(z^c, T), f_{ego})$ + +$$ +\operatorname {p r e d} _ {\text {c o a r s e}} = \operatorname {M L P} \left(h ^ {c}\right) +$$ + +Fine Stream: $h^f = \mathrm{GRU}(\mathrm{repeat}(z^f, T), f_{ego})$ + +$$ +\operatorname {p r e d} _ {\text {f i n e}} = \mathrm {M L P} \left(h ^ {f}\right) \tag {14} +$$ + +where $z^c$ and $z^f$ represent the approximate motion pattern and the refined trajectory representation, respectively. $T$ denotes the number of future timesteps for planning. $h^c$ and $h^f$ correspond to the latent feature sequences for coarse and fine trajectory prediction. $pred_{coarse}$ and $pred_{fine}$ denote the planned coarse and fine trajectories, respectively. + +# 3.6. Loss Function + +Our composite loss function integrates five key components for effective hierarchical learning: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {o v e r a l l}} = \underbrace {\mathcal {L} _ {\text {b a s e l i n e}}} _ {\text {B a s e m o d e l}} + \underbrace {\lambda_ {0} \mathcal {L} _ {\text {s i m}} + \lambda_ {1} \mathcal {L} _ {\text {g t - r e c o n}}} _ {\text {S t r a g e t i c R e a s o n i n g I n j e c t i o n L o s s}} \tag {15} \\ + \underbrace {\lambda_ {2} \mathcal {L} _ {\text {c o a r s e}} + \lambda_ {3} \mathcal {L} _ {\text {h i e r - k l}}} _ {\text {H i e r a c h i c a l T r a j e c t o r y D e c o d e r L o s s}} \\ \end{array} +$$ + +Base Model. We retain all baseline model losses, denoted as $\mathcal{L}_{\mathrm{baseline}}$ + +Strategic Reasoning Injection Loss. For the Strategic Reasoning Injection Module, we introduce two losses: + +1. $\mathcal{L}_{sim}$ enforces consistency between strategy semantics and planning dynamics, as defined in 3.3 +2. $\mathcal{L}_{gt\_ recon}$ supervises GT trajectory encoding to ensure proper feature representation: + +$$ +\mathcal {L} _ {g t \_ r e c o n} = \| \operatorname {D e c} (F _ {g t}) - \mathrm {g t} _ {t r a j} \| _ {2} \tag {16} +$$ + +where $\mathrm{Dec}(\cdot)$ denotes our hierarchical trajectory decoder. + +Hierarchical Trajectory Decoder Loss. To ensure effective hierarchical trajectory planning, we introduce two key loss functions for the Hierarchical Trajectory Decoder: + +1. $\mathcal{L}_{\mathrm{coarse}}$ applies the baseline objective to the predicted coarse trajectory, $pred_{\mathrm{coarse}}$ , and the Bezier curve-fitted [34] coarse trajectory ground truth $gt_{\mathrm{coarse}}$ . + +$$ +\mathcal {L} _ {\text {c o a r s e}} = \mathcal {L} _ {\text {b a s e l i n e}} \left(\operatorname {p r e d} _ {\text {c o a r s e}}, g t _ {\text {c o a r s e}}\right) \tag {17} +$$ + +2. $\mathcal{L}_{\mathrm{hier - kl}}$ enforces consistency across the hierarchical latent spaces using a two-level KL divergence formulation: + +$$ +\mathcal {L} _ {\mathrm {K L}} = \beta_ {c} \cdot \underbrace {D _ {K L} \left(q _ {\phi} \left(z _ {\text {c u r r}} ^ {c} \| z _ {\text {f u t}} ^ {c}\right)\right)} _ {\text {C o a r s e - l e v e l}} + \beta_ {f} \cdot \underbrace {D _ {K L} \left(q _ {\phi} \left(z _ {\text {c u r r}} ^ {f} \| z _ {\text {f u t}} ^ {f}\right)\right)} _ {\text {F i n e - l e v e l}} \tag {18} +$$ + +For hierarchy level $l \in \{c, f\}$ , the KL divergence is computed as: + +$$ +D _ {K L} ^ {l} = \frac {1}{2} \left(\log \frac {\sigma_ {\text {f u t} , l} ^ {2}}{\sigma_ {\text {c u r r} , l} ^ {2}} - 1 + \frac {\sigma_ {\text {c u r r} , l} ^ {2} + \left(\mu_ {\text {c u r r} , l} - \mu_ {\text {f u t} , l}\right) ^ {2}}{\sigma_ {\text {f u t} , l} ^ {2}}\right) \tag {19} +$$ + +The overall hierarchical KL loss is then computed as: + +$$ +\mathcal {L} _ {\text {h i e r}} = D _ {K L} ^ {c} + D _ {K L} ^ {f} \tag {20} +$$ + +# 4. Experiments + +# 4.1. Settings + +Baseline. Our method supports different end-to-end planning networks and VLMs. In this paper, we use VAD [20] and UniAD [15] as the baselines, and select MiniCPM-Llama3-2.5V [54] and Qwen-VL [2] as VLMs. + +Datasets. We evaluate open-loop planning on the nuScenes dataset [3], which contains 1,000 20-second scenes annotated at $2\mathrm{Hz}$ and serves as a key benchmark for E2E autonomous driving. For both open-loop and closed-loop evaluation, we use Bench2Drive [19], featuring 2M frames from 13,638 clips across 44 scenarios, 23 weather conditions, and 12 CARLA v2 towns. Its rigorous closed-loop protocol assesses E2E-AD models on 220 routes, ensuring a fair and comprehensive performance evaluation. + +Evaluation Metrics. For open-loop evaluation, we use L2 error and collision rate. L2 error measures the distance between planned and ground truth trajectories, while collision rate quantifies collisions with traffic participants. By default, we evaluate at 1, 2, and 3 seconds using VAD metrics [20]. For closed-loop evaluation, following [19], we use Driving Score and Success Rate. Driving Score reflects route completion with infractions, while Success Rate is the percentage of routes completed without violations. + +Implementation Details. For baselines, we use the official codes and adhere to the hyperparameters specified in their official implementations. $\lambda_0$ , $\lambda_1$ , $\lambda_2$ , $\lambda_3$ are set as 1.0, 0.5, 0.5 and 1.0. $\beta_c$ is 0.5 and $\beta_f$ is 1.0. Models are trained on 8 NVIDIA A40 GPUs or 4 NVIDIA 4090 GPUs using the PyTorch framework. + +# 4.2. Main Results + +Open-loop evaluation. To better assess the effectiveness of our framework, we compare our method with several SOTA methods on nuScenes and Bench2Drive datasets. To better evaluate the effectiveness of our framework, we conducted experiments on two datasets, nuScenes and Bench2Drive, and compared it with several state-of-the-art methods. As shown in Tab. 1 and 2, our method achieves significant improvements over our baseline methods VAD and UniAD, especially in terms of L2 error and collision rate, which can reach more than $30\%$ improvement. It is worth noting that our performance is also better than other VLM-assessed methods such as VLP and VLM-AD that use the same baseline, achieving the lowest average L2 error $(0.48\mathrm{m})$ and collision rate $(0.15\%)$ on nuScenes, and $0.84\mathrm{m}$ and $0.12\%$ on Bench2Drive. This shows that the introduction of human decision-making process makes the network more effective in learning driving ability. + +Closed-loop evaluation. Although the open-loop metrics provide partial performance results, we conduct closed-loop evaluation on Bench2Drive to assess real-world applicability. The results show that after integrating our framework, both the driving score and the number of completed paths have been significantly improved compared to the baseline, indicating that the introduction of human thinking processes has successfully improved driving mastery. + +
ModelL2(m) ↓Collision Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
NMP [55]0.531.252.671.480.040.120.870.34
FF [13]0.551.202.541.430.060.171.070.43
EO [22]0.671.362.781.600.040.090.880.33
ST-P3 [14]1.332.112.902.110.230.621.270.71
UniAD† [15]0.480.961.651.030.050.170.710.31
OccNet† [43]1.292.132.992.140.210.591.370.72
VAD-Base [20]0.410.701.050.720.070.170.410.22
PARA-Drive† [48]0.400.771.310.830.070.250.600.30
GenAD [57]0.280.490.780.520.080.140.340.19
UAD [11]0.390.811.500.900.010.120.430.19
DriveVLM* [42]0.150.290.480.310.050.080.170.10
VLP-UniAD [33]0.360.681.190.740.030.120.320.16
VLP-VAD [33]0.300.530.840.550.010.070.380.15
VLM-AD (UniAD) [51]0.390.821.430.880.050.110.430.19
VLM-AD (VAD) [51]0.240.460.750.480.120.170.410.23
Senna [21]0.370.540.860.590.090.120.330.18
Ours (VAD & MiniCPM-V)0.300.480.670.480.070.100.280.15
Ours (UniAD &. MiniCPM-V)†0.400.711.140.770.020.120.370.17
Ours (VAD & Qwen-VL)0.350.530.710.530.090.120.310.17
Ours (UniAD & Qwen-VL)†0.390.751.090.740.020.140.390.17
+ +Table 1. Open-loop planning results on nuScenes dataset. “*” indicates ego-vehicle status features are used as input. “†” denotes results evaluated using UniAD metrics. The best result is in bold, while the second-best is underlined. (Methods that use ego-vehicle status features as input are not considered in this ranking) + +
ModelOpen-loop MetricClosed-loop Metric
L2(m) ↓Collision Rate (%) ↓Driving Score ↑Success Rate(%) ↑
1s2s3sAvg.1s2s3sAvg.
UniAD-tiny [20]0.510.911.370.930.842.293.442.1932.009.54
UniAD-base [20]0.380.721.120.740.541.653.361.8537.729.54
VAD [20]0.450.911,470.940.100.200.290.2039.4210.00
AD-MLP [56]1.983.635.293.395.928.039.707.8817.910.00
TCP [49]1.251.70-1.483.313.00-3.1532.439.54
Ours (VAD & MiniCPM-V)0.380.841.320.840.030.080.200.1141.1711.36
Ours (UniAD &. MiniCPM-V)0.320.671.080.690.391.222.581.4039.1310.00
Ours (VAD & Qwen-VL)0.410.841.370.870.060.150.250.1540.7610.93
Ours (UniAD & Qwen-VL)0.330.651.120.700.371.142.531.3438.8710.00
+ +Table 2. Open-loop and Closed-loop Planning Results on the Bench2Drive dataset. We reimplement models using the official code provided by the Bench2Drive paper. In the code, TCP plans for the next two seconds, while the other methods plan for the next three seconds. The best result is shown in bold, while the second-best is underlined. + +# 4.3. Ablation Study + +We conduct an ablation study on the Bench2Drive validation set to evaluate our proposed modules. Refer to VAD, we adopt the two-stage training strategy here to accelerate experimentation. All ablation models share the same stage-1 checkpoint for fair comparison and all experiments use NVIDIA 4090 GPUs with VAD-base and MiniCPM-Llama3-2.5V baselines. + +Effectiveness of Strategic Reasoning Injector. To as- + +sess the effectiveness of the Strategic Reasoning Injector module, we conducted an ablation study by removing this component. As shown in Table 3 (settings 0 and 1), its absence led to an approximately $12\%$ increase in average L2 error and a $19\%$ increase in average collision rate. These highlight that strategic decision-making guides ego-query modeling, allowing it to adaptively prioritize information relevant to the current scene decision-making process, improving overall planning performance. + +
SettingSRITRIHTDL2 (m) ↓Collision Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
00.510.981.551.010.130.190.310.21
1✓0.440.861.360.890.080.160.260.17
2✓0.520.931.450.970.080.140.270.16
3✓0.460.891.460.930.050.160.210.14
4✓✓0.540.921.390.950.050.110.290.15
5✓✓0.420.861.390.890.080.140.230.15
6✓✓0.440.931.500.960.080.110.240.14
7✓✓✓0.410.851.400.890.040.130.200.12
+ +Effectiveness of Tactical Reasoning Integrator. As shown in Table 3, Settings 2 and 6 validate the effectiveness of this module. In Setting 2, we concatenate the embedded command features with the modeled ego-query features, and the final output is obtained through the decoder. Comparing Settings 0 and 2, we observe a $0.14\mathrm{m}$ reduction in average L2 error and a $0.05\%$ decrease in average collision rate. These results suggest that tactical commands provide guidance that is closer to planning than strategic decisions, reducing the complexity of the learning space and enabling the network to make more informed decisions. + +Effectiveness of Hierarchical Trajectory Decoder. To highlight the importance of the Hierarchical Trajectory Decoder, we replace it with an MLP that directly predicts future trajectory based on ego-features. This modification leads to a $0.07\mathrm{m}$ increase in L2 error and a $0.07\%$ rise in the collision rate (Setting 0 vs. 3), demonstrating the challenges of directly decoding fine-grained trajectories. The absence of hierarchical decoding—from easy to difficult and coarse to fine—hinders the model's ability to refine trajectory predictions, ultimately suboptimal performance. + +# 4.4. More Analysis + +In this section, we further analyze our designed module on the Bench2Drive validation set. Settings are consistent with those in the ablation study. More analytical experiments and visualization for qualitative assessment of performance and interpretability, are provided in the Appendix. + +Dissusion on the use of similarity loss for encoded driving strategy features. As shown in Tab. 4, we removed the component of encoding ground truth trajectory and the similarity loss. We can observe that both L2 loss and collision rate increase, which can be attributed to the gap between the strategic strategy text features encoded by the VLM text encoder and the perception features used in the end-to-end network. The similarity loss helps the adapter bridge this gap, aligning the text features more closely with those required for trajectory prediction. This, in turn, better guides the learning of ego-query and improves performance. + +Table 3. Ablation study of each module on Bench2Drive dataset. "SRI" refers to Strategic Reasoning Injector module, "TRI" refers to Tactical Reasoning Integrator module and "HTD" means Hierarchical Trajectory Decoder module. The best result is in bold + +
ModelL2(m) ↓Collision Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
w/o loss0.450.881.430.920.090.160.270.18
w. loss0.440.861.360.890.080.160.260.17
+ +Table 4. Ablation study to evaluate the use of similarity loss for encoded driving strategy features. + +
#layersL2(m) ↓Collision Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
10.510.951.510.990.070.150.300.17
20.460.891.460.930.050.160.210.14
30.480.931.520.970.050.170.240.15
+ +Table 5. Ablation study for different numbers of layers in the hierarchical trajectory decoder. + +Dissusion on different numbers of layers in the hierarchical trajectory decoder. As shown in Tab. 5, we experiment with different numbers of layers in decoder. One layer represents the direct output of the fine-grained trajectory, while two layers follow the described methods. In the three-layer setting, we adjust the decoder to three layers and use Bezier curve fitting on the coarse trajectory to generate a coarser ground truth for joint supervision. Experiments show that increasing from one to two layers significantly improves performance, but too many layers introduce unnecessary complexity, leading to overfitting and loss of fine-grained details critical for accurate prediction. + +# 5. Conclusions + +In conclusion, our paper presents ReAL-AD, a reasoning-augmented learning framework that enhances end-to-end autonomous driving by leveraging Vision-Language Models (VLMs) for structured reasoning across strategy, decision, and operation levels. By modeling human-like hierarchical decision-making, ReAL-AD integrates strategic decisions with tactical commands and trajectory refinement. Extensive experiments on NuScenes and Bench2Drive demonstrate its state-of-the-art performance in trajectory planning accuracy and driving safety. + +# 6. Acknowledgement + +This work was supported by NSFC (No.62206173), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (ShangHAI), and MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (KLIP-HuMaCo). + +# 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. 1, 3 +[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, 2023.1, 3, 6 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 2, 6 +[4] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 3 +[5] Long Chen, Oleg Sinavski, Jan Hunermann, 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 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 3 +[6] Shaoyu Chen, Bo Jiang, Hao Gao, Benchcheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2 +[7] Zhili Chen, Maosheng Ye, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Ppad: Iterative interactions of prediction and planning for end-to-end autonomous driving. 2 +[8] DeepSeek-AI. Deepseek II: Scaling open-source language models with longtermism. arXiv preprint arXiv:2401.02954, 2024. 1, 3 +[9] Simon Doll, Niklas Hanselmann, Lukas Schneider, Richard Schulz, Marius Cordts, Markus Enzweiler, and Hendrik Lensch. Dualad: Disentangling the dynamic and static world for end-to-end driving. In CVPR, pages 14728-14737, 2024. 2 +[10] Team GLM et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools, 2024. 3 +[11] Mingzhe Guo, Zhipeng Zhang, Yuan He, Ke Wang, and Liping Jing. End-to-end autonomous driving without costly modularization and 3d manual annotation. arXiv preprint arXiv:2406.17680, 2024. 2, 7 + +[12] 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 +[13] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In CVPR, pages 12732-12741, 2021. 7 +[14] 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, pages 533-549. Springer, 2022. 1, 2, 7 +[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, pages 17853-17862, 2023. 1, 2, 6, 7 +[16] Xin Huang, Eric M Wolff, Paul Vernaza, Tung Phan-Minh, Hongge Chen, David S Hayden, Mark Edmonds, Brian Pierce, Xinxin Chen, Pratik Elias Jacob, et al. Drivegpt: Scaling autoregressive behavior models for driving. arXiv preprint arXiv:2412.14415, 2024. 3 +[17] Zhijian Huang, Tao Tang, Shaoxiang Chen, Sihao Lin, Zequn Jie, Lin Ma, Guangrun Wang, and Xiaodan Liang. Making large language models better planners with reasoning-decision alignment. In European Conference on Computer Vision, pages 73-90. Springer, 2024. +[18] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262, 2024. 3 +[19] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. arXiv preprint arXiv:2406.03877, 2024. 2, 6 +[20] Bo Jiang, Shaoyu Chen, Qing Xu, Benchcheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, pages 8340-8350, 2023. 1, 2, 6, 7 +[21] Bo Jiang, Shaoyu Chen, Benchcheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving. arXiv preprint arXiv:2410.22313, 2024. 3, 7 +[22] Tarasha Khurana, Peiyun Hu, Achal Dave, Jason Ziglar, David Held, and Deva Ramanan. Differentiable raycasting for self-supervised occupancy forecasting. In ECCV, pages 353-369. Springer, 2022. 7 +[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 3 +[24] Peidong Li and Dixiao Cui. Does end-to-end autonomous driving really need perception tasks? arXiv preprint arXiv:2409.18341, 2024. 2 +[25] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end + +autonomous driving with latent world model. arXiv preprint arXiv:2406.08481, 2024. 2 +[26] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In CVPR, 2024. 3 +[27] Bencheng Liao, Shaoyu Chen, Haoran Yin, Bo Jiang, Cheng Wang, Sixu Yan, Xinbang Zhang, Xiangyu Li, Ying Zhang, Qian Zhang, et al. Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving. arXiv preprint arXiv:2411.15139, 2024. 2 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 3 +[29] Jiaqi Liu, Peng Hang, Xiao Qi, Jianqiang Wang, and Jian Sun. Mtd-gpt: A multi-task decision-making gpt model for autonomous driving at unsignalized intersections. In 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC), pages 5154-5161. IEEE, 2023. 3 +[30] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. In ECCV, pages 403-420. Springer, 2024. 3 +[31] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 3 +[32] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 1, 3 +[33] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. Vlp: Vision language planning for autonomous driving. In CVPR, pages 14760-14769, 2024. 1, 3, 7 +[34] Tim A Pastva. Bezier curve fitting. PhD thesis, Monterey, California. Naval Postgraduate School, 1998. 6 +[35] Kangan Qian, Zhikun Ma, Yangfan He, Ziang Luo, Tianyu Shi, Tianze Zhu, Jiayin Li, Jianhui Wang, Ziyu Chen, Xiao He, et al. Fasionad: Fast and slow fusion thinking systems for human-like autonomous driving with adaptive feedback. arXiv preprint arXiv:2411.18013, 2024. 3 +[36] 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. 3 +[37] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In CVPR, pages 15120-15130, 2024. +[38] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beißwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In ECCV, pages 256-274. Springer, 2024. 3 +[39] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 1, 2 + +[40] 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. 3 +[41] Ran Tian, Boyi Li, Xinshuo Weng, Yuxiao Chen, Edward Schmerling, Yue Wang, Boris Ivanovic, and Marco Pavone. Tokenize the world into object-level knowledge to address long-tail events in autonomous driving. arXiv preprint arXiv:2407.00959, 2024. 3 +[42] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 3, 7 +[43] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, pages 8406-8415, 2023. 7 +[44] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 3 +[45] 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. 1, 3 +[46] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1, 3 +[47] 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. arXiv preprint arXiv:2309.16292, 2023. 3 +[48] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In CVPR, pages 15449-15458, 2024. 1, 2, 7 +[49] Penghao 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. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 7 +[50] Shuo Xing, Chengyuan Qian, Yuping Wang, Hongyuan Hua, Kexin Tian, Yang Zhou, and Zhengzhong Tu. Openemma: Open-source multimodal model for end-to-end autonomous driving. arXiv preprint arXiv:2412.15208, 2024. 3 +[51] Yi Xu, Yuxin Hu, Zaiwei Zhang, Gregory P Meyer, Siva Karthik Mustikovela, Siddhartha Srinivasa, Eric M Wolff, and Xin Huang. Vlm-ad: End-to-end autonomous driving through vision-language model supervision. arXiv preprint arXiv:2412.14446, 2024. 1, 3, 7 + +[52] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 1, 3 +[53] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 3 +[54] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 3, 6 +[55] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In CVPR, pages 8660-8669, 2019. 7 +[56] Jiang-Tian Zhai, Ze Feng, Jinhao 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 +[57] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. In ECCV, pages 87-104. Springer, 2024. 1, 2, 5, 7 \ No newline at end of file diff --git a/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/images.zip b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..144d1360bf4b2d501cc5c3041b7b17bdade25b3d --- /dev/null +++ b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cc5524dfde4ba29e5101c1eeabf8a29ec370b8ee5b81aea0a800cbcbdf1dd65 +size 632615 diff --git a/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/layout.json b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe92d5d3de91d2ad2685127c8c115024bed9860c --- /dev/null +++ b/ICCV/2025/ReAL-AD_ Towards Human-Like Reasoning in End-to-End Autonomous Driving/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2131afa12452e9c77c3d530babaf2c9d7a27325a0cb6f977f8d7cc020f24105c +size 411533 diff --git a/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/1a50cae9-992f-4437-a481-423135a77892_content_list.json b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/1a50cae9-992f-4437-a481-423135a77892_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f7c95d444df3828f5d6f364da2b150ef06938570 --- /dev/null +++ b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/1a50cae9-992f-4437-a481-423135a77892_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ecb1ec8c8f3bc0809e7288223276a7eeb01ced7bba9f03fa5dc6df36e70996f +size 80930 diff --git a/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/1a50cae9-992f-4437-a481-423135a77892_model.json b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/1a50cae9-992f-4437-a481-423135a77892_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ab596364a4fe2bd683365ce8ced96f2f477ce17 --- /dev/null +++ b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/1a50cae9-992f-4437-a481-423135a77892_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:423d0aa11d29e03379508171ae11b68e25146276210c42148d713c80fa6ff540 +size 103377 diff --git a/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/1a50cae9-992f-4437-a481-423135a77892_origin.pdf b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/1a50cae9-992f-4437-a481-423135a77892_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..454b26a2fc7ade188de7ca1137ab168fbe3685b1 --- /dev/null +++ b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/1a50cae9-992f-4437-a481-423135a77892_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59d55553efeff25d6e0659eb43ee0b11b2e1fdb18c30c8212aa2aed911d8ab80 +size 3111254 diff --git a/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/full.md b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a84b0c8045235c3013212be568d754a85b1efacd --- /dev/null +++ b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/full.md @@ -0,0 +1,318 @@ +# ReCamMaster: Camera-Controlled Generative Rendering from A Single Video + +Jianhong Bai $^{1*}$ , Menghan Xia $^{2\dagger}$ , Xiao Fu $^{3}$ , Xintao Wang $^{2}$ , Lianrui Mu $^{1}$ , Jinwen Cao $^{2}$ , Zuozhu Liu $^{1}$ , Haoji Hu $^{1\dagger}$ , Xiang Bai $^{4}$ , Pengfei Wan $^{2}$ , Di Zhang $^{2}$ + +1Zhejiang University, 2Kling Team, Kuaishou Technology, 3CUHK, 4HUST + +Project webpage: https://jianhongbai.github.io/ReCamMaster/ + +![](images/f9c162ba7903ce60c43155be350cc20c9d31286b9678c99412a7de1c9fd02934.jpg) + +![](images/657dc5a1929bddb652aebfd38b5043eed1c0a5ff69c289924df6478fea4f3cdc.jpg) + +![](images/21b386480550edca3f61bf287a64dd9b1bfd5f9094477cf648eb2d7f3e2b8434.jpg) +Figure 1. Examples synthesized by ReCamMaster. ReCamMaster re-shoots the source video with novel camera trajectories. We visualized the novel camera trajectories alongside the video frames. Video results are on our project page. + +# Abstract + +Camera control has been actively studied in text or image conditioned video generation tasks. However, altering camera trajectories of a given video remains under-explored, despite its importance in the field of video creation. It is non-trivial due to the extra constraints of maintaining multiple-frame appearance and dynamic synchronization. To address this, we present ReCamMaster, a camera-controlled generative video re-rendering framework that reproduces the dynamic scene of an input video at novel camera trajectories. The core innovation lies in harnessing the generative capabilities of pre-trained text-to-video models + +through a simple yet powerful video conditioning mechanism—its capability is often overlooked in current research. To overcome the scarcity of qualified training data, we construct a comprehensive multi-camera synchronized video dataset using Unreal Engine 5, which is carefully curated to follow real-world filming characteristics, covering diverse scenes and camera movements. It helps the model generalize to in-the-wild videos. Lastly, we further improve the robustness to diverse inputs through a meticulously designed training strategy. Extensive experiments show that our method substantially outperforms existing state-of-the-art approaches. Our method also finds promising applications in video stabilization, super-resolution, and outpainting. Our code and dataset are publicly available at: https://github.com/KwaiVGI/ReCamMaster. + +# 1. Introduction + +Camera movement is a fundamental element in film production, profoundly shaping the audience's visual experience and conveying both emotional depth and narrative intentions. For instance, a dolly-in shot can masterfully emphasize a character, building tension or guiding viewers' attention to crucial details. Similarly, crane shots excel at showcasing expansive landscapes, often employed at the opening to dramatically reveal the full scope of a setting. Despite its artistic significance, achieving professional-level camera movement remains a challenge for amateur videographers, because of hardware limitations (such as stabilization issues in handheld recordings) and technical skill gaps. To address this dilemma, we explore innovative approaches to modify camera trajectories in post-production, empowering creators with the ability to enhance their original footage by displaying the dynamic scenes with more compelling and polished camera trajectories when needed. + +Previous research on camera-controlled generation has primarily focused on text-to-video (T2V) or image-to-video (I2V) generation [18, 51]. Recently, GCD [45] pioneered camera-controlled video-to-video generation, achieving promising results on domain-specific videos synthesized by the Kubric simulator [14]. However, its effectiveness on real-world videos is limited due to the narrow domain of training data and its inferior video conditioning mechanism. As a concurrent work, ReCapture [65] introduces a novel two-stage approach: first generating a camera-controlled anchor video using existing image-level multiview diffusion models, then optimizing the result by fine-tuning spatial and temporal LoRA [21] layers over the input and anchor videos respectively. While showing promising performance, its requirement for per-video optimization constrains practical applications. Similarly, other approaches [5, 15] achieve video re-rendering through explicit 4D reconstruction followed by video post-optimization. However, their performance is significantly restricted by the challenge of single video-based 4D reconstruction techniques [26, 53]. + +To address these challenges, we present ReCamMaster, a camera-controlled generative video re-rendering framework that can regenerate in-the-wild videos at novel camera trajectories. The core innovation is to utilize the generative capabilities of pre-trained text-to-video models through an elegant yet powerful video conditioning mechanism. Compared to alternatives, our method shows superior performance in preserving both the visual and dynamic characteristics of the input video. Notably, this conditioning mechanism indeed exhibits remarkable potential as a versatile solution for conditional generation tasks, which is overlooked in current research. Since no qualified training data are publicly available, we develop a large-scale multi-camera synchronized video dataset using Unreal Engine 5, which contains 136K realistic videos shot from 13.6K different + +dynamic scenes in 40 high-quality 3D environments with $122\mathrm{K}$ different camera trajectories. Particularly, it is carefully curated to simulate real-world filming characteristics, which is proven to bring an advantage to striking domain alignment to in-the-wild videos. Lastly, we further improve the robustness of our model to diverse inputs through a meticulously designed training strategy. + +Experimental results show that our method outperforms state-of-the-art approaches and strong baselines by a large margin. Ablation studies verified the effectiveness of our key designs. Furthermore, ReCamMaster demonstrates its potential in scenarios like video stabilization, video superresolution, and video outpainting. Our contribution can be summarized as follows: + +- We introduce a high-quality multi-camera synchronized video dataset featuring diverse camera trajectories. This dataset has been publicly released to advance research in camera-controlled video generation, 4D reconstruction, and related fields. +- We conduct an in-depth investigation and validation of an effective video conditioning mechanism for text-to-video generation models. It significantly outperforms other alternatives employed in baseline methods. +- Extensive experiments show that our proposed ReCam-Master significantly advances the state-of-the-art in video recapturing. Furthermore, it finds promising applications across multiple real-world scenarios. + +# 2. Related Works + +Camera-Controlled Video Generation. With the success of text-to-video generation models [6, 7, 10, 35, 62], the introduction of other conditional signals for controllable video generation has been widely studied [11, 16, 56, 63]. In camera-controlled video generation [2, 61, 67, 68], researchers aim to incorporate camera parameters into video generation models to control the viewpoint of the output video.AnimateDiff [17] introduces various motion LoRAs [21] to learn specific patterns of camera movements. MotionCtrl [51] encodes 6DoF camera extrinsics and injects them into the diffusion model, fine-tuning on video-camera pair data to achieve video generation with arbitrary trajectory control. CameraCtrl [18] further improves the accuracy and generalizability of single-sequence camera control with a dedicatedly designed camera encoder. CVD [28] achieves multi-sequence camera control with the proposed cross-video synchronization module. AC3D [1] conducts an in-depth investigation into camera motion knowledge within diffusion transformers, achieving camera-controlled generation with enhanced visual quality. Meanwhile, training-free methods have also been explored [20, 22, 30]. + +Video-to-Video Generation. Video-to-video generation [34, 48, 49] is explored in various tasks, including video + +editing [33, 60], video outpainting [8, 46], video superresolution [58, 69], etc. Most relevant to our work is synthesizing novel viewpoint videos based on a given video and camera parameters [5, 15, 39, 45, 65]. GCD [45] pioneered camera-controlled video-to-video generation by training a video generation model with paired video data created by the Kubric simulator [14]. Although it performs well on in-domain data, GCD's generalization capability is limited by the significant domain gap between the training data and real-world videos. Recapture [65] proposes a practical solution with per-video optimization. Several concurrent works [5, 15, 39, 54, 64], achieve 4D-consistent video-to-video generation by extracting dynamic information from the original video using 3D point tracking [26, 53] and incorporating it as a condition into the video generator. This paradigm is promising when synchronized multi-view videos are unavailable. Nevertheless, the generation quality is limited by the accuracy of the point-tracking methods. In the field of 4D object generation, video-to-video generation can be achieved by 1) training a multi-view video generator [52, 55], or 2) following a reconstruct-and-render pipeline via 4D reconstruction methods [29, 66]. [32, 42] achieve 3D/4D scene reconstruction with videos generated with diffusion models. In this paper, we focus on open-domain camera-controlled video generation. To achieve this, we first create a large multi-camera synchronized dataset with diverse camera trajectories. We also propose a novel video conditioning method to achieve better generation ability. + +# 3. Multi-Cam Video: A High-Quality Multi-Camera Synchronized Video Dataset + +To re-shoot input videos based on novel camera trajectories, paired video data is required for training the video generation model. Specifically, the training data should include multiple shots captured in the same scene simultaneously, allowing the model to learn 4D consistent generation. Acquiring such data in real-world scenarios is extremely costly, and publicly available multi-view synchronized datasets [13, 24, 25, 40, 59] are also limited by the diversity of scenes and constrained camera movements, making them unsuitable for our task. Therefore, we chose to use a rendering engine to generate the training data. The advantages of this approach are: 1) precise camera trajectories can be obtained, 2) complete synchronization in the time dimension can be achieved, and 3) the data volume can be more easily scaled up. + +We build the entire data rendering pipeline in Unreal Engine 5 [12]. Specifically, we first collect multiple 3D environments as "backgrounds". Then, we place animated characters within these environments as the "main subjects" of the videos. We then position multiple cameras facing the subjects and moving along predefined trajectories to simulate the process of simultaneous shooting. This allows us to + +![](images/734f1d1fd69b98c47fcaeb58615d8f59fefd56909bffc0e7de2afd97ab5470c2.jpg) +(a) 3D Environments + +![](images/84e2bf6805b66a3bfc435cfb5327b8d4c0463aa0b81da8a02149f94a17e61b89.jpg) +(b) Characters + +![](images/d7be86212865d36f58f221d57ab7b230975953193aa6fb1de7de4eedeb176e08.jpg) +(c) Animations + +![](images/be641963c4518780eb203a9e7d4df41e0f8b4a81d316b6cee7f115aeef42b00d.jpg) +(d) Camera Trajectories +Figure 2. Illustration of the dataset construction process. We build the multi-camera synchronized training dataset by rendering in Unreal Engine 5. This is achieved using 3D environments, characters, animations collected from the internet, and our designed massive camera trajectories. + +render datasets with synchronized cameras that include dynamic objects. To scale up the data volume, we construct a set of camera movement rules for automatically batch generation of natural and diverse camera trajectories. Additionally, we randomly combined different characters and actions across different video sets. In total, we obtain 136K visually-realistic videos shot from 13.6K different dynamic scenes in 40 high-quality 3D environments with 122K different camera trajectories. Fig. 8 in the Appendix shows some of the rendered video frames. In Appendix C.1, we conduct an ablation study comparing model performance when trained on a dataset with limited scenes and camera trajectories versus our comprehensive dataset. This highlights the significance of our high-quality dataset. For more details about the dataset, please refer to Appendix B. + +# 4. Camera-Controlled Video Re-Generation + +Given a source video $V_{s} \in \mathbb{R}^{f \times c \times h \times w}$ , we aim to synthesize a target video $V_{t} \in \mathbb{R}^{f \times c \times h \times w}$ sharing the same dynamic scene but presenting at specified camera trajectories denoted by $\mathrm{cam}_t \coloneqq [R, t] \in \mathbb{R}^{f \times 3 \times 4}$ . Particularly, $V_{t}$ should conform to the multiple-frame appearance and synchronized dynamics of $V_{s}$ . To achieve this, we propose to harness the generative capability of pre-trained text-to-video diffusion models [7, 62] by imposing dual conditions, i.e., the source video and target camera trajectories through a meticulously designed framework. The overview of the model is depicted in Fig. 3. + +# 4.1. Preliminary: Text-to-Video Base Model + +Our study is conducted over an internal pre-trained text-to-video foundation model. It is a latent video diffusion model, consisting of a 3D Variational Auto-Encoder (VAE) [27] and a Transformer-based diffusion model (DiT) [37]. Typ + +![](images/de0fc10d4ae44d01dfb123efdc5361c9fe394a506245e43a14ea1536fa99c0dd.jpg) +Figure 3. Overview of ReCamMaster. Left: The training pipeline of ReCamMaster. A latent diffusion model is optimized to reconstruct the target video $V_{t}$ , conditioned on the source video $V_{s}$ , target camera pose $cam_{t}$ , and target prompt $p_t$ . Right: Comparison of different video condition techniques. (a) Frame-dimension conditioning used in our paper; (b) Channel-dimension conditioning used in baseline methods [5, 45]; (c) View-dimension conditioning in [3]. We omit the text prompt $p_t$ in (a)-(c) for simplicity. + +![](images/e82dc49cdf5b8ad14e2312fc4e795c4cb74e3ebcd932c42e698beda2c0ba880b.jpg) + +![](images/242ad2a66fb4f62222bc2fe8163e839df7940130a25d588b542e3e00dd494d86.jpg) + +ically, each Transformer block is instantiated as a sequence of spatial attention, 3D (spatial-temporal) attention, and cross-attention modules. The generative model adopts Rectified Flow framework [9] for the noise schedule and denoising process. The forward process is defined as straight paths between data distribution and a standard normal distribution, i.e. + +$$ +z _ {t} = (1 - t) z _ {0} + t \epsilon , \tag {1} +$$ + +where $\epsilon \in \mathcal{N}(0,I)$ and $t$ denotes the iterative timestep. To solve the denoising processing, we define a mapping between samples $z_{1}$ from a noise distribution $p_1$ to samples $z_{0}$ from a data distribution $p_0$ in terms of an ordinary differential equation (ODE), namely: + +$$ +d z _ {t} = v _ {\Theta} \left(z _ {t}, t\right) d t, \tag {2} +$$ + +where the velocity $v$ is parameterized by the weights $\Theta$ of a neural network. For training, we regress a vector field $u_{t}$ that generates a probability path between $p_0$ and $p_1$ via Conditional Flow Matching [31]: + +$$ +\mathcal {L} _ {L C M} = \mathbb {E} _ {t, p _ {t} (z, \epsilon), p (\epsilon)} | | v _ {\Theta} \left(z _ {t}, t\right) - u _ {t} \left(z _ {0} \mid \epsilon\right) | | _ {2} ^ {2}, \tag {3} +$$ + +where $u_{t}(z,\epsilon) \coloneqq \psi_{t}^{\prime}(\psi_{t}^{-1}(z|\epsilon)|\epsilon)$ with $\psi (\cdot |\epsilon)$ denotes the function of Eq. 1. For inference, we employ Euler discretization for Eq. 2 and perform discretization over the timestep interval at [0, 1], starting at $t = 1$ . We then processed with iterative sampling with: + +$$ +z _ {t} = z _ {t - 1} + v _ {\Theta} \left(z _ {t - 1}, t\right) * \Delta t. \tag {4} +$$ + +# 4.2. Conditional Video Injection Mechanism + +To modify the camera trajectories of a given video using generative models, it is essential to condition the generation process on the source video. Our experiments reveal that the video injection mechanism is crucial for + +overall performance. We analyze and compare several widely-adopted video conditioning approaches from previous works [3, 5, 45], alongside our proposed simple yet powerful video injection mechanism. We provide an intuitive explanation of our design's superior performance and its potential impact. + +Channel Dimension Conditioning. Recent attempts [5, 45] incorporate the source video by concatenating the latent of the source video $x_{s}$ with the noised latent of the target video $x_{t}$ along the channel dimension, and add additional input channels to the input convolutional layer. To be specific, a variational autoencoder with encoder $\mathcal{E}$ is utilized to project the source video $V_{s}$ and target video $V_{t}$ to the latent space, $z_{s} = \mathcal{E}(V_{s}), z_{t} = \mathcal{E}(V_{t})$ , where $z_{s}, z_{t} \in \mathbb{R}^{b \times f \times c \times h \times w}$ are the latent of source video and target video with $f$ frames, $c$ channels, and spatial size of $h \times w$ . Then, the source latent is concentrated with the noised target latent along the channel dimension, and patched into $f \times h \times w$ tokens with several convolutional layers: + +$$ +x _ {t} = \operatorname {p a t c h i f y} \left(\left[ z _ {s}, z _ {t} \right] _ {\text {c h a n n e l - d i m}}\right), \tag {5} +$$ + +where $x_{t} \in \mathbb{R}^{b \times f \times s \times d}$ , $s = h \times w$ , $d$ is the channel dimension for the latent diffusion model. We denote this conditioning technique as the "channel-dimension conditioning" and illustrate in Fig. 3(b). + +View Dimension Conditioning. To achieve multi-view video generation, [3] proposes to introduce a plug-and-play module for feature aggregation across views. Given a source video latent $z_{s} = \mathcal{E}(V_{s})$ and a target video latent $z_{t} = \mathcal{E}(V_{t})$ , and patchify and forward to the diffusion transformer respectively: + +$$ +x _ {s} = \text {p a t c h i f y} \left(z _ {s}\right), x _ {t} = \text {p a t c h i f y} \left(z _ {t}\right), \tag {6} +$$ + +where $x_{s}, x_{t} \in \mathbb{R}^{b \times f \times s \times d}$ . To achieve synchronization and content consistency across views, an additional attention layer (denoted as view-attention) is added in each basic transformer block, it performs self-attention between each frame in multiple views: + +$$ +\overline {{F}} _ {s} ^ {i}, \overline {{F}} _ {t} ^ {i} = \mathrm {a t t n \_ v i e w} (F _ {s} ^ {i}, F _ {t} ^ {i}), \tag {7} +$$ + +where $F_{s}^{i}$ and $F_{t}^{i}$ are the $i$ -th frame feature of $V_{s}$ and $V_{t}$ respectively, and $\overline{F}_s^i$ and $\overline{F}_t^i$ are the output features. We denote this conditioning technique as the "view-dimension conditioning" and illustrated in Fig. 3(c). + +Frame Dimension Conditioning (ours). To achieve better synchronization and content consistency with the source video, we propose to concatenate the source video tokens with the target video tokens along the frame dimension: + +$$ +\left\{ \begin{array}{l l} x _ {s} = \operatorname {p a t c h i f y} \left(z _ {s}\right), & x _ {t} = \operatorname {p a t c h i f y} \left(z _ {t}\right), \\ x _ {i} = \left[ x _ {s}, x _ {t} \right] _ {\text {f r a m e - d i m}}, \end{array} \right. \tag {8} +$$ + +where $x_{i} \in \mathbb{R}^{b \times 2f \times s \times d}$ is the input of diffusion transformer. In other words, the input token number is doubled compared to the vanilla text-to-video generation process. Moreover, we do not introduce additional attention layers for feature aggregation across the source video and the target video, since self-attention is performed between all tokens in the 3D (spatial-temporal) attention layers. We denote this conditioning technique as the "frame-dimension conditioning" and illustrate in Fig. 3(a). + +Comparison and Discussion. Our experimental results demonstrate that the frame dimension conditioning approach provides substantial advantages in utilizing condition information effectively, as illustrated by the comparative analysis in Fig. 5. We attribute this superior performance to the inherent flexibility of our frame dimension conditioning method, which enables spatio-temporal interaction between the conditional tokens and target tokens through all blocks of the base model. This approach provides a more robust mechanism for understanding the correlations between video pairs. While similar token concatenation strategies have proven effective in image-to-image tasks [43], their potential in video controllable generation has remained largely unexplored. Our study provides compelling evidence of the effectiveness, highlighting the potential of this underappreciated technique as a versatile solution for conditional generation tasks. + +# 4.3. Camera Pose Conditioning + +To achieve camera-controlled video generation, a natural idea is to condition the model on the source and target video camera trajectories $\mathrm{cam}_s$ and $\mathrm{cam}_t$ to help the model + +better understand the 4D space. Unfortunately, during inference, even with state-of-the-art structure-from-motion (SfM) methods for camera parameter estimation, it is challenging to obtain accurate camera trajectory information for the input video [1]. Therefore, we alternate to only conditioning the model on the target camera $\mathrm{cam}_t$ , while relying on the model to interpret the camera trajectory of the input video. Furthermore, we condition the model using camera extrinsics, specifically the rotation and translation matrices for each frame's camera. We chose not to include camera intrinsics as a condition because, while they are easily obtainable in the rendered training set, accurately estimating the intrinsics of real-world video cameras remains challenging. This limitation would complicate parameter provision in practical applications, as users are unlikely to have access to the source video's camera intrinsics. However, our method can be readily adapted to incorporate intrinsics as input with minimal modifications. + +Given a $f$ frames target camera sequence $\mathsf{cam}_t\in \mathbb{R}^{f\times (3\times 4)}$ , we project it to have the same channels with the video tokens through a learnable camera encoder $\mathcal{E}_c$ and add it to the visual features: + +$$ +F _ {t} = F _ {o} + \mathcal {E} _ {c} (\operatorname {c a m} _ {t}), \tag {9} +$$ + +where $F_{o}$ is the output feature of the spatial-attention layer, $F_{i}$ is the input feature of the 3D-attention layer, $\mathcal{E}_c$ is instantiated as a fully connected layer with an input dimension of 12 and an output dimension of $d$ in practice. The camera encoder is inserted into each transformer block for fine-grained camera control. + +# 4.4. Training Strategy + +Enhancing Generalization Capabilities. Using the annotated dataset described in Sec. 3, ReCamMaster is trained to process input videos with target camera parameters. To preserve the base T2V model's native capability, we fine-tune only the camera encoder and 3D-attention layers while keeping other parameters frozen. To mitigate Unreal Engine's synthetic characteristics, we apply moderate noise (200-500 noise scheduling steps) to the conditional video latent during training, reducing the domain gap between synthetic and real-world data at inference. + +Improving Generation Capability by Unifying Camera Control Tasks. To encourage content generation capability, we implement T2V camera-controlled generation [18, 51] with a $20\%$ probability and I2V camera-controlled generation [57] with a $20\%$ probability during training. Specifically, we replace the latent representations of all $f$ frames with Gaussian noise for T2V generation, and replace $f - 1$ frames starting from the second frame for I2V generation. Our experiments demonstrate that this strategy prompts performance in synthesizing coherent objects invisible in the source video. As a byproduct, our model + +![](images/01ea366e5ab16aa297a91a57d5adaaf2e481c0a3e99d5e6ce7b2c8cf71e796ab.jpg) +Figure 4. Comparison with state-of-the-art methods. It shows that ReCamMaster generates videos that maintain appearance consistency and temporal synchronization with the source video. + +inherently supports T2V, I2V, and V2V camera-controlled generation, as shown in Fig. 9. + +# 5. Experimental Results + +# 5.1. Experiment Settings + +Implementation Details. We train ReCamMaster on the rendered dataset introduced in Sec. 3. During training, we randomly select 2 cameras from each 10 synchronized cameras, designating one as the source video and the other as the target video. We use $\{\mathrm{source~video}V_s\}$ , target camera $\mathrm{cam}_t$ , target prompt $p_t\}$ as conditions input to the model, to reconstruct the target video $V_t$ as introduced in Eq. 3. We train the model for 10K steps at the resolution of $384\times 672$ with a learning rate of 0.0001, batch size 40. The camera encoder and the projector are zero-initialized. + +Evaluation Metrics. We mainly evaluate the proposed method in terms of camera accuracy, source-target synchronization, and visual quality. For camera accuracy, we use GLOMAP [36] to extract the camera pose sequence of the generated videos, and calculate the rotation error and translation error, denoted as RotErr and TransErr respectively [18]. In terms of synchronization, we utilize the state-of-the-art image matching method GIM [41] to calculate the number of matching pixels with confidence greater than the threshold, denoted as Mat. Pix.. Furthermore, we calculate the FVD-V score in SV4D [55], and the average CLIP similarity between source and target frames at the same timestamp, denoted as CLIP-V [28]. For visual quality, we divide it into fidelity, coherence with text, and temporal consistency, and quantify them with Fréchet Image Distance + +[19] (FID) and Fréchet Video Distance [44] (FVD), CLIP-T, and CLIP-F, respectively. CLIP-T refers to the average CLIP [38] similarity of each frame and its corresponding text prompt, and CLIP-F is the average CLIP similarity of adjacent frames. We also evaluate our method on the widely used VBench [23] metrics. + +Evaluation Set. We construct the evaluation set with 1000 random videos from WebVid [4] and 10 different camera trajectories, including pan, tilt, vertical translation, zoom in, zoom out, and horizontal arc trajectories. Note that since we automatically generated 122K different camera trajectories when constructing the training set, ReCamMaster can support a variety of camera trajectories as input. However, since the baseline method was trained on specific basic trajectories, we opted to use trajectories that closely align with the baseline for comparison. + +# 5.2. Comparison with State-of-the-Art Methods + +Baselines. We compare the proposed ReCamMaster with state-of-the-art camera-controlled video-to-video generation methods [15, 45, 54]. GCD [45] pioneered camera-controlled video-to-video generation by training a video generation model with paired data created by the Kubric simulator [14]. Although it performs well on in-domain data, GCD's generalization capability is limited by its inferior video conditioning method and the significant domain gap between the training data and real-world videos. Trajectory-Attention [54] and DaS [15] extract dynamic information from the source video using 3D point tracking and incorporate it as a condition into the video generator. + +Table 1. Quantitative comparison with state-of-the-art methods on visual quality, camera accuracy, and view synchronization. + +
MethodVisual QualityCamera AccuracyView Synchronization
FID ↓FVD ↓CLIP-T ↑CLIP-F ↑RotErr ↓TransErr ↓Mat. Pix.(K) ↑FVD-V ↓CLIP-V ↑
GCD72.83367.3232.8695.662.275.51639.39365.7585.92
Trajectory-Attention69.21276.0633.4396.522.185.32619.13256.3088.65
DaS63.25159.6033.0598.321.455.59633.53154.2587.33
ReCamMaster57.10122.7434.5398.741.224.85906.0390.3890.36
+ +Table 2. Quantitative comparison with state-of-the-art methods on VBench [23] metrics. + +
MethodAesthetic Quality ↑Imaging Quality ↑Temporal Flickering ↑Motion Smoothness ↑Subject Consistency ↑Background Consistency ↑
GCD38.2141.5695.8198.3788.9492.00
Trajectory-Attention38.5051.0095.5298.2190.6092.83
DaS39.8651.5597.4499.1490.3492.03
ReCamMaster42.7053.9797.3699.2892.0593.83
+ +Qualitative Results. We present synthesized examples of ReCamMaster in Fig. 1 (additional examples in Fig. 12 of Appendix C). Please visit our project page for more videos. ReCamMaster demonstrates the ability to: 1) generate consistent content from novel camera trajectories of the same scene; 2) achieve excellent temporal synchronization with the source video; and 3) generate plausible video content in areas not visible in the original video. Note that, due to the use of $122\mathrm{K}$ different random trajectories during training, ReCamMaster also supports complex trajectories as input, such as zigzag paths. Please refer to the results on the project page. In the paper, to visually capture the camera movement trajectory more clearly, we do not showcase results generated along complex trajectories. + +We compare ReCamMaster with state-of-the-art methods in Fig. 4 and Fig. 13. It's observed that baselines produce content with notable artifacts and temporal desynchronization. Similar phenomena were noted when training ReCamMaster with other video conditioning techniques (as shown in Fig. 5), highlighting the importance of the novel video conditioning method proposed in the paper. + +Quantitative Results. We quantitatively evaluate ReCamMaster against baselines using various automatic metrics, the summarized results are in Tab. 1 and 2. For visual quality assessment, we used the widely adopted FID, FVD, and CLIP metrics as shown in Tab. 1, along with the VBenchmark metrics in Tab. 2. In terms of camera trajectory accuracy, we calculate the rotation error and translation error by following previous work in camera-controlled T2V and I2V generation [18, 57]. For view synchronization, we calculate the clip similarity score and FVD between video frames of different viewpoints within one scene, denoted as CLIP-V and FVD-V. It's observed that ReCamMaster outperforms baselines across multiple dimensions of metrics. + +![](images/fad86149e06377e7e62708d3a908a75b6864770a32db88ea17caa68007a70d7f.jpg) +Figure 5. Ablation on video conditioning techniques. We compared the channel-/view- concatenation schemes proposed by previous methods and frame-concatenation in ReCamMaster. We observed that both channel-conditioning and view-conditioning schemes suffer from significant artifacts, content inconsistency, and asynchronous dynamics with respect to the original video. + +# 5.3. More Analysis and Ablation Studies + +Ablation on Video Conditioning Techniques. In Sec. 4.2, we introduce a novel video conditioning scheme that concatenates the tokens of a source video with the target video tokens along the frame dimension. To verify its effectiveness, we compare our "frame concatenation" technique with the "channel concatenation" used in baseline methods [5, 45] and the "view concatenation" in [3]. We performed both qualitative and quantitative comparisons, as shown in Fig. 5 and Tab. 3. We only altered the video conditioning method while keeping the camera injection method, training data, and training steps unchanged. The results clearly show that our conditioning technique significantly enhances the model's performance. In the second and third rows of Fig. 5, the generated videos exhibit the rendering style of the + +Table 3. Quantitative comparison on the video conditioning strategy. + +
MethodVisual QualityCamera AccuracyView Synchronization
FID ↓FVD ↓CLIP-T ↑CLIP-F ↑RotErr ↓TransErr ↓Mat. Pix.(K) ↑FVD-V ↓CLIP-V ↑
Channel dim.74.09187.9433.7098.671.284.98521.10148.5184.62
View dim.80.51194.4734.6698.711.425.77573.92177.6883.40
Frame dim. (ours)57.10122.7434.5398.741.224.85906.0390.3890.36
+ +Table 4. Ablation on our training strategies. + +
MethodFID ↓FVD ↓Aesthetic Quality ↑Imaging Quality ↑
Baseline66.67171.8040.0251.93
+ Add noise65.17164.0440.3652.22
+ 3D-Attn. tuning59.47132.5843.0852.80
+ Drop latent62.39149.5241.4752.65
+ All57.10122.7442.7053.97
+ +training dataset, and the hand movements become unsynchronized with the original video. In contrast, our method preserves the person's identity and maintains synchronization even during rapid and complex movements. + +The Effectiveness of the Training Strategies. In Section 4.4, we enhance ReCamMaster's robustness with improved training strategies. We fine-tune only the spatial-temporal (3D) attention layers and freeze other parameters. Additionally, we drop $f$ frames and $f - 1$ frames of the source video with a $20\%$ probability to improve generation capability. Table 4 shows the improvement using visual quality metrics. The 'baseline' is vanilla training with reconstruction loss. Results indicate that all techniques improve visual quality, with their combination yielding the best performance. Dropping the source latent during training also allows our method to support camera-controlled T2V, I2V, and V2V generation simultaneously, as shown in Fig. 9. + +# 6. Applications of ReCamMaster + +We find that ReCamMaster also has promising performances in several traditional tasks, as shown in Fig. 6. + +Video Stabilization. It is usually not easy to obtain stable videos when recording video while moving. Video stabilization techniques [50] aim to smooth out camera movements to produce easy-to-watch videos, which can be achieved by adjusting the camera trajectories via ReCamMaster. We evaluate our method using unsteady videos from the DeepStab [47] dataset. It can be observed that the model stabilizes the video while preserving the content from the original video. Best viewed on the project page. + +Video Super-Resolution. $^{1}$ Thanks to the generation capability of the diffusion model, we can input 'zoom-in' camera trajectories into ReCamMaster to achieve video local super-resolution, as more details are observed in Fig. 6. + +![](images/56c167c51990a4270ca2d4f5bc76e444faf092b42f19b1f8ca6141ab815d7f61.jpg) +Figure 6. Applications of ReCamMaster. From top to bottom: video stabilization, video super-resolution, and video outpainting. + +Video Outpainting. Similarly, we can input zoom-out trajectories to achieve video outpainting. As shown in the last row of Fig. 6 that areas not visible in the original video, such as feet and the ground, have been generated. + +# 7. Conclusion and Limitations + +In this paper, we propose ReCamMaster to reproduce dynamic scenes from input videos with new camera trajectories. We develop an innovative video conditioning technique to enhance pre-trained text-to-video models and curate a large-scale multi-camera synchronized video dataset using Unreal Engine 5, covering diverse scenes and camera movements. Our method also shows promise in video stabilization, super-resolution, and outpainting. There are nevertheless some limitations. Concatenating source and target video tokens improves generation quality but increases computational demands. Additionally, ReCamMaster inherits limitations from the pre-trained T2V models, such as less effective hand generation, as shown in Fig. 11. + +# Acknowledgments + +We thank Jinwen Cao, Yisong Guo, Haowen Ji, Jichao Wang, and Yi Wang from Kuaishou Technology for their invaluable help in constructing the multi-view video dataset. Jianhong Bai would like to thank Quande Liu and Xiaoyu Shi for fruitful discussions. We thank Qinghe Wang and Yawen Luo for their help in method evaluation. + +This work is supported by the National Natural Science Foundation of China (Grant No. 62441615, 62225603, U21B2004, 62476241), the Natural Science Foundation of Zhejiang Province, China (Grant No. LZ23F020008), and the Zhejiang University-Angelalign Inc. R&D Center for Intelligent Healthcare. + +# References + +[1] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. +[2] Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, et al. Vd3d: Taming large video diffusion transformers for 3d camera control. arXiv preprint arXiv:2407.12781, 2024. +[3] Jianhong Bai, Menghan Xia, Xintao Wang, Ziyang Yuan, Xiao Fu, Zuozhu Liu, Haoji Hu, Pengfei Wan, and Di Zhang. Syncammaster: Synchronizing multi-camera video generation from diverse viewpoints. arXiv preprint arXiv:2412.07760, 2024. +[4] Max Bain, Arsha Nagrani, Gül 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. +[5] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yijin Li, FuYun Wang, and Hongsheng Li. Gs-dit: Advancing video generation with pseudo 4d gaussian fields through efficient dense 3d point tracking. arXiv preprint arXiv:2501.02690, 2025. +[6] 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. +[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. +[8] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-your-canvas: Higher-resolution video outpainting with extensive content generation. arXiv preprint arXiv:2409.01055, 2024. + +[9] 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 (ICML), 2024. +[10] Weijie Kong et. al. Hunyuanvideo: A systematic framework for large video generative models, 2024. +[11] Xiao Fu, Xian Liu, Xintao Wang, Sida Peng, Menghan Xia, Xiaoyu Shi, Ziyang Yuan, Pengfei Wan, Di Zhang, and Dahua Lin. 3dtrajmaster: Mastering 3d trajectory for multi-entity motion in video generation. arXiv preprint arXiv:2412.07759, 2024. +[12] Epic Games. Unreal engine 5. https://www.unrealengine.com/en-US/unreal-engine-5, 2022. +[13] 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. +[14] 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 Nowrouzezahrai, 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. 2022. +[15] Zekai Gu, Rui Yan, Jiahao Lu, Peng Li, Zhiyang Dou, Chenyang Si, Zhen Dong, Qifeng Liu, Cheng Lin, Ziwei Liu, et al. Diffusion as shader: 3d-aware video diffusion for versatile video generation control. arXiv preprint arXiv:2501.03847, 2025. +[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933, 2023. +[17] 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. +[18] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. +[19] 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. + +[20] Chen Hou, Guoqiang Wei, Yan Zeng, and Zhibo Chen. Training-free camera control for video generation. arXiv preprint arXiv:2406.10126, 2024. +[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. +[22] Teng Hu, Jiangning Zhang, Ran Yi, Yating Wang, Hongrui Huang, Jieyu Weng, Yabiao Wang, and Lizhuang Ma. Motionmaster: Training-free camera motion transfer for video generation. arXiv preprint arXiv:2404.15789, 2024. +[23] 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. +[24] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. +[25] 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. +[26] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European Conference on Computer Vision, pages 18-35. Springer, 2024. +[27] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations (ICLR), 2014. +[28] Zhengfei Kuang, Shengqu Cai, Hao He, Yinghao Xu, Hongsheng Li, Leonidas Guibas, and Gordon Wetzstein. Collaborative video diffusion: Consistent multi-video generation with camera control. arXiv preprint arXiv:2405.17414, 2024. +[29] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast, and robust structure and motion from casual dynamic videos. arXiv preprint arXiv:2412.04463, 2024. +[30] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. Motionclone: Training-free motion cloning for controllable video generation. arXiv preprint arXiv:2406.05338, 2024. +[31] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In International Conference on Learning Representations (ICLR), 2023. +[32] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. + +[33] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8599–8608, 2024. +[34] Arun Mallya, Ting-Chun Wang, Karan Sapra, and Ming-Yu Liu. World-consistent video-to-video synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 359-378. Springer, 2020. +[35] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7038-7048, 2024. +[36] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, pages 58-77. Springer, 2024. +[37] 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. +[38] 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. +[39] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed world-consistent video generation with precise camera control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. +[40] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. +[41] Xuelun Shen, Zhipeng Cai, Wei Yin, Matthias Müller, Zijun Li, Kaixuan Wang, Xiaozhi Chen, and Cheng Wang. Gim: Learning generalizable image matcher from internet videos. In The Twelfth International Conference on Learning Representations, 2024. +[42] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. +[43] Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. *Omnincontrol: Minimal and universal control for diffusion transformer.* arXiv preprint arXiv:2411.15098, 2024. +[44] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. Openreview, 2019. +[45] Basile Van Hoorick, Rundi Wu, Ege Ozguroglu, Kyle Sargent, Ruoshi Liu, Pavel Tokmakov, Achal Dave, Changxi + +Zheng, and Carl Vondrick. Generative camera dolly: Extreme monocular dynamic novel view synthesis. In European Conference on Computer Vision, pages 313-331. Springer, 2024. +[46] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. In European Conference on Computer Vision, pages 153-168. Springer, 2024. +[47] Miao Wang, Guo-Ye Yang, Jin-Kun Lin, Song-Hai Zhang, Ariel Shamir, Shao-Ping Lu, and Shi-Min Hu. Deep online video stabilization with multi-grid warping transformation learning. IEEE Transactions on Image Processing, 28(5): 2283-2292, 2018. +[48] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. arXiv preprint arXiv:1808.06601, 2018. +[49] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot video-to-video synthesis. arXiv preprint arXiv:1910.12713, 2019. +[50] Yiming Wang, Qian Huang, Chuanxu Jiang, Jiwen Liu, Mingzhou Shang, and Zhuang Miao. Video stabilization: A comprehensive survey. Neurocomputing, 516:205-230, 2023. +[51] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. +[52] 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. +[53] 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. +[54] Zeqi Xiao, Wenqi Ouyang, Yifan Zhou, Shuai Yang, Lei Yang, Jianlou Si, and Xingang Pan. Trajectory attention for fine-grained video motion control. In The Thirteenth International Conference on Learning Representations, 2025. +[55] Yiming Xie, Chun-Han Yao, Vikram Voleti, Huaizu Jiang, and Varun Jampani. Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470, 2024. +[56] Jinbo Xing, Menghan Xia, Yuxin Liu, Yuechen Zhang, Yong Zhang, Yingqing He, Hanyuan Liu, Haoxin Chen, Xiaodong Cun, Xintao Wang, et al. Make-your-video: Customized video generation using textual and structural guidance. IEEE Transactions on Visualization and Computer Graphics, 2024. +[57] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. +[58] Yiran Xu, Taesung Park, Richard Zhang, Yang Zhou, Eli Shechtman, Feng Liu, Jia-Bin Huang, and Difan Liu. + +Videoigigagan: Towards detail-rich video super-resolution. arXiv preprint arXiv:2404.12388, 2024. +[59] Zhen Xu, Yinghao Xu, Zhiyuan Yu, Sida Peng, Jiaming Sun, Hujun Bao, and Xiaowei Zhou. Representing long volumetric video with temporal gaussian hierarchy. ACM Transactions on Graphics, 43(6), 2024. +[60] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. +[61] Shiuyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. +[62] 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. +[63] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. +[64] Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. arXiv preprint arXiv:2503.05638, 2025. +[65] David Junhao Zhang, Roni Paiss, Shiran Zada, Nikhil Karnad, David E Jacobs, Yael Pritch, Inbar Mosseri, Mike Zheng Shou, Neal Wadhwa, and Nataniel Ruiz. Recapture: Generative video camera controls for user-provided videos using masked video fine-tuning. arXiv preprint arXiv:2411.05003, 2024. +[66] 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. +[67] Guangcong Zheng, Teng Li, Rui Jiang, Yehao Lu, Tao Wu, and Xi Li. Cami2v: Camera-controlled image-to-video diffusion model. arXiv preprint arXiv:2410.15957, 2024. +[68] Sixiao Zheng, Zimian Peng, Yanpeng Zhou, Yi Zhu, Hang Xu, Xiangru Huang, and Yanwei Fu. Vidcraft3: Camera, object, and lighting control for image-to-video generation. arXiv preprint arXiv:2502.07531, 2025. +[69] 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. \ No newline at end of file diff --git a/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/images.zip b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6ded504c09ed041ccd3d3f0b79fa86d16650a4bc --- /dev/null +++ b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e327ef31038147c5e875aea3e1112889365fddc769535025dd913596204fbab +size 833021 diff --git a/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/layout.json b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f0add08048205b744b973169ee68bb30f7eac735 --- /dev/null +++ b/ICCV/2025/ReCamMaster_ Camera-Controlled Generative Rendering from A Single Video/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:459481769842c38143cecefa55136ff714e2ca847978628a5affd37a6a4f8043 +size 401981 diff --git a/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/a371b18d-1a91-428e-9232-7291cc721803_content_list.json b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/a371b18d-1a91-428e-9232-7291cc721803_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..76de3f4528b540ed298244c7d9377f4ee6de55e4 --- /dev/null +++ b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/a371b18d-1a91-428e-9232-7291cc721803_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddc946fafb5ff3cf4d8d9c1ee9e85547f0a85295758b8f2b89641cf782a7ab74 +size 82687 diff --git a/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/a371b18d-1a91-428e-9232-7291cc721803_model.json b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/a371b18d-1a91-428e-9232-7291cc721803_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bbe65a5dd112e1ddadbc194d8d5bdd67db4e1cbb --- /dev/null +++ b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/a371b18d-1a91-428e-9232-7291cc721803_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b13e26fb830fde8a36e6d8f40b1976d23c0b6d568a2b21a86485cba95daac586 +size 104440 diff --git a/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/a371b18d-1a91-428e-9232-7291cc721803_origin.pdf b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/a371b18d-1a91-428e-9232-7291cc721803_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..56d083282cc6370d174bc0719b9ef78f8bd70004 --- /dev/null +++ b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/a371b18d-1a91-428e-9232-7291cc721803_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0ba93c58076ae9efa799ad269ef9d9712fd72c22b35a5b61d51238719d81855 +size 630225 diff --git a/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/full.md b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e8a061a217a860c879f32d14357051667ee4f0fe --- /dev/null +++ b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/full.md @@ -0,0 +1,354 @@ +# ReCoT: Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models + +Mengxue $\mathbf{Q}\mathbf{u}^{1,3*}$ Yibo Hu2 Kunyang Han1,3 Yunchao Wei1,3\* Yao Zhao1,3 + +$^{1}$ Institute of Information Science, Beijing Jiaotong University, China $^{2}$ NIO, China $^{3}$ Visual Intelligence + X International Joint Laboratory of the Ministry of Education, China + +{qumengxue, kunyanghan, yunchao.wei, yzhao}@bjtu.edu.cn, huyibo871079699@gmail.com + +# Abstract + +Recent advancements in Large Vision-Language Models (LVLMs) have greatly improved their ability to understand both visual and text information. However, a common problem in LVLMs is confirmation bias, where models tend to repeat previous assumptions and follow earlier viewpoints instead of reflecting and correcting themselves. This problem is more common in smaller-scale LVLMs, as they are usually fine-tuned with training data that is mostly positive, focusing on generating coherent dialogue. To address this issue, we introduce ReCoT, a method designed to mitigate confirmation bias in smaller-scale LVLMs through Reflective Self-Correction Training. The method follows a two-stage SFT-DPO paradigm: the first SFT stage aims to cultivate the model's reflective correction abilities, while the DPO stage focuses on enhancing the consistency between answers and reflections. Specifically, we construct dialogue-based reflective samples, which serve as adversarial samples during SFT. In this process, the model is initially presented with a potentially incorrect answer, followed by a reflection and correction phase to generate the final answer. To enhance answer-reflection consistency, we propose the consistency direct preference optimization. To comprehensively evaluate the effectiveness of our ReCoT, we introduce a set of novel metrics to measure the accuracy of the reflection and correction process. Extensive experiments show that ReCoT enables LVLM to engage in robust self-reflection and error correction and reduce confirmation bias. Code is available at ReCoT. + +# 1. Introduction + +In recent years, the development of Large Vision-Language Models (LVLMs) [9, 35, 37] has significantly enhanced their capability to understand both visual and textual in- + +![](images/df6992f70472a1fdc3daff0788f65a70ee3852c8ca19991d46e471a65c421ab1.jpg) +(a) REVERIE Positive Rationale Format + +![](images/3a7fcb3fee423b6225c70c4a66dfd72f1526654af904ad0936022b7341482f2b.jpg) +(b) REVERIE Negative Rationale Format + +![](images/ea51adde4d406f72dd7615d96ae533c8e43c7f937f927d4d1087ecf45a6f5db5.jpg) +(c) Confirmation Bias Example + +![](images/71d59fead04341360336e0f653d8f62d1f51ddb79dbfb6ccb4f17afde7989673.jpg) +(d) Reflective Correction Example +Figure 1. Different reflective data formats. (a) and (b) expect a model to make a choice, then explain its implicit reason (correctness or incorrectness of an option). This often leads to (c), where confirmation bias causes the model to make excuses for the previous answer. (d) is the ideal reflective response with correction. + +formation, leading to widespread applications [5, 16, 18, 19, 21, 33, 34, 42] in tasks such as dialogue generation and knowledge-based question answering. Models like OpenAI-o1 [30], DeepSeek-R1 [11] have demonstrated strong reasoning abilities. Notably, DeepSeek-R1 introduces the concept of “aha moment”, which highlights the model's capacity for self-reflection and error correction. + +The ability of models to self-reflect and correct mistakes is crucial for improving robustness and trustworthiness. In complex reasoning tasks, LVLMs often generate responses based on prior assumptions, and without self-reflection, they may reinforce incorrect reasoning patterns. Several works [40, 43, 48, 50, 56, 58] have begun exploring the reasoning capabilities of LVLMs. Specifically, studies + +such as [10, 50, 56] investigate how incorporating reflective data into training can enhance the model's reasoning ability. + +However, LVLMs suffer from confirmation bias [31], a phenomenon where models tend to reaffirm their initial assumptions rather than critically rethinking and correcting their mistakes. This issue is particularly pronounced in LVLMs with fewer than 10B parameters. Taking the REVERIE model as an example, Figures 1 (a) and (b) illustrate its training data format, where the model is required to answer a question and subsequently provide rationales explaining why the correct answer is correct (positive rationale) and why incorrect answers are incorrect (negative rationales). To examine the model's robustness, we conduct a naive experiment by deliberately modifying correct answers to incorrect ones and prompting the model to justify the altered answers. The results indicate that the model often rationalizes incorrect answers instead of recognizing and correcting the mistake. This behavior arises because many LVLMs are trained through Supervised Fine-Tuning (SFT) on data distilled from larger models like o1 [30], where the training objective emphasizes generating coherent dialogues. Moreover, training data is often skewed towards positive and user-aligned examples, making models more inclined to affirm user perspectives rather than question or correct them. + +To address this issue, we propose Reflective Self-Correction Training (ReCoT), a method designed to mitigate confirmation bias in LVLMs by enhancing their capacity for self-reflection and error correction. The ReCoT framework begins by constructing dialogue-based reflective samples, iteratively generating samples from incorrect answers that require correction. Each sample consists of an initial direct response followed by a reflection and correction phase, providing a structured format to evaluate the model's self-correction ability. Based on these reflective samples, we conduct Reflective Fine-tuning (RFT), training the model to develop self-reflection and correction skills. During this training process, loss computation is restricted to the incorrect portions of responses, heightening the model's sensitivity to errors and reinforcing its capacity for correction. + +Nevertheless, a critical challenge persists: inconsistencies between self-reflections and corrected answers. Without explicit alignment mechanisms, models may produce self-reflections that fail to meaningfully contribute to answer correction. To address this, we introduce a consistency direct preference optimization (Consist-DPO), whose key idea is to construct a consistency correction dataset, ensuring stronger alignment between reflections and final answers. Furthermore, we propose a set of novel evaluation metrics, including final answer accuracy, correction success rate, and reflection-answer consistency, to rigorously assess the accuracy of the reflection-correction process and + +the model's ability to overcome confirmation bias. + +We conduct extensive experiments to verify the effectiveness of ReCoT. Our results on ScienceQA [28], MME [14], and MMStar [4] benchmarks demonstrate that ReCoT significantly enhances LVLMs' ability to engage in critical self-reflection and error correction, effectively mitigating confirmation bias. Our main contributions are as follows: + +- We identify confirmation bias phenomenon in LVLMs and analyze the limitations of current models in self-reflection and error correction. We further reveal how training data distribution and objective functions contribute to this bias. +- We propose a novel approach combining adversarial sample training and reflection-answer alignment, along with a comprehensive evaluation framework tailored for assessing self-correction abilities in LVLMs. +- We conduct extensive experiments to validate whether models can effectively self-correct when confronted with adversarial modifications or misleading contexts, providing new insight and improvement direction for LVLM applications. + +# 2. Related Work + +# 2.1. Vision-Language Reasoning + +LVLMs [1, 25-27] have achieved remarkable progress in recent years, demonstrating robust capabilities across various domains, including visual understanding [23, 35], mathematics [22], college-level question answering [9], and scientific reasoning. Some models emphasize visual comprehension by utilizing formatted bounding boxes [32, 44] or region-based encoders [52, 59] to achieve regional grounding. Others enhance linguistic context by introducing hierarchical feature pyramids [36] for regional prompting or incorporating the Chain-of-Thought (CoT) procedure [2] for improved contextual understanding. + +Despite their strong performance on general benchmarks [8, 24], LVLMs still face challenges in integrating linguistic cues with visual information, reasoning about novel concepts, and solving complex problems. The CoT method [45] employs explicitly crafted, step-by-step instructions as prompts. Building on this, methods such as GoT [3] and ToT [51] predict the next thought from multiple candidates, selecting the most confident option. Furthermore, several approaches have explored collecting rationale data to fine-tune models for step-by-step reasoning. G-llava [15] and Mavis [57] augment existing mathematical datasets with rationales distilled from GPT models, while MathGLM [49] enhances performance through manually curated CoT annotations. Additionally, other works [47, 54, 55] implement search algorithms such as Monte Carlo Tree Search (MCTS) to generate Tree-of + +Thought datasets using large language models themselves. + +# 2.2. Reflection and Correction + +Prompting a LLM to reflect on its previous steps is a widely adopted strategy to enhance output quality and success rates. Early studies [17, 20, 60] decompose complex problems into simpler subproblems, enabling LLMs to plan in a manner analogous to human cognitive processes. Some approaches [13, 38, 46] focus on collecting relevant data from human annotators, expert models, and other sources. And some other methods introduce iterative frameworks for self-training [53] and inference-refinement [29]. + +Despite these advancements, limited attention has been given to this problem within the context of LVLMs. DRESS [6] aims to improve models' ability to revise responses, which focuses on open-ended QA with subjective human alignment feedback. Mulberry [50] integrates collective learning with MCTS to generate tree-structured reasoning nodes and identify negative nodes, thereby constructing reflective reasoning paths. R3V [10] employs an iterative bootstrapping approach to generate both positive and negative solutions for reasoning datasets, subsequently optimizing the model's reflection ability through a self-refine loss. REVERIE [56] introduces ReflRctiVE and RationalE instructions, curating a dataset containing both expression and explanation annotations for positive and negative rationales. Distinct from these methods, which primarily address reflection issues, this paper explores both reflection and correction strategies, aiming to further enhance the reasoning capabilities of LVLMs. + +# 2.3. Vision-Language Alignment + +The success of ChatGPT [41] highlights the significance of Reinforcement Learning from Human Feedback (RLHF), which enables language models to align a model trained on a general corpus of text data with complex human values. However, RLHF necessitates training a model to serve as a value function within the Proximal Policy Optimization (PPO) [39] algorithm. To mitigate this complexity, Direct Preference Optimization (DPO) has been introduced to streamline the alignment process. DPO optimizes models to produce human-preferred responses by directly training with preference data. Iterative DPO [7] reveals that optimizing the model with a static batch of offline-generated data yields limited improvements. To address this, it proposes a strategy of updating preference pairs at regular intervals throughout the training process using the evolving model. + +Building on these advancements, Insight-V [12] aims to alleviate the hallucination problem within reasoning agents, enhancing both stability and response quality. Similarly, LLaVA-Reasoner-DPO [58] uses a two-fold approach to collect a dataset containing more rationale information, and generate preference pairs based on this dataset for DPO + +training. Different from current research that primarily focuses on mitigating hallucinations and improving reasoning capabilities. In this paper, we leverage DPO to achieve answer-reflection alignment, fostering a more robust and thoughtful model. + +# 3. Method + +Our proposed Reflective Self-Correction Training (ReCoT) aims to enhance the robustness of LVLMs and mitigate confirmation bias. ReCoT follows a two-stage training paradigm: (1) In the Reflective Fine-Tuning (RFT) stage, we train the model to develop reflection and self-correction abilities. (2) In the Consistency Direct Preference Optimization (Consist-DPO) stage, we optimize the alignment between the model's final answer and its reflection, ensuring consistency. The overall framework is illustrated in Fig. 2. In Sec. 3.1, we first introduce the motivation and preliminary observations that led to our approach. Sec.s 3.2 and 3.3 then detail the two key components of ReCoT: Reflective Fine-Tuning and Consistency Direct Preference Optimization, respectively. + +# 3.1. Preliminaries + +Given an image $I$ and a question $x$ , the reasoning LVLM is expected to generate both a Chain-of-Thought (CoT) rationale $r$ and a final answer $a$ . As illustrated in Fig. 2(a), existing vision-language reasoning and reflection methods mainly adopt two formats: + +Answer-then-Rationale, where the model first provides an answer and then explains its correctness(e.g., REVERIE [56]): + +$$ +D = \{(I, x) | (a, r) \} \tag {1} +$$ + +Rationale-then-Answer, where the model first performs CoT reasoning before generating the final answer(e.g., Mulberry [50]): + +$$ +D = \{(I, x) \mid (r, a) \} \tag {2} +$$ + +where the USER prompt and ASSISTANT response are presented on the left and right side respectively. + +We first observe confirmation bias in the REVERIE model [56], a phenomenon where the model tends to justify its initial answer rather than critically reassess and correct errors. As shown in Fig. 1(a) and (b), the REVERIE training data format requires the model to first provide an answer and then generate a positive rationale (explaining why the correct answer is right) and a negative rationale (explaining why incorrect answers are wrong). To test the model's ability to reflect and self-correct, we performed a naive intervention: we manually alter correct answers to incorrect ones and ask the model to justify the modified (incorrect) answer. Surprisingly, instead of detecting the mistake, the + +![](images/f68cba7ae85c43ef1e83d8cf878fdfe7ef47831b09e317c982e3d1632f6f904e.jpg) +(a) Original Reasoning Data Format + +![](images/cecad743ea5367426269bbdee320c7a18f4549c2a96cf61fe93c5ff1db483d01.jpg) +(b) Standardized Data Format + +![](images/bf61b7927858ed7489cbd7651af565b7fd21d56bbcc8cd8534322c77ee15e7c3.jpg) +(d) Data Filtered by Consistency Rules + +![](images/c641092ab2789c08abf056d8fc816d5e4f27c222fee3fe27deadd3e124c265ea.jpg) + +![](images/4f65e35cbc68ae01a196b30a07d8acc114a0d397a8846f4785b203576f6c99fe.jpg) +(c) Correction-required Data Format + +![](images/fcb516d720d347a959f09ab5322969511c40bdda6c8f7cded6adef1a44f34551.jpg) +Figure 2. Overall framework. The upper section illustrates various data formats and converting progress in our methods: (a) presents two common data formats referenced in Eq.s 1 and 2; (b) represents our standardized data used for self-correction training and evaluation; (c) corresponds to data of Eq. 3 constructed by altering $a_0$ ; and (d) contains filtered data utilized for DPO training. The lower section depicts the ReCoT training paradigm: initially, data (b) and (c) are employed for reflective fine-tuning (RFT), followed by the use of data (c) and (d) for Consist-DPO training. Best viewed in color. Green denotes the correct ground truth (GT) answer. + +model often rationalizes the incorrect answer as if it's correct, as shown in Fig. 1(c). Ideally, we expect the model to recognize inconsistencies and correct its previous response in subsequent turns like Fig. 1(d). + +We further validate this bias on Mulberry [50], modifying its Rationale-then-Answer format by incorporating a small proportion of direct-answer samples. Despite the majority of training data following the standard CoT reasoning format, robustness testing reveals that the model still exhibits lazy reasoning: its final answer tended to simply copy the initial answer, even when the reasoning contradicts it. This suggests that LVLMs, often derived from instruction supervised fine-tuning, prioritize dialogue coherence over critical reassessment. Since most training samples are positive, user-aligned examples, these models tend to confirm previous viewpoints rather than challenge or correct them. This observation motivated us to explore data-driven training strategies that encourage self-correction and reflection. + +# 3.2. Reflective Fine-Tuning + +To enhance the model's ability to reflect and correct the possible mistakes, we first standardize the data format to sup- + +port self-correction evaluation. Specifically, we structure training samples as "direct answer + reflective correction", ensuring that the model learns to first provide an initial answer and then critically assess and revise it if necessary. The standardized format is as follows: + +$$ +D _ {\mathrm {S F T}} = \{(I, x) \mid (a _ {0}, r, a _ {f}) \} \tag {3} +$$ + +where $a_0$ represents the initial answer, which is the model's response without reasoning; $r$ denotes the reflection process, where the model reviews its initial response and analyzes potential errors; $a_f$ is the final corrected answer, which is the model's refined response after reflection, $a_0$ and $a_f$ are both assigned as ground-truth answers. For REVERIE, $r$ is the reflection rationales; for Mulberry, $r$ is the reasoning rationales. + +To construct reflection samples, we systematically traverse incorrect answers and generate correction-required instances, prompting the model to detect and fix errors. In detail, for multiple-choice tasks, we traverse incorrect options $a_{i}(i \in [0, N], a_{i} \neq a_{GT})$ and replace the initial answer $a_{0}$ to construct correction-required data: + +$$ +D _ {\mathrm {R F T}} = \{(I, x) \mid (a _ {i}, r, a _ {f}) \} \tag {4} +$$ + +where $a_{GT}$ represents the ground-truth answer. For short-answer or open-ended dialogue tasks: if corresponding reflection data is available (e.g., REVERIE [56]), we replace the initial answer with incorrect responses from the reflection data; if no reflection data is available (e.g., Mulberry [50]), we construct only positive examples without explicitly introducing errors. We do not introduce incorrect answers for short-answer and open-ended tasks in the Mulberry dataset because our goal is to ensure a fair experimental comparison without introducing additional data. + +The optimization objective of Reflective Fine-tuning (RFT) consists of two steps: (1) supervised fine-tuning for correction format alignment, which ensures the model generates outputs in the desired format, and (2) reflective fine-tuning with selective loss masking, where we prevent direct-answer loss computation on incorrect initial responses, the final loss $\mathcal{L}_{\mathrm{RFT}}$ is presented as follows: + +$$ +\mathcal {L} _ {\mathrm {R F T}} = \mathcal {L} _ {\mathrm {C E}} (r) + \mathcal {L} _ {\mathrm {C E}} \left(a _ {f}\right) \tag {5} +$$ + +where $\mathcal{L}_{\mathrm{CE}}$ is the cross-entropy loss. By restricting gradient updates on erroneous parts, we ensure that the model does not merely memorize mistakes but instead focuses on identifying and correcting them. This approach significantly enhances the model's sensitivity to errors and strengthens its self-correction capabilities. + +# 3.3. Consistency Direct Preference Optimization + +Although RFT improves self-correction, we observe that reflection and final answers still suffer from inconsistencies. For instance, in some cases, the model correctly identifies an error in its rationale but still fails to revise its final answer accordingly. This inconsistency suggests that while RFT encourages self-reflection, it does not explicitly optimize for answer-reflection alignment. + +Inspired by Direct Preference Optimization (DPO) [7], we propose Consistency Direct Preference Optimization (Consist-DPO) to enforce logical coherence between a model's reflection and its final answer. The key idea is to construct a consistency correction dataset by identifying instances where the reflection contradicts the final answer. + +Specifically, we traverse correct and incorrect initial answers and check whether the model's self-correction aligns across different instances. If the model detects errors inconsistently or fails to update its final answer accordingly, we label these cases for direct preference optimization. The example data formats filtered by consistency rules are illustrated as Fig. 2(d). + +We then apply a DPO-based loss to encourage answer-reflection consistency. The optimization objective is formulated as: + +$$ +\mathcal {L} _ {\text {C o n s i s t - D P O}} = - \mathbb {E} _ {(x, y ^ {+}, y ^ {-})} \left[ \log \sigma \left(f _ {\theta} \left(y ^ {+}\right) - f _ {\theta} \left(y ^ {-}\right)\right) \right] \tag {6} +$$ + +where $y^{+}$ represents the consistent correction, and $y^{-}$ denotes the inconsistent response. By training the model to prefer outputs where self-correction and final answers are aligned, we significantly improve logical consistency and ensure that reflection effectively guides answer revision. + +Through this two-stage training paradigm, ReCoT not only enhances self-correction capabilities but also ensures coherence between reasoning and final predictions, mitigating confirmation bias in LVLMs. + +# 4. Experiment + +To validate the effectiveness of our Reflective Self-Correction Training (ReCoT) framework, we conduct extensive experiments across multiple vision-language reasoning tasks. In Sec. 4.1 presents the datasets used for training and benchmarking. Sec. 4.2, we introduce key evaluation metrics to systematically assess the model's self-correction capability. Sec. 4.3 details the implementation. Sec. 4.4 and Sec. 4.5 report the main results and ablation studies, comparing ReCoT with the baseline methods and analyzing its impact. + +# 4.1. Datasets + +We conduct training on two vision-language reasoning datasets and evaluate our model's reflective self-correction ability on three multimodal benchmarks. + +# 4.1.1. Training Datasets + +REVERIE [56] is a large-scale instruction-tuning dataset with reflective rationale annotations. It comprises 115k machine-generated reasoning instructions, each meticulously annotated with a corresponding pair of correct and confusing responses, alongside comprehensive rationales elucidating the justification behind the correctness or erroneousness of each response. Each instance consists of an image $I$ and a question $x$ where the model is required to generate an answer $a$ along with a rationale $r$ . The dataset is structured as follows: + +$$ +D _ {\text {R E V E R I E}} = \{(I, x) \mid (a, r ^ {\pm}) \} \tag {7} +$$ + +where $r^{\pm} \in \{r^{+}, r^{-}\}$ , $r^{+}$ is the positive rationale, explaining why the correct answer $a$ is valid, and $r^{-}$ is the negative rationale, explaining why incorrect answers are invalid. Since the original REVERIE dataset focuses on justification rather than self-correction, models trained on it tend to exhibit confirmation bias, reinforcing initial responses without critically reassessing them. + +Mulberry-260k [50] is a multimodal dataset with a tree of rich, explicit and well-defined reasoning nodes for each question. It follows the Chain-of-Thought (CoT) reasoning paradigm, where the model first generates a reasoning process $r$ before producing the final answer $a$ : + +$$ +D _ {\text {M u l b e r r y}} = \{(I, x) \mid (r, a) \} \tag {8} +$$ + +To evaluate whether LVLMs can effectively reflect and self-correct, we modify the dataset by introducing direct answer + reflection correction samples and constructing correctionrequired data. The modified format is as Eq. 3. + +# 4.1.2. Evaluation Benchmarks + +ScienceQA [28] is a large-scale visual question-answering benchmark covering diverse domains such as science, mathematics, and social sciences. + +MME [14] is designed to evaluate LVLMs' generalization across various tasks, including visual reasoning, logical inference, and cross-modal understanding. + +MMStar [4] is a challenging multimodal benchmark designed to assess LVLMs on complex multimodal tasks, including counterfactual reasoning and knowledge-based inference. Its format is similar to MME but contains harder reasoning challenges that require more robust self-correction mechanisms. + +# 4.2. Metrics + +To comprehensively evaluate the effectiveness of the ReCoT framework, we introduce two key evaluation metrics in addition to standard reasoning accuracy: Error Correction Ability and Reflection-Answer Consistency. These metrics are designed to quantitatively assess the model's capability to identify and correct errors, as well as maintain consistency between its self-reflection and final answer. + +Error Correction Ability Evaluation. One of the core objectives of ReCoT is to enable models to recognize and correct errors in their initial predictions. To measure this ability, we compare the model's initial answer $a_0$ and its final corrected answer $a_f$ against the ground-truth answer $a_{GT}$ . We define the correction success rate (CSR) as follows: + +$$ +\mathrm {C S R} = \frac {\sum 1 \left(a _ {0} \neq a _ {G T} \wedge a _ {f} = a _ {G T}\right)}{\sum 1 \left(a _ {0} \neq a _ {G T}\right)} \tag {9} +$$ + +where $1(\cdot)$ is an indicator function, and $\wedge$ denotes a logical AND operation. A high CSR indicates that the model can effectively revise incorrect predictions after reflection. + +To comprehensively evaluate the model's error correction capability, we assess CSR under two distinct conditions: (1) CSR on Self-Made Mistakes, denoted as $\mathsf{CSR}_{\text {self }}$ . In this case, we evaluate whether the model can self-correct when its own initial prediction $a_0$ is incorrect. That is, instead of artificially modifying $a_0$ , we consider natural errors made by the model itself. A high $\mathsf{CSR}_{\text {self }}$ indicates that the model can recognize its own mistakes and correct them through reflection, demonstrating robust self-correction capabilities. (2) CSR under Controlled Perturbation, denoted as $\mathsf{CSR}_{\text {ctrl }}$ . In this setting, we deliberately modify the model's initial answer $a_0$ to an incorrect value, simulating cases where the model initially makes a mistake. + +The model then undergoes a reflection step and produces a revised answer $a_{f}$ . This metric quantifies the model's ability to identify externally introduced errors and correct them after reasoning. + +Reflection-Answer Consistency Evaluation. To evaluate whether the model's reflection aligns with its initial answer and final answer, we introduce the Reflection-Answer Consistency Score (RCS). This metric assesses whether the model correctly judges the correctness of its initial prediction and whether the final answer aligns with the provided explanation. The evaluation is based on the structured prompt (details are in the supplementary), which is provided to GPT-4 for evaluation, leveraging its advanced reasoning and comprehension capabilities to assess the quality of the model's responses. It decomposes the reflection into two key aspects: (1) Initial Answer Judgment (IAJ): Determines whether the explanation explicitly assesses the correctness of the initial answer. If the explanation correctly identifies the correctness of $a_0$ , then $IAJ = 1$ , otherwise $IAJ = 0$ . (2) Final Answer Judgment (FAJ): Determines whether the explanation correctly supports the correctness of the final answer. If the reflection correctly affirms $a_f$ , then $FAJ = 1$ , otherwise $FAJ = 0$ . + +We define the Reflection-Answer Consistency Score (RCS) as follows: + +$$ +\mathrm {R C S} = \frac {\sum (\mathrm {I A J} \wedge \mathrm {F A J})}{N} \tag {10} +$$ + +where $N$ is the total number of evaluated samples, and $\wedge$ denotes a logical AND operation, meaning only cases where both IAJ and FAJ are correct contribute to the score. + +A high RCS indicates that the model consistently produces explanations that correctly assess both the initial answer and the final answer, thereby demonstrating strong reflection consistency. Conversely, a low RCS suggests that the model either fails to correctly identify initial mistakes or does not adequately justify its final answer, which could contribute to inconsistencies in reasoning. + +By using this metric, we quantitatively measure the alignment between a model's reflection and its reasoning steps, ensuring that self-correction is not only accurate but also well-justified. + +# 4.3. Implementation Details + +We implement ReCoT on three LVLMs: REVERIE-1.5-7B [56], Mulberry-2B and Mulberry-7B [50]. We adopt the REVERIE model based on LLaVA-1.5 [25] and the Mulberry model based on Qwen2-VL [35]. Following the original experimental setups of REVERIE and Mulberry, we train LLaVA-1.5-7B using the REVERIE dataset, while Qwen2-VL-2B and Qwen2-VL-7B are trained on Mulberry-260k. We sample 10k standardized instances, which are then mixed with the original REVERIE or Mul + +
MethodReflection?ScienceQAMMEMMStar
Acc.\(CSR_{self}\)\(CSR_{ctrl}\)Score\(CSR_{self}\)\(CSR_{ctrl}\)Acc.\(CSR_{self}\)\(CSR_{ctrl}\)
REVERIE-1.5-7B [56]
SFTX79.23--1827.89--35.33--
SFT✓79.230.0040.871827.890.399.9535.330.006.79
ReCoT✓82.5058.1186.531841.3413.1439.5937.8720.1235.13
Mulberry-2B [50]
SFTX84.78--1851.80--46.07--
SFT✓84.9811.0436.351869.000.616.1146.402.104.81
ReCoT✓86.5151.6985.682070.5823.9165.6350.2026.3348.90
Mulberry-7B [50]
SFTX91.77--2158.80--55.13--
SFT✓93.2122.2931.692249.817.879.2757.2017.225.79
ReCoT✓95.1475.0094.462262.4031.0582.7059.8726.8657.94
+ +Table 1. The main results of three LVLMs (REVERIE-1.5-7B [56], Mulberry-2B [50], and Mulberry-7B [50]) trained with our ReCoT. The results demonstrate significant improvements in accuracy and correction success rate (CSR) across three benchmarks: ScienceQA [28], MME [14], and MMStar [4]. "Reflection?" indicates whether the trained model use reflective instruction to encourage reflection during inference. $\mathrm{CSR}_{\text{self}}$ measures the proportion of cases where the model initially produces an incorrect answer but corrects itself after reflection. $\mathrm{CSR}_{\text{ctrl}}$ evaluates whether the model can correct a given incorrect answer after reflection. + +
MethodScienceQAMMEMMStar
Formatted SFT85.031869.0046.40
+ RFT86.511931.7049.40
+ Consist-DPO86.372070.5850.20
+ +berry data for training. For this stage, we adopt full-parameter fine-tuning. + +During inference, we evaluate models with two baseline answering strategies: (1) direct answering and (2) answerthen-reflection, which reflect the prevalent approaches in existing vision-language models. We then apply ReCoT to further enhance the models' reflective self-correction capability, which consists of SFT and DPO two stages. Unlike the initial full fine-tuning, both SFT and DPO are trained using LoRA to improve efficiency. + +To construct the DPO training dataset, we first fine-tune models using SFT and leverage the trained models to generate reflection-based correction data. Specifically, for each question, we consider both correct and incorrect candidate answers as initial responses and observe whether the model consistently corrects erroneous responses. Those samples exhibit inconsistent corrections, are collected for DPO training to improve answer-reflection consistency. + +# 4.4. Main Results + +The experimental result in Table 1 demonstrates the effectiveness of ReCoT across three base models (REVERIE-1.5-7B [56], Mulberry-2B, and Mulberry-7B [50]) on three benchmarks (ScienceQA [28], MME [14], and MMStar [4]). The results show notable improvements in accu + +Table 2. Effectiveness of RFT and Consist-DPO in our ReCoT. + +
Training StageMulberry-2BREVERIE-7B
Before Consist-DPO89.8387.81
After Consist-DPO97.8891.04
+ +Table 3. Effectiveness of Consist-DPO. The table shows the RCS score before and after applying Consist-DPO. + +racy and correction success rate (CSR), indicating that ReCoT effectively mitigates the confirmation bias in LVLMs. According to the results, we found that even though traditional SFT-trained models achieve reasonable accuracy, they fail to reflect effectively when prompted with reflective instructions, as evidenced by their low CSR. Compared to REVERIE, the Mulberry model has a more detailed step-by-step reasoning process, rather than just analyzing why an answer is correct or incorrect. As a result, Mulberry models exhibit some intrinsic error correction ability, but their CSR still indicate significant room for improvement. After applying ReCoT, all three models show substantial increases in both $\mathrm{CSR}_{\text {self }}$ and $\mathrm{CSR}_{\text {ctrl }}$ , demonstrating that the models can effectively identify and correct errors after reflection. The improvement is particularly pronounced in Mulberry-7B, which achieves the highest CSR scores across all benchmarks, suggesting that larger models with reasoning rationales benefit the most from ReCoT. + +# 4.5. Ablation Study + +Effectiveness of RFT and Consist-DPO We conduct a series of experiments to evaluate the effectiveness of the two training stages of ReCoT. We evaluate the model trained on Mulberry-2B using three datasets, with the results summarized in Table 2. "Formatted SFT" refers to training + +![](images/bf2cd2913ecf3688e2472d67f3c78229881b250c6222aef8d0c944a486ff8fcc.jpg) +Figure 3. Performance comparison of the Mulberry model trained with ReCoT on the MME benchmark across three response strategies: reason-then-answer, answer-then-reflect, and direct answer. Best viewed in color. + +
Data SizeScienceQAMMEMMStar
All (~100k)81.801909.9945.87
30k85.372059.8349.27
10k86.372070.5850.20
+ +Table 4. Impact of DPO Training Data Size. The table presents the performance of models trained with different amounts of DPO training data across various benchmarks. + +solely on the formatted data shown in Fig. 2 (b). Reflective Fine-Tuning enhances the model's performance by 1.48, 62.7, and 3.0 points on the ScienceQA, MME, and MMStar datasets, respectively. With the addition of ConsistDPO training, the model further improves its performance on the MME and MMStar datasets. These results collectively demonstrate that ReCoT significantly boosts overall model performance. + +Effectiveness of Consist-DPO To further examine the impact of the proposed Consist-DPO approach, we conduct an ablation study. As shown in Table 3, the Consist-DPO algorithm yields substantial performance improvements on the validation splits of both training datasets, highlighting its effectiveness in enhancing model consistency. + +Data Size of Consist-DPO To improve DPO training efficiency and prevent overfitting, which can lead to performance degradation for both chosen and rejected responses, we experimented with training on subsets of DPO data. The results show that using only 10k randomly sampled training samples achieves the best performance across all benchmarks (ScienceQA, MME, MMStar), outperforming models trained on the full 100k dataset. This suggests that excessive DPO training may lead to diminishing returns or overfitting. + +Performance of Different Response Strategy Fig. 3 presents a performance comparison of the Mulberry model, trained with ReCoT, on the MME benchmark across three response strategies: reason-then-answer, answer-then-reflect, and direct answer. The results indicate that each + +strategy has its own strengths and weaknesses depending on the type of question. Based on these findings, we can conclude that no single response strategy universally outperforms the others across all question types, and the optimal strategy depends on the specific nature of the task or query. This insight suggests a promising direction for future research, potentially guiding the development of more effective reasoning and response mechanisms in multimodal models. Further exploration could focus on adaptive strategies that dynamically select the most suitable response approach based on the context or complexity of the question, thereby enhancing the overall performance and versatility of multimodal models like Mulberry. + +# 5. Conclusion + +In this paper, we propose ReCoT, a Reflective Self-Correction Training framework to mitigate confirmation bias in small-scale LVLMs. ReCoT follows a two-stage training paradigm, where SFT enhances reflective error correction, and DPO improves answer-reflection consistency. We construct dialog-based reflective samples, introducing incorrect initial answers to encourage self-correction, and propose a consistency correction mechanism to refine DPO training data. Experiments show that ReCoT significantly improves self-reflection and correction abilities, reducing confirmation bias. We also introduce new evaluation metrics to assess reflection accuracy and bias resistance. Our findings highlight the potential of reflective training in improving LVLM reasoning. + +**Limitation** Our work demonstrates the effectiveness of reflective self-correction training in mitigating confirmation bias in small-scale LVLMs. However, challenges remain. We observe that direct and reasoning-based answers perform optimally in different scenarios, suggesting the need for further exploration into how to optimize LVLMs to dynamically decide when to engage in reasoning and reflection and determine the optimal length of such processes. + +Acknowledgements This work was supported by the National Natural Science Foundation of China (No.92470203), Beijing Natural Science Foundation (No.L242022), the Fundamental Research Funds for the Central Universities (2024XKRC082). + +# References + +[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. 2023. 2 +[2] Xiaoyi Bao, Siyang Sun, Shuailei Ma, Kecheng Zheng, Yuxin Guo, Guosheng Zhao, Yun Zheng, and Xingang Wang. Cores: Orchestrating the dance of reasoning and segmentation. In European Conference on Computer Vision, pages 187-204. Springer, 2024. 2 +[3] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Michal Podstawski, Hubert Niewiadomski, Piotr Nczyk, and Torsten Hoefler. Graph of Thoughts: Solving Elaborate Problems with Large Language Models. Proceedings of the AAAI Conference on Artificial Intelligence, 38(16):17682-17690, 2024. 2 +[4] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? Advances in Neural Information Processing Systems, 37:27056-27087, 2024. 2, 6, 7 +[5] Xiaodong Chen, Kunlang He, Wu Liu, Xinchen Liu, Zheng-Jun Zha, and Tao Mei. Clam: An open-source library for performance evaluation of text-driven human motion generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11194-11197, 2024. 1 +[6] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14239-14250, 2024. 3 +[7] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 3, 5 +[8] 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, 67(12):220101, 2024. 2 +[9] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 2 + +[10] Kanzhi Cheng, Yantao Li, Fangzhi Xu, Jianbing Zhang, Hao Zhou, and Yang Liu. Vision-language models can self-improve reasoning via reflection. arXiv preprint arXiv:2411.00855, 2024. 2, 3 +[11] DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. 1 +[12] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 9062–9072, 2025. 3 +[13] Chengyu Du, Jinyi Han, Yizhou Ying, Aili Chen, Qianyu He, Haokun Zhao, Sirui Xia, Haoran Guo, Jiaqing Liang, Zulong Chen, et al. Think thrice before you act: Progressive thought refinement in large language models. arXiv preprint arXiv:2410.13413, 2024. 3 +[14] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 2, 6, 7 +[15] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023. 2 +[16] Kunyang Han, Yibo Hu, Mengxue Qu, Hailin Shi, Yao Zhao, and Yunchao Wei. Rose: Revolutionizing open-set dense segmentation with patch-wise perceptual large multimodal model. arXiv preprint arXiv:2412.00153, 2024. 1 +[17] Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. Reasoning with language model is planning with world model. arXiv preprint arXiv:2305.14992, 2023. 3 +[18] Jiannan Huang, Mengxue Qu, Longfei Li, and Yunchao Wei. Adgpt: Explore meaningful advertising with chatgpt. ACM Transactions on Multimedia Computing, Communications and Applications, 21(4):1-23, 2025. 1 +[19] Weitai Kang, Mengxue Qu, Jyoti Kini, Yunchao Wei, Mubarak Shah, and Yan Yan. Intent3d: 3d object detection in rgb-d scans based on human intention. In The Thirteenth International Conference on Learning Representations. 1 +[20] Tushar Khot, Daniel Khashabi, Kyle Richardson, Peter Clark, and Ashish Sabharwal. Text modular networks: Learning to decompose tasks in the language of existing models. arXiv preprint arXiv:2009.00751, 2020. 3 +[21] Alicia Li, Xiaodong Chen, Bohao Liang, Qian Bao, and Wu Liu. M-adaptor: Text-driven whole-body human motion generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2604-2613, 2025. 1 +[22] Zhenwen Liang, Tianyu Yang, Jipeng Zhang, and Xiangliang Zhang. Unimath: A foundational and multimodal mathematical reasoner. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7126-7133, 2023. 2 +[23] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, + +and Song Han. Vila: On pre-training for visual language models, 2023. 2 +[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. 2 +[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 2, 6 +[26] 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. +[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 2 +[28] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 2, 6, 7 +[29] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegrefe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36, 2024. 3 +[30] OpenAI. Openai o1 system card, 2024. 1, 2 +[31] Daniel E. O'Leary. Confirmation and specificity biases in large language models: An explorative study. IEEE Intelligent Systems, 40(1):63-68, 2025. 2 +[32] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. ArXiv, abs/2306, 2023. 2 +[33] Mengxue Qu, Yu Wu, Wu Liu, Xiaodan Liang, Jingkuan Song, Yao Zhao, and Yunchao Wei. Rio: A benchmark for reasoning intention-oriented objects in open environments. Advances in Neural Information Processing Systems, 36: 43041-43056, 2023. 1 +[34] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvgt: Video temporal grounding via chat with video dialogue large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1847-1856, 2024. 1 +[35] QwenTeam. Qwen2-vl: To see the world more clearly. Wwen Blog, 2024. 1, 2, 6 +[36] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. Glamm: Pixel grounding large multimodal model. CVPR, 2024. 2 +[37] 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. 1 +[38] Jérémy Scheurer, Jon Ander Campos, Tomasz Korbak, Jun Shern Chan, Angelica Chen, Kyunghyun Cho, and Ethan + +Perez. Training language models with language feedback at scale. arXiv preprint arXiv:2303.16755, 2023. 3 +[39] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. CoRR, abs/1707.06347, 2017. 3 +[40] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models, 2024. 1 +[41] Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. In Advances in Neural Information Processing Systems, pages 3008-3021. Curran Associates, Inc., 2020. 3 +[42] Yuhao Sun, Lingyun Yu, Hongtao Xie, Jiaming Li, and Yongdong Zhang. Diffam: Diffusion-based adversarial makeup transfer for facial privacy protection. In CVPR, pages 24584-24594. IEEE, 2024. 1 +[43] Qwen Team. Qvq: To see the world with wisdom, 2024. 1 +[44] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. In The Twelfth International Conference on Learning Representations. 2 +[45] 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. 2 +[46] Sean Welleck, Ximing Lu, Peter West, Faeze Brahman, Tianxiao Shen, Daniel Khashabi, and Yejin Choi. Generating sequences by learning to self-correct. In The Eleventh International Conference on Learning Representations, 2023. 3 +[47] Yuxi Xie, Anirudh Goyal, Wenyue Zheng, Min-Yen Kan, Timothy P. Lillicrap, Kenji Kawaguchi, and Michael Shieh. Monte carlo tree search boosts reasoning via iterative preference learning, 2024. 2 +[48] Guowei Xu, Peng Jin, Hao Li, Yibing Song, Lichao Sun, and Li Yuan. Llava-cot: Let vision language models reason step-by-step, 2024. 1 +[49] Zhen Yang, Jinhao Chen, Zhengxiao Du, Wenmeng Yu, Wei Han Wang, Wenyi Hong, Zhihuan Jiang, Bin Xu, Yuxiao Dong, and Jie Tang. Mathglm- vision: Solving mathematical problems with multi-modal large language model. arXiv preprint arXiv:2409.13729, 2024. 2 +[50] Huanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang, Yibo Wang, Shunyu Liu, Yingjie Wang, Yuxin Song, Haocheng Feng, Li Shen, et al. Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte carlo tree search. arXiv preprint arXiv:2412.18319, 2024. 1, 2, 3, 4, 5, 6, 7 +[51] 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, 2023. 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. 2 +[53] Siyu Yuan, Zehui Chen, Zhiheng Xi, Junjie Ye, Zhengyin Du, and Jiecao Chen. Agent-r: Training language model agents to reflect via iterative self-training. arXiv preprint arXiv:2501.11425, 2025. 3 +[54] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems, 35:15476-15488, 2022. 2 +[55] Dan Zhang, Sining Zhoubian, Ziniu Hu, Yisong Yue, Yuxiao Dong, and Jie Tang. Rest-mcts*: Llm self-training via process reward guided tree search, 2024. 2 +[56] Jinrui Zhang, Teng Wang, Haigang Zhang, Ping Lu, and Feng Zheng. Reflective instruction tuning: Mitigating hallucinations in large vision-language models. In European Conference on Computer Vision, pages 196-213. Springer, 2024. 1, 2, 3, 5, 6, 7 +[57] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Yichi Zhang, Ziyu Guo, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, Shanghang Zhang, et al. Mavis: Mathematical visual instruction tuning. arXiv preprint arXiv:2407.08739, 2024. 2 +[58] Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqing Sun, Zhe Gan, Yinfei Yang, Ruoming Pang, and Yiming Yang. Improve vision language model chain-of-thought reasoning. arXiv preprint arXiv:2410.16198, 2024. 1, 3 +[59] 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. In European conference on computer vision, pages 52-70. Springer, 2024. 2 +[60] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc V Le, et al. Least-to-most prompting enables complex reasoning in large language models. In The Eleventh International Conference on Learning Representations. 3 \ No newline at end of file diff --git a/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/images.zip b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0ba5dfab72fb8a23ce906d00bc180e6727fc28f5 --- /dev/null +++ b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40e0125cb103190e29242fabedc4a0264ec4eae7bcfdda5f482820013645073b +size 391941 diff --git a/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/layout.json b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4828c2b57b51553aa6882f23ff6805e042bc5a7a --- /dev/null +++ b/ICCV/2025/ReCoT_ Reflective Self-Correction Training for Mitigating Confirmation Bias in Large Vision-Language Models/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c85a45cd40851bcf34e8f4eedaf8b9168e0299342bda1bd1bb0b01df3eb033f +size 386287 diff --git a/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/2530398c-cd9d-4f17-964e-d2a853017942_content_list.json b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/2530398c-cd9d-4f17-964e-d2a853017942_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4662ff3a7a17a24b3e996196e4e81e4ee35d4d95 --- /dev/null +++ b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/2530398c-cd9d-4f17-964e-d2a853017942_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c3d0ec3afe224e4a49927dd6299450efea74a1cc85ce64e68adf6494d810ff8 +size 69014 diff --git a/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/2530398c-cd9d-4f17-964e-d2a853017942_model.json b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/2530398c-cd9d-4f17-964e-d2a853017942_model.json new file mode 100644 index 0000000000000000000000000000000000000000..85231f901181a6108234d1635e5c0f4f89958a39 --- /dev/null +++ b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/2530398c-cd9d-4f17-964e-d2a853017942_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b1564fbf43199b14c245b4381560ccc53673cecc3f673ec62bffca9ef08477e +size 86019 diff --git a/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/2530398c-cd9d-4f17-964e-d2a853017942_origin.pdf b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/2530398c-cd9d-4f17-964e-d2a853017942_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..104a58f8af4a29f0e6c257c5aeb0a9436b87850f --- /dev/null +++ b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/2530398c-cd9d-4f17-964e-d2a853017942_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bf7ed859ef3824992fa3863150e6af3b43076a801cf893bf69b8a944bda07cf +size 5380074 diff --git a/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/full.md b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d1deb1a4ab3bf942a8e09d9507fcfff5f8c6e7dd --- /dev/null +++ b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/full.md @@ -0,0 +1,263 @@ +# ReFlex: Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation + +Jimyeong Kim1 Jungwon Park1 Yeji Song1 Nojun Kwak1,2 Wonjong Rhee1,2,3 +Department of Intelligence and Information1 & IPAI2, & RICS3 Seoul National University +{wlaud1001, quoded97, ldynx, nojunk, wrhee}@snu.ac.kr + +![](images/36fa6631637ea5e0c3cff90185a396254b4072262a8a48b8548c3e6f7c30877d.jpg) +Figure 1. Given a source image and a target prompt, our method can preserve core information of the source image, including structure and background, while adapting high-level attributes in accordance with the target prompt. + +# Abstract + +Rectified Flow text-to-image models surpass diffusion models in image quality and text alignment, but adapting ReFlow for real-image editing remains challenging. We propose a new real-image editing method for ReFlow by analyzing the intermediate representations of multimodal transformer blocks and identifying three key features. To extract these features from real images with sufficient structural preservation, we leverage mid-step latent, which is inverted only up to the mid-step. We then adapt attention during injection to improve editability and enhance alignment to the target text. Our method is training-free, requires no user-provided mask, and can be applied even without a source prompt. Extensive experiments on two benchmarks with nine baselines demonstrate its superior performance over prior methods, further validated by human evaluations confirming a strong user preference for our approach. + +# 1. Introduction + +The emergence of text-to-image foundation models [28, 30, 33] has not only empowered image generation but also en + +abled text-guided editing of real images [2-4, 11, 13, 22, 23, 36]. While early methods on text-guided real-image editing were predominantly developed using U-Net based diffusion models (DMs), such as Stable Diffusion [30], their effectiveness is often constrained by the inherent performance limitations of DMs, leading to compromised image quality and weaker text alignment in the edited images. Recently, Rectified Flow (ReFlow)-based text-to-image models, such as FLUX [17], have demonstrated substantially higher image quality and stronger text alignment than DMs by leveraging flow matching [19, 20] and multi-modal diffusion transformer (MM-DiT) [8, 26]. However, existing ReFlow-based image editing methods [7, 15, 32, 39] often struggle with real image editing, particularly in preserving the source structure and handling significant modifications, such as color changes. While techniques developed for DM-based editing methods provide a useful starting point, adapting them to ReFlow models has been challenging. This difficulty arises from the differences in training process and architecture between DMs and ReFlow, meaning that a simple plug-and-play approach cannot be applied. + +In DM-based editing, feature-based methods have emerged as the most effective and widely adopted approach + +for image editing [4, 11, 25, 36]. Within this framework, cross- and self-attention in U-Net have been identified as the key features, encoding text-image relationships and image structural information, respectively [11, 36]. However, in ReFlow models, these features are not directly applicable. This is because, unlike U-Net, MM-DiT entangles text and image information within the architecture through joint self-attention [17, 32]. As a workaround, previous studies [7, 39] have explored the use of value features in MM-DiT, but these approaches often suffer from appearance leakage. As of today, it remains unclear which MM-DiT features can enable effective editing. To address this problem, we conduct a broad investigation of MM-DiT features. Specifically, we decompose MM-DiT's joint self-attention map into four components based on query-key modality and examine them. Furthermore, we investigate residual and identity features of image embeddings. Through our investigations, we identify three key features of MM-DiT that are effective for real-image editing: two components of the joint self-attention map that are associated with image query and the residual feature. + +Even after identifying the three key features, another major challenge lies in extracting the features reliably from real images and injecting them into the target image in an effective manner. This process remains highly challenging as it requires inverting the image into the noise and reconstructing it back to its original image [11, 22, 36]. Various inversion methods have been explored in DMs [13, 22, 34], but they cannot be directly applied to ReFlow due to differences in sampling process. Furthermore, existing ReFlow-based inversion methods [7, 32, 39] often fail to accurately reconstruct the original image. To address these challenges, we introduce a simple yet effective approach, mid-step feature extraction: instead of relying on the reconstruction process that starts from the fully inverted latent, we extract features from the mid-step latent, which is inverted only up to the mid-step. As shown in Fig. 2, the mid-step latent exhibits significantly higher reconstructability than the fully inverted latent, thereby providing more accurate source information for feature extraction. While the use of mid-step feature in our method is crucial for source structure preservation, this approach also causes a reduction in editability because of its excessive preservation of source information. To counter this effect and recover editability, we introduce adaption techniques that can be effectively applied in the injection step to the two key features that are associated with joint self-attention map. For instance, we strengthen the feature with the query-key modality of image-text. For the third key feature, residual feature, we inject it without any modification. For local editing task, we extract an editing mask from the joint self-attention map and apply latent blending [2] to mitigate unintended alterations in non-edit regions. + +Our method, ReFlex (Recified Flow via Mid-Step Fea + +![](images/d960434df85382a3eb6d6b17e1cca01d75bdcf85cb6957314f0414f098f9b92d.jpg) + +![](images/219dbe052b242f144eeee132d44f8ade876cbf2182c5d11924a8c63cb68da64e.jpg) + +![](images/be71ada42adaf24f9a1e0cb62ce84f40593ba52647a1df15ff2481c9b1ed5ed5.jpg) +red short and blue in a grassy field. + +![](images/7d6a71af9f01d3a32288a583611cae318f9eee9e12c2044822ebddb12ffde7f7.jpg) + +![](images/cd13fc8c640cefba7e92f68d8d1e53887dd7ba992105aca1ac3ae4d20ae85637.jpg) +a man wearing a red shirt and blue pants is jumping in a grassy field. + +![](images/723a71149a2864dceddada44e5413bc5f9cf915995549b6cc453f5b431cf5fb6.jpg) + +![](images/f07f7da4e14a0cd556ad2595db120a21ff4b18577ba9be0910c80a10260fd45a.jpg) +a photo of a pizzasuperhero in NYC +Figure 2. Mid-step feature extraction. The vanilla ReFlow inversion extracts features starting from the fully inverted latent, during the image reconstruction process. In contrast, our method extracts features at the mid-step of the inversion. While ReFlow models struggle with structural preservation in the case of fully inverted latents, our mid-step approach effectively resolves this problem. + +ture Extraction and Attention Adaptation), uses the state-of-the-art ReFlow-based model FLUX [17] as its backbone and demonstrates strong performance in real image editing. We compare ReFlex against nine baseline methods, including four FLUX-based and five Stable Diffusion (SD)-based methods, using two widely used benchmarks: PIE-Bench [14] and Wild-TI2I-Real [36]. ReFlex achieves text alignment improvements of $1.69\%$ to $7.11\%$ on PIE-Bench and $3.21\%$ to $16.46\%$ on Wild-TI2I-Real. Furthermore, in human evaluation, ReFlex is preferred in $68.2\%$ of cases against four FLUX-based methods, while the second-best method is preferred in only $11.2\%$ of cases. Similarly, in the comparison with four SD-based methods, ReFlex is preferred in $61.0\%$ of cases, with the second-best method preferred in only $11.4\%$ of cases. These results demonstrate the effectiveness of ReFlex in producing high-quality results. + +# 2. Related work + +# 2.1. Image editing for diffusion models + +Diffusion Model (DM)-based image editing methods fall into three categories: training-based, mask-based, and feature-based. Training-based methods [3, 16, 23] require large datasets and overheads in optimization time. Mask-based methods [2, 6, 23, 40] require user-specified mask and often lead to inconsistencies between edited and unedited regions. Feature-based methods [4, 11, 25, 36] are the most widely used, leveraging intermediate diffusion features from source image generation to inject structural information into the target image. For example, P2P [11] uses cross-attention maps, while PnP [36] employs self-attention maps and spatial features extracted from residual blocks. However, existing feature-based approaches have been developed specifically for the U-Net architecture of DMs, with limited exploration of newer architectures like MM-DiT in ReFlow models. To address this limitation, we carefully analyze the features of ReFlow models and leverage the identified features to significantly enhance editing performance. To extend feature-based editing to real-image editing, several DM-based inversion methods [13, 22, 31, 34] have been + +![](images/e80e41e9e81509b2ca505a5f6a915cbb9d7d5a48f012666a2fca0f6121fe7d5b.jpg) +Figure 3. Overview of our method, ReFlex: (a) We extract three key features—I2T-CA, I2I-SA, and the residual feature—from a mid-step latent, while the fully inverted latent serves as initial noise for target image generation. To enhance text alignment, we propose two adaptation methods, respectively for two attention maps, (b) I2T-CA and (c) I2I-SA, whereas residual feature is injected without modification. (d) For local editing, we generate an editing mask from the source I2T-CA. This injection process is applied only during the early timesteps of target image generation. + +![](images/d2fd628aa5e8da4ebea9d64ea586467c0dd8557b35c8a30b12f18ce91636e704.jpg) + +developed. These methods often incorporate optimization process [22, 31] or are not directly applicable to the ReFlow [13, 22, 34] due to differences in the sampling process. In contrast, our mid-step feature extraction eliminates the need for optimization, enabling effective feature extraction from real images in ReFlow model. + +# 2.2. Image editing for rectified flow models + +With the advancement of ReFlow-based text-to-image models [8, 17], recent studies have explored their application to image editing [7, 15, 32, 39]. RF-Inversion [32] constructs a controlled ODE by leveraging a conditional vector field formed through the interpolation of the source image and noise, heavily relying on the source image. RF-Edit [39] and FireFlow [7] introduce second-order Taylor expansion to reduce reconstruction error and utilize value features extracted during inversion for editing, but they still struggle to achieve effective editing. In addition, the use of value features often results in appearance leakage. FlowEdit [15] bypasses inversion by constructing a direct ODE between source and target images, minimizing changes during editing. In our work, we found that previous methods typically fail to precisely preserve the overall source structure, or struggle with large modifications, such as color changes. In contrast, our method introduces mid-step feature extraction for source preservation and applies two adaptation techniques to balance the source and target information. + +# 3. Preliminaries + +Rectified flow (ReFlow)-based models differ from diffusion models (DM) in two key aspects: training and architecture. In terms of training, ReFlow leverages flow matching [19, 20], which is explained in detail in Appendix A.1. Regarding architecture, ReFlow models employ the multimodal diffusion transformer (MM-DiT) [8, 17, 26]. Unlike + +U-Net, which consists of cross-/self-attention and residual blocks, MM-DiT is composed of modulation layers, MLP layers, and joint self-attention. The joint self-attention processes text and image tokens jointly by concatenating them within a self-attention as $Q = [Q_{text};Q_{image}]$ , with the same operation applied to keys and values, ensuring direct interaction between modalities. MM-DiT also incorporates a residual connection [10] at the end of each block. An explanation on our backbone model, FLUX [17], is provided in Appendix A.2. + +# 4. Method + +Given a real image $\mathcal{I}$ , a corresponding source prompt $P_{S}$ and a target prompt $P_{T}$ , our goal is to generate an image that aligns with $P_{T}$ while preserving the overall structure of $\mathcal{I}$ . To achieve this, we analyze MM-DiT's features and identify three key features that effectively preserve the source structure in Sec. 4.1. In Sec. 4.2, we propose an effective method for extracting identified features from real images. Sec. 4.3 introduces our feature injection process along with attention adaptation methods to enhance editability, while Sec. 4.4 describes the editing mask generation for latent blending. An overview of our method is presented in Fig. 3. + +# 4.1. Observations: three key features in MM-DiT + +Unlike the well-studied diffusion features in U-Net [11, 36], ReFlow features in MM-DiT remain largely unexplored. To address this, we analyze ReFlow features using generated images, which is more straightforward than using real images. Since diffusion features are mainly found in attention maps and residual blocks [11, 36], we focus on two key sources in MM-DiT: (1) joint self-attention maps and (2) the residual connection. In the following, we identify three key features, with more details in Appendix B.1. + +![](images/30f87e736f7b1c70b0d018a3ded8619474acc55726779ea1a4d02bdfad30ef8f.jpg) +(a) Four components in the MM-DiT's attention. + +![](images/3a4866f3bc7788aac80fc1faba630b0131d7aba3b46b3b9f8ed93e74525601d9.jpg) +(b) Residual and identity features in the MM-DiT. +Figure 4. Analysis on MM-DiT block. (a) Upper row: The PCA result of I2I-SA and the attention maps of I2T-CA and T2I-CA for two words. Lower row: Edited results generated by injecting features, indicated above each image. (b) PCA results of residual and identity features, along with edited results generated by injecting these features from MM-DiT blocks. + +Two key features from the joint self-attention map: Diffusion models' U-Net employs two types of attention layers: cross-attention, which captures text-image relations, and self-attention, which captures image structure. ReFlow's MM-DiT replaces these with joint self-attention, processing image and text embeddings together. To better analyze the joint self-attention map $Q \cdot K^T$ , we decompose it into four components based on query-key modalities: I2I-SA, I2T-CA, T2I-CA, and T2T-SA. For example, I2T-CA (image-to-text cross-modality attention) corresponds to $Q_{\mathrm{image}} \cdot K_{\mathrm{text}}^T$ . The upper row of Fig. 4(a) visualizes a PCA of I2I-SA alongside I2T-CA and T2I-CA for two words ('bear' and 'floral'), showing that I2I-SA encodes structural information, while I2T-CA and T2I-CA capture text-image relationships. To evaluate their role in editing, we inject each component during the target image generation and examine the results in the lower row of Fig. 4(a). This reveals that injecting I2I-SA and I2T-CA preserve the source image structure in the target output, whereas T2I-CA and T2T-SA do not. Two factors may explain this: (1) The output of the attention is aggregated along the query dimension [30, 37], meaning I2I-SA and I2T-CA directly influence image token embeddings, while T2I-CA and T2T-SA directly influence text token embeddings. (2) Only image token embeddings from the final layer are passed to the image decoder for generation. Based on these findings, we identify I2I-SA and I2T-CA as key features for editing. + +A key feature from residual connection: Unlike U-Net architecture, MM-DiT lacks residual blocks. Therefore, we analyze two features from MM-DiT's residual connections: residual and identity features. MM-DiT applies residual connections at the end of each block, producing outputs + +![](images/726609d26fae0fbf23b48e7830d52fc4f25c5f26bb57e9c7516b71cc7917a3eb.jpg) +Figure 5. Image reconstruction from intermediate latents at varying timesteps. Latents near noise $(t = T)$ struggle to fully preserve the source structure, while the mid-way latent $(t = T / 2)$ achieves near-perfect source preservation. This suggests that features extracted from earlier timesteps $(t \approx T)$ may contain inaccurate information about the source image. Here, $T = 28$ . + +as MM-DiT(x) = $f(x) + x$ , where $x = [x_{\mathrm{text}}; x_{\mathrm{image}}]$ . Since image token embeddings are crucial for structure preservation, we define the residual and identity features as $f(x)_{\mathrm{image}}$ and $x_{\mathrm{image}}$ , respectively. Fig. 4(b) shows PCA visualizations of these features and their impact when injected during target image generation. While both capture structural information, the identity feature retains excessive appearance details, limiting editing flexibility. Therefore, we identify only the residual feature as a key for editing. + +# 4.2. Mid-step feature extraction + +Prior DM-based methods [11, 36] typically extract features at every timestep during the reconstruction of the fully inverted latent and inject them at early timesteps in target image generation. However, applying this conventional approach to ReFlow-based models can degrade editing results, as the extracted features may be less relevant to the source image. This is because ReFlow models struggle to reconstruct images from inverted latents, leading to significant deviations from the original images. An example is shown in Fig. 5, where images are reconstructed from intermediate latents $\{z_{t}\}_{t = 0}^{T}$ at varying timesteps. The figure demonstrates that latents at timesteps earlier than the midpoint $(t < T / 2)$ fail to fully preserve the source structure, which can possibly cause features extracted from these intermediate latents to contain inaccurate information about the source image. In contrast, the mid-way latent $(t = T / 2)$ achieves near-perfect source preservation, resolving this issue. Motivated by this, we propose mid-step feature extraction, where features are extracted from a single mid-step latent $z_{t'}$ at $t' = T / 2$ . The effect of varying $t'$ on editing performance is discussed in Sec. 5.4, supporting $t' = T / 2$ as the optimal extraction point. + +# 4.3. Two feature adaptation techniques + +The main idea of our method is to extract three key features from the mid-step latent $z_{t'}$ and then to inject them during the early timesteps of target image generation. However, + +![](images/5fcdfcc860b2025aa8923cf39abb369aa3fe7ac847752ecd7e4ca499f715de40.jpg) +Figure 6. Visualization of the I2I-SA values for each query, marked with a colored $+$ in the source image, on both a linear and log scale, shown in the first and second rows, respectively. + +careful consideration is required because the timestep at which features are extracted is different from the timesteps at which features are injected. To address this issue, we introduce two feature adaptation techniques that balance the source and target information. + +I2T-CA adaptation: We define a mapping function $f$ that takes a text token index from the target prompt $P_T$ and returns its corresponding text token index in the source prompt $P_S$ . If no corresponding token is found in $P_S$ , $f$ returns $\emptyset$ . For example, given $P_S =$ "a goat and a cat" and $P_T =$ "a horse and a cat" the token "horse" has no corresponding token in $P_S$ . Thus, when $f$ receives the token index of "horse", it outputs $\emptyset$ . For each target text token index $i$ , if $f(i)$ exists, the target's I2T-CA of the $i$ th target text token is replaced with the source's I2T-CA of the $f(i)$ th source text token. If $f(i)$ returns $\emptyset$ , the target's I2T-CA of the $i$ th target text token is scaled by a factor of $\alpha > 1$ and used instead. Thus, our final formulation is given as follows: + +$$ +C A ^ {\prime} [:, i ] = \left\{ \begin{array}{l l} \alpha \times C A _ {T} [:, i ], & \quad \text {i f} f (i) = \emptyset \\ C A _ {S} [:, f (i) ], & \quad \text {o t h e r w i s e} \end{array} \right. +$$ + +Here, $CA^{\prime}$ denotes the adapted I2T-CA, $CA_{T}(:,i]$ is the target's I2T-CA of $i$ th target text token, $CA_{S}(:,i]$ is the source's I2T-CA of $i$ th source text token. We then inject $CA^{\prime}$ instead of $CA_{S}$ during target image generation. If the source prompt is not provided, or if the source and target prompts have different sentence structures, defining the mapping function $f$ becomes challenging. In such cases, we define $f(i)$ as $\emptyset$ for all $i$ . The I2T-CA adaptation is inspired by the prompt refinement and attention re-weighting proposed in P2P [11]. + +I2I-SA adaptation: We found that directly injecting I2ISA without modification often results in poor edits, especially when the target prompt specifies large structural changes. To better understand this issue, we visualize how each query attends to different pixel locations in I2I-SA, as shown in Fig. 6. The four columns on the right display I2I-SA values for four different queries, corresponding to four different rows of I2I-SA (i.e., $Q_{\mathrm{image}} \cdot K_{\mathrm{image}}^T$ ), with each query highlighted with a distinct color on the left image. For better visualization, we reshape each row vector + +of I2I-SA into a 2D square format and overlay it onto the resized original image. The four columns on the right side present these visualizations using two different scales. From the linear scale in the upper row, we observe that each query overly concentrates on a small number of pixels near its own location. In contrast, the log scale in the lower row reveals that I2I-SA effectively captures the overall structure of the source image at the same time. Motivated by this, we propose replacing the overly concentrated top- $k$ attention values in the source's I2I-SA with the corresponding values from the target's I2I-SA. This approach prevents source injection from overly preserving local structures while maintaining the global structure of the source image. The replaced target values are normalized for each row to ensure that the scale of each row remains unchanged after the adaptation. This adaptation is formulated as below: + +$$ +S A ^ {\prime} [ i, j ] = \left\{ \begin{array}{l l} S A _ {T} [ i, j ] \times \frac {\sum_ {j ^ {\prime} \in \mathcal {K} (i)} S A _ {S} [ i , j ^ {\prime} ]}{\sum_ {j ^ {\prime} \in \mathcal {K} (i)} S A _ {T} [ i , j ^ {\prime} ]}, & j \in \mathcal {K} (i) \\ S A _ {S} [ i, j ], & \text {o t h e r w i s e} \end{array} \right. +$$ + +Here, $SA_{S}^{\prime}$ denotes the adapted I2I-SA, $SA_{S}[i,j]$ is the $(i,j)$ component of the source's I2I-SA, $SA_{T}[i,j]$ is the $(i,j)$ component of the target's I2I-SA, and $\mathcal{K}(i)$ is the set of indices corresponding to the top- $k$ attention values in the $i$ -th row of $SA_{S}$ . We then inject $SA^{\prime}$ instead of $SA_{S}$ during target image generation. + +# 4.4. Mask generation for latent blending + +Latent blending [2] is a powerful technique for local editing, enabling modifications confined to specific regions of an image. To utilize this approach, we propose a mask generation process that operates when source prompts are provided. The process begins with the user specifying a blended word from the source prompt, which defines the subject of the edit—for example, in the source prompt 'photo of a red car' and the target prompt 'photo of a blue car', the blended word is 'car'. We then extract the I2T-CA of the blended word, apply Gaussian smoothing, and use Otsu's thresholding method [24] to generate a binary mask $M_{t}$ for the initial $m$ timesteps of target image generation. This mask is used to compute the blended latent $z_{t}^{\text{blend}}$ as $z_{t}^{\text{blend}} = M_{t} \odot z_{t} + (1 - M_{t}) \odot z_{t}^{\text{source}}$ , where $z_{t}$ is the target latent at timestep $t$ , $z_{t}^{\text{source}}$ is the source latent, and $\odot$ denotes the Hadamard product. Instead of using $z_{t}$ , we input $z_{t}^{\text{blend}}$ into the subsequent ReFlow iteration for the first $m$ timesteps of target image generation. Since this mask generation process requires specifying the blended word from the source prompt, it is applied only when source prompts are available. While blended words could be detected automatically, we currently rely on predefined or manually selected words and leave the development of an automatic detection method for future work. + +![](images/1ec90a22e6796496dc45e214ea32b67f79a9c950a1628e37fa15e5f6ddaeb3f2.jpg) +(a) PIE-Bench +Figure 7. Qualitative evaluations on (a) PIE-Bench [14] and (b) Wild-TI2I-Real [36]. When a source prompt is available (PIE-Bench), the removed text from the source prompt is indicated with a striking through (e.g., removed) and the text added in the target prompt is indicated in bold within brackets (e.g., [added]). If no source prompt is provided (Wild-TI2I-Real), only the target prompt is displayed. + +# 5. Experiments + +Datasets: We evaluated our method, ReFlex, using two widely used benchmarks: PIE-Bench [14] and Wild-TI2I-Real [36]. PIE-Bench consists of 700 images across 9 tasks. It provides source and target prompts along with blended words, which can be used in latent blending. Wild-TI2I-Real comprises 78 pairs of images and target prompts but does not provide source prompts. + +Implementation Details: We used FLUX [17], the state-of- + +the-art text-to-image generation model, as our base model. Throughout the comparisons, we fixed our hyperparameters: $T = 28$ for sampling steps, $t' = T / 2$ (i.e., 14) for feature extraction, $\alpha = 4$ for I2T-CA adaptation, and $k = 20$ for I2I-SA adaptation. When source prompts are provided, we inject I2T-CA for the initial $0.4T$ steps, I2I-SA for $0.25T$ , and residual features for $0.15T$ steps. When source prompts are not provided, we omitted the I2T-CA injection and use $0.4T$ steps for I2I-SA injection and $0.25T$ steps for residual features injection. In PIE-bench, we applied la + +![](images/e1c4afbc42d23642ba297b0bf6e4870d93dc30da9becfb3d6f15875027eab4b0.jpg) +(a) PIE-Bench + +![](images/10397da97ada491038815637476c5cca5b2c25b4210303584753f76029abc73d.jpg) +(b) Wild-TI2I-Real +Figure 8. Trade-off plot between editability (text alignment) and source preservation (PSNR and IoU). + +
Comparison with Flux-based methods
MethodRF-InversionRF-EditFireFlowFlowEditOurs
User Preference7.8%7.3%5.5%11.2%68.2%
Comparison with SD-based methods
MethodDI+PnPDDIM+PnPSDEditP2P-ZeroOurs
User Preference11.4%7.5%9.4%10.6%61.0%
+ +Table 1. User study results comparing our method with four FLUX-based and four SD-based methods. + +tent blending using the provided blended words, whereas in Wild-TI2I-Real, we omitted latent blending. More implementation details are provided in Appendix B. + +Baseline methods for comparison: We compared our method against editing methods designed for both FLUX [17] and SD [30] as backbone models. For FLUX-based methods, we considered RF-Inversion [32], RF-Edit [39], FireFlow [7], and FlowEdit [15] across all benchmarks. For SD-based methods, we evaluated P2P [11] + Direct Inversion [14], P2P + Null Text Inversion [22], and PnP [36] + Direct Inversion for PIE-Bench. For Wild-TI2I-Real, P2P-based methods were excluded due to the lack of source prompt. Instead, we compared with PnP + DDIM, PnP + Direct Inversion, P2P-Zero [25], and SDEdit [21]. + +# 5.1. Qualitative evaluation + +Qualitative results are shown in Fig. 7. The first four rows present results from the PIE-bench dataset, while the last four rows correspond to the Wild-TI2I-Real dataset. Unlike our method, both FLUX-based and SD-based approaches struggle to accurately follow the target prompt. For instance, all compared models fail to transform the heart shape into a circle in the first row and to change the yellow center to red in the third row. Additionally, SD-based models exhibit lower overall image quality than FLUX-based models due to their inferior backbone, as evidenced by the artifacts in the fifth and seventh rows. In contrast, our method successfully edits the source image with precise adjustments and color modifications that align with the target prompt while maintaining high image quality and preserving the overall source structure. Further qualitative comparisons for both benchmark datasets can be found in Fig. S.3 and Fig. S.4, while additional diverse edited results of our method are provided in Fig. S.5 and Fig. S.6 of Appendix C. + +![](images/de7637ace280e9b58086c0ee2299e9e09e46bc259bf65df125affafd42840b7d.jpg) +Figure 9. Ablation examples for assessing the impact of each technique in our method. + +# 5.2.Quantitative evaluation + +We conducted a quantitative comparison from two perspectives: (1) text alignment and (2) source preservation. Text alignment evaluates how well the edited images align with the target prompt, while source preservation quantifies the preservation of the structure and content from the source image. For text alignment, we measured the CLIP text score [27]. To assess source preservation, we used two different metrics for Wild-TI2I-Real [36] and PIE-Bench [14]. In WILD-TI2I-Real, we extracted segmentation masks from both source and target images using Grounded-SAM [29] and computed their IoU, following [1]. For PIE-Bench, where tasks involve complex edits like object addition and removal, IoU is inapplicable. Instead of IoU, we used the provided editing mask to compute PSNR in non-edit regions, assessing the preservation of unchanged parts of the source image, as done in [14]. We visualized the results using a 2D plot in Fig. 8, where the $x$ -axis represents text alignment and the $y$ -axis represents source preservation. Our method outperforms all the others in text alignment, showing improvements of $1.69\%$ to $7.11\%$ in PIE-Bench and $3.21\%$ to $16.46\%$ in WILD-TI2I-Real. Our method effectively balances the trade-off between text alignment and source preservation, consistently achieving Pareto optimality. Additional results can be found in Appendix D. + +# 5.3. User study + +We conducted two user studies to compare our method against four FLUX-based methods and four SD-based methods using the full Wild-TI2I-Real dataset. Each participant was presented with five images, generated from the same source image and target prompt but using different methods, and asked to select the one that best represents the + +![](images/3a7bc72770ef69a686bbe3e4abff1a7ef5da2ce0bf7f0fabb720e5edae6ff183.jpg) +(watercolor of) a woman in sunglasses and leather pants sitting on a bench + +![](images/e50ba3930a7ae16db4d259ad4e468769935fb8c275a3ad9b001e50148aebcace.jpg) +Figure 10. Effect of the $t'$ selection in mid-step feature extraction, where $t'$ is the timestep of the latent from which features are extracted. + +![](images/9b53c2151bf5690bd2a1c5cb63e91d5345f4e83fba82720706f1659fedad4b1f.jpg) +a painting of a cabin (car) in the snow with mountains in the background + +![](images/db054a136f2e672f1df0fe387d124c07d7e433e188c0d11d51ec01a34d3235f7.jpg) +a photo of a silver robot in the snow +Figure 11. Effect of varying $k$ in I2I-SA adaptation, where $k$ denotes the number of top attention values replaced. + +intended edit. Using Amazon Mechanical Turk (MTurk), we collected 1,410 responses from 97 valid participants for the FLUX-based comparison and 1,530 responses from 102 valid participants for the SD-based comparison. As shown in Tab. 1, our method significantly outperforms the others, receiving $68.2\%$ of votes in the FLUX-based comparison and $61.0\%$ in the SD-based comparison. Additional details on the user study, including survey questions and participant filtering criteria, can be found in Appendix D.3. + +# 5.4. Ablations + +Role of key techniques. We conducted an ablation study on three key techniques of our method: mid-step feature extraction, I2T-CA adaptation, and I2I-SA adaptation. First, we replaced mid-step extraction with the conventional feature extraction approach used in previous DM-based editing methods, where features are extracted at every early timestep during the reconstruction of the fully inverted latent. For the other two techniques, we simply removed them during editing. The star-marked points in Fig. 8 (b) present the quantitative results, while Fig. 9 provides a qualitative comparison. Without mid-step extraction, much of the source structure is lost in the edited images, resulting in a low IoU score. This may stem from extracting features from latents that contain inaccurate information about the source image, as shown in Fig. 5. Meanwhile, without the two adaptation techniques, the edits fail to fully align with the target prompt, leading to a lower text-alignment score. These highlight the importance of balancing source and target information in editing, which our adaptation techniques effectively achieve. Additional qualitative comparisons can be found in Fig. S.9 of Appendix E. + +![](images/e2462462f6955b1340103148f057f759207f4265584ce2ce3acb9679e349a05e.jpg) +(a) Incomplete source preservation +Figure 12. Two examples illustrating the limitations of our method: (a) loss of fine-grained source details, and (b) introduction of artifacts due to inaccuracies in the generated editing mask. + +![](images/107b4d4e9401b7f220f7792d6c080e83ca22d358193f669bfadee801aca75c27.jpg) +(b) Limitation of editing mask generation + +Effect of varying $t'$ in mid-step feature extraction: Fig. 10 shows the editing results with varying $t'$ , the timestep at which feature extraction occurs. Features from the fully inverted latent $(t' = 28)$ fail to preserve the source structure. On the other hand, features from a latent too close to the image $(t' = 4)$ degrade overall image quality, most likely because of the large gap between the timestep where feature extraction occurs and the timesteps where injection takes place. Features from the mid-step latent $(t' = 14)$ yield the best editing results with high image quality. More comparisons can be found in Fig. S.10 of Appendix E. + +Effect of varying $k$ in I2I-SA adaptation: Fig. 11 illustrates the effect of varying $k$ in I2I-SA adaptation. When only the source I2I-SA is used, the source structure remains unchanged. Increasing $k$ adjusts the image structure to improve alignment with the target prompt. However, if $k$ is set too high, it may result in the loss of source structure information. Overall, setting $k = 20$ achieves the best balance between structure preservation and editability, though the optimal value of $k$ may vary depending on the sample. Additional results can be found in Fig. S.11 of Appendix E. + +Effect of varying $\alpha$ in I2T-CA adaptation can be found in Fig. S.12 of Appendix E. + +# 6. Limitations and Conclusion + +Our method has several limitations. First, when the edited region overlaps with the subject, it may unintentionally change other features of the subject, as shown in Fig. 12 (a). In this example, while our model successfully removes the woman's glasses, it also alters her hair, which was not specified in the target prompt. Second, the editing mask generated from I2T-CA may not perfectly localize the editing region, leading to artifacts in the edited image, such as misaligned trees in Fig. 12 (b). This issue could potentially be mitigated with more advanced mask generation techniques. Additionally, in Appendix F, we present another limitation related to variability in editing results, along with further examples of the two previously mentioned limitations. + +In conclusion, we address the challenge of real-image editing in ReFlow by identifying key features in MM-DiT and proposing a simple yet effective approach for feature extraction, complemented by two adaptation techniques. We hope our findings serve as a building block for future advancements in real-image editing. + +# 7. Acknowledgments + +This work was supported by the following grants funded by the Korea government: NRF (2022R1A6A1A03063039, 2021R1A2C3006659) and IITP ([NO.RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)], [No. RS-2023-00235293]). + +# References + +[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 7 +[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 2, 5 +[3] 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. 2 +[4] 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, pages 22560-22570, 2023. 1, 2 +[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] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2 +[7] Yingying Deng, Xiangyu He, Changwang Mei, Peisong Wang, and Fan Tang. Fireflow: Fast inversion of rectified flow for image semantic editing. arXiv preprint arXiv:2412.07517, 2024. 1, 2, 3, 7 +[8] 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. 1, 3 +[9] fallenshock. FlowEdit. https://github.com/fallenshock/FlowEdit.2 +[10] 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 +[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 3, 4, 5, 7 + +[12] HolmesShuan. FireFlow-Fast-Inversion-of-Rectified-Flow-for-Image-Semantic-Editing. https://github.com/HolmesShuan/FireFlow-Fast-Inversion-of-Rectified-Flow-for-Image-Semantic-Editing.2 +[13] 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. 1, 2, 3 +[14] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2, 6, 7, 3 +[15] Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. arXiv preprint arXiv:2412.08629, 2024. 1, 3, 7, 2 +[16] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2 +[17] Black Forest Labs. Announcing black forest labs, 2025. 1, 2, 3, 6, 7 +[18] 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 +[19] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 1, 3 +[20] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 1, 3 +[21] 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. arXiv preprint arXiv:2108.01073, 2021. 7, 2 +[22] 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. 1, 2, 3, 7 +[23] Alex 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. arXiv preprint arXiv:2112.10741, 2021. 1, 2 +[24] Nobuyuki Otsu et al. A threshold selection method from gray-level histograms. Automatica, 11(285-296):23-27, 1975. 5 +[25] 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. 2, 7 +[26] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF Inter- + +national Conference on Computer Vision, pages 4195-4205, 2023. 1, 3 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 3 +[28] 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. 1 +[29] 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. 7 +[30] 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, 4, 7 +[31] Litu Rout, Yujia Chen, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Beyond first-order tweedie: Solving inverse problems using latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9472-9481, 2024. 2, 3 +[32] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inversion and editing using rectified stochastic differential equations. arXiv preprint arXiv:2410.10792, 2024. 1, 2, 3, 7 +[33] 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 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3 +[35] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 3 +[36] 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, pages 1921-1930, 2023. 1, 2, 3, 4, 6, 7 +[37] 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. 4 +[38] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, + +Dhruv Nair, Sayak Paul, Steven Liu, William Berman, Yiyi Xu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers.2 +[39] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 1, 2, 3, 7 +[40] Qian Wang, Biao Zhang, Michael Birsak, and Peter Wonka. Instructedit: Improving automatic masks for diffusion-based image editing with user instructions. arXiv preprint arXiv:2305.18047, 2023. 2 +[41] 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. 3 +[42] wangjiangshan0725. RF-Solver-Edit. https://github.com/wangjiangshan0725/RF-Solver-Edit. 2 +[43] 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. 3 \ No newline at end of file diff --git a/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/images.zip b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0ae5bb1253232127c1af2f604be248cdbfcadc62 --- /dev/null +++ b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce144b1e744fac9811d2fdc69dfde5ef2238573febe51a5249636d7250f34d59 +size 886090 diff --git a/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/layout.json b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..82a7545aceccdefa751486086b9662304f10c036 --- /dev/null +++ b/ICCV/2025/ReFlex_ Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e351796c81f78bf53d8b9515e84eb818c18da5e540b6fb42432c72639a9f6e5f +size 387260 diff --git a/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/3032e5d9-b631-4933-a4e0-7d455ca4d844_content_list.json b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/3032e5d9-b631-4933-a4e0-7d455ca4d844_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..54d86d3d7391e5fd2ca7274f7adc9b1e1aaf8812 --- /dev/null +++ b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/3032e5d9-b631-4933-a4e0-7d455ca4d844_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34c9d2aca7af77396561c2a9d518205b1a46dacc80acbfe8ad57b7144fa5c691 +size 88800 diff --git a/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/3032e5d9-b631-4933-a4e0-7d455ca4d844_model.json b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/3032e5d9-b631-4933-a4e0-7d455ca4d844_model.json new file mode 100644 index 0000000000000000000000000000000000000000..39510c47c94cfcc4013ade7fed0f10889b5c8d43 --- /dev/null +++ b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/3032e5d9-b631-4933-a4e0-7d455ca4d844_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:296beef3486f4dc88bf781809f16d1f5e231f24bf78d46c70f8e820605326b4d +size 114172 diff --git a/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/3032e5d9-b631-4933-a4e0-7d455ca4d844_origin.pdf b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/3032e5d9-b631-4933-a4e0-7d455ca4d844_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ca69b32be5edddbe225599a732eb9af3696cfad5 --- /dev/null +++ b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/3032e5d9-b631-4933-a4e0-7d455ca4d844_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:086b0b3ae08639d14c9545d16fdf71f6e1494a1a09f1820c9961b659bf4b0106 +size 13671773 diff --git a/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/full.md b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4959fd7689c6bbaac5ed6b53c79cc610cb60ff14 --- /dev/null +++ b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/full.md @@ -0,0 +1,315 @@ +# ReME: A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation + +Xiwei Xuan, Ziquan Deng, and Kwan-Liu Ma +University of California, Davis + +{xwxxuan, ziqdeng, klma}@ucdavis.edu + +# Abstract + +Training-free open-vocabulary semantic segmentation (OVS) aims to segment images given a set of arbitrary textual categories without costly model fine-tuning. Existing solutions often explore attention mechanisms of pre-trained models, such as CLIP, or generate synthetic data and design complex retrieval processes to perform OVS. However, their performance is limited by the capability of reliant models or the suboptimal quality of reference sets. In this work, we investigate the largely overlooked data quality problem for this challenging dense scene understanding task, and identify that a high-quality reference set can significantly benefit training-free OVS. With this observation, we introduce a data-quality-oriented framework, comprising a data pipeline to construct a reference set with well-paired segment-text embeddings and a simple similarity-based retrieval to unveil the essential effect of data. Remarkably, extensive evaluations on ten benchmark datasets demonstrate that our method outperforms all existing training-free OVS approaches, highlighting the importance of data-centric design for advancing OVS without training. Our code is available here. + +# 1. Introduction + +Open-vocabulary semantic segmentation (OVS) aims to segment images according to a set of arbitrary textual categories. While conventional approaches rely on training or fine-tuning large models with segment-text [10, 13, 19, 33, 36, 40, 45, 70] or image-text [8, 9, 59, 64, 67-69, 77] datasets, these methods incur significant annotation and computational costs. Recent advancements in vision-language models (VLMs), such as CLIP [47], have spurred interest in training-free OVS methods that leverage these pre-trained models without additional fine-tuning. + +Some approaches assume VLMs have sufficient classification capabilities to recognize semantics, and attempt to enhance their pixel-level localization by modifying the at + +![](images/686f2930eda0f67fb008dfe7ee31cd5354f523097ca025e9af9dcb1e335a91c8.jpg) +(a) + +![](images/07a0b324301292e32d911715622f0f651bde8013dbaef855b3b761adac4c1f6f.jpg) +(b) +Figure 1. (a) Qualitatively compare ReME with SCLIP [58] (a representative CLIP-attention-based method) on an example from ADE20K [81] with 150 classes. Both methods have no mask post-processing. ReME produces higher-quality masks with a simple segmentation algorithm, yet SCLIP generates noisy masks paired with many irrelevant classes, such as "curtain" and "microwave", etc. (b) Performance comparison between SCLIP and retrieval-based OVS using different reference sets. GT: ground-truth segment-label of COCO Stuff [5]; Syn: synthetic reference set from FreeDA [3]; Ours: our reference set from real images with quality refinement. + +tention mechanisms [29, 30, 53, 56, 58, 60, 82]. However, VLMs are trained under weak image-text supervision [32], and merely tweaking their attention modules post-hoc cannot break through the inherent ceiling imposed by coarse training signals [21]. As a result, such approaches often struggle with low-quality segmentation masks and noisy labels. As shown by an example in Fig. 1 (a), methods relying on CLIP attention often fail to even match the segmentation quality of an algorithm-based segmenter [17], which, despite being class-agnostic, can propose perceptually meaningful image regions. (Refer to Fig. 5 for more examples). + +Confronted with such limitations of VLMs, the retrieval paradigm offers a key inspiration: correcting the vulnerabilities of VLMs by reference substances in an external knowl + +edge base. We investigate this from the data-centric perspective – assuming there exists a high-quality reference set for OVS, can we unlock the potential of retrieval for better results? Specifically, we experiment with ground truth (GT) segment-text pairs from COCO Stuff [5], viewing it as a “high-quality” reference set. Using a class-agnostic segmentation algorithm [17] for test images, we apply a simple similarity-based strategy (refer to Sec. 3.3) for retrieving and aggregating mask labels. No models or complex retrieving algorithms are involved to unveil the fundamental data capabilities. As shown in Fig. 1 (b), retrieval from GT annotations consistently surpasses SCLIP [58], which is a well-performed method using CLIP attention. This observation indicates the significant potential of data for enhancing OVS with retrieval. + +However, in the absence of labor-intensive GT annotations, despite the anticipated performance gain, existing retrieval-based methods remain modest compared to attention-based approaches. To fill the data gap, they often resort to diffusion models (DMs) [50] to generate synthetic images, utilizing text-corresponding attention masks to construct reference sets [2, 3, 24, 62]. However, synthetic data often lacks the realism and richness of real images, making them a suboptimal retrieval resource [15, 55, 74]. Fig. 1 (b) compares retrieval results from the COCO Stuff GT and a state-of-the-art synthetic reference set from FreeDA [3], constructed from COCO captions. The comparison indicates that the synthetic reference set falls significantly short of its GT counterpart, highlighting its limitations compared to real images. Some methods employ a quick fix of data bottleneck by designing sophisticated retrieval strategies to tolerate suboptimal data [62], which is less preferred as the central issue remains unaddressed [73]. + +In this work, we aim to construct a well-aligned, rich, and contextually relevant set of segment-text pairs from real images, investigating the capabilities of such data to benefit training-free OVS. Specifically, we introduce ReME, a data-centric framework that Refines Multi-modal Embeddings for retrieval-based, training-free OVS. + +ReME harnesses the intrinsic capabilities of VLMs with the aggregation effect - while they may not yield satisfactory results for individual instances, VLMs demonstrate a certain degree of effectiveness when analyzing data at a collective level [12, 42, 46, 51, 52, 80, 83]. Using only images as input, the data pipeline of ReME starts by constructing a semantically rich base set with VLM-paired image segments and textual labels. With the observation of superior discriminativeness in same-modal data features, ReME leverages their collective patterns to recalibrate cross-modal misalignments introduced by the VLM, producing a reference set with enhanced quality. In the inference, we apply a basic class-agnostic segmentation algorithm to test images, then align and aggregate the proposed + +masks into final predictions by referring to our reference set with a simple similarity-based strategy. + +As shown in Fig. 1 (b), compared to using GT annotations for reference, our reference set indicates even better OVS capacity. Specifically, governed by our data quality consideration in semantic richness and alignment correctness, ReME brings significantly further performance gains for challenging benchmarks, such as PC-459 [39], and domain-specific scenarios, such as Cityscapes [11]. + +Extensive experiments on ten OVS benchmarks demonstrate the superior performance of ReME over existing training-free methods. Overall, our main contributions are summarized as follows: + +- We introduce ReME, a data-centric framework for training-free OVS by constructing high-quality segment-text data from real images, with a simple retrieval process to perform OVS without any model fine-tuning. +- We identify the critical yet underexplored role of data quality, and provide a data pipeline for refining multimodal data collectively from an intra-modal perspective, which unveils the data potential to regularize or complement pre-trained models. +- Extensive experiments on ten benchmark datasets demonstrate the consistent performance gains of ReME over 14 training-free OVS baselines. + +# 2. Related Work + +Training-Free OVS. Many training-free OVS methods analyze attention mechanisms [57, 71, 72] of VLMs or diffusion models to produce segmentation results [4, 29, 37, 49, 53, 60, 65, 82]. However, their reliant models trained without pixel-level supervision may produce inferior results, constraining the upper bound of these approaches. To improve the situation, methods such as CaR [56] leverage multiple CLIP models to iteratively refine segmentation results, which are less practical due to significant inference costs. Another line of methods [2, 3, 20, 24, 54, 62] constructs a reference set with extensive segment-text pairs, which serves as reference knowledge to support better alignment between test segments and text. Most of such methods either use synthetic data [2, 3, 24, 62] that falls short of diversity and realism [15, 55], or opt for sophisticated retrieval strategy or shortcut like adding GT classes to bypass the data quality issues [41, 62]. To fill the gap, our framework investigates the importance of data for this challenging task. Multi-Modal Datasets Curation. High-quality datasets have driven remarkable progress in multi-modal models [8, 16, 38, 44, 61, 66, 75]. Recent techniques of image-text data curation typically involve two branches: data filtering [6, 18] that removes noisy data, and data improvement [28, 83] that refines multi-modal alignment. Despite training better VLMs is often the objective of data curation, these models themselves are also used for curating better + +![](images/dfc08ae7ac3b37272cca65235eaa346e3f06b3991da1123d3ce5e3a6ce3d30f9.jpg) +Figure 2. The data pipeline of ReME to construct a high-quality reference set. The pipeline includes two phases: the initial pairing that produces pre-matched segment-text pairs using images as input; and the data enhancing that performs group-based filtering according to more discriminative intra-modal similarity among segment embeddings and semantic enriching with similar labels. + +data. For instance, CLIP is frequently used for data filtering with CLIP score (i.e., similarity between visual and text features) [46, 51, 52], and some methods use BLIP-2 to reduce semantic noise [42, 83]. As both data filtering and data improvement address essential aspects of data quality, we consider both of them in our approach for more comprehensive data enhancement. In addition, multi-modality data curation in fine-grain such as segment-text pairs remains underexplored, requiring further research to support more precise, context-aware applications. Beyond OVS, our proposed data cleaning methodology has the potential for broader applicability across other contexts. + +Pre-Training Models. Pre-training models often have high reusability and adaptability, which can be easily reused in various approaches without further fine-tuning [24, 37, 54, 76, 78]. For instance, Vision-Language Models (VLMs) like CLIP [47] and ALIGN [23] link text with images for zero-shot image classification. Multimodal Large Language Models (MLLMs), such as LLaVA [35], BLIP-2 [31], and GPT-Vision [1], support diverse tasks such as captioning, reasoning, and visual chat. Vision models like DINO [7] and DINOv2 [43] provide universal visual feature representations. Despite their success, such models are not almighty solutions, having various limitations for specified tasks [27]. In this work, we study how to effectively harness the value of data to compensate for the shortcomings of such models in a training-free manner. + +# 3. Our Approach + +The goal of an OVS method is to segment an input image given an arbitrary set of textual labels. Our training-free framework ReME consists of a data pipeline with two key + +phases, initial pairing and data enhancing, to construct a high-quality reference set of segment-text pairs (Fig. 2); followed by a simple similarity based retrieval process to unveil the performance gains enabled by our data (Fig. 4). + +# 3.1. Initial Pairing + +Our data pipeline starts by constructing a diverse base set with segment-text pairs, where we aim to ensure broad coverage of various objects and contextual elements. + +As shown in Fig. 2, for each input image, a segmenter generates class-agnostic segment masks, and an image description generator produces semantically rich descriptions. Following [44, 66], we extract noun phrases (i.e., nouns with descriptive modifiers such as "a lovely white rabbit") from the descriptions to produce candidate textual labels, which reserve richer semantics than pure nouns. Then, we pair the segments and labels using a capable VLM. Despite the noises, CLIP still provides a certain degree of correctness for segment labeling, as evidenced by CLIP-based OVS methods [20, 29, 53, 56, 82]. We thus leverage CLIP to perform initial pairing. Given an image, we extract CLIP embeddings of its segments and noun phrases, $S \in \mathbb{R}^{m \times d}$ and $L \in \mathbb{R}^{n \times d}$ , respectively, where $m$ , $n$ is the number of segments and labels, and $d$ is the embedding dimension. Note that all embeddings across the paper are L2 normalized so that $X \cdot Y^T$ is equivalent to their cosine similarity $\langle X, Y \rangle$ . We then compute $sim = S \cdot L^T \in \mathbb{R}^{m \times n}$ to pair each label with its closest segment, where $sim_{ij}$ is the similarity between $S_i$ and $L_j$ . The top-matching $S_{i*}$ of $L_j$ is identified by: $i^* = \arg \max_i sim_{ij}$ . The unpaired segments are dropped to reduce the redundancy in class-agnostic image segments. + +Despite including diverse segments paired with noun + +![](images/44948002b09319976f7f7fc7fbfbc39d2631cf3781b349f067f7e4f80dcf67a8.jpg) +Figure 3. The superiority of intra-modality over cross-modality for data issue detection. The plot provides the UMAP projection of segment embeddings in the base set labeled as "dog", colored by cross-modal similarity scores (CLIP scores). Blue boxes highlight misalignments detected by our filtering; orange boxes are those detected by low CLIP scores, which remove correct pairings while leaving many misalignments unaddressed. + +phrase labels, the base set still has unaddressed quality concerns, which is our major focus in the following phase. + +# 3.2. Data Enhancing + +As shown in Fig. 2, through initial pairing, we obtain a base set with paired segments and labels. Various factors negatively affect its quality: (1) irrelevant textual labels caused by object hallucinations of the description generator, (2) meaningless or partial segments from over-segmentation, and (3) segment-text pairing errors. These challenges are essentially introduced by the ambiguity in cross-modality. We opt for navigating them with intramodality operations—which we find to be more discriminative to effectively clean and enrich such data. + +Unlike the common practice of dropping segment-label pairs with low cross-modal similarity scores (i.e., CLIP scores) [46, 51, 52], we reexamine data features and test whether CLIP scores can favorably clean noisy data. As shown in Fig. 3, our observation brings two key insights: (1) CLIP scores tend to mistakenly remove correct pairs while leaving many misalignments unaddressed; (2) Visual features, i.e., embeddings of image segments, have favorable discriminative power for data issue detection. (Please refer to Sec. 4.3 for extensive discussions.) + +Group-Based Filtering. Guided by our observation, in this phase, we detect misalignments by leveraging the discriminativeness of intra-modal features. Specifically, we aim to leverage the visual features of segments to identify outliers in their own modality. First, for our labels in the format of noun phrases, we view those with the same "root" noun as an identical label. Next, we group each single segment-label pair according to identical labels. For example, segments labeled by "a small dog" or "an adorable dog" belong to the same group denoted by the root "dog." Note that we reserve all descriptive modifiers in the phrases to keep se + +mantic richness; "roots" are just used for grouping. + +This is an important step for detecting data misalignments. Since each group of segments corresponds to an identical label, their visual features should be inherently consistent, and misaligned data would be automatically highlighted as outliers. As shown in Fig. 3, the majority of segments marked by $①$ accurately depict dogs, while others are misalignments located as outliers in the projection, indicating their distinctive features. + +In detail, we define the segment center $S_{center}$ of each group, computed by the median of all visual features. $S_{center}$ denotes the representative features for correctly labeled segments. We then compute the intra-modal similarity between each segment in the current group and $S_{center}$ , indicating the possibility of a segment being mistakenly labeled. $\delta_{filter}$ percent of segments with the lowest scores are considered mistakenly paired, where we remove the corresponding noun phrases from their labels. By performing this filtering guided by intra-modality feature similarity, we effectively reduce misalignment in the base set. + +Semantic Enriching. By clearing outliers in each group, we have significantly mitigated segment-text misalignments. However, data enhancement remains incomplete in terms of semantic diversity - While statistical analysis of the entire label corpus of our base set reveals the textual modality includes different words for the same concept (e.g., "bike" and "bicycle"), such alternatives are missing in individual segments. This results in a lack of semantic richness, as we expect our method to recognize concepts in various representations. Consistently, we design an intra-modality-based approach to diversify our labels. As shown in Fig. 2, we collect embeddings of identical labels, i.e., the "root" noun of our noun phrase labels, denoted as $\{L_1,\ldots ,L_n\}$ . Then, we compute their pair-wise cosine similarity $\langle L_i,l_j\rangle$ , and identify $k$ top-similar pairs to be treated as synonyms. Next, label semantics are enriched by adding these synonyms, so, if a segment has a label "a small cat", we add "a small kitten" to become its label as well. Those synonym-enriched noun phrases are our finalized labels. This process further diversifies our textual descriptions by leveraging the global semantic richness present within the data. Note that, an LLM could produce even better semantic enriching, while we opt for a lightweight-yet-effective solution to avoid the additional overhead of large models. + +# 3.3. Similarity-Based Retrieval + +Following Sec. 3.2, we have constructed a reference set with well-paired segments and labels and stored their embeddings. At inference time, given a test image and a set of target classes, our goal is to assign pixel-level labels to the test image by referring to the reference set. Inspired by Tip-Adapter [79], we perform a simple retrieval strategy + +![](images/8f84729bd4ca375f32d0494acaa1dfa800f024e2216920d13a7b227e4830fc6f.jpg) +Figure 4. The process of similarity-based retrieval from the reference set given a test image and classes. Note that the same modality embeddings (i.e., reference & test segments; reference & test labels) are encoded by the same feature encoder, respectively. + +grounded in feature similarity, where labels of test segments are estimated by their closest matches in the reference set. + +As shown in Fig. 4, the reference set includes embeddings of $m$ segments $\mathbf{S}_{\mathrm{ref}} \in \mathbb{R}^{m \times d_1}$ and $n$ labels $\mathbf{L}_{\mathrm{ref}} \in \mathbb{R}^{n \times d_2}$ . The label assignments are encoded by a binary matrix $\mathbf{O}_{\mathrm{ref}} \in \mathbb{R}^{m \times n}$ , with an entry 0 indicating the corresponding label is absent in the segment and 1 otherwise. + +Given a test image, we segment it into $k$ class-agnostic masks, denoted by $M_{\mathrm{seg}} \in \mathbb{R}^{k \times h \times w}$ , where $h \times w$ is the image size. We extract the segment and test class embeddings $S_{\mathrm{test}} \in \mathbb{R}^{k \times d_1}$ and $L_{\mathrm{test}} \in \mathbb{R}^{c \times d_2}$ , using the same encoders applied to the reference set, where $d_1$ and $d_2$ are the embedding dimensions of the image and text modalities, respectively. The affinity between test segments and reference set labels is computed as: + +$$ +A _ {1} = \operatorname {S o f t m a x} \left(\boldsymbol {S} _ {\text {t e s t}} \cdot \boldsymbol {S} _ {\text {r e f}} ^ {T}\right) \cdot \boldsymbol {O} _ {\text {r e f}}, \tag {1} +$$ + +where $A_{1} \in \mathbb{R}^{k \times n}$ , and Softmax normalizes the affinities to formulate the probability distribution of labels for each test segment. Similarly, the affinity matrix between the reference labels and test classes is represented as: + +$$ +A _ {2} = \operatorname {S o f t m a x} \left(\boldsymbol {L} _ {\text {r e f}} \cdot \boldsymbol {L} _ {\text {t e s t}} ^ {T}\right), \tag {2} +$$ + +where $A_{2} \in \mathbb{R}^{n \times c}$ . Ensembling the affinities, for $k$ test segments, the predicted logit is defined as: $P_{\mathrm{seg}} = A_{1} \cdot A_{2}, P_{\mathrm{seg}} \in \mathbb{R}^{k \times c}$ . Then, we calculate pixel-wise label probabilities $P_{\mathrm{test}} \in \mathbb{R}^{h \times w \times c}$ by aggregating ensembled affinities according to the segment masks $M_{seg}$ , where the probability of class $j$ at coordinates $(x,y)$ is: + +$$ +\boldsymbol {P} _ {\text {t e s t}} ^ {(x, y, j)} = \sum_ {i = 1} ^ {k} \boldsymbol {P} _ {\text {s e g}} ^ {(i, j)} \cdot \boldsymbol {M} _ {\text {s e g}} ^ {(i, x, y)}. \tag {3} +$$ + +Lastly, the predicted mask $\hat{l} \in \mathbb{R}^{h \times w}$ is determined by: $\hat{l}^{(x,y)} = \arg \max_{j} P_{\mathrm{test}}^{(x,y,j)}$ , where $\hat{l}^{(x,y)} \in [0, c - 1]$ indicates class predictions, which conclude our retrieval phase. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets for standard benchmarks. We conducted experiments on 10 widely-used OVS benchmarks to evaluate ReME, including validation splits of Pascal VOC (VOC) [14], Pascal Context (PC) [39], COCO Object (Object) [34], COCO Stuff (Stuff) [5], Cityscapes (City) [11], and ADE20K [81]. Specifically, VOC has 20 object classes (VOC-20). PC involves PC-459 with 459 classes and PC-59 with 59 frequent classes. Object and Stuff provide COCO2017 image annotations of 81 and 171 classes, respectively. City captures urban street scenes with 19 classes. ADE20K includes A-150 and A-849 with 150 and 847 classes. For VOC-20 and PC-59, we consider pixels not belonging to any class as "background", represented by VOC-21 and PC-60, respectively. We use the standard mean Intersection-over-Union (mIoU) to measure OVS performance. + +Implementation. We obtain textual descriptions of input images with LLaVA-1.5 [35]. For class-agnostic segmenter, we deploy Felzenszwalb's algorithm by default, which is a lightweight superpixel-based algorithm, following the same settings as [3]. The initial pairing is performed by CLIP [47] with ViT-B. In the data-enhancing and retrieval phases, we use CLIP-encoded text features, and employ DINOv2 with ViT-L as the default visual feature encoder. Hyperparameters $\delta_{filter}$ , $k_{sim}$ are both set to 30. To compare with SAM-involved OVS baselines, we optionally use SAM as the segmenter, where we prompt a $32 \times 32$ point grid to obtain masks. More implementation details are in supplementary. + +Our reference set construction requires only images as input—no GT captions, classes, or masks, and we default to using COCO-2017 [5, 34] images, which depict everyday scenes with objects in their natural contexts. + +# 4.2. Comparison to the state-of-the-arts + +Baselines. We compare ReME with 14 training-free OVS approaches, including ReCo [54], MaskCLIP [82], SCLIP [58], NACLIP [22], GEM [4], PnP [37], FreeDA [3], RIM [62], OVDiff [24], CLIPtrase [53], Diff-Segmenter [60], CaR [56], ProxyCLIP [29], and CorrCLIP [78]. Since post-processing techniques for mask refinements, such as DenseCRF [26], enhance OVS performance but introduce additional computational overhead, we indicate whether each method applies such refinements. Moreover, as SAM is widely regarded as a strong backbone for segment-related tasks, we conduct our comparison across both SAM-free and SAM-involved approaches. + +Comparison. Table 1 presents the quantitative comparison results, where we observe that ReME consistently outperforms training-free OVS baselines across ten evaluated benchmarks. Notably, ReME achieves superior performance in challenging benchmarks with complex scenarios + +![](images/65faf1ba848f8e348601fd04e1baa0319ed221d688d2f2b9a6dfd4cdd193d5de.jpg) +Figure 5. Qualitative results of ReME in comparison with other training-free OVS methods. Examples are from ADE20K [81] (w/ 150 classes) and COCO Stuff [5] (w/ 171 classes), respectively. SCLIP is based on CLIP attention; ProxyCLIP enhances CLIP attention with DINO features; FreeDA and ReME are retrieval-based methods, adopting the same superpixel-algorithm [17] for class-agnostic segmentation. We observe increasing quality of OVS results from left to right, with less noise in both masks and assigned labels. + +
MethodsPost-processingmIoU
VOC-20VOC-21CityPC-59PC-60ObjectStuffA-150PC-459A-847
Training-free Methods without SAM
GEM [4]X46.224.7-32.621.2-15.110.14.63.7
MaskCLIP [82]✓74.938.812.625.523.620.614.69.8--
ReCo [54]✓62.427.223.224.721.917.316.312.4--
SCLIP [58]✓83.561.734.136.131.532.123.917.89.36.1
CaR [56]✓91.467.615.139.530.536.611.217.711.55.0
NACLIP [22]✓83.064.138.338.435.036.225.719.19.06.5
CLIPtrase [53]✓81.253.021.134.930.839.624.117.09.95.9
PnP [37]✓79.151.319.331.028.036.217.914.25.54.2
FreeDA [3]✓87.955.436.743.538.337.428.822.410.25.3
ProxyCLIP [29]X83.260.640.137.734.539.225.622.611.26.7
DiffSegmenter [60]✓71.460.1-27.525.137.9----
OVDiff [24]✓80.968.423.432.931.236.220.314.112.06.6
ReME (Ours)X92.379.650.444.941.645.533.126.114.18.4
ReME (Ours - VOC)X84.775.043.940.938.740.822.625.212.88.3
ReME (Ours - ADE)X84.372.342.144.039.735.827.026.013.28.6
Training-free Methods with SAM
RIM [62]X77.8--34.3-44.5----
CaR w/ SAM [56]X-70.216.940.531.137.612.417.911.85.7
CLIPtrase w/ SAM [53]X82.357.1-36.432.044.224.817.210.66.0
ProxyCLIP w/ SAM [29]X80.459.337.037.033.635.425.019.16.94.8
CorrCLIP [78]X91.674.147.745.540.343.630.6---
ReME w/ SAM (Ours)X93.282.259.053.144.648.233.328.215.88.8
+ +Table 1. Comparison to state-of-the-art training-free OVS approaches. The best results are bolded, with the second-best underlined. We also analyze data robustness by varying the image resources from the default COCO-2017 to VOC and ADE, respectively (w/o SAM version), where leading performances over SAM-free baselines are bolded. + +and large numbers of classes, including A-150, PC-459, and A-847. On Cityscapes, an increase of $8.4\mathrm{mIoU}$ points (11.3 mIoU increase for SAM-involved comparison) further indicates that our approach excels in domain-specific scenarios. We also provide qualitative results of ReME in Fig. 5. + +Data robustness. To evaluate the data robustness of ReME, we vary image resources of our reference set, replacing the default COCO-2017 images with those from VOC and ADE. As shown in Table 1, SAM-free ReME maintains + +strong performance over baselines. Across 10 benchmarks, ReME achieves the highest mIoU in 6 benchmarks with VOC and 7 with ADE. These results highlight the effectiveness and flexibility of our approach in diverse settings. + +# 4.3. Analysis, Ablation, and Discussion + +Hyperparameters. Our data-enhancement process involves two key hyperparameters: the drop ratio per group $(\delta_{filter})$ in group-based filtering (i), and the number of top-similar label pairs $(k_{sim})$ in semantic enriching (ii). To de + +![](images/40f720aa3b1c92381df03a742d3bcfa295d19b2c84df622d5a116156cb53b9d4.jpg) +Figure 6. Hyperparameter analysis. $\delta_{filter}$ is the drop ratio in group-based filtering; $k_{sim}$ is the number of top-similar pairs in semantic enriching. + +
ComponentsmIoU
VOC-20PC-59ObjectA-150
Base set (no enhancement)70.0335.4239.3822.03
w/ (i) Group-based filtering91.1040.6642.4824.09
w/ (ii) Semantic enriching79.5036.6939.9223.41
w/ (i) and (ii)92.3444.8945.5026.13
+ +termine their values, we perform a grid search on $1k$ $(1\%)$ randomly sampled images from the training split of COCO Stuff (with no overlap with evaluation data), as shown in Fig. 6. Initially, increasing both parameters improves performance, as more misaligned data is filtered out and label diversity is enhanced. However, beyond a certain point, further increases cause a decline in performance: a high drop ratio removes correct pairings and reduces data diversity, while an excessive $k_{sim}$ introduces noise by including less-similar labels. We use 30 as the default value for both parameters in our implementation. + +Data enhancing component analysis. We analyze the effects of our data enhancement components: (i) group-based filtering and (ii) semantic enriching. As shown in Table 2, the base set without enhancement yields relatively low performance, highlighting the critical role of data quality. We also observe higher individual effectiveness of (i), while enriching labels alone provides modest gains due to data noises. Notably, the performance improves significantly when both are integrated, demonstrating the complementary benefits of our data enhancement components. + +Analysis of different data-cleaning approaches. We apply group-based filtering to remove $\delta_{filter}\%$ of outliers from each label-identified group (refer to Sec. 3.2). This experiment examines the impact of different data-cleaning strategies by comparing three methods: (a) the common approach of dropping segment-label pairs with the globally lowest cross-modal similarity scores (CLIP scores); (b) a variant that also uses CLIP scores, removing segment-label pairs within each constructed groups; (c) our group-based filtering based on intra-modal feature similarity. By only changing this module, we evaluate OVS performance as + +Table 2. Impact of data enhancement components. + +
Data cleaningmIoU
VOC-20PC-59ObjectA-150
Global CLIP score(a)79.3439.7940.7921.18
Group-based CLIP score(b)80.0541.8441.8122.79
Ours(c)92.3444.8945.5026.13
+ +Table 3. Analysis of different data filtering approaches. + +
Feature encodermIoU
VOC-20PC-59ObjectA-150
CLIPB91.6136.5139.8224.72
CLIPL92.1637.6340.5225.56
DINOv2B91.7243.6544.7125.29
DINOv2L92.3444.8945.5026.13
+ +Table 4. Analysis of visual encoder variations. + +![](images/409794948022dd8face597887d830b948ef28884c978e398e7d4f3584ebd7a42.jpg) +Figure 7. Image descriptions from different resources. Red text highlights concepts uniquely present in the LLaVA description. + +shown in Table 3, where we keep the drop ratio fixed at $30\%$ . The results indicate that (a) and (b) perform much worse than (c), suggesting that filtering based on CLIP scores is less effective, likely due to its weaker alignment with actual pairing quality. + +Analysis of different visual encoders. We investigate the impact of different visual feature encoders, including CLIP and DINOv2 with ViT-B and ViT-L architectures, denoted by $\mathrm{CLIP}_B$ , $\mathrm{CLIP}_L$ , $\mathrm{DINOv2}_B$ , and $\mathrm{DINOv2}_L$ , respectively. As shown in Table 4, upgrading from the base to large variants within the same encoder family results in slight performance improvements. When comparing CLIP and DINOv2 at the same scale, DINOv2 achieves better performance as expected, given its specialization in visual representation learning. Notably, by using CLIP alone and removing DINOv2 from our framework, we still achieve performance that is comparable to or better than baselines using similar backbones. This suggests that our approach is robust to variations in visual encoders. + +Analysis of the description generator. Providing semantically rich image descriptions is important for our data construction. However, not all descriptions are equally informative. Fig. 7 presents an image from the COCO-2017 [5] + +
Description generatormIoU
VOC-20PC-59ObjectA-150
LLaVA [35]92.3444.8945.5026.13
BLIP-2 [31]89.4140.8537.6424.31
GT Caption89.0240.1537.7724.09
+ +Table 5. Ablation study of the image description generator. + +
SegmentermIoU
VOC-20PC-59ObjectA-150
Superpixel [17]92.3444.8945.5026.13
SAM [25]93.1553.1048.2128.21
SAM2 [48]93.1852.0448.4028.21
+ +dataset alongside its descriptions from three resources: (1) GT captions, (2) BLIP-2 [31], and (3) LLaVA [35], where we highlight the detected visual concepts. We can observe common concepts presented in all three, "sandwich", "fries", and "plate". However, concepts such as "dining table", "burger", "knife", "bottle", and "person" are missing from both (1) and (2). To further study the impact of using BLIP-2 and LLaVA as description generators, we conduct a quantitative evaluation. As shown in Table 5, LLaVA demonstrates superior performance, achieving significantly higher mIoU compared to BLIP-2, where BLIP-2 captions yield only marginal improvements over GT captions. These results highlight LLaVA's ability to generate richer, more detailed descriptions and quantitatively confirm its effectiveness in enhancing training-free OVS with retrieval. + +Analysis of the segmenter. We also investigate the impact of different segmenters in Table 6 by replacing our default superpixel algorithm-based segmenter with advanced segmentation models, SAM [25] and SAM2 [48]. While a more advanced segmentation model leads to slightly better performance, our data-quality-enhancement pipeline enables even a simple segmentation algorithm to achieve strong results compared to baseline methods. + +Analysis of data effectiveness. We evaluate data effectiveness by comparing ReME with FreeDA [3], a representative retrieval-based method that also utilizes COCO-2017 dataset [5, 34]. Unlike our approach, which takes images as input and generates textual descriptions, FreeDA uses GT captions to generate synthetic images for constructing its reference set. As shown in Table 7, despite FreeDA incorporating five times more input samples, its reference set contains fewer unique labels than ReME, due to the limited semantic diversity of GT captions (refer to Fig. 7). Our data-enhancement strategy improves data quality while reducing dataset size, resulting in a reference set with over a million fewer segment-text pairs than FreeDA. Using this considerably smaller reference set, ReME achieves supe + +Table 6. Ablation study of the segmenter. + +
MethodsInput dataReference set
FormatSize#labels#pairsStorage
OursImages118k41k1,023k4GB
FreeDA [3]Captions591k19k2,167k17GB
+ +Table 7. Comparing data effectiveness with FreeDA, a retrieval-based OVS baseline that uses a synthetic reference set. + +
MethodsPC-59A-150PC-459A-847
ProxyCLIP [29]0.170.270.210.32
SCLIP [58]0.200.340.570.53
CaR [56]5.0614.6733.4675.09
FreeDA [3]0.890.894.895.01
Ours0.290.340.310.34
+ +Table 8. Inference time comparison (seconds/image). + +rior performance even with a much simpler retrieval mechanism, highlighting the effectiveness of high-quality, semantically rich data over quantity. + +Analysis of inference time. Inference time for training-free OVS varies due to factors such as the number of classes, image resolution, and inference strategies. To evaluate this, we compare our method with four representative approaches: (1) ProxyCLIP [29], which augments CLIP attention with DINO features, (2) SCLIP [58], which exploits CLIP self-attention followed by PAMR post-processing [63], (3) CaR [56], which iteratively queries two CLIP models for mask proposal and classification, and (4) FreeDA [3], a retrieval-based method. All experiments are conducted on two NVIDIA 4090 GPUs, measuring total inference time per dataset to compute the average. As shown in Table 8, (3) is the most time-consuming due to its iterative processing. Our method surpasses the retrieval-based baseline (4), benefiting from a smaller reference set and streamlined retrieval design, while remaining competitive with (1) and (2), which do not rely on reference sets. + +# 5. Conclusion + +In this work, we introduce ReME, a data-centric framework for training-free OVS by refining multi-modal embeddings. We observe the overlooked challenge of data quality, and demonstrate its critical impact on this dense scene understanding task. Following our data pipeline, we produce a reference set with well-aligned, rich, and contextually relevant segment-text pairs. Extensive experimental results highlight that enhancing data quality can be more beneficial than relying on complex retrieval mechanisms or model-specific adaptations. We hope this work can inspire future research for exploring data-centric strategies to further improve training-free OVS. + +Acknowledgements This research is supported in part by + +the National Institute of Biomedical Imaging and Bioengineering of NIH under Grant No. P41-EB032840. + +# 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. 3 +[2] Luca Barsellotti, Roberto Amoroso, Lorenzo Baraldi, and Rita Cucchiara. Fossil: Free open-vocabulary semantic segmentation through synthetic references retrieval. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1464-1473, 2024. 2 +[3] 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. 1, 2, 5, 6, 8 +[4] 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. 2, 5, 6 +[5] 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. 1, 2, 5, 6, 7, 8 +[6] Liangliang Cao, Bowen Zhang, Chen Chen, Yinfei Yang, Xianzhi Du, Wencong Zhang, Zhiyun Lu, and Yantao Zheng. Less is more: Removing text-regions improves clip training efficiency and robustness. arXiv preprint arXiv:2305.05095, 2023. 2 +[7] 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 +[8] 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. 1, 2 +[9] Jun Chen, Deyao Zhu, Guocheng Qian, Bernard Ghanem, Zhicheng Yan, Chenchen Zhu, Fanyi Xiao, Sean Chang Culatana, and Mohamed Elhoseiny. Exploring open-vocabulary semantic segmentation from clip vision encoder distillation only. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 699-710, 2023. 1 +[10] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Catseg: Cost aggregation for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4113-4123, 2024. 1 + +[11] 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. 2, 5 +[12] Ziquan Deng, Xiwei Xuan, Kwan-Liu Ma, and Zhaodan Kong. A reliable framework for human-in-the-loop anomaly detection in time series. arXiv preprint arXiv:2405.03234, 2024. 2 +[13] 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. 1 +[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 5 +[15] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 2 +[16] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander T Toshev, and Vaishaal Shankar. Data filtering networks. In The Twelfth International Conference on Learning Representations, 2024. 2 +[17] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 1, 2, 6, 8 +[18] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36, 2024. 2 +[19] 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. 1 +[20] Zhongrui Gui, Shuyang Sun, Runjia Li, Jianhao Yuan, Zhaochong An, Karsten Roth, Ameya Prabhu, and Philip Torr. kNN-CLIP: Retrieval enables training-free segmentation on continually expanding large vocabularies. Transactions on Machine Learning Research, 2024. 2, 3 +[21] Yong Guo, David Stutz, and Bernt Schiele. Robustifying token attention for vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17557-17568, 2023. 1 +[22] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. arXiv preprint arXiv:2404.08181, 2024. 5, 6 +[23] 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. 3 + +[24] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for open-vocabulary segmentation. In European Conference on Computer Vision, pages 299-317. Springer, 2025. 2, 3, 5, 6 +[25] 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. 8 +[26] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 5 +[27] 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. 3 +[28] Zhengfeng Lai, Haotian Zhang, Bowen Zhang, Wentao Wu, Haoping Bai, Aleksei Timofeev, Xianzhi Du, Zhe Gan, Jiulong Shan, Chen-Nee Chuah, et al. Veclip: Improving clip training via visual-enriched captions. In European Conference on Computer Vision, pages 111–127. Springer, 2025. 2 +[29] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, 2024. 1, 2, 3, 5, 6, 8 +[30] Jingyao Li, Pengguang Chen, Shengju Qian, and Jiaya Jia. Tagclip: Improving discrimination ability of open-vocabulary semantic segmentation. arXiv preprint arXiv:2304.07547, 2023. 1 +[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 International conference on machine learning, pages 19730-19742. PMLR, 2023. 3, 8 +[32] 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. 1 +[33] 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. 1 +[34] 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, 8 +[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3, 5, 8 +[36] 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. 1 +[37] Jiayun Luo, Siddhesh Khandelwal, Leonid Sigal, and Boyang Li. Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4029-4040, 2024. 2, 3, 5, 6 +[38] Anas Mahmoud, Mostafa Elhoushi, Amro Abbas, Yu Yang, Newsha Ardalani, Hugh Leather, and Ari S Morcos. Sieve: Multimodal dataset pruning using image captioning models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22423-22432, 2024. 2 +[39] 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. 2, 5 +[40] 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. 1 +[41] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[42] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Swoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[43] 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, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 3 +[44] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 3 +[45] Yu Q. and et al. Convolutions die hard: Ovs with single frozen convolutional clip. In NeurIPS, 2023. 1 +[46] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6967-6977, 2023. 2, 3, 4 + +[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 International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 5 +[48] 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 Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 8 +[49] Pitchaporn Rewatbowornwong, Nattanat Chathee, Ekapol Chuangsuwanich, and Supasorn Suwajanakorn. Zero-guidance segmentation using zero segment labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1162-1172, 2023. 2 +[50] 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 +[51] 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. 2, 3, 4 +[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, 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 Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2, 3, 4 +[53] Tong Shao, Zhuotao Tian, Hang Zhao, and Jingyong Su. Explore the potential of clip for training-free open vocabulary semantic segmentation. arXiv preprint arXiv:2407.08268, 2024. 1, 2, 3, 5, 6 +[54] 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. 2, 3, 5, 6 +[55] 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 +[56] 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. 1, 2, 3, 5, 6, 8 +[57] 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. 2 +[58] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In European Conference on Computer Vision, pages 315–332. Springer, 2024. 1, 2, 5, 6, 8 +[59] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3635–3647, 2024. 1 +[60] 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. 1, 2, 5, 6 +[61] Xiaoqi Wang, Wenbin He, Xiwei Xuan, Clint Sebastian, Jorge Piazentin Ono, Xin Li, Sima Behpour, Thang Doan, Liang Gou, Han-Wei Shen, et al. Use: Universal segment embeddings for open-vocabulary image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4187–4196, 2024. 2 +[62] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3952-3963, 2024. 2, 5, 6 +[63] Anne S Wannenwetsch and Stefan Roth. Probabilistic pixel-adaptive refinement networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11642-11651, 2020. 8 +[64] Ji-Jia Wu, Andy Chia-Hao Chang, Chieh-Yu Chuang, Chun-Pei Chen, Yu-Lun Liu, Min-Hung Chen, Hou-Ning Hu, Yung-Yu Chuang, and Yen-Yu Lin. Image-text codecomposition for text-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26794-26803, 2024. 1 +[65] Monika Wysoczańska, Michael Ramamonjisoa, Tomasz Trzciński, and Oriane Simeoni. Clip-diy: Clip dense inference yields open-vocabulary semantic segmentation for-free. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1403-1413, 2024. 2 +[66] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4818–4829, 2024. 2, 3 +[67] Yun Xing, Jian Kang, Aoran Xiao, Jiahao Nie, Ling Shao, and Shijian Lu. Rewrite caption semantics: Bridging semantic gaps for language-supervised semantic segmentation. Advances in Neural Information Processing Systems, 36, 2024. 1 +[68] 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. +[69] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2935-2944, 2023. 1 +[70] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang 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. 1 +[71] Xiwei Xuan, Xiaoyu Zhang, Oh-Hyun Kwon, and Kwan-Liu Ma. VAC-CNN: A visual analytics system for comparative studies of deep convolutional neural networks. IEEE Transactions on Visualization and Computer Graphics, 28 (6):2326-2337, 2022. 2 +[72] Xiwei Xuan, Ziquan Deng, Hsuan-Tien Lin, Zhaodan Kong, and Kwan-Liu Ma. Suny: A visual interpretation framework for convolutional neural networks from a necessary and sufficient perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8371-8376, 2024. 2 +[73] Xiwei Xuan, Ziquan Deng, Hsuan-Tien Lin, and Kwan-Liu Ma. SLIM: Spuriousness mitigation with minimal human annotations. In European Conference on Computer Vision, pages 215-231. Springer, 2024. 2 +[74] Xiwei Xuan, Jorge Piazentin Ono, Liang Gou, Kwan-Liu Ma, and Liu Ren. AttributionScanner: A visual analytics system for model validation with metadata-free slice finding. IEEE Transactions on Visualization and Computer Graphics, 2025. 2 +[75] Xiwei Xuan, Xiaoqi Wang, Wenbin He, Jorge Piazentin Ono, Liang Gou, Kwan-Liu Ma, and Liu Ren. VISTA: A visual analytics framework to enhance foundation model-generated data labels. IEEE Transactions on Visualization and Computer Graphics, 2025. 2 +[76] Xinyuan Yan, Xiwei Xuan, Jorge Piazentin Ono, Jiajing Guo, Vikram Mohanty, Shekar Arvind Kumar, Liang Gou, Bei Wang, and Liu Ren. Vislix: An xai framework for validating vision models with slice discovery and analysis. In Computer Graphics Forum, page e70125. Wiley Online Library, 2025. 3 +[77] Muyang Yi, Quan Cui, Hao Wu, Cheng Yang, Osamu Yoshie, and Hongtao Lu. A simple framework for text-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7071-7080, 2023. 1 +[78] Dengke Zhang, Fagui Liu, and Quan Tang. Corrclip: Reconstructing correlations in clip with off-the-shelf foundation models for open-vocabulary semantic segmentation. arXiv preprint arXiv:2411.10086, 2024. 3, 5, 6 +[79] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European conference on computer vision, pages 493-510. Springer, 2022. 4 + +[80] Xiaoyu Zhang, Xiwei Xuan, Alden Dima, Thurston Sexton, and Kwan-Liu Ma. Labelvizier: Interactive validation and relabeling for technical text annotations. In 2023 IEEE 16th Pacific Visualization Symposium (PacificVis), pages 167-176. IEEE, 2023. 2 +[81] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 1, 5, 6 +[82] 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. 1, 2, 3, 5, 6 +[83] Deyao Zhu, Jun Chen, Kilichbek Haydarov, Xiaogian Shen, Wenxuan Zhang, and Mohamed Elhoseiny. Chatgpt asks, blip-2 answers: Automatic questioning towards enriched visual descriptions. arXiv preprint arXiv:2303.06594, 2023. 2, 3 \ No newline at end of file diff --git a/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/images.zip b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5aa2af4f716c68d8a7cc8b72f3c6e064385c6162 --- /dev/null +++ b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efe9847edc99b73e8c287fb7f999f96aeaf3e9e20d7a2c257143238130bb0ef9 +size 703449 diff --git a/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/layout.json b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6b2b9f941b8d4fd2d1dcadc586f2d3b351c22efa --- /dev/null +++ b/ICCV/2025/ReME_ A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:139d7b281be90d69338266d9d131ff3103f3e41b1dacfe01c959415520a208a3 +size 402929 diff --git a/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/bcba052f-fc64-4717-8010-aec1a7ee0c99_content_list.json b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/bcba052f-fc64-4717-8010-aec1a7ee0c99_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..886ad9d7d7c20f60dfe3fe969c1c246297edcbe6 --- /dev/null +++ b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/bcba052f-fc64-4717-8010-aec1a7ee0c99_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:498c05239c0cf482f0463171409193daf3ccd1c96e23e7f1905ff1cddd8366d3 +size 81454 diff --git a/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/bcba052f-fc64-4717-8010-aec1a7ee0c99_model.json b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/bcba052f-fc64-4717-8010-aec1a7ee0c99_model.json new file mode 100644 index 0000000000000000000000000000000000000000..83d58239f9667249e791aa815c653fd563d73e21 --- /dev/null +++ b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/bcba052f-fc64-4717-8010-aec1a7ee0c99_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba2f933a17cd1b3618ca7a7f47e0bf2beac45cefac50fed38927bb461835b792 +size 96724 diff --git a/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/bcba052f-fc64-4717-8010-aec1a7ee0c99_origin.pdf b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/bcba052f-fc64-4717-8010-aec1a7ee0c99_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5985bad763d780cd7df57985011841fd477b7fac --- /dev/null +++ b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/bcba052f-fc64-4717-8010-aec1a7ee0c99_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1933856a3224d438fa09dafb9c45b914e58ce5b4faa18beafc109c77c1e8d257 +size 5642366 diff --git a/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/full.md b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..34cf49ff9a80d7ed54f4e56cfc1eaef58fb76b8a --- /dev/null +++ b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/full.md @@ -0,0 +1,361 @@ +# ReMP-AD: Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection + +Hongchi Ma +Harbin Institute of Technology +1201020518@stu.hit.edu.cn + +Guanglei Yang* +Harbin Institute of Technology +yangguanglei@hit.edu.cn + +Debin Zhao +Harbin Institute of Technology +dbzhao@hit.edu.cn + +Yanli Ji Sun Yat-Sen University jiyanli82@gmail.com + +Wangmeng Zuo +Harbin Institute of Technology +wmzuo@hit.edu.cn + +# Abstract + +Industrial visual inspection is crucial for detecting defects in manufactured products, but it traditionally relies on human operators, leading to inefficiencies. Industrial Visual Anomaly Detection (IVAD) has emerged as a promising solution, with methods such as zero-shot, few-shot, and reconstruction-based techniques. However, zero-shot methods struggle with subtle anomalies, and reconstruction-based methods fail to capture fine-grained details. Few-shot methods, which use limited samples and prompts, offer a more efficient approach. Despite their promise, challenges remain in managing intra-class variation among references and in effectively extracting more representative anomaly features. This paper presents Retrieval-enhanced Multimodal Prompt Fusion Anomaly Detection (ReMP-AD), a framework that introduces Intra-Class Token Retrieval (ICTR) to reduce noise in the memory bank and Vision-Language Prior Fusion (VLPF) to guide the encoder in capturing more distinctive and relevant features of anomalies. Experiments on the VisA and MVtec-AD datasets demonstrate that ReMP-AD outperforms existing methods, achieving $97.8\% / 94.1\%$ performance in 4-shot anomaly segmentation and classification. Our approach also shows strong results on the PCB-Bank dataset, highlighting its effectiveness in few-shot industrial anomaly detection. Code is available at https://github.com/cshcma/ReMP-AD.git + +# 1. Introduction + +Within the manufacturing process, industrial visual inspection is the final but most crucial step in identifying defects in produced goods. This step is significant for cost savings by detecting faults as early as possible in the produc + +![](images/5048a3f8ea5194bc4c33e4cb41442ed456db2795468a090296b73bf17b46ed4c.jpg) +(a)Query +Figure 1. A comparison of the visualization results under the 2-shot setting using different references. (a) Original input with ground truth. (b) The first reference. (c) The second reference. (d) Visualization results. + +![](images/2715771110d92581a5d7a8334ab6069a130b9fc448e983f5c68d08fd513c2db6.jpg) +(b)ReferenceA + +![](images/610cab58327bf50cf3c8aa151456d6fcbf41e65da988ac74970bc1850378369d.jpg) +(c)ReferenceB + +![](images/f1345f7a07659537c934c0290a173bb851c74f7c14feff4df70a3ff41f494fe6.jpg) +(d)Result + +tion stages. Moreover, it prevents the dispatch of defective products to customers, thereby maintaining product quality and preserving the company's reputation. Traditionally, this task heavily relies on human operators, which incurs a high human resource burden and results in low production efficiency. To this end, Industrial Visual Anomaly Detection (IVAD) [2, 5, 6, 14, 15, 19, 24] has garnered increased attention in recent years. IVAD aims to detect and localize anomalies in industrial product images automatically using detection algorithms. Relevant works can be categorized into zero-shot methods [3, 17, 28], few-shot methods [5, 6, 10, 13, 19] and reconstruction methods [8, 15, 25]. Zero-shot methods mainly utilize pretrained vision-language models to exploit the correlation between visual and textual features for anomaly detection. However, due to the ambiguity of the textual modality, these methods may fail to detect subtle and irregular anomalies. + +Reconstruction-based methods measure reconstruction error between the reconstructed image and the input image for anomaly detection. Nonetheless, they may struggle to capture fine-grained details from hidden patterns, resulting in substantial reconstruction errors for normal samples. Few-shot methods leverage both textual prompts and a small number of normal reference samples, requiring less computational effort while achieving relatively good results, which is regarded as a more reasonable solution. + +In the context of few-shot anomaly detection (FSAD), existing models typically build the memory bank by treating all reference samples with equal confidence, regardless of inherent differences between them, such as background noise, lighting conditions, or variations in placement. This approach fails to account for the fact that certain references may be more relevant than others, leading to a lack of differentiation in how the model processes these samples. Consequently, the presence of irrelevant or noisy references, coupled with low-quality embedding features in the retrieval memory bank, can obscure the most relevant pathways for anomaly detection. As shown in Figure 1, selecting dissimilar references results in ambiguous and dispersed anomaly detection heatmaps. In contrast, using references closely matching the query produces precise anomaly localization and more accurate detection. This phenomenon demonstrates selecting reference images similar to the query significantly improves anomaly localization accuracy in few-shot visual anomaly detection, underscoring the importance of managing intra-class variability. Moreover, the task of anomaly detection is complicated by the indistinct and localized nature of anomalies themselves. Anomalies often present as subtle and context-specific deviations, making it difficult for the pre-trained visual encoder to generalize and capture the underlying patterns that differentiate abnormal from normal instances. + +To address the above challenges, we propose the Retrieval-enhanced Multi-modal Prompt Fusion Anomaly Detection (ReMP-AD), which includes Intra-Class Token Retrieval (ICTR) and Vision-Language Prior Fusion (VLPF). To effectively reduce the impact of background noise in the reference images and prevent being disturbed by low-relevance images, we propose an ICTR mechanism, which filters the relevant normal prototypical pattern of reference patches according to similarities, ensuring that only sufficiently similar reference patches are retrieved. Furthermore, to maximize the contribution of the most relevant examples to the memory bank, we assign the greatest weight to this reference when retrieving prototypical patterns. To use prior knowledge to guide the encoder to extract more representative features of anomalies, we propose VLPF to integrate visual and language prompts. In VLPF, we employ the potential abnormal region-level feature as visual prompts to enhance related patch-level features for fusing + +both vision and language knowledge. At the same time, a vision mask attention is integrated into the image encoder, which is termed the VLPF image encoder. This integration with visual prompts helps the model to focus on the likely abnormal regions while reducing attention to less probable areas. + +Our contributions can be summarized as follows: + +- This paper proposes a novel Intra-Class Token Retrieval (ICTR) mechanism for mitigating intra-class variation and reducing noise from the memory bank. In ICTR, Global Prototype Retrieval (GPR) is adapted to integrate the global correlation of features. Moreover, Class Token Balancing (CTB) addresses the insufficient consideration of variation among references. +- This paper proposes Vision-Language Prior Fusion (VLPF), employing both the vision principle and vision-textual principle for Industrial Visual Anomaly Detection. VLPF takes advantage of visual prompts to make the network extract representative anomaly features. +- Comprehensive experiments on VisA [30], MVtec-AD [1] datasets demonstrate ReMP-AD's superior performance on few-shot Industrial Image Anomaly Detection. Especially, in the VisA dataset, we achieve $97.8\% / 94.1\%$ in 4-shot anomaly segmentation and classification. For further validation, the proposed method is applied to a real-world printed circuit board dataset (PCB-Bank [25]) and achieves impressive performance. + +# 2. Related Work + +# 2.1. Visoin-prior anomaly detection + +Vision-prior methods leverage visual cues to identify anomalous regions, which can be categorized into reconstruction-based methods [8, 15, 25] and feature embedding-based methods [5, 6, 19]. Reconstruction-based approaches utilize the reconstruction error between the reconstructed image and the original to detect anomalies. Notable examples include MemAE [8], OCR-GAN [15], and GLAD [25], employing diverse reconstruction network architectures such as autoencoders, GANs, and diffusion models. In contrast, feature embedding methods focus on extracting normal features and storing them in a memory bank, where the average similarity distance between input embeddings and stored normal features is used to detect anomalies. For instance, SPADE [5] computes anomaly scores using the average Euclidean distance from the K nearest normal images. In comparison, Padim [6] derives scores based on the Gaussian distribution of patches at the same position, applying the Mahalanobis distance. PatchCore [19] enhances efficiency through a greedy coreset subsampling technique to reduce redundancy in the feature bank, utilizing patches from various positions and conduct + +ing a nearest neighbor search for anomaly localization. + +# 2.2. Language-prior anomaly detection + +Vision-language models (VLMs) excel at understanding the interrelationships between image and text data. A notable example is CLIP [18], which was pre-trained on 400 million image-text pairs from the internet. It demonstrates exceptional generalization across various downstream tasks [7, 10, 27], and exhibits strong zero-shot transfer capabilities [20, 26]. Additionally, CLIP's effectiveness in detecting out-of-distribution data [21, 23] has motivated us to use CLIP for language-prior IVAD. + +WinCLIP+ [10] first introduces CLIP into IVAD and proposes a language-prior paradigm by constructing two sets of prompts for normal and abnormal objects. APRILGAN [4] simplifies WinCLIP's network and further enhances its performance. Recent works [13, 28] modify the attention computation to better refine the local visual space. By introducing an innovative attention mechanism, these approaches enhance the model's focus on local features while maintaining the original structure. Moreover, VCP-CLIP [17] employ cross-attention to improve language-prior anomaly detection, whereas InCTRL [29] introduces in-context residual learning for anomaly classification. Furthermore, AdaCLIP [3] and PromptAD [13] leverage prompt learning to enhance language-prior IVAD performance by increasing the discrepancy between normal and abnormal textual features. AdaCLIP [3] leverages language prior detection results to integrate visual features, but it focuses only on the fusion of image-level features. Aside from CLIP, FiLo [9] enhances text prompts with positional information for Grounding DINO [16] to obtain preliminary bounding boxes. However, these methods overlook the critical role of visual context in fine-grained multimodal alignment or integrating visual prior knowledge only at the image level. This limitation hinders their effectiveness in addressing complex industrial anomaly segmentation scenarios. + +# 3. Methodology + +Anomaly detection primarily relies on two sources: data from the memory bank and vision-language information. ReMP-AD leverages a vision encoder to embed visual features and stores them in the memory bank. At the same time, it takes advantage of both language and vision information via a CLIP-based backbone. + +# 3.1. Overview of ReMP-AD + +The overall pipeline of the proposed ReMP-AD is illustrated in Fig. 2. This framework consists of two novel core components: Intra-Class Token Retrieval (ICTR) and Vision-Language Prior Fusion (VLPF). Specifically, ICTR effectively leverages reference features by extracting and + +comparing class-specific prototypes for each query patch. Additionally, ICTR addresses variability among references by incorporating global variation analysis. VLPF integrates language and vision prior through a unified approach. It utilizes vision-prior masks as vision prompts to guide the transformer extract anomaly features and employs connected area analysis to transform these masks into coherent regions generated by language prior, enabling a comprehensive extension from patch-level to region-level feature analysis. During training, VLPF is kept frozen, and only the MLP module is trained. + +# 3.2. Preliminaries + +CLIP-based anomaly detection leverages the semantic differences between visual features and the textual features corresponding to "normal" and "abnormal" concepts to recognize anomalies. Two textual prompts are formulated: "a photo of a normal [CLASS]" and "a photo of an abnormal [CLASS]", where [CLASS] represents the name of the target class. The corresponding text features are denoted as $F_{t} = [F_{t}^{n}, F_{t}^{a}]$ , where $F_{t} \in \mathbb{R}^{2 \times C}$ . Here, $F_{t}^{n}$ and $F_{t}^{a}$ represent the text features for normal and abnormal descriptions, respectively. Given an image, the visual feature $F_{v} \in \mathbb{R}^{1 \times C}$ is obtained via the vision encoder. Additionally, the intermediate layer feature map $F_{patch} \in \mathbb{R}^{H \times W \times C_{patch}}$ and the class token $F_{cls} \in \mathbb{R}^{1 \times C_{patch}}$ are extracted from the visual encoder. For computational convenience, the spatial dimensions of $F_{patch}$ are concatenated, resulting in $F_{patch} \in \mathbb{R}^{HW \times C_{patch}}$ . Here, $H$ and $W$ represent height and width, respectively. It is important to note that the dimension $C_{patch}$ , representing the feature dimension of the intermediate layer, differs from $C$ . The anomaly classification probability score $S_{vl}$ is computed as follows: + +$$ +S _ {v l} = \frac {e x p \langle F _ {v} , F _ {t} ^ {a} / \tau \rangle}{e x p \langle F _ {v} , F _ {t} ^ {n} / \tau \rangle + e x p \langle F _ {v} , F _ {t} ^ {a} / \tau \rangle} (1) +$$ + +Here, $\langle ,\rangle$ represents the similarity measure by computing the cosine similarity between vectors. And $\tau$ is the temperature hyperparameter. + +# 3.3. Intra-Class Token Retrieval + +To effectively handle noisy and diverse feature embeddings within the memory bank, the Intra-Class Token Retrieval (ICTR) algorithm is introduced to leverage comprehensive reference features, effectively reducing interference from noisy or misleading information. Its primary function is to identify the normal prototypical patterns within query patches, enabling accurate detection of abnormal regions. Additionally, intra-class variations are balanced by utilizing image-level class tokens. The ICTR algorithm comprises two sub-processes: Class Token Balancing (CTB) and Global Prototype Retrieval (GPR). These processes are designed to extract prototypes of query features at both the + +![](images/e78375e297c666fe27f4bbe460ae6e983ede63b70b82d384f264ff4144ca952e.jpg) +Figure 2. The computation of the anomaly map consists of two branches. The memory-guided branch capitalizes on the feature discrepancy between the query image and the reference image. Meanwhile, the language-guided branch leverages the paradigm that abnormal vision features are more similar to textual features from abnormal prompts than to those from normal prompts. The results from both branches are combined to produce the final anomaly map. This illustration of ReMP-AD showcases two novel modules: ICTR and VLPF. For the memory bank, ICTR outputs prototypes for comparison with query features. Additionally, intermediate patch features are input into the Vision-Language Prior Fusion (VLPF), where vision prior knowledge masks serve as visual prompts for the VLPF image encoder. Regions enriched with vision-language prompts are further computed, extending feature representation to the regional level. The MLP module is the only trainable part of the model. And the entire process is conducted over $P$ layers. + +image and patch levels. In a few-shot scenario, $K$ reference images are processed through a pre-trained visual encoder to generate patch-level features and class tokens. These features are then concatenated and stored within a memory bank, resulting in patch-level features $B_{patch} \in \mathbb{R}^{L \times C_{patch}}$ and class tokens $B_{cls} \in \mathbb{R}^{K \times C_{patch}}$ , where $L = K \times H \times W$ represents the total number of patch features. + +Class Token Balancing (CTB) To address intra-class variations and to enhance confidence among references across different classes, the Class Token Balancing (CTB) method is proposed. The confidence score for each reference is computed based on the similarity between the query class token and the reference tokens, as well as the similarities among the references. As illustrated in Algorithm $1$ , $B_{cls}^{i}$ is the class token of the $i$ -th reference, and $W_{cls}$ measures the similarity between the query and various references. It is assumed that a higher value of $W_{cls}^{i}$ reflects greater confidence in the $i$ -th reference. Additionally, the minimum similarity across other references is considered as the confidence bias $W_{ref}$ . To further enhance the differences in confidence scores, the squared $\ell_2$ norm is applied. + +Global Prototype Retrieval (GPR) For an input query feature $F_{patch}$ , the memory bank utilizes GPR to retrieve its corresponding prototype feature $F_{pro}$ . The procedure first involves calculating the similarity between the test and reference features to derive patch-level weights + +# Algorithm 1: CTB Algorithm + +Input: Input Features $F_{cls}$ , $B_{cls}$ + +Output: Output image-level weights $\hat{W}_{cls}$ + +1 Compute the similarity score $W_{cls}^{i}$ between the test image class token and the reference class tokens: + +$$ +W _ {c l s} ^ {i} \leftarrow \left\langle F _ {c l s}, B _ {c l s} ^ {i} \right\rangle ; +$$ + +2 Calculate the confidence score $W_{ref}^{i}$ for each reference: + +$$ +W _ {r e f} ^ {i} \leftarrow \min \left(\frac {B _ {c l s} ^ {i} \cdot B _ {c l s} ^ {T}}{\| B _ {c l s} ^ {i} \| _ {2} \cdot \| B _ {c l s} \| _ {2}}\right); +$$ + +3 Compute the final weights $\hat{W}_{cls}$ by evaluating the discrepancy between the similarity matrix and the reference confidence scores. Specifically, for each reference $i$ , calculate: + +4 $\hat{W}_{cls}^{i}\gets \left\| 1 + W_{cls}^{i} - W_{ref}^{i}\right\|_{2}^{2};$ + +5 $\hat{W}_{cls} \gets \text{concat}\left(\hat{W}_{cls}^{i}\right) (i = 1 \dots K)$ ; + +6 return $\hat{W}_{cls}$ + +$W_{patch} \in \mathbb{R}^{HW \times L}$ as follows: + +$$ +W _ {p a t c h} = \frac {F _ {p a t c h} \cdot B _ {p a t c h} ^ {T}}{\| F _ {p a t c h} \| _ {2} \cdot \| B _ {p a t c h} \| _ {2}} \tag {2} +$$ + +Subsequently, for $L = KHW$ , the $W_{patch}$ is reshaped into $\hat{W}_{patch} \in \mathbb{R}^{K \times H^2 W^2}$ . Next, the image-level weights + +Table 1. Comparison of few-shot anomaly detection on MVTec-AD and VisA datasets. The results are presented as mean ± standard deviation, based on three runs conducted with different random seeds. The best results are in bold, and the second-best results are underlined. + +
SetupMethodMVTec-ADVisA
I-AUCI-F1P-AUCP-F1PROI-AUCI-F1P-AUCP-F1PRO
1-shotPadim [6]76.6±3.188.2±1.189.3±0.940.2±2.173.3±2.062.8±5.475.3±1.289.9±0.817.4±1.764.3±2.4
PatchCore [19]83.4±3.090.5±1.592.0±1.050.4±2.179.7±2.079.9±2.981.7±1.695.4±0.638.0±1.980.5±2.5
WinCLIP+ [10]93.1±2.093.7±1.195.2±0.555.9±2.787.1±1.283.8±4.083.1±1.796.4±0.441.3±2.385.1±2.1
APRIL-GAN [4]92.0±0.392.4±0.295.1±0.154.2±0.090.6±0.291.2±0.886.9±0.696.0±0.038.5±0.390.0±0.1
PromptAD [13]94.6±1.7-95.9±0.5-87.9±1.086.9±2.3-96.7±0.4-85.1±2.5
ReMP-AD(ours)94.7±0.194.3±0.195.6±0.156.8±0.190.2±0.391.3±0.487.0±0.297.6±0.044.7±0.292.2±0.1
2-shotPadim [6]78.9±3.189.2±1.191.3±0.743.7±1.578.2±1.867.4±5.175.7±1.892.0±0.721.1±2.470.1±2.6
PatchCore [19]86.3±3.392.0±1.593.3±0.653.0±1.782.3±1.381.6±4.082.5±1.896.1±0.541.0±3.982.6±2.3
WinCLIP+ [10]94.4±1.394.4±0.896.0±0.358.4±1.788.4±0.984.6±2.483.0±1.496.8±0.343.5±3.386.2±1.4
APRIL-GAN [4]92.4±0.392.6±0.195.5±0.055.9±0.591.3±0.192.2±0.387.7±0.396.2±0.039.3±0.290.1±0.1
PromptAD [13]95.7±1.5-96.2±0.3-88.5±0.888.3±2.0-97.1±0.3-85.8±2.1
ReMP-AD(ours)95.9±0.595.1±0.196.2±0.158.9±0.391.6±0.392.9±0.188.2±0.797.8±0.147.1±0.992.6±0.0
4-shotPadim [6]80.4±2.590.2±1.292.6±0.746.1±1.881.3±1.972.8±2.978.0±1.293.2±0.524.6±1.872.6±1.9
PatchCore [19]88.8±2.692.6±1.694.3±0.555.0±1.984.3±1.485.3±2.184.3±1.396.8±0.343.9±3.184.9±1.4
WinCLIP+ [10]95.2±1.394.7±0.896.2±0.359.5±1.889.0±0.887.3±1.884.2±1.697.2±0.247.0±3.087.6±0.9
APRIL-GAN [4]92.8±0.292.8±0.195.9±0.056.9±0.191.8±0.192.6±0.488.4±0.596.2±0.040.0±0.190.2±0.1
PromptAD [13]96.6±0.9-96.5±0.2-90.5±0.789.1±1.7-97.4±0.4-86.2±1.7
ReMP-AD(ours)96.8±0.295.5±0.196.6±0.161.1±0.692.8±0.294.1±0.289.7±0.697.8±0.048.2±1.092.9±0.1
+ +$\hat{W}_{cls}$ , as defined in Algorithm 1, are incorporated with the patch-level weights to form a hierarchical weight matrix $W_{pro}$ as follows: + +$$ +W _ {p r o} ^ {i} = \hat {W} _ {p a t c h} ^ {i} \cdot \hat {W} _ {c l s} ^ {i} \tag {3} +$$ + +where $i$ indexes first dimension of $W_{pro}$ . + +However, unrelated regions might still be integrated into the prototype due to a dense $W_{pro}$ . To address this issue, a filtering operation is applied with a hyperparameter $\lambda$ to promote sparsity. + +$$ +\hat {W} _ {p r o} = \left\{ \begin{array}{l l} W _ {p r o}, & W _ {p r o} > \lambda , \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {4} +$$ + +Anomalies are detected by measuring the discrepancy between the query feature and its prototype. The prototype feature should resemble its normal query features while differing from the abnormal ones. After reshaping, the weight $\hat{W}_{pro} \in \mathbb{R}^{HW \times L}$ is used to integrate the prototype $F_{pro}$ in the following equation: + +$$ +F _ {p r o} = \frac {\hat {W} _ {p r o} \cdot B _ {p a t c h}}{\| \hat {W} _ {p r o} \| _ {1}} \tag {5} +$$ + +The $i$ -th patch localization result $M_{mem}^{i}$ from memory bank can be computed as follows: + +$$ +\begin{array}{l} M _ {m e m} ^ {i} = \left(1 - \langle F _ {p a t c h} ^ {i}, F _ {p r o} ^ {i} \rangle\right) \\ + \left(1 - \max _ {j} \langle F _ {p a t c h} ^ {i}, B _ {p a t c h} ^ {j} \rangle\right) \tag {6} \\ \end{array} +$$ + +# 3.4. Vision-Language Prior Fusion + +In designing the visual prompt, it is assumed that visually salient regions are more likely to contain anomalies. These potentially anomalous regions are identified and utilized as visual prompts. To quantitatively assess the anomaly likelihood of each patch, the Vision Prior Knowledge (VPK) score is calculated as follows: + +$$ +S _ {v p k} ^ {i} = 1 - \frac {1}{m _ {k}} \sum_ {j = 1} ^ {m _ {k}} \left(T o p _ {m _ {k}} \langle F _ {p a t c h} ^ {i}, F _ {p a t c h} ^ {j} \rangle\right) \tag {7} +$$ + +where $\langle F_{patch}^{i},F_{patch}^{j}\rangle$ represents the calculation of cosine similarity matrix between the $i$ -th and $j$ -th patch features, and $Top_{m_k}$ indicates the top $m_{k}$ highest values in the similarity matrix, where $m_{k}$ is a hyperparameter. The indices $i$ and $j$ range from 1 to $HW$ . + +The vision prior knowledge score $S_{vpk}$ is first converted into a binary VPK mask $M_{vpk}$ using a hyperparameter $m_{\lambda}$ . This hyperparameter sets a threshold for the conversion: values above the threshold are assigned a weight $w_{1}$ , and values equal to or below the threshold are assigned a weight $w_{2}$ . This approach enhances the focus on potential abnormal regions and reduces the emphasis on less likely ones. + +$$ +M _ {v p k} = \left\{ \begin{array}{l l} 1, & S _ {v p k} > m _ {\lambda}, \\ 0, & \text {o t h e r w i s e .} \end{array} , \tilde {M} _ {v p k} = \left\{ \begin{array}{l l} w _ {1}, & S _ {v p k} > m _ {\lambda}, \\ w _ {2}, & \text {o t h e r w i s e .} \end{array} \right. \right. \tag {8} +$$ + +Subsequently, the vision prior knowledge weighted mask + +$\tilde{M}_{vpk}$ , is multiplied with $F_{patch}$ : + +$$ +\tilde {F} _ {\text {p a t c h}} = F _ {\text {p a t c h}} \circ \tilde {M} _ {v p k} \tag {9} +$$ + +where $\circ$ is the Hadamard product. Then, the resulting weighted feature map $\tilde{F}_{patch}$ is fed to the next layer of the pre-trained image encoder, referred to as the VLPF image encoder. This operation is applied at $\hat{p}$ th layer in the VLPF image encoder. + +To align with the textual features, an MLP layer is trained to generate the post-processed feature $F_{patch}^{\prime} \in \mathbb{R}^{HW \times C}$ from $F_{patch} \in \mathbb{R}^{HW \times C_{patch}}$ . A language prior knowledge score $S_{lpk}$ and its corresponding binary mask $M_{lpk}$ are as follows: + +$$ +S _ {l p k} = \frac {\exp \left\langle F _ {p a t c h} ^ {\prime} , F _ {t} ^ {a} / \tau \right\rangle}{\exp \left\langle F _ {p a t c h} ^ {\prime} , F _ {t} ^ {n} / \tau \right\rangle + \exp \left\langle F _ {p a t c h} ^ {\prime} , F _ {t} ^ {a} / \tau \right\rangle} \tag {10} +$$ + +$$ +M _ {l p k} = \left\{ \begin{array}{l l} 1, & S _ {l p k} > m _ {\lambda}, \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {11} +$$ + +Then the vision-language prompt fusion mask $M_{vlpf}$ is obtained as the intersection of $M_{vpk}$ and $M_{lpk}$ . Next, connected area analysis is then applied to the $M_{vlpf}$ to produce the binary region $R_{vlpf} \in \mathbb{R}^{N \times HW}$ . To extend from patch-level to region-level, the average feature of patches within the same region is utilized as the region feature $F_{region}$ . + +$$ +F _ {r e g i o n} = \sum_ {n = 1} ^ {N} \frac {R _ {v l p f} ^ {n} \circ F _ {p a t c h} ^ {\prime}}{\| R _ {v l p f} ^ {n} \| _ {1}} \tag {12} +$$ + +This feature is then combined with the patch feature $F_{patch}^{\prime}$ using the hyperparameter $m_{\alpha}$ : + +$$ +F _ {r e g i o n} ^ {\prime} = m _ {\alpha} F _ {r e g i o n} + \left(1 - m _ {\alpha}\right) F _ {p a t c h} ^ {\prime} \tag {13} +$$ + +Next, the language-prior segmentation map $M_{vl}$ is generated using a binary classification approach: + +$$ +M _ {v l} = \frac {\exp \left\langle F _ {r e g i o n} ^ {\prime} , F _ {t} ^ {a} / \tau \right\rangle}{\exp \left\langle F _ {r e g i o n} ^ {\prime} , F _ {t} ^ {n} / \tau \right\rangle + \exp \left\langle F _ {r e g i o n} ^ {\prime} , F _ {t} ^ {a} / \tau \right\rangle} \tag {14} +$$ + +During inference, the final segmentation result $M_{pix}$ is obtained by fusing the language-guided map $M_{vl}$ and the memory-guided map $M_{mem}$ , as follows: + +$$ +M _ {p i x} = \frac {1}{2 P} \sum_ {p = 1} ^ {P} U P \left(M _ {v l} ^ {p}\right) + \frac {1}{2 P} \sum_ {p = 1} ^ {P} U P \left(M _ {m e m} ^ {p}\right) \tag {15} +$$ + +where $p$ denotes the result computed at the $p$ -th layer of the ViT, with a total of $P$ layers. The operator $UP(\cdot)$ represents the upsampling operation. + +For the final classification outcome, the language-prior classification score $S_{vl} \in [0,1]$ is combined with the maximum pixel value derived from the anomaly map. + +$$ +S _ {i m g} = \frac {1}{2} S _ {v l} + \frac {1}{2} \max \left(M _ {p i x}\right) \tag {16} +$$ + +# 4. Experiments + +We comprehensively compared ReMP-AD against recent state-of-the-art approaches across 1-shot, 2-shot, and 4-shot scenarios. Furthermore, ablation studies were carried out to + +examine the individual contributions of the VLPF and ICTR enhancements. + +# 4.1. Experiments setup + +# 4.1.1. Dataset descriptions + +Our benchmarks include the MVTec-AD [1], VisA [30] and PCB-Bank [25] datasets. The MVTec-AD dataset comprises 15 objects, with image resolutions ranging from $700 \times 700$ to $900 \times 900$ pixels. It includes 3,629 samples in the training set and 1,725 in the test set. Similarly, the VisA dataset contains 12 objects, each approximately $1.5\mathrm{K} \times 1\mathrm{K}$ pixels in size, with 9,621 normal samples and 1,200 anomalous samples. PCB-Bank is a dataset of printed circuit boards (PCBs) encompassing seven different PCB types. It consists of 4,214 normal samples in the training set and 2,253 test samples. Following APRIL-GAN [4], we train on the VisA test set and evaluate on MVTec-AD. Additionally, we train on the MVTec-AD test set and evaluate on the other datasets. + +# 4.1.2. Implementation details + +In our experiments, we utilize the CLIP model with ViT-L-14-336 pretrained backbone. The threshold $\lambda$ in (4) is 0.72 for the VisA dataset, 0.64 for the MVTec-AD dataset. In (7), $m_{k}$ is set to 100. The hyperparameters for generating $\tilde{M}_{vpk}$ in (8) include $m_{\lambda}$ , $w_{1}$ , and $w_{2}$ , with optimal settings of $m_{\lambda} = 0.5$ , $w_{1} = 1.2$ , and $w_{2} = 0.8$ . The attention layers $\hat{p}$ in the VLPF image encoder are implemented at the 12th, 18th, and 24th layers. The hyperparameter $m_{\alpha}$ in (13) is defined as 0.2. We configure the number of layers $P$ in (15) for extracting patch-level features to 4, selecting features evenly from layers 6, 12, 18, 24 of the 24 transformer layers in the image encoder. All images are resized to $518 \times 518$ pixels before being input into the encoder. The Adam is employed as an optimizer with an initial learning rate of 1e-3. The model is trained for 3 epochs on MVTec-AD and 15 epochs on VisA, using a batch size of 8. All experiments are conducted on a single NVIDIA GeForce RTX 4060 Ti, with three runs performed using different random seeds, and the results are averaged. + +# 4.1.3. Evaluation metrics + +Industrial Visual Anomaly Detection is evaluated by three key metrics: the area under the receiver operating characteristic curve (AUC), the F1-score at the optimal threshold (F1), and the per-region overlap (PRO). For image-level anomaly classification, I-AUC and I-F1 are used, while pixel-level anomaly segmentation is assessed using P-AUC, P-F1, and PRO. + +# 4.2. Image/Pixel-level comparison results + +Table 1 presents image-level comparative results for ReMP-AD and existing methods, including PaDiM [6] and Patch-Core [19], which adapt traditional full-shot techniques for + +![](images/fc1dff5bd7696f2850e40dfef4ad0e95f54ac0d87f845f7d2be564b68aa41da0.jpg) +Figure 3. Qualitative segmentation results of 4-shot anomaly detection on MVTec-AD and VisA dataset, compared with PatchCore [19], WinCLIP+ [10], APRIL-GAN [4], and PromptAD [13]. The redder the area on the segmentation map, the higher the corresponding anomaly score at that location. From the results, it is clear that our method excels at reducing false positives and effectively detecting large abnormal areas. + +![](images/f25eb9e2e0d30be1c9a391351d08e379905a3fd0eadc6b3c13487459f3496017.jpg) + +Table 2. Comparison of few-shot anomaly detection on PCB-Bank dataset. Metrics are I-AUC/I-F1 at first raw and P-AUC/P-F1/PRO at second raw. + +
SettingWinCLIP+ [10]APRIL-GAN [4]ReMP-AD(ours)
1-shot62.8/71.077.6/77.779.0/77.6
88.3/11.5/62.791.9/25.6/75.893.8/32.4/80.5
2-shot66.4/72.879.7/77.982.3/80.1
89.0/13.0/64.192.6/26.1/77.295.6/36.6/84.9
4-shot69.6/73.083.5/80.586.9/84.1
90.8/17.4/68.192.7/26.8/77.895.4/38.8/85.4
+ +Table 3. Results of integrating the proposed method into the many-shot framework (MuSc). LP represents the branch of language-prior anomaly detection (APRIL-GAN). * indicates that only the VLPF image encoder is used. Metrics are P-AUC/P-F1/I-AUC/I-F1. + +
MethodMVTec-ADVisA
Baseline97.1/62.2/97.8/97.498.7/49.0/92.6/88.9
w/ VLPF*(ours)97.2/63.0/97.9/97.398.7/49.1/92.6/89.1
w/ LP97.1/62.5/98.2/97.798.7/52.2/94.7/90.8
w/ LP + VLPF(ours)97.2/63.4/98.5/97.698.8/52.7/94.9/91.1
+ +few-shot scenarios. These methods demonstrate limited anomaly detection (AD) performance. In contrast, WinCLIP+ [10] and APRIL-GAN [4], which first introduce CLIP, significantly enhance image-level AD in few-shot settings. PromptAD [13] further leverages prompt learning and a new attention mechanism, achieving state-of-the-art performance. Notably, ReMP-AD shows substantial improvements across all three settings in both benchmarks compared to the aforementioned methods. Except for comparable performance on the MVtec-AD dataset in the 1-shot setting, ReMP-AD surpasses the state-of-the-art in both benchmarks. Moreover, to ensure a fair comparison, Table 11 reports the results using identical backbones, since PromptAD employs smaller backbones, which may impact the performance. + +To further prove its effectiveness in a real-world application, ReMP-AD is applied to a printed circuit board dataset (PCB-Bank). Table 2 presents the results on the PCB-Bank + +Table 4. Results of ablation study under 4-shot setting on VisA dataset. The best results are in **bold**, and the second-best results are in **underlined**. + +
ICTRVLPFP-AUCP-F1I-AUCI-F1
GPRCTB
×××96.439.791.587.4
✓××97.145.494.089.3
✓✓×97.646.994.089.5
✓✓✓97.848.094.189.9
+ +dataset in comparison with WinCLIP+ [10] and APRILGAN [4]. The proposed method demonstrates superior performance compared to the two language-prior methods and exhibits strong capability in detecting PCB defects. Furthermore, the proposed method demonstrates strong generalizability on the MPDD [11] and Real-IAD [22] datasets, as shown in Table 10. + +To provide a more intuitive understanding, the quantitative results of anomaly localization are presented in Fig. 3. For example, in the third row of Fig. 3 (a), ReMP-AD's results show comprehensive recognition of the "tape" sticking to the tile. In the third row of Fig. 3 (b), it exhibits fewer false positives for the wick of the candle compared to APRIL-GAN [4]. In conclusion, compared to WinCLIP+ [10] and APRIL-GAN [4], ReMP-AD demonstrates superior anomaly localization capability and reduced false positives for both objects and textures in the 4-shot setting. + +We also validate our method on a many-shot setting, which is proposed by MuSc [12]. In Table 3, the VLPF module is applied to MuSc to evaluate the efficiency of the method, with mask attention alone improving P-F1 by $0.8\%$ on MVTec-AD and $0.1\%$ on VisA. As MuSc is a vision-prior method, the language-prior branch of APRIL-GAN is introduced to further demonstrate the efficiency of the VLPF module. When VLPF is applied to this branch, additional gains of $0.9\%$ in P-F1 on MVTec-AD and $0.5\%$ on VisA are observed. It shows the effectiveness of the proposed method in integrating vision and language prior knowledge. + +Table 5. Comparison of results with and without CTB on VisA dataset in few-shot settings.Metrics are P-AUC/P-F1/I-AUC/I-F1. + +
Settingw/o CTBw CTB
1-shot97.3/44.5/91.2/86.697.6/44.8/90.9/86.8
2-shot97.4/45.1/92.5/88.497.6/46.3/93.0/89.1
4-shot97.4/46.6/93.6/88.197.8/48.0/94.1/89.9
8-shot97.3/47.0/94.3/89.497.7/48.1/94.7/89.9
16-shot97.4/49.4/95.2/90.897.6/50.3/95.3/90.9
+ +Table 6. Results of integrating the state-of-the-art method PromptAD [13] with the intra-class token retrieval module on MVTec-AD and VisA datasets. Metrics are P-AUC/P-F1/I-AUC/I-F1. + +
MethodSettingPromptADPromptAD+ICTR
MVTec-AD1-shot95.4/52.7/91.8/93.194.9/53.1/92.8/93.6
2-shot95.8/54.5/94.5/95.595.8/54.9/94.7/95.4
4-shot96.1/55.9/95.8/96.096.1/56.6/96.0/96.0
VisA1-shot95.9/34.8/85.4/82.296.3/35.5/86.2/82.8
2-shot96.4/35.9/84.6/81.796.3/36.6/85.2/82.1
4-shot96.8/36.5/87.9/83.996.7/37.4/88.4/84.0
+ +# 4.3. Ablation study + +The proposed method incorporates two key components: Intra-Class Token Retrieval (ICTR) and Vision-Language Prior Fusion (VLPF), the former comprising Global Prototype Retrieval (GPR) and Class Token Balancing (CTB). In this section, ablation studies are conducted to evaluate the effectiveness of each component within the overall framework. + +In Table 4, the first and second rows show that applying GPR improves pixel-level and image-level results (AUC/F1) by $0.7\% / 5.7\%$ and $2.5\% / 1.9\%$ , respectively. This demonstrates that GPR effectively integrates global patterns of normality by capturing and incorporating global feature correlations. Additionally, as shown in Table 4, the second and third rows show that the inclusion of CTB leads to further improvements over GPR, increasing P-AUC and P-F1 by $0.5\%$ and $1.5\%$ , respectively. This suggests that CTB enhances the robustness of the retrieved prototypes, particularly when variations exist among the reference samples. After using VLPF shown in Table 4, the pixel-level and image-level results (AUC/F1) show improvements, respectively, indicating that VLPF effectively enhances localization ability. VLPF leverages vision-prior knowledge to guide the network's attention toward patches with potential anomalies. The results demonstrate that this method for selecting potential abnormal areas and achieving region-level anomaly localization is noteworthy. + +To further explore CTB's efficiency in the few-shot setting, Table 5 shows that the presence of CTB significantly enhances performance in multiple-shot scenarios compared to its absence. Improvements from 1-shot to 16-shot demonstrate that CTB's ability to adjust the confidence of references is valuable in the few-shot setting. + +Additionally, some other few-shot methods [10, 13, 19] did not use external datasets for pre-training. For a fair + +![](images/c023c3b46a8f2334a37c05a6e6cf85612d182a81fdf5f8711b48155da8ca7e20.jpg) +(a) I-AUC + +![](images/812f8cf1d99849f0e9cc8ef4e60ec9a19f412f1026421443efa883912af44d9e.jpg) +(b) I-F1 +Figure 4. Results under different $m_{k}$ and $m_{\lambda}$ applied to VLPF on the VisA dataset in a 4-shot setting. + +comparison, we incorporated the ICTR module into PromptAD [13], following a standard few-shot setup. Table 6 shows that adding the ICTR module improves performance in the few-shot setting, confirming its effectiveness. The performance gains from 1-shot to 4-shot further highlight the effectiveness of the ICTR module in enhancing feature retrieval and adaptation. + +# 4.4. Hyper-parameter analysis + +To evaluate the impact of the hyperparameters $m_{k}$ and $m_{\lambda}$ on the generation of the VPK mask, Fig. 4 illustrates their effects. The hyperparameter $m_{k}$ exhibits minimal influence on the I-AUC, whereas the I-F1 score reaches its peak at $m_{k} = 100$ . In contrast, $m_{\lambda}$ significantly affects the results, with an optimal peak observed at $m_{\lambda} = 0.5$ . Therefore, we adopt the default settings of $m_{k} = 100$ and $m_{\lambda} = 0.5$ . Furthermore, we also investigate additional hyperparameters, including $\hat{p}$ , which refers to the attention layers used in VLPF, $w_{1}$ and $w_{2}$ , which denote the attention weights within VLPF, and $\lambda$ , which represents the filtering threshold in ICTR. The effects of these hyperparameters are analyzed in the supplementary material: the impact of $\hat{p}$ is shown in Table 7, the influences of $w_{1}$ and $w_{2}$ are discussed in Table 8, and the effect of $\lambda$ is detailed in Table 9. + +# 5. Conclusion + +In this paper, we introduce ReMP-AD, a novel anomaly detection approach that integrates prototypical patterns from reference samples and fuses vision-language prior knowledge for few-shot anomaly detection. First, we propose a prototype retrieval method that extracts normal prototypes from the references at both pixel and global levels. Furthermore, to better exploit potential abnormal regions, we introduce vision-language prior fusion, applying mask attention in the vision encoder to capture abnormal features and extend anomaly detection from the patch level to the region level. Moreover, extensive experiments confirm the superiority of ReMP-AD in few-shot industrial image anomaly detection. The application in real-world printed circuit board anomaly detection further proves its effectiveness. In the future, we intend to apply this approach to a broader spectrum of 3D industrial scenarios. + +# 6. Acknowledgement + +This work was supported by the National Key R&D Program of China under Grant No. 2023YFA1008500. Guanglei Yang is supported by the Postdoctoral Fellowship Program of CPSF under Grant Number GZC20233458 and the Postdoctoral Fund of Harbin Institute of Technology under Grant Number AUGA5630115523. + +# References + +[1] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 2, 6 +[2] Yunkang Cao, Xiaohao Xu, Zhaoge Liu, and Weiming Shen. Collaborative discrepancy optimization for reliable image anomaly localization. IEEE Transactions on Industrial Informatics, 19(11):10674-10683, 2023. 1 +[3] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, 2024. 1, 3 +[4] Xuhai Chen, Yue Han, and Jiangning Zhang. A zero-/fewshot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad. arXiv preprint arXiv:2305.17382, 2(4), 2023. 3, 5, 6, 7 +[5] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 1, 2 +[6] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 1, 2, 5, 6 +[7] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3 +[8] 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. 1, 2 +[9] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high-quality localization. In ACM Multimedia 2024, 2024. 3 +[10] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision + +and Pattern Recognition, pages 19606-19616, 2023. 1, 3, 5, 7, 8 +[11] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), pages 66-71. IEEE, 2021. 7, 1 +[12] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. Musc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. In The Twelfth International Conference on Learning Representations, 2024. 7 +[13] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompted: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838-16848, 2024. 1, 3, 5, 7, 8 +[14] Yuanze Li, Haolin Wang, Shihao Yuan, Ming Liu, Debin Zhao, Yiwen Guo, Chen Xu, Guangming Shi, and Wangmeng Zuo. Myriad: Large multimodal model by applying vision experts for industrial anomaly detection. arXiv preprint arXiv:2310.19070, 2023. 1 +[15] Yufei Liang, Jiangning Zhang, Shiwei Zhao, Runze Wu, Yong Liu, and Shuwen Pan. Omni-frequency channel-selection representations for unsupervised anomaly detection. IEEE Transactions on Image Processing, 2023. 1, 2 +[16] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, 2024. 3 +[17] Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. Vcp-clip: A visual context prompting model for zero-shot anomaly segmentation. In European Conference on Computer Vision, 2024. 1, 3 +[18] 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 +[19] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2022. 1, 2, 5, 6, 7, 8 +[20] 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. 3 +[21] Yang Shu, Xingzhuo Guo, Jialong Wu, Ximei Wang, Jian-min Wang, and Mingsheng Long. Clipood: Generalizing clip to out-of-distributions. In International Conference on Machine Learning, pages 31716-31731. PMLR, 2023. 3 + +[22] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22883-22892, 2024. 7, 1 +[23] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1802-1812, 2023. 3 +[24] Haiming Yao, Yunkang Cao, Wei Luo, Weihang Zhang, Wenyong Yu, and Weiming Shen. Prior normality prompt transformer for multiclass industrial image anomaly detection. IEEE Transactions on Industrial Informatics, 20(10): 11866-11876, 2024. 1 +[25] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. In European Conference on Computer Vision, 2024. 1, 2, 6 +[26] 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 +[27] 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. 3 +[28] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In The Twelfth International Conference on Learning Representations, 2024. 1, 3 +[29] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 3 +[30] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 2, 6 \ No newline at end of file diff --git a/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/images.zip b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6831d12d2afa095e80ee2231a0a856bdd11833d4 --- /dev/null +++ b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea47c39aabb9c72a8c91e26ae492fd359143395a6790f7b081a98caf20d4c7fa +size 679684 diff --git a/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/layout.json b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a5edd541a213595fb8f35f62fa27ddc01190a3a6 --- /dev/null +++ b/ICCV/2025/ReMP-AD_ Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7edcfa8efd2c9bcb450f2493dbd8112c3a77cbc6825a0d32328975355ca7b81b +size 417540 diff --git a/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/40353887-4f73-4a0a-b981-5f923a53c68c_content_list.json b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/40353887-4f73-4a0a-b981-5f923a53c68c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..09248f9d33c7cca8e994c9d4f5d927365cc314c8 --- /dev/null +++ b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/40353887-4f73-4a0a-b981-5f923a53c68c_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5520a51a97c944b472d748382cd9750efa4ab2c1c7493e10f50da2896ba7bc5b +size 105243 diff --git a/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/40353887-4f73-4a0a-b981-5f923a53c68c_model.json b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/40353887-4f73-4a0a-b981-5f923a53c68c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6b1d2a97d9376a273b7ce42b243bf025eec562e8 --- /dev/null +++ b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/40353887-4f73-4a0a-b981-5f923a53c68c_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecc044aa19a0f0793bca68b0a9aeaa62ce2a43a4c636b3402558436370095b31 +size 128219 diff --git a/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/40353887-4f73-4a0a-b981-5f923a53c68c_origin.pdf b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/40353887-4f73-4a0a-b981-5f923a53c68c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4e506dc030e95bf264ffe3662594e9b0be30512d --- /dev/null +++ b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/40353887-4f73-4a0a-b981-5f923a53c68c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fce82d910cc04f829f3f1aa5cc3daacba7f045628bd6f3f23c6b1783c3a931fc +size 871628 diff --git a/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/full.md b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/full.md new file mode 100644 index 0000000000000000000000000000000000000000..33c2d5c664caa483895249bd4740d6b8267af915 --- /dev/null +++ b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/full.md @@ -0,0 +1,319 @@ +# RePoseD: Efficient Relative Pose Estimation With Known Depth Information + +Yaqing Ding $^{1}$ , Viktor Kocur $^{2}$ , Václav Vávra $^{1}$ , Zuzana Berger Haladová $^{2}$ , Jian Yang $^{3}$ , Torsten Sattler $^{4}$ and Zuzana Kukelova $^{1}$ + +Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague + +$^{2}$ Faculty of Mathematics, Physics and Informatics, Comenius University in Bratislava + +3 PCA Lab, VCIP, College of Computer Science, Nankai University, Tianjin, China + +$^{4}$ Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague + +# Abstract + +Recent advances in monocular depth estimation methods (MDEs) and their improved accuracy open new possibilities for their applications. In this paper, we investigate how monocular depth estimates can be used for relative pose estimation. In particular, we are interested in answering the question whether using MDEs improves results over traditional point-based methods. We propose a novel framework for estimating the relative pose of two cameras from point correspondences with associated monocular depths. Since depth predictions are typically defined up to an unknown scale or even both unknown scale and shift parameters, our solvers jointly estimate the scale or both the scale and shift parameters along with the relative pose. We derive efficient solvers considering different types of depths for three camera configurations: (1) two calibrated cameras, (2) two cameras with an unknown shared focal length, and (3) two cameras with unknown different focal lengths. Our new solvers outperform state-of-the-art depth-aware solvers in terms of speed and accuracy. In extensive real experiments on multiple datasets and with various MDEs, we discuss which depth-aware solvers are preferable in which situation. The code is available at https://github.com/kocurvik/mdrp. + +# 1. Introduction + +Relative pose estimation is a core problem in a wide range of computer vision applications, including Structure-from-Motion (SfM) [57], visual localization [55, 61, 71], and autonomous robots [43]. Typically, the relative pose of cameras is estimated from 2D-2D image point correspondences that are established using feature matching. Given 2D-2D matches, the relative pose [28, 37, 38, 51, 59, 60] of two cameras can be estimated using the epipolar geometry constraints [29]. For two calibrated cameras, 5 point correspond + +![](images/dcbfd1ec816ab792aa2fed2ef9b7dbef0efe487d0b4f2e0f782b274ddc425a94.jpg) +Figure 1. Given point matches and their corresponding depths from a pair of images, we propose the use of different minimal solvers for relative pose estimation, according to the properties of the depth data. For local optimization, the Sampson error is generally more robust across varying depth conditions, while the reprojection error may yield better results when the depth measurements are highly reliable. + +dences are needed to estimate the 5 degrees of freedom of the relative pose (3 for rotation and two for the translation up to an unknown scale) [28, 37, 51, 60]. The relative pose of two cameras with unknown shared focal length can be estimated from 6 correspondences [28, 37, 38, 59]. If the full intrinsic calibration is unknown and/or both cameras have different calibrations, the relative pose can be estimated from 7 or 8 correspondences [29]. + +Due to noise in the measurements and the presence of outlier correspondences, algorithms for relative pose estimation are usually employed in RANSAC-style hypothesize-and-verify frameworks [23]. The number of RANSAC iterations grows exponentially with the number of correspondences needed to estimate the pose. As such, a considerable amount of work focuses on reducing the number of correspondences necessary for pose estimation. This is typically done by using additional information, such as point coplanarity [10, 29], the gravity direction from an Inertial Measurement Unit (IMU) [15-17, 24, 36, 50, 62], or information about the local feature geometry such as the scale and rotation of features [3, 27, 46] or local affine frames [5, 6, 22]. + +Recently, learning-based methods [9, 26, 30, 48, 53, 64, 67, 68] have achieved significant success in predicting depth + +from a single image. The increasing precision of Monocular Depth Estimation methods (MDEs) opens new possibilities of using depth information in downstream tasks. Recent studies have shown a growing focus on the use of dense depth estimates to improve the performance of various geometry-based applications, including structure-from-motion/SLAM [45, 49, 73], dense reconstruction [66, 70], and novel view synthesis [31, 42]. + +Depth naturally provides additional information for camera pose estimation. Thus, several recent works [2, 4, 20, 69] used monocular depths from MDEs to improve the relative pose estimation of cameras. The depth information provides additional constraints on the geometry, and thus reduces the number of correspondences necessary for the estimation. However, it also introduces new challenges: (1) Learning-based depth estimates are usually much more noisy and erroneous than image measurements. (2) MDEs typically produce scale/affine-invariant depth [9, 34, 48, 64], affine-invariant inverse depth [8, 54, 67, 68] or metric depth [7, 30, 53], meaning that the estimated depth or inverse depth is only defined up to an unknown scale factor or both unknown scale and shift parameters. (3) The unknown shift and scale can be different for different images of the same scene. Due to these challenges and the fact that the solvers proposed in [2, 20] use only relative depths, do not model unknown shifts, and have been tested only with depths from one network (Depth Anything [68] in [20], and a network for estimating the relative depth proposed directly in [2]), [2, 20] do not convincingly show the improvement of their depth-aware solvers over traditional 2D-2D point-based relative pose solvers [29, 51, 60]. The solvers proposed in [69], which model scale and shift, have been integrated into hybrid RANSAC [11] with Ceres-based optimization [1]. While this approach outperforms traditional methods, it is very time-consuming. Thus, prior work does not fully answer the question of whether recent MDEs provide useful information for the relative pose estimation task. + +This paper attempts to provide an answer to this question by presenting a comprehensive comparison of different depth-aware solvers for calibrated, shared focal length and different focal lengths cases, under different conditions, i.e., different sources of depth information, different feature detectors and matchers, different types of datasets (interior/exterior) and different optimization schemes inside RANSAC. To this end, we provide novel solvers for three different types of monocular depths: metric depth, scale-invariant depth, and affine-invariant depth. We introduce novel formulations that jointly estimate the relative pose of two cameras along with the scale or affine parameters of the monocular depths. The proposed solvers use the full information from the depth estimates, resulting in fewer correspondences necessary for the relative pose estimation com + +pared to solvers based on relative depths [20]. The main contributions of the paper are: + +- We integrate scale and affine parameters of monocular depth into solvers for relative pose estimation. These parameters are estimated together with the relative pose. +- We discuss all possible combinations of known and unknown shift and scale parameters. We propose several practical and efficient minimal solvers for calibrated cameras, two cameras with shared unknown focal length, and two cameras with different unknown focal lengths. Our solvers outperform state-of-the-art depth-aware solvers in terms of efficiency and accuracy. +- We evaluate different RANSAC schemes to account for potentially erroneous and noisy depth estimates. +- In extensive real experiments, we answer the questions which state-of-the-art MDEs are suitable for relative pose estimation, which require modeling unknown scale or affine parameters, and which provide the best improvement in terms of efficiency and accuracy over traditional, purely point correspondence-based solvers. + +# 2. Related Work + +Liwicki and Zach [46] show that the relative scales of SIFT [47] features can approximate the inverse relative depths of the features. In this case, one point correspondence provides two constraints for relative pose estimation, and the 5-DOF problem can be solved using only three point correspondences with two relative depths. In the same setting, Guan and Zhao [27] solve the generalized relative pose problem using three SIFT correspondences. Astermark et al. [2] propose a closed-form minimal solution to the relative pose of two calibrated pinhole cameras with known relative depths. In addition, they propose a neural network to estimate the relative depths. In [19], relative depths approximated from SIFT scales are used to estimate the calibrated homography. Barath and Sweeney [4] propose a 2-point solver with depth information for relative pose estimation. However, as pointed out in [69], this 2-point solver is degenerate for rigid alignments due to rank deficiency. In [20], the authors propose new minimal solvers that use relative depths for fundamental matrix estimation. They discuss different variants of solvers with known and unknown focal lengths. Their solvers work for monocular depths with unknown different scales. However, the solvers do not model unknown shifts in the depths, assuming that the shifts are zero in both images and thus can be ignored. All of the methods and minimal solvers discussed above use only the information from relative depths. Thus, they do not utilize the full information from the two monocular depth maps. + +Dust3r [65] is a recent 3D foundation model for two-view geometry estimation tasks, including depth prediction and relative pose estimation. The relative pose is estimated by aligning two point maps via a similarity transformation. + +Trained on large-scale datasets, Dust3r is able to handle scenarios where there is little overlap between the two images. Mast3r [41] further improves performance through feature matching. Mast3r minimizes both 3D-3D matching errors and 3D-2D reprojection errors to refine depths, relative poses, and intrinsic parameters. + +In this paper, in contrast to the minimal solvers mentioned above that use only relative depths, we solve the problem using the full information from two depth maps. Concurrently to our work, Yu et al. [69] developed three novel minimal relative pose solvers for affine-invariant depths. Their calibrated, shared-focal, and two-focal length solvers solve the same formulation as our solvers with unknown shift and scales. In order to deal with potentially noisy and erroneous depth information, Yu et al. propose to use their depth-aware solvers jointly with classical 2D-2D point-based solvers in a hybrid LO-MSAC [11, 63] framework. In their Madpose framework, they combine the symmetric depth-induced reprojection error and the Sampson error for scoring pose hypotheses and for local optimization. The main differences between our work and [69] are the following: (1) Our solutions to configurations with unknown shifts and scales, i.e., affine-invariant depths, result in smaller and thus faster solvers. For a detailed comparison of sizes and run-times of the proposed and the MADPose solvers [69], see Table 2. (2) In addition to affine-invariant solvers, we also derive novel minimal solvers for cases with known/zero shifts and either unknown or known/same scales. Moreover, we derive a calibrated solver for inverse depth and discuss the number of solutions for all other variants of known/unknown/same/different shifts and scales (see Table 1). (3) By testing more variants of solvers, including novel solvers with zero shifts, we show that in some scenarios it is not necessary to model unknown shifts. This is in contrast to the observation made by [69], who, by missing some variants of the solvers in their experiments, concluded that modeling the unknown shift is always beneficial (even for "metric" depths). (4) In contrast to using a complex hybrid RANSAC scheme and a combination of reprojection and Sampson errors that results in slower running times, we look at a simple and more efficient local optimization scheme based on using a single solver with either reprojection or Sampson error. + +# 3. Problem Statement + +Assume that a set of 3D points $\Omega$ is observed by two cameras with projection matrices $\mathbf{K}_1[\mathbf{I} \mid \mathbf{0}]$ and $\mathbf{K}_2[\mathbf{R} \mid \mathbf{T}]$ . Let $\{\mathbf{p}_i, \mathbf{q}_i\}, i = 1, \dots, n$ be a set of $n$ 2D point correspondences, i.e., the projections of the 3D points from $\Omega$ in the first and the second camera, respectively. Then we have + +$$ +\lambda_ {i} \mathbf {K} _ {2} ^ {- 1} \mathbf {q} _ {i} = \eta_ {i} \mathbf {R} \mathbf {K} _ {1} ^ {- 1} \mathbf {p} _ {i} + \mathbf {T}, \tag {1} +$$ + +where $\lambda_{i}$ and $\eta_{i}$ are the depths of the points $\mathbf{q}_i$ and $\mathbf{p}_i$ . + +MDEs provide either scale/affine-invariant depth [9, 34, 48, 64], affine-invariant inverse depth [8, 54, 67, 68], or metric depth [7, 30, 53]. In the most general case of affine-invariant depth, the true depths $\lambda_{i}$ and $\eta_{i}$ in (1) can be expressed as + +$$ +\eta_ {i} = s _ {1} \left(\alpha_ {i} + u\right), \lambda_ {i} = s _ {2} \left(\beta_ {i} + v\right), \tag {2} +$$ + +where $\alpha_{i},\beta_{i}$ are estimated affine-invariant depths, and $\{s_1,s_2\} ,\{u,v\}$ are the unknown scales and shifts of the depths. Substituting (2) into (1), we have + +$$ +s _ {2} (\beta_ {i} + v) \mathbf {K} _ {2} ^ {- 1} \mathbf {q} _ {i} = s _ {1} (\alpha_ {i} + u) \mathbf {R} \mathbf {K} _ {1} ^ {- 1} \mathbf {p} _ {i} + \mathbf {T}, \quad (3) +$$ + +Dividing (3) by $s_1$ , results in + +$$ +s (\beta_ {i} + v) \mathbf {K} _ {2} ^ {- 1} \mathbf {q} _ {i} = (\alpha_ {i} + u) \mathbf {R} \mathbf {K} _ {1} ^ {- 1} \mathbf {p} _ {i} + \mathbf {t}, \qquad (4) +$$ + +where $s = s_2 / s_1$ and $\mathbf{t} = \mathbf{T} / s_1$ , i.e., two different scales are replaced by the relative scale $s$ . Note that in this case, the scale of the translation $\mathbf{t}$ is defined w.r.t. $s_1$ , thus we can not further fix the scale of $\mathbf{t}$ , i.e., the translation vector is not estimated up to scale like in the standard relative pose solvers [51], but needs to be estimated as a 3 degrees of freedom (3-DOF) vector. This means that for calibrated cameras, i.e., for known calibration matrices $\mathbf{K}_1$ and $\mathbf{K}_2$ , we have a 9-DOF problem w.r.t. $\{s, u, v, \mathbf{R}, \mathbf{t}\}$ . Since one 3D-3D point correspondence provides three constraints of the form (4), we need at least 3 point correspondences with their monocular depths to solve this problem for the calibrated case. For the case of unknown focal lengths $f_1$ and $f_2$ in calibration matrices $\mathbf{K}_1$ and $\mathbf{K}_2$ , we have 10 and 11 DoF, for cases $f_1 = f_2$ and $f_1 \neq f_2$ respectively. + +In addition to affine-invariant depths, some MDE networks provide scale-invariant depths or even metric depths; moreover, in some applications the depth can be estimated directly using depth sensors. Thus in some scenarios, we can assume known/zero shift and/or equal scales and shifts of two depths. There are 6 possible combinations of known and unknown shifts and scales that result in different DOF problems. Table 1 lists all these combinations together with the number of solutions for different problem settings, with novel solvers to interesting practical setups that are tested in the main paper and in the Supplemental Material (SM) marked in gray. Note that for calibrated case the setups with known shifts and scales $\{s,u,v\} = \{1,0,0\}$ , and with known shifts and unknown scale $\{s,u,v\} = \{s,0,0\}$ can be solved using the well-known P3P solver [18, 25, 52]. The P3P solver can handle scale invariant depth, since it only uses single-side depth. Due to space limits, we only + +
SolverPoint CorrespondenceScale RatioShiftDOFNo. of Solutions
3D-3D3D-2D2D-2D
Calibrated5-point [51]--5--510
Rel3PT* [2]2-11(0, 0)54
P3P [18]-3-s(0, -)64
3PT1uu1201(u, u)78
3PT1uv2101(u, v)88
3PTsuu210s(u, u)88
3PTsuv300s(u, v)94
3PTsuv (inverse)300s(u, v)910
Equal f6-point [28]--6--615
3p3d* [20]3--1(0, 0)64
3PT100f1201(0, 0)76
3PT1uuf2101(u, u)86
3PT1uvf3001(u, v)94
3PTs00f210s(0, 0)84
3PTsuvf300s(u, u)96
4PTsuvf301s(u, v)108
Varying \( f_{1} \),\( f_{2} \)7-point [29]--7--73
4p4d* [20]4--s(0, 0)71
3PT100\( f_{1,2} \)2101(0, 0)83
3PT1uu\( f_{1,2} \)3001(u, u)94
4PT1uv\( f_{1,2} \)3011(u, v)1016
3PTs00\( f_{1,2} \)300s(0, 0)91
4PTsuv\( f_{1,2} \)301s(u, u)1018
4PTsuv\( f_{1,2} \)310s(u, v)114
+ +Table 1. Different combinations of using 3D-3D, 3D-2D, and 2D-2D point correspondences for focal length problems. The solvers marked gray are used in our experiments. Note that subscripts $\text{suv}$ in names of the solvers represent different configurations of known/unknown/same/different scales and shifts in the depth maps. *Relative depth was used, where one partial 3D-3D correspondence provides only 2 constraints. + +discuss two practical scenarios in the main paper: the two depths up to unknown scale with two different shifts, and the two depth maps up to unknown scale. Other cases are discussed in the SM. + +# 3.1. Calibrated Case + +In this section, we first discuss the solver for the calibrated case, where the intrinsic parameters of cameras are known, i.e., $\mathbf{K}_1$ and $\mathbf{K}_2$ are known. We use $\tilde{\mathbf{q}}_i = \mathbf{K}_2^{-1}\mathbf{q}_i$ , $\tilde{\mathbf{p}}_i = \mathbf{K}_1^{-1}\mathbf{p}_i$ to represent the normalized image points. + +Given 3 point correspondences and their monocular depths, substituting $\tilde{\mathbf{q}}_i,\tilde{\mathbf{p}}_i,i = 1,2,3$ into (4) we have nine equations in nine unknowns + +$$ +s \left(\beta_ {i} + v\right) \tilde {\mathbf {q}} _ {i} = \left(\alpha_ {i} + u\right) \mathbf {R} \tilde {\mathbf {p}} _ {i} + \mathbf {t}. \tag {5} +$$ + +By eliminating the translation from the above equations, i.e., by subtracting pairs of equations, we have + +$$ +s \left(\beta_ {i} + v\right) \tilde {\mathbf {q}} _ {i} - s \left(\beta_ {j} + v\right) \tilde {\mathbf {q}} _ {j} = \mathbf {R} \left(\left(\alpha_ {i} + u\right) \tilde {\mathbf {p}} _ {i} - \left(\alpha_ {j} + u\right) \tilde {\mathbf {p}} _ {j}\right), +$$ + +for $(i,j) = (1,2),(1,3),(2,3)$ + +Since applying the rotation matrix on a vector preserves the length of this vector, we have the following constraints + +$$ +\left\| s \left(\beta_ {i} + v\right) \tilde {\mathbf {q}} _ {i} - s \left(\beta_ {j} + v\right) \tilde {\mathbf {q}} _ {j} \right\| = \left\| \left(\alpha_ {i} + u\right) \tilde {\mathbf {p}} _ {i} - \left(\alpha_ {j} + u\right) \tilde {\mathbf {p}} _ {j} \right\|, +$$ + +
SolverG-JEigenTime (μs)
3PTsuv(Ours)3 × 6Closed-form1.46
3PTsuv(M)12 × 164 × 44.45
4PTsuvf(Ours)24 × 328 × 812.5
4PTsuvf(M)36 × 448 × 823.6
4PTsuvf1,2(Ours)20 × 244 × 46.45
4PTsuvf1,2(M)40 × 444 × 420.2
+ +Table 2. Efficiency comparison between the proposed affine-invariant solvers and those from MADPose [69], denoted as (M). + +for $(i,j) = (1,2),(1,3),(2,3)$ . This is also known as the rigid motion constraint, which preserves the distance between two 3D points under a rigid transformation, and has been widely used in the literature [18, 23, 52, 72]. In this case, the rotation is eliminated and we obtain three equations in three unknowns. Since the equations only contain $s^2$ , we can let $c = s^2$ to reduce the degrees and eliminate the symmetries. In short, if $s$ is a solution, then $-s$ is also a solution. However, $s$ should be positive since it presents the relative scale of two depths. Thus a solution to $c = s^2$ gives us only one geometrically feasible solution for $s$ . + +The three above equations can be written in a matrix form $\mathbf{M}[cv^2, cv, c, u^2, u, 1]^\top = 0$ , where $\mathbf{M}$ is a $3 \times 6$ coefficient matrix. After Gauss-Jordan (G-J) elimination of this matrix, the three monomials $\{cv^2, cv, c\}$ can be expressed as quadratic functions of $u$ , i.e., $cv^2 = g_1(u)$ , $cv = g_2(u)$ , $c = g_3(u)$ , where $g_1, g_2, g_3$ are polynomials in $u$ of degree 2. Since $(cv)^2 = (c)(cv^2)$ , we have $g_2^2 = g_1g_3$ , resulting in a quartic equation in $u$ . This equation can be solved in closed form. We denote the final solver as $3\mathrm{PT}_{suv}$ . Note that the inverse depth model results in a more complex solver without improving performance in our experiments (details provided in the SM). Therefore, we do not discuss the inverse depth model in the main paper. + +# 3.2. Focal Length Problems + +In many practical scenarios, the intrinsic parameters of the cameras may not be available. However, it is often reasonable to assume that modern cameras have square-shaped pixels, and the principal point coincides with the image center [28]. This is a widely used assumption in many camera geometry solvers, where the only unknown intrinsic parameters are focal lengths, i.e., $\mathbf{K}_i = \mathrm{diag}(f_i,f_i,1)$ . + +# Shared Unknown Focal Length + +First, let us assume that the two cameras have a shared unknown focal length, i.e., $\mathbf{K}_1 = \mathbf{K}_2 = \mathrm{diag}(f,f,1)$ , which is a common scenario, e.g., when estimating the motion of a single uncalibrated camera. This is a 10-DOF problem w.r.t. $\{s,u,v,\mathbf{R},\mathbf{t},f\}$ , and we need at least three 3D-3D point correspondences and one 2D-2D point correspondence, i.e., four 2D-2D correspondences from which three have monocular depths estimated in both images and + +thus can be considered as 3D-3D correspondences. Note that, here the 3D-3D point correspondences depend on the unknown focal length parameter. + +In this case, we may obtain the following system of 6 equations in 6 unknowns $\{s,u,v,f,\eta_4,\lambda_4\}$ + +$$ +\begin{array}{l} \left\| s \mathbf {K} _ {2} ^ {- 1} \left(\left(\beta_ {i} + v\right) \mathbf {q} _ {i} - s \left(\beta_ {j} + v\right) \mathbf {q} _ {j}\right) \right\| = \left\| \mathbf {K} _ {1} ^ {- 1} \left(\left(\alpha_ {i} + u\right) \mathbf {p} _ {i} - \left(\alpha_ {j} + u\right) \mathbf {p} _ {j}\right) \right\|, \\ \left\| s \mathbf {K} _ {2} ^ {- 1} \left(\left(\beta_ {k} + v\right) \mathbf {q} _ {k} - \lambda_ {4} \mathbf {q} _ {4}\right) \right\| = \left\| \mathbf {K} _ {1} ^ {- 1} \left(\left(\alpha_ {k} + u\right) \mathbf {p} _ {k} - \eta_ {4} \mathbf {p} _ {4}\right) \right\|, \\ \end{array} +$$ + +for $(i,j) = (1,2),(1,3),(2,3)$ and $k = 1,2,3$ . To solve such a system, we need to perform G-J elimination of a large matrix. However, the system has up to 28 real solutions. Thus, its solver is not practically relevant. + +To find a more efficient and practical solution, we use four 2D-2D point correspondences with their four depths in both images. In this case, we obtain the following 6 equations in 4 unknowns $\{s, u, v, f\}$ + +$$ +\left\| s \mathbf {K} _ {2} ^ {- 1} \left(\left(\beta_ {i} + v\right) \mathbf {q} _ {i} - \left(\beta_ {j} + v\right) \mathbf {q} _ {j}\right) \right\| = \left\| \mathbf {K} _ {1} ^ {- 1} \left(\left(\alpha_ {i} + u\right) \mathbf {p} _ {i} - \left(\alpha_ {j} + u\right) \mathbf {p} _ {j}\right) \right\|, +$$ + +where $i,j = 1,2,3,4$ , $i \neq j$ . In general, it is an overconstrained system. One way to solve it is to use four of these six equations in four unknowns. The system of four polynomials can be solved using the Gröbner basis method [40], where the final solver performs G-J elimination of a $24 \times 32$ matrix and extracts up to 8 real solutions from the eigenvalues and eigenvectors of an $8 \times 8$ matrix. We denote this solver as $4\mathrm{PT}_{\text{suv}}f$ . Alternatively, by using all six equations, we can derive a significantly smaller solver with up to two real solutions. This solver only needs to perform the G-J elimination of a $6 \times 8$ matrix. However, it is more sensitive to noise. Further details on this simplified focal length solver are provided in the SM. + +With scale invariant depth, we have an 8-DOF problem w.r.t. $\{s,\mathbf{R},\mathbf{t},f\}$ . We need at least two 3D-3D point correspondences and one 3D-2D point correspondence. The problem can be solved similarly as the calibrated case shown in Sec. 3.1. Due to the space limits, the details are shown in the SM, where we show a general solution to all problem configurations using three point correspondences. + +# Different and Unknown Focal Lengths + +Finally, we consider the case where $\mathbf{K}_i = \mathrm{diag}(f_i,f_i,1)$ , $i = 1,2$ , with $f_{1}\neq f_{2}$ . This is a 11-DOF problem w.r.t. $\{s,u,v,\mathbf{R},\mathbf{t},f_1,f_2\}$ . Thus, we need at least three 3D-3D point correspondences and one 3D-2D point correspondence, i.e., a correspondence where we have monocular depth only in one image. Similarly to the equal and unknown focal length case, we can formulate this problem using a system of 6 equations in 6 unknowns with up to 14 real solutions and solve it using [40]. + +However, a more efficient solver can be obtained using four 3D-3D point correspondences, which yields six equations in five unknowns: $s, u, v, f_1, f_2$ . This is an overconstrained system that can again be solved by using only + +five of these six equations. The five equations can be rewritten as $\mathbf{M}[1, c, f_1, u f_1, u^2 f_1, c f_2, c v f_2, c v^2 f_2]^\top = 0$ , where $\mathbf{M}$ is a $5 \times 8$ coefficient matrix. Here, $f_2$ always appears together with $c$ , thus, we let $c f_2 = \tilde{f}_2$ to simplify the polynomials. We obtain a solver that performs the G-J elimination of a $20 \times 24$ matrix and results in up to four solutions. We denote this solver as $4\mathrm{PT}_{suv}f_{1,2}$ . Alternatively, we can use all six equations to derive a smaller solver, which is faster, but more sensitive to noise (see SM for more details). + +With scale-invariant depth, we have a 9-DOF problem w.r.t. $\{s,\mathbf{R},\mathbf{t},f_1,f_2\}$ . We need at least three 3D-3D point correspondences. This problem can be solved similarly as the calibrated case shown in Sec. 3.1. Due to space limits, the details on this $3\mathrm{PT}_{s00}f_{1,2}$ solver are shown in the SM. + +# 4. Experiments + +In this section, we present a comprehensive comparison of different depth-aware and point-based solvers for calibrated, shared focal length, and different focal length cases, under different conditions, i.e., different sources of depth information, different feature detectors and matchers, different type of datasets (interior/exterior), and different optimization schemes inside RANSAC. + +Solvers. (1) For calibrated relative pose estimation, we compare our $3\mathrm{PT}_{\mathrm{suv}}$ (ours) solver with the solver $3\mathrm{PT}_{\mathrm{suv}}(\mathbf{M})$ from [69], the Rel3PT solver [2], the P3P solver [18], and the 5PT point-based solver [51]. (2) For the unknown shared focal length case, we compared the proposed $3\mathrm{PT}_{s00}f$ (ours) and $4\mathrm{PT}_{\mathrm{suv}}f$ (ours) solvers with the $4\mathrm{PT}_{\mathrm{suv}}f(\mathbf{M})$ solver from [69], the 3p3d solver [20], and the 6PT point-based solver [40]. (3) For the different focal length case, we compare our $3\mathrm{PT}_{s00}f_{1,2}$ (ours) and $4\mathrm{PT}_{\mathrm{suv}}f_{1,2}$ (ours) solvers with the $4\mathrm{PT}_{\mathrm{suv}}f_{1,2}(\mathbf{M})$ solver from [69], the 4p4d solver [20], and the 7PT solver [29]. + +Datasets. In order to test the proposed solvers on real-world data, we choose the Phototourism [32], the ETH3D [58], and the ScanNet [13] datasets. The images of the Phototourism dataset were collected from multiple cameras obtained at different times, from different viewpoints, and with occlusions. It is a challenging outdoor dataset that is commonly used as a benchmark dataset [32] for camera geometry methods. We used five test scenes with 24,750 image pairs from this dataset. The ETH3D multi-view stereo and 3D reconstruction benchmark [58] includes a range of indoor and outdoor scenes. Ground truth geometry was acquired using a high-precision laser scanner. Images were captured with a DSLR camera and a synchronized multicamera rig featuring cameras with varying fields of view. In total, 4,144 image pairs were used from the ETH3D dataset, and we used triangulation to obtain the ground truth depths. The ScanNet dataset is an indoor RGB-D video dataset containing 2.5 million views in more than 1,500 scans, an + +notated with camera poses and 3D surface reconstructions. We followed the evaluation setup in [56], and used 1,500 pairs from ScanNet for testing. + +Feature Detection and Matching. We used three different image features, the first is the popular sparse SuperPoint (SP) features [14], widely used in SfM and localization pipelines [55]. For SuperPoint matching, we employ LightGlue (LG) [44], a state-of-the-art deep learning-based method that demonstrates significant improvements over traditional matching methods. The second is RoMa [21], a state-of-the-art dense matcher. The last is Mast3r [41], where feature matching is considered as a 3D task with DUSt3R [65], a recent and powerful 3D reconstruction framework based on Transformers. + +Depth Estimation. We use five state-of-the-art MDE methods to obtain depth data, including Depth Anything v2 (DA v2) [68], UniDepth [53], MoGe [64], and Mast3r [41]. Experiments for MiDas [8] are provided in the SM. These MDE methods represent scale/affine-invariant [8, 41, 64, 68] as well as metric depths [53]. + +Robust Estimation Frameworks. We evaluated all solvers using the LO-RANSAC framework [12] within PoseLib [39] with either the Sampson (2 px threshold) or the reprojection error (16 px threshold) for LO and scoring, and a fixed number of 1,000 RANSAC iterations. In the main paper we show the results for the better scoring function for a given solver. Additional experiments are provided in the SM. We also evaluated selected solvers using the Hybrid RANSAC framework presented in [69] with 2 px Sampson error and 16 px reprojection error thresholds with equal weight for both types of error and fixed 1,000 iterations. For Mast3r [41], we use their non-linear optimization with 500 iterations for coarse refinement and 200 iterations for fine refinement. We observe that Mast3r performs better in our evaluation compared to the results reported in [69]. We attribute this potentially to using more refinement iterations, which in our experience improves performance. + +Evaluation Metrics. In our experiments, we report the median pose error $\epsilon$ calculated as the maximum of rotation and translation errors in degrees. We also report the mean average accuracy (mAA) [32] with a threshold of 10 degrees. For the focal length solvers we use $f_{err} = |f_{est} - f_{gt}| / f_{gt}$ as the focal length error for one camera. When two cameras with different focal lengths are considered, we use the geometric mean of their errors. We report the median focal length errors denoted as $\epsilon_f$ . We also report the mean average accuracy for focal lengths [35] at the $10\%$ error threshold, denoted as $\mathrm{mAA}_f$ . Run-times $\tau$ are reported in milliseconds. Run-times for all methods except Mast3r are reported per one CPU of an Intel Xeon Gold 6338. The run-times of Mast3r are reported on the CPU of an Intel I7-11700KF with an RTX 3090 GPU. The run-times of all the methods do not include feature matching and depth es + +
DepthSolverOpt.SP+LG [14, 44]RoMA [21]
ε(°) ↓mAA↑τ(ms) ↓ε(°) ↓mAA ↑τ(ms) ↓
-5PT [51]S0.9187.6748.140.5691.10184.36
Real DepthRel3PT [2]S0.8888.21103.180.5291.38532.02
P3P [18]S0.8388.8829.680.5291.33141.39
3PTsuv(M) [69]S0.7988.5541.810.4591.39145.59
3PTsuv(ours)S0.8088.6029.590.4791.37127.81
3PTsuv(M) [69]H [69]0.5291.39549.590.3992.731505.19
3PTsuv(ours)H [69]0.5291.42543.480.3992.721490.93
DA v2 [68]Rel3PT [2]S5.5768.4635.743.9080.56145.85
P3P [18]S0.9086.2523.260.7290.6193.99
3PTsuv(M) [69]S0.9087.1932.500.5690.9988.60
3PTsuv(ours)S0.9187.0721.680.5691.0175.87
3PTsuv(M) [69]H [69]0.9785.52605.630.5490.621493.75
3PTsuv(ours)H [69]0.9885.56593.950.5490.621477.73
MoGe [64]Rel3PT [2]S4.7472.0842.192.7482.04170.29
P3P [18]S0.9187.6725.720.5491.16111.74
3PTsuv(M) [69]S0.8987.7133.450.5391.0498.20
3PTsuv(ours)S0.8987.6722.410.5491.0584.24
3PTsuv(M) [69]H [69]0.8688.26566.160.5091.171414.96
3PTsuv(ours)H [69]0.8588.24554.790.4991.231401.61
UniDepth [53]Rel3PT [2]S1.3678.8249.700.7088.25207.86
P3P [18]S0.8888.0025.930.5691.11112.65
3PTsuv(M) [69]S0.9487.4233.900.5591.0497.85
3PTsuv(ours)S0.9587.4922.580.5591.0183.72
3PTsuv(M) [69]H [69]0.8688.03558.610.5391.331402.46
3PTsuv(ours)H [69]0.8688.03550.740.5391.331392.71
+ +
DepthSolverOpt.Mast3r [41]
ε(°) ↓mAA ↑τ(ms) ↓
-5PT [51]S0.6690.29126.77
Mast3r [41]Rel3PT [2]S0.6790.30104.46
P3P [18]S0.6790.2456.52
3PTsuv(M) [69]S0.6790.2042.49
3PTsuv(ours)S0.6790.3530.90
3PTsuv(M) [69]H [69]0.9287.962647.02
3PTsuv(ours)H [69]0.9287.982635.27
+ +Table 3. Comparison of different methods on the ETH3D dataset [58] for the calibrated case. Opt.: S - PoseLib [39] implementation using Sampson error, H - hybrid RANSAC from [69]. + +timation since they are equal for all depth-based solvers. For a fair comparison, for Mast3r, we only measure the run-times of the non-linear optimization, i.e., time without feature matching, and initial depth estimation. For datasets with multiple scenes, we report the mean values across the scenes. + +# 4.1. Calibrated Relative Pose Estimation + +Table 3 shows the comparison of different methods on the ETH3D dataset. The results for the Phototourism and ScanNet datasets are provided in the SM. The results show that using depth in pose estimation results in more accurate poses when good depth estimates (e.g. MoGe [64], UniDepth [53]) are available. In this case, the overall best-performing methods use the hybrid RANSAC proposed in [69], with our solver $3\mathrm{PT}_{\mathrm{su}}v$ (ours) slightly outperforming $3\mathrm{PT}_{\mathrm{su}}v(\mathbf{M})$ [69]. We note that the use of the hybrid RANSAC scheme comes at a significant increase in computational cost over the use of PoseLib [39]. As the authors of [69] point out, their implementation of the hybrid RANSAC scheme, which we use in our experiments, could be optimized for computational efficiency. However, even with optimizations the method requires significantly more computation per one iteration than standard RANSAC. + +Based on the results for PoseLib [39], when good depth + +is available, P3P [18] and $3\mathrm{PT}_{\mathrm{suv}}$ have similar performance. When less accurate depth estimates are available, 5PT performs best especially with good matches (e.g., Mast3r [41], RoMA [21]). + +# 4.2. Shared Unknown Focal Length + +Table 4 shows the comparison of the methods on the ETH3D dataset [58] for two cameras with a shared unknown focal length. When a good depth estimate is available, the best results are obtained utilizing the hybrid RANSAC scheme from [69]. We show that our solver $3\mathrm{PT}_{s00}f$ outperforms $4\mathrm{PT}_{suv}f(\mathbf{M})$ [69] for many combinations of MDEs and matches despite not modeling shift. Our $3\mathrm{PT}_{s00}f$ solver also performs best when comparing solvers incorporated within PoseLib [39] and used in combination with MoGe [64], UniDepth [53], and Mast3r [41]. For DA v2 [68] and MiDas [8] (results provided in SM) modeling shift seems to be beneficial, and we show that our $4\mathrm{PT}_{suv}f$ solver performs on par with $4\mathrm{PT}_{suv}f(\mathbf{M})$ while being $2x$ faster. Our observation that modeling an unknown shift is not always necessary is in contradiction to the observation made in [69]. The reason is that authors in [69] have not tested a scale-invariant 3PT solver and have not incorporated other depth-aware solvers into the hybrid RANSAC scheme. We also note that despite performing expensive non-linear optimization to obtain the poses, Mast3r [41] performs worse than approaches based on RANSAC when either good depth (UniDepth [53], MoGe [64]) or matches (RoMA [21]) are available. + +A comparison of the speed-accuracy trade-off of various methods is shown in Figure 2 for the combination of RoMA [21] and UniDepth [53]. The figure shows that when fast computation is necessary, our solver $3\mathrm{PT}_{s00}f$ implemented in PoseLib [39] performs the best. However, in general, the best accuracy is achieved using the same solver within the hybrid RANSAC scheme from [69]. + +# 4.3. Different Focal Lengths + +Table 5 shows the comparison of the methods on the Phototourism dataset for two cameras with different unknown focal lengths. As for the previous two cases, when good depth estimates are available the best results are obtained with the hybrid RANSAC scheme [69] in which our solver $3\mathrm{PT}_{s00}f_{1,2}$ outperforms $4\mathrm{PT}_{suv}f_{1,2}(\mathbf{M})$ [69], showing that it is not necessary to model shift for good performance. This is also confirmed by the experiments in PoseLib [39] which show that $3\mathrm{PT}_{s00}f$ outperforms other depth-based solvers across all combinations of matches and MDEs in terms of pose accuracy while also being significantly faster than solvers which incorporate shift. Figure 3 shows that $3\mathrm{PT}_{s00}f$ clearly outperforms the alternatives in terms of speed and pose estimation accuracy. The results for this case also show that when good depth estimates are + +
MatchesDepthSolverOptETH
ε(°)↓εf(°)↓mAAT↑mAAf↑τ(ms)↓
SP+LG [14, 44]-6PT [40]S2.450.0475.5761.5280.02
Real Depth3p3d [20]S2.060.0478.0062.8330.70
4PTsucf(M) [69]S1.830.0378.8663.71112.34
4PTsucf(ours)S1.720.0378.9063.5651.10
3PTs00f(ours)S1.750.0379.1763.6325.11
4PTsucf(M) [69]H [69]1.070.0282.0175.631502.02
3PTs00f(ours)H [69]1.070.0282.1975.761411.05
DA v2 [68]3p3d [20]S4.540.0861.6048.7627.01
4PTsucf(M) [69]S2.100.0474.6060.77104.75
4PTsucf(ours)S2.020.0474.6760.4146.62
3PTs00f(ours)S2.200.0473.6660.9321.62
4PTsucf(M) [69]H [69]2.340.0570.0758.001190.89
3PTs00f(ours)H [69]2.330.0570.1658.311216.89
MoGe [64]3p3d [20]S3.180.0668.2454.6427.70
4PTsucf(M) [69]S2.100.0476.0661.81107.30
4PTsucf(ours)S2.150.0475.5960.8857.91
3PTs00f(ours)S1.990.0476.9462.6624.72
4PTsucf(M) [69]H [69]1.500.0379.2366.34956.33
3PTs00f(ours)H [69]1.410.0380.2467.42967.52
UniDepth [53]3p3d [20]S3.490.0769.4755.5727.57
4PTsucf(M) [69]S2.210.0475.6161.37106.73
4PTsucf(ours)S1.920.0476.1862.0046.97
3PTs00f(ours)S2.040.0476.8962.4223.47
4PTsucf(M) [69]H [69]1.270.0381.6869.641107.04
3PTs00f(ours)H [69]1.280.0381.9969.281149.25
RoMA [21]-6PT [40]S1.150.0285.2375.03147.48
Real Depth3p3d [20]S0.930.0285.9874.98113.30
4PTsucf(M) [69]S0.990.0286.3775.21167.49
4PTsucf(ours)S1.030.0286.2575.04107.05
3PTs00f(ours)S0.990.0286.6874.9979.96
4PTsucf(M) [69]H [69]2.120.0276.6972.513218.22
3PTs00f(ours)H [69]2.280.0276.4272.433067.57
DA v2 [68]3p3d [20]S2.070.0278.8668.2991.39
4PTsucf(M) [69]S1.210.0285.3375.09146.35
4PTsucf(ours)S1.200.0285.6175.4388.03
3PTs00f(ours)S1.230.0284.8975.0067.61
4PTsucf(M) [69]H [69]1.280.0381.5668.722161.36
3PTs00f(ours)H [69]1.300.0381.5768.682245.95
MoGe [64]3p3d [20]S1.560.0280.9870.1998.44
4PTsucf(M) [69]S1.020.0285.9175.61151.65
4PTsucf(ours)S1.100.0285.8575.8092.92
3PTs00f(ours)S1.050.0286.0475.8375.12
4PTsucf(M) [69]H [69]0.890.0286.9976.661923.31
3PTs00f(ours)H [69]0.910.0287.2076.502043.91
UniDepth [53]3p3d [20]S1.890.0282.5071.8997.99
4PTsucf(M) [69]S1.120.0285.6375.66151.31
4PTsucf(ours)S1.120.0285.5975.6491.92
3PTs00f(ours)S1.040.0285.7875.6975.62
4PTsucf(M) [69]H [69]0.830.0287.3077.341997.88
3PTs00f(ours)H [69]0.820.0287.5377.792133.46
-6PT [40]S1.230.0382.9969.0085.89
Mast3r [41]Mast3r [41]3p3d [20]S1.370.0381.3067.3138.96
4PTsucf(M) [69]S1.580.0379.6866.0595.89
4PTsucf(ours)S1.360.0380.8667.1648.36
3PTs00f(ours)S1.340.0382.0968.2634.54
4PTsucf(M) [69]H [69]2.430.0572.2858.323825.11
3PTs00f(ours)H [69]2.390.0472.4558.614079.21
-M [41]1.320.0185.6482.954800.37
+ +Table 4. Comparison of different methods on the ETH3D dataset [58] for the equal and unknown focal length case. Opt.: S - PoseLib [39] implementation using Sampson error, H - hybrid RANSAC from [69], M - non-linear optimization used in [41]. + +available it is possible to obtain significantly better results by using reprojection errors. This is due to the fact that the problem has a very commonly occurring degeneracy when only point correspondences are used [33]. + +Additional Experiments. We provide additional experiments in the SM, which, in addition to results on additional datasets, also provide an evaluation of alternative scoring and LO strategies within PoseLib, as well as the proposed $3\mathrm{PT}_{100}f$ and $3\mathrm{PT}_{100}f_{1,2}$ solvers. We also show that it is possible to use Mast3r [41] for matches in combination with MDE to improve estimated poses and intrinsics. + +Limitations. Although using depth maps does not add computational overhead in pose estimation, obtaining these depth maps using learning-based methods introduces additional costs. This is usually a few milliseconds per im + +![](images/ae4dbc56331ded637dca116741cd8ad41cd5d7b0f46672b6e5beca022d126ef7.jpg) +Figure 2. Speed-accuracy evaluation for the case of two cameras with shared unknown focal length using RoMA matches [21] and UniDepth [53] depths on the ETH3D dataset [58]. We evaluated mAA and runtimes $(\tau)$ for each method running for 50, 100, 200, 500, and 1,000 iterations, with the exception of Mast3r [41], which ran with a maximum of 500 iterations. + +![](images/03f9cd9525b8710ebdce6d8400fd6f3bfa506caf661281720d49db40139bdbcb.jpg) +Figure 3. Speed-accuracy evaluation for the different unknown focal lengths case using SP+LG matches [14, 44] and MoGe [64] depths on the Phototourism dataset [32]. We evaluated mAA and runtimes $(\tau)$ for each method running for 50, 100, 200, 500, and 1,000 iterations within PoseLib [39]. + +age [69]. However, in many applications, MDE methods are used for multiple tasks. Thus, the computational overhead is not solely attributable to the pose estimation method. + +Compared to standard point-based focal length solvers, the proposed solvers encounter fewer degenerate cases. The formulation allows us to handle pure rotation effectively. However, there are still several degenerate cases similar to those in point-based solvers; for example, four coplanar points and cases of pure translation can introduce degeneracies in focal length recovery [33]. + +# 5. Conclusion + +We address the problem of estimating the relative pose of two cameras using monocular depth predictions. We propose novel solvers that jointly estimate the scale and shift (or only one of them), and the relative pose. We consider solvers for calibrated cameras, and cameras with shared un + +
MatchesDepthSolverOptPhototourism
ε(°)↓εf(°)↓mAA↑mAAf↑τ(ms)↓
-7PT [29]S8.030.1738.1423.7824.80
Real Depth4pd4 [20]S7.580.1639.3424.6523.69
4PTsuvf1,2(M) [69]R2.270.0369.1763.33106.32
4PTsuvf1,2(ours)R2.280.0369.1463.3124.35
3PTsuvf1,2(ours)R2.230.0369.5063.6612.05
4PTsuvf1,2(M) [69]H [69]2.860.0463.8853.461922.12
4pd4 [20]H [69]2.890.0563.2152.871921.03
3PTsuvf1,2(ours)H [69]2.820.0464.1953.731901.23
SP+LG [14, 44]4pd4 [20]S10.400.2034.7222.0822.02
4PTsuvf1,2(M) [69]S8.170.1837.6223.32113.20
4PTsuvf1,2(ours)S8.190.1837.6223.2733.14
3PTsuvf1,2(ours)S7.660.1638.9023.5720.44
4PTsuvf1,2(M) [69]H [69]9.550.2132.8820.882490.39
4pd4 [20]H [69]9.890.2232.1420.412402.27
3PTsuvf1,2(ours)H [69]9.480.2132.4920.652494.66
MoGe [64]4pd4 [20]S7.710.1639.1224.4422.82
4PTsuvf1,2(M) [69]R6.600.0847.1639.0698.69
4PTsuvf1,2(ours)R6.710.0846.9838.9631.64
3PTsuvf1,2(ours)R5.730.0748.5940.4110.22
4PTsuvf1,2(M) [69]H [69]3.880.0757.2142.801953.23
4pd4 [20]H [69]4.190.0755.7641.761881.29
3PTsuvf1,2(ours)H [69]3.830.0757.6243.252031.28
UniDepth [53]4pd4 [20]S7.670.1639.1824.6123.63
4PTsuvf1,2(M) [69]R4.900.0752.6443.12106.30
4PTsuvf1,2(ours)R4.890.0752.5943.1224.46
3PTsuvf1,2(ours)R4.820.0752.8543.2812.89
4PTsuvf1,2(M) [69]H [69]3.380.0660.1547.021945.50
4pd4 [20]H [69]3.450.0659.4946.631940.62
3PTsuvf1,2(ours)H [69]3.350.0660.3847.241957.65
-7PT [29]S4.300.1053.1634.7375.10
Real Depth [68]4pd4 [20]S4.240.1053.5135.0177.49
4PTsuvf1,2(M) [69]R1.250.0279.7573.91136.12
4PTsuvf1,2(ours)R1.260.0279.6873.8454.16
3PTsuvf1,2(ours)R1.210.0280.3774.4948.02
4PTsuvf1,2(M) [69]H [69]1.600.0375.1064.073518.89
4pd4 [20]H [69]1.630.0374.6363.623199.77
3PTsuvf1,2(ours)H [69]1.580.0375.4164.373531.28
DA v2 [68]4pd4 [20]S4.610.1151.2433.3073.93
4PTsuvf1,2(M) [69]S4.430.1052.9034.36170.68
4PTsuvf1,2(ours)S4.420.1052.8334.3590.36
3PTsuvf1,2(ours)S4.200.1053.9534.5681.59
4PTsuvf1,2(M) [69]H [69]6.530.1643.2527.714026.79
4pd4 [20]H [69]6.560.1643.2127.743676.08
3PTsuvf1,2(ours)H [69]6.520.1643.3027.794113.61
MoGe [64]4pd4 [20]S4.220.1053.6134.9977.14
4PTsuvf1,2(M) [69]R4.870.0754.2143.35128.82
4PTsuvf1,2(ours)R4.900.0754.0543.3448.16
3PTsuvf1,2(ours)R4.400.0655.3844.3346.38
4PTsuvf1,2(M) [69]H [69]2.410.0569.0852.393663.43
4pd4 [20]H [69]2.440.0568.7552.183330.63
3PTsuvf1,2(ours)H [69]2.390.0569.2952.573803.60
UniDepth [53]4pd4 [20]S4.200.1053.8435.2477.78
4PTsuvf1,2(M) [69]R4.200.0656.7647.14135.10
4PTsuvf1,2(ours)R4.200.0656.7247.1053.40
3PTsuvf1,2(ours)R4.060.0657.1647.3254.73
4PTsuvf1,2(M) [69]H [69]2.300.0469.9553.903440.95
4pd4 [20]H [69]2.310.0469.6753.643114.21
3PTsuvf1,2(ours)H [69]2.290.0470.1854.093533.26
Mast3r [41]-7PT [29]S4.390.0954.0139.53
Mast3r [41]4pd4 [20]S5.200.1149.6936.90
4PTsuvf1,2(M) [69]S7.020.1443.9527.62108.64
4PTsuvf1,2(ours)S7.030.1444.1027.8144.88
3PTsuvf1,2(ours)S5.400.1249.2530.8837.40
4PTsuvf1,2(M) [69]H [69]10.110.2435.1620.564871.24
4pd4 [20]H [69]32.030.761.332.496492.27
3PTsuvf1,2(ours)H [69]10.130.2535.1720.514980.42
-M [41]2.710.0466.5456.434903.10
+ +Table 5. Comparison of different methods on the Phototourism dataset [32] for the two unknown focal lengths case. Opt.: S/R - PoseLib [39] implementation using Sampson error (S) and reprojection error (R), H - hybrid RANSAC from [69], M - non-linear optimization used in [41]. + +known or different unknown focal lengths. Efficient solvers that outperform state-of-the-art depth-aware solvers are proposed for all three cases. In extensive experiments, we discuss which solvers are preferable in different situations, e.g., for precise or imprecise depths. We find that when depth estimates are accurate, modeling only scale tends to yield better performance. In such cases, it's also more effective to use reprojection error for the focal length problems. + +Acknowledgments. This work was supported by the Czech Science Foundation (GACR) JUNIOR STAR, Grant No. 22-23183M, the Student Grant Competition of the Czech Technical University in Prague, Grant No. SGS23/173/OHK3/3T/13, the EU NextGenerationEU through the Recovery and Resilience Plan for Slovakia under the project "InnovAite Slovakia, Illuminating Pathways for AI-Driven Breakthroughs" No. 09I02-03-V01-00029, and the NSFC under Grant Nos. 62361166670 and U24A20330. The access to the computational infrastructure of the OP VVV funded project CZ.02.1.01/0.0/0.0/16_019/0000765 "Research Center for Informatics" is also gratefully acknowledged. Part of the results was obtained using the computational resources procured in the project National competence centre for high performance computing (project code: 311070AKF2) funded by European Regional Development Fund, EU Structural Funds Informatization of society, Operational Program Integrated Infrastructure. + +# References + +[1] Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres Solver, 2023. 2 +[2] Jonathan Astermark, Yaqing Ding, Viktor Larsson, and Anders Heyden. Fast relative pose estimation using relative depth. In International Conference on 3D Vision (3DV), 2024. 2, 4, 5, 6 +[3] Daniel Barath and Zuzana Kukelova. Relative pose from sift features. In European Conference on Computer Vision, pages 454-469. Springer, 2022. 1 +[4] Daniel Barath and Chris Sweeney. Relative pose solvers using monocular depth. In 2022 26th International Conference on Pattern Recognition (ICPR), 2022. 2 +[5] Daniel Barath, Tekla Toth, and Levente Hajder. A minimal solution for two-view focal-length estimation using two affine correspondences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1 +[6] Jacob Bentolila and Joseph M Marcos. Conic epipolar constraints from affine correspondences. Computer Vision and Image Understanding, 2014. 1 +[7] 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. 2, 3 +[8] Reiner Birkl, Diana Wofk, and Matthias Müller. Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460, 2023. 2, 3, 6, 7 +[9] 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 preprint arXiv:2410.02073, 2024. 1, 2, 3 +[10] Shen Cai, Zhanhao Wu, Lingxi Guo, Jiachun Wang, Siyu Zhang, Junchi Yan, and Shuhan Shen. Fast and interpretable + +2d homography decomposition: Similarity-kernel-similarity and affine-core-affine transformations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1 +[11] Federico Camposeco, Andrea Cohen, Marc Pollefeys, and Torsten Sattler. Hybrid Camera Pose Estimation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3 +[12] Ondrej Chum, Jiří Matas, and Josef Kittler. Locally optimized ransac. In Pattern Recognition: 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003. Proceedings 25, pages 236-243. Springer, 2003. 6 +[13] 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 +[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, 2018. 6, 7, 8 +[15] Yaqing Ding, Jian Yang, Jean Ponce, and Hui Kong. An efficient solution to the homography-based relative pose problem with a common reference direction. In International Conference on Computer Vision (ICCV), 2019. 1 +[16] Yaqing Ding, Daniel Barath, and Zuzana Kukelova. Minimal solutions for panoramic stitching given gravity prior. In International Conference on Computer Vision (ICCV), pages 5579-5588, 2021. +[17] Yaqing Ding, Jian Yang, Jean Ponce, and Hui Kong. Homography-based minimal-case relative pose estimation with known gravity direction. Trans. Pattern Analysis and Machine Intelligence (PAMI), 2022. 1 +[18] Yaqing Ding, Jian Yang, Viktor Larsson, Carl Olsson, and Kalle Åström. Revisiting the p3p problem. In Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4, 5, 6, 7 +[19] Yaqing Ding, Jonathan Astermark, Magnus Oskarsson, and Viktor Larsson. Noisy one-point homographies are surprisingly good. In Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[20] Yaqing Ding, Václav Várvá, Snehal Bhayani, Qianliang Wu, Jian Yang, and Zuzana Kukelova. Fundamental matrix estimation using relative depths. In European Conference on Computer Vision (ECCV), 2024. 2, 4, 5, 7, 8 +[21] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6, 7, 8 +[22] Ivan Eichhardt and Daniel Barath. Relative pose from deep learned depth and a single affine correspondence. In European Conference on Computer Vision (ECCV), 2020. 1 +[23] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 1, 4 + +[24] Friedrich Fraundorfer, Petri Tanskanen, and Marc Pollefeys. A minimal case solution to the calibrated relative pose problem for the case of two known orientation angles. In European Conference on Computer Vision (ECCV), 2010. 1 +[25] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE transactions on pattern analysis and machine intelligence, 2003. 3 +[26] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In International Conference on Computer Vision (ICCV), 2019. 1 +[27] Banglei Guan and Ji Zhao. Relative pose estimation for multi-camera systems from point correspondences with scale ratio. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 1, 2 +[28] Richard Hartley and Hongdong Li. An efficient hidden variable approach to minimal-case camera motion estimation. Trans. Pattern Analysis and Machine Intelligence (PAMI), 2012. 1, 4, 8 +[29] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 2, 4, 5, 8 +[30] 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. 1, 2, 3 +[31] 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. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2 +[32] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. International Journal of Computer Vision, 2020. 5, 6, 8 +[33] Fredrik Kahl and Bill Triggs. Critical motions in euclidean structure from motion. In Computer Vision and Pattern Recognition (CVPR), 1999. 7, 8 +[34] 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, 3 +[35] Viktor Kocur, Daniel Kyselica, and Zuzana Kukelova. Robust self-calibration of focal lengths from the fundamental matrix. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5220-5229, 2024. 6 +[36] Zuzana Kukelova, Martin Bujnak, and Tomas Pajdla. Closed-form solutions to minimal absolute pose problems with known vertical direction. In Asian Conference on Computer Vision (ACCV), 2010. 1 +[37] Zuzana Kukelova, Martin Bujnak, and Tomas Pajdla. Polynomial eigenvalue solutions to minimal problems in com + +puter vision. Trans. Pattern Analysis and Machine Intelligence (PAMI), 2012. 1 +[38] Zuzana Kukelova, Joe Kileel, Bernd Sturmfels, and Tomas Pajdla. A clever elimination strategy for efficient minimal solvers. In Computer Vision and Pattern Recognition (CVPR), 2017. 1 +[39] Viktor Larsson and contributors. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 6, 7, 8 +[40] Viktor Larsson, Kalle Åström, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Computer Vision and Pattern Recognition (CVPR), 2017. 5, 7 +[41] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision. Springer, 2024. 3, 6, 7, 8 +[42] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 2 +[43] Hyon Lim, Sudipta N. Sinha, Michael F. Cohen, Matt Uytendaele, and H. Jin Kim. Real-time Monocular Image-based 6-DoF Localization. *IJRR*, 34(4-5):476-492, 2015. 1 +[44] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In International Conference on Computer Vision (ICCV), 2023. 6, 7, 8 +[45] Sheng Liu, Xiaohan Nie, and Raffay Hamid. Depth-guided sparse structure-from-motion for movies and tv shows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[46] Stephan Liwicki and Christopher Zach. Scale exploiting minimal solvers for relative pose with calibrated cameras. In BMVC, 2017. 1, 2 +[47] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004. 2 +[48] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think, 2024. 1, 2, 3 +[49] Nathaniel Merrill, Patrick Geneva, Saimouli Katragadda, Chuchu Chen, and Guoquan Huang. Fast monocular visual-inertial initialization leveraging learned single-view depth. Robotics: Science and Systems (RSS) 2023, 2023. 2 +[50] Oleg Naroditsky, Xun S Zhou, Jean Gallier, Stergios I Roumeliotis, and Kostas Daniilidis. Two efficient solutions for visual odometry using directional correspondence. Trans. Pattern Analysis and Machine Intelligence (PAMI), 2012. 1 +[51] David Nister. An efficient solution to the five-point relative pose problem. Trans. Pattern Analysis and Machine Intelligence (PAMI), 2004. 1, 2, 3, 4, 5, 6 +[52] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (p3p) solver. In Proceedings of the European conference on computer vision (ECCV), pages 318-332, 2018. 3, 4 + +[53] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 1, 2, 3, 6, 7, 8 +[54] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In International Conference on Computer Vision (ICCV), 2021. 2, 3 +[55] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Computer Vision and Pattern Recognition (CVPR), 2019. 1, 6 +[56] 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, 2020. 6 +[57] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016. 1 +[58] 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 Computer Vision and Pattern Recognition (CVPR), 2017. 5, 6, 7, 8 +[59] Henrik Stewénius, David Nister, Fredrik Kahl, and Frederik Schaffalitzky. A minimal solution for relative pose with unknown focal length. In Computer Vision and Pattern Recognition (CVPR), 2005. 1 +[60] Henrik Stewenius, Christopher Engels, and David Nistér. Recent developments on direct relative orientation. ISPRS Journal of Photogrammetry and Remote Sensing, 60(4):284-294, 2006. 1, 2 +[61] Linus Svärn, Olof Enqvist, Fredrik Kahl, and Magnus Oskarsson. City-scale localization for cameras with known vertical direction. Trans. Pattern Analysis and Machine Intelligence (PAMI), 2016. 1 +[62] Chris Sweeney, John Flynn, and Matthew Turk. Solving for relative pose with a partially known rotation is a quadratic eigenvalue problem. International Conference on 3D Vision (3DV), 2014. 1 +[63] Philip HS Torr and Andrew Zisserman. Mlesac: A new robust estimator with application to estimating image geometry. Computer vision and image understanding, 2000. 3 +[64] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 1, 2, 3, 6, 7, 8 +[65] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 6 +[66] Chen Yang, Sikuang Li, Jiemin Fang, Ruofan Liang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Gaussianobject: High-quality 3d object reconstruction from four views + +with gaussian splatting. ACM Transactions on Graphics (TOG), 2024. 2 +[67] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3 +[68] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 1, 2, 3, 6, 7, 8 +[69] Yifan Yu, Shaohui Liu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. Relative pose estimation through affine corrections of monocular depth priors. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16706-16716, 2025. 2, 3, 4, 5, 6, 7, 8 +[70] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 2022. 2 +[71] Bernhard Zeisl, Torsten Sattler, and Marc Pollefeys. Camera pose voting for large-scale image-based localization. In International Conference on Computer Vision (ICCV), 2015. 1 +[72] Xinyue Zhang, Liangzu Peng, Wanting Xu, and Laurent Kneip. Accelerating globally optimal consensus maximization in geometric vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4 +[73] Yunwen Zhou, Abhishek Kar, Eric Turner, Adarsh Kowdle, Chao X Guo, Ryan C DuToit, and Konstantine Tsotsos. Learned monocular depth priors in visual-inertial initialization. In European conference on computer vision, 2022. 2 \ No newline at end of file diff --git a/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/images.zip b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5b0d8ddba2d24b69ec8cfbd8c7a0617a2ad12e13 --- /dev/null +++ b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3412620fa51965301c0031ea37a769f8d3eab6287cea66aba4efed91729daa27 +size 557403 diff --git a/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/layout.json b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c2df4e8bb8c8d3910ec6a24a670e907f98e5e9b8 --- /dev/null +++ b/ICCV/2025/RePoseD_ Efficient Relative Pose Estimation With Known Depth Information/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c29ad7810213f13a10456a641ecf5f8662beec0d5acff617cd68355411840598 +size 459146 diff --git a/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/42118481-e852-4ebc-8501-2c156127aa2c_content_list.json b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/42118481-e852-4ebc-8501-2c156127aa2c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8a0ff73c3c95a3901e028e1bbb7d4a77c36ee7bd --- /dev/null +++ b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/42118481-e852-4ebc-8501-2c156127aa2c_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee370b8bb8c463b7231d05c6b23eb0fec1a070eb4e641f29d49b4c3a89e9bd40 +size 74849 diff --git a/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/42118481-e852-4ebc-8501-2c156127aa2c_model.json b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/42118481-e852-4ebc-8501-2c156127aa2c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..07b572b3126b24e2bb0de43bfc8a1b067686d950 --- /dev/null +++ b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/42118481-e852-4ebc-8501-2c156127aa2c_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:613b24e28768d98f4e699553dcc355a8b095e2e8f5b147a5654d5b04561650d3 +size 93774 diff --git a/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/42118481-e852-4ebc-8501-2c156127aa2c_origin.pdf b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/42118481-e852-4ebc-8501-2c156127aa2c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..da465aa660ceedcab6fdb27f4972a00806e2a5fe --- /dev/null +++ b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/42118481-e852-4ebc-8501-2c156127aa2c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88c6878cfa7cefd173b18c0b0817ae9e33b0a51a9f996b22c35bd9a148f362b1 +size 6261115 diff --git a/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/full.md b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..627259aa34ac0c89e337b912714cc96189383ddd --- /dev/null +++ b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/full.md @@ -0,0 +1,244 @@ +# ReTracker: Exploring Image Matching for Robust Online Any Point Tracking + +Dongli Tan $^{1,2*}$ Xingyi He $^{1,2*}$ Sida Peng $^{1}$ Yiqing Gong $^{1}$ Xing Zhu $^{2}$ +Jiaming Sun $^{1,3}$ Ruizhen Hu $^{4}$ Yujun Shen $^{2}$ Hujun Bao $^{1}$ Xiaowei Zhou $^{1\dagger}$ $^{1}$ Zhejiang University $^{2}$ Ant Research $^{3}$ 4DV.ai $^{4}$ Shenzhen University +*Equal contribution †Corresponding author + +![](images/9bc667e253d782a1c976dc04581234090aa456fc520af38ed9d7d166bf371396.jpg) +Figure 1. Comparison Between Existing Online Trackers and the Proposed ReTracker. Retracking target points in a video sequence after long-term occlusions is challenging due to significant scene changes, where current state-of-the-art online trackers often struggle. In contrast, the proposed ReTracker pipeline achieves robust and accurate continuous tracking of target points upon reappearance, effectively handling large viewpoint and appearance variations. + +![](images/9f574f169a59b57000bd828f3679ed278d681d2dee1f39c187821c278a584a69.jpg) + +![](images/4d1961c5e43b1f8148068e4fca325952782a95131d02cec0fb18030406d92ede.jpg) + +# Abstract + +This paper aims to establish correspondences for a set of 2D query points across a video sequence in an online manner. Recent methods leverage future frames to achieve smooth point tracking at the current frame, but they still struggle to find points with significant viewpoint changes after long-term occlusions and inherently cannot achieve online tracking. To overcome these challenges, we develop a novel online tracking framework, named ReTracker, that integrates two advances in image matching with tracking-specific designs. First, a decoder network with a global receptive field is incorporated with a temporal attention module to robustly track points undergoing large location changes. Second, the decoder network is adapted to pretrain on large-scale two-view matching data, which offers significantly greater diversity and volume than tracking data, to learn general matching priors. This pretraining strategy effectively enhances our tracker's ability to handle viewpoint and appearance varia + +tions after long-term occlusions. Experiments demonstrate that our method outperforms recent online trackers across multiple benchmarks and achieves competitive or superior performance compared to offline methods. Furthermore, we collect an ego-centric, occlusion-heavy dataset to illustrate the retracking capabilities of our approach. Project page: re-tracker.github.io. + +# 1. Introduction + +Point tracking is a fundamental problem in computer vision, which aims to find pixel correspondences of a set of 2D query points in a sequence of video frames. The sequentially tracked 2D point locations along the frames can be used in various downstream tasks, such as SLAM [4, 31, 32], object tracking [29, 38], video generation [21], robotic manipulation [16, 20, 51], etc. In this paper, we address the problem of online tracking, which aims to predict tracking results for each newly captured frame instantly, without relying on + +future frames. The immediate tracking without frame delay makes online tracking particularly valuable for latency-sensitive applications such as SLAM and robotic manipulation. + +Recent tracking approaches, such as CoTracker3 [22] and TAPTRv2 [25], achieve robust point tracking by a sliding window-based approach. For a sequence of new frames to be tracked, they first initialize the tracking locations by replicating the last tracked locations or by conducting global feature correlation between point features from the initial frame with the entire feature maps of the query frames to predict correspondences. Then, they apply iterative local refinement within a sequence window to gradually obtain temporally consistent and accurate tracking results, which works well in occlusion-free scenarios. + +However, this paradigm encounters two main challenges. First, it is difficult to maintain continuous tracking of target points when they reappear after prolonged occlusions, because the spatial locations of reappearing points, along with the surrounding scene context, may differ significantly from the frames in which they last appeared. Current approaches [9, 12], which rely solely on feature correlation using extracted features, perform poorly in such scenarios. Moreover, the scarcity of tracking training data for occlusion-heavy scenes is another factor leading to the poor performance of existing methods. The second challenge lies in the online tracking setting, as existing strong window-based approaches [22, 23, 25, 26] rely on information from multiple future frames to achieve robust and consistent tracking results, making them unsuitable for online settings where future information is inaccessible. + +In this paper, we propose a novel tracking framework, named ReTracker, that can reliably retrack points after long-term occlusion and achieve online tracking. Our key insight is to harness the advances in two-view matching for point tracking, considering its impressive progress [15, 19, 44, 54] on handling large viewpoint changes and appearance variations. To this end, we leverage the architecture of image matching methods, which is designed with a global receptive field for robust correspondence estimation. Additionally, we take advantage of their ability to train on large-scale two-view datasets, which offer significantly better diversity than tracking data. + +Specifically, we first extract features from both the initial and current frames. A memory attention module is designed to accumulate information from intermediate history frames to enhance the image features, which are then fed into a decoder network with the global receptive field to establish correspondences between the initial and current frames. To exploit the large-scale two-view data, we develop an approach to pretrain our decoder network on the image matching task, effectively learning the global matching priors. Subsequently, the decoder is jointly optimized + +with the temporal attention mechanism on the point tracking task to acquire temporal priors. Thanks to temporal feature enhancement, a decoder network with a global receptive field, and pre-training on two-view data, our method can effectively re-track points even after long-term occlusion and significant viewpoint changes. Additionally, as our approach does not rely on future frames for prediction, it naturally supports online tracking. + +The previous global tracking module uses low-resolution features for efficiency, limiting tracking accuracy. To address this, we propose an efficient local refinement module that enhances tracking precision. This module utilizes multiscale feature patches around initially tracked points, capturing rich contextual information. By iteratively regressing local offsets, the module progressively refines the initial tracking locations. To enable historical information propagation, we correlate history and current feature patches to construct feature volumes. These volumes are further refined through iterative temporal and spatial attention, producing the final features used to update tracking locations. + +Results show that our method achieves state-of-the-art performances on various public benchmarks, where the proposed online pipeline can achieve competitive or even better results compared with strong offline methods [11, 22]. To further evaluate and demonstrate the capability of our method on continued tracking after long-term occlusion, we annotate a new evaluation dataset that focuses on the egocentric occlusion-heavy scenes and is composed of 30 sequences with ground-truth annotations. Thanks to the design of our strong global tracking and local refinement pipeline with spatial and temporal attention for global context extraction, our online tracking method can accurately track points for challenging egocentric scenes. Fig. 1 presents some examples. + +In summary, this paper has the following contributions: + +- An online tracking framework that leverages the advances in two-view matching to enable robust retracking after long-term occlusions. +- A global tracking module and a local refinement module incorporating temporal and spatial attention to capture global context and handle significant scene changes effectively. +- A training strategy that leverages two-view matching priors learned from diverse datasets to enhance tracking performance. +- An ego-centric occlusion-heavy dataset with ground-truth tracking annotations. + +# 2. Related Work + +# 2.1. Tracking Any Points + +Early methods like PIPs [18] introduced reliable tracking within a sliding window but were limited by the window's + +duration. TAP-Vid [9] advanced the field by providing a new benchmark and a simple baseline, while TAPIR [12] improved tracking performance through a two-stage approach combining matching and refinement strategies. LocoTrack [6] extends TAPIR by incorporating local 4D correlation. Recently, BootsTAP [11] achieves impressive state-of-the-art performance by leveraging a bootstrapped training approach based on TAPIR using large-scale unlabeled video data. MFT [34] and MFTIQ [43] employ dense long-term correspondence methods with advanced chaining strategies. OmniMotion [53], Dino-tracker [49] and protracker [57] refined correspondences by leveraging test-time optimization. These kinds of methods require substantial computational resources and time during inference, limiting real-time applicability. Inspired by PIPs, CoTracker [23] proposes a transformer-based point tracker that jointly tracks a group number of 2D points in long video sequences by leveraging their dependencies. SpatialTracker [55] lifts 2D pixels to 3D space using monocular depth estimators and represents each frame with a triplane structure. TAPTR [26] proposes a DETR-like [1] framework for tracking arbitrary points, treating the task from a detection perspective. TAPTRv2 [25] improves TAPTR by eliminating the reliance on cost-volume. CoTracker3 [22] builds on CoTracker by further enhancing the tracking of multiple points in long video sequences. These methods [1, 22, 24-26, 35, 55, 58] employ a sliding window on long videos and optimize the predictions across all frames within the window jointly, thereby enabling robust prediction of positions and occlusions. However, since these methods require access to frames beyond the current one for prediction, they are not suitable for online tracking. + +# 2.2. Two-View Image Matching + +Image matching aims to establish accurate correspondences between image pairs with significant viewpoint and illumination variations. A typical feature matching pipeline [8, 13, 30, 41] involves detecting and describing keypoints in each image, followed by matching via nearest-neighbor search or learning-based matchers [28, 42]. In contrast to sparse matching methods that depend on keypoint detection, recent detector-free approaches [3, 5, 14, 15, 45, 46], which directly match dense image features, have exhibited impressive performances in challenging scenarios such as textureless scenes. Early methods [39, 40] utilize 4D correlation volumes to achieve robust matching. LoFTR [45] introduced transformer architectures [50] into detector-free matching to effectively model long-range dependencies and produce semi-dense correspondences through a coarse-to-fine strategy. Subsequent methods have further enhanced matching accuracy by leveraging multi-scale attention and flow guidance [3], or by constraining hierarchical attention to relevant spatial regions [46]. Recent dense match + +ing approaches [14, 15, 48], designed to estimate comprehensive correspondences between images, demonstrate remarkable robustness to large perspective variations. Furthermore, dense matching methods trained on large-scale datasets [19, 44] show outstanding generalization capabilities and improved performance across various tasks. The advancements in image matching methods motivate exploring them to address the challenges in online tracking. + +# 3. Method + +Given a video stream $\mathcal{V} = \{\mathbf{I}_0, \mathbf{I}_1, \dots, \mathbf{I}_t\}$ and a set of query points $\mathbf{Q} \in \mathbb{R}^2$ in the initial frame $\mathbf{I}_0$ , the proposed online point tracker predicts the corresponding 2D points $\mathbf{P} \in \mathbb{R}^2$ in the newly captured frame $\mathbf{I}_t$ , without access to future frames. Our method achieves this through a two-stage approach, consisting of a global tracking module (Sec. 3.1) that robustly identifies point locations combining tracking-specific designs and two-view matching advancements, and an iterative refinement module (Sec. 3.2) that further refines these tracking results to achieve high accuracy. An overview of our pipeline is shown in Fig. 2. + +# 3.1. Global Tracking with Matching Advancements + +Dense two-view matching methods [15, 19] have demonstrated impressive performance in establishing correspondences between images with large viewpoint and appearance changes. The capabilities are achieved through network architectures with global attention mechanisms to model long-range dependencies and enhance feature distinctiveness. Additionally, the models are trained on large-scale real-world image pairs with severe viewpoint changes, enabling them to learn strong matching priors in diverse scenes. Our insight is to adapt these advancements to the point tracking framework for solving the re-tracking problem after long-term occlusions. + +However, directly applying two-view matchers for tracking is not trivial, as sequential pairwise matching accumulates errors [47], leading to drift over time. Additionally, performing matching between the first frame and the current frame is hindered by substantial temporal scene changes, limiting tracking accuracy. + +Our global point tracker integrates temporal information from intermediate history frames to bridge appearance and scene context gaps using the proposed temporal attention mechanism. Correspondences between the initial and current frames are then established by a decoder network with a global receptive field. Furthermore, to leverage large-scale realistic two-view matching data, we introduce a pretraining strategy in which our network is first trained on the two-view image matching task to learn global matching priors. It is then further trained on tracking datasets to acquire temporal priors. + +![](images/c80b00330b6a8d03d4a96d14a2c0708f47a72c1b4a6684c0624aee969fd9bfd3.jpg) + +![](images/38c0911566e48ec9e737e8d067af463072d5beaf5496e6eaec1caf1b894e449c.jpg) +Figure 2. Overview of ReTracker. (a) The Global Point Tracking module establishes correspondences between the initial query frame and the current frame. For extracted dense image features, we enhance them with History Tokens to capture temporal dependencies from historical frames. The match distribution maps are then decoded with spatial attention to incorporate the overall scene context and enhance tracking robustness. Coarse tracking locations are extracted from the distribution maps. (b) The Multi-scale local refinement module further refines the coarse tracking locations for high accuracy. At each scale, local feature patches are cropped around the current tracking locations. 4D correlation and encoding operations are then performed between the current frame and a set of historical frames to encode local contextual relationships. The encoded tokens are then processed through spatial and temporal attentions to predict local offsets, occlusion and certainty. + +Concretely, the dense feature maps of initial frame $\mathbf{I}_0$ and current frame $\mathbf{I}_t$ are firstly extracted. In our global tracking phase, we utilize the DINOv2 [36] backbone to capture rich high-level contextual information. The extracted feature maps are flattened and subsequently enhanced through our temporal and spatial feature attention to facilitate the tracking point decoding. Notably, since the selected query points in the initial frame are often floating points instead of lying on grid points, we additionally extract their features using grid sampling on the extracted feature map. These sampled features are then concatenated with the flattened feature vectors of the initial frame, providing a more precise and informative representation for query points. + +Temporal Feature Transformation. Given the extracted features of the initial and current frames, we introduce a temporal transformation module to bridge the context gap between these two distant frames. This module leverages temporal features from intermediate history frames to enhance feature distinctiveness and improve tracking robustness. As we use dense image features for transformation and tracking point decoding, preforming cross-attention between + +the current frame features $\mathbf{F}_t$ with multiple history frame features $\mathbf{F}_{t - 1},\mathbf{F}_{t - 2},\ldots$ is computationally expensive, bar-ricading the latency-sensitive applications. + +To address this issue, we propose a recurrently updated memory token mechanism that maintains history tokens for both the initial and current frames. By directly attending to the initial and current frame features with history tokens, our approach effectively integrates temporal information and enhances the stability of prediction among frames. Specifically, we create an initial frame memory $\mathbf{M}_0$ and a current frame memory $\mathbf{M}_t$ , initializing them with the feature of the first frame $\mathbf{F}_0$ and the feature of the second frame $\mathbf{F}_1$ . These history tokens are then separately attended to the initial and current features through a cross-attention block, generating temporal residuals that are added to the original features to obtain enhanced features. After the subsequent spatial attention, the final features are applied to update the history tokens for the next frame. + +Spatial Feature Transformation and Token Update. With the previously enhanced temporal features, we utilize a spatial attention module to further improve feature distinct + +tiveness, inspired by two-view matching [15, 45]. This module attends to the entire image space, capturing global relationships between dense pixels in both the initial and current frames. Incorporating long-range dependencies strengthens the model's ability to handle challenging tracking scenarios, ensuring robust performance even under significant scene structure changes. Subsequently, the history tokens are updated with the enhanced features from the spatial attention module, which will be used to provide temporal information for the next frame. + +Tracking Points Prediction. Given the enhanced features from the preceding temporal and spatial attentions, we utilize a transformer decoder to generate a distribution map for each query point in the initial frame. The tracking points are then determined by identifying the locations corresponding to the maximum values in the distribution map. Specifically, for each query point, we predict a distribution map with a resolution of $64 \times 64$ , where the total classification anchor number $K = 4096$ . The transformer decoder predicts the flattened distribution map vector $D \in \mathbb{R}^{N \times K}$ , where $N$ is the number of query points and $K$ is the total number of anchors for classification. The predicted distribution maps are then reshaped back into 2D grids, and the tracking points are determined by applying the Argmax operator over them. + +# Global Matching Prior from Two-View Pretraining. + +Training tracking models to address the real-world retracking problem is challenging due to the scarcity of diverse real-world tracking data. On one hand, real-world tracking data is limited because annotating dynamic scenes is difficult, particularly in cases where significant scene context changes occur after long-term occlusions. On the other hand, synthetic data generated from physics simulations [17, 58] is widely used for training tracking models. However, the diversity of simulated object motions is constrained and cannot fully replicate real-world viewpoint variations and scene context dynamics. + +To overcome this limitation, we propose a two-view matching pretraining strategy, enabling the model to learn global matching priors from large-scale realistic image pairs with significant real-world viewpoint changes and diverse scene contexts. Our key observation is that the spatial attention module and the correspondence decoder network inherently exhibit a two-view nature, as they operate on the previously enhanced initial and query frame features. Therefore, we pretrain these components on two-view datasets to learn global matching priors. Subsequently, we introduce the temporal attention module and train it jointly on tracking datasets to acquire temporal priors. More training details are provided in Sec. 3.3. + +# 3.2. Multi-Scale Local Refinement + +Based on global tracking, we obtain the corresponding locations of query points in the new frame. However, to reduce the computation cost of global attention, the resolution of the feature maps is limited, leading to poor tracking accuracy. In this section, we devise a local refinement module to refine the tracking results for high accuracy. Moreover, to enhance the robustness of our method against noisy initial tracking results, we introduce the multi-scale refinement mechanism that iteratively regresses local offsets within a local search window surrounding initial locations across different feature scales. At the low-resolution level, the large search space enables the model to explore a broader range of potential correspondences, effectively correcting deviations from the tracking. In contrast, the high-resolution level provides finer details, ensuring precise and accurate predictions. Furthermore, similar to the motivation in the global tracking phase, we devise a spatial and temporal transformation mechanism to further enhance feature distinctiveness and capture global relationships between pixels within the local search window. Since the transformations in the refinement phase are performed on local feature patches with a small number of tokens, the history tokens are attended to by the current patch features. + +Specifically, we first extract multi-scale, fine-grained feature maps for the new frame and $n$ selected history frames, including the initial frame and $n - 1$ recent frames. Then, for a scale at level $l$ , the local feature patches with size of $w \times w$ centered at all tracking points are extracted, where the new frame patch is 4D correlated with each history frame patch to get $n$ history correlation maps. Then, we employ an encoder to reduce the dimensionality of the 4D correlation output to tokens with the size of $w \times w \times c$ . Subsequently, spatial attention is applied to the tokens using a self-attention mechanism, allowing the model to capture global pixel relationships between the query feature and its context feature. Next, temporal attention is performed across different history tokens to capture temporal dependencies. The spatial- and temporal-attention are interleaved by $k$ times. Finally, the spatial and temporal attended tokens are aggregated by weighted average and then used to decode local offset for tracked locations in the new frame, along with the certainty $c$ and occlusions scores $o$ indicating the reliability of the predictions. The offsets $\Delta x$ , $\Delta y$ are then added to initial tracking points for location updates and then used to perform the next iteration of local refinement. + +# 3.3. Training + +We separately train our global tracking module and local refinement module to reduce GPU memory consumption. + +Global Tracking. We propose a two-stage training strategy for our global tracking module, which is partially + +pre-trained on two-view matching datasets and then fully fine-tuned on tracking datasets. Specifically, we use MegaDepth [27] and ScanNet++ [56] as two-view matching pretraining datasets, covering indoor and outdoor scenes with significant viewpoint and illumination changes. During this phase, only the global transformation and decoder networks are trained, while the temporal attention module is skipped. To enhance the model's adaptability to dynamic scenarios and learn temporal priors, we subsequently train the entire model using synthetic point tracking dataset Kubric [17]. We employ Focal Loss to optimize the prediction of correspondences. + +Local Refinement. To train our local refinement module separately, we generate pseudo coarse tracking points as input. Specifically, these pseudo coarse tracking locations are initialized using tracking locations from the previous frame. To introduce a more challenging initialization and encourage the refinement module to correct large errors, we perturb the initial tracking locations with Gaussian noise, defined as $g \sim \alpha \cdot \mathcal{N}(0,1)$ , resulting in the final coarse tracking points. We employ the Huber Loss $L_{Huber}$ as the objective function for optimizing refined tracking locations. To supervise the certainty estimation, we follow [12] and set the ground truth certainty label to True if the pixel error is below a threshold of $\delta = 6$ . The certainty and occlusion loss are supervised using the BCE loss, denoted as $L_{exp}$ and $L_{occ}$ , respectively. The overall refinement loss is computed as the weighted average of $L_{Huber}$ , $L_{exp}$ , and $L_{occ}$ . More details about the loss function can be found in the supplementary material. + +# 4. Experiments + +In the experiment section, we first introduce the datasets used to evaluate our method (Sec. 4.1), followed by a detailed description of the important implementation details of the model (Sec. 4.2). Subsequently, we present our results across several datasets (Sec. 4.3). Finally, we conduct extensive ablation studies to verify the design choices of our pipeline (Sec. 4.4). + +# 4.1. Datasets and Metrics + +To evaluate the retrack capabilities of different tracking methods, we collect a new evaluation set named Ego-retrack dataset. Unlike EgoPoints [7], which supplies ground-truth labels for three selected frames per sequence, we follow the per-frame annotation protocol introduced by TAPNet [9]. The proposed dataset comprises 30 egocentric videos, each consisting of 300-600 frames that are annotated with 5-10 keypoints with ground-truth locations and visibility. It is designed to challenge tracking methods with frequent object reappearance, introducing significant viewpoint shifts and scene content variations. These factors make it particularly + +suitable for evaluating robustness in re-tracking scenarios. More details of this dataset can be found in the supplementary material. Besides, we conduct experiments on the common Tap-vid benchmarks and RoboTAP [52] dataset. Tap-vid benchmarks include labeled videos from Kinetics [2], RGB-Stacking [9], and DAVIS [37] datasets. Kinetics is a comprehensive dataset comprising over 1000 labeled videos, including challenges such as scene cuts and camera shake. RGB Stacking is a synthetic dataset of 50 robotics videos that have many textureless areas. DAVIS is a tracking benchmark encompassing complex motion and significant variations in object scale. RoboTAP consists of real-world robotic manipulation videos. Following [12, 23], videos are resized to $256 \times 256$ pixels before being processed by the model. Following [9], we evaluate different methods using the average proportion of points $\delta_{avg}^{x}$ . Average Jaccard (AJ), and Occlusion Accuracy (OA). More details can be found in the supplementary material. + +# 4.2. Implementation details + +For the training of the global tracking module, we sample $N = 400$ tracks for each video sequence during training. For the local refinement module, we set the level of refinement $L$ to 3, incorporating feature maps from the $16 \times 8 \times 2$ and $2 \times 2$ resolution scales. $N = 50$ tracks per sequence are sampled for training. The model is trained using 32 NVIDIA A100 GPUs with a batch size of 32 for 50,000 steps. Further details about pre-training and network structure can be found in the supplementary material. At inference time, the refinement module determines whether to adopt the position from the previous frame or the corrected position provided by the global matching module, based on the corresponding confidence score of the predicted coarse tracking location. + +# 4.3. Evaluation Results + +Ego-Retrack Dataset. Tab. 1 shows the performance of various point-tracking methods on the Ego-Retrack dataset. The upper part of the table presents the results of offline tracking methods, while the lower part shows the performance of online tracking methods. Our method outperforms all online baselines and remains competitive compared to the previously best-performing offline method. Specifically, our approach achieves a $5.7\%$ improvement in $\delta_{avg}^{x}$ over the online version of BootsTAPIR, along with higher accuracy in occlusion prediction. Compared to the offline version of BootsTAPIR, our method demonstrates superior performance on most metrics. The slightly lower performance on $\delta_{8}^{x}$ and $\delta_{16}^{x}$ metrics is primarily due to the offline model's ability to leverage future temporal context during inference, allowing it to reference more frames and jointly optimize predictions across multiple frames. Notably, "online" refers to the capability of making predictions without relying on future observations. To ensure this comparison + +Table 1. Quantitative comparison on the Ego-ReTrack dataset. All methods are evaluated with the first query mode. Online mode means tracks are predicted w/o subsequent frames. Here, CoTracker3 uses checkpoints trained with sliding window lengths of 60 and 16, respectively. We denote these variants as CoTracker3† and CoTracker3‡. + +
MethodAJ↑<δxavg↑OA↑<δ2↑<δ4↑<δ8↑<δx16↑
OfflineTAPIR [12]38.449.378.739.054.463.769.8
BootsTAPIR [11] (+15M videos)43.556.281.644.761.672.979.0
CoTracker3†[22] (+15K videos)41.952.780.146.059.564.968.6
CoTracker3‡(+15K videos)46.056.083.049.363.169.073.3
OnlineTAPIR33.843.376.134.449.156.662.0
BootsTAPIR (+15M videos)38.150.978.440.156.866.271.9
CoTracker3‡(Step Size=1)27.643.571.837.748.653.257.4
Ours44.356.682.247.063.170.777.5
+ +![](images/f58a1bd65d8b50ea96f1a92bffd290218de98d1b759e7bb16513b8580e392a71.jpg) +Figure 3. Qualitative results. Our method is qualtatively compared with CoTracker3 (sliding window size=16, step=8) on the Ego-ReTrack dataset. $\times$ and represent ground-truth and predicted tracking locations, respectively. + +setting, we configured CoTracker3's sliding window to advance one step at a time. Without accessing future frames, CoTracker3's performance declines. Our proposed method achieved the best performance across all metrics in the online tracking, demonstrating its robustness and effectiveness. Qualitative comparisons are shown in Fig. 3. + +TAP-Vid and RoboTAP Datasets. Results are presented in Tab. 2. Our method achieves the best performance in the online setting on the Tap-Vid and RoboTAP benchmarks. Besides, our approach not only excels in the online setting but remains competitive with offline methods. Specifically, when compared to the offline version of CoTracker3, our online version still outperforms it on both DAVIS and RoboTAP datasets. This demonstrates the effectiveness of our approach in online scenarios while maintaining strong performance relative to offline models. Specifically, under the same point tracking training data (Kubric, 24 frames) and the same online setting, our method outperforms TAPIR by 2.6 points on the Kinetic dataset and 7.2 points on the + +DAVIS benchmark. Notably, BootsTAPIR scales up the model by leveraging an additional 15 million real-world videos with pseudo-labels. Similarly, CoTracker3 scales up its model using 15,000 real-world video sequences, each comprising 64 frames. We believe that incorporating more real-world data and longer sequences could further unleash the potential of matching pre-trained models and enhance the performance of our approach. + +# 4.4. Ablation Studies + +To demonstrate the effectiveness of our method, we conducted extensive ablation studies to validate our design choices on Tapvid-DAVIS and Ego-ReTrack datasets. + +Global Tracking. As shown in Tab. 3 (1), eliminating the global tracking module leads to a significant decrease in performance on the Ego-Tracking dataset. The results reveal that the global tracking module is crucial for challenging scenarios, such as those involving significant viewpoint changes or complex motions in occlusion-heavy scenarios. + +Table 2. Quantitative comparison on the TAP-Vid and RoboTap datasets. All methods are evaluated with the first query mode. "Online" means tracks are predicted w/o subsequent frames. Best performance is bolded. + +
MethodKineticsDAVISRGB-StackingRoboTap
AJ↑<δxavg↑OA↑AJ↑<δxavg↑OA↑AJ↑<δxavg↑OA↑AJ↑<δxavg↑OA↑
OfflineTAP-Net [10]38.554.480.633.048.678.853.568.186.345.162.182.9
TAPIR [12]49.664.285.056.270.086.555.569.788.059.673.487.0
TAPTR [26]49.664.285.063.076.191.160.876.287.060.175.386.9
TAPTRv2 [25]49.764.285.763.575.991.453.470.581.260.974.687.7
CoTracker [23]48.764.386.562.675.489.371.683.389.658.670.687.0
SpaTracker [55]50.165.986.961.176.389.563.577.688.2---
LocalTrack [6]52.966.885.363.075.387.269.783.289.562.376.287.1
CoTracker3 [22]53.566.586.463.376.288.074.084.990.559.973.487.1
BootsTAPIR [11] (+15M videos)54.768.586.361.474.088.4---64.980.186.3
CoTracker3 (+15K videos)54.767.887.464.476.991.274.385.292.464.778.089.4
OnlineMFT [33]39.660.472.747.366.877.8------
TAPIR-online [12]51.564.485.256.770.285.767.7--59.1--
CoTracker (Step Size=1)---55.968.783.759.571.982.4---
CoTracker3 (Step Size=1)---56.070.887.064.380.988.652.970.182.6
Ours54.567.086.463.777.588.268.280.785.865.578.189.9
+ +Table 3. Ablation of key components. The ablations explore the impact of removing global tracking, two-view pre-training, and temporal attention, as well as memory length and refine levels. + +
DAVISEgo-Retrack
AJ↑<δxavg↑OA↑AJ↑<δxavg↑OA↑
Full63.577.688.044.356.682.2
(1) w/o Global Tracking60.773.986.731.635.972.5
(2) w/o Two-View Pre-training62.775.987.441.653.880.1
(3) w/o Temporal Attention62.676.187.639.352.578.6
(4) History Lengthm=360.975.686.240.853.381.0
m=662.877.287.343.555.081.5
m=963.177.587.443.055.581.1
m=1263.577.688.044.356.682.2
m=1563.377.287.744.156.682.0
(5) Refine LevelL=136.349.382.721.131.775.2
L=256.070.486.739.451.680.9
L=363.577.688.044.356.682.2
+ +Two-View Matching Pre-training. Tab. 3 (2) demonstrates the advantages of two-view matching pre-training in our approach. The results show that incorporating pretraining on two-view matching datasets consistently enhances model performance across both datasets, highlighting the effectiveness of learned global two-view matching priors for improving the tracking task. + +Temporal Attention Module. Tab. 3 (3) shows the benefits of the temporal attention module in the global tracking phase. We removed the temporal attention and directly utilize the spatial attention and decoder network to obtain tracking locations Without temporal information, the performance of the model decreases due to the large scene context gap between the initial frame and current frame. The results indicate that spatial attention effectively bridges the context gap by leveraging intermediate frames, thereby enhancing the model's tracking performance. + +Number of History Frames. In the local refinement stage, we employ $m$ historical feature patches for each tracking point to compute a 4D correlation with patches from the current frame, followed by temporal-spatial attention to decode local offsets. Tab. 3 (4) illustrates the impact of varying + +patch counts. Results show consistent performance gains as more feature patches are incorporated, with the highest inference accuracy obtained at $m = 12$ . This improvement can be attributed to the model's enhanced capability to leverage richer temporal and spatial context. + +Number of Refinement Level. Our local refinement module iteratively refines tracking results using features from different levels to achieve higher accuracy. Tab. 3 (5) reports the performance after each refinement iteration. Results indicate progressive improvements in tracking accuracy with each refinement layer, achieving the best performance after three iterations. This highlights the effectiveness of our design, particularly in challenging scenarios such as occlusions or rapid motion, where iterative refinement corrects coarse predictions and enhances precision. + +# 5. Conclusions + +We propose an online tracking framework that robustly retracks points after long-term occlusions, leveraging recent advancements in two-view matching to enhance tracking under significant viewpoint changes. Our approach includes a global tracking model that establishes correspondences between the initial query frame and the current frame by integrating temporal and spatial attention mechanisms. Additionally, we introduce a pretraining strategy that first learns general global matching priors from large-scale, realistic two-view data and then fine-tunes on tracking data to acquire temporal priors. To further improve accuracy, we incorporate a multi-scale local refinement pipeline to refine tracked locations. Extensive experiments demonstrate that our framework outperforms existing online tracking baselines across all datasets and evaluation metrics. We believe the proposed online tracker can serve as a valuable tool for latency-sensitive and occlusion-heavy applications. + +# Acknowledgement + +This work was partially supported by the following grants: National Key R&D Program of China (No.2024YFB2809105), NSFC (No.U24B20154, 62402427), Zhejiang Provincial Natural Science Foundation of China (No.LR25F020003), Ant Research, and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University. + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Quentin Le, and Priya Dehghani. End-to-end object detection with transformers. arXiv preprint arXiv:2005.12872, 2020.3 +[2] João Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 6 +[3] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In Proc. ECCV, 2022. 3 +[4] Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry. CVPR, pages 19844–19853, 2024. 1 +[5] Zihao Chen, Kunhong Li, Haoran Li, Zhiheng Fu, Hanmo Zhang, and Yulan Guo. Metric localization for lunar rovers via cross-view image matching. Visual Intelligence, 2(1):12, 2024. 3 +[6] Seong Hun Cho, Seokju Hong, and Seungryong Kim. Local all-pair correspondence for point tracking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 366-383, Cham, 2024. Springer. 3, 8 +[7] Ahmad Darkhalil, Rhodri Guerrier, Adam W Harley, and Dima Damen. Egopoints: Advancing point tracking for egocentric videos. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8556-8565. IEEE, 2025. 6 +[8] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. CVPRW, 2018. 3 +[9] 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, 2022. 2, 3, 6 +[10] 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. Advances in Neural Information Processing Systems, 35:13610–13626, 2022. 8 +[11] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 2, 3, 7, 8 +[12] 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. 2, 3, 6, 7, 8 +[13] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. In CVPR, 2019. 3 +[14] Johan Edstedt, Ioannis Athanasiadis, Marten Wadenback, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In CVPR, 2023. 3 +[15] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5 +[16] Kuan Fang, Fangchen Liu, Pieter Abbeel, and Sergey Levine. Moka: Open-world robotic manipulation through mark-based visual prompting. Robotics: Science and Systems (RSS), 2024. 1 +[17] 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. 5, 6 +[18] 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 +[19] Xingyi He, Hao Yu, Sida Peng, Dongli Tan, Zehong Shen, Hujun Bao, and Xiaowei Zhou. Matchanything: Universal cross-modality image matching with large-scale pre-training. In arXiv preprint arXiv:2501.07556, 2025. 2, 3 +[20] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 1 +[21] Hyeonho Jeong, Chun-Hao Paul Huang, Jong Chul Ye, Niloy J. Mitra, and Duygu Ceylan. Track4gen: Teaching video diffusion models to track points improves video generation. ArXiv, abs/2412.06016, 2024. 1 +[22] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 3, 7, 8 +[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. 2, 3, 6, 8 +[24] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: Connecting the dots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19187-19197, June 2024. 3 +[25] 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. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 8 +[26] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, + +Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. arXiv preprint arXiv:2403.13042, 2024. 2, 3, 8 +[27] 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 (CVPR), pages 2041-2050, 2018. 6 +[28] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 3 +[29] Chang Liu, Yongsheng Yuan, Xin Chen, Huchuan Lu, and Dong Wang. Spatial-temporal initialization dilemma: towards realistic visual tracking. Visual Intelligence, 2(1):35, 2024. 1 +[30] G LoweDavid. Distinctive image features from scale-invariant keypoints. IJCV, 2004. 3 +[31] Raul Mur-Artal, José M. M. Montiel, and Juan D. Tardós. Orb-slam: A versatile and accurate monocular slam system. IEEE Transactions on Robotics, 2015. 1 +[32] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 2016. 1 +[33] Michal Neoral, Jonás Šerzych, and Jiří Matas. Mft: Long-term tracking of every pixel. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6837-6847, 2024. 8 +[34] Michal Neoral, Jonás Šerzych, and Jiří Matas. Mft: Long-term tracking of every pixel, 2023. 3 +[35] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video. arXiv preprint arXiv:2410.24211, 2024. 3 +[36] 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. arXiv, 2023. 4 +[37] F. Peruzzi, 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 Proc. CVPR, 2016. 6 +[38] Frano Rajic, Lei Ke, Yu-Wing Tai, Chi-Keung Tang, Martin Danelljan, and Fisher Yu. Segment anything meets point tracking. arXiv preprint arXiv:2307.01197, 2023. 1 +[39] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. Efficient neighbourhood consensus networks via submanifold sparse convolutions. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 605-621. Springer, 2020. 3 +[40] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in neural information processing systems, 31, 2018. 3 +[41] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary R. Bradski. Orb: An efficient alternative to sift or surf. ICCV, + +2011.3 +[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 3 +[43] Jonas Serych, Michal Neoral, and Jiri Matas. Mftiq: Multiflow tracker with independent matching quality estimation. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 8068-8078, February 2025. 3 +[44] Xuelun Shen, Zhipeng Cai, Wei Yin, Matthias Müller, Zijun Li, Kaixuan Wang, Xiaozhi Chen, and Cheng Wang. Gim: Learning generalizable image matcher from internet videos. arXiv preprint arXiv:2402.11095, 2024. 2, 3 +[45] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 3, 5 +[46] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. *ICLR*, 2022. 3 +[47] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proc. ECCV, 2020. 3 +[48] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5714-5724, 2021. 3 +[49] 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 (ECCV), 2024. 3 +[50] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurlPS, 2017. 3 +[51] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation. arXiv preprint arXiv:2308.15975, 2023. 1 +[52] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation, 2023. 6 +[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. 3 +[54] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In arXiv preprint arXiv:2312.14132, 2024. 2 +[55] 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 (CVPR), pages 20406-20417, June 2024. 3, 8 +[56] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision (ICCV), pages 12-22, 2023. 6 +[57] Tingyang Zhang, Chen Wang, Zhiyang Dou, Qingzhe Gao, Jiahui Lei, Baoquan Chen, and Lingjie Liu. Protracker: Probabilistic integration for robust and accurate point tracking. arXiv preprint arXiv:2501.03220, 2025. 3 +[58] 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. 3, 5 \ No newline at end of file diff --git a/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/images.zip b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f1a4448fe61e07dc3ebb11a41f417aa2a27cd7cc --- /dev/null +++ b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79bfa02b9c2991a75a8949dff04a52c150ffd9dd3b8645e29fc88952eaae958d +size 598643 diff --git a/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/layout.json b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cc15a73fb05a3d4a779f612a4dad853ee9b27222 --- /dev/null +++ b/ICCV/2025/ReTracker_ Exploring Image Matching for Robust Online Any Point Tracking/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:463f67c44ae27a8aee46529798fb763410c17d277d700ed185371281c1494c81 +size 328399 diff --git a/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/f297df5f-7239-483e-af8f-937fc815c2f4_content_list.json b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/f297df5f-7239-483e-af8f-937fc815c2f4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a8c514d24ef5c7c379867b8a2a89580380a76de0 --- /dev/null +++ b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/f297df5f-7239-483e-af8f-937fc815c2f4_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c81eeaa555e784688cb49864fc3f2aa57f8c4fe54ed5e447f3ec82c94f8ddc55 +size 104481 diff --git a/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/f297df5f-7239-483e-af8f-937fc815c2f4_model.json b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/f297df5f-7239-483e-af8f-937fc815c2f4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1f210e57f7e097503bc9806d505e2b97b89fc1d0 --- /dev/null +++ b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/f297df5f-7239-483e-af8f-937fc815c2f4_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d7a815f9b1ed84914cc98581fd3f3beb02b7a65d7a94b6e43d8fc4cd67aa844 +size 137271 diff --git a/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/f297df5f-7239-483e-af8f-937fc815c2f4_origin.pdf b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/f297df5f-7239-483e-af8f-937fc815c2f4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..728af6f0a61c79453408f6ab0fb30240cfea04df --- /dev/null +++ b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/f297df5f-7239-483e-af8f-937fc815c2f4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e5d8beef997183743dbda49c03d2a90d9cd65734a30f708567999b4138ec82f +size 2257243 diff --git a/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/full.md b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cb47be5a39caa466dd2bbe2372d1b191968dc8d8 --- /dev/null +++ b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/full.md @@ -0,0 +1,436 @@ +# Real3D: Towards Scaling Large Reconstruction Models with Real Images + +# Hanwen Jiang Qixing Huang Georgios Pavlakos +The University of Texas at Austin + +Project & Code: https://hwjiang1510.github.io/Real3D/ + +![](images/b591f597a9c6d980f930afc835a7482bd6789f48712d715f87d0b6ef9e8a0ab9.jpg) +Figure 1. Real3D reconstructions of in-the-wild instances. We show the input view and two novel views of our reconstruction, including a novel side view and a novel back view. Real3D can reconstruct diverse shapes from in-the-wild images, including those in uncommon poses and shape categories. + +# Abstract + +Training single-view Large Reconstruction Models (LRMs) follows the fully supervised route, requiring multi-view supervision. However, the multi-view data typically comes from synthetic 3D assets, which are hard to scale further and are not representative of the distribution of real-world object shapes. To address these limitations, we introduce Real3D, the first LRM that uses single-view real images for training, benefiting from their scalability and capturing the real-world shape distribution. Real3D introduces a novel self-training framework, including unsupervised losses at the pixel- and semantic-level, enabling LRMs to learn from these single-view images without multi-view supervision. Simultaneously, to deal with the noise of real data, Real3D also presents an automatic data curation approach to gather high-quality examples that have positive impact on training. Our experiments show that Real3D consistently outperforms prior work in diverse evaluation settings that include real and synthetic data, as well as both in-domain and out-of-domain shapes. + +# 1. Introduction + +The scaling law is the secret sauce of large foundation models [33]. By scaling both model parameters and training data, the foundation models demonstrate impressive emerging ca + +pabilities [2, 60]. Recently, the same recipe has been applied to build large reconstruction models (LRMs) [23], i.e., foundation models for single-view 3D reconstruction. LRMs benefit from training on multi-view images from 3D/video data, where the increased dataset size [9, 10, 98] is the key of the improved model performance. However, the excessive reliance on multi-view supervision creates two bottlenecks for LRMs. First, expanding the scale of such data is hard - either creating synthetic 3D assets [9] or intentional video captures [98] is laborious. This leads to a limited scale of training data of LRMs, compared with LLMs [2, 76] and VLMs [36, 101], ceiling their scalability. Second, the multi-view data is biased towards shapes that are easy to model by artists or easy to capture in the round table setting [29, 62, 98], creating a domain gap with the real-life shapes we want to reconstruct in most applications. + +To address these limitations, we propose Real3D. Our main objective is to expand the training data of LRMs to single-view images. This is motivated by the potential of using single-view images to solve the two aforementioned problems. First, single-view images are readily available in existing large-scale datasets [12, 66, 74] or in-the-wild resources, leading to a large-scale set of training examples. Second, by leveraging a large number of collected real images for training, we can capture the distribution of real objects more faithfully (Figure 1), closing the training-inference gap and + +![](images/ddd1f4f956125bbe902bf4ba527158d88e4a456218936f28e8ec1d315f43300d.jpg) +Figure 2. Overview of Real3D. (Left) Real3D adopts a standard LRM model architecture with Triplane 3D representation $\mathbf{T}$ that can render a novel view $I_{\Phi}$ given a target camera pose/viewpoint $\Phi$ (Sec. 3). The LRM is first initialized with medium-scale multi-view synthetic data using supervised losses $\mathcal{L}_{\mathrm{mv}}^S$ . (Middle) To enable scaling, Real3D extends the training data to single-view real images, benefiting from their scalability and faithful capture of the real-world shape distribution. Real3D proposes a novel self-training method for improving LRM on this single-view data, including pixel-level, $\mathcal{L}_{\mathrm{pix}}^R$ , and semantic-level, $\mathcal{L}_{\mathrm{sem}}^R$ , self-supervision (Sec. 4.1). For the pixel-level supervision we leverage a cycle consistency rendering loss, while for the semantic-level self-supervision, we maximize the semantic similarity between rendered novel views of the reconstruction and the input image. (Right) To further benefit from self-training, Real3D automatically selects the high-quality instances from large-scale noisy images (Sec. 4.2), benefiting model accuracy and generalization capability. + +![](images/bff4d39adcd782aeec5fb573ecca127e6ef335b20f5db7af18dd351c4e15c269.jpg) +Func. 1 Self-Supervised Learning $(L_{\text{mix}}^R \& L_{\text{sem}}^R)$ + +![](images/af3a30c683a25e48ba494dffef4f0f4e4a57bda54884f70fb8617c50b6628386.jpg) + +improving generalization [38, 44, 94]. + +Given this abundance of single-view images, the key challenge is how to incorporate them in the training of LRMs. To achieve this, Real3D proposes a novel self-training framework. More specifically, we introduce unsupervised training losses, i.e., both pixel-level and semantic-level self-supervision, mimicking how LRMs are trained with multi-view supervision. The pixel-level self-supervision leverages a cycle consistency rendering loss, enabling training with photometric losses on merely the single-view input image. The semantic-level self-supervision maximizes the semantic similarity between rendered novel views of the reconstruction and the input image (Figure 2, middle). + +Although the large-scale image collections are beneficial, they tend to be highly noisy, leading to limited improvements for our system. To address this, Real3D introduces an automated data curation method for selecting high-quality examples, i.e., unoccluded instances. Eventually, we train jointly on the curated in-the-wild data using self-training and on the synthetic data using supervised losses. This strategy enriches the model with knowledge from real single-view data while preventing divergence. + +We evaluate Real3D on a diverse set of datasets, spanning both real and synthetic data and covering both in-domain and out-of-domain shapes. Experimental results highlight Real3D's three key strengths: i) Superior performance. Real3D outperforms prior works in all evaluations, with an average relative improvement of $4.2\%$ PSNR (15% MSE), $6.3\%$ LPIPS, $13.5\%$ FID, and $14.3\%$ mesh accuracy. ii) Effective use of real data. Real3D demonstrates greater improvement using single view real-data than what the previous methods achieved using multi-view real data. iii) Scalability. Performance improves as more data is incorporated, demonstrating Real3D's potential of scaling. We will make our code, models, and data available upon publication. + +# 2. Related Work + +Single-view 3D Reconstruction. Reconstructing 3D scenes and objects from a single image is a core task in 3D Computer Vision. One line of work focuses on developing better 3D representations to improve reconstruction quality. For example, different explicit representations, e.g., voxels [8, 78], point clouds [14, 30], multiplane images [51, 77], meshes [16, 43], and 3D Gaussians [35, 71, 73], as well as implicit representations, e.g., SDFs [49, 56, 69], and radiance fields [27, 52, 63, 97], have been explored. Recent methods explore incorporating various guidance to improve single-view 3D reconstruction. MCC [85] and ZeroShape [25] use depth guidance; however, accurate depth inputs are not always available. RealFusion [48], Make-it-3D [72] and Magic123 [58] harness diffusion priors as guidance [57]. However, they require slow per-shape optimization. Zero-1-to-3 [42] fine-tunes diffusion models for direct novel view generation. With multiple views, it is easier to get a 3D reconstruction. However, this route suffers from limited reconstruction quality, caused by inconsistency between the generated novel views [45, 68, 81]. Adversarial guidance is explored to distinguish reconstruction from input view [24, 90]. However, adversarial training is generally unstable and difficult to scale. Moreover, semantic guidance leverages the image-to-text inversion model and the similarity calculation model to supervise novel reconstruction views [11, 72]. This improves the semantic consistency of the reconstructions but harms the reconstruction details [15, 60]. In contrast, our method, Real3D, proposes two complementary unsupervised losses to improve both the semantic and spatial consistency of our reconstruction, enabling us to train using single-view images. + +Large Reconstruction Models. Large reconstruction models are proposed for generalizable and fast feed-forward + +3D reconstruction. They use scalable model architectures, e.g., Transformers [23, 26, 80] or Convolutional U-Nets [64, 73, 83] to encode diverse shape and texture priors and directly map 2D information to 3D representations. The models are trained with multi-view rendering losses, assuming access to 3D ground-truth. For example, LRM [23] uses triplane tokens to query information from 2D image features. Other methods improve reconstruction quality by exploiting better representations [84, 99, 102] and introducing generative priors to first generate multi-view images [18, 46, 73, 83, 91, 92]. However, one shortcoming of these models is that they require a normalized coordinate system and canonicalized input camera pose, which limits the scalability and effectiveness of training with multi-view real data. To solve this problem, we enable the model to perform self-training on real single images, without overly relying on multi-view supervision. + +Unsupervised 3D Learning from Real Images. Learning to perform 3D reconstruction typically requires 3D ground-truth or multi-view supervision, which makes scaling up more challenging. To solve this problem, a promising avenue would be learning from massive unannotated data. Early works in this direction leverage category-level priors, where the reconstruction can benefit from category-specific templates and the definition of a canonical coordinate frame [3, 13, 21, 32, 34, 37, 39, 41, 53-55, 86]. Recent works extend this paradigm to general categories by adjusting adversarial losses [4, 40, 50, 96], multi-category distillation [1], synergy between multiple generative models [88], knowledge distillation [70], depth regularization [65], and multi-view compression [28]. However, these methods learn 3D reconstruction from scratch, without leveraging the available 3D annotations due to limitations of their learning frameworks. Thus, their 3D accuracy and viewpoint range are limited. In contrast, our model enables initialization with available 3D ground-truth from synthetic examples and is jointly trained with in-the-wild images using unsupervised losses, which improves the reconstruction quality. + +Model Self-Training. Self-training helps improve performance when labeled data is limited or not available [61, 67, 89]. For example, contrastive learning methods use different image augmentations to learn visual representations [5, 6, 19]. This strategy has also been applied to 3D computer vision tasks, e.g., hand pose estimation [44], detection [93], segmentation [31] and depth estimation [94, 95]. In this paper, we propose novel losses to perform self-training on real images to improve 3D reconstruction. + +# 3. Preliminaries + +Large Reconstruction Model (LRM). Let $I \in \mathbb{R}^{H \times W \times 3}$ be an input image of the target object. The LRM outputs a 3D representation of the object, $\mathbf{T} = \mathrm{LRM}(I)$ , where $\mathbf{T} \in \mathbb{R}^{3 \times h \times w \times c}$ is the latent triplane. LRM performs vol + +ume rendering [4] to produce novel views. The rendering module is formulated as $\hat{I}_{\Phi} = \pi (\mathbf{T},\Phi)$ , where $\hat{I}_{\Phi}$ is the rendered image under a target camera pose $\Phi \in \mathrm{SE}(3)$ , and $\pi$ represents the rendering process. + +Training LRM on Synthetic Multi-view Images. The standard training of LRMs requires multi-view image supervision. For each training object, we have access to several views of it with the corresponding camera poses. We denote the views and poses as $\{(I_i^{gt},\Phi_i)|i = 1,\dots,n\}$ , where $n$ images are collected. These multi-view images are usually obtained from synthetic 3D assets [9] using rendering tools. The loss function on these multiple views is: + +$$ +\mathcal {L} _ {\mathrm {m v}} (I) = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(\mathcal {L} _ {\mathrm {M S E}} \left(\hat {I} _ {\Phi_ {i}}, I _ {i} ^ {g t}\right) + \lambda \cdot \mathcal {L} _ {\mathrm {L P I P S}} \left(\hat {I} _ {\Phi_ {i}}, I _ {i} ^ {g t}\right)\right), \tag {1} +$$ + +where $\lambda$ balances the two losses, $\mathcal{L}_{\mathrm{MSE}}$ and $\mathcal{L}_{\mathrm{LPIPS}}$ [100] that provide pixel-level and semantic-level supervision, respectively. + +Coordinate Frame System. LRM has a constant SE(3) camera pose, $\phi$ , that renders the reconstruction triplane $\mathbf{T}$ at the viewpoint of input images. This camera pose $\phi$ has an identity rotation and a fixed translation [23]. Since this constant camera pose is shared across any input image, we denote it as the canonical pose. This canonical pose is used in Sec. 4.1 for designing our self-training techniques. + +# 4. Real3D + +We propose a novel framework that enables training LRMs using unlabeled real-world single-view images, which are easier to collect/scale and can better capture the distribution of real object shapes. As shown in Fig. 2, we initialize an LRM on a synthetic dataset. Then we collect real-world object instances. We jointly train the model on synthetic multi-view data (Eq. 1) and perform self-training on real single-view data. The former prevents the model from diverging with the help of supervision from ground-truth novel views. The latter introduces new data in the model training, which improves the reconstruction quality and generalization capability (Sec. 4.1). We also propose an automatic data curation method to collect high-quality instances from real images to benefit self-training (Sec. 4.2). + +# 4.1. Self-Training on Real Images + +The effectiveness of the multi-view training loss (Eq. 1) originates from applying supervision for improving both pixel-level and semantic-level similarity between reconstruction and ground-truth novel views. Following this philosophy, we develop novel unsupervised pixel-level and semantic-level supervision when training the model on single-view images, where ground-truth multiple views are not available. For our base model, we use a fine-tuned TripoSR [75], an LRM without input pose and intrinsics conditioning [23]. + +![](images/02461a53406cf360e82b846b47dd3fc36ac791bb2993f3a65d438a4459a1cd23.jpg) +Figure 3. Pixel-level supervision with cycle-consistency loss. We use LRM to reconstruct the object $(T^R)$ and render a novel view $\hat{I}_{\Phi}^{R}$ at viewpoint/pose $\Phi$ . We again apply LRM on this rendered novel view to reconstruct the object $(T^{R})$ and render it from the viewpoint of the original input, denoted as $\dot{I}^{R}$ . The difference between $I^{R}$ and $\dot{I}^{R}$ provides the loss for training. We apply stop gradient on intermediate result $\hat{I}_{\Phi}^{R}$ to avoid model degradation. + +# 4.1.1. Pixel-level Supervision + +The pixel-level supervision has three components: 1) the cycle consistency loss for providing pixel-level supervision, 2) the stop gradient regularizer, which avoids minimizing the loss by learning trivial solutions, and 3) the camera pose sampling curriculum to improve stability. + +Cycle-consistency Loss. The loss is illustrated in Fig. 3. It provides pixel-level supervision using the original input image. We formulate it as follows. + +We input the model with an image $I^R$ that contains a real-world shape instance and then reconstruct the triplane: + +$$ +\mathbf {T} ^ {R} = \operatorname {L R M} \left(I ^ {R}\right), I ^ {R} \in \mathbb {R} ^ {H \times W \times 3}. \tag {2} +$$ + +We sample a camera pose $\Phi$ to render a novel view of the reconstruction and create the second input $\hat{I}_{\Phi}^{R}$ in the cycle: + +$$ +\hat {I} _ {\Phi} ^ {R} = \pi (\mathbf {T} ^ {R}, \Phi), \Phi = \phi \cdot \Delta \Phi , \tag {3} +$$ + +where the sampled camera pose $\Phi$ is associated with the relative pose $\Delta \Phi$ between the sampled target view and the input view. Here, $\phi$ is the constant canonical pose of the input view (introduced in Sec. 3). We input this synthesized novel view back to the LRM for reconstruction, and render a image at the viewpoint of the original input $I^R$ : + +$$ +\dot {\mathbf {T}} ^ {R} = \operatorname {L R M} \left(\hat {I} _ {\Phi} ^ {R}\right), \text {a n d} \tag {4} +$$ + +$$ +\dot {I} ^ {R} = \pi (\dot {\mathbf {T}} ^ {R}, \dot {\Phi}), \text {w h e r e}, \dot {\Phi} = \phi \cdot (\Delta \Phi) ^ {- 1}. \tag {5} +$$ + +Here $\dot{I}^R$ is the image that closes the cycle and $\dot{\Phi}$ is the camera pose to render the image at the viewpoint of original input $I^R$ . The pixel-level cycle consistency loss is defined as: $\mathcal{L}_{\mathrm{pix}}^R = \mathcal{L}_{\mathrm{MSE}}(\dot{I}^R, I^R) + \lambda \cdot \mathcal{L}_{\mathrm{LPIPS}}(\dot{I}^R, I^R)$ . + +Stop-gradient Regularizer. We observe that simply training with the cycle consistency loss leads to degraded performance. In detail, $\mathcal{L}_{\mathrm{pix}}^R$ , which is defined on 2D images, + +![](images/4193ce7f76627bf40129f4fa12f19654e37e27432077ac8276e13e2c1fc02747.jpg) +Pose Sampling w. Curriculum +Figure 4. Pose sampling curriculum. By guiding the sampling process of pose $\Phi$ , we control the training difficulty. At the beginning of training, we ensure a small variation between sampled pose $\Phi$ and canonical pose $\phi$ , achieving steady training. We gradually enlarge the upper bound of this variation ( $\Phi_n > \ldots > \Phi_2 > \Phi_1$ ) during training to increase difficulty and improve learning. + +![](images/903c21db605fb2881f2c8811e549b39d4f595554dfba0b838402a9ad4c5266b5.jpg) +Epoch 1 +I + +![](images/2ef63888e51d9bb0f3db41f09ee5ffe93c8db417f6c93fafc122c41f3d364b19.jpg) +Epoch 2 +I + +![](images/233c74a648b64d6956fec067296769a94bf1cb6cc8293a20d0eb837f53907c29.jpg) +Epoch 3 +I + +can be minimized by adopting trivial solutions that do not improve 3D geometry quality. Please see Appendix for visualization. To solve the problem, we apply a stop-gradient regularization on the intermediate rendering result $\hat{I}_{\Phi}^{R}$ , as shown in the right side of Fig. 3. The stop-gradient operation is widely used in self-supervised learning to improve training effectiveness by avoiding trivial solutions [6, 17, 79]. + +This stop-gradient operator divides the cycle into two parts. The second half of the cycle (from $\hat{I}_{\Phi}^{R}$ to $\hat{I}^{R}$ with $\mathcal{L}_{\mathrm{pix}}^{R}$ ) is the normal method for training LRM with novel view rendering losses. The first half of the cycle (from $I^{R}$ to $\hat{I}_{\Phi}^{R}$ ) can be interpreted as the process of collecting the input image $\hat{I}_{\Phi}^{R}$ of the second half. + +Learning with Curriculum. When using the stop-gradient regularizer, the training performance is dependent on the quality of the intermediate rendered view $\hat{I}_{\Phi}^{R}$ . However, since the model is imperfect, the created novel view $\hat{I}_{\Phi}^{R}$ can be inaccurate, limiting the effectiveness of self-training. + +To address this problem, we introduce a curriculum learning approach. This curriculum is based on our observation that the rendered novel view tends to be more accurate when the camera pose has a smaller variation with the input image. Thus, we can control the learning difficulty and uncertainty by varying the pose change between the sampled pose $\Phi$ , which is used for rendering, and the canonical pose $\phi$ . As shown in Fig. 4, initially, the model learns on simpler cases with a small pose variation, where the rendered view $\hat{I}_{\Phi}^{R}$ tends to be more accurate. We formulate the camera sampling method under the curriculum as: + +$$ +\Phi = \phi \cdot \Delta \Phi , \Delta \Phi \sim U n i f o r m \left(- \Delta \Phi_ {\max } ^ {j}, \Delta \Phi_ {\max } ^ {j}\right), \tag {6} +$$ + +where we denote Uniform as the uniform sampling in the SE(3) pose space, achieved by parametrizing poses with the polar coordinate system. $\Delta \Phi_{\mathrm{max}}^{\mathrm{j}}$ is the maximum sampling range of relative camera pose at the training iteration $j$ . Note $\Delta \Phi_{\mathrm{max}}^{j}$ is parameterized by the relative azimuth $\theta_{\mathrm{max}}^{j}$ + +and elevation $\varphi_{\mathrm{max}}^j$ , which are + +$$ +\theta_ {\max } ^ {j} = j / j _ {\max } \cdot \left(\theta_ {\max } - \theta_ {\min }\right) + \theta_ {\min }, \text {a n d} \tag {7} +$$ + +$$ +\varphi_ {\max } ^ {j} = j / j _ {\max } \cdot \left(\varphi_ {\max } - \varphi_ {\min }\right) + \varphi_ {\min }, \tag {8} +$$ + +where $j_{\mathrm{max}}$ is the number of total training iterations. We start the curriculum with $\theta_{\mathrm{min}} = \varphi_{\mathrm{min}} = 15^{\circ}$ , and finalize with $\theta_{\mathrm{max}} = \varphi_{\mathrm{max}} = 90^{\circ}$ . The camera pose sampling range keeps increasing in the curriculum during the training process. + +# 4.1.2. Semantic-level Supervision + +The semantic supervision is performed between the novel view of the reconstruction and the input view. We leverage CLIP to compute the semantic similarity loss, as CLIP is trained on image-text pairs and can capture high-level image semantics. The image semantic similarity loss is $\mathcal{L}_{\mathrm{CLIP}}(\hat{I}, I) = -\langle f(\hat{I}), f(I) \rangle$ , where $f$ represents the CLIP visual encoder for predicting normalized image features. + +A simple strategy for applying the loss is rendering multiple novel views of a reconstruction and calculating the loss on all of them. However, this leads to the multi-head problem, a trivial solution to minimize the loss (see visualization in Appendix). The reason is that CLIP is not fully invariant to the camera viewpoint. + +To address this issue, we apply the loss to a single rendered novel view, specifically the one least similar to the input view among multiple candidates, leveraging hard negative mining. Besides, we avoid rendering novel views that are too far from the input viewpoint, as the back of the shape may exhibit semantic differences from the input view. + +We formulate the semantic-level supervision as follows. For the latent triplane $\mathbf{T}^R$ of real image $I^R$ , we render $m$ novel views using $m$ sampled rendering camera poses as: + +$$ +\hat {I} _ {\Phi_ {i}} ^ {R} = \pi \left(\mathbf {T} ^ {R}, \Phi_ {i}\right), \text {f o r} i \in \{1, \dots , m \}. \tag {9} +$$ + +We sample the camera poses using a similar strategy in Eq. 6: + +$$ +\Phi_ {i} = \phi \cdot \Delta \Phi_ {i}, \Delta \Phi_ {i} \sim U n i f o r m (- \Delta \Phi_ {\max }, \Delta \Phi_ {\max }), \tag {10} +$$ + +where $\Delta \Phi_{\mathrm{max}}$ is parameterized by $\theta_{\mathrm{max}}^{\prime} = 120^{\circ}$ and $\varphi_{\mathrm{max}}^{\prime} = 45^{\circ}$ . Unlike the sampling of pixel-level supervision (Eq. 6), this sampling does not depend on the training iteration $j$ . + +Finally, we calculate the semantic-level supervision as: + +$$ +\mathcal {L} _ {\mathrm {s e m}} ^ {R} = \mathcal {L} _ {\mathrm {C L I P}} \left(\hat {I} _ {\Phi_ {k}} ^ {R}, I ^ {R}\right), \text {w h e r e} \tag {11} +$$ + +$$ +k = \underset {i \in \{1, \dots , m \}} {\arg \max } L _ {\mathrm {C L I P}} \left(\hat {I} _ {\Phi_ {i}} ^ {R}, I ^ {R}\right). \tag {12} +$$ + +Training Target. The losses applied in self-training on real data can be defined as: + +$$ +\mathcal {L} _ {\text {s e l f}} ^ {R} = \lambda_ {\text {i n}} ^ {R} \cdot \mathcal {L} _ {\text {i n}} ^ {R} + \lambda_ {\text {p i x}} ^ {R} \cdot \mathcal {L} _ {\text {p i x}} ^ {R} + \lambda_ {\text {s e m}} ^ {R} \cdot \mathcal {L} _ {\text {s e m}} ^ {R}, \tag {13} +$$ + +where $\mathcal{L}_{\mathrm{in}}^R$ is the rendering loss on the input view, and $\hat{I}_{\phi}^{R}$ is rendered using the canonical pose of input as $\hat{I}_{\phi}^{R} = \pi (\mathbf{T}^{R},\phi)$ . The final losses on both synthetic and real-world data can be defined as $\mathcal{L} = \mathcal{L}_{mv}^{S} + \mathcal{L}_{self}^{R}$ . + +![](images/a037992c8f289ef221b3c8da1ab99bf984ac02cce00f7a2c706af46e1603a244.jpg) +Figure 5. Pipeline for automatic selection of clean instances. We use an instance segmentation model (DECOLA [7]) to detect masks of object instances and a monocular depth model (DepthAnything [94]) to estimate depth maps $D$ . For each instance, we sample points on its boundary and select pixels along the direction of the inner and outer boundary normal $(p_{in}$ and $p_{out}$ respectively). If $D(p_{in}) < D(p_{out})$ , we infer that the instance "owns" this boundary. If this inequality holds for all boundary points, we conclude that the object is unoccluded and is considered a clean instance. + +# 4.2. Automatic Data Curation + +Our data curation aims to identify unoccluded instances from real images. This is important because occluded instances have incomplete side-view information and lead to limited model improvement (verified in Table 8). We use models for open-vocabulary instance segmentation (DECOLA [7]) and monocular depth estimation (Depth Anything [94]) to predict instance masks and depth maps (Fig. 5). From the instance mask, we detect contacting boundaries between different object instances. We sample points on the boundary, and query the predicted depth map along the direction of the inner and outer normal. We infer occlusion if the depth of the inner point exceeds the depth of the outer point. We also incorporate robust estimation methods by dense point sampling and voting. Please see Appendix for more details. + +# 5. Experiments + +We introduce our evaluation results on diverse datasets, including real images in controlled and in-the-wild settings, for comparison with previous work. + +Implementation Details. On the synthetic data, we supervise the model on $n = 4$ renderings. On the real data, we render $m = 4$ novel views to apply the semantic supervision. We use 1 view to apply the pixel-level supervision. All evaluations are performed with a rendering resolution of 224. We set the learning rate as $4e - 5$ using the AdamW optimizer [47]. We train the model with $j = 40,000$ iterations with a cosine learning rate scheduler. We use a batch size of 80, where we have half synthetic samples and half real-data samples. We set $\lambda, \lambda^{R}, \lambda_{pix}^{R}, \lambda_{sem}^{R}$ as 1.0, 0.3, 5.0 and 1.0. Please see more details in the Appendix. + +Datasets. We train Real3D on our collected WildImages real single-view data and the synthetic multi-view renderings of Objaverse [9] jointly. We evaluate the models on test splits of WildImages, MVImageNet, CO3D, and OmniObject3D. + +- WildImages contains 300K single-view segmented + +![](images/a4fd8d3e3b6602cb50fd404ab158ec4bf2746af55e487c5d8d62b3bd7976a1fa.jpg) +Figure 6. Comparison with TrioSR. We compare Real3D with the base model TrioSR. Real3D uses real images for training and performs better on real images. Red boxes highlight failure cases of TrioSR due to the limited data coverage of its synthetic training data. + +objects, collected from datasets in diverse domains [12, 38, 98], which is filtered from more than 3M real object instances. We keep aside a test split with 1000 images. + +- Objaverse [9] is a large-scale synthetic dataset. We use a filtered subset consisting of 260K high-quality shapes for training. We use renderings from [42, 59]. We note this is a widely used subset [59, 73], as Objaverse contains many low-quality instances, which harm learning. +- MVImgNet [98], CO3D [62] and OmniObject3D [87] are used for evaluation, containing real data and out-of-domain synthetic data. + +Metrics. We use two sets of metrics to evaluate the models on data with and without multi-view annotations: + +- Novel View Synthesis (NVS) Metrics. Following prior work, we evaluate reconstructions on data with multiview information, using standard metrics, including PSNR, SSIM [82], and LPIPS [100]. +- Semantic and Self-Consistency Metrics. To evaluate reconstruction quality on single-view data without ground-truth multi-view images, we introduce novel metrics. First, we render novel views of a reconstruction and measure the semantic similarity with the input view. In detail, we render 7 views where the azimuths are uniformly sampled in range [0, 360] and no elevations. We use semantic metrics, including LPIPS [100], CLIP similarity [60], and FID score [22]. Second, we evaluate the self-consistency of the reconstruction. We render a designated novel view of the reconstruction, use it as input for the LRM, and render the second reconstruction from the viewpoint of the original input. We evaluate the consistency using NVS metrics. + +- Mesh Quality Metrics. We report the mesh quality using the Chamfer-L1 Distance (CD) with mesh scale 2.0. To get mesh from the triplane representation, we use marching cubes with resolution 256 for TripoSR and Real3D. + +Baselines. We compare Real3D with OpenLRM [20], TripoSR [75], LGM [73], CRM [83] and InstantMesh [91]. We use OpenLRM [20], an open-sourced LRM for comparisons, as LRM is close-sourced. All models are trained on Objaverse [9] unless noted otherwise. Specifically, InstantMesh uses the larger synthetic dataset Objaverse-XL [10]. Moreover, we finetune TripoSR, as it predicts reconstruction with random scales on different inputs with non-clean backgrounds, which leads to inferior evaluation results (see Appendix) and failure of self-training. TripoSR is a base model directly comparable to ours1. For all baselines, we use the official checkpoints. Furthermore, we follow the specific settings of each model to normalize target camera poses. + +# 5.1. Experimental Results + +Qualitative Results. We show Real3D reconstruction examples in Fig. 6, showing that Real3D can recover the geometry with high fidelity. Compared with the baseline method TripoSR, Real3D demonstrates less reconstruction error on objects from uncommon categories or under uncommon camera poses that lead to ambiguity. Specifically, Real3D shows better accuracy for reconstructing side-view and back of objects. This performance gain shows that Real3D leverages the semantic knowledge learned from large-scale real data to resolve shape ambiguity, verifying the effectiveness of our self-training and data curation methods. Please also see Appendix for comparisons with baselines on evaluation sets. + +Quantitative Results. Real3D consistently outperforms prior works on all four test sets of our evaluation. As shown in Table 1 to Table 4, Real3D showcases a 0.74 dB PSNR + +
MethodEval. on Input ViewEval. on GT Novel Views
Semantic SimilaritySelf-ConsistencyNovel View Synthesis Quality
CLIP↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
METHODS W. GENERATIVE PRIORS
LGM [73]0.8200.204158.414.200.8330.22715.950.8130.181
CRM [83]0.8230.179172.515.720.8730.16817.540.8530.142
InstantMesh [91]0.8730.188153.514.810.8610.17114.700.8060.197
METHODS W/O GENERATIVE PRIORS (DETERMINISTIC)
OpenLRM [20]0.8680.160147.617.690.8730.14019.750.8640.112
TripoSR [75]0.8600.157129.717.720.8740.14319.810.8640.116
Real3D (ours)0.8920.147116.919.800.8930.12520.530.8710.107
+ +Table 1. Evaluation results on the real-world MVImageNet dataset. We evaluate the performance both on input view and novel views. We bold and highlight the best results. + +
MethodEval. on GT Novel Views
Novel View Synthesis Quality
PSNR↑SSIM↑LPIPS↓
METHODS W. GENERATIVE PRIORS
LGM [73]15.140.8020.187
CRM [83]16.380.8400.153
InstantMesh [91]13.990.7890.199
METHODS W/O GENERATIVE PRIORS
OpenLRM [20]18.310.8490.126
TripoSR [75]18.440.8480.127
Real3D (ours)19.180.8550.119
+ +Table 2. Evaluation of novel view synthesis quality on real-world CO3D data. + +
MethodEval. on Input View
Semantic SimilaritySelf-Consistency
CLIP↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓
METHODS W. GENERATIVE PRIORS
LGM [73]0.8430.188146.014.130.8250.210
CRM [83]0.8070.174162.415.670.8620.160
InstantMesh [91]0.8410.182152.514.680.8540.163
METHODS W/O GENERATIVE PRIORS
OpenLRM [20]0.8470.149144.918.090.8720.129
TripoSR [75]0.8770.148128.518.180.8740.125
Real3D (ours)0.8920.144106.519.000.8820.117
+ +Table 3. Evaluation results on WildImages test set. Due to the absence of ground-truth novel views, we perform evaluation on the input view. + +
MethodEval. on GT Novel Views
Novel View Synthesis Quality
PSNR↑SSIM↑LPIPS↓
METHODS W. GENERATIVE PRIORS
LGM [73]15.830.7910.197
CRM [83]16.750.8230.182
InstantMesh [91]15.830.7910.197
METHODS W/O GENERATIVE PRIORS
OpenLRM [20]18.200.8310.144
TripoSR [75]19.430.8470.128
Real3D (ours)20.170.8550.119
+ +Table 4. Evaluation of novel view synthesis quality on synthetic out-of-domain OmniObject3D data. + +(4% relatively, equivalent to 16% better MSE) improvement, a 6.3% relative LPIPS improvement, and a 13.5% relative FID improvement on average over the directly comparable TripoSR model, demonstrating the effectiveness of our self-training method using real data. Results in Table 3 also highlight the advantage of our self-training method by using a broader data distribution to handle in-the-wild test data. + +Additionally, we observe that methods with generative priors do not perform well on out-of-distribution data. These methods generate novel views and use those views to perform sparse-view reconstruction. We conjecture that the reason is the compounding error of the novel view synthesis and reconstruction stages. This is another argument in favor of single-stage methods, like ours. + +Mesh Quality. We report that the CD for InstantMesh (the best baseline with generative prior) and TripoSR (the best deterministic baseline) are 0.395 and 0.321 respectively. In contrast to that, our method Real3D achieves a CD of 0.275, an improvement by $14.3\%$ . We include the visualization of the mesh reconstructions in Appendix. We observe that both baselines perform worse than Real3D, particularly in cases with non-common object shapes, while InstantMesh specifically struggles to faithfully reconstruct thin structures. + +# 5.2. Ablation Studies + +With our ablation studies, we analyze the proposed losses and examine key design choices for semantic- and pixel-level supervision. Additionally, we evaluate the effectiveness of + +our data curation procedure. + +Ablation of proposed losses. As shown in Table 5, when we use only the $\mathcal{L}_{in}^{R}$ loss, we observe slight improvements of PSNR, but SSIM and LPIPS have limited improvements. This pattern implies that $\mathcal{L}_{in}^{R}$ can only help the model render more realistic pixels by learning the real-image pixel distribution, but it can not improve 3D reconstruction quality. Further adding the proposed semantic- and pixel-level unsupervised losses helps improve the performance. + +Semantic-level Supervision Designs. As shown in Table 6, naively applying the CLIP-based semantic loss on all rendered novel views degrades the performance. We present visualization results in Appendix, observing the multi-head problem of reconstructions. We conjecture that the reason is that copying the input view to all other views is a trivial solution to minimize the loss. This motivates us to incorporate regularization as we do with our semantic supervision, which improves the performance across all metrics. + +Pixel-level Supervision Designs. As shown in Table 7, the pixel-level supervision is only useful when stopping the gradient of intermediate rendering and applying a training curriculum. The result demonstrates the importance of each proposed component. We present more visualization for the ablation in Appendix. + +Effectiveness of Data Curation. As shown in Table 8, simply training with all data leads to smaller performance gains, compared to using clean data filtered from real images + +
\( {\mathcal{L}}_{\text{in }}^{R} \)\( {\mathcal{L}}_{\text{sem }}^{R} \)\( {\mathcal{L}}_{\text{pix }}^{R} \)Eval. on GT Novel Views Novel View Synthesis Quality
\( PSNR_{\uparrow} \)\( SSIM_{\uparrow} \)\( LPIPS_{\downarrow} \)
XXX18.440.8480.127
✓XX18.600.8500.127
✓✓X18.810.8530.125
✓✓✓19.180.8550.119
+ +Table 5. Ablation study of proposed losses, including the input view rendering loss $\left( {\mathcal{L}}_{\text{in }}^{R}\right)$ , semantic-level supervision $\left( {\mathcal{L}}_{\text{sem }}^{R}\right)$ and pixellevel supervision $\left( {\mathcal{L}}_{\text{pix }}^{R}\right)$ . + +
Eval. on GT Novel Views
Novel View Synthesis Quality
PSNR↑SSIM↑LPIPS↓
No sem. loss18.600.8500.127
Naive sem. loss17.890.8300.151
\( \mathcal{L}_{\text{sem}}^R \)18.810.8530.125
+ +Table 6. Ablation study of semantic-level supervision. "Naive" means simply applying CLIP-semantic loss on all novel views. We do not apply the pixel-level supervision $(\mathcal{L}_{\mathrm{pix}}^{R})$ for these experiments. + +
Stop Grad.Curri-culumEval. on GT Novel Views
Novel View Synthesis Quality
PSNR↑SSIM↑LPIPS↓
X✓17.780.8210.140
✓X18.630.8480.125
✓✓19.180.8550.119
+ +Table 7. Ablation study of pixel-level supervision with cycle-consistency. "Stop Grad." means stopping the gradient on intermediate rendering results. ${\mathcal{L}}_{\text{sem }}^{R}$ is incorporated for these experiments. + +
Eval. on GT Novel Views
Novel View Synthesis Quality
PSNR↑SSIM↑LPIPS↓
All data18.790.8520.123
Clean data19.180.8550.119
+ +Table 8. Ablation study of data cur- ration for getting clean data. All losses are incorporated for these ex-periments + +
MethodEval. on GT Novel ViewEval. on Input View
MVIImgNetCO3DOmniObect3DWildImages
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Δmulti-view0.4100.0030.0070.5100.0030.0070.3300.0060.0060.2600.0000.000
Δours0.7200.0070.0090.7400.0070.0080.7400.0080.0090.7200.0080.008
+ +Table 9. Performance gain comparison between: i) using multi-view real data with multi-view rendering losses ( $\Delta$ multi-view) and ii) using our real images with self-training losses ( $\Delta$ ours). $\Delta$ multi-view uses 260K MVImgNet videos with about 8 million images. In contrast, our method uses 300K high-quality images selected from about 3 million real images for self-training. + +![](images/60b95cb2ef840a6f3aaf01117e87c4d33cd3252431f81d89b42fad4a934b8154.jpg) +(a) MVImgNet + +![](images/380a6f79584c66d6a63ab65dca6bec38f37a4d13b77f4a4f4d501828b1c1e567.jpg) +(b) CO3D + +![](images/fc11afd6d81009ebfe2208972a95d1a74c7322526be8c2907e7ff2d496a0c317.jpg) +(c) OmniObject3D +Figure 7. Real3D performance (PSNR) using different amounts of real data for training. The PSNR is evaluated on novel views for (a)-(c), and it is evaluated on (d) with self-consistency. + +![](images/30a33bb7400c393b5c9172c4da9b0ba89320aa35282eb2fe5028c87a9f9522aa.jpg) +(d) WildImages + +with our method. This highlights both the importance of data curation and the effectiveness of our method. + +# 5.3. Scalability and Effectiveness of Self-Training + +We provide experimental results to further understand the scalability and effectiveness of our self-training method. + +Scalability - Data Amount Analysis. Here, we evaluate the effect of scaling up the training data. As shown in Fig. 7, Real3D achieves consistent improvements when more real images are incorporated for training. This performance gain curve demonstrates the potential benefits of further scaling up the real data for our self-training method. + +Effectiveness - Performance Gain Analysis. We intend to further validate the effectiveness of using single-view images for self-training. For this purpose, we compare the + +performance gain of using self-training with another method to scale up the training data, i.e., using captured multi-view real images. We report the performance gain of training with multi-view images by comparing LRM variants, i.e., an LRM when trained on synthetic data only and when trained with both synthetic and real-world MVImgNet multi-view capture [20]. We report this gain as $\Delta$ multi-view. We also report the gain of our self-training method as $\Delta$ ours. We report the results on the four datasets. As shown in Table 9, our method achieves a larger performance gain while using much fewer images for training, showing its effectiveness. The limited improvement of $\Delta$ multi-view also indicates the limitation of leveraging multi-view real data for training. We include a detailed discussion and report the detailed numbers for computing $\Delta$ multi-view in Appendix. + +More interestingly, $\Delta$ multi-view uses only MVImgNet data for training and has limited, nearly zero, improvement on in-the-wild images. In contrast, our method achieves larger improvements by leveraging more in-the-wild data, demonstrating its success at improving generalization. + +# 6. Conclusion + +We present Real3D, the first LRM that can leverage single-view real images for training. This has the major advantage of enabling training on a seemingly endless data source, which is representative of the general object shape distribution. We propose a self-training framework using unsupervised losses to leverage these images. Additionally, we develop an automatic data curation method to deal with the noisy real-world data. Compared to previous work, Real3D shows consistent improvements in diverse evaluation sets and highlights the potential to improve LRMs by training in large-scale image collections. + +Acknowledgment. We acknowledge the support of NSF-2047677, NSF-2413161, NSF-2504906, NSF-2515626, and Gifts from Adobe and Google. We would like to thank UT GenAI Center for generous support of GPU hours. + +# References + +[1] Kalyan Vasudev Alwala, Abhinav Gupta, and Shubham Tulsiani. Pre-train, self-train, distill: A simple recipe for supersizing 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3773-3782, 2022. +[2] 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. +[3] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. +[4] 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. +[5] 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. +[6] 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. +[7] Jang Hyun Cho and Philipp Krahenbuhl. Language-conditioned detection transformer. arXiv preprint arXiv:2311.17902, 2023. +[8] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 628-644. Springer, 2016. +[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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. +[10] 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. +[11] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. + +Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20637-20647, 2023. +[12] 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. +[13] Shivam Duggal and Deepak Pathak. Topologically-aware deformation fields for single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1536-1546, 2022. +[14] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. +[15] 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. +[16] Georgia Gkioxari, Jitendra Malik, and Justin Johnson. Mesh r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9785-9795, 2019. +[17] Jean-Bastien Grill, Florian Strub, Florent Altché, Coretin 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. +[18] Junlin Han, Filippos Kokkinos, and Philip Torr. Vfusion3d: Learning scalable 3d generative models from video diffusion models. In European Conference on Computer Vision, pages 333-350. Springer, 2025. +[19] 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. +[20] Zexin He and Tengfei Wang. Openlrm: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. +[21] Paul Henderson, Vagia Tsiminaki, and Christoph H Lampert. Leveraging 2d data to learn textured 3d mesh generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7498-7507, 2020. +[22] 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. +[23] 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. +[24] Tao Hu, Liwei Wang, Xiaogang Xu, Shu Liu, and Jiaya Jia. Self-supervised 3d mesh reconstruction from single images. + +In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6002-6011, 2021. +[25] Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M Rehg. Zeroshape: Regression-based zero-shot shape reconstruction. arXiv preprint arXiv:2312.14198, 2023. +[26] Hanwen Jiang, Zhenyu Jiang, Yue Zhao, and Qixing Huang. Leap: Liberate sparse-view 3d modeling from camera poses. arXiv preprint arXiv:2310.01410, 2023. +[27] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown categories and camera poses. In 2024 International Conference on 3D Vision (3DV), pages 31-41. IEEE, 2024. +[28] Hanwen Jiang, Hao Tan, Peng Wang, Haian Jin, Yue Zhao, Sai Bi, Kai Zhang, Fujun Luan, Kalyan Sunkavalli, Qixing Huang, et al. Rayzer: A self-supervised large view synthesis model. arXiv preprint arXiv:2505.00702, 2025. +[29] Hanwen Jiang, Zexiang Xu, Desai Xie, Ziwen Chen, Haian Jin, Fujun Luan, Zhixin Shu, Kai Zhang, Sai Bi, Xin Sun, et al. Megasynth: Scaling up 3d scene reconstruction with synthesized data. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16441-16452, 2025. +[30] Li Jiang, Shaoshuai Shi, Xiaojuan Qi, and Jiaya Jia. Gal: Geometric adversarial loss for single-view 3d-object reconstruction. In Proceedings of the European conference on computer vision (ECCV), pages 802-816, 2018. +[31] Ying Jin, Jiaqi Wang, and Dahua Lin. Semi-supervised semantic segmentation via gentle teaching assistant. Advances in Neural Information Processing Systems, 35:2803-2816, 2022. +[32] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. +[33] 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. +[34] Abhishek Kar, Shubham Tulsiani, Joao Carreira, and Jitendra Malik. Category-specific object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1966-1974, 2015. +[35] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. +[36] 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. +[37] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 452-461, 2020. + +[38] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. +[39] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 677-693. Springer, 2020. +[40] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In European Conference on Computer Vision, pages 515-534. Springer, 2022. +[41] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. Sdfsrn: Learning signed distance 3d object reconstruction from static images. Advances in Neural Information Processing Systems, 33:11453-11464, 2020. +[42] 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. +[43] 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. +[44] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. +[45] 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. +[46] 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. +[47] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. +[48] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8446-8455, 2023. +[49] 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. + +[50] Lu Mi, Abhijit Kundu, David Ross, Frank Dellaert, Noah Snavely, and Alireza Fathi. im2nerf: Image to neural radiance field in the wild. arXiv preprint arXiv:2209.04061, 2022. +[51] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. +[52] 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. +[53] Tom Monnier, Matthew Fisher, Alexei A Efros, and Mathieu Aubry. Share with thy neighbors: Single-view reconstruction by cross-instance consistency. In European Conference on Computer Vision, pages 285-303. Springer, 2022. +[54] KL Navaneet, Ansu Mathew, Shashank Kashyap, Wei-Chih Hung, Varun Jampani, and R Venkatesh Babu. From image collections to point clouds with self-supervised shape and pose networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1132-1140, 2020. +[55] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588-7597, 2019. +[56] 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. +[57] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. +[58] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. +[59] 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. arXiv preprint arXiv:2311.16918, 2023. +[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. +[61] Ilija Radosavovic, Piotr Dollar, Ross Girshick, Georgia Gkioxari, and Kaiming He. Data distillation: Towards omnisupervised learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4119-4128, 2018. + +[62] 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. +[63] Konstantinos Rematas, Ricardo Martin-Brualla, and Vittorio Ferrari. Sharf: Shape-conditioned radiance fields from a single view. arXiv preprint arXiv:2102.08860, 2021. +[64] 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. +[65] Kyle Sargent, Jing Yu Koh, Han Zhang, Huiwen Chang, Charles Herrmann, Pratul Srinivasan, Jiajun Wu, and Deqing Sun. Vq3d: Learning a 3d-aware generative model onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4240-4250, 2023. +[66] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, 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. +[67] Henry Scudder. Probability of error of some adaptive pattern-recognition machines. IEEE Transactions on Information Theory, 11(3):363-371, 1965. +[68] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. +[69] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems, 32, 2019. +[70] Ivan Skorokhodov, Aliaksandr Siarhin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. arXiv preprint arXiv:2303.01416, 2023. +[71] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. arXiv preprint arXiv:2312.13150, 2023. +[72] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22819–22829, 2023. +[73] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. +[74] 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. Communications of the ACM, 59(2):64-73, 2016. +[75] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian + +Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. +[76] 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. +[77] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. +[78] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2626-2634, 2017. +[79] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. +[80] 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. +[81] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. +[82] 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. +[83] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. +[84] 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. +[85] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9065-9075, 2023. +[86] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. +[87] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. + +[88] Jianfeng Xiang, Jiaolong Yang, Binbin Huang, and Xin Tong. 3d-aware image generation using 2d diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2383-2393, 2023. +[89] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. +[90] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In European Conference on Computer Vision, pages 736-753. Springer, 2022. +[91] 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. +[92] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. +[93] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10368-10378, 2021. +[94] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891, 2024. +[95] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. +[96] Yufei Ye, Shubham Tulsiani, and Abhinav Gupta. Shelf-supervised mesh prediction in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8843-8852, 2021. +[97] 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. +[98] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. +[99] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. arXiv preprint arXiv:2404.19702, 2024. +[100] 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. +[101] Yue Zhao, Long Zhao, Xingyi Zhou, Jialin Wu, Chun-Te Chu, Hui Miao, Florian Schroff, Hartwig Adam, Ting Liu, Boqing Gong, et al. Distilling vision-language models on millions of videos. arXiv preprint arXiv:2401.06129, 2024. +[102] 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. arXiv preprint arXiv:2312.09147, 2023. \ No newline at end of file diff --git a/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/images.zip b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c2785a97410d86fc9291819018981ef2b5b7ddf1 --- /dev/null +++ b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a26cb0d838672089fdbb08f926865263377da7f935e59c1d9bc8cfc0097d18fb +size 634693 diff --git a/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/layout.json b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7adb1e8bec9e9c21b9a45f3277801abf2349dbc5 --- /dev/null +++ b/ICCV/2025/Real3D_ Towards Scaling Large Reconstruction Models with Real Images/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ba3b580114cf56cb65ff18a53de1b7c2895919cb24825cce0af3446b3393b26 +size 548195 diff --git a/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/b63fbb92-0f5b-4445-8e06-8024ee574ac9_content_list.json b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/b63fbb92-0f5b-4445-8e06-8024ee574ac9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..db2f834029f6f64cb7d5626d937053e525a8c14b --- /dev/null +++ b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/b63fbb92-0f5b-4445-8e06-8024ee574ac9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:975e596d2e12fc768c9e77364fc7e823fe872222792345cf576d5518abb7fd0d +size 93050 diff --git a/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/b63fbb92-0f5b-4445-8e06-8024ee574ac9_model.json b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/b63fbb92-0f5b-4445-8e06-8024ee574ac9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e0f3b2ec98c9f57e1ef88063a61805b035576a3a --- /dev/null +++ b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/b63fbb92-0f5b-4445-8e06-8024ee574ac9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a9f86a732a663418e08196fbf4303ef2400a2c79341762edfabda9b5aa4cc16 +size 122906 diff --git a/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/b63fbb92-0f5b-4445-8e06-8024ee574ac9_origin.pdf b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/b63fbb92-0f5b-4445-8e06-8024ee574ac9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3d1e2b030f6c2803d1194fff55826144e2d924a7 --- /dev/null +++ b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/b63fbb92-0f5b-4445-8e06-8024ee574ac9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a2bac54d056fe23bb2f78a3c7b01b3eb72becbe023f6e294604982332c2649f +size 4810321 diff --git a/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/full.md b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d0f7e706a97fc124537afe80d0d5d2269590346a --- /dev/null +++ b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/full.md @@ -0,0 +1,406 @@ +# RealCam-I2V: Real-World Image-to-Video Generation with Interactive Complex Camera Control + +Teng Li $^{1*}$ , Guangcong Zheng $^{1*}$ , Rui Jiang $^{1}$ , Shuigen Zhan $^{1}$ , Tao Wu $^{1}$ , Yehao Lu $^{1}$ , Yining Lin $^{3}$ , Chuanyun Deng $^{2}$ , Yepan Xiong $^{2}$ , Min Chen $^{2}$ , Lin Cheng $^{2}$ , and Xi Li $^{1\dagger}$ + +1 College of Computer Science & Technology, Zhejiang University +2 Central Media Technology Institute, 2012 Lab, Huawei +3 Supremind + +![](images/d2862e44bbb090478e87680c8b9caf1d68259460db74cac507753798e1f84822.jpg) +Figure 1. We propose RealCam-12V, a camera controllable image-to-video generation framework for complex real-world camera control and extra applications including camera-controlled loop video generation, generative frame interpolation, and smooth scene transitions. + +# Abstract + +Recent advancements in camera-trajectory-guided image-to-video generation offer higher precision and better support for complex camera control compared to text-based approaches. However, they also introduce significant usability challenges, as users often struggle to provide precise camera parameters when working with arbitrary real-world images without knowledge of their depth nor scene scale. To address these real-world application issues, we propose RealCam-I2V, a novel diffusion-based video generation framework that integrates monocular metric depth estimation to establish 3D scene reconstruction in a preprocessing step. During training, the reconstructed 3D + +scene enables scaling camera parameters from relative to metric scales, ensuring compatibility and scale consistency across diverse real-world images. In inference, RealCam-I2V offers an intuitive interface where users can precisely draw camera trajectories by dragging within the 3D scene. To further enhance precise camera control and scene consistency, we propose scene-constrained noise shaping, which shapes high-level noise and also allows the framework to maintain dynamic and coherent video generation in lower noise stages. RealCam-I2V achieves significant improvements in controllability and video quality on the RealEstate10K and out-of-domain images. We further enables applications like camera-controlled looping video generation and generative frame interpolation. Project page: zgctroy.github.io/RealCam-I2V. + +# 1. Introduction + +Recent advancements in image-to-video generation [9, 10, 20, 24, 82] have significantly improved controllability over synthesized videos. However, challenges remain in achieving realistic, controllable camera movement within complex real-world scenes. Text-based camera-control methods [4, 20, 29, 31, 38, 66], like traditional diffusion-based video generation, are intuitive and straightforward but lack precision in explicit control over camera parameters, such as angle, scale, and movement direction. This limitation has spurred the development of camera-trajectory-guided approaches, which attempt to address these issues by offering finer control over camera movement. + +Current camera-trajectory-guided methods typically rely on relative camera trajectories, as seen in models like MotionCtrl [69], CameraCtrl [22], CamCo [78], and CamI2V [96]. While these approaches provide more control than text-based models, they are fundamentally limited by their reliance on relative scale trajectories. Training on relative scales results in inconsistencies when applied to real-world scenes, where metric scale is crucial for realistic depth perception. Additionally, without access to depth information, users find it challenging to draw precise trajectories, making these methods difficult to use effectively. + +To overcome these limitations, we propose RealCam-I2V, a video generation framework that integrates monocular depth estimation as a preprocessing step to construct a robust, metric-scale 3D scene. Our approach leverages the Depth Anything v2 [80] model (metric version) to predict the metric depth of a user-provided reference image, reprojecting its pixels back into camera space to create a stable 3D representation. This 3D scene serves as the foundation for camera control, providing a consistent and metric scale that is critical for real-world applications. + +In the training stage, we align the reconstructed 3D scene of the reference image with the point cloud of each video sample, reconstructed using COLMAP [54], a structure-from-motion (SfM) method. This alignment allows us to rescale COLMAP-annotated camera parameters to the Depth Anything metric, providing an metric, stable, and robust scale across training data. By aligning relative-scale camera parameters to metric scales, we can condition the video generation model on accurately scaled camera trajectories, achieving greater control and scene consistency across diverse real-world images. + +During inference, RealCam-I2V provides an interactive interface where users can intuitively design camera trajectories by drawing within the reconstructed 3D scene of the reference image. This interface renders preview videos of the trajectory in a static scene, offering users real-time feedback and greater control over camera movement. This interactive feature enhances usability, allowing precise trajectory control even for users without specialized knowledge of scene + +depth. To further improve video quality and control precision, we introduce scene-constrained noise initialization as a mechanism to shape the camera movement in its high-noise stages. By using the preview video of the static 3D scene, RealCam-I2V injects scene-visible regions with controlled noise, guiding the video diffusion model's early generation stages. This high-noise feature constrains the initial layout and camera dynamics, providing a strong foundation for the remaining denoising stages. As denoising progresses, the condition-based approach, trained on metric-scale camera trajectories, preserves global layout and completes the dynamic scene in previously unseen regions. This approach maintains the video diffusion model's capacity for dynamic content generation while ensuring accurate, coherent camera control. + +Our experimental results show that RealCam-I2V achieves significant performance gains in video quality and controllability. When relative scales are aligned to metric scales, models such as MotionCtrl, CameraCtrl, and CamI2V see substantial improvements in video quality. Furthermore, with the introduction of scene-constrained noise initialization, RealCam-I2V surpasses state-of-the-art performance benchmarks, particularly on datasets like RealEstate10K [100] and out-of-domain images. These results demonstrate the effectiveness of our approach in both controlled and diverse real-world settings. In summary, our contributions are as follows: + +- We identify scale inconsistencies and real-world usability challenges in existing trajectory-based methods and introduce a simple yet effective monocular 3D reconstruction into the preprocessing step of the generation pipeline, serving as a reliable intermediary reference for both training and inference. +- With reconstructed 3D scene, we enable metric-scale training and provide an interactive interface during inference to easily design camera trajectories with preview feedback, along with proposed scene-constrained noise shaping to significantly enhance scene consistency and camera controllability. +- Our method overcomes critical real-world application challenges and achieves substantial improvements on the RealEstate10K dataset, establishing a new sota benchmark both in video quality and control precision. + +# 2. Related Works + +Diffusion-based Video Generation. The advancements in diffusion models [52, 53, 94] have led to significant progress in video generation, gradually forming two major categories of foundational tasks: text-to-video and imaged-to-video. Due to the lack of high-quality video-text datasets [4, 5], previous outstanding works such asAnimateDiff [20], Align Your Latents [5], PYoCo [17], Emu Video [18], LVDM [24], VideoCrafter [9, 10], ModelScope [62], + +![](images/0080dbcb09349487678ec1b40b18a35b0d5e56d68ca30e1aad4e57048ba39f77.jpg) +Figure 2. RealCam-I2V pipeline. For training, we align camera parameters from relative scale to metric scale. For inference, we use metric depth estimation to construct the point cloud for users to interactively draw the camera trajectory. Due to the metric scale alignment, the user-given camera trajectory in the 3D scene shares the same scene scale as those in real world. + +LAVIE [67], and VideoFactory [65] have facilitated text-to-video generation by incorporating Motion Blocks into existing text-to-image [51, 53, 85] models. Building on these T2V efforts, several works introduced images as control signals, leading to a series of I2V models. Representative examples include SVD [4], SEINE [13], Pixel-Dance [87], PIA [92], I2VGen-XL [91], DynamicCrafter [76], and Moonshot [88]. With the improvement in video data [12, 34, 47, 63] and the rapid development of the DiT architecture [42, 49, 86], recent works such as CogVideoX [82], Sora [6], HunyuanVideo [35], GoKu [11], Open-Sora [98], Open-Sora-Plan [40], Vidu [3], Lumina-T2X [16], Vchitect-2.0 [14], RepVideo [55] and Step-Video-T2V[41] have significantly enhanced the quality of video generation. These advancements demonstrate the potential of Video Diffusion Models as realistic world simulators. + +Controllable Generation. Controllable generation has become a central focus in both image [32, 45, 50, 57, 72, 75, 83, 90, 95] and video [19, 21, 33, 88] generation, enabling users to direct the output through various types of control. A wide range of controllable inputs has been explored, including text descriptions, pose [27, 43, 64, 79], audio [23, 58, 59], identity representations [7, 37, 68, 73, 74, 93], trajectory [8, 39, 46, 71, 84]. + +Text-based Camera Control. Text-based camera control methods use natural language descriptions to guide camera motion in video generation.AnimateDiff [20] and SVD [4] fine-tune LoRAs [26] for specific camera movements based on text input. Image conductor[38] proposed to separate different camera and object motions through camera LoRA weight and object LoRA weight to achieve more precise motion control. In contrast, MotionMaster [29] and Peekaboo [31] offer training-free approaches for generating coarse-grained camera motions, though with limited precision. VideoComposer [66] adjusts pixel-level motion vectors to provide finer control, but challenges remain in + +achieving precise camera control. + +Trajectory-based Camera Control. MotionCtrl [69], CameraCtrl [22], and Direct-a-Video [81] use camera pose as input to enhance control, while CVD [36] extends CameraCtrl for multi-view generation, though still limited by motion complexity. To improve geometric consistency, Pose-guided diffusion [60], CamCo [78], and CamI2V [96] apply epipolar constraints for consistent viewpoints. VD3D [2] introduces a ControlNet[90]-like conditioning mechanism with spatiotemporal camera embeddings, enabling more precise control. CamTrol [25] offers a training-free approach that renders static point clouds into multi-view frames for video generation. Cavia [77] introduces view-integrated attention mechanisms to improve viewpoint and temporal consistency, while I2VControl-Camera [15] refines camera movement by employing point trajectories in the camera coordinate system. Recently, 4DiM [70] and AC3D [1] also leverage monocular metric depth estimator to tackling the issues of scale inconsistency. Despite these advancements, challenges in maintaining camera control and scene-scale consistency remain, which our method seeks to address. + +# 3. Method + +Our overall pipeline is shown in Fig. 2. For training, we align camera parameters from relative scale to metric scale. For inference, we use metric depth estimation to construct the point cloud for users to interactively draw the camera trajectory. Due to the metric scale alignment, the user-given camera trajectory in the 3D scene shares the same scene scale as those in real world. + +# 3.1. Camera-controlled Image-to-Video Model + +Instead of directly modeling the video $x$ , the latent representation $z = \mathcal{E}(x)$ is used for training. The diffusion model $\epsilon_{\theta}$ learns to estimate the noise $\epsilon$ added at each timestep $t$ , con + +![](images/3148199296b6a24bea477f8c7b877c85b31c80dd21e5aac686e13f7953802796.jpg) +Figure 3. Scene scale mismatch. Point clouds reconstructed from metric depth estimation (RGB) are robust and unified, whereas SfM reconstructions (yellow) are relative-scale that may vary across frames. Our alignment enables relative-to-metric conversion of scene scale for real-world applications. + +![](images/c10ce943fa7f1cb71fb391006cfafacec188114eb363ae0b167de659b9021ebb.jpg) +Figure 4. Camera trajectory ambiguity. The relative scene-scale measurement fundamentally hinders models from learning physically consistent camera motion. + +ditioned on both a text prompt $c_{\mathrm{txt}}$ , a reference image $c_{\mathrm{img}}$ and camera condition $c_{\mathrm{cam}}$ , with $t \in \mathcal{U}(0,1)$ . The training objective simplifies to a reconstruction loss defined as: + +$$ +\mathcal {L} = \mathbb {E} _ {z, c _ {\mathrm {t x t}}, c _ {\mathrm {i m g}}, c _ {\mathrm {c a m}}, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, c _ {\mathrm {t x t}}, c _ {\mathrm {i m g}}, c _ {\mathrm {c a m}}, t\right) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $z \in \mathbb{R}^{F \times H \times W \times C}$ represents the latent code of a video, with $F, H, W, C$ corresponding to frame count, height, width, and channel dimensions. The noise-corrupted latent code $z_{t}$ , derived from the ground-truth latent $z_{0}$ , is expressed as: + +$$ +z _ {t} = \alpha_ {t} z _ {0} + \sigma_ {t} \epsilon , \tag {2} +$$ + +where $\sigma_{t} = \sqrt{1 - \alpha_{t}^{2}}$ . Here, $\alpha_{t}$ and $\sigma_{t}$ are hyperparameters governing the diffusion process. + +# 3.2. Metric Scene-scale Alignment + +Metric Depth Estimation. To obtain a depth map from a given input image, we use a metric depth predictor $f_{\mathrm{depth}}$ which takes the RGB image $I$ as input and outputs the corresponding depth map $D(u,v)$ . The prediction process is formulated as: + +$$ +D (u, v) = f _ {\text {d e p t h}} (I), +$$ + +where $I$ is the input RGB image and $D(u,v)$ is the predicted depth value for each pixel at coordinates $(u,v)$ . This predicted depth map $D(u,v)$ serves as the foundation for projecting the image into 3D space, allowing us to construct a point cloud in the camera coordinate system. The camera + +![](images/3fe445a347c41e6f453e4bd5ad926d213eb0996585a6c55487f293f80e4af073.jpg) +Figure 5. One-round generation versus multi-round generation. Our framework decouples camera adjustment via interactive 3D scenes, enabling fast feedback before slow generation. + +intrinsic matrix $K$ is defined as: + +$$ +K = \left[ \begin{array}{c c c} f _ {x} & 0 & c _ {x} \\ 0 & f _ {y} & c _ {y} \\ 0 & 0 & 1 \end{array} \right], +$$ + +where $f_{x}$ and $f_{y}$ are the focal lengths along the $x$ and $y$ axes, $(c_{x}, c_{y})$ is the principal point of the camera. + +Metric Scene-Scale Alignment. To convert camera extrinsics from world-to-camera to an metric-scale camera-to-world representation, we defines that the world-to-camera extrinsics matrix $F_{\mathrm{w2c}} \in \mathbb{R}^{4 \times 4}$ is inverted to obtain the corresponding camera-to-world matrix $F_{\mathrm{c2w}} = F_{\mathrm{w2c}}^{-1}$ . To express the transformations relative to the first frame, each $F_{\mathrm{c2w}}$ is left-multiplied by the camera-to-world matrix of the inverse of first frame $F_{\mathrm{c2w},1}$ : + +$$ +c _ {\mathrm {c a m}} = F _ {\mathrm {c 2 w , 1}} ^ {- 1} \cdot F _ {\mathrm {c 2 w}}. +$$ + +Here, $c_{\mathrm{cam}} \in \mathbb{R}^{F \times 4 \times 4}$ represents the camera-to-world transformations aligned relative to the first frame. However, the translation component of $c_{\mathrm{cam}}$ remains in a relative scene scale. To convert the relative translation to an metric scale, we align the metric 3D point cloud reconstructed by Depth Anything (metric version) with the 3D point cloud reconstructed by COLMAP (Structure-from-Motion), as shown in Fig. 3 and Fig. 4. The alignment process yields a scale factor $\alpha$ and is applied to the translation component of $c_{\mathrm{cam}}$ , resulting in an metric-scale camera-to-world transformation: + +$$ +c _ {\mathrm {c a m}} ^ {\mathrm {m e t r i c}} = \left[ \begin{array}{c c} R & \alpha \cdot T \\ 0 & 1 \end{array} \right], +$$ + +where $R$ is the rotation matrix, $T$ is the relative translation vector. The resulting $c_{\mathrm{cam}}^{\mathrm{metric}} \in \mathbb{R}^{F \times 4 \times 4}$ represents the camera-to-world transformations with metric scene scale, enabling robust and accurate real-world applications. + +# 3.3. One-round Generation in Inference + +Users are allowed to draw camera trajectory in the constructed 3D scene and can easily preview the rendering video of expected camera trajectory, free from the costly generation via video diffusion models, as shown in Fig. 5. + +![](images/ff46680a88bbc77938881a36987d6137792256cd9e588e9967372646f20ea479.jpg) +(a) One of the Preview Frames (Camera Views) + +![](images/ba580e23b48295841464a8b4fe0ac00e43ddfdeb03a8c04a579496163a53ec3c.jpg) +(b) Interactive 3D Scene + +![](images/cf4cb2c73c4d4f297b0fde94630f02e48f4b3c510a842976b4e736a03405affd.jpg) +(c) One of the Preview Frames (Camera Views) + +![](images/1d61c6ac50b703ebe9929f555053e1138ac877b3ec0a369512aa43f23bc3c11d.jpg) +Invisible Pixel + +Noise Shaping? + +Visible Pixel + +YES + +without Invisible Pixels in 5x5 + +YES + +Visible Pixel + +with Invisible Pixels in 5x5 + +NO + +Noise Shaping: + +Paste to the clean part of noised latents + +3D Point Projection For Interaction. Given a depth map $D(u,v)$ , the projected 3D coordinates in the camera coordinate system, $\mathbf{p}_c = (x_c,y_c,z_c)^T$ , are computed as: + +$$ +\mathbf {p} _ {c} = \mathbf {D} (u, v) \cdot K ^ {- 1} \cdot \left[ \begin{array}{c} u \\ v \\ 1 \end{array} \right]. +$$ + +Here $(u,v,1)$ represents the homogeneous coordinates of the pixel, $K^{-1}$ is the inverse of the intrinsic matrix, which maps pixel coordinates to normalized image coordinates. By applying this transformation to all pixels in the depth map, we obtain a set of 3D points $\{\mathbf{p}_c\}$ in the camera coordinate system. + +Preview Video Rendering for Camera Motion. With constructed 3D point cloud, we can render a preview video of camera traces. As shown in Fig. 7, the preview video functions as a reference video for validation. + +Scene-constrained Noise Shaping. Inspired by previous works SDEdit [44] and DDIM inversion [56], noised features $z_{t}$ can be used for shaping the layout, camera control of the entire image, especially at timestep with high-level noise. We propose scene-constrained noise shaping, which utilizes preview videos generated along user-defined trajectories in the interactive 3D scene. Each frame of the preview video is treated as a reference frame and provided to the generation process during the high-noise stage. The reference frame's pixels are overlaid onto the model-predicted $z_{0}$ to achieve the shaping effect. + +Next, we detail the process for selecting the pixels to be referenced. As illustrated in Fig. 6, the primary criterion is that a pixel must be visible under the current camera viewpoint in the preview video. To mitigate issues such as holes caused by inaccurate depth predictions, we apply an additional filtering rule: if a visible pixel's $k \times k$ neighborhood + +![](images/e634f90026b6ae2d22e3bd75966c0155acfc6c20b928e92e8c17b150f147dc48.jpg) +Figure 7. Preview video rendering for user-expected camera trajectory. It also serves as the reference video in our proposed scene-constrained noise shaping. + +![](images/6539308ab4f7f870d39feca7b1beaa091bf91b2ebccb35be92b5309b13a1486b.jpg) + +![](images/db936ba951d69d54fbfe122490bfdbaf0e6591c60982193d0e1f9c8175471b08.jpg) + +![](images/a21d7f5a67cc90411f17592ca0f058a2ba7e09268378bfc6c644b3f255ff8ddf.jpg) +Figure 6. Pixel selection for noise shaping. Selected pixels are pasted onto the clean part (predicted $z_0$ ) of noised latent $z_t$ . Empirically, applying noise shaping at high noise levels $t > 0.9$ achieves the trade-off of camera control and dynamics. +Figure 8. Concatenation strategies for different tasks. Basic mode, interpolation mode, and continuation mode can be supported with only minor modification. + +contains any invisible pixels, it is considered to lie on an object's edge and potentially affected by depth prediction errors. Such pixels are excluded from selection. Finally, we define the noise shaping process as the formula: + +$$ +z _ {t} = m \cdot \left(\alpha_ {t} z _ {\text {p r e v i e w}} + \sigma_ {t} \epsilon\right) + (1 - m) \cdot z _ {t}, +$$ + +where $m$ identifies the selected reference pixels, $z_{\text{review}}$ represents the clean latent features from the preview video. It should be noted that resampling $\epsilon$ in each timestep is important because the motivation of noise shaping is to utilize the clean feature of preview feature to shape the low-frequency information of the generation process while the random sampled $\epsilon$ may cover some of the valid information in each timestep if it is sampled to be a large value. + +# 3.4. Interpolation, Loop and Continuation + +To support different tasks, including interpolation, looping, and continuation for long video generation, we train video diffusion model with different input concatenation mode, as shown in Fig. 8. Given a video latents $z \in \mathbb{R}^{F \times H \times W \times C}$ , we define the noised latents of $f$ -th frame at timestep $t$ as $z_{t}^{f}$ . We then select $i$ -th clean frame as the condition frame $z_{0}^{i}$ . For interpolation mode, we define $z_{0}^{f-1}$ as the end condition frame. For continuation mode, we define all $1$ $i$ -th as condition frame. + +
MethodRotErr ↓TransErr ↓CamMC ↓FVD ↓
Relative ScaleMetric ScaleRelative ScaleMetric ScaleVideoGPTStyleGAN
DynamiCrafter [76]3.34159.802414.13511.62515.726106.0292.196
MotionCtrl* [69]1.05272.28606.81822.93127.227270.29260.845
CameraCtrl* [22]0.73731.76195.50902.16445.764869.20258.900
CamI2V* [96]0.41201.34093.29341.52913.423062.43953.361
RealCam-I2V (Ours)0.38841.29432.23171.46282.360953.71845.460
w.r.t. CamI2V+5.73%+3.48%+32.24%+4.33%+31.03%+13.97%+14.81%
+ +Table 1. Quantitative comparison with SOTA methods. Our approach excels all baselines on both relative and metric results, while coherently improve visual quality of generated videos. We observe over $30\%$ improvement on metric scale results and over $10\%$ improvement on FVD. * denotes our reproduced results on DynamiCrafter [76]. Best and second best results are highlighted respectively. + +# 4. Experiments + +# 4.1. Setup + +Dataset. We train our model on RealEstate10K [100], which contains 70,000 video clips with well-annotated relative-scale camera poses. For metric scene scale alignment, we run COLMAP [54] point triangulator on each video clip with fixed camera intrinsics and extrinsics directly from RealEstate10K, obtaining the sparse point cloud of the reconstructed scene. For metric depth predictor, we choose Depth Anything V2 [80] Large Indoor, which is fine-tuned on metric depth estimation. We then calculate per-point relative-to-metric scaling factor against metric depth prediction. We term the median value of scaling factors in a frame as the frame-level factor for robustness. To make stable training, we eliminate outliers of video clips whose maximum frame-level scaling factors are among the first $2\%$ with too small values or the last $2\%$ with too large values, assuming sorted in ascending order. The same quantile filtering strategy is applied on the minimum frame-level scaling factors of video clips. It remains 58,000 video clips for training and another 6,000 for testing. + +Implementation Details. We choose DynamiCrafter [76] as our image-to-video (I2V) base model and seamlessly integrate proposed RealCam-I2V into it as a plugin. During depth-aligned training, we freeze all parameters of the base model and the depth predictor, while only parameters of proposed method are trainable. We follow DynamiCrafter to sample 16 frames from each single video clip while perform resizing, keeping the aspect ratio, and center cropping to fit in our training scheme. We train the model with a random frame stride ranging from 1 to 10 and take random condition frame as data augmentation. We fix the frame stride to 8 and always use the first frame as the condition frame for inference. We supervise $\epsilon$ -prediction on the model of $256 \times 256$ resolution and $v$ -prediction on the model of $512 \times 320$ resolution respectively, following the pre-training scheme of DynamiCrafter. We apply the Adam optimizer with a constant learning rate of $1 \times 10^{-4}$ with mixed-precision fp16 and ZeRO-1. + +# 4.2. Metrics + +We follow previous works [22, 69, 78, 96] to evaluate camera-controllability by RotErr, TransErr and CamMC on their estimated camera poses using structure-from-motion (SfM) methods, e.g. COLMAP [54] and GLOMAP [48]. We convert the camera pose of each frame in a video clip to be relative to the first frame as canonicalization. We denote the $i$ -th frame relative camera-to-world matrix of ground truth as $\{R_i^{3\times 3}, T_i^{3\times 1}\}$ , and that of generated video as $\{\tilde{R}_i^{3\times 3}, \tilde{T}_i^{3\times 1}\}$ . We randomly select 1,000 samples from test set for evaluation. We sum up per-frame errors as the scene-level result for camera metrics. Inspired by Zheng et al. [96], we repetitively conduct 5 individual trials on each video clips for camera-control metrics to reduce the randomness introduced by SfM tools. Metrics of one video clip are averaged on successful trials at first for later sample-wise average to get final results. + +RotError [22, 78, 96]. We calculate camera rotation errors by the relative angle between generated videos and ground truths in radians for rotation accuracy. + +$$ +\operatorname {R o t E r r} = \sum_ {i = 1} ^ {n} \arccos \frac {\operatorname {t r} \left(\tilde {R} _ {i} R _ {i} ^ {\mathrm {T}}\right) - 1}{2} \tag {3} +$$ + +TransError [22, 78, 96]. We normalize the camera positions by the scene scale of generated video $\tilde{s}_i$ and growth truth $s_i$ individually for relative TransErr. The scene scale is calculated as the $\mathcal{L}_2$ distance from the first camera to the farthest one for each video clip. For metric TransErr, we normalize both by ground truth video, i.e. $\tilde{s}_i = s_i$ . + +$$ +\operatorname {T r a n s E r r} = \sum_ {i = 1} ^ {n} \left\| \frac {\tilde {T} _ {i}}{\tilde {s} _ {i}} - \frac {T _ {i}}{s _ {i}} \right\| _ {2} \tag {4} +$$ + +CamMC [69]. We perform the same normalization strategy as TransError, and evaluate the overall camera pose accuracy by directly calculating the $\mathcal{L}_2$ similarity. + +$$ +\mathrm {C a m M C} = \sum_ {i = 1} ^ {n} \left\| \left[ \tilde {R} _ {i} \mid \frac {\tilde {T} _ {i}}{\tilde {s} _ {i}} \right] ^ {3 \times 4} - \left[ R _ {i} \mid \frac {T _ {i}}{s _ {i}} \right] ^ {3 \times 4} \right\| _ {2} \tag {5} +$$ + +
MethodMSASNSRotErr ↓TransErr ↓CamMC ↓FVD ↓
Relative ScaleMetric ScaleRelative ScaleMetric ScaleVideoGPTStyleGAN
DynamiCrafter [76]✓3.34159.802414.13511.62515.726106.0292.196
1.51636.63928.46077.21088.950571.94265.014
MotionCtrl* [69]✓1.05272.28606.81822.93127.227270.29260.845
0.86552.33424.22182.80834.598467.13058.311
✓0.63732.07253.23082.37713.472158.88550.111
CameraCtrl* [22]✓0.73731.76195.50902.16445.764869.20258.900
0.70421.94773.82182.30074.082960.31451.918
✓0.54361.79543.18452.03363.362055.00446.702
CamI2V* [96]✓0.41201.34093.29341.52913.423062.43953.361
0.42431.35962.65241.55392.800860.51651.581
RealCam-I2V (Ours)✓✓0.38841.29432.23171.46282.360953.71845.460
+ +Table 2. Ablation study of RealCam-I2V plugins. Metric Scene-scale Alignment (MSA) mitigates scale inconsistency for real-world applications, indicating a more stable and unified camera control. Scene-constrained Noise Shaping (SNS) solely provides substantial improvements on the base model but is less effective than the combined approach (ours). * denotes our reproduced results on Dynami-Crafter [76]. Best and second best results are highlighted respectively. + +
MethodTotal ScoreI2V SubjectI2V BackgroundCamera MotionSubject ConsistencyBackground ConsistencyMotion SmoothnessDynamic DegreeAesthetic QualityImaging Quality
DynamiCrafter [76]84.2295.9395.4322.6795.4498.5498.0834.1558.9862.35
RealCam-I2V (Ours)85.7196.1495.2793.3293.9697.5897.6635.7759.7963.08
w.o. SNS84.3794.7393.4487.4290.9996.2397.3646.7558.3762.91
w.o. MSA83.7794.7192.6591.4891.2496.0697.3536.9958.3263.03
+ +Table 3. Ablation study on Vbench-I2V [30], investigating how camera-conditioned fine-tuning exclusively on static RealEstate10K [100] data affects the generation quality and camera motion of the base model. Notably, even fine-tuned on a static dataset, it preserves dynamics (Dynamic Degree) without compromise. Introducing dynamic datasets will better enhance dynamics, we leave it for future work. + +FVD [61]. We also assess the visual quality of generative videos by the distribution distance between generated videos and ground truths. + +# 4.3. Comparison with SOTA Methods + +We compare RealCam-I2V with models that either lack camera-condition training (DynamiCrafter [76]) or incorporate camera-condition training, namely MotionCtrl [69] $(3\times 4$ camera extrinsics), CameraCtrl [22] (Plucker embedding constructed from camera extrinsic and intrinsic as side input), and CamI2V [96] (current SOTA using Plucker embedding and epipolar attention between all frames), as shown in Tab. 1. We report all results using DynamiCrafter as the base model. + +Our method demonstrates significant improvements in visual quality (FVD) and camera control metrics (TransErr, RotErr, CamMC), with particularly over $30\%$ gains on metric scale results. However, these improvements are not fully captured by the RealEstate10K dataset, which contains mostly static scenes. + +# 4.4. Ablation Study + +To validate our proposed method, we conducted a comprehensive ablation study on Metric Scene-scale Alignment (MSA) and Scene-constrained Noise Shaping (SNS). + +Effect of MSA only. As shown in Tab. 2, compared to + +relative scale training, metric scale training mitigates scale ambiguity in Fig. 4 and yields notable improvements, especially on metric scale results. It implies that models trained on metric-scale data can more accurately capture true-to-scale translations and better understand camera rotations within a realistic spatial framework. The metric scene scale enhances robustness and compatibility, ensuring that the framework adapts effectively to real-world input and applications. This approach allows for interaction within a unified scale, enabling intuitive user control over camera actions. + +Effect of MSA together with SNS. Adding scene-constrained noise shaping to a model with metric scale training yields substantial gains in video quality and camera controllability. This improvement is evident in both camera metrics and FVD. The synergy of metric-scale training and noise shaping ensures robust and precise control in diverse scenarios. This combined approach delivers noticeably better dynamics compared to using noise shaping alone. Large camera movements, rotations, and rapid transitions, which previously struggled to maintain consistency and realism, now work seamlessly. This improvement underscores the strength of integrating metric scale training with noise shaping for complex motion scenarios. + +Effect of SNS only. As shown in Tab. 2, scene-constrained noise shaping can be used as the only method for camera + +![](images/db2b83c9281b1e727346c46972602a5450d3b5654ab1a8f315a5573dcb23299d.jpg) + +![](images/37ad049805c823214cb2f5f6818060a3a6ee961ee9a8cba9e53d17ee9411c11e.jpg) +Figure 9. Visualization on various domains in real life scenarios. Despite training on RealEstate10K [100], our method can generalize naturally to out-of-domain images, including pets, landscape, anime, food and etc. + +![](images/d3240b437204c39a58eaf14acbba6a8352ca1e8352b4a2a0c62e1ce119f4380a.jpg) + +![](images/df0c619092b4e16a19b7d252cd77ff62b0bf6e229223ebf243e8d0b9ceba3c93.jpg) + +![](images/659d8e929e254bcf6e002a8a2ef143a01e9b8746163564e954120275df499b60.jpg) + +![](images/91a8838367a106816301c819d587bc3d71224e97e19ee7daae687308604a9dce.jpg) + +![](images/7c7aff1c489f9671b258e7a17980731e75b1f9995a53c77c3717c95e1388b446.jpg) + +![](images/b535461e6eee56eb8b1f16645fb330ae05e129061892bdacb94a6fc190f1e134.jpg) + +![](images/c518df976cefdead1ff41ef47b110aebdc38a48146258e870c4daf181e89477b.jpg) +Figure 10. Visualization on complex trajectory, large movement and video dynamism. Our framework achieves precise trajectory adherence for complex camera motion paths while preserving high-fidelity video generation with dynamics. + +control when applied to a base model not trained with any camera conditions. It provides nearly $50\%$ reduction on DynamiCrafter. However, this method performs poorly compared to the combined method with metric scale training. It also introduces challenges in parameter selection: Applying noise shaping only in the high noise stages limits camera control in lower noise stages, while extending noise shaping to mid-noise phase can suppress video dynamics, resulting in static video output. This limitation affects the fluidity and responsiveness of camera movements, making our combined approach preferable for applications requiring natural dynamics. + +# 4.5. Application + +As illustrated in Fig. 9 and Fig. 10, we demonstrate the versatility of our method through visualization results across various applications. Additionally, our results include camera-controlled loop video generation, generative frame interpolation, and smooth scene transitions, highlighting the robustness of our approach. These visualizations showcase two major breakthroughs: first, our method achieves a real-world application breakthrough by addressing challenges like training-inference scale inconsistency and low usability, ensuring improved robustness and compatibility with real-world images. Second, our framework exhibits superior performance in complex camera motions, handling + +large and rapid movements, rotations, and dynamics more effectively than existing methods. + +# 5. Limitation Analysis and Future Work + +Despite training in static RealEstate10K, as discussed in Fig. 9 and Fig. 10, our method demonstrate well generalization to videos with dynamic scenes, various styles and domains due to the designed algorithm to best preserve the knowledge of foundation model. However, the current model still suffers from real-world application mainly due to the limitation of data and there's no suitable dataset designed for camera-controlled video generation, which requires both dynamic scenes and camera movement. + +# 6. Conclusion + +In this paper, we address the scale inconsistencies and real-world usability challenges in existing trajectory-based camera-controlled image-to-video generation methods. We introduce a simple yet effective monocular 3D reconstruction into the preprocessing step of the generation pipeline, serving as a reliable intermediary reference for both training and inference. With reconstructed 3D scene, we enable absolute-scale training and provide an interactive interface during inference to easily design camera trajectories with preview feedback, along with proposed scene-constrained noise shaping to significantly enhance scene consistency and camera controllability. Our method overcomes critical real-world application challenges and achieves substantial improvements on the RealEstate10K dataset, establishing a new sota both in video quality and control precision. + +# 7. Acknowledgement + +This work is supported in part by National Science Foundation for Distinguished Young Scholars under Grant 62225605, Zhejiang Provincial Natural Science Foundation of China under Grant LD24F020016, Project 12326608 supported by NSFC, Ningbo Science and Technology Special Projects under Grant No. 2025Z028 and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 3 +[2] Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, et al. Vd3d: Taming large video diffusion transformers for 3d camera control. arXiv preprint arXiv:2407.12781, 2024. 3 +[3] 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. 3 +[4] 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, 3 +[5] 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, pages 22563-22575, 2023. 2 +[6] 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. 3 +[7] Hila Chefer, Shiran Zada, Roni Paiss, Ariel Ephrat, Omer Tov, Michael Rubinstein, Lior Wolf, Tali Dekel, Tomer Michaeli, and Inbar Mosseri. Still-moving: Customized video generation without customized video data. arXiv preprint arXiv:2407.08674, 2024.3 +[8] Changgu Chen, Junwei Shu, Lianggangxu Chen, Gaoqi He, Changbo Wang, and Yang Li. Motion-zero: Zero-shot moving object control framework for diffusion-based video generation. arXiv preprint arXiv:2401.10150, 2024. 3 +[9] 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 +[10] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models. In CVPR, pages 7310-7320, 2024. 2 +[11] Shoufa Chen, Chongjian Ge, Yuqi Zhang, Yida Zhang, Fengda Zhu, Hao Yang, Hongxiang Hao, Hui Wu, Zhichao Lai, Yifei Hu, et al. Goku: Flow based video generative foundation models. arXiv preprint arXiv:2502.04896, 2025.3 + +[12] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, 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. 3 +[13] 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. 3 +[14] Weichen Fan, Chenyang Si, Junhao Song, Zhenyu Yang, Yinan He, Long Zhuo, Ziqi Huang, Ziyue Dong, Jingwen He, Dongwei Pan, et al. Vchitect-2.0: Parallel transformer for scaling up video diffusion models. arXiv preprint arXiv:2501.08453, 2025. 3 +[15] Wanquan Feng, Jiawei Liu, Pengqi Tu, Tianhao Qi, Mingzhen Sun, Tianxiang Ma, Songtao Zhao, Siyu Zhou, and Qian He. 12vcontrol-camera: Precise video camera control with adjustable motion strength. 2024. 3 +[16] Peng Gao, Le Zhuo, Dongyang Liu, Ruoyi Du, Xu Luo, Longtian Qiu, Yuhang Zhang, Chen Lin, Rongjie Huang, Shijie Geng, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv preprint arXiv:2405.05945, 2024.3 +[17] 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 Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 2 +[18] 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 +[19] Litong Gong, Yiran Zhu, Weijie Li, Xiaoyang Kang, Biao Wang, Tiezheng Ge, and Bo Zheng. Atomovie: High fidelity image-to-video generation. arXiv preprint arXiv:2403.01800, 2024. 3 +[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. 2, 3 +[21] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 3 +[22] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 2, 3, 6, 7, 1 +[23] Xu He, Qiaochu Huang, Zhensong Zhang, Zhiwei Lin, Zhiyong Wu, Sicheng Yang, Minglei Li, Zhiyi Chen, Song- + +cen Xu, and Xiaofei Wu. Co-speech gesture video generation via motion-decoupled diffusion model. In CVPR, pages 2263-2273, 2024. 3 +[24] 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 +[25] Chen Hou, Guoqiang Wei, Yan Zeng, and Zhibo Chen. Training-free camera control for video generation. arXiv preprint arXiv:2406.10126, 2024.3 +[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. arXiv preprint arXiv:2106.09685, 2021.3 +[27] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In CVPR, pages 8153-8163, 2024. 3 +[28] 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. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[29] Teng Hu, Jiangning Zhang, Ran Yi, Yating Wang, Hongrui Huang, Jieyu Weng, Yabiao Wang, and Lizhuang Ma. Motionmaster: Training-free camera motion transfer for video generation. arXiv preprint arXiv:2404.15789, 2024. 2, 3 +[30] Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, Yaohui Wang, Xinyuan Chen, YingCong Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Vbench++: Comprehensive and versatile benchmark suite for video generative models. arXiv preprint arXiv:2411.13503, 2024. 7 +[31] Yash Jain, Anshul Nasery, Vibhav Vineet, and Harkirat Behl. Peekaboo: Interactive video generation via masked-diffusion. In CVPR, pages 8079-8088, 2024. 2, 3 +[32] Rui Jiang, Guang-Cong Zheng, Teng Li, Tian-Rui Yang, Jing-Dong Wang, and Xi Li. A survey of multimodal controllable diffusion models. Journal of Computer Science and Technology, 39(3):509-541, 2024. 3 +[33] Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In CVPR, pages 6689-6700, 2024. 3 +[34] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions. Advances in Neural Information Processing Systems, 37:48955-48970, 2024. 3 +[35] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 3 +[36] Zhengfei Kuang, Shengqu Cai, Hao He, Yinghao Xu, Hongsheng Li, Leonidas Guibas, and Gordon Wet + +zstein. Collaborative video diffusion: Consistent multivideo generation with camera control. arXiv preprint arXiv:2405.17414, 2024.3 +[37] Hengjia Li, Haonan Qiu, Shiwei Zhang, Xiang Wang, Yu-jie Wei, Zekun Li, Yingya Zhang, Boxi Wu, and Deng Cai. Personalvideo: High id-fidelity video customization without dynamic and semantic degradation. arXiv preprint arXiv:2411.17048, 2024. 3 +[38] Yaowei Li, Xintao Wang, Zhaoyang Zhang, Zhouxia Wang, Ziyang Yuan, Liangbin Xie, Yuexian Zou, and Ying Shan. Image conductor: Precision control for interactive video synthesis. arXiv preprint arXiv:2406.15339, 2024. 2, 3 +[39] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In CVPR, pages 24142-24153, 2024. 3 +[40] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024.3 +[41] Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, et al. Step-video-t2v technical report: The practice, challenges, and future of video foundation model. arXiv preprint arXiv:2502.10248, 2025. 3 +[42] 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 +[43] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024. 3 +[44] 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. arXiv preprint arXiv:2108.01073, 2021.5 +[45] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 3 +[46] Koichi Namekata, Sherwin Bahmani, Ziyi Wu, Yash Kant, Igor Gilitschenski, and David B Lindell. Sg-i2v: Self-guided trajectory control in image-to-video generation. arXiv preprint arXiv:2411.04989, 2024.3 +[47] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 3 +[48] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes Lutz Schonberger. Global Structure-from-Motion Revisited. In European Conference on Computer Vision (ECCV), 2024. 6, 2 + +[49] 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. 3 +[50] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-Chang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 3 +[51] 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. 3 +[52] 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 +[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3 +[54] 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, 6 +[55] Chenyang Si, Weichen Fan, Zhengyao Lv, Ziqi Huang, Yu Qiao, and Ziwei Liu. Repvideo: Rethinking cross-layer representation for video generation. arXiv preprint arXiv:2501.08994, 2025.3 +[56] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5 +[57] Kunpeng Song, Yizhe Zhu, Bingchen Liu, Qing Yan, Ahmed Elgammal, and Xiao Yang. Moma: Multimodal llm adapter for fast personalized image generation. arXiv preprint arXiv:2404.05674, 2024.3 +[58] Zineng Tang, Ziyi Yang, Chenguang Zhu, Michael Zeng, and Mohit Bansal. Any-to-any generation via composable diffusion. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3 +[59] 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, 2024.3 +[60] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, pages 16773-16783, 2023. 3 +[61] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7 +[62] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2 +[63] Qiuheng Wang, Yukai Shi, Jiarong Ou, Rui Chen, Ke Lin, Jiahao Wang, Boyuan Jiang, Haotian Yang, Mingwu Zheng, + +Xin Tao, et al. Koala-36m: A large-scale video dataset improving consistency between fine-grained conditions and video content. arXiv preprint arXiv:2410.08260, 2024. 3 +[64] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for referring human dance generation in real world. arXiv e-prints, pages arXiv-2307, 2023. 3 +[65] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation, 2024. 3 +[66] 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. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[67] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 3 +[68] Zhao Wang, Aoxue Li, Enze Xie, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. Customvideo: Customizing text-to-video generation with multiple subjects. arXiv preprint arXiv:2401.09962, 2024. 3 +[69] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6, 7, 1 +[70] Daniel Watson, Saurabh Saxena, Lala Li, Andrea Tagliasacchi, and David J Fleet. Controlling space and time with diffusion models. arXiv preprint arXiv:2407.07860, 2024. 3 +[71] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024.3 +[72] Tao Wu, Xuewei Li, Zhongang Qi, Di Hu, Xintao Wang, Ying Shan, and Xi Li. Spherediffusion: Spherical geometry-aware distortion resilient diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6126-6134, 2024. 3 +[73] Tao Wu, Yong Zhang, Xiaodong Cun, Zhonggang Qi, Junfu Pu, Huanzhang Dou, Guangcong Zheng, Ying Shan, and Xi Li. Videomaker: Zero-shot customized video generation with the inherent force of video diffusion models. arXiv preprint arXiv:2412.19645, 2024. 3 +[74] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. arXiv preprint arXiv:2408.13239, 2024. 3 +[75] Yinwei Wu, Xianpan Zhou, Bing Ma, Xuefeng Su, Kai Ma, and Xinchao Wang. Ifadapter: Instance feature con + +trol for grounded text-to-image generation. arXiv preprint arXiv:2409.08240, 2024.3 +[76] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccrafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 3, 6, 7, 1 +[77] Dejia Xu, Yifan Jiang, Chen Huang, Liangchen Song, Thorsten Gernoth, Liangliang Cao, Zhangyang Wang, and Hao Tang. Cavia: Camera-controllable multi-view video diffusion with view-integrated attention. arXiv preprint arXiv:2410.10774, 2024.3 +[78] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 2, 3, 6 +[79] 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 CVPR, pages 1481–1490, 2024. 3 +[80] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv:2406.09414, 2024. 2, 6, 1 +[81] Shiuyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3 +[82] 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, 3 +[83] 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 +[84] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 3 +[85] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 3 +[86] Sihyun Yu, Weili Nie, De-An Huang, Boyi Li, Jinwoo Shin, and Anima Anandkumar. Efficient video diffusion models via content-frame motion-latent decomposition. arXiv preprint arXiv:2403.14148, 2024. 3 +[87] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In CVPR, pages 8850-8860, 2024. 3 +[88] David Junhao Zhang, Dongxu Li, Hung Le, Mike Zheng Shou, Caiming Xiong, and Doyen Sahoo. Moonshot: Towards controllable video generation and editing with multi- + +modal conditions. arXiv preprint arXiv:2401.01827, 2024.3 +[89] 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. 2 +[90] 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. 3 +[91] 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. 3 +[92] Yiming Zhang, Zhening Xing, Yanhong Zeng, Youqing Fang, and Kai Chen. Pia: Your personalized image animatior via plug-and-play modules in text-to-image models. In CVPR, pages 7747-7756, 2024. 3 +[93] Ruisi Zhao, Zechuan Zhang, Zongxin Yang, and Yi Yang. 3d object manipulation in a single image using generative models. arXiv preprint arXiv:2501.12935, 2025. 3 +[94] Guangcong Zheng, Shengming Li, Hui Wang, Taiping Yao, Yang Chen, Shouhong Ding, and Xi Li. Entropy-driven sampling and training scheme for conditional diffusion generation. In European Conference on Computer Vision, pages 754-769. Springer, 2022. 2 +[95] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, pages 22490-22499, 2023. 3 +[96] Guangcong Zheng, Teng Li, Rui Jiang, Yehao Lu, Tao Wu, and Xi Li. Cami2v: Camera-controlled image-to-video diffusion model. arXiv preprint arXiv:2410.15957, 2024. 2, 3, 6, 7, 1 +[97] Guangcong Zheng, Teng Li, Xianpan Zhou, and Xi Li. Realcam-vid: High-resolution video dataset with dynamic scenes and metric-scale camera movements. arXiv preprint arXiv:2504.08212, 2025. 2 +[98] 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. arXiv preprint arXiv:2412.20404, 2024. 3 +[99] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 2 +[100] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018, 2, 6, 7, 8, 1 \ No newline at end of file diff --git a/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/images.zip b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dccbd8939188c5f82b630d90e7743650fb9ca458 --- /dev/null +++ b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:118c20139c34ba4ff05eb7c0c2b999eb63c8cd02ec5f1b6b3a95d5d3a4d0350e +size 739098 diff --git a/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/layout.json b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1b9413d6f10c50933899c8015d5d071e5faa6d86 --- /dev/null +++ b/ICCV/2025/RealCam-I2V_ Real-World Image-to-Video Generation with Interactive Complex Camera Control/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83a35383f66c5c90ae577ea479cfce2265675534ace4f3375f0d11f9431fe364 +size 498210 diff --git a/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/c08ceab1-72c4-4e59-ba66-a43eddf28bb5_content_list.json b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/c08ceab1-72c4-4e59-ba66-a43eddf28bb5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eb9365353b0278454b2c24bf9829eb98aa57d252 --- /dev/null +++ b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/c08ceab1-72c4-4e59-ba66-a43eddf28bb5_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16382f312ec3b9793c89a84ac478d33513c7bdf83b0e8d22b17f424ae477b8b4 +size 77444 diff --git a/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/c08ceab1-72c4-4e59-ba66-a43eddf28bb5_model.json b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/c08ceab1-72c4-4e59-ba66-a43eddf28bb5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a54903134985b1d82a2622ee62f48de9a9c3d865 --- /dev/null +++ b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/c08ceab1-72c4-4e59-ba66-a43eddf28bb5_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f7f4fcbe66f1338e81b0c2a60d76160b45e0618dd48d1a6753b367bc2594b38 +size 96770 diff --git a/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/c08ceab1-72c4-4e59-ba66-a43eddf28bb5_origin.pdf b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/c08ceab1-72c4-4e59-ba66-a43eddf28bb5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a1b3dcd016e7c4512be8d161ead15b5506ab068 --- /dev/null +++ b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/c08ceab1-72c4-4e59-ba66-a43eddf28bb5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db13f7deeb8891024a5960329489e16c9a3d9c8ebf91c877c15791eae9bfb8f6 +size 22321458 diff --git a/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/full.md b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b52c7c69c6e13097eaef1643679a261affafe1b9 --- /dev/null +++ b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/full.md @@ -0,0 +1,313 @@ +# RealGeneral: Unifying Visual Generation via Temporal In-Context Learning with Video Models + +Yijing Lin, Mengqi Huang, Shuhan Zhuang, Zhendong Mao† +University of Science and Technology of China + +{lyijing, huangmq, zhuangsh}@mail.ustc.edu.cn, zdmao@ustc.edu.cn + +![](images/32d3d0ac342b269b9b903c7d1aa0cbdae1418939f042a4efe69b19701e44a2cd.jpg) + +![](images/8dffdcc77d25bf29607cd93096ea3a01da9fba57cbc3584f5e0a6ee72ad3e450.jpg) + +![](images/70050fa4368ccd612d3f18fa693e31434521785057525b1cb3e08c29e02d7f82.jpg) + +![](images/42564ed8f5db8931ad188bd2e6b8b66f1ee053b19fbd226bfee077fc895c84c8.jpg) +Figure 1. Results from RealGeneral, demonstrate the ability to produce high-quality images from diverse input conditions. + +# Abstract + +Unifying diverse image generation tasks within a single framework remains a fundamental challenge in visual generation. While large language models (LLMs) achieve unification through task-agnostic data and generation, existing visual generation models fail to meet these principles. Current approaches either rely on per-task datasets and large-scale training or adapt pre-trained image models with task-specific modifications, limiting their generalizability. In this work, we explore video models as a foundation for unified image generation, leveraging their inherent ability to model temporal correlations. We introduce RealGen + +eral, a novel framework that reformulates image generation as a conditional frame prediction task, analogous to incontext learning in LLMs. To bridge the gap between video models and condition-image pairs, we propose (1) a Unified Conditional Embedding module for multi-modal alignment and (2) a Unified Stream DiT Block with decoupled adaptive LayerNorm and attention mask to mitigate cross-modal interference. RealGeneral demonstrates effectiveness in multiple important visual generation tasks, e.g., it achieves a $14.5\%$ improvement in subject similarity for customized generation and a $10\%$ enhancement in image quality for canny-to-image task. Project Page: realgeneral_web; GitHub Link: https://github.com/Lyne1/RealGeneral + +# 1. Introduction + +With the rapid development of generative modeling [5, 15, 16, 20, 21, 28, 54], image generation has gained considerable attention in various application domains. Although foundation models such as Stable Diffusion [10, 35] and FLUX [24] have achieved remarkable success in unimodal text-to-image generation, processing diverse inputs (e.g., text, depth maps, conditional images) and generating corresponding images within a unified framework remains a significant challenge. While large language models (LLMs) have unified text generation [1, 4, 51], an analogous unified model for the visual domain is still lacking. + +Upon closer examination of the core principles underlying LLMs' success as a unified text generation architecture, we identify two fundamental pillars, i.e., (1) a task-agnostic data principle during pre-training that facilitates the utilization of web-scale corpora for general linguistic knowledge acquisition, and (2) a task-agnostic generation principle that coherently unifies all kinds of textual tasks through next-token prediction. Analogously, a promising and effective architecture for unified visual generation should also exhibit (1) a task-agnostic visual data principle for general visual knowledge acquisition, and (2) a task-agnostic visual generation principle that could unify all kinds of multi-modal visual tasks (e.g., customization generation, canny-to-image generation, etc.), as shown in Fig. 2. + +Recent research on unified visual generation models mainly evolves along two streams, i.e., (1) the training-from-scratch stream [2, 27, 36, 45, 48], which curates pertask datasets and trains unified models through de novo optimization, and (2) the image-model-adaptation stream [6, 57], which reformulates generation tasks as partial image synthesis by leveraging pre-trained generative priors (e.g., FLUX). The first stream usually requires vast amounts of high-quality data and massive computational resources [8, 50]. For example, OmniGen[53] constructs a dataset of 0.1 billion images for all tasks and trains a unified model using 104 A800 GPUs. The second stream leverages the rich, general representations of pre-trained image generative models by fine-tuning them for specific tasks, thereby reducing computational costs. Some work [53, 56, 58] designs additional condition encoders or task-specific adapters to handle new condition types, while other work [46] revises only certain modules in the pre-trained models to adapt them to new tasks. Though much progress has been made, both of the streams fail to meet either the task-agnostic visual data principle or the task-agnostic visual generation principle. To be specific, the first stream heavily relies on the image dataset, inherently violating the task-agnostic visual data principle. Analogous to LLMs that learn by modeling relationships between sequential tokens, only video data inherently captures inter-frame relationships essential for general visual knowledge acquisition, while single im + +![](images/5f5511870949095b132b028dd624b215fb2ed0732e3a7f5264237bab34a58be4.jpg) +Figure 2. LLMs unify text generation through task-agnostic data and task-agnostic generation. Therefore, unified visual generation models should adhere to similar principles through visual task-agnostic data and task-agnostic visual generation principles. + +ages lack this attribution. Moreover, the second stream not only depends on image data but also adopts specialized architecture for certain tasks, violating both principles. + +To address the limitations of existing methods, we further explore whether large-scale pre-trained video generation models can overcome these challenges and provide a more robust framework for unified visual generation. In this study, we posit that such models, which are structurally endowed with both short-term micro-coherence and long-term macro-consistency, serve as a more effective and general foundation for this task. On the one hand, short-term micro-coherence captures fine-grained temporal dynamics, ensuring precise local details. On the other hand, long-term macro-consistency inherently preserves global structure and semantic consistency across entire sequences. Though great potential, how to effectively harness the pre-trained video models' inherent consistent generation capabilities for unified visual generation remains unexplored. + +Building on this insight, we argue that video models, with their inherent capability to model temporal correlations across frames, provide a more natural foundation for unified visual generation. Inspired by in-context learning in LLMs, we reformulate image generation as a conditional frame prediction task. In this formulation, input images are treated as preceding frames in a video sequence, analogous to previous tokens in LLMs, and generate the subsequent frame as the target output. This unified formulation recasts diverse visual generation tasks as instances of conditional frame prediction, eliminating the need for task-specific architectural modifications and thereby satisfying the task-agnostic visual generation principle. Furthermore, video models are trained on task-agnostic data (i.e., videos), adhering to the task-agnostic visual data principle. Therefore, the pre-trained video model only needs a small amount of corresponding task data for efficient fine-tuning, which can elicit the ability to handle various visual generation tasks, analogous to in-context learning in LLMs. + +In this study, we propose a new framework termed Re- + +alGeneral, adapting video diffusion models to image generation by proposing a minimalist framework centered on one-condition-image to one-target-image synthesis. In visual generative tasks, effectively fusing multi-modal inputs without conflating them with the generated image is challenging. To address these challenges, we introduce two core innovations. First, the Unified Conditional Embedding module fuses multi-modal inputs, aligning multi-modal conditions through both inter-modal conditional semantic alignment and intra-modal generative semantic distinction. Second, the Unified Stream DiT Block, a redesigned video model transformer block that incorporates (1) triple-branch adaLN to decouple feature modulation among text, condition frames, and target frames, eliminating the interference between the condition frame and target frame, and (2) attention mask that prevents text-to-condition interactions while preserving full condition-to-target visual attention to avoid the blending of textual and visual information. + +Our main contributions are summarized as follows: + +Conceptual contribution. We point out two crucial principles underlying the unification of visual generation. Furthermore, we propose a novel formulation that recasts multi-condition image generation as temporal in-context learning within video models, thereby establishing a unified framework for image generation. + +Technical contribution. (1) We propose Unified Conditional Embedding module for aligning multi-modal inputs. (2) We propose Unified Stream DiT Block with decoupled adaLN and attention mask to mitigate interference between different modalities. + +Experimental contribution. We validate our framework across multiple tasks. For instance, RealGeneral outperforms existing methods in the customized generation, achieving a $14.5\%$ improvement in subject similarity compared to state-of-the-art models. In the canny-to-image task, RealGeneral enhances image quality by $10\%$ , while in the depth-to-image task, it attains image quality comparable to existing models. Furthermore, training RealGeneral requires significantly fewer computational resources and less data than both unified models and specialized models. + +# 2. Related work + +# 2.1. Image Generation + +Image generation has rapidly advanced by combining NLP techniques with transformer architectures. Autoregressive models utilize transformers to predict sequences of discrete codebook codes [10, 38, 43]. Meanwhile, continuous diffusion models have emerged as a powerful framework for image generation [9, 15, 40]. The integration of large-scale transformer architectures has led to the development of advanced models, such as DiT [24, 34]. Many efforts have been made to extend the capabilities of generation mod + +els. For example, T2I-Adapter [31] introduces additional encoding layers to improve controllability, while RealCustom [22, 29] disentangles subject similarity from textual controllability by restricting subject influence to semantically relevant regions through a train-inference decoupled framework. However, these methods are designed for specific tasks, enhancing the capabilities of generative models through architectural modifications. + +# 2.2. Video Generation + +The field of text-to-video models has witnessed remarkable progress. Early efforts to pre-train and scale Transformers for generating videos from text, such as CogVideo [18] and Phenaki [47], demonstrate significant potential. At the same time, diffusion models have recently made groundbreaking strides in video generation [3, 12, 17, 26, 42]. With the introduction of DiT, text-to-video generation has reached a new level, as highlighted by the impressive performances of Sora [32] and CogVideoX [55]. + +# 2.3. Unified Image Generation + +Current research on unified image generation can be mainly categorized into two streams: training-from-scratch, and image-model-adaptation. Many works focus on the first stream [2, 8, 13, 30, 36, 45, 48]. For example, PixWizard [27] employs a flow-based DiT [34] model that gradually injects conditional and textual information during cross-attention. OmniGen [53] jointly models text and images, applying causal attention to each token in the sequence and bidirectional attention within each image sequence. The image-model-adaptation stream leverages pretrained image models to improve efficiency. For instance, ControlNet [58] uses trainable copies and zero convolution layers, enabling Stable Diffusion [10, 35] to adapt specific tasks. OmniControl [46] builds on FLUX.1 [24], proposing a multi-modal attention mechanism that facilitates direct interaction between conditions and images. + +Our concurrent work, UniReal [8], also reformulates images as discrete frames within a video generation framework. However, there are three critical distinctions between UniReal and our approach. First, the motivations differ: UniReal adopts the training-from-scratch paradigm, only modeling the relationship between discrete frames, whereas RealGeneral introduces two visual principles underlying the unification of visual generation, recasting multi-condition image generation as temporal in-context learning within video models. Second, the design choices diverge: UniReal relies on complex prompt engineering, while RealGeneral achieves controllability without such designs. Third, RealGeneral is significantly more efficient, fine-tuning a pre-trained video model using LoRA with only $0.1\%$ of UniReal's 360 million training samples, resulting in substantially reduced computational cost. + +![](images/f92362f1680c3f2c7b535a678ffc8d0d698dc86a5ca6e9dbdd3e6458a011cbe0.jpg) +Figure 3. The overview framework of RealGeneral, recasts the image conditional generation task as a next-frame prediction. We show two frames here for simplicity. First, two images are separately encoded by VAE. The Unified Condition Embedding Module integrates textual and global task priors into the condition frame while adding noise to the target frame. Then all tokens are concatenated into a sequence entering the Unified Stream DiT Block. The Separated Condition AdaLN Module modulates text, condition, and target tokens independently via three distinct branches. The Frame-Condition Decoupling Module prevents confusion between text and condition information. + +# 3. Methodology + +Inspired by LLMs that unify textual generation through unlabeled large-scale text pretraining and in-context learning abilities, we argue that the video foundation model pre-trained on massive continuous visual data can similarly unify the visual generation tasks. Analogous to how LLMs autoregressively predict tokens conditioned on previous context, video models can naturally extend this concept to temporal visual frames, predicting subsequent frames from preceding ones. Specifically, given previous frames as the condition frames $\{\mathbf{c}_1,\mathbf{c}_2,\mathbf{c}_3\ldots \ldots \}$ , video models predicts subsequent frames $\{\mathbf{x}_1,\mathbf{x}_2,\mathbf{x}_3\ldots \ldots \}$ , by exploiting is temporal modeling capabilities. + +In this work, we focus on a basic two-frame setting: the condition image is considered as the first frame, while the target image (accompanied by descriptive text) forms the second frame, as illustrated in Fig. 3. In this section, we first introduce our base model in Sec. 3.1 and then describe three key modules in detail. Sec. 3.2 details Unified Conditional Embedding (UCE) module for aligning multi-modal inputs. Sec. 3.3 explains Separated Condition AdaLN (SC-AdaLN) module enforcing semantic separation between condition and target frames. And Sec. 3.4 describes Frame-Condition Decoupling (FCD) module that employs an attention mask + +to prevent interactions between the condition image and text. Finally, we describe our task-specific LoRA in Sec. 3.5 + +# 3.1. Preliminaries + +Our proposed model is based on CogVideoX1.5 [55], a video generation model comprising three core components: (1) a T5 text encoder [37] for text processing, (2) a 3D causal VAE for spatio-temporal compression, and (3) a transformer based on the DiT [34] architecture. + +The 3D VAE encoder $\mathcal{E}(\cdot)$ compresses both spatial and temporal dimension through temporally causal convolutions, mapping an input video $\mathbf{x} \in \mathbb{R}^{(4f + 1) \times 8h \times 8w \times c}$ to a latent representation $\mathbf{z}_{\text{vision}} = \mathcal{E}(\mathbf{x}) \in \mathbb{R}^{(f + 1) \times h \times w \times d}$ , where $d$ denotes the hidden dimension. + +The transformer processes a concatenated sequence $\mathbf{z} = [\mathbf{z}_{text},\mathbf{z}_{\text{vision}}]$ , where text tokens $\mathbf{z}_{\text{text}}$ are text embeddings. It utilizes 3D full attention for joint spatio-temporal modeling, incorporating the 3D rotary position embedding (3D-RoPE) during the computation of the query and key. Additionally, the model implements modality-specific adaptive layernorm, where video and text tokens are normalized independently and then modulated by learned scale and shift factors conditioned on their respective modalities. + +The denoising objective is defined in Eq. (1): + +$$ +\mathcal {L} = \left\| \epsilon - \epsilon_ {\theta} \left(\mathbf {z} _ {t}, t, \mathbf {c}\right) \right\| ^ {2}, \tag {1} +$$ + +where $\epsilon$ denotes the noise added to the latent representations, $t$ represents the diffusion timestep, and $\mathbf{c}$ represents the conditioning signal. + +# 3.2. Unified Conditional Embedding Module + +In various visual generative tasks, inputs are often multimodal(e.g., text and image), while output is image modality. Therefore, the key to unified visual generation lies in how to effectively fuse multi-modal conditions and avoid confusing conditional and generated images with each other. In Real-General, we design the UCE module to align multi-modal conditions, containing two complementary pathways, i.e., the Subject-Specific Embedding layer for inter-modal conditional semantic alignment and the Condition Embedding layer for intra-modal generative semantic distinction. + +As depicted in Fig. 3, 3D VAE separately encodes the condition and target images, producing latent codes $\mathbf{c}_{cond} \in \mathbb{R}^{h \times w \times d}$ and $\mathbf{x}_{target} \in \mathbb{R}^{h \times w \times d}$ , respectively. Then in the UCE module, given a text embedding, we first extract subject-related embeddings $\mathbf{c}_{\text{instance}} \in \mathbb{R}^k$ using keyword matching (e.g., "pot" in Fig. 3). The Subject-Specific Embedding layer then maps these embeddings to the visual latent space, denoted as $\mathbf{c}'_{\text{instance}} \in \mathbb{R}^d$ . This process explicitly anchors textual subject descriptions to corresponding spatial regions in the condition image, thereby aligning textual and visual conditions. Additionally, the Condition Embedding layer applies a learnable task-specific bias $\mathbf{b}_c \in \mathbb{R}^d$ to reposition the condition embedding distribution, thereby better aligning the condition image with the target image. The enhanced condition token is obtained by combining the original embedding $\mathbf{c}_{\text{cond}}$ with these adjustments, as Eq. (2) formulated. + +$$ +\mathbf {c} _ {\text {c o n d}} = \mathbf {c} _ {\text {c o n d}} + \mathbf {c} _ {\text {i n s t a n c e}} ^ {\prime} + \mathbf {b} _ {c}. \tag {2} +$$ + +Before patchify operation, the vision sequence $\mathbf{z}_{\text{vision}} \in \mathbb{R}^{2 \times h \times w \times d}$ is formed by concatenating $\mathbf{c}_{\text{cond}}$ and $\mathbf{x}_{\text{target}}$ . The patchify operation in CogVideoX divides the input sequence along spatial and temporal dimensions using patch factors $p = 2$ . This process compresses the temporal dimension, leading to the intermingling of the latent tokens for $\mathbf{z}_{\text{cond}}$ and $\mathbf{z}_{\text{target}}$ . Such mixing obscures the distinct temporal features inherent to each frame. To mitigate this, we propose a simple replication strategy, i.e., we replicate the latent tokens for both $\mathbf{c}_{\text{cond}}$ and $\mathbf{x}_{\text{target}}$ along the temporal dimension. Finally, $\mathbf{c}_{\text{cond}}$ , $\mathbf{x}_{\text{target}}$ are concatenated into the input sequence for the Unified Stream DiT Block: + +$$ +\mathbf {z} = \left[ \mathbf {c} _ {\text {t e x t}}; \mathbf {c} _ {\text {c o n d}}; \mathbf {x} _ {\text {t a r g e t}} \right]. \tag {3} +$$ + +# 3.3. Separated Condition AdaLN Module + +We introduce the SC-AdaLN module, a novel mechanism to address the inherent conflict between multi-modal conditions in image generation tasks. Existing video generation frameworks employ two AdaLN modules to handle text and video inputs separately. This leads to interference between the condition frame and target frame due to their shared video AdaLN parameters. Our SC-AdaLN module resolves this by decoupling this process into three independent branches, explicitly disentangling three distinct modalities: + +1. Text AdaLN leverages the semantic information from text embeddings to provide global semantic guidance for the generation process, ensuring text-image alignment. This branch remains unchanged from the original. +2. Condition Frame AdaLN inherits the parameters from the original video AdaLN to preserve spatial-temporal modeling capabilities. It specializes in extracting style, structure, and local details from the condition image. +3. Target frame AdaLN also inherits the parameters from the original video AdaLN, which learns transition patterns between the condition and target frames. + +To modulate each branch according to the current timestep $t$ , we compute adaptive scale and shift parameters and residual gate via linear layers. Specifically, for each branch $\mathbf{k} \in \{\mathbf{c}_{text}, \mathbf{c}_{cond}, \mathbf{x}_{target}\}$ , we have: + +$$ +[ \gamma_ {\mathbf {k}}, \beta_ {\mathbf {k}}, g _ {\mathbf {k}} ] = f _ {\mathbf {k}} (t), \tag {4} +$$ + +where $f_{\mathbf{k}}$ denotes a MLP that outputs the modulation coefficients, scale $\gamma_{\mathbf{k}}\in \mathbb{R}^{d}$ , shift $\beta_{\mathbf{k}}\in \mathbb{R}^{d}$ , and a residual gate $g_{\mathbf{k}}\in \mathbb{R}^{d}$ conditioned on $t$ . + +Then for each branch, the normalized feature is calculated as: + +$$ +\operatorname {A d a L N} _ {\mathbf {k}} (\mathbf {k}; t) = \operatorname {L N} (\mathbf {k}) \odot (1 + \gamma_ {\mathbf {k}}) + \beta_ {\mathbf {k}}, \tag {5} +$$ + +where $\mathrm{LN}(\cdot)$ denotes standard layer normalization and $\odot$ represents element-wise multiplication. After applying AdaLN to each branch, the outputs are concatenated into a unified sequence. + +# 3.4. Frame-Condition Decoupling Module + +We propose the FCD module to mitigate multi-modal condition confusion, mainly containing the Frame-Condition Decoupling Attention (FCD Attention), which addresses the cross-modal condition interference. + +In general-purpose image generation tasks, textual input primarily conveys the semantic details of the target frame, while the condition frame provides structural or stylistic guidance. In CogVideoX, text and visual tokens are jointly processed using 3D full attention, allowing unrestricted cross-modal interactions. However, this direct interaction + +![](images/bea5f4f0df9fe5bdbd0b918f8e0c61d44a6d97517995879b7d6e77847fc1da96.jpg) + +![](images/eb1ccd9bc0f92e12fb60e14b9a082bf5e4d162845615d38accfd8835252d4ed2.jpg) + +![](images/a2bc51f5f805b88f961536a2feb1ca94998e17f9b492ddc20d16beebfd5bdbe2.jpg) +Figure 4. Various attention mask strategies. Red for interaction, and gray for blocked. (a) constrains the interaction between text and condition image. Based on (a), (b) further limits the cross-attention from the target image to the condition image, and (c) restricts the self-attention of the condition image. (d) is the original configuration without any mask. + +![](images/b1d742cdf9ba75733658a77f61313bc14186488e21518405ed38bf66e4096091.jpg) + +between text tokens and condition tokens leads to the blending of semantic and structural information, negatively affecting the quality of the generated frame. + +To address this issue, we introduce the FCD Attention mechanism, which explicitly restricts interactions between text and condition tokens. Specifically, we employ a custom attention mask, $M$ , to control these interactions precisely. The mask assigns an extremely negative value to the similarity scores between text tokens and condition tokens during attention computation, effectively decoupling the multi-modal condition interactions. This ensures that textual input influences only the target frame, while the condition frame preserves its structural properties without being overwhelmed by semantic details. + +To validate the effectiveness of our design, we test various masking strategies, as illustrated in Fig. 4. Among these, the configuration shown in $(a)$ achieves the best performance. The mask is defined as: + +$$ +M _ {i j} = \left\{ \begin{array}{c} - \infty , \forall (i, j) \in \left(\mathbf {c} _ {\text {t e x t}} \times \mathbf {c} _ {\text {c o n d}}\right) \cup \left(\mathbf {c} _ {\text {c o n d}} \times \mathbf {c} _ {\text {t e x t}}\right). \\ 0, \text {o t h e r w i s e}. \end{array} \right. \tag {6} +$$ + +Formally, the modified attention operation is defined as: + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}} + M\right) V. \tag {7} +$$ + +# 3.5. Task-Specific LoRA + +To efficiently adapt the billion-parameter CogVideoX, we employ LoRA [19] for fine-tuning. Specifically, for each task, we inject LoRA parameters exclusively into our proposed Separated Condition AdaLN Module and Frame-Condition Decoupling Module. + +# 4. Experiments + +# 4.1. Implementation Details + +Task. In our experimental evaluation, we selected three tasks to evaluate our framework: subject-driven text-to-image generation, canny-to-image, and depth-to-image. The subject-driven generation is a challenging task, explored in many works [7, 11, 23, 41], evaluating the model's ability to generate images related to a specific subject based on conditional inputs. The canny-to-image and depth-to-image tasks assess spatial understanding through edge and depth conditioning, respectively, thereby evaluating structural fidelity and three-dimensional awareness. These tasks demonstrate the model's versatility and precision in processing diverse input modalities and generating corresponding outputs, highlighting its generalizability. + +Dataset. For subject-driven text-to-image generation, we use FLUX to generate paired images of the same object with variations in scenes and poses, then filter Subject200k [46] to obtain 40K high-quality pairs, yielding a 260K two-frame video dataset. For spatially-aligned tasks (e.g., canny-to-image, depth-to-image), we extract a subset from 512-2M [46], forming 310K videos, fewer than those used in previous models. All videos are resized to $512 \times 512$ . Details refer to supplementary material. + +Benchmark. For customized generation, we follow previous work [7, 52], using 20 testing images in Dream-Bench [41] and 20 prompts. For spatially-aligned tasks, we use the COCO2017 validation set (5000 images) resized to $512 \times 512$ , with descriptions generated by ChatGPT-4o. + +Metric. For subject-driven generation, we use CLIP-I, DINO [33] to calculate subject similarity and CLIP-T to measure semantic consistency. To mitigate background interference, we compute the subject similarity after segmenting the reference and generated subjects using Grounded-SAM [39]. For spatially-aligned tasks, we use FID [14] and SSIM to measure image quality. + +Training and Inference. Our model, based on CogVideoX1.5 and fine-tuned using LoRA with a rank and alpha of 256, comprises 892 million parameters. Training is performed over 4000 iterations across 2 days on 2 A800 GPUs, with a batch size of 44. We utilize an AdamW optimizer with a learning rate of 0.0001 and betas set to 0.9 and 0.95. During inference, we apply classifier-free guidance for the text condition, with $\omega$ values set to 6 for customization and 2 for other tasks. The denoising step is set at 50. + +
MethodBase-ModelCLIP-I↑DINO↑CLIP-T↑
BLIP-Diffusion [25]SD 1.50.7680.5350.278
ELITE [52]SD 1.50.7620.5330.291
SSR-Encoder [59]SD 1.50.7670.5240.303
IP-Adapter [56]SDXL0.7900.5700.289
EMU2 [44]SDXL0.7950.5830.298
MS-Diffusion [49]SDXL0.7720.5400.312
OursCogVideoX0.8490.6680.314
+ +Table 1. Quantitative results with existing methods on customization task. We highlight the best and second-best values for each metric. + +# 4.2. Main Results + +Customized Image generation. In Fig. 5, we present qualitative comparisons with existing models, including both unified image generation models (OmniControl, OmniGen) and customization-specific models (MS-Diffusion, Emu2, SSR-Encoder). Compared to other methods, RealGeneral produces more precise details and better adherence to the prompt, underlining its effectiveness in both subject consistency and text controllability. + +The quantitative results are shown in Tab. 1. Compared to existing methods, our method demonstrates significant improvement in image similarity. The image similarity is notably higher, with our approach achieving a CLIP-I score of 0.849 and a DINO score of 0.668, surpassing the second-best results by 0.054 and 0.085, respectively. At the same time, our method attains a CLIP-T score of 0.314, outperforming other methods. + +Spatially-aligned tasks. In Fig. 6, we compare with existing methods. RealGeneral exhibits enhanced consistency with the provided canny edge or depth maps, generating more detailed and controllable images. Notably, because RealGeneral is based on a large-scale pre-trained video model, it essentially acquires more general visual knowledge and has a better understanding of visual relationships than those methods using image base model, as shown in Fig. 6 row 1 (i.e., dog case). + +The quantitative results are presented in Tab. 2. For the canny-to-image task, RealGeneral attains superior FID and SSIM scores compared to existing models. For the depth-to-image task, RealGeneral achieves a balanced performance in terms of FID and SSIM, attaining the second-best for both metrics. In addition, because the depth and canny maps contain spatial information, the CLIP-T of the generated images is very close (both 0.328). We don't show this metric in the table. + +# 4.3. Base Model Performance + +We compare the image generation performance between CogVideoX1.5 and image base models (SD1.5, SDXL, FLUX), as shown in Fig. 7. The results present that + +
ConditionMethodBase-ModelFID↓SSIM↑
CannyControlNet [58]SD1.518.790.28
ControlNet [58]FLUX98.680.25
OmniGen [53]-22.510.34
OmniControl [46]FLUX20.630.40
OursCogVideoX17.500.44
DepthControlNet [58]SD1.525.120.24
ControlNet [58]FLUX62.200.26
OmniGen [53]-21.620.25
OmniControl [46]FLUX27.260.39
OursCogVideoX23.400.35
+ +Table 2. Quantitative results with existing method on canny-to-image, depth-to-image generation tasks. + +
SettingCLIP-I↑DINO↑CLIP-T↑
w/o UCE0.8330.6660.320
w/o SC-AdaLN0.8300.6660.316
w/o FCD0.8270.6650.318
full version0.8490.6680.314
+ +Table 3. Effect of Unified Conditional Embedding(UCE), Separated Condition AdaLN(SC-AdaLN), Frame-Condition Decoupling(FCD) modules. Removing any of the above modules causes a significant reduction in subject consistency while attaining a subtle increase in textual controllability. + +CogVideoX1.5 achieves comparable image quality with SD1.5, but significantly weaker than SDXL and FLUX. Nevertheless, RealGeneral still attains better or comparable performance compared to existing methods based on pretrained image models, demonstrating the effectiveness of the pre-training video model and the viability of unifying visual generation through task-agnostic videos. + +# 4.4. Ablation Studies + +Effect of each module. To validate the effectiveness of our proposed modules, we conduct the ablation experiments on the Unified Conditional Embedding(UCE), Separated Condition AdaLN(SC-AdaLN), and Frame-Condition Decoupling(FCD) modules, with the results shown in Tab. 3. Removing UCE module results in a significant decrease in CLIP-I $(0.849\rightarrow 0.833)$ , indicating its critical role in aligning multi-modal semantics. Similarly, the removal of SC-AdaLN or FCD module leads to reductions in CLIP-I by 0.019 and 0.022, respectively, showcasing the importance of decoupling multi-modal conditions. Although these modifications cause a slight decrease in CLIP-T, the improvement in topic similarity remains noteworthy. + +Impact of various attention mask strategies. To assess the impact of different attention mask strategies on model performance, we conduct ablation experiments as shown in Tab. 4. Among the several attention mask strategies de + +![](images/da3266d95dc01c9aac62442e074cee4ef4f73c75ccf5387297ff9427ca1fcc2a.jpg) +Figure 5. Qualitative comparison with existing methods, consisting of unified image generation models and customization-specific models. The results demonstrate that RealGeneral exhibited superior performance in terms of subject consistency and text consistency. + +![](images/711037de3ff208934494260aac48dda07719a3bec7b7946a1ae1470a7dcfa7bd.jpg) +Figure 6. Qualitative comparison with existing methods on canny-to-image task. Our method is more consistent with the given depth or canny map than other methods. + +![](images/f31fb412f71649e66cb2dfeaf8b61c381186befcc151d550526ced818052433b.jpg) +Figure 7. Comparison of image generation results between CogVideoX1.5 and other image generation models. + +picted in Fig. 4, Mask A, which blocks the interaction between the textual input and the condition image, achieves the highest scores in subject similarity with only a minimal reduction in textual similarity. In contrast, both Mask B and the no-mask result in lower subject consistency, and Mask C significantly impairs the model's performance in both subject consistency and textual controllability by restricting the condition image's self-attention. The qualitative results are + +
SettingCLIP-I↑DINO↑CLIP-T↑
Mask B0.8270.6620.321
Mask C0.7790.6500.312
no mask0.8270.6650.318
Mask A0.8490.6680.314
+ +Table 4. Ablation study on various attention masks. + +shown in supplementary material. + +# 5. Conclusion & Future Work + +In this paper, we propose RealGeneral, a novel framework that bridges the gap between video generation models and unified visual synthesis tasks. We unlock the inherent potential of pre-trained video models for diverse image generation tasks by reformulating conditional image generation as temporal in-context learning through temporal modeling. This is achieved through three core modules: (1) UCE module that aligns multi-modal semantics, (2) SC-AdaLN that disentangles feature modulation across text, conditional frames, and target frames, (3) FCD module that prevents conditional semantic interference. Together, these components form a solution for unified image generation tasks. + +Future Work. Future research will explore several directions to further advance RealGeneral. First, we plan to validate the framework on video models with higher generation quality and larger pre-training to assess its generalizability. Second, we will focus on developing enhanced video foundation models that generate higher-quality image synthesis. Finally, we aim to develop a fully integrated, universal framework instead of task-specific LoRA fine-tuning. + +# Acknowledgment + +This research is supported by Artificial Intelligence National Science and Technology Major Project 2023ZD0121200, and National Natural Science Foundation of China under Grant 62222212 and 623B2094. + +# 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] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22861-22872, 2024. 2, 3 +[3] 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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563-22575, 2023. 3 +[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] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2 +[6] Haodong Chen, Lan Wang, Harry Yang, and Ser-Nam Lim. Omnicreator: Self-supervised unified generation with universal editing. arXiv preprint arXiv:2412.02114, 2024. 2 +[7] Nan Chen, Mengqi Huang, Zhuowei Chen, Yang Zheng, Lei Zhang, and Zhendong Mao. Customcontrast: A multilevel contrastive perspective for subject-driven text-to-image customization. arXiv preprint arXiv:2409.05606, 2024. 6 +[8] Xi Chen, Zhifei Zhang, He Zhang, Yuqian Zhou, Soo Ye Kim, Qing Liu, Yijun Li, Jianming Zhang, Nanxuan Zhao, Yilin Wang, et al. Unireal: Universal image generation and editing via learning real-world dynamics. arXiv preprint arXiv:2412.07774, 2024. 2, 3 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3 +[10] 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. 2, 3 +[11] 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. 6 +[12] 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. 3 +[13] Zhen Han, Zeyinzi Jiang, Yulin Pan, Jingfeng Zhang, Chaojie Mao, Chenwei Xie, Yu Liu, and Jingren Zhou. Ace: All-round creator and editor following instructions via diffusion transformer. arXiv preprint arXiv:2410.00086, 2024. 3 +[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, 3 +[16] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 2 +[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3 +[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3 +[19] 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. 6 +[20] Mengqi Huang, Zhendong Mao, Penghui Wang, Quan Wang, and Yongdong Zhang. Dse-gan: Dynamic semantic evolution generative adversarial network for text-to-image generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4345-4354, 2022. 2 +[21] Mengqi Huang, Zhendong Mao, Zhuowei Chen, and Yongdong Zhang. Towards accurate image coding: Improved autoregressive image generation with dynamic vector quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22596-22605, 2023. 2 +[22] Mengqi Huang, Zhendong Mao, Mingcong Liu, Qian He, and Yongdong Zhang. Realcustom: Narrowing real text word for real-time open-domain text-to-image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7476-7485, 2024. 3 +[23] 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. 6 + +[24] Black Forest Labs. Flux: Official inference repository for flux.1 models. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-12. 2, 3 +[25] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36:30146-30166, 2023. 7 +[26] Muyang Li, Ji Lin, Chenlin Meng, Stefano Ermon, Song Han, and Jun-Yan Zhu. Efficient spatially sparse inference for conditional gans and diffusion models. Advances in neural information processing systems, 35:28858-28873, 2022. 3 +[27] Weifeng Lin, Xinyu Wei, Renrui Zhang, Le Zhuo, Shitian Zhao, Siyuan Huang, Junlin Xie, Yu Qiao, Peng Gao, and Hongsheng Li. Pixwizard: Versatile image-to-image visual assistant with open-language instructions. arXiv preprint arXiv:2409.15278, 2024. 2, 3 +[28] Tao Liu, Kai Wang, Senmao Li, Joost van de Weijer, Fahad Shahbaz Khan, Shiqi Yang, Yaxing Wang, Jian Yang, and Ming-Ming Cheng. One-prompt-one-story: Free-lunch consistent text-to-image generation using a single prompt. arXiv preprint arXiv:2501.13554, 2025. 2 +[29] Zhendong Mao, Mengqi Huang, Fei Ding, Mingcong Liu, Qian He, and Yongdong Zhang. Realcustom++: Representing images as real-word for real-time customization. arXiv preprint arXiv:2408.09744, 2024. 3 +[30] David Mizrahi, Roman Bachmann, Oguzhan Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, and Amir Zamir. 4m: Massively multimodal masked modeling. Advances in Neural Information Processing Systems, 36:58363-58408, 2023. 3 +[31] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296-4304, 2024. 3 +[32] OpenAI. Sora. https://openai.com/index/sora/, 2024.3 +[33] Maxime Oquab, Timothee Darcet, Theo 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. 6 +[34] 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. 3, 4 +[35] 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 +[36] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 2, 3 + +[37] 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. 4 +[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. 3 +[39] 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, 2024.6 +[40] 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 +[41] 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. 6 +[42] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3 +[43] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3 +[44] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 7 +[45] Zeyi Sun, Ziyang Chu, Pan Zhang, Tong Wu, Xiaoyi Dong, Yuhang Zang, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. X-prompt: Towards universal in-context image generation in auto-regressive vision language foundation models. arXiv preprint arXiv:2412.01824, 2024. 2, 3 +[46] Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. *Omnincontrol: Minimal and universal control for diffusion transformer.* arXiv preprint arXiv:2411.15098, 2024. 2, 3, 6, 7 +[47] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 3 +[48] 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, 3 +[49] Xierui Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. MS-diffusion: Multi-subject zero-shot image personalization with layout guidance. In The Thirteenth International Conference on Learning Representations, 2025. 7 +[50] Zhaoqing Wang, Xiaobo Xia, Runnan Chen, Dongdong Yu, Changhu Wang, Mingming Gong, and Tongliang Liu. Lavin-dit: Large vision diffusion transformer. arXiv preprint arXiv:2411.11505, 2024. 2 +[51] 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. 2 +[52] 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. 6, 7 +[53] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 2, 3, 7 +[54] Xingqian Xu, Jiayi Guo, Zhangyang Wang, Gao Huang, Irfan Essa, and Humphrey Shi. Prompt-free diffusion: Taking" text" out of text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8682-8692, 2024. 2 +[55] 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. 3, 4 +[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. 2, 7 +[57] Qifan Yu, Wei Chow, Zhongqi Yue, Kaihang Pan, Yang Wu, Xiaoyang Wan, Juncheng Li, Siliang Tang, Hanwang Zhang, and Yueting Zhuang. Anyedit: Mastering unified high-quality image editing for any idea. arXiv preprint arXiv:2411.15738, 2024. 2 +[58] 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, 3, 7 +[59] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8069–8078, 2024. 7 \ No newline at end of file diff --git a/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/images.zip b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..249c9a3f4aca1d5d839ad4fcbf3c5e7e4732799b --- /dev/null +++ b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33b88020d6e6729462ec081cce01902b032c8e49fe9956675c13c7d9af19f396 +size 728821 diff --git a/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/layout.json b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dfdaf4de5559b20ec1e178431f4e03ce3b95ddf0 --- /dev/null +++ b/ICCV/2025/RealGeneral_ Unifying Visual Generation via Temporal In-Context Learning with Video Models/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f292932264153a133fc3e335ccb367eaa21132aa075c31498897def926573a68 +size 350279 diff --git a/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/642933ff-f262-4046-94fb-8c22ab2e3a58_content_list.json b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/642933ff-f262-4046-94fb-8c22ab2e3a58_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..54b7b50ec0501f2796fb052464db858e33a989db --- /dev/null +++ b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/642933ff-f262-4046-94fb-8c22ab2e3a58_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f313ec3279826b740fe496f93b1bcfc53e64e211a4d0d8a84e1aacdf96a2992d +size 84457 diff --git a/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/642933ff-f262-4046-94fb-8c22ab2e3a58_model.json b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/642933ff-f262-4046-94fb-8c22ab2e3a58_model.json new file mode 100644 index 0000000000000000000000000000000000000000..68229b55285e556d5b146d22da7f037c96ff483f --- /dev/null +++ b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/642933ff-f262-4046-94fb-8c22ab2e3a58_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bc431c5554f7cb9ddfd585c65d9583877fe849ac71eb5a7bc9adb666543a9db +size 111964 diff --git a/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/642933ff-f262-4046-94fb-8c22ab2e3a58_origin.pdf b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/642933ff-f262-4046-94fb-8c22ab2e3a58_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..655d8c9a2d83e907a4f631471f7c728d5c5daf83 --- /dev/null +++ b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/642933ff-f262-4046-94fb-8c22ab2e3a58_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8d6ddd55f9cc00da3a63b692e79b9739c3702fe550665a158735ba55244279c +size 39018221 diff --git a/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/full.md b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f19f1bce61d81662c05ccec6a40b105377d24ce1 --- /dev/null +++ b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/full.md @@ -0,0 +1,327 @@ +# Reangle-A-Video: 4D Video Generation as Video-to-Video Translation + +Hyeonho Jeong $^{1,2,*}$ + +Suhyeon Lee $^{1,*}$ + +Jong Chul Ye1 + +$^{1}$ KAIST $^{2}$ Adobe Research + +{hyeonho.jeong, suhyeon.lee,jong.ye}@kaist.ac.kr + +![](images/33c92f96295a7b852f82f1e92d39ad05071cf02853692a9f4177c2b4187d9b58.jpg) +Input video + +![](images/84cb8383b5cac9639867d2ff57227f1ae77eac68a55f5ffdc487b4bef9ec8484.jpg) +Output videos (Static View Transport) + +![](images/6bfccd5b5f23e5ef6641d90255093441ff6dddd0673aa0b0dbf1f847797071c9.jpg) + +![](images/e7627444fe851e300972a1f3667881b4dbcc887aae4486ad099d2b827d74c0ab.jpg) +Figure 1. From a single monocular video of any scene, Reangle-A-Video generates synchronized videos from diverse camera viewpoints or movements without relying on any multi-view generative prior—using only single fine-tuning of a video generator. The first row shows the input video, while the rows below present videos generated by Reangle-A-Video. (Left): Static view transport results. (Right): Dynamic camera control results. Full video examples are available on our project page: hyeonho99.github.io/reangle-a-video + +![](images/bc729a0b83391c0c3782e30843403b240b2f8049b3a0141b5cb617c720bc6aa4.jpg) +Input video + +![](images/31d20d6b18c80d761f8d547b19d01fe7b4933fae2ed13ab1a33fee33718ede8c.jpg) +Output videos (Dynamic Camera Control) + +![](images/9b65dea3af41888b7ec5946943c26e482b728d53253de599a60ed524d07baee8.jpg) + +![](images/7e134cc0dbeebbdc3bee76b690db9b92ac712be704c06863e22d1c94c5f8ac5c.jpg) + +# Abstract + +We introduce Reangle-A-Video, a unified framework for generating synchronized multi-view videos from a single input video. Unlike mainstream approaches that train multiview video diffusion models on large-scale 4D datasets, our method reframes the multi-view video generation task as video-to-videos translation, leveraging publicly available image and video diffusion priors. In essence, Reangle-A-Video operates in two stages. (1) Multi-View Motion Learning: An image-to-video diffusion transformer is synchronously fine-tuned in a self-supervised manner to distill view-invariant motion from a set of warped videos. (2) Multi-View Consistent Image-to-Images Translation: The first frame of the input video is warped and inpainted into + +various camera perspectives under an inference-time cross-view consistency guidance using DUSt3R, generating multiview consistent starting images. Extensive experiments on static view transport and dynamic camera control show that Reangle-A-Video surpasses existing methods, establishing a new solution for multi-view video generation. Our codes and data are available at our project page. + +# 1. Introduction + +Diffusion-based video generators [5, 7, 9, 39, 44, 57, 73, 84, 91] are rapidly advancing, enabling the generation of visually rich and dynamic videos from text or visual inputs. Recent progress in video generative models highlights the + +Figure 2. Qualitative results on static view transport (left) & dynamic camera control (right). Click with Acrobat Reader to play videos. + +growing need for user control over object appearance [12, 41, 48, 76], object motion [10, 26, 38, 40, 55, 85, 93], and camera pose [30, 59, 74, 78, 81, 83, 87, 90]. However, a video itself inherently captures only a partial perspective of the world, which exists as a dynamic 4D environment. + +To obtain dynamic 4D generative priors, previous works [50, 52, 79, 96] have extended 3D or video generators into 4D generative models using rendered synthetic assets [21, 22] that focus on animated objects or rigged characters. As a result, these methods are limited to object-level 4D synthesis and fail to generalize to real-world scenes. Recent studies have addressed this limitation by training foundational multi-view video diffusion models [4, 64, 67, 69, 71, 75, 77] on hybrid datasets combining indoor/outdoor static scenes [20, 53, 95], general single-view videos, and realistic scenes rendered by simulation engines [21, 22, 28, 62]. Although promising, these approaches often face challenges in real-world scenarios [69], are limited to specific domains (e.g., human-centric synthesis [64]), and most are not publicly available. More importantly, most existing methods generate multi-view videos from image or text inputs, rather than from user-input videos. + +To address this, here we present Reangle-A-Video, an alternative solution for synchronized multi-view video generation that does not require specialized multi-view generative priors. Given any input video, we frame the task as video-to-videos translation, capitalizing on publicly available image and video diffusion models. Our unified framework supports both static view transport—resimulating a video from target viewpoints—and dynamic camera control, where the video gradually transitions to the target viewpoints. Both approaches offer six degrees of freedom (see Fig. 1 and 2). + +In essence, our approach is based on the decomposition of a dynamic 4D scene into view-specific appearance (the starting image) and view-invariant motion (image-to-video generation). To capture view-invariant motion of the scene, we augment the training dataset with warped videos generated via repeated point-based warping of a single monocular video—these warped videos provide strong hints about the camera's perspectives. We then fine-tune + +a pre-trained image-to-video model [84] using a synchronized few-shot training strategy building on a masked diffusion loss [2, 17, 63, 90]. After training, dynamic camera control over video (Fig.1,2-right) is achieved by generating videos using the original first frame input with text that specifies the desired camera movement. On the other hand, static view transport of a video requires viewpoint-transported starting images. To achieve this, our method addresses several key technical challenges. First, we generate the starting images by inpainting warped images with image diffusion prior [49]. Second, inspired by test-time compute scaling [42, 70, 86], we enforce multi-view consistency in inference-time, using a stochastic control guidance with an off-the-shelf multi-view stereo reconstruction network [72]. We demonstrate the effectiveness of our approach on a variety of real-world scenes and metrics. + +# 2. Related Work + +Denoising-based Image and Video Generation. The introduction of Diffusion Transformers (DiT) [56] revolutionized image and video generation by replacing the traditional U-Net with a transformer-based backbone. Combined with larger-scale, high-quality, curated datasets, this shift enabled more efficient and scalable diffusion generators. In image generation, the PixArt family [13-15] extended DiT to text-to-image synthesis, showcasing its versatility. Stable Diffusion 3 [23] and Flux [49] further advanced the field with Multi-modal Diffusion Transformers (MM-DiT), enabling bidirectional text-image interactions. Following this trend, recent video diffusion models also integrate MM-DiT with spatio-temporal VAEs, achieving high-quality, long-form video generation. In our work, we adopt Flux [49] for image diffusion and CogVideoX [84] for video diffusion, both built on the MM-DiT architecture. + +Multi-View Video Generation. Previous efforts on multiview (4D) video generation [50, 52, 79, 96] have mainly focused on synthesizing 4D assets—animated objects and rigged characters [21, 22]. A common approach is to combine motion priors from video diffusion models with 3D priors from multi-view image models and train on curated, synthetic datasets of dynamic 3D objects [50, 79]. While + +![](images/cd78e0d2cf0a78d8c2aeb9850b31cad508d413e0e47e97e985bacc69adc25850.jpg) + +![](images/8a7ce55f89d55802154e59740ac295c35eaf15cac2c986f6548ee9d37d98a6cf.jpg) +Figure 3. Multi-view motion learning pipelines for (a) Static view transport and (b) Dynamic camera control. For both tasks, we distill view-robust motion of the underlying scene to a pre-trained MM-DiT video model [84], using all visible pixels within the sampled videos. This few-shot, self-supervised training optimizes only the LoRA layers [35, 61], enabling lightweight training. + +promising, these methods are limited to object-level 4D synthesis—rendered animations of single objects or animals—and produce multi-view videos with a static camera pose per video, failing to generalize to real-world videos and arbitrary viewpoints. More recently, to capture open-domain real-world scenes, several works [4, 45, 67, 69, 71, 77] have built foundational multi-view video generative priors by training on multi-view images/videos of real-world static scenes [53, 95], general single-view video datasets, and photorealistic scenes rendered by various engines [21, 28, 62]. These efforts aim to overcome the scarcity of dynamic, in-the-wild multi-view video data. + +In contrast, our work reinterprets multi-view video generation as a video-to-multi-view-videos translation task. We demonstrate that synchronized few-shot fine-tuning of a video diffusion prior with visible pixels across arbitrary views is sufficient for multi-view video generation in real-world settings, without the need for expensive 4D diffusion models trained on large datasets. + +Camera Controllable Video Generation. Our framework is also related to video generation methods that incorporate controllable camera trajectories [3, 16, 30, 34, 46, 51, 59, 69, 74, 78, 80, 81, 83, 87, 89, 90]. Most approaches fine-tune pre-trained text/image-to-video diffusion models on multi-view datasets of static scenes (e.g., RealEstate10K [95], ScanNet [20], DL3DV-10K [53]) to integrate additional camera control signals. They often rely on ControlNet-like hypernetworks [26, 29, 30, 74, 92], raw camera extrinsics [74], camera ray coordinates [30, 75], Plücker coordinate embeddings [3, 46, 81], or combinations of multiple conditions [59]. + +Since these methods derive multi-view priors primarily from static scenes, they typically support camera-controlled generation from text or image inputs and are not designed to produce multiple videos that are mutually consistent. In contrast, our work focuses on (i) enabling camera control over an input video and, more importantly, (ii) generating multiple videos that remain consistent with each other. This is more challenging, as our framework must replicate the inherent dynamics of the input video, such as object motion, while ensuring consistency across the generated videos. + +# 3. Reangle-A-Video + +In this section, we provide a comprehensive discussion of our framework. Given an input video describing an arbitrary 4D scene, we aim to generate synchronized multiview videos of the same scene without relying on multiview generative priors. Leveraging pre-trained latent image and video diffusion models (Sec. 3.1), our approach decomposes the dynamic 4D scene into view-specific appearance (starting image) and view-invariant motion (imageto-video), addressing each component separately. We first embed the scene's view-invariant motion into a pre-trained video diffusion model using our novel self-supervised training with data augmentation strategy. Initially, to capture diverse perspectives from a single monocular video, we repeatedly perform point-based warping to generate a set of warped videos (Sec. 3.2). These videos, together with the original video, form the training dataset for fine-tuning a pre-trained image-to-video diffusion model with a masked diffusion objective (Sec. 3.3). To achieve $(b)$ dynamic camera control, we sample videos using the fine-tuned model + +with the original first frame as input. In contrast, for (a) static view transport, we generate view-transported starting images by inpainting the warped first frames under an inference-time view consistency guidance using an off-the-shelf multi-view stereo reconstruction network (Sec. 3.4). + +# 3.1. Preliminary: Latent Diffusion Models + +Diffusion models [33, 65, 66] generate clean images or videos from Gaussian noise via an iterative denoising process. This process reverses a fixed, time-dependent forward diffusion process, which gradually corrupts the data by adding Gaussian noise. In latent diffusion models [8, 60], this process operates in the lower-dimensional latent space of a pre-trained VAE [43], comprised of an encoder $\mathcal{E}(\cdot)$ and a decoder $\mathcal{D}(\cdot)$ . Given a clean sample, $\pmb{x}_0\sim p_{\mathrm{data}}(\pmb {x})$ it is first compressed to a latent representation $z_0 = \mathcal{E}(x_0)$ . Gaussian noise $\epsilon \sim \mathcal{N}(0,I)$ is then added to produce intermediate noisy latents via the forward process $z_{t} = \alpha_{t}z_{0} + \sigma_{t}\epsilon$ , where $t$ denotes the diffusion timestep, and $\alpha_{t}$ , $\sigma_t$ are noise scheduler parameters. The training objective is to learn a denoiser network $\epsilon_{\theta}$ that minimizes: + +$$ +\mathbb {E} _ {\boldsymbol {\epsilon} \sim \mathcal {N} (0, I), \boldsymbol {z} _ {t} \sim p, t, c} \left[ \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {z} _ {t}, t, c) \| _ {2} ^ {2} \right], \tag {1} +$$ + +with condition $c$ provided by a text prompt, image, or both. Once trained, the diffusion model generates clean samples by iteratively denoising a pure Gaussian noise. + +# 3.2. Stage I: Point-based Warping for Training Data Augmentation + +Given an input video $\pmb{x}^{1:N}$ of $N$ frames, our goal is to construct a training dataset for the subsequent fine-tuning stage, by lifting pixels into time-aware 3D point clouds and reprojecting them onto the image plane with target perspectives. + +For each frame $\pmb{x}^i, i \in \{1, \dots, N\}$ , we first estimate its depth map $\mathbf{D}^i$ using a monocular depth estimator [82]. The corresponding point cloud $\mathcal{P}^i$ is then generated from the RGBD image $[\pmb{x}^i, \mathbf{D}^i]$ as follows: + +$$ +\mathcal {P} ^ {i} = \phi_ {2 \rightarrow 3} ([ \boldsymbol {x} ^ {i}, \mathbf {D} ^ {i} ], \mathbf {K}, \mathbf {P} _ {\mathrm {s r c}} ^ {i}), \tag {2} +$$ + +where $\mathbf{K}$ is the camera intrinsic matrix, $\mathbf{P}_{\mathrm{src}}^i$ is the extrinsic matrix for frame $x^i$ , and $\phi_{2\rightarrow 3}$ is the function that lifts the RGBD image into a point cloud. Note that $\mathbf{K}$ and $\mathbf{P}_{\mathrm{src}}^i$ are set by convention as in [18], since they are intractable in open-domain videos. + +Next, we define $M$ target extrinsic matrix trajectories $\Phi_{1:M}$ , where each trajectory is given by + +$$ +\Phi_ {j} = \left\{\mathbf {P} _ {j} ^ {1}, \dots , \mathbf {P} _ {j} ^ {N} \right\}, j \in \{1, \dots , M \}. \tag {3} +$$ + +Each extrinsic matrix $\mathbf{P}_j^i$ comprises a rotation matrix $\mathbf{R} \in \mathbb{R}^{3 \times 3}$ and a translation vector $\mathbf{t} \in \mathbb{R}^{3 \times 1}$ , which together transform the point cloud $\mathcal{P}^i$ into the target camera coordinate system. In the static view transport setting, each target trajectory $\Phi_j$ is constant across all frames, i.e., $\mathbf{P}_j^1 = \mathbf{P}_j^2 =$ + +$\ldots = \mathbf{P}_j^N$ ; for dynamic camera control, each frame's camera pose $\mathbf{P}_j^i$ is determined by incrementally moving and rotating from the previous pose $\mathbf{P}_j^{i - 1}$ , with the first target pose set to the pose of the first input frame $(\mathbf{P}_j^1 = \mathbf{P}_{\mathrm{src}}^1)$ . + +Finally, we reproject each point cloud $\mathcal{P}^i$ to the image plane under the target perspective using $\mathbf{K}$ and $\mathbf{P}_j^i$ via the function $\phi_{3\rightarrow 2}$ : + +$$ +\left(\hat {\boldsymbol {x}} _ {j} ^ {i}, \boldsymbol {m} _ {j} ^ {i}\right) = \phi_ {3 \rightarrow 2} \left(\mathcal {P} ^ {i}, \mathbf {K}, \mathbf {P} _ {j} ^ {i}\right), \tag {4} +$$ + +where $\hat{\pmb{x}}_j^i$ is the rendered warped image and $m_j^i$ is the corresponding visibility mask (1 for visible surfaces, 0 for invisible regions). Repeating this process constructs our training dataset $\Omega$ , which consists of $M$ pairs of warped videos and corresponding visibility masks: $\Omega = \{(\hat{\pmb{x}}_j^{1:N},\pmb {m}_j^{1:N})\mid j = 1,\dots,M\}$ . We also add the input video $\pmb{x}^{1:N}$ with uniform visibility masks (all pixels set to 1), resulting in a total of $M + 1$ video-mask pairs. + +# 3.3. Stage II: Multi-View Motion Learning + +In this stage, we fine-tune a pre-trained image-to-video diffusion transformer to learn view-invariant motion from the dataset $\Omega$ that captures diverse scene perspectives and corresponding pixels. Here, "motion" refers to the nuanced dynamics of each object—including its type, direction, and speed—as well as any inherent camera movement when the original video's camera pose is not fixed. + +To achieve lightweight fine-tuning while preserving the original model's prior, we employ Low-Rank Adaptation (LoRA) [35, 61]. LoRA augments an attention layer by adding a residual path comprised of two low-rank matrices, $\theta_B \in \mathbb{R}^{d \times r}$ , $\theta_A \in \mathbb{R}^{r \times k}$ , to the original weight $\theta_0 \in \mathbb{R}^{d \times k}$ , $r \ll \min(d, k)$ . The modified forward pass is defined as: + +$$ +\theta = \theta_ {0} + \alpha \Delta \theta = \theta_ {0} + \alpha \theta_ {B} \theta_ {A}, \tag {5} +$$ + +where $\alpha$ controls the strength of the LoRA adjustment. We implement this approach on a MM-DiT-based image-to-video diffusion model [84]—which incorporates 3D full-attention in its MM-DiT blocks—by attaching LoRA layers to these 3D full-attention layers. + +We then randomly sample a video $\pmb{x}^{1:N}$ and its corresponding visibility mask $m^{1:N}$ from the dataset $\Omega$ prepared in Sec. 3.2. Note that if $\pmb{x}^{1:N}$ is the original input video, its mask $m^{1:N}$ is uniformly filled with 1. The video is compressed into latent space $z_0^{1:N} = \mathcal{E}(x^{1:N})$ , and the mask is downsampled to $m_{\mathrm{down}}^{1:N}$ using spatio-temporal downsampling (see Sup. Sec.A.4). The video latents are then noised, patched, and unfolded into a long sequence to form video tokens. These, along with the text tokens, are fed into the video denoiser $\epsilon_{\theta}$ . Since most sampled videos are warped and contain significant black regions, naive LoRA fine-tuning with the standard diffusion loss + +![](images/01b0e9cd8849632e863dd939e31acddd18c9180b19bbb5ec2c4a3833062cde1d.jpg) +Figure 4. Multi-view consistent image inpainting using stochastic control guidance. In experiments, we set $S = 25$ . + +(Eq. (1)) inevitably degrades the original model prior and makes the model generate warped videos (see Supp. Sec. B.2). Thus, leveraging the compositionality of diffusion objectives [2, 17, 25, 63, 90, 94], we fine-tune the LoRA layers $\Delta \theta$ using a masked diffusion loss that excludes invisible regions in its loss computation: + +$$ +\mathbb {E} _ {\boldsymbol {\epsilon}, \boldsymbol {z} _ {t} ^ {1: N}, \boldsymbol {m} _ {\text {d o w n}} ^ {1: N}, t, c} \left[ \left\| \boldsymbol {\epsilon} \odot \boldsymbol {m} _ {\text {d o w n}} ^ {1: N} - \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {z} _ {t} ^ {1: N}, t, c\right) \odot \boldsymbol {m} _ {\text {d o w n}} ^ {1: N} \right\| _ {2} ^ {2} \right], \tag {6} +$$ + +where condition $c$ includes both the input text and image, with the latter being the first frame of the sampled video. + +For text input, $(b)$ dynamic camera control necessitates explicit specification of the camera movement type. As shown in Fig. 3-(b), during the dynamic camera control training, the image-to-video model always starts with the same first frame—the original video's first frame—which is uniformly used across warped videos (see Sec. 3.2). As a result, the model cannot infer the desired camera movement from the uniform starting frame alone. To resolve this, we explicitly include the camera movement type in the text input during both training and inference. Our LoRA layers, attached to the 3D full-attention modules, enable interactions between text and video tokens, learning to distinguish between different camera movements and map the corresponding text tokens to the appropriate video tokens. + +The proposed training pipelines are outlined in Fig. 3. This synchronized few-shot training strategy prevents the video model from overfitting to a specific view. Instead, it learns general scene motion, enabling video generation even from viewpoints that were unseen during the training (see Sec. 4.3), while ensuring that all output videos are synchronized (e.g., with consistent object motion speeds). + +Once training is complete, for the $(b)$ dynamic camera control setup, we sample videos directly using the original video's first frame as the input image and text specifying the desired camera movement. However, to achieve $(a)$ static view transport, we require a starting input image that captures the scene from the desired viewpoint—a process addressed in the next stage. + +![](images/3eeee986fe9d618614859326954a1cc3a47089d8a6fc2bd9a204e66565ec0d3b.jpg) +Figure 5. Qualitative inpainting comparisons. We compare naive inpainting to inpainting with stochastic control guidance. + +# 3.4. Stage III: Multi-View Image Inpainting + +We achieve multi-view consistent starting image generation using a warp-and-inpaint approach. In Stage I, given an original image $\pmb{x}^{1}$ (the first frame of the input video), we render a set of warped images $\hat{\pmb{x}}_{1:M}^{1}$ from $M$ target viewpoints, along with corresponding binary masks $m_{1:M}^{1}$ that indicate invisible surfaces. These missing regions are then filled via inpainting using image diffusion prior. While state-of-the-art image generators [15, 23, 49] yield plausible inpainting results on single warped images, independently inpainting each warped view fails to ensure cross-view consistency (see Fig. 5, 2nd column). Motivated by inference-time scaling practices [36, 42, 70, 86], we address the emerging inconsistencies by introducing a stochastic control guidance that enforces multi-view consistency during the denoising process using the DUSt3R multi-view stereo reconstruction model [72]. Specifically, we sequentially inpaint warped images by leveraging previously inpainted views for consistency, guided by stochastic control guidance. At each denoising step, we generate $S$ clean estimates from $S$ stochastic paths and compute a multi-view consistency score for each estimate with respect to the previously inpainted images $\{\tilde{\pmb{x}}_1^1,\dots,\tilde{\pmb{x}}_I^1\}$ . Following [1], the score is calculated by projecting the images into a shared view using DUSt3R [72] and then computing the DINO feature similarity [11] on the projected features. We select the path with the highest score, add new noise to generate $S$ new paths, and iterate until the inpainting is complete. We present detailed algorithms in Supp. Alg. 1 and 2. + +![](images/962ba0fbecbe6a710d975ed0e63f9b8dfc0da73c81a515d98d4f07416d94dac7.jpg) +Figure 6. Qualitative comparisons. Top half shows (a) Static view transport and bottom half presents (b) Dynamic camera control results. The first row in each half displays the input videos, and for each input video, two generated videos corresponding to target cameras (1 and 2) are shown for each method. Across baseline, same camera parameters were used for each 1,2. Visit our page for full-video results. + +Table 1. Quantitative comparison results. Reangle-A-Video is evaluated in two modes (a) Static View Transport and (b) Dynamic Camera Control—against the baselines. VBench metrics [37] are presented on the left, other metrics [1, 31, 68] appear on the right. + +
(a) Static View TransportSubject ConsistencyBackground ConsistencyTemporal FlickeringMotion SmoothnessDynamic DegreeAesthetic QualityImaging QualityMEt3R↓FID↓FVD↓
Generative Camera Dolly [69]0.88490.91070.87310.91190.76080.40520.55240.1237155.155264.7
Vanilla CogVideoX [84]0.94480.93980.97380.98590.72860.49970.64750.053979.6213664.2
Reangle-A-Video (Ours)0.95160.93270.97640.99070.76570.51600.64140.041253.4482690.9
(b) Dynamic Camera Control
NVS-Solver [87]0.90370.93250.90490.95210.88090.50230.64110.109095.8153516.5
Trajectory Attention [78]0.89840.92880.93420.96580.88890.48540.59900.0965109.203624.9
Reangle-A-Video (Ours)0.91400.93640.93860.97940.88840.52380.62710.064874.1943019.7
+ +# 4. Experiments + +# 4.1. Implementation Details + +We experiment with 28 publicly sourced videos [27, 58], covering diverse scenes with varying objects, scene motions, and physical environments. Each video consists of 49 frames at a resolution of $480 \times 720$ . On average, we generate videos from 3.5 different camera viewpoints/movements per input, resulting in a total of 98 generated videos. + +For Stage I, Depth Anything V2 model [82] is used to estimate the depth maps of the input videos. For Stage II and III, we use the CogVideoX-5b image-to-video diffusion model [84], and FLUX text-to-image model [49] with a pre-trained inpainting ControlNet [19] attached to it. + +Upon video model fine-tuning, we set the LoRA rank to 128 and optimize the LoRA layers over 400 steps, where the fine-tuned parameters account for only about $2\%$ of the original Video DiT parameters. We use AdamW [54] optimizer with a learning rate of $1e - 4$ and a weight decay of $1e - 3$ . The fine-tuning takes about an hour. For the static view transport mode, we set $M = 12$ warped videos; for dynamic camera control, we set $M = 6$ (see Supp. Sec. A.2). For video sampling, we apply 40 sampling steps with a CFG [32] scale of 6.0. During multi-view image inpainting, we first resize the input to $1024\times 1024$ and then restore it to its original resolution. We use MEt3R [1] as the reward function for stochastic control with $S = 25$ . To introduce stochasticity for the path control, we employ an SDE-based sampler with 50 steps, ensuring the same marginal distribution as the original ODE of the flow trajectory (see Supp. Sec. A.6). All experiments, including inference and fine-tuning, are performed using 40GB A100 GPUs. + +# 4.2. Comparisons + +# 4.2.1. Baselines + +Multi-view/camera synchronized video generation from an input video remains largely underexplored. For (a) Static view transport, the closest works include Generative Camera Dolly [69] and the closed-source multi-view video foundation model CAT4D [77]. Although GS-DiT [6] + +aligns with our framework, its code was unavailable during our research. In addition to GCD, we evaluate a baseline using naive CogVideoX-I2V inference (Vanilla CogVideoX)[84], which employs the same input frame as our approach. For (b) Dynamic camera control over videos, we compare against two state-of-the-art methods, NVS-Solver [87] and Trajectory Attention [78]. While Recapture [90] is closely related, its code is not available, and it employs the proprietary multi-view image foundation model CAT3D [24]. + +# 4.2.2. Qualitative Results + +Comprehensive qualitative results are shown in Fig. 6. For (a) Static View Transport, Generative Camera Dolly [69] struggles with real-world videos (trained on synthetic data [28]), while Vanilla CogVideoX [84] fails to capture the input video's motion. In contrast, Reangle-A-Video accurately reproduces the input motion from the target viewpoint. For (b) Dynamic Camera Control, NVS-Solver [87] and Trajectory Attention [78] either confuse foreground objects (e.g., a girl's and a cat's head in the first video), fail to capture precise motion (e.g., neck movement in the second video), or miss background elements (e.g., a bench in the third video). Our method faithfully regenerates the input video's motion, preserves object appearance, and accurately follows the target camera movement. + +# 4.2.3. Quantitative Results + +Automatic metrics. For automated evaluation, we first use VBench [37], which assesses generated videos across various disentangled dimensions. As shown in Tab. 1-left, our method outperforms baselines in most metrics for both static view transport and dynamic camera control modes. Notably, compared to vanilla CogVideoX I2V—which uses the same input image—our approach maintain robust performance even when fine-tuned with warped videos. Additionally, Tab. 1-right presents evaluation results using FID [31], FVD [68], and the recently proposed multi-view consistency metric MEt3R [1]. + +Human evaluation. We further assess Reangle-A-Video against baselines via a user study involving 36 participants who compare our results with randomly selected baselines. For the static view transport evaluation, participants rate: (i) accuracy of the transported viewpoint, (ii) preservation of the input video's motion. For the dynamic camera control + +![](images/ee11eefa5f0e62a58388c00432b9840b2dc31e73424381d957110dcd61b29a0b.jpg) +Figure 7. User study results. Top: Static view transport results. Bottom: Dynamic camera control results. + +Table 2. Quantitative evaluation of multi-view consistency in image inpainting with and without stochastic control guidance. + +
Warped image inpaintingMEt3R ↓SED ↓TSED ↑
w/o stochastic control guidance0.14311.19660.5241
w/ stochastic control guidance0.11841.18440.5588
+ +evaluation, participants assess: (i) the accuracy of the target camera movement, (ii) the preservation of the input video's motion in the output. As reported in Fig. 7, our method outperforms the baselines in all aspects. + +# 4.3. Ablation Studies + +First, we ablate the stochastic control guidance by comparing it with naive inpainting (where each image is inpainted independently). Tab. 2 evaluates multi-view accuracy using the MEt3R [1], SED, and TSED [88] metrics (with $T_{e} = 1.25$ and $T_{m} = 10$ for SED). Fig. 5 shows that our method fills target (invisible) regions consistently across multiple views. Next, we ablate the necessity of our data augmentation strategy (i.e., the use of warped videos) in learning view-invariant motion. Fig. 8 compares fine-tuning with only the original input video versus using both the original and warped videos. The results indicate that relying solely on the original video fails to accurately capture motion—for example, the rhino moving in front of the tree. For quantitative assessment on input video's motion preservation, we employ user study (see Supp. Sec. B.1.) In Fig. 9-top, we demonstrate unseen view video generation. We exclude specific warped view videos (vertical up/down orbits) from the training dataset and fine-tune the video model without them. Then, using an inpainted first frame as input, we generate a video from that omitted view. In Fig. 9-bottom, we showcase novel view video generation guided by an edited first frame. Starting with an inpainted image representing the scene from a target viewpoint, we apply FlowEdit [47] to modify the image, then generate the novel-view video with our fine-tuned model. In both cases, the generated videos faithfully follow the input video's motion, demonstrating that our few-shot training strategy is robust in both view and appearance. + +![](images/b598cce5a3c9043acddf780994a0d38bd66ca48689002230b647b811afa0794a.jpg) +Figure 8. Novel view video generation with and without using warped video for training (target viewpoint: dolly zoom in). + +![](images/6146b050f8932d0b9926569bdca9a59ce6b132ea0bfd5f55a697159d6e55d4d1.jpg) +Figure 9. Top: Unseen view video generation. Bottom: Novel view video generation using an appearance-edited first frame. + +# 5. Conclusion + +We have presented an approximate yet effective solution for generating synchronized multi-view videos from a single monocular real-world video—without relying on any multiview generative priors. Our approach enables an image-to-video diffusion model to learn view-invariant scene motion through self-supervised fine-tuning with video augmentation. Then, we sample videos from the fine-tuned model using a set of multi-view consistent first-frame images, where the images are generated via a warp-and-inpaint process that enforces multi-view consistency during inference-time. Limitations and future work. Input image quality is crucial in image-to-video generation, as the output cannot exceed the first frame's quality. While our approach ensures multi-view consistency in the warp-and-inpaint scheme, the warping stage is inherently prone to artifacts caused by camera and depth errors (Supp. Sec. C). We believe that improved depth models could enhance both the warped and inpainted images. Another limitation of our work is the need for scene-specific model tuning to achieve 4D video synthesis. A promising future direction is to extend existing video datasets with warped videos to train 4D foundation models. + +# Acknowledgments + +We would like to thank Geon Yeong Park and Mingi Kwon for their valuable discussions at the early stage of the project. This work was supported by the National Research Foundation of Korea under Grant RS-2024-00336454, and by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No. RS-2024-00457882, AI Research Hub Project; No. RS-2025-02304967, AI Star Fellowship (KAIST)). + +# References + +[1] Mohammad Asim, Christopher Wewer, Thomas Wimmer, Bernt Schiele, and Jan Eric Lenssen. Met3r: Measuring multi-view consistency in generated images. arXiv preprint arXiv:2501.06336, 2025. 5, 7, 8 +[2] 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, 5 +[3] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 3 +[4] Jianhong Bai, Menghan Xia, Xintao Wang, Ziyang Yuan, Xiao Fu, Zuozhu Liu, Haoji Hu, Pengfei Wan, and Di Zhang. Syncammaster: Synchronizing multi-camera video generation from diverse viewpoints. arXiv preprint arXiv:2412.07760, 2024. 2, 3 +[5] 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 space-time diffusion model for video generation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 1 +[6] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yijin Li, FuYun Wang, and Hongsheng Li. Gs-dit: Advancing video generation with pseudo 4d gaussian fields through efficient dense 3d point tracking. arXiv preprint arXiv:2501.02690, 2025.7 +[7] 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. 1 +[8] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 4 +[9] 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 + +[10] Ryan Burgert, Yuancheng Xu, Wenqi Xian, Oliver Pilarski, Pascal Clausen, Mingming He, Li Ma, Yitong Deng, Lingxiao Li, Mohsen Mousavi, et al. Go-with-the-flow: Motion-controllable video diffusion models using real-time warped noise. arXiv preprint arXiv:2501.08331, 2025. 2 +[11] 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. 5 +[12] Hila Chefer, Shiran Zada, Roni Paiss, Ariel Ephrat, Omer Tov, Michael Rubinstein, Lior Wolf, Tali Dekel, Tomer Michaeli, and Inbar Mosseri. Still-moving: Customized video generation without customized video data. ACM Transactions on Graphics (TOG), 43(6):1-11, 2024. 2 +[13] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2 +[14] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91. Springer, 2024. +[15] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-δ: Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 2, 5 +[16] Soon Yau Cheong, Duygu Ceylan, Armin Mustafa, Andrew Gilbert, and Chun-Hao Paul Huang. Boosting camera motion control for video diffusion transformers. arXiv preprint arXiv:2410.10802, 2024. 3 +[17] Gene Chou, Kai Zhang, Sai Bi, Hao Tan, Zexiang Xu, Fujun Luan, Bharath Hariharan, and Noah Snavely. Generating 3d-consistent videos from unposed internet photos. arXiv preprint arXiv:2411.13549, 2024. 2, 5 +[18] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 4 +[19] Alimama Smart Creative and AI Application Team. Flux inpainting controlnet, 2025. URL https://huggingface.co/alamama-creative/FLUX.1-dev-Controlnet-Inpainting-Beta. 7 +[20] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 3 +[21] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3 + +[22] 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. 2 +[23] 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, 5 +[24] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 7 +[25] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Mdtv2: Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2023. 5 +[26] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, et al. Motion prompting: Controlling video generation with motion trajectories. arXiv preprint arXiv:2412.02700, 2024. 2, 3 +[27] Canva Germany GmbH. pexels, 2024. URL https://www.pexels.com/videos/. 7 +[28] 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, 3, 7 +[29] Zekai Gu, Rui Yan, Jiahao Lu, Peng Li, Zhiyang Dou, Chenyang Si, Zhen Dong, Qifeng Liu, Cheng Lin, Ziwei Liu, et al. Diffusion as shader: 3d-aware video diffusion for versatile video generation control. arXiv preprint arXiv:2501.03847, 2025. 3 +[30] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 2, 3 +[31] 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. 7 +[32] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 7 +[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4 +[34] Chen Hou, Guoqiang Wei, Yan Zeng, and Zhibo Chen. Training-free camera control for video generation. arXiv preprint arXiv:2406.10126, 2024. 3 +[35] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. + +Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3, 4 +[36] Yujia Huang, Adishree Ghature, Yuanzhe Liu, Ziniu Hu, Qinsheng Zhang, Chandramouli S Sastry, Siddharth Gururani, Sageev Oore, and Yisong Yue. Symbolic music generation with non-differentiable rule guided diffusion. arXiv preprint arXiv:2402.14285, 2024. 5 +[37] Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, et al. Vbench++: Comprehensive and versatile benchmark suite for video generative models. arXiv preprint arXiv:2411.13503, 2024. 7 +[38] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similar score distillation for zero-shot video editing. arXiv preprint arXiv:2403.12002, 2024. 2 +[39] Hyeonho Jeong, Chun-Hao Paul Huang, Jong Chul Ye, Niloy Mitra, and Duygu Ceylan. Track4gen: Teaching video diffusion models to track points improves video generation. arXiv preprint arXiv:2412.06016, 2024. 1 +[40] Hyeonho Jeong, Geon Yeong Park, and Jong Chul Ye. Vmc: Video motion customization using temporal attention adaptation for text-to-video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9212-9221, 2024. 2 +[41] Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6689-6700, 2024. 2 +[42] Jaemin Kim, Bryan S Kim, and Jong Chul Ye. Free 2 guide: Gradient-free path integral control for enhancing text-to-video generation with large vision-language models. arXiv preprint arXiv:2411.17041, 2024. 2, 5 +[43] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4 +[44] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1 +[45] Zhengfei Kuang, Shengqu Cai, Hao He, Yinghao Xu, Hongsheng Li, Leonidas Guibas, and Gordon Wetzstein. Collaborative video diffusion: Consistent multi-video generation with camera control. arXiv preprint arXiv:2405.17414, 2024.3 +[46] Zhengfei Kuang, Shengqu Cai, Hao He, Yinghao Xu, Hongsheng Li, Leonidas J Guibas, and Gordon Wetzstein. Collaborative video diffusion: Consistent multi-video generation with camera control. Advances in Neural Information Processing Systems, 37:16240-16271, 2025. 3 +[47] Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. arXiv preprint arXiv:2412.08629, 2024. 8 + +[48] Gihyun Kwon and Jong Chul Ye. Tweediemix: Improving multi-concept fusion for diffusion-based image/video generation. arXiv preprint arXiv:2410.05591, 2024. 2 +[49] Black Forest Labs. Flux, 2024. URL https://blackforestlabs.ai/. 2, 5, 7 +[50] Bing Li, Cheng Zheng, Wenxuan Zhu, Jinjie Mai, Biao Zhang, Peter Wonka, and Bernard Ghanem. Vivid-zoo: Multi-view video generation with diffusion model. arXiv preprint arXiv:2406.08659, 2024. 2 +[51] Teng Li, Guangcong Zheng, Rui Jiang, Tao Wu, Yehao Lu, Yining Lin, Xi Li, et al. Realcam-i2v: Real-world image-to-video generation with interactive complex camera control. arXiv preprint arXiv:2502.10059, 2025. 3 +[52] 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. 2 +[53] 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, 3 +[54] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7 +[55] Geon Yeong Park, Hyeonho Jeong, Sang Wan Lee, and Jong Chul Ye. Spectral motion alignment for video motion transfer using diffusion models. arXiv preprint arXiv:2403.15249, 2024. 2 +[56] 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. 2 +[57] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, ChihYao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1 + +[58] 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. 7 +[59] Stefan Popov, Amit Raj, Michael Krainin, Yuanzhen Li, William T Freeman, and Michael Rubinstein. Camctrl3d: Single-image scene exploration with precise 3d camera control. arXiv preprint arXiv:2501.06006, 2025. 2, 3 +[60] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 4 +[61] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. Low-rank adaptation for fast text-to-image diffusion fine-tuning, 3, 2023. 3, 4 +[62] Andrew Sanders. An introduction to Unreal engine 4. AK Peters/CRC Press, 2016. 2, 3 +[63] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. Advances in Neural Information Processing Systems, 36:39443-39469, 2023. 2, 5 +[64] Ruizhi Shao, Youxin Pang, Zerong Zheng, Jingxiang Sun, and Yebin Liu. 360-degree human video generation with 4d diffusion transformer. ACM Transactions on Graphics (TOG), 43(6):1-13, 2024. 2 +[65] 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. 4 +[66] 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. 4 +[67] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 2, 3 +[68] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7 +[69] Basile Van Hoorick, Rundi Wu, Ege Ozguroglu, Kyle Sargent, Ruoshi Liu, Pavel Tokmakov, Achal Dave, Changxi Zheng, and Carl Vondrick. Generative camera dolly: Extreme monocular dynamic novel view synthesis. In European Conference on Computer Vision, pages 313-331. Springer, 2024. 2, 3, 7 +[70] Bram Wallace, Akash Gokul, Stefano Ermon, and Nikhil Naik. End-to-end diffusion latent optimization improves classifier guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7280-7290, 2023. 2, 5 +[71] Chaoyang Wang, Peiye Zhuang, Tuan Duc Ngo, Willi Menapace, Aliaksandr Siarohin, Michael Vasilkovsky, Ivan Skorokhodov, Sergey Tulyakov, Peter Wonka, and Hsin-Ying + +Lee. 4real-video: Learning generalizable photo-realistic 4d video diffusion. arXiv preprint arXiv:2412.04462, 2024. 2, 3 +[72] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 5 +[73] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, pages 1-20, 2024. 1 +[74] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3 +[75] Daniel Watson, Saurabh Saxena, Lala Li, Andrea Tagliasacchi, and David J Fleet. Controlling space and time with diffusion models. arXiv preprint arXiv:2407.07860, 2024. 2, 3 +[76] Yujie Wei, Shiwei Zhang, Hangjie Yuan, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Feng Liu, Zhizhong Huang, Jiaxin Ye, et al. Dreamvideo-2: Zero-shot subject-driven video customization with precise motion control. arXiv preprint arXiv:2410.13830, 2024. 2 +[77] 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. 2, 3, 7 +[78] Zeqi Xiao, Wenqi Ouyang, Yifan Zhou, Shuai Yang, Lei Yang, Jianlou Si, and Xingang Pan. Trajectory attention for fine-grained video motion control. arXiv preprint arXiv:2411.19324, 2024. 2, 3, 7 +[79] Yiming Xie, Chun-Han Yao, Vikram Voleti, Huaizu Jiang, and Varun Jampani. Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470, 2024. 2 +[80] Dejia Xu, Yifan Jiang, Chen Huang, Liangchen Song, Thorsten Gernoth, Liangliang Cao, Zhangyang Wang, and Hao Tang. Cavity: Camera-controllable multi-view video diffusion with view-integrated attention. arXiv preprint arXiv:2410.10774, 2024.3 +[81] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 2, 3 +[82] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 4, 7 +[83] Shiyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2, 3 + +[84] 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. 1, 2, 3, 4, 5, 7 +[85] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8466-8476, 2024. 2 +[86] Po-Hung Yeh, Kuang-Huei Lee, and Jun-Cheng Chen. Training-free diffusion model alignment with sampling demons. arXiv preprint arXiv:2410.05760, 2024. 2, 5 +[87] Meng You, Zhiyu Zhu, Hui Liu, and Junhui Hou. Nvs-solver: Video diffusion model as zero-shot novel view synthesizer. arXiv preprint arXiv:2405.15364, 2024. 2, 3, 7 +[88] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7094-7104, 2023. 8 +[89] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3 +[90] David Junhao Zhang, Roni Paiss, Shiran Zada, Nikhil Karnad, David E Jacobs, Yael Pritch, Inbar Mosseri, Mike Zheng Shou, Neal Wadhwa, and Nataniel Ruiz. Recapture: Generative video camera controls for user-provided videos using masked video fine-tuning. arXiv preprint arXiv:2411.05003, 2024. 2, 3, 5, 7 +[91] 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. International Journal of Computer Vision, pages 1-15, 2024. 1 +[92] 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. 3 +[93] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In European Conference on Computer Vision, pages 273–290. Springer, 2024. 2 +[94] Hongkai Zheng, Weili Nie, Arash Vahdat, and Anima Anandkumar. Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305, 2023. 5 +[95] 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, 3 +[96] Qi Zuo, Xiaodong Gu, Lingteng Qiu, Yuan Dong, Zhengyi Zhao, Weihao Yuan, Rui Peng, Siyu Zhu, Zilong Dong, Liefeng Bo, et al. Videomv: Consistent multi-view generation based on large video generative model. arXiv preprint arXiv:2403.12010, 2024. 2 \ No newline at end of file diff --git a/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/images.zip b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..93650c88c248ed88a3a0f1ba99c679611e6bf929 --- /dev/null +++ b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:547f63c7677828830ad841f7ed843f37fa89c4a8ca35ebac4c96f7e312a79c73 +size 1177451 diff --git a/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/layout.json b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9d11ee1be19b228d79c1b36acfcdd7aeec9dfef5 --- /dev/null +++ b/ICCV/2025/Reangle-A-Video_ 4D Video Generation as Video-to-Video Translation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae68517925aa7e75891817aee7a13e843bb9aca995778e079e3578b3b8966083 +size 446495 diff --git a/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/68918c4c-e885-4f4a-8dd7-71574a6a8b2b_content_list.json b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/68918c4c-e885-4f4a-8dd7-71574a6a8b2b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a32053a4082fd7750430b6c175165a11153dad1e --- /dev/null +++ b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/68918c4c-e885-4f4a-8dd7-71574a6a8b2b_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72927730c6fd92c9b148f8f72961a1f08ec3be928d147de8aacfdfd0fe87887 +size 73804 diff --git a/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/68918c4c-e885-4f4a-8dd7-71574a6a8b2b_model.json b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/68918c4c-e885-4f4a-8dd7-71574a6a8b2b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0e3897a3336c2f2c788c984fd5ffd5b4c5a2039a --- /dev/null +++ b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/68918c4c-e885-4f4a-8dd7-71574a6a8b2b_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90d6a2f091835ba87cbb64c36f3f581b558d636ea6f06cd4f6315f1e2c1131b8 +size 87633 diff --git a/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/68918c4c-e885-4f4a-8dd7-71574a6a8b2b_origin.pdf b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/68918c4c-e885-4f4a-8dd7-71574a6a8b2b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c88f0e8506b835c3eefee3e6310209895a8b86a3 --- /dev/null +++ b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/68918c4c-e885-4f4a-8dd7-71574a6a8b2b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:818f5bdd8416908bcd745702c700a4f5e91ad02920cb5ac8337b80c9e19c2b54 +size 2387775 diff --git a/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/full.md b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cf6fbccc9f34eb2821363fa42916772e84fa943d --- /dev/null +++ b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/full.md @@ -0,0 +1,254 @@ +# ReasonVQA: A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering + +Duong T. Tran1,2 + +Duong.Tran@de.bosch.com + +Trung-Kien Tran + +TrungKien.Tran@de.bosch.com + +Manfred Hauswirth $^{2,3}$ + +manfred.hauswirth@tu-berlin.de + +Danh Le Phuoc $^{2,3}$ + +danh.lephuoc@tu-berlin.de + +1Bosch Center for AI, Germany 2Technical University of Berlin 3Fraunhofer FOKUS + +# Abstract + +In this paper, we propose a new dataset, ReasonVQA, for the Visual Question Answering (VQA) task. Our dataset is automatically integrated with structured encyclopedic knowledge and constructed using a low-cost framework, which is capable of generating complex, multi-hop questions. We evaluated state-of-the-art VQA models on ReasonVQA, and the empirical results demonstrate that ReasonVQA poses significant challenges to these models, highlighting its potential for benchmarking and advancing the field of VQA. Additionally, our dataset can be easily scaled with respect to input images; the current version surpasses the largest existing datasets requiring external knowledge by more than an order of magnitude. + +# 1. Introduction + +In recent years, significant advancements have been made in the field of Visual Question Answering (VQA) on standard VQA datasets [1, 5, 11, 39]. Initially, these datasets focused mainly on simple questions related to object identification and attributes, such as name, shape, color, and position. Towards the goal of general-purpose artificial intelligence, VQA models are expected to answer questions that require a deeper understanding of the world, fine-grained visual recognition, and multi-step reasoning. Recently, several additional VQA datasets [8, 12, 15, 23, 30, 34] have been introduced to challenge VQA systems to handle more complex questions. However, there are limitations associated with these datasets. Some datasets are entirely synthetic, while others rely heavily on manual human effort. + +![](images/6d9afde55955d3a323fcaee26f05c75692deb401747418616513282cb4f6cce8.jpg) + +(attribute) + +How tall is this church? + +(architecture-building) + +What is the architectural style of this church? + +location, geography) + +In which country is this church situated? + +(geography) + +What is the capital of the country where + +this church is located? + +![](images/099c5a238ca6ff3d77f4154020f6fee31a1e07434eaf6abf29b9e4f75a1d9425.jpg) +Figure 1. Sample image and questions from ReasonVQA. Using an existing image, a question is formulated by reasoning through one or multiple hops over the knowledge graph. The generated questions span diverse domains. + +In this paper, we propose a new dataset called ReasonVQA, which was developed from our low-cost and scalable framework. ReasonVQA focuses on integrating external (world) knowledge associated with objects in the images and multi-hop reasoning. For example, in Figure 1, the question "How tall is this church?" requires not only the identification of + +the church but also additional specific facts, i.e. its height. And the question "What is the capital of the country where this church is located?" additionally requires multi-hop knowledge. Such information is often dispersed across multiple paragraphs in the training text, presenting considerable challenges for VQA models. + +Our main contributions are summarized as follows: (1) We introduce a new high quality VQA benchmark, called ReasonVQA, which consists of multi-hop questions that require external knowledge to answer; (2) We propose a scalable, low-cost framework for the construction of ReasonVQA without involving too much manual effort. The largest version of ReasonVQA comprises 4.2 million questions, making it one to two orders of magnitude larger than similar existing datasets. The quality of our dataset is verifiable via a user study. Moreover, the framework is designed to be scalable with respect to the number of input images; and, (3) We evaluated various state-of-the-art VQA models, including those based on foundation models, on ReasonVQA. Our experiments demonstrate that ReasonVQA presents substantial challenges for both current VQA models and multimodal methods, highlighting its potential for benchmarking and advancing the field of VQA. + +# 2. Related Work + +Standard VQA datasets [5, 11, 21, 39] usually focus on the explicit content in the input images with different complex levels of the questions. In this section, we focus mainly on related works that incorporate external knowledge. + +Over the last few years, more datasets with external knowledge have been proposed. Noticeably, KVQA [30] contains questions that require multi-entity, multi-relation, and multi-hop reasoning over a large knowledge graph. The persons' information and images were collected from Wikidata [33], however, they were filtered and identified by human annotators. Marino et al. [23] provided OK-VQA, which includes only questions that require external resources for answering them. OK-VQA used random images from the COCO dataset [20] as the input for two rounds of crowdsourcing and the final dataset contains more than 14K images and questions. While OK-VQA is well-balanced and covers a variety of knowledge categories, its construction process is time-intensive and demands considerable human effort. Moreover, there is no control over the reasoning complexity, i.e. the number of hops in questions. + +To tackle the issue of relying on human annotations, Gao et al. [7] proposed a commonsense-based VQA system for templated-based question generation, called CRIC. Starting from dynamic templates, questions are generated automatically using functional programming with the fusion of the scene graph and the commonsense facts from the external knowledge graph. The dataset utilizes the images of Visual Genome [16] with their corresponding Scene Graph annota + +tions. In contrast to our dataset containing specific and detailed knowledge, CRIC only extracted the commonsense facts mostly from ConceptNet [31]. Recently, LORA [8] was introduced as a dataset that requires formal and complex Description Logic reasoning. It includes 200,000 automatically generated questions. However, the dataset is confined to food and kitchen scenes, with synthetic images. As a result, it is challenging to extend LORA to new domains of knowledge or to leverage various existing image sources. + +Recent efforts to incorporate external knowledge into datasets, like INFOSEEK [4] and Encyclopedic-VQA [24], consistently reveal the high cost and limited scalability of human annotation. While automation enables scale, these projects show significant human oversight remains crucial for data reliability. For instance, INFOSEEK relied on semi-automation due to cost, and Encyclopedic-VQA's automated output still required extensive human validation. This underscores the enduring need for labor-intensive quality control, especially for complex or nuanced information. + +# 3. The ReasonVQA Dataset + +The ReasonVQA dataset is created to assess the capabilities of VQA models to answer multi-hop questions and to utilize external knowledge. It contains a large number of questions categorized into three levels of complexity: 1-hop, 2-hop, and 3-hop. To answer these questions, models must possess a robust ability to reason across multiple external facts. The overview of our framework, illustrated in Figure 2, to create ReasonVQA consists of three steps: (1) External Knowledge Integration, (2) Question Generation, (3) Dataset Construction. In the External Knowledge Integration step (Sect.3.1) we "link" annotated objects in the images to corresponding entities in external knowledge graphs. In the Question Generation step, multiple-hop questions are generated semi-automatically by traveling through the knowledge graph (Sect.3.2). In the Dataset Construction step (Sect.3.3), the question distribution is balanced to reduce bias, and all the images and questions are split into train set and test set with a similar answer distribution. The detailed workflow are described as algorithms in the supplementary material. + +# 3.1. External Knowledge Integration + +Although our framework is adaptable to any knowledge base and annotated vision dataset, we chose Wikidata [33], one of the most complete structure knowledge bases, as the external knowledge source. We utilize SPARQL to seamlessly integrate image sources, such as Visual Genome (VG) [16], which contains over 108K images along with various descriptions of objects and their relationships within images, e.g., region descriptions, objects, attributes, relationships, region graphs, scene graphs, and question-answer pairs. VG has been used as a main resource for construction + +![](images/0c59c4fc8daa33b4129cf390e202cd734e32459b6155763569add4a82276662a.jpg) +Figure 2. Overview of the question generation process. The multiple choice question and answer are generated given the input image and the knowledge base. Our approach first utilizes the annotation from the dataset to link the object from the input image to a concept in knowledge base, then assembles the question from a predefined template, and finally generate the multiple choices. + +of many other VQA datasets like GQA [12] and CRIC [7]. In this paper, we leverage such existing images, questions, object descriptions and scene graphs from VG to build a more robust question generation process. By representing the relationship between an object and other elements, scene graphs allow for a more comprehensive interpretation of the image, avoiding the loss of contextual information from the visual component. The annotations of objects in VG are canonicalized to WordNet [25] synset names, which can be used to retrieve the corresponding concepts from Wikidata by using Natural Language Toolkit (NLTK) [3] and SPARQL queries. + +Wikidata has rich information about well-known entities such as landmarks and public figures. Hence, we choose Google Landmark Dataset v2 (GLDv2) [36] as another image source to expand the knowledge about spatial objects in our candidate images. GLDv2 is a benchmark for largescale, fine-grained instance recognition and image retrieval in the domain of human-made and natural landmarks. The dataset consists of over 5M images and 200k distinct instance labels to mix with VG. Since the images are annotated with Wikimedia URLs enabling us to retrieve Wikidata's knowledge, e.g linked concepts and relationships to such URLs. From such URLs, we can extract the landmark names, and the corresponding knowledge in Wikidata. + +# 3.2. Question Generation + +# 3.2.1. Template-based Question Generation + +Figure 3 shows the overview of our question generation process. We generate questions through a template-based approach, utilizing predefined templates with fillable placeholders. This method supports the creation of low-cost, scalable systems by enabling a single template to yield many questions. Although the templates are manually crafted, they are designed to be generic enough to provide + +![](images/761f4ca5dbcf9f3bddd70de5366af79f9ffa853142d95eeca839a0d20d5f03f4.jpg) +Figure 3. The process of question generation step-by-step from a Wikidata concept to the multiple choice question. The class hierarchy and all properties of the concept are queried and filtered down to a subset. Then each of the remaining properties is used to generate a multiple choice question. + +extensive coverage. To boost linguistic diversity, we also develop linguistic variations for each template. A key advantage of our proposed method is its ability to combine multiple templates into a single question, further enhancing system scalability. Furthermore, as we construct potential templates based on knowledge gleaned from the knowledge base (KB), the diversity of the templates aligns with the breadth of knowledge available in the KB. + +One-hop Question Generation. Let us assume we are constructing questions around an object in an image, referred to as the main object. This main object corresponds to a Wikidata entity, from which we extract factual knowledge to generate questions. After obtaining external facts from Wikidata for the main object as described in Section 3.1, we use a predefined template for each property (or relation, interchangeably) occurring in these facts. Each template contains a placeholder, which will be filled later by the class name of the main object, resulting in 1-hop question. For Wikidata, the class name of an object is simply the value of either the "instance of" or "subclass of" prop- + +erty. The Figure 4 shows an example of 1-hop template question. In this case, the template is "Who designed ____?" and the class name "skyscraper" is used to fill in the placeholder, producing the complete question "Who designed this skyscraper?". Note that during the process of retrieving facts about the main object from Wikidata, we also obtain the answer to the question. For example, the answer for the mentioned question is "César Pelli". + +Multi-hop Question Generation. Multi-hop questions about a main object are constructed by filling the placeholders in 1-hop question templates, which we will now refer to as the main template, with a sub-clause. This sub-clause is formed using a sub-clause template, which is also manually designed for each relation. For example, the sub-clause template for the relation "architect" is "the architect of _. The placeholder in a sub-clause template can be filled by either the class name of the main object or another sub-clause. This is the cornerstone of our multi-hop question construction mechanism. We additionally leverage the scene graph annotations in VG to construct a sub-clause, e.g. "___ parked next to the sidewalk". This enables us to better integrate semantic visual information into the questions. Figure 4 show examples of the templates and the corresponding questions for 2-hop and 3-hop questions. + +Additionally, we categorize each question into one or more domains based on the property used to generate the question. For instance, a question about the birthday of a famous person, derived from the property "date of birth," can be classified under both the "person" and "history-events" domains. This enables the system to receive input as one or multiple specified domains and generate questions accordingly. This mechanism is particularly useful for narrowing down the dataset to a specific domain, facilitating more focused and refined results. For this, we predefined 20 domains that encompass all aspects of human knowledge. A detailed list of these domains can be found in Section 2 of the supplementary material. + +# 3.2.2. False Choice Generation + +To generate challenging multiple choice questions, in addition to the correct answer, we need to create non-trivial false choices. First, we assign a question into one of four categories based on its correct answer: fixed, date, number, and literal. The fixed category contains questions that have an enclosed set of choices, such as questions about gender, continent, country, and language. The date, number and literal categories contain questions about date time, numeric values, and literal values, respectively. For the questions in the fixed categories, the false choices are randomly selected from the corresponding set of choices. For the date category, the false choices are generated by randomly selecting $N$ date values within $[-10, + 10]$ years. For the number category, the false choices are generated by randomly selecting $N$ numbers within a specific range $[m,n]$ , where $m = \frac{i}{2}$ + +$n = \max (1.5i, m + 2N)$ with $i$ being the correct answer. For the literal category, the false choice is generated by retrieving $N$ values that come from the same property as the correct answer. In practice, we set $N = 3$ , meaning that there is up to three false choices for each correct answer. + +# 3.3. Dataset Construction + +# 3.3.1. Answer Distribution Balancing + +One of the prevalent problems of existing VQA datasets is the strong bias toward specific answers that allows ML models to assume the output without completely understanding the input. Therefore, we look at how to smooth out the answer distribution to alleviate this issue. Inspired by GQA [12], we compute the distribution of all the true answers for each group, where a group is defined by its corresponding property. We sort the answers by the number of occurrences and select only the top ten most frequent answers in each group for the next step. Then, we create a smoother answer distribution using the following process: We start by going through the answers in a specific group in decreasing frequency order and repeatedly remove ones, together with their questions, to make the head size comparable to the tail size. The head and tail are respectively the group of questions having the most and the least number of answers. Essentially, answers with higher frequency will be discarded more often. Thus, this process helps the benchmark become less biased, more balanced and also more challenging to VQA models. While repeating this operation, we maintain the frequency order by keeping minimum and maximum ratios between each pair of consecutive answers. We also prioritize removing answers from images with the most number of questions. Doing that helps us to equalize the number of questions between images and avoid discarding images as much as possible. The balancing process is applied in an iterative manner, with the answer distribution becoming more and more uniform after each iteration. The visualization of our balancing process is provided in Section 1.3 of the supplementary material. + +# 3.3.2. Dataset Splitting + +Since it is necessary to maintain the answer distribution between the train set and test set, the dataset is split through the following procedure: We distribute images into categories and then divide each category randomly by the same train/test ratio. For each image, we first compute the frequency of every answer and select answers that appear at least $1\%$ in the answer pool, which is the set of all the answers from the dataset. Next, we assign each image to a category, where the category name is constructed by concatenating its top two most frequent answers. We also combine all the categories containing only a single image. Finally, images in each category are split randomly into $70\%$ train and $30\%$ test, together with their question-answer pairs. + +![](images/d87bd13d0a011b6c39147e33d82aef4ea44c362cc41392fde5a993f98fedca42.jpg) +Figure 4. Example of constructing questions using template-based approach. A 1-hop question is constructed by filling the class name of the main object into a template. Meanwhile, 2-hop and 3-hop questions are constructed in a similar manner by nesting two or more templates. + +![](images/f96db1c697b10c19ea01e2ebedf32cd3385eebe17f55b6275373617161e97435.jpg) + +The resulting sets are guaranteed not to have overlapping images, and all the questions about a given image appear in the same split. Please refer to Section 1.4 in the supplementary material for visualization depicting the similarity in answer distribution between the two sets. + +# 4. Dataset Analysis and Experiments + +# 4.1. Dataset Statistics and Quality Evaluation + +The latest version of ReasonVQA consists of more than 598K images and about 4.2M questions. The scalability of our dataset depends on two key factors: (1) numerous objects share common attributes; and (2) multi-hop questions are constructed by nesting multiple templates. For instance, since many facts about building share the "height" attribute, we only need to define a single template for this property during question generation. To assess state-of-the-art approaches, we derived a subset version of our dataset, ReasonVQA-U. The main objective is to create a manageable subset of our dataset to assess and compare these models efficiently. The answer distribution balancing process was applied on this subset to obtain the balanced version ReasonVQA-B. + +In Table 1, we compare our dataset with various VQA datasets. In terms of the number of images, ours has the highest quantity. Moreover, we have a significantly larger number of questions compared to all the others, except GQA. To assess the quality of our dataset, we conducted a user study over 1000 randomly selected question and image pairs. The study involves 20 participants, who use English as the main working and/or studying language, to evaluate the correctness of the answers and naturalness of 50 randomly selected questions each. They were asked to rate the + +
Dataset#Images#QuestionsKB
DAQUAR [21]1.4K12K✘
COCO-QA [29]69K117K✘
VQA v2 [9]204K1.1M✘
CLEVR [15]100K999K✘
GQA [12]113K22M✘
PQA [27]32.9K157K✘
OK-VQA [23]14K14K✓
KVQA [30]24K183K✓
CRIC [7]96K494K✓
LORA [8]100K200K✓
InfoSeek [4]8.9K1.35M✓
Encyclopedic VQA [24]514K1M✓
ReasonVQA-U13.3K78K✓
ReasonVQA-B8.7K51.9K✓
ReasonVQA Full598.5K4.2M✓
+ +Table 1. Comparison between ReasonVQA and previous VQA datasets. The top section contains VQA datasets without knowledge component. The middle and bottom sections contain knowledge-based VQA datasets. + +naturalness on a four-level scale: (1) very unnatural; (2) unnatural; (3) natural; and (4) very natural. Additionally, participants also marked questions with grammatical errors. The user study showed that $96\%$ of them are correct. The primary sources of error comes from numerical values, e.g. the area of cities, and ambiguities in the height or length of + +certain constructions. We also observed that $2.20\%$ of the questions were rated as “very unnatural,” $13.9\%$ as “unnatural,” $58.1\%$ as “natural,” and $25.8\%$ as “very natural,” with only $2.5\%$ containing grammatical errors. This user study indicates that more than $83\%$ of the questions are natural to humans, demonstrating that our dataset already has a good level of naturalness but still has room for improvement. Additional dataset statistics and results of the user study can be found in the Section 2 of the supplementary material. + +We categorize each question into one or more domains based on the property used to generate the question. For instance, a question about the birthday of a famous person, derived from the property "date of birth," can be classified under both the "person" and "history-events" domains. This enables the system to receive input as one or multiple specified domains and generate questions accordingly. We predefined 20 domains covering a wide range of human knowledge. During question generation, we select only properties associated with the queried domains, filtering out others to optimize efficiency. Figure 5 visualizes the domain distribution. The complete list of domains is available in Section 2 of the supplementary material. + +![](images/9610ff26fcc5b76118791b257e9df3f803e9576dc47ce72d9dca1f49bc98064b.jpg) +Figure 5. Distribution of questions across domains. Each question is categorized into specific domains based on the property from which it was generated. + +# 4.2. Baseline Experiments and Analysis + +To rigorously evaluate the challenges posed by our dataset in knowledge-based visual question answering tasks, we perform comprehensive assessments using both the unbalanced and balanced versions of our dataset across a wide range of state-of-the-art vision-language models: BLIP-2 [19], InstructBLIP [6], mPLUG-Owl2 [38], Idefics2 [17], Mantis-SigLIP [14], Mantis-Idefics2 [14], mPLUG-Owl3 [37], GPT-4o [13], LLaVAOV [18], Qwen2.5-VL [2], PaliGemma-2 [32], PaliGemma-2-Mix [26], SmolVLM-Instruct [22]. + +Zero-shot Evaluation and Dataset Comparison. We assessed the zero-shot performance of all models across two scenarios: open-ended and multiple choice. For each + +model, we constructed prompts directly using the questions from ReasonVQA, with the associated images serving as the input. In open-ended scenario, we collect and evaluate the text answers generated by each LLM. For the multiple-choice setup, we carefully design prompts to instruct each LLM to respond with only a single letter that corresponds to its predicted choice. + +Fine-tuning and Performance Gains. To assess the effectiveness of fine-tuning, we compared the performance of three vision-language models of varying scales before and after training on our dataset. Models were fine-tuned using direct image, question, and answer inputs. This experiment evaluates how well models can adapt to the dataset and improve their understanding of knowledge-based visual question answering. By analyzing performance gains, we demonstrate the potential of fine-tuning in enhancing model capabilities beyond their zero-shot performance. + +Impact of Dataset Scale on Model Performance. To assess the performance and scalability of LLMs, we also benchmark them using the ReasonVQA at multiple size increments. By gradually increasing the number of questions and images, we aim to observe how each model's accuracy relative to the amount of data. This approach will allow us to identify any limitations in model capacity, potential bottlenecks, and the overall robustness of each model under different data volumes. The results and visualizations are presented in Section 3 of the supplementary material. + +# 4.3. Metrics + +We report the accuracy of the LLMs to assess their performance. In open-ended scenario, where the LLM outputs are in free-text format, comparing responses directly with the ground truth answers can be challenging. To address this, we tested several string-matching methods to compare the output with the ground truth answer. We employ three approaches - exact match, substring, and semantic similarity - which offer different strategies for aligning and scoring the outputs based on varying degrees of match tolerance and answer format flexibility. The exact match score is 1 if the answer provided by a model precisely matches the ground truth, and 0 otherwise. However, since foundation models often provide free-text answers, exact match may not accurately reflect performance. The substring score is defined as 1 if the answer provided by the model is contained in the list of words from the ground truth answer, and 0 otherwise. Nevertheless, this approach falls short in many cases; for example, when the ground truth is "male" and the prediction is "a man," the substring score is 0, which is incorrect. Therefore, the semantic similarity score is necessary to capture the semantic quality of the generated answer and provide a more meaningful evaluation of the LLMs. The score is based on the similarity of the generated answer and the ground truth computed using a widely recognized simi + +
ModelR-UR-BOK-VQAVQAv2GQA
BLIP-246.446.145.965.044.7
InstructBLIP43.743.966.369.847.9
mPLUG-Owl222.122.257.779.456.1
Idefics250.849.754.670.3-
Mantis-SigLIP37.637.255.474.9-
Mantis-Idefics243.742.952.677.6-
mPLUG-Owl347.944.860.182.165.0
GPT-4o62.860.871.8-
LLaVA-OV50.550.666.981.1-
Qwen2.5-VL59.358.184.988.6-
PaliGemma-240.039.863.671.8-
PaliGemma-2-Mix43.741.886.890.4-
SmolVLM-Instruct34.531.749.135.4-
Mantis-SigLIP67.867.699.295.5-
Mantis-Idefics268.768.598.997.3-
mPLUG-Owl368.968.199.197.8-
GPT-4o76.673.496.7--
LLaVA-OV56.852.494.696.1-
PaliGemma-2-Mix65.664.795.897.8-
+ +Table 2. The benchmark scores across various models on our datasets, compared to those on previous VQA datasets. The top section presents scores in the open-ended scenario, while the bottom section shows scores in the multiple-choice scenario. The accuracies are computed using the semantic similarity string-matching method. We emphasize the lowest scores for each row. "R-U" and "R-B" refer to ReasonVQA-U and ReasonVQA-B. + +larity all-MiniLM-L6-v2 [28], which is a compact, efficient model for semantic textual similarity tasks. The reported scores are the average values over the entire dataset. Section 3 of the supplementary material provides a more comprehensive report of the inference results. + +# 4.4. Experimental Results + +Table 2 presents the accuracies of tested models on two versions (balanced and unbalanced) of our dataset and on three datasets: VQAv2, OK-VQA, and GQA. We observed that in most experiments, the scores decreased, sometimes significantly, when evaluating models on our datasets compared to their performance on previous VQA datasets. In addition, the accuracies on our balanced version are slightly lower to those on the unbalanced version. + +In the zero-shot open-ended scenario, older models like BLIP-2, InstructBLIP, and mPLUG-Owl2 show moderate performance, with a notable drop on ReasonVQA compared to VQAv2 and OK-VQA and a slight decrease relative to GQA. mPLUG-Owl2, in particular, struggles significantly, indicating a substantial gap between the model's general + +VQA capabilities and its ability to handle the complexities of our datasets. Moving to more recent models, GPT-4o, Qwen2.5-VL, and PaliGemma-2-Mix emerge as the top-performing approaches. GPT-4o demonstrates its advanced reasoning and multimodal capabilities by achieving the highest accuracy. Qwen2.5-VL also performs strongly, surpassing other models across benchmarks. PaliGemma-2-Mix performs exceptionally well on previous datasets but sees a sharp drop on ReasonVQA, suggesting that even state-of-the-art models may not generalize effectively to our datasets. SmolVLM-Instruct, as the most lightweight model among the recent ones, exhibits the lowest performance across all datasets. Its accuracy on our dataset is particularly low, highlighting its limited capability in handling the complexity of our benchmark. On the other hand, all tested models were able to achieve remarkably higher accuracies in the zero-shot multiple choice scenario. Nevertheless, the scores on ReasonVQA experienced a significant drop compared to those on other datasets. + +For recent models, we also reported in Figure 6 the accuracies in four string-matching approaches across different categories: the number of hop in the questions (1-hop, 2-hop, 3-hop); the availability of scene graph and the image source. Our results revealed that exact matching yields low overall accuracies because these models treat questions as open-ended, generating varied, free-text responses that, while often not entirely incorrect, lack the specificity required for exact matches (e.g., "very high" instead of "390 meters"). The substring approach, which considers all words from both the predicted answers and the ground truth, yields more positive scores. Semantic similarity matching consistently achieves higher scores, as it considers synonymous expressions, offering a more adaptable evaluation that closely aligns with human judgment. + +Regarding the number of hops in questions, it is consistent across all tested models that they perform much worse on 3-hop questions compared to 1-hop and 2-hop questions. This confirms the high complexity of 3-hop questions and the need for proper reasoning steps to answer them accurately. However, it is worth noting that the performance of most models on 1-hop questions is only slightly better than on 2-hop questions, and in some cases, even worse. One possible reason for this is that while 2-hop questions are indeed more complex, they are generally longer and provide more contextual information. Consequently, certain models are able to effectively utilize the additional context in 2-hop questions, resulting in better performance. Figure 6 also presents scores for two types of questions: those generated with information from scene graphs and those generated without such information. Generally, the scores for questions with scene graph information are lower than the other. This suggests that incorporating scene graphs to generate questions indeed increases the complexity of Reason- + +![](images/f4918567ab1b1f1570f24fd328276c1684599b819af8b565b91d8976a0a59a38.jpg) + +![](images/86ae0757e041d7f6856cef6d81729eb7b1d50e35a4b89ccdc42d9bfb58f5e26f.jpg) + +![](images/3a33b5221249ef52c1e21a9d01b0a04baead5b56873ce59668417fb3e8500e5a.jpg) +Figure 6. The evaluation results in four string-matching strategies by the number of hops, by the availability of scene graph, and by the source CV dataset. "Acc." refers to the overall accuracy. "SG" refers to scene graph. "VG" and "GLDv2" refer to Visual Genome and Google Landmarks Datasets v2 respectively. + +![](images/740980c91ce06063a7bff7471360e53627193c9cbfb7478134871ac656507ab4.jpg) + +
ModelZero-shot Acc.Fine-tuned Acc.
Qwen2-VL-7B-Instruct59.065.0 (+10.1%)
PaliGemma-2-3B-Mix40.166.8 (+66.5%)
PaliGemma-2-10B-Mix65.674.5 (+13.5%)
+ +Table 3. Fine-tuning performance of three LLMs on our dataset. The table compares accuracy before and after fine-tuning, highlighting the improvement achieved through training. Improvement percentages are shown in parentheses. + +VQA. This valuable insight helps to guide future efforts to further increase the complexity of the dataset. + +For the fine-tuning experiment, we fine-tune three vision-language models, Qwen2-VL-Instruct 7B [35], PaliGemma-2-Mix 3B, and PaliGemma-2-Mix 10B [26, 32], on our dataset to evaluate its effectiveness in adapting pre-trained models to knowledge-based VQA. Finetuning is conducted using LoRA [10] to reduce computational overhead while preserving model capacity. For each model, we compare performance before and after finetuning using ReasonVQA train set. Table 3 presents the results on the test set, showing a significant improvement in accuracy across these models, demonstrating the dataset's + +effectiveness in knowledge-based reasoning. Notably, the PaliGemma-2-3B-Mix achieves significant gains with lower resource, suggesting that smaller models have more room for enhancement when adapted to a specific dataset. + +# 5. Conclusions + +We have introduced a novel, cost-effective, and scalable VQA dataset that requires external knowledge for answering questions, demanding minimal manual effort. Our construction framework leverages existing CV dataset annotations and external knowledge base Wikidata. While currently built from specific image sources, the framework is generic, allowing expansion to integrate images from other CV datasets and maintain question diversity. Users can also tailor the dataset to specific interests by selecting from 20 predefined knowledge domains. ReasonVQA features complex questions derived from scene and multi-hop knowledge graphs, proving highly challenging for state-of-the-art VQA models. We provide source code for easy integration of new images and knowledge sources, enabling researchers to expand and customize their own datasets. We believe this resource will be valuable for evaluating existing models and advancing more robust VQA systems. + +# Acknowledgements + +This work was partially funded by the European Union's programme under grant agreement No.101092908 (SMARTEDGE), by the Chips Joint Undertaking (JU), European Union (EU) HORIZON-JU-IA, under grant agreement No. 101140087 (SMARTY) and by the German Research Foundation (DFG) under the COSMO project (ref. 453130567). + +# References + +[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6077-6086. Computer Vision Foundation / IEEE Computer Society, 2018. 1 +[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Ming-Hsuan Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report. CoRR, abs/2502.13923, 2025. 6 +[3] Steven Bird, Ewan Klein, and Edward Loper. Natural Language Processing with Python. O'Reilly, 2009. 3 +[4] Yang Chen, Hexiang Hu, Yi Luan, Haitian Sun, Soravit Changpinyo, Alan Ritter, and Ming-Wei Chang. Can pre-trained vision and language models answer visual information-seeking questions? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pages 14948-14968. Association for Computational Linguistics, 2023. 2, 5 +[5] Zhuo Chen, Jiaoyan Chen, Yuxia Geng, Jeff Z. Pan, Zonggang Yuan, and Huajun Chen. Zero-shot visual question answering using knowledge graph. In The Semantic Web - ISWC 2021 - 20th International Semantic Web Conference, ISWC 2021, Virtual Event, October 24-28, 2021, Proceedings, pages 146-162. Springer, 2021. 1, 2 +[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 6 +[7] Difei Gao, Ruiping Wang, Shiguang Shan, and Xilin Chen. CRIC: A VQA dataset for compositional reasoning on vision and commonsense. IEEE Trans. Pattern Anal. Mach. Intell., 45(5):5561-5578, 2023. 2, 3, 5 +[8] Jingying Gao, Qi Wu, Alan Blair, and Maurice Pagnucco. Lora: A logical reasoning augmented dataset for visual ques + +tion answering. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1, 2, 5 +[9] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6325-6334. IEEE Computer Society, 2017. 5 +[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 The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. Open-Review.net, 2022. 8 +[11] Ronghang Hu, Amanpreet Singh, Trevor Darrell, and Marcus Rohrbach. Iterative answer prediction with pointer-augmented multimodal transformers for textvqa. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 9989-9999. Computer Vision Foundation / IEEE, 2020. 1, 2 +[12] Drew A. Hudson and Christopher D. Manning. GQA: A new dataset for real-world visual reasoning and compositional question answering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 6700-6709. Computer Vision Foundation / IEEE, 2019. 1, 3, 4, 5 +[13] Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrom, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Madry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, Alex Nichol, Alex Paino, Alex Renzin, Alex Tachard Passos, Alexander Kirillov, Alexi Christakis, Alexis Conneau, Ali Kamali, Allan Jabri, Allison Moyer, Allison Tam, Amadou Crookes, Amin Tootoonchian, Ananya Kumar, Andrea Vallone, Andrej Karpathy, Andrew Braunstein, Andrew Cann, Andrew Codispoti, Andrew Galu, Andrew Kondrich, Andrew Tulloch, Andrey Mishchenko, Angela Baek, Angela Jiang, Antoine Pelisse, Antonia Woodford, Anuj Gosalia, Arka Dhar, Ashley Pantuliano, Avi Nayak, Avital Oliver, Barret Zoph, Behrooz Ghorbani, Ben Leimberger, Ben Rossen, Ben Sokolowsky, Ben Wang, Benjamin Zweig, Beth Hoover, Blake Samic, Bob McGrew, Bobby Spero, Bogo Giertler, Bowen Cheng, Brad Lightcap, Brandon Walkin, Brendan Quinn, Brian Guarraci, Brian Hsu, Bright Kellogg, Brydon Eastman, Camillo Lugaresi, Carroll L. Wainwright, Cary Bassin, Cary Hudson, Casey Chu, Chad Nelson, Chak Li, Chan Jun Shern, Channing Conger, Charlotte Barette, Chelsea Voss, Chen Ding, Cheng Lu, Chong Zhang, Chris Beaumont, Chris Hallacy, Chris Koch, Christian Gibson, Christina Kim, Christine Choi, Christine McLeavey, Christopher Hesse, Claudia Fischer, Clemens Winter, Coley Czarnecki, Colin Jarvis, Colin Wei, Constantin Koumouzelis, and Dane Sherburn. Gpt-4o system card. CoRR, abs/2410.21276, 2024. 6 + +[14] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. Trans. Mach. Learn. Res., 2024, 2024. 6 +[15] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross B. Girshick. CLEVR: A diagnostic dataset for compositional language and elementary visual reasoning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1988-1997. IEEE Computer Society, 2017. 1, 5 +[16] 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. Int. J. Comput. Vis., 123(1):32-73, 2017. 2 +[17] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 6 +[18] 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. Trans. Mach. Learn. Res., 2025, 2025. 6 +[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 19730-19742. PMLR, 2023. 6 +[20] Tsung-Yi Lin, Michael Maire, Serge J. 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, pages 740-755. Springer, 2014. 2 +[21] Mateusz Malinowski and Mario Fritz. A multi-world approach to question answering about real-world scenes based on uncertain input. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 1682-1690, 2014. 2, 5 +[22] Andrés Marafioti, Orr Zohar, Miquel Farre, Merve Noyan, Elie Bakouch, Pedro Cuenca, Cyril Zakka, Loubna Ben Allal, Anton Lozhkov, Nouamane Tazi, Vaibhav Srivastav, Joshua Lochner, Hugo Larcher, Mathieu Morlon, Lewis Tunstall, Leandro von Werra, and Thomas Wolf. Smolvlm: Redefining small and efficient multimodal models. arXiv preprint arXiv:2504.05299, 2025. 6 +[23] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. OK-VQA: A visual question answering benchmark requiring external knowledge. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 3195-3204. Computer Vision Foundation / IEEE, 2019. 1, 2, 5 + +[24] Thomas Mensink, Jasper R. R. Uijlings, Lluís Castrejón, Arushi Goel, Felipe Cadar, Howard Zhou, Fei Sha, André Araújo, and Vittorio Ferrari. Encyclopedic VQA: visual questions about detailed properties of fine-grained categories. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3090-3101. IEEE, 2023. 2, 5 +[25] George A. Miller. Wordnet: A lexical database for english. Commun. ACM, 38(11):39-41, 1995. 3 +[26] Merve Noyan, Aritra Roy Gosthipaty, and Andreas P. Steiner. Paligemma 2 mix - new instruction vision language models by google. arXiv preprint arXiv:2412.03555, 2025. 6, 8 +[27] Yonggang Qi, Kai Zhang, Aneeshan Sain, and Yi-Zhe Song. PQA: perceptual question answering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 12056-12064. Computer Vision Foundation / IEEE, 2021. 5 +[28] 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. 7 +[29] Mengye Ren, Ryan Kiros, and Richard S. Zemel. Exploring models and data for image question answering. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 2953-2961, 2015. 5 +[30] Sanket Shah, Anand Mishra, Naganand Yadati, and Partha Pratim Talukdar. KVQA: knowledge-aware visual question answering. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 8876-8884. AAAI Press, 2019. 1, 2, 5 +[31] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, February 4-9, 2017, San Francisco, California, USA, pages 4444-4451. AAAI Press, 2017. 2 +[32] Andreas Steiner, André Susano Pinto, Michael Tschannen, Daniel Keysers, Xiao Wang, Yonatan Bitton, Alexey A. Gritsenko, Matthias Minderer, Anthony Sherbondy, Shangbang Long, Siyang Qin, R. Reeve Ingle, Emanuele Bugliarello, Sahar Kazemzadeh, Thomas Mesnard, Ibrahim Alabdul-mohsin, Lucas Beyer, and Xiaohua Zhai. Paligemma 2: A family of versatile vlms for transfer. CoRR, abs/2412.03555, 2024. 6, 8 +[33] Denny Vrandecic and Markus Krötzsch. Wikidata: a free collaborative knowledgebase. Commun. ACM, 57(10):78-85, 2014. 2 +[34] Peng Wang, Qi Wu, Chunhua Shen, Anthony R. Dick, and Anton van den Hengel. FVQA: fact-based visual question answering. IEEE Trans. Pattern Anal. Mach. Intell., 40(10): 2413-2427, 2018. 1 + +[35] 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. CoRR, abs/2409.12191, 2024. 8 +[36] Tobias Weyand, André Araujo, Bingyi Cao, and Jack Sim. Google landmarks dataset v2 - A large-scale benchmark for instance-level recognition and retrieval. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 2572-2581. Computer Vision Foundation / IEEE, 2020. 3 +[37] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplugowl3: Towards long image-sequence understanding in multimodal large language models. CoRR, abs/2408.04840, 2024. 6 +[38] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. CoRR, abs/2311.04257, 2023. 6 +[39] Jing Yu, Zihao Zhu, Yujing Wang, Weifeng Zhang, Yue Hu, and Jianlong Tan. Cross-modal knowledge reasoning for knowledge-based visual question answering. Pattern Recognit., 108:107563, 2020. 1, 2 \ No newline at end of file diff --git a/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/images.zip b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f293960a78d6d91581c8b22c1030a3609e0d22de --- /dev/null +++ b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99261e0bcdbe524787c62c407b9b1fb8db29db1a937a1d818dadbcf68effaedb +size 565263 diff --git a/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/layout.json b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fa02d31daf5e9f64afcb938c4461aa77f163584a --- /dev/null +++ b/ICCV/2025/ReasonVQA_ A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad9249041a290eb6d57068c8b07b350fada3d638d5aeb70394ed12322adf4f92 +size 275727 diff --git a/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/cb93cbc0-8a2b-4340-917d-d03dcb8e957f_content_list.json b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/cb93cbc0-8a2b-4340-917d-d03dcb8e957f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..11d565c13072928487b2ceb908ea9ff3eb272b73 --- /dev/null +++ b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/cb93cbc0-8a2b-4340-917d-d03dcb8e957f_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c97c2589635d20efcc35351f388a43ce6da06e9e1e0b52be4b47d1a2a4d60ebe +size 72924 diff --git a/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/cb93cbc0-8a2b-4340-917d-d03dcb8e957f_model.json b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/cb93cbc0-8a2b-4340-917d-d03dcb8e957f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..396af2593e58023fc8adc2fb8d68eef64832b7a5 --- /dev/null +++ b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/cb93cbc0-8a2b-4340-917d-d03dcb8e957f_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fdc59ba543e8528ec1983905f252b042b0c34bb6063f145a3d9b7c4044bf5f49 +size 88298 diff --git a/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/cb93cbc0-8a2b-4340-917d-d03dcb8e957f_origin.pdf b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/cb93cbc0-8a2b-4340-917d-d03dcb8e957f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7d086f7c0f46c340963e20cd3802701170d2980b --- /dev/null +++ b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/cb93cbc0-8a2b-4340-917d-d03dcb8e957f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5917cf9128a6e14aac3f504ed639d4a36554d8615402de3436a86bfb630b384a +size 18112499 diff --git a/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/full.md b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cbd2c6be57541fb6ba922789eb0b25d0d3be3922 --- /dev/null +++ b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/full.md @@ -0,0 +1,278 @@ +# ReassembleNet: Learnable Keypoints and Diffusion for 2D Fresco Reconstruction + +Adeela Islam $^{1,2}$ Stefano Fiorini $^{1}$ Stuart James $^{3}$ Pietro Morerio $^{1}$ Alessio Del Bue $^{1}$ $^{1}$ Fondazione Istituto Italiano di Tecnologia $^{2}$ University of Genova $^{3}$ Durham University +{adeela.islam, pietro.morerio}@iit.it + +# Abstract + +The task of reassembly is a significant challenge across multiple domains, including archaeology, genomics, and molecular docking, requiring the precise placement and orientation of elements to reconstruct an original structure. In this work, we address key limitations in state-of-the-art Deep Learning methods for reassembly, namely i) scalability; ii) multimodality; and iii) real-world applicability: beyond square or simple geometric shapes, realistic and complex erosion, or other real-world problems. We propose ReassemblyNet, a method that reduces complexity by representing each input piece as a set of contour keypoints and learning to select the most informative ones by Graph Neural Networks pooling inspired techniques. ReassemblyNet effectively lowers computational complexity while enabling the integration of features from multiple modalities, including both geometric and texture data. Further enhanced through pretraining on a semi-synthetic dataset. We then apply diffusion-based pose estimation to recover the original structure. We improve on prior methods by $57\%$ and $87\%$ for RMSE Rotation and Translation, respectively. + +# 1. Introduction + +Reassembly requires placing each element in its correct position and orientation to form the original shape as a whole – whether it be a 2D or 3D object. This ability is a form of spatial intelligence, which refers to the capacity to accurately perceive the visual-spatial environment and manipulate that perceived space [9]. This skill is typically evaluated through reassembly tasks, where individual components must be arranged and connected to form a coherent, functional whole—such as solving 2D jigsaw puzzles or assembling 3D structures with LEGO blocks. + +Since the advent of the first puzzle solver [5], reassembly tasks have posed a significant challenge to the machine learning community due to their inherent combinatorial complexity. These challenges are further underscored by + +![](images/aff52f622ba33ccc0428d7a1f6235fd582d23d794aca225c1fe15f50594b8ae3.jpg) +Figure 1. We introduce ReassembleNet as a method for fresco reassembly. Our ReassembleNet addresses key challenges that have been ignored by traditional methods, including complex geometry, texture and erosion, missing pieces, often in a data-scarcity scenario. + +their wide range of applications, including genomics [21], assistive technologies [39], and molecular docking [3]. + +In recent years, AI techniques for reassembling objects have gained increasing traction in the heritage field, particularly in fresco reconstruction [38]. This growing application of AI is driven by the fundamental challenge archaeologists face in reconstructing the past. After extensive and painstaking work involving site surveys and excavations, they are often confronted with the daunting task of reassembling countless fragments of varying sizes, shapes, and appearances to recreate ancient artifacts or artworks. Depending on the complexity of the artifact, this reassembly process can span months, years, or even decades. In cases involving a vast number of pieces, the task may become nearly insurmountable, regardless of the skill of the experts. Despite recent advances, state-of-the-art methods for solving this task remain far from achieving a satisfactory accuracy [36]. The challenges persist not only in the 3D domain but also in the 2D space, where the reduced degrees of + +freedom do not significantly alleviate the complexity of the problem. Reduction to 2D has been used for a long time to aid in finding the solutions as many heritage objects, e.g. frescos, are planar on the dominant surface [6]. Traditional approaches use keypoints or patches [4, 22, 25, 35] benefiting from optimization strategy allowing them to scale to complex irregular shapes, in contrast to deep learning approaches. Three primary factors contribute to the difficulty of fresco reassembly for deep learning approaches. First, the inherent complexity of irregular pieces and the diverse textures present in the fragments (Figure 1). Secondly, as a byproduct of the first, complex polygonal shapes are difficult to model in deep learning approaches, where comparisons need to be made across the points of one fragment edge to another fragment. Finally, deep learning methods typically require large-scale datasets to effectively learn meaningful patterns [32], a requirement that is difficult to meet in most applications domains of reassembly. + +To address these challenges, we propose ReassembleNet, a scalable deep learning method that integrates a diffusion process with multi-modal feature fusion. Our approach enables scalability on irregular 2D shapes through keypoint selection, a crucial capability for highly fragmented structures that existing methods struggle to process [18]. We introduce a semi-synthetic dataset specifically designed to highlight the characteristics of fresco pieces. Through finetuning, we demonstrate that pre-training on a semi-synthetic dataset significantly enhances model performance on real-world frescos. We combine the keypoint selection with geometry and texture representation to constrain the matching problem modeled through an inter and intra-piece attention block. Finally, we wrap this model in a diffusion process to iteratively denoise the pose of pieces into the final locations. + +The main contributions of this work are: i) We propose a scalable end-to-end deep learning method designed to solve reassembly tasks with 2D irregular input shapes. ii) We introduce a 2D learnable keypoint selector that identifies the most significant keypoints along the borders of 2D irregular shapes. This approach reduces complexity, enhances scalability, and overcomes the computational limitations of SOTA models that struggle when using all points of the contours. iii) We use multimodal features that incorporate geometric features, local and global texture features, to provide a richer representation and aid solving the task. iv) We address the limitation of scarce training data by introducing a novel pipeline with a reduced sym-to-real gap that generates a semi-synthetic dataset for pre-training. This strategy enhances overall performance when solving real puzzles. + +# 2. Related Works + +We review key literature on reassembly tasks for both square and irregular shape puzzles. + +# 2.1. Reassembly of Square-based Puzzles + +Square-based jigsaws have been extensively studied in the literature, with early approaches using greedy methods. For instance, [24] introduced a three-phase greedy approach involving placement, segmentation, and shifting, relying on compatibility functions and estimation metrics. While effective for large puzzles, it requires predefined orientations and does not handle missing pieces. Similarly, [7] proposed a tree-based algorithm for puzzles with unknown orientations and locations, using Mahalanobis Gradient Compatibility (MGC) for boundary analysis. However, it struggles with color-based similarity and missing pieces. [23] presented a greedy strategy for puzzles without prior information on piece sizes or orientations, refining the initial configuration. It works with mixed or missing pieces but struggles with noisy images. Other methods, such as [19], combine edge and content similarity using MGC, but face challenges with large-scale puzzles. [29] used a genetic algorithm based on color similarity to solve puzzles with unknown locations and orientations. + +In recent years, several deep learning approaches have been developed for puzzle reconstruction. One prominent direction leverages generative models to reconstruct a complete image from an unordered set of puzzle pieces [33, 34]. These methods aim to infer the global structure of the puzzle, effectively generating missing spatial relationships. However, they are not applicable when the input pieces are rotated, as they assume a fixed orientation. Another line of research explores diffusion-based processes to iteratively refine the placement of puzzle pieces step by step [10, 26]. These approaches have demonstrated effectiveness in solving the oriented puzzle problem, where piece orientations are known. However, they are restricted to regular puzzle settings, where pieces conform to a structured grid. For real-world applications, square-based approaches are often impractical, as broken objects typically have irregular shapes, noisy visual content, and eroded or missing edges between pieces. A naive solution is to create a regular patch that encloses the broken object; however, this drastically hinders the effectiveness. + +# 2.2. Reassembly of Irregular Pieces + +To overcome the limitations of approaches restricted to square-based shapes, several recent studies have introduced methods designed explicitly for irregular-shaped object pieces. + +Before the adoption of deep learning, various techniques were developed to address this problem. For instance, [25] proposed an artifact reassembly method using inpainting, texture synthesis, and FFT-based image registration. Their approach aligns fragments by maximizing correlation through FFT shift theory, effectively handling damaged edges. Similarly, [22] uses a divide-and-conquer strategy, + +grouping pieces based on texture and color similarity. The method utilizes RGB and HIS color spaces for color matching and employs co-occurrence matrices to extract texture features. In [35], the authors introduced a computer-aided method consisting of four key steps. Their approach reconstructs fragmented images by identifying adjacent pieces, matching contours using a Smith-Waterman-based method with color similarity, refining alignment through Iterative Closest Point, and assembling the final image based on optimized alignment angles. More recently, [4] proposed a patch-based optimization method for artifact reconstruction, leveraging color, gradients, and geometric transformations to match fragments. + +In recent years, significant progress has been made in applying deep learning to handle irregular input shapes. The first work in this area is [30], which introduced DNN-Buddies, a deep neural network model for evaluating piece compatibility. In [20], the authors enhanced puzzle piece matching using CNNs and adaptive boosting. They proposed a multi-graph search algorithm to replace greedy strategies, enabling the handling of missing pieces and low-texture areas. [1] developed a classification network for evaluating the compatibility of fragment pairs and matching irregular puzzle pieces. However, their approach requires square-shaped pieces and cannot accommodate irregular outputs. PuzzleFusion [17] is based on a diffusion model and treats puzzle pieces as simple polygons with a limited number of vertices. This method is designed to work only on toy problems, as it processes all keypoints of the input polygons. As a result, it becomes impractical for real-world datasets, where a large number of keypoints lead to exponential memory consumption. Finally, PairingNet [40] is a learning-based image fragment pair-searching and -matching approach. Their method employs a graph-based network for feature extraction, a linear transformer-based module for fragment pair-searching, and a weighted fusion module for pair-matching. By formulating a similarity matrix to infer adjacent segments. + +Unlike previous works, which have several limitations for real-world applications, our solution is designed for practical scenarios and is capable of handling irregular inputs without scalability issues, thanks to the keypoint selection module. It also distinctively leverages multi-modal features by including both geometry and texture during training and inference. + +# 3. ReassembleNet Pipeline + +In this paper, we focus on reassembling fragmented 2D objects where the input of our method is a set of images representing fragments with irregular shapes. Following [18], our method takes as input $m$ images of unordered pieces. We extract keypoints from the images by applying Harris Corner Detector [13]. Points act as a helpful cue in the ar + +rangement of the pieces as subsets of the points on each piece can be aligned to recover the overall piece pose. + +For each of the $m$ pieces, let the set of keypoints be defined as $K^{m} = \{\mathbf{k}_{i}^{m}\}_{i = 1}^{d^{m}}$ , where $d^{m}$ is the number of keypoints in each piece $m$ . As this results in a significant number of points, direct comparison between all pieces is infeasible; therefore, our method applies an end-to-end learnable keypoint selector, which is trained to identify the most informative keypoints from the fragments. These keypoints are crucial for guiding the reassembly process. + +Next, we aim at extracting a multi-modal keypoint feature designed to capture both the geometric and textural properties of each fragment. This allows us to fully embed the shape and visual characteristics of the pieces, which are essential for proper alignment. Finally, we utilize a combination of attention blocks to model the relationships between the fragments. This solution helps the method better fit the pieces together by analyzing their spatial and contextual connections, ensuring that they are aligned in the correct configuration. The model is wrapped within a diffusion process to progressively learn the correct alignment in terms of translation $\mathbf{s}_i^m\in \mathbb{R}^2$ and rotation $\mathbf{r}_i^m = [\cos (\theta_i^m),\sin (\theta_i^m)]^\top$ [41]. + +# 3.1. KeyPoint Feature Representation + +Each keypoint $\mathbf{k}_i^m$ is represented by a feature vector $\mathbf{h}_i^m$ that combines several components. In challenging reassembly tasks, extracting meaningful characteristics is crucial for providing the network with valuable prior knowledge and improving its performance. Therefore, we focus on extracting multi-modal features from the input - specifically, geometric and texture attributes. + +Geometric Features. We focus on two specific geometric features: Curvature, which measures the curvature of the boundary at the keypoint. Curvature indicates whether a corner is a sharp bend or a more gentle curve, and it can be approximated by fitting a curve (e.g., a circle or spline) around the keypoint and estimating its curvature. It is analyzed by second-order derivatives and is represented as a single scalar value per keypoint. Edge Angles involve calculating the angle of the tangent to the edge at each boundary keypoint. It is a 1D feature (one angle per keypoint) and is measured in degrees. This angle helps distinguish between different types of boundary segments, such as straight edges versus curved ones. + +Texture Features. We have also focused on extracting features specifically related to the textures of the pieces. In particular, we extracted global texture features that summarize the general characteristics of the element, as well as local texture features that capture more detailed, localized information. For global texture features, we leveraged a pre-trained ResNet18 [15] to capture high-level texture representations across the entire element. For local texture + +![](images/cb17884107ba329a925774ca3aa4bc07f8422ca5dceb07090f9be369d79028b1.jpg) +Figure 2. Framework of our proposed ReassembleNet. We begin by extracting keypoints from the input pieces, followed by computing global and local texture features alongside geometric features. Using the geometric features and keypoint coordinates, we then select the most relevant $k$ keypoints. To model the reassembly process, we employ a Diffusion Probabilistic Model, formulating a Markov chain that gradually injects noise into the keypoints' positions and orientations. At timestep $t = 0$ , the pieces are correctly aligned, whereas at timestep $t = T$ , their keypoints are randomly translated and rotated (note that for visualization purpose, we compute the average translation and rotation of keypoints within each piece at every step in the chain). At each timestep $t$ , our attention module processes the keypoints—incorporating their coordinates, orientations, and extracted features—to predict a less noisy version of their positions and orientations, $\{\hat{X}_{t-1}^m\}_{m=1}^M$ , iteratively refining them toward the correct configuration. + +features, we adopted a similar approach using the same pretrained ResNet18. Specifically, we extract $32 \times 32$ patches centered on each selected keypoint. These localized patches are processed through ResNet18, where we remove the final classification layer and utilize the penultimate layer for feature extraction. This combination of global and local texture features allows our model to capture both the overall texture characteristics of the element as well as more detailed texture variations, which are crucial for accurately solving reassembly tasks. + +# 3.2. Keypoint Selector Module + +A key component of our pipeline is the selection of $k \in \mathbb{R}$ keypoints, which act as anchors to ensure scalability. To maintain consistency and simplicity, we select the same number of $k$ keypoints for all pieces (ablated on in Sec. 5). Identifying the most relevant keypoints is essential for handling irregular shapes with an arbitrary number of points. To achieve this, we explore two strategies: i) a nonlearnable pre-selection method and ii) a learnable approach. Non-Learnable Algorithm. For the non-learnable strategy, we employ the heuristic Farthest Point Sampling (FPS) [11] to perform an initial pre-selection of the $k$ -points. FPS is a well-established technique in the field of 3D point cloud processing and geometric sampling. It iteratively selects the point that is furthest from the current set, ensuring maximum dispersion throughout the entire data space. This ap + +proach guarantees that the selected keypoints are evenly distributed, capturing the underlying structure and geometry of the input data effectively. By leveraging FPS, our method can robustly cover the entire image or spatial domain with a fixed number of representative points. + +Learnable Algorithm. We frame the keypoint selection task as a graph sparsification problem [14]. In our formulation, the set of keypoints $K^m$ corresponds to the set of vertices $V^m$ of a complete graph $G^m$ (i.e. fully connected). Graph sparsification is then defined as the process of selecting a subset of nodes, i.e., keypoints or edges from $G^m$ to produce a sparser graph $\hat{G}^m$ . In other words, the elements of $\hat{G}^m$ are a subset of those in $G^m$ . + +To address this problem, we designed an architecture that learns to extract $k$ keypoints in a data-driven manner. As illustrated in Figure 3, the architecture comprises three main components: + +Projection Layer. The input features—comprising the coordinates and geometric attributes—are first projected into a higher-dimensional space. We rely solely on coordinates and geometric features, excluding texture information, as the pre-training is specifically designed to focus on the input shape (see Equation (2)). + +Graph Transformer. The projected features are then processed by a Graph Transformer [28] to aggregate and refine the input information. + +Pooling Layer. Finally, we apply a pooling layer [8], where + +we select $k$ nodes from the original graph. The selection of nodes to drop is guided by a projection score computed against a learnable vector $\mathbf{p}$ . To ensure that gradients propagate into $\mathbf{p}$ , these projection scores also serve as gating values, allowing nodes with lower scores to retain fewer features. Defined as: + +$$ +\begin{array}{l} \mathbf {y} = \frac {D \mathbf {p}}{\| \mathbf {p} \|} \quad \mathbf {j} = \operatorname {t o p - k} (\mathbf {y}, k) \\ \hat {D} = \left(D \odot \tanh (\mathbf {y})\right) _ {j} \quad \hat {A} = A _ {j, j} \tag {1} \\ \end{array} +$$ + +where $D$ represents the feature matrix of the input graph $G^{m}$ , $\hat{D}$ denotes the output feature matrix of the subset graph $\hat{G}^{m}$ , $\|\cdot\|$ denotes the L2 norm, top- $k$ selects the top $k$ indices from a given input vector, $\odot$ represents element-wise multiplication, and $j$ is an indexing operation that extracts slices at the indices specified by $\mathbf{j}$ . This operation involves only a point-wise projection and slicing of the original feature and adjacency matrices, thereby preserving sparsity. + +The algorithm benefits from a pre-training phase to enable the model to identify, based on a general criterion, which keypoints to retain and which to discard in an unsupervised manner. This is helpful because we have no prior knowledge of the most relevant keypoints and aim for a starting point that is not task-specific. Therefore, we train the model to select $k$ nodes while maximizing the preservation of both the perimeter and the area. The keypoints are chosen to best maintain the overall shape of the object. For this reason, we employ the following two losses: + +$$ +\mathcal {L} _ {\text {a r e a}} = \left(\frac {A _ {\text {t o t a l}} - A _ {\text {s e l}}}{A _ {\text {t o t a l}}}\right) ^ {2}; \mathcal {L} _ {\text {p e r}} = \left(\frac {P _ {\text {t o t a l}} - P _ {\text {s e l}}}{P _ {\text {t o t a l}}}\right) ^ {2}, \tag {2} +$$ + +where $A_{\mathrm{total}}$ , $A_{\mathrm{sel}}$ , $P_{\mathrm{total}}$ , and $P_{\mathrm{sel}}$ represent the area and perimeter measurements. Specifically, $A_{\mathrm{total}}$ and $P_{\mathrm{total}}$ correspond to the area and perimeter computed using all initial keypoints, while $A_{\mathrm{sel}}$ and $P_{\mathrm{sel}}$ refer to the area and perimeter computed using only the selected $k$ keypoints. The final loss function is defined as: + +$$ +\mathcal {L} = \lambda_ {\text {a r e a}} \mathcal {L} _ {\text {a r e a}} + \lambda_ {\text {p e r}} \mathcal {L} _ {\text {p e r}}, \tag {3} +$$ + +where $\lambda_{\mathrm{area}}$ and $\lambda_{\mathrm{per}}$ are the regularizing parameters. This architecture allows us to effectively select representative keypoints by leveraging both the geometric information and the relational structure captured by the graph. + +# 3.3. Estimating Position and Rotation with Diffusion + +We adopt Diffusion Probabilistic Models as defined in Denoising Diffusion Implicit Models (DDIM) [31]. For each piece $m$ , we apply an initial transformation, consisting of a translation $\mathbf{s}^m$ and a rotation $\mathbf{r}^m$ , that is replicated identical for all its keypoints. To represent this initial transformation compactly, we define a concatenated vector for each + +keypoint $\mathbf{x}_{i0}^{m} = [\mathbf{s}_{i0}^{m^{\top}},\mathbf{r}_{i0}^{m^{\top}}]^{\top}$ , where $\mathbf{s}_{i0}^{m}$ and $\mathbf{r}_{i0}^{m}$ denote the initial translation and rotation applied to all keypoints of the piece. Since this transformation is applied globally, all keypoints within the same piece share these initial transformation parameters. + +At training time, we iteratively add noise sampled from a Gaussian distribution $\mathcal{N}(0,I)$ to the poses of each keypoints (Forward Process). Following that, for the denoising process, we train ReassembleNet to reverse the noising process (Reverse Process) and predict the initial poses of all keypoints for every piece. We denote the initial poses of all the keypoints as $X_0 = \{X_0^m\}_{m=1}^M$ , where for each piece, $X_0^m = \{\mathbf{x}_{i0}^m\}_{i=1}^K$ , represents the set of initial keypoints poses. Additionally, we denote the predicted keypoints pose at timestep $t-1$ as $\hat{X}_{t-1} = \{\hat{X}_{t-1}^m\}_{m=1}^M$ . For the final result at time $t=0$ , we take the average polygon Euclidean center position and the polygon rotation. Further details on the diffusion process we use can be found in Supplementary Material C. + +The Architecture. To process the task, we aim to capture both intra-piece information and inter-piece (i.e. global information between pieces). The key idea is to handle these two types of information separately but in parallel. To achieve this, we employ a combination of attention layers. Intra-piece information is processed using a sparse self-attention block [2], which helps the network focus on obtaining consistent positions and rotations, across different keypoints within the same piece. While, inter-piece information is handled using a standard self-attention block [37] between pieces. + +Losses. Following [16] and standard practice in Diffusion Models, we train ReassembleNet to directly predict $\hat{X}_0$ rather than $\hat{X}_{t-1}$ . We employ two loss functions to reconstruct the initial pose of each piece's keypoints. + +Translation Loss. This loss measures the average discrepancy between the ground truth translation vectors and the predicted translations $\hat{\mathbf{s}}_{i0}^{m}$ : + +$$ +\mathcal {L} _ {t r} = \frac {1}{M} \frac {1}{K} \sum_ {m = 1} ^ {M} \sum_ {i = 1} ^ {K} \| \mathbf {s} ^ {m} - \hat {\mathbf {s}} _ {i 0} ^ {m} \| _ {2} ^ {2}, +$$ + +where $\| \cdot \|_2^2$ denotes the squared L2 norm. + +Rotation Loss. This loss quantifies the average discrepancy between the ground truth rotation and the predicted rotations $\hat{\mathbf{r}}_{i0}^{m}$ : + +$$ +\mathcal {L} _ {r t} = \frac {1}{M} \frac {1}{K} \sum_ {m = 1} ^ {M} \sum_ {i = 1} ^ {K} \| \mathbf {r} ^ {m} - \hat {\mathbf {r}} _ {i 0} ^ {m} \| _ {2} ^ {2}. +$$ + +# 3.4. Semi-Synthetic Dataset Creation and Fine-Tuning + +To increase training data while reducing the sym-to-real gap of the synthetic data, we introduce a large-scale + +![](images/982c2aa742a95c2d7d9baeed8e05be357e03a78000eeadc219ab66303f429391.jpg) +Figure 3. An illustration of the Learnable KeyPoint Selector Module where keypoints are projected into a high dimensional space, then use a graph transformer to predict scores which is then used to identify the top- $k$ and pooled to identify the most important keypoints. + +semi-synthetic dataset based on RePAIR [36]. We build from [40], where frescoes are divided into fragments with varied break patterns. The segmentation begins by selecting two points on the fresco's circumscribed circle. The segmentation line is then randomly divided into straight or curved segments, with curves created using Fourier bases for realistic edges. The generated pieces of this algorithm are not capable of emulating the complexity of fresco datasets. To overcome this limitation, we enhance the approach by incorporating two additional features: applying erosion operation on the piece and applying a slight random rotations and translations. Rather than introducing noise, these adjustments are essential for accurately modeling realistic fragmentation (see Section 5 and Supp. Mat. H for more details). + +Based on this new semi-synthetic dataset, we pre-train our model on it and fine-tune the method to a real-world dataset with fewer samples. This approach leverages the knowledge acquired from a large-scale dataset to reduce the need for extensive supervised training, thereby enhancing sample efficiency and accelerating convergence, while reducing overfitting and enhancing generalization. + +# 4. Evaluation + +We evaluate our method on the RePAIR benchmark dataset [36], following the same experimental protocol. The dataset evaluation procedures are detailed in Section 4.1, while the performance of all methods is assessed in Section 4.2. In Section 5, we present an ablation study on the scalability, the selection of the number of keypoints and the semi-synthetic dataset creation. Additional ablation studies on ReassembleNet's configuration options Supp. Mat. G and results on the semi-synthetic dataset Supp. Mat. H. + +# 4.1. Dataset, Metrics and Baseline Methods + +RePAIR Dataset. The dataset serves as a challenging benchmark for testing modern computational and data-driven puzzle-solving methods. It features realistic 2D and 3D fragments of frescos that, due to natural and human-made impacts undergone over time, exhibit erosion, missing pieces, and irregular shapes. The dataset is multi-modal, in + +cluding high-resolution images and archaeologist-annotated ground truth and metadata. The RePAIR 2D dataset consists of 121 puzzle samples, with 97 for training and 24 for testing. The total number of fragments in RePAIR is 957, with an average of 7.91 fragments per puzzle. + +Semi-Synthetic Dataset The semi-synthetic dataset has 5000 samples with a total of 45834 pieces, resulting in an average of around 9 pieces per puzzle. Following the Repair 2D train-test split (80:20 ratio), we divided the semi-synthetic dataset into 4000 training samples and 1000 testing samples. + +Evaluation Metrics. Following [10], we evaluated the methods using the Root Mean Square Error (RMSE) for both translation (in millimeters) and rotation (in degrees), computed with respect to the ground truth. Specifically, we predict the final pose of each keypoint and compute the average polygon translation and rotation as $\pmb{\mu}_{\hat{t}}^{m} = \frac{1}{K}\sum_{i=1}^{K}\hat{\mathbf{s}}_{i0}^{m}$ and $\pmb{\mu}_{\hat{r}}^{m} = \frac{1}{K}\sum_{i=1}^{K}\hat{\mathbf{r}}_{i0}^{m}$ , respectively. We also evaluated the performance of the methods using the $Q_{pos}$ metric [36], which quantifies the overlap between the ground truth fragment poses (translation and rotation) and the reconstructed solution. More details about these metrics can be found in the Supplementary Material E. + +Baseline Methods. We compare ReassembleNet against both non-learnable and learnable SOTA approaches. + +Non-learnable Approaches: $i$ ) the Archaeological Puzzle Solver [4], which applies a greedy "next best piece" algorithm based on texture. However, they use an outdated extrapolation process, which was replaced by [12] with the stable-diffusion extrapolation method. ii) The Genetic Algorithm reconstruction uses a fitness function based on geometry, specifically the area of the puzzle's bounding rectangle and the intersection area of overlapping pieces. While perfect solutions minimize both values, this does not guarantee an optimal solution. iii) Greedy geometric matching, which, starting from a random seed fragment, iteratively extends the fragment pose in a greedy fashion based on geometry, in contrast to [4], which relies on texture compatibility. + +Learnable Approaches: i) DiffAssemble [26] is a GNN-based architecture designed to tackle reassembly tasks using a diffusion model formulation. In this approach, pieces + +are treated as elements within a set, represented as nodes in a spatial graph. In the 2D scenario, these pieces are modeled as regular patches. ii) PairingNet [40] is a learning-based approach for fragment pair-searching and matching. It uses a graph-based network to extract features, integrates them via a linear transformer module, and employs contrastive loss for global encoding. A weighted fusion module then computes similarity scores to infer adjacent segments. + +# 4.2. RePAIR Dataset Evaluation + +Details. We compare ReassembleNet with both non-learnable and learnable methods. We train our model with Adafactor as the optimization algorithm [27] and initialize the learning rate with 0.001. We set a batch size of 4. Our method involves selecting $k$ keypoints (sec. 3.2, which is set to $k = 20$ . For the learnable keypoint selection module, we pre-train to enhance its performance. Further details on the pre-training process are provided in the Supp. Mat. F. We assess the performance of ReassembleNet in the three different configurations for keypoint selection: (i) a strategy based on the no-learnable keypoint selection, (ii) a frozen learnable module, and (iii) a trainable keypoint selection module optimized for the given task. + +Results. Table 1 presents the results on the RePAIR dataset. ReassembleNet with learnable keypoint selection and fine-tuning outperforms the other methods. This result emphasizes that representing irregular pieces as points, while selecting the best keypoints as done by ReassembleNet, is effective. It leads to improvements over the second best performing method, Greedy Geom Match [36], by $57\%$ and $87\%$ for RMSE rotation and translation, respectively. + +Regarding the keypoint selection comparison, as shown in the table, ReassembleNet with the learnable keypoint selector achieves the best performance. These results demonstrate the benefits of using a learnable module instead of a non-learnable one (i.e. Furthest Point Sampling), as it can adapt to extract the most significant keypoints. + +Additionally, when comparing the performance of the learnable models with and without fine-tuning, we observe a clear benefit from applying fine-tuning. All methods show improved performance in rotation and translation, while their performance in $Q_{pos}$ decreases. These results emphasize the effectiveness of our dataset construction, as it successfully preserves the patterns from the original dataset. + +Regarding the results of other methods, DiffAssemble approximates irregular shapes as squared pieces, making it challenging to accurately identify the correct matches. Furthermore, in the original 2D configuration reported in [26], the method relies on a regular grid to arrange the pieces in the output. Since no such grid is present in our case, this introduces additional limitations. PairingNet performs poorly because it employs a pair-matching loss to align contour points, which is unsuitable for puzzle pieces with erosion + +![](images/af41d4c0fcea768b492eb004dd5e59e950f9a89fab88844d84a1e868a1e4ee9c.jpg) +Figure 4. Qualitative results on RePAIR, showing the reassembly outcomes on four freecoes. + +or gaps between them. Non-learnable approaches, as also stated in [36], achieve low performance, reinforcing the advantages of learned strategies. + +PuzzleFusion does not solve the task as its architecture lacks mechanisms to reduce the number of keypoints, therefore, it is computationally intractable on complex geometry pieces. + +Figure 4 reports qualitative results of ReassembleNet on four example frescos. More quantitative are presented in Supplementary Material I. + +# 5. Ablation Study + +Scalability. In this experiment, we aim to demonstrate the scalability of ReassembleNet. To achieve this, we generate $n$ different datasets, where the number of pieces grows exponentially, i.e., $2^n$ . We set $n = 7$ and evaluate both models. + +![](images/dc962780b5412f000a7d40f342d711e86f191299349d1b73c6e84e0bb578d4ae.jpg) +Figure 5. GPU memory consumption as a function of the number of puzzle pieces on RePAIR dataset. + +In Figure 5, we present the memory consumption trends of ReassembleNet as the number of puzzle pieces increases. As shown, ReassembleNet successfully scales with larger puzzle sizes. This highlights the robustness and efficiency of our method in handling more complex puzzle configurations, making it suitable for larger-scale realistic problems. + +
CategoryMethodQpos↑RMSE (R°) ↓RMSE (Tmm) ↓
Non-learnableDerech et al. [4]0.0480.96139.49
Genetic Optimization [36]0.0585.63151.71
Greedy Geom Match [36]0.0276.99135.95
Learnable & No Fine-TuningDiffAssemble [26]0.10131.36283.55
PairingNet [40]0.1698.45390.43
ReassembleNet w/ no Learnable KP selection0.3555.0116.12
ReassembleNet w/ Frozen Learnable KP selection0.3951.9626.67
ReassembleNet w/ Learnable KP selection0.2747.6119.16
Learnable & Fine-TuningDiffAssemble [26]0.10123.42280.76
PairingNet [40]0.1391.54364.62
ReassembleNet w/ no Learnable KP selection0.1642.9818.11
ReassembleNet w/ Frozen Learnable KP selection0.1739.1218.41
ReassembleNet w/ Learnable KP selection0.2132.9117.18
+ +The Number of Keypoints. We evaluate the impact of the number of keypoints by conducting experiments with Re-assembleNet, using learnable keypoint selection on the RePAIR dataset without fine-tuning. + +Table 1. Results on RePAIR dataset [36] using $Q_{pos}$ from [36] for groundtruth overlap, Root Mean Square Error (RMSE), in terms of, rotation ( $\mathcal{R}^{\circ}$ ) and translation ( $T_{mm}$ ). Comparing against non-learnable optimization methods [4, 36] and learning (i.e., deep learning) methods [26, 40]. We do not include PuzzleFusion [18] because it runs out-of-memory due to highly irregular shape. + +
KeyPointsMemory (MB)Qpos↑RMSE (R°) ↓RMSE (Tmm) ↓
3226300.07103.2650.89
5247910.1086.7437.42
10284340.1478.9123.37
15342090.1767.6124.25
20367010.2747.6119.16
25[OOM][OOM][OOM][OOM]
+ +Table 2 reports the results of the keypoint number evaluation. It is clear that setting $k = 20$ yields the best performance while efficiently maximizing the available resources (40 GB of GPU memory). The reported VRAM reflects the entire pipeline, including the keypoint selector and two attention modules. + +Semi-Synthetic Dataset Creation Process. As discussed in Section 3.4, we modify the algorithm proposed by [40]. In this experiment, we compare our semi-synthetic dataset creation method with the original algorithm introduced by [40]. + +Table 2. Comparison of the parameter $k$ for the number of Keypoints selected by the Keypoint selector (sec. 3.2) + +
StrategyQpos↑RMSE (R°) ↓RMSE (Tmm) ↓
[40]0.1856.5426.42
Our0.2132.9117.18
+ +Table 3. Results on the impact of the Semi-Synthetic Dataset Creation process in contrast to [40]. + +Table 3 reports the results, showing that ReassembleNet trained on our semi-synthetic dataset outperforms Reassem- + +bleNet trained on the dataset generated using [40]. This validates the effectiveness of the modifications we introduced to the original algorithm. + +# 6. Conclusion + +In this work, we introduce ReassembleNet, a scalable deep learning approach for reassembly tasks. We propose the first 2D keypoint selector module designed to identify the most relevant keypoints that represent the pieces contour. Additionally, we integrate multimodal features, including both geometric and texture-based information, to better capture the characteristics of the pieces. We demonstrate that pre-training and fine-tuning on a semi-synthetic dataset, specifically created for this study, enhance ReassembleNet's ability to generalize to real-world fresco datasets, such as RePAIR. + +Our experiments show that ReassembleNet outperforms existing methods in the fresco domain and is suitable for real-world applications. Through an extensive ablation study, we highlight the benefits of using ReassembleNet in terms of memory consumption, as well as the importance of incorporating a learnable keypoint selector. This integration allows the network to adapt its selection based on the task and input dataset, improving overall performance. + +The limitations of ReassembleNet include the use of a fixed number of keypoints, whereas the selection should ideally vary depending on the complexity of the pieces, as well as the use of a non-rotation equivariant backbone for extracting texture features. + +Our ReassembleNet demonstrates a move towards usable reassembly for real-world problems, enabling the possibility for archaeologists and other application domains to integrate automatic assembly into workflows. + +# References + +[1] Yangjie Cao, Zheng Fang, Hui Tian, and Ronghan Wei. 2d irregular fragment reassembly with deep learning assistance. IEEE Access, 2024. 3 +[2] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 5 +[3] Gabriele Corso, Hannes Stärk, Bowen Jing, Regina Barzilay, and Tommi Jaakkola. Diffdock: Diffusion steps, twists, and turns for molecular docking. International Conference on Learning Representations (ICLR), 2023. 1 +[4] Niv Derech, Ayellet Tal, and Ilan Shimshoni. Solving archaeological puzzles. Pattern Recognition, 119:108065, 2021. 2, 3, 6, 8 +[5] Herbert Freeman and L Garder. Apictorial jigsaw puzzles: The computer solution of a problem in pattern recognition. IEEE Transactions on Electronic Computers, (2):118-127, 1964. 1 +[6] Thomas Funkhouser, Hijung Shin, Corey Toler-Franklin, Antonio Garcia Castañeda, Benedict Brown, David Dobkin, Szymon Rusinkiewicz, and Tim Weyrich. Learning how to match fresco fragments. Journal on Computing and Cultural Heritage (JOCCH), 4(2):1-13, 2011. 2 +[7] Andrew C Gallagher. Jigsaw puzzles with pieces of unknown orientation. In 2012 IEEE Conference on computer vision and pattern recognition, pages 382-389. IEEE, 2012. 2 +[8] Hongyang Gao and Shuiwang Ji. Graph u-nets. In international conference on machine learning, pages 2083-2092. PMLR, 2019. 4 +[9] Howard E Gardner. Frames of mind: The theory of multiple intelligences. Basic books, 2011. 1 +[10] Francesco Giuliani, Gianluca Scarpellini, Stuart James, Yiming Wang, and Alessio Del Bue. Positional diffusion: Ordering unordered sets with diffusion probabilistic models. arXiv preprint arXiv:2303.11120, 2023. 2, 6 +[11] Teofilo F. Gonzalez. Clustering to minimize the maximum intercluster distance. Theoretical Computer Science, 38: 293-306, 1985. 4 +[12] Peleg Harel, Ofir Itzhak Shahar, and Ohad Ben-Shahar. Pictorial and apictorial polygonal jigsaw puzzles from arbitrary number of crossing cuts. International Journal of Computer Vision, 2024. 6 +[13] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, pages 10-5244. CiteSeer, 1988. 3 +[14] Mohammad Hashemi, Shengbo Gong, Juntong Ni, Wenqi Fan, B. Aditya Prakash, and Wei Jin. A comprehensive survey on graph reduction: Sparsification, coarsening, and condensation. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, pages 8058–8066. International Joint Conferences on Artificial Intelligence Organization, 2024. Survey Track. 4 +[15] 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 + +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 5, 1 +[17] Sepidehsadat Hosseini, Mohammad Amin Shabani, Saghar Irandoust, and Yasutaka Furukawa. Puzzle fusion: Unleashing the power of diffusion models for spatial puzzle solving. 3 +[18] Sepidehsadat Hosseini, Mohammad Amin Shabani, Saghar Irandoust, and Yasutaka Furukawa. Puzzle fusion: Unleashing the power of diffusion models for spatial puzzle solving, 2023. 2, 3, 8 +[19] Sou-Young Jin, Suwon Lee, Nur Aziza Azis, and Ho-Jin Choi. Jigsaw puzzle image retrieval via pairwise compatibility measurement. In 2014 International Conference on Big Data and Smart Computing (BIGCOMP), pages 123–127. IEEE, 2014. 2 +[20] Canyu Le and Xin Li. Jigsawnet: Shredded image reassembly using convolutional neural network and loop-based composition. IEEE Transactions on Image Processing, 28(8): 4000-4015, 2019. 3 +[21] William Marande and Gertraud Burger. Mitochondrial dna as a genomic jigsaw puzzle. Science, 318(5849):415-415, 2007. 1 +[22] Ture R Nielsen, Peter Drewsen, and Klaus Hansen. Solving jigsaw puzzles using image features. Pattern Recognition Letters, 29(14):1924-1933, 2008. 2 +[23] Genady Paikin and Ayellet Tal. Solving multiple square jigsaw puzzles with missing pieces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4832-4839, 2015. 2 +[24] Dolev Pomeranz, Michal Shemesh, and Ohad Ben-Shahar. A fully automated greedy square jigsaw puzzle solver. In CVPR 2011, pages 9-16. IEEE, 2011. 2 +[25] Mahmut Samil Sagiroglu and Aytül Erçil. A texture based matching approach for automated assembly of puzzles. In 18th International Conference on Pattern Recognition (ICPR'06), pages 1036-1041. IEEE, 2006. 2 +[26] Gianluca Scarpellini, Stefano Fiorini, Francesco Giuliani, Pietro Moreiro, and Alessio Del Bue. Diffassemble: A unified graph-diffusion model for 2d and 3d reassembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28098-28108, 2024. 2, 6, 7, 8, 1 +[27] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In International Conference on Machine Learning, pages 4596-4604. PMLR, 2018. 7 +[28] Yunsheng Shi, Zhengjie Huang, Shikun Feng, Hui Zhong, Wenjin Wang, and Yu Sun. Masked label prediction: Unified message passing model for semi-supervised classification. arXiv preprint arXiv:2009.03509, 2020. 4 +[29] Dror Sholomon, Omid David, and Nathan Netanyahu. A generalized genetic algorithm-based solver for very large jigsaw puzzles of complex types. In Proceedings of the AAAI Conference on Artificial Intelligence, 2014. 2 +[30] Dror Sholomon, Omid E David, and Nathan S Netanyahu. Dnn-buddies: A deep neural network-based estimation met + +ric for the jigsaw puzzle problem. 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 170-178. Springer, 2016. 3 +[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5, 1 +[32] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 2 +[33] Davide Talon, Alessio Del Bue, and Stuart James. Ganzzle: Reframing jigsaw puzzle solving as a retrieval task using a generative mental image. In 2022 IEEE International Conference on Image Processing (ICIP), pages 4083-4087. IEEE, 2022. 2 +[34] Davide Talon, Alessio Del Bue, and Stuart James. Ganzzle++: Generative approaches for jigsaw puzzle solving as local to global assignment in latent spatial representations. Pattern Recognition Letters, 187:35-41, 2025. 2 +[35] Efthymia Tsamoura and Ioannis Pitas. Automatic color based reassembly of fragmented images and paintings. IEEE Transactions on Image Processing, 19(3):680-690, 2009. 2, 3 +[36] Theodore Tsesmelis, Luca Palmieri, Marina Khoroshiltseva, Adeela Islam, Gur Elkin, Ofir I Shahar, Gianluca Scarpellini, Stefano Fiorini, Yaniv Ohayon, Nadav Alali, et al. Reassembling the past: The repair dataset and benchmark for real world 2d and 3d puzzle solving. Advances in Neural Information Processing Systems, 37:30076-30105, 2025. 1, 6, 7, 8 +[37] 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 +[38] Ariana M Villegas-Suarez, Cristian Lopez, and Ivan Sipiran. Matchmakernet: Enabling fragment matching for cultural heritage analysis. Preprint. Accessed: No DOI available at the time of retrieval. 1 +[39] E Yurteri, Sabri Derin, Fatma Polat, Murat Canpolat, Zekiye Barman, C Polat, Emel Kaya, and Selim Güllü. Obstacle maps. New Era Journal, 1(1):1-15, 2022. 1 +[40] Rixin Zhou, Ding Xia, Yi Zhang, Honglin Pang, Xi Yang, and Chuntao Li. Pairingnet: A learning-based pair-searching and-matching network for image fragments. In European Conference on Computer Vision, pages 234-251. Springer, 2024. 3, 6, 7, 8, 1, 2 +[41] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. 3 \ No newline at end of file diff --git a/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/images.zip b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..45e855598431a3ef3cbe9562c7a54d4310eada50 --- /dev/null +++ b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b4947ce830ce17b3fa16d272444984477a5fba000b44a400398b13de035d71e +size 331688 diff --git a/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/layout.json b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..989c83afaaf3105e3a3e8febdf76200255e4920d --- /dev/null +++ b/ICCV/2025/ReassembleNet_ Learnable Keypoints and Diffusion for 2D Fresco Reconstruction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb25b88a7ba7da274c4bcf88815f7ac3174042469e19303b36a49e2335ba081d +size 346382 diff --git a/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/c5c90ef8-4e0a-470c-b38e-81e675e75ccc_content_list.json b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/c5c90ef8-4e0a-470c-b38e-81e675e75ccc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4c1f7eeba0c30472114e3c234801ef905fe10f12 --- /dev/null +++ b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/c5c90ef8-4e0a-470c-b38e-81e675e75ccc_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99ee8b8a066f78be95fcbd07b2d4220d19bd674f26cbbad5883528ac6a68624a +size 81114 diff --git a/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/c5c90ef8-4e0a-470c-b38e-81e675e75ccc_model.json b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/c5c90ef8-4e0a-470c-b38e-81e675e75ccc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..33a91bc3a6e8115a0f8ebea3bd1699213b9c01d4 --- /dev/null +++ b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/c5c90ef8-4e0a-470c-b38e-81e675e75ccc_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc62f1fb7e0b0e80da73016a4f3f6739ed593d7d68804a69335b2213795068ea +size 94577 diff --git a/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/c5c90ef8-4e0a-470c-b38e-81e675e75ccc_origin.pdf b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/c5c90ef8-4e0a-470c-b38e-81e675e75ccc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ecb3d77b36915937c0bcd09cea6ddab7bd2a5793 --- /dev/null +++ b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/c5c90ef8-4e0a-470c-b38e-81e675e75ccc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e859e224c699a6c8b7a1e7abe65d16340e3c8d5caab99a5bc13e4cbabacda18 +size 7315511 diff --git a/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/full.md b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cc4afb6a1df8f39d733929422437d62e3a56ba93 --- /dev/null +++ b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/full.md @@ -0,0 +1,327 @@ +# Recognizing Actions from Robotic View for Natural Human-Robot Interaction + +Ziyi Wang1 Peiming Li1 Hong Liu1 Zhichao Deng2 Can Wang3 Jun Liu4 Junsong Yuan5 Mengyuan Liu1† + +$^{1}$ State Key Laboratory of General Artificial Intelligence, Peking University, Shenzhen Graduate School + $^{2}$ Sun Yat-sen University $^{3}$ Kiel University $^{4}$ Lancaster University $^{5}$ State University of New York at Buffalo +{ziyiwang, lipeiming1001}@stu.pku.edu.cn hongliu@pku.edu.cn dengzhch3@mail2.sysu.edu.cn canwang@pku.edu.cn j.liu81@lancaster.ac.uk jsyuan@buffalo.edu liumengyuan@pku.edu.cn + +# Abstract + +Natural Human-Robot Interaction (N-HRI) requires robots to recognize human actions at varying distances and states, regardless of whether the robot itself is in motion or stationary. This setup is more flexible and practical than conventional human action recognition tasks. However, existing benchmarks designed for traditional action recognition fail to address the unique complexities in N-HRI due to limited data, modalities, task categories, and diversity of subjects and environments. To address these challenges, we introduce ACTIVE (Action from Robotic View), a large-scale dataset tailored specifically for perception-centric robotic views prevalent in mobile service robots. ACTIVE comprises 30 composite action categories, 80 participants, and 46,868 annotated video instances, covering both RGB and point cloud modalities. Participants performed various human actions in diverse environments at distances ranging from $3\mathrm{m}$ to $50\mathrm{m}$ , while the camera platform was also mobile, simulating real-world scenarios of robot perception with varying camera heights due to uneven ground. This comprehensive and challenging benchmark aims to advance action and attribute recognition research in N-HRI. Furthermore, we propose ACTIVE-PC, a method that accurately perceives human actions at long distances using Multilevel Neighborhood Sampling, Layered Recognizers, Elastic Ellipse Query, and precise decoupling of kinematic interference from human actions. Experimental results demonstrate the effectiveness of ACTIVE-PC. Our code is available at: https://github.com/wangzy01/ ACTIVE-Action-from-Robotic-View. + +# 1. Introduction + +A robust Natural Human-Robot Interaction (N-HRI) system can provide a more personalized and intelligent interaction + +![](images/cf3689af00b8961685bd82d9769e3b673bb9614b2b7a491260f5b81ab6897e9e.jpg) +Figure 1. Data Collection Process and Samples of the ACTIVE Dataset: The movable camera and LiDAR capture human motion at distances ranging from 3 to 50 meters, with scene shaking introduced during the recording process. For visualization purposes, the samples have been scaled and cropped. + +experience by adapting to changes in user needs and evolving action patterns. It has a wide range of applications, such as in service robots [19], public safety surveillance [10], and medical diagnostics [3, 32]. N-HRI requires robots to have the ability to adapt in real-time, not only understanding human actions but also considering changes in human position (e.g., walking, standing) and their own motion. Therefore, compared to traditional human action recognition, human action recognition in N-HRI involves handling more complex human-robot interaction scenarios. + +Compared to traditional human action recognition videos, videos captured from perception-centric robotic views in N-HRI scenarios exhibit diverse viewpoints, greater kinematic disturbances, and varied subject resolutions due to changing human-robot positions, robot movements, and uneven ground. While different from manipulator-mounted or end-effector perspectives common in industrial robots, this perception-centric viewpoint aligns closely with real-world applications, such as mobile assistants and patrol robots sensing humans at distances of 3–50 meters using forward-facing sensors. These challenges necessitate specialized methods tailored specifically to the + +Table 1. A comparison of the ACTIVE dataset, designed for N-HRI, with other widely used general action and micro-action datasets. ACTIVE is the first large-scale action recognition dataset specifically focused on N-HRI. + +
General ActionFine-Grained ActionActions from Robotic View
Dataset AttributeMSR Action3DMultiview 3D EventNTU RGB+D 120UAV HumanPA HMDB51HOMAGEiMiGUESMGACTIVE (Ours)
Action Recog.# Videos5673,815114,48067,4285151,75218,4993,71246,868
# Subjects2010106119-27724080
# Classes10101201555175321630
Attribute Recog.# Samples×××22,476515×××46,868
Data ModalityDRGB+DRGB+D+IRRGB+D+IRRGBRGB+IRRGBRGB+DRGB+Point Cloud
SensorsKinectKinectKinect V2Azure DKNARGB Cam.RGB Cam.Kinect V2LiDAR, RGB Cam.
Capturing Scenarios# Sites1NA345NANANA16
Outdoor××✓✓××××✓
Shooting TimeNight×××✓××××✓
Composite Action×××××✓✓✓✓
Dynamic Distance××××××××✓
Platform Motion×××✓××××✓
SourceCVPRICCVTPAMICVPRTPAMICVPRCVPRIJCV-
Year20102013201920212020202120212023-
+ +unique characteristics of robotic views in N-HRI. + +Existing work has already demonstrated the importance of using large, comprehensive, and challenging datasets to develop and evaluate state-of-the-art deep learning methods. However, in the field of human action understanding for N-HRI scenarios, current datasets [13, 21] have limitations in several aspects, including: 1) Limited Sample Size: Large-scale datasets are often critical for mitigating overfitting and enhancing the generalization ability of models developed on them. 2) Fixed Capture Modes: Most existing datasets rely on fixed relative positions and backgrounds, with subjects typically standing still and performing predefined actions or postures. In contrast, humans in N-HRI are usually mobile. 3) Constrained Capture Scenes: In real-world applications, N-HRI needs to function in diverse environments (e.g., indoor, outdoor) and across various time periods (e.g., day, night). However, samples in current datasets are often collected under similar conditions, which oversimplifies the challenges in real-world N-HRI scenarios. 4) Limited Capture Distances: N-HRI often involves frequent changes in the human-robot position, leading not only to significant motion blur but also to considerable resolution changes. Yet, in most existing datasets, humans only exhibit slow and subtle movements, with fixed capture distances. 5) Limited Data Modalities: In real-world applications, different types of sensors are often deployed to collect data under varying conditions. For instance, LiDAR sensors can be used for long-distance N-HRI understanding, while RGB cameras are typically used for close-range scenes that require rich detail. This highlights the importance of collecting multimodal data to analyze human action under different conditions. However, most existing human action datasets only provide traditional RGB video samples or depth modalities. Devices like Kinect suffer from limited capture range and poor performance under daylight conditions. These limitations emphasize the need for a more robust and diverse dataset that better reflects the dynamic and real-world complexities of N-HRI scenarios. + +The aforementioned limitations of existing datasets + +clearly highlight the need for a larger, more challenging, and comprehensive dataset for human action analysis for N-HRI. To this end, we introduce ACTIVE, a large-scale multimodal dataset in this field. Specifically, ACTIVE includes 30 action categories with disturbing motion (e.g., Heading + Nodding), involving 80 participants and attribute information such as gender and clothing. It consists of a total of 46,868 video instances, captured using sensors including LiDAR and RGB cameras. During data capture, participants perform various human actions in different environments (indoor and outdoor, day and night) at varying capture distances (ranging from $3\mathrm{m}$ to a long-range $50\mathrm{m}$ ). Additionally, we simulate the robot's motion by introducing movement to the camera platform. The use of different sensors allows our dataset to provide rich data modalities, including both RGB and point cloud data. Overall, ACTIVE serves as a more challenging dataset, encompassing action recognition video instances in N-HRI scenarios and introducing diverse kinematic disturbances, which can promote research on human action understanding in N-HRI settings. Tab. 1 presents a statistical comparison between ACTIVE and existing datasets. More details about ACTIVE will be discussed in Sec. 3. + +As shown in Fig. 1, unlike general actions, ACTIVE incorporates variations in human-robot states (e.g., changes in relative distance and position through disturbing motion like walking and camera platform movement). Recognizing actions from robotic view aims to understand human actions in N-HRI scenarios, with the goal of filtering out the influence of kinematic disturbances from the robot's perspective and recognizing human actions to guide N-HRI. The main challenges of ACTIVE are as follows: 1) At long distances (e.g., 50 meters), human actions are subtle and rapid, making it even more difficult to capture the fine details of these actions. 2) In dynamic interaction scenarios, large-scale kinematic disturbances (such as the motion of the robot platform, environmental changes like lighting and obstacles, and the relative motion between the human and the robot) further complicate action recognition. + +Recognizing human actions from a robotic view presents several challenges, particularly when dealing with long distances and the interference of robotic motion. Traditional methods often fail to capture local details across varying scales and cannot effectively differentiate subtle human actions from kinematic disturbances caused by robot platform movements. To address these issues, we propose ACTIVE-PC, a comprehensive framework that utilizes Multilevel Neighborhood Sampling (MNS), Layered Recognizers (LR), and Elastic Ellipse Query (EEQ). MNS ensures both global coverage and retention of critical local details, while LR decouple features at different layers to preserve both fine-grained human actions and global kinematic variations. EEQ introduces axis-specific adaptive scaling to handle varying interaction distances, enhancing action differentiation by adjusting to the spatial distribution of motion. Experimental results demonstrate that this method improves action recognition performance in N-HRI scenarios. + +# 2. Related Work + +# 2.1. Action Recognition Datasets + +Early 3D action recognition datasets, such as MSR-Action3D [13], Multiview 3D Event [31], and NTU RGB+D [21], have played a crucial role in advancing action recognition technology. The MSR-Action3D dataset [13] primarily consists of sequences of gaming actions and was one of the first to incorporate depth information for action analysis. Many methods have been evaluated on this dataset, but due to the limited dataset size, recent approaches have achieved near saturation in accuracy. To address the limited sample size, the Multiview 3D Event dataset [31] employs multiple depth sensors to capture multi-view representations of the same action. The NTU RGB+D dataset [21], the first large-scale RGB+D action recognition dataset, features a broader range of samples, more intra-class variations, and more camera views, capturing RGB video, depth sequences, and skeleton data. In recent years, research has shifted towards fine-grained action recognition tasks. The iMiGUE dataset [15] aims to interpret athletes' emotional states by analyzing their micro-gestures, recording spontaneous micro-actions after a match. Similarly, the SMG dataset [4] focuses on both micro-gesture and emotion recognition, documenting participants' micro-actions while narrating stories. However, action recognition for N-HRI involves more complex and diverse interaction states. Existing datasets typically rely on fixed relative positions and backgrounds, which are not ideal for action recognition in real-world human-machine interactions. In contrast, ACTIVE is the first large-scale action recognition dataset focused on N-HRI, capturing a variety of actions with disturbing motions in complex interaction scenarios. + +# 2.2. Action Recognition Methods + +RGB-based methods, such as TSN [27], VideoMAE [24], UniFormer [11], and InternVideo [29], have achieved high accuracy in general action recognition. Similarly, various methods have been developed for general action recognition from point cloud videos [16, 22, 34]. To enhance the receptive field, P4Transformer [6] employs transformers to avoid explicit point tracking, although it does not fully encode the spatiotemporal structure. PST-Transformer [9] improves upon this by encoding features based on the spatiotemporal displacement between reference points and all points in the video. However, existing point cloud-based methods are typically limited to simple, single-action recognition and perform poorly on actions with disturbing motions. Our approach focuses on recognizing actions with disturbing motions in N-HRI, while remaining compatible with general action recognition. + +Currently, there is limited research specifically addressing micro-action recognition from point cloud videos. In RGB-based methods, Yonetani et al. [33] extracted handcrafted features from both first-person and third-person videos for micro-action recognition. However, this approach requires both first-person and third-person videos, along with manual synchronization. Mi et al. [20] proposed a segment-level temporal pyramid for micro-action videos. Nonetheless, current micro-action recognition methods typically rely on fixed relative positions and backgrounds. In contrast, our approach adapts to different complex interaction scenarios, making it more suitable for action recognition in real-world human-machine interactions. + +# 3. Dataset + +# 3.1. The Action from Robotic View Dataset + +ACTIVE is the first large-scale human behavior understanding dataset specifically designed for N-HRI scenarios (shown in Fig. 2), featuring data from 80 participants along with annotated attribute information. It covers six distinct environments (including both indoor and outdoor settings, under varying daylight and nighttime conditions) and includes 30 action (e.g., Walking forward + Waving) categories. The dataset comprises 46,868 video instances, with an average duration of 3.2 seconds, totaling 41.7 hours. It includes two modalities: RGB and point cloud data. The point cloud data is captured using the HESAI AT128P Li-DAR, with a horizontal field of view of $120^{\circ}$ , a horizontal angular resolution of $0.1^{\circ}$ , a vertical field of view of $25.4^{\circ}$ , and a vertical resolution of $0.2^{\circ}$ . The scanning frame rate is set at $10\mathrm{Hz}$ . The RGB video has a resolution of $1280 \times 720$ and a frame rate of 30 fps. Tab. 1 compares ACTIVE dataset with existing human action recognition datasets. Below are some key features of ACTIVE. + +![](images/18343be76cfa1524b5f80028ea2c94038394685c94678d88906788f7ab4c914d.jpg) +(a) + +![](images/46763d87d9359d6d1a2f3950172ea3386cebb43e097f3612ca5cb10dae0cd4cc.jpg) + +![](images/cbb4400bfe50e8207b7d5b41a1e47f10cc77a9b29ef11916ef06c0ec5bc395ba.jpg) + +![](images/bdbd795740220750dc76eb10aa716077d066a457ed988df1e5fd71960ac882a3.jpg) + +![](images/5e153bc6c3c29c49a9de830b92bb42a7bd29676d49768fadb0bfbd65e6713e1e.jpg) + +![](images/a8897b5edfa47a458ecdb7aec374a92da5d0fd39ec3df5789dad5325d5e6e4e8.jpg) + +![](images/09aaf2f0b85ca0aa0762705027e9248ae94296acc43e9bcf25559f705f3aed46.jpg) + +![](images/c89d2bbab3b97e7390457620d0a2171303990b00f43b7aa0dc977b951c8b5e5b.jpg) + +![](images/e8fdc422a1dca34ff125d1948ad81bb77b9c849a198f283c73160027da558212.jpg) + +![](images/d2cdff20ea0f5fd2d61d462b0cd8cd0a98624d3e5a8df843a287324de08a68da.jpg) +(b) +Raising Arms + +![](images/a22a2fbe06677841deed73fe9c73222ef6b8cb653f21f6776f3f8d8bb43f5348.jpg) + +![](images/af9f29ee17cbfe3861a52483f2f5c5301aa4d16351333bda94b9b836f62b536b.jpg) + +![](images/79dfcd36f28a96788862b8b6af94ad5e44ef6d9f590815e1a5ce7bb745fce5e6.jpg) + +![](images/582912c7d1b3f40e8619ca91240f32e4bf2a21ebbc872b666070f68d7b7ee148.jpg) +Pointing Up +Figure 2. Details of ACTIVE: (a) Action Labels: Three categories are included: Base Motions, Oriented Actions, and Social Gestures. (b) The first row illustrates the significant variations in the captured results of the same action at different human-robot distances. The second row presents the raw point cloud captured by LiDAR and its corresponding RGB image. The point cloud modality retains more detailed geometric information, enabling long-range N-HRI understanding with strong robustness to interference. In contrast, the RGB modality complements this by providing rich texture and semantic information at close distances. The final row showcases additional action samples from our dataset. (c) Shooting Scenes: Six shooting scenes used in ACTIVE, covering both indoor and outdoor settings, as well as day and night conditions. + +![](images/b32d9913da65bf0ca1be8927583b72b3598c2cb2039bdf26e57e49b48fcd9915.jpg) +Clapping + +![](images/703b8bdfd4849383c743398a8c00e2a8adcf3b60162a75d42f8d05260013ffcd.jpg) +Grasping + +![](images/2b3ec17029b6ffc18f0b610fa406559d00b545fa6acb437fb5154c39e964e4ca.jpg) +Rightward + +![](images/c8ae21757570163f68cf69ece26c1ece8830163734c26bccb80794f7bb88c957.jpg) +Looking Left + +![](images/d280526f4e9ccfa64e2196f9a8ea02edec7cb83f5ea49f642dc3f6cead11f94a.jpg) +Thumbs Up + +![](images/c31e4dea74a96bca2d3cc5f45a9b5e38802147c32b5fb89b21ebc2fb9d0df026.jpg) + +![](images/bd58214cf7e14ab11dedcfef5528297979710baa3cd94701f806b9e494286342.jpg) + +![](images/a28d30e259ace34f745b3ef654518e0c11a04eeb3f429b96ccc3d8fcc99600a8.jpg) + +![](images/991f5a98a3c441af9055eb65f6dff6a1341d8d68ffc79b797e3f3cd2291abf84.jpg) + +![](images/d699fddc95a4a8e50d7bb6d961634ab925faf1832a5d3aa8fa9e7c24ed3d1b4f.jpg) +(c) + +![](images/9ad06fb3bee0f56914adaf471b6541f11ecb1118fcb828917d16042f5fb2619b.jpg) + +Diversity of Capture States and Behavior Patterns. Traditional human behavior datasets typically focus on fixed capture modes and distances, often emphasizing a single behavior. In contrast, human behavior understanding in N-HRI requires handling more complex human-robot interaction scenarios. Therefore, ACTIVE collects data from subjects performing various actions under different human-robot states. In different settings, participants are instructed to perform multiple natural actions at varying distances. The linear distance between participants and the data capture platform ranges from 3 meters to 50 meters, with additional movement and vibrations to simulate changes in the robot's state. Minor disturbances from vehicles and pedestrians are also allowed during recording. + +More Relevant Data Modalities. Most existing human behavior datasets only provide traditional RGB video samples or depth modalities. While depth data can be converted into point clouds, it often lacks detailed geometric struc + +ture. Additionally, devices like Kinect suffer from short capture distances and poor performance under daylight conditions. In contrast, point cloud data captured by LiDAR retains richer geometric details, supports long-range N-HRI understanding, and is more robust to interference. The RGB modality complements this by offering rich texture and semantic information at close range. This underscores the importance of collecting diverse data modalities. Consequently, our varied data sources are designed to foster the development of robust models for human behavior analysis in complex N-HRI scenarios. + +Multiple Human Behavior Understanding Tasks. To comprehensively analyze human behavior and actions in N-HRI, ACTIVE currently provides annotations for two primary tasks: human action recognition and human attribute recognition. Detailed annotations for the pose estimation task will be released at a later stage. + +# 3.2. Benchmark Evaluations + +Action Recognition. To ensure consistency in benchmarking, we define the Cross-Subject Evaluation protocol and use classification accuracy (percentage) as the evaluation metric. A total of 80 participants were divided into a training group and a testing group, consisting of 53 and 27 subjects, respectively. + +Attribute Recognition. For attribute recognition, we use the same dataset partitioning strategy as for action recognition. The performance of attribute recognition is evaluated by measuring the classification accuracy for each attribute. + +# 4. Methodologies + +The ACTIVE dataset captures natural human-robot interactions from robotic view, encompassing both large-scale global motion disturbances and subtle local postural changes in the human body. In this work, we primarily aim to address two key challenges posed by the ACTIVE dataset: (1) At long distances, human actions are subtle and rapid, making it a complex task to recognize these fine movements. (2) In dynamic interaction scenarios, large-scale global motion disturbances further increase the difficulty of action recognition. To address these challenges in ACTIVE, we introduce ACTIVE-PC. In Sec. 4.1, we define the RARV task. In Sec. 4.2, we present the pipeline of ACTIVE-PC. In Sec. 4.3 and Sec. 4.4, we propose Multilevel Neighborhood Sampling (MNS) and Layered Recognizers (LR) to address challenge (1) and (2), and Elastic Ellipse Query to tackle challenge (2). + +# 4.1. Task Definition of RARV + +Consider an input video, $\mathbf{X} = \{I_1, I_2, \dots, I_T\}$ , where $T$ is the length of the video. Each frame contains $N$ points, and the video includes both human actions (e.g., raising arms) and kinematic disturbances (e.g., human-robot proximity and robot movement). Our goal is to filter out the influence of kinematic disturbances from the robot's perspective, in order to obtain the human action label $Y$ . + +# 4.2. Pipeline Overview + +The pipeline of ACTIVE-PC is shown in Fig. 3. The input video $\mathbf{X} \in \mathbb{R}^{T \times N \times 3}$ first undergoes MNS, producing point clouds with different densities: $\mathbf{X}_1 \in \mathbb{R}^{T/2 \times N/4 \times 3}$ , $\mathbf{X}_2 \in \mathbb{R}^{T/2 \times N/16 \times 3}$ and $\mathbf{X}_3 \in \mathbb{R}^{T/2 \times N/32 \times 3}$ . These point clouds, $\mathbf{X}_1$ , $\mathbf{X}_2$ and $\mathbf{X}_3$ are then processed through EEQ for adaptive neighbor querying, enabling spatiotemporal tracking and differentiation of human actions and kinematic disturbances, resulting in feature maps $\mathbf{F}_1$ , $\mathbf{F}_2$ and $\mathbf{F}_3$ . + +Low-density point clouds, which capture deeper features $\mathbf{F}_2$ and $\mathbf{F}_3$ , are passed through the Kinematic Interpreter for human-robot state understanding. High-density point clouds, containing finer local details in $\mathbf{F}_1$ , are input to the + +Action Recognizer for human-action recognition. $\mathbf{Y}_{\text {human }}$ and $\mathbf{Y}_{\text {kinematic }}$ are fused to obtain the fine-grained action score $\mathbf{Y}$ . The process is formulated as follows: + +$$ +\begin{array}{l} \mathbf {X} _ {1}, \mathbf {X} _ {2}, \mathbf {X} _ {3} = \operatorname {M N S} (\mathbf {X}), \\ \mathbf {F} _ {1}, \mathbf {F} _ {2}, \mathbf {F} _ {3} = \operatorname {P o i n t T u b e} _ {\mathbf {E E Q}} \left(\mathbf {X} _ {1}, \mathbf {X} _ {2}, \mathbf {X} _ {3}\right), \\ \mathbf {Y} _ {\text {k i n e m a t i c}} = \operatorname {I n t e r p r e t e r} _ {\text {k i n e m a t i c}} \left(\mathbf {F} _ {2}, \mathbf {F} _ {3}\right), \tag {1} \\ \mathbf {Y} _ {h u m a n} = \operatorname {R e c o n g i z e r} _ {\text {h u m a n}} (\mathbf {F} _ {1}), \\ \mathbf {Y} = \frac {\left(\mathbf {Y} _ {h u m a n} + \mathbf {Y} _ {k i n e m a t i c}\right)}{2}. \\ \end{array} +$$ + +# 4.3. Multilevel Neighborhood Sampling + +To address the core challenges of the ACTIVE dataset, which include the precise perception of subtle human motions at long distances and the accurate decoupling of kinematic interference from human actions in dynamic environments, Multilevel Neighborhood Sampling (MNS, Sec. 4.3) and Layered Recognizers (LR, Sec. 4.4) work in tandem. + +Traditional methods typically apply multi-scale or different sampling rates of farthest point sampling (FPS) directly to the entire point cloud. While this approach ensures globally uniform sampling, it often neglects the correlation between local details across different scales, which is critical in the ACTIVE dataset. In situations where the humanrobot distance is large, high-density local sampling can better capture small yet distinguishable motion features. + +To address this, we propose Multilevel Neighborhood Sampling (MNS), which builds hierarchical dependencies using local neighborhood constraints, ensuring both global coverage and retention of local details. The implementation of MNS is shown in Fig. 3. The black points represent the anchor points obtained through downsampling, while the red points represent the neighboring points. The process is as follows: First, for the input point cloud video data $\mathbf{X} \in \mathbb{R}^{T \times N \times 3}$ , we apply FPS to sample $N_{1}$ points globally, obtaining the first layer of anchor points: + +$$ +\mathbf {X} _ {1} = \operatorname {F P S} (\mathbf {X}; N _ {1}). \tag {2} +$$ + +For the $l$ -th layer, instead of directly sampling globally, we construct a candidate set within the local neighborhood of the previous layer's sampled points $\mathbf{x}_{l-1,j}$ (where $j = 1, \ldots, N_{l-1}$ ). Using a radius $r$ and a candidate number $k$ , we define the spherical neighborhood as: + +$$ +\mathcal {N} \left(\mathbf {x} _ {l - 1, j}\right) = \mathbf {x} _ {i}: | \mathbf {x} _ {i} - \mathbf {x} _ {l - 1, j} | _ {2} \leq r ^ {(l)}, | \mathcal {N} | \leq k ^ {(l)}. \tag {3} +$$ + +Then, the union of the neighborhoods across all $j$ is taken and duplicates are removed, forming the local candidate set: + +$$ +\widehat {\mathbf {X}} _ {l} = \bigcup \mathcal {N} \left(\mathbf {x} _ {l - 1, j}\right). \tag {4} +$$ + +Through MNS, we can capture global structure in the sparse layers (counteracting the kinematic interference + +![](images/bc6bd06ec7f40c3b9eb07d205ec2f5535bf05a39ac88b41a75fbb68e95b7e9a4.jpg) +Figure 3. ACTIVE-PC architecture. First, the point cloud sequence undergoes Multilevel Neighborhood Sampling, followed by feature extraction using a point tube with Elastic Ellipse Query. The dense features are fed into Action Recognizer, while the sparse features are input into Kinematic Interpreter. The resulting human-action and kinematic scores are then fused to obtain the fine-grained action score. + +caused by platform motion from robotic view), while finetuning the extraction of subtle human motion changes in the denser layers. This approach precisely aligns with the ACTIVE dataset's high demand for fine-grained local information when capturing complex human-robot dynamics. + +# 4.4. Layered Recognizers + +Considering the characteristics of the ACTIVE dataset, human action recognition requires not only capturing fine-grained, small-scale movements but also filtering out global kinematic variations caused by robot platform motion. To address these challenges, we design Layered Recognizers, which decouples features from different layers by sending them through specialized sub-networks. The core of the design includes two branches: + +Kinematic Interpreter: For features obtained from larger sampling scales (i.e., sparser layers, such as $\mathbf{F}_2$ and $\mathbf{F}_3$ ), this branch focuses on global kinematic impacts caused by platform motion and changes in the human-robot spatial relationship from the robot's perspective. Since these global features are relatively dispersed in the temporal dimension, we use Global Average Pooling (GAP) to effectively aggregate information from all spatiotemporal regions, ensuring that contributions from different regions are considered. The process is as follows: + +$$ +\begin{array}{l} \mathbf {Y} _ {\text {k i n e m a t i c}} = \operatorname {I n t e r p r e t e r} _ {\text {k i n e m a t i c}} \left(\mathbf {F} _ {2}, \mathbf {F} _ {3}\right) \\ = \frac {1}{2} (\mathbf {M L P} (\mathbf {G A P} (\mathbf {T r a n s f o r m e r} (\mathbf {F} _ {2}))) \quad (5) \\ + \mathbf {M L P} (\mathbf {G A P} (\mathbf {T r a n s f o r m e r} (\mathbf {F} _ {3})))) \\ \end{array} +$$ + +Human Action Recognizer: For features extracted from lower layers with denser point clouds (e.g., $\mathbf{F}_1$ ), these features more directly reflect subtle human actions. Here, we use Global Max Pooling (GMP) to select the most significant spatiotemporal points, emphasizing key fine-grained action details. The calculation is as follows: + +$$ +\begin{array}{l} \mathbf {Y} _ {\text {h u m a n}} = \operatorname {R e c o n g i z e r} _ {\text {h u m a n}} \left(\mathbf {F} _ {1}\right) \tag {6} \\ = \mathbf {M L P} (\mathbf {G M P} (\mathbf {T r a n s f o r m e r} (\mathbf {F} _ {1})). \\ \end{array} +$$ + +Finally, we fuse the classification scores from both branches, averaging their contributions. This approach retains the robustness of global kinematic interpretation while capturing the high discriminability of subtle action changes. + +# 4.5. Elastic Ellipse Query + +Let the 4D point cloud set be $P = \{p_1, p_2, \ldots, p_N\}$ , where each point $p_n = (x_n, y_n, z_n)$ represents the neighboring point $n$ of a query point. Given a query point $p_q$ and a radius $r$ , the objective of the Ball Query is to find all points whose distance from the query point $p_q$ is less than or equal to $r$ . Let $d(p_n, p_q)$ denote the Euclidean distance between point $p_n$ and query point $p_q$ : + +$$ +d \left(p _ {n}, p _ {q}\right) = \sqrt {\left(x _ {n} - x _ {q}\right) ^ {2} + \left(y _ {n} - y _ {q}\right) ^ {2} + \left(z _ {n} - z _ {q}\right) ^ {2}}, \tag {7} +$$ + +where $p_{q} = (x_{q},y_{q},z_{q})$ is the query point. + +The goal of Ball Query is to identify all points $p_n$ satisfying the following condition: + +$$ +\mathcal {B} \left(p _ {q}, r\right) = \left\{p _ {n} \mid d \left(p _ {n}, p _ {q}\right) \leq r \right\}. \tag {8} +$$ + +That is, the set of points within the spherical neighborhood $\mathcal{B}(p_q,r)$ consists of all points $p_n$ such that $d(p_n,p_q)\leq r$ . + +In N-HRI, conventional spherical neighborhood queries (e.g., Ball Query) exhibit inherent limitations when confronted with kinematic-induced geometric distortions. The relative motion between mobile robots and humans introduces anisotropic deformations in the spatiotemporal point cloud distribution. Three critical challenges emerge: 1) Scale Discrepancy: Human actions (e.g., hand gestures) manifest as localized deformations $(0.1 - 0.5\mathrm{m})$ within the robot's egocentric coordinates, while kinematic disturbances induce macroscopic displacements (up to $50\mathrm{m}$ ) along the xy-plane. 2) Axis-wise Heterogeneity: Vertical displacements (along the z-axis) primarily reflect genuine human joint movements, while horizontal movements (in the xy-plane) are more susceptible to artifacts caused by the robot's self-motion and changes in the human-robot spatial positioning. 3) Dynamic Occlusion: Transient point cloud density variations caused by viewpoint changes violate the isotropic neighborhood assumption. + +For RARV, visual similarities influenced by kinematic disturbances from robotic view complicate the differentiation, causing deformations in human-actions and irregular distributions in the spatiotemporal domain. Unlike kinematic disturbances, which have a large range of movement, human-actions are constrained to smaller areas. For instance, while walking involves significant movement in the xy-plane, human-actions like waving hands are restricted to a smaller range. Due to the influence of kinematic, points with large xy distance at different times are more likely to represent the same point, while points with large differences in the vertical dimension are more likely to be distinct. + +To address these challenges, we propose the Elastic Ellipse Query. Considering the distribution differences between datasets, we introduce axis-specific adaptive query scale learning. This mechanism establishes an anisotropic metric space through learnable axis-specific scaling, enabling partial tracking of points and differentiation of actions. Given a query point $p_{q}$ and a neighborhood point $p_{n}$ , our adaptive distance metric is formulated as follows: + +$$ +d \left(p _ {n}, p _ {q}\right) = \sqrt {\alpha \left(x _ {n} - x _ {q}\right) ^ {2} + \beta \left(y _ {n} - y _ {q}\right) ^ {2} + \gamma \left(z _ {n} - z _ {q}\right) ^ {2}}, \tag {9} +$$ + +where $\Omega = \{\alpha, \beta, \gamma\}$ are trainable scaling parameters that automatically adjust the neighborhood aspect ratios. For the ACTIVE dataset, the final values of $\alpha$ , $\beta$ , and $\gamma$ are (3.5632, 3.6789, 2.8038), consistent with the expected scale ratios. This elliptic metric space provides following distinctive advantages: 1) Kinematic-Aware Adaptation: The horizontal scaling factors $(\alpha, \beta)$ suppress planar displacements induced by the robot while preserving fine-grained human motions. In our experiments with the ACTIVE dataset, the optimized parameters $(\alpha = 3.56$ , $\beta = 3.68)$ effectively compress the neighborhood range in the xy-plane relative to the z-axis $(\gamma = 2.80)$ . The vertical scaling factor $\gamma$ remains sensitive to subtle posture changes, + +Table 2. Action recognition accuracy $(\%)$ of ACTIVE-PC on ACTIVE under different dataset settings. + +
MethodVenueSetting
8 × 51212 × 51212 × 768
PSTNet [7]ICLR'2118.8526.4934.10
P4Transformer [6]CVPR'2137.9740.5643.88
PSTNet++ [8]TPAMI'2234.9838.8140.29
PST-Transformer [9]TPAMI'2341.6142.9246.76
3DinAction [2]CVPR'2444.8846.2349.19
ACTIVE-PC w/o EEQ-52.2953.3255.11
ACTIVE-PC (Ours)-58.7159.1860.10
+ +Table 3. Action recognition accuracy $(\%)$ of ACTIVE-PC on NTU RGB+D. CS and CV represent the cross-subject and cross-view evaluation protocols, respectively. + +
MethodVenueCSCV
PSTNet [7]ICLR’202190.596.5
P4Transformer [6]CVPR’202190.296.4
PointCMP [22]CVPR’202388.5-
PointCPSC [23]ICCV’202388.0-
PST-Transformer [9]TPAMI’202391.096.4
ACTIVE-PC (Ours)-91.796.8
+ +Table 4. Action recognition accuracy (\%) of ACTIVE-RGB on ACTIVE under different dataset settings. + +
MethodVenueSetting
8 × 22412 × 224
VideoSwin [17]CVPR'2241.0744.20
MViT V2 [14]CVPR'2238.1541.83
VideoMAE [24]NeurIPS'2242.4645.63
UniFormer [11]TPAMI'2342.5545.60
VideoMAE V2 [28]CVPR'2345.3947.98
UniFormerv2 [12]ICCV'2348.2150.74
InternVideo2 [30]ECCV'2452.5254.88
ACTIVE-RGB (Ours)-55.9057.56
+ +such as a 0.2-meter arm elevation, while the horizontal scaling factors adaptively filter out irrelevant positional shifts. 2) Dynamic Focus Adjustment: Unlike fixed elliptical kernels, our parameters evolve during training to handle varying interaction distances (3-50m). + +# 5. Experiment + +In this section, we evaluate the performance of state-of-the-art methods and ACTIVE-PC, on the ACTIVE dataset across both point cloud and RGB modalities. We also conduct ablation experiments on the proposed components. + +# 5.1. Metrics and Implementation Details + +For evaluation, we use standard action recognition metrics, specifically classification accuracy. In the point cloud modality, we first apply the TransFusion[1] method to detect human subjects within the scene. The point cloud sequence is then segmented into fixed-length clips, with each frame sampled to 768 points. For training, video-level labels are assigned as clip-level labels. The color informa + +Table 5. Impact of Layered Recognizers, MNS and EEQ of ACTIVE-PC on ACTIVE. + +
Layered RecognizersMNSEEQAccuracy (%)
✘✘✘46.76
✓✘✘51.98
✓✓✘55.11
✓✘✓55.63
✓✓✓60.10
+ +Table 6. Influence of the spatial radius $r$ . EEQ reduces the sensitivity to the hyperparameter $r$ . + +
EEQSpatial radius rsVariance
0.10.20.511.52
X54.8355.1155.3954.5254.2153.510.456
✓59.9860.1060.0459.7359.5459.300.100
+ +![](images/64b45f9db885db0221e9ebbfe81e61f818a891ca2bd2b6ba116fb04c82675579.jpg) +Figure 4. t-SNE distribution of PST-Transformer [6] and ACTIVE-PC on ACTIVE dataset. Different colors represent different action categories, with each point corresponding to a video sample. All 11 categories from the Base Motions are visualized. + +tion of the points is not used. For evaluation, the video-level prediction is obtained by averaging the clip-level predicted probabilities. The model is trained using the SGD optimizer for 50 epochs, with a learning rate of 0.01, decaying by a factor of 0.1 at epochs 20 and 30. For the RGB modality, human bodies are detected using the object detection method [26]. Each video is evenly split into 8 or 12 segments, with one frame randomly sampled from each segment. Each frame is resized to $224 \times 224$ . We use the AdamW optimizer [18] for 50 epochs, with a learning rate of 0.0002. All experiments are conducted on the PyTorch platform, utilizing an NVIDIA RTX 4090 GPU. + +# 5.2. Point Cloud Modality Evaluation + +We first validate the performance of ACTIVE-PC on the ACTIVE dataset. As shown in Tab. 2, Transformer-based models, such as P4Transformer and PST-Transformer, outperform CNN-based models achieving Acc-Top1 scores of $43.88\%$ and $46.76\%$ , respectively. ACTIVE-PC surpasses PST-Transformer by $13.34\%$ , reaching an Acc-Top1 of $60.10\%$ , demonstrating the effectiveness of our method. To validate the robustness of the model, we conduct experiments on two general action recognition datasets, MSRAction3D and NTU RGB+D. Tab. 3 presents the results on NTU RGB+D. + +# 5.3. RGB Modality Evaluation + +ACTIVE-RGB is built upon the ViT [5], augmented with Layered Recognizers. Tab. 4 presents a performance comparison of ACTIVE-RGB with other state-of-the-art methods in the RGB modality. ACTIVE-RGB outperforms the leading method, the baseline model InternVideo2, by $2.68\%$ , validating the effectiveness of our approach. + +# 5.4. Ablation Study + +Evaluation of different Components: We conduct ablation experiments on the key components of our model, as well as the number of sampled frames per point cloud video. Results in Tab. 5 show that each proposed modification contributes positively. Specifically, w/o EEQ and w/o MNS show performance drops of $4.99\%$ and $4.47\%$ , respectively, compared to ACTIVE-PC. The significant performance degradation observed with w/o Layered Recognizers highlights the effectiveness of fusing high-frequency and low-frequency features. Allowing the model to simultaneously focus on both action and kinematic variations enhances its ability to handle complex environments. + +Evaluation of the Stability of EEQ: Tab. 6 compares the performance of the full ACTIVE-PC with its variant, which excludes the EEQ module, under different neighborhood radius settings. Traditional neighborhood query methods are constrained by fixed neighborhood radii, whereas EEQ adapts by learning the intrinsic geometric properties of the data distribution, allowing for flexible scaling of the spherical neighborhood, thereby enhancing robustness. The variance of accuracy (acc) for ACTIVE-PC across different neighborhood radii is 0.100, significantly outperforming the variant without EEQ, which exhibits a variance of 0.456. + +# 5.5. Visualization and Analysis + +We utilize t-SNE [25] visualization to analyze feature distributions from the point cloud modality, as shown in Fig. 4. ACTIVE-PC distinctly separates similar actions (e.g., "Raising Arms" vs. "Scratching Head" and "Arms Crossed" vs. "Hands on Waist"), whereas PST-Transformer [6] produces more overlapping clusters. This demonstrates ACTIVE-PC's superior ability to differentiate subtle action categories from a robotic viewpoint. + +# 6. Conclusion + +To the best of our knowledge, our dataset is the first human action recognition dataset specifically designed for Natural Human-Robot Interaction scenarios, laying the foundation for more accurate and detailed understanding of human actions in N-HRI. We also propose ACTIVE-PC designed for action recognition in N-HRI scenarios. Experimental results demonstrate the effectiveness of this approach. + +# References + +[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 7 +[2] Yizhak Ben-Shabat, Oren Shrout, and Stephen Gould. 3d: In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 7 +[3] Elif Bozkaya-Aras, Tolga Onel, Levent Eriskin, and Mumtaz Karatas. Intelligent human activity recognition for healthcare digital twin. *Internet of Things*, 2025. 1 +[4] Haoyu Chen, Henglin Shi, Xin Liu, Xiaobai Li, and Guoying Zhao. Smg: A micro-gesture dataset towards spontaneous body gestures for emotional stress state analysis. International Journal of Computer Vision, 2023. 3 +[5] 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, 2020. 8 +[6] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point 4d transformer networks for spatio-temporal modeling in point cloud videos. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 7, 8 +[7] Hehe Fan, Xin Yu, Yuhang Ding, Yi Yang, and MohanS. Kankanhalli. Pistnet: Point spatio-temporal convolution on point cloud sequences. Cornell University - arXiv, Cornell University - arXiv, 2022. 7 +[8] Hehe Fan, Xin Yu, Yi Yang, and Mohan Kankanhalli. Deep hierarchical representation of point cloud videos via spatiotemporal decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 7 +[9] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point spatiotemporal transformer networks for point cloud video modeling. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 7 +[10] Shiming Ge, Jia Li, Qiting Ye, and Zhao Luo. Detecting masked faces in the wild with lle-cnns. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1 +[11] Kunchang Li, Yali Wang, Junhao Zhang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Jiao Qiao. Uniformer: Unifying convolution and self-attention for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3, 7 + +Acknowledgements This work was supported by National Natural Science Foundation of China (No. 62473007), Natural Science Foundation of Guangdong Province (No. 2024A1515012089), Shenzhen Innovation in Science and Technology Foundation for The Excellent Youth Scholars (No. RCYX20231211090248064). +[12] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Unlocking the potential of image vits for video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 7 +[13] Wanqing Li, Zhengyou Zhang, and Zicheng Liu. Action recognition based on a bag of 3d points. In 2010 IEEE computer society conference on computer vision and pattern recognition-workshops, 2010. 2, 3 +[14] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 7 +[15] Xin Liu, Henglin Shi, Haoyu Chen, Zitong Yu, Xiaobai Li, and Guoying Zhao. imgue: An identity-free video dataset for micro-gesture understanding and emotion analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 3 +[16] Yunze Liu, Junyu Chen, Zekai Zhang, Jingwei Huang, and Li Yi. Leaf: Learning frames for 4d point cloud sequence understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[17] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 7 +[18] I Loshchilov. Decoupled weight decay regularization. 2017. 8 +[19] Marcus Mast, Michael Burmester, Birgit Graf, Florian Weisshardt, Georg Arbeiter, Michal Spanel, Zdenek Materna, Pavel Smrz, and Gernot Kronreif. Design of the human-robot interaction for a semi-autonomous service robot to assist elderly people. 2015. 1 +[20] Yang Mi and Song Wang. Recognizing micro actions in videos: learning motion details via segment-level temporal pyramid. In 2019 IEEE International Conference on Multimedia and Expo (ICME), 2019. 3 +[21] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2016. 2, 3 +[22] Zhiqiang Shen, Xiaoxiao Sheng, Longguang Wang, Yulan Guo, Qiong Liu, and Xi Zhou. Pointcmp: Contrastive mask prediction for self-supervised learning on point cloud videos. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 7 +[23] Xiaoxiao Sheng, Zhiqiang Shen, Gang Xiao, Longguang Wang, Yu Kuen Guo, and Hehe Fan. Point contrastive prediction with semantic clustering for self-supervised learning on point cloud videos. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 7 +[24] 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, 2022. 3, 7 + +[25] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 2008. 8 +[26] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Yolov10: Real-time end-to-end object detection. arXiv preprint arXiv:2405.14458, 2024. 8 +[27] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 2018. 3 +[28] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 7 +[29] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. Intervideo: General video foundation models via generative and discriminative learning. ArXiv, 2022. 3 +[30] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, 2024. 7 +[31] Ping Wei, Yibiao Zhao, Nanning Zheng, and Song-Chun Zhu. Modeling 4d human-object interactions for event and object recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2013. 3 +[32] Sravani Yenduri, Nazil Perveen, Vishnu Chalavadi, and C KrishnaMohan. Fine-grained action recognition using dynamic kernels. Pattern Recognit., 2022. 1 +[33] Ryo Yonetani, Kris M Kitani, and Yoichi Sato. Recognizing micro-actions and reactions from paired egocentric videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016. 3 +[34] Jianqi Zhong, Kaichen Zhou, Qingyong Hu, Bing Wang, Niki Trigoni, and A. Markham. No pain, big gain: Classify dynamic point cloud sequences with static models by fitting feature-level space-time surfaces. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 \ No newline at end of file diff --git a/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/images.zip b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e4a3a2f26476369b5e74407859b050556a0101e --- /dev/null +++ b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb1064bd2c33fbe34ce2f3c64ec1e157bf18f5c8ef182f45320aceaa126d6c66 +size 665325 diff --git a/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/layout.json b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c3b811471c4e85413a6655cd3c982bd0120b36aa --- /dev/null +++ b/ICCV/2025/Recognizing Actions from Robotic View for Natural Human-Robot Interaction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f647f6590006e8744e6be8829f6afaaaeec3cf581dc35fa89b76246c7a5a2c5 +size 382526 diff --git a/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/062120a6-c966-496f-b5f3-707befbc6aaf_content_list.json b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/062120a6-c966-496f-b5f3-707befbc6aaf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2c92eda217ce8c5677ad91f325459b72ea6b4e39 --- /dev/null +++ b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/062120a6-c966-496f-b5f3-707befbc6aaf_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a00671c1d876a328217e07ecfe5e2b3131d912c1d1145b55d065f050f00fc0d2 +size 74986 diff --git a/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/062120a6-c966-496f-b5f3-707befbc6aaf_model.json b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/062120a6-c966-496f-b5f3-707befbc6aaf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bb28115182e3eb0fd05b883aeae63d7d831cd80c --- /dev/null +++ b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/062120a6-c966-496f-b5f3-707befbc6aaf_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e87841044dd5aa4343e01030f9d581bc3f8bc369fe4f53bdfbfa20f4073b29ff +size 94392 diff --git a/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/062120a6-c966-496f-b5f3-707befbc6aaf_origin.pdf b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/062120a6-c966-496f-b5f3-707befbc6aaf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ef15ed65013d0afb9984c9c50c9f49ae072b28b7 --- /dev/null +++ b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/062120a6-c966-496f-b5f3-707befbc6aaf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de26ba813d71a10a00e17e2367cc60605dee6ae8d35b3f88164e717f7dbd3b16 +size 9643357 diff --git a/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/full.md b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8da7b02056bc9914fe65b2d6bd05f4b515ddd814 --- /dev/null +++ b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/full.md @@ -0,0 +1,289 @@ +# ReconDreamer++: Harmonizing Generative and Reconstructive Models for Driving Scene Representation + +Guosheng Zhao $^{1,2,3*}$ Xiaofeng Wang $^{1,2,3*}$ Chaojun Ni $^{3,4*}$ Zheng Zhu $^{3*}$ Wenkang Qin $^{3}$ Guan Huang $^{3}$ Xingang Wang $^{1,5\dagger}$ + +$^{1}$ Institute of Automation, Chinese Academy of Sciences + +$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences $^{3}$ GigaAI $^{4}$ Peking University $^{5}$ Luoyang Institute for Robot and Intelligent Equipment + +![](images/a32647c0149f62bb734bb8ff43c6526488e17642772832d560b58b7be5902c02.jpg) + +![](images/6087cb64992a55e0893a86fa424d9526801d7e6952b2f9c57b999f21987465ea.jpg) + +![](images/9896cb06c61823bc8cc7e5ec8564a1363d90306c2e36ccd2e92776c049181ca4.jpg) + +![](images/a5cb0aa623ecb0a25b5ca1a0fe49ef512f8f8feb83e02d16ecc703edb7d77e1e.jpg) +Street Gaussians + +![](images/f15c4af2c2d8a4a93b6925440485d5066094e74cf918b2445e846129e1659f29.jpg) +ReconDreamer + +![](images/fa2d4237efca6a99ec2fa6dbe3f2397c7be2e8ded8e8baff16e9aff759f451ce.jpg) +ReconDreamer++ + +![](images/8c0109ee5f3fabbaeccedfafe97c836a52ec58e826bd13a3c6988a26a5e9d62d.jpg) + +![](images/0a5f66e03ee2df88e8fff91f96f45495601d426928229fb743b6fa6179aee6bf.jpg) +Figure 1. Comparison of ReconDreamer++ with SOTA methods, Street Gaussians [46] and ReconDreamer [29], on original and novel trajectories. Left: ReconDreamer++ demonstrates superior rendering performance for both vehicle foregrounds and road surfaces compared to existing SOTA methods. Right: ReconDreamer++ significantly improves performance on novel trajectories while maintaining high rendering quality on the original trajectory. + +# Abstract + +Combining reconstruction models with generative models has emerged as a promising paradigm for closed-loop simulation in autonomous driving. For example, ReconDreamer has demonstrated remarkable success in rendering large-scale maneuvers. However, a significant gap remains between the generated data and real-world sensor observations, particularly in terms of fidelity for structured elements, such as the ground surface. To address these challenges, we propose ReconDreamer++, an enhanced framework that significantly improves the overall rendering quality by mitigating the domain gap and refining the representation of the ground surface. Specifically, ReconDreamer++ introduces the Novel Trajectory Deformable Network (NTDNet), which leverages learnable spatial de + +formation mechanisms to bridge the domain gap between synthesized novel views and original sensor observations. Moreover, for structured elements such as the ground surface, we preserve geometric prior knowledge in 3D Gaussians, and the optimization process focuses on refining appearance attributes while preserving the underlying geometric structure. Experimental evaluations conducted on multiple datasets (Waymo, nuScenes, PandaSet, and EUVS) confirm the superior performance of ReconDreamer++. Specifically, on Waymo, ReconDreamer++ achieves performance comparable to Street Gaussians for the original trajectory while significantly outperforming ReconDreamer on novel trajectories. In particular, it achieves substantial improvements, including a $6.1\%$ increase in NTA-IoU, a 23.0% improvement in FID, and a remarkable $4.5\%$ gain in the ground surface metric NTL-IoU, highlighting its effectiveness in accurately reconstructing structured elements such as the road surface. + +# 1. Introduction + +Robust driving scene representation is of paramount importance for autonomous driving simulation systems. With advances in Neural Radiance Fields (NeRF) [12, 27, 38, 48] and 3D Gaussian Splitting (3DGS) [6, 17, 20, 42, 46, 50], substantial progress has been achieved in reconstructing autonomous driving scenes along recorded trajectories. By incorporating generative priors [4, 11, 32, 34, 40, 58], the performance of scene representation [29, 47, 51, 54, 57] in novel trajectories has been significantly improved. These developments make the realization of closed-loop simulation for autonomous driving systems increasingly feasible. + +Specifically, DriveDreamer4D [57] introduces a world model [58] to directly generate videos for novel trajectories, which are then used in mixed training with 4DGS models. This significantly improves the quality of the reconstruction model under novel trajectories. Furthermore, ReconDreamer [29] leverages a crafted world model for video restoration, enhancing the spatial consistency of reconstructed outputs under large maneuvers. However, existing methods that mix generated data with real-world sensor observations to train the same set of Gaussian parameters often overlook the inherent gap between these two data sources. This oversight has two critical consequences. First, it results in significant quality degradation on the original trajectory, as illustrated in Fig. 1, where ReconDreamer, although it performs well under a lane shift of 3 meters, exhibits a noticeable performance decline on the original trajectory. Second, this neglect exacerbates the low-quality rendering of structured elements within the scene, such as the ground surface. A key challenge lies in the fact that the ground surface is typically parallel to the camera's optical axis, presenting significant difficulties for traditional reconstruction methods and leading to poor robustness in modeling the ground surface. As a result, inconsistencies between generated and real-world data are further exacerbated in this region, causing suboptimal modeling quality and increased errors. + +To address the aforementioned challenges, we propose ReconDreamer++, an innovative approach that significantly improves the overall rendering quality by mitigating the domain gap and refining the representation of the ground surface. Specifically, we introduce the Novel Trajectory Deformable Network (NTDNet) to systematically bridge the inherent domain gap between generated data and recorded observations. This network is specifically designed to model and learn spatial deformations, enabling a coherent mapping between synthesized and real-world data. Moreover, we independently model the road surface by incorporating 3D geometric priors into the ground Gaussian model, with an emphasis on optimizing appearance attributes while preserving the underlying geometric structure. As depicted in Fig. 1, experimental results demonstrate the superior + +performance of ReconDreamer++ along original and novel trajectories. Furthermore, the exceptional performance across four widely-used datasets—Waymo [36], nuScenes [5], PandaSet [44], and EUVS [14]—demonstrates the robustness of ReconDreamer++. Specifically, on the Waymo dataset, ReconDreamer++ achieves significant improvements over ReconDreamer, with increases of $6.1\%$ in NTA-IoU, $4.5\%$ in NTL-IoU, and $23.0\%$ in FID. + +The main contributions of this paper are summarized as follows: + +- We propose ReconDreamer++, an innovative approach that significantly improves the overall rendering quality by mitigating the domain gap and refining the representation of the ground surface. +- To bridge the domain gap between generated and recorded data, we introduce the NTDNet, designed to learn spatial deformations between synthesized data and real-world observations. Additionally, the ground surface is modeled independently by incorporating 3D geometric priors into the ground Gaussian model. The optimization process focuses on appearance while preserving geometric fidelity, effectively alleviating the issue of cumulative errors between reconstruction and generative models. +- Experimental results demonstrate the superior performance of ReconDreamer++ across multiple popular datasets (Waymo [36], nuScenes [5], PandaSet [44], EUVS [14]). Specifically, in Waymo, ReconDreamer++ outperforms Street Gaussians by $6.4\%$ in the LPIPS metric on the original trajectory. On novel trajectories, it surpasses ReconDreamer by $6.1\%$ in NTA-IoU, $4.5\%$ in NTL-IoU, and $23.0\%$ in FID. + +# 2. Related Work + +# 2.1. Driving Scene Reconstruction + +NeRF and 3D Gaussian Splatting (3DGS) have become two of the most influential paradigms for 3D scene representation. NeRF models [2, 3, 27, 28] leverage multilayer perceptron (MLP) networks to represent continuous volumetric scenes, enabling highly detailed reconstructions with exceptional rendering quality. On the other hand, 3DGS [20, 53] adopts a different approach by defining anisotropic Gaussians in the 3D space, using adaptive density control to produce high-quality renderings from the input of a sparse point cloud. These techniques have been adapted and extended for autonomous driving scenarios [6, 12, 17, 19, 24, 26, 33, 37, 48, 49, 60], where the dynamic and complex nature of driving environments presents unique challenges. Street Gaussians [46] achieves remarkable results by decoupling dynamic scenes into foreground moving objects and static background. OmniRe [7] further refines the dynamic foreground by distinguishing between rigid vehicles and non-rigid elements such as pedestrians + +![](images/834b03388e89c218ec13d6d967b9ef045c7e412ce079edcc9e2303df0e555f5b.jpg) +Figure 2. The overall framework of ReconDreamer++. The driving scene is decomposed into three components: the ground surface, non-ground static background, and dynamic objects. For camera poses on the original trajectory, rendering is performed directly by skipping NTDNet. For camera poses on novel trajectories, the Gaussian parameters are refined through NTDNet to perform rendering. + +and cyclists, thereby significantly enhancing the quality of dynamic driving scene reconstruction. Despite these advancements, significant limitations remain. Existing methods [1, 8, 10, 21-23, 25, 30, 31, 35, 38, 43, 45, 59] are highly dependent on the density and distribution of the input sensor data. When applied to scenarios that deviate from the distribution of training data, such as lane shifts, their performance tends to degrade significantly. + +# 2.2. Modeling Driving Scene with Diffusion Prior + +Recent advancements in image and video generation models [4, 16, 32, 34, 41, 52, 61] have paved the way for a new paradigm in enhancing the representation of driving scenes by leveraging generative priors [9, 13, 18, 39, 54]. DriveDreamer4D [57] pioneers the use of a world model [40, 58] to directly generate novel trajectory videos for the purpose of jointly training reconstruction models. ReconDreamer [29] focuses on leveraging video generation models to refine the rendering results of novel viewpoints. By enhancing the spatial consistency between the generated outputs and the recorded sensor data, ReconDreamer achieves a significant improvement in performance when reconstructing scenes from previously unseen perspectives. These advancements collectively underscore the potential of generative approaches in addressing the challenges of novel view synthesis in autonomous driving scenarios. However, these approaches often overlook the inherent domain gap between the generated data and the original sensor observa + +tions. This neglect not only results in degraded rendering quality for the original trajectory but also amplifies the low-quality rendering of structured elements within the scene, such as the ground surface. + +# 3. Method + +# 3.1. Preliminary: ReconDreamer + +ReconDreamer [29] proposes the incremental integration of world model knowledge, significantly enhancing the model's performance for large maneuvers. The training process of ReconDreamer is divided into two stages. In the first stage, a partially trained reconstruction model is utilized to render videos along the original trajectory. This process naturally introduces ghosting artifacts or blurriness, a direct consequence of the model's underfitting. These degraded outputs are then paired with recorded sensor data to form training pairs, which are used to optimize the video restoration model, referred to as DriveRestorer [29]. In the second stage, degraded videos of novel trajectories are first rendered, and the fine-tuned DriveRestorer is subsequently employed to refine these rendering results, enhancing both their quality and spatial consistency. These restored outputs, combined with the original video data, are then integrated into the training pipeline to further enhance the performance and robustness of the reconstruction model. This two-stage approach ensures that ReconDreamer achieves superior performance across the novel trajectories. + +# 3.2. Overview of ReconDreamer++ + +The general pipeline of ReconDreamer++ is illustrated in Fig. 2. First, the driving scene is decomposed into three components: ground surface, non-ground static background, and dynamic objects. By independently modeling the road surface, the robustness of the reconstruction model is improved, strengthening spatial consistency between generated data and recorded observations. Next, the Novel Trajectory Deformable Network (NTDNet) is introduced to systematically address the domain gap between generated and observed data. Finally, the ReconDreamer framework is adopted to restore the rendered videos along novel trajectories. The repaired results are then used to supervise and further refine the reconstruction model. + +# 3.3. Driving Scene Representation + +ReconDreamer [29] builds upon Street Gaussians [46], separately modeling dynamic foreground objects and static background elements while jointly optimizing them, achieving notable success. However, as illustrated in Fig. 1, this approach exhibits a lack of 3D consistency between the rendered lane markings in novel trajectories and the original data. This limitation is largely attributed to the fact that in autonomous driving scenarios, the camera's optical axis is generally parallel to the ground. Consequently, when viewed from novel perspectives, the reconstruction of the road surface often becomes suboptimal, frequently introducing streak-like or blurry artifacts. Using these rendered results, which lack 3D consistency constraints, as conditions for DriveRestorer [29] can further amplify these errors, leading to a degradation in overall performance. + +To alleviate this issue, we separately model the road surface, with the entire driving scene representation consisting of three components: the ground, non-ground static background, and dynamic objects. + +Ground Model. The ground model is represented by a set of Gaussians $\mathcal{G}^g$ in the global coordinate system. Each Gaussian [20] is parameterized by its central position $x$ , opacity $\gamma$ , covariance $\Sigma$ , and view-dependent RGB color $c$ , controlled via spherical harmonics. For stability, each covariance matrix $\Sigma$ is decomposed by: + +$$ +\boldsymbol {\Sigma} = \boldsymbol {R} \boldsymbol {S} \boldsymbol {S} ^ {T} \boldsymbol {R} ^ {T}, \tag {1} +$$ + +where scaling matrix $S$ and a rotation matrix $R$ are learnable parameters, represented by scaling $s$ and quaternion $r$ . To integrate 3D structural priors into the ground model, we employ [55] to segment road point clouds, which are used to initialize the ground model. During training, the position $x$ of the ground model $\mathcal{G}^g$ is fixed, and only the remaining parameters are optimized. Moreover, due to the rich availability of multi-frame fused ground point cloud data, we directly fix the number of Gaussians in the ground model, + +thereby avoiding any modifications to the count of Gaussian spheres. This explicit incorporation of 3D LiDAR priors significantly reduces the search space for the reconstruction model while enhancing its generalization capabilities. + +Non-ground Background Model. In the world coordinate system, the non-ground background model $\mathcal{G}^{bg}$ is initialized using non-ground point cloud data, excluding moving objects. To further enhance the expressive capacity of the model, additional points are randomly generated for initialization, following the strategy outlined in [6]. In contrast to the ground model, all parameters of the background model $\mathcal{G}^{bg}$ are fully optimized during training. + +Object Model. The dynamic object models are defined by $\{\mathcal{G}_i^o, 1 \leq i \leq N\}$ in the local object coordinate system, where $N$ denotes the number of dynamic objects in a specific scenario. Since the object models are defined in a different coordinate system than the ground and non-ground background models, they need to be transformed into the world coordinate system during rendering based on the vehicle's pose. Suppose $\boldsymbol{x}_o$ and $\boldsymbol{q}_o$ denote the position and rotation quaternion of the Gaussian in the local object coordinate system, respectively. The pose of the dynamic object at timestep $t$ is represented by the rotation matrix $R_t$ and the translation vector $T_t$ . Then, the position $\boldsymbol{x}_w$ and rotation quaternion $\boldsymbol{q}_w$ of the corresponding Gaussian in the world coordinate system can be obtained as follows: + +$$ +\boldsymbol {x} _ {w} = R _ {t} \boldsymbol {x} _ {o} + T _ {t}, \quad \boldsymbol {q} _ {w} = \operatorname {R O T} \left(R _ {t}, \boldsymbol {q} _ {o}\right), \tag {2} +$$ + +where $\mathrm{ROT}(\cdot)$ denotes the operation of rotating the quaternion by the rotation matrix. Additionally, multi-frame fused object point clouds are used to initialize the dynamic object models. + +# 3.4. Novel Trajectory Deformable Network + +Existing methods [29, 47, 51, 57] improve performance along novel trajectories by leveraging generative priors. However, these approaches often experience a noticeable decline in performance when applied to the original trajectories. To effectively bridge the domain gap between generated results and original observations, the Novel Trajectory Deformable Network (NTDNet) is proposed. As illustrated in Fig. 2, NTDNet comprises a pose feature module $\mathcal{F}_{\phi}$ and a temporal field module $\mathcal{F}_{\theta}$ . Given the original trajectory of the ego car $\mathcal{T}^{\mathrm{ori}} = \{p_t^{\mathrm{ori}}, 1 \leq t \leq F\}$ and the novel trajectory $\mathcal{T}^{\mathrm{novel}} = \{p_t^{\mathrm{novel}}, 1 \leq t \leq F\}$ , the corresponding delta pose $\Delta p_t$ can be calculated by: + +$$ +\Delta p _ {t} = \frac {p _ {t} ^ {\mathrm {n o v e l}} - p _ {t} ^ {\mathrm {o r i}}}{L}, \tag {3} +$$ + +where $L$ refers to a hyperparameter used to normalize the delta pose. This computed delta pose $\Delta p_{t}$ serves as the input for the pose feature module $\mathcal{F}_{\phi}$ . For the temporal field network $\mathcal{F}_{\theta}$ , its inputs are the timestep $t$ and the Gaussian + +![](images/52decaded00fa51edeb2714a9731bbbcae282219ea2d93852db6a7cb2f02dd40.jpg) +Figure 3. Qualitative comparisons of different trajectory renderings on Waymo [36]. The orange boxes highlight that ReconDreamer++ significantly enhances the rendering quality across various SOTA methods [29, 39, 46, 57]. + +parameters $g$ . Then we can obtain the delta Gaussian parameters $\Delta g$ : + +$$ +\Delta g = \mathcal {F} _ {\text {o u t}} \left(\mathcal {F} _ {\phi} \left(\Delta p _ {t}\right) + \mathcal {F} _ {\theta} (\mathrm {P E} (g), \mathrm {P E} (t))\right), \tag {4} +$$ + +where $\mathrm{PE}(\cdot)$ represents the positional encoding [50], and $\mathcal{F}_{out}$ refers to the output linear layer. Additionally, both $\mathcal{F}_{\phi}$ and $\mathcal{F}_{\theta}$ are MLP layers. Finally, the updated Gaussian parameters $g^{\prime} = g + \Delta g$ are used to render the video along the novel trajectory. Notably, the NTDNet is applied exclusively to the novel trajectory. In contrast, for the original trajectory, rendering is performed directly using the Gaussian parameters $g$ without any pose adjustments. + +# 3.5. Optimization + +Different from ReconDreamer [29], we introduce depth supervision for synthesized data. Due to the domain gap between generated and recorded data, the direct use of point cloud re-projection of a single frame may result in performance degradation [57]. To mitigate this issue, we leverage multi-frame fused static point clouds to provide depth constraints for novel viewpoints. Specifically, we first fuse the point clouds from all frames to obtain $Pts_{\text{use}}$ : + +$$ +P t s _ {f u s e} = \bigcup_ {t = 1} ^ {F} \left\{p \in P t s _ {t} \mid \operatorname {I s S t a t i c} (p) \right\}, \tag {5} +$$ + +where $Pts_{t}$ represents the point cloud of the $t$ -th frame in the global coordinate system, and IsStatic( $\cdot$ ) denotes a function that determines whether a point cloud is static. + +Then using the camera pose $C_{novel}$ of the novel viewpoint, the coarse depth map for the novel trajectory can be obtained. Finally, by leveraging the dynamic object mask $M_{novel}$ in the novel viewpoint, a static background depth map $D_{novel}$ can be obtained by filtering out the depth of dynamic objects as follows: + +$$ +D _ {n o v e l} = \operatorname {P r o j} \left(P t s _ {f u s e}, C _ {n o v e l}\right) \odot (1 - M _ {n o v e l}), \tag {6} +$$ + +where $\operatorname{Proj}(\cdot)$ denotes the camera projection transformation, and $\odot$ represents the element-wise multiplication. + +The overall training loss is: + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {\mathrm {R G B}} + \lambda_ {2} \mathcal {L} _ {\mathrm {S S I M}} + \lambda_ {3} \mathcal {L} _ {\mathrm {D e p t h}}, \tag {7} +$$ + +where $\mathcal{L}_{\mathrm{RGB}}$ , $\mathcal{L}_{\mathrm{SSIM}}$ , $\mathcal{L}_{\mathrm{Depth}}$ are reconstruction losses typically used in the Gaussian Splitting optimization [46], and $\lambda_1, \lambda_2$ and $\lambda_3$ are loss weights. + +# 4. Experiments + +In this section, we present our experimental setup, which encompasses the datasets, baselines, implementation details, and evaluation metrics. Subsequently, both quantitative and qualitative results are provided to demonstrate the + +
MethodLane Shift @ 0mLane Shift @ 3mLane Shift @ 6m
PSNR ↑SSIM ↑LPIPS ↓NTA-IoU ↑NTL-IoU ↑FID↓NTA-IoU ↑NTL-IoU ↑FID↓
Street Gaussians [46]36.500.95700.11490.49853.19130.750.37449.27213.04
FreeVS [39]26.590.80570.13250.50556.84104.230.46555.37121.44
DriveDreamer4D with PVG [57]34.370.94470.13210.45753.30113.450.15950.09261.81
ReconDreamer [29]34.310.94310.15190.53954.5893.560.46752.58149.19
ReconDreamer++36.290.95730.10760.57257.0672.020.48956.57111.92
+ +Table 1. Comparison of different lane shifts on the Waymo dataset with various methods [29, 39, 46, 57]. + +
MethodLane Shift @ 0mLane Shift @ 3mLane Shift @ 6m
PSNR ↑SSIM ↑LPIPS ↓NTA-IoU ↑NTL-IoU ↑FID↓NTA-IoU ↑NTL-IoU ↑FID↓
Street Gaussians [46]34.910.95280.11850.21948.72149.500.13246.57247.73
ReconDreamer [29]33.850.94970.12350.32549.72125.430.24647.08197.52
ReconDreamer++34.690.95330.11610.36551.33118.270.26849.26187.51
+ +Table 2. Comparison of different lane shifts on the nuScenes dataset with various methods [29, 46]. + +superior performance of the proposed ReconDreamer++. Finally, we perform experiments to investigate the effectiveness of the depth loss, ground model, and NTDNet. + +# 4.1. Experiment Setup + +Dataset. We conduct our experiments on various datasets, including Waymo [36], nuScenes [5], PandaSet [44] and EUVS [14]. For Waymo, we follow the configuration of ReconDreamer [29]. For nuScenes, we select eight scenes rich in dynamic objects and complex lane structures to test large maneuver capabilities. For PandaSet, we adopt the ten scenes used in Unisim [49] and NeuRAD [38] for fair comparison. For EUVS, we select four scenes from Level 1 for evaluation. Specifically, we use only a single video segment from the training set for training and employ the test set (offset by one lane from the training set) for evaluation. These datasets enable comprehensive performance analysis across diverse driving scenarios. + +Baselines. Since existing methods typically conduct experiments on one or two datasets with varying scene selections, we chose different state-of-the-art (SOTA) methods for comparison across different datasets. Specifically, on Waymo, we compare with FreeVS [39], Street Gaussians [46], DriveDreamer4D [57], and ReconDreamer [29]. For nuScenes and EUVS, we reproduce and evaluate Street Gaussians and ReconDreamer on our selected scenes. On PandaSet, we include comparisons with Street Gaussians, ReconDreamer, UniSim [49], and NeuRAD [38]. + +Implementation details. The training strategies and hyperparameters are configured to align with those of ReconDreamer [29], with a total of 50,000 training steps. Both the temporal field module and pose feature module are lightweight networks, implemented using 8 layers of MLPs with a hidden dimension of 256. + +Metrics. For the evaluation of the original trajectory, we use PSNR, SSIM, and LPIPS [56] as metrics. For novel trajectories, we follow the evaluation protocol of DriveDreamer4D [57], utilizing three key metrics: Novel Trajectory Agent IoU (NTA-IoU) to assess foreground quality, Novel Trajectory Lane IoU (NTL-IoU) to evaluate background lane markings, and Fréchet Inception Distance (FID) [15] to measure overall rendering fidelity. + +# 4.2. Main Results + +Waymo. For Waymo dataset [36], we compare our method with several SOTA approaches [29, 39, 46, 57]. Notably, the generation results of FreeVS [39] rely on LiDAR point clouds, which means that areas in the upper part of the image that are typically not scanned by LiDAR are ignored by this method. Therefore, when calculating metrics, we crop the corresponding regions of the ground truth images accordingly. The results shown in Tab. 1 demonstrate that our ReconDreamer++ not only achieves performance comparable to traditional reconstruction methods like Street Gaussians [46] on the original trajectory but also improves the LPIPS metric by $6.4\%$ . However, FreeVS, ReconDreamer [29], and DriveDreamer4D [57] all exhibit significant quality degradation on the original trajectory. Additionally, on novel trajectories, ReconDreamer++ comprehensively outperforms the SOTA method ReconDreamer. Specifically, for the lane shift of 3 meters, it achieves improvements of $6.1\%$ in NTA-IoU, $4.5\%$ in NTL-IoU, and $23.0\%$ in FID. For the lane shift of 6 meters, it achieves gains of $4.7\%$ in NTA-IoU, $7.6\%$ in NTL-IoU, and $25.0\%$ in FID. Fig. 3 demonstrates that ReconDreamer++ achieves results comparable to Street Gaussians on the original trajectory, while surpassing all SOTA methods on novel trajectories, achieving the best performance. Specifically, under a 3-meter + +![](images/5f716d47cd2a3d96663b10c5db047442eb08c07c94397bcd44c2f4c813ad69a5.jpg) +Figure 4. Ablation studies on the depth loss, ground model and NTDNet. When none of these components are employed, the model corresponds to ReconDreamer. In contrast, when all of these components are integrated, the model represents ReconDreamer++. This highlights the incremental contributions of each component in enhancing the overall performance. + +
MethodNo ShiftLane 2mLane 3mVert. 1m
UniSim [49]-74.797.5-
NeuRAD [38]25.072.393.976.3
Street Gaussians [46]14.666.380.778.4
ReconDreamer [29]16.165.474.967.7
ReconDreamer++13.861.971.762.4
+ +Table 3. FID scores for ego vehicle pose shifts in PandaSet. + +
MethodTest Set
PSNR ↑SSIM ↑LPIPS ↓
Street Gaussians [46]15.080.60430.3428
ReconDreamer [29]20.140.69140.3286
ReconDreamer++20.520.70330.3057
+ +Table 4. Comparison of various methods [29, 46] on the EUVS test set from Level 1. + +lane shift, ReconDreamer++ maintains spatial consistency of lane markings with respect to the ground truth. In contrast, ReconDreamer fails to preserve this consistency, highlighting the superiority of ReconDreamer++ in handling challenging scenarios. + +nuScenes. As shown in Tab. 2, we reproduce Street Gaussians [46] and ReconDreamer [29] on nuScenes for comparison with ReconDreamer++. The results align with conclusion on Waymo [36], where ReconDreamer++ achieves performance comparable to Street Gaussians on the original trajectory. Specifically, the metrics are as follows: PSNR of 34.69 versus 34.91, SSIM of 0.9533 versus 0.9528, and LPIPS of 0.1161 versus 0.1185. For novel trajectories with a 3-meter lateral shift, ReconDreamer++ shows noteworthy improvements over ReconDreamer, achieving NTA-IoU of + +0.365 (a $12.3\%$ improvement), NTL-IoU of 51.33 (a $3.2\%$ improvement), and FID of 118.27 (a $5.7\%$ improvement). Similarly, for trajectories with a 6-meter lateral shift, ReconDreamer++ attains NTA-IoU of 0.268 (a $8.9\%$ improvement), NTL-IoU of 49.26 (a $4.6\%$ improvement), and FID of 187.51 (a $5.1\%$ improvement). + +PandaSet. For PandaSet [44], we conduct a comprehensive comparison of our method with several SOTA approaches, including UniSim [49], NeuRAD [38], Street Gaussians [46], and ReconDreamer [29]. Notably, the results for Street Gaussians and ReconDreamer are based on our own reproduction. To ensure consistency with the evaluation protocols of UniSim and NeuRAD, we focus exclusively on the FID metric for assessing performance. Specifically, we evaluate the FID scores on the original trajectory, as + +
Depth lossGround modelNTDNetLane Shift @ 0mLane Shift @ 3mLane Shift @ 6m
PSNR ↑SSIM ↑LPIPS ↓NTA-IoU ↑NTL-IoU ↑FID↓NTA-IoU ↑NTL-IoU ↑FID↓
---34.310.94310.15190.53954.5893.560.46752.58149.19
✓--35.130.95260.12070.56155.9177.510.47353.55124.52
✓✓-34.930.95180.11740.56656.8975.220.48455.34121.46
✓✓✓36.290.95730.10760.57257.0672.020.48956.57111.92
+ +Table 5. Ablation study of the depth loss, ground model and NTDNet on the Waymo dataset. + +well as for novel trajectories with lateral shifts of $2\mathrm{m}$ and $3\mathrm{m}$ , and a vertical shift of $1\mathrm{m}$ . As shown in Tab. 3, our method achieves SOTA results, demonstrating superior performance in terms of rendering quality and generalization capabilities across all evaluated scenarios. These results highlight the effectiveness of our approach in addressing the challenges posed by complex driving scenes and diverse trajectory variations. + +EUVS. We reproduce Street Gaussians [46] and ReconDreamer [29] using the EUVS dataset [14] and report the performance comparison with ReconDreamer++ on the test set metrics in Tab. 4. As shown in Tab. 4, ReconDreamer++ demonstrates outstanding performance on the EUVS test set, achieving remarkable results across multiple evaluation metrics. Specifically, it scores 20.52 in PSNR, 0.7033 in SSIM, and 0.3057 in LPIPS, showcasing its ability to produce high-fidelity reconstructions. Compared to ReconDreamer, these results represent significant improvements of $1.9\%$ in PSNR, $1.7\%$ in SSIM, and an impressive $7.0\%$ in LPIPS. These enhancements highlight the effectiveness of ReconDreamer++ in capturing finer details and improving perceptual quality, marking a substantial advancement over its predecessor in handling complex driving scenarios. + +# 4.3. Ablation Study + +ReconDreamer serves as our baseline method, and the improvements achieved by ReconDreamer++ can be attributed to three key aspects: the introduction of depth supervision for novel views, the separate modeling of the road surface, and the mitigation of the gap between generated and real data. To evaluate the individual contributions of these components, we conduct an ablation study on the depth loss, ground model, and NTDNet. The experimental results presented in Tab. 5 and Fig. 4 provide valuable insight into the contributions of each component. + +Depth loss. The introduction of depth loss specifically for novel trajectories strengthens the model's spatial understanding, resulting in notable improvements in the overall performance of the reconstruction process. As illustrated in Fig. 4, the introduction of depth supervision for novel viewpoints significantly enhances the model's ability to accurately model vehicles, resulting in sharper and more detailed reconstructions. This improvement directly translates into better performance across all metrics reported in Tab. 5. + +Ground model. As shown in Fig. 4, the introduction of the ground model notably improves reconstruction quality, especially in maintaining geometric consistency of lane markings relative to the ground truth. This enhancement can be credited to the independent modeling of the ground surface, which effectively reduces the cumulative error between the generative and reconstruction models. This is further validated by the quantitative results in Tab. 5. Notably, for large lateral shifts of 6 meters, the model achieves a $3.3\%$ improvement in NTL-IoU, highlighting its effectiveness in addressing challenging scenarios. + +NTDNet. Both Fig. 4 and Tab. 5 demonstrate that the integration of NTDNet further elevates the model's performance across both original and novel trajectories. These results indicate that NTDNet not only bridges the gap between original observations and synthetic data but also unlocks the model's potential by enabling robust and high-quality reconstructions in diverse scenarios. + +In summary, the ablation study highlights the complementary roles of depth loss, the ground model, and NTDNet in improving accuracy, consistency, and generalization. + +# 5. Discussion and Conclusion + +Owing to the rapid advancements in generative models, existing reconstruction methods leveraging generative priors have made remarkable strides in rendering novel trajectories, showcasing significant potential for closed-loop simulation in autonomous driving. Despite these achievements, a critical limitation persists: the domain gap between generated and recorded data is frequently overlooked. This neglect can result in inaccuracies that severely undermine spatial consistency and compromise the overall quality of scene reconstruction, particularly in complex driving scenarios. To tackle this challenge, we introduce ReconDreamer++, a novel framework that integrates NTDNet and adopts a dedicated modeling approach for the ground surface. By explicitly addressing the domain gap and enhancing the representation of structured elements, ReconDreamer++ achieves substantial improvements in both original and novel trajectories. Extensive experiments conducted on multiple datasets (Waymo, nuScenes, PandaSet, and EUVS) highlight the effectiveness and robustness of our method across a wide range of scenarios. + +# References + +[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In CVPR, 2023. 3 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2 +[4] 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, 3 +[5] 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. 2, 6 +[6] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2, 4 +[7] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 2 +[8] Kai Cheng, Xiaoxiao Long, Wei Yin, Jin Wang, Zhiqiang Wu, Yuexin Ma, Kaixuan Wang, Xiaozhi Chen, and Xuejin Chen. Uc-nerf: Neural radiance field for under-calibrated multi-view cameras in autonomous driving. arXiv preprint arXiv:2311.16945, 2023. 3 +[9] Lue Fan, Hao Zhang, Qitai Wang, Hongsheng Li, and Zhaoxiang Zhang. Freesim: Toward free-viewpoint camera simulation in driving scenes. arXiv preprint arXiv:2412.03566, 2024.3 +[10] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 3 +[11] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 2 +[12] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2 +[13] Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, and Chunxia Xiao. Ggs: Generalizable gaussian splatting for lane switching in autonomous driving. arXiv preprint arXiv:2409.02382, 2024. 3 + +[14] Xiangyu Han, Zhen Jia, Boyi Li, Yan Wang, Boris Ivanovic, Yurong You, Lingjie Liu, Yue Wang, Marco Pavone, Chen Feng, and Yiming Li. Extrapolated urban view synthesis benchmark. arXiv preprint arXiv:2412.05256, 2024. 2, 6, 8 +[15] 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. NeurIPS, 2017. 6 +[16] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3 +[17] Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. $s^3$ gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323, 2024. 2 +[18] Sungwon Hwang, Min-Jung Kim, Taewoong Kang, Jayeon Kang, and Jaegul Choo. VEGs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors. In ECCV, 2024. 3 +[19] Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Vitor Guizilini, Thomas Kollar, Adrien Gaidon, Zsolt Kira, and Rares Ambrus. Neo 360: Neural fields for sparse view synthesis of outdoor scenes. In ICCV, 2023. 2 +[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM ToG, 2023. 2, 4 +[21] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2407.02598, 2024. 3 +[22] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, 2022. +[23] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 3 +[24] Zhuopeng Li, Yilin Zhang, Chenming Wu, Jianke Zhu, and Liangjun Zhang. Ho-gaussian: Hybrid optimization of 3d gaussian splatting for urban scenes. arXiv preprint arXiv:2403.20032, 2024. 2 +[25] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia, 2022. 3 +[26] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In ICCV, 2023. 2 +[27] 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. CACM, 2021. 2 + +[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM ToG, 2022. 2 +[29] Chaojun Ni, Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Wenkang Qin, Guan Huang, Chen Liu, Yuyin Chen, Yida Wang, Xueyang Zhang, et al. Recondreamer: Crafting world models for driving scene reconstruction via online restoration. arXiv preprint arXiv:2411.19548, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[30] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, 2021. 3 +[31] 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. 3 +[32] 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 +[33] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3 +[35] 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, 2023. 3 +[36] 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 CVPR, 2020. 2, 5, 6, 7 +[37] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 2 +[38] Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In CVPR, 2024. 2, 3, 6, 7 +[39] Qitai Wang, Lue Fan, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. Freevs: Generative view synthesis on free driving trajectory. arXiv preprint arXiv:2410.18079, 2024. 3, 5, 6 +[40] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 3 + +[41] Xiaofeng Wang, Kang Zhao, Feng Liu, Jiayu Wang, Guosheng Zhao, Xiaoyi Bao, Zheng Zhu, Yingya Zhang, and Xingang Wang. Egovid-5m: A large-scale video-action dataset for egocentric video generation. arXiv preprint arXiv:2411.08380, 2024. 3 +[42] 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 CVPR, 2024. 2 +[43] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In ICAI, 2023. 3 +[44] 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 ITSC, 2021. 2, 6, 7 +[45] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. arXiv preprint arXiv:2303.00749, 2023. 3 +[46] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 1, 2, 4, 5, 6, 7, 8 +[47] Yunzhi Yan, Zhen Xu, Haotong Lin, Haian Jin, Haoyu Guo, Yida Wang, Kun Zhan, Xianpeng Lang, Hujun Bao, Xiaowei Zhou, et al. Streetcrafter: Street view synthesis with controllable video diffusion models. arXiv preprint arXiv:2412.13188, 2024. 2, 4 +[48] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 2 +[49] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, 2023. 2, 6, 7 +[50] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 2, 5 +[51] Zeyu Yang, Zijie Pan, Yuankun Yang, Xiatian Zhu, and Li Zhang. Driving scene synthesis on free-form trajectories with generative prior. arXiv preprint arXiv:2412.01717, 2024. 2, 4 +[52] 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. 3 +[53] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, 2024. 2 +[54] Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming + +Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. arXiv preprint arXiv:2403.20079, 2024. 2, 3 +[55] Dimitris Zermas, Izzat Izzat, and Nikolaos Papanikolopoulos. Fast segmentation of 3d point clouds: A paradigm on lidar data for autonomous vehicle applications. In ICRA, 2017. 4 +[56] 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. 6 +[57] Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, Wenjun Mei, and Xingang Wang. Drivedreamer4d: World models are effective data machines for 4d driving scene representation. arXiv preprint arXiv:2410.13571, 2024. 2, 3, 4, 5, 6 +[58] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: LIm-enhanced world models for diverse driving video generation. arXiv preprint arXiv:2403.06845, 2024. 2, 3 +[59] Hongyu Zhou, Longzhong Lin, Jiabao Wang, Yichong Lu, Dongfeng Bai, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugsim: A real-time, photo-realistic and closed-loop simulator for autonomous driving. arXiv preprint arXiv:2412.01718, 2024. 3 +[60] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In CVPR, pages 21634-21643, 2024. 2 +[61] Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Ni-anchen Deng, Min Dou, Yuqi Wang, Botian Shi, Kai Wang, Chi Zhang, et al. Is sora a world simulator? a comprehensive survey on general world models and beyond. arXiv preprint arXiv:2405.03520, 2024. 3 \ No newline at end of file diff --git a/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/images.zip b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c67a245d044b0d5ad36a01b5e1d606c5a9e85637 --- /dev/null +++ b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f8774e570500caf6b2a31acd6dc00eb18db0b8493f6f2dd0172027ed55a6ac2 +size 689349 diff --git a/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/layout.json b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..629535fb3fe73f25c573ede0519646fdef2d6ce4 --- /dev/null +++ b/ICCV/2025/ReconDreamer++_ Harmonizing Generative and Reconstructive Models for Driving Scene Representation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54580f7b4424800d31f62f198779aa93e4563bd2999b0bc460399e91207ac241 +size 383045 diff --git a/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/b9ad909b-077c-4950-9bf7-53f3aac3aea6_content_list.json b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/b9ad909b-077c-4950-9bf7-53f3aac3aea6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4b70338905149bb4e79af21f7c67adbea977cd17 --- /dev/null +++ b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/b9ad909b-077c-4950-9bf7-53f3aac3aea6_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9a9731d1acd501de44ac31a42d945015d3b9269929fec144c19ce5b7ba4ea1a +size 82032 diff --git a/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/b9ad909b-077c-4950-9bf7-53f3aac3aea6_model.json b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/b9ad909b-077c-4950-9bf7-53f3aac3aea6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..af4e9ea09d1bea71251d8c6b7be5c6fcefdbddac --- /dev/null +++ b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/b9ad909b-077c-4950-9bf7-53f3aac3aea6_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a24bcd0efead8b9a12a9016b41b813bbf57e3611c97d1edd239f82c6161e0b3e +size 107291 diff --git a/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/b9ad909b-077c-4950-9bf7-53f3aac3aea6_origin.pdf b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/b9ad909b-077c-4950-9bf7-53f3aac3aea6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9696b6782053ec5f9a7007c4f9abe8668d930c57 --- /dev/null +++ b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/b9ad909b-077c-4950-9bf7-53f3aac3aea6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b040a65f78a93924fba8778f09f093b5e333f51bc06d90390157c14fa71d9e8b +size 3233499 diff --git a/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/full.md b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..673cf58fec7fd03af799012517f2c9e116b9b106 --- /dev/null +++ b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/full.md @@ -0,0 +1,343 @@ +# Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior + +Renzhi He1, Haowen Zhou2, Yubei Chen1, Yi Xue1 + +1 University of California, Davis 2 California Institute of Technology + +{cubhe, yxxue, ybchen}@ucdavis.edu, hzhou7@caltech.edu + +# Abstract + +Volumetric reconstruction of label-free living cells from non-destructive optical microscopic images reveals cellular metabolism in native environments. However, current optical tomography techniques require hundreds of 2D images to reconstruct a 3D volume, hindering them from intravital imaging of biological samples undergoing rapid dynamics. This poses the challenge of reconstructing the entire volume of semi-transparent biological samples from sparse views due to the restricted viewing angles of microscopes and the limited number of measurements. In this work, we develop Neural Volumetric Prior (NVP) for high-fidelity volumetric reconstruction of semi-transparent biological samples from sparse-view microscopic images. NVP integrates explicit and implicit neural representations and incorporates the physical prior of diffractive optics. We validate NVP on both simulated data and experimentally captured microscopic images. Compared to previous methods, NVP significantly reduces the required number of images by nearly 50-fold and processing time by 3-fold while maintaining state-of-the-art performance. NVP is the first technique to enable volumetric reconstruction of label-free biological samples from sparse-view microscopic images, paving the way for real-time 3D imaging of dynamically changing biological samples. Project Page + +# 1. Introduction + +Tomography is widely used to reconstruct 3D biomedical structures from multiple 2D image views. Computed Tomography (CT), for instance, produces images of the human body based on varying absorption of X-ray by different tissues, while optical tomography produces microscale 3D images of label-free cellular structures based on varying diffraction of visible light by intracellular structures. Numerous advanced optical tomography systems and computational algorithms [1-16] have been developed for cultured cells or thin tissue slices, where 3D intracellular structures + +![](images/7b2a45765426088c383aa150629635243955d5c1a84c0377217f035ed3e2e899.jpg) +a. Explicit +Non-parametric method + +$$ +\hat {n} = W _ {e n r} (x, y, z) +$$ + +![](images/ab468df517307301eaa6b1ad82a03cda40cbe316169ccefc369fd8bfc7069eb9.jpg) +b. Implicit +Fully parametric method + +$$ +\hat {n} = F _ {i n r} (x, y, z) +$$ + +![](images/1928339b6664d492fede52c6ddc15ba1c9594c33548acec5fa257b5a2f5ea356.jpg) +c. Triplane +Non-parametric for surface +parametric +Figure 1. Illustrations of different neural representation methods. (a) Explicit neural representation: A non-parametric method in which the refractive index (RI) distribution $\hat{n}$ is directly reconstructed by the projection function $W_{\mathrm{enr}}(x,y,z)$ , which provide a one-to-one mapping from spatial coordinates to RI values [17, 18]. (b) Implicit neural representation: Instead of directly reconstructing $\hat{n}$ , this method optimizes the parameters of a multi-layer perceptron (MLP) $F_{\mathrm{inr}}$ , which is then used to predict $\hat{n}$ [19]. (c) Triplane: A hybrid method combining non-parametric and parametric components by solving both the triplane features $W_{xy}, W_{xz}, W_{yz}$ and the neural model $F_{\mathrm{tri}}$ to reconstruct $\hat{n}$ [20]. (d) NVP: Our proposed hybrid approach reconstructs $\hat{n}$ by integrating an uncompressed volumetric prior $W_{xyz}$ into the neural model $F_{\mathrm{nvp}}$ . + +$$ +\hat {n} = F _ {t r i} \left(W _ {x y} (x, y) + W _ {x z} (x, z) + W _ {y z} (y, z)\right) +$$ + +![](images/12bab2901aa9cadd6313d87b49c465da9042b683ab458e7587d5b957a2c61de0.jpg) +d. Neural volumetric prior +Non-parametric for volumetric +parametric + +$$ +\hat {n} = F _ {n v p} \left(W _ {x y z} (x, y, z)\right) +$$ + +are reconstructed from diffracted laser or LED light as it transmits through the sample with heterogenous refractive index (RI). However, when the sample is too thick for visible light to transmit through, such as intact, bulky specimens or living animals, these imaging techniques cannot reconstruct 3D structures because the transmission side is inaccessible. To address this challenge, Fluorescence Diffraction Tomography (FDT) [17, 18] has been developed as a promising technique for intravital optical tomography. It reconstructs 3D label-free structures from diffracted fluo + +rescence without requiring access to the transmission side, using intrinsic or exogenous fluorescence within the sample as internal light sources. Despite its potential, FDT still requires hundreds of 2D images to reconstruct a 3D volume, requiring that the object remain static during image acquisition (seconds to minutes). This limitation restricts FDT's application in capturing dynamic cellular or tissue structural changes in real time, where high-speed image acquisition is essential [21]. + +To achieve high-speed volumetric imaging with FDT, it is necessary to reduce the required number of 2D images while preserving the accuracy and resolution of the reconstructed 3D volumes. Compared to 3D rendering of natural scenes, rendering semi-transparent biological samples from diffracted microscopic images is more challenging due to fewer measurements and more unknowns. First, the range of viewing angles in microscopy is constrained by the numerical aperture of the system and the fixed position of the camera, unlike natural scenes that can be imaged from full 360-degree perspectives. Second, the spatial distribution of fluorescence sources in real samples further limits the number of available viewing angles, resulting in a "sparse view" setting that complicates volumetric reconstruction. Third, microscopic tomography aims to reconstruct the entire volume of semi-transparent biological samples, unlike surface reconstruction of opaque objects in nature scenes, leading to a much larger number of unknown voxels. + +To solve the unique challenges of volumetric rendering with microscopic images, the unknown voxels can be represented by both implicit [22] and explicit [23] neural fields [18, 19, 24-26] (Fig. 1a,b) and reconstructed from multiple 2D images. However, previous approaches [18] require hundreds of 2D images for the volumetric rendering, limiting their application in reconstructing dynamic structures from sparse views. Even though 3D rendering of the surface of opaque nature scenes from sparse views has been achieved by many methods [27-32], they cannot directly apply to volumetric reconstruction of semi-transparent biological samples. One potential approach to reduce the requires number of 2D images for biological imaging is to employ the network's structural prior with radial fields [33], such as triplane and hexplane methods [20, 34], which has been applied for MRI imaging [35]. However, because the low-rank decomposition, it can introduce grid artifacts when representing volumetric structures.(Fig. 1c). + +In this work, we develop NVP (Fig. 1d), a hybrid neural representation model that incorporates explicit neural representations and implicit neural fields as neural volumetric priors to achieve physically accurate volumetric reconstruction of label-free, semi-transparent biological samples from sparse views. The NVP regularize the inverse problem to enable accurate reconstruction of large-scale volumes from only $1.5\%$ of the 2D images required by previous work [18]. + +The key innovations of our method are: + +i) We demonstrate, for the first time, the capability to reconstruct volumetric RI of semi-transparent biological samples from diffracted fluorescence images with limited angles and sparse views, validated through both simulations and real-world experiments. This work opens up a new avenue in diffraction-informed neural volumetric representations. +ii) We demonstrate the effectiveness and efficiency of NVP in optical tomography, reducing the required number of images by nearly 50-fold and processing time by 3-fold compared to previous methods [18] in our demonstrated experiments. +iii) We leverage the physical prior of light diffraction to achieve physically accurate rendering and quantitative RI reconstruction of volumetric objects, overcoming the limitations of ray-optics models at microscale imaging and broadening the applicability of neural fields in microscopic imaging. + +# 2. Related Work + +# 2.1. 3D Reconstruction from Sparse Viewpoints + +To accurately reconstruct 3D volumes from sparse views, various priors have been explored, such as local geometry [36-38], multiview consistency [39-42], signed distance functions [43, 44], depth regularization [45, 46], and diffusion-based probabilistic generation [47, 48]. Even though these approaches produce high-accuracy 3D nature scenes [49-53], they model light propagation using ray optics, assuming that light travels in straight lines. While this assumption is valid for natural scenes at the macroscale, it becomes inaccurate at the microscale — where object sizes are comparable to the wavelength of light — as diffraction effects pose challenges in applying these priors to microscopy. To overcome this challenge, we propose developing a physics-informed rendering equation based on diffraction optics for high-accuracy reconstruction of microscale objects from microscopic images. + +# 2.2. 3D Reconstruction using Diffraction Optics + +Cameras can directly measure the intensity of light but not its phase, whereas heterogeneous RI alters the phase of light. To reconstruct the RI of transparent objects, many computational models based on diffractive optics (i.e., wave optics) have been developed. The most classic model is the Born approximation [54], which applies in the weak scattering regime. For the multiple scattering regime, multi-layer Born approximation model has been developed for 3D phase microscopy [55]. Recently, several deep learning-based algorithms have been applied to phase recovery [56, 57], including CNN [58-60], implicit neural fields [19], and explicit neural fields [18]. However, most + +![](images/d313db835652b65f611c529dfaadf12a884646a4f251f7eb9034860c99247b4b.jpg) +Figure 2. Overview of our method for 3D reconstruction. a. Neural volumetric prior (NVP): Predefined 3D grids are reshaped and processed by MLPs to generate the predicted 3D RI volume $\hat{n}$ . b. Multi-slice rendering equation: The multi-slice model calculates light propagation through the volumetric sample from the fluorescence sources (white spots at the bottom) by accounting for light diffraction at each slice. Each illumination configuration of the fluorescence sources (e.g., $F_{1}, F_{i}, F_{n}$ ) interacts with the volume to produce a corresponding set of rendered images $(\hat{I}_{1}, \hat{I}_{i}, \hat{I}_{n})$ . The illumination configurations are jointly optimized with the RI volume. c. Loss functions: The predicted images $\hat{I}_{\mathrm{pred}}$ are aligned with ground truth images $I_{\mathrm{GT}}$ and masked based on light coherence, resulting in $\hat{I}_{\mathrm{mask}}$ and $I_{\mathrm{mask}}$ . Loss functions, including $L1$ , $L2$ , and SSIM, are calculated over the masked regions of $\hat{I}_{\mathrm{pred}}$ and $I_{\mathrm{GT}}$ , and a total variation (TV) regularizer $R(\hat{n})$ is applied. + +of these models are designed for transmission geometry under coherent plane wave illumination (except [18]), whereas FDT uses partially coherent spherical wave illumination. Furthermore, both image acquisition and reconstruction are time-consuming, as hundreds of images from varying viewpoints must be acquired and processed. To address these challenges, we propose to develop a hybrid representation model with a multi-layer Born approximation to reconstruct the 3D RI from sparse viewpoints. + +# 3. Method + +# 3.1. Prior in Neural Representation + +# 3.1.1. Preliminaries + +Implicit Neural Representation Implicit neural representations [22] are a type of parametric model that employ a multi-layer perceptron (MLP) to reconstruct 3D properties such as color and density to coordinate systems. These methods have achieved remarkable success across multiple domains, including graphic rendering and 3D scene reconstructions [61-63]. Recently, implicit neural representations have also been utilized to enhance microscopic volume reconstructions [19, 64]. Here, we represent the estimated RI of sample $\hat{n}$ by $\hat{n}(x,y,z) = F_{inr}(x,y,z)$ . where $F_{inr}$ denotes the learnable network. That is, the MLP maps a 3-D spatial coordinate directly to a scalar RI value. + +Explicit Neural Representation Explicit neural representations are a non-parametric model first introduced by Plenoxel [23]. This approach directly solves object + +properties at predefined grid points $W_{\mathrm{enr}}$ without an MLP: $\hat{n}(x,y,z) = W_{\mathrm{enr}}(x,y,z)$ . This non-parametric strategy is computationally efficient but weakly encodes spatial correlations, hence typically needs more views for high-fidelity results. + +Triplane Representation Triplane representation [20, 34] models 3D objects using a low rank grid consisting of three orthogonal feature planes $(W_{xy}, W_{xz}, W_{yz})$ , then fused by a shallow MLP $F_{\mathrm{tri}}: \hat{n}(x,y,z) = F_{\mathrm{tri}}(W_{xy}(x,y) + W_{xz}(x,z) + W_{yz}(y,z))$ . This hybrid design is memory-efficient and excels in surface reconstruction [20, 65-67], though faithfully representing fine volumetric details at the microscale remains challenging. + +# 3.1.2. Neural Volumetric Prior + +![](images/65b398169d55146a0cf3a0243dbe46557218f1d7ee55b788f89007998b7ef031.jpg) +Figure 3. NVP architecture. Adaptive feature grids $(H, W, L, F)$ are first reshaped into feature vectors of size $(N, F)$ , where $H, W, L$ are the volume dimensions and $F$ is the feature dimension. These vectors are then passed through MLP layers to predict the RI of each voxel. + +To enhance the representation capability, we develop a + +hybrid model, NVP, that integrates explicit and implicit representations, enabling the network to capture implicit spatial correlations while retaining the sparsity prior of the explicit grid (Fig. 2a, Fig. 3). The explicit representations store the unknown features of an imaging object in a non-parametric, structural grid, providing direct access to the features at each sampled voxel. The feature grid is adaptive—that is, its resolution dynamically adjusts according to the spatial varying distribution of the RI. Meanwhile, the implicit representations extract dependencies of the features from the explicit neural fields, refining the reconstructed RI and filling in missing values at unmeasured voxels. This hybrid strategy decouples reconstruction quality (i.e., resolution, accuracy) from the number of measurements, effectively overcoming overfitting issues due to sparse views and significantly reducing computational time. + +Specifically, we first define a sparse 3D grid for the object with randomly initialized features at each voxel, which are then mapped to corresponding RI by a three-layer MLP (Fig. 2a, Fig. 3): + +$$ +\hat {n} (x, y, z) = F _ {\mathrm {n v p}} \left(W _ {x y z} (x, y, z)\right), \tag {1} +$$ + +where the non-parametric grid $W_{xyz}(x,y,z) \in \mathbb{R}^F$ stores a learnable $F$ -dimensional feature vector at each voxel, and the MLP $F_{\mathrm{nvp}}$ converts each vector into a single RI value. + +The output of the MLP is a volumetric object with estimated RI, which becomes the input to the physics-informed rendering equation, along with self-calibrated fluorescent illumination (Fig. 2b). The loss function is calculated between the predicted and measured images (Fig. 2c). Details of the rendering equation and loss function are presented in the following subsections. + +# 3.2. Physical prior in the Rendering Equation using Diffraction Optics + +# How is physically accurate 3D RI reconstruction of microscale objects achieved from sparse views? + +We model the rendering process using a multi-slice Born approximation that accounts for light diffraction at the microscale [17, 55, 68-70]. In diffractive optics, when light propagates through a transparent sample with spatially varying RI (Figure 2b), each voxel alters the phase of the wavefront. As light continues to propagate, these phase modulations accumulate and interfere, resulting in diffraction patterns that carry information about the 3D RI distribution. The rendering equation describes this process by modeling the imaging volume as a stack of $N_z$ thin slices with spatially varying RI. The light field $\hat{E}_{k,i}(\mathbf{r})$ at $k^{th}$ slice illuminated by the $i^{th}$ fluorescence source is defined as: + +$$ +\hat {E} _ {k, i} (\mathbf {r}) = \mathcal {P} _ {\Delta z} \left\{t _ {k} (\mathbf {r}) \cdot \hat {E} _ {k - 1, i} (\mathbf {r}) \right\}, \tag {2} +$$ + +where $\mathcal{P}_{\Delta z}$ denotes the propagation operator, and $t_k(\mathbf{r})$ is the transmission function of the $k^{th}$ slice and is related to + +its RI. After transmitting through the final slice, the accumulated light field $\hat{E}_{N_z}(\mathbf{r})$ is captured by a camera as an intensity-only image $\hat{I}_i(\mathbf{r})$ : + +$$ +\hat {I} _ {i} (\mathbf {r}) = \left| \hat {E} _ {N _ {z}, i} (\mathbf {r}) \right| ^ {2}. \tag {3} +$$ + +The intensity images under illumination from each fluorescence source $\hat{I}_1, \hat{I}_2, \dots, \hat{I}_n$ are computed using this procedure, which are the sparse views. Further details on the mathematical formulation are provided in the Supplementary 1. + +We implemented this diffraction-based multi-slice rendering equation on a GPU (see details in Section 4), achieving fast rendering through parallel computing and predefined propagation kernels. This rendering equation reforms conventional ray optics by introducing a more accurate, physics-informed prior while maintaining comparable computational efficiency. + +# 3.3. Essential Module for Processing Experimentally Captured Data + +Coherent Alignment The diffraction of coherent light in the rendering equation is different from the partially coherent or incoherent light (e.g., fluorescence) used in the experiment. To mitigate the model mismatching, we incorporate a coherent mask in the model that accurately and efficiently aligns measured fluorescence images with the model-predicted images. More details are provided in Supplementary 2. + +Self-Calibration of Viewpoints Calibrating viewpoints is crucial [71-73] for 3D rendering with the multi-slice model, as the quality and accuracy of the 3D reconstruction depend on precise localization of the illumination sources. We propose a self-supervised calibration method that estimates fluorescence source positions from the fluorescence images by Gaussian fitting and jointly optimizes them with the MLP parameters to improve 3D reconstruction quality. We have conducted ablation study on the self-calibration of viewpoints. More details are provided in Supplementary 7 and Table 7. + +# 4. Experiments + +# 4.1. Experimental Settings + +Computational setting We train our model on an NVIDIA A6000 GPU with 48 GB memory using the PyTorch framework. The Adam optimizer is used with an initial learning rate of $5 \times 10^{-3}$ , a momentum decay rate of 0.9, and a squared gradient decay rate of 0.99. The batch size is set to 10. For the results presented in this paper, the typical computation time for 3D RI reconstruction is around 20 minutes. More details on the implementation are provided in Supplementary 6. + +![](images/b858aa33f30c4971e8c9b9ea066dd6bb2a6ecd491af242d88618a39eb4e8292b.jpg) +a. 3D refractive index + +![](images/c207d2e89165d405a0149818a64ff4f589a9570b8a3303797fff9370c52ed1a1.jpg) +b. Images + +![](images/e0db0db8b45c0e555aaef968a3456b9aaa6d00b8e2b10740f749cf39fa75d26a.jpg) + +![](images/0f0e3d081f71c5b3d185d13823b0436c314e931eead453af862e24dbfd9a5d6e.jpg) + +![](images/84f661cbe35baeda4cda1bb4f193471444a4b1d0e83ab30eea718ab7ad1be439.jpg) + +![](images/7bc5e40817ebc8177e95b2196343da27cc24e5e406748544c4559a392c1e9c41.jpg) + +![](images/0f17dd7dd6c1f3988a433c09afe0d3b7d6043fc1ba6f3de14fe5b5d1c5e2cbeb.jpg) + +![](images/7efadda9d7917f2a59134e954462bdbd40d32b249c72302b1a856e9a0286235e.jpg) + +![](images/088033f5610411cc0047d6ad2f586f94693b38e73910257d018a42c685fc9b16.jpg) +c. Different cross-sections of reconstructed 3D refractive index +Figure 4. NVP reconstructs the 3D RI of the synthetic cell dataset with micron-scale resolution and excellent optical sectioning. (a) Reconstructed 3D RI distribution showing the spatial distribution of cells within a $154\mu \mathrm{m}\times 154\mu \mathrm{m}\times 24\mu \mathrm{m}$ volume. (b) Predicted fluorescence images under the illumination from two representative fluorescence sources, demonstrating the accuracy of the reconstructed model compared to the ground-truth data. (c) Cross-sectional comparison of RI reconstructed by NVP (top) and ground-truth (bottom) along the axial direction, illustrating consistency between the reconstructed and ground-truth RI distribution across the sample volume. Color bars indicate intensity ranges for predicted images, absolute errors, and RI values. + +Image Acquisition We implement the optical setup as described in [18, 74, 75]. The imaging objects are label-free cultured cells in a petri dish coated with fluorescent paint on the outer surface. Fluorescence from a single spot is excited by focusing a laser beam, which is modeled as a point source. The excitation positions are randomly selected. As the fluorescence propagated through the cells, it is diffracted and captured by a high-sensitivity sCMOS camera (Kinetix22). + +Loss and Metrics Design We formulate the loss function to minimize the difference between the predicted and ground-truth images, while applying a regularization term on the predicted 3D RI to promote smoothness and preserve fine structural details. The overall loss is expressed as $\mathcal{L} = \mathcal{L}_{\mathrm{img}} + \tau \mathcal{R}_{\mathrm{ri}}$ , where $\mathcal{L}_{\mathrm{img}}$ combines multiple image-level loss terms such as $L_{1}$ , $L_{2}$ , $SSIM$ , and $\mathcal{R}_{\mathrm{ri}}$ is the loss associated with total-variation regularization (Supplementary 4). + +# 4.2. Data Preparation + +We evaluate our method on two in-house synthetic 3D datasets: synthetic cells and synthetic tissue. The synthetic cells are designed to assess the resolution across varying RI and the optical sectioning ability across different axial planes (Fig. 4). Each synthetic cell is modeled as a 3D Gaussian function $\mathcal{N}(\mu, \sigma)$ , where $\mu$ denotes its 3D position and $\sigma$ denotes its radius ( $\sigma = 5 - 10$ pixels). The synthetic cells are randomly distributed within a volume of $512 \times 512 \times 12$ voxels with a voxel size of $\{0.3, 0.3, 2\} \mu m$ along the $x, y, z$ axes. The synthetic tissue dataset consists of blood vessels and neurons spanning a 3D volume and is designed to assess reconstruction performance across different morphological structures (spatial frequencies). The 3D volume consists of $512 \times 512 \times 15$ voxels, with a voxel size of $\{0.9, 0.9, 3\} \mu m$ . In both datasets, 100 ground-truth fluorescence images ( $512 \times 512$ pixels) are generated by the rendering equation with the illumination of 100 distinct + +![](images/8e7e6b46a1a2d7ca8667044ab25e128f5103e851b61b5b848ca52ebaf919d73a.jpg) +Figure 5. NVP achieves more accurate reconstruction compared to the explicit and triplane methods. The top row shows the RI maps produced by each method, with the ground-truth RI map included for reference. The bottom row displays the corresponding error maps, highlighting the reconstruction errors in each method. NVP demonstrates superior performance, producing smoother reconstructions and lower errors that closely match the ground truth. + +
Synthetic cells6 illuminations7 illuminations20 illuminations
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
ExpRI28.62030.85400.182128.72680.84690.177728.88030.86500.1188
IMG58.25880.99990.000157.35990.99990.000255.42240.99960.0003
TriRI28.72880.71280.215230.43420.76190.138530.61490.96210.0620
IMG55.40190.99970.000356.16540.99970.000247.16650.99700.0028
OursRI30.73180.89100.103430.96430.89660.089731.37870.89590.0543
IMG71.04921.00000.000069.97481.00000.000068.62981.00000.0000
+ +Table 1. Reconstruction quality metrics for synthetic cells dataset across different illuminations and various methods in RI and predicted image results. + +
Synthetic tissue7 illuminations
PSNR ↑SSIM ↑LPIPS ↓
ExpRI11.61440.29540.5676
IMG35.73190.91680.0329
TriRI9.06320.13230.7122
IMG46.48420.99560.0040
NVPRI11.64120.47750.5038
IMG62.79140.99990.0001
+ +Table 2. Reconstruction quality metrics for synthetic tissue. + +fluorescence sources. It is worth noting that NVP typically uses only 6-7 fluorescence images randomly selected from this set. + +# 4.3. Comparison of Neural Representation Methods + +Baseline Since our task is to reconstruct 3D RI from microscopic images using a diffraction-based optical model, few existing methods are available for direct comparison. We implement the explicit approach [18] as a baseline in our experiment, as it is the only closely related method. Furthermore, we adapt the triplane approach to this task for comparison. + +Quantitative Reconstruction Results We first assess the resolution and optical sectioning capability of NVP by reconstructing the 3D RI of synthetic cells (Figure 4). The synthetic cells, with sizes ranging from $3.53\mu m$ to 7.05 $\mu m$ (full-width-half-maximum), are randomly distributed within a 3D volume. The NVP model is configured with a feature dimension of 16 and 6 network layers (see details in Supplementary Section 6, Table 5, and Table 6). In the reconstructed volume (Figure 4a), all cells are clearly distinguishable from one another and from the background, indicating the high-resolution reconstruction of RI. The predicted fluorescence images also closely match the ground + +truth images (Figure 4b). Furthermore, the reconstructed RI on each $Z$ -plane is nearly identical to the ground-truth RI at the corresponding plane, indicating excellent optical sectioning performance (see Supplementary Section 3 for details). These results demonstrate that NVP accurately reconstructs high-resolution 3D RI. + +Next, we quantitatively compare the reconstruction results of the synthetic cells using different neural representation methods with the ground-truth RI (Figure 5). The RI reconstructed by NVP is more accurate and contains fewer artifacts compared to the results from the explicit and triplane methods, especially in fine structures. The error maps further support this conclusion, showing that NVP yields notably lower errors, particularly in dense or complex regions, indicating its ability to preserve fine details while minimizing noise. + +To evaluate NVP's reconstruction performance across different morphological structures, we reconstruct the RI of synthetic tissue using NVP and compare the results with those from the explicit and triplane methods, as well as the ground truth (Figure 6). The explicit method exhibits structural artifacts and lacks detail in complex areas, while the triplane method reduces some artifacts but introduces grid effects, particularly in dense regions. In contrast, NVP + +![](images/b24d6610ab4bf52577121af1617ba1a155dbf5afd7bffbe718e07a89cd02d444.jpg) + +![](images/9d13f197778e01a542159b84f6cf0d0587a164c3cee080784ce561be3103c6ed.jpg) + +![](images/e0265ed846fb9e56983537043a0a203c89bbe5ef51703e3a1c2f80f1e148527a.jpg) +Figure 6. NVP reconstructs the RI of varying morphological structures in the synthetic tissue dataset and outperforms the explicit and triplane methods. The synthetic tissue contains both continuous structures (blood vessels, red) and sparse structures (neurons, blue). Left: the 3D view of the reconstructed RI. Right: a representative $Z$ -slice and zoomed-in views of the reconstructed RI. Scale bars denote $50~\mu \mathrm{m}$ . + +![](images/323a9b0e91cfbe2174a0fd3f6833c26afc1d00235e3fff91d69d41504ae1c616.jpg) + +closely matches the ground truth, achieving a significantly higher SSIM of 0.4775 for RI, compared to 0.2954 for explicit method and 0.1323 for triplane method (Table 2). NVP preserves fine details and minimizes artifacts in both sparse and continuous structures, demonstrating superior robustness and accuracy. These results highlight NVP's effectiveness for high-fidelity reconstruction. + +# 4.4. Effect of Varying Number of Views + +To evaluate reconstruction from sparse views, we reconstruct the RI of synthetic cells using 6, 7, and 20 fluorescence images with NVP, the explicit method, and the triplane method. We quantitatively evaluate the results by comparing PSNR, SSIM [76], and LPIPS [77] for both the RI and the predicted images (Table 1). With as few as 6 fluorescence images, NVP outperforms the other methods in RI reconstruction, achieving a PSNR of 30.73, SSIM of 0.8910, and LPIPS of 0.1034, demonstrating strong structural fidelity and perceptual quality. When the number of fluorescence images is reduced from 20 to 6, the performance of NVP shows minimal performance degradation, with only a 0.65 decrease in PSNR and a 0.005 decrease in SSIM. Furthermore, NVP with 6 images achieves results comparable to explicit and triplane methods using 20 illuminations. For predicted images, NVP consistently achieves near-perfect scores, with SSIM reaching 1.0000 and LPIPS approaching zero across all cases, indicating excellent alignment with the ground truth. We also evaluate + +the PSNR, SSIM, and LPIPS for reconstructed synthetic tissue with these methods and observe similar conclusions (Table 2, Table 4). More details are provided in the Supplementary 5. These results highlight NVP's robustness to sparse views, making it highly suitable for applications with limited data acquisition, such as imaging cardiomyocyte contractions and embryo development. The ability to achieve high-quality 3D reconstruction with fewer 2D images enables faster imaging and reduces photodamage to biological specimens. + +# 4.5. Experiments on Real Biological Samples + +We apply the explicit, triplane, and NVP methods to reconstruct the 3D RI of living MDCK cells from 19 fluorescence images captured with our custom-built microscope system described in Section 4.1 (Figure 7, Table 3). Among the three methods, NVP reconstructs cellular structures more accurately than the explicit and triplane methods, since it uses the full-dimensional representatives informed by prior knowledge of features. The RI reconstructed by the explicit method shows discontinuities and noise due to absence of prior knowledge of data distribution, making it more prone to local minima. Meanwhile, the triplane method displays severe grid artifacts, as its three-plane decomposition limits its ability to reconstruct high-dimensional representation. This limitation has been observed in previous applications, such as MRI volume reconstruction [35]. The predicted images and error maps further confirm that NVP sur + +![](images/cc27bafc5d99bbcdae8d161c799efee0a7ebb48bda19c33e04c86347b0441397.jpg) +Figure 7. NVP reconstructs the 3D RI of living MDCK cells with superior accuracy and fewer artifacts compared to the results from the explicit and triplane methods. RI maps (top row), predicted images (bottom row, left), and error maps (bottom row, right) are shown for the explicit method, the triplane method, and NVP. The zoomed-in views (purple and green boxes) of the RI maps are shown in the top row, right. Scale bars denote $20\mu \mathrm{m}$ . + +![](images/7c9d3abefd6877c48b6cbc5504424c9917dba1650b5e294ec748f673c8bdb60a.jpg) + +![](images/5609c870e8090742a115fddcc365958bcb67925917f3a533bf63c2cece235d9c.jpg) + +passes both baselines in the accuracy of predicted images. Quantitative metrics show that NVP achieves the highest SSIM (0.9977) and PSNR (40.7043), and the lowest LPIPS (0.0015), indicating superior structural accuracy and perceptual fidelity compared to the other methods (Table 3). While the explicit method slightly outperforms the triplane method in structural metrics, it has higher overall errors. In summary, these results demonstrate that NVP is able to quantitatively reconstruct the 3D RI of real biological samples from limited fluorescence images with high accuracy and robustness. At the same resolution, our method converges in 20 minutes to SSIM $>0.99$ , whereas the explicit baseline [18] requires 60 minutes. + +
MethodsSSIM↑LPIPS↓PSNR↑
Explicit0.99440.005136.6540
Triplane0.97620.028532.6921
NVP0.99770.001540.7043
+ +Table 3. Comparison of NVP, explicit, and triplane methods based on the quality of predicted images of MDCK cells. + +# 5. Conclusion and Discussion + +We present a new neural representation method, NVP, for efficient and accurate volumetric RI representation from limited and sparse-view microscopic images acquired using FDT. NVP uses hybrid representations by integrating the explicit adaptive feature grid with nonlinear MLP to enhance the quality and efficiency of 3D reconstruction of the RI spatial distribution. NVP also incorporates a rendering equation based on diffraction optics, which accurately models phase changes as light interacts with biological sam + +plies. This enables high-fidelity, high resolution rendering of microscale structures and quantitative volumetric reconstruction, while maintaining comparable processing time. To address model mismatches when processing experimentally captured images, we incorporate coherent alignment and self-calibration. We validate NVP using experiments on both synthesized data and real biological samples. NVP demonstrates superior performance over the explicit and triplane methods, achieving a 50-fold reduction in the number of required measurements and a 3-fold increase in processing speed compared to previous work [18]. These results highlight NVP's potential for high-fidelity, high-speed volumetric reconstruction using FDT. + +Compared to the explicit method, NVP introduces a neural prior that is significantly more effective than the sparse prior used in the explicit method. Compared to the triplane method, NVP retains the full representation dimensions, allowing it to better preserve high-dimensional representations and produce fewer artifacts. NVP can be extended to other imaging modalities and datasets by adjusting the underlying physics model. Its performance may be further improved by integrating 3D convolution kernels to capture spatial relationships and off-grid features, or leveraging data priors such as masked auto-encoders [78] or diffusion models [79, 80]. + +# 6. Acknowledgement + +We thank Dr. Sara Fridovich-Keil at Georgia Institute of Technology for insightful discussions on representation methods. We thank Ziqiao Ma at the University of Michigan for proofreading the manuscript. Research reported in this publication was supported by the National + +Institute of General Medical Sciences of the National Institutes of Health 1R35GM155193-01 and the National Science Foundation CAREER Award 2443604 to Dr. Yi Xue. This work was also supported by Dr. Yi Xue's startup funds from the Department of Biomedical Engineering at the University of California, Davis. + +# References + +[1] YongKeun Park, Gabriel Popescu, Kamran Badizadegan, Ramachandra R. Dasari, and Michael S. Feld. Diffraction phase and fluorescence microscopy. Opt. Express, 14(18):8263-8268, Sep 2006. 1 +[2] Kyoohyun Kim, Wei Sun Park, Sangchan Na, Sangbum Kim, Taehong Kim, Won Do Heo, and YongKeun Park. Correlative three-dimensional fluorescence and refractive index tomography: bridging the gap between molecular specificity and quantitative bioimaging. Biomed. Opt. Express, 8(12):5688-5697, Dec 2017. +[3] Etienne Shaffer, Nicolas Pavillon, and Christian Depeursinge. Single-shot, simultaneous incoherent and holographic microscopy. Journal of microscopy, 245(1):49-62, 2012. +[4] Baptiste Marthy, Maëlle Bénéfice, and Guillaume Baffou. Single-shot quantitative phase-fluorescence imaging using cross-grating wavefront microscopy. Scientific Reports, 14(1):2142, 2024. +[5] Xiangyu Quan, Manoj Kumar, Sudheesh K. Rajput, Yosuke Tamada, Yasuhiro Awatsuji, and Osamu Matoba. Multimodal microscopy: Fast acquisition of quantitative phase and fluorescence imaging in 3d space. IEEE Journal of Selected Topics in Quantum Electronics, 27(4):1-11, 2021. +[6] Shilpa Tayal, Veena Singh, Tejinder Kaur, Neetu Singh, and Dalip Singh Mehta. Simultaneous fluorescence and quantitative phase imaging of mg63 osteosarcoma cells to monitor morphological changes with time using partially spatially coherent light source. Methods and Applications in Fluorescence, 8(3):035004, 2020. +[7] Sudheesh K Rajput, Osamu Matoba, Manoj Kumar, Xiangyu Quan, Yasuhiro Awatsuji, Yosuke Tamada, and Enrique Tajahuence. Multi-physical parameter cross-sectional imaging of quantitative phase and fluorescence by integrated multimodal microscopy. IEEE Journal of Selected Topics in Quantum Electronics, 27(4):1-9, 2021. +[8] Yang Liu, Jinli Suo, Yuanlong Zhang, and Qionghai Dai. Single-pixel phase and fluorescence microscope. Opt. Express, 26(25):32451-32462, Dec 2018. +[9] Thanh an Pham, Emmanuel Soubies, Ferreol Soulez, and Michael Unser. Optical diffraction tomography from single-molecule localization microscopy. Optics Communications, 499:127290, 2021. +[10] Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Seungeun Oh, Niyom Lue, Ramachandra R Dasari, and Michael S Feld. Tomographic phase microscopy. Nat. Methods, 4(9):717-719, September 2007. +[11] Yongjin Sung, Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Ramachandra R. Dasari, and Michael S. + +Feld. Optical diffraction tomography for high resolution live cell imaging. Opt. Express, 17(1):266-277, Jan 2009. +[12] Laura Waller, Lei Tian, and George Barbastathis. Transport of intensity phase-amplitude imaging with higher order intensity derivatives. Opt. Express, 18(12):12552-12561, June 2010. +[13] Lei Tian and Laura Waller. 3D intensity and phase imaging from light field measurements in an LED array microscope. Optica, OPTICA, 2(2):104-111, February 2015. +[14] Shwetadwip Chowdhury, Will J. Eldridge, Adam Wax, and Joseph A. Izatt. Structured illumination multimodal 3d-resolved quantitative phase and fluorescence sub-diffraction microscopy. Biomed. Opt. Express, 8(5):2496-2518, May 2017. +[15] Li-Hao Yeh, Shwetadwip Chowdhury, and Laura Waller. Computational structured illumination for high-content fluorescence and phase microscopy. Biomed. Opt. Express, 10(4):1978-1998, Apr 2019. +[16] Dashan Dong, Xiaoshuai Huang, Liuju Li, Heng Mao, Yanquan Mo, Guangyi Zhang, Zhe Zhang, Jiayu Shen, Wei Liu, Zeming Wu, Guanghui Liu, Yanmei Liu, Hong Yang, Qihuang Gong, Kebin Shi, and Liangyi Chen. Super-resolution fluorescence-assisted diffraction computational tomography reveals the three-dimensional landscape of the cellular organelle interactome. Light Sci Appl, 9:11, January 2020. 1 +[17] Yi Xue, David Ren, and Laura Waller. Three-dimensional bi-functional refractive index and fluorescence microscopy (brief). Biomed. Opt. Express, 13(11):5900-5908, Nov 2022. 1, 4 +[18] Renzhi He, Yucheng Li, Junjie Chen, and Yi Xue. Fluorescence diffraction tomography using explicit neural fields, 2024. 1, 2, 3, 5, 6, 8 +[19] Renhao Liu, Yu Sun, Jiabei Zhu, Lei Tian, and Ulugbek S Kamilov. Recovery of continuous 3d refractive index maps from discrete intensity-only measurements using neural fields. Nature Machine Intelligence, 4(9):781-791, 2022. 1, 2, 3 +[20] Sara Fridovich-Keil and Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 1, 2, 3 +[21] Venkatakaushik Voleti, Kripa B Patel, Wenze Li, Cittlali Perez Campos, Srinidhi Bharadwaj, Hang Yu, Caitlin Ford, Malte J Casper, Richard Wenwei Yan, Wenxuan Liang, et al. Real-time volumetric microscopy of in vivo dynamics and large-scale samples with scape 2.0. Nature methods, 16(10):1054-1062, 2019. 2 +[22] 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. 2, 3 +[23] 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 (CVPR), pages 5501–5510, June 2022. 2, 3 + +[24] Haowen Zhou, Brandon Y. Feng, Haiyun Guo, Siyu (Steven) Lin, Mingshu Liang, Christopher A. Metzler, and Changhuei Yang. Fourier ptychographic microscopy image stack reconstruction using implicit neural representations. Optica, 10(12):1679-1687, Dec 2023. 2 +[25] Iksung Kang, Qinrong Zhang, Stella X Yu, and Na Ji. Coordinate-based neural representations for computational adaptive optics in widefield microscopy. Nature Machine Intelligence, pages 1-12, 2024. +[26] Iksung Kang, Hyeonggeon Kim, Ryan Natan, Qinrong Zhang, Stella X Yu, and Na Ji. Adaptive optical correction in in vivo two-photon fluorescence microscopy with neural fields. bioRxiv, pages 2024-10, 2024. 2 +[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2 +[28] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields, 2022. +[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4), July 2022. +[30] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8456-8465, June 2023. +[31] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3295–3306, October 2023. +[32] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields, 2023. 2 +[33] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 2 +[34] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. CVPR, 2023. 2, 3 +[35] Annesha Ghosh, Gordon Wetzstein, Mert Pilanci, and Sara Fridovich-Keil. Volumetric reconstruction resolves off-resonance artifacts in static and dynamic propeller mri, 2023. 2, 7 +[36] Kevin Raj, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Spurfies: Sparse surface reconstruction using local geometry priors, 2024. 2 +[37] Yuanhao Cai, Jiahao Wang, Alan Yuille, Zongwei Zhou, and Angtian Wang. Structure-aware sparse-view x-ray 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11174-11183, June 2024. + +[38] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs, 2021. 2 +[39] Shoukang Hu, Kaichen Zhou, Kaiyu Li, Longhui Yu, Lanqing Hong, Tianyang Hu, Zhenguo Li, Gim Hee Lee, and Ziwei Liu. Consistentnerf: Enhancing neural radiance fields with 3d consistency for sparse view synthesis, 2023. 2 +[40] Zi-Xin Zou, Weihao Cheng, Yan-Pei Cao, Shi-Sheng Huang, Ying Shan, and Songjie Zhang. Sparse3d: Distilling multiview-consistent diffusion for object reconstruction from sparse views. ArXiv, abs/2308.14078, 2023. +[41] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis, 2021. +[42] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting, 2024. 2 +[43] Xiaoxiao Long, Chu-Hsing Lin, Peng Wang, T. Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. ArXiv, abs/2206.05737, 2022. 2 +[44] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. Sparsecraft: Few-shot neural reconstruction through stereopsis guided geometric linearization, 2024. 2 +[45] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free, 2024. 2 +[46] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction, 2022. 2 +[47] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion, 2023. 2 +[48] 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. 2 +[49] Jiawei Yao, Chen Wang, Tong Wu, and Chuming Li. Geometry-guided ray augmentation for neural surface reconstruction with sparse views. ArXiv, abs/2310.05483, 2023. 2 +[50] Liyang Zhou, Zhuang Zhang, Hanqing Jiang, Han Sun, H. Bao, and Guofeng Zhang. Dp-mvs: Detail preserving multi-view surface reconstruction of large-scale scenes. Remote. Sens., 13:4569, 2021. +[51] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12588-12597, 2022. +[52] Mude Hui, Zihao Wei, Hongru Zhu, Fei Xia, and Yuyin Zhou. Microdiffusion: Implicit representation-guided diffusion for 3d reconstruction from limited 2d microscopy projections, 2024. +[53] Zilong Li, Chenglong Ma, Jie Chen, Junping Zhang, and Hongming Shan. Learning to distill global representation for sparse-view ct, 2023. 2 + +[54] Max Born and Emil Wolf. Principles of Optics: Electromagnetic Theory of Propagation, Interference and Diffraction of Light. Elsevier, June 2013. 2 +[55] Michael Chen, David Ren, Hsiou-Yuan Liu, Shwetadwip Chowdhury, and Laura Waller. Multi-layer born multiple-scattering model for 3d phase microscopy. Optica, 7(5):394-403, May 2020. 2, 4 +[56] Kaiqiang Wang, Li Song, Chutian Wang, Zhenbo Ren, Guangyuan Zhao, Jiazhen Dou, Jianglei Di, George Barbastathis, Renjie Zhou, Jianlin Zhao, et al. On the use of deep learning for phase recovery. Light: Science & Applications, 13(1):4, 2024. 2 +[57] Jonathan Dong, Lorenzo Valzania, Antoine Maillard, Thanhan Pham, Sylvain Gigan, and Michael Unser. Phase retrieval: From computational imaging to machine learning: A tutorial. IEEE Signal Processing Magazine, 40(1):45-57, 2023. 2 +[58] Ulugbek S. Kamilov, Ioannis N. Papadopoulos, Morteza H. Shoreh, Alexandre Goy, Cedric Vonesch, Michael Unser, and Demetri Psaltis. Learning approach to optical tomography. Optica, 2(6):517-522, Jun 2015. 2 +[59] Xiaofeng Wu, Ziling Wu, Sibi Chakravarthy Shanmugavel, Hang Z. Yu, and Yunhui Zhu. Physics-informed neural network for phase imaging based on transport of intensity equation. Opt. Express, 30(24):43398-43416, Nov 2022. +[60] Alex Matlock, Jiabei Zhu, and Lei Tian. Multiple-scattering simulator-trained neural network for intensity diffraction tomography. Opt. Express, 31(3):4094-4107, Jan 2023. 2 +[61] Towaki Takikawa, Joey Litalien, K. Yin, Karsten Kreis, Charles T. Loop, D. Nowrouzezahrai, Alec Jacobson, M. McGuire, and S. Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11353-11362, 2021. 3 +[62] Justus Thies, M. Zollhöfer, and M. Nießner. Deferred neural rendering. ACM Transactions on Graphics (TOG), 38:1 - 12, 2019. +[63] Ziqi Lu, Jianbo Ye, Xiaohan Fei, Xiaolong Li, Jiawei Mo, Ashwin Swaminathan, and Stefano Soatto. Fast sparse view guided nef update for object reconfigurations, 2024. 3 +[64] Iksung Kang, Qinrong Zhang, Stella X. Yu, and Na Ji. Coordinate-based neural representations for computational adaptive optics in widefield microscopy. Nature Machine Intelligence, 6(6):714-725, June 2024. 3 +[65] 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 multiview diffusion using 3d large reconstruction model. ArXiv, abs/2311.09217, 2023. 3 +[66] J. Shue, Eric Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2022. +[67] Anchit Gupta and Anchit Gupta. 3dgen: Triplane latent diffusion for textured mesh generation. ArXiv, abs/2303.05371, 2023. 3 + +[68] Lei Tian and Laura Waller. Quantitative differential phase contrast imaging in an led array microscope. Opt. Express, 23(9):11394-11403, May 2015. 4 +[69] Shwetadwip Chowdhury, Michael Chen, Regina Eckert, David Ren, Fan Wu, Nicole Repina, and Laura Waller. High-resolution 3d refractive index microscopy of multiple-scattering samples from intensity images. Optica, 6(9):1211-1219, Sep 2019. +[70] Michael Chen, David Ren, Hsiou-Yuan Liu, Shwetadwip Chowdhury, and Laura Waller. Multi-layer born multiple-scattering model for 3d phase microscopy. Optica, 7(5):394-403, May 2020. 4 +[71] 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. 4 +[72] 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. +[73] 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. 4 +[74] Yi Xue, Kalen P Berry, Josiah R Boivin, Christopher J Rowlands, Yu Takiguchi, Elly Nedivi, and Peter T C So. Scanless volumetric imaging by selective access multifocal multiphoton microscopy. Optica, 6(1):76-83, January 2019. 5 +[75] Yucheng Li and Yi Xue. Two-photon bi-functional refractive index and fluorescence microscopy (2P-BRIEF). In Three-Dimensional and Multidimensional Microscopy: Image Acquisition and Processing XXXI, volume PC12848, page PC128480A. SPIE, March 2024. 5 +[76] 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. 7 +[77] 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. 7 +[78] Adam R Kosiorek, Heiko Strathmann, Daniel Zoran, Pol Moreno, Rosalia Schneider, Sona Mokra, and Danilo Jimenez Rezende. Nerf-vae: A geometry aware 3d scene generative model. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 5742-5752. PMLR, 18-24 Jul 2021. 8 +[79] Jiatao Gu, Alex Trevithick, Kai-En Lin, Josh Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion, 2023. 8 + +[80] 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 +[81] Martin J Bastiaans. Application of the wigner distribution function to partially coherent light. JOSA A, 3(8):1227-1238, 1986. 1 +[82] Adam Wax, Changhuei Yang, Vadim Backman, Kamran Badizadegan, Charles W. Boone, Ramachandra R. Dasari, and Michael S. Feld. Cellular organization and substructure measured using angle-resolved low-coherence interferometry. Biophysical Journal, 82(4):2256-2264, 2002. 2 \ No newline at end of file diff --git a/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/images.zip b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..411351fff446851377bac8b0e5135f9f0fa51589 --- /dev/null +++ b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e15eff8c6586c717c1d37e20f7d3ca3c29cf504cb40182d7be684a17c487a03 +size 747079 diff --git a/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/layout.json b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..50e818e5286de4e3a4c91ed16d8c4496d570289a --- /dev/null +++ b/ICCV/2025/Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eaf4c56a0e8e9c0e2056469d15bd7a473bdc936f9185e5c083e977ed5ffd46b3 +size 432109 diff --git a/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/0d6707dd-8891-4089-87e1-808c4b20461d_content_list.json b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/0d6707dd-8891-4089-87e1-808c4b20461d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..97e821bf66a93eed55f3dea0b2bc568b953bedfb --- /dev/null +++ b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/0d6707dd-8891-4089-87e1-808c4b20461d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7463b3e06cf12e62a4a3417482c0e0aabc472ebcbf968aebf87c01efc3044d4 +size 84805 diff --git a/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/0d6707dd-8891-4089-87e1-808c4b20461d_model.json b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/0d6707dd-8891-4089-87e1-808c4b20461d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a26155c791d4f97d30454b0ed478b2c5c2a952fd --- /dev/null +++ b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/0d6707dd-8891-4089-87e1-808c4b20461d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d250801196b5088f90b91f891902e858c2a8901aff881dfe7594685790c9f5c +size 102709 diff --git a/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/0d6707dd-8891-4089-87e1-808c4b20461d_origin.pdf b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/0d6707dd-8891-4089-87e1-808c4b20461d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2c6a3875200427b3e4752d69397c7f895d60a22a --- /dev/null +++ b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/0d6707dd-8891-4089-87e1-808c4b20461d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0162a0189b858da4a2ba6c83b201e33151b522c2422eca54e4a2f9effca007e2 +size 11220957 diff --git a/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/full.md b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d13cb210cb7224780b77cdf5fb11acc3d25a3f4 --- /dev/null +++ b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/full.md @@ -0,0 +1,317 @@ +# Recovering Parametric Scenes from Very Few Time-of-Flight Pixels + +Carter Sifferman, $^{\dagger}$ Yiquan Li, $^{\dagger}$ Yiming Li, Fangzhou Mu, Michael Gleicher, Mohit Gupta, Yin Li University of Wisconsin-Madison + +![](images/5bec0851307a3bbcdaed46ab7386b700dfad8c3ad445f064ba858a8f39b171fe.jpg) +Figure 1. We introduce a method for recovering the geometry of parametric 3D scenes, such as the 6D pose of a known object, from a distributed set of very few (e.g., 15), diffuse (i.e., wide field-of-view) single-pixel ToF Sensors. Methods based on traditional depth sensors suffer poor performance under a low-pixel-count regime due to their sparse coverage. Our method outperforms a point cloud-based baseline by utilizing the entirety of data recovered by a diffuse ToF sensor. See Sec. 4.4 for details of the experiment. + +![](images/71ae7b473195360bb75769a9b6d40d8f051bd75fccba9aec3b722b4ef041441c.jpg) + +![](images/9a0e0f6e99d67480a94a4187314b6a99df81441144a5e45e41d872ae321d6e5c.jpg) + +# Abstract + +We aim to recover the geometry of 3D parametric scenes using very few depth measurements from low-cost, commercially available time-of-flight sensors. These sensors offer very low spatial resolution (i.e., a single pixel), but image a wide field-of-view per pixel and capture detailed time-of-flight data in the form of time-resolved photon counts. This time-of-flight data encodes rich scene information and thus enables recovery of simple scenes from sparse measurements. We investigate the feasibility of using a distributed set of few measurements (e.g., as few as 15 pixels) to recover the geometry of simple parametric scenes with a strong prior, such as estimating the 6D pose of a known object. To achieve this, we design a method that utilizes both feed-forward prediction to infer scene parameters, and differentiable rendering within an analysis-by-synthesis framework to refine the scene parameter estimate. We develop hardware prototypes and demonstrate that our method effectively recovers object pose given an untextured 3D model in both simulations and controlled real-world captures, and show promising initial results for other parametric scenes. We additionally conduct experiments to explore the limits and capabilities of our imaging solution. Our project webpage is available at cpsiff.github.io/recovering parametric Scenes + +# 1. Introduction + +Time-of-flight (ToF) cameras such as LiDARs are a key technology for modern 3D vision, enabling tasks like pose + +estimation, shape reconstruction, and object recognition, with applications spanning fields such as robotics, augmented reality, and autonomous driving. Most current methods for inference from ToF imagery depend on dense 3D data, usually represented as a point cloud captured by high-resolution camera(s). It is generally accepted that a dense collection of depth measurements (e.g., ToF pixels as points) is vital for precise 3D vision. While certain applications do require high-resolution geometry, can some vision tasks be accomplished with only sparse 3D measurements? + +This question is particularly relevant given the recent emergence of low-cost ( $<\$ 3$ USD per unit), miniature ( $<5$ mm across) ToF sensors [1, 30]. These sensors, already deployed in mobile [41] and wearable applications [9], are often implemented with a single photon avalanche diode (SPAD) array [31, 33], featuring very limited pixel counts (even a single pixel) yet a wide field-of-view (FoV) per pixel (e.g., $30^{\circ}$ ). They capture raw ToF data with a transient histogram—a 1D waveform that encodes the intensity of light returning from the scene at pico-to-nanosecond timescales, integrated over a pixel's entire FoV. Traditionally, these histograms are processed into a point cloud by detecting and converting their peaks into depth estimates. However, this processing reduces the high-dimensional histogram to a single number, eliminating potentially useful information. Our key hypothesis, inspired by recent studies [3, 15, 24, 32], is that while these sensors cannot recover high-resolution point clouds, even a few transient histograms encode rich scene information sufficient for various downstream 3D vision tasks. An example is recovering scenes with low geometric complexity or scenarios where a strong geometric prior (e.g., a low-dimensional parametric shape model) is available. Our question is, under these con + +ditions, what is the minimal number of depth measurements required to recover 3D scenes? + +As a first step towards addressing this question, we investigate the recovery of simple parametric 3D scenes using very few ToF measurements, each with a wide FoV (see Fig. 1), e.g., as few as 15 pixels captured by spatially distributed, low-cost single-pixel SPAD sensors. We assume a 3D Lambertian scene defined by a parametric shape model and aim to recover the parameters of that model using a limited number of transient histograms captured from known fixed poses. We place special emphasis on the task of 6D pose estimation, which is a specific case of parametric scene recovery. In this case, the parameters that we aim to recover are the position and orientation of a known object mesh. We focus on 6D pose estimation because it is a well-defined problem with practical applications in robotics and augmented reality. This makes it a good testbed to tackle the fundamental challenge: utilizing very low resolution sensor data. With very few pixels, each of which integrates over a wide FoV, recovering 6D pose is challenging. + +To solve this problem, we present an analysis-by-synthesis based approach. Our method integrates (1) a learning-based feedforward model which predicts an initial estimate of scene parameters; (2) a differentiable renderer that synthesizes sensor measurements given scene parameters in the parametric model; and (3) an optimization-based refiner that iteratively renders sensor measurements to optimize scene parameters using the differentiable renderer. To address the scarcity of real-world imaging data, we re-use our renderer to generate a large-scale synthetic dataset for training our feedforward model, and explore its ability to transfer to real-world captures. + +We develop hardware prototypes for real-world capture and evaluate our approach using both simulated and real-world data. In real-world tests, our approach successfully estimates poses of even non-Lambertian objects using only 15 ToF pixels and an untextured object mesh. Moreover, leveraging our approach and hardware, we also briefly investigate two other forms of parametric scene recovery: parametric shape recovery (i.e., the position and scale of a known spherical object), and human hand pose recovery (i.e., pose and articulation), for which we demonstrate encouraging preliminary results in real-world settings. + +Scope and Limitations. While our problem formulation is general, we focus on 6D pose estimation, with a limited exploration of two other forms of parametric scene recovery. Our main objective is to establish feasibility rather than present an immediately deployable solution. To simplify real-world experiments, we make key assumptions, such as Lambertian surfaces, co-located sensor and light source, and known sensor poses. While robustness to varying scene reflectance and imperfect sensor poses are assessed in our experiments, our approach and prototype are not yet prac + +tical for widespread use. Moreover, we rely on currently available consumer hardware, which has a restricted sensing range, limiting our experiments to tabletop scenes. + +# 2. Related Work + +Time-of-flight (ToF) Imaging with SPADs. A ToF camera emits light pulses and measures the return time of incident photons to estimate distance. SPAD sensors have increasingly been adopted for ToF imaging, typically combined with a co-located light source (e.g., laser) [31]. This setup has been successfully applied to fluorescence lifetime imaging [38], novel view synthesis [29], and non-line-of-sight (NLOS) imaging [11, 14, 17, 18, 48]. Many of these systems rely on large, costly SPAD arrays (\(>\\)10K USD) with high spatial and temporal resolution. Recent works have explored low-cost SPAD sensors (<\\(3 USD) with limited pixel counts and lower temporal resolution for applications such as NLOS imaging [6, 50], shape reconstruction [32], human pose estimation [37], and SLAM [25]. Our work also explores low-cost SPAD sensors for ToF imaging; however, our primary focus is on the feasibility of using a minimal number of SPAD sensors for parametric scene recovery. + +A key component in SPAD imaging is modeling the image formation process. Sifferman et al. [39] introduce a simple model for miniature ToF sensors. Recent works model the SPAD image formation process with differentiable functions for laboratory-grade [27, 28] or commodity [32] sensors, enabling gradient-based optimization and facilitating 3D tasks such as pose estimation and shape reconstruction. Our work modifies the sensor model in [32] to accommodate differentiable mesh rendering. + +3D Vision with Low-Cost SPADs. Prior works have explored feedforward neural networks for inference from transient histograms. Pixels2Pose [37] proposes a learning-based method to estimate whole-body human pose from a single $4 \times 4$ pixel transient histogram. DELTAR [24] and Jungerman et al. [15] both predict high resolution depth images from ToF transient(s) plus an RGB image. The neural networks in these prior works are generally trained on real-world data only, or on simulated data generated from a simple sensor model. In this work, we train on simulated data generated via a high-fidelity sensor model. + +Recent works have considered an analysis-by-synthesis (AbS) paradigm, which uses differentiable rendering to align a set of underlying scene parameters with observed transient measurements. Sifferman et al. [39] design an AbS pipeline to recover 3DoF plane pose and albedo from a set of $3 \times 3$ transient measurements. Mu et al. [32] present a method to recover arbitrary 3D scenes from a distributed set of $>100$ single-pixel transient measurements. Behari et al. [3] leverage AbS to reconstruct arbitrary 3D scenes from a miniature ToF sensor plus an RGB camera. Liu et al. [25] build a neural radiance field for dense SLAM by + +![](images/9210853765e182c4b8b18e8968adcd5023e555d90f3c8f791bb875db6219061b.jpg) +Figure 2. Overview of our method as applied to 6D pose estimation. Our method consists of two components: 1) a pose prediction module, where a feedforward network estimates initial object pose from a sparse set of input transient histograms; and 2) a pose refiner, where a differentiable renderer is integrated into an analysis-by-synthesis framework to iteratively optimize the pose estimates. Yellow boxes indicate inputs. Green boxes indicate (intermediate) outputs. The optimization loop is illustrated with green arrows. + +integrating data from a miniature ToF sensor with an RGB camera. Luo et al. [27] reconstruct 3D scenes from transient measurements from very few viewpoints, however they use a high-resolution scanning LiDAR with higher fidelity and data rate than the miniature ToF sensors we consider. + +Our work shares a similar goal to prior studies that use multiple sensors for 3D scene recovery. However, our focus is on leveraging strong geometric priors to push the limits of scene recovery using only a few low-fidelity sensors. + +6D Pose Estimation. 6D pose estimation aims to determine the 6 degree-of-freedom pose of a known rigid object relative to the camera, given its 3D mesh model. Recent approaches use supervised learning to directly regress object pose from RGB and/or depth images [5, 20, 34, 46, 47]. The predicted pose can be further refined via an AbS pipeline [19, 20, 43, 46]. We take inspiration from the success of these works in designing our approach, integrating a feedforward network for initial pose prediction and an AbS pipeline for pose refinement. + +# 3. Scene Recovery from a Few ToF Pixels + +Problem Formulation. We aim to recover 3D geometry of a Lambertian scene specified by a set of parameters $\mathbf{P}$ from a distributed set of $n$ ToF sensors with known poses. We make two key assumptions regarding the sensing setup and scene representation. First, we assume that each ToF sensor operates via a co-located diffuse light source with a finite field-of-view, and reports a transient histogram $\pmb{h}$ which captures the intensity of light returning from the scene after a controlled pulse of illumination. This assumption covers a range of ToF sensors, including flash LiDAR and the low-cost SPAD sensor considered in this paper. Second, we utilize a mesh-based 3D scene representation, where the scene is modeled as a polygonal mesh composed of interconnected triangles that define its shape and surface. Mesh-based representations are widely used in graphics and many shape models are built on meshes [23, 26, 36, 49]. In this case, $\mathbf{P}$ can be the 6 DoF pose of a 3D object mesh or parameters for a mesh-based shape model. Our goal is to esti + +mate $\mathbf{P}$ from the set of measured histograms $\{\pmb {h}_i\}_{i = 1}^n$ + +Method Overview. Our method consists of two steps: 1) given a set of input transient histograms $\{h_i\}_{i=1}^n$ , a feedforward network outputs a prediction $\mathbf{P}_{\mathrm{FF}}$ of the scene parameters, and 2) an analysis-by-synthesis based refiner takes $\mathbf{P}_{\mathrm{FF}}$ as an initial estimate, alongside camera pose and any other scene prior (e.g., a parametric model), and iteratively optimizes $\mathbf{P}_{\mathrm{FF}}$ to minimize the difference between the measured histograms $\{h_i\}_{i=1}^n$ and histograms synthesized by our differentiable renderer. An illustration of our method as applied to the task of 6D pose estimation is shown in Fig. 2. In what follows we introduce the SPAD image formation process and present each component of our method. + +# 3.1. Background: Transient Formation Model + +We utilize physics-based sensor modeling to accurately render the transient histograms $\{h_i\}_{i=1}^n$ . For each captured histogram, the laser source emits $N_{\mathrm{emit}}$ photons. Assuming that the source is co-located with the sensor at the origin $o$ , the rays of the emitted photons can be parametrized by a direction $\omega$ . As in [15, 32, 39], we ignore high-order light paths and consider one-bounce paths only. Therefore, each photon travels from $o$ to a point on the scene $x$ and then back to $o$ , where $x = x(\omega, \mathcal{M})$ is the first intersection between the ray in the direction of $\omega$ and the scene $\mathcal{M}$ . + +Following prior work [15, 32], the expected number of photons $N[i]$ received by the sensor within the $i$ -th bin, in its angular integral form, is + +$$ +N [ i ] = N _ {\mathrm {e m i t}} \int_ {\Omega} I (\boldsymbol {\omega}) \frac {\rho (\boldsymbol {x})}{\pi} \frac {\left\langle - \boldsymbol {\omega} , \hat {\boldsymbol {n}} (\boldsymbol {x}) \right\rangle}{\| \boldsymbol {x} \| ^ {2}} W \left(\frac {2 \| \boldsymbol {x} \|}{c}, t _ {i}\right) \mathrm {d} \boldsymbol {\omega}, \tag {1} +$$ + +where $\Omega$ is the space of solid angles within the FoV of the sensor. $I(\omega)$ encodes the intensity of the laser along the direction $\omega$ . $\rho(\pmb{x})$ is the albedo, and $\hat{\pmb{n}}(\pmb{x})$ is the normal of the surface at $\pmb{x}$ . $t_i = i\Delta t$ corresponds to the time of the leading edge of the $i^{\mathrm{th}}$ bin, with $\Delta t$ as the bin width. Lastly, + +$$ +W (t, t _ {i}) = \left\{ \begin{array}{l l} 1 & \text {i f} t \in [ t _ {i}, t _ {i} + \Delta t), \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {2} +$$ + +bins the photons. Due to hardware effects, the binning of photons is not perfect in reality. $\pmb{h}[i]$ in fact might also detect photon arrivals near but outside that bin, and is affected by the shape of the outgoing laser pulse. Therefore, we convolve the expected photon numbers with an empirically derived discrete jitter kernel $s$ to account for this effect + +$$ +\boldsymbol {h} [ i ] = \Sigma_ {j} N [ i + \Delta i - j ] \boldsymbol {s} [ j ], \tag {3} +$$ + +where $\Delta i$ is sensor-specific, constant temporal offset to the histogram, which needs to be calibrated. + +Pile-up Correction. Previous works [13, 32] model the pile-up effect, which can significantly alter the measured histogram at high levels of ambient or active flux. To mitigate this effect, many existing sensors pre-process the histogram with algorithms like Coates' correction [8]. With this correction and under reasonable lighting conditions, the pile-up effect is often negligible. We thus do not include pile-up in our imaging model. + +# 3.2. Differentiable Rendering + +We numerically integrate Eq. (1) using the weighted sum + +$$ +N [ i ] \approx \frac {N _ {\mathrm {e m i t}}}{\pi} \sum_ {\boldsymbol {\omega} \in \mathcal {W}} Q (\boldsymbol {\omega}) I (\boldsymbol {\omega}) \rho (\boldsymbol {x}) \frac {\left\langle - \boldsymbol {\omega} , \hat {\boldsymbol {n}} (\boldsymbol {x}) \right\rangle}{\| \boldsymbol {x} \| ^ {2}} W (\frac {2 \| \boldsymbol {x} \|}{c}, t _ {i}), (4) +$$ + +where $\mathcal{W}$ is a specified set of rays, and $Q(\omega)$ is the associated quadrature. To make the rendering differentiable, we calculate $\partial x / \partial \mathcal{M}$ and $\partial \hat{n} /\partial \mathcal{M}$ using off-the-shelf differentiable rendering libraries. Specifically, we set $\mathcal{W}$ to a $h\times w$ grid of rays, fully covering the FoV and resembling the rendering of classical pixels, and the rasterization computes the rays' $\pmb {x},\hat{\pmb{n}}$ , and the gradients. Suppose that the center of the FoV is the $\mathbf{z}$ -axis and the imaging plane is $z = 1$ , each pixel has area $A_{\mathrm{pixel}} = (2\tan (FoV / 2))^2 /(h\cdot w)$ Assuming $\omega = (\omega_{x},\omega_{y},\omega_{z})$ , the quadrature $Q(\omega)$ transforms the square pixel areas to solid angle differentials by + +$$ +Q (\omega) = \frac {A _ {\mathrm {p i x e l}} \langle \omega , \hat {z} \rangle}{(1 / \omega_ {z}) ^ {2}} = \frac {4 \tan^ {2} (F o V / 2) \omega_ {z} ^ {3}}{h \cdot w}. \tag {5} +$$ + +The binning function $W$ in Eq. (2) is discontinuous. We thus approximate it with the sigmoid function $\sigma(x)$ by + +$$ +W (t, t _ {i}) = \sigma (k (t - t _ {i})) - \sigma (k (t - t _ {i} - \Delta t)), \tag {6} +$$ + +where $k$ is a hand-picked constant to balance smoothness and realism. Further, the intensity map $I(\omega)$ is discontinuous under an idealized diffuse laser source, since it provides uniform illumination within its FoV and zero illumination elsewhere. However, real-world lasers exhibit a nonuniform distribution, where intensity is highest at the center and gradually decreases toward the edges of the FoV. This allows us to approximate $I(\omega)$ using a differentiable, spatially-varying function. We fit this function using real-world sensor properties in our experiments (see Sec. 4.2). + +# 3.3. Feedforward Estimation of Scene Parameters + +We learn a neural network $f_{\theta}$ to predict initial scene parameters $\mathbf{P}_{\mathrm{FF}}$ , which will be further refined. This is given by + +$$ +f _ {\theta} \left(\left\{\boldsymbol {h} _ {i} \right\} _ {i = 1} ^ {n}\right)\rightarrow \mathbf {P} _ {\mathrm {F F}}, \tag {7} +$$ + +where $\theta$ is the network weights learned from data, $\{h_i\}_{i=1}^n$ is the input of $n$ transient histograms from ToF sensors. Namely, this feedforward network directly regresses the scene parameters based on sensor data. + +Network Architecture. Specifically, $f_{\theta}$ is a standard Transformer model [45]. The input transient histograms are first normalized, and then embedded using an MLP. These embeddings are added to positional embeddings, and further processed by a stack of Transformer blocks (4 in our implementation). The output embeddings are concatenated and fed into another MLP to predict scene parameters $\mathbf{P}$ . This network is trained with full supervision, and the loss function varies depending on the scene parameterization used, as described in the supplement. + +Sim-to-Real Transfer. A key challenge for training is the lack of real-world sensor data. We explore training on a large-scale synthetic dataset and transfer the learned model directly to real-world captures. This is made possible thanks to our efficient renderer in Sec. 3.2 and availability of 3D models [7]. We demonstrate strong results using this sim-to-real transfer in our experiments. + +Discussion. Our network assumes fixed sensor poses and requires re-training for every sensor configuration. This design is highly tailored for our current hardware prototypes, yet can be easily extended to accommodate varying sensor poses, e.g., encoding sensor pose as part of the input [22]. + +# 3.4. Parameter Refinement + +Given an estimate $\mathbf{P}_{\mathrm{FF}}$ of the scene parameters from the feedforward network and the differentiable renderer $\mathcal{R}$ , we propose an analysis-by-synthesis approach to further refine $\mathbf{P}_{\mathrm{FF}}$ . This is done by directly optimizing $\mathbf{P}$ to minimize the difference between the measured histograms $\{h_i\}_{i=1}^n$ and rendered histograms $\mathcal{R}(\mathbf{P})$ , given by + +$$ +\arg \min _ {\mathbf {P}} \Sigma_ {i = 1} ^ {n} \| \mathcal {R} (\mathbf {P}) _ {i} - \boldsymbol {h} _ {i} \|. \tag {8} +$$ + +Since the renderer $\mathcal{R}$ is fully differentiable, gradient descent can be used to solve this optimization locally. Specifically, this optimization starts from the initial estimate $\mathbf{P}_{\mathrm{FF}}$ and applies gradients steps until convergence. Since the rendering process $\mathcal{R}$ is highly nonlinear, the quality of the solution depends on the accuracy of initial estimate $\mathbf{P}_{\mathrm{FF}}$ . + +# 4. Experiments on 6D Pose Estimation + +To adopt our method for 6D pose recovery, we set the scene parameters $\mathbf{P}$ to a rotation $\mathbf{R}$ and translation $\mathbf{T}$ which trans + +![](images/5fcd556454323dcb5412c3ea6fe0373b1761f0c64385ead502fbdda8c26e9ecc.jpg) +Figure 3. Left: Illustration of our capture setup, where a ToF sensor is mounted on a robot arm and moved between a set of positions. Right: 15 sensor poses used for our experiments. + +form a known object mesh to its position in a global coordinate frame. We evaluate our method for 6D pose estimation in simulation and on real-world captures. + +# 4.1. Hardware and Real-world Capture + +Hardware Prototype. Our imaging system assumes known relative position of the ToF sensors. In practical deployment, this may be achieved by placing multiple sensors in a stationary position in the environment, or by attaching them each to a rigid object, like a mobile device. To allow flexibility for our experiments, we instead place a single sensor on an industrial robot arm, and move the sensor to multiple positions by controlling the robot while the scene remains static. We rely on the robot's kinematics, which are quoted as repeatable to $\pm 0.1\mathrm{mm}$ [44], to record sensor pose. + +We use the AMS TMF8820 SPAD-based ToF sensor. Like other sensors of its class, the sensor is very small $(12.8\mathrm{mm}^3, < 1\mathrm{g})$ and low-power $(< 100\mathrm{mW})$ [1]. The illumination source is an integrated low-power $940\mathrm{nm}$ VCSEL laser. We operate the sensor in "low range, high accuracy mode," and 4 million iterations per measurement, giving it a maximum range of $1.5\mathrm{m}$ and a bin size equivalent to $\sim 1.4\mathrm{cm}$ . We interface with the sensor via an attached microcontroller, which forwards transient histogram measurements from the sensor to a connected computer. + +Capture Setup. We sample random sensor positions between $30\mathrm{cm}$ and $80\mathrm{cm}$ from the workspace center (within the range of TMF8820), and random orientations which face the camera to the workspace center $(\pm 15^{\circ})$ . For a fair comparison, the same 15 randomly sampled sensor poses are used for all real-world experiments. This number (15) allows for practical data capture, and was chosen to strike a balance between estimation accuracy and information sparsity based on our simulation results (see Sec. 4.4). An Intel RealSense D405 depth-from-stereo camera is affixed next to the ToF sensor for ground truth capture. We utilize the merged point cloud from the depth camera views alongside ICP [4] and manual registration to generate ground truth object poses. We measure the geometry of the known background (the tabletop) by touching the robot to the surface at multiple points. This capture setup is shown in Fig. 3. + +Data Capture. We capture each object at 25 poses (10 for the highly symmetric basketball, softball, and tennis ball) + +by manually positioning the object such that the object center is within $15\mathrm{cm}$ of the workspace center. Effort is made to distribute the object poses uniformly within the workspace and to vary the object orientation uniformly. Because the objects are placed on a tabletop, we are restricted to orientations that provide stable support on the surface. + +# 4.2. Implementation Details + +TMF8820 Modeling. The TMF8820 reports transient histograms for nine separate fields-of-view, called "zones", each of which correspond to different sets of pixels on the SPAD array. Significant bloom artifacts are present between zones, and the exact zone dimensions are poorly defined [3, 39]. To address this challenge, and in line with our focus on very-low-pixel-count regimes, we aggregate the zones into one by summing across the zone dimension, yielding a single 128-bin histogram per sensor measurement. This approach is consistent with prior work [32, 40]. + +Further, in the TMF8820 datasheet, the laser illumination is reported as non-uniform across the FoV [1] $I(\omega)$ . We therefore approximate the intensity map using + +$$ +I (\boldsymbol {\omega}) = K _ {1} \exp \left(- K _ {2} \left(\omega_ {x} ^ {2} + \omega_ {y} ^ {2}\right) - K _ {3} \left(\omega_ {x} ^ {4} + \omega_ {y} ^ {4}\right)\right). \tag {9} +$$ + +We calibrate the constants $K$ to match the intensity map shown in the TMF8820 datasheet. A visualization of $I(\omega)$ is included in the supplement (Fig. D). We find that pile-up is not very apparent even at high ambient light levels. + +Jitter Kernel. The TMF8820 reports the shape of the outgoing laser pulse in a "reference histogram" alongside each measurement, which is measured from a SPAD sensor inside the laser cavity. Because this reference histogram is itself captured by a SPAD sensor, it encapsulates the shape of the outgoing laser pulse and the temporal response function of the SPAD itself. We make use of this histogram as the jitter kernel $s$ in our imaging model. + +Sensor Calibration. The reference histogram reported by the TMF8820 is not reported at the same temporal resolution as the transient histogram [32, 39]. We resample the reference histogram by a factor of $s_{\mathrm{scale}}$ before use in our imaging model. Additionally, the temporal resolution $\Delta t$ and a constant temporal offset to the histogram $\Delta i$ are not known. Following prior work, we recover the parameters $s_{\mathrm{scale}}, \Delta t$ , and $\Delta i$ by calibrating on some set of reference captures of a planar surface with known geometry. To do so, we keep scene geometry fixed, and optimize the sensor parameters to minimize the loss between captured and rendered histograms, akin to Eq. (8). + +Feedforward Network. We train the network described in Sec. 3.3 to predict 6D pose. For non-symmetric objects, we use a combination of rotation loss, translation loss, and a point matching loss. For symmetric objects, we use the ADD-S loss [47]. See the supplement (Sec. A) for details + +![](images/07fd3432b99b7e212b2a6e654782c14db8c755f2d64d62323a3eb67dc7b1d18f.jpg) +Figure 4. Visualization of 6D pose estimation results with 3D-printed objects using our method (feedforward + refiner). For each object, the pose prediction with the median pose error (ADD) over the 25-capture dataset is shown. + +of the loss functions and training parameters. A single-instance forward pass takes $\sim 4.6\mathrm{ms}$ on Nvidia RTX 4080. + +Pose Refinement. We use adaptive gradient descent following Adam [16] to solve Eq. (8). We set the step size (i.e., learning rate) to 0.01 for $\mathbf{R}$ and 0.001 for $\mathbf{T}$ . Additionally, we optimize the albedo of both the object ( $\rho_{\mathrm{obj}}$ ) and the planar surface ( $\rho_{\mathrm{plane}}$ ). These albedos are not predicted by the feedforward network; instead they are initialized to 1 and further optimized. We find empirically that optimizing albedos improves performance. We set the number of optimization steps to 200. We represent $\mathbf{R}$ using the 6D rotation representation proposed in [51]. Differentiable rendering is implemented via Nvdiffrast [21] and PyTorch [2]. + +# 4.3. Experiment Protocol + +Datasets. We evaluate our method for 6D pose estimation on two sets of objects: 1) 3D printed test objects and 2) seven readily available objects from the YCB dataset [7] — a standard benchmark for 6D pose recognition. See the supplement (Fig. I) for images of the objects. For YCB objects, the high-resolution "Google 16k" meshes provided by the YCB dataset are used for data generation and as input to the refiner. A child's basketball is used in place of the child's soccer ball in the YCB object set, along with a manually constructed spherical mesh. + +Synthetic Training Data. We generate synthetic data with our renderer to train the feedforward model. For each object, we synthesize 200K samples by limiting the object center within $15\mathrm{cm}$ of the workspace center. Object orientations are randomly sampled. To ensure physical plausibility, the object height is adjusted so that at least one vertex of the mesh lies on the planar surface, preventing the object from appearing to float in space, though this configuration may not correspond to a stable resting pose. This setup imposes a conservative prior on object placement. The planar surface is included in the scene when rendering synthetic data. We also perform domain randomization [42]; we add Gaussian noise with a standard deviation of $1.5\mathrm{cm}$ to the sensor positions independently for each sample, and vary the albedo of the object and the planar surface. + +![](images/a8a2c79286ac15c9023eadfe1428cb72a1b666d7fb3a0b9b93db1f27e3aa2ba2.jpg) +Figure 5. Visualization of 6D pose estimation results with objects from the YCB dataset using our method (feedforward + refiner). For each object, the pose prediction with the median pose error (ADD-S) over the 25-capture dataset is shown. + +Evaluation Metrics. We follow standard metrics [47] for evaluation, including ADD for non-symmetric objects, and ADD-S for symmetric objects. ADD captures the average distance between corresponding points on the predicted and ground truth object. ADD-S captures the average distance between a point on the predicted object and the nearest point on the ground truth object. We report the AUC-ADD(-S) in order to capture the distribution of scores over the entire dataset, with the maximum threshold in calculating AUC set to $10\mathrm{cm}$ . Note that ADD(-S) score is highly dependent on the scale and geometry of a specific object, thus should not be compared between objects. + +Point Cloud Baselines. We implement a point-cloud based baseline which represents an upper bound on point-cloud based system performance. To avoid pitfalls of any one particular depth camera, we simulate idealized sensor measurements; for each sensor pixel we project rays to get points of intersection with scene geometry, which are combined to form a point cloud. Points which lie on the planar surface are removed. The result is a noise-free point cloud of points on the object mesh. We then use ICP [4] to align the object mesh to this point cloud. To maximize the chance of success, the ground truth pose is used as initialization for ICP. Therefore, this baseline represents the best possible performance of a vanilla ICP-based registration system. We consider two variants: single-pixel, in which the same number of imaging pixels are used as in our system (one per view), and $16^2$ , in which a $16 \times 16$ point cloud sampled from a pixel grid spanning an FoV matching the diffuse sensor is generated per view (yielding $256 \times$ imaging pixels of our system). + +RGB(D) Baselines. For reference, we compare to two recent deep models: FoundPose [52], which uses single-view RGB input, and FoundationPose [46], which uses single-view RGB-D input. For both methods, the RGB image from viewpoint 15 is used as it provides a wide view. For FoundationPose, a high resolution depth camera view is simulated from the ground truth object pose. + +
DataTotal +PixelsMethodCrackersMustardChipsAUC-ADD-S (↑)
SPAMBasketballTennis BallSoftballMean
Sim.151 Px Point Cloud†78.3682.1277.8385.0782.9288.0986.3682.96
Real15Ours: Feedforward88.0290.0488.3890.0495.1596.2694.9591.83
Real15Ours: FF + Refiner90.0490.0788.5090.0095.7696.0694.9592.20
Sim.384016² Point Cloud†95.1797.2397.2397.1997.6797.5797.3797.06
Real407KSingle-View RGB [52]60.7187.9340.1258.9565.4677.6872.4266.18
Real407KSingle-View RGB-D* [46]90.4992.1092.5493.8094.2486.6794.2492.01
+ +Table 1. Results of 6D pose estimation with (symmetric) objects from the YCB object set. gray: methods using additional pixels; †: methods using oracle ground-truth pose; *: methods using simulated high-resolution point cloud data. See details in Sec. 4.3. + +
DataTotal +PixelsMethodAUC-ADD (↑)
TwoPLBunnyArmadilloMean
Sim.151 Px Point Cloud†73.8759.6555.1156.1453.8959.73
Real15Ours: Feedforward74.6777.3169.1167.7865.9370.96
Real15Ours: FF + Refiner83.4784.9477.7577.6879.7180.71
Sim.3840162 Point Cloud†97.5596.1895.1796.1896.9096.39
Real407KSingle-View RGB [52]56.9965.4051.5164.6386.2864.96
Real407KSingle-View RGB-D* [46]86.5785.5182.7288.3087.5886.14
+ +Table 2. Results of 6D pose estimation with (non-symmetric) 3D printed objects. gray: methods using additional pixels; †: methods using oracle ground-truth pose; *: methods using simulated high-resolution point cloud data. See details in Sec. 4.3. + +# 4.4.6D Pose Estimation Results + +Main Results. Our main results are presented in Tab. 1 and Tab. 2. Visualization of sample results with median pose errors are shown in Fig. 4 and Fig. 5. In all cases, our method outperforms the best-case performance of using a single-pixel point cloud. Additionally, our refiner improves performance in most cases. We also notice that our refiner yields a much larger performance improvement over the feedforward network on 3D printed objects. We hypothesize that this is due to the ambiguity of symmetric objects, and the symmetric loss can easily converge to a local minimum. Despite reasonable metrics, the SPAM is a failure case for our method; because the object is very small and near-symmetric along many axes, a high ADD-S score can be achieved by matching object translation. However, qualitative results (Fig. 5) show that orientation is not predicted reliably. Our method approaches the performance of the RGB-D baseline, while exceeding the performance of the RGB baseline which struggles due to the lack of metric depth information and/or lack of object texture. + +Varying Viewpoint Count. In simulation, we evaluate the performance of our method with varying number of views. We compare to the point cloud-based baseline described in Sec. 4.3. We evaluate on synthetic data of the 25 real "2" poses. Camera poses are sampled via the same process as real poses ( Sec. 4.1). Results are shown in the rightmost panel of Fig. 1. See the supplement for full results (Tab. D). Our approach exceeds the performance of the point cloud-based baseline for 5-100 total pixels. The point cloud-based method suffers with very few input pixels because, despite some variation in sensor orientation, the poses do not achieve good coverage of the scene. + +Additional Results. Due to space limit, the following are included in the supplement: (1) results with varying scene reflectance and ambient light; (2) ablations on the sensor model and feedforward network training; and (3) experiments on sensor interference. + +# 5. Exploration Beyond 6D Pose Estimation + +# 5.1. Size and Position of Spherical Objects + +We experiment with recovering the size and location of a sphere resting on a planar surface. We use an identical method to that used for 6D pose estimation (Sec. 4), except the predicted parameters $\mathbf{P}$ consist of the center point and diameter of a sphere, rather than the rotation and translation of a known mesh. Both parameters are predicted by the feedforward network and optimized during refinement. We evaluate on our pre-existing captures of the basketball, softball, and tennis ball objects. The results of this experiment are shown in Tab. 3. We find that our method can recover the position and diameter of the sphere with an average error of $< 1\mathrm{cm}$ in all cases, with often $< 0.5\mathrm{cm}$ of error, despite the temporal resolution of an individual SPAD being $\sim 1.4\mathrm{cm}$ . + +# 5.2. Human Hand Pose + +We recover human hand pose (absolute pose and articulation) from a ring of eight sensors encircling the wrist, as shown in Fig. 6. We modify the feedforward prediction to include pose, shape, global translation, and rotation of the human hand. The hand pose and shape is represented using the parameters of the MANO hand model [36]. + +Data Capture. As an initial feasibility study, we gather 250 measurements of a single individual's hand from the + +
ObjectMean Error in Diameter (cm) (↓)Mean Error in Position (cm) (↓)
Basketball0.35 (1.9%)0.84
Softball0.28 (2.9%)0.24
Tennis Ball0.33 (4.8%)0.39
+ +Table 3. Results of recovering position / size of spherical objects. + +
MethodPA-MPJPE (mm) (↓)
ToF-based Prior Work† [10]11.96
Trained on Sim. Data Only (ours)19.56
Trained on Real Data Only (ours)9.98
P.T. on Sim., F.T. on Real (ours)8.18
RGB-Based Method† [35]6.0
+ +Table 4. Results of hand pose and shape estimation. Related works are provided for context only; metrics are over a different dataset and should not be directly compared. + +ring of sensors. To sample hand poses, we prompt the user to match their hand pose to a random hand pose selected from the DART dataset [12]. RGB-D cameras are mounted above and below the hand to capture ground truth, which is provided by the RGB-based method HaMeR [35], and aligned to a fused point cloud from the two depth maps via ICP [4]. We reserve 50 captures for testing. + +Results. We report Procrutes aligned mean per joint position error (PA-MPJPE), a standard metric for hand tracking [10, 35]. PA-MPJPE captures the average distance between corresponding joints in the predicted and ground truth hand pose. The results of our experiment are shown in Tab. 4. We find that, in this setting, training on simulated data alone yields unsatisfactory results. A closer inspection reveals that the simulated histograms become inaccurate at distances below $15\mathrm{cm}$ . We attribute this to unmodeled sensor effects, such as unmodeled effects such as gating and/or pile-up from the high intensity of returning light. While we attempted to mitigate this issue by modeling these effects and learning a custom gating function, these efforts did not lead to improved performance. For the same reason, our refiner module is also not effective for hand pose estimation. + +We observe significantly improved results when training on real data, with the best results achieved through a two-stage process: pre-training on simulated data followed by fine-tuning on real data. Despite the limited realism of the simulated data, pre-training still provides benefits, likely because the network is able to learn transferable high-level features. To contextualize our results, we include comparisons to related works in Tab. 4. However, results from prior works are based on different datasets and experimental conditions, and thus are not directly comparable. + +# 6. Discussion + +Our work demonstrate that a few (e.g., 15) diffuse ToF pixels are sufficient to recover simple scene geometry. More + +![](images/5adf1b5fe4a78ec4bd1f4f48e7ba814e376aad9dd5c2b580b87f782f74d211bd.jpg) +Figure 6. Top: Setting for hand pose capture, in which eight ToF sensors encircle the wrist. Bottom: Visualization of results of applying our method to hand pose estimation, corresponding to method "Pretrained on Sim., F.T. on Real" in Tab. 4. + +over, we showcased the potential of our approach for more complex geometry, including recovering the position and scale of a spherical object, and human hand pose estimation. Our work offers an initial step toward enabling a range of practical applications and open up several promising directions for future research. + +Practical Implications. While still in its early stage, our approach has great potential for 3D vision applications that benefit from low-cost, low-power, and distributed sensing. One promising application domain is wearable computing. Inspired by our experiments on hand tracking, we envision that an array of miniature ToF sensors could be deployed in head-mounted or wrist-worn devices to track a user's body motion (e.g., arm and hand pose), enabling gesture-based user interfaces. Another key application domain is robotics. Imagine a robotic arm or drone equipped with a distributed array of lightweight, energy-efficient ToF sensors. These sensors could function as a network of spatially distributed cameras, reconstructing the environment from an inside-out perspective, enhancing tasks like grasping, navigation, and human-robot interaction. + +Future Directions. Our work demonstrates the estimation of 6D pose of rigid objects in a tabletop setting. Future work should aim to improve robustness to environmental factors such as ambient light and varied surface reflectance. This could be achieved by explicitly modeling these factors or developing methods that are inherently invariant to them. Additionally, future work should explore recovery of more complex scene geometries at larger scales, e.g., multiple deformable, articulated objects in room- or playground-sized environments. A promising future direction is learning from large-scale synthetic data. Encouragingly, our results have demonstrated that effective sim-to-real transfer is possible with ToF histogram data. We are hopeful that large-scale synthetic data can be applied to a range of inference tasks. + +Acknowledgements: This work was supported in part by the National Science Foundation under Grant Numbers 2152163 (NRT), 2333491 (CPS Frontier), 2442739 (CAREER), 1943149 (CAREER), 2107060 (CNS Core), by ARL under contract number W911NF-2020221, and by ONR grant N000142412155. + +# References + +[1] AMS OSRAM AG. TMF882X Datasheet. AMS OSRAM AG. 1, 5 +[2] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Chourdia, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshitteej Kalambarkar, Laurent Kirsch, Michael Lazos, Mario Lezcano, Yanbo Liang, Jason Liang, Yinghai Lu, C. K. Luk, Bert Maher, Yunjie Pan, Christian Puhrsch, Matthias Reso, Mark Saroufim, Marcos Yukio Siraichi, Helen Suk, Shunting Zhang, Michael Suo, Phil Tillet, Xu Zhao, Eikan Wang, Keren Zhou, Richard Zou, Xiaodong Wang, Ajit Mathews, William Wen, Gregory Chanan, Peng Wu, and Soumith Chintala. Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2, page 929-947, New York, NY, USA, 2024. Association for Computing Machinery. 6 +[3] Nikhil Behari, Aaron Young, Siddharth Somasundaram, Tzofi Klinghoffer, Akshit Dave, and Ramesh Raskar. Blurred LiDAR for Sharper 3D: Robust Handheld 3D Scanning with Diffuse LiDAR and RGB, 2024. arXiv:2411.19474 [eess]. 1, 2, 5 +[4] P.J. Besl and Neil D. McKay. A method for registration of 3-D shapes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 14(2):239-256, 1992. Conference Name: IEEE Transactions on Pattern Analysis and Machine Intelligence. 5, 6, 8 +[5] Dingding Cai, Janne Heikkia, and Esa Rahtu. OVE6D: Object Viewpoint Encoding for Depth-based 6D Object Pose Estimation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6793-6803, New Orleans, LA, USA, 2022. IEEE. 3 +[6] Clara Callenberg, Zheng Shi, Felix Heide, and Matthias B. Hullin. Low-cost SPAD sensing for non-line-of-sight tracking, material classification and depth imaging. ACM Transactions on Graphics, 40 (4):1-12, 2021. 2 + +[7] Berk Calli, Arjun Singh, Aaron Walsman, Siddhartha Srinivasa, Pieter Abbeel, and Aaron 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, 2015. 4, 6 +[8] PB Coates. The correction for photonpile-up'in the measurement of radiative lifetimes. Journal of Physics E: Scientific Instruments, 1968. 4 +[9] Dan O'shea. UK start-up Singular Photonics tout SPAD sensors, Meta collab, 2025. Accessed: 2025-03-07. 1 +[10] Nathan Devrio and Chris Harrison. discoBand: Multiview Depth-Sensing Smartwatch Strap for Hand, Body and Environment Tracking. In Proceedings of the 35th Annual ACM Symposium on User Interface Software and Technology, pages 1-13, New York, NY, USA, 2022. Association for Computing Machinery. 8 +[11] Daniele Faccio, Andreas Velten, and Gordon Wetzstein. Non-line-of-sight imaging. Nature Reviews Physics, 2(6):318-327, 2020. 2 +[12] Daiheng Gao, Yuliang Xiu, Kailin Li, Lixin Yang, Feng Wang, Peng Zhang, Bang Zhang, Cewu Lu, and Ping Tan. DART: Articulated Hand Model with Diverse Accessories and Rich Textures, 2022. arXiv:2210.07650 [cs]. 8 +[13] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-Flooded Single-Photon 3D Cameras. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6763–6772, Long Beach, CA, USA, 2019. IEEE. 4, 5 +[14] Felix Heide, Matthew O'Toole, Kai Zang, David B Lindell, Steven Diamond, and Gordon Wetzstein. Non-line-of-sight imaging with partial occluders and surface normals. ACM Transactions on Graphics (ToG), 38(3):1-10, 2019. 2 +[15] Sacha Jungerman, Atul Ingle, Yin Li, and Mohit Gupta. 3D Scene Inference from Transient Histograms. In Computer Vision – ECCV 2022, pages 401–417. Springer Nature Switzerland, Cham, 2022. Series Title: Lecture Notes in Computer Science. 1, 2, 3 +[16] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization, 2017. arXiv:1412.6980 [cs]. 6 +[17] Ahmed Kirmani, Tyler Hutchison, James Davis, and Ramesh Raskar. Looking around the corner using transient imaging. In 2009 IEEE 12th International Conference on Computer Vision, pages 159-166. IEEE, 2009. 2 +[18] Tzofi Klinghoffer, Xiaoyu Xiang, Siddharth Somasundaram, Yuchen Fan, Christian Richardt, Ramesh + +Raskar, and Rakesh Ranjan. Platonerf: 3d reconstruction in plato's cave via single-view two-bounce lidar. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14565-14574, 2024. 2 +[19] Alexander Krull, Eric Brachmann, Frank Michel, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Learning Analysis-by-Synthesis for 6D Pose Estimation in RGB-D Images. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 954-962, Santiago, Chile, 2015. IEEE. 3 +[20] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare, 2022. arXiv:2212.06870 [cs]. 3 +[21] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 6 +[22] Ruilong Li, Brent Yi, Junchen Liu, Hang Gao, Yi Ma, and Angjoo Kanazawa. Cameras as relative positional encoding. arXiv preprint arXiv:2507.10496, 2025. 4 +[23] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3 +[24] Yijin Li, Xinyang Liu, Wenqi Dong, Han Zhou, Hujun Bao, Guofeng Zhang, Yinda Zhang, and Zhaopeng Cui. DELTAR: Depth Estimation from a Light-Weight ToF Sensor and RGB Image. In Computer Vision – ECCV 2022, pages 619–636. Springer Nature Switzerland, Cham, 2022. Series Title: Lecture Notes in Computer Science. 1, 2 +[25] Xinyang Liu, Yijin Li, Yanbin Teng, Hujun Bao, Guofeng Zhang, Yinda Zhang, and Zhaopeng Cui. Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1-11, 2023. 2 +[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 3 +[27] Weihan Luo, Anagh Malik, and David B. Lindell. Transientangelo: Few-Viewpoint Surface Reconstruction Using Single-Photon Lidar, 2024. arXiv:2408.12191. 2, 3 +[28] Anagh Malik, Parsa Mirdehghan, Sotiris Nousias, Kyros Kutulakos, and David Lindell. Transient Neural + +Radiance Fields for Lidar View Synthesis and 3D Reconstruction. Advances in Neural Information Processing Systems, 36:71569-71581, 2023. 2 +[29] Anagh Malik, Noah Juravsky, Ryan Po, Gordon Wetzstein, Kiriakos N Kutulakos, and David B Lindell. Flying with photons: Rendering novel views of propagating light. In European Conference on Computer Vision, pages 333-351. Springer, 2024. 2 +[30] ST Microelectronics. VL6180X Proximity and Ambient Light Sensing Module Datasheet. ST Microelectronics. 1 +[31] Kazuhiro Morimoto, Andrei Ardelean, Ming-Lo Wu, Arin Can Ulku, Ivan Michel Antolovic, Claudio Bruschini, and Edoardo Charbon. Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica, 7(4):346, 2020. 1, 2 +[32] Fangzhou Mu, Carter Sifferman, Sacha Jungerman, Yiquan Li, Mark Han, Michael Gleicher, Mohit Gupta, and Yin Li. Towards 3D Vision with Low-Cost SinglePhoton Cameras. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5302-5311, Seattle, WA, USA, 2024. IEEE. 1, 2, 3, 4, 5 +[33] C. Niclass, A. Rochas, P.-A. Besse, and E. Charbon. Design and characterization of a CMOS 3-D image sensor based on single photon avalanche diodes. IEEE Journal of Solid-State Circuits, 40(9): 1847-1854, 2005. Conference Name: IEEE Journal of Solid-State Circuits. 1 +[34] Kiru Park, Timothy Patten, and Markus Vincze. Pix2Pose: Pixel-Wise Coordinate Regression of Objects for 6D Pose Estimation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7667-7676, 2019. arXiv:1908.07433 [cs]. 3 +[35] Georgios Pavlakos, Dandan Shan, Iija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing Hands in 3D with Transformers. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9826-9836, Seattle, WA, USA, 2024. IEEE. 8 +[36] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 3, 7, 1 +[37] Alice Ruget, Max Tyler, Germán Mora-Martín, Stirling Scholes, Feng Zhu, Istvan Gyongy, Brent Hearn, Steve McLaughlin, Abderrahim Halimi, and Jonathan Leach. Pixels2Pose: Super-Resolution Time-of-Flight Imaging for 3D Pose Estimation. In Imaging and Applied Optics Congress 2022 (3D, AOA, COSI, ISA, pcAOP), page ITh5D.5, Vancouver, British Columbia, 2022. Optica Publishing Group. 2 + +[38] David Eric Schwartz, Edoardo Charbon, and Kenneth L Shepard. A single-photon avalanche diode array for fluorescence lifetime imaging microscopy. IEEE journal of solid-state circuits, 43(11):2546-2557, 2008. 2 +[39] Carter Sifferman, Yeping Wang, Mohit Gupta, and Michael Gleicher. Unlocking the Performance of Proximity Sensors by Utilizing Transient Histograms. IEEE Robotics and Automation Letters, 8(10):6843-6850, 2023. 2, 3, 5 +[40] Carter Siffterman, William Sun, Mohit Gupta, and Michael Gleicher. Using a Distance Sensor to Detect Deviations in a Planar Surface. IEEE Robotics and Automation Letters, 9(10):8515-8522, 2024. Conference Name: IEEE Robotics and Automation Letters. 5 +[41] TechInsights. iPhone 15 Pro Max Rear LiDAR Camera Process Flow Analysis, 2023. Accessed: 2025-03-07. 1 +[42] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23-30. IEEE, 2017. 6 +[43] Jonathan Tremblay, Bowen Wen, Valts Blukis, Balakumar Sundaralingam, Stephen Tyree, and Stan Birchfield. Diff-DOPE: Differentiable Deep Object Pose Estimation, 2023. arXiv:2310.00463 [cs]. 3 +[44] Universal Robots. Ur5 technical specifications. Online, 2016. Accessed: 2025-02-28. 5 +[45] 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. 4 +[46] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17868-17879, Seattle, WA, USA, 2024. IEEE. 3, 6, 7 +[47] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes, 2018. arXiv:1711.00199 [cs]. 3, 5, 6, 1 +[48] Shumian Xin, Sotiris Nousias, Kiriakos N. Kutulakos, Aswin C. Sankaranarayanan, Srinivasa G. Narasimhan, and Ioannis Gkioulekas. A Theory of Fermat Paths for Non-Line-Of-Sight Shape Reconstruction. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 6800–6809, 2019. 2 + +[49] Wenqiang Xu, Zhenjun Yu, Han Xue, Ruolin Ye, Siqiong Yao, and Cewu Lu. Visual-Tactile Sensing for In-Hand Object Reconstruction. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8803-8812, Vancouver, BC, Canada, 2023. IEEE. 3 +[50] Aaron Young, Nevindu M. Batagoda, Harry Zhang, Akshit Dave, Adithya Pediredla, Dan Negrut, and Ramesh Raskar. Enhancing Autonomous Navigation by Imaging Hidden Objects using Single-Photon LiDAR, 2024. arXiv:2410.03555 [cs]. 2 +[51] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the Continuity of Rotation Representations in Neural Networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5738-5746, 2019. ISSN: 2575-7075. 6, 1 +[52] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. FoundPose: Unseen Object Pose Estimation with Foundation Features. In Computer Vision – ECCV 2024, pages 163–182. Springer Nature Switzerland, Cham, 2025. Series Title: Lecture Notes in Computer Science. 6, 7 \ No newline at end of file diff --git a/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/images.zip b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a99c1a8a7bb047f29fa00848b7787791d6464025 --- /dev/null +++ b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c2e4445266b76d2d06d0892c854eddd475cf6985b95eea5be83483c7c8dcf71 +size 378504 diff --git a/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/layout.json b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..edd3d61dc69c30660b22b70822cb772c7f9a112c --- /dev/null +++ b/ICCV/2025/Recovering Parametric Scenes from Very Few Time-of-Flight Pixels/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09b90d6de06d6ee2e253a389e6aa1960e6e6bc4a73e9fcf33b1aeb9fc2ea300a +size 404021 diff --git a/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/e6100218-5fd4-4b92-a47e-baa795e53261_content_list.json b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/e6100218-5fd4-4b92-a47e-baa795e53261_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..abac607e87a1dbfef9df09df198b225a17bee7cb --- /dev/null +++ b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/e6100218-5fd4-4b92-a47e-baa795e53261_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6440b00a5eb57f755f56e22e006b05bd5607960535f198b73d5bd6b778e27324 +size 92243 diff --git a/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/e6100218-5fd4-4b92-a47e-baa795e53261_model.json b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/e6100218-5fd4-4b92-a47e-baa795e53261_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ea85e458e94d78c81c979d58b16edba8f008ecf3 --- /dev/null +++ b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/e6100218-5fd4-4b92-a47e-baa795e53261_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be3457db11779b25d7d5e231532b289f19120a8d7cb872ea0085238519b24c15 +size 112394 diff --git a/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/e6100218-5fd4-4b92-a47e-baa795e53261_origin.pdf b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/e6100218-5fd4-4b92-a47e-baa795e53261_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..94b4a4c949edb58d96d931b5219f6bf5d6c8bdf7 --- /dev/null +++ b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/e6100218-5fd4-4b92-a47e-baa795e53261_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99057c384900d30df5bb244a1f9b7b79be025d4f5f2fc5d79d82e2de5d54362c +size 1220466 diff --git a/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/full.md b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2b926f4cea028f58884b370b2ed953d698536f29 --- /dev/null +++ b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/full.md @@ -0,0 +1,387 @@ +# Rectifying Magnitude Neglect in Linear Attention + +Qihang Fan $^{1,2}$ , Huaibo Huang $^{1*}$ , Yang Ai $^{1,2}$ , Ran He $^{1,2}$ + +$^{1}$ MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences, Beijing, China + $^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China +fanqihang.159@gmail.com, huaibo.huang@cripac.ia.ac.cn, shallowdream555@gmail.com, rhe@nlpr.ia.ac.cn + +# Abstract + +As the core operator of Transformers, Softmax Attention exhibits excellent global modeling capabilities. However, its quadratic complexity limits its applicability to vision tasks. In contrast, Linear Attention shares a similar formulation with Softmax Attention while achieving linear complexity, enabling efficient global information modeling. Nevertheless, Linear Attention suffers from a significant performance degradation compared to standard Softmax Attention. In this paper, we analyze the underlying causes of this issue based on the formulation of Linear Attention. We find that, unlike Softmax Attention, Linear Attention entirely disregards the magnitude information of the Query(Q or $\phi (Q)$ ). This prevents the attention score distribution from dynamically adapting as the Query scales. As a result, despite its structural similarity to Softmax Attention, Linear Attention exhibits a significantly different attention score distribution. Based on this observation, we propose Magnitude-Aware Linear Attention (MALA), which modifies the computation of Linear Attention to fully incorporate the Query's magnitude. This adjustment allows MALA to generate an attention score distribution that closely resembles Softmax Attention while exhibiting a more well-balanced structure. We evaluate the effectiveness of MALA on multiple tasks, including image classification, object detection, instance segmentation, semantic segmentation, natural language processing, speech recognition, and image generation. Our MALA achieves strong results on all of these tasks. Code will be available at https://github.com/qhfan/MALA. + +# 1. Introduction + +Since the introduction of the Transformer [9, 50] into the vision domain, it has gained increasing attention. Its exceptional global modeling capability has enabled Vision Trans + +![](images/9451333ff51bddde643b9621bbeda854d02a26360303a279c7cf0e921b74a053.jpg) +Figure 1. Comparison between Softmax Attention and Linear Attention. While linear attention offers linear complexity and high computational efficiency, its modeling capability falls short compared to Softmax Attention. + +formers to achieve outstanding performance in various visual tasks, such as image classification, object detection, and semantic segmentation, fully demonstrating the Transformer's potential in vision applications [20, 37]. + +However, the core operator of the Transformer, Softmax Attention, has a quadratic complexity with respect to the number of tokens $N$ , resulting in high computational costs that significantly hinder its widespread adoption in the vision domain. Many models reduce the computational cost of Softmax Attention by decreasing the number of tokens involved in its computation, bringing its complexity closer to or even achieving linearity [8, 13, 18, 37, 51, 52]. However, these methods, which limit the number of tokens, often fail to accurately model the relationships between all tokens globally, preventing the Transformer from fully leveraging its original advantages. + +Unlike these improvements to Softmax Attention, linear attention fundamentally eliminates the Softmax operation. As shown in Fig. 1, by removing the Softmax operation, the computation order of $Q$ , $K$ , and $V$ is rearranged, resulting in a linear complexity with respect to the number of tokens $N$ . Although Linear Attention and Softmax Attention share a very similar form, the removal of the Softmax operation introduces several challenges, often leading to significantly + +![](images/6c14ddb589b32119ce4cc76823457c8e6929fd3e3ad8a8951ef418327f1495fc.jpg) +Figure 2. Comparison of attention score distributions across different mechanisms. As the magnitude of the Query $(Q)$ or $\phi(Q))$ increases, the attention score distribution in Softmax Attention becomes increasingly spiky, concentrating more attention on keys that originally have higher scores. In contrast, Linear Attention maintains an unchanged distribution or exhibits only minimal variation, resulting in a relatively smooth attention score distribution. Our MALA retains the spiky characteristic of Softmax Attention while preventing it from becoming excessively sharp, achieving a more balanced distribution. + +inferior performance compared to Softmax Attention. + +In this paper, we analyze the computational formulation of Linear Attention and observe that it entirely disregards the magnitude information of the Query $(Q$ or $\phi (Q))$ , preserving only its directional component. Consequently, Linear Attention exhibits a substantial discrepancy in attention score distribution compared to Softmax Attention. Specifically, as illustrated in Fig. 2, for a fixed direction, the attention scores in Softmax Attention become increasingly spiky as the Query magnitude increases, concentrating more attention on keys that originally have higher attention scores. In contrast, due to the inherent limitations of its computation, Linear Attention either maintains a fixed attention score distribution or undergoes only minimal variation, and the distribution remains consistently smooth. This phenomenon may account for its weak local perception and the tendency to produce overly smooth attention scores [4, 20, 21, 42]. + +To address this issue and better align the attention score distribution of Linear Attention with that of Softmax Attention, we propose Magnitude-Aware Linear Attention (MALA). MALA fully integrates the magnitude information of the Query, mimicking the variation trend of Softmax Attention while achieving a more balanced and reasonable allocation of attention. As a result, MALA outperforms Softmax Attention while preserving its linear complexity. To demonstrate the effectiveness of MALA, we conduct extensive experiments on image classification, object detection, instance segmentation, semantic segmentation, natural language processing, speech recognition, and image generation. Strong results across all these tasks demonstrate the effectiveness of proposed MALA. + +Our contributions can be summarized as follows: + +- We analyze the computational formulation of Linear Attention and reveal that it entirely disregards variations in + +the Query $(Q$ or $\phi (Q))$ 's magnitude. This omission leads to a substantial discrepancy between the attention score distributions of Linear Attention and Softmax Attention. + +- To bridge this gap, we propose Magnitude-Aware Linear Attention (MALA), which fully incorporates the Query's magnitude information. MALA mimics the variation trend of Softmax Attention while achieving a more balanced and principled attention score distribution. +- Based on MALA, we develop the Magnitude-Aware Vision Transformer (MAViT). We also test MALA on other tasks, such as natural language processing, speech recognition, and image generation. All models achieve promising results. + +# 2. Related Works + +Vision Transformers. Vision Transformer (ViT) is a powerful foundational vision model inspired by advancements in natural language processing (NLP) [9, 50]. It demonstrates remarkable performance across various vision tasks. However, the core operator of the Transformer, Softmax Attention, has a quadratic complexity with respect to the number of tokens $N$ , imposing a computational burden that limits the application of Transformers in vision tasks. Many works have proposed improvements to address this issue. One approach adopts a grouping strategy, where tokens are divided into multiple groups, reducing the computational burden at the cost of sacrificing the global receptive field of ViT [7, 8, 23, 37, 57]. Another approach directly downsamples the tokens, preserving ViT's global perception capability but compromising its fine-grained representation [11, 18, 35, 46, 51, 52]. Some methods integrate convolution with Transformers to enhance model's efficiency [10, 28, 31]. However, most of these approaches still rely on the quadratic complexity of Softmax Attention. + +Linear Attention. Linear Attention assumes that the exponential function can be approximated by the product of kernel functions. This decomposition reformulates the computation of attention scores, reducing the complexity of Attention to linear time. However, this improvement in efficiency comes at the cost of performance degradation. Many works have explored ways to bridge the gap between Softmax Attention and Linear Attention [4, 20, 22, 39, 44]. Among them, MILA [22], inspired by Mamba, incorporates Mamba's macro architecture into the design of Linear Attention. EfficientViT [4] and Flatten Transformer [20] integrate Linear Attention with convolution to compensate for its limitations in capturing local features. In contrast to these methods, we directly address the computational form of Linear Attention and the distribution of attention scores, aiming to align the behavior of Linear Attention with that of Softmax Attention. + +# 3. Method + +# 3.1. Preliminary + +Given an input token sequence $X \in \mathbb{R}^{N \times d}$ of length $N$ and dimension $d$ , the output of the $i$ th token $X_{i}$ after attention processing can be expressed as: + +$$ +Q = X W _ {Q}, K = X W _ {K}, V = X W _ {V}, +$$ + +$$ +Y _ {i} = \sum_ {j = 1} ^ {N} \frac {\operatorname {S i m} \left(Q _ {i} , K _ {j}\right)}{\sum_ {m = 1} ^ {N} \operatorname {S i m} \left(Q _ {i} , K _ {m}\right)} V _ {j}; \tag {1} +$$ + +Where $W_{Q}, W_{K}, W_{V} \in \mathbb{R}^{d \times d}$ are learnable matrices, $\operatorname{Sim}(.,.)$ is the similarity function. In classical Softmax Attention, $\operatorname{Sim}(Q_i, K_j) = \exp(Q_i K_j^T / \sqrt{d})$ . This requires computing the exponential value for each pair of query and key, resulting in a complexity of $O(N^2)$ . + +In Linear Attention, this situation changes. It employs a kernel function $\phi(.)$ to approximate the similarity function and maps $Q$ and $K$ into positive real numbers. and leading to $\mathrm{Sim}(Q_i,K_j) = \phi(Q_i)\phi(K_j)^T$ . Based on this transformation, the formulation of Linear Attention can be rewritten as: + +$$ +\begin{array}{l} Y _ {i} = \sum_ {j = 1} ^ {N} \frac {\phi \left(Q _ {i}\right) \phi \left(K _ {j}\right) ^ {T}}{\sum_ {m = 1} ^ {N} \phi \left(Q _ {i}\right) \phi \left(K _ {m}\right) ^ {T}} V _ {j} \tag {2} \\ = \frac {\phi (Q _ {i}) \left(\sum_ {j = 1} ^ {N} \phi (K _ {j}) ^ {T} V _ {j}\right)}{\phi (Q _ {i}) \left(\sum_ {m = 1} ^ {N} \phi (K _ {m}) ^ {T}\right)}; \\ \end{array} +$$ + +in this computational form, the order of operations for $Q$ , $K$ , and $V$ changes from $(QK^T)V$ to $Q(K^TV)$ , eliminating the need to compute the result for each query-key pair. This reduces the complexity with respect to the number of tokens $N$ from $O(N^2)$ to $O(N)$ . However, the reduction in complexity also leads to a decline in performance. + +# 3.2. Magnitude Neglect in Linear Attention + +We define + +$$ +\phi (Q _ {i}) = \| \phi (Q _ {i}) \| \vec {\alpha} _ {i}; \tag {3} +$$ + +where $\| \phi (Q_i)\|$ represents the magnitude of $\phi (Q_{i})$ , and $\vec{\alpha}_{i}$ denotes its direction vector. Substituting this expression into the formulation of Linear Attention, we obtain: + +$$ +\begin{array}{l} Y _ {i} = \frac {\left| \left| \phi \left(Q _ {i}\right) \right| \right| \vec {\alpha} _ {i} \left(\sum_ {j = 1} ^ {N} \phi \left(K _ {j}\right) ^ {T} V _ {j}\right)}{\left| \left| \phi \left(Q _ {i}\right) \right| \right| \vec {\alpha} _ {i} \left(\sum_ {m = 1} ^ {N} \phi \left(K _ {m}\right) ^ {T}\right)} \tag {4} \\ = \frac {\vec {\alpha} _ {i} (\sum_ {j = 1} ^ {N} \phi (K _ {j}) ^ {T} V _ {j})}{\vec {\alpha} _ {i} (\sum_ {m = 1} ^ {N} \phi (K _ {m}) ^ {T})}; \\ \end{array} +$$ + +from this equation, we observe that the magnitude information of $\phi (Q)$ in Linear Attention is completely ignored. As a result, as long as $\vec{\alpha}$ remains fixed, the attention score distribution of Linear Attention remains unchanged. + +This phenomenon leads to a significant discrepancy between the attention score distributions of Linear Attention and Softmax Attention. In Softmax Attention, the magnitude of $Q_{i}$ is fully taken into account. Given $Q_{i}$ , the ratio of its attention scores for two different keys, $K_{m}$ and $K_{n}$ , is given by + +$$ +\frac {\exp \left(Q _ {i} K _ {m} ^ {T} / \sqrt {d}\right)}{\exp \left(Q _ {i} K _ {n} ^ {T} / \sqrt {d}\right)} = p; \tag {5} +$$ + +We assume that $Q_{i}$ assigns a higher attention weight to $K_{m}$ , i.e., $p > 1$ . When the direction of $Q_{i}$ remains unchanged and its magnitude is scaled by a factor of $a > 1$ , the ratio of its attention scores for $K_{m}$ and $K_{n}$ becomes: + +$$ +\frac {\exp \left(a Q _ {i} K _ {m} ^ {T} / \sqrt {d}\right)}{\exp \left(a Q _ {i} K _ {n} ^ {T} / \sqrt {d}\right)} = \frac {\exp \left(Q _ {i} K _ {m} ^ {T} / \sqrt {d}\right) ^ {a}}{\exp \left(Q _ {i} K _ {n} ^ {T} / \sqrt {d}\right) ^ {a}} = p ^ {a} = p _ {s}; \tag {6} +$$ + +Since $p > 1$ and $a > 1$ , it follows that $p_s > p$ . Given that the attention scores of $Q_i$ across all $K$ 's sum to 1, Eq. 5 and Eq. 6 imply that as the magnitude $\| Q_i \|$ increases, the attention of $Q_i$ becomes more concentrated on keys with higher original attention scores, while the attention assigned to keys with lower initial scores diminishes. + +However, this situation does not occur in Linear Attention. The ratio of $Q_{i}$ 's attention to $K_{m}$ and $K_{n}$ is given by: + +$$ +\frac {\phi (Q _ {i}) \phi (K _ {m}) ^ {T}}{\phi (Q _ {i}) \phi (K _ {n}) ^ {T}} = \frac {| | \phi (Q _ {i}) | | \vec {\alpha} _ {i} \phi (K _ {m}) ^ {T}}{| | \phi (Q _ {i}) | | \vec {\alpha} _ {i} \phi (K _ {n}) ^ {T}} = \frac {\vec {\alpha} _ {i} \phi (K _ {m}) ^ {T}}{\vec {\alpha} _ {i} \phi (K _ {n}) ^ {T}}; \tag {7} +$$ + +This indicates that regardless of the changes in the magnitude of $\phi(Q_i)$ , the attention scores in Linear Attention remain in the same distribution and do not concentrate on specific keys. This distinction explains why the attention scores learned by Linear Attention are less spiky compared to those of Softmax Attention and why the learned features exhibit weaker locality [4, 20, 21, 42]. + +
ModelSoftmaxQ' = Q/||Q||Softmax→Linear
Acc(%)72.270.069.8
+ +Table 1. Discarding magnitude information in Softmax Attention. + +![](images/b587852ddacfc40a760379609e772b012f6a01009b14a74c2e2cd5445c0a195d.jpg) + +![](images/41e3a0bd6171c3046153f876c70e65e3e453cf7aa6bbcaa67a24a06f13ef5108.jpg) + +![](images/39e972851f30492eb49ce255a82b230feb7aff3595382977c4c5c319c785784d.jpg) + +![](images/f3d513b4a0fb78df32f5327a2f608a2ef94535c9f410a9474feb2ad31dea2686.jpg) +Figure 3. Attention scores of different models. When $Q$ is replaced with $Q / ||Q||$ , Softmax Attention exhibits a distribution similar to that of Linear Attention, becoming much smoother and losing locality. +Softmax Attention + +![](images/bf8e46679c798e0d269016338b8444fab82d3117b7b3859fe34027949a11eb4b.jpg) +Enhanced Linear Attention + +![](images/f418eedb448e829d0e14da5dfba6e37e745bb16d7929720679580a12adf3ce88.jpg) +InLine Attention +Figure 4. Visualization of attention scores on DeiT-T setting. Softmax Attention's score is too spiky and primarily focuses on local regions, while Linear Attention's score is too smooth and excessively disregards local information. In contrast, MALA effectively balances both aspects. The visualizations on natural images are provided in the Appendix. + +![](images/0f513f61b7eceb83c4c4bab7ae52d64a465e823f7bf6dd486b210299ef70f834.jpg) +MALA + +In addition to the theoretical analysis above, we also conduct experimental validation. As shown in Tab. 1, based on DeiT-T, we rewrite the $Q$ in Softmax Attention as $Q / ||Q||$ , thereby disregarding the magnitude information. We observe a significant drop in the model's performance, which becomes similar to that of the model based on Linear Attention. We visualize the attention scores in Fig. 3 and find that the distribution converges to that of Linear Attention, becoming much smoother and losing locality. + +# 3.3. Magnitude-Aware Linear Attention + +To bridge the gap between Linear Attention and Softmax Attention, we aim for Linear Attention to incorporate the magnitude information $\| \phi (Q_i)\|$ and exhibit similar variation trend as Softmax Attention. + +In our Magnitude-Aware Linear Attention(MALA), building upon the original Linear Attention, we introduce + +a scaling factor and an offset term while discarding the division-based normalization in favor of an addition-based normalization: + +$$ +\operatorname {A t t n} \left(Q _ {i}, K _ {j}\right) = \beta \phi \left(Q _ {i}\right) \phi \left(K _ {j}\right) ^ {T} - \gamma ; \tag {8} +$$ + +Where: + +$$ +\beta = 1 + \frac {1}{\phi (Q _ {i}) \sum_ {m = 1} ^ {N} \phi (K _ {m}) ^ {T}}, +$$ + +$$ +\gamma = \frac {\phi (Q _ {i}) \sum_ {m = 1} ^ {N} \phi (K _ {m}) ^ {T}}{N}, +$$ + +$$ +\sum_ {j = 1} ^ {N} \operatorname {A t t n} \left(Q _ {i}, K _ {j}\right) = \beta \sum_ {j = 1} ^ {N} \phi \left(Q _ {i}\right) \phi \left(K _ {j}\right) ^ {T} - N \gamma = 1; \tag {9} +$$ + +When considering all attention scores as positive values, the ratio of $Q_{i}$ 's attention scores for $K_{m}$ and $K_{n}$ is given by: + +$$ +\frac {\beta \phi \left(Q _ {i}\right) \phi \left(K _ {m}\right) ^ {T} - \gamma}{\beta \phi \left(Q _ {i}\right) \phi \left(K _ {n}\right) ^ {T} - \gamma} = p; \tag {10} +$$ + +We assume that $Q_{i}$ assigns a higher attention score to $K_{m}$ , i.e., $\beta \phi(Q_{i})\phi(K_{m}) > \beta \phi(Q_{i})\phi(K_{n})$ and $p > 1$ . When the direction of $\phi(Q_{i})$ remains unchanged and its magnitude is scaled by a factor of $a > 1$ , it is straightforward to derive that the new $\beta$ and $\gamma$ can be written as: + +$$ +\beta_ {n e w} = \frac {\beta + a - 1}{a}, \tag {11} +$$ + +$$ +\gamma_ {n e w} = a \gamma ; +$$ + +At this point, the ratio of the attention scores of $a\phi(Q_i)$ for $K_m$ and $K_n$ becomes: + +$$ +\begin{array}{l} \beta_ {n e w} a \phi (Q _ {i}) \phi (K _ {m}) ^ {T} - \gamma_ {n e w} \\ \beta_ {n e w} a \phi (Q _ {i}) \phi (K _ {n}) ^ {T} - \gamma_ {n e w} \\ = \frac {\beta \phi \left(Q _ {i}\right) \phi \left(K _ {m}\right) ^ {T} - \frac {a \beta}{\beta + a - 1} \gamma}{\beta \phi \left(Q _ {i}\right) \phi \left(K _ {n}\right) ^ {T} - \frac {a \beta}{\beta + a - 1} \gamma} = p _ {m}; \tag {12} \\ \end{array} +$$ + +Since $\beta > 1$ and $a > 1$ , it is straightforward to prove that $\frac{a\beta}{a + \beta - 1} > 1$ . From this, we can further easily prove that when considering all attention scores as positive values, $p_m > p$ (Details can be found in the Appendix). Moreover, since $\sum_{j=1}^{N}\operatorname{Attn}(Q_i,K_j) = 1$ , as the magnitude of $\phi(Q_i)$ increases, MALA concentrates more attention on the keys that originally received higher attention, while allocating less attention to the keys that originally had lower attention. This behavior is similar to Softmax Attention. + +Although both Softmax Attention and MALA exhibit a trend of more concentrated attention score distributions as the magnitude of $Q_{i}$ or $\phi(Q_{i})$ increases, the rate at which this concentration occurs differs between the two. From the comparison between Eq. 6 and Eq. 12, it can be observed + +that in Softmax Attention, the ratio $p$ of attention scores exhibits an exponential growth with respect to the scaling factor $a$ of $\| Q \|$ . In contrast, in MALA, the ratio $p$ follows a fractional growth pattern with respect to the scaling factor $a$ of $\| \phi(Q) \|$ . The variation of $p$ in MALA is smaller than that in Softmax Attention, which may contribute to the superior performance of MALA over Softmax Attention. As shown in Fig. 4, we visualize the attention scores of different mechanisms. It can be observed that Softmax Attention's score is too spiky and primarily focuses on local regions. In contrast, Linear Attention's score is too smooth and excessively disregards local information [4, 20, 21]. MALA, however, effectively balances both aspects. This indicates that the gradual variation of $p$ in MALA leads to a more appropriately distributed attention score. + +As for the occurrence of negative/zero attention scores in MALA, in our experiments (image classification, object detection, instance segmentation and semantic segmentation), we find that although negative/zero attention scores are theoretically possible, their actual frequency of occurrence is equal zero. We do not observe any negative or zero attention scores. So we do not introduce additional considerations. + +When the attention scores are applied to the values, the complete formulation of MALA is expressed as: + +$$ +\begin{array}{l} Y _ {i} = \sum_ {j = 1} ^ {N} (\beta \phi (Q _ {i}) \phi (K _ {j}) ^ {T} - \gamma) V j \tag {13} \\ = \beta \phi (Q _ {i}) \sum_ {j = 1} ^ {N} \phi (K _ {j}) ^ {T} V _ {j} - \gamma \sum_ {j = 1} ^ {N} V j; \\ \end{array} +$$ + +Where: + +$$ +\begin{array}{l} \beta = 1 + \frac {1}{\phi \left(Q _ {i}\right) \sum_ {m = 1} ^ {N} \phi \left(K _ {m}\right) ^ {T}}, \tag {14} \\ \gamma = \frac {\phi (Q _ {i}) \sum_ {m = 1} ^ {N} \phi (K _ {m}) ^ {T}}{N}; \\ \end{array} +$$ + +# 4. Experiments + +We conduct extensive experiments on image classification, object detection, instance segmentation, semantic segmentation, natural language processing, speech recognition and image generation. Additionally, we perform ablation studies to validate the impact of MALA. More visualization results and details can be found in the Appendix. + +# 4.1. Image Classification + +Settings. We follow the same training strategy in previous works with the only supervision being classification loss [8, 13, 22, 27, 49, 57]. We train our models on ImageNet-1K [6] from scratch. The maximum rates of increasing stochastic depth [26] are set to $0.1 / 0.15 / 0.4 / 0.55$ for MAViT-T/S/B/L, respectively. The batch size is set to + +
CostModelTypeParmas (M)FLOPs (G)Top1-acc (%)
Tiny model ~ 2.5GNAT-M [23]Trans202.781.8
FAT-B2 [11]Trans142.081.9
GC-ViT-XT [24]Trans202.682.0
RMT-T [13]Trans142.582.4
MSVMamba-M [45]Mamba121.579.8
Flatten-PVTv2-B1 [20]Linear132.279.5
RAVLT-T [14]Linear152.482.8
MAViT-TLinear162.582.9
Small model ~ 4.5GMogaNet-S [32]CNN255.083.4
SG-Former-S [16]Trans234.883.2
FAT-B3 [11]Trans294.483.6
SMT-S [34]Trans204.883.7
RMT-S [13]Trans274.584.1
SECViT-S [12]Trans274.684.3
Vmamba-T [36]Mamba304.982.6
MSVMamba-T [45]Mamba325.183.0
Flatten-CSwin-T [20]Linear214.383.1
RAVLT-S [14]Linear264.684.4
MAViT-SLinear274.684.7
Base model ~ 10.0GConvNeXT-S [38]CNN508.783.1
InternImage-S [53]CNN508.084.2
MogaNet-B [32]CNN449.984.3
BiFormer-B [57]Trans579.884.3
RMT-B [13]Trans549.785.0
SECViT-B [12]Trans579.885.2
Vmamba-S [36]Mamba508.783.6
MSVMamba-S [45]Mamba508.884.1
MILA-S [22]Linear437.384.4
RAVLT-B [14]Linear489.985.5
MAViT-BLinear509.985.7
Large model ~ 15.0GMogaNet-L [32]CNN8315.984.7
InterImage-B [53]CNN9716.084.9
SG-Former-B [16]Trans7815.684.7
STViT-L [27]Trans9515.685.3
RMT-L [13]Trans9518.285.5
Vmamba-B [36]Mamba8915.483.9
MSVMamba-B [45]Mamba9116.384.4
SOFT-Huge [39]Linear8716.383.3
InLine-CSwin-B [21]Linear7314.984.5
RAVLT-L [14]Linear9516.085.8
MAViT-LLinear9816.186.0
+ +Table 2. Comparison with the state-of-the-art on ImageNet-1K classification. We use "CNN" to refer to convolutional neural networks, "Trans" to refer to Vision Transformers, "Mamba" to refer to visual state space model, and "Linear" to refer to models based on Linear Attention. + +1024 and the max learning rate is 1e-3. We train all models for 300 epochs. + +Results. we compare the performance of various models in Tab. 2. Under models of comparable size, MAViT achieves the best results. Specifically, with 98M parameters and 16.1G FLOPs, MAViT-L achieves the accuracy of $86.0\%$ . This performance surpasses MILA, another Linear Attention method, by $0.7\%$ . Moreover, MAViT-S achieves an accuracy of $84.7\%$ with only 27M parameters and 4.6G FLOPs, surpassing the larger MILA-S. + +
BackboneTypeParams FLOPs\(AP^b\)\(AP_{50}^b\)\(AP_{75}^b\)\(AP^m\)\(AP_{50}^m\)\(AP_{75}^m\)
(M)(G)
Mask R-CNN 3×+MS
NAT-T [23]Trans4825847.869.052.642.666.045.9
SMT-S [34]Trans4026549.070.153.443.467.346.7
RMT-S [13]Trans4626250.771.955.644.969.148.4
Vmamba-T [36]Mamba5027148.8--43.7--
MILA-T [22]Linear4425548.871.053.643.868.046.8
MAViT-SLinear4426251.472.656.245.569.849.2
InternImage-S [53]CNN6934049.771.154.544.568.547.8
SMT-B [34]Trans5232849.871.054.444.068.047.3
RMT-B [13]Trans7337352.272.957.046.170.449.9
Vmamba-S [36]Mamba7034949.9--44.2--
MILA-S [22]Linear6331950.571.855.244.969.148.2
MAViT-BLinear6737253.274.158.547.071.551.1
InternImage-B [53]CNN11550150.371.455.344.868.748.0
Swin-B [37]Trans10749648.670.053.443.367.146.7
CSwin-B [8]Trans9752650.872.155.844.969.148.3
MAViT-LLinear11450153.674.358.747.271.551.4
Cascade Mask R-CNN 3×+MS
HorNet-T [43]CNN8072852.471.656.845.669.149.6
GC-ViT-T [24]Trans8577051.670.456.144.667.848.3
CSWin-T [8]Trans8075752.571.557.145.368.848.9
RMT-S [13]Trans8374153.272.057.846.169.849.8
FL-Swin-T [20]Linear8774750.869.655.144.167.048.1
MAViT-SLinear8274154.272.658.647.070.551.1
NAT-S [23]Trans10880951.970.456.244.968.248.6
UniFormer-B [31]Trans10787853.872.858.546.469.950.4
RMT-B [13]Trans11185254.572.859.047.270.551.4
FL-Swin-S [20]Linear10884152.271.256.845.468.349.4
InLine-Swin-S [21]Linear10983552.471.056.945.468.849.6
MAViT-BLinear10585155.574.060.448.071.752.5
ConvNeXt-B [38]CNN14596452.771.357.245.668.949.5
Swin-B [37]Trans14598251.970.956.545.068.448.7
CSwin-B [8]Trans135100453.972.658.546.470.050.4
MAViT-LLinear15297956.074.660.948.472.452.9
+ +Table 3. Comparison to other backbones on "3 × +MS" schedule. + +# 4.2. Object Detection and Instance Segmentation + +Settings. Following previous works [13, 37, 57], we use RetinaNet [33], Mask-RCNN [25] and Cascade Mask R-CNN [5] to evaluate our models. We use the commonly used "1×" (12 epochs) setting for the RetinaNet and Mask R-CNN and "3×+MS" (36 epochs) for Mask R-CNN and Cascade Mask R-CNN. + +Results. We show the experimental results in Tab. 3 (" $3 \times +\mathrm{MS}$ ) and Tab. 4 (" $1 \times$ "). MAViT demonstrates significant advantages over other models based on Linear Attention. Moreover, it surpasses models utilizing Softmax Attention across all model scales. Specifically, MAViT-B achieves $55.5AP^{b}$ and $48.0AP^{m}$ , which even surpass the larger CSwin-B ( $53.9AP^{b}$ and $46.4AP^{m}$ ) under the framework of Cascade Mask R-CNN. + +# 4.3. Semantic Segmentation + +Settings. Follow previous works [18, 37, 46], we adopt SemanticFPN [29] and UperNet [54] to evaluate our models. For SemanticFPN, we train the models for 80K iterations [51, 52], while for UperNet, we train them for 160K + +![](images/b41cb6be6a1d103756ab140f8a5247ff0a9c968d973b426712333f71e95b50d8.jpg) +Figure 5. Comparison of general backbones' inference speed on low resolution task (image classification, resolution $224 \times 224$ ). The inference speed are measured on A100, batch size 64. + +![](images/ca78e7a209d759d23a28d940f481e09487184e7e4cf8d55f26d06daf196e3f8a.jpg) +Figure 6. Comparison of general backbones' inference speed on high resolution task (semantic segmentation with UperNet, resolution $512 \times 2048$ ). The inference speed are measured on A100, batch size 1. + +iterations. The batch sizes are set to 16 for all models. The input images are cropped to $512 \times 512$ . + +Results. We present the results in the Tab. 5. MAViT surpasses other models across various sizes. Specifically, MAViT-B achieves $52.8\mathrm{mIoU}$ under the framework of UperNet, which surpasses larger MILA. MAViT-L can even achieve $53.6\mathrm{mIoU}$ . + +# 4.4. Inference Efficiency + +Settings. To thoroughly assess the inference efficiency of MAViT, we evaluate its performance on both low-resolution task (image classification with a resolution of $224 \times 224$ ) and high-resolution task (semantic segmentation with a resolution of $512 \times 2048$ ). For the low-resolution task, we measure model throughput using a batch size of 64. For the high-resolution task, we evaluate the frame per second (FPS) performance with a batch size of 1. + +Results. The results are presented in Fig. 5 and Fig. 6. + +
BackboneTypeParams (M)FLOPs (G)Mask R-CNN 1×Params (M)FLOPs (G)RetinaNet 1×
APbAPb50APb75APmAPm50APm75APbAPb50APb75APSAPMAPbL
PVTv2-B1 [52]Trans3324341.854.345.938.861.241.62322541.261.943.925.444.554.3
MPViT-XS [30]Trans3023144.266.748.440.463.443.42021143.865.047.128.147.656.5
FAT-B2 [11]3321545.267.949.041.364.644.02319644.065.247.227.547.758.8
MSVmamba-M [45]Mamba3220143.865.847.739.962.942.9------
MAViT-TLinear3321947.669.552.542.966.546.42420145.666.748.928.949.761.1
CMT-S [18]Trans4524944.666.848.940.763.943.44423144.365.547.527.148.359.1
FAT-B3 [11]49-47.669.752.343.166.446.239-45.966.949.529.350.160.9
RMT-S [13]Trans4626249.070.853.943.967.847.43624447.869.151.832.151.863.5
VMamba-T [36]Mamba5027147.369.352.042.766.445.9------
MILA-T [22]Linear4425546.869.551.542.166.445.0------
MAViT-SLinear4426250.271.755.344.768.747.93424448.369.452.231.852.664.0
CSWin-S [8]Trans5434247.970.152.643.267.146.2------
STViT-B [27]Trans7035949.771.754.744.868.948.7------
MSVmamba-S [15]Mamba7034948.170.152.843.267.346.5------
SOFT++-medium [40]Linear6934246.667.851.242.064.845.25932244.364.747.429.048.259.9
MLLA-S [22]Linear6331949.271.553.944.268.547.2------
MAViT-BLinear6737251.773.357.046.170.650.15735349.971.153.833.754.565.5
InternImage-B [53]CNN11550148.870.954.044.067.847.4------
MPViT-B [30]Trans9550348.270.052.943.567.146.88548247.068.450.829.451.361.5
RMT-L [13]Trans11455751.673.156.545.970.349.810453749.470.653.134.253.965.2
MILA-B [22]Linear11550250.572.055.445.069.348.6------
MAViT-LLinear11450152.573.657.846.571.050.610448250.671.754.934.155.365.6
+ +Table 4. Comparison to other backbones with "1×" schedule. + +
ModelTypeSemantic FPN 80K Params FLOPsUpernet 160K Params FLOPs
(M)(G)(%)(M)(G)(%)
VAN-B1 [19]CNN1814042.9---
PVTv2-B1 [52]Trans1813642.5---
RMT-T [13]Trans1713646.4---
MSVmamba-M [45]Mamba---4287545.1
MAViT-TLinear1813647.64489348.4
MogaNet-S [32]CNN2918947.75594649.2
SMT-S [34]Trans---5093549.2
RMT-S [13]Trans3018049.45693749.8
Vmamba-T [36]Mamba---6294947.9
Fl-Swin-T [20]Linear---6094644.8
MAViT-SLinear2818050.75593751.0
MogaNet-B [32]CNN---74105050.1
RMT-B [13]Trans5729450.483105152.0
Vmamba-S [36]Mamba---82102850.6
FL-Swin-S [20]Linear---82103848.1
MAViT-BLinear5129251.577105052.8
MogaNet-L [32]CNN---113117650.9
CSWin-B [8]Trans8146449.9109122251.1
SGFormer-B [16]Trans8147550.6109130452.0
RMT-L [13]Trans9848251.4125124152.8
Vmamba-B [36]Mamba---122117051.0
MILA-B [22]Linear---128118351.9
MAViT-LLinear9842452.8125118253.6
+ +Fig. 5 illustrates the inference efficiency of different models on the low-resolution task, where MAViT achieves the best balance between throughput and accuracy. Similarly, for high-resolution tasks, the results in Fig. 6 further highlight MAViT's superior efficiency. This demonstrates that MALA not only has a significantly lower theoretical complexity than Softmax Attention but also achieves high inference speed in practice. + +Table 5. Comparison with the state-of-the-art on ADE20K. + +
LMB↑PIQA↑Hella↑Wino↑
Transformer [50]31.063.334.050.4
RetNet [47]28.663.533.552.5
GLA [55]30.364.834.551.4
Mamba [15]30.665.035.450.1
MALA31.065.034.551.9
+ +# 4.5. Natural Language Processing + +Settings. Following previous works, we train the 0.3B MALA based model on 15B tokens, and evaluate the model on several commonly used benchmarks. + +Results. We show the results in the Tab. 6. In the four commonly used benchmarks (LMB, PIQA, Hella, and Wino), our MALA exhibits strong performance. + +# 4.6. Speech Recognition + +Settings. Our evaluation on speech recognition is based on the previous Conformer [17]. We replace the Softmax Attention in the Conformer with 1) vanilla Linear Attention and 2) our MALA. All training settings are the same as Conformer. + +Table 6. MALA in NLP. + +
ModelParamsWER Without LM testclean↓testother↓WER With LM testclean↓testother↓
Conformer(S)10.32.76.32.15.0
Linear Attn10.33.410.22.67.3
MALA10.32.45.31.94.2
+ +Table 7. MALA in speech recognition. + +Results. We show the results in Tab. 7. The results demon + +strate that MALA perform better than Softmax Attention and vanilla Linear Attention. + +# 4.7. Image Generation + +Settings. Following previous works [1, 2, 41], we train the models for 400K iterations with the batch size of 256 and learning rate of 1e-4. + +
ModelFLOPsThroughput↑FID↓IS↑
DiT-S/2(400K) [41]250×6.06G4.9imgs/s68.40-
DiG-S/2(400K) [58]250×4.30G3.8imgs/s62.0622.81
DiC-S/2(400K) [48]250×5.90G-58.6825.82
MALA (400K)250×4.26G5.6imgs/s49.6232.18
+ +Results. We show the results in Tab. 8. Compare to other methods based on ConvNet/Transformer, our model based on MALA exhibits better performance and faster speed, which demonstrate the superiority of MALA. + +# 4.8. Ablation Study + +Comparison with Other Linear Attentions. To ensure a fair comparison with previous state-of-the-art Linear Attention mechanisms, we adopt three model settings: DeiT-T, Swin-T, and Swin-S. Under these settings, we replace all instances of Softmax Attention with our proposed Magnitude-Aware Linear Attention while keeping all other components unchanged to maintain absolute fairness. The results are presented in Tab. 9. From the results, we observe that MALA achieves a significant improvement over previous linear attention mechanisms. Specifically, under the Swin-S setting, MALA outperforms InLine Attention by $+1.7$ in accuracy. + +Kernel Function. In MAViT, we employ $\phi(.)=\mathrm{ELU}(.)+1$ as the kernel function to ensure the non-negativity of $\phi(Q)$ and $\phi(K)$ . To evaluate the impact of the kernel function, we conduct ablation studies based on MAViT-T. The results are presented in Tab. 10. Our MALA is not sensitive to the choice of kernel function, as almost any non-negative kernel function can achieve comparable performance. + +$\beta$ and $\gamma$ . $\beta$ and $\gamma$ are the core design elements of our model, endowing MALA with outstanding properties. We conduct ablation studies on these two parameters, and the results are presented in Tab. 11. We separately remove $\beta$ and $\gamma$ , leading to a sharp decline in model performance, with some cases even resulting in NaN values. We also replace $\beta$ and $\gamma$ with learnable parameters, and the model's performance significantly deteriorates. + +# 5. Conclusion + +In this paper, we observe that the attention scores of Softmax Attention and Linear Attention exhibit distinct variation patterns as the magnitude of the Query $(Q$ or $\phi (Q))$ + +Table 8. MALA for diffusion. + +
Linear AttentionParams(M)FLOPs(G)Top1-acc(%)
Comparison on DeiT-T Setting
DeiT-T [49]61.172.2
Hydra Attn [3]61.168.3
Efficient Attn [44]61.170.2
Linear Angular Attn [56]61.170.8
Enhanced Linear Attn [4]61.172.9
Focused Linear Attn [20]61.174.1
InLine Attn [21]71.174.5
Magnitude-Aware Linear Attn61.175.1
Comparison on Swin-T Setting
Swin-T [37]294.581.3
Hydra Attn [3]294.580.7
Efficient Attn [44]294.581.0
Linear Angular Attn [56]294.579.4
Enhanced Linear Attn [4]294.581.8
Focused Linear Attn [20]294.582.1
InLine Attn [21]304.582.4
Magnitude-Aware Linear Attn294.583.7
Comparison on Swin-S Setting
Swin-S [37]508.783.0
Focused Linear Attn [20]518.783.5
InLine Attn [21]508.783.6
Magnitude-Aware Linear Attn508.785.3
+ +Table 9. Comparison of different Linear Attentions based on DeiT, Swin-T, and Swin-S. MALA surpasses others by a large margin. + +
φ(.)Elu(.) + 1ReLU(.)exp(.)
Acc(%)82.982.882.9
mIoU47.647.747.4
+ +Table 10. Effect of different kernel functions. + +
ModelAcc(%)APbAPmmIoU
MAViT-T82.947.642.947.6
w/o β52.324.618.722.2
w/o γNaN---
Learnable71.734.331.831.9
+ +Table 11. Effect of $\beta$ and $\gamma$ . + +changes. From a formulation perspective, we analyze the underlying cause of this behavior and design Magnitude-Aware Linear Attention (MALA), which ensures that the attention scores of Linear Attention display a variation pattern similar to, yet more reasonable than, that of Softmax Attention. Based on MALA, we construct the Magnitude-Aware Vision Transformer (MAViT) and perform extensive experiments, which demonstrate the superior performance and high efficiency of MALA. + +# 6. Acknowledgements + +This work is partially funded by Beijing Natural Science Foundation (4252054), Youth Innovation Promotion Association CAS(Grant No.2022132), Beijing Nova Program(20230484276), and CCF-Kuaishou Large Model Explorer Fund (NO. CCF-KuaiShou 2024005). + +# References + +[1] Yuang Ai, Qihang Fan, Xuefeng Hu, Zhenheng Yang, Ran He, and Huaibo Huang. Dico: Revitalizing convnets for scalable and efficient diffusion modeling. arXiv preprint arXiv:2505.11196, 2025.8 +[2] Yang Ai, Huaibo Huang, Tao Wu, Qihang Fan, and Ran He. Breaking complexity barriers: High-resolution image restoration with rank enhanced linear attention. arXiv preprint arXiv:2505.16157, 2025.8 +[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, and Judy Hoffman. Hydra attention: Efficient attention with many heads, 2022. 8 +[4] Han Cai, Junyan Li, Muyan Hu, Chuang Gan, and Song Han. Efficientvit: Lightweight multi-scale attention for high-resolution dense prediction. In ICCV, pages 17302-17313, 2023. 2, 3, 5, 8 +[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 6 +[6] Jia Deng, Wei Dong, Richard Socher, et al. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5 +[7] Mingyu Ding, Bin Xiao, Noel Codella, et al. Davit: Dual attention vision transformers. In ECCV, 2022. 2 +[8] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, et al. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 1, 2, 5, 6, 7 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2 +[10] Qihang Fan, Huaibo Huang, Jiyang Guan, and Ran He. Rethinking local perception in lightweight vision transformer. arXiv preprint arXiv:2303.17803, 2023. 2 +[11] Qihang Fan, Huaibo Huang, Xiaogiang Zhou, and Ran He. Lightweight vision transformer with bidirectional interaction. In NeurIPS, 2023. 2, 5, 7 +[12] Qihang Fan, Huaibo Huang, Mingrui Chen, and Ran He. Semantic equitable clustering: A simple and effective strategy for clustering vision tokens. arXiv preprint arXiv:2405.13337, 2024. 5 +[13] Qihang Fan, Huaibo Huang, Mingrui Chen, Hongmin Liu, and Ran He. Rmt: Retentive networks meet vision transformers. In CVPR, 2024. 1, 5, 6, 7 +[14] Qihang Fan, Huaibo Huang, and Ran He. Breaking the low-rank dilemma of linear attention. In CVPR, 2025. 5 +[15] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 7 +[16] SG-Former: Self guided Transformer with Evolving Token Reallocation. Sucheng ren, xingyi yang, songhua liu, xinchao wang. In ICCV, 2023. 5, 7 +[17] Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. Conformer: Convolution-augmented transformer for speech recognition. In INTERSPEECH, 2020. 7 +[18] Jianyuan Guo, Kai Han, Han Wu, Chang Xu, Yehui Tang, Chunjing Xu, and Yunhe Wang. Cmt: Convolutional neural + +networks meet vision transformers. In CVPR, 2022. 1, 2, 6, 7 +[19] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 7 +[20] Dongchen Han, Xuran Pan, Yizeng Han, Shiji Song, and Gao Huang. Flatten transformer: Vision transformer using focused linear attention. In ICCV, 2023. 1, 2, 3, 5, 6, 7, 8 +[21] Dongchen Han, Yifan Pu, Zhuofan Xia, Yizeng Han, Xuran Pan, Xiu Li, Jiwen Lu, Shiji Song, and Gao Huang. Bridging the divide: Reconsidering softmax and linear attention. In NeurIPS, 2024. 2, 3, 5, 6, 8 +[22] Dongchen Han, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yifan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and Gao Huang. Demystify mamba in vision: A linear attention perspective. In NeurIPS, 2024. 3, 5, 6, 7 +[23] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 2, 5, 6 +[24] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In ICML, 2023. 5, 6 +[25] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask r-cnn. In ICCV, 2017. 6 +[26] Gao Huang, Yu Sun, and Zhuang Liu. Deep networks with stochastic depth. In ECCV, 2016. 5 +[27] Huaibo Huang, Xiaogiang Zhou, Jie Cao, Ran He, and Tieniu Tan. Vision transformer with super token sampling. In CVPR, 2023. 5, 7 +[28] Manjin Kim, Paul Hongsuck Seo, Cordelia Schmid, and Minsu Cho. Learning correlation structures for vision transformers. In CVPR, 2024. 2 +[29] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 6 +[30] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In CVPR, 2022. 7 +[31] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning, 2022. 2, 6 +[32] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z. Li. Moganet: Multi-order gated aggregation network. In ICLR, 2024. 5, 7 +[33] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, and Kaiming He and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 6 +[34] Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, and Lianwen Jin. Scale-aware modulation meet transformer. In ICCV, 2023. 5, 6, 7 +[35] Kai Liu, Tianyi Wu, Cong Liu, and Guodong Guo. Dynamic group transformer: A general vision transformer backbone with dynamic group attention. In IJCAI, 2022. 2 +[36] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. In NeurIPS, 2024. 5, 6, 7 + +[37] 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. 1, 2, 6, 8 +[38] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, et al. A convnet for the 2020s. In CVPR, 2022. 5, 6 +[39] Jiachen Lu, Li Zhang, Junge Zhang, Xiatian Zhu, Jianfeng Feng, and Tao Xiang. Softmax-free linear transformers. IJCV, 2024. 3, 5 +[40] Jiachen Lu, Li Zhang, Junge Zhang, Xiatian Zhu, Jianfeng Feng, and Tao Xiang. Softmax-free linear transformers. IJCV, 2024. 7 +[41] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 8 +[42] Zhen Qin, Xiaodong Han, Weixuan Sun, Dongxu Li, Lingpeng Kong, Nick Barnes, and Yiran Zhong. The devil in linear transformer. In EMNLP, 2022. 2, 3 +[43] Yongming Rao, Wenliang Zhao, Yansong Tang, Jie Zhou, Ser-Lam Lim, and Jiwen Lu. Hornet: Efficient high-order spatial interactions with recursive gated convolutions. In NeurIPS, 2022. 6 +[44] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In WACV, 2021. 3, 8 +[45] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. In NeurIPS, 2024. 5, 7 +[46] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xin chao Wang, and Shuicheng YAN. Inception transformer. In NeurIPS, 2022. 2, 6 +[47] Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, and Furu Wei. Retentive network: A successor to Transformer for large language models. ArXiv, abs/2307.08621, 2023. 7 +[48] Yuchuan Tian, Jing Han, Chengcheng Wang, Yuchen Liang, Chao Xu, and Hanting Chen. Dic: Rethinking conv3x3 designs in diffusion models. In CVPR, 2025. 8 +[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, et al. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 5, 8 +[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, et al. Attention is all you need. In NeurIPS, 2017. 1, 2, 7 +[51] 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 ICCV, 2021. 1, 2, 6 +[52] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):1-10, 2022. 1, 2, 6, 7 +[53] 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 CVPR, 2023. 5, 6, 7 +[54] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 6 + +[55] Songlin Yang, Bailin Wang, Yikang Shen, Rameswar Panda, and Yoon Kim. Gated linear attention transformers with hardware-efficient training. In ICML, 2024. 7 +[56] Haoran You, Yunyang Xiong, Xiaoliang Dai, Bichen Wu, Peizhao Zhang, Haoqi Fan, Peter Vajda, and Yingyan Lin. Castling-vit: Compressing self-attention via switching towards linear-angular attention during vision transformer inference. In CVPR, 2023. 8 +[57] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bformer: Vision transformer with bi-level routing attention. In CVPR, 2023. 2, 5, 6 +[58] Lianghui Zhu, Zilong Huang, Bencheng Liao, Jun Hao Liew, Hanshu Yan, Jiashi Feng, and Xinggang Wang. Dig: Scalable and efficient diffusion models with gated linear attention. In CVPR, 2025. 8 \ No newline at end of file diff --git a/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/images.zip b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..db9aebed032e73d61004cd0bc5435d072011d947 --- /dev/null +++ b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9db65a5e8f68e1deb76b6695938fed965ccd6225723cd51368696d29348d6db7 +size 1084004 diff --git a/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/layout.json b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..62edd789f8be4c721211807c4ed8a79d1c262054 --- /dev/null +++ b/ICCV/2025/Rectifying Magnitude Neglect in Linear Attention/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:479f3dfa802ffbb471a85edd460adf9d775408a67ac7d3e51b1bb87c449aa4dd +size 469712 diff --git a/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/3f13f2ff-b34f-4073-be92-b7b0aa04657f_content_list.json b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/3f13f2ff-b34f-4073-be92-b7b0aa04657f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..56a18aed5bd75270cb17bac194fd16b6a01ff317 --- /dev/null +++ b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/3f13f2ff-b34f-4073-be92-b7b0aa04657f_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d8532945d7178a82e380e0882d1d8e1f03cd5f80b9b9f7a9c2aa444a3d61172 +size 84407 diff --git a/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/3f13f2ff-b34f-4073-be92-b7b0aa04657f_model.json b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/3f13f2ff-b34f-4073-be92-b7b0aa04657f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..62f41a255e48df76236736ca20cb745a76776e36 --- /dev/null +++ b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/3f13f2ff-b34f-4073-be92-b7b0aa04657f_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:182f5bfd22249c4dc66ea747707aa0459b862eeba9657a594cf641c2af7c30ca +size 109162 diff --git a/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/3f13f2ff-b34f-4073-be92-b7b0aa04657f_origin.pdf b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/3f13f2ff-b34f-4073-be92-b7b0aa04657f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..03f1c4d0010f7a9efb44d5f970bd5ecd1ab925d7 --- /dev/null +++ b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/3f13f2ff-b34f-4073-be92-b7b0aa04657f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95b2cdbf2c87713d77a99a4e15166aa4478fc23ab971d3015c221c2d3d9f1783 +size 3623465 diff --git a/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/full.md b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bc42c68abd5adb63ebcca112c91c09390c1155ed --- /dev/null +++ b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/full.md @@ -0,0 +1,343 @@ +# Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization + +Xu Zheng $^{1,2}$ Yuanhuiyi Lyu $^{1}$ Lutao Jiang $^{1}$ Danda Pani Paudel $^{2}$ Luc Van Gool $^{2}$ Xuming Hu $^{1,3,*}$ + +$^{1}$ AI Thrust, HKUST(GZ) $^{2}$ INSAIT, Sofia University “St. Kliment Ohridski” $^{3}$ CSE, HKUST + +# Abstract + +Fusing and balancing multi-modal inputs from novel sensors for dense prediction tasks, particularly semantic segmentation, is critically important yet remains a significant challenge. One major limitation is the tendency of multimodal frameworks to over-rely on easily learnable modalities, a phenomenon referred to as unimodal dominance or bias. This issue becomes especially problematic in real-world scenarios where the dominant modality may be unavailable, resulting in severe performance degradation. To this end, we apply a simple but effective plug-and-play regularization term based on functional entropy, which introduces no additional parameters or modules. This term is designed to intuitively balance the contribution of each visual modality to the segmentation results. Specifically, we leverage the log-Sobolev inequality to bound functional entropy using functional-Fisher-information. By maximizing the information contributed by each visual modality, our approach mitigates unimodal dominance and establishes a more balanced and robust segmentation framework. A multi-scale regularization module is proposed to apply our proposed plug-and-paly term on high-level features and also segmentation predictions for more balanced multimodal learning. Extensive experiments on three datasets demonstrate that our proposed method achieves superior performance, i.e., $+13.94\%$ , $+3.25\%$ and $+3.64\%$ , without introducing any additional parameters. + +# 1. Introduction + +Multi-modal dense prediction tasks, especially for semantic segmentation [3-16], rely on integrating complementary inputs from diverse sensors (e.g., RGB, depth, thermal, event cameras) to achieve robust scene understanding [2, 17, 18]. While such fusion promises resilience to real-world variability, a critical challenge persists: models often exhibit + +![](images/2e0f5c83001f3f86dc5a576d76e10251c201e818b7e46414126f7dde8d4c1cda.jpg) +(a) Previous Methods + +![](images/0f9eb1bb5d3bb7bbd4ac1d6b81c8369be6a7f4410d25a4330a4780f588c75a32.jpg) +(b) Ours + +![](images/4703a19e70595185695a6928046a96e1386fc56fad0e10ddda5d85152f826f70.jpg) +(c) Performance Comparison + +![](images/a1cae5b96334db1d8d216aaddb992d252bb1d7ad5ae8135e0cbf7e50bbd52d21.jpg) +Figure 1. (a) Previous methods use complex architectures to fuse multi-modal data. (b) Our approach simplifies the architecture and use plug-and-play regularization terms. The results show significant improvements of ours than CMNeXt [1] and MAGIC [2]. + +![](images/67ebcafeda5f12dee96e977c21c2c403fb30088b62e427003318fc23c174ef12.jpg) + +unimodal bias, disproportionately favoring one modality over others [19, 20]. For instance, in RGB-thermal segmentation, RGB inputs—common in training pipelines—often dominate predictions, even in low-light conditions where thermal data is crucial. Similarly, systems combining event cameras with conventional sensors may under-utilize the motion-aware edge information from event cameras in dynamic scenes, defaulting to static RGB inputs. This imbalance occurs when models prioritize the most easily learned or information-rich modality, neglecting others. As a result, performance degrades when the dominant modality is missing (e.g., RGB corruption) or unreliable (e.g., thermal sensor noise), posing risks in safety-critical applications like autonomous navigation. + +Prior multi-modal methods often fail to explicitly regulate how modalities contribute to predictions [1], leaving models vulnerable to over-reliance on dominant in + +puts. Functional entropy addresses this gap by quantifying the uncertainty in a model's dependence on modalities: a system with high entropy distributes its "attention" evenly across sensors, while low entropy indicates undue reliance on a single modality. Inspired by functional entropy which has been successfully applied in visual question answering task [21], we apply this concept to the challenging multimodal semantic segmentation. However, directly applying the regularization is non-trivial because the multi-modal semantic segmentation task presents significant challenges in enabling the model to learn pixel-wise multi-modal representations while balancing the contributions of multiple modalities, in contrast to the dual-modality and VQA task addressed in [21]. + +In this paper, we apply the simple yet effective plug-and-play regularization terms at multiple scales of the segmentation model to balance various visual modalities while encourage equal multi-modal learning. As shown in Fig. 1 (c), framing modality balance as an entropy maximization problem, these regularization terms reduce bias and ensure equitable sensor utilization, crucial for segmentation accuracy when modalities degrade or fail. Unlike complex architectures [2, 18], we draw from recent multi-modal learning advancements [22], which show that intricate connectors are increasingly unnecessary. This suggests that sophisticated fusion mechanisms are not required. We propose a simpler approach: instead of adding fusion modules, we treat all input modalities equally within a mini-batch, achieving fusion by averaging predictions, avoiding unnecessary complexity in the segmentation backbone. + +To implement the proposed regularization term in multimodal semantic segmentation tasks, we introduce a multiscale regularization module that takes both multi-scale features and predictions into consideration. The prediction-level regularization is to ensure that predictions across different modalities are consistent, thereby encouraging the model to integrate information across modalities in a balanced manner. Meanwhile, the regularization with the multi-scale feature aims to enforce smoothness and balance, making sure that no single scale or modality disproportionately influences the final output. + +Extensive experiments on real-world and synthetic benchmarks demonstrate the superior robustness and performance of our method compared to existing state-of-the-art approaches, achieving mIoU improvements of $+13.94\%$ , $+3.25\%$ and $+3.64\%$ respectively. Additionally, we conduct series of ablation studies to show how the proposed regularization terms work. Our contributions are as follows: (I) We introduce functional entropy-based regularization to mitigate unimodal bias and ensure balanced multi-modal learning, reducing reliance on any single modality. (II) We propose a straightforward multi-modal fusion framework that averages input modalities within a mini-batch, avoid- + +ing complex fusion modules and simplifying the segmentation backbone. (III) Our regularization terms are applied at both feature and prediction levels, effectively balancing multi-modal training. Extensive experiments show robust performance across various benchmarks. + +# 2. Related Work + +# 2.1. Multi-modal Semantic Segmentation + +Multi-modal Semantic Segmentation seeks to combine RGB with complementary modalities such as depth [23-36], thermal [37-47], events [48-52], and LiDAR [53-60]. Advances in sensor technology have driven significant progress in multi-modal fusion [23, 51, 61-67], evolving from dual-modality to comprehensive multi-modal systems, like MCubeSNet [68], which improves scene understanding by utilizing richer sensor data. + +From an architectural standpoint, multi-modal fusion models are generally classified into three categories: separate branches [38, 69-71], joint branches [30, 55], and asymmetric branches [1, 72]. A common approach in these models is to treat RGB as the primary modality, while auxiliary sensors provide additional information. For instance, CMNeXT [1] prioritizes RGB, incorporating other sensors to supplement the data. However, RGB alone may be insufficient, particularly in challenging conditions like low light or nighttime. This limitation highlights the need for more robust fusion models that leverage the strengths of multiple modalities, minimizing dependence on any single sensor. In this context, Liu et al. [73] introduced the concept of modality-incomplete scene segmentation, addressing deficiencies at both the system and sensor levels. + +# 2.2. Unimodal Bias + +As shown in Fig.2, a key challenge in multi-modal tasks—especially semantic segmentation—is unimodal bias, where models overly rely on one modality, resulting in suboptimal performance. For example, in Tab.2, the frame camera dominates the performance of most methods: results improve markedly when it is present but degrade sharply when absent. This bias occurs when models favor the most informative or easily learnable modality under certain conditions, overlooking the complementary value of others. Consequently, performance may deteriorate when the dominant modality is corrupted or unavailable [2, 18, 73], which poses risks in safety-critical applications like autonomous navigation. While several strategies aim to mitigate this issue [2, 18], many fusion models still lack mechanisms to regulate modality contributions, often leading to over-reliance on a single input. Recent works, such as MAGIC [2] and Any2Seg [18], address this by incorporating carefully designed training objectives. + +In multi-modal learning, methods [19, 20] often draw + +![](images/bb25fbc29557fc543de301eb3ae552a4436783adf2b93a819b8f0097f5f9e438.jpg) +Figure 2. Visualization of Fisher information during training with real-world MUSES (right) and synthetic DELIVER (left) datasets. Adding a regularization term reduces the gap between modality Fisher information values. With regularization, the model utilizes both input modalities, resulting in closer modality distances. Without regularization, the model relies more on RGB in the first two rows and Event in the last two, increasing the modality distance (red arrows). + +from information theory, particularly the concept of functional entropy [21]. Entropy-based approaches quantify a model's uncertainty in its reliance on different modalities. A system with high entropy distributes its "attention" evenly across modalities, while low entropy signals an overreliance on a single modality. Building on the application of functional Fisher information in visual question answering tasks [21], directly applying this approach to semantic segmentation presents challenges. Unlike visual question answering, which focuses on interpreting questions based on a single modality, segmentation tasks require the integration of both spatial and contextual information across multiple modalities. To address this, we introduce multi-scale regularization terms at both the feature and prediction levels. These regularization terms provide a principled framework to mitigate unimodal bias, encouraging a more balanced utilization of all available sensors. This ensures that models maintain accuracy even in scenarios where certain modalities degrade or fail—an essential consideration in environments with unreliable or variable sensor performance. + +# 3. Method + +This paper focuses on multi-modal visual data from novel sensors, where each instance in a data batch $x$ consists of multiple modalities. For instance, in multi-modal semantic segmentation [1], RGB images provide color and texture information, depth maps capture geometric details, event cameras offer motion cues, and LiDAR provides accurate point clouds. A mini-batch of data from these modalities can be represented as: $x = \{x_{\mathrm{r}}, x_{\mathrm{d}}, x_{\mathrm{e}}, x_{\mathrm{l}}\}$ , where each modality resides in its own Euclidean space, i.e., $x_{i} \in \mathbb{R}^{d}$ . Here, $x_{\mathrm{r}}$ represents the RGB, $x_{\mathrm{d}}$ represents the depth, $x_{\mathrm{e}}$ represents the event, and $x_{\mathrm{l}}$ represents the LiDAR modality, with each $x_{i}$ being a vector in a $d$ -dimensional space. The goal of multi-modal semantic segmentation is to optimize the parameters $w$ of a function that assigns a probability score to each possible label $y$ given the input mini-batch $x$ . + +# 3.1. Reducing Unimodal Dominance / Bias + +In the context of multi-modal semantic segmentation, existing methods [1, 2, 18] introduce additional modules and parameters to encourage balanced training across multiple visual modalities. However, neural networks often tend to overly rely on a single modality, leading to significant challenges in real-world scenarios. For instance, as shown in Tab. 2, the frame camera has a dominant effect on the performance of most methods. Specifically, when the frame camera is included in the evaluation, the performance significantly improves. However, its absence results in a sharp decline in performance. This phenomenon, commonly referred to as unimodal dominance or bias, occurs when the learned function disproportionately emphasizes the most easily learnable modality, potentially neglecting critical information from other modalities. + +# 3.1.1. Functional Entropy + +In multi-modal learning, functional entropy quantifies the uncertainty and variability in the model's multi-modal predictions and representation [21]. Let $f(x)$ represent the model's prediction for a multi-modal input mini-batch $x = \{x_{\mathrm{r}}, x_{\mathrm{d}}, x_{\mathrm{e}}, x_{\mathrm{l}}\}$ , where $x \in \mathbb{R}^d$ includes RGB, Depth, Li-DAR, and Event data. The functional entropy of a non-negative function $f(x) \geq 0$ is defined as: + +$$ +\begin{array}{l} \operatorname {E n t} _ {\mu} (f) = \int_ {\mathbb {R} ^ {d}} f (x) \log f (x) d \mu (x) - \\ \left(\int_ {\mathbb {R} ^ {d}} f (x) d \mu (x)\right) \log \left(\int_ {\mathbb {R} ^ {d}} f (x) d \mu (x)\right), \tag {1} \\ \end{array} +$$ + +where $f(x)$ represents the model's predictions in the form of segmentation logits and multi-scale features, which reflect how the model combines different modalities. Here, $\mu$ represents a probability measure defined over the Euclidean space $\mathbb{R}^d$ . This measure describes the distribution of the multi-modal inputs, quantifying how likely each point (pixel in prediction logits or element from high-level features) in the input space is, typically based on a distribution over $\mathbb{R}^d$ . As shown in Fig. 2, a low entropy suggests + +![](images/54bb8f07ff83a27043474b785a362d42d4f41713f0252de0d506d8b86e068b63.jpg) +(a) Calculate Regularization Term + +![](images/ea60dc73ea2a456df3223b5d6da214830c35de4da84faa627ac5a2ca11756b26.jpg) +(b) Forward Propagation of Overall Framework +Figure 3. (a) The proposed regularization term for multi-modal input and (2) overall framework of our proposed multi-scale regularization terms' implementation. + +over-reliance on a single modality (e.g., RGB), while higher entropy indicates better modality integration (e.g., RGB, Depth, LiDAR, Event data). + +During training, we apply random modality dropout to expose the model to diverse input combinations. Let $x' = \{x_1, x_2, \ldots, x_i\}$ represent the available modalities during training, where $1, 2, \ldots, i \in \{r, d, e, l\}$ are a subset of modalities (RGB, Depth, Event, LiDAR), and $n \leq 4$ . This accounts for missing modalities by excluding corresponding terms from the input. For a non-negative function $f(x')$ , the functional entropy is bounded as: + +$$ +\operatorname {E n t} _ {\mu} (f) \leq \frac {1}{2} \int_ {\mathbb {R} ^ {d}} \frac {\left\| \nabla f \left(x ^ {\prime}\right) \right\| ^ {2}}{f \left(x ^ {\prime}\right)} d \mu \left(x ^ {\prime}\right), \tag {2} +$$ + +where $\| \nabla f(x') \|$ is the gradient of $f(x')$ . Large Fisher information values indicate a model's dependence on specific modalities (e.g., RGB), suggesting a need for better modality integration, even with missing inputs. + +Different from [21], which focuses solely on the image-level visual question answering task using only image and text modalities, we aim to address the more complex pixel-level understanding task by incorporating multiple visual modalities. Simply increasing the utilization of the visual modality in visual question answering is straightforward; however, when multiple visual modalities are integrated with pixel-level tasks such as semantic segmentation, the problem becomes far more challenging. These challenges stem not only from improving the utilization of the visual modality but also from effectively balancing multiple modalities. To better handle multi-modal data, we decompose the overall functional entropy across modalities, allowing the model to independently assess the contribution of each modality. This is crucial when some modalities are missing during training or inference. By isolating the + +impact of each modality, the model can adapt to modality-missing scenarios more effectively. For a multi-modal segmentation model, the functional entropy is bounded as: + +$$ +\operatorname {E n t} _ {\mu} \left(f ^ {x}\right) \leq \sum_ {i = 1} ^ {n} \operatorname {E n t} _ {\mu_ {i}} \left(f \left(x _ {i}\right)\right), \tag {3} +$$ + +where $\mu_{i}$ is the marginal distribution for modality $x_{i}$ . This decomposition enables the model to focus on each modality's role, improving efficiency and flexibility. As a result, the model provides more accurate predictions even when some modalities are missing or incomplete. + +# 3.1.2. Regularization + +To enhance multi-modal integration and prevent modality dominance, the regularization term is: + +$$ +\mathcal {R} = \lambda \sum_ {i = 1} ^ {n} \left(\int_ {\mathbb {R} ^ {d _ {i}}} \frac {\left\| \nabla_ {x _ {i}} \mathrm {C E} \left(p _ {w} (\cdot | x _ {i}) , p _ {w} (\cdot | x)\right) \right\| ^ {2}}{\mathrm {C E} \left(p _ {w} (\cdot | x _ {i}) , p _ {w} (\cdot | x)\right)} d \mu_ {i} \left(x _ {i}\right)\right) ^ {- 1} \tag {4} +$$ + +where $\lambda$ controls regularization strength. This regularization term serves to promote smoother predictions and ensures a more balanced contribution from all modalities. By discouraging over-reliance on any single modality—such as the frame camera, which may underperform in low-light conditions—it encourages the use of complementary modalities. For instance, depth data aids in capturing geometric information, while LiDAR offers enhanced precision. However, directly estimating functional entropy from the equation presents practical challenges. To address this, we leverage the log-Sobolev inequality for Gaussian measures, which provides a bound on functional entropy via functional Fisher information. This approach implicitly encourages the maximization of information across modali- + +ties. To implement this, we utilize the inverse of the information derived from Equation 4, thereby promoting a more equitable contribution from each modality. The efficacy of this regularization method is demonstrated in Fig. 2, where it improves the stability and robustness of the multi-modal segmentation process. + +# 3.2. Overall Framework + +# 3.2.1. Overview + +Functional entropy and Fisher information offer a theoretical basis for analyzing and improving multi-modal semantic segmentation. By quantifying modality integration and balance, these measures help address modality dominance, leading to more robust and generalizable predictions. This is especially crucial in multi-modal tasks, where leveraging complementary modality strengths is key to accurate segmentation. Recent advances, such as Qwen2-VL [22], suggest that complex connectors in multi-modal architectures are becoming less necessary. Early models like ViLT showed that simple linear projections could effectively process visual inputs. More recent lightweight models use MLPs in place of dedicated visual encoders, reinforcing that sophisticated fusion mechanisms are not essential for effective multi-modal integration. Building on this insight, we adopt a simple yet effective strategy: instead of introducing additional fusion modules, we treat all visual modalities equally within a mini-batch. Fusion is performed by averaging the input modalities—without modifying the backbone semantic segmentation framework. + +In the forward pass, the multi-modal mini-batch $x = \{x_{\mathrm{r}}, x_{\mathrm{d}}, x_{\mathrm{e}}, x_{\mathrm{l}}\}$ is processed directly by the encoder within the backbone, producing multi-scale, high-level feature representations $\{f_r^j, f_d^j, f_l^j, f_e^j\}_{j=1}^4$ as well as the prediction segmentation maps $\{p_r, p_d\}$ for the respective modalities. This operation is formulated as: + +$$ +\left\{f _ {r} ^ {j}, f _ {d} ^ {j} \right\} _ {j = 1} ^ {4}, \left\{p _ {r}, p _ {d} \right\} = F (\{r, d \}), \tag {5} +$$ + +where $j$ corresponds to the feature level derived from the $j$ -th transformer block of the encoder. The final prediction $p$ for the input multi-modal mini-batch $x$ is obtained by averaging the individual modality predictions: $p = \mathrm{Mean}(p_r, p_d)$ . + +# 3.2.2. Multi-scale Regularization Module + +To implement the proposed regularization term in multimodal semantic segmentation tasks, we introduce a multiscale regularization module that takes both multi-scale features and predictions into consideration. This module encourages balanced learning at both the feature and prediction levels, ensuring that all modalities contribute equally to the final segmentation output across all scales. + +Regularization with Predictions: We introduce a regularization term at the prediction level. The objective is to ensure that predictions at different scales are consis + +Table 1. Results with dual (RGB-Depth) modalities on DELIVER dataset. † denotes the MMS method from [73] and ‡ follows [74]. + +
MethodRGB-Depth FusionMean
RGB-DepthRGB-DpethRGB-Depth
CMNeXt [1]1.601.4463.5822.81
CMNeXt‡32.9748.5361.9347.81
CMNeXt†53.3953.7362.2456.45
MultiMAE [75]24.6038.5558.9440.70
MultiMAE‡19.2442.5456.6239.47
MultiMAE†52.4645.6258.6852.25
FPT [73]50.7339.6057.3849.24
MAGIC [2]37.2659.0266.8954.39
Ours55.0459.6066.4660.37
+ +tent, thereby encouraging the model to integrate information across modalities in a balanced manner. The regularization term for predictions can be written as: + +$$ +\mathcal {R} _ {\mathrm {p}} = \lambda_ {\mathrm {p}} \sum_ {i = 1} ^ {n} \left(\int_ {\mathbb {R} ^ {d _ {i}}} \frac {\| \nabla_ {x _ {i}} \mathbf {C E} (p , g t) \| ^ {2}}{\mathbf {C E} (p , g t)} d \mu_ {i} (x _ {i})\right) ^ {- 1}, \tag {6} +$$ + +where $gt$ represents the ground truth for the semantic segmentation, and $p$ is the predicted output. $\lambda_{\mathrm{p}}$ is a regularization parameter controlling the strength of the prediction-level regularization. This term penalizes inconsistencies in predictions across different modalities, encouraging the model to produce balanced outputs. + +Regularization with Features: As Fig. 3 shows, we take the multi-scale features $\{f_r^j, f_d^j\}_{j=1}^4$ from the segmentation backbone model to perform feature-level functional entropy-based regularization. This regularization enforces smoothness and balance across the multi-scale features, ensuring that no single scale or modality disproportionately influences the final output. The regularization term for features can be formulated as: + +$$ +\mathcal {R} _ {\mathrm {f}} = \lambda_ {\mathrm {f}} \sum_ {j = 1} ^ {4} \sum_ {i = 1} ^ {n} \left(\int_ {\mathbb {R} ^ {d _ {i}}} \frac {\| \nabla_ {f _ {i}} \mathrm {C E} (\{f _ {r} ^ {j} , f _ {d} ^ {j} \} , f _ {m} ^ {j}) \| ^ {2}}{\mathrm {C E} (\{f _ {r} ^ {j} , f _ {d} ^ {j} \} , f _ {m} ^ {j})} d \mu_ {i} (f _ {i} ^ {j})\right) ^ {- 1}, \tag {7} +$$ + +where $\lambda_{\mathrm{f}}$ is a regularization parameter that controls the strength of the feature-level regularization, and $f_{j}$ represents the features extracted at the $j$ -th scale. The regularization encourages consistency and smoothness in feature representations across different scales, preventing any one scale from dominating the learning process. + +# 3.3. Training Objectives + +The supervised loss uses the cross-entropy between the predicted labels $p$ and the ground truth $gt$ , as follows: + +$$ +\mathcal {L} _ {s u p} = \mathrm {C E} (p, g t). \tag {8} +$$ + +To ensure that the model generalizes effectively and integrates multi-modal information in a balanced manner, the + +Table 2. Results of segmentation validation with three modalities (frame: F, event: E, and LiDAR: L) on MUSES [76]. + +
MethodTraining with All Modalities and Evaluation with Different CombinationsMean
FELFEFLELFEL
CMX [77]2.522.353.0141.1541.252.5642.2719.30
CMNeXt [1]3.502.772.646.6310.283.1446.6610.80
MAGIC [2]43.222.6822.9543.5149.0522.9849.0233.34
Any2Seg [18]44.403.1722.3344.5149.9622.6350.0033.86
Ours60.3333.1542.5947.1953.4739.5947.8947.80
+ +Table 3. Results of segmentation validation with four modalities (RGB: R, Depth: D, Event: E, and LiDAR: L) on DELIVER [1]. + +
MethodTraining with All Modalities and Evaluation with Different CombinationsMean
RDELRDRERLDEDLELRDERDLRELDELRDEL
CMNeXt [1]0.860.490.660.3747.069.9713.752.631.732.8559.0359.1814.7359.1839.0720.77
MAGIC [2]32.6055.060.520.3963.3233.0233.1255.1655.170.2663.3763.3633.3255.2663.4040.49
Any2Seg [18]39.0260.112.070.3168.2139.1139.0460.9260.151.9968.2468.2239.0660.9568.2545.04
Ours55.5657.7432.4034.2659.6947.7047.7848.9950.2634.3753.4254.1245.2447.3351.6548.29
+ +![](images/b33e723fa5e946f7b3e23057b5e278d372378fa3efe3f48f46a2616810f2e4c4.jpg) +Figure 4. Visualization on DELIVER [1] with [1] and [2]. + +overall training objective is formulated as a combination of the supervised loss and the regularization terms: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {s u p}} + \mathcal {R} _ {\mathrm {p}} + \mathcal {R} _ {\mathrm {f}}, \tag {9} +$$ + +where $\mathcal{R}_p$ and $\mathcal{R}_f$ are the regularization terms that ensure consistency and smoothness in the model's predictions and features, respectively. + +# 4. Experiments + +Datasets. We evaluate our method on both synthetic and real-world multi-sensor datasets. MUSES[78] provides high-resolution frame, event, and MEMS LiDAR data, along with high-quality 2D panoptic labels for robust multimodal semantic segmentation. DELIVER[1] contains RGB, depth, LiDAR, and event data across 25 categories, capturing diverse environments and sensor failures. + +MCubeS [68] is a material segmentation dataset with 20 categories and paired RGB, Near-Infrared (NIR), DoLP, and AoLP images. Implementation Details. All experiments were conducted on 8 NVIDIA GPUs. The initial learning rate was set to $6 \times 10^{-5}$ and adjusted using a polynomial decay strategy with a power of 0.9 over 200 epochs. More details can be found in the suppl. mat.. + +# 4.1. Results + +Table 1 presents results on DELIVER [1] using dual modalities. Our method achieves the highest mean score of $60.37\%$ , outperforming MAGIC by $+7.98\%$ in RGB-Depth fusion and $+6.99\%$ in the overall mean. Table 2 shows results on MUSES [78]. Our method achieves mIoUs of $60.33\%$ (frame), $33.15\%$ (event), and $42.59\%$ (LiDAR), with a mean of $47.80\%$ , representing a $+13.94\%$ improvement over the previous state of the art. Table 3 further reports DELIVER results across single and dual modalities, demonstrating strong cross-modal performance. In Table 4, our method achieves $41.64\%$ mIoU on MCubeS, outperforming MAGIC $(38.00\%)$ and CMNeXt $(25.17\%)$ . Notably, we observe significant gains on combinations such as Image+Aolp, with an average improvement of $+3.64\%$ over previous methods. Figure 4 provides qualitative results on DELIVER, illustrating our method's effectiveness across diverse modalities and conditions. + +# 5. Ablation + +Effectiveness of Loss Functions Tab. 5 presents the results of different loss function combinations on the MUSES + +Table 4. Results of segmentation validation with four modalities (I: Image, Aolp: A, Dolp: D, and Nir: N) on MCubeS [68]. + +
MethodTraining with All Modalities and Evaluation with Different CombinationsMean
IADNIAIDINADANDNIADIANIDNADNIADN
CMNeXt [1]1.861.542.512.2847.9643.6745.906.997.589.9550.0748.7748.838.0651.5425.17
MAGIC [2]51.910.3234.522.6652.2452.1652.571.9836.0137.0952.4852.8252.7337.5753.0138.00
Ours46.2137.6640.4940.5643.9343.1143.7340.1039.1940.7241.7841.5342.2140.6342.7341.64
+ +![](images/fa77f9b14ea52ca40477e36dfe33eded107af3067a1fa8efd72a4aca8f45b028.jpg) +Figure 5. Comparison of model performance with (w/) and without (w/o) our proposed prediction level regularization term across different modalities (RGB-Depth, RGB, and RGB-Event). + +dataset. Introducing $\mathcal{L}_{Rp}$ improves performance, particularly for paired modalities. The mean mIoU increases to $50.02\%$ , with a $+29.44\%$ gain for the event modality. Although the frame modality slightly decreases to $59.63\%$ , the overall result improves due to substantial gains in the event $(+29.44\%)$ and both modalities $(+61.00\%)$ . Including $\mathcal{L}_{Rf}$ , yields a consistent performance improvement across all modalities. The mean mIoU rises to $50.44\%$ , with gains of $+29.66\%$ for the frame, $+61.02\%$ for event, and $+50.44\%$ for both modalities. + +Effectiveness of Regularization Term. As shown in Fig. 5, the results demonstrate consistent improvement with regularization. For example, in the RGB-Depth combination, the model with regularization (w/ Reg) reaches $65\%$ at 150 epochs, outperforming the model without regularization (w/o Reg), which peaks at $62\%$ . A similar trend is observed in RGB-Event. Fisher Information plots further confirm these results, showing that without regularization, the model over-relies on individual modalities, as indicated by higher Fisher information values for RGB and Event. With regularization, Fisher information is more balanced across modalities, indicating better utilization of both inputs. This balance is particularly evident in RGB-Depth and RGB-Event combinations, where modality distances decrease with regularization. Overall, these findings highlight the effectiveness of regularization in promoting bal + +Table 5. Ablation of different loss combinations on MUSES. + +
OptimizationmIoUMeanAccMean
LsupRpRfFrameEventBothFrameEventBoth
✓--60.636.4860.7042.6071.3414.9271.3252.53
✓✓-59.6329.4461.0050.0268.8939.3974.0160.76
✓✓✓60.6329.6661.0250.4469.9839.9874.4961.48
+ +anced modality usage, improving stability, and enhancing performance in multi-modal tasks. + +Ablation of Multi-scale Feature-level Regularization The results in Fig. 6 compare the performance of two regularization strategies: the Pred-level Reg-Term (left) and the combined Feat- & Pred-level Reg-Term (right). The comparison clearly shows that incorporating multi-scale regularization results in a smoother reduction of Fisher information over time, as seen in the right plot. In contrast, the left plot with only prediction-level regularization exhibits more fluctuations, indicating less stable convergence. This ablation emphasizes the significance of multi-scale feature-level regularization, as it improves stability and effectiveness, leading to smoother training dynamics and enhanced generalization across modalities. + +Hyper-parameters Tables 6 and 7 show the impact of $\lambda_{p}$ and $\lambda_{f}$ on the MUSES dataset. In Table 6, $\lambda_{p} = 0.3$ + +Table 6. Sensitivity analysis on the effect of different values of $\lambda_p$ on MUSES [78]. + +
λpmIoUMeanAccMean
FEFEFEFE
0.057.9328.6858.8448.4867.9537.9872.5859.50
0.158.4529.6459.4349.1768.2338.7873.0560.02
0.359.0629.9560.6749.8968.4839.9673.9360.79
0.557.8229.3159.8148.9867.0438.2873.2659.53
0.757.6029.0459.0148.5567.1938.8972.7659.61
+ +Table 7. Ablation on the effect of different ${\lambda }_{f}$ on MUSES [78]. + +
λpmIoUMeanAccMean
FEFEFEFE
0.059.0629.9560.6749.8968.4839.9673.9360.79
0.0260.6329.6661.0250.4469.8839.9874.4961.45
0.0459.9630.0960.7450.2669.3840.4474.2961.37
0.0658.8431.0860.1650.0368.6240.9674.0561.21
+ +achieves the best mIoU (59.06%) and accuracy (60.79%), outperforming both lower (0.0, 0.1) and higher values (0.5, 0.7), which especially hurt frame modality performance. Table 7 shows that $\lambda_{f} = 0.02$ yields the highest mIoU (60.63%) and accuracy (61.45%), while larger values (0.04, 0.06) lead to slight drops. These results underscore the importance of tuning, with $\lambda_{p} = 0.3$ and $\lambda_{f} = 0.02$ offering the best overall performance. + +# 6. Discussion + +Differences with [21]. Unlike [21], which focuses on image-level visual question answering with image and text modalities, we address the more complex pixel-level task by incorporating multiple visual modalities. Challenges arise from both improving modality utilization and balancing multiple modalities. When we directly apply the regularization term from [21], the resulting mIoU on MUSES is $42.89\%$ , significantly lower than our $50.44\%$ . + +Unimodal Bias. As in Fig. 7, compared to MAGIC [2] (green line), which exhibits a notable unimodal bias, our method (red line) demonstrates consistent performance across modality combinations. MAGIC shows a mean mIoU of $40.49\%$ with a high standard deviation of 22.36, reflecting sensitivity to certain modality pairs. In contrast, our method achieves a higher mean mIoU of $48.05\%$ with a smaller standard deviation of 7.91, highlighting its robustness and stability. This trend emphasizes the absence of unimodal bias in our approach, as it handles all modality combinations effectively, unlike MAGIC, which suffers from significant fluctuations. The shaded areas around the curves further emphasize the reduced variability of our method, + +![](images/bc4afe55f2d2c3cf54933e9f044b1d68489f3521a6d8358d505b71c1e83030ae.jpg) +Figure 6. Ablation of the effectiveness of multi-scale feature-level regularization. + +![](images/da331d183626513991fbe2a0e9b54dbde9b7e2ca665a9fa4b16ff990405c8860.jpg) +Figure 7. Balanced multi-modal performance comparison between ours and MAGIC [2] on (a) DELIVER and (b) MCubeS. + +demonstrating its stability and reliability in achieving balanced performance across modalities. + +Training Efficiency We compare the training efficiency of the model with and without regularization terms across 8 NVIDIA GPUs. The results show a slight increase in training time, with the model using regularization taking 2.476 hours, compared to 2.462 hours without regularization—an increase of 0.014 hours (approximately $0.57\%$ ). Despite this minimal increase, the performance improvements justify the inclusion of regularization, offering a favorable trade-off between efficiency and accuracy. Thus, regularization proves to be a valuable investment with negligible impact on training time. + +# 7. Conclusion + +We presented a simple yet effective regularization term based on functional entropy to mitigate unimodal bias in multi-modal semantic segmentation. Our approach ensures balanced utilization of all input modalities, improving robustness in scenarios where some modalities are unavailable or degraded. Extensive experiments on synthetic and real-world benchmarks demonstrated significant performance improvements, achieving better results without introducing extra parameters or complexity. Key contributions include introducing a novel regularization technique, proposing a simple fusion framework, and applying entropy maximization for balanced multi-modal learning. Our method outperforms existing approaches, offering a practical solution for robust multi-modal integration with multiple visual modalities. Future work could extend this approach to other multimodal tasks for robust and stable scene understanding. + +# 8. Acknowledgement + +# References + +[1] J. Zhang, R. Liu, H. Shi, K. Yang, S. Reiβ, K. Peng, H. Fu, K. Wang, and R. Stiefelhagen, “Delivering arbitrary-modal semantic segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1136–1147, 2023. 1, 2, 3, 5, 6, 7 +[2] X. Zheng, Y. Lyu, J. Zhou, and L. Wang, "Centering the value of every modality: Towards efficient and resilient modality-agnostic semantic segmentation," in ECCV, pp. 192–212, Springer, 2025. 1, 2, 3, 5, 6, 7, 8 +[3] Y. Miao, I. Armeni, M. Pollefeys, and D. Barath, "Volumetric semantically consistent 3d panoptic mapping," 2024. 1 +[4] Y. Miao, F. Engelmann, O. Vysotska, F. Tombari, M. Pollefeys, and D. B. Barath, "Scenegraphloc: Cross-modal coarse visual localization on 3d scene graphs," 2024. +[5] A.-M. Halacheva, Y. Miao, J.-N. Zaech, X. Wang, L. V. Gool, and D. P. Paudel, "Holistic understanding of 3d scenes as universal scene description," 2025. +[6] E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, "Segformer: Simple and efficient design for semantic segmentation with transformers," Advances in neural information processing systems, vol. 34, pp. 12077-12090, 2021. +[7] J. Chen, D. Deguchi, C. Zhang, X. Zheng, and H. Murase, "Frozen is better than learning: A new design of prototype-based classifier for semantic segmentation," Pattern Recognition, vol. 152, p. 110431, 2024. +[8] X. Zheng, Y. Luo, P. Zhou, and L. Wang, "Distilling efficient vision transformers from cnns for semantic segmentation," arXiv:2310.07265, 2023. +[9] X. Zheng, C. Fu, H. Xie, J. Chen, X. Wang, and C. Sham, "Uncertainty-aware deep co-training for semi-supervised medical image segmentation," Comput. Biol. Medicine, vol. 149, p. 106051, 2022. +[10] J. Chen, C. Fu, H. Xie, X. Zheng, R. Geng, and C. Sham, "Uncertainty teacher with dense focal loss for semi-supervised medical image segmentation," Comput. Biol. Medicine, vol. 149, p. 106034, 2022. + +This work was supported by the Guangdong Provincial Department of Education Project (Grant No.2024KQNCX028); CAAI-Ant Group Research Fund; Scientific Research Projects for the Higher-educational Institutions (Grant No.2024312096), Education Bureau of Guangzhou Municipality; Guangzhou-HKUST(GZ) Joint Funding Program (Grant No.2025A03J3957), Education Bureau of Guangzhou Municipality. +This research was partially funded by the dAIedge project (HORIZON-CL4-2022-HUMAN-02-02, Grant Agreement Number: 101120726) and the Ministry of Education and Science of Bulgaria (support for INSAIT, part of the Bulgarian National Roadmap for Research Infrastructure). It was also supported with computational resources provided by Google Cloud Platform (GCP). +[11] J. Zhou, X. Zheng, Y. Lyu, and L. Wang, "Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more," CoRR, vol. abs/2403.12534, 2024. +[12] H. Xie, C. Fu, X. Zheng, Y. Zheng, C. Sham, and X. Wang, "Adversarial co-training for semantic segmentation over medical images," Comput. Biol. Medicine, vol. 157, p. 106736, 2023. +[13] X. Zheng, P. Zhou, A. V. Vasilakos, and L. Wang, "Semantics, distortion, and style matter: Towards source-free UDA for panoramic segmentation," in IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pp. 27885-27895, IEEE, 2024. +[14] B. Ren, Y. Li, J. Liang, R. Ranjan, M. Liu, R. Cucchiara, L. V. Gool, M.-H. Yang, and N. Sebe, "Sharing key semantics in transformer makes efficient image restoration," Advances in Neural Information Processing Systems, vol. 37, pp. 7427-7463, 2024. +[15] B. Ren, Y. Liu, Y. Song, W. Bi, R. Cucchiara, N. Sebe, and W. Wang, “Masked jigsaw puzzle: A versatile position embedding for vision transformers,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20382-20391, 2023. +[16] B. Ren, G. Mei, D. P. Paudel, W. Wang, Y. Li, M. Liu, R. Cucchiara, L. Van Gool, and N. Sebe, "Bringing masked autoencoders explicit contrastive properties for point cloud self-supervised learning," in Proceedings of the Asian Conference on Computer Vision, pp. 2034-2052, 2024. 1 +[17] X. Zheng, Y. Liu, Y. Lu, T. Hua, T. Pan, W. Zhang, D. Tao, and L. Wang, “Deep learning for event-based vision: A comprehensive survey and benchmarks,” arXiv:2302.08890, 2023. 1 +[18] X. Zheng, Y. Lyu, and L. Wang, "Learning modality-agnostic representation for semantic segmentation from any modalities," in European Conference on Computer Vision, pp. 146–165, Springer, 2024. 1, 2, 3, 6 +[19] Y. Zhang, P. E. Latham, and A. M. Saxe, "Understanding unimodal bias in multimodal deep linear networks," in International Conference on Machine Learning, 2024. 1, 2 +[20] X. Zheng, H. Xue, J. Chen, Y. Yan, L. Jiang, Y. Lyu, K. Yang, L. Zhang, and X. Hu, "Learning robust anymodal segmentor with unimodal and cross-modal distillation," arXiv preprint arXiv:2411.17141, 2024. 1, 2 +[21] I. Gat, I. Schwartz, A. Schwing, and T. Hazan, “Removing bias in multi-modal classifiers: Regularization by maximizing functional entropies,” Advances in Neural Information Processing Systems, vol. 33, pp. 3197–3208, 2020. 2, 3, 4, 8 +[22] P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, et al., “Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution,” arXiv preprint arXiv:2409.12191, 2024. 2, 5 +[23] Y. Lyu, X. Zheng, D. Kim, and L. Wang, “Omnibind: Teach to build unequal-scale modality interaction for omni-bind of all,” arXiv:2405.16108, 2024. 2 +[24] Y. Lyu, X. Zheng, J. Zhou, and L. Wang, "Unibind: Llmaugmented unified and balanced representation space to bind + +them all," in Proceedings of IEEE/CVF CVPR, pp. 26752-26762, 2024. +[25] Y. Lyu, X. Zheng, and L. Wang, "Image anything: Towards reasoning-coherent and training-free multi-modal image generation," arXiv:2401.17664, 2024. +[26] Y. Wang, F. Sun, M. Lu, and A. Yao, “Learning deep multimodal feature representation with asymmetric multi-layer fusion,” in Proceedings of the 28th ACM MM, pp. 3902–3910, 2020. +[27] H. Zhou, L. Qi, Z. Wan, H. Huang, and X. Yang, “Rgb-d coattention network for semantic segmentation,” in Proceedings of the ACCV, 2020. +[28] Y. Wang, W. Huang, F. Sun, T. Xu, Y. Rong, and J. Huang, "Deep multimodal fusion by channel exchanging," NeurIPS, vol. 33, pp. 4835-4845, 2020. +[29] J. Cao, H. Leng, D. Lischinski, D. Cohen-Or, C. Tu, and Y. Li, "Shapeconv: Shape-aware convolutional layer for indoor rgb-d semantic segmentation," in Proceedings of the IEEE/CVF ICCV, pp. 7088-7097, 2021. +[30] L.-Z. Chen, Z. Lin, Z. Wang, Y.-L. Yang, and M.-M. Cheng, "Spatial information guided convolution for real-time rgbd semantic segmentation," IEEE TIP, vol. 30, pp. 2313-2324, 2021. 2 +[31] X. Ying and M. C. Chuah, “Uctnet: Uncertainty-aware cross-modal transformer network for indoor rgb-d semantic segmentation,” in ECCV, pp. 20–37, Springer, 2022. +[32] M. Lee, C. Park, S. Cho, and S. Lee, "Spsn: Superpixel prototype sampling network for rgb-d salient object detection," in ECCV, pp. 630-647, Springer, 2022. +[33] R. Cong, Q. Lin, C. Zhang, C. Li, X. Cao, Q. Huang, and Y. Zhao, "Cir-net: Cross-modality interaction and refinement for rgb-d salient object detection," IEEE TIP, vol. 31, pp. 6800-6815, 2022. +[34] W. Ji, G. Yan, J. Li, Y. Piao, S. Yao, M. Zhang, L. Cheng, and H. Lu, "Dmra: Depth-induced multi-scale recurrent attention network for rgb-d saliency detection," IEEE TIP, vol. 31, pp. 2321-2336, 2022. +[35] F. Wang, J. Pan, S. Xu, and J. Tang, "Learning discriminative cross-modality features for rgb-d saliency detection," IEEE TIP, vol. 31, pp. 1285-1297, 2022. +[36] M. Song, W. Song, G. Yang, and C. Chen, "Improving rgb-d salient object detection via modality-aware decoder," IEEE TIP, vol. 31, pp. 6124-6138, 2022. 2 +[37] S. S. Shivakumar, N. Rodrigues, A. Zhou, I. D. Miller, V. Kumar, and C. J. Taylor, "Pst900: Rgb-thermal calibration, dataset and segmentation network," in 2020 IEEE ICRA, pp. 9441-9447, IEEE, 2020. 2 +[38] Q. Zhang, S. Zhao, Y. Luo, D. Zhang, N. Huang, and J. Han, "Abmdrnet: Adaptive-weighted bi-directional modality difference reduction network for rgb-t semantic segmentation," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2633-2642, 2021. 2 +[39] W. Wu, T. Chu, and Q. Liu, "Complementarity-aware cross-modal feature fusion network for rgb-t semantic segmentation," Pattern Recognition, vol. 131, p. 108881, 2022. +[40] G. Liao, W. Gao, G. Li, J. Wang, and S. Kwong, "Cross-collaborative fusion-encoder network for robust rgb-thermal + +salient object detection," IEEE TCSVT, vol. 32, no. 11, pp. 7646-7661, 2022. +[41] W. Zhou, H. Zhang, W. Yan, and W. Lin, “Mmsmcnet: Modal memory sharing and morphological complementary networks for rgb-t urban scene semantic segmentation,” IEEE TCSVT, 2023. +[42] Z. Xie, F. Shao, G. Chen, H. Chen, Q. Jiang, X. Meng, and Y.-S. Ho, "Cross-modality double bidirectional interaction and fusion network for rgb-t salient object detection," IEEE TCSVT, 2023. +[43] G. Chen, F. Shao, X. Chai, H. Chen, Q. Jiang, X. Meng, and Y.-S. Ho, "Modality-induced transfer-fusion network for rgb-d and rgb-t salient object detection," IEEE TCSVT, vol. 33, no. 4, pp. 1787-1801, 2022. +[44] Y. Pang, X. Zhao, L. Zhang, and H. Lu, "Caver: Cross-modal view-mixed transformer for bi-modal salient object detection," IEEE TIP, vol. 32, pp. 892-904, 2023. +[45] T. Hui, Z. Xun, F. Peng, J. Huang, X. Wei, X. Wei, J. Dai, J. Han, and S. Liu, "Bridging search region interaction with template for rgb-t tracking," in Proceedings of IEEE/CVF CVPR, pp. 13630-13639, 2023. +[46] T. Zhang, H. Guo, Q. Jiao, Q. Zhang, and J. Han, "Efficient rgb-t tracking via cross-modality distillation," in Proceedings of IEEE/CVF CVPR, pp. 5404-5413, 2023. +[47] J. Zhao, F. Teng, K. Luo, G. Zhao, Z. Li, X. Zheng, and K. Yang, "Unveiling the potential of segment anything model 2 for rgb-thermal semantic segmentation with language guidance," CoRR, vol. abs/2503.02581, 2025. 2 +[48] I. Alonso and A. C. Murillo, “Ev-segnet: Semantic segmentation for event-based cameras,” in Proceedings of IEEE/CVF CVPR Workshops, pp. 0–0, 2019. 2 +[49] J. Zhang, K. Yang, and R. Stiefelhagen, "Issafe: Improving semantic segmentation in accidents by fusing event-based data," in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1132-1139, IEEE, 2021. +[50] X. Zheng and L. Wang, “Eventdance: Unsupervised source-free cross-modal adaptation for event-based object recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17448–17458, 2024. +[51] J. Cao, X. Zheng, Y. Lyu, J. Wang, R. Xu, and L. Wang, "Chasing day and night: Towards robust and efficient all-day object detection guided by an event camera," arXiv:2309.09297, 2023. 2 +[52] J. Zhou, X. Zheng, Y. Lyu, and L. Wang, "Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 18633-18643, 2024. 2 +[53] Z. Zhuang, R. Li, K. Jia, Q. Wang, Y. Li, and M. Tan, "Perception-aware multi-sensor fusion for 3d lidar semantic segmentation," in Proceedings of the IEEE/CVF ICCV, pp. 16280-16290, 2021. 2 +[54] X. Yan, J. Gao, C. Zheng, C. Zheng, R. Zhang, S. Cui, and Z. Li, “2dpass: 2d priors assisted semantic segmentation on lidar point clouds,” in ECCV, pp. 677–695, Springer, 2022. + +[55] Y. Wang, X. Chen, L. Cao, W. Huang, F. Sun, and Y. Wang, "Multimodal token fusion for vision transformers," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12186-12195, 2022. 2 +[56] Y. Li, A. W. Yu, T. Meng, B. Caine, J. Ngiam, D. Peng, J. Shen, Y. Lu, D. Zhou, Q. V. Le, et al., "Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection," in Proceedings of IEEE/CVF CVPR, pp. 17182-17191, 2022. +[57] S. Borse, M. Klingner, V. R. Kumar, H. Cai, A. Almuzairee, S. Yogamani, and F. Porikli, "X-align: Cross-modal cross-view alignment for bird's-eye-view segmentation," in Proceedings of the IEEE/CVF WACV, pp. 3287-3297, 2023. +[58] B. Zhang, Z. Wang, Y. Ling, Y. Guan, S. Zhang, and W. Li, "Mx2m: Masked cross-modality modeling in domain adaptation for 3d semantic segmentation," in Proceedings of the AAAI, vol. 37, pp. 3401-3409, 2023. +[59] H. Liu, T. Lu, Y. Xu, J. Liu, W. Li, and L. Chen, "Camliflow: Bidirectional camera-lidar fusion for joint optical flow and scene flow estimation," in Proceedings of IEEE/CVF CVPR, pp. 5791-5801, 2022. +[60] J. Li, H. Dai, H. Han, and Y. Ding, "Mseg3d: Multi-modal 3d semantic segmentation for autonomous driving," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 21694-21704, 2023. 2 +[61] W. Zhang, Y. Liu, X. Zheng, and L. Wang, "Goodsam: Bridging domain and capacity gaps via segment anything model for distortion-aware panoramic semantic segmentation," in IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pp. 28264-28273, IEEE, 2024. 2 +[62] W. Zhang, Y. Liu, X. Zheng, and L. Wang, "Good-sam++: Bridging domain and capacity gaps via segment anything model for panoramic semantic segmentation," CoRR, vol. abs/2408.09115, 2024. +[63] X. Zheng, Y. Lyu, L. Jiang, J. Zhou, L. Wang, and X. Hu, "MAGIC++: efficient and resilient modality-agnostic semantic segmentation via hierarchical modality selection," CoRR, vol. abs/2412.16876, 2024. +[64] C. Liao, X. Zheng, Y. Lyu, H. Xue, Y. Cao, J. Wang, K. Yang, and X. Hu, "Memorysam: Memorize modalities and semantics with segment anything model 2 for multi-modal semantic segmentation," CoRR, vol. abs/2503.06700, 2025. +[65] C. Liao, K. Lei, X. Zheng, J. Moon, Z. Wang, Y. Wang, D. P. Paudel, L. V. Gool, and X. Hu, "Benchmarking multimodal semantic segmentation under sensor failures: Missing and noisy modality robustness," CoRR, vol. abs/2503.18445, 2025. +[66] D. Zhong, X. Zheng, C. Liao, Y. Lyu, J. Chen, S. Wu, L. Zhang, and X. Hu, “Omnisam: Omnidirectional segment anything model for UDA in panoramic semantic segmentation,” CoRR, vol. abs/2503.07098, 2025. +[67] X. Zheng and L. Wang, “Eventdance++: Language-guided unsupervised source-free cross-modal adaptation for event-based object recognition,” CoRR, vol. abs/2409.12778, 2024. 2 +[68] Y. Liang, R. Wakaki, S. Nobuhara, and K. Nishino, "Multimodal material segmentation," in Proceedings of IEEE/CVF CVPR, pp. 19800-19808, 2022. 2, 6, 7 + +[69] T. Broedermann, C. Sakaridis, D. Dai, and L. Van Gool, "Hrfuser: A multi-resolution sensor fusion architecture for 2d object detection," in IEEE International Conference on Intelligent Transportation Systems, pp. 4159-4166, 2023. 2 +[70] S. Wei, C. Luo, and Y. Luo, "Mmanet: Margin-aware distillation and modality-aware regularization for incomplete multimodal learning," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20039-20049, 2023. +[71] Y. Man, L.-Y. Gui, and Y.-X. Wang, “Bev-guided multimodality fusion for driving perception,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 21960–21969, 2023. 2 +[72] J. Zhang, H. Liu, K. Yang, X. Hu, R. Liu, and R. Stiefelhagen, "Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers," arXiv:2203.04838, 2022. 2 +[73] R. Liu, J. Zhang, K. Peng, Y. Chen, K. Cao, J. Zheng, M. S. Sarfraz, K. Yang, and R. Stiefelhagen, "Fourier prompt tuning for modality-incomplete scene segmentation," in IEEE Intelligent Vehicles Symposium, pp. 961-968, 2024. 2, 5 +[74] Y.-L. Lee, Y.-H. Tsai, W.-C. Chiu, and C.-Y. Lee, "Multimodal prompting with missing modalities for visual recognition," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14943-14952, 2023. 5 +[75] R. Bachmann, D. Mizrahi, A. Atanov, and A. Zamir, “Multimae: Multi-modal multi-task masked autoencoders,” in European Conference on Computer Vision, pp. 348–367, Springer, 2022. 5 +[76] T. Brodermann, D. Bruggemann, C. Sakaridis, K. Ta, O. Liagouris, J. Corkill, and L. Van Gool, “Muses: The multi-sensor semantic perception dataset for driving under uncertainty,” in ECCV, pp. 21–38, Springer, 2025. 6 +[77] J. Zhang, H. Liu, K. Yang, X. Hu, R. Liu, and R. Stiefelhagen, "Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers," IEEE Transactions on intelligent transportation systems, 2023. 6 +[78] T. Brodermann, D. Bruggemann, C. Sakaridis, K. Ta, O. Liagouris, J. Corkill, and L. Van Gool, "Muses: The multi-sensor semantic perception dataset for driving under uncertainty," in Proceedings of the European Conference on Computer Vision, 2024. 6, 8 \ No newline at end of file diff --git a/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/images.zip b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b6bff17629508125cff69dfed5eb67830685a129 --- /dev/null +++ b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15baf9a1069afbf204140d7067484a522b687e8c80a872afc798e18cac19f333 +size 693828 diff --git a/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/layout.json b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..28519694d93f293ce984f4e8c4bfcb024b0a9a94 --- /dev/null +++ b/ICCV/2025/Reducing Unimodal Bias in Multi-Modal Semantic Segmentation with Multi-Scale Functional Entropy Regularization/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1416459e1eb243e762ee23ed2f90b11de7ca20a1875b3e6cd852a3243722297 +size 440704 diff --git a/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/43bba521-e7da-412c-921a-566f4f8a1b53_content_list.json b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/43bba521-e7da-412c-921a-566f4f8a1b53_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fca790597c84d927357dc1e728b071aa2de02ed5 --- /dev/null +++ b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/43bba521-e7da-412c-921a-566f4f8a1b53_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4de2fbc9bf314569b5bbfeaad1f0dbf352f446dbe672406586d61760af7d0ea1 +size 81668 diff --git a/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/43bba521-e7da-412c-921a-566f4f8a1b53_model.json b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/43bba521-e7da-412c-921a-566f4f8a1b53_model.json new file mode 100644 index 0000000000000000000000000000000000000000..af4b166a22ed55a3da8256cfa71d48973a4f2fc7 --- /dev/null +++ b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/43bba521-e7da-412c-921a-566f4f8a1b53_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93dbc1eb2351a3abddf2c535400a5edcef8bd9cbf829155b659583984329db46 +size 99012 diff --git a/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/43bba521-e7da-412c-921a-566f4f8a1b53_origin.pdf b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/43bba521-e7da-412c-921a-566f4f8a1b53_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7e061cab6e35c82ca399fec82f3203499a53e333 --- /dev/null +++ b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/43bba521-e7da-412c-921a-566f4f8a1b53_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dd9c07274d9dddf1802429732407b1a423d1fd1497758791e1c6ef17aed21e1 +size 16117385 diff --git a/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/full.md b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b4387f6f4e182b004b0ede4354aedba6eca7acb5 --- /dev/null +++ b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/full.md @@ -0,0 +1,363 @@ +# RefEdit: A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions + +Bimsara Pathiraja* Maitreya Patel* Shivam Singh Yezhou Yang Chitta Baral Arizona State University + +{bpathirl, maitreya.patel, ssing631, yz.yang, chitta}@asu.edu + +http://refedit.verbcl.app + +![](images/48ebc644336f27334dca1bd2f5dab57bc2fe9cfc0ae6207e119eee2a166a8837.jpg) +Figure 1. RefEdit is a referring expression-based image editing benchmark and a finetuned model. Our proposed RefEdit model can accurately identify the entity of interest and perform accurate edits. + +![](images/c27b05d809bdce351a32b26b78d6031eadcc3d1a58d7ada628dc22794bfbea02.jpg) + +![](images/047860080f4f893a820eeff4ecb4d1fc5404a928631d05ff78c363f4356f728c.jpg) + +# Abstract + +Despite recent advances in inversion and instruction-based image editing, existing approaches primarily excel at editing single, prominent objects but significantly struggle when applied to complex scenes containing multiple entities. To quantify this gap, we first introduce RefEdit-Bench, a rigorous real-world benchmark rooted in RefCOCO, where even baselines trained on millions of samples perform poorly. To overcome this limitation, we introduce RefEdit - an instruction-based editing model trained on our scalable synthetic data generation pipeline. Our RefEdit, trained on only 20,000 editing triplets, outperforms the Flux/SD3 model-based baselines trained on millions of data. Extensive evaluations across various benchmarks demonstrate that our model not only excels in referring expression tasks but also enhances performance on traditional benchmarks, achieving + +state-of-the-art results comparable to closed-source methods. We release data & checkpoint for reproducibility. + +# 1. Introduction + +The field of AI for content creation (AICC) stands at the forefront of technological innovation, propelled by unprecedented advances in visual generative models [10, 31, 32, 36, 38-40]. Techniques such as diffusion [9, 41] and flow-based models [21, 24, 34] have revolutionized image editing, enabling precise semantic transformations guided by user queries. Early efforts centered on inversion-based strategies—like Prompt-to-Prompt [12], Null-Text Inversion [29], and edit-friendly DDPM-inversion [14] which map reference image latents into noise and deterministically reconstruct the original. While these methods laid the critical groundwork, they are hindered by significant computational demands, prolonged processing times, and the need for exten + +sive hyperparameter tuning. In response, instruction-based approaches, including InstructPix2Pix [5], UltraEdit [50], and OmniEdit [44], have gained prominence by fine-tuning diffusion models on millions of image editing triplets derived from inversion-based techniques, offering a more efficient alternative despite inheriting some foundational limitations. + +These advancements, however, reveal a persistent weakness: current methods struggle to maintain precision in complex scenarios, particularly when images feature multiple similar entities (i.e., "two cats"). Inversion-based strategies, though innovative, are computationally intensive and falter when tasked with isolating specific regions, a flaw that propagates to instruction-based models trained on their outputs. This imprecision often results in failed edits or unintended modifications spilling into adjacent areas. Synthetic data has been leveraged extensively across various domains beyond image editing [25, 33]; however, existing pipelines within image editing rarely address multi-object complexity, further exacerbating their shortcomings. As a result, the field faces an unresolved bottleneck that limits the practical utility of image editing technologies in real-world applications. + +We identify this bottleneck as the referring expression challenge—a problem well-explored in discriminative tasks like segmentation [45] but largely neglected in the generative domain of image editing. To confront this, we introduce RefEdit-Bench, a rigorous benchmark designed to evaluate image editing under both straightforward and demanding conditions. By repurposing the RefCOCO dataset [45], we ensure real-world relevance, capturing intricate scenarios with multiple entities. As evidenced in Table 3, leading baselines—such as MagicBrush [48], InstructDiffusion [11], and UltraEdit [50]—experience marked performance declines on our benchmark, highlighting its complexity. + +To address this challenge, we propose a scalable synthetic data generation pipeline tailored to the referring expression challenge, harnessing GPT-4o [1] to craft sophisticated editing tasks and integrating Grounded Segment Anything [37] with primarily a Flux-based inversion-free editing methodology (FlowChef [34]) for precise, controlled edits. Using this pipeline, we fine-tune diffusion models—Stable Diffusion v1.5 [38] and Stable Diffusion 3 [10]—to create RefEdit, a model that redefines the state-of-the-art in image editing. Extensive experiments, human evaluations, and detailed analyses confirm RefEdit's superiority, with our Stable Diffusion v1.5 variant outperforming baselines trained on vastly larger datasets and more advanced architectures like SDXL [35] and SD3. This breakthrough not only overcomes the referring expression challenge but also sets a new standard for efficiency and effectiveness in visual generative modeling. In summary, our contributions are: + +- We present RefEdit-Bench, a novel benchmark for realistic, referring expression-driven image editing. +- We introduce a novel, FlowChef and Inpaint Anything + +[46]-powered pipeline for automated, scalable synthetic data generation. +- Through comprehensive evaluation, we establish RefEdit as the most human-preferred model, achieving exceptional success rates and outperforming existing methods. + +# 2. Related works + +Inversion-based editing. Inversion-based image editing leverages generative models to modify existing images by mapping them into a latent space, applying desired alterations, and then reconstructing the edited images. Prompt-to-Prompt [12] leverages cross attention maps to edit images by solely modifying the textual prompts. Negative-text Inversion [28] and Edict [43] uses DDIM/DDPM inversions to preserve the original image while editing the image. Other methods like LEdits [42], Sega [4], Null-text inversion [29], Direct Inversion [16] provides more efficient and reliable inversion based image editing methods. Still, the inversion based techniques suffer from leaking the editing to background areas and poor localization abilities. Notably we utilize FlowChef [34] for our pipeline as it provides a gradient-and inversion-free framework for image editing. + +Region-based controlled editing. Region-based editing approaches perform image modifications within user-provided masked regions guided by textual instructions. Blended Diffusion [2, 3] facilitates local inpainting based on user-supplied captions and masks. SDEdit [27] leverages pre-trained models to iteratively apply noise and denoising to real images guided by text prompts. DiffEdit [8] integrates DDIM inversion for automatic mask generation to preserve image backgrounds during editing. MAGEdit [26] optimizes latent noise features via mask-based cross-attention constraints, progressively improving local alignment with desired textual prompts. Despite their effectiveness, region-based methods require additional user effort in mask annotation. + +Instruction-based editing. Instruction-based editing methods allow users to modify images solely through textual descriptions of intended changes. InstructPix2Pix [6] utilizes image prompts and GPT-3-generated editing instructions [7], alongside image synthesis techniques from Prompt2Prompt [13]. Subsequently, MagicBrush [48] introduced manually annotated datasets for InstructPix2Pix coupled with fine-tuned model checkpoints. InstructDiffusion [11] extends beyond editing, incorporating broader computer vision tasks like segmentation and keypoint detection. OmniEdit [44] applies a specialist-to-generalist training strategy, consolidating diverse specialized editing capabilities into a single instruction-based model. UltraEdit [50] employs a large-scale, automatically generated + +![](images/1328811d8368c61ff28968bfe032ca040bdf4f81862d7a2737965d910a016a2e.jpg) +Figure 2. Three samples from each of Easy and Hard categories and the overlayed mask of the interested object are shown. + +![](images/2e59480454e872c91c85fdde62df9d64b47ab621629366f15ca8d3823fc5b01b.jpg) + +
DatasetsAutomatic generated?#Edits#Editing typesMultiple similar entities?Avg. #similar entitiesUse of referring expressions?
InstructPix2Pix [6]✓450K4✕1✕
MagicBrush [48]✕10K7✕1✕
HQ-Edit [15]✓200K7✕1✕
UtraEdit [50]✓4M9+✕1✕
InstructDiffusion [11]✓425K6✕1✕
HIVE [49]✓1.1M5+✕1✕
OmniEdit [44]✓1.2M7✕1✕
RefEdit-Data (ours)✓20K5✓3✓
+ +Table 1. Comparison of different image editing datasets. + +dataset containing approximately 4 million samples for enhanced instruction-based image editing. Even though ReferDiffusion [22] addresses referring expressions based image editing, the training dataset is based on 400 initial images and its data generation pipeline is primarily manual, hindering scalability. Additionally, the work does not provide a benchmark for the community to facilitate comparative evaluations. AURORA [18] creates a dataset and a benchmark using videos and simulation engines but we find that the number of real-world relevant referring expression-based image editing tasks are very limited and mixed with other tasks in the benchmark. + +# 3. Method + +This section introduces the RefEdit synthetic data creation framework to train the RefEdit model that can perform complex real-world edits. First, we will detail the preliminaries about diffusion-based instruction-guided image editing. + +Next, we will present our synthetic data creation pipeline. Finally, we will share our training details. + +# 3.1. Preliminaries + +Diffusion models are generative models that create images by reversing a diffusion process. In the forward process, Gaussian noise is incrementally added to an image across multiple time steps until it becomes pure noise. The reverse process, parameterized by a neural network, begins with noise and iteratively denoises it to produce a realistic image. The model is trained to predict the noise added at each step, minimizing the loss function: + +$$ +L = \mathbb {E} [ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, c) \| ^ {2} ], \tag {1} +$$ + +where $z_{t}$ represents the noisy latent at time $t$ , $\epsilon$ is the actual noise, and $c$ denotes conditioning information, such as a text prompt. + +# (a) Generate image and edit prompts + +Caption + +generation + +template + +![](images/ef86f0ed627ec27382faa0fd45edfda35e7aa5bce81532132c663c9d62effaa8.jpg) + +GPT-40 + +![](images/7475582a27b6e6b6f2903aa6e7499fee5666e9b36d301438dd716119eab034f2.jpg) + +Editing + +Instruction + +types + +(change color) + +Image prompt - "... the chef cutting vegetables has a white + +shirt, and the chef stirring a pot has a black shirt.' + +Edited prompt - "... the chef cutting vegetables has a red shirt, and the chef stirring a pot has a black shirt." + +Editing instruction - "Change the color of the shirt of the + +chef cutting vegetables to red" + +Editing object - "shirt" + +Referring expression - "Shirt of the chef cutting + +vegetables" + +Descriptive referring expression - "White shirt of the chef cutting vegetables" + +# (b) Generate image and mask + +![](images/650b58062113e61c09a55d662a583b042583cc46bbc878b08ddf6de6ebbfc6e6.jpg) + +FLUX.1 [dev] + +Image prompt + +Grae rree + +Anything. + +...the chef + +cutting vegetables + +has a white + +and the chef + +stirring a pot has a + +black shirt." + +![](images/e894e70db58c2214a2367c033c41f9535d92a77636af6b2f1147b8cee4708c8c.jpg) + +the chef cutting vegetables" + +Descriptive referring + +expression + +![](images/72850f724b7bfb49667b2e5f765c17ac77c6dcd3b8e2779187983228e0aadc9c.jpg) + +![](images/bcbd9f16226ba0dbc5e04c09a931b3b16198a2887593e6cbd38da23dd3a04ee3.jpg) + +# (c) Generate the edited image + +![](images/7de79f62230c63a72772f2f874d081128fbb63e88c71d009a1b9a49e4d5ac3cc.jpg) + +...thechefcutting + +vegetables has a + +red shirt, and the + +chef stirring a pot + +has a black shirt.' + +FlowChef + +Edited prompt + +![](images/307f77aa0e2633839bea65d087b3087717177b7ef2dda5de7b27f7f23a0b6be2.jpg) +Figure 3. Overview of our pipeline for generating controlled image edits. GPT-4o [30] produces the textual information. Based on that, FLUX [20] and Grounded Segment Anything [37] identify masks based on these expressions. At last, we use FlowChef [34], mask-based inversion-free technique, for editing the images. + +...the chef cutting + +vegetables has a + +white shirt, and the + +chef stirring a pot + +has a green shirt." + +![](images/47210341b796f8da8d53ff5d9b84f73656b0a8fe02dd9d89b38409aa9523b1af.jpg) + +Instructpix2pix extends Stable Diffusion to enable text-guided image editing by conditioning the diffusion process on both an input image $c_{I}$ and a text instruction $c_{T}$ . This modification allows the model to generate edited images that align with textual descriptions while retaining relevant aspects of the original image. The training objective is adapted to: + +$$ +L = \mathbb {E} _ {\mathcal {E} (x), \mathcal {E} (c _ {I}), c _ {T}, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, \mathcal {E} \left(c _ {I}\right), c _ {T}\right) \| _ {2} ^ {2} \right] \tag {2} +$$ + +where $\mathcal{E}(c_I)$ is the encoded input image. By leveraging a synthetic dataset comprising image pairs and corresponding edit instructions, Instructpix2pix learns to apply precise, instruction-driven modifications, offering an efficient and flexible approach to image editing without requiring perexample fine-tuning. + +# Remove the child flying a bright yellow kite + +![](images/d61b9d13a80f1566265f985cb7024b5797166e99d28a078ec0d3c8638bb363bd.jpg) + +![](images/76c258f61c4cf975f4cc1ea2de2deda2fb8814b6e6bd2072fe07f94827c77b3a.jpg) + +![](images/903a16ca3cdbdf087eac31beb83f409120d5a85ce3a2846680d00c587f2b3be9.jpg) + +# Change the dog in the middle to a dalmatian + +![](images/ae020ee897d8603e0506ff48d0d6c09a1cb6883bba406c8c001249c05764ce15.jpg) + +![](images/d615fe64fd951a975d7c5c4a87cb66c1f71b4355a14a5882bec63fdd8299c517.jpg) + +![](images/654057cc4533f5c685c52c059f304affafbe2bbc2c2b351eba5168242c5fcd01.jpg) + +# Change the material of the costume worn by the dancer with the red fan + +![](images/6d545a94636f9c4cafa4b63061100ad5608477d92c1978a92cf84343a85800fe.jpg) +Figure 4. Two samples from our training dataset. From left to right: the initial image, the bounding box highlighting the object of interest, and the edited image. + +![](images/5ef0201c3437274d8a05cff84ed3cebdc3493327f3e5ee8d0c5ae7af8b2e2790.jpg) + +![](images/6eded745e67a33164ce2675a6e78b02f1e998827002c4da9a8338179ebefa66b.jpg) + +![](images/5fc0433e46f3dc3d020bfa2d305df4938c8bfbe2744cff8328c4d0c4ffce48ce.jpg) +Figure 5. An overview of keywords in edit instructions. The inner circle depicts the types of edits and outer circle showcases the most frequent words used within each type. + +# 3.2. RefEdit-Data: Synthetic data generation pipeline + +The central challenge of instruction-based image editing is the data. Therefore, existing works rely on some form of synthetic data creation (see Table 1). One way to create such editing triplets is inversion-based editing methods, which do + +
Editing TasksDefinitionInstruction example
Change colorDescribes the new color of an object instance. The initial color may or may not be given.Can the horse which has a blue color rein be white? +Change the shirt color of the tennis player on right to red.
Change objectDescribes the object to be replaced with a new object in its placeLet the coffee drink be a matcha drink. +Replace the computer monitor with two green lines on it with a computer monitor displaying a vibrant beach scene.
Add objectDescribes a new object to add by the location the object should be added.Put a red cap on the elderly man with a blue apron. +Add a bird to the bench next to the bench with the bird on it.
Remove objectDescribes the object instance that should be removed.Remove the half full mug of beer to the left of the pizza. +Get rid of the woman in a yellow shirt.
Change textureDescribes how to change the texture, pattern or material of an object instanceTurn the man wearing surfing suit and holding surfboard into a silver statue. +Let the man with the red t-shirt who is playing on his phone wear a stripped red shirt.
+ +Table 2. Task Definitions and examples from RefEdit-Bench. + +not require any training. However, inference-time editing is a very ill-posed problem statement. It requires careful hyperparameter selection for each prompt, and inversion is not guaranteed to work. Despite these challenges, existing approaches utilize these methods to create million-scale diverse editing triplets. Furthermore, these inversion-based editing frameworks fail measurably in referring expressions (see Table 3). Hence, the challenges of inversion-based editing methods limit instruction-based editing and require a lot of careful data creation process. Additionally, we found that existing referring expression datasets (e.g., RefMSCOCO) were very difficult for the existing method, and the failure rate was very high, which made them useless. Therefore, we propose a new and simplified data generation framework. + +In Figure 3, we outline the four key steps of our framework. For training data generation we start with generating all the text data. We utilize GPT-4o [30] to generate the image prompt, edited prompt editing instruction, edited object, referring expression, and a more expanded version of the same referring expression. We group the image editing task into five different categories, as detailed in Table 2. Importantly, each category focuses on the referring expression. We usually generate two sets of editing information per the same image prompt. In the template, we specify how to generate the image prompts, such as if there are two or more instances of the same object but with some differences. The difference could be due to the color ("A white shirt" or a + +"black shirt"), a separate item ("person holding a knife" or "a person holding a bag"), or an activity like in this case ("chef cutting vegetables" or "chef stirring the pot"). We generate a $1024 \times 1024$ image using FLUX [20]. The referring expression is used to generate the editing instruction, but we ask the LLM to generate a descriptive referring expression for ease of mask generation. Grounded SAM [37] uses Segment Anything [17] and Grounding DINO [23] to generate accurate masks of objects using referring expressions. The initial image, the editing instructions, and the mask are used for generating the edited image. We utilized Inpaint Anything [47] for object removals and FlowChef [34] for other editing tasks. Adding object data were created by swapping the initial image and the edited image. With the help of this pipeline, we create a $20000+$ paired training dataset. Figure 4 shares the three examples of our training dataset. Each image contains multiple similar entities that require some form of referring expression to pinpoint. Figure 5 shows that our RefEdit data is very diverse and well distributed across the editing categories. + +# 3.3. Training Details + +We use the finetuning configuration of MagicBrush [48] where the training starts with the pre-trained InstructPix2Pix model. To avoid overfitting on our RefEdit-Data and to ensure that our method can generalize to other tasks, we combine MagicBrush data with our own generated data dur + +
MethodEasyHard
\(SC_{avg}\)↑\(PQ_{avg}\)↑\(O_{avg}\)↑\(SC_{avg}\)↑\(PQ_{avg}\)↑\(O_{avg}\)↑
P2P [12]2.75.462.311.715.31.66
P2P (Direct Inversion)2.965.842.652.425.92.3
PnP4.255.593.552.455.622.27
InstructPix2Pix3.525.662.823.806.023.22
MagicBrush4.186.103.674.116.163.56
HIVE3.155.262.423.225.522.47
InstructDiffusion4.164.713.314.094.943.16
HQ-Edit2.585.392.372.955.212.6
OmniEdit*4.186.153.684.076.283.53
AURORA4.015.273.234.356.053.69
InstructDiffusion-HA5.015.884.214.566.13.75
RefEdit (ours)5.475.854.684.516.483.93
CosXLEdit3.26.732.883.446.652.88
FLUX-OmniEdit1.952.461.311.551.920.89
UltraEdit3.686.983.42.936.232.59
RefEdit-SD3 (ours)5.466.364.994.476.213.98
+ +ing the training. This results in the 30K amount of combined training data. Specifically, we train the RefEdit Stable Diffusion v1.5 variant on this data following the Eq. (1) for 24 epochs on $2 \times 80$ GB NVIDIA A100 GPUs. Similarly, we train RefEdit-SD3 (UltraEdit) for 6000 iterations on $8 \times 80$ GB NVIDIA A100 GPUs. + +# 4. RefEdit-Bench + +To evaluate the models' performance on more complex and semantically rich image editing tasks, we introduce the RefEdit Benchmark, a novel evaluation benchmark dataset designed to assess the capability of models in handling referring expression-based editing instructions. The benchmark leverages images sourced from the RefCOCO dataset [45], which provides a rich collection of images annotated with referring expressions. Each image in RefEdit is carefully selected to ensure diversity in scenes, objects, and contexts. Additionally, we manually craft detailed and varied editing instructions for each image, encompassing tasks such as changing color, changing objects, adding objects, removing objects, and changing texture. + +The benchmark is divided into Easy and Hard categories, in which each category has 100 images. In the Easy category, the images had primarily one object. Even if there were multiple instances of the same object, mostly the interested object occupies a significant portion of the image or can easily be identified. In the Hard category, there are multiple instances of the same object, and mostly those instances occupy the same areas of the image, making isolating the + +correct instance hard for the model. Some chosen examples from both Easy and Hard are shown in Figure 2. + +# 5. Experiments + +In this section, we first discuss the baselines and evaluation setup. Later, we make extensive comparisons on RefEdit-Bench and PIE-Bench [16] along with human evaluations. + +# 5.1. Experimental setup + +Table 3. Modified VIEScore evaluation results on RefEdit benchmark for both Easy and Hard categories. The best value is bolded and the second-best value is underlined. $\mathbf{O}_{\mathrm{avg}}$ is the overall VIEScore. GPT-4o is used as the MLLM. + +
ModelSCavg↑PQavg↑Oavg↑
InstructPix2Pix4.006.623.96
MagicBrush5.026.804.79
InstructDiffusion4.495.514.06
OmniEdit5.187.225.12
InstructDiffusion-HA4.696.864.54
RefEdit (ours)5.656.465.21
CosXLEdit5.127.785.11
FLUX-OmniEdit2.373.592.14
RefEdit-SD3 (ours)6.046.475.70
+ +Table 4. VIEScore evaluation results on PIE-bench. The best value is bolded and the second-best value is underlined. GPT-4o is used as the MLLM. + +Baselines. We evaluate multiple instruction-based image editing models across the various base models at scale. We + +![](images/087b7acb471baf6c26f223025e767d76fe2d6f75f0c4009be23d85c225d72d48.jpg) +Human Evaluation: RefEdit-SD3 vs Baselines + +![](images/4fb14d5dc7fe16853a63d98e028fd6b709fdeba891d7a2f68191a64586faea59.jpg) +Figure 6. Human preference analysis for image editing, RefEdit-SD3 vs. all. Here, win indicates the winning rate for RefEdit-SD3. + +select pre-trained models built on top of Stable Diffusion v1.5, such as InstructPix2Pix, MagicBrush, InstructDiffusion, HQ-Edit, AURORA and HIVE. Most of these baselines use the InstructPix2Pix as the base model and finetune it using respective proposed data. Similarly, we train additional baseline OmniEdit on SDv1.5 using the released 1.2M editing dataset. Apart from this, we consider bigger baselines such as CosXLEdit (Stable Diffusion XL), UltraEdit (Stable Diffusion 3), and Flux-OmniEdit. As OmniEdit model is not public, we utilize the open-source replication*. We present two variants, RefEdit & RefEdit-SD3, trained on our proposed synthetic data. + +Evaluation Metrics. We perform extensive evaluations on our proposed RefEdit-Bench and PIE-Bench for referring expression and general tasks, respectively. We incorporate PIE-Bench as it is, which utilizes several different metrics focusing on background preservation and image editing. Importantly, it utilizes the ground truth mask to determine the region of interest. However, as CLIP is not good at referring expression, we instead utilize VIEScore [19] as an alternative evaluation metric for both benchmarks, which is a training-free visual instruction-guided metric leveraging multimodal LLMs (MLLMs). It contains two metrics: Semantic Consistency $(SC)$ and Perceptual Quality $(PQ)$ . $SC$ determines the alignment of the edited image with the editing instruction and $PQ$ determines the authenticity and the naturalness of the image. Following prompting templates from OmniEdit [44], we prompt GPT-4o [30] on $SC$ and $PQ$ . And we report the average performance. At last, we calculate the overall score as: $O = \sqrt{SC \times PQ}$ . To further improve the referring expression ability of GPT-4o, after PIE-Bench, we incorporate the ground truth mask to extract the region of interest. We call this metric Modified VIEScore. + +# 5.2.Quantitative Evaluation + +RefEdit-Bench Evaluations. We provide the VIEScore based evaluations in Table 3 for easy and hard categories. It can be observed that the RefEdit (SDv1.5) variant consistently outperforms the baselines by significant margins on both categories. Interestingly, MagicBrush trained on small amount of human annotated data performs similar to OmniEdit's 1.2 million synthetic data. While RefEdit consistently tops the benchmark. This shows the importance of small but high-quality datasets. Interestingly, bigger models (CosXLEdit, UltraEdit, etc.) perform subpar even worse than the SDv1.5 based methods and observes the biggest drop. We attribute this behavior to potential overfitting. As existing training datasets do not contain complex image editing often requiring referring expression, larger models tend to overfit. Importantly, adding only 20000 of our synthetic data improves the performance over UltraEdit significantly across both categories. + +PIE-Bench Evaluations. Apart from RefEdit-Bench, we evaluate the methods on standard image editing tasks using PIE-Bench (see Table 4). It can be observed that RefEdit further improves the performance over the baselines despite not containing any additional training data. Importantly, OmniEdit achieves a 5.12 score with the help of a 1.2M dataset while RefEdit gets 5.21 with only 20000 additional training data. Our RefEdit-SD3 variant achieves the SOTA performance. Importantly, all methods drop performance on RefEdit-Bench compared to the PIE-Bench. That shows the difficulty of our RefEdit-Bench tasks. + +# 5.3. Qualitative Evaluations. + +We perform human evaluations, where we ask the human annotators to pick the model that performs the task accurately or whether both models at hand lose or win. Specifically, randomly selected 400 pairs from RefEdit-Bench and asked the annotators to conduct A/B testing. Figure 6 shows the human evaluation results. It can be observed that our RefEdit-SD3 + +![](images/825e60304b6bab4235350194f7b1fde4ecfc16e0aa094ff03419e9f382010794.jpg) +Figure 7. Qualitative results on image editing. The top 3 samples are from the Easy category and the bottom 3 samples are from the Hard category. As illustrated, our method attains the SOTA performance on comparison of all the methods. + +model significantly outperforms all baselines and is preferred consistently. Furthermore, Figure 7 shows the qualitative results from RefEdit-Bench in both Easy and Hard categories. While baselines either leak the edit to background regions or don't perform edits at all, our model RefEdit-SD3 can recognize the object of interest and consistently performs edits while maintaining the background information. We provide more examples in the appendix. + +# 6. Conclusion + +In this paper, we introduce RefEdit-Bench, a challenging benchmark designed specifically for evaluating image edit + +ing models on referring expressions, and RefEdit, a novel model that significantly surpasses the existing state-of-the-art in this domain. By leveraging a carefully designed synthetic data generation pipeline, we demonstrate that training RefEdit on a relatively small yet highly specialized dataset effectively outperforms existing baselines trained on considerably larger datasets. Our results underscore the potential of targeted, quality-focused synthetic data for improving model precision in complex editing scenarios. Future research could extend our pipeline to additional editing scenarios and explore further integration with multimodal AI frameworks. + +# Acknowledgment + +This work was supported by NSF RI grant #2132724. We thank the NSF NAIRR initiative, Research Computing (RC) at Arizona State University (ASU), and cr8dl.ai for their generous support in providing computing resources. The views and opinions of the authors expressed herein do not necessarily state or reflect those of the funding agencies and employers. + +# 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] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 18187-18197. IEEE, 2022. 2 +[3] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Transactions on Graphics, 42(4):1-11, 2023. 2 +[4] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36: 25365-25389, 2023. 2 +[5] 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. 2 +[6] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions, 2023. 2, 3 +[7] 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. 2 +[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1 +[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 *Forty-first International Conference on Machine Learning*, 2024. 1, 2 + +[11] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Han Hu, Dong Chen, and Baining Guo. Instructdiffusion: A generalist modeling interface for vision tasks, 2023. 2, 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. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 6 +[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022. 2 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[15] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 3 +[16] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2, 6 +[17] 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. 5 +[18] Benno Krojer, Dheeraj Vattikonda, Luis Lara, Varun Jampani, Eva Portelance, Chris Pal, and Siva Reddy. Learning action and reasoning-centric image editing from videos and simulation. Advances in Neural Information Processing Systems, 37:38035-38078, 2024. 3 +[19] Max Ku, Dongfu Jiang, Cong Wei, Xiang Yue, and Wenhu Chen. Viescore: Towards explainable metrics for conditional image synthesis evaluation, 2024. 7 +[20] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2023.4.5 +[21] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. 1 +[22] Chang Liu, Xiangtai Li, and Henghui Ding. Referring image editing: Object-level image editing via referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13128-13138, 2024. 3 +[23] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5 +[24] Xingchao Liu, Chengyue Gong, and qiang liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, 2023. 1 +[25] Lin Long, Rui Wang, Ruixuan Xiao, Junbo Zhao, Xiao Ding, Gang Chen, and Haobo Wang. On llms-driven synthetic data generation, curation, and evaluation: A survey. arXiv preprint arXiv:2406.15126, 2024. 2 + +[26] Qi Mao, Lan Chen, Yuchao Gu, Zhen Fang, and Mike Zheng Shou. Mag-edit: Localized image editing in complex scenarios via mask-based attention-adjusted guidance. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6842-6850, 2024. 2 +[27] 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. arXiv preprint arXiv:2108.01073, 2021. 2 +[28] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 2 +[29] 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. 1, 2 +[30] OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Madry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, Alex Nichol, Alex Paino, Alex Renzin, Alex Tachard Passos, Alexander Kirillov, Alexi Christakis, Alexis Conneau, Ali Kamali, Allan Jabri, Allison Moyer, Allison Tam, Amadou Crookes, Amin Tootoochian, Amin Tootoonchian, Ananya Kumar, Andrea Vallone, Andrej Karpathy, Andrew Braunstein, Andrew Cann, Andrew Codispoti, Andrew Galu, Andrew Kondrich, Andrew Tulloch, Andrey Mishchenko, Angela Baek, Angela Jiang, Antoine Pelisse, Antonia Woodford, Anuj Gosalia, Arka Dhar, Ashley Pantuliano, Avi Nayak, Avital Oliver, Barret Zoph, Behrooz Ghorbani, Ben Leimberger, Ben Rossen, Ben Sokolowsky, Ben Wang, Benjamin Zweig, Beth Hoover, Blake Samic, Bob McGrew, Bobby Spero, Bogo Giertler, Bowen Cheng, Brad Lightcap, Brandon Walkin, Brendan Quinn, Brian Guerraci, Brian Hsu, Bright Kellogg, Brydon Eastman, Camillo Lugaresi, Carroll Wainwright, Cary Bassin, Cary Hudson, Casey Chu, Chad Nelson, Chak Li, Chan Jun Shern, Channing Conger, Charlotte Barette, Chelsea Voss, Chen Ding, Cheng Lu, Chong Zhang, Chris Beaumont, Chris Hallacy, Chris Koch, Christian Gibson, Christina Kim, Christine Choi, Christine McLeavey, Christopher Hesse, Claudia Fischer, Clemens Winter, Coley Czarnecki, Colin Jarvis, Colin Wei, Constantin Koumouzelis, Dane Sherburn, Daniel Kappler, Daniel Levin, Daniel Levy, David Carr, David Farhi, David Mely, David Robinson, David Sasaki, Denny Jin, Dev Valladares, Dimitris Tsipras, Doug Li, Duc Phong Nguyen, Duncan Findlay, Edede Oiwoh, Edmund Wong, Ehsan Asdar, Elizabeth Proehl, Elizabeth Yang, Eric Antonow, Eric Kramer, Eric Peterson, Eric Sigler, Eric Wallace, Eugene Brevdo, Evan Mays, Farzad Khorasani, Felipe Petroski Such Filippo Raso, Francis Zhang, Fred von Lohmann, Freddie Sulti Gabriel Goh Gene Oden Geoff Salmon Giulio Starace Greg Brockman,Hadi Salman,Haiming Bao,Haitang Hu Hannah Wong,Haoyu WangHeather SchmidtHeather Whitney Heewoo Jun Hendrik KirchnerHenrique Ponde de Oliveira Pinto Hongyu Ren Huiwen Chang Hyung Won + +Chung, Ian Kivlichan, Ian O'Connell, Ian O'Connell, Ian Osband, Ian Silber, Ian Sohl, Ibrahim Okuyucu, Ikai Lan, Ilya Kostrikov, Ilya Sutskever, Ingmar Kanitscheider, Ishaan Gulrajani, Jacob Coxon, Jacob Menick, Jakub Pachocki, James Aung, James Betker, James Crooks, James Lennon, Jamie Kiros, Jan Leike, Jane Park, Jason Kwon, Jason Phang, Jason Teplitz, Jason Wei, Jason Wolfe, Jay Chen, Jeff Harris, Jenia Varavva, Jessica Gan Lee, Jessica Shieh, Ji Lin, Jiahui Yu, Jiayi Weng, Jie Tang, Jieqi Yu, Joanne Jang, Joaquin Quinonero Candela, Joe Beutler, Joe Landers, Joel Parish, Johannes Heidecke, John Schulman, Jonathan Lachman, Jonathan McKay, Jonathan Uesato, Jonathan Ward, Jong Wook Kim, Joost Huizinga, Jordan Sitkin, Jos Kraaijeveld, Josh Gross, Josh Kaplan, Josh Snyder, Joshua Achiam, Joy Jiao, Joyce Lee, Juntang Zhuang, Justyn Harriman, Kai Fricke, Kai Hayashi, Karan Singhal, Katy Shi, Kavin Karthik, Kayla Wood, Kendra Rimbach, Kenny Hsu, Kenny Nguyen, Keren Gu-Lemberg, Kevin Button, Kevin Liu, Kiel Howe, Krithika Muthukumar, Kyle Luther, Lama Ahmad, Larry Kai, Lauren Itow, Lauren Workman, Leher Pathak, Leo Chen, Li Jing, Lia Guy, Liam Fedus, Liang Zhou, Lien Mamitsuka, Lilian Weng, Lindsay McCallum, Lindsey Held, Long Ouyang, Louis Feuvrier, Lu Zhang, Lukas Kondraciuk, Lukasz Kaiser, Luke Hewitt, Luke Metz, Lyric Doshi, Mada Aflak, Maddie Simens, Madelaine Boyd, Madeleine Thompson, Marat Dukhan, Mark Chen, Mark Gray, Mark Hudnall, Marvin Zhang, Marwan Aljubeh, Mateusz Litwin, Matthew Zeng, Max Johnson, Maya Shetty, Mayank Gupta, Meghan Shah, Mehmet Yatbaz,Meng Jia Yang,Mengchao Zhong Mia GlaeseMianna Chen Michael Janner Michael Lampe Michael Petrov Michael Wu,Michele Wang,Michelle Fradin Michelle Pokrass,Miguel Castro,Miguel Oom Temudo de Castro,Mikhail Pavlov,Miles Brundage,Miles Wang,Minal Khan,Mira Murati Mo Bavarian Molly LinMurat Yesildal Nacho Soto,Natalia Gimelshein,Natalie Cone,Natalie Staudacher,Natalie Summers,Natan LaFontaine,Neil Chowdhury Nick Ryder Nick Stathas Nick Turley,Nik Tezak Nik Felix Nithanth Kudige Nitish Keskar Noah Deutsch Noel Bundick,Nora Puckett Ofir Nachum,Ola Okelola,Oleg Boiko,Oleg MurkOliver JaffeOlivia WatkinsOlivier Godement Owen Campbell-Moore Patrick ChaoPaul McMillan,Pavel Belov,Peng Su,Peter Bak,Peter Bakkum,Peter Deng,Peter Dolan,Peter Hoeschele,Peter Welinder Phil Tillet Philip Pronin Philippe Tillet Prafulla Dhariwal Qiming Yuan,Rachel Dias,Rachel Lim,Rahul Arora,Rajan Troll,Randall Lin,Rapha Gontijo Lopes,Raul PuriReah Miyara Reimar Leike,Renaud Gaubert Reza Zamani Ricky Wang,Rob Donnelly,Rob Honsby,Rocky Smith,Rohan Sahai,Rohit Ramchandani,Romain Huet,Rory Carmichael Rowan Zellers,Roy Chen,Ruby Chen,Ruslan Nigmatullin Ryan Cheu,Saachi Jain,Sam Altman,Sam Schoenholz,Sam Toizer,Samuel Miserendino,Sandhini Agarwal,Sara CulverScott Ethersmith Scott Gray Sean GroveSean Metzger. Shamez Hermani Shantanu Jain Shengjia Zhao Sherwin WuShino JomotoShirong WuShuaiqiXiaSonia Phene Spencer Papay,Srinivas Narayanan Steve CoffeySteve Lee Stewart Hall Suchir Balaji Tal Broda Tal Stramer Tao Xu Tarun Gogineni,Taya ChristiansonTed Sanders Tejal Pat + +wardhan, Thomas Cunninghamman, Thomas Degry, Thomas Dimson, Thomas Raoux, Thomas Shadwell, Tianhao Zheng, Todd Underwood, Todor Markov, Toki Sherbakov, Tom Rubin, Tom Stasi, Tomer Kaftan, Tristan Heywood, Troy Peterson, Tyce Walters, Tina Eloundou, Valerie Qi, Veit Moeller, Vinnie Monaco, Vishal Kuo, Vlad Fomenko, Wayne Chang, Weiyi Zheng, Wenda Zhou, Wesam Manassra, Will Sheu, Wojciech Zaremba, Yash Patil, Yilei Qian, Yongjik Kim, Youlong Cheng, Yu Zhang, Yuchen He, Yuchen Zhang, Yujia Jin, Yunxing Dai, and Yury Malkov. Gpt-4o system card, 2024. 4, 5, 7 +[31] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. λ-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. ArXiv, abs/2402.05195, 2024. 1 +[32] Maitreya Patel, Changhoon Kim, Sheng Cheng, Chitta Baral, and Yezhou Yang. Eclipse: A resource-efficient text-to-image prior for image generations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9069-9078, 2024. 1 +[33] Maitreya Patel, Naga Sai Abhiram Kusumba, Sheng Cheng, Changhoon Kim, Tejas Gokhale, Chitta Baral, et al. Triplet-clip: Improving compositional reasoning of clip via synthetic vision-language negatives. Advances in neural information processing systems, 37:32731-32760, 2024. 2 +[34] Maitreya Patel, Song Wen, Dimitris N. Metaxas, and Yezhou Yang. Steering rectified flow models in the vector field for controlled image generation, 2024. 1, 2, 4, 5 +[35] 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 +[36] 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. 1 +[37] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 2, 4, 5 +[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, 2 +[39] 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. +[40] 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. 1 +[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1 +[42] Linoy Tsaban and Apolinário Passos. Ledits: Real image editing with ddpm inversion and semantic guidance. arXiv preprint arXiv:2307.00522, 2023. 2 +[43] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22532-22541, 2023. 2 +[44] Cong Wei, Zheyang Xiong, Weiming Ren, Xinrun Du, Ge Zhang, and Wenhu Chen. Omniedit: Building image editing generalist models through specialist supervision. arXiv preprint arXiv:2411.07199, 2024. 2, 3, 7 +[45] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling context in referring expressions, 2016. 2, 6 +[46] 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. 2 +[47] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting, 2023. 5 +[48] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing, 2024. 2, 3, 5 +[49] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9026-9036, 2024. 3 +[50] Haozhe Zhao, Xiaojian Shawn Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. Advances in Neural Information Processing Systems, 37:3058-3093, 2025. 2, 3 \ No newline at end of file diff --git a/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/images.zip b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c0e16bf81e33af73dbcfc9af2974a03966fdb1a3 --- /dev/null +++ b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f280be485f90ffa1242d591ebb7c3c49aadf7362e49f6bb5592eacbfc1596be +size 1156798 diff --git a/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/layout.json b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d296e726c7e7121262b74b362e510bafdc9aeca4 --- /dev/null +++ b/ICCV/2025/RefEdit_ A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expressions/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca26f29c65a10618146ad70ee7641709c3ecea3ca343be673339e2f771c2127e +size 356765 diff --git a/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/980e7edd-bb53-44da-b915-887c539e6e01_content_list.json b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/980e7edd-bb53-44da-b915-887c539e6e01_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5b4c5da257e20353b41bf48200e5c9dc7f393b2d --- /dev/null +++ b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/980e7edd-bb53-44da-b915-887c539e6e01_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96b30292d4aaf28df430bc6e3e9552567e25ee4fa88e92229e67c550642dfe17 +size 82297 diff --git a/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/980e7edd-bb53-44da-b915-887c539e6e01_model.json b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/980e7edd-bb53-44da-b915-887c539e6e01_model.json new file mode 100644 index 0000000000000000000000000000000000000000..62cf72681341c5344a3bcc5c0b72941b1210de48 --- /dev/null +++ b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/980e7edd-bb53-44da-b915-887c539e6e01_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1336117b5fdfc4a90976580021625f0fcf2e78fde6c28df8152b6f6762179146 +size 107463 diff --git a/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/980e7edd-bb53-44da-b915-887c539e6e01_origin.pdf b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/980e7edd-bb53-44da-b915-887c539e6e01_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2284a24c8785c80fe5fa817cd6f561dc14ac3bcf --- /dev/null +++ b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/980e7edd-bb53-44da-b915-887c539e6e01_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94039862603bf681ca353f9ad83abab871c073bae5407c9c9a5847548cb607af +size 1546292 diff --git a/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/full.md b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6d95679c4192659a5a6f9187ea30472d1e43167f --- /dev/null +++ b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/full.md @@ -0,0 +1,308 @@ +# Refer to Any Segmentation Mask Group With Vision-Language Prompts + +Shengcao Cao1 Zijun Wei2 Jason Kuen2 Kangning Liu2 Lingzhi Zhang2 +Jiuxiang Gu2 HyunJoon Jung2 Liang-Yan Gui1* Yu-Xiong Wang1* +1University of Illinois Urbana-Champaign 2Adobe + +{cao44,lgui,yxw}@illinois.edu {zwei,kuen,kangning1,lingzzha,jigu,hjung}@adobe.com + +![](images/a102e9f3abcea18301c9e6d218bc1c7023770e5bac1801b34ebd2165b1b6a6cd.jpg) +Select all spools that are not green. + +![](images/581b8380acef2fb43c2654f6e6b64022f5efe69c49f7f01f62840e16d7acc5db.jpg) +Find all players wearing red jackets. + +![](images/51c1cc12b8166000f156222522de1aac208691d08dc4c8199484ee4db0086a3e.jpg) +Select the deer decorations. +Figure 1. Omnimodal referring expression segmentation (ORES) according to arbitrary vision-language prompts. (Left) Our approach, Refer to Any Segmentation Mask Group (RAS), can understand a complex text prompt involving multiple conditions. (Middle) Reference visual entities can be included as visual prompts to enhance expressivity, addressing the challenge of describing the same details using language alone. (Right) The grouped segmentation masks conveniently enable various fine-grained downstream applications, such as object removal and editing. In each pair of images, the left one is the input and the right one is the output. Best viewed on an electronic device with zoom-in functionality. + +![](images/ef2ca3b756b0528e78ad63d4d5e71e24a794cb21c1bd5df4c3288ce7ed5ba113.jpg) +Water bottles similar to . + +![](images/d3c89f29b9193af5c218e8b8b03180ad86095b567b4fa562b9b34590b16f48f6.jpg) +All objects on . + +![](images/bb78b3640ff30c9925b197ac489130aa1cd1637d11d97cb2eaa3aeaff407edc0.jpg) +Locate all pillows on . + +![](images/5de67900ecad0400ec153ed036bdc8a5a582ad74892ad4f0af14e3777e868528.jpg) +Remove them. +Similar pattern + +# Abstract + +Recent image segmentation models have advanced to segment images into high-quality masks for visual entities1, and yet they cannot provide comprehensive semantic understanding for complex queries based on both language and vision. This limitation reduces their effectiveness in applications that require user-friendly interactions driven by vision-language prompts. To bridge this gap, we introduce a novel task of omnimodal referring expression segmentation (ORES). In this task, a model produces a group of masks based on arbitrary prompts specified by text only or text plus reference visual entities. To address this new challenge, we propose a novel framework to "Refer to Any Segmentation Mask Group" (RAS), which augments segmentation models with complex multimodal interactions and + +comprehension via a mask-centric large multimodal model. For training and benchmarking ORES models, we create datasets MASKGROUPS-2M and MASKGROUPS-HQ to include diverse mask groups specified by text and reference entities. Through extensive evaluation, we demonstrate superior performance of RAS on our new ORES task, as well as classic referring expression segmentation (RES) and generalized referring expression segmentation (GREs) tasks. Project page: https://Ref2Any.github.io. + +# 1. Introduction + +Referring expression segmentation (RES) [18, 34, 69] enables language-based object segmentation by relating a text prompt with a segmentation mask for the referred target, and can be generalized (i.e., GRES [34]) for multiple targets. However, real-world applications, such as autonomous driving [9, 52], robotics [14, 37, 64], augmented + +reality [25], and image editing [15, 41, 53], require additional flexibility in the prompt. As exemplified in Figure 1, a user may want to specify a segmentation instruction that involves a relationship/comparison/interaction with a reference visual entity (see red regions in the middle column). Expressing the reference via visual and text prompts together is usually preferred over text-only descriptions in such cases, because language may not be able to concisely and accurately locate the reference entity or describe its intricate characteristics in a complex image. + +In this work, we introduce a novel task of omnimodal referring expression segmentation (ORES) with vision-language prompts (Figure 1). For a given image, the system generates a group of relevant masks that satisfy a user-specified instruction, which can be a) a text-only prompt describing a property (e.g., category, attribute, position, or their combination) of the targets, or b) a vision-enriched prompt that provides masks of reference entities and expresses a complex property involving the reference entities. + +Unfortunately, it is not straightforward to extend existing models for this new challenge (Table 1). Interactive segmentation models (e.g., SEEM [77]) accept both text and visual prompts, but visual prompts only lead to the directly indicated entity rather than any related ones. In contrast, ORES aims to return a group of relevant masks based on the prompt, providing a more contextually cohesive response to the user's input. The visual prompt can be provided to conveniently describe a relationship between the target(s) and a reference visual entity that is difficult to describe verbally. Some grounding large multimodal models (LMMs) accept mask or region prompts [49, 68, 73], but they are designed for descriptive tasks [4, 71], not for segmentation. Another limitation of most existing models (except GRES models [34]) is that they generate only one target per query, even if multiple targets are available in the image. + +
ParadigmPromptTarget
TextMaskMaskGroup
Interactive segmentation [24, 77]○○✓×
RES [18, 65], GRES [34]✓×✓○
LMM [10, 35, 75]✓×××
Grounding LMM [27, 49, 73]✓○✓○
ORES (Ours)✓✓✓✓
+ +Table 1. Comparison with existing paradigms. Our omnimodal referring expression segmentation (ORES) task poses new challenges for all prior methods, including allowing mask-based visual prompts for reference visual entities and predicting a group of masks. $\checkmark$ : supported, $\circ$ : partially supported, $\times$ : unsupported. + +To address the ORES challenge, we adopt a simple yet effective approach that enables vision-language comprehension at the mask level. We extend segmentation foundation models [24] with multimodal semantic understanding + +of segmentation masks to leverage the strengths of both segmentation foundation models and LMMs: Segmentation foundation models, such as SAM [24], benefit from large-scale training data with fine-grained mask annotations, but have limited semantic understanding of the produced masks; LMMs excel at language-based comprehension of visual inputs, but datasets that can train LMMs for pixel-level grounding are much smaller in scale [27]. With such insight, we propose a framework, Refer to Any Segmentation Mask Group (RAS), to bridge the strengths of both sides. We first leverage the visual entity masking ability of segmentation models to propose a pool of candidate masks, which effectively covers the true targets, as we will show in Section 5.3. Then, we introduce a mask-centric LMM with enhanced semantic understanding of each visual entity encapsulated by candidate masks. + +Specifically, our RAS framework a) employs a segmentation foundation model to propose candidate masks, b) extracts semantic-rich visual feature maps with an ensemble of visual backbones [56], c) produces entity-level visual features by aggregating the features within each masked region to form mask tokens, and d) aligns mask tokens with a language model through visual instruction tuning [35]. Notably, reference masks in vision-enriched prompts can be naturally converted into mask tokens as part of the input. In this mask-centric formulation, each mask token is designed to encode one visual entity instead of a fixed-size image patch. This approach is more suitable for modeling the semantics of individual entities and their interactions. + +ORES requires the model to output a group of target masks, which is essentially a set prediction problem [51] and is known to pose difficulties in model optimization [54]. To facilitate optimization, we adopt a non-autoregressive decoding [2, 59] procedure in RAS. Instead of letting the model output the selected masks one by one autoregressively [27, 62, 73] in the prediction stage, we feed all candidate mask tokens into the model and learn to perform binary classification on each contextualized mask token to decide whether this candidate should be included in the group or not. With this design, we avoid directly predicting a sequence of mask embeddings [27, 73] and effectively convert the set prediction problem into an easy-to-optimize per-mask binary classification problem. + +To learn RAS, we construct a large-scale instruction-tuning dataset $\mathsf{MASKGROUPS - 2M}$ by automatically repurposing object-level annotations from existing datasets [16, 26, 32, 34, 69]. Based on labeled categories, attributes, and relationships of objects, we create 2 million mask groups for visual instruction tuning. Furthermore, in order to align RAS with user preferences and evaluate its performance in real-world applications, we collect a high-quality mask grouping dataset $\mathsf{MASKGROUPS - HQ}$ by requesting expert human annotators to propose meaningful visual en + +tity groups and select the corresponding masks. In summary, our main contributions include: + +- We introduce the omnimodal referring expression segmentation (ORES) task, which extends the classic RES and GRES tasks with vision-language prompts for more flexible and practical use cases. +- We propose the Refer to Any Segmentation Mask Group (RAS) framework to strengthen the semantic understanding of segmentation masks with a mask-centric LMM and produce mask groups for vision-language prompts. +- We build a large-scale dataset MASKGROUPS-2M for instruction tuning of RAS and curate MASKGROUPS-HQ for alignment with human preferences and evaluation. +- Empirical results demonstrate state-of-the-art performance of our solution on the newly proposed ORES dataset, as well as classic RES and GRES benchmarks. + +# 2. Related Work + +Referring expression segmentation (RES) aims to segment one object at a time based on descriptions in natural language [18, 20, 40, 69]. Earlier approaches focused on combining visual and language features [3, 13, 19, 29, 33, 67] and incorporating transformer models [21, 61, 65, 66]. Recent advancements [34] have expanded the classic RES task to include multi-target and no-target queries, referred to as generalized RES (GRES). Building on this progress, our work further enhances GRES by enabling more effective and user-friendly interactions through the flexible integration of visual and textual inputs. + +Large multimodal models (LMMs) extend large language models (LLMs) [11, 47, 58] with vision-language capabilities via visual instruction tuning [10, 35, 75]. Early LMMs are mainly based on CLIP [48] patch-level visual features and show weaknesses in object-level comprehension and reasoning [30, 55-57]. LMMs can be equipped with grounding capabilities for generating bounding boxes [4, 28, 43, 44, 60, 68] or segmentation masks [27, 49, 50, 73] via training on converted datasets with semantic-pixel alignment. Unlike prior LMMs, our RAS is not trained for text generation, because text responses are unnecessary in the task and applications we consider (Figure 1), and high-quality mask groups are prioritized. + +Grounding LMMs provide grounded vision-language understanding, and achieve state-of-the-art performance in RES and GRES [5, 8, 27, 49, 63, 73, 74]. Among them, Groundhog [73] is most related to our work, which also performs RES by selecting from mask proposals. Our RAS differs from Groundhog in these critical aspects: a) We adopt a non-autoregressive decoding procedure, outperforming the traditional autoregressive decoding used by Groundhog (Section 5.3). b) We accept mask-based visual prompts for the complex ORES task, while for Groundhog, mask prompts are only effective in region description tasks. + +# 3. Refer to Any Segmentation Mask Group + +Extending the decoupling strategy in open-vocabulary segmentation [17, 31, 70], our approach leverages a segmentation model to propose candidate masks for a given image. As the segmentation model does not directly comprehend the complex vision-language prompt, we design a mask-centric large multimodal model (LMM) to address the new challenge of understanding and grouping these masks. Our proposed framework, Refer to Any Segmentation Mask Group (RAS, illustrated in Figure 2), includes several specialized components: a segmentation model that proposes candidates, a mask projector that encodes mask features, a binary selection classifier that determines which masks to include, and a non-autoregressive decoding procedure for more effective model optimization. We introduce the model designs in Section 3.1 and training procedure in Section 3.2. + +# 3.1. Architecture Designs + +LMM as meta-architecture. The widely adopted LMM architecture [35] uses a CLIP visual encoder [48] to extract features from a given image, and then maps the visual features into the language feature space via a lightweight image projector. The converted visual tokens are concatenated with text tokens to form a sequence and fed into a large language model (LLM) to generate output responses autoregressively [47]. Although LMMs have acquired strong image-level vision-language capabilities, they are not originally designed for tasks focusing on understanding fine-grained visual entities. Therefore, to perform the mask grouping task, we enhance LLaVA-1.5 [36] (finetuned from Vicuna-13B [7]) with the ability to encode mask representations and select masks according to input prompts. + +Mask tokenization. After segmenting the image into candidate masks, we tokenize the masks into individual elements for the LLM to understand. Given a segmentation mask (either proposed by segmentation models or specified by users) plus feature maps extracted by visual encoders from the entire image, we perform mask pooling to aggregate visual features within the mask. More specifically, the mask is downsampled to the same spatial size as each visual feature map, and visual features within the downsampled mask are averaged to produce the mask-level feature. Then, a lightweight mask projector converts the concatenated mask-level features into the language feature space, and finally we consider these converted features as mask tokens. This procedure is depicted in the mask tokenization block in Figure 2. Furthermore, we preprocess a learnable special token $<\text{mask-pool-pre}>$ to each token that corresponds to a candidate mask. This special token indicates that the following token will be a mask token converted from a continuous embedding of a mask in the pool of candidates. These mask tokens are concatenated with the global visual tokens and text tokens as the LLM inputs. + +![](images/df899c62fd014d38280fc0c940df6313d0372ceec366dbc61235eeb6f4c2a74f.jpg) +Figure 2. Overview of our Refer to Any Segmentation Mask Group (RAS) framework. We extend LLaVA-1.5 [36] with a segmentation model, a visual encoder ensemble, mask tokenization, and a binary selection classifier for mask grouping. The decoding procedure of the LLM is non-autoregressive [2], as the input tokens are given as candidate mask tokens rather than predicted from previous tokens. + +Reference mask representation. In the input, reference masks are encoded similarly to the candidate masks. For instance, in the prompt "Select all objects with the same color as ," the special token would be replaced by the actual token of the reference mask specified by the user. We can reuse the mask tokenization for encoding candidate mask tokens, but one minor difference is that we pretend different special tokens to candidate mask tokens and reference mask tokens, in order to distinguish their roles. For candidate masks, we pretend to indicate that the next token will be an embedding for a candidate mask, and the model should interpret it as one possible choice for the mask grouping task. For reference masks, we pretend to indicate that the next token will be a reference mask embedding, which allows the model to extract related information for the mask grouping prompt. + +Visual feature ensemble. Recent studies [56, 57] suggest that the CLIP visual encoder [48] has several inherent weaknesses such as unlocalized features. Therefore, CLIP alone is not ideal for our mask grouping task, as we need localized features to represent and distinguish different masks. Following Cambrian-1 [56], we employ an ensemble of four visual encoders: CLIP [48], SigLIP [72], ConvNeXt-based CLIP [6, 38], and DINOv2 [42]. In addition, we create a position-aware feature map from 2D sinusoidal position embeddings [12] to explicitly encode positions. To tokenize each mask, after computing mask-pooled features from all encoders, we concatenate them all to produce an aggregated mask feature, and then use the mask projector to map the feature into a language-aligned token. Since the mask pool + +ing operation is performed per feature map, we allow different resolutions for each visual encoder, which better preserves the original capabilities of each encoder. + +Mask group decoding. Given a set of candidate masks, our model needs to predict a subset (group) of masks according to the prompt. For mask group prediction, a straightforward solution may be directly predicting continuous mask embeddings one by one in an autoregressive manner [27, 62, 73]. However, this is challenging because a) LLMs are originally trained to model a distribution over discrete tokens, and b) learning to predict an unordered set is inherently hard due to the instable bipartite matching between predictions and the ground truth [54]. For consistency with the discrete nature of LLMs and effective model optimization, we reformulate the mask group prediction problem as a per-mask binary classification problem (Figure 2). More specifically, we learn to make a binary prediction for each candidate mask to indicate whether it should be included in the mask group based on the input prompt. We first provide all the tokens that encode the context, and then feed the candidate mask tokens again to the LLM to capture their output hidden states. Leveraging the strong semantic understanding and reasoning capabilities of the LLM, the output hidden state can indicate whether a candidate mask is positively related to the user prompt. Finally, a learnable binary selection classifier is applied on top of the hidden states to produce the binary predictions. + +Note that our LLM decoding does not follow the autoregressive paradigm—the inference-time input to the LLM is fixed as the candidate mask tokens, instead of using the previous output tokens. As we will show in Section 5.3, this + +simple and direct decoding strategy greatly outperforms traditional autoregressive decoding. Meanwhile, we can perform binary classification on all candidate masks in one pass to boost inference efficiency, while autoregressive generation can output only one token at a time. + +# 3.2. Multi-stage Training + +Training stages. To efficiently train the model containing both pretrained weights and randomly initialized weights, we divide the training into two stages following the practice of LLaVA [35]. The first stage is mask projector pretraining, during which we freeze all modules except the mask projector, as it is a new module that cannot inherit from LLaVA model weights. We design a pretext task where we provide only mask tokens and text prompts (without global visual tokens) to the LLM and let it predict the image-level description, which is similar to the original LLaVA pretraining task but uses mask tokens instead. This task aligns mask tokens with the LLM, so that they can jointly produce image descriptions. We find it viable to caption images using only mask tokens and the loss converges to a level similar to LLaVA pretraining, because mask tokens can capture and describe the major objects, which is enough to caption object-centric images. For this stage, we reuse the pretraining data from LLaVA, a set of image-caption pairs, and add SAM-generated [24] masks to each image. + +In the second stage, visual instruction tuning, we finetune all modules except the visual encoders for the mask grouping task. Given an image and a set of candidate masks, the model learns to predict the correct subset of masks based on the input vision-language prompt. The model can be further finetuned on task-specific data to adapt the model for greater specialization in downstream tasks like RES. + +Training objectives. In the pretraining stage, the expected model outputs are text-only, so we can train the model using cross-entropy loss [35, 47]. During the visual instruction tuning stage, the learning task transitions to mask grouping. Therefore, we change the training objective from maximizing the likelihood of the caption to optimizing for permask binary classification. More specifically, we compute the mean binary cross-entropy loss, averaged over all candidate masks. Among numerous mask candidates, usually only a few should be selected in the group. Due to this imbalanced distribution of positive and negative samples, we assign a larger loss weight to positive candidates. + +# 4. Data for Model Training and Evaluation + +To effectively train and evaluate RAS, we build two datasets: MASKGROUPS-2M, which is a dataset containing 2 million samples automatically generated from existing datasets with object-level annotations, and MASKGROUPS-HQ, which is a smaller, high-quality and diverse dataset annotated by human annotators. + +# 4.1. MASKGROUPS-2M: Data Repurposed for Visual Instruction Tuning + +Each training sample of the mask grouping task consists of an image, a set of candidate masks, a prompt (described by free-form text and optional reference masks), and a target mask group containing an arbitrary number of masks. No existing datasets provide all of these elements. To build MASKGROUPS-2M, we convert object-level annotations into the mask grouping format with templates. Table 2 summarizes the sources of each component in MASKGROUPS-2M. More details are in the supplementary material. + +
Sourcew/ow/
Cat.Att.Pos.Free.Cat.Att.Pos.Free.
MS-COCO+LVIS [16, 32]166K---166K---
VG [26]224K149K132K-224K149K392K34K
(G)RES [34, 69]---474K----
+ +Table 2. Composition of MASKGROUPS-2M. We collect mask groups based on categories, attributes, positions, and other freeform descriptions by converting object-level annotations from MS-COCO [32], LVIS [16], Visual Genome [26], and (generalized) referring expression segmentation datasets [34, 69]. + +Category-based groups. Given categorical annotations, we find same-category objects in each image, and form prompts in templates like "Select all " or "Segment everything of the same class as ." These groups originate from MS-COCO [32], LVIS [16], and Visual Genome [26]. LVIS uses the same images as MS-COCO but annotates more object categories with improved mask quality. Therefore, we merge MS-COCO and LVIS annotations before proposing category-based groups. Attribute-based groups. Visual Genome [26] includes annotations for object attributes (e.g., colors, materials). Similar to category-based groups, we collect objects with the same attribute and formulate groups like "Select all objects" or "Find all the objects with the same attribute as in the image." + +Position-based groups. The bounding box annotations provide positional information of objects. We form groups based on absolute positions (e.g., "Locate all the items on the left side of the image.") or relative positions (e.g., "Find all the objects above .) by comparing the coordinates of the bounding boxes. + +Other free-form prompts. In addition to the mask grouping criteria introduced above, we include groups with diverse free-form descriptions. RES (RefCOCO, RefCOCO+, and RefCOCOg [69]) and GRES (gRefCOCO [34]) datasets contain free-form phrases for localizing specific objects, which can be converted into text-only prompts and mask groups (e.g., "Select the in the image"). Visual Genome [26] contain annotations for object relationships. When there are + +
Modelw/o <mask-ref>w/ <mask-ref>Overall
gIoUcIoUgIoUcIoUgIoUcIoU
ReLA [34]34.9343.22----
PSALM1.3B [74]36.9237.33----
GSVA13B [62]41.9849.55----
RAS 13B, SAM (Ours)55.8260.1235.9137.7750.9853.93
RAS 13B, SAM, ORES-FT (Ours)66.7174.5958.7268.7764.7773.13
+ +Table 3. Results on our ORES dataset MASKGROUPS-HQ. Existing GRES models are unable to process reference masks as part of the input prompt ('- in the table). Given text-only prompts, RAS shows significantly stronger performance, which can be further improved by ORES finetuning. For LLM-based models, we mark the LLM scales in the subscript. + +![](images/22eaf75710b8a72ca7a2adbf5a08911ab398dd73c86c24e0eb3dcde00dd0d275.jpg) + +![](images/5bfd7c1d30aee31c6b6bd93fcd01376a63ef20a57fa76be9ba5e4c62b6c80d20.jpg) +The smallest elephant. + +![](images/75531525077709c4d6c87eed08458d98795a15724b21608f9e8744936d0cd4b7.jpg) +All animals with eyes showing in the picture. +Objects on + +![](images/586842a50dd3c749b8607d1933906a3c2dc256cf5ec77960f609a4772944d76b.jpg) +All those holding + +![](images/ad5ea98222833639d0a5d3697231a88f563f975b1f4cc32752fafab9355bdf86.jpg) +All paper products smaller than . +Figure 3. Examples of MASKGROUPS-HQ. Diverse vision-language prompts are included, involving object categories, attributes, positions, comparisons, interactions, etc. Best viewed on an electronic device with zoom-in functionality. + +![](images/28fdf7b2783ffe21cb08ce378067d61dd6ef4e8615dd33873479b459647f6104.jpg) +All blue letters lower than . + +multiple objects sharing the same relationship with the same subject, we group these objects (e.g., "Select all objects that ". + +Avoiding data contamination. The validation and test sets of RES/GRES datasets use images from MS-COCO and Visual Genome training data. To avoid data contamination, we exclude such images from MASKGROUPS-2M. + +# 4.2. MASKGROUPS-HQ: Human-Annotated Data for Finetuning and Evaluation + +Although MASKGROUPS-2M is large enough for instruction tuning, its mask groups in pre-defined templates cannot cover all possible criteria that human users may use for grouping, and may introduce noises due to inaccurate annotations in the source data. To further improve and evaluate the generalizability of RAS, we manually annotate a high-quality dataset MASKGROUPS-HQ (visualized in Figure 3). We start from EntitySeg [46], an image segmentation dataset containing high-resolution images and category-agnostic masks. Notably, the images are from various sources, not only MS-COCO. Then, human annotators inspect the images and masks, and annotate several mask groups by proposing a reasonable vision-language prompt and labeling the IDs of masks that should be included. Our + +quality check ensures that the proposed mask groups are agreed upon by different users. In total, 100,299 mask groups are annotated. Finally, we split MASKGROUPS-HQ into 96,697 samples over 18,368 images for finetuning, and 3,599 samples over 661 images for evaluation. $28\%$ of the samples include reference mask(s) in their prompts. + +# 5. Experiments + +In this section, we thoroughly test RAS in various tasks, including omnimodal referring expression segmentation (ORES, Section 5.1), and classic and generalized referring expression segmentation (RES and GRES, Section 5.2). We then analyze the candidate mask quality and design choices of RAS (Section 5.3), and finally demonstrate its applications (Section 5.4). Following prior practice in RES, we mainly consider cumulative/generalized intersection over union (cIoU/gIoU) metrics; for GRES we also report the accuracy of identifying "no-target" samples (N-acc.) [34]. + +Due to limited space, we include a) additional results on comparison with SEEM [77], finetuning GSVA [62] on our data, and converting ORES visual prompts into language, b) ablation study on mask tokenization and LLM scales, and c) qualitative results in the supplementary material. + +# 5.1. Omnimodal RES + +Our new ORES task poses new challenges to existing GRES models, because it uses mask-based visual prompts to describe relationships with reference entities. In Table 3, we compare RAS (both before and after finetuning on MASKGROUPS-HQ) based on off-the-shelf SAM [24] for proposing candidate masks, with state-of-the-art GRES models [34, 62, 74]. Unlike all GRES baselines, RAS is able to accept reference masks and process all prompts in MASKGROUPS-HQ, demonstrating a stronger prompting flexibility. Meanwhile, for text-only prompts, our results show significantly improved generalizability. Further finetuning RAS on MASKGROUPS-HQ brings more performance gains, especially enhancing the ability for understanding reference masks. Quantitative results are shown in Figure D in the supplementary material. + +
ModelRefCOCORefCOCO+RefCOCOgAvg.
valtestAtestBvaltestAtestBvaltest
ReLA [34]73.876.570.266.071.057.765.066.068.3
LISA13B,FT[27]74.979.172.365.170.858.167.970.669.9
MagNet [8]76.678.372.268.173.661.867.869.371.0
Groundhog7B[73]78.579.975.770.575.064.974.174.674.2
GSVA13B,FT[62]79.281.777.170.373.863.675.777.074.8
GLaMM7B,FT[49]79.583.276.972.678.764.674.274.975.6
u-LLaVA7B[63]80.482.777.872.276.666.874.875.675.9
SAM4MLLM8B[5]79.882.774.774.680.067.275.576.476.4
UNINEXT-H[65]82.283.481.372.576.466.274.676.476.6
PSALM1.3B[74]83.684.781.672.975.570.173.874.477.1
RAS13B,Co-DETR(Ours)79.482.675.972.277.364.773.274.575.0
RAS13B,Co-DETR,RES-FT(Ours)81.083.579.075.180.070.376.077.577.8
+ +Table 4. Results on referring expression segmentation (RES). With Co-DETR, an instance segmentation model specialized for MS-COCO (retrained to avoid data leakage), we establish new state of the art in RES. Models that are finetuned again for RES after training on mixed data are labeled with the subscript $\mathrm{{FT}}$ . + +
ModelvaltestAtestBAvg. cIoU
gIoUcIoUN-acc.gIoUcIoUN-acc.gIoUcIoUN-acc.
LAVT [66]58.4057.6449.3265.9065.3249.2555.8355.0448.4659.33
ReLA [34]63.6062.4256.3770.0369.2659.0261.0259.8858.4063.85
LISA13B,FT [27]65.2463.9657.4969.9971.0055.4362.1162.2956.3465.75
HDC [39]68.2865.4263.3872.5271.6065.2963.8562.7960.6866.60
GSVA13B,FT [62]70.0466.3866.0273.2972.7964.7265.4563.2062.4767.46
SAM4MLLM7B [5]71.8667.8366.0874.1572.2263.9265.2963.4259.9967.82
RAS 13B, Co-DETR (Ours)68.8664.4457.1974.8374.4054.8666.7464.5154.1367.78
RAS 13B, Co-DETR, GRES-FT (Ours)74.6470.4869.0577.4576.9964.6269.4267.9062.9271.79
+ +Table 5. Results on gRefCOCO for generalized referring expression segmentation (GRES). Our approach achieves the best overall performance compared with baselines. + +# 5.2. RES and GRES + +Since our new task extends the classic and generalized referring expression segmentation (RES and GRES), RAS can readily tackle both earlier tasks. We evaluate RAS on RefCOCO, RefCOCO+, RefCOCOg, and gRefCOCO datasets [34, 69], and compare it with prior state-of-the-art models [5, 8, 27, 34, 39, 49, 63, 65, 66, 73, 74]. The segmentation modules in prior models are finetuned for the RES and GRES tasks. Therefore, to ensure a fair comparison, we employ a model trained for MS-COCO instance segmentation, Co-DETR [76]. To avoid data leakage, we retrain the Co-DETR instance segmentation model on MScOCO excluding all RES/GRES validation and test images. As RAS is not limited to a specific segmentation model, we can produce candidate masks with Co-DETR and seamlessly apply RAS to its proposed masks. + +As shown in Tables 4 and 5, RAS demonstrates competitive results on both RES and GRES. Following prior models + +that are also trained on mixed data [27, 49, 62], we further finetune RAS on RES/GRES training data to adapt RAS for these tasks, which leads to state-of-the-art performance. + +# 5.3. Analysis and Ablation Study + +Quality of candidate masks. RAS extends existing segmentation models with complex vision-language interactions. To validate that the segmentation models, Co-DETR and SAM, can propose candidate masks that sufficiently cover the referred targets, we analyze the proposed masks by selecting the ones with the highest IoU with ground-truth masks and computing their cIoU on all three tasks. As shown in Table 6, even without finetuning on RES tasks, the best candidates by Co-DETR and SAM achieve significantly better cIoU than existing RES models, and the proposals indeed include most of the true targets ( $>85$ cIoU). By addressing the challenge of understanding and selecting high-quality candidates, RAS achieves the best final results. + +
ORESRESGRES
ModelcIoUModelcIoUModelcIoU
Previous state of the art
GSVA13B49.55PSALM1.3B77.1SAM4MLLM7B67.82
Best candidates proposed by segmentation models in RAS
SAMOracle86.39Co-DETROracle87.2Co-DETROracle87.60
Final performance of RAS (Ours)
RAS 13B, SAM74.59RAS 13B, Co-DETR77.8RAS 13B, Co-DETR71.79
+ +Table 6. Analysis of candidate mask quality. "Oracle" denotes a setting where the ground-truth targets are known and the closest candidates are chosen. Compared with previous models, the best candidates proposed by the segmentation models in RAS already obtain much higher mask quality, even without finetuning on RES tasks. Building upon the high-quality candidates, RAS delivers the strongest final performance. + +Due to limited computation, models in the following ablation are only trained on 0.5M samples of MASKGROUPS-2M and tested on our ORES dataset MASKGROUPS-HQ. Non-autoregressive vs. autoregressive decoding. RAS uses a simple yet effective decoding strategy, where the LLM inputs are directly from the candidate mask tokens rather than the previously predicted tokens. We compare our non-autoregressive formulation with the traditional autoregressive paradigm adopted by previous LMMs [27, 62, 73]. The autoregressive baseline learns to predict continuous mask embeddings of the selected masks in a sequential manner, and we collect the candidate masks whose embeddings are closest to these predicted embeddings as the output. Table 7 shows that our non-autoregressive formulation improves performance and enables more efficient inference. + +
Decoding ParadigmcIoU↑Latency↓
Autoregressive45.342.13
Non-Autoregressive53.750.56
+ +Table 7. Comparison between decoding paradigms. Our non-autoregressive formulation leads to more effective training and more efficient inference. + +Visual encoders. We use four visual encoders as a feature ensemble for mask tokenization. In Table 8, we compare RAS with its variants that encode mask tokens with a single encoder, as well as the previously best model GSVA [62]. Even with one single encoder, RAS outperforms GSVA. Combining all four encoders leads to the best results. + +# 5.4. Applications + +By addressing the ORES task, RAS improves a range of applications requiring fine-grained localization of multiple visual entities. As shown in Figure 4, the predicted segmentation mask groups can be seamlessly integrated with generative models (e.g., Adobe Photoshop Generative Fill [1]) to remove or edit multiple targets conveniently and efficiently. + +
Modelw/o <mask-ref>w/ <mask-ref>Overall cIoU
GSVA13B [62]49.55--
RAS 13B, CLIP, SAM58.1337.6152.44
RAS 13B, ConvCLIP, SAM56.8344.0653.53
RAS 13B, SigLIP, SAM54.2432.0948.07
RAS 13B, DINOv2, SAM57.4021.7047.71
RAS 13B, Ensemble, SAM57.7344.4753.75
+ +Table 8. Comparison of RAS with different visual encoders. Our ensemble of four visual encoders yields the best visual features for mask tokenization. + +![](images/66293d54e21b8b645df4e066d09cb6cb8c562ff4f4e8ba5e3508d9f3bb6c4956.jpg) +Figure 4. Fine-grained image content manipulation enabled by our approach. In each row we visualize the original image, the predicted segmentation masks, and the object removal (first two rows) or editing (last two rows) results. Best viewed on an electronic device with zoom-in functionality. + +# 6. Conclusion + +This work introduces a novel task, omnimodal referring expression segmentation (ORES), which extends RES with more sophisticated interactions through visual and textual prompts. We present a simple yet effective solution, RAS, to achieve complex multimodal comprehension of segmentation masks. We demonstrate state-of-the-art performance compared to various baselines, not only in ORES but also in existing tasks (i.e., RES and GRES). + +Future directions include leveraging the potentials of LMMs to enable capabilities such as generating textual justifications for predicted mask groups and supporting multi-round interactions. Furthermore, we plan to explore improving the synergy between the segmentation model and the LMM, along with developing compact models variants tailored specifically for ORES, based on computation-efficient LLMs. + +Acknowledgments. This work was supported in part by NSF Grant 2106825 and NIFA Award 2020-67021-32799. This work used computational resources, including the NCSA Delta and DeltaAI supercomputers through allocations CIS230012, CIS230013, CIS240133, and CIS240428 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, as well as the TACC Frontera supercomputer, Amazon Web Services, and OpenAI API through the National Artificial Intelligence Research Resource (NAIRR) Pilot. + +# References + +[1] Adobe Inc. Photoshop: Generative Fill, 2023. 8 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2, 4 +[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In ICCV, 2019. 3 +[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal LLM's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3 +[5] 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 ECCV, 2024. 3, 7 +[6] 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. 4, 12 +[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing GPT-4 with $90\%$ ChatGPT quality, 2023. 3, 12, 14 +[8] Yong Xien Chng, Henry Zheng, Yizeng Han, Xuchong Qiu, and Gao Huang. Mask grounding for referring image segmentation. In CVPR, 2024. 3, 7 +[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A survey on multimodal large language models for autonomous driving. In WACV, 2024. 1 +[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2, 3 +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NACCL, 2019. 3 + +[12] 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 ICLR, 2021. 4, 12 +[13] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In CVPR, 2021. 3 +[14] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. In ICRA, 2024. 1 +[15] Yuhan Guo, Hanning Shao, Can Liu, Kai Xu, and Xiaoru Yuan. PromptTHis: Visualizing the process and influence of prompt editing during text-to-image creation. IEEE Transactions on Visualization and Computer Graphics, 2024. 2 +[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 2, 5, 12 +[17] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In ICCV, 2023. 3 +[18] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In ECCV, 2016. 1, 2, 3 +[19] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Ti-niu Tan. Locate then segment: A strong pipeline for referring image segmentation. In CVPR, 2021. 3 +[20] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. ReferItGame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 3 +[21] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. ReSTR: Convolution-free referring image segmentation using transformers. In CVPR, 2022. 3 +[22] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 12 +[23] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019. 1 +[24] 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. In ICCV, 2023. 2, 5, 6, 12 +[25] Mikhail Konenkov, Artem Lykov, Daria Trinitatova, and Dzmitry Tsetserukou. VR-GPT: Visual language model for intelligent virtual reality applications. arXiv preprint arXiv:2405.11537, 2024. 2 +[26] 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 Fei-Fei Li. Visual genome: Connecting language and vision using crowdsourced dense image annotations. *IJCV*, 123: 32–73, 2017. 2, 5, 12 +[27] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning segmenta + +tion via large language model. In CVPR, 2024. 2, 3, 4, 7, 8 +[28] Junyan Li, Delin Chen, Yining Hong, Zhenfang Chen, Peihao Chen, Yikang Shen, and Chuang Gan. CoVLM: Composing visual entities and relationships in large language models via communicative decoding. In ICLR, 2024. 3 +[29] 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. 3 +[30] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, 2023. 3 +[31] 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 CVPR, 2023. 3 +[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 ECCV, 2014. 2, 5, 12 +[33] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In ICCV, 2017. 3 +[34] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized referring expression segmentation. In CVPR, 2023. 1, 2, 3, 5, 6, 7, 12, 13, 14 +[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 5, 12 +[36] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 3, 4, 12, 14 +[37] Sichao Liu, Jianjing Zhang, Robert X. Gao, Xi Vincent Wang, and Lihui Wang. Vision-language model-driven scene understanding and robotic object manipulation. In CASE, 2024. 1 +[38] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In CVPR, 2022. 4, 12 +[39] Zhuoyan Luo, Yinghao Wu, Yong Liu, Yicheng Xiao, Xiao-Ping Zhang, and Yujiu Yang. HDC: Hierarchical semantic decoding with counting assistance for generalized referring expression segmentation. arXiv preprint arXiv:2405.15658, 2024. 7 +[40] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L. Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 3 +[41] Thao Nguyen, Yuheng Li, Utkarsh Ojha, and Yong Jae Lee. Visual instruction inversion: Image editing via image prompting. In NeurlPS, 2023. 2 +[42] Maxime Oquab, Timothee Darct, 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, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael + +Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. TMLR, 2023. 4, 12 +[43] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. In ICLR, 2024. 3 +[44] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, and Tong Zhang. DetGPT: Detect what you need via reasoning. In EMNLP, 2023. 3 +[45] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Philip Torr, Zhe Lin, and Jiaya Jia. Open world entity segmentation. TPAMI, 45(7):8743-8756, 2022. 1 +[46] Lu Qi, Jason Kuen, Tiancheng Shen, Jiuxiang Gu, Wenbo Li, Weidong Guo, Jiaya Jia, Zhe Lin, and Ming-Hsuan Yang. High quality entity segmentation. In ICCV, 2023. 6, 13 +[47] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018. 3, 5 +[48] 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 ICML, 2021. 3, 4, 12 +[49] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. GLaMM: Pixel grounding large multimodal model. In CVPR, 2024. 2, 3, 7 +[50] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. PixelLM: Pixel reasoning with large multimodal model. In CVPR, 2024. 3 +[51] S. Hamid Rezatofighi, Vijay Kumar B.G., Anton Milan, Ehsan Abbasnejad, Anthony Dick, and Ian Reid. DeepSet-Net: Predicting sets with deep neural networks. In ICCV, 2017. 2 +[52] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. LMDrive: Closed-loop end-to-end driving with large language models. In CVPR, 2024. 1 +[53] Tiancheng Shen, Jun Hao Liew, Long Mai, Lu Qi, Jiashi Feng, and Jiaya Jia. Empowering visual creativity: A vision-language assistant to image editing recommendations. arXiv preprint arXiv:2406.00121, 2024. 2 +[54] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M. Kitani. Rethinking transformer-based set prediction for object detection. In ICCV, 2021. 2, 4 +[55] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, Kurt Keutzer, and Trevor Darrell. Aligning large multimodal models with factually augmented RLHF. In ACL Findings, 2024. 3 +[56] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, + +Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal LLMs. In NeurIPS, 2024. 2, 4, 12 +[57] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? Exploring the visual shortcomings of multimodal LLMs. In CVPR, 2024. 3, 4 +[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and efficient foundation language models. 2023. 3 +[59] Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In NeurIPS, 2015. 2 +[60] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, and Jifeng Dai. VisionLLM: Large language model is also an open-ended decoder for vision-centric tasks. In NeurIPS, 2023. 3 +[61] 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 +[62] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. GSVA: Generalized segmentation via multimodal large language models. In CVPR, 2024. 2, 4, 6, 7, 8, 13, 14 +[63] Jinjin Xu, Liwu Xu, Yuzhe Yang, Xiang Li, Fanyi Wang, Yanchun Xie, Yi-Jie Huang, and Yaoqian Li. u-LLaVA: Unifying multi-modal tasks via large language model. In ECAI, 2024. 3, 7 +[64] Zhiyuan Xu, Kun Wu, Junjie Wen, Jinming Li, Ning Liu, Zhengping Che, and Jian Tang. A survey on robotics with foundation models: toward embodied AI. arXiv preprint arXiv:2402.02385, 2024. 1 +[65] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, 2023. 2, 3, 7 +[66] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip H.S. Torr. LAVT: Language-aware vision transformer for referring image segmentation. In CVPR, 2022. 3, 7 +[67] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In CVPR, 2019. 3 +[68] 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. In ICLR, 2024. 2, 3 +[69] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling context in referring expressions. In ECCV, 2016. 1, 2, 3, 5, 7, 12 +[70] Qihang Yu, Xiaohui Shen, and Liang-Chieh Chen. Towards open-ended visual recognition with large language models. In ECCV, 2024. 3 + +[71] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In CVPR, 2019. 2 +[72] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 4, 12 +[73] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. Groundhog: Grounding large language models to holistic segmentation. In CVPR, 2024, 2, 3, 4, 7, 8 +[74] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. PSALM: Pixelwise segmentation with large multi-modal model. In ECCV, 2024. 3, 6, 7, 13, 14 +[75] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024. 2, 3 +[76] Zhuofan Zong, Guanglu Song, and Yu Liu. DETRs with collaborative hybrid assignments training. In ICCV, 2023. 7 +[77] 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. In NeurIPS, 2024. 2, 6, 13 \ No newline at end of file diff --git a/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/images.zip b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..401837acd576b048b66ad34f54d91d5a4456054b --- /dev/null +++ b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:585c05d815b757788efbd86a2bd59a2989189bffc8b003ffd0d3fb4dfd16d6c9 +size 695668 diff --git a/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/layout.json b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..167b832b70f98fbb36c4245039c7043c52daaddf --- /dev/null +++ b/ICCV/2025/Refer to Any Segmentation Mask Group With Vision-Language Prompts/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:959cd28562f2ff4437b929558831d1652ef31d6f198f669a3a94e3b5aced2e11 +size 358198 diff --git a/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/20afbaf7-7893-4bfa-8b37-12b93f2016db_content_list.json b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/20afbaf7-7893-4bfa-8b37-12b93f2016db_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ba63f2c2679fb8e1cc2bb026f8da07ae57a6708e --- /dev/null +++ b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/20afbaf7-7893-4bfa-8b37-12b93f2016db_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a6dbd164f28cb4ff5aec264087ee2da83afcce7677a53d225fcc0d0a8e864ca +size 80469 diff --git a/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/20afbaf7-7893-4bfa-8b37-12b93f2016db_model.json b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/20afbaf7-7893-4bfa-8b37-12b93f2016db_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2a7e26313c5e4ff8ae56288c7e985d1550107837 --- /dev/null +++ b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/20afbaf7-7893-4bfa-8b37-12b93f2016db_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3929e826560eb4084859968c3176c01e00969ba7bf2b6286df62fbafc0ba4265 +size 98948 diff --git a/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/20afbaf7-7893-4bfa-8b37-12b93f2016db_origin.pdf b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/20afbaf7-7893-4bfa-8b37-12b93f2016db_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a050096f2b69f047e7b1bdb25734e19ee95a154 --- /dev/null +++ b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/20afbaf7-7893-4bfa-8b37-12b93f2016db_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a30699e4867fdc287c090e1535df5b1b581b7bd1288aba44149d1fca0938fba6 +size 3290347 diff --git a/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/full.md b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..391b7789dbb37b9a08a5a2b0c67538cb355465c5 --- /dev/null +++ b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/full.md @@ -0,0 +1,293 @@ +# ReferDINO: Referring Video Object Segmentation with Visual Grounding Foundations + +Tianming Liang $^{1}$ Kun-Yu Lin $^{1}$ Chaolei Tan $^{1}$ Jianguo Zhang $^{2}$ Wei-Shi Zheng $^{1}$ Jian-Fang Hu $^{1*}$ + +$^{1}$ Sun Yat-sen University $^{2}$ Southern University of Science and Technology + +liangtm@mail2.sysu.edu.cn, hujf5@mail.sysu.edu.cn + +Project page: https://isee-laboratory.github.io/ReferDINO + +# Abstract + +Referring video object segmentation (RVOS) aims to segment target objects throughout a video based on a text description. This is challenging as it involves deep vision-language understanding, pixel-level dense prediction and spatiotemporal reasoning. Despite notable progress in recent years, existing methods still exhibit a noticeable gap when considering all these aspects. In this work, we propose ReferDINO, a strong RVOS model that inherits region-level vision-language alignment from foundational visual grounding models, and is further endowed with pixel-level dense perception and cross-modal spatiotemporal reasoning. In detail, ReferDINO integrates two key components: 1) a grounding-guided deformable mask decoder that utilizes location prediction to progressively guide mask prediction through differentiable deformation mechanisms; 2) an object-consistent temporal enhancer that injects pretrained time-varying text features into inter-frame interaction to capture object-aware dynamic changes. Moreover, a confidence-aware query pruning strategy is designed to accelerate object decoding without compromising model performance. Extensive experimental results on five benchmarks demonstrate that our ReferDINO significantly outperforms previous methods (e.g., $+3.9\%$ $\mathcal{J}\& \mathcal{F}$ on RefYouTube-VOS) with real-time inference speed (51 FPS). + +# 1. Introduction + +Referring Video Object Segmentation (RVOS) [3, 40] aims to segment the target object in a video, referred by a given text description. This emerging task is potentially beneficial for many interactive video applications, and has therefore attracted great attention in the computer vision community. Compared to the unimodal video segmentation tasks [26, 41], RVOS is more challenging since it requires + +![](images/991343b2f9c172a257d6e867c6d127c1ddf968af29234b695a81974dc2f5b68a.jpg) + +![](images/8144e7ef02d9ae96f80c3e22c6fd9930241d0620f229de07ac69d7fe09acf856.jpg) +Figure 1. (a) The SOTA method [23] in RVOS struggles to distinguish similar objects properly based on compositional attribute descriptions (e.g., shapes + colors), while our ReferDINO overcomes this limitation. (b) GroundingDINO [22] fails to identify the target cat based on the motion reference "swinging its tail", while our ReferDINO corrects the prediction through cross-modal spatiotemporal reasoning and generates precise object masks. + +a strong vision-language capability to understand complicated descriptions and associate visual objects with text. + +Despite significant progress in recent years, existing RVOS models [3, 9, 23, 25, 40, 45] still suffer from many common problems. For example, they often struggle to handle the queries involving complex appearance, relative locations and attributes, as shown in Figure 1 (a). These problems are mainly caused by the insufficient vision-language capabilities of current models, which, in turn, result from the limited scale and diversity of available RVOS data, as densely annotating video-text data is extremely expensive. + +The recent progress in visual grounding field [5, 19, 22] offers unique potential to address this data limita- + +tion. Benefiting from large-scale pretraining on image-text data, the representative foundational models (e.g., GroundingDINO [22]) showcase profound capabilities in object-level vision-language understanding. However, simply employing them to tackle RVOS task is impracticable, as shown in Figure 1 (b). There are mainly two challenges: (1) These models well-designed for region-level regression lack the ability of pixel-level dense prediction. (2) Although they excel in understanding static attributes that can be observed in single frames (e.g., "the white cat"), they fail to identify the target object described by dynamic attributes (e.g., "the cat swinging its tail"). + +To address these challenges, we propose ReferDINO, a strong end-to-end RVOS approach that inherits static object perception from GroundingDINO [22], and is further endowed with pixel-wise dense segmentation and spatiotemporal reasoning capabilities. First, we elaborate a grounding-guided deformable mask decoder to produce high-quality object masks on each frame. Instead of simply adding a mask prediction branch in parallel with the original box prediction branch [9, 18, 23, 40], our mask decoder cascades the two branches as a grounding-deformation-segmentation pipeline. It utilizes the pretrained box prediction as location prior, progressively refining the mask prediction through deformable attention mechanisms. This process is differentiable, enabling mask learning to feed back into the box prediction branch for collaborative task learning. Second, we present an object-consistent temporal enhancer that injects pretrained time-varying text features into inter-frame interaction for cross-modal temporal reasoning. These two modules work together to bridge the gap between visual grounding and RVOS, effectively overcome the limitations of both foundational visual grounding models and existing RVOS methods, as shown in Figure 1. + +Another limitation of these foundational models is their huge computational overhead, which is commonly unacceptable for video-scale training and inference. To address this, we further propose a confidence-aware query pruning strategy to reduce per-frame computations without compromising pretrained knowledge, by progressively identifying and pruning the low-confidence object queries. + +Extensive experiments on five public RVOS benchmarks demonstrate that ReferDINO significantly outperforms state-of-the-art (SOTA) methods. Moreover, compared to other baselines built upon GroundingDINO, our approach achieves remarkable performance improvements, with $+2.6 \sim 3.6\%$ $\mathcal{J}\& \mathcal{F}$ on the Ref-Youtu-VOS dataset and and $+1.4 \sim 4.6\%$ $\mathcal{J}\& \mathcal{F}$ on the MeViS. In addition, our query pruning strategy reduces $40.6\%$ FLOPs and $41.3\%$ memory usage, achieving comparable performance with a $10 \times$ speedup ( $4.9 \rightarrow 51.0$ FPS). These results demonstrate ReferDINO's potential for real-time video applications. + +Overall, our contributions are summarized as follows: + +- We propose ReferDINO, a strong RVOS approach that inherits static object perception from pretrained foundational models and extends their capabilities to pixelwise dense segmentation and spatiotemporal reasoning through a grounding-guided deformable mask decoder and an object-consistent temporal enhancer. +- We introduce a confidence-aware query pruning strategy, which significantly improves training and inference efficiency without compromising performance. +- ReferDINO establishes new SOTA performance on five benchmarks, and our proposed components provide significant improvements in performance and efficiency. + +# 2. Related Works + +Referring Video Object Segmentation. RVOS [6, 7, 31] aims to segment objects throughout the video based on text descriptions. MTTR [3] firstly introduces the DETR paradigm [4] into RVOS. Furthermore, ReferFormer [40] proposes to produce the queries from the text description. On the top of this pipeline, follow-up works [8, 23, 25, 45] focus on modular improvements. Despite notable progress on specific datasets, these models are limited by insufficient vision-language understanding, and often struggle in unseen objects or scenarios. Recently, some works [1, 13] attempt to utilize GroundingDINO to identify objects in single frames and then apply SAM2 [28] to generate dense masks. However, such a manner of model ensemble is inefficient and non-differentiable, preventing further refinement of models' task-specific capabilities. In contrast, our ReferDINO is an end-to-end adaptation approach that benefits from both open-world knowledge of GroundingDINO and specific knowledge from RVOS data. + +Visual Grounding Foundational Model. Visual grounding [10, 12, 20, 21, 30, 34-36] aims to localize object regions in an image for a given text. Recent works [5, 19, 22, 44] focus on unifying detection and image-text datasets to increase the training vocabulary at scale. For example, GroundingDINO [22] incorporates grounded pretraining into DINO [46] to achieve impressive performance. These models showcase strong object-level vision-language understanding, and have been explored in many downstream works [27, 29, 38, 43, 47]. However, in most of these works, foundational models are primarily used as preprocessing tools to extract object regions, and their further potential in end-to-end task adaptation remains largely under explored. Recently, Video-GroundingDINO [38] extend GroundingDINO to produce temporal boundaries by simply inserting several temporal self-attention modules. To the best of our knowledge, our work is the first end-to-end approach that adapts GroundingDINO to RVOS task. Besides, our designs upon foundational models can provide insights for other downstream fields. + +![](images/19869007a5cb067e1c1d668ecb7d360ff7bf71b21dcf27ff91c66d14dd7c3816.jpg) +Figure 2. Overall architecture of ReferDINO. Modules colored in blue are borrowed from GroundingDINO, while those in red are newly introduced in this work. First, a confidence-aware query pruning strategy is employed to progressively prune low-confidence queries, deriving only a compact set of important object features $\mathcal{O}^t$ . After collecting the object features from all frames $\{\mathcal{O}^t\}_{t=1}^T$ , the object-consistent temporal enhancer performs cross-modal temporal reasoning by injecting time-varying text embeddings. Finally, our grounding-guided deformable mask decoder takes the box predictions as location priors to refine the mask features through deformation attention and cross-modal attention mechanisms. Note that this process is end-to-end differentiable, enabling segmentation gradients to propagate backward into the box head for collaborative learning. Best viewed in color. + +# 3. Background: GroundingDINO + +Our approach builds on a visual grounding foundational model, GroundingDINO [22], which we briefly revisit here. GroundingDINO is a DETR-based object detector, which introduces language into an object detector to achieve visual grounding. It mainly consists of an image backbone, a text backbone, a cross-modal encoder-decoder Transformer architecture, a box head, and a classification head. + +Given an image-text pair, GroundingDINO adopts the dual backbones to extract vanilla features, which are then fed into the cross-modal encoder to derive enhanced image features $F_{\mathrm{img}}$ and text features $F_{\mathrm{text}}$ . These enhanced cross-modal features are used to initialize a large set of query embeddings $\mathcal{Q}$ , which are individually fed into a cross-modal decoder to generate object features $\mathcal{O}$ . Note that the shapes of $\mathcal{O}$ and $\mathcal{Q}$ are equal. Finally, each object feature is passed to the box head and the classification head to predict bounding boxes and scores. Here, the scores are defined as the similarities between each object feature and the text tokens. For RVOS, we define the binary classification probability as the maximum score over all tokens. + +# 4. ReferDINO + +We illustrate our ReferDINO in Figure 2. Apart from GroundingDINO, our model consists of three main components: grounding-guided deformable mask decoder (§4.1), object-consistent temporal enhancer (§4.2) and confidence-aware query pruning (§4.3). First, given a $T$ -frame video + +and a language query, GroundingDINO is separately applied to each frame to obtain the object features. To improve per-frame efficiency, we use the confidence-aware query pruning strategy to progressively remove low-confidence object queries, achieving a very compact set of object features. Then, collecting the object features of all frames, we utilize object-consistent temporal enhancer to perform cross-modal temporal reasoning and ensure temporal consistency. After that, we employ the grounding-guided deformable mask decoder, which takes each object's box prediction as location condition, progressively refining its feature through the deformable cross-attention and cross-modal attention. Finally, the output feature corresponding to each object is used to produce a mask sequence $\{\mathbf{m}^t\}_{t=1}^T$ by dot-producting with the per-frame feature maps. + +# 4.1. Grounding-guided Deformable Mask Decoder + +This decoder takes an object feature as input and generates corresponding pixel-wise predictions on each frame. Instead of simply adding a mask head in parallel with the original box head, our mask decoder formulates the box and mask predictions as a grounding-deformation-segmentation pinline. It utilizes the box prediction as location prior to iteratively enhance the mask predictions, and incorporates the text prompts for further refinement. Moreover, this process is differentiable, enabling mask learning to feed back into the box head for collaborative task learning. + +For simplicity, we use $\tilde{\pmb{o}}$ to represent an arbitrary object feature. Before the mask decoder, we feed $\tilde{\pmb{o}}$ into + +the box head to predict its bounding box $\pmb{b} \in \mathbb{R}^4$ , where $\pmb{b} = \{b_x, b_y, b_w, b_h\}$ encodes the normalized box center coordinates, box height and width. Meanwhile, we employ a feature pyramid network (FPN) upon $\pmb{F}_{\mathrm{img}}$ , producing a high-resolution feature map $\pmb{F}_{\mathrm{seg}} \in \mathbb{R}^{\frac{H}{4} \times \frac{W}{4} \times d}$ , where $H$ and $W$ denote the height and width of the raw video frames. + +Then, we input these features and box predictions into the mask decoder for location-aware object refinement and mask generation. Our mask decoder consists of $L_{m}$ blocks, with each block comprising two components: deformable cross-attention [48] and cross-modal attention. When applying attention on image feature maps (termed memory), deformable cross-attention aggregates the features by adaptively sampling a small set of sampling points around a reference point. Unlike the vanilla deformable cross-attention [48] that generates reference points through MLPs, we directly use the predicted box center $\{b_{x}, b_{y}\}$ as the reference point. Specifically, we treat $\tilde{o}$ as the query, $F_{\mathrm{seg}}$ as the memory, and the normalized box center $\{b_{x}, b_{y}\}$ as the reference point. The sampling process is implemented by bilinear interpolation [16], thus it is end-to-end differentiable. In this way, the pretrained grounding knowledge is naturally incorporated to refine the mask prediction, and the segmentation gradients are also backward to optimize the object grounding. A concern is that noisy features (e.g. background) can also be involved during adaptive sampling, which potentially compromises the mask quality. To mitigate this, we further apply cross-modal attention to integrate the text conditions, by taking $\tilde{o}$ as query and $F_{\mathrm{text}}$ as key & value. These two attention mechanisms work together to ensure that mask prediction is tightly coupled with both text prompts and object locations. Finally, for each object query, we obtain a refined mask embedding $o_{m} \in \mathbb{R}^{d}$ , which is then dot-produced with the high-resolution feature map $F_{\mathrm{seg}}$ to generate an instance mask $m$ . + +Discussion. The dynamic mask head [33] also integrates object location information into mask prediction by separately concatenating the feature map with each object's relative coordinates, and it has been adopted in many previous works [3, 23, 25, 39, 40, 45]. However, the memory cost of this approach is dramatically high, due to the storage demand of per-object high-resolution feature maps. This is particularly unacceptable for foundational models, which typically uses a large number of object queries. In contrast, our mask decoder efficiently integrates the location information of different objects by location-guided sampling on a shared feature map. This design requires no extra memory burden, making our mask decoder particularly suitable for adaptation of foundational models. + +# 4.2. Object-consistent Temporal Enhancer + +Although GroundingDINO can detect referring objects from single images, this is not reliable enough for RVOS. + +![](images/70b92cbb54be35a5f5dba1dfba3869bb8b3fdce398f67ab5be639b3d8100b16f.jpg) +Figure 3. Illustration of our object-consistent temporal enhancer, where $f_{cls}^{t}$ is the cross-modal sentence feature of $t$ -th frame. + +First, it fails to understand the dynamic attributes in descriptions (e.g., action or temporal relationship). Second, video frames often contain camera motion blur and constrained perspectives, greatly undermining its temporal consistency. Therefore, we introduce an object-consistent temporal enhancer to enable corss-modal temporal reasoning. As shown in Figure 3, this module consists of a memory-augmented tracker and a cross-modal temporal decoder. It receives two inputs: all the object embeddings $\{\mathcal{O}^t\}_{t=1}^T$ , and time-varying sentence features $\{f_{\mathrm{cls}}^t\}_{t=1}^T$ , where $f_{\mathrm{cls}}^t$ corresponds to the [CLS] token in $F_{\mathrm{text}}^t$ . + +Memory-augmented Tracker. Before temporal interaction, we align the objects across different frames with a tracker mechanism. Let $\mathcal{M}^t$ indicates the memory for $t$ -th frame, and $\mathcal{M}^1 = \mathcal{O}^1$ . Our tracker consists of two steps: object alignment and memory updating. In the first step, we compute the cosine similarity between $\mathcal{M}^{t-1}$ and $\mathcal{O}^t$ as the assignment cost, and apply the Hungarian algorithm [17] to align the objects with the memory: + +$$ +\hat {\mathcal {O}} ^ {t} = \text {H u n g a r i a n} \left(\mathcal {M} ^ {t - 1}, \mathcal {O} ^ {t}\right), \tag {1} +$$ + +where $\hat{\mathcal{O}}^t$ indicates the aligned object embeddings. In the second step, these embeddings are used to update the memory in a momentum manner. Meanwhile, we incorporate the text relevance to prevent object-invisible frames from disturbing the long-term memory. Formally, the memory is updated as follows: + +$$ +\mathcal {M} ^ {t} = \left(1 - \alpha \cdot \boldsymbol {c} ^ {t}\right) \cdot \mathcal {M} ^ {t - 1} + \alpha \cdot \boldsymbol {c} ^ {t} \cdot \hat {\mathcal {O}} ^ {t}, \tag {2} +$$ + +where $\alpha$ is the momentum coefficient, and $\pmb{c} \in \mathbb{R}^{N_s}$ is the cosine similarity between $\hat{\mathcal{O}}^t \in \mathbb{R}^{N_s \times d}$ and $\pmb{f}_{\mathrm{cls}}^t \in \mathbb{R}^d$ . + +Cross-modal Temporal Decoder. This module takes the time-varying text embeddings as frame proxies to perform inter-frame interaction and temporal enhancement. Specifically, this module comprises $L_{t}$ blocks. In each block, given the aligned object embeddings $\{\hat{\mathcal{O}}^t\}_{t = 1}^T$ and the sentence embeddings $\{\pmb {f}_{\mathrm{cls}}^{t}\}_{t = 1}^{T}$ , we employ self-attention along the temporal dimension to achieve inter-frame interaction. Next, we extract dynamic information with a cross-attention module, which takes the sentence embeddings as + +query and the object embeddings as key and value, deriving temporal-enhanced object features $\{\mathcal{O}_v^t\}_{t = 1}^T$ . These features containing effective temporal information are used to enhance the frame-wise object embeddings as follows: + +$$ +\tilde {\mathcal {O}} ^ {t} = \operatorname {L a y e r N o r m} \left(\hat {\mathcal {O}} ^ {t} + \mathcal {O} _ {v} ^ {t}\right). \tag {3} +$$ + +Discussion. Our temporal enhancer improves existing temporal modules of RVOS models in two aspects. 1) Previous works either ignore object tracking [23, 25, 32, 37] or only consider it in adjacent frames [6, 9, 39], while we employ a memory-augment tracker for stable long-term consistency. 2) In cross-modal temporal interaction, prior methods [9, 23, 32, 40] always use a static text embedding from the text encoder (e.g., BERT) for all objects across different frames, while our temporal enhancer uses time-varying text embeddings to better capture fine-grained temporal dynamics. These two designs make our temporal enhancer fully exploit the pretrianed object knowledge and cross-modal features from foundational models for effective spatiotemporal reasoning and object consistency. + +# 4.3. Confidence-aware Query Pruning + +Foundational visual grounding models typically utilize a large set of query embeddings to store extensive object information, e.g., GroundingDINO uses $N_{q} = 900$ queries. Iteratively processing such a large amount of queries significantly limits the efficiency, especially in video processing. However, directly reducing these queries can compromise the well-pretrained object knowledge. To solve this dilemma, we design a confidence-aware query pruning strategy to progressively identify and reject low-confidence queries at each decoder layer, as shown in Figure 4. + +Specifically, the cross-modal decoder is stacked by $L$ layers, and each layer consists of a self-attention, a cross-attention with image features and a cross-attention with text features. Let $Q_{l}\in \mathbb{R}^{N_{l}\times d}$ denote the output query embeddings of the $l$ -th decoder layer, where $N_0 = N_q$ . We reuse the attention weights in the decoder layer to compute a confidence score for each query as follows: + +$$ +s _ {j} = \frac {1}{N _ {l} - 1} \sum_ {i = 1, i \neq j} ^ {N _ {l}} \boldsymbol {A} _ {i j} ^ {s} + \max _ {k} \boldsymbol {A} _ {k j} ^ {c}, \tag {4} +$$ + +where $s_j$ is the confidence of $j$ -th query, $A^s \in \mathbb{R}^{N_l \times N_l}$ denotes the self-attention weight, $A^c \in \mathbb{R}^{K \times N_l}$ denotes the transposed cross-attention weight with $K$ text tokens. The former term represents the average attention received by $j$ -th query from other queries. A query receiving high attention from others typically implies that it is irreplaceable. The latter maximum term measures the probability that $j$ -th object query is mentioned in the text. The combination of these two terms comprehensively reflects the query importance. Based on this score, we retain only $1 / k$ high + +![](images/6dd63959517e339ceecf7f93f02dc977be6f5ad7087ac95623e4842600d8aff6.jpg) +Figure 4. Illustration of our confidence-aware query pruning, where $q_{i}$ and $t_{i}$ indicate $i$ -th query and text token, respectively. + +confidence object queries at each layer, finally yielding a compact set of $N_{s}$ object embeddings, where $N_{s} \ll N_{q}$ . This strategy significantly reduces the computation costs while preserving the pretrianed object knowledge, enabling ReferDINO to achieve real-time efficiency. + +Time Complexity Analysis. Let $N = N_{q}$ , total time complexity of the original decoder is $O\left(L(N^{2}d + Nd^{2})\right)$ . If we retain only $1 / k$ queries per layer, the total query number would decrease exponentially, and the total time complexity becomes $O\left((\frac{k^2}{k^2 - 1})N^2 d + (\frac{k}{k - 1})Nd^2\right)$ , independent of the decoder depth $L$ . In practice, this computational improvement is significant. For GroundingDINO with $N = 900$ , $L = 6$ , $d = 256$ , even when $k = 2$ , the decoder computational cost can be reduced to $24.7\%$ . More detailed derivation is presented in the supplementary A. + +# 4.4. Training and Inference + +ReferDINO ultimately produces $N_{s}$ object prediction sequences $\pmb{p} = \{\pmb{p}_i\}_{i=1}^{N_s}$ for a video-text pair, and each sequence is represented by $\pmb{p}_i = \{\pmb{c}_i^t, \pmb{b}_i^t, \pmb{m}_i^t\}_{t=1}^T$ , which denotes the binary classification probability, bounding box and mask of the $i$ -th object query on $t$ -th frame. + +Training. Suppose the ground truth object sequence to be $\mathbf{y} = \{\pmb{c}^t, \pmb{b}^t, \pmb{m}^t\}_{t=1}^T$ . Following the previous works [3, 4, 23], we apply the Hungarian algorithm for instance matching. Specifically, we select the instance sequence with the lowest matching cost as the positive and assign the remaining sequences as negative. The matching cost is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} (\boldsymbol {y}, \boldsymbol {p} _ {i}) = \lambda_ {\mathrm {c l s}} \mathcal {L} _ {\mathrm {c l s}} (\boldsymbol {y}, \boldsymbol {p} _ {i}) + \lambda_ {\mathrm {b o x}} \mathcal {L} _ {\mathrm {b o x}} (\boldsymbol {y}, \boldsymbol {p} _ {i}) \tag {5} \\ + \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {m a s k}} \left(\boldsymbol {y}, \boldsymbol {p} _ {i}\right). \\ \end{array} +$$ + +This matching cost is computed on individual frames and normalized by the frame number. Here, $\mathcal{L}_{\mathrm{cls}}$ is the focal loss that supervises the binary classification prediction. $\mathcal{L}_{\mathrm{box}}$ sums up the L1 loss and GIoU loss. $\mathcal{L}_{\mathrm{mask}}$ is the combination of DICE loss, binary mask focal loss and projection loss [33]. $\lambda_{\mathrm{cls}}$ , $\lambda_{\mathrm{box}}$ and $\lambda_{\mathrm{mask}}$ are scalar weights of individual losses. The model is optimized end-to-end by minimiz- + +
MethodVenueRef-YouTube-VOSRef-DAVIS17A2D-SentencesJHMDB-Sentences
J&FJFJ&FJFmAPoIoUmIoUmAPoIoUmIoU
Video-Swin-T/Swin-T
ReferFormer [40]CVPR'2259.458.060.959.656.562.752.877.669.642.271.971.0
HTML [8]ICCV'2361.259.563.0---53.477.669.242.7--
SgMg [25]ICCV'2362.060.463.561.959.064.856.178.070.444.472.871.7
SOC [23]NeurIPS'2362.461.163.763.560.266.754.878.370.642.772.771.6
LoSh [45]CVPR'2463.762.065.462.960.165.757.679.371.6---
DsHmp [9]CVPR'2463.661.865.464.060.867.257.279.071.344.973.172.1
G-DINO+SH-63.962.065.862.058.365.957.879.071.044.672.771.6
G-DINO+DH-64.262.466.165.161.369.157.378.269.945.173.271.8
ReferDINO (ours)ICCV'2567.565.569.666.762.970.758.980.272.345.674.273.1
Video-Swin-B/Swin-B
ReferFormer [40]CVPR'2262.961.364.661.158.164.155.078.670.343.773.071.8
HTML [8]ICCV'2363.461.565.262.159.265.156.779.571.244.2--
SgMg [25]ICCV'2365.763.967.463.360.666.058.579.972.045.073.772.5
SOC [23]NeurIPS'2366.064.167.964.261.067.457.380.772.544.673.672.3
LoSh [45]CVPR'2467.265.469.064.361.866.859.981.273.1---
DsHmp [9]CVPR'2467.165.069.164.961.768.159.881.172.945.873.973.0
G-DINO+SH-66.264.268.365.261.668.858.480.972.645.574.272.7
G-DINO+DH-66.764.569.067.663.671.358.780.271.845.374.173.2
ReferDINO (ours)ICCV'2569.367.071.568.965.172.961.182.173.646.674.273.2
+ +ing the total loss $\mathcal{L}_{\mathrm{total}}$ for positive sequences and only the classification loss $\mathcal{L}_{\mathrm{cls}}$ for negative sequences. + +Inference. At inference, we select the best sequence with the highest average classification score as follows: + +$$ +\sigma = \underset {i \in [ 1, N _ {s} ]} {\arg \max } \frac {1}{T} \sum_ {t = 1} ^ {T} \boldsymbol {c} _ {i} ^ {t}. \tag {6} +$$ + +Finally, the output mask sequence is formed as $\{\pmb{m}_{\sigma}^{t}\}_{t = 1}^{T}$ . + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets and Metrics. We evaluate ReferDINO on five public RVOS benchmarks: Ref-YouTube-VOS [31] covers 3,978 videos with 15K descriptions. Ref-DAVIS17 [15] contains 90 videos and 1.5K descriptions. A2DSentences [7] contains 3.7K videos and 6.6K descriptions. JHMDB-Sentences [7] contains 928 videos and 928 descriptions. MeViS [6] contains 2K videos and 28K descriptions. On Ref-YouTube-VOS, Ref-DAVIS17 and MeViS, we use region similarity $\mathcal{I}$ , contour accuracy $\mathcal{F}$ , and their average $\mathcal{J} \& \mathcal{F}$ as metrics. For A2D-Sentences and JHMDB-Sentences, we employ mAP, overall IoU (oIoU), and mean IoU (mIoU) metrics. + +Protocols. We follow the protocols established in previous works, and all evaluations are conducted using the official code or online platforms. Specifically, the model trained on the training set of Ref-YouTube-VOS is directly evaluated on both Ref-YouTube-VOS and Ref-DAVIS17. Similarly, the model trained on A2D-Sentences is directly evaluated on A2D-Sentences and JHMDB-Sentences. The model is first pretrained on image datasets RefCOCO+/g [14, 24], and then trained on the RVOS datasets, except for MeViS + +Table 1. Comparison on Ref-YouTube-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences. + +
MethodVenueJ&FJF
Video-Swin-T / Swin-T
MTTR [3]CVPR'2230.028.831.2
ReferFormer [40]CVPR'2231.029.832.2
LMPM [6]ICCV'2337.234.240.2
DsHmp [9]CVPR'2446.443.049.8
G-DINO-SH-43.439.847.1
G-DINO-DH-45.441.948.9
ReferDINO (ours)ICCV'2548.043.652.3
Video-Swin-B / Swin-B
G-DINO-SH-45.041.348.9
G-DINO-DH-47.944.351.5
ReferDINO (ours)ICCV'2549.344.753.9
+ +Table 2. Comparison on MeViS. + +where the model is trained directly, following [6, 9]. + +Implementation Details. Our model is built upon the pretrained GroundingDINO, which uses Swin Transformer as the image backbone and BERT as the text backbone. The official source releases two GroundingDINO checkpoints: Swin-T and Swin-B, both of which are covered in our experiments. We freeze the backbones and finetune the cross-modal Transformer with LoRA techniques [11], where the rank is set to 32. We set $\alpha = 0.1$ , $L_{t} = 3$ , and $L_{m} = 3$ . In the mask decoder, we set the number of sampling points to 16. In the query pruning, we set the drop rate as $50\%$ . For MeViS that involves multiple target objects, we follow the practice in [6, 9] to select multiple object trajectories with classification scores higher than a threshold $\sigma = 0.3$ . For the other datasets, we select the best object trajectory. Extra training details can be found in the supplementary B. + +# 5.2. Main Results + +Comparisons with SOTA Methods. As shown in Tables 1 and 2, our ReferDINO significantly and consistently outperforms SOTA methods across all five RVOS datasets. + +Specifically, with the Swin-T backbone, our ReferDINO achieves $48.0\%$ $\mathcal{J}\& \mathcal{F}$ on MeViS, outperforming existing SOTA by $1.6\%$ . On Ref-YouTube-VOS, our ReferDINO achieves $66.4\%$ $\mathcal{J}\& \mathcal{F}$ , improving SOTA performance by $2.8\%$ . When a larger backbone Swin-B is applied, ReferDINO further improves $\mathcal{J}\& \mathcal{F}$ to $69.3\%$ on Ref-YouTube-VOS, surpassing SOTA by more than $2.2\%$ . Consistent performance improvements are observed across the other datasets, which demonstrates the superiority of ReferDINO. + +Comparisons with Baselines. To further evaluate Refer-DINO's effectiveness, we develop two enhanced variants of GroundingDINO as baselines: G-DINO+SH (static mask head) and G-DINO+DH (dynamic mask head [33]). The former directly dot products the query embedding with the feature map, while the latter uses object-wise feature maps as discussed in Section 4.1. These two mask heads are generic and commonly used in previous RVOS models [3, 9, 23, 25, 40, 45]. Moreover, both baselines are augmented with temporal self-attention modules for cross-frame interaction, as in VideoGroundingDINO [38]. As shown in Tables 1 and 2, although these baselines achieve competitive results on specific datasets (e.g., on Ref-DAVIS17, G-DINO-SH surpasses SOTAs by $+0.3\%$ $\mathcal{J}\& \mathcal{F}$ and G-DINO-DH attains $+2.7\%$ ), they fail to exhibit significant or consistent improvements over existing models across datasets. These results demonstrate that trivial adaptation cannot fully unleash the pretrained capabilities of foundational models. In contrast, our ReferDINO demonstrates robust superiority across different backbones and datasets. The advantages become particularly significant on large benchmarks (Ref-Youtu-VOS and MeViS), where our ReferDINO with Swin-T can even outperform the baselines with Swin-B. More comparison results of GroundingDINO baselines are presented in the supplementary C. These results demonstrate the effectiveness of our approach in adapting foundational models to RVOS task. + +# 5.3. Ablation Studies + +In this section, we mainly analyze the effectiveness of the three proposed components. + +Mask Decoder. Our grounding-guided deformable mask decoder utilizes deformable cross-attention (DCA) for grounding-guided adaptive sampling and cross-modal attention (CMA) to further mitigate the impacts of sampling noise. To validate the effectiveness of these two modules, we conduct ablation studies. As shown in Table 3, removing CA and DCA results in performance drops of $0.4\%$ and $2.7\%$ in $\mathcal{J}\&\mathcal{F}$ , respectively. These results demonstrate the complementary advantages of the two attention mechanisms, where DCA plays a dominant role in improving mask quality and CA provides further refinement. + +Temporal Enhancer. Our object-consistent temporal enhancer consists of a memory-augmented tracker and a + +
MethodJ&FJF
ReferDINO48.043.652.3
w/o CMA47.643.152.0
w/o DCA45.341.848.7
+ +Table 3. Ablation study of the mask decoder on MeViS. + +
MethodJ&FJF
ReferDINO48.043.652.3
w/o Tracker47.643.252.1
w/o Decoder45.842.748.8
+ +Table 4. Ablation study of the temporal enhancer on MeViS. + +
Drop RateJ&FJFFLOPsMemoryFPS
067.665.669.7840.3G24.0G4.9
15%67.765.769.8672.7G18.9G26.5
30%67.665.669.6572.4G15.9G47.2
50%67.565.569.6499.3G14.1G51.0
Random 50%38.337.139.4499.3G14.1G51.0
+ +Table 5. Ablation study of query pruning on Ref-YouTube-VOS. FLOPs and Memory are measured with batch size of 1 on 6-frame 360P videos, and Pytorch checkpointing disabled. FPS is measured at 360P, following the official evaluation protocol [31]. + +cross-modal temporal decoder. We analyze their individual contributions in Table 4. First, while the object queries in GroundingDINO is naturally aligned across frames to some extent, adding a tracker can still improve the temporal consistency, leading to a $0.4\%$ improvement in $\mathcal{J}\& \mathcal{F}$ . Second, in the temporal decoder is essential for understanding temporal dynamics and object motion, providing $2.2\%$ improvements in $\mathcal{J}\& \mathcal{F}$ . We additionally present visualizations to show the effects of memory-augmented tracker in the supplementary E. These results demonstrate the effectiveness of our object-consistent temporal enhancer. + +Query Pruning. We explore the impact of pruning different numbers of queries. As training with more than $50\%$ of queries on MeViS is highly time-consuming, we conduct this validation on Ref-YouTube-VOS, and present the training costs, inference speed and performance in Table 5. It shows that our query pruning strategy can effectively maintain the performance while significantly reducing the computational costs. For instance, reducing $50\%$ queries with our strategy can dramatically reduce $40.6\%$ FLOPs and $41.2\%$ memory usage, making video-scale training on smaller GPUs possible. Meanwhile, our strategy achieves 51.0 FPS, increasing the inference speed by $10 \times$ while maintaining consistent performance. In contrast, randomly pruning the same number of queries results in a substantial $29.3\%$ decrease in $\mathcal{J} \& \mathcal{F}$ . With our strategy, Refer-DINO can be efficiently trained on large RVOS datasets and achieve real-time video processing. + +# 5.4. Qualitative Analysis + +In Figure 5, we provide the qualitative comparisons with the SOTA model [23]. In Figure 6, we present several visualizations of ReferDINO processing various text references within a video. More qualitative results are presented in the supplementary F. These scenarios involve multiple sim + +![](images/3cf3a4af6ecc50a12f291c2f58031c5e243a37ae9dfa4fb65181e0ab1975c16a.jpg) + +![](images/2b14683653e4ee7a824cd43e7466a2ae220c488459273a8d8f7348dc028e63ea.jpg) +Figure 5. Qualitative results of our ReferDINO with the SOTA method [23]. ReferDINO performs much better in understanding complex object description involving motion and spatial relationships. + +![](images/8da192823e9912c2e2ae763977dc7999d3f45893238a2b141023fe478acc88dd.jpg) +"Cat turning around and playing with toy" + +![](images/445ac0cd202902738bf963b0733e569f0f942578124a50b895a9a98024e408a1.jpg) +"The bird that was stepped on by another bird" + +![](images/1ce1ea0c063961d1a9e45115d9ddcfa79e74d89c071fd906e6bd05ac70736c69.jpg) +"Red fish with the most pronounced up-and-down motion" "Yellow fish that swims forward and turns right" "Fish swimming up and down at the far left" + +![](images/085dec5d5998c54f6f49448aab08084ce8443e7e8ac3d87473aaa8ae28dd3cb0.jpg) +"The motionless lamb standing on the right side, next to the wall" +"The lamb with black head and brown wool" +Figure 6. Visualization of our ReferDINO for multiple text references. + +GroundingDINO, and is further endowed with pixel-level object segmentation and cross-modal spatiotemporal reasoning capabilities, through a grounding-guided deformable mask decoder and an object-consistent temporal enhancer. Moreover, we propose a confidence-aware query pruning strategy to accelerate the decoding process. Extensive experiments across five public benchmarks demonstrate that ReferDINO significantly outperforms existing RVOS methods. Moreover, the proposed query pruning strategy significantly improves both training and inference efficiency without compromising performance. + +ilar objects and complex object descriptions, including rich composite attributions, spatial relationships, and temporal motion. The results clearly demonstrate the effectiveness of ReferDINO in video object segmentation, spatio-temporal reasoning, and vision-language understanding. + +# 6. Conclusion + +In this work, we propose ReferDINO for RVOS task to address the challenges in vision-language understanding, dense perception and spatiotemporal reasoning. ReferDINO inherits region-level vision-language alignment from + +Acknowledgements. This work was supported partially by the NSFC (U22A2095, 6247072922), Guangdong Natural Science Funds Project (2023B1515040025), Guangdong NSF for Distinguished Young Scholar (2022B15-15020009), Guangdong Provincial Key Laboratory of Information Security Technology (2023B1212060026), open research fund of Key Laboratory of Machine Intelligence and System Control, Ministry of Education (No. MISC-202407). We are grateful to the National Supercomputing Center in Guangzhou for providing computing resources. + +# References + +[1] Grounded-SAM-2. https://github.com/IDEA-Research/Grounded-SAM-2.2, 12 +[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. +[3] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4985-4995, 2022. 1, 2, 4, 5, 6, 7, 12 +[4] 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. 2, 5 +[5] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 1, 2 +[6] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2694-2703, 2023. 2, 5, 6 +[7] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5958-5966, 2018. 2, 6 +[8] Mingfei Han, Yali Wang, Zhihui Li, Lina Yao, Xiaojun Chang, and Yu Qiao. HtmI: Hybrid temporal-scale multimodal learning framework for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13414-13423, 2023. 2, 6 +[9] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13332-13341, 2024. 1, 2, 5, 6, 7 +[10] Richang Hong, Daqing Liu, Xiaoyu Mo, Xiangnan He, and Hanwang Zhang. Learning to compose and reason with lan + +guage tree structures for visual grounding. IEEE transactions on pattern analysis and machine intelligence, 44(2): 684-696, 2019. 2 +[11] 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. 6 +[12] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1115-1124, 2017. 2 +[13] 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. arXiv preprint arXiv:2408.15876, 2024. 2, 12 +[14] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 6 +[15] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 6 +[16] Earl J Kirkland and Earl J Kirkland. Bilinear interpolation. Advanced computing in electron microscopy, pages 261-263, 2010. 4 +[17] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 4 +[18] 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 +[19] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 1, 2 +[20] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10880-10889, 2020. 2 +[21] Kun-Yu Lin, Hongjun Wang, Weining Ren, and Kai Han. Panoptic captioning: Seeking an equivalency bridge for image and text. arXiv preprint arXiv:2505.16334, 2025. 2 +[22] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded + +pre-training for open-set object detection. In European Conference on Computer Vision, 2024. 1, 2, 3 +[23] 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. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 26425-26437, 2023. 1, 2, 4, 5, 6, 7, 8, 12 +[24] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6 +[25] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920-930, 2023. 1, 2, 4, 5, 6, 7, 12 +[26] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 1 +[27] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. Detgpt: Detect what you need via reasoning. In The 2023 Conference on Empirical Methods in Natural Language Processing. 2 +[28] 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 Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 12 +[29] 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. 2 +[30] Arka Sadhu, Kan Chen, and Ram Nevatia. Zero-shot grounding of objects from natural language queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4694-4703, 2019. 2 +[31] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 2, 6, 7 +[32] Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15466-15476, 2023. 5 +[33] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In Computer Vision- + +ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 282-298. Springer, 2020. 4, 5, 7 +[34] An-Lan Wang, Bin Shan, Wei Shi, Kun-Yu Lin, Xiang Fei, Guozhi Tang, Lei Liao, Jingqun Tang, Can Huang, and Wei-Shi Zheng. Pargo: Bridging vision-language with partial and global views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7491-7499, 2025. 2 +[35] Liwei Wang, Yin Li, Jing Huang, and Svetlana Lazebnik. Learning two-branch neural networks for image-text matching tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2):394–407, 2018. +[36] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2019. 2 +[37] 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/CVF conference on computer vision and pattern recognition, pages 8741-8750, 2021. 5 +[38] Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Videogrounding-dino: Towards open-vocabulary spatio-temporal video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18909-18918, 2024. 2, 7 +[39] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 4, 5 +[40] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 1, 2, 4, 5, 6, 7, 12 +[41] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 1 +[42] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by multi-modality: A unified temporal transformer for video object segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6449-6457, 2024. +[43] 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. 2 +[44] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pre + +training for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 2 +[45] Linfeng Yuan, Miaojing Shi, Zijie Yue, and Qijun Chen. Losh: Long-short text joint prediction network for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14001-14010, 2024. 1, 2, 4, 6, 7 +[46] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2 +[47] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 2 +[48] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4, 12 \ No newline at end of file diff --git a/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/images.zip b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..27e25582fd8108033475ba2ec64ffce3e1f91505 --- /dev/null +++ b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4dae42508801e295f16dcf90f59c0ccd603ebafb339f26217d3f151eb44bfef +size 783806 diff --git a/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/layout.json b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1646685a2dc533a10a828b320d1479966288579d --- /dev/null +++ b/ICCV/2025/ReferDINO_ Referring Video Object Segmentation with Visual Grounding Foundations/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d57c498ac1e75825356e90438430088a385a6fa2bad02f50174b847f89b84e6a +size 421483 diff --git a/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/f9eb3d3c-0357-4236-b785-119874e96abc_content_list.json b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/f9eb3d3c-0357-4236-b785-119874e96abc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3e76740396ef3f62079f27017115e82cd49cff22 --- /dev/null +++ b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/f9eb3d3c-0357-4236-b785-119874e96abc_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f562bc0a5d8ab34db7e29e72a5d3bd5dc9d67d439381db9ca3ed6afdcaa4c7e +size 77206 diff --git a/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/f9eb3d3c-0357-4236-b785-119874e96abc_model.json b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/f9eb3d3c-0357-4236-b785-119874e96abc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b0e0c2eaef4e9ab4eb10a556531f4de751f5c497 --- /dev/null +++ b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/f9eb3d3c-0357-4236-b785-119874e96abc_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f01495d50b0cb434a3c166947a96278b3cfe27e020d705cc4df64ebd2c3cc6c4 +size 97548 diff --git a/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/f9eb3d3c-0357-4236-b785-119874e96abc_origin.pdf b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/f9eb3d3c-0357-4236-b785-119874e96abc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..57e354831b74a93da111b639c5c9a032de2835a3 --- /dev/null +++ b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/f9eb3d3c-0357-4236-b785-119874e96abc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3e574f6144c651855aa6f83f27b2425160a5629641387665d3c12c8bce8af78 +size 9846924 diff --git a/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/full.md b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..06e747ee2c7906767ba93d37fd372a52f7da43fb --- /dev/null +++ b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/full.md @@ -0,0 +1,311 @@ +# ReferEverything: Towards Segmenting Everything We Can Speak of in Videos + +Anurag Bagchi1 Zhipeng Bao1 Yu-Xiong Wang2 Pavel Tokmakov3† Martial Hebert1† +1Carnegie Mellon University 2University of Illinois Urbana-Champaign 3Toyota Research Institute https://refereverything.github.io/ + +![](images/5aedb250770d2299a8a5faee5a86ab739f56f84b0ddbc8b8f1303444724afa8d.jpg) + +![](images/83b1e9ab81806463514897fd53bc8a9cd450aa87ae2c32b716f530d5841507bd.jpg) + +![](images/fe027ad827e48cef4a807d2e69c065792ecc58a5b3febe1de22c0bf1e505c246.jpg) +Figure 1. Results of our method, REM, which leverages visual-language representations learned by video diffusion models to segment a wide range of concepts based on natural language descriptions (shown in grey boxes). REM exhibits zero-shot generalization to challenging, dynamic concepts, such as raindrops or shattering glass. Video visualizations are available on the project page. + +![](images/0fef79fd70c49e5d834b0adf9d8ce749090c9b71f728a9c7875e15d48f0583e0.jpg) + +![](images/76918f6c4b4e87c0a6aa22dbfb1bacf439f8827bafe1a3c0d6ff25bdf68ac8eb.jpg) + +![](images/6fd43ca383afa28377c7d9cf95f1eda9568d32ec5d465a2bddb49e7f9501825e.jpg) + +![](images/b4da5fe171e6ada96f763d64af03bd7bfd7355083757f733c339e44600312dc9.jpg) + +![](images/fac534393ed5e4400a9f817ba60bf4e97e0bed1486d5262df1eca9c7ea505ff5.jpg) + +![](images/77499a1bfb281da47237cef63bdcfd92f19c4f10910f6c653d596a197446d14c.jpg) + +# Abstract + +We present REM, a framework for segmenting a wide range of concepts in video that can be described through natural language. Our method leverages the universal visual-language mapping learned by video diffusion models on Internet-scale data by fine-tuning them on small-scale Referring Object Segmentation datasets. Our key insight is to preserve the entirety of the generative model's architecture by shifting its objective from predicting noise to predicting mask latents. The resulting model can accurately segment rare and unseen objects, despite only being trained on a limited set of categories. Additionally, it can effortlessly generalize to non-object dynamic concepts, such as smoke or raindrops, as demonstrated in our new benchmark for Referring Video Process Segmentation (Ref-VPS). REM performs on par with the state-of-the-art on in-domain datasets, like Ref-DAVIS, while outperforming them by up to + +12 IoU points out-of-domain, leveraging the power of generative pre-training. We also show that advancements in video generation directly improve segmentation. + +# 1. Introduction + +One of the most remarkable properties of natural language lies in its ability to convey the richness and complexity of human visual experience. From fleeting moments, like raindrops rolling down the window or smoke dissipating from a cigarette (see row 1 in Figure 1), to dynamic processes, such as a glass shattering or a whirlpool forming in the water (row 2 in Figure 1), language enables us to describe and reference events with precision. Crucially, if an event can be expressed verbally, it can often be accurately localized in both space and time. This universal mapping between the discrete, symbolic realm of language and the continuous, ever-changing visual world is developed through a lifetime of visual-linguistic interaction [4, 37]. + +![](images/1a2673e22d1b695cae5f753a078078eb1acc5e50cb0c7ed8d907759b3c38f852.jpg) +"Explode colorful smoke coming out" + +![](images/3311db5bc9e1a9b80d751aecc6cd84436e1ae15c24800dd810396b04f789df0a.jpg) +"Time-lapse of a blooming flower on a stem " +Figure 2. Through Internet-scale pre-training, video diffusion models learn to generate videos capturing the diversity of the dynamic visual world (samples shown above). We leverage their powerful visual-language representation for open-world referring video segmentation. + +The corresponding task in computer vision - Referring Video Segmentation (RVS) [15, 19], is defined as the task of segmenting a specific region in a video based on a natural language expression. However, virtually all existing literature focuses on a narrow subset of RVS - Referring Video Object Segmentation (RVOS) [45, 53]. This emphasis stems from data constraints: RVOS datasets were built by annotating object tracking benchmarks [36, 59], which are inherently object-centric and limited in scale. Recent advances in large visual-language datasets [2, 44] and generative models [41, 50, 51] present opportunities to overcome these limitations. Visual-language representations learned by these models from Internet data demonstrated strong generalization in image-based segmentation [32, 68], but their potential in videos remains underexplored [71]. + +In this work, we introduce REM - a novel approach to RVS that enables spatio-temporal localization of a wide range of concepts [16] in video that can be described through natural language, beyond conventional object tracking (shown in Figure 1). A key factor behind REM's success is preserving the universal visual-language mapping learned by generative models during Web-scale pre-training (see Figure 2). To this end, we retain the original model architecture and fine-tune it on small-scale RVOS datasets, adjusting the output to generate target mask latents instead of Gaussian noise. As shown in Table 1, this model, originally designed and trained for video generation, demonstrates competitive performance against specialized RVOS methods on popular benchmarks [36, 59]. More significantly, it exhibits much stronger generalization. + +To quantify this effect, we report zero-shot evaluation results on the open-world object tracking benchmark - BURST [1], and the non-object 'Stuff' segmentation dataset - VSPW [29], as well as on a newly collected benchmark that expands the focus of RVS to include dynamic process in Section 4. We define the latter as temporally evolving events, where subjects undergo continuous changes in state, shape, or appearance (see examples in Figure 1). Our new benchmark, which we call Ref-VPS for Referring Video Process Segmentation, consists of 145 videos that are labeled with referring expressions and masks at 6 fps and span 39 unique concepts. Experiments in Section 5.2 demonstrate that existing approaches, including the very recent method of Zhu et al. [71], fail to generalize outside of the + +narrow training distribution, whereas our method effortlessly segments a wide spectrum of targets (see Figures 1 and 5). + +The primary contribution of this paper is to demonstrate that Web-scale video diffusion models have learned universal visual-language mapping that can be repurposed for open-world referring video segmentation. We further introduce a new benchmark for Referring Video Process Segmentation in Section 4, expanding the focus of RVS beyond conventional object tracking. Finally, we provide a detailed analysis of our approach in Section 5.3, demonstrating that retaining the full architecture of the generative model, rather than isolating the de-noising network as a feature extractor, is key to unlocking the strongest generalization in RVS. + +# 2. Related Work + +Referring Video Segmentation (RVS) involves segmenting specific regions in a video based on a natural language description [15, 23, 45]. Most benchmarks for this task were developed by adding referring expression annotations to existing Video Object Segmentation (VOS) datasets, such as DAVIS'17 [36] or YouTube-VOS [59]. Consequently, the role of language in these benchmarks is limited to providing an interface for user-initialized object tracking [34, 55]. While segmenting objects is valuable, it addresses only a narrow subset of the possible interactions between language and the space-time continuum of videos. Equally important is the ability of RVS methods to segment video concepts beyond common object categories. To address this gap, we introduce a new benchmark focused on segmenting dynamic processes, which we term Referring Video Process Segmentation (Ref-VPS). + +Earlier RVOS approaches [5, 20, 30] generally employed a bottom-up strategy: first, image-level methods [8, 35, 43, 63] were applied to obtain frame masks, followed by spatiotemporal reasoning, such as mask propagation [45], to refine the segmentation across frames. More recently, with the success of cross-attention-based methods [28, 49, 66] in object segmentation and tracking, query-based architectures have been introduced to RVOS, leading to significant improvements [52, 53, 62]. The limited scale of paired video-language data with segmentation annotations has always been a major limitation, causing most methods to train + +jointly on videos and images [21, 22]. The latest approaches go further and unify all object localization tasks in a single framework [11, 54, 60]. However, while these models excel in object tracking, they struggle to generalize to more dynamic concepts. In contrast, we demonstrate that generative pre-training on Internet-scale data [2, 44] results in a universal (i.e., not limited to one domain) mapping between the space of language and the ever-changing visual world. + +Diffusion models have emerged as the de facto standard for generative learning in computer vision [18, 46] and beyond [12]. Among them, the Denoising Diffusion Probabilistic Model (DDPM) [18] leverages neural network components to model the denoising process. Stable Diffusion (SD) [41] shifts the denoising process into the latent space of a pre-trained autoencoder [24], allowing for model scaling. Expanding from images to videos, diffusion models have seen success in text-to-video (T2V) generation [6, 9, 10, 51, 69]. In addition to the capacity to generate high-fidelity images based on text prompts, the T2V diffusion models implicitly learn the mapping from linguistic descriptions to video regions, providing an opportunity to repurpose them for RVOS. Among current T2V methods, ModelScope [51] and Wan [50] stand out for their open-source implementations and top performance. + +Visual-language pre-training for perception: in addition to being highly effective in image and video generation, diffusion models have been shown to learn a strong representation of the natural image manifold. Several works demonstrated that they can be re-purposed for computer vision problems, including semantic segmentation [57, 67, 68] and pixel-level correspondence [47], achieving an impressive degree of generalization. Others have shown that image diffusion models learn powerful object representations, enabling open-world novel view synthesis [27] and amodal segmentation [32]. Most recently, Zhu et al. [71] also leverages pre-trained T2V models for RVOS, however, our analysis shows that their approach fails to fully capitalize on the universal visual-language mapping learned in generative pre-training. In this work, we explore the application of video diffusion models to RVS, demonstrating how to maintain a high-level generalizability during fine-tuning. + +In a separate line of work, visual-language representations learned with contrastive objectives [3, 38] have been adapted for referring image [25, 39, 58, 64] and video segmentation [70]. However, their performance remains limited compared to both generative models and classical referring segmentation approaches. + +# 3. Method + +# 3.1. Visual-language mapping via video denoising + +Text-to-Video (T2V) diffusion models [10, 51, 69] generate videos that align with a given language description, starting + +from Gaussian noise. The process can be formalized as: + +$$ +\hat {x} = f _ {\mathrm {v d m}} \left(x _ {T}, c, T\right), \tag {1} +$$ + +where $\hat{x}$ is the generated video, $T$ denotes the maximum timestep specified by the video diffusion model $f_{\mathrm{vdm}}$ , $x_{T}$ is a sample drawn from a Gaussian distribution $\mathcal{N}(\mu_T,\sigma_T^2)$ predefined by the video diffusion model, and $c$ is the conditioning prompt. To reduce computational complexity, these models often perform denoising in the latent space [41]. Specifically, a pre-trained Variational Autoencoder (VAE) [24] is employed to map the video $x$ from pixel space into latent space, denoted as $\mathcal{E}(x) = z$ , while a decoder reconstructs it from the latent space, $\mathcal{D}(z)\approx x$ . Thus, the generation process becomes: + +$$ +\hat {x} = \mathcal {D} \left(f _ {\mathrm {v d m}} \left(z _ {T}, c, T\right)\right). \tag {2} +$$ + +During training, rather than denoising from pure Gaussian latents, T2V models denoise from partially noisy video latents and optimize the following latent diffusion objective: + +$$ +\min _ {\theta} \mathbb {E} _ {z \sim \mathcal {E} (x), t, \epsilon \sim \mathcal {N} (0, 1)} \| \epsilon - \epsilon_ {\theta} (z _ {t}, e _ {c}, t) \| _ {2} ^ {2}, \tag {3} +$$ + +where $\epsilon$ is the Gaussian noise added to the clean video latent, $z_{t}$ represents the noisy video latent at timestep $t$ derived by the diffusion forward pass [18, 41], and $e_{c}$ is the conditional embedding generated from $c$ using a text encoder [38]. The denoising network $\epsilon_{\theta}(z_t,e_c,t)$ , typically a U-Net [42] or a Diffusion Transformer (DiT) [33], is tasked with predicting the noise $\epsilon$ . In this network, the conditional embedding $e_{c}$ interacts with the latent representations through attention mechanisms, guiding the model to generate diverse, semantically accurate videos. + +# 3.2. From language-conditioned denoising to RVS + +Referring Video Segmentation (RVS) involves segmenting an entity in a video across spatial and temporal dimensions, guided by a natural language description. Formally, the task is defined as: + +$$ +\hat {m} = f _ {\mathrm {R V S}} (x, c), \tag {4} +$$ + +where $f_{\mathrm{RVS}}$ is the RVS model, $x$ represents a video sequence, $c$ denotes a referring text prompt, and $\hat{m}$ is the binary masks produced as output. This task aligns naturally with T2V models, which establish a robust mapping between the entities described in the text and the corresponding spatial-temporal regions in the video by optimizing the denoising objective in Equation 3. + +Several prior works have explored the alignment of diffusion models with referring segmentation [57, 68, 71], typically employing these models as feature extractors. Specifically, they adjust the input format of a referring segmentation model to match that of the denoising network $\epsilon_{\theta}$ , and + +![](images/a7ebd357896aedf03fcbed321b04232a492b9f3bce5d269aa470eacfa8d1628e.jpg) +Figure 3. The model architecture of Refer Everything with Diffusion Models (REM). Like a video diffusion model it is based on, our approach takes video frames with added noise and a language expression as input. Our key insight is preserving the entirety of the generative model's architecture by shifting its objective from predicting noise to predicting mask latents. + +pass the resulting features to a task-specific decoder $f_{\mathrm{dec}}$ (e.g., a convolutional network) to predict the target masks: + +$$ +\hat {m} = f _ {\mathrm {d e c}} \left(\epsilon_ {\theta} ^ {(n)} \left(z _ {t}, e _ {c}, t\right)\right), \tag {5} +$$ + +where $z_{t}$ is the noisy latent representation of the input images at timestep $t$ , $e_{c}$ is a feature embedding of the referring expression $c$ , and $\epsilon_{\theta}^{(n)}$ denotes the intermediate feature at the $n^{th}$ layer. In practice, $t$ is usually set to a small value (e.g., 50), and $n$ is set to the later layer indexes to obtain the optimal performance. The entire model is then trained in a conventional discriminative learning setup. However, replacing parts of the generative model with newly initialized layers can disrupt alignment between the model's representation from pre-training and the new features learned on narrow-domain datasets, leading to a substantial loss of generalization capabilities. + +In our approach, shown in Figure 3, we propose to instead preserve the architecture of the video diffusion model in its entirety. Specifically, rather than using intermediate features $\epsilon_{\theta}^{(n)}$ , REM repurposes the whole denoising network $\epsilon_{\theta}$ (together with the VAE) by shifting its objective from predicting noise to predicting mask latents (shown on the right in Figure 3): + +$$ +\hat {m} = \mathcal {D} \left(\epsilon_ {\theta} \left(z _ {t}, e _ {c}, t\right)\right), \tag {6} +$$ + +where $\mathcal{D}$ denotes the (frozen) VAE decoder used to produce the actual binary segmentation masks from the predicted latents. That is, instead of learning the decoder network $f_{\mathrm{dec}}$ from scratch, we reuse the VAE from the video diffusion + +model. This subtle yet powerful modification allows the model to better preserve its universal visual-language mapping learned on Internet-scale data during generative pretraining while adapting to the task of RVS. + +Training and optimization. During training, to encode the ground-truth segmentation masks with the VAE, we broadcast the single-channel mask into three channels by simply duplicating it (shown in the top right of Figure 3). For simplicity, we still denote this three-channel mask representation as $m$ . The pre-trained VAE can then map the mask sequence into the latent space via $\mathcal{E}(m) = z^m$ and decode the masks back from predicted latents via $\mathcal{D}(z^m) \approx m$ . For the noisy latent $z_t$ and timestep $t$ , we prioritize using latents that remain as clean as possible. Therefore, we always set the timestep to its minimum value, $t = 0$ . To train the model, we supervise the predicted mask latents using an $\mathcal{L}_2$ loss (shown in the center-right of Figure 3) by minimizing: + +$$ +\min _ {\theta} \mathbb {E} _ {z ^ {m} \sim \mathcal {E} (m), t = 0} \| z ^ {m} - \epsilon_ {\theta} (z _ {t}, e _ {c}, t) \| _ {2} ^ {2}. \tag {7} +$$ + +Model inference. During inference, we follow Equation 6, with $t = 0$ , to decode the predicted mask latent and generate three-channel mask predictions. We then compute the single-channel masks by averaging the pixel values of the three channels and applying a constant threshold of 0.5 to binarize the result (shown in the bottom right of Figure 3). Notably, the inference is non-iterative (the mask is predicted in a single forward pass), making the computational cost of REM on par with other approaches in the literature. + +# 4. Benchmark Design and Collection + +Existing datasets only allow for quantifying the generalization of RVS models to rare object categories [1] or static 'Stuff' [29]. However, covering the entire spectrum of concepts that can be spoken of in videos in a single benchmark would be extremely costly. Instead, we narrow our focus to the most salient subset that necessitates joint modeling of language and temporal dynamics. Specifically, we target dynamic processes, defined as temporally evolving events, where the subjects undergo continuous changes in state, shape, or appearance. Importantly, these subjects are not limited to objects but include any spatio-temporally localizable phenomena, such as light or fire. Our new benchmark, Referring Video Process Segmentation (Ref-VPS), is built by selecting representative videos and annotating them with referring expressions and segmentation masks. + +# 4.1. Video selection + +To source the videos, we require a large, public, and diverse database that supports natural language queries and permits content redistribution for research. Based on these criteria, we selected TikTok — a platform with tens of millions of daily uploads that capture a wide range of dynamic scenarios. Moreover, TikTok's policies generally allow free redistribution of content, with individual users retaining the option to opt out. + +To identify a representative pool of videos, we established a taxonomy of five broad, possibly overlapping concepts (e.g., 'object transformations', or 'pattern evolution'; see Section A in the supplementary for definitions). Although these do not cover every possible dynamic process, we selected them to create a clear, focused framework for sourcing representative samples. For each concept, we use ChatGPT [31] to generate concrete examples and search queries for TikTok (e.g., 'a wax candle melting' for 'object transformations'), resulting in 120 fine-grained concepts. The queries retrieved over 1,000 samples, but many were unsuitable - either because the events (e.g., 'glaciers melting over time') occur over extended periods and are rarely captured on TikTok, or because of ambiguous search terms. After this step, we obtained a set of 342 samples. + +We then manually filtered the videos using these criteria: (1) exclude videos lacking significant dynamic changes (e.g., mostly stationary clouds); (2) exclude events that occur too rapidly to label a sufficient number of non-empty frames (e.g., lightning flashes); and (3) exclude videos with frequent shot changes that prevent extraction of a continuous clip capturing the event. Additionally, for videos that are compilations of similar events, we split them into individual clips and treat each independently. The final dataset comprises 145 clips covering 39 concepts. As it is intended for zero-shot evaluation, no additional splits are defined. Representative samples with results are shown in Figure 1. + +![](images/99d994d6a8821ecf643f80f2599448532baccc82c2a7595002391368bbcbf284.jpg) +Figure 4. Our mask annotation pipeline. We first use SAM2 to interactively segment the region of interest in the video (shown above). We then manually refine the masks where SAM2 fails and label ambiguous regions as Ignore (shown in yellow below). + +# 4.2. Annotation collection and evaluation + +To label the selected videos, we first adjust each clip's temporal boundaries to focus on the event of interest and avoid shot changes, ensuring that the event, along with some contextual frames before and after, is captured in its entirety. The clips are then exported as frames at 24 FPS. + +To collect referring expressions, we first manually identify the target entity in each clip and then instruct two independent annotators to provide descriptions. Each annotator contributes two expressions per target, yielding a total of four distinct expressions per clip. Following standard protocol [23, 45], evaluation is conducted on all queries, and results are reported as the average performance across them. + +Next, we label the identified targets with segmentation masks at 6 FPS using a semi-automatic pipeline. To this end, we leverage the recently introduced SAM2 [40] foundational model for interactive video segmentation. In particular, as shown in the top part of Figure 4, we first provide positive and negative clicks to a frame of the video. SAM2 then automatically segments the entity of interest in the frame, as well as propagates the mask across the entire clip. We interactively improve segmentation quality by providing additional clicks as needed. + +As SAM2 cannot always accurately segment the challenging and often ambiguous entities featured in Ref-VPS even with a large number of clicks, we manually refine the masks in frames where it fails (Figure 4, bottom). We additionally label ambiguous regions, such as the stem of the dandelion, as Ignore (shown in yellow). Visualizations + +
MethodPre-training DataMask/Box SupervisionRef-DAVISRef-YTB
J&FJFJ&FJF
Referformer [53]ImageNet + Kinetics + SSv2Ref-COCO+/g + Ref-YTB61.158.164.162.961.364.6
MUTR [62]ImageNet + Kinetics + SSv2Ref-YTB + AVS68.064.871.368.466.470.4
VLMO-L [70]UnknownRef-COCO+/g + Ref-YTB70.266.374.167.665.369.8
UNINEXT [60]Object36510+ Image/Video datasets72.568.276.870.167.672.7
VD-IT [71]LAION5B + WebVidRef-COCO+/g + Ref-YTB69.466.272.666.564.468.5
REM (MS-1.4B)LAION5B + WebVidRef-COCO+/g + Ref-YTB72.669.975.368.467.169.7
REM (Wan-14B)Internal + Public Images/VideosRef-COCO+/g + Ref-YTB75.071.378.771.769.274.3
+ +Table 1. Comparison to the state of the art on the validation set of the Ref-DAVIS and the test set of Ref-YTB benchmarks using the standard metrics. Even the base version of our method performs on par with the strong UNINEXT approach, despite not being specifically designed for object localization and having access to only a fraction of the localization labels used by that method. + +of Ref-VPS annotations, together with additional statistics, are included in the supplementary. For evaluation, we follow Tokmakov et al. [48] and only report region similarity $\mathcal{I}$ [14] as contour accuracy $\mathcal{F}$ [34] is often not well defined for dynamic entities like smoke or light. Pixels inside the Ignore regions are not included in the metric calculation. + +# 5. Experiments + +Datasets and evaluation. We evaluate our method on six benchmarks in total. Ref-YTB [45], Ref-DAVIS [23], and MeViS [13] are standard RVOS benchmarks, with MeViS focusing on challenging, motion-guided referring expressions. For evaluating generalization to rare objects and 'Stuff' categories, we use BURST [1] and VSPW [29] datasets, respectively. Finally, we evaluate REM and the strongest baselines on our newly introduced Ref-VPS (detailed in Section 4). All the datasets, except Ref-YTB and MeViS, are only used for evaluation in a zero-shot manner. + +For Ref-YTB [45], Ref-DAVIS [23], and MeViS [13], we use the standard evaluation metrics - Region Similarity $(\mathcal{I})$ [14], Contour accuracy $(\mathcal{F})$ [34] and their mean $(\mathcal{J}\& \mathcal{F})$ [36]. For all other evaluations, we use the $\mathcal{I}$ metric. The evaluations on Ref-YTB and MeViS are done on the official servers, and we use the official metric implementation of Ref-DAVIS for all the other benchmarks. + +Implementation details. Our approach builds upon two state-of-the-art text-to-video diffusion architectures: ModelScope [51] and Wan [50]. Additional video diffusion backbones are evaluated in the supplementary. ModelScope comprises 1.4 billion parameters and extends Stable Diffusion [6] with temporal modules. We adopt a two-stage training protocol following Zhu et al. [71]: in Stage I, we fine-tune only the spatial weights on Ref-COCO image-text pairs[65] for one epoch; in Stage II, we fine-tune all network weights for 40 epochs using Ref-YTB video-text examples [45] supplemented with 12K Ref-COCO images converted into pseudo-videos following Wu et al. [53]. In contrast, Wan employs a 14 billion parameter diffusion transformer that jointly models spatial and temporal information, without dedicated temporal modules [33]. Accordingly, we + +train this variant in a single stage on the combined Ref-COCO and Ref-YTB datasets for 80k iterations. Throughout training, the text encoder and VAE remain frozen. Unless otherwise stated, all models are trained and evaluated at a resolution of $512 \times 512$ . + +More details about the datasets, evaluation setup on BURST and VSPW, implementation, as well as ablations of our training strategy and runtime analysis are included in the supplementary. + +# 5.1. Referring video object segmentation results + +In this section, we compare REM to the state of the art on the standard RVOS benchmarks. We report results on the validation set of Ref-DAVIS [23] and the test set of Ref-YTB [45] in Table 1. The basic variant of our method, which is based on the ModelScope video diffusion model (denoted as 'REM (MS-1.4B)' in the table), outperforms the state of the art in terms of $\mathcal{J}$ on Ref-DAVIS and is only second to UNINEXT [60] on Ref-YTB. Note that this approach is specifically designed for object segmentation and trained on more than 10 datasets with localization annotations. In contrast, REM adopts an architecture of a video generation model and is only fine-tuned on one image- and one video-segmentation dataset. Despite this, our method is competitive with UNINEXT on standard RVOS benchmarks, and, as we will show next, outperforms it by up to 21 points out-of-domain in terms of $\mathcal{J}$ . + +Another notable observation is that this variant of REM outperforms VD-IT [71], which is built on top of the same MS-1.4B backbone, on both datasets. This result demonstrates the effectiveness of our approach to preserving the generative model's architecture, which will become even more evident in out-of-domain evaluation. Using the recent, large-scale Wan T2V diffusion model [50] (denoted as 'REM (Wan-14B)') further improves performance, achieving state-of-the-art results on both benchmarks. + +We also benchmark our REM against the top entrants on the MeViS leaderboard [13], a dataset specifically designed for segmenting objects based on referring expressions describing their motion. As shown in Table 2, our base vari + +
MethodMeViS
J&FJF
Referformer [53]31.029.832.2
VISA-13B [61]44.541.847.1
DsHmp [17]46.443.049.8
GLUS [26]51.348.554.2
REM (MS-1.4B)53.550.856.3
REM (Wan-14B)60.357.263.4
+ +ant already outperforms the strongest baseline, GLUS [26], which combines a pre-trained multi-modal LLM with the SAM2 foundational video segmentation model [40], by 2.2 points in $\mathcal{J}\&\mathcal{F}$ . Our Wan-14B variant further improves the performance to 60.3, setting the new state-of-the-art for this benchmark. These results validate our method's ability to reuse the motion-language priors encoded in video diffusion models by preserving their full generative architecture. + +# 5.2. Out-of-domain generalization + +Rare objects and 'Stuff'. We begin by performing a generalization study on the existing open-world tracking BURST dataset [1] as well as on the 'Stuff' categories [7] from VSPW [29] in Table 3. BURST is a large-scale, open-world video object segmentation benchmark with diverse scenes and rare objects, whereas VSPW tests the ability to generalize to non-object categories. We report zero-shot evaluation results on the validation set of VSPW and combined validation and test sets of BURST and compare to the top-performing methods in Table 1. + +We observe that on both datasets our method outperforms the baselines by significant margins. The improvements are especially noticeable on BURST, demonstrating that REM successfully preserves the object representation learned in generative pre-training. In contrast, VD-IT loses this generalization capacity. We analyze per-category performance in Section B.8 in supp., showing that REM is especially effective for the most challenging objects. On the 'Stuff' categories, all the methods do relatively poorly, reflecting the challenge of generalizing to more amorphous 'Stuff'. Here, VD-IT maintains a lead over entirely object-centric UNINEXT, but REM still outperforms both. + +Dynamic processes. We compare REM to the top-performing RVS baselines on our Ref-VPS benchmark in Table 4. As before, all the evaluations are zero-shot. Our approach outperforms all baselines by up to 12.1 points in Region Similarity (31.9% relative improvement), and notably surpasses the top RVOS method, UNINEXT, by 21.3 points (74.2% relative improvement). While generative pretraining enhances VD-IT's generalization ability compared + +Table 2. Comparison to the state of the art on the MeViS benchmark. REM outperforms approaches specifically designed for this dataset without any modifications. + +
BenchmarkMUTRUNINEXTVD-ITREM (Ours)
MS-1.4BWan-14B
VSPW10.510.112.715.218.5
BURST27.230.229.037.540.9
+ +Table 3. Comparison against the state of the art on the 'Stuff' categories in VSPW and on the open-world object-tracking BURST benchmark shows that REM achieves substantially stronger generalization. In particular, it outperforms VD-IT, which shares the same underlying diffusion backbone. + +
MUTRUNINEXTVD-ITGLUSREM (Ours)
MS-1.4BWan-14B
25.428.737.934.649.050.0
+ +Table 4. Comparison to the state of the art on the new Ref-VPS benchmark. REM shows much stronger zero-shot generalization to challenging, dynamic concepts in this dataset compared to the baselines by effectively capitalizing on Internet-scale pre-training. + +to UNINEXT, it struggles to preserve its representations as effectively as REM. The recent GLUS approach, which was designed for MeViS, also fails to generalize to Ref-VPS, highlighting the complementarity of the two benchmarks. + +A qualitative comparison of REM (MS-1.4B) with VD-IT and UNINEXT is provided in Figure 5. In the first row, our approach is able to track the sponge (which was never seen in training) in a challenging sequence from BURST, whereas other methods focus on foreground objects. In the second sequence from VSPW REM successfully generalizes to the non-object 'wall' category, whereas UNINEXT focuses on a nearby object and VD-IT fails entirely. The following examples from Ref-VPS illustrate that both baselines exhibit object-centric bias, as in the examples with the lizard skin in row 3 and blue smoke in row 6. While VD-IT shows better generalization, it often latches on the dominant region (rows 4 and 7 in Figure 5). In contrast, REM demonstrates both good coverage of rare concepts and high precision with respect to the language prompt. See more examples of highly dynamic sequences in the supplementary. + +# 5.3. Ablation analysis + +We now assess how effectively our REM transfers generative representations to RVS. We report results on the in-distribution Ref-YTB dataset and our Ref-VPS benchmark, using the MS-1.4B variant of our model. We further ablate the effect of the generative pre-training strategy in the supplementary. For efficiency, we train our model on a reduced subset of $\sim 12\mathrm{K}$ image and video samples for these experiments, which accounts for the slightly lower performance compared to the full-data results reported above. + +Is supervising in the latent space important? We evaluate the effectiveness of our key design decision to supervise mask prediction in the latent space of the frozen VAE decoder in Table 5. Instead of the latent space, we supervised the masks in the raw pixels space (RGB) by propagating + +![](images/ef2b5007f660ad6ed2802ea5509f50b5de8a20db2b28f9f495feb782f3c19d79.jpg) +Figure 5. Qualitative results of REM (MS-1.4B) and state-of-the-art baselines on BURST (top row left), VSPW (top row right), and Ref-VPS (bottom three rows) benchmarks. Our method demonstrates both superior coverage of rare, dynamic concepts, such as smoke or waves, and higher segmentation precision (e.g., only capturing the skin on the lizard). Video comparisons are available here. + +the gradients through the VAE decoder (rows two and three in the table). This increases memory requirements during training, forcing a lower $256 \times 256$ input resolution to fit on an A100 GPU. The results show that latent-space supervision is crucial for maintaining generalization, as evidenced by the performance drop of the RGB-based variants on RefVPS, even when the VAE decoder is additionally finetuned. + +Is it better to train a mask decoder from scratch? Several prior works have proposed to use a de-noising network as a feature extractor and learn a mask decoder head from scratch [68, 71]. Our approach uses a pre-trained VAE instead, which is not only a highly effective, general-purpose image encoder, but is also used in the de-noising network pre-training. We now ablate this design choice by replacing the VAE with a dedicated mask prediction module. We ablate both a CNN mask decoder from [68] and an MLP decoder from [56] and train them jointly with the rest of the model at the full resolution (last 2 rows in Table 5). Removing the pre-trained VAE has a moderately negative effect on performance on Ref-YTB, and, notably, destroys the model's ability to generalize to the out-of-distribution RefVPS (10.6 and 6.9 points drop for CNN and MLP, respectively). These results underscore the main takeaway of our analysis - preserving the entirety of the generative model's architecture is key for maximizing generalization in RVS. + +
SupervisionResolutionDecoderRef-YTB (J&F)Ref-VPS J
Latent256×256VAE (Frozen)61.435.8
RGB256×256VAE (Frozen)58.431.6
RGB256×256VAE (Fine-tuned)60.432.4
Latent512×512VAE (Frozen)63.540.0
RGB512×512CNN59.629.4
RGB512×512MLP59.333.1
+ +Table 5. Analysis of the effect of the fine-tuning strategy on Ref-YTB and Ref-VPS. The key to the success of REM is in preserving the entirety of the generative model's architecture. + +# 6. Discussion + +We introduced REM, a framework that leverages Intemet-scale video-language representations learned by diffusion models to enable open-world referring video segmentation. By preserving the full generative model architecture, REM preserves the universal visual-language mapping, allowing it to generalize beyond object-centric segmentation to dynamic, non-object concepts. Our experiments show strong generalization on standard datasets and our new Ref-VPS benchmark for dynamic video processes. Despite limited training on object masks, REM outperforms prior methods by up to 12 points in region similarity out-of-domain. Notably, we also show that advances in video diffusion models directly improve video segmentation performance. + +# Acknowledgement + +This project is supported in part by Toyota Research Institute, NSF Grant 2106825, and NIFA Award 2020-67021-32799. + +# References + +[1] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. BURST: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. +[2] 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. +[3] Zhipeng Bao, Martial Hebert, and Yu-Xiong Wang. Generative modeling for multi-task visual learning. In ICML, 2022. +[4] Lawrence W Barsalou. Perceptual symbol systems. Behavioral and brain sciences, 22(4):577-660, 1999. +[5] Miriam Bellver, Carles Ventura, Carina Silberer, Ioannis Kazakos, Jordi Torres, and Xavier Giro-i Nieto. A closer look at referring expressions for video object segmentation. Multimedia Tools and Applications, 82(3):4419-4438, 2023. +[6] 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. +[7] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and stuff classes in context. In CVPR, 2018. +[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. +[9] 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. +[10] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, 2024. +[11] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In CVPR, 2023. +[12] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In RSS, 2023. +[13] 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. +[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The PASCAL visual object classes (VOC) challenge. International journal of computer vision, 88:303-338, 2010. + +[15] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In CVPR, 2018. +[16] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. NeurIPS, 2019. +[17] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In CVPR, 2024. +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. +[19] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In ECCV, 2016. +[20] Tianrui Hui, Shaofei Huang, Si Liu, Zihan Ding, Guanbin Li, Wenguan Wang, Jizhong Han, and Fei Wang. Collaborative spatial-temporal modeling for language-queried video actor segmentation. In CVPR, 2021. +[21] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In ICCV, 2013. +[22] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. ReferItGame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. +[23] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In ACCV, 2019. +[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. +[25] 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. +[26] Lang Lin, Xueyang Yu, Ziqi Pang, and Yu-Xiong Wang. GLUS: Global-local reasoning unified into a single large language model for video segmentation. In CVPR, 2025. +[27] Ruoshi Liu, Rundi We, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. In ICCV, 2023. +[28] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. TrackFormer: Multi-object tracking with transformers. In CVPR, 2022. +[29] Jiaxu Miao, Yunchao Wei, Yu Wu, Chen Liang, Guangrui Li, and Yi Yang. VSPW: A large-scale dataset for video scene parsing in the wild. In CVPR, 2021. +[30] Ke Ning, Lingxi Xie, Fei Wu, and Qi Tian. Polar relative positional encoding for video-language segmentation. In IJ-CAI, 2020. +[31] OpenAI. Chatgpt, 2023. +[32] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: Amodal segmentation by synthesizing wholes. In CVPR, 2024. +[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. +[34] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. + +[35] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30K entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. +[36] 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. +[37] Sara F Popham, Alexander G Huth, Natalia Y Bilenko, Fatma Deniz, James S Gao, Anwar O Nunez-Elizalde, and Jack L Gallant. Visual and linguistic semantic representations are aligned at the border of human visual cortex. Nature neuroscience, 24(11):1628-1636, 2021. +[38] 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. +[39] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. GLaMM: Pixel grounding large multimodal model. In CVPR, 2024. +[40] 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. In ICLR, 2025. +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. +[42] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. +[43] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. GrabCut interactive foreground extraction using iterated graph cuts. ACM TOG, 2004. +[44] 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. +[45] Seonguk Seo, Joon-Young Lee, and Bohyung Han. URVOS: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, 2020. +[46] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. +[47] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. NeurIPS, 2023. +[48] Pavel Tokmakov, Jie Li, and Adrien Gaidon. Breaking the "object" in video object segmentation. In CVPR, 2023. +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. + +[50] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. +[51] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. +[52] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. OnlineRefer: A simple online baseline for referring video object segmentation. In CVPR, 2023. +[53] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, 2022. +[54] Junfeng Wu, Yi Jiang, Qihao Liu, Zehuan Yuan, Xiang Bai, and Song Bai. General object foundation model for images and videos at scale. In CVPR, 2024. +[55] Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang. Online object tracking: A benchmark. In CVPR, 2013. +[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. In NeurIPS, 2021. +[57] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. ODISE: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. +[58] Jiarui Xu, Xingyi Zhou, Shen Yan, Xiuye Gu, Anurag Arnab, Chen Sun, Xiaolong Wang, and Cordelia Schmid. Pixel-aligned language model. In CVPR, 2024. +[59] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. YouTube-VOS: Sequence-to-sequence video object segmentation. In ECCV, 2018. +[60] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, 2023. +[61] 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. +[62] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by multi-modality: A unified temporal transformer for video object segmentation. In AAAI, 2024. +[63] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In CVPR, 2019. +[64] 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. In ICLR, 2024. +[65] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling context in referring expressions, 2016. + +[66] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. MOTR: End-to-end multiple-object tracking with transformer. In ECCV, 2022. +[67] 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. NeurIPS, 2023. +[68] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In ICCV, 2023. +[69] 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, 2024. +[70]Zikun Zhou, Wentao Xiong, Li Zhou, Xin Li, Zhenyu He, and Yaowei Wang. Driving referring video object segmentation with vision-language pre-trained models. arXiv preprint arXiv:2405.10610, 2024. +[71] Zixin Zhu, Xuelu Feng, Dongdong Chen, Junsong Yuan, Chunming Qiao, and Gang Hua. Exploring pre-trained text-to-video diffusion models for referring video object segmentation. In ECCV, 2024. \ No newline at end of file diff --git a/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/images.zip b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..af35da1614469e398d4e6f1ccb01e75f2f4905e1 --- /dev/null +++ b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c208ebf0abccea9d1bf794aa3731ef4d8f1d69212325052fb5866441600cda5 +size 758442 diff --git a/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/layout.json b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d467451b20b5cbab198229ce67f89ee74fb22083 --- /dev/null +++ b/ICCV/2025/ReferEverything_ Towards Segmenting Everything We Can Speak of in Videos/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e6a77d6b67ba73588d231d7ab66ccbac4c7fc67c6d7b6e947b6869cf83a2162 +size 363557 diff --git a/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/31188455-5d8d-4a90-8d1d-7770f1fdc725_content_list.json b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/31188455-5d8d-4a90-8d1d-7770f1fdc725_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..645efcdf48008d937b09e90038a9dc21ff932695 --- /dev/null +++ b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/31188455-5d8d-4a90-8d1d-7770f1fdc725_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce402d2fa098c4ac2364408aa9d66bfcbdc5f4d9db95c428ffca56b937cafaf6 +size 85581 diff --git a/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/31188455-5d8d-4a90-8d1d-7770f1fdc725_model.json b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/31188455-5d8d-4a90-8d1d-7770f1fdc725_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4118a486568abf48b5629a24e4e1714f8bca19b0 --- /dev/null +++ b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/31188455-5d8d-4a90-8d1d-7770f1fdc725_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5883b0275f923ed219c23f09a5f04ea8e57e653bba297a791b1efc25fb5214c8 +size 109130 diff --git a/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/31188455-5d8d-4a90-8d1d-7770f1fdc725_origin.pdf b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/31188455-5d8d-4a90-8d1d-7770f1fdc725_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7c905627362307a9d8ff9a403e7f439cef346fcd --- /dev/null +++ b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/31188455-5d8d-4a90-8d1d-7770f1fdc725_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d3aba92b6a920ea2dd5ef90789c25a16e13838292d09601e4a75511e047b3e2 +size 49766161 diff --git a/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/full.md b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5a9c288a00df4de5caeda2f7b06923ebe857e83c --- /dev/null +++ b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/full.md @@ -0,0 +1,381 @@ +# Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion + +Byeonghun Lee $^{1*}$ +Soo Min Kang $^{2}$ $^{1}$ Korea Uni + +Hyunmin Cho $^{1*}$ +Iljun Ahn $^{2}$ $^{2}$ Inc + +Hong Gyu Choi² +Kyong Hwan Jin¹ +Researcher + +{byeonghun_lee, hyun_cho, kyong_jin}@korea.ac.kr,{hgc950909, kang.soomin, ijahn0828}@gmail.com + +# Abstract + +Most existing diffusion models have primarily utilized reference images for image-to-image translation rather than for super-resolution (SR). In SR-specific tasks, diffusion methods rely solely on low-resolution (LR) inputs, limiting their ability to leverage reference information. Prior reference-based diffusion SR methods have shown that incorporating appropriate references can significantly enhance reconstruction quality; however, identifying suitable references in real-world scenarios remains a critical challenge. Recently, Retrieval-Augmented Generation (RAG) has emerged as an effective framework that integrates retrieval-based and generation-based information from databases to enhance the accuracy and relevance of responses. Inspired by RAG, we propose an image-based RAG framework (iRAG) for realistic super-resolution, which employs a trainable hashing function to retrieve either real-world or generated references given an LR query. Retrieved patches are passed to a restoration module that generates high-fidelity super-resolved features, and a hallucination filtering mechanism is used to refine generated references from pre-trained diffusion models. Experimental results demonstrate that our approach not only resolves practical difficulties in reference selection but also delivers superior performance over existing diffusion and non-diffusion RefSR methods. Code is available at https://github.com/ByeonghunLee12/iRAG. + +# 1. Introduction + +Single image super-resolution (SISR) aims to reconstruct a high-resolution (HR) image from a low-resolution (LR) input, presenting a long-standing challenge in computer vision. Recent deep learning methods [6, 10, 17, 32, 40, 62] have significantly improved the performance of SISR, but + +![](images/e732035806e3b313e938aab5f35aafca4ab36ca85335d5514ddce8287bec9432.jpg) +Figure 1. **Concept of image-based RAG (iRAG).** Our framework comprises two key components: image restoration and data retrieval. The image restoration process is facilitated by two modules—the Reference-Aware Restoration Module and the RAG Restoration Module. The data retrieval component manages internal communication through hash keys, enabling efficient and rapid retrieval. Additionally, the database is constructed using a subset of real datasets combined with synthetically generated data produced by a pre-trained generative model. + +![](images/c2169e6a6af902b2b2fb9041460ebb2b4df77bc6541a63f7503ce5b6971c95a3.jpg) + +often result in over-smoothed details when optimized only for fidelity metrics (e.g., PSNR, SSIM [61]). To address this limitation, leveraging adversarial training emphasizes perceptually plausible high-frequency details, often leveraging adversarial training [4, 59, 67]. However, GAN-based models can be unstable during training and can introduce unnatural artifacts [34], spurring interest in alternative generative formulations such as diffusion probabilistic models [22, 46]. + +Diffusion probabilistic models have gained prominence by capturing complex, multimodal data distributions without the mode collapse problems often associated with GANs [20, 51]. By modeling a forward noise-adding and reverse-denoising process, these models excel in high-quality image generation [13, 23] and have been extended to inpainting, colorization, and super-resolution [47, 52]. Although operating in pixel space achieves state-of-the-art fidelity, it demands substantial computational resources in both training and inference [13]. To address these con + +![](images/6ee23f07cc9f6df9878f87fece3b1dea40fa355dae01e17b916a24aa084475d8.jpg) +Figure 2. Flowchart of image-based RAG. (a) A database of high-resolution references is hashed for efficient querying, and retrieved references are used to generate additional training samples via a diffusion model. (b) The low-resolution input is fused with the retrieved reference patches through a dedicated restoration module, leveraging spatial feature transform (SFT) layers. (c) The final stage refines the intermediate reconstruction to produce high-fidelity outputs. + +straints, latent diffusion models (LDMs) [30] project images into a lower-dimensional latent space, enabling efficient training and sampling while retaining high-quality reconstructions. This approach democratizes large-scale diffusion training and broadens its applicability to various conditional image-to-image transformations [38, 68]. + +In parallel, reference-based super-resolution (RefSR) utilizes HR auxiliary images with semantically or texturally similar regions to guide the reconstruction of the LR input. Early RefSR methods relied on optical flow [71] or patch similarity [48, 70] to align and transfer relevant textures from the reference image. Contemporary techniques employ attention mechanisms [65] or contrastive learning [24, 63] to handle complex alignments and improve texture fidelity. While existing RefSR methods assign HR auxiliary images that are semantically or structurally similar to LR input images, they face two key challenges: (1) matching difficulty, especially when the reference image differs substantially from the input LR image in illumination or pose [24, 41]; and (2) robust texture transfer, to ensure that only relevant high-frequency details are mapped to the LR input while minimizing artifacts from mismatched regions. + +Recently, the Retrieval-Augmented Generation (RAG) framework has been widely used to enhance the quality of retrieved responses in NLP tasks [19, 33]. In the text-to-image generation domain, Re-Imagen [9] leverages crossattention with retrieved images from an image database to improve semantic fidelity. We observe the analogy between retrieval-augmented systems and reference-based SR in terms of using additional retrieved patches for their own tasks. + +In this work, we propose a novel reference-based superresolution diffusion model follows RAG that unites the rich generative priors of latent diffusion [30] with the targeted detail enhancement strategy of RefSR [24, 41] which fol + +lows RAG [33] dedicated to the image domain such as [9]. Our approach tackles the reference selection and matching challenges by: + +- We introduce a hashing code vector strategy to efficiently retrieve relevant references within a compact latent space, making it robust against illumination and style variations between LR and HR images. +- We integrate additional modules within our diffusion-based super-resolution framework to refine domain alignment between the LR input and the selected reference, thereby efficiently leveraging reference information. +- Extensive experiments show that our diffusion-based RefSR framework outperforms previous methods and provides a robust real-world RAG-based solution even when exact references are absent. + +# 2. Related Works + +Diffusion Probabilistic Models Diffusion probabilistic models have emerged as powerful generative frameworks for high-fidelity image synthesis. Earlier generative models such as variational autoencoders and flow-based models [14, 15, 28] focused on likelihood-based training and efficient sampling, yet they often fell short of the visual quality achieved by Generative Adversarial Networks (GANs) [20]. Autoregressive models [7, 11] offered strong density estimation, but were hampered by slow sampling. + +Diffusion models [51] have since demonstrated excellent sample quality [13] and robust density estimation [27]. However, directly operating in pixel space is computationally expensive. Latent diffusion models (LDMs) [30] address this challenge by projecting images into a compressed latent space, reducing computational demands while preserving the quality of the synthesis. This approach supports a wide range of conditional and unconditional image generation tasks, including applications like super-resolution. + +Realistic Image Super-Resolution Realistic image super-resolution (Real-ISR) targets high-fidelity, perceptually convincing, and artifact-free outputs in real-world settings. Conventional methods [16, 36] optimizing fidelity metrics (e.g., PSNR, SSIM) tend to produce over-smoothed details, whereas adversarial approaches [20, 31, 58] enhance texture sharpness via discriminative training, albeit with potential instability and artifacts [34]. Recently, diffusion models [22, 46] have been employed to capture the natural image distribution more effectively, resulting in finer textures and improved training stability. By integrating diffusion priors with tailored degradation models and objectives, these methods demonstrate superior restoration performance under challenging conditions. + +Reference-Based Image Super-Resolution and RAG Reference-based image super-resolution (RefSR) improves reconstruction quality by transferring high-frequency details from an auxiliary high-resolution (HR) reference image to the low-resolution (LR) input. Traditional methods align the reference to the LR image using optical flow [71] or patch matching [48, 70], while advanced techniques employ attention mechanisms, contrastive learning, or teacher-student distillation for more robust texture alignment [24, 63, 65]. Recently, Retrieval-Augmented Generation (RAG) [19, 33] has emerged as a promising paradigm that dynamically sources and integrates external references, addressing the challenges of precise matching in real-world scenarios and further enriching the reconstruction process. RAG has been applied across multiple task domains: text-to-image generation [9], vision-language models [21], autonomous driving [60], etc. + +Data Retrieval via Neural Hash Network Neural hashing has advanced data retrieval by learning compact binary codes that capture both semantic and structural information. Supervised methods [18, 43, 64, 66] leverage label information to ensure similar images yield similar codes, preserving fine details for various vision tasks. Meanwhile, unsupervised approaches [37, 44] derive representations directly from the data distribution, revealing intrinsic patterns. + +Hallucination in Diffusion Models Diffusion models are well-known for generating high-fidelity images; however, they can also produce anomalous outputs—often termed hallucinations—such as images where hands exhibit extra fingers. These irregularities not only compromise image quality but may also undermine model robustness. Prior works [49, 50] have demonstrated that training on recursively generated data can erode critical, rare features, potentially leading to model collapse. Other works [2, 26] have examined the origins of these hallucinations and proposed various strategies for mitigation. + +# 3. Methodology + +In the context of image super-resolution (SR), an auxiliary high-resolution (HR) image, containing semantically or texturally similar information to the input low-resolution (LR) image, is often leveraged to guide the restoration of fine details and structural integrity. However, obtaining such a reference image for real-world datasets is challenging. Large datasets, such as ImageNet [12], typically consist of single-view images, making it difficult for direct use as reference patches. Furthermore, the process of curating a reference image from these extensive datasets is computationally onerous. To address this challenge, we propose a reference-based SR method that involves three key steps: (i) augmenting the existing dataset by enriching it with auxiliary HR images, (ii) retrieving a relevant HR image from a large database to match the target LR image, and (iii) generating a high-quality HR image by integrating LR and reference features into a diffusion model, as illustrated in Fig. 2. + +# 3.1. Data Augmentation + +The collection and annotation of real-world data can incur significant expenses, pose privacy concerns, and be subject to bias, potentially compromising the diversity of auxiliary data useful for LR image super-resolution [3, 45]. To address these challenges, we employ a diffusion model [46, 54] for data augmentation, thereby expanding the available dataset. Specifically, we sample latent noise $\sigma$ from $\mathcal{N}(0,1)$ and use a pre-trained diffusion editing model, $\mathcal{G}$ [42], to generate a synthesized image $\mathbf{I}_{\mathrm{gen}}$ conditioned on a real image $\mathbf{I}_{\mathrm{HR}}$ as follows: + +$$ +\mathbf {I} _ {\text {g e n}} = \mathcal {G} \left(\tilde {\mathbf {z}} _ {\mathrm {H R}}\right), \quad \tilde {\mathbf {z}} _ {\mathrm {H R}} \triangleq \mathbf {z} _ {\mathrm {H R}} + \alpha \cdot \sigma , \tag {1} +$$ + +where $\tilde{\mathbf{z}}_{\mathrm{HR}}$ is noisy latent obtained by adding noise $\sigma$ to $\mathbf{z}_{\mathrm{HR}}$ , an encoded latent of $\mathbf{I}_{\mathrm{HR}}$ . + +Recent analyses indicate that generative models are prone to producing out-of-distribution artifacts, commonly referred to as "hallucinations." These hallucinations can degrade downstream performance if used naively as training data [49, 50]. To mitigate these risks, we have incorporated a variance-based filtering method [2] to remove extreme hallucinations from the $\mathbf{I}_{\mathrm{gen}}$ . This method is based on the assumption that the high variance of a sampling trajectory can lead to the generation of hallucinated samples. Suppose $\{\hat{x}_0^{(i)}\}_{i = T_1}^{T_2}$ denotes the sequence of predicted images at timestep $i$ during the reverse diffusion process, and $\hat{\bar{x}}_0$ denotes their mean. We leverage the hallucination metric proposed in [2]: + +$$ +\operatorname {H a l} (x) = \frac {1}{| T _ {2} - T _ {1} |} \sum_ {i = T _ {1}} ^ {T _ {2}} \left(\hat {x} _ {0} ^ {(i)} - \bar {\hat {x}} _ {0}\right) ^ {2}. \tag {2} +$$ + +
Training LossPixel LossGAN LossDiffusion Loss
Using References✓✓✘✘✘✘✘✓
BenchmarksMetricsTTSR [65]DATSR [5]Real-ESRGAN [59]BSRGAN [67]DASR [35]LDM [46]StableSR [56]Ours
DIV2K Valid [1]PSNR ↑27.56027.84323.49923.86124.74622.62922.04422.323
SSIM ↑0.7830.7880.6860.6830.7070.6610.6260.632
LPIPS ↓0.2640.2570.2390.2530.1720.2280.2480.228
CLIP-IQA ↑0.4510.4520.5340.5800.6120.6180.6720.702
MUSIQ ↑58.70259.03663.38764.81765.34066.91067.06668.248
RealSR [4]PSNR ↑31.14431.54625.34726.34928.02124.22823.70223.951
SSIM ↑0.9020.9060.7890.7990.8380.7730.7210.717
LPIPS ↓0.1760.1290.1940.1990.1400.2230.2620.263
CLIP-IQA ↑0.4580.4540.5260.5740.5540.6570.6640.678
MUSIQ ↑64.29263.49367.53669.66067.96469.90270.19670.524
DRealSRPSNR ↑36.42637.09029.89330.41834.53224.22827.04927.144
SSIM ↑0.9490.9520.8650.8650.9050.7730.7650.776
LPIPS ↓0.1100.1050.1990.2210.1160.2230.3050.300
CLIP-IQA ↑0.4250.4220.5160.5730.4550.6570.6870.682
MUSIQ ↑49.23747.57560.05663.02752.98469.90264.33864.861
+ +Table 1. Quantitative comparison on test benchmarks. The best and second results are in red and orange. All the models on the table are trained with Flickr2K, DIV2K-train and OST. + +This metric ensures that latent codes exhibiting unusually high variance are flagged as hallucinations and excluded from the augmentation set. As illustrated in Sec. 4.2, the method successfully filters out hallucinated samples (i.e., unrealistic data), enabling robust data augmentation. + +# 3.2. Hashing for Data Retrieval + +To search for the HR exemplar relevant to the input LR image from an external database, our pipeline incorporates a retrieval mechanism [44]. Specifically, for each input $\mathbf{I}_{\mathrm{LR}}$ , we retrieve $\mathbf{I}_{\mathrm{ref}}$ from ref.DB, the database of reference images, that best matches the content or style of $\mathbf{I}_{\mathrm{LR}}$ under a compact hashing scheme. To achieve this, we first project the image into an embedding space, where image $\mathbf{I}$ is mapped to vector $\mathbf{h} = H_{\phi}(\mathbf{I})$ . We then retrieve the reference image $\mathbf{I}_{\mathrm{ref}}$ from ref.DB whose hash code is closest to the encoded query image's hash code $\mathbf{h}_{\mathrm{LR}} = H_{\phi}(\mathbf{I}_{\mathrm{LR}})$ , as follows: + +$$ +\mathbf {I} _ {\mathrm {r e f}} = \arg \min _ {\mathbf {I} _ {\mathrm {D B}} \in \mathrm {r e f . D B}} \left(\operatorname {d i s t} \left(\mathbf {h} _ {\mathrm {L R}}, H _ {\phi} \left(\mathbf {I} _ {\mathrm {D B}}\right)\right)\right). +$$ + +To facilitate contrastive learning, we construct positive pairs by generating correlated views for each image. Following standard convention, we apply arbitrary transformations (e.g., rotation, reflection, etc.) to each image, thereby producing two correlated views for the $k$ -th image $x^{(k)}$ in the database: + +$$ +v _ {i} ^ {(k)} := \mathcal {T} (x ^ {(k)}), \quad i \in \{1, 2 \}, +$$ + +where $\mathcal{T}$ denotes an arbitrary transform operator. Next, these views are passed through a pre-trained VGG encoder network $f(\cdot)$ to obtain the corresponding latent representations: + +$$ +z _ {i} ^ {(k)} = f (v _ {i} ^ {(k)}), \quad h _ {i} ^ {(k)} = g _ {\phi} (z _ {i} ^ {(k)}), +$$ + +where $z_{i}^{(k)}$ represents the encoded latent vector of the $i$ -th view of the $k$ -th image and $h_{i}^{(k)}$ denotes the corresponding binary hash code generated by the learnable function $g_{\phi}(\cdot)$ . The primary objective of our learning process is to minimize the distance between the binary hash codes $h_1^{(k)}$ and $h_2^{(k)}$ derived from the same image. We quantify the similarity between binary hash codes [8] using a cosine similarity function $\mathcal{C}$ . To amplify the differences between positive and negative pairs, we exponentiate the cosine similarity, yielding: + +$$ +\mathcal {S} (h _ {1} ^ {(k)}, h _ {2} ^ {(k)}) := e ^ {\mathcal {C} (h _ {1} ^ {(k)}, h _ {2} ^ {(k)})} / \tau , +$$ + +where $\tau$ is a normalization term. Next, we compute the log-likelihood of correctly identifying the positive match among all candidate pairs: + +$$ +\ell_ {1} ^ {(k)} := - \log \frac {\mathcal {S} \left(h _ {1} ^ {(k)} , h _ {2} ^ {(k)}\right)}{\mathcal {S} \left(h _ {1} ^ {(k)} , h _ {2} ^ {(k)}\right) + \sum_ {i , n \neq k} \mathcal {S} \left(h _ {1} ^ {(k)} , h _ {i} ^ {(n)}\right)}. +$$ + +The summation in the denominator aggregates the similarities between the anchor hash code $h_1^{(k)}$ and all negative samples. To symmetrize the objective, we also compute an analogous term $\ell_2^{(k)}$ for the second view. The final contrastive loss is defined as the average over all instances: + +$$ +\mathcal {L} _ {c l} := \frac {1}{N} \sum_ {k = 1} ^ {N} \left(\ell_ {1} ^ {(k)} + \ell_ {2} ^ {(k)}\right). +$$ + +By minimizing $\mathcal{L}_{cl}$ , the model learns an embedding space where semantically related samples cluster together. + +# 3.3. Reference-based Latent Diffusion Model + +To reconstruct the LR image using a retrieved reference, we proceed in two stages. First, we generate an intermediate reconstruction, $\mathbf{I}_{\mathrm{inter}}$ , using a transformer-based reference SR + +![](images/61edf749cb0e5e7b7c5a9f13a663847c85148f6549002a4f5fed1d90e0a33419.jpg) +Figure 3. Qualitative comparison on the test benchmarks with respect to our reference DB (Sec. 4.1). We compare DATSR [5], DASR [35], LDM [46], StableSR [56], and our method. Zoomed-in regions highlight differences in detail preservation. + +module [5, 65], which contains enriched features from the $\mathbf{I}_{\mathrm{LR}}$ . Then, we feed both the $\mathbf{I}_{\mathrm{LR}}$ and $\mathbf{I}_{\mathrm{inter}}$ into a diffusion-based refinement module to guide the final reconstruction. + +# 3.3.1. Intermediate Reference-based SR + +Given $\mathbf{I}_{\mathrm{LR}}$ and its retrieved reference $\mathbf{I}_{\mathrm{ref}}$ , we employ a reference-based SR network, [5, 65] $R_{\phi}$ , to produce an intermediate super-resolved image: + +$$ +\mathbf {I} _ {\text {i n t e r}} = R _ {\phi} \left(\mathbf {I} _ {\mathrm {L R}}, \mathbf {I} _ {\text {r e f}}\right). \tag {3} +$$ + +We minimize a combined loss that incorporates both L1 loss and a perceptual loss to ensure that the network captures the global structure and texture of the reference image. The resulting image, $\mathbf{I}_{\mathrm{inter}}$ , enriches the feature representation of $\mathbf{I}_{\mathrm{LR}}$ , ensuring that the diffusion model benefits from faithful structural and texture cues. + +Next, we feed both $\mathbf{I}_{\mathrm{LR}}$ and $\mathbf{I}_{\mathrm{inter}}$ into a feature encoder $\varepsilon_{\varphi}$ . Specifically, we adapt SFT layers [56] to condition the reverse diffusion process on the features extracted from $\mathbf{I}_{\mathrm{LR}}$ and $\mathbf{I}_{\mathrm{inter}}$ . Formally, let: + +$$ +\boldsymbol {F} _ {\text {c o n d}} \triangleq \operatorname {c o n c a t} \left(\varepsilon_ {\varphi} \left(\mathbf {I} _ {\text {L R}}\right), \varepsilon_ {\varphi} \left(\mathbf {I} _ {\text {i n t e r}}\right)\right), \tag {4} +$$ + +be the condition features encoded from both images, which are injected into each diffusion block through the SFT layers. We integrate the pre-cleaned features into the diffusion process via SFT layers, which perform affine transformations on the diffusion model's intermediate features. Let $\pmb{F}_{dif}^{n}$ be the feature map in the $n$ -th residual block of the Stable Diffusion U-Net. The SFT layer computes two affine + +parameters, $\alpha^n$ and $\beta^n$ , based on both the LR feature $F_{LR}^n$ and the reference feature $F_{inter}^n$ . Formally, these parameters are computed as: + +$$ +\alpha^ {n} (\pmb {F} _ {L R} ^ {n}, \pmb {F} _ {i n t e r} ^ {n}), \beta^ {n} (\pmb {F} _ {L R} ^ {n}, \pmb {F} _ {i n t e r} ^ {n}) = \mathcal {K} _ {\theta} ^ {n} \big (\pmb {F} _ {L R} ^ {n}, \pmb {F} _ {i n t e r} ^ {n} \big), +$$ + +where $\mathcal{K}_{\theta}^{n}(\cdot)$ is a small network (e.g., a series of convolutional and activation layers) that learns to predict the affine parameters. The transformed diffusion feature map $\hat{\pmb{F}}_{dif}^{n}$ is then given by: + +$$ +\hat {\pmb {F}} _ {d i f} ^ {n} = \alpha^ {n} (\pmb {F} _ {L R} ^ {n}, \pmb {F} _ {i n t e r} ^ {n}) \odot \pmb {F} _ {d i f} ^ {n} + \beta^ {n} (\pmb {F} _ {L R} ^ {n}, \pmb {F} _ {i n t e r} ^ {n}), +$$ + +where $\odot$ denotes element-wise multiplication. By injecting $F_{inter}^{n}$ and $F_{LR}^{n}$ into the diffusion process, the network can selectively emphasize or suppress certain features, effectively leveraging the intermediate features extracted from the reference image to guide super-resolution. During reverse diffusion, we iteratively denoise a latent $\mathbf{x}_T\sim \mathcal{N}(0,\mathbf{I})$ down to $\mathbf{x}_0$ , guided by the condition features $\pmb{F}_{\mathrm{cond}}$ . The final super-resolved output is: + +$$ +\mathbf {I} _ {\text {f i n a l}} = \mathcal {D} \left(\mathcal {U} \left(\mathbf {x} _ {T}, \boldsymbol {F} _ {\text {c o n d}}\right)\right), \tag {5} +$$ + +where $\mathcal{U}$ and $\mathcal{D}$ denote the pre-trained denoising U-Net and the latent decoder from [46], respectively. + +# 3.3.2. Training Objectives + +We supervise both the intermediate and final outputs using the following losses. Specifically, we define the loss for intermediate reconstruction as: + +$$ +\mathcal {L} _ {\mathrm {i n t e r}} = \left\| \mathbf {I} _ {\mathrm {i n t e r}} - \mathbf {I} _ {\mathrm {H R}} \right\| _ {1} + \lambda \cdot \operatorname {G A N} \left(\mathbf {I} _ {\mathrm {i n t e r}}, \mathbf {I} _ {\mathrm {H R}}\right), +$$ + +![](images/ac56ded9a41998f7b5b22b48e8e2ad863c2f1af6b8c5213c1743bb0fca9d5e9c.jpg) +Reference + +![](images/e4b1899cf91d500bc18201a55c7b3870b8d24b6b2f900bf2eeb0fe9c7ea28461.jpg) +TTSR [65] + +![](images/bd64ee75750e6252eff9ac2bd64648d4bfb43e5aa0b21edb9a466cda1ce48d02.jpg) +DATSR [5] + +![](images/259a43f02f6f98177d03f9810b542373cfb358fb2642162820ef6f29b0aac349.jpg) +DASR [35] + +![](images/148d121a9f8188c7365ffa7b3ffd2f427c404372fa3efc713e84a9f815167cea.jpg) +Ours + +![](images/9d21045d8c31cf981fffc851c5a8db5559089d9d35bd417864926ebf5cbfd99e.jpg) +Ground Truth + +
Quant. Result: Evaluation on CUFED5 (i.e., Reference-Provided setting)
MetricsModels for Eval
TTSR [65]DATSR [5]DASR [35]Ours
PSNR↑25.22426.7120.6920.56
SSIM↑0.7820.8380.6490.686
LPIPS↓0.2220.1520.1830.219
CLIP-IQA↑0.3720.4350.5360.693
MUSIQ↑66.32969.14968.35573.573
+ +Table 2. Model performance comparison on reference-provided settings to assess the intrinsic performance of the models. All models were trained on DF2K-OST. + +and the final loss as: + +$$ +\mathcal {L} _ {\text {f i n a l}} = \left\| \mathbf {I} _ {\text {f i n a l}} - \mathbf {I} _ {\mathrm {H R}} \right\| _ {2}. \tag {6} +$$ + +The overall loss is a weighted sum of the intermediate and final losses: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {i n t e r}} + \alpha \cdot \mathcal {L} _ {\text {f i n a l}}, \tag {7} +$$ + +where $\alpha$ is a hyperparameter that balances the importance of intermediate reconstruction versus final refinement. By jointly optimizing these two stages, the network learns to leverage both reference-based SR cues and diffusion refinement to produce high-quality high-resolution outputs. + +# 4. Experiments + +# 4.1. Experiment and Evaluation Settings + +We trained our model using the Adam optimizer [29]. All experiments were performed on NVIDIA RTX 3090 GPUs. + +![](images/469be52745e551c8eeb161aa62fa222ce8648fcc44733f998d1bf323142107b7.jpg) +Figure 4. Qualitative comparison on the CUFED5 dataset with respect to CUFED5 references. We compare TTSR [65], DATSR [5], DASR [35] and our method against the ground truth. +Figure 5. Composition of the database. We partition the real-domain dataset into a training set (75% of the data) and a reference set (25%). We then augment the reference set at a 1:1 ratio. + +# Reference Database (ref. DB) and Training Detail + +Fig. 5 illustrates the configuration of the dataset. We split the DF2K-OST dataset [1, 57] such that $75\%$ of the images form the training set for image restoration. The remaining $25\%$ is used to build the reference database, which is further augmented with synthetic images. To enrich this reference database, we add synthetic images in a one-to-one ratio with the real patches. Specifically, from each reference image, we extracted $512 \times 512$ patches (using reflective padding if needed) and trained an unsupervised hashing model on these patches using positive pairs generated via standard data augmentation [8]. Synthetic images were generated using the diffusers pipeline [54] and SDEdit [42], with guidance scales randomly chosen from [7, 10] and noise scales from [0.55, 0.65]. A hallucination threshold (average variance of 0.03) was applied; if exceeded, the generation was repeated up to 10 times, retaining the sample with the lowest variance. All processes utilized SD2.1 [46]. + +Evaluation Settings We evaluated the performance of our model on three benchmark datasets: the DIV2K validation set [1], the RealSR dataset [4], and the DRealSR dataset. For each dataset, we computed quality metrics—including PSNR, LPIPS [69], SSIM [61], CLIP-IQA [55], and MUSIQ [25]—to assess performance. These metrics comprehensively assess both the pixel-level fidelity and the perceptual quality of the generated images. + +Quantitative & Qualitative Results We validated our proposed approach using two distinct datasets: (i) a reference-provided dataset and (ii) a real-world image dataset as described in Fig. 5. As illustrated in Tab. 2, our method demonstrates superior perceptual quality in the matched dataset compared to other baselines. In particular, as shown in Fig. 4, fine details and realistic textures are preserved more effectively. Furthermore, as demonstrated in Fig. 3 and Tab. 1, our method consistently outperforms baseline models in real-world scenarios, thereby confirming its robustness and practical applicability. + +Ablation Study In Tab. 3, we compare our full method (Ours) with three ablated variants on the RealSR dataset. In Ours (-H), we remove hashing-based reference matching during training and instead use random pairing. This leads to a noticeable drop in performance (e.g., PSNR from 23.957 to 23.614, SSIM from 0.715 to 0.698), underscoring the importance of accurate reference selection. In Ours (-RM), the reference restoration module is removed, causing further degradation in visual quality (e.g., PSNR decreases to 23.374, SSIM to 0.680), indicating that this module is essential for effectively leveraging reference patches. Finally, in Ours (-R), hashing-based reference retrieval is replaced with random patch selection during inference, which also impairs reconstruction quality (PSNR drops to 23.709, SSIM to 0.708). As shown in Tab. 4, varying the lossbalancing weight $\alpha$ indicates that $\alpha = 1.0$ strikes the best trade-off between distortion metrics and perceptual quality. + +
Quant. Result: Evaluation on each modules
AblationPSNR↑SSIM↑LPIPS↓CLIP-IQA↑MUSIQ↑
Ours23.9570.7150.2640.67770.521
Ours (-H)23.6140.6980.3020.63966.146
Ours (-RM)23.3740.6800.2890.64568.122
Ours (-R)23.7090.7080.2760.66169.182
+ +Table 3. Ablation study on different variants of the proposed method on the RealSR dataset. + +
Quant. Result: Ablation study on weight α
AblationPSNR↑SSIM↑LPIPS↓CLIP-IQA↑MUSIQ↑
Ours (α = 0.7)22.4310.6370.2290.66567.102
Ours (α = 1.0)22.3230.6320.2280.70268.248
Ours (α = 1.3)22.0070.6170.2390.71667.913
+ +Table 4. Impact of the loss-balancing weight $\alpha$ (Eq. (7)) on distortion and perceptual quality. + +![](images/a0b1e73164682329b787c149df3d253013a29bbeffb674ee829d4397ffd50348.jpg) + +![](images/0ff82414761d9f83867659d014f510dab3c7f109492fa95b0ea042411e8e4ec3.jpg) +Figure 6. Effectiveness of variance (Hallucination) in diffusion trajectories during sampling [39]. Our empirical results indicate that the effect shown in [2] is also observed in the latent domain. Furthermore, the high variance sampling process even shows the oscillation in trajectory. + +# 4.2. Robust Data Augmentation via Generation + +We apply rejection sampling, as described in Eq. (2), to generate robust image samples. As illustrated in Fig. 6, higher variance in the sampling trajectory tends to yield unrealistic images (e.g., faces appearing on mountains), whereas lower variance produces images that preserve the core realistic attributes while incorporating desired edits. Furthermore, samples generated with high variance tend to exhibit persistently high variance values and lack a decaying trend during the initial timesteps. In contrast, samples generated with low variance exhibit an almost monotonic decrease in the trajectory, indicating smooth sampling in the vector field that likely guides the sample toward the proper mode. + +# 4.3. Augmented Data Improves Hashing + +We evaluate our approach by measuring the hit rate of the hash function, a key retrieval indicator. For each real image, augmented samples are generated as in Fig. 5; each CUFED query is paired with four references ranked by similarity. As shown in Tab. 5, generated-only training underperforms real-only training, but combining the two $(\text{Real} + \text{Gen})$ yields the highest hit rates and retrieves the correct patches in proper order (Fig. 7). These results show that training with diverse patches strengthens invariant representations and overall retrieval performance. Crucially, even when we restrict the real set to just $10\%$ of its original size, augmenting it with a nine-fold amount of generated patches $(\text{Real}_{\text{lim}} + \text{Gen})$ almost restores—and in some ranks surpasses—the full-data baseline, highlighting generation's value in data-scarce scenarios. + +Quant. Result: Hit rate $(\%)$ of correct CUFED ref. patch retrieval + +
Train ConfigurationRank of the ref. patch Similarity
Train. DB#samplesR1R2R3R4R5
Real13K4559647075
Gen13K4052596366
Real+Gen26K5260667173
Reallim1K4855606367
Reallim + Gen10K5262636771
+ +Table 5. Effectiveness of generated samples. Evaluation is conducted on the CUFED dataset to compare the hit rate for models trained on Real vs. Gen patches, as well as a combination of both. + +![](images/f61edf45c26e6c487fd12ac8eda808e6b3b617bd6964473434272d084ff42585.jpg) +Figure 7. Effectiveness of generated samples. Hash trained with both (Gen+Real) dataset retrieves the most similar patches. + +# 4.4. Hashing Time and Resource Usage + +Tab. 6 shows that our hashing-based retrieval approach reduces both storage memory and matching cost compared to VGG-based feature matching. We assume 26,500 reference images $(3 \times 512 \times 512)$ , where each reference feature or hash code is precomputed, and 80,000 query images are matched in batches of 64. By compressing the code length from 4096 to just 16, we shrink the memory footprint from more than $800\mathrm{MB}$ to less than $52\mathrm{KB}$ while also cutting the retrieval time from $4338.2\mathrm{ms}$ to $67.8\mathrm{ms}$ . + +Quant. Result: Computational Cost + +
AlgorithmCode LengthEvaluation Metric Storage MemoryCost
VGG Matching4096828.13 M4338.2 ms
Hashing1651.76 KB67.8 ms
+ +# 4.5. Integrating Hashing into Existing Ref-SR + +As described in Tab. 8 and Fig. 8, integrating hashing into conventional Ref-SR models [65] efficiently retrieves structurally correlated regions from references, enabling more + +precise alignment and fusion for higher-fidelity reconstructions. In diffusion-based Ref-SR, CoSeR [53] first generates a reference from the LR input; replacing or augmenting it with our hash-selected references yields consistent gains in PSNR/SSIM and no-reference IQA (CLIP-IQA, MUSIQ), as shown in Tab. 7. + +Table 6. Effectiveness of Hash on computational cost. Comparison conducted with VGG-based matching and our hashing method. We assume 26,500 reference images with precomputed codes, and 80,000 query images matched. + +
Quant. Result: Reference adaptation in CoSeR
Ref. TypePSNR↑SSIM↑LPIPS↓CLIP-IQA↑MUSIQ↑
Gen20.230.5170.4050.56256.419
Hash20.300.5190.4080.56657.686
+ +Table 7. Impact of hash-retrieved references on diffusion-based RefSR [53]; metrics compare generated (Gen) and hash-selected (Hash) reference strategies. + +
Quant. Result: PSNR/SSIM
model: TTSRRef. at Eval
DatasetRef. at TrainRandomHash
DIV2KRandom28.26 / 0.77128.25 / 0.770
Hash29.08 / 0.79329.16 / 0.794
CUFEDRandom24.74 / 0.73624.73 / 0.736
Hash25.34 / 0.74825.65 / 0.762
+ +Table 8. Effectiveness of Hash for conventional RefSR [65]. Evaluation was conducted using different reference-selection strategies on DIV2K and CUFED. Each cell shows the average PSNR/SSIM under two factors: reference selection (Random vs Hash) at training and evaluation. + +![](images/d38f5bc5b4e81cad2b5f53c063c0cfc1d4b79a1cd541a5ab9cbb96139ef14241.jpg) +Figure 8. Effectiveness of Hash. The two groups on the left compare reference patches selected randomly or those retrieved using a hashing. Hashing mechanism enables the RefSR model to perform effectively even when an exact reference patch is unavailable. + +# 5. Conclusion + +We proposed a novel image-based Retrieval Augmented Generation framework that combines latent diffusion models with an efficient hashing code vector strategy achieving robust reference matching and realistic reference-based SR. Operating in a compact latent space by short binary hash codes, our method addressed the challenges of reference selection and improves domain consistency between low-resolution inputs and high-resolution references. Experiments on real-world datasets demonstrate that our approach outperforms existing diffusion-based superresolution methods and reference-based methods in terms of fidelity, perceptual quality, and computational efficiency. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 4, 6, 7 +[2] Sumukh K Aithal, Pratyush Maini, Zachary Lipton, and J Zico Kolter. Understanding hallucinations in diffusion models through mode interpolation. Advances in Neural Information Processing Systems, 37:134614-134644, 2025. 3, 7 +[3] Jerone Andrews, Dora Zhao, William Thong, Apostolos Modas, Orestis Papakyriakopoulos, and Alice Xiang. Ethical considerations for responsible data curation. Advances in Neural Information Processing Systems, 36, 2024. 3 +[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 1, 4, 7 +[5] Jiezhang Cao, Jingyun Liang, Kai Zhang, Yawei Li, Yulun Zhang, Wenguan Wang, and Luc Van Gool. Reference-based image super-resolution with deformable attention transformer. In European conference on computer vision, pages 325-342. Springer, 2022. 4, 5, 6 +[6] Hao-Wei Chen, Yu-Syuan Xu, Min-Fong Hong, Yi-Min Tsai, Hsien-Kai Kuo, and Chun-Yi Lee. Cascaded local implicit transformer for arbitrary-scale super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18257-18267, 2023. 1 +[7] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2 +[8] 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. 4, 6 +[9] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491, 2022. 2, 3 +[10] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 1 +[11] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2 +[12] 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 +[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2 +[14] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: + +Non-linear independent components estimation. arXiv preprint arXiv:1410.8516, 2014.2 +[15] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 2 +[16] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 3 +[17] Huiyuan Fu, Fei Peng, Xianwei Li, Yejun Li, Xin Wang, and Huadong Ma. Continuous optical zooming: A benchmark for arbitrary-scale image super-resolution in real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3035-3044, 2024. 1 +[18] Hua Gao, ChenChen Hu, Guang Han, Jiafa Mao, Wei Huang, and Kaiyuan Wan. Hashneck is a boosting tool for deep learning to hashing. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 83-91, 2024. 3 +[19] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2, 2023. 2, 3 +[20] 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. 1, 2, 3 +[21] 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 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3 +[23] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 1 +[24] Yuming Jiang, Kelvin CK Chan, Xintao Wang, Chen Change Loy, and Ziwei Liu. Robust reference-based super-resolution via c2-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2103-2112, 2021. 2, 3 +[25] 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. 7 +[26] Seunghoi Kim, Chen Jin, Tom Diethe, Matteo Figini, Henry FJ Tregidgo, Asher Mullokandov, Philip Teare, and Daniel C Alexander. Tackling structural hallucination in image translation with local diffusion. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 3 +[27] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 2 +[28] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 + +[29] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[30] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 2 +[31] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 3 +[32] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1929-1938, 2022. 1 +[33] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459-9474, 2020. 2, 3 +[34] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 1, 3 +[35] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 4, 5, 6 +[36] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 3 +[37] Kevin Lin, Jiwen Lu, Chu-Song Chen, and Jie Zhou. Learning compact binary descriptors with unsupervised deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1183-1192, 2016. 3 +[38] Hanyuan Liu, Jinbo Xing, Minshan Xie, Chengze Li, and Tien-Tsin Wong. Improved diffusion-based image colorization via piggybacked models. arXiv preprint arXiv:2304.11105, 2023. 2 +[39] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 7 +[40] Xiaoyi Liu and Hao Tang. Diffnno: Diffusion fourier neural operator. arXiv preprint arXiv:2411.09911, 2024. 1 +[41] Liying Lu, Wenbo Li, Xin Tao, Jiangbo Lu, and Jiaya Jia. Masa-sr: Matching acceleration and spatial adaptation for reference-based image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6368-6377, 2021. 2 +[42] 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. arXiv preprint arXiv:2108.01073, 2021. 3, 6 +[43] Xiushan Nie, Xingbo Liu, Jie Guo, Letian Wang, and Yi-long Yin. Supervised discrete multiple-length hashing for image retrieval. IEEE Transactions on Big Data, 9(1):312-327, 2022. 3 +[44] Zexuan Qiu, Qinliang Su, Zijing Ou, Jianxing Yu, and Changyou Chen. Unsupervised hashing with contrastive information bottleneck. arXiv preprint arXiv:2105.06138, 2021. 3, 4 +[45] Vikram V Ramaswamy, Sing Yu Lin, Dora Zhao, Aaron Adcock, Laurens van der Maaten, Deepti Ghadiyaram, and Olga Russakovsky. Geode: a geographically diverse evaluation dataset for object recognition. Advances in Neural Information Processing Systems, 36, 2024. 3 +[46] 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, 4, 5, 6 +[47] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 1 +[48] Gyumin Shim, Jinsun Park, and In So Kweon. Robust reference-based super-resolution with similarity-aware deformable convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8425-8434, 2020. 2, 3 +[49] Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, and Ross Anderson. The curse of recursion: Training on generated data makes models forget. arXiv preprint arXiv:2305.17493, 2023. 3 +[50] Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal. Ai models collapse when trained on recursively generated data. Nature, 631 (8022):755-759, 2024. 3 +[51] 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. 1, 2 +[52] 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. 1 +[53] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. arXiv preprint arXiv:2311.16512, 2023. 8 +[54] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.3,6 + +[55] 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. 7 +[56] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. 2024. 4, 5 +[57] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 6 +[58] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 3 +[59] 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. 1, 4 +[60] Yujin Wang, Quanfeng Liu, Jiaqi Fan, Jinlong Hong, Hongqing Chu, Mengjian Tian, Bingzhao Gao, and Hong Chen. RAC3: Retrieval-augmented corner case comprehension for autonomous driving with vision-language models. arXiv preprint arXiv:2412.11050, 2024. 3 +[61] 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. 1, 7 +[62] Min Wei and Xuesong Zhang. Super-resolution neural operator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18247-18256, 2023. 1 +[63] Yanchun Xie, Jimin Xiao, Mingjie Sun, Chao Yao, and Kaizhu Huang. Feature representation matters: End-to-end learning for reference-based image super-resolution. In European Conference on Computer Vision, pages 230-245. Springer, 2020. 2, 3 +[64] Chenggang Yan, Biao Gong, Yuxuan Wei, and Yue Gao. Deep multi-view enhancement hashing for image retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(4):1445-1451, 2020. 3 +[65] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5791-5800, 2020. 2, 3, 4, 5, 6, 8 +[66] Li Yuan, Tao Wang, Xiaopeng Zhang, Francis EH Tay, Zequn Jie, Wei Liu, and Jiashi Feng. Central similarity quantization for efficient image and video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3083-3092, 2020. 3 +[67] 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. 1, 4 +[68] 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 +[69] 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. 7 +[70] Zhifei Zhang, Zhaowen Wang, Zhe Lin, and Hairong Qi. Image super-resolution by neural texture transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7982-7991, 2019. 2, 3 +[71] Haitian Zheng, Mengqi Ji, Haoqian Wang, Yebin Liu, and Lu Fang. Crossnet: An end-to-end reference-based super resolution network using cross-scale warping. In Proceedings of the European conference on computer vision (ECCV), pages 88-104, 2018. 2, 3 \ No newline at end of file diff --git a/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/images.zip b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b52f10252fceb520099fdb340007199e26145e4b --- /dev/null +++ b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:543dcccc872c9fddeac01ace6c4c5bb7e67a28a2f9917bf645c836e3abb897ec +size 1097681 diff --git a/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/layout.json b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1fcae5b48570fb2529a9e510390e607691215920 --- /dev/null +++ b/ICCV/2025/Reference-based Super-Resolution via Image-based Retrieval-Augmented Generation Diffusion/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8657fda7e3f9691fe8b5fa7cde2cda13db068cb6ec49122d4b91ce9b37584edd +size 438498 diff --git a/ICCV/2025/Referring Expression Comprehension for Small Objects/9584ea89-42d5-450f-a1e2-eadc02ff019e_content_list.json b/ICCV/2025/Referring Expression Comprehension for Small Objects/9584ea89-42d5-450f-a1e2-eadc02ff019e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b091d3b5779da77bba79e7c2cb49386caafca22 --- /dev/null +++ b/ICCV/2025/Referring Expression Comprehension for Small Objects/9584ea89-42d5-450f-a1e2-eadc02ff019e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcadfa3d7ac0c22364507372322bbba3ca553ba4b1bb85977db533f52ce56fae +size 101210 diff --git a/ICCV/2025/Referring Expression Comprehension for Small Objects/9584ea89-42d5-450f-a1e2-eadc02ff019e_model.json b/ICCV/2025/Referring Expression Comprehension for Small Objects/9584ea89-42d5-450f-a1e2-eadc02ff019e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..98b07ccc05048769123353edb750c8b3e3e5b77e --- /dev/null +++ b/ICCV/2025/Referring Expression Comprehension for Small Objects/9584ea89-42d5-450f-a1e2-eadc02ff019e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03a02f54736a80d581e2c38b3f71f0c17a6342c55d9f9fbc0432f5b04a296a5d +size 130912 diff --git a/ICCV/2025/Referring Expression Comprehension for Small Objects/9584ea89-42d5-450f-a1e2-eadc02ff019e_origin.pdf b/ICCV/2025/Referring Expression Comprehension for Small Objects/9584ea89-42d5-450f-a1e2-eadc02ff019e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..060bb3507194206a09a7b539e374855ff18b2c15 --- /dev/null +++ b/ICCV/2025/Referring Expression Comprehension for Small Objects/9584ea89-42d5-450f-a1e2-eadc02ff019e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10fe39fe00c1ddc0472b2780e7cef097def665c809a42f0254dc5b8d05f2a131 +size 12363502 diff --git a/ICCV/2025/Referring Expression Comprehension for Small Objects/full.md b/ICCV/2025/Referring Expression Comprehension for Small Objects/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e261dd046832a79062206a1598a9290fdec0a9b5 --- /dev/null +++ b/ICCV/2025/Referring Expression Comprehension for Small Objects/full.md @@ -0,0 +1,430 @@ +# Referring Expression Comprehension for Small Objects + +Kanoko Goto $^{1*}$ Takumi Hirose $^{1*}$ Mahiro Ukai $^{1}$ Shuhei Kurita $^{2,1}$ Nakamasa Inoue $^{1}$ + +$^{1}$ Institute of Science Tokyo $^{2}$ National Institute of Informatics + +Project Page: https://github.com/mmaiLab/sorec + +# Abstract + +Referring expression comprehension (REC) aims to localize the target object described by a natural language expression. Recent advances in vision-language learning have led to significant performance improvements in REC tasks. However, localizing extremely small objects remains a considerable challenge despite its importance in real-world applications such as autonomous driving. To address this issue, we introduce a novel dataset and method for REC targeting small objects. First, we present the small object REC (SOREC) dataset, which consists of 100,000 pairs of referring expressions and corresponding bounding boxes for small objects in driving scenarios. Second, we propose the progressive-iterative zooming adapter (PIZA), an adapter module for parameter-efficient fine-tuning that enables models to progressively zoom in and localize small objects. In a series of experiments, we apply PIZA to GroundingDINO and demonstrate a significant improvement in accuracy on the SOREC dataset. Our dataset, codes and pre-trained models are publicly available on the project page. + +# 1. Introduction + +Object localization in images has been a long-term research topic in the field of computer vision. Early studies introduced image datasets such as Pascal VOC [14] and COCO [39], which involve bounding box annotations for predefined object categories, leading to the development of object detection models including CNN-based models [15, 16, 19, 60] and Transformer-based models [1, 21, 29, 41, 50, 64, 82, 85, 92]. For more detailed and flexible object localization, referring expression comprehension (REC) aims to localize a specific object referred to by a natural language description. REC uses queries like "the red car parked in front of the coffee shop" as input and requires locating this unique object in an input image. Ref- + +![](images/44cfa18dee852df89416a87da36c93bc962bf8ba496fa013f28b111e2521aee0.jpg) + +![](images/edf0d62c56901520280f0d0cd21bd7b04577c79a4c8edd0c666f36a00c03cdb8.jpg) + +# Referring expression $t$ + +The sandwich board with white text advertising breakfast, lunch, sandwiches, paninis, tea, and coffee is placed on the sidewalk next to a red rectangular sign with rounded corners, in front of a storefront. + +learn to localize the target + +![](images/02f6ead8e1870333c29ff678618c2a1d7eba0263d5db5065ff6cbe336226a608.jpg) + +(a) Existing approach + +learn to zoom in and localize the target + +![](images/cacfff3006541d9b91f24de225450297e3efdb05801738a2da3fb51e32bf12ca.jpg) + +(b) Our approach + +Input image $x$ + +![](images/91ce642d8500aa771783328c1bf16d259d273525e0ae01b8e7ca434f09028a0c.jpg) +Figure 1. The SOREC dataset consists of pairs of referring expressions and bounding boxes for extremely small objects. (a) Existing approach fine-tunes a model $F$ to localize the target. (b) Our approach fine-tunes $F$ to progressively zoom in and localize the target in an autoregressive manner. (c) Example of prediction in three zooming steps. + +![](images/df8ae9d9e1fed7fe1cbc2bfcce4494ed0d5822d1b6207e3ecaa71adb7749e0a5.jpg) + +Zoom $x_{2}$ + +![](images/f48305bc19ef8730278591bd0660f7d04a09d28210be1995a6ec7a4f823be780.jpg) + +(c) Autoregressive prediction + +COCO, RefCOCO+ and RefCOCOg [53, 78] are the most popular datasets for REC, providing referring expressions for images in the COCO dataset along with Flickr30K entities [57]. Over the past decade, deep learning architectures that bridge visual contents and natural language descriptions have been investigated. Examples include one-stage architectures [38, 48, 62, 75] and two-stage architectures based on CNN-LSTM [24, 40, 49, 52, 54, 78, 79, 83] and attention mechanism [8, 12, 25, 43, 48, 66, 80, 94]. + +With advances in large-scale vision-language learning, recent models have become capable of precisely understanding object attributes and relations described in natural language [35, 42, 69, 70, 73, 84, 87]. As a result, these models have achieved high accuracy in REC tasks, with accuracy rates of over $90\%$ on the RefCOCO test sets. However, localizing small objects remains a significant challenge. The lack of REC datasets targeting small objects has impeded further progress in this area, despite its critical role in real-world applications such as autonomous driving, where the ability to localize small objects is essential for ensuring safety and facilitating precise decision-making in complex environments. + +To address this issue, the present study makes two significant contributions. First, we introduce the small object REC (SOREC) dataset, a new dataset for REC targeting small objects in autonomous driving scenarios. Second, we propose the progressive-iterative zooming adapter (PIZA), an adapter module for parameter-efficient fine-tuning that enables models to progressively zoom in and localize small objects. Below we highlight each contribution. + +1) Dataset contribution. We propose the SOREC dataset, which consists of 100,000 pairs of referring expressions and corresponding bounding boxes for extremely small objects in road, highway, rural-area, and off-road scenes. As shown in Figure 1, the typical size of a bounding box is approximately $0.1\%$ of the input image size, making it challenging to localize these objects. To the best of our knowledge, this is the first dataset for REC targeting small objects in autonomous driving scenarios. The availability of this dataset promotes further research advancements in the area. + +2) Technical contribution. We propose PIZA, a lightweight learnable module for parameter-efficient fine-tuning. Through fine-tuning with PIZA, the model learns to localize small objects in an autoregressive manner, where a zoomed image is fed into the model iteratively, as shown in Figure 1 (b-c). This approach significantly improves accuracy, as demonstrated in Table 3. + +# 2. Related work + +# 2.1. Tasks and datasets + +Referring expression comprehension. Considerable efforts have been dedicated to constructing datasets of referring expressions on images over the past decade. ReferItGame [30] was a pioneering large-scale dataset for REC, consisting of 130k expressions for 20k images collected from ImageCLEF and SAIAPR. RefCOCO, RefCOCO+ [78] and RefCOCOg [53] provided expressions for images in COCO [39], and are the most popular benchmarking datasets. CLEVR-Ref [45] is a diagnostic dataset focusing on compositional language understanding using synthetic images. Refer360 [71] is a dataset for referring ex + +pression recognition in 360-degree images. REVERIE [58] offers a dataset for remote embodied visual referring expressions in real indoor environments. RefEgo [33] focuses on egocentric REC in first-person videos. + +Small object detection. The importance of small object detection has been recognized in various object detection scenarios. Examples of datasets involving small objects include WiderFace [74] for face detection, TinyPerson [81] for person detection, TT100K [93] for traffic sign detection, VisDrone [91] for drone-based detection, and DOTA [11] for remote sensing detection. SODA [5] is the latest large-scale dataset for small object detection in automatic driving scenarios. Compared with normal-sized object detection datasets such as Pascal VOC [14] and COCO [39], creating datasets for small object detection is generally more expensive as annotating small objects is more challenging. + +Other related tasks. Visual grounding (VG) also aims to localize objects given natural language descriptions. In VG, each image may contain multiple target objects, typically described with shorter phrases than those used in REC. Example VG datasets include SK-VG [4] and GigaGrounding [51]. Open-vocabulary object detection aims to detect objects that are not seen in the training dataset. For training open-vocabulary detection models, large object detection datasets are recently used such as O365 [63], GoldG [29], GRIT [56] and V3Det [68]. + +# 2.2. Models + +REC models have evolved significantly over the years, transitioning from traditional CNN-LSTM architectures to attention and transformer-based architectures. One-stage REC models [38, 48, 62, 75] integrated object detection and language grounding into a unified architecture, allowing for end-to-end training. Two-stage approaches [24, 40, 49, 52, 54, 78, 79, 83] utilized region proposals generated by object detectors and applied LSTM to encode the referring expressions. Attention mechanisms were later incorporated to improve the alignment between image regions and referring expressions [8, 12, 25, 43, 48, 66, 80, 94]. + +To cover multiple vision tasks recent studies have demonstrated the effectiveness of large-scale vision-language pre-training [7, 36, 42, 69, 70, 73, 84, 87]. For example, open-set detection models such as GLIPv2 [84] and Grounding DINO [42] can handle both object detection and REC. We chose GroundingDINO as the baseline model because it is pre-trained on a union of datasets, including those for relatively small object detection and REC. + +# 2.3. Parameter efficient fine-tuning + +Prompt-based fine-tuning. Inspired by prompt tuning methods for natural language processing tasks [27, 44, 65], prompt-based fine-tuning methods have been proposed for computer vision tasks. Context optimization (CoOp) [89] + +![](images/9197d166162e6d8ba642ba8b17832960931642fef58f2ad1c2c017bf6bb1f8ba.jpg) +(a) RefCOCO + +![](images/8c79527e62aa9133c6df13b91392322f7aac1911d253ee6c0201496b6e172bd2.jpg) + +![](images/48222fec5f1168a4213eedeccde1ea8175738c193d72f29722bb2b70c0d70d1b.jpg) + +(b) SREC Dataset (ours) +![](images/70870f4312727c583d982174e09b5243c9a85b535d37f5e361a24213df7c374e.jpg) +The circular sign with a large white "P" in the center is situated above a yellow and brown sign that reads "ACIONAMENTO". + +Figure 2. Dataset comparison. (a) RefCOCO is a representative REC dataset consisting of expressions and bounding boxes for normalized objects. (d) SOREC is our dataset, consisting of relatively longer expressions compared to RefCOCO, to identify small objects. (c-e) Comparison of word count, image size, and relative bounding box size distributions on test sets. +![](images/e24d47e15ce1bcfabbd0a94ca10fdfdbd471d220dbcdfeead884b3ba64f88e1d.jpg) +The yellow traffic sign with a blue square and a white triangle, depicting a pedestrian crossing, is located near a bus stop sign and in front of a building with visible text. + +![](images/44802dc726a90b6e4072b75a421577e4da999658cd8937a080ec162aa28a9b7e.jpg) + +![](images/63a1c063ece7eaa4e77b1a2908ac3f1402c9552e3823c8ab4e28306ef5e74ae9.jpg) + +![](images/54657adc4f3c9b943395e234282c4173fdddf4ad5f060b46cd4d9d835d33562f.jpg) + +![](images/0b2e0b1bfcd77a5ce4546b15da6a92033a988ead1c8d0a2d4fec0718e31eb63a.jpg) + +![](images/1ae09759ae5344ad65e70530f16ade68b98749b313bc0c4a45ebb09e694fb972.jpg) +The person wearing a yellow helmet and dark clothing is standing on a red platform in the water, surrounded by several white buoys. + +![](images/8029c95f0cb82fde24ce6735ac71b292ea2234425d4387fe13a20880d488e0e0.jpg) + +![](images/18a9af755f73c7e57fc3d8f5f0dfd7c9cdf3b7f7a0ff6ed26be9644833742fc8.jpg) + +and visual prompt tuning (VPT) [26] are two representative methods. CoOp incorporated learnable embeddings into the text encoder of CLIP [59]. CoCoOp [88] introduced prompts conditioned by image features. VPT incorporated learnable embeddings into the vision transformers [13]. Further extension includes distribution learning, multi-modal learning and various techniques to leverage pre-trained knowledge [6, 9, 31, 32, 47, 76, 77, 86, 90]. + +LoRA-based fine-tuning. LoRA [23] introduces low-rank adaptations to the weight matrices of a pre-trained model. LoRA reduces the number of trainable parameters by decomposing the weight updates into low-rank matrices. For further improving parameter efficiency, quantization techniques are also introduced [10, 72]. + +Adapter-based fine-tuning. Adapters are lightweight learnable modules incorporated into a frozen pretrained model. The first adapter architecture [22] was proposed for Transformers, which inserts adapters into the attention module and the feed forward network module in each encoder layer. There have been substantial efforts in architectural adapter design for various neural networks for computer vision tasks [2, 3, 17, 18, 28, 67]. Adapter+ [67] is a well-designed adapter architecture for vision transformers, which we employ in our experiments. + +# 3. SOREC Dataset + +The SOREC dataset consists of 100,000 pairs of referring expressions and corresponding bounding boxes for small objects in road, highway, rural, and off-road images. As shown in Figure 2 (b), referring expressions describe both the characteristics of the target object and its spatial relationships with surrounding objects, in order to locate target objects. Each bounding box typically occupies approximately $0.05\%$ of the entire image area. Compared to exist- + +ing datasets such as RefCOCO [78] in Figure 2 (a), SOREC presents a particularly challenging task due to the extremely small bounding boxes. This challenge is critical for advancing real-world applications including autonomous driving and surveillance, where detecting small objects is essential. + +# 3.1. Dataset Construction + +The SOREC dataset is semi-automatically created in the following five steps. + +1) Source selection. We selected the SODA-D dataset [5] as our source dataset. It consists of 24,828 high-quality images for small object detection, collected primarily from the Mapillary Vistas dataset [55]. The average resolution of these images is $3407 \times 2470$ pixels. +2) Segmentation. To extract small object regions, we applied Semantic-SAM [34] to image patches of size $800 \times 800$ pixels in a sliding window manner, with the granularity prompt level set to 3. Bounding boxes of object region were also computed. We excluded object regions whose bounding boxes occupied more than $2\%$ of the image. +3) Filtering. Through manual inspection of the extracted object regions, we found many instances of trees and windows that are not suitable for REC. We filtered them out by computing CLIP scores of each bound box region using a prompt of "tree, forest, window." Subsequently, we sorted the object regions based on the score $S \exp(-|a - 1|) p$ , where $S$ is the bounding box size, $a$ is the aspect ratio and $p$ is the stability score obtained from Semantic-SAM. From the top 200,000 results, we excluded any remaining meaningless objects and selected the top 150,000 results through crowdsourcing. +4) Referring expression generation. For each object region, we cropped an image centered on its bounding box, with a random height between 2.5 and 3.5 times the height + +![](images/77b3df34e85b4dc7da73bbf329cd117b338fbb3daeca35348ca8666ead7c151f.jpg) +Figure 3. Word clouds for RefCOCO (left) and SOREC (right). + +
SplitImagesExpressions
Train-S1,44610,000
Train-L13,49461,369
Validation2,38210,712
Test-A4,10710,815
Test-B5,15317,104
+ +Table 1. Dataset split + +of the bounding box, and a width between 1.5 and 3 times its height. We drew a red bounding box on the image with a line thickness of 2 pixels. We input these images into GPT-40 to generate initial referring expressions using the prompt "Write a sentence that refers to the object in the red frame by mentioning its details, colors, relative relationship to surrounding objects." We excluded cases where the generated descriptions did not include references to surrounding objects by counting nouns. + +5) Quality control. Finally, using crowdsourcing again, we excluded object regions that could not be uniquely identified by the corresponding expression. This step was a binary decision on whether the object could be identified by the expression, allowing for minor errors in descriptions of surrounding objects, and chose 100,000 high-quality pairs of bounding boxes and expressions. For test sets, we further asked annotators to revise expressions if they involve errors. As a result, $18.45\%$ of sentences were found to contain minor errors related to color, spatial relations, and similar attributes. + +# 3.2. Dataset statistics + +Description length. Figure 2 (c) shows the distribution of the number of words per expression. As shown, the average number of words is 25.5, which is approximately seven times longer than the 3.52 words in RefCOCO. This is due to the need for more detailed information to identify small objects, which is a key characteristic of this dataset. Figure 3 compares word distributions by word clouds. As shown, SOREC dataset contains words related to positioning, such as 'positioned' and 'situated'. + +Bounding box size. Figure 2 (d) shows the distribution of bounding box sizes relative to the image sizes. As indicated, all the target bounding boxes occupy less than $1\%$ of the image area, presenting a challenging REC task. Since pretraining is often performed on datasets that feature normal sized objects, fine-tuning is necessary to bridge this gap for localizing small target object. + +Image size. Figure 2 (e) shows the distribution of image sizes. As shown, the SOREC dataset consists of high-resolution images that are sufficient for performing REC targeting small objects. + +Data split. We created training, validation, test splits as summarized in Table 1. The train-L set is the full training + +set, and the train-S set is a small subset consisting of 10,000 expressions. The validation set consists of 10,712 expressions, with no overlap in images with the training sets. The test-A and -B sets contain expressions for traffic objects and the other objects. + +# 4. Method + +This section describes PIZA, a lightweight adapter module for parameter-efficient fine-tuning that enables models to localize target small objects by progressively and iteratively zooming into them. + +# 4.1. Preliminary + +Problem settings. Let $\boldsymbol{x} \in \mathbb{R}^{W \times H \times C}$ be an input image, where $W$ is the width, $H$ is the height, and $C$ is the number of channels. We denote by $\boldsymbol{b} = (x_0, y_0, x_1, y_1) \in \mathbb{R}^4$ a bounding box, where $(x_0, y_0)$ is the coordinate of the top-left corner, and $(x_1, y_1)$ is the coordinate of the bottom-right corner of the box. Given a natural language expression $t$ , the goal of REC is to localize the object corresponding to $t$ in the image. As such, the model $F$ takes as input $(\boldsymbol{x}, t)$ and learns to predict a bounding box as $\hat{\boldsymbol{b}} = F(\boldsymbol{x}, t)$ , so that the predicted bounding box $\hat{\boldsymbol{b}}$ matches the ground truth bounding box $\boldsymbol{b}^*$ . This work focuses on the setting where the size of $\boldsymbol{b}^*$ is significantly smaller than the image size, i.e., we assume that $|\boldsymbol{b}^*| \ll WH$ , where $|\boldsymbol{b}| = (x_1 - x_0)(y_1 - y_0)$ indicates the size of $\boldsymbol{b}$ . + +Vision-language pre-training. We assume that a pretrained model $F$ is given and explore parameter-efficient fine-tuning methods by which the model quickly adapts to localize small objects. + +Difficulty. The main difficulty lies in the gap between pretraining and fine-tuning regarding the bounding box sizes and expression lengths. Recent object localization models, such as GroundingDINO [42], are capable of both REC and open-set object detection because they are pre-trained on a large union set of REC and object detection datasets. However, a challenge arises when dealing with a combination of long sentences and extremely small objects. We aim to address this challenge in parameter-efficient fine-tuning. + +![](images/654fa4d20849ba914659db6e89e404db5fbee564fba8fc31f3fca42301e1b163.jpg) +Figure 4. Fine-tuning with PIZA. Given a pre-trained model $F$ , PIZA produces a model $F_{\text{总}}$ that zooms in to localize small objects in an autoregressive manner through fine-tuning. In the inference phase, bounding boxes $b_{0}, b_{1}, \dots, b_{T}$ indicating zooming steps are predicted to localize the target at the end. + +# 4.2. Progressive-integrative zooming adapter + +Overview. Zooming in to localize a small target object can be understood as a search problem over an image. We model a search process $P$ as a sequence of bounding boxes: + +$$ +P = \left(\boldsymbol {b} _ {0}, \boldsymbol {b} _ {1}, \dots , \boldsymbol {b} _ {T}\right), \tag {1} +$$ + +where $\pmb{b}_0 = (0,0,W,H)$ indicates the bounding box covering the entire image, $\pmb{b}_T$ is the final small bounding box for localizing the target, $T$ is the number of zooming steps, and $\pmb{b}_i$ covers $\pmb{b}_j$ if $i < j$ , as shown in Figure 1 (c). + +Through fine-tuning, models learn to predict search processes so that the final bounding box matches the ground truth, i.e., $\boldsymbol{b}_T \simeq \boldsymbol{b}^*$ . To this end, PIZA extends the pretrained model $F$ to a model $F_*$ that predicts a search process in an autoregressive manner as + +$$ +\hat {\boldsymbol {b}} _ {i + 1} = F _ {*} \left(\boldsymbol {x} _ {i}, \boldsymbol {t}, \boldsymbol {b} _ {0: i}\right), \tag {2} +$$ + +where $\pmb{x}_i$ is the cropped image region corresponding to $\pmb{b}_i$ , $\pmb{t}$ is an input expression, $\pmb{b}_{0:i} = (\pmb{b}_0, \dots, \pmb{b}_i)$ is a subsequence of bounding boxes and $\circledast$ indicates that PIZA is applied. A visualization of this procedure is shown in Figure 4. Since $F$ is a function that accepts two inputs, $\pmb{x}$ and $\pmb{t}$ , $F_{\text{串}}$ is built up by incorporating a module that can take $\pmb{b}_{0:i}$ as an additional input into $F$ . Below we describe the module architecture. + +PIZA module. Inspired by time-step embeddings in diffusion models [20, 61] that represent stages of the diffusion process, the PIZA module learns zooming-step embeddings that represent progress of the search process. Specifically, the zooming-step embedding $h \in \mathbb{R}^d$ is extracted from the input sequence of bounding boxes $b_{0:i} \in \mathbb{R}^{4 \times (i+1)}$ in two steps. First, a sequence of low-level features $l_{0:i} = (l_0, l_1, \dots, l_i)$ is extracted. Each feature $l_j$ is a 6-dimensional vector, and its elements are listed in Table 2. Second, $h$ is extracted by feeding $l_{0:i}$ into a small learnable + +
FeatureDefinition
Normalized sizesj=|bj|/|b0|
Relative sizer0=1, rj=|bj|/|bj-1|
Normalized widthwj=(x1(j)-x0(j))/W
Normalized heighthj=(h1(j)-h0(j))/H
Center position (x-axis)x̅j=(x0(j)+x1(j))/(2W)
Center position (y-axis)ŷj=(y0(j)+ŷ1(j))/(2H)
+ +Table 2. Low-level features extracted from the sequence of bounding boxes $\mathbf{b}_j = (x_0^{(j)}, x_1^{(j)}, y_0^{(j)}, y_1^{(j)})$ . $W$ and $H$ denotes the width and height of the input image. + +module. Figure 5 shows the architecture consisting of a sequence of learnable Fourier embeddings [37], a transformer encoder and an average pooling layer. The embeddings are trained with two heads: an EOS head and a progress head. The EOS head predicts a binary label $\hat{y}_{i+1} \in \{[\text{CONT}], [\text{EOS}]\}$ indicating either "continue to search (CONT)" or "end of search (EOS)". The progress head predicts progress of search $\hat{z}_{i+1} \in [0,1]$ expressed as a real value, where 0.0 indicates the start and 1.0 indicates the end of the search process. The binary cross-entropy loss and mean squared error loss are applied to these heads, respectively, on the extended training dataset described in Section 4.4. In the inference phase, search is stopped when the EOS label is predicted. The number of parameters of this module is 0.27M and the feature dimension is set to 16. The detailed architectural hyperparameters are provided with our code. + +# 4.3. Parameter-efficient fine-tuning with PIZA + +To perform fine-tuning with the PIZA module, we incorporate the embeddings $\pmb{h}$ into parameter-efficient fine-tuning methods. Here, we propose prompt-based, LoRA-based and adapter-based fine-tuning with PIZA. Their architectures are shown in Figure 6. + +PIZA-CoOp. CoOp [89] is a prompt-based fine-tuning method, which prepends learnable embeddings to the input text prompts as $G(\boldsymbol{x}, \boldsymbol{t}) = F(\boldsymbol{x}, [\boldsymbol{e}, \boldsymbol{t}])$ , where $e = (e_0, e_1, \dots, e_L)$ is a sequence of learnable embeddings. PIZA-CoOp inserts $h$ as $G_{\otimes}(\boldsymbol{x}, \boldsymbol{t}) = F(\boldsymbol{x}, [\boldsymbol{e}, H(\boldsymbol{h}), \boldsymbol{t}])$ , where $H$ is a learnable linear layer as shown in Figure 6 (a). + +PIZA-LoRA. LoRA [23] is a low-rank adaptation method that injects trainable low-rank matrices into linear projections as $W\mathbf{x} + BA\mathbf{x}$ , where $\mathbf{x}$ is an input, $W$ is a frozen weight matrix and $A, B$ are learnable low-rank matrices. PIZA-LoRA integrates the embeddings $h$ into the bottleneck of LoRA as shown in Figure 6 (b), resulting in $W\mathbf{x} + BA\mathbf{x} + BCh$ , where $C$ is a newly added learnable matrix. We set the rank to 16. + +PIZA-Adapter+. Adapter+ [67] utilizes the post-adapter architecture, channel-wise scaling and Houlsby initialization [22]. PIZA-Adapter+ adds the embeddings $h$ to the output of the channel-wise scaling layer as shown in Fig- + +![](images/dde191bd6a98289a47e740af29bb37a5d7129e3c6a1dbab0c0682bbdf60900f5.jpg) +Figure 5.PIZA module. + +Figure 6. Parameter efficient fine-tuning with PIZA. +![](images/921c2fbdc85aebafb2e7747dc44705b9580305328c798bb42025499ef009caf7.jpg) +: Frozen modules and original inputs + +![](images/dfc45af6e7b9ddac165b0c4dcdea9e97e6afc721e97685b56d507c002399c141.jpg) +(b) PIZA-LoRA + +ure 6 (c). We set the bottleneck dimension to 256. + +# 4.4. Training + +Given a training dataset $\mathcal{D}$ consisting of images, expressions and ground truth bounding boxes, we construct an extended dataset $\mathcal{E}$ that involves ground truth search processes for training with PIZA. + +Overview. Each ground truth search process $P^{*}$ represents zooming steps to localize the target object. Specifically, it is given by $P^{*} = (b_{0}^{*}, b_{1}^{*}, \dots, b_{T^{*}}^{*})$ , where $b_{0}^{*} = (0, 0, W, H)$ indicates the entire image and $b_{T^{*}}^{*} = b^{*}$ is the ground truth bounding box. To facilitate efficient fine-tuning, we generate $P^{*}$ so that the distribution of inverse zoom factors $(z_{j}^{*})^{-1} = |b_{j}^{*}| / |b_{j-1}^{*}|$ match to the distribution of bounding box area ratios $p(r)$ in pre-training. + +Distribution $p(r)$ . Along with a pre-trained model, the distribution $p(r)$ is pre-estimated by applying kernel density estimation to the pre-training dataset (we use the union of O365 and GoldG in our experiments). We randomly sample 100,000 bounding boxes to compute area ratios $r = |b^{*}| / (WH)$ and apply a Gaussian kernel. + +Width and height for $b_{j}^{*}$ . The width and height for each bounding box in $P^{*}$ are randomly sampled in three steps. First, ratios $r_k$ are randomly sampled from $p(r)$ for $k = 1,2,\dots ,T_{\mathrm{max}}$ , where $T_{\mathrm{max}}$ is a sufficiently large constant. Second, the number of zooming steps $T^{*}$ is determined such that the cumulative product of $r_k$ is closest to the area ratio $r^{*}$ of the ground truth bounding box: + +$$ +T ^ {*} = \underset {T} {\operatorname {a r g m i n}} \left(\frac {1}{r ^ {*}} \prod_ {k = 1} ^ {T} r _ {k} - 1\right), r ^ {*} = \frac {| \boldsymbol {b} ^ {*} |}{W H}, \tag {3} +$$ + +where $\pmb{b}^{*}$ is a bounding box sampled from $\mathcal{D}$ , and $W, H$ are the width and height of the corresponding image. Then, the zoom factors $z_{j}^{*}$ and the size of bounding boxes $S_{j}^{*}$ are determined as follow: + +$$ +z _ {j} ^ {*} = \left(\frac {1}{r ^ {*}} \prod_ {k = 1} ^ {T ^ {*}} r _ {k} ^ {\omega_ {k} ^ {- 1}}\right) r _ {j} ^ {- 1}, S _ {j} ^ {*} = \left(\prod_ {k = 1} ^ {T ^ {*} - j} z _ {T ^ {*} - k}\right) | \boldsymbol {b} ^ {*} |, \tag {4} +$$ + +(c) PIZA-Adapter+ +![](images/0a7bc01a0691be29202f17d0a9c39eac99181cb0ef2622bb7dd17b1fe7a7d602.jpg) +: Learnable modules and embeddings : Modules and embeddings for PIZA + +where $\omega_{k} = \lambda_{1}e^{-\lambda_{2}k} / \sum_{k^{\prime} = 1}^{T^{*}}\lambda_{1}e^{-\lambda_{2}k^{\prime}}$ are weights derived from an exponential distribution. This weighting ensures that $z_{j}^{*}\simeq r_{j}^{-1}$ with smaller $j$ , encouraging more precise bounding box predictions at the initial zooming step. Finally, the width and height of $\pmb{b}_{j}^{*}$ are determined by $w_{j}^{*} = a_{j}\sqrt{S_{j}^{*}}$ , $h_j^* = a_j^{-1}\sqrt{S_j^*}$ where $a_{j}$ is an interpreted aspect ratio + +$$ +a _ {j} = \frac {W}{H} - \left(\frac {W}{H} - 1\right) \frac {W H - S _ {j} ^ {*}}{W H - S _ {T ^ {*} - 1} ^ {*}}. \tag {5} +$$ + +Centers. The center of each bounding box is aligned with the center coordinates of the target object. If the region extends beyond the image boundaries, the bounding box is minimally shifted to remain entirely within the image. + +Labels. Finally, for the generated search process $P^{*}$ , the sequence of binary labels $\mathbf{y}^{*} = (y_{0}^{*}, y_{1}^{*}, \dots, y_{T^{*}}^{*})$ and zooming step labels $z = (z_{0}^{*}, z_{1}^{*}, \dots, z_{T^{*}}^{*})$ are attached. Each label is given by + +$$ +y _ {j} ^ {*} = \left\{ \begin{array}{l l} [ \text {C O N T} ] & (0 \leq j < T ^ {*}) \\ [ \text {E O S} ] & (j = T ^ {*}) \end{array} , \quad z _ {j} ^ {*} = \frac {j}{T ^ {*}} \right. \tag {6} +$$ + +Loss function. The loss for fine-tuning is computed in three steps. First, a mini-batch of quadruplets $(\boldsymbol{x},\boldsymbol{t},P^{*},\boldsymbol{y}^{*})$ is drawn from $\mathcal{E}$ . Second, for each quadruplet, index $i\in \{1,2,\dots ,T^{*} - 1\}$ is randomly drawn to compute the forward process $\hat{\pmb{b}}_{i + 1} = F_{\oplus}(\pmb {x},\pmb {t},\pmb{b}_{0:i})$ . Finally, the loss depending on the pre-trained model is applied. In the experiments, we implement PIZA over the GroundingDINO model; thus the loss consists of the contrastive loss and the localization loss [42]. + +# 5. Experiments + +# 5.1. Experimental settings + +Dataset and metrics. The SOREC dataset is used for training and evaluation. We report the mean accuracy (mAcc) over IoU thresholds from 0.50 to 0.95 in increments of 0.05, as well as the accuracy at IoU of 0.50 $(\mathrm{Acc}_{50})$ and 0.75 $(\mathrm{Acc}_{75})$ . + +
Method#ParamsTrain-STrain-L
ValTest-ATest-BmAcc Acc50 Acc75mAcc Acc50 Acc75mAcc Acc50 Acc75ValTest-ATest-B
mAccAcc50Acc75mAccAcc50Acc75mAccAcc50Acc75mAccAcc50Acc75
Zero-shot00.20.60.00.31.00.10.00.20.00.20.60.00.31.00.10.0
Full fine-tuning173.0M29.551.730.235.958.638.823.043.621.937.463.739.243.869.648.030.5
CoOp0.1M20.236.120.424.240.125.815.529.614.622.641.622.027.546.528.717.5
PIZA-CoOp (Ours)0.9M26.339.129.729.441.234.221.933.824.329.844.133.533.446.838.724.4
LoRA1.3M21.638.521.826.243.128.117.032.515.925.244.525.330.750.233.019.7
PIZA-LoRA (Ours)1.5M30.944.734.933.846.639.225.838.728.934.549.939.139.354.045.529.0
Adapter+3.3M26.048.124.832.055.033.320.340.417.934.659.535.740.765.944.427.6
PIZA-Adapter+(Ours)3.5M36.853.541.843.159.650.130.445.934.139.060.642.945.166.251.731.7
+ +Pre-trained model. We selected GroundingDINO [42] using Swin-T [46] as a baseline model, and used the improved version provided as MM-GroundingDINO in the mmdetection library [87]. This model is pre-trained on the union of the following four datasets: O365 [63], GoldG [29], GRIT [56] and V3Det [68]. + +Baselines. We implemented three baselines for parameter-efficient fine-tuning: CoOp [89], LoRA [23] and Adapter+ [67]. PIZA is applied to each method as described in Section 4.3. Zero-shot baseline results are also reported. + +Implementation details. The AdamW optimizer is used for 5 epochs with a learning rate of $2 \times 10^{-4}$ , which is decayed by a factor of 0.5 at epoch 3. The hyperparameters for AdamW are set to their default values in PyTorch. The batch size is set to 16. LoRA is applied to each self-attention and cross-attention module. Adapter+ modules are inserted after each self-attention and feed-forward network module. Further details are provided in Appendix. + +# 5.2. Experimental results + +Main results. Table 3 summarizes the parameter-efficient fine-tuning results. As shown, PIZA significantly improved the performance for all methods. PIZA-Adapter+ achieved the best performance in terms of mAcc, surpassing the full fine-tuning baseline while reducing the number of learnable parameters from 173.0M to 3.5M. Prompt-tuning methods (CoOp and PIZA-CoOP) were more efficient but less effective than Adapter+ and LoRA methods. This is likely due to the low performance of the zero-shot baseline, which suggests that prompt-tuning alone may struggle to bridge the gap between the vision-language pre-training task and the REC task for small objects. When comparing Test-A and Test-B, all models exhibited higher accuracy on Test-A, which consists of traffic objects. This result is understandable, as objects such as traffic lights and road signs are designed with colors and shapes that make them easily to detect. When comparing training dataset sizes, the larger dataset (Train-L) consistently demonstrated higher performance, suggesting that further increasing the dataset size + +Table 3. Parameter-efficient fine-tuning results. #Params indicates the number of fine-tuned parameters. Best results are underlined. + +
Method#Prm.ValTest-ATest-B
PIZA-Adapter+w/oemb.insertion3.5M36.8/53.5/41.843.1/59.6/50.130.4/45.9/34.1
3.5M36.7/53.2/41.742.8/59.2/49.930.3/45.8/34.0
w/oPIZA module3.3M26.0/48.1/24.832.0/55.0/33.320.3/40.4/17.9
d=2563.5M36.8/53.5/41.843.1/59.6/50.130.4/45.9/34.1
d=1282.4M36.4/52.8/41.641.8/57.9/48.530.3/45.5/34.2
d=641.9M36.6/52.9/41.842.2/58.1/49.029.9/45.1/33.6
d=321.6M35.1/51.0/40.140.8/56.3/47.429.0/43.7/32.5
+ +Table 4. Ablation and hyperparameter studies for PIZA-Adapter+. "w/o embedding insertion" omits the connection colored in red in Figure 6(c). $d$ is the bottleneck dimension of the adapter. Each triplet of values indicates mAcc/Acc50/Acc75. Train-S is used for training. + +could be beneficial. + +Ablation and hyperparameter studies. Tables 4 and 5 show the results of ablation and hyperparameter studies for PIZA-Adapter+, PIZA-LoRA, and PIZA-CoOp. As shown, autoregressive prediction with the PIZA module is essential, and incorporating the zooming-step embedding further improved the performance. For PIZA-Adapter+, performance improves as the bottleneck dimensions increase. + +Necessity of pre-training. Table 6 compares training from scratch and full fine-tuning. Although the zero-shot baseline performance was low, the results confirmed that pretraining is necessary. + +Zooming steps. Table 7 presents the tradeoff between the number of zooming steps and performance by comparing our best results, which resulted in 2.11 steps on average, with those obtained by enforcing the number of steps $T^{*}$ to 1, 2 and 3 when creating the extended training dataset. As shown, our method performed the best among the tested configurations. Some qualitative examples are shown in Figure 7. As shown, for the SOREC dataset, 2 or 3 zooming steps were sufficient in most cases. + +Comparison with greedy approaches. To validate the necessity of the multi-step inference, we compare PIZA with sliding window and grid-like separation approaches. Figure 8 (a) compares our method with the eight fully + +Figure 7. Qualitative examples. +![](images/67a64d09a5c0497e335fe08dca9d4f14099f9df574e4d5982a22b85009ce774d.jpg) +$\hat{b}_{1}$ $\hat{b}_{2}$ $\hat{b}_{3}$ : Predicted bounding boxes : Ground truth bounding boxes + +
Method#Prm.ValTest-ATest-B
PIZA-LoRA1.5M30.9/44.7/34.933.8/46.6/39.225.8/38.7/28.9
w/oemb.insertion1.5M30.2/43.9/34.033.5/46.4/38.625.3/38.1/28.3
w/oPIZA module1.3M21.6/38.5/21.826.2/43.1/28.117.0/32.5/15.9
PIZA-CoOp0.9M26.3/39.1/29.729.4/41.2/34.221.9/33.8/24.3
w/oemb.insertion0.3M26.1/38.7/29.029.3/40.9/33.721.6/33.2/23.9
w/oPIZA module0.1M20.2/36.1/20.424.2/40.1/25.815.5/29.6/14.6
+ +Table 5. Ablation study for PIZA-LoRA and PIZA-CoOp (Train-S). Each triplet of values indicates mAcc/Acc $_{50}$ /Acc $_{75}$ . + +
Method#Prm.ValTest-ATest-B
Scratch173M0.00/0.00/0.000.00/0.01/0.000.00/0.00/0.00
Full fine-tuning173M29.5/51.7/30.235.9/58.6/38.823.0/43.6/21.9
+ +Table 6. Comparison with training from scratch (Train-S). + +
MethodStepsValTest-ATest-B
PIZA-Adapter+2.1136.8/53.5/41.843.1/59.6/50.130.4/45.9/34.1
w/ step enforcing1.025.8/47.5/25.531.7/54.2/33.320.0/39.3/18.1
w/ step enforcing2.036.1/53.1/40.841.7/57.8/48.629.6/45.6/32.9
w/ step enforcing3.034.3/50.3/39.039.8/55.3/46.327.1/41.8/30.1
+ +Table 7. Zooming step analysis (Train-S). + +fine-tuned sliding window baselines using combinations of four window sizes $W \in \{500, 750, 1000, 2000\}$ and two strides $S \in \{W, W/2\}$ . Apparently, the sliding window approach can improve the performance. However, all of these baselines indeed yield significantly lower performance and higher computational cost than our method. Although smaller windows and smaller strides capture finer details, they exponentially increase computational cost and often lead to false positive detections. Our method addresses these limitations, being $7.3 \times$ faster and using $49.4 \times$ fewer learnable parameters than the best sliding window baseline. Figure 8 (b) shows comparison with the tile-grid baselines. Our method outperformed them for the same reason discussed for the sliding window approach, suggesting that our method with average zooming step of 2.11 is reason- + +![](images/3f8167f042c7fa542718c63c166b3e2b58650792349ee06a526e4bc196e7026f.jpg) +Fig. 8. Comparison with sliding window and tile-grid baselines. + +able compared to these greedy grid-like separation counterparts. The typical window size used in the first step of our method was around $500 \times 500$ , while that for the tile grid approach on a typical high-resolution image in our dataset was $3407 \times 2470$ , resulting in a 36-fold increase in inference cost for the greedy tile grid approach. + +# 6. Conclusion + +We introduced the SOREC dataset, a new dataset for referring expression comprehension targeting small objects. Furthermore, we proposed PIZA tuning, a novel parameter-efficient fine-tuning approach that allows models to progressively zoom in and localize small objects based on natural language expressions. + +Future work and limitations. This work focused on localizing small objects in autonomous driving scenarios since detecting such objects is critical for ensuring safety and improves the overall reliability. Extending the proposed dataset to include more diverse environments and object types would remain an interesting future research direction. In addition, extending this work to video data and applying PIZA tuning to architectures for video processing would also be a promising next step. We believe that this work contributed to the computer vision community from both dataset and technical perspectives. + +Acknowledgment. This work was supported by JSPS KAKENHI 25K03135, 23H00490, JST CRONOS JPMJCS24K6, and JST BOOST Program JPMJBY24C5, and was carried out using the TSUBAME4.0 supercomputer at Institute of Science Tokyo. + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. European Conference on Computer Vision (ECCV), pages 213-229, 2020. 1 +[2] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. AdaptFormer: Adapting vision transformers for scalable visual recognition. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 25428-25440, 2022. 3 +[3] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. In Proc. International Conference on Learning Representations (ICLR), 2023. 3 +[4] Zhihong Chen, Ruifei Zhang, Yibing Song, Xiang Wan, and Guanbin Li. Advancing visual grounding with scene knowledge: Benchmark and method. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15039-15049, 2023. 2 +[5] Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, and Junwei Han. Towards large-scale small object detection: Survey and benchmarks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13467-13488, 2023. 2, 3 +[6] Eulrang Cho, Jooyeon Kim, and Hyunwoo J Kim. Distribution-aware prompt tuning for vision-language models. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 22004-22013, 2023. 3 +[7] Ming Dai, Lingfeng Yang, Yihao Xu, Zhenhua Feng, and Wankou Yang. SimVG: A Simple Framework for Visual Grounding with Decoupled Multi-modal Fusion. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), 2024. 2 +[8] Chao Deng, Qi Wu, Qingyao Wu, Fei Hu, Feng Lyu, and Mingkui Tan. Visual grounding via accumulated attention. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7746-7755, 2018. 1, 2 +[9] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor Guilherme Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Bayesian prompt learning for image-language model generalization. Proc. IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3 +[10] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), 2023. 3 +[11] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge J Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Object detection in aerial images: A large-scale benchmark and challenges. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(11):7778-7796, 2022. 2 +[12] Yang Ding, Jing Yu, Bang Liu, and Yue Hu. Revisiting counterfactual problems in referring expression comprehension. + +In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2 +[13] 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 ICLR, 2021. 3 +[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV), 88(2):303-338, 2010. 1, 2 +[15] Ross Girshick. Fast r-cnn. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 1440-1448. IEEE, 2015. 1 +[16] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 580-587, 2014. 1 +[17] Tianxiang Hao, Hui Chen, Yuchen Guo, and Guiguang Ding. Consolidator: Mergeable adapter with grouped connections for visual adaptation. In Proc. International Conference on Learning Representations (ICLR), 2023. 3 +[18] Haoyu He, Jianfei Cai, Jing Zhang, Dacheng Tao, and Bohan Zhuang. Sensitivity-aware visual parameter-efficient finetuning. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 11791-11801, 2023. 3 +[19] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross B. Girshick. Mask R-CNN. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 1 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 5 +[21] Xiuquan Hou, Meiqin Liu, Senlin Zhang, Ping Wei, and Badong Chen. Salience detr: Enhancing detection transformer with hierarchical salience filtering refinement. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17574-17583, 2024. 1 +[22] 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 Proc. International Conference on Machine Learning (ICML), pages 2790-2799, 2019. 3, 5 +[23] 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 Proc. International Conference on Learning Representations (ICLR), 2022. 3, 5, 7 +[24] Ronghang Hu, Huijuan Xu, Marcus Rohrbach, Jiashi Feng, Kate Saenko, and Trevor Darrell. Natural language object retrieval. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4555-4564, 2016. 1, 2 + +[25] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1115-1124, 2017. 1, 2 +[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proc. European Conference on Computer Vision (ECCV), 2022. 3 +[27] Zhengbao Jiang, Frank F Xu, Junichi Araki, and Graham Neubig. How can we know what language models know? In Proc. Annual Meeting of the Association for Computational Linguistics (ACL), pages 5191-5200, 2020. 2 +[28] Shibo Jie and Zhi-Hong Deng. FacT: Factor-tuning for lightweight adaptation on vision transformer. In Proc. AAAI Conference on Artificial Intelligence, pages 1060-1068, 2023. 3 +[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR-modulated detection for end-to-end multi-modal understanding. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 1780-1790, 2021. 1, 2, 7 +[30] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitagame: Referring to objects in photographs of natural scenes. In Proc. Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 787-798, 2014. 2 +[31] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19113-19122, 2023. 3 +[32] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 15190-15200, 2023. 3 +[33] Shuhei Kurita, Naoki Katsura, and Eri Onami. Refego: Referring expression comprehension dataset from first-person perception of ego4d. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 3421-3431, 2023. 2 +[34] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, Lei Zhang, and Jianfeng Gao. Semantic-SAM: Segment and recognize anything at any granularity. In Proc. European Conference on Computer Vision (ECCV), 2024. 3 +[35] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10965–10975, 2022. 2 +[36] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and + +Jianfeng Gao. Grounded language-image pre-training. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[37] Yang Li, Si Si, Gang Li, Cho-Jui Hsieh, and Samy Bengio. Learnable fourier features for multi-dimensional spatial positional encoding. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 15816-15829, 2021. 5 +[38] Yunhua Liao, Shuang Liu, Guanbin Li, Fei Wang, Yunchao Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 10877-10886, 2020. 1, 2 +[39] 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 Proc. European Conference on Computer Vision (ECCV), pages 740-755, 2014. 1, 2 +[40] Jiajun Liu, Liang Wang, and Ming-Hsuan Yang. Referring expression generation and comprehension via attributes. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 4856-4864, 2017. 1, 2 +[41] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic anchor boxes are better queries for DETR. In Proc. International Conference on Learning Representations (ICLR), 2022. 1 +[42] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In Proc. European Conference on Computer Vision (ECCV), 2024. 2, 4, 6, 7 +[43] Xiaohui Liu, Zhizhong Wang, Jing Shao, Xiaogang Wang, and Hongsheng Li. Improving referring expression grounding with cross-modal attention-guided erasing. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1950-1959, 2019. 1, 2 +[44] Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proc. Annual Meeting of the Association for Computational Linguistics (ACL), pages 61-68, 2022. 2 +[45] Yining Liu, Somak Aditya, and Yonatan Bisk Lee. Clevr-ref+: Diagnosing visual reasoning with referring expressions. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[46] 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 Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 7 +[47] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[48] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collabora + +tive network for joint referring expression comprehension and segmentation. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10034-10043, 2020. 1, 2 +[49] Renjie Luo and Greg Shakhnarovich. Comprehension-guided referring expressions. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7102-7111, 2017. 1, 2 +[50] Wenyu Lv, Shangliang Xu, Yian Zhao, Guanzhong Wang, Jinman Wei, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1 +[51] Tao Ma, Bing Bai, Haozhe Lin, Heyuan Wang, Yu Wang, Lin Luo, and Lu Fang. When visual grounding meets gigapixel-level large-scale scenes: Benchmark and approach. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22119-22128, 2024. 2 +[52] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L. Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11-20, 2016. 1, 2 +[53] Jiayuan Mao, Jing Huang, Alexander Toshev, Oana-Maria Camburu, Alan Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11-20, 2016. 1, 2 +[54] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Proc. European Conference on Computer Vision (ECCV), pages 792-807, 2016. 1, 2 +[55] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 5000-5009, 2017. 3 +[56] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 7 +[57] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 2641–2649, 2015. 1 +[58] 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 Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9982-9991, 2020. 2 +[59] 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 Proc. International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 3 +[60] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 91–99, 2015. 1 +[61] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 5 +[62] Arka Sadhu, Kan Chen, and Ram Nevatia. Zero-shot grounding of objects from natural language queries. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 4694-4703, 2019. 1, 2 +[63] 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 Proc. IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 7 +[64] Liu Shilong, Liang Yaoyuan, Huang Shijia, Li Feng, Zhang Hao, Su Hang, Zhu Jun, and Zhang Lei. DQ-DETR: Dual query detection transformer for phrase extraction and grounding. In Proc. AAAI Conference on Artificial Intelligence, 2023. 1 +[65] Taylor Shin, Yasaman Razeghi, Robert L. Logan, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In Proc. Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4222-4235, 2020. 2 +[66] Sijie Song, Xudong Lin, Jiaying Liu, Zongming Guo, and Shih-Fu Chang. Co-grounding networks with semantic attention for referring expression comprehension in videos. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1346-1355, 2021. 1, 2 +[67] Jan-Martin O. Steitz and Stefan Roth. Adapters strike back. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23449-23459, 2024. 3, 5, 7 +[68] Jiaqi Wang, Pan Zhang, Tao Chu, Yuhang Cao, Yujie Zhou, Tong Wu, Bin Wang, Conghui He, and Dahua Lin. V3det: Vast vocabulary visual detection dataset. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 19844-19854, 2023. 2, 7 +[69] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In Proc. International Conference on Machine Learning (ICML), pages 23318-23340, 2022. 2 +[70] Peng Wang, Shijie Wang, Junyang Lin, Shuai Bai, Xiaohuan Zhou, Jingren Zhou, Xinggang Wang, and Chang Zhou. One-peace: Exploring one general representation model toward unlimited modalities. arXiv preprint arXiv:2305.11172, 2023. 2 + +[71] Zhilin Xie, Kyungjae Lee, Howard Chen, Shuang Li, and Angel Chang. Refer360: Referring expression comprehension in $360^{\circ}$ images. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[72] Yuhui Xu, Lingxi Xie, Xiaotao Gu, Xin Chen, Heng Chang, Hengheng Zhang, Zhengsu Chen, Xiaopeng Zhang, and Qi Tian. Qa-lora: Quantization-aware low-rank adaptation of large language models. In Proc. International Conference on Learning Representations (ICLR), 2024. 3 +[73] Bin Yan, Yifan Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7042-7052, 2023. 2 +[74] Shuo Yang, Ping Luo, Chen Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5525-5533, 2016. 2 +[75] Zhongjie Yang, Boqing Gong, Liangliang Wang, Wei Huang, Dong Yu, and Jiebo Luo. A fast and accurate one-stage approach to visual grounding. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 4683-4693, 2019. 1, 2 +[76] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6757-6767, 2023. 3 +[77] Hantao Yao, Rui Zhang, and Changsheng Xu. Tcp: Textual-based class-aware prompt tuning for visual-language model. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[78] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Proc. European Conference on Computer Vision (ECCV), pages 69-85, 2016. 1, 2, 3 +[79] Licheng Yu, Hao Tan, Mohit Bansal, and Tamara L Berg. A joint speaker-listener-reinforcer model for referring expressions. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7282-7290, 2017. 1, 2 +[80] Licheng Yu, Zhe Lin, Xin Shen, Jimei Yang, Xiaohui Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1307-1315, 2018. 1, 2 +[81] Xuehui Yu, Yuxuan Gong, Nianlong Jiang, Qixiang Ye, and Zhenjun Han. Scale match for tiny person detection. In Proc. IEEE Conference on Winter Conference on Applications of Computer Vision (WACV), pages 1257-1265, 2020. 2 +[82] Chunyuan Zhang, Junjun Li, Shoufa Chen, et al. Detrs with collaborative hybrid assignments training. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1 +[83] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4158-4166, 2018. 1, 2 + +[84] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), 2022. 2 +[85] 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 Proc. International Conference on Learning Representations (ICLR), 2023. 1 +[86] Ji Zhang, Shihan Wu, Lianli Gao, Hengtao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[87] X. Zhang, Y. Li, G. Huang, Z. Wang, et al. An open and comprehensive pipeline for unified object grounding and detection. arXiv preprint arXiv:2401.02361, 2024. 2, 7 +[88] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16816-16825, 2022. 3 +[89] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 2, 5, 7 +[90] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pages 15659-15669, 2023. 3 +[91] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Haibin Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(11):7380-7399, 2021. 2 +[92] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: Deformable transformers for end-to-end object detection. In Proc. International Conference on Learning Representations (ICLR), 2020. 1 +[93] Z. Zhu, D. Liang, S. Zhang, X. Huang, B. Li, and S. Hu. Traffic-sign detection and classification in the wild. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2110-2118, 2016. 2 +[94] Bohan Zhuang, Qi Wu, Chunhua Shen, Ian Reid, and Anton van den Hengel. Parallel attention: A unified framework for visual object discovery through dialogs and queries. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4252-4261, 2018. 1, 2 \ No newline at end of file diff --git a/ICCV/2025/Referring Expression Comprehension for Small Objects/images.zip b/ICCV/2025/Referring Expression Comprehension for Small Objects/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..48ffae836da8a0dfe3b761a9c497e2db7593a893 --- /dev/null +++ b/ICCV/2025/Referring Expression Comprehension for Small Objects/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5885f3fd438de6d7122b859959df01a661beaa5a349e3be3f12e7518230de8a +size 911407 diff --git a/ICCV/2025/Referring Expression Comprehension for Small Objects/layout.json b/ICCV/2025/Referring Expression Comprehension for Small Objects/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4e52e17cb0ea89af6e2c0f2a614c27728c9b1375 --- /dev/null +++ b/ICCV/2025/Referring Expression Comprehension for Small Objects/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2185528ff65ee74022d045952c693118fe4f296c7add408b38591cdf4d8e6a7 +size 572926 diff --git a/ICCV/2025/Referring to Any Person/4946b02a-8894-40fb-97d2-3ddacd05e326_content_list.json b/ICCV/2025/Referring to Any Person/4946b02a-8894-40fb-97d2-3ddacd05e326_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..315864d8b293928fa7fbbca4f6addddeb0b86077 --- /dev/null +++ b/ICCV/2025/Referring to Any Person/4946b02a-8894-40fb-97d2-3ddacd05e326_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7186d3e060c824f3e5aabe4c901398fa50ffe2d01e53386f14bb0b005361c19 +size 93810 diff --git a/ICCV/2025/Referring to Any Person/4946b02a-8894-40fb-97d2-3ddacd05e326_model.json b/ICCV/2025/Referring to Any Person/4946b02a-8894-40fb-97d2-3ddacd05e326_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cc3f9c3bd0b8c68f6099ea3d3669d605aa28b716 --- /dev/null +++ b/ICCV/2025/Referring to Any Person/4946b02a-8894-40fb-97d2-3ddacd05e326_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6745c716d5f460ae8ad4f698fddfe740c07c32678e778d4a29839dea6e7201e +size 120604 diff --git a/ICCV/2025/Referring to Any Person/4946b02a-8894-40fb-97d2-3ddacd05e326_origin.pdf b/ICCV/2025/Referring to Any Person/4946b02a-8894-40fb-97d2-3ddacd05e326_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..571a75b5b940a38aaa62f5dce5b5558acc413d10 --- /dev/null +++ b/ICCV/2025/Referring to Any Person/4946b02a-8894-40fb-97d2-3ddacd05e326_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58797269a312f7628ee50e4406360ef55f2d68c014996dce9bf7fa37bb63d03f +size 5805740 diff --git a/ICCV/2025/Referring to Any Person/full.md b/ICCV/2025/Referring to Any Person/full.md new file mode 100644 index 0000000000000000000000000000000000000000..faadffc52db31345bafa26c23b9184e23008d980 --- /dev/null +++ b/ICCV/2025/Referring to Any Person/full.md @@ -0,0 +1,374 @@ +# Referring to Any Person + +Qing Jiang $^{1,2}$ , Lin Wu $^{1,2}$ , Zhaoyang Zeng $^{2}$ , Tianhe Ren $^{2}$ , Yuda Xiong $^{2}$ , Yihao Chen $^{2}$ , Liu Qin $^{2}$ , Lei Zhang $^{1,2\dagger}$ + +$^{1}$ South China University of Technology + +$^{2}$ International Digital Economy Academy (IDEA) + +mountchicken@outlook.com，leizhang@idea.edu.cn + +![](images/d98bb73b106b86a4b007c3cd50ee67a284a4b8cc3805fa438b79220366e02a4c.jpg) +Figure 1. We introduce referring to any person, a task that requires detecting all individuals in an image which match a given natural language description, and a new model RexSeek designed for this task with strong perception and understanding capabilities that effectively captures attributes, spatial relations, interactions, reasoning, celebrity recognition, etc. + +# Abstract + +Humans are undoubtedly the most important participants in computer vision, and the ability to detect any individual given a natural language description, a task we define as referring to any person, holds substantial practical value. However, we find that existing models generally fail to achieve real-world usability, and current benchmarks are limited by their focus on one-to-one referring, which hin + +ders progress in this area. In this work, we revisit this task from three critical perspectives: task definition, dataset design, and model architecture. We first identify five aspects of referable entities and three distinctive characteristics of this task. Next, we introduce HumanRef, a novel dataset designed to tackle these challenges and better reflect real-world applications. From a model design perspective, we integrate a multimodal large language model with an object detection framework, constructing a robust referring model named RexSeek. Experimental results reveal that state-of-the-art models, which perform well on commonly used benchmarks like RefCOCO+/g, struggle with HumanRef due to their inability to detect multiple individuals. In con + +trast, RexSeek not only excels in human referring but also generalizes effectively to common object referring, making it broadly applicable across various perception tasks. Code is available at https://github.com/IDEA-Research/RexSeek. + +# 1. Introduction + +Humans are central to computer vision [4, 10-12, 16, 21, 27, 28, 30, 45, 46, 63, 65, 84, 85], and the ability to identify and detect specific individuals based on natural language descriptions, a task we define as referring to any person, is crucial for numerous applications, including human-robot interaction, industrial automation, healthcare, etc. + +However, we argue that progress in this area has been hindered by unclear task definitions and a lack of high-quality data. Our findings show that despite achieving state-of-the-art performance on referring benchmarks RefCO $\mathrm{CO} / + / \mathrm{g}$ [50, 75], most models remain impractical for real-world applications, as illustrated in Figure 2. To address this challenge, we revisit this task from three perspectives: task definition, dataset construction, and model design. + +We begin by formally defining the task of referring to any person: given a natural language description and an input image, the model needs to detect all individuals in the image who match the description. To comprehensively capture the scope of this task, we identify five key aspects that define how humans can be referred to: i) Attributes: Encompassing intrinsic characteristics such as gender, age, action, clothing, etc. ii) Position: Describing spatial relationships both among individuals and between individuals and their surroundings. iii) Interaction: Accounting for human-to-human, human-to-object, and human-to-environment interactions. iv) Reasoning: Involving multi-step inference that considers multiple objects to resolve complex expressions. v) Celebrity Recognition: Identifying specific individuals, whether by their real names or characters names. + +Next, we identify three crucial characteristics that define this task: i) Multi-Instance Referring: A referring expression can correspond to multiple individuals. While mainstream referring datasets $\mathrm{RefCOCO} / + / \mathrm{g}$ [50, 75] typically assume that each expression refers to a single object, this does not align with real-world scenarios. We find through experiments that most models experience significant performance degradation when tasked with identifying more than one individual. ii) Multi-Instance Discrimination: The image should contain multiple individuals in addition to the target person. This setting ensures that the model fully comprehends the referring expression to identify the correct individual rather than simply detecting all people in the image. iii) Rejection of Non-existence: If the referred person is not present in the image, the model should refuse to generate a result rather than produce a hallucinated output. + +![](images/e5e5b193a8db93125f5704bfa440c3677d5b4047bc6c9adc0f43dd59f312822d.jpg) +Figure 2. Visualization results of Qwen2.5-VL [3], InternVL2.5 [14], and DeepSeek-VL2 [70] on the human referring task. Despite achieving strong results on referring benchmarks RefCO:CO/+/g [50, 75], state-of-the-art models struggle when tasked with identifying multiple individuals as they output an insufficient number of bounding boxes. + +Based on the task definition, we manually constructed a novel dataset for human referring, named HumanRef. Unlike the traditional ReferItGame [29] annotation approach, where one annotator describes an object and another finds it based on the description, we adopt a different annotation methodology. Our process begins with annotators listing the key properties of individuals in an image according to the predefined referable entities. Next, for each person, they determine whether these properties apply and result in a property dictionary. Finally, a large language model [71] composes these properties into referring expressions. HumanRef comprises 103,028 referring statements, with each expression referring to an average of 2.2 instances. We also split a benchmark from HumanRef with 6,000 referring expressions spanning six subsets, ensuring comprehensive coverage across all referable properties. + +From the model design perspective, we argue that a robust referring model should possess two key characteristics: i) Robust Perception Ability: The model should be capable of detecting all individuals in an image. ii) Strong Language Comprehension: The model should effectively interpret complex language descriptions of people. To address these requirements, we introduce RexSeek, a detection-oriented multimodal large language model specifically designed for this task. Inspired by ChatRex [25], we formulate referring as a retrieval-based task. RexSeek integrates a person detector [60] as its box input, ensuring strong perception capabilities while incorporating Qwen2.5 [71] as the LLM to enhance language comprehension. We adopt a multi-stage training approach that progressively refines both detection and comprehension skills, equipping RexSeek with strong referring capabilities. + +Experimental results indicate that most state-of-the-art models [3, 9, 14, 25, 49, 60, 70, 74] exhibit performance + +degradation on the HumanRef benchmark, despite achieving strong results on RefCOCO/+/g. The primary limitation is that these models typically detect only a single instance, as they are trained on datasets that assume one-to-one referring. In contrast, RexSeek, trained on HumanRef, exhibits strong referring capabilities. Additionally, benefiting from the multi-stage training approach, RexSeek also emerges with the ability to refer to generalized objects, extending its applicability beyond human-centric tasks. To summarize, our contributions are threefold: + +- We introduce referring to any person with a clear definition by identifying five aspects of referable entities and three key characteristics that distinguish this task. +- We introduce HumanRef, a novel referring dataset, and establish a challenging benchmark to drive progress in human-centric referring expression research. +- We propose RexSeek, a detection-oriented multimodal large language model trained through a multi-stage process, demonstrating strong referring capabilities for both humans and general objects. + +# 2. Related Work + +Referring Expression Comprehension Task. Referring Expression Comprehension (REC) [29, 36, 48, 50, 56, 72, 75, 76, 76, 81] involves interpreting a natural language expression to localize specific objects within an image. Unlike open-vocabulary object detection [15, 26, 34, 42, 52, 60, 61, 69, 73, 78] or phrase grounding [18, 23, 31, 54, 68], which identify objects based on brief category names or short phrases, REC requires understanding complex, freeform descriptions. This task necessitates not only recognizing object attributes and relationships but also comprehending spatial configurations and interactions, making it inherently more challenging. In this work, we systematically analyze the referable entities and the critical characteristics that define this task. + +REC Datasets and Benchmarks. The first large-scale Referring Expression Comprehension (REC) dataset, ReferItGame [29], was created through a two-player game in which one annotator describes an object, and another selects it. This was later followed by more sophisticated datasets [7, 13, 17, 19, 54, 55], such as RefCOCO [75], RefCOCO+ [75], and RefCOCOg [50], which leverage MSCOCO [37] images to provide more complex referring expressions. Beyond these general datasets, others address specific challenges. CLEVR-Ref+ [40] focuses on geometric object referring. RefCrowd [57] targets person detection in crowded scenes. Ref-L4 [8] handles longer and more detailed descriptions. GRES [77] introduces multi-target referring expression segmentation. However, existing datasets typically assume a one-to-one correspondence between a referring expression and a single instance, which fails to reflect real-world scenarios. To address this gap, we + +
domainsub-domainsexamples
attributegender, age, race, profession, posture, appearance, clothing and accessories, actionmale, female, white man, the police officer, person with a shocked expression, person wearing a mask, person standing
positioninner position (human to human), outer position (human to environment)the second person from left to right, person at the right, person closest to the microphone, person sitting in the chair
interactioninner interaction (human with human), outer interaction (human with environment)two people holding hands, people locked in each other's gaze, the person holding a gun, person holding the certificate in hand
reasoninginner position reasoning, outer position reasoning, attribute reasoningall the people to the right of the person closest to the glass, person wearing a lab coat but not putting their hand on the board
celebrity recognitionactor, character, athlete, entrepreneur, scientist, politician, singerBrad Pitt, Bruce Wayne, Cristiano Ronaldo, Rihanna, Elon Musk, Albert Einstein, Donald Trump
rejectionattribute, position, interaction, reasoninga man in red hat, three women in a circle
+ +Table 1. The primary annotation domains and their corresponding sub-domains within HumanRef. + +refine the referring task and introduce HumanRef, a dataset specifically designed to support multi-instance referring and advance research in this domain. + +MLLM-based REC Methods Multimodal Large Language Models (MLLMs) [1-3, 14, 20, 32, 33, 35, 44, 47, 53, 64, 66, 70, 82] have demonstrated strong capabilities in both text and image comprehension, motivating efforts to integrate referring expression understanding into these models. A common approach involves outputting bounding box coordinates as tokens [3, 9, 14, 51, 67, 70, 74, 79, 80, 83]. Alternatively, methods like Groma [49] and ChatRex [25] frame detection as a retrieval task, where a proposal model generates bounding boxes, and the LLM selects the index of the relevant box based on the referring expression. While these MLLM-based methods achieve high performance on RefCOCO $+ / \mathrm{g}$ , our experiments reveal that they remain inadequate for practical applications due to low recall rate on multi-instance referrals. + +# 3. HumanRef Dataset + +In this section, we present the design philosophy, data acquisition process, annotation pipeline, and dataset statistics of the proposed HumanRef dataset. + +# 3.1. Data Design Philosophy + +We define five key aspects that determine how humans can be referred to using natural language, including attribute, position, interaction, reasoning, and celebrity recognition. These categories are further elaborated with definitions and examples in Table 1. A key distinction between Human-Ref and existing referring datasets is its focus on multi-instance referring rather than one-to-one object referring. Our dataset ensures that a single referring expression can correspond to multiple individuals, providing a more realistic and practical reflection of real-world scenarios. + +# 3.2. Data Acquisition + +The HumanRef dataset is designed to capture human presence across diverse contexts, including natural environments, industrial settings, healthcare, sports, films, animations, etc. To ensure dataset diversity, we sourced images + +![](images/96cf5226a5a601fb6558074359715a2d097448b7db2428af73470a31df1dbd08.jpg) +Figure 3. Overview of the manual annotation pipeline of the HumanRef dataset. + +containing humans from the web image dataset [5]. To filter candidate images, we first retained those with a resolution larger than $1000 \times 1000$ pixels to ensure high-quality content. Next, we use an open-set object detector DINO-X [60] to detect human instances. To align with the multi-instance discrimination requirement, we retain only images containing at least four individuals. + +To assist the annotator in writing properties, we prompt the QwenVL-2.5 [3] model to create a structured property dictionary for each person in the image, capturing details such as gender, clothing, actions, etc. Ultimately, this phase produced image, person box, and person description triples used for further annotation. + +# 3.3. Manual Annotation + +For attribute, position, interaction, and reasoning subsets, we adopt manual annotation. This annotation process consists of three main steps: property listing, property assignment, and referring style rewriting. Given an image, along with the corresponding person boxes and pre-labeled property dictionary, the annotation system will randomly select one annotation type from attribute, location, interaction, and reasoning to assign to the annotator. The following annotation process is then carried out: + +Property Listing: The annotator examines all individuals in the image, considering both their visual appearance, action, position, interaction, and the pre-labeled property dictionary. Based on these observations, the annotator compiles a list of properties. To enhance dataset richness, annotators are encouraged to label attributes shared by multiple individuals while avoiding those common to all. Additionally, we monitor the word frequency of labeled referring expressions and restrict the use of high-frequency words to improve data diversity. + +Property Assignment: Once the properties are listed, annotators systematically assign them to the correspond + +ing individuals. This interactive process involves selecting a property value and clicking on the associated bounding boxes to link it to the correct person. The final output is a structured dictionary, where keys represent property names and values contain lists of bounding box indices corresponding to the individuals possessing each property. + +Referring Style Rewriting: In the final step, we prompt Qwen2.5-14B [71] to reformulate the structured attribute dictionary into short, natural language referring expressions. The final annotated data also undergoes a thorough review process to ensure its quality. + +# 3.4. Automatic Annotation + +For celebrity recognition and rejection referring, we employ two efficient and effective automatic annotation pipelines. + +Celebrity Recognition: We first categorize celebrities into seven distinct fields: actors, film characters, athletes, singers, entrepreneurs, scientists, and politicians. For each field, we identify the most well-known individuals, compiling a final list of 636 names, which we then used as prompts to retrieve images via the Bing Search API. The collected images include both individual and group photos, necessitating a method to accurately associate each celebrity name with the correct person in the image. To achieve this, we first use the DINO-X [60] model to detect all human faces and persons, linking each detected face to its corresponding person box based on overlap measurements. If an image contains only one person, we assume this individual is the target celebrity. For images featuring multiple individuals, we use a Python face recognition library, leveraging a single-person image as a recognition template to match and identify the same person in such images. + +Rejection Referring: The objective of this sub-dataset is to ensure that when a referring description targets a per + +![](images/141e25d62316f3c74ba471e4dad0e28f9d2b6afc6e6508735f3898a59a25e0f5.jpg) +Attribute + +![](images/ceb467b07a81566fcfa7f0cd6fd38c831321e92f86efed6df8f0e0eb15df6281.jpg) +Position + +![](images/3e8edcd53ea7956cb80c4c857fa5aca183c44707b71340245105a1572c0855d1.jpg) + +![](images/74aaf90a4370908c9b09e67a9f3226d2d40f18efbe8921be6f915fc6ab112fae.jpg) + +![](images/cf220ac7f4b72fe96fac07069f9a64904aa7bafa3e560c34b47d5029237ed767.jpg) +Interaction + +![](images/6508079a6ed91f816fc34052b23ff06d09310f7a05a3db80b7506f03abee1865.jpg) +Reasoning + +![](images/0b9d4feecef584b74afcf7f34e8cfa9085b7a3c0a1a13a5cebaada4cb2aebcb6.jpg) +Celebrity Recognition +Figure 4. Visualization of the six subsets in the HumanRef Benchmark. + +son who does not exist in the input image, the model rejects the referring request instead of hallucinating and outputting an incorrect bounding box. To construct this dataset, we first extract referring expressions from the attribute, position, interaction, and reasoning subsets. We then prompt Qwen2.5 [71] to modify these descriptions, transforming them into similar but semantically altered versions. For instance, a description such as "the person wearing a blue hat" may be changed to "the person wearing a red hat". To validate the generated descriptions, we prompt Molmo [20] to detect the modified referring expression. If no matching object is found in the output, the data is retained. + +# 3.5. HumanRef Benchmark + +To construct the HumanRef Benchmark, we sample 1,000 referring expressions from each of the four manually annotated subsets. Additionally, for the celebrity and rejection subsets, we conduct a separate manual annotation process to create 1,000 new referring expressions for each category, ensuring high-quality and challenging evaluation data. To further support advancements in referring expression segmentation, we utilize SAM2 [59] to generate masks for each ground truth bounding box. Figure 4 presents example cases from the HumanRef Benchmark, illustrating the diversity and complexity of the dataset. + +# 3.6. Statistics + +We first present the basic statistics of the HumanRef dataset and its subsets in Table 2, and then illustrate the characteristics of multi-instance referring and multi-instance discrimination in HumanRef in Figure 5. Additionally, Table 3 compares the HumanRef Benchmark with widely used referring benchmarks, including RefCOCO, RefCOCO+, and RefCOCOg. A key distinction of HumanRef is its higher + +
HumanRef Train
typeattributepositioninteractionreasoningcelebrityrejectiontotal
images8,6147,5771,6324,4744,9907,51934,806
referrals52,51322,4962,9116,8084,99013,310103,028
avg. boxes/ref2.91.93.13.01.002.2
HumanRef Benchmark
typeattributepositioninteractionreasoningcelebrityrejectiontotal
images8389729409821,0001,0005,732
referrals1,0001,0001,0001,0001,0001,0006,000
avg. boxes/ref2.82.12.12.71.102.2
+ +Table 2. Main statistics of the HumanRef dataset, including the number of images, the number of referring expressions, the average word count per referring expression, and the average number of instances associated with each referring expression. + +
Datasetsimagesrefsvocabsavg. sizeavg. person/imageavg. words/refavg. boxes/ref
RefCOCO [75]1,51910,7711,874593x4845.723.431
RefCOCO+ [75]1,51910,9082,288592x4845.723.341
RefCOCOg [50]1,5215,2532,479585x4802.739.071
HumanRef5,7326,0002,7141432x10748.606.692.2
+ +Table 3. Comparison of the HumanRef Benchmark with RefCO-CO/+/g. For a fair comparison, we present only the statistics related to human referring in RefCOCO/+/g. + +![](images/0e1b2969a283afbaaa0da9e45d0cb81f8588826cac2c9174b63280cd0be8ace3.jpg) +Figure 5. Distribution of the number of individuals per image and the number of individuals referenced by each referring expression. + +image resolution and larger number of individuals per image, requiring models to precisely identify all correct individuals among multiple people. Unlike traditional benchmarks, where each referring expression corresponds to a + +![](images/1c70c44963c3b6ad9ae654145dc39074dada2f3e5e807a39705c1575dcde3a60.jpg) +Figure 6. Overview of the RexSeek model. RexSeek is a retrieval-based model built upon ChatRex [25]. By integrating a person detection model, RexSeek transforms the referring task from predicting box coordinates to retrieving the index of input boxes. + +single person, HumanRef supports multi-instance referring, offering a more realistic and challenging evaluation setting for referring expression comprehension. + +# 4. RexSeek Model + +# 4.1. Model Design Philosophy + +From a model design perspective, we argue that a robust referring model should have two essential capabilities: i) robust perception ability, where the model can reliably detect all individuals in an image, and ii) strong language comprehension, where the model can accurately interpret complex natural language descriptions of people. + +For the first capability, modern object detection models [26, 41, 60, 61] are highly effective at identifying people within images. However, these models often lack the necessary language comprehension abilities to process intricate and nuanced referring expressions. On the other hand, while MLLMs are proficient in understanding natural language, they often struggle with fine-grained object detection tasks. Inspired by ChatRex [25], we propose a hybrid framework, RexSeek, which integrates the strengths of both object detection models and LLMs. RexSeek combines a high-performance detection model with a multimodal LLM to achieve both accurate detection and effective language understanding. + +# 4.2. Architecture + +Following ChatRex, we formulate the referring task as a retrieval-based process [25, 49]. As illustrated in Figure 6, RexSeek consists of three main components: vision encoders, a person detector, and a large language model. Given an input image, we first pass it through a dual vision encoder module used in ChatRex. This module consists of a CLIP [58] to extract low-resolution image features $\mathcal{F}_{\mathrm{low}}$ and a ConvNeXt [43] to extract high-resolution image features $\mathcal{F}_{\mathrm{high}}$ . We adjust the input resolutions for both vision en + +coders to ensure they generate the same number of tokens at the last scale. The final vision tokens $\mathcal{F}$ is obtained by concatenating these features at the channel dimension: + +$$ +\mathcal {F} = \operatorname {C o n c a t} \left(\mathcal {F} _ {\text {l o w}}, \mathcal {F} _ {\text {h i g h}}\right) +$$ + +Next, we prompt DINO-X [60] to get the bounding boxes of persons $\{B_i\}_{i=1}^K$ in the image. For each bounding box, we extract its RoI features $\mathcal{C}_i$ and add their positional embeddings to generate object tokens $\mathcal{O}_i$ , which capture both the content and spatial context of each detected person: + +$$ +\mathcal {O} _ {i} = \mathcal {C} _ {i} + \operatorname {P E} (B _ {i}) +$$ + +Specifically, the RoI feature is extracted from the high-resolution vision features using a multi-scale RoI Align operation [24]. The positional embedding is computed by encoding the bounding box coordinates $(x,y,w,h)$ using a sinusoidal encoding function and concatenating the encoded values along the channel dimension. + +Finally, the vision tokens $\mathcal{F}$ , object tokens $\mathcal{O}$ , and text tokens $\mathcal{T}$ are projected using different MLPs and then fed into the LLM. By default, we use Qwen2.5 [71] as the LLM. The LLM decodes the input to produce the corresponding box indices $\mathcal{I}$ : + +$$ +\mathcal {I} = \operatorname {L L M} (\mathcal {F}, \mathcal {O}, \mathcal {T}) +$$ + +The output $\mathcal{I}$ consists of object indices that correspond to the bounding boxes of the target persons corresponding to the referring. This sequence is structured as follows: + +$$ +< g > \text {r e f e r r i n g} < / g > < o > < \text {o b j m} > \dots < \text {o b j n} > < / o > +$$ + +Here, $\langle \text{objm} \rangle$ and $\langle \text{objn} \rangle$ refer to specific object index tokens that correspond to the detected persons. The special tokens $\langle g \rangle$ , $\langle /g \rangle$ , $\langle o \rangle$ , and $\langle /o \rangle$ are used to format the output, linking the referring expression with the relevant object indices. + +# 4.3. Four Stage Training + +Similar to other VLMs, we adopt a pretraining followed by supervised fine-tuning approach [39]. Our training process consists of four stages. In the first stage, we align the visual and textual modalities using image-captioning data. In the second stage, we focus on perception training with detection-oriented data, enabling the model to retrieve relevant objects from input bounding boxes. In the third stage, we incorporate multimodal data to enhance the model's general understanding abilities. Finally, in the fourth stage, we fine-tune the model using the HumanRef dataset, resulting in the final RexSeek model. The data, task, and trainable modules for each stage are shown in Table 5. + +# 5. Experiments + +In this section, we first introduce the evaluation metrics used in our study and assess the performance of multimodal models on HumanRef. We perform a comprehensive analysis to + +
MethodAttributePositionInteractionReasoningCelebrityAverageRejection
RPDF1RPDF1RPDF1RPDF1RPDF1RPDF1Score
Baseline†100.037.224.2100.028.515.9100.032.519.4100.042.630.3100.014.44.9100.031.018.90.0
DINOX [60]59.528.820.978.828.117.667.328.518.976.232.122.294.148.037.075.233.123.336.0
InternVL-2.5-8B [14]23.539.027.123.028.024.327.840.131.317.522.818.957.459.358.029.837.831.954.9
Ferret-7B [74]27.944.430.430.236.229.830.841.831.219.733.722.863.260.057.534.443.234.32.0
Groma-7B [49]67.547.838.663.243.137.266.648.140.659.141.434.873.263.359.165.948.742.10.0
ChatRex-7B [25]44.378.051.848.066.752.549.674.856.536.665.142.873.776.574.250.472.255.60.0
Qwen2.5-VL-7B [3]49.171.354.450.261.752.848.266.353.234.661.240.380.381.980.152.568.556.27.1
DeepSeek-VL2-small [70]52.378.057.756.466.158.155.475.760.746.661.750.185.974.370.759.371.259.53.1
Molmo-7B-D* [20]82.786.476.378.080.672.469.977.766.172.180.465.585.987.582.977.782.572.668.6
RexSeek-7B87.286.881.586.186.383.884.884.680.787.884.781.583.486.584.285.985.882.354.1
+ +Table 4. Benchmarking multimodal models on HumanRef Benchmark. R, P, and DF1 represent Recall, Precision, and DensityF1, respectively. † A simple baseline that uses the bounding boxes of all persons in the image as results, simulating a person detection model that does not follow the referring description. *Molmo-7B-D predicts point coordinates as output and use point-in-mask evaluation criteria. + +
StageTrainable ModulesTask# SamplesDatasets
Stage1MLPsImage Captioning976KALLAVA-4V-Caption [6]
Stage2MLPs + LLM + Vision EncodersGrounding & Region Understanding2.07MCOCO [37], LVIS [22], O365 [62], Rexverse-2M [25]
Stage3MLPs + LLM + Vision EncodersGeneral Knowledge & Grounding & Region Understanding2.15MLLAVA-665K [38] Rexverse-2M [25]
Stage4MLPs + LLM + Vision EncodersReferring103KHumanRef
+ +Table 5. Data, task, and trainable modules for each stage. + +explore the challenges faced by existing models in handling the referring task. Additionally, we perform ablation experiments on RexSeek for model design choices. + +# 5.1. Metrics + +We evaluate the referring task using Precision, Recall, and DensityF1 Score. Given a referring expression, the model predicts one or more bounding boxes, and a prediction is considered correct if its IoU with any ground truth box exceeds a predefined threshold. Following the evaluation protocol in COCO [37], we report the average performance across IoU thresholds from 0.5 to 0.95 in increments of 0.05. For models that only output points, such as Molmo [20], a prediction is considered correct if the predicted point falls within the mask of the corresponding instance. However, this evaluation is less strict than the IoU-based metric, as point-in-mask criteria impose looser spatial constraints, making direct comparisons less fair. For the rejection subset, we calculate the number of referring expressions that the model does not predict any boxes and divide it by the number of total expressions. + +To penalize models that indiscriminately detect all persons in an image to achieve a high F1 score through high recall, we introduce the DensityF1 Score, which modifies the standard F1 Score with a density-aware penalty: + +$$ +\text {D e n s i t y F 1} = \frac {1}{N} \sum_ {i = 1} ^ {N} 2 \times \frac {\text {P r e c i s i o n} _ {i} \times \text {R e c a l l} _ {i}}{\text {P r e c i s i o n} _ {i} + \text {R e c a l l} _ {i}} \times D _ {i} \tag {1} +$$ + +![](images/e76448e6bb7f0f7930e59218aa9a531b5a1dde2f74d048afcf3071a82aa4bb7b.jpg) +Figure 7. Visualizing the trend of recall and precision variations across different models as the number of instances corresponding to each referring expression increases. + +![](images/7e1ccedf8849aa36895eaecf8f53514bc69284ee5a4e6a8e4cddb0c9cd7c8287.jpg) + +where $D_{i}$ is the density penalty factor, defined as: + +$$ +D _ {i} = \min \left(1. 0, \frac {\mathrm {G T C o u n t} _ {i}}{\text {P r e d i c t e d C o u n t} _ {i}}\right) \tag {2} +$$ + +Here, GT Count is the total number of persons in an image, and Predicted Count is the number of predicted boxes for a given referring expression. This penalty discourages over-detection by reducing the score when the predicted box count significantly exceeds the ground truth count. + +# 5.2. Benchmarking on HumanRef + +In Table 4, we evaluate the performance of various multimodal models on the HumanRef benchmark. While these models perform well on the widely used RefCOCO, RefCOCO+, and RefCOCOg benchmarks, their performance significantly degrades on HumanRef. Our analysis reveals two common issues among these models: + +Low Recall for Multi Instance: We observe a common issue among most models: when a referring expression corresponds to multiple instances, recall drops significantly, as shown in Figure 7. This suggests that when multiple objects need to be detected, most models tend to predict only a few bounding boxes, limiting their applicability in real-world scenarios. A key factor contributing to this behavior is the nature of the training data. Most multimodal models are trained on RefCOCO, RefCOCO+, and RefCOCOg, + +
ModelWith Rejection DataRejection Score
RexSeek-7BNo0
RexSeek-7BYes541
+ +Table 6. Rejection score comparison under different model scales with and without rejection data during training. + +
Loading StageHumanRef Average
RPDF1
stage173.973.568.2
stage277.077.372.2
stage377.978.073.0
+ +Table 7. Ablation experiments on multi-stage training by loading models from different training stages and fine-tuning them on the HumanRef dataset. We Qwen2.5-3B as the base LLM. + +
MethodRefCOCOg
valtest
Shikra-7B [9]82.382.2
InternVL2-8B [14]82.782.7
Grounding DINO-L [42]86.187.0
Qwen2.5-VL-7B [3]87.287.2
MM1.5-7B [82]-87.1
ChatRex-7B [25]88.888.6
RexSeek-7B84.084.4
+ +Table 8. Zero-shot evaluation of RexSeek on RefCOCOg. We use the open-set detector DINOX to detect the subject object in the image and use the detected bounding box as input to RexSeek. + +where referring expressions rarely correspond to multiple instances. As a result, these models become biased toward single-instance predictions. In contrast, RexSeek has been trained on datasets that explicitly include multi-instance referring expressions, demonstrate a significantly improved ability to handle these real-world cases. + +Hallucination Issue: On the rejection subset, we observe that most models perform poorly with low rejection score. This indicates that regardless of whether the referred object is actually present in the image, these models tend to predict a bounding box, exhibiting a severe hallucination issue. In real-world referring applications, such as referring in video streams, it is crucial for models to accurately determine whether the specified object exists in the image. Additionally, we find that the rejection capability can be significantly improved by incorporating appropriate training data. As shown in Table 6, when trained without the rejection data in HumanRef, RexSeek also demonstrates strong hallucination tendencies. This highlights the critical role of dataset design in the referring task, as inadequate dataset construction can lead to overconfident predictions. + +# 5.3. Ablations on RexSeek + +Ablation of Multi-stage Training: We analyzed the impact of the four-stage training approach used in RexSeek. As shown in Table 7, we conducted supervised fine-tuning on the HumanRef dataset after each training stage. The re + +![](images/053bedc7c6fd8f82bf101c581412a026d2513d61bf791d010350a26e7026abff.jpg) + +![](images/c2da178494ccc6014d6e80c836e68c2dae0ab5170eff06a9cd2de97b8bf42cbe.jpg) +unripe tomatoes + +![](images/57c26ac50a80f8ac505cb599bf5916d9d1ebb332da89a26d508f2646173aa577.jpg) + +![](images/aedd3564962d6a236954d30a716fbb8641f913886bd35427666119f1846c0235.jpg) +Figure 8. RexSeek can refer to arbitrary objects beyond person. + +![](images/3781c911788fc91b29ecf89f3e48a1c73e313e3ae2375cbaedc0b732bacd8536.jpg) + +![](images/b55d37431f878003f25dae012ce78849f287f5d093cd0800623357f100458805.jpg) + +sults demonstrate that the model achieves its best performance after undergoing SFT with general multimodal data (LLaVA-665K [38]). We attribute this improvement to the model acquiring richer general knowledge from multimodal data, which enhances its ability to accurately refer to persons in complex scenarios. + +Generalization to Any Object Referring: Although RexSeek is trained exclusively on human-related referring data, we find that it also demonstrates the ability to refer to arbitrary objects. We first evaluate the performance of RexSeek on RefCOCOg. Given a referring expressions, we apply DINO-X to detect the main object in the image, using the detected bounding box as input to RexSeek. As shown in Table 8, RexSeek achieves competitive performance on RefCOCO/+/g, despite not being explicitly trained on general object referring. Additionally, Figure 8 presents visualizations illustrating that RexSeek can also detect multiple instances even for non-human objects. We attribute this generalization ability to our multi-stage training approach, where perception and multimodal understanding training develop object comprehension, and fine-tuning on Human-Ref effectively extends it to arbitrary objects. + +# 6. Conclusion + +In this work, we identify the fundamental limitations of existing referring datasets and models, demonstrating that they fail to meet real-world application demands, particularly in multi-instance referring. To address this, we introduce HumanRef, a large-scale benchmark reflecting real-world complexity, and propose RexSeek, a retrieval-based detection MLLM integrating person detection with a language model. Our multi-stage training approach equips RexSeek with strong generalization capabilities, allowing it to excel in human-centric referring while extending effectively to arbitrary object referring. Extensive evaluations highlight the struggles of state-of-the-art models with multi-instance detection and hallucination, underscoring the importance of dataset design and training strategies for more reliable and generalizable referring expression models. + +# References + +[1] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 3 +[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. NeurIPS, 35: 23716-23736, 2022. +[3] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 3, 4, 7, 8 +[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 561-578. Springer, 2016. 2 +[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.4 +[6] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model. arXiv preprint arXiv:2402.11684, 2024. 7 +[7] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. Touchdown: Natural language navigation and spatial reasoning in visual street environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12538-12547, 2019. 3 +[8] Jierun Chen, Fangyun Wei, Jinjing Zhao, Sizhe Song, Bohuai Wu, Zhuoxuan Peng, S-H Gary Chan, and Hongyang Zhang. Revisiting referring expression comprehension evaluation in the era of large multimodal models. arXiv preprint arXiv:2406.16866, 2024.3 +[9] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3, 8 +[10] Ling-Hao Chen, Jiawei Zhang, Yewen Li, Yiren Pang, Xiaobo Xia, and Tongliang Liu. Humanmac: Masked motion completion for human motion prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9544-9555, 2023. 2 +[11] Weihua Chen, Xiaotang Chen, Jianguo Zhang, and Kaiqi Huang. Beyond triplet loss: a deep quadruplet network for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 403–412, 2017. +[12] Weihua Chen, Xianzhe Xu, Jian Jia, Hao Luo, Yaohua Wang, Fan Wang, Rong Jin, and Xiuyu Sun. Beyond appear + +ance: a semantic controllable self-supervised learning framework for human-centric visual tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15050-15061, 2023. 2 +[13] Zhenfang Chen, Peng Wang, Lin Ma, Kwan-Yee K Wong, and Qi Wu. Cops-ref: A new dataset and task on compositional referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10086-10095, 2020. 3 +[14] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhang-wei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 2, 3, 7, 8 +[15] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xing-gang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 3 +[16] Yuanzheng Ci, Yizhou Wang, Meilin Chen, Shixiang Tang, Lei Bai, Feng Zhu, Rui Zhao, Fengwei Yu, Donglian Qi, and Wanli Ouyang. Unihcp: A unified model for human-centric perceptions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17840-17852, 2023. 2 +[17] Volkan Cirik, Taylor Berg-Kirkpatrick, and Louis-Philippe Morency. Refer360: A referring expression recognition dataset in 360 images. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7189–7202, 2020. 3 +[18] Samyak Datta, Karan Sikka, Anirban Roy, Karuna Ahuja, Devi Parikh, and Ajay Divakaran. Align2ground: Weakly supervised phrase grounding guided by image-caption alignment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2601-2610, 2019. 3 +[19] Harm De Vries, Florian Strub, Sarath Chandar, Olivier Pietquin, Hugo Larochelle, and Aaron Courville. Guesswhat?! visual object discovery through multi-modal dialogue. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5503-5512, 2017. 3 +[20] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 3, 5, 7 +[21] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10884-10894, 2019. 2 +[22] 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. 7 +[23] Tanmay Gupta, Arash Vahdat, Gal Chechik, Xiaodong Yang, Jan Kautz, and Derek Hoiem. Contrastive learning for + +weakly supervised phrase grounding. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 3 +[24] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6 +[25] Qing Jiang, Yuqin Yang, Yuda Xiong, Yihao Chen, Zhaoyang Zeng, Tianhe Ren, Lei Zhang, et al. Chatrex: Taming multimodal llm for joint perception and understanding. arXiv preprint arXiv:2411.18363, 2024. 2, 3, 6, 7, 8 +[26] Qing Jiang, Feng Li, Zhaoyang Zeng, Tianhe Ren, Shilong Liu, and Lei Zhang. T-rex2: Towards generic object detection via text-visual prompt synergy. In European Conference on Computer Vision, pages 38-57. Springer, 2025. 3, 6 +[27] Xuan Ju, Ailing Zeng, Jianan Wang, Qiang Xu, and Lei Zhang. Human-art: A versatile human-centric dataset bridging natural and artificial scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 618-629, 2023. 2 +[28] Xuan Ju, Ailing Zeng, Chenchen Zhao, Jianan Wang, Lei Zhang, and Qiang Xu. Humansd: A native skeleton-guided diffusion model for human image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15988-15998, 2023. 2 +[29] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 2, 3 +[30] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2024. 2 +[31] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 3 +[32] 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. arXiv preprint arXiv:2410.05993, 2024. 3 +[33] 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. 3 +[34] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 3 +[35] Zhiqi Li, Guo Chen, Shilong Liu, Shihao Wang, Vibashan VS, Yishen Ji, Shiyi Lan, Hao Zhang, Yilin Zhao, Subhashree Radhakrishnan, et al. Eagle 2: Building posttraining data strategies from scratch for frontier vision + +language models. arXiv preprint arXiv:2501.14818, 2025.3 +[36] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10880-10889, 2020. 3 +[37] 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 ECCV, pages 740-755, 2014. 3, 7 +[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv: 2310.03744, 2023. 7, 8 +[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 6 +[40] Runtao Liu, Chenxi Liu, Yutong Bai, and Alan L Yuille. Clevr-ref+: Diagnosing visual reasoning with referring expressions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4185-4194, 2019. 3 +[41] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6 +[42] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 3, 8 +[43] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 6 +[44] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, et al. Nvila: Efficient frontier visual language models. arXiv preprint arXiv:2412.04468, 2024.3 +[45] 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. 2 +[46] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978, 2023. 2 +[47] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 3 +[48] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020. 3 + +[49] Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiaojuan Qi. Groma: Localized visual tokenization for grounding multimodal large language models. arXiv preprint arXiv:2404.13013, 2024. 2, 3, 6, 7 +[50] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L. Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016. 2, 3, 5 +[51] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, Anton Belyi, Haotian Zhang, Karanjeet Singh, Doug Kang, Ankur Jain, Hongyu He, Max Schwarzer, Tom Gunter, Xiang Kong, Aonan Zhang, Jianyu Wang, Chong Wang, Nan Du, Tao Lei, Sam Wiseman, Guoli Yin, Mark Lee, Zirui Wang, Ruoming Pang, Peter Grasch, Alexander Toshev, and Yinfei Yang. MM1: methods, analysis & insights from multimodal LLM pretraining. arXiv: 2403.09611, 2024. 3 +[52] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European conference on computer vision, pages 728-755. Springer, 2022. 3 +[53] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.3 +[54] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 3 +[55] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 3 +[56] Yanyuan Qiao, Chaorui Deng, and Qi Wu. Referring expression comprehension: A survey of methods and datasets. IEEE Transactions on Multimedia, 23:4426-4440, 2020. 3 +[57] Heqian Qiu, Hongliang Li, Taijin Zhao, Lanxiao Wang, Qingbo Wu, and Fanman Meng. Refcrowd: Grounding the target in crowd with referring expressions. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4435-4444, 2022. 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 ICML, pages 8748-8763, 2021. 6 +[59] 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. 5 + +[60] Tianhe Ren, Yihao Chen, Qing Jiang, Zhaoyang Zeng, Yuda Xiong, Wenlong Liu, Zhengyu Ma, Junyi Shen, Yuan Gao, Xiaoke Jiang, et al. Dino-x: A unified vision model for openworld object detection and understanding. arXiv preprint arXiv:2411.14347, 2024. 2, 3, 4, 6, 7 +[61] Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, et al. Grounding dino 1.5: Advance the" edge" of open-set object detection. arXiv preprint arXiv:2405.10300, 2024. 3, 6 +[62] 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 Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 7 +[63] Shixiang Tang, Cheng Chen, Qingsong Xie, Meilin Chen, Yizhou Wang, Yuanzheng Ci, Lei Bai, Feng Zhu, Haiyang Yang, Li Yi, et al. Humanbench: Towards general human-centric perception with projector assisted pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21970-21982, 2023. 2 +[64] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2024. 3 +[65] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1653-1660, 2014. 2 +[66] 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. 3 +[67] 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. 3 +[68] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 3 +[69] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15254-15264, 2023. 3 +[70] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 2, 3, 7 +[71] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, + +Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2, 4, 5, 6 +[72] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4644-4653, 2019. 3 +[73] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 3 +[74] 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. 2, 3, 7 +[75] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling context in referring expressions. In ECCV, pages 69-85, 2016. 2, 3, 5 +[76] 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 Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 3 +[77] Zhihan Yu and Ruifan Li. Revisiting counterfactual problems in referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13438-13448, 2024. 3 +[78] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14393-14402, 2021. 3 +[79] Yufei Zhan, Yousong Zhu, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon v2: Advancing multimodal perception with high-resolution scaling and visual-language co-referring. arXiv preprint arXiv:2403.09333, 2024. 3 +[80] Yufei Zhan, Yousong Zhu, Zhiyang Chen, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon: Spelling out all object locations at any granularity with large language models. In European Conference on Computer Vision, pages 405-422. Springer, 2025. 3 +[81] Chao Zhang, Weiming Li, Wanli Ouyang, Qiang Wang, Woo-Shik Kim, and Sunghoon Hong. Referring expression comprehension with semantic visual relationship and word mapping. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1258-1266, 2019. 3 +[82] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, et al. Mm1. 5: Methods, analysis & insights from multimodal llm fine-tuning. arXiv preprint arXiv:2409.20566, 2024. 3, 8 +[83] Haotian Zhang, Haoxuan You, Philipp Duffer, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 3 + +[84] Liang Zheng, Hengheng Zhang, Shaoyan Sun, Manmohan Chandraker, Yi Yang, and Qi Tian. Person re-identification in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1367–1376, 2017. 2 +[85] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. Deephuman: 3d human reconstruction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7739-7749, 2019. 2 \ No newline at end of file diff --git a/ICCV/2025/Referring to Any Person/images.zip b/ICCV/2025/Referring to Any Person/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c8759ad60861726c8e92039d7fb4fe94cb0b4743 --- /dev/null +++ b/ICCV/2025/Referring to Any Person/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:282f1f7e53487dfc31bb2ab1464c3e42314ff54ce68ae7c729fe6f082453e906 +size 1011073 diff --git a/ICCV/2025/Referring to Any Person/layout.json b/ICCV/2025/Referring to Any Person/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f44801ec6c14c30d7895139d5d0ab59421196c27 --- /dev/null +++ b/ICCV/2025/Referring to Any Person/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efa59d530ef3ec6bc0234deceefaeb013ae598664211f0a478cb4b97a1074d86 +size 424204 diff --git a/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/9bbce2c6-ee4f-46f6-a472-68b3931775ff_content_list.json b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/9bbce2c6-ee4f-46f6-a472-68b3931775ff_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3fbeec9ec8da17b23c7e2d90d590664d5508ba87 --- /dev/null +++ b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/9bbce2c6-ee4f-46f6-a472-68b3931775ff_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56ca2251940ecf5c952f79dda457d55f9191e838342a4eb4e29895ccd784abe5 +size 78580 diff --git a/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/9bbce2c6-ee4f-46f6-a472-68b3931775ff_model.json b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/9bbce2c6-ee4f-46f6-a472-68b3931775ff_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ac73b9860f945e98aed504c4a5c544d23fa36bfc --- /dev/null +++ b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/9bbce2c6-ee4f-46f6-a472-68b3931775ff_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20cc3066071503a4332a687594b8f8433eb1e6ab8290c61c4a4e62cd6c893315 +size 101799 diff --git a/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/9bbce2c6-ee4f-46f6-a472-68b3931775ff_origin.pdf b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/9bbce2c6-ee4f-46f6-a472-68b3931775ff_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..36ac7cab445aeaa249c9ae2636a7dfd35f5a5ee1 --- /dev/null +++ b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/9bbce2c6-ee4f-46f6-a472-68b3931775ff_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fd5a200b2307855ade177217f78f5757aec19a1bdc6739a014d818f4cf514e6 +size 6518640 diff --git a/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/full.md b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..95739e04dbbc8a36581943c494d2897f7d57a470 --- /dev/null +++ b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/full.md @@ -0,0 +1,309 @@ +# Reflect-DiT: Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection + +Shufan Li\*1, Konstantinos Kallidromitis\*2, Akash Gokul\*3, Arsh Koneru\*1 Yusuke Kato2, Kazuki Kozuka2, Aditya Grover1 1UCLA 2Panasonic AI Research 3Salesforce AI Research + +*Equal Contribution + +Correspondence to jacklishufan@cs.ucla.edu + +# Abstract + +The predominant approach to advancing text-to-image generation has been training-time scaling, where larger models are trained on more data using greater computational resources. While effective, this approach is computationally expensive, leading to growing interest in inference-time scaling to improve performance. Currently, inference-time scaling for text-to-image diffusion models is largely limited to best-of- $N$ sampling, where multiple images are generated per prompt and a selection model chooses the best output. Inspired by the recent success of reasoning models like DeepSeek-R1 in the language domain, we introduce an alternative to naive best-of- $N$ sampling by equipping text-to-image Diffusion Transformers with in-context reflection capabilities. We propose Reflect-DiT, a method that enables Diffusion Transformers to refine their generations using in-context examples of previously generated images alongside textual feedback describing necessary improvements. Instead of passively relying on random sampling and hoping for a better result in a future generation, Reflect-DiT explicitly tailors its generations to address specific aspects requiring enhancement. Experimental results demonstrate that Reflect-DiT improves performance on the GenEval benchmark $(+0.19)$ using SANA-1.0-1.6B as a base model. Additionally, it achieves a new state-of-the-art score of 0.81 on GenEval while generating only 20 samples per prompt, surpassing the previous best score of 0.80, which was obtained using a significantly larger model (SANA-1.5-4.8B) with 2048 samples under the best-of- $N$ approach. $^{1}$ + +# 1. Introduction + +Text-to-image diffusion models have made significant progress by training larger architectures with more data + +![](images/cf6e012da6179686aacdd06dc5c715e24d54130ddca0d12935b745cef7813221.jpg) +Figure 1. Reflect-DiT iteratively refines image generation by using a vision-language model (VLM) to critique generations and a Diffusion Transformer (DiT) to self-improve using past generations and feedback. Specifically, at each generation step N, feedback from previous iterations (N-3, N-2, N-1, ...) are incorporated to progressively improve future generations. Unlike traditional best-of-N sampling, Reflect-DiT actively corrects errors in object count, position, and attributes, enabling more precise generations with fewer samples. + +[39, 42, 43, 60]. However, scaling the training of such models, i.e. using more data and/or larger models, is computationally expensive, as exponential increases in compute are required to achieve near-linear performance gains [28]. Recently, works studying Large Language Models (LLMs), which exhibit similar scaling laws, have explored inference-time scaling as another means of improving model performance [5, 48, 59, 69]. These approaches introduce additional compute resources during inference, either by generating many samples and employing a selection mechanism to find the best output [29] or by prompting the LLM + +to generate longer reasoning traces, often involving self-verification and self-correction, before reaching an answer [7, 57]. These methods have shown promising results, offering substantial performance gains with a relatively moderate increase in compute compared to training-time scaling. + +More recently, several works have attempted to apply the concept of inference-time scaling to text-to-image diffusion models [34, 60]. They primarily focus on two key aspects: scaling denoising steps per sample and scaling the number of samples. For the latter, a commonly used framework is to generate N random samples per prompt and select the best result using a reward or judge model (best-of-N). In particular, SANA-1.5 [61] achieved state-of-the-art performance on the GenEval benchmark [14] by generating a very large number of samples per prompt $(N = 2048)$ using a 4.8B Diffusion Transformer (DiT). Its performance surpassed previous results achieved by significantly larger models, highlighting the benefits of inference-time scaling. + +Despite their encouraging success, these methods still have considerable room for improvement in terms of efficiency. For example, generating 2048 samples per prompt is impractical for real-world applications. In this work, we propose Reflect-DiT, an effective framework to improve the inference-time scaling of Diffusion Transformers by equipping DiTs with the ability to refine future generations through reflecting upon its past generated images and natural language feedback. Reflect-DiT draws inspiration from the recent success of reasoning models, such as DeepSeek-R1 [15], which exhibit self-verification and reflection capabilities. Rather than relying on random sampling to produce a better output in the next generation, these models utilize their long context windows to reason about a given problem. This process allows the model to search, reflect on, and refine potential solutions before providing a final output. Unlike autoregressive LLMs, text-to-image diffusion models currently lack the ability to reason about past generations and feedback, as they condition solely on the input prompt. We argue that this limitation—specifically, the inability to reference and learn from past generations—prevents them from achieving inference-time scaling benefits beyond naive best-of-N sampling strategies. To address this shortcoming, Reflect-DiT incorporates in-context reflection, expanding the conditioning signals to include past image generations and natural language feedback (Fig. 1). Reflect-DiT can evaluate its past generations, identify misalignments with the input prompt (e.g. object count, spatial positioning, etc.) and refine subsequent generations to correct these issues. + +Concretely, Reflect-DiT consists of (1) a Vision-Language Model (VLM) that serves as a judge that evaluates generated images with respect to input prompts and provides natural language feedback, and (2) a Diffusion Transformer that refines its generations based on previous + +generations and corresponding feedback. Previously generated images and text feedback are first encoded with vision and text encoders into modality-specific embedding spaces, then processed by a lightweight Context Transformer to obtain a set of conditional embeddings that are passed to the cross-attention layers of the DiT. To ensure scalability, Reflect-DiT maintains a fixed context length, limiting the number of past generations it considers. When the total number of past generations exceeds the context limit, we employ a selection mechanism to stochastically sample a subset of past generations as the context. The Reflect-DiT framework is illustrated in Figure 1. + +We conduct extensive experiments to test the effectiveness of Reflect-DiT. Compared to the naive best-of-N approach, Reflect-DiT improves performance by $+0.19$ on the GenEval benchmark using SANA-1.0-1.6B as the base model. Reflect-DiT also establishes a new absolute state-of-the-art score of 0.81 on GenEval while generating only 20 samples per prompt during inference, surpassing the previous best score of 0.80 that was obtained using a significantly larger model (SANA-1.5-4.8B) with 2048 samples under the naive best-of-N approach. These results highlight Reflect-DiT as a more effective and efficient alternative to best-of-N sampling for inference-time scaling of DiTs. + +# 2. Related Works + +# 2.1. Inference-time scaling of LLMs + +Traditionally, stronger performance on language tasks was achieved by training larger language models on more data [6, 38]. However, recent works have explored leveraging additional compute at inference time to improve performance [9, 23, 36, 41, 56, 63]. Brown et al. [5] demonstrated the benefits of combining repeated sampling and a selection mechanism, such as automatic verifiers or reward models. Beyond parallel sampling, works have also explored enabling self-corrective behavior [7, 23, 35, 36, 41, 56]. Compared to related works, such as Welleck et al. [56] and Madaan et al. [36], which explore self-correction at inference time via feedback for language tasks, Reflect-DiT focuses on enabling self-correction capabilities within text-to-image models. More recently, Snell et al. [48] discovered that adaptive search and iterative self-refinement are more effective than naive Best-of-N sampling given a fixed compute budget. Wu et al. [59] showed that with proper inference strategies, scaling test-time compute can be more efficient than scaling model parameters. Building on these insights, Reflect-DiT explores inference-time scaling beyond naive Best-of-N sampling by enabling text-to-image models to iteratively refine their generations by reflecting on past generations and feedback. + +![](images/0534c4895cddb26656a0b74dab12365c0efb3aa970326328f204af526b2fedfc.jpg) +Figure 2. Architecture of Reflect-DiT. Given a prompt, past images and feedback, we first encode the images into a set of vision embeddings $[V_1, V_2, \ldots]$ using a vision encoder, and encode text feedback to a set of text embeddings $[E_1, E_2 \ldots]$ . We then concatenate these embeddings into a single sequence $M$ , and pass it through the Context Transformer to obtain $M'$ . The extra context $M'$ is concatenated directly after the standard prompt embeddings and passed into the cross-attention layers of the Diffusion Transformer (DiT). + +# 2.2. Inference-time scaling of Text-to-Image Diffusion Models + +Inspired by the success of inference-time scaling of LLMs, recent works have explored similar strategies for text-to-image diffusion models. Ma et al. [34] explored several search strategies and concluded that random search combined with Best-of-N selection is currently the best strategy for improving performance. SANA-1.5 [61] applied Best-of-N sampling to a frontier text-to-image model and established a new state of the art on the GenEval benchmark through Best-of-N sampling (N=2048). Concurrent to this work, Singhal et al. [47] proposed a more structured search method that extends beyond random sampling. + +# 2.3. Self-Correction for Text-to-Image Generation + +Before the recent advancements in inference-time scaling for text-to-image models, earlier works explored using LLM agents to enable self-verification and self-improvement (SLD[58] and GenArtist[53]). These methods use LLMs or VLMs as controllers to generate a series of image operations or function calls to specialized models [4, 22, 31, 66]. Unlike Reflect-DiT, these works do not enable text-to-image diffusion models to learn from natural feedback; instead, they rely on a predefined and limited set of operations. (e.g. object manipulation in latent space [59]). Furthermore, the performance of SLD and GenArtist is highly dependent on the successful execution of each + +submodule/operation, many of which involve heuristics or are not up-to-date. Because of such limitations, these approaches are not ideal for inference time scaling of frontier text-to-image models. We consider these works to be tangentially related to Reflect-DiT rather than directly comparable. Further discussion can be found in the Appendix A.5. + +# 2.4. Controllable Text-to-Image Generation + +Text-to-image generation is inherently constrained by language as it relies solely on the input prompt. Thus, limiting its ability to represent concepts that may be difficult to capture in words. In the case of subject-driven generation, i.e. generating an image of a specific subject such as one's pet, conditioning solely on text has shown to be ineffective. To address this, subject-driven generation methods introduce learnable embeddings [12, 24, 45, 51], or provide visual conditioning signals [26, 33, 40, 55]. Furthermore, previous works [1, 8, 20, 27, 37, 44, 46, 62, 68, 71] have introduced mechanisms to expand the conditioning signal of text-to-image diffusion models to include a variety of signals to enable control beyond text. Unlike Reflect-DiT, these works focus on zero-shot controllable image generation and do not use conditioning signals to allow the model to learn from iterative feedback. Prior works have also explored using natural language instructions to improve image editing [4, 11, 13, 21, 70]. We provide further comparisons with these works in Section A.4. + +# 3. Method + +# 3.1. Overall Framework + +Similar to its counterparts in the language domain, Reflect-DiT iteratively refines its generation by performing a verification-reflection loop. Reflect-DiT consists of a vision-language model (VLM) feedback judge $F_{j}$ that generates text feedback $T$ for an input image $X$ , and a Diffusion Transformer (DiT) text-to-image generator $F_{g}$ that maps a text prompt $P$ and a set of past generations and feedback $C = \{(X_{i}, T_{i}) \mid i = 1, 2, \ldots\}$ to a new output image $X_{j}$ . Given an input prompt $P$ , we first generate an image $X_{0} = F_{g}(P)$ without any additional context, and obtain the initial feedback $T_{0} = F_{j}(P, X_{0})$ . At each subsequent iteration $i$ , we obtain a new generation $X_{i}$ and its corresponding feedback $T_{i}$ . The reflection context is then updated, as $C_{i} = \{(X_{j}, T_{j}) \mid j = 1, 2, \ldots, i\}$ , to include all past generations and feedback. If the size of $C_{i}$ is larger than a predefined max context length $K$ , we randomly sample $K$ past generations and their corresponding feedback as the input context $C_{k}$ . Otherwise, we use all pairs in $C_{i}$ to form input context $C_{k}$ . The generator then produces an updated image $X_{i+1} = F_{g}(P, C_{k})$ , incorporating past feedback. This loop continues until the VLM feedback judge produces null feedback (indicating no further improvements) or a maximum number of iterations $N$ is reached. The full procedure is formally described in Algorithm 1. + +# 3.2. VLM Feedback + +The goal of the VLM judge, $F_{j}$ , is to provide natural language feedback for generated images. We use Qwen2.5-VL 3B[2] as the judge model and finetune it following the setup of SANA 1.5 [61]. Specifically, the training data is curated by generating a large number of images from synthetic prompts and using an object detector to judge whether the desired objects are present in the image and whether their counts and attributes agree with the prompt. Feedback data used in VLM training are generated using structured templates. For example, if an object X is missing, we use the template "There is no {X} in the image." If the count of objects is incorrect, we use the template "There should be {N} {X} in the image, but only {K} exist." The feedback is intentionally concise to improve VLM inference efficiency and minimize memory overhead for the DiT. We provide additional details of VLM training in Appendix C.1. + +# 3.3. Diffusion Transformer + +We implement Reflect-DiT using SANA-1.0-1.6B [60], which offers competitive performance while being $106 \times$ faster than open-source alternatives[25, 39]. This improvement in efficiency is due to its smaller size and adoption of a linear attention mechanism, making it an ideal choice for inference-time scaling which requires generating a large + +Algorithm 1 Iterative Image Refinement with Verification-Reflection Loop +Require: Text prompt $P$ , Feedback Judge VLM $F_{j}$ , DiT Image Generator $F_{g}$ , Max Context Length $K$ , Max Iterations $N$ +1: Initialize $X_{0} \gets F_{g}(P)$ ▷ Generate initial image +2: $T_{0} \gets F_{j}(P, X_{0})$ ▷ Obtain initial feedback +3: $C_{0} \gets \{(X_{0}, T_{0})\}$ ▷ Initialize reflection context +4: for $i = 1$ to $N$ do +5: if $T_{i-1} = \emptyset$ then ▷ Stop if no more improvements +6: break +7: end if +8: Construct $C_{i} = \{(X_{j}, T_{j}) | j = 1, 2, \dots, i\}$ +9: if $|C_{i}| > K$ then +10: Sample $K$ elements from $C_{i}$ to obtain $C_{k}$ +11: else +12: $C_{k} = C_{i}$ +13: end if +14: Generate new image $X_{i} \gets F_{g}(P, C_{k})$ +15: Obtain feedback $T_{i} \gets F_{j}(P, X_{i})$ +16: Update reflection context $C_{i} \gets C_{i} \cup \{(X_{i}, T_{i})\}$ +17: end for +18: return Image Trajectory $\{X_{0}, X_{1}, \dots, X_{n}\}$ + +number of samples per prompt. SANA consists of consecutive Linear-DiT blocks, each containing a self-attention layer, a cross-attention layer, and a feed forward network (FFN). We incorporate past generated images and text feedback as additional context for the cross-attention layer. + +Concretely, given prompt $P$ and past context $C = \{(X_1, T_1), (X_2, T_2), \ldots\}$ , we first encode the images $[X_1, X_2, \ldots]$ into a sequence of vision embeddings $[V_1, V_2, \ldots]$ using a vision encoder. We use SigLIP-Large [67] as our vision encoder, which encodes each image to a feature map of size $24 \times 24$ . We downsample this feature map to $8 \times 8$ before flattening it into a 1D sequence, reducing the sequence length of image embeddings from 576 to 64. This compression minimizes the additional memory needed, making it computationally feasible to fit multiple images in the context during training and inference. To encode text feedback, we use Gemma-2-2B [49] and convert the text feedback $[T_1, T_2 \dots]$ to a set of 1D embeddings $[E_1, E_2 \dots]$ . Finally, we concatenate these image and text embeddings into a single sequence $M = \text{Concat}([V_1, E_1, V_2, E_2, \dots])$ . Since each image and text feedback are encoded separately, these features are not aligned with the features of the base DiT and there is no mechanism to associate each image with its corresponding feedback. Hence, we first process $M$ through a small two-layer Transformer [50], which we refer to as the Context Transformer, to obtain intermediate output $M'$ . The extra context $M'$ is concatenated directly after the standard + +
GeneratorParamsOverallSingleTwoCountingColorPositionAttribution
SDXL[39]2.6B0.550.980.740.390.850.150.23
DALLE 3[3]-0.670.960.870.470.830.430.45
SD3[10]8B0.740.990.940.720.890.330.60
Flux-Dev[25]12B0.680.990.850.740.790.210.48
Playground v3[30]-0.760.990.950.720.820.500.54
SANA-1.5-4.8B[61]‡4.8B0.760.990.950.720.820.500.54
+ Best-of-2048‡4.8B0.800.990.880.770.900.470.74
SANA-1.0-1.6B† [60]1.6B0.620.980.830.580.860.190.37
+ Best-of-201.6B0.750.990.870.730.880.540.55
+ Reflect-DiT(N=20)1.6B + 0.1B0.810.980.960.800.880.660.60
(Δ vs Baseline)-+0.19+0.00+0.13+0.22+0.02+0.47+0.23
Image-Gen-CoT[16]1.3B0.770.990.860.670.840.660.64
+ +Table 1. Results on the GenEval benchmark [14]. Reflect-DiT achieves the highest overall score (0.81) with only 20 samples per prompt, outperforming all other models despite having significantly fewer parameters. Compared to the base SANA-1.0-1.6B, Reflect-DiT demonstrates consistent improvements across all evaluation categories, with a notable overall gain of +0.19. While SANA-1.5-4.8B achieves competitive performance, it requires substantially more computational resources and is not open-sourced at the time of writing. † Evaluated using the released checkpoint of SANA-1.0. ‡ SANA-1.5 updated the results after this paper was written. (March 23, 2025, arXiv v3) We cite results of their Arxiv v2. + +prompt embeddings of SANA-1.0 and passed into the crossattention layers of the DiT. This is illustrated in Figure 2. + +The same synthetic dataset is used to train both the VLM judge and the Diffusion Transformer. Specifically, each training data sample used for training the DiT consists of a "good" image $X_w$ as the positive sample, and a set of "bad" images $[X_1, X_2, \ldots]$ and corresponding feedback $[T_1, T_2, \ldots]$ as reflection context. The VLM training data consists of multiple image-feedback pairs per prompt. We select images that pass the object detector test (and hence use the feedback template "This image is correct") as positive samples and the remaining images as in-context feedback samples. Prompts that lead to only "good" images or "bad" images are excluded from the training data. Following SANA, we finetune the DiT using the flow-matching objective as follows: + +$$ +\mathbb {E} _ {t \in \operatorname {U n i f} [ 0, 1 ]} w (t) \| \epsilon - x _ {w} - F _ {g} \left(x _ {w} ^ {t}, t, C\right) \| ^ {2} \tag {1} +$$ + +where $\epsilon$ is an i.i.d Gaussian noise, $x_w$ is the image latent corresponding to the "good" image $X_w$ and $x_w^t = (1 - t)x_w + t\epsilon$ is a noised image latent, $C$ is the set of in-context image-feedback pairs, and $w(t)$ is a weighting function. We use a logit-normal weighting scheme following SANA. + +# 4. Experiments + +# 4.1. Setup + +# 4.1.1. Dataset + +We generate prompts using the GenEval templates [14] and filter out those present in the test set, resulting in 6,000 + +prompts. We generate 20 images per prompt and obtain synthetic feedback using object detectors [14]. This results in a dataset of $78.5\mathrm{k}$ image-feedback pairs. Our pipeline follows the setup of SANA-1.5 but produces a smaller dataset (2M images in SANA-1.5) due to computational constraints. + +# 4.1.2. Training + +We train the VLM judge for 1 epoch with a learning rate of 1e-5. We train the DiT and Context Transformer for 5,000 steps with a learning rate of 1e-5 and batch size of 48. We freeze the image and text encoders, and finetune the DiT and Context Transformer end-to-end. Training is conducted on Nvidia A6000 GPUs and takes approximately one day. + +# 4.2. Sampling + +For all experiments, we use the DPM-Solver++ sampler proposed by SANA. We use 20 sampling steps per image and set the maximum number of images per prompt to 20 $(N = 20)$ . For Reflect-DiT, we set the maximum number of in-context feedback to 3 $(K = 3)$ unless stated otherwise. We provide further details on inference speed in Appendix A.2. + +# 4.2.1. Baselines + +We compare against several training-free baselines and finetuning-based methods. For training-free baselines, we compare against best-of-N sampling employed by SANA-1.5. For finetuning baselines, we consider supervised finetuning (SFT) and Diffusion-DPO[52]. For both methods, we use SANA-1.0-1.6B as the base model. Since these finetuning methods do not enable additional inference-time scaling capabilities, we combine them with best-of-N sampling to equalize test-time compute. + +![](images/50307150dacbdb6f1d312d5410eaf047e82e77f62f4cb73abcd814cc111b8591.jpg) +Reflect-DiT + +Best-of-N +![](images/db988f713c018548d941715f537476a6d42a260103f2abe8ab57b399fa741437.jpg) +"A photo of a cup left of an umbrella" + +![](images/ca960e1fec5186abfa6d31b1588badfbfc848418d3a76c4b83f6ee43cf56a756.jpg) +Reflect-DiT + +Best-of-N +C +![](images/04d4979199972e3de81b25a14b68ff72e437e1677c6be84919c766deaec26b35.jpg) +"Five monarch butterflies flying around a single dandelion above the grass" + +![](images/ecf368d0c9dc58d39d12bdc2c5b386b8e1680bee6b5578db96bc451b9c4227a2.jpg) +Reflect-DiT +Best-of-N + +![](images/1d86169cdf3afbce104429dc699128121d2f0f24b334eca12df69dbf61121a98.jpg) +"Crown shaped splash created by a raindrop falling into a forest puddle" + +# 4.3. Main results + +We report results on the GenEval benchmark in Table 1. Reflect-DiT achieves the highest overall score (0.81) using at most 20 samples per prompt, outperforming all other models, including those with substantially more $(3\times)$ parameters. Compared to the SANA-1.0-1.6B baseline, Reflect-DiT demonstrates consistent improvements across all evaluation categories, with an overall gain of $+0.19$ . Notably, we establish a new state-of-the-art (0.81), surpassing the previous best (0.80), which was achieved using best-of-N sampling with N=2048 using SANA-1.5-4.8B. Improvements are especially pronounced for prompts requiring complex reasoning over multiple objects (e.g. counting, spatial positioning, and attribute binding). As prompt complexity increases, the baseline model struggles to generate high-quality samples, since the probability of satisfying multiple constraints is low. + +# 4.4. Qualitative Examples + +We present samples from the generation trajectory of Reflect-DiT in Figure 3. Compared with best-of-N sampling with random search, Reflect-DiT improves image quality more efficiently by tailoring future generations according to prior feedback. For example, in the rightmost trajectory of Figure 3 the initial generations fail to render a "crown shaped splash", but as the Reflect-DiT trajectory progresses, the rendered splash resembles the shape of a crown (top row). In contrast, best-of-N sampling (bottom row) fails to find a generation with a crown shaped splash. + +Figure 6 showcases some examples of the self-correction process of Reflect-DiT. The vision-language model cor + +![](images/247af57c561391d47ee77d32bdaf68e648547f08d7d6d13678df90ca24121f41.jpg) +Figure 3. Side-by-side qualitative comparison of Reflect-DiT and best-of-N sampling. Reflect-DiT leverages feedback to iteratively refine image generations, resulting in more accurate and visually coherent outputs. In the first example, Reflect-DiT progressively adjusts object positions to better satisfy the prompt "a cup left of an umbrella," achieving significantly better image-text alignment than best-of-N sampling. The second example demonstrates how Reflect-DiT corrects multiple counting constraints ("five monarch butterflies" and "a single dandelion") over successive iterations, gradually converging to the correct solution. Lastly, in the rightmost example, Reflect-DiT uses in-context feedback to refine object shapes, producing a more precise and intentional design compared to best-of-N. +Figure 4. Comparison of Reflect-DiT with other finetuning methods. We find that Reflect-DiT is able to consistently outperform finetuning methods, like supervised finetuning (SFT) and Diffusion-DPO (DPO). Using only 4 samples, Reflect-DiT can outperform related finetuning baselines using best-of-20 sampling. + +rectly identifies misalignments between prompts and generated images, enabling Reflect-DiT to improve subsequent generations based on this feedback. For example, the top-most row of Figure 6 shows generations from a prompt involving multiple entities and relative positioning constraints. Here, the VLM judge accurately identifies issues in generated images, such as inconsistent positioning and missing subjects, and Reflect-DiT is able to improve future generations based on this feedback. + +
Number of Samples
48121620
Baseline0.7050.7310.7430.7490.751
+SFT0.7360.7540.7600.7670.767
+DPO0.7130.7410.7470.7540.765
+Refl. DiT0.7760.7920.7990.7990.807
+ +![](images/3e740dbf33a760907dacf1c4c52ee4a473d7339d7bf13658f957458d4acfdab7.jpg) +Figure 5. Human evaluation win-rate $(\%)$ on PartiPrompts dataset. We perform a user study to evaluate the effectiveness of Reflect-DiT in broadly improving text-to-image generation. Results show that human evaluators consistently prefer generations from Reflect-DiT over best-of-N sampling. + +# 4.5. Comparison with Finetuning + +We additionally compare with several finetuning methods, including supervised-finetuning (SFT) and Diffusion-DPO [52]. We report results on GenEval with varying number of samples in Table 2 and Figure 4. Since SFT and DPO do not enable additional inference-time scaling, we combine them with best-of-N sampling. Reflect-DiT outperforms both baselines by a considerable margin, especially at large number of samples. Using just 4 samples, Reflect-DiT outperforms both the SFT and DPO baselines with 20 samples, saving $80\%$ of the compute budget. Among the baselines, SFT and DPO have nearly identical performance with $\mathrm{N} = 20$ samples, but SFT outperforms DPO on fewer samples, presumably because of the KL-divergence penalty. + +# 4.6. Human Evaluation + +To evaluate Reflect-DiT's effectiveness in real-world use cases beyond GenEval-style prompts, we conduct additional evaluations using 100 randomly sampled prompts from PartiPrompts[65]. We conduct human evaluations and + +ask evaluators to compare Reflect-DiT against best-of-N sampling, with a maximum of 20 samples per prompt for each method. We attempted to evaluate generations using frontier VLMs such as GPT-4o, but encountered similar problems as [61], where the API gives inconsistent outputs and exhibits a strong bias towards the first presented image. Evaluators are given a pair of images and a prompt and asked to determine which generated image is better, without any prior knowledge of how they were generated. Three responses are collected per image pair. Our results demonstrate that Reflect-DiT significantly outperforms the best-of-N baseline, with evaluators selecting Reflect-DiT $53.3\%$ of the time compared to $32.3\%$ for best-of-N (Fig. 5). We provide additional details in Appendix C.3. + +# 4.7. Ablation Studies + +We measure the effect of various design choices by conducting a series of ablation studies on the number of in-context feedback, Transformer layers, and image embeddings. + +# 4.7.1. Number of Feedback + +One of the key innovations of Reflect-DiT is the ability to utilize multi-turn feedback. We experiment with varying amounts of feedback and find that an increase in-context feedback ( $K$ in Algorithm 1) yields better performance, with the 5-feedback setup achieving the best result.(Tab. 3). To achieve a balance of speed and quality, we pick $N = 3$ , since increasing $N = 3$ to $N = 5$ only leads to $+0.01$ score while increasing the number of tokens by $1.6 \times$ . + +Table 2. GenEval performance using different finetuning methods. Results show that Reflect-DiT consistently outperforms supervised finetuning (SFT), and Diffusion-DPO (DPO) across a varying number of samples at inference. We use best-of-N sampling for the base model, SFT, and DPO baselines and in-context reflection for Reflect-DiT. + +
Number of Samples
48121620
Baseline0.7050.7310.7430.7490.751
K = 10.7380.7550.7610.7720.766
K = 20.7430.7650.7650.7740.781
K = 30.7760.7920.7990.7990.807
K = 50.7760.7940.8040.8160.816
+ +Table 3. Ablation study on the number of in-context imagefeedback pairs (K). + +# 4.7.2. Number of Image Embeddings + +We also explore different down-sampling sizes for image embeddings (default $8 \times 8$ ) and present the results in Table 4. Results show that using more tokens to represent past generations leads to better performance. + +![](images/302875619e347a91dc75610455bffa046481bf3b5e2b4e3fb1166afb3e6fd756.jpg) +Figure 6. Illustration of the iterative refinement process in Reflect-DiT. Reflect-DiT starts with an initial image generated from the prompt and progressively refines it based on textual feedback until the final output meets the desired criteria, demonstrating the effectiveness of our reflection-based approach. In the first sequence, Reflect-DiT handles a complex scene by gradually repositioning multiple objects—“woman,” “tree,” “cat,” and “dog”—to achieve correct spatial alignment. Additionally, it recognizes subtle object misclassifications, such as changing the second “dog” to a “cat” based on feedback. The second example demonstrates a counting problem, where the model iteratively adjusts the number of detached seeds until it converges to the correct count. The final example presents a particularly challenging scenario: the prompt requires the “dog” to be positioned to the “right of a tie”, an unusual object to appear independently. + +
Number of Samples
48121620
Baseline0.7050.7310.7430.7490.751
4 × 40.7700.7790.7830.7860.786
6 × 60.7520.7810.7950.7950.801
8 × 80.7760.7920.7990.7990.807
+ +# 4.8. Generalization + +To further validate the effectiveness of Reflect-DiT beyond Gen-Eval style prompts, we report additional results on DPG-Bench [18] to highlight the effectiveness of Reflect-DiT on a broad range of prompts in Tab. 5. Reflect-Dit demonstrate strong performance and was able to improve the generation of the Sana baseline. We include further discussion in Appendix A.1. + +# 5. Conclusion + +In this work, we presented Reflect-DiT, a framework for inference-time scaling of text-to-image diffusion models that leverages reflection on past generations and natural language feedback. Results show that Reflect-DiT significantly outperforms naive best-of-N sampling, which to date has been the most effective inference-time scaling method, and establishes a new state-of-the-art on the + +Table 4. Ablation study on the size of image embeddings. + +
Parm.AllHard-246Hard-56
SD3-Medium2B84.1--
DALLE 3-83.5--
FLUX-dev12B84.0--
SANA-1.54.8B85.0--
SANA-1.01.6B84.156.124.2
+Best-of-201.6B85.663.441.0
+Reflect-DiT1.7B86.169.551.8
(Δ vs Base.)-+2.0+13.4+27.6
+ +Table 5. Additional quantitative results on DPG-Bench [19] + +GenEval benchmark, outperforming larger models with substantially higher sampling budgets at test-time. While the Reflect-DiT framework has shown to be a promising approach to improve text-to-image generation, it is important to note that our method inherits the biases and shortcomings of both the base text-to-image model and the VLM judge that guides refinement. For example, the VLM judge, like other VLMs, may generate incorrect feedback due to hallucination. We also find it challenging for the VLM to recognize small objects. Future work should focus on auditing these limitations and developing safeguards to ensure responsible deployment of Reflect-DiT in applications. Further discussion of limitations is provided in Appendix B. + +# 6. Acknowledgement + +AG was supported by NSF CAREER award 2341040 + +# References + +[1] 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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18370-18380, 2023. 3 +[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 4, 15, 16 +[3] 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.openuai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 5 +[4] 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 (CVPR), pages 18392-18402, 2023. 3, 13 +[5] Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christopher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling. arXiv preprint arXiv:2407.21787, 2024. 1, 2 +[6] 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. Advances in Neural Information Processing Systems, 33:1877-1901, 2020. 2 +[7] Jiefeng Chen, Jie Ren, Xinyun Chen, Chengrun Yang, Ruoxi Sun, and Sercan Ö Arik. Sets: Leveraging self-verification and self-correction for improved test-time scaling. arXiv preprint arXiv:2501.19306, 2025. 2 +[8] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-imagen: Retrieval-augmented text-to-image generator. In The Eleventh International Conference on Learning Representations, 2023. 3 +[9] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 2 +[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 *Forty-first international conference on machine learning*, 2024. 5 + +[11] Yutong Feng, Biao Gong, Di Chen, Yujun Shen, Yu Liu, and Jingren Zhou. Ranni: Taming text-to-image diffusion for accurate instruction following. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4744-4753, 2024. 3 +[12] 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, 2022.3 +[13] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 12709-12720, 2024. 3, 13 +[14] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36:52132-52152, 2023. 2, 5 +[15] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 2, 14 +[16] Ziyu Guo, Renrui Zhang, Chengzhuo Tong, Zhizheng Zhao, Peng Gao, Hongsheng Li, and Pheng-Ann Heng. Can we generate images with cot? let's verify and reinforce image generation step by step, 2025. 5, 14 +[17] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In European Conference on Computer Vision, pages 289-305. Springer, 2024. 16 +[18] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu Ella. Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 5(7): 16, 2024. 8, 13 +[19] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 8 +[20] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: creative and controllable image synthesis with composable conditions. In Proceedings of the 40th International Conference on Machine Learning, pages 13753-13773, 2023. 3 +[21] Ying Jin, Pengyang Ling, Xiaoyi Dong, Pan Zhang, Jiaqi Wang, and Dahua Lin. Reasonpix2pix: instruction reasoning dataset for advanced image editing. arXiv preprint arXiv:2405.11190, 2024.3 +[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. 3 +[23] Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, + +Colton Bishop, Rebecca Roelofs, et al. Training language models to self-correct via reinforcement learning. arXiv preprint arXiv:2409.12917, 2024. 2 +[24] 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. 3 +[25] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.4, 5 +[26] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36:30146-30166, 2023. 3 +[27] 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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511-22521, 2023. 3 +[28] Zhengyang Liang, Hao He, Ceyuan Yang, and Bo Dai. Scaling laws for diffusion transformers. arXiv preprint arXiv:2410.08184, 2024. 1 +[29] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2023. 1 +[30] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Chase Lambert, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 5 +[31] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 3 +[32] Yang Luo, Xiaozhe Ren, Zangwei Zheng, Zhuo Jiang, Xin Jiang, and Yang You. CAME: Confidence-guided adaptive memory efficient optimization. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4442-4453, Toronto, Canada, 2023. Association for Computational Linguistics. 16 +[33] 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, New York, NY, USA, 2024. Association for Computing Machinery. 3 +[34] Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, et al. Inference-time scaling for diffusion models beyond scaling denoising steps. arXiv preprint arXiv:2501.09732, 2025. 2, 3 +[35] Ruotian Ma, Peisong Wang, Cheng Liu, Xingyan Liu, Jiaqi Chen, Bang Zhang, Xin Zhou, Nan Du, and Jia Li. $s^2 r$ : + +Teaching llms to self-verify and self-correct via reinforcement learning. arXiv preprint arXiv:2502.12853, 2025. 2 +[36] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36:46534-46594, 2023. 2 +[37] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 3 +[38] OpenAI. Gpt-4 technical report, 2023. arXiv preprint arXiv:2303.08774. 2 +[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. arXiv preprint arXiv:2307.01952, 2023. 1, 4, 5, 14 +[40] Senthil Purushwalkam, Akash Gokul, Shafiq Joty, and Nikhil Naik. Bootpig: Bootstrapping zero-shot personalized image generation capabilities in pretrained diffusion models. arXiv preprint arXiv:2401.13974, 2024. 3 +[41] Yuxiao Qu, Tianjun Zhang, Naman Garg, and Aviral Kumar. Recursive introspection: Teaching language model agents how to self-improve. Advances in Neural Information Processing Systems, 37:55249-55285, 2024. 2 +[42] 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. 1 +[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. 1, 14 +[44] Robin Rombach, Andreas Blattmann, and Björn Ommer. Text-guided synthesis of artistic images with retrieval-augmented diffusion models. arXiv preprint arXiv:2207.13038, 2022. 3 +[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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3 +[46] Shelly Sheynin, Oron Ashual, Adam Polyak, Uriel Singer, Oran Gafni, Eliya Nachmani, and Yaniv Taigman. Knndiffusion: Image generation via large-scale retrieval. arXiv preprint arXiv:2204.02849, 2022. 3 +[47] Raghav Singhal, Zachary Horvitz, Ryan Teehan, Mengye Ren, Zhou Yu, Kathleen McKeown, and Rajesh Ranganath. A general framework for inference-time scaling and steering of diffusion models. arXiv preprint arXiv:2501.06848, 2025.3, 14 +[48] Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute optimally can be more + +effective than scaling model parameters. arXiv preprint arXiv:2408.03314, 2024. 1, 2 +[49] Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhopatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024. 4, 16 +[50] 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. 4 +[51] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $\mathfrak{p}+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3 +[52] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 5, 7, 14 +[53] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal llm as an agent for unified image generation and editing. Advances in Neural Information Processing Systems, 37:128374-128395, 2024. 3, 13, 14 +[54] 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. 14 +[55] 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. 3 +[56] Sean Welleck, Ximing Lu, Peter West, Faeze Brahman, Tianxiao Shen, Daniel Khashabi, and Yejin Choi. Generating sequences by learning to self-correct. ArXiv, abs/2211.00053, 2022. 2 +[57] Yixuan Weng, Minjun Zhu, Fei Xia, Bin Li, Shizhu He, Shengping Liu, Bin Sun, Kang Liu, and Jun Zhao. Large language models are better reasoners with self-verification. arXiv preprint arXiv:2212.09561, 2022. 2 +[58] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 3, 13, 14 +[59] Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. Inference scaling laws: An empirical analysis of compute-optimal inference for problem-solving with language models. arXiv preprint arXiv:2408.00724, 2024. 1, 2, 3 +[60] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, + +Yao Lu, et al. Sana: Efficient high-resolution text-to-image synthesis with linear diffusion transformers. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 4, 5 +[61] Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng Yu, Ligeng Zhu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, Han Cai, et al. Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer. arXiv preprint arXiv:2501.18427, 2025. 2, 3, 4, 5, 7, 13, 14, 15, 16 +[62] 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 +[63] Yuxi Xie, Kenji Kawaguchi, Yiran Zhao, James Xu Zhao, Min-Yen Kan, Junxian He, and Michael Xie. Self-evaluation guided beam search for reasoning. Advances in Neural Information Processing Systems, 36:41618-41650, 2023. 2 +[64] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Qimai Li, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. arXiv preprint arXiv:2311.13231, 2023. 14 +[65] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 7 +[66] 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 +[67] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 4, 15 +[68] 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. 3 +[69] Lunjun Zhang, Arian Hosseini, Hritik Bansal, Mehran Kazemi, Aviral Kumar, and Rishabh Agarwal. Generative verifiers: Reward modeling as next-token prediction, 2024. URL https://arxiv.org/abs/2408.15240, 2024. 1 +[70] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9026-9036, 2024. 3 +[71] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:11127-11150, 2023. 3 \ No newline at end of file diff --git a/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/images.zip b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..004c35c7d75856748b06a3ce3fcb65ac99223666 --- /dev/null +++ b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0cbc23ae945d3b0a48b4bf6cb3e26dcddc2fb64bc0ca672dd24ddea4c9a4a0c +size 604400 diff --git a/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/layout.json b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3cc6e640624bdd82c5f9dcf5965bb60833c7fdd8 --- /dev/null +++ b/ICCV/2025/Reflect-DiT_ Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a668aab9c271a6129886072ba20259646c0e9be083c7ebd46d5a33fe34167f31 +size 413532