CVPR 2025 论文列表

Spatial understanding is a crucial capability that enables robots to perceive their surroundings, reason about their environment, and interact with it meaningfully. In modern robotics, these capabilities are increasingly provided by vision-language models. However, these models face significant challenges in spatial reasoning tasks, as their training data are based on general-purpose image datasets that often lack sophisticated spatial understanding. For example, datasets frequently do not capture reference frame comprehension, yet effective spatial reasoning requires understanding whether to reason from ego-, world-, or object-centric perspectives. To address this issue, we introduce RoboSpatial, a large-scale dataset for spatial understanding in robotics. It consists of real indoor and tabletop scenes, captured as 3D scans and egocentric images, and annotated with rich spatial information relevant to robotics. The dataset includes 1M images, 5k 3D scans, and 3M annotated spatial relationships, and the pairing of 2D egocentric images with 3D scans makes it both 2D- and 3D- ready. Our experiments show that models trained with RoboSpatial outperform baselines on downstream tasks such as spatial affordance prediction, spatial relationship prediction, and robotics manipulation.

Recent advancements in robotic manipulation have highlighted the potential of intermediate representations for improving policy generalization. In this work, we explore grounding masks as an effective intermediate representation, balancing two key advantages: (1) effective spatial guidance that specifies target objects and placement areas while also conveying information about object shape and size, and (2) broad generalization potential driven by large-scale vision-language models pretrained on diverse grounding datasets. We introduce \method, a grounding-aware robotic manipulation policy that leverages grounding masks as an intermediate representation to guide policy networks in object manipulation tasks. To further explore and enhance generalization, we propose an automated pipeline for generating large-scale, simulated data with a diverse set of objects and instructions. Extensive experiments show the value of our dataset and the effectiveness of grounding masks as intermediate guidance, significantly enhancing the generalization abilities of robot policies.

In skeleton-based action recognition, a key challenge is distinguishing between actions with similar trajectories of joints due to the lack of image-level details in skeletal representations. Recognizing that the differentiation of similar actions relies on subtle motion details in specific body parts, we direct our approach to focus on the fine-grained motion of local skeleton components. To this end, we introduce ProtoGCN, a Graph Convolutional Network (GCN)-based model that breaks down the dynamics of entire skeleton sequences into a combination of learnable prototypes representing core motion patterns of action units. By contrasting the reconstruction of prototypes, ProtoGCN can effectively identify and enhance the discriminative representation of similar actions. Without bells and whistles, ProtoGCN achieves state-of-the-art performance on multiple benchmark datasets, including NTU RGB+D, NTU RGB+D 120, Kinetics-Skeleton, and FineGYM, which demonstrates the effectiveness of the proposed method. The code is available at https://github.com/firework8/ProtoGCN.

Deepfake detection is a long-established research topic vital for mitigating the spread of malicious misinformation. Unlike prior methods that provide either binary classification results or textual explanations separately, we introduce a novel method capable of generating both simultaneously. Our method harnesses the multi-modal learning capability of the pre-trained CLIP and the unprecedented interpretability of large language models (LLMs) to enhance both the generalization and explainability of deepfake detection. Specifically, we introduce a multi-modal face forgery detector (M2F2-Det) that employs tailored face forgery prompt learning, incorporating the pre-trained CLIP to improve generalization to unseen forgeries. Also, M2F2-Det incorporates an LLM to provide detailed textual explanations of its detection decisions, enhancing interpretability by bridging the gap between natural language and subtle cues of facial forgeries. Empirically, we evaluate M2F2-Det on both detection and explanation generation tasks, where it achieves state-of-the-art performance, demonstrating its effectiveness in identifying and explaining diverse forgeries. Source code and models are available at this link.

While vision-language models like CLIP have shown remarkable success in open-vocabulary tasks, their application is currently confined to image-level tasks, and they still struggle with dense predictions. Recent works often attribute such deficiency in dense predictions to the self-attention layers in the final block, and have achieved commendable results by modifying the original query-key attention to self-correlation attention, (e.g., query-query and key-key attention). However, these methods overlook the cross-correlation attention (query-key) properties, which capture the rich spatial correspondence. In this paper, we reveal that the cross-correlation of self-attention in non-final layers of CLIP also exhibits localization properties. Therefore, we propose the Residual Cross-correlation Self-attention (RCS) module, which leverages the cross-correlation self-attention from intermediate layers to remold the attention in the final block. The RCS module effectively reorganizes spatial information, unleashing the localization potential within CLIP for dense vision-language inference. Furthermore, to enhance the focus on regions of the same categories and local consistency, we propose the Semantic Feedback Refinement (SFR) module, which utilizes semantic segmentation maps to further adjust the attention scores. By integrating these two strategies, our method, termed ResCLIP, can be easily incorporated into existing approaches as a plug-and-play module, significantly boosting their performance in dense vision-language inference. Extensive experiments across multiple standard benchmarks demonstrate that our method surpasses state-of-the-art training-free methods, validating the effectiveness of the proposed approach. Code is available at https://github.com/yvhangyang/ResCLIP.

Contrastive Language-Image Pre-training (CLIP) models as a milestone of modern multimodal intelligence, its generalization mechanism grasped massive research interests in the community. While existing studies limited in the scope of pre-training knowledge, hardly underpinned its generalization to countless open-world concepts absent from the pre-training regime. This paper dives into such Out-of-Pre-training (OOP) generalization problem from a holistic perspective. We propose LAION-Beyond benchmark to isolate the evaluation of OOP concepts from pre-training knowledge, with regards to OpenCLIP and its reproducible variants derived from LAION datasets. Empirical analysis evidences that despite image features of OOP concepts born with significant category margins, their zero-shot transfer significantly fails due to the poor image-text alignment. To this, we elaborate the "name-tuning" methodology with its theoretical merits in terms of OOP generalization, then propose few-shot name learning (FSNL) and zero-shot name learning (ZSNL) algorithms to achieve OOP generalization in a data-efficient manner. Their superiority have been further verified in our comprehensive experiments.

The zero-shot capabilities of Vision-Language Models (VLMs) have been widely leveraged to improve predictive performance. However, previous works on transductive or test-time adaptation (TTA) often make strong assumptions about the data distribution, such as the presence of all classes. Our work challenges these favorable deployment scenarios and introduces a more realistic evaluation framework, including (i) a variable number of effective classes for adaptation within a single batch, and (ii) non-i.i.d. batches of test samples in online adaptation settings. We provide comprehensive evaluations, comparisons, and ablation studies that demonstrate how current transductive or TTA methods for VLMs systematically compromise the models' initial zero-shot robustness across various realistic scenarios, favoring performance gains under advantageous assumptions about the test sample distributions. Furthermore, we introduce StatA, a versatile method that can handle a wide range of deployment scenarios, including those with a variable number of effective classes at test time. Our approach incorporates a novel regularization term designed specifically for VLMs, which acts as a statistical anchor preserving the initial text-encoder knowledge, particularly in low-data regimes. Code available at https://github.com/MaxZanella/StatA.

Large language models (LLMs) excel at retrieving information from lengthy text, but their vision-language counterparts (VLMs) face difficulties with hour-long videos, especially for temporal grounding. Specifically, these VLMs are constrained by frame limitations, often losing essential temporal details needed for accurate event localization in extended video content. We propose ReVisionLLM, a recursive vision-language model designed to locate events in hour-long videos. Inspired by human search strategies, our model initially targets broad segments of interest, progressively revising its focus to pinpoint exact temporal boundaries. Our model can seamlessly handle videos of vastly different lengths--from minutes to hours. We also introduce a hierarchical training strategy that starts with short clips to capture distinct events and progressively extends to longer videos. To our knowledge, ReVisionLLM is the first VLM capable of temporal grounding in hour-long videos, outperforming previous state-of-the-art methods across multiple datasets by a significant margin (e.g., +2.6% R1@0.1 on MAD). The code is available in the supplementary and will be released.

This paper presents a groundbreaking approach - the first online automatic geometric calibration method for radar and camera systems. Given the significant data sparsity and measurement uncertainty in radar height data, achieving automatic calibration during system operation has long been a challenge. To address the sparsity issue, we propose a Dual-Perspective representation that gathers features from both frontal and bird's-eye views. The frontal view contains rich but sensitive height information, whereas the bird's-eye view provides robust features against height uncertainty. We thereby propose a novel Selective Fusion Mechanism to identify and fuse reliable features from both perspectives, reducing the effect of height uncertainty. Moreover, for each view, we incorporate a Multi-Modal Cross-Attention Mechanism to explicitly find location correspondences through cross-modal matching. During the training phase, we also design a Noise-Resistant Matcher to provide better supervision and enhance the robustness of the matching mechanism against sparsity and height uncertainty. Our experimental results, tested on the nuScenes dataset, demonstrate that our method significantly outperforms previous radar-camera auto-calibration methods, as well as existing state-of-the-art LiDAR-camera calibration techniques, establishing a new benchmark for future research.

Agglomerative models have recently emerged as a powerful approach to training vision foundation models, leveraging multi-teacher distillation from existing models such as CLIP, DINO, and SAM. This strategy enables the efficient creation of robust models, combining the strengths of individual teachers while significantly reducing computational and resource demands. In this paper, we thoroughly analyze state-of-the-art agglomerative models, identifying critical challenges including resolution mode shifts, teacher imbalance, idiosyncratic teacher artifacts, and an excessive number of output tokens. To address these issues, we propose several novel solutions: multi-resolution training, mosaic augmentation, and improved balancing of teacher loss functions. Specifically, in the context of Vision Language Models, we introduce a token compression technique to maintain high-resolution information within a fixed token count. We release our top-performing variants at multiple scales (-B, -L, -H, and -g), along with inference code and pretrained weights.

This paper explores using large language models (LLMs) as low-level action planners for embodied tasks. While LLMs excel as the robot's "brain" for high-level planning, they face challenges in directly controlling the "body" by generating precise low-level actions. This limitation arises from LLMs' strength in high-level conceptual understanding but their inability to handle spatial perception effectively, restricting their potential in embodied tasks. To address this, we bridge the gap by enabling LLMs to not only comprehend complex instructions but also produce actionable, low-level plans. We introduce Room to Chessboard (R2C), a novel semantic representation that maps environmental states onto a grid-based chessboard, empowering LLMs to generate specific low-level coordinates and guide the robot in a manner akin to playing a game of chess. To further enhance decision-making, we propose the Chain-of-Thought Decision (CoT-D) paradigm, which improves LLMs' interpretability and context-awareness in spatial reasoning. By jointly training LLMs for high-level task decomposition and low-level action generation, we create a unified "brain-body" system capable of handling complex, free-form instructions while producing precise low-level actions, allowing the robot to flexibly control its movements and adapt to varying tasks. We validate R2C using both fine-tuned open-source LLMs and GPT-4, demonstrating effectiveness on the challenging ALFRED benchmark. Results show that with our R2C framework, LLMs can effectively act as low-level planners, generalizing across diverse settings and open-vocabulary robotic tasks. The code and demonstrations are available at: https://vipl-vsu.github.io/Room2Chessboard.

We present a simple approach to make pre-trained Vision Transformers (ViTs) interpretable for fine-grained analysis, aiming to identify and localize the traits that distinguish visually similar categories, such as bird species. Pre-trained ViTs, such as DINO, have demonstrated remarkable capabilities in extracting localized, discriminative features. However, saliency maps like Grad-CAM often fail to identify these traits, producing blurred, coarse heatmaps that highlight entire objects instead. We propose a novel approach, Prompt Class Attention Map (Prompt-CAM), to address this limitation. Prompt-CAM learns class-specific prompts for a pre-trained ViT and uses the corresponding outputs for classification. To correctly classify an image, the true-class prompt must attend to unique image patches not present in other classes' images (i.e., traits). As a result, the true class's multi-head attention maps reveal traits and their locations. Implementation-wise, Prompt-CAM is almost a "free lunch," requiring only a modification to the prediction head of Visual Prompt Tuning (VPT). This makes Prompt-CAM easy to train and apply, in stark contrast to other interpretable methods that require designing specific models and training processes. Extensive empirical studies on a dozen datasets from various domains (e.g., birds, fishes, insects, fungi, flowers, food, and cars) validate the superior interpretation capability of Prompt-CAM. The source code and demo are available at https://github.com/Imageomics/Prompt_CAM.

Contrastive language image pre-training (CLIP) is an essential component of building modern vision-language foundation models. While CLIP demonstrates remarkable zero-shot performance on downstream tasks, the multi-modal feature spaces still suffer from a modality gap, which is a gap between image and text feature clusters and limits downstream task performance. Although existing works attempt to address the modality gap by modifying pre-training or fine-tuning, they struggle with heavy training costs with large datasets or degradations of zero-shot performance. This paper presents CLIP-Refine, a post-pre-training method for CLIP models at a phase between pre-training and fine-tuning. CLIP-Refine aims to align the feature space with 1 epoch training on small image-text datasets without zero-shot performance degradations. To this end, we introduce two techniques: random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). RaFA aligns the image and text features to follow a shared prior distribution by minimizing the distance to random reference vectors sampled from the prior. HyCD updates the model with hybrid soft labels generated by combining ground-truth image-text pair labels and outputs from the pre-trained CLIP model. This contributes to achieving both maintaining the past knowledge and learning new knowledge to align features. Our extensive experiments with multiple classification and retrieval tasks show that CLIP-Refine succeeds in mitigating the modality gap and improving the zero-shot performance. Code is available at https://github.com/yshinya6/clip-refine.

Building a generalizable self-correction system is crucial for robots to recover from failures. Despite advancements in Multimodal Large Language Models (MLLMs) that empower robots with semantic reflection ability for failure, translating semantic reflection into how to correct fine-grained robotic actions remains a significant challenge. To address this gap, we build the Phoenix framework, which leverages motion instruction as a bridge to connect high-level semantic reflection with low-level robotic action correction. In this motion-based self-reflection framework, we start with a dual-process motion adjustment mechanism with MLLMs to translate the semantic reflection into coarse-grained motion instruction adjustment. To leverage this motion instruction for guiding how to correct fine-grained robotic actions, a multi-task motion-conditioned diffusion policy is proposed to integrate visual observations for high-frequency robotic action correction. By combining these two models, we could shift the demand for generalization capability from the low-level manipulation policy to the MLLMs-driven motion adjustment model and facilitate precise, fine-grained robotic action correction. Utilizing this framework, we further develop a lifelong learning method to automatically improve the model's capability from interactions with dynamic environments. The experiments conducted in both the RoboMimic simulation and real-world scenarios prove the superior generalization and robustness of our framework across a variety of manipulation tasks.

Anomaly Detection (AD) and Anomaly Localization (AL) are crucial in fields that demand high reliability, such as medical imaging and industrial monitoring. However, current AD and AL approaches are often susceptible to adversarial attacks due to limitations in training data, which typically include only normal, unlabeled samples. This study introduces PatchGuard, an adversarially robust AD and AL method that incorporates pseudo anomalies with localization masks within a Vision Transformer (ViT)-based architecture to address these vulnerabilities.We begin by examining the essential properties of pseudo anomalies, and follow it by providing theoretical insights into the attention mechanisms required to enhance the adversarial robustness of AD and AL systems. We then present our approach, which leverages Foreground-Aware Pseudo-Anomalies to overcome the deficiencies of previous anomaly-aware methods. Our method incorporates these crafted pseudo-anomaly samples into a ViT-based framework, with adversarial training guided by a novel loss function designed to improve model robustness, as supported by our theoretical analysis.Experimental results on well-established industrial and medical datasets demonstrate that PatchGuard significantly outperforms previous methods in adversarial settings, achieving performance gains of 53.2% in AD and 68.5% in AL, while also maintaining competitive accuracy in non-adversarial settings.

Large Vision-Language Models (VLMs) have been extended to understand both images and videos. Visual token compression is leveraged to reduce the considerable token length of visual inputs. To meet the needs of different tasks, existing high-performance models usually process images and videos separately with different token compression strategies, limiting the capabilities of combining images and videos. To this end, we extend each image into a "static" video and introduce a unified token compression strategy called Progressive Visual Token Compression (PVC), where the tokens of each frame are progressively encoded and adaptively compressed to supplement the information not extracted from previous frames. Video tokens are efficiently compressed with exploiting the inherent temporal redundancy. Images are repeated as static videos, and the spatial details can be gradually supplemented in multiple frames. PVC unifies the token compressing of images and videos. With a limited number of tokens per frame (64 tokens by default), spatial details and temporal changes can still be preserved. Experiments show that our model achieves state-of-the-art performance across various video understanding benchmarks, including long video tasks and fine-grained short video tasks. Meanwhile, our unified token compression strategy incurs no performance loss on image benchmarks, particularly in detail-sensitive tasks.

Vision language models (VLMs) respond to user-crafted text prompts and visual inputs, and are applied to numerous real-world problems. VLMs integrate visual modalities with large language models (LLMs), which are well known to be prompt-sensitive. Hence, it is crucial to determine whether VLMs inherit this instability to varying prompts. We therefore investigate which prompt variations VLMs are most sensitive to and which VLMs are most agnostic to prompt variations. To this end, we introduce PARC (Prompt Analysis via Reliability and Calibration), a VLM prompt sensitivity analysis framework built on three pillars: (1) plausible prompt variations in both the language and vision domain, (2) a novel model reliability score with built-in guarantees, and (3) a calibration step that enables dataset- and prompt-spanning prompt variation analysis. Regarding prompt variations, PARC's evaluation shows that VLMs mirror LLM language prompt sensitivity in the vision domain, and most destructive variations change the expected answer. Regarding models, outstandingly robust VLMs among 22 evaluated models come from the InternVL2 family. We further find indications that prompt sensitivity is linked to training data. https://github.com/NVlabs/PARC

Vision-language model (VLM) is one of the most important models for multi-modal tasks. Real industrial applications often meet the challenge of adapting VLMs to different scenarios, such as varying hardware platforms or performance requirements. Traditional methods involve training or fine-tuning to adapt multiple unique VLMs or using model compression techniques to create multiple compact models. These approaches are complex and resource-intensive. This paper introduces a novel paradigm called Once-Tuning-Multiple-Variants (OTMV). OTMV requires only a single tuning process to inject dynamic weight expansion capacity into the VLM with dynamic expansion capacity. This tuned VLM can then be expanded into multiple variants tailored for different scenarios in inference. The tuning mechanism of OTMV is inspired by the mathematical series expansion theorem, which helps to reduce the parameter size and memory requirements while maintaining accuracy for VLM. Experiment results show that OTMV-tuned models achieve comparable accuracy to baseline VLMs across various visual-language tasks. The experiments also demonstrate the dynamic expansion capability of OTMV-tuned VLMs, outperforming traditional model compression and adaptation techniques in terms of accuracy and efficiency.