CVPR 2025 Papers

Improving semantic grounding in Vision-Language Models (VLMs) often involves collecting domain-specific training data, refining the network architectures, or modifying the training recipes. In this work, we venture into an orthogonal direction and explore self-correction in VLMs focusing on semantic grounding. We find that VLMs can correct their own semantic grounding mistakes when properly prompted and framed for the task, without any fine-tuning or even access to oracle feedback. We also introduce a self-correction framework in an iterative setting which consistently improves performance across all models investigated. Overall, we show that iterative self-correction consistently improves VLM performance in semantic grounding by up to 8.4 accuracy points across all models investigated, without requiring fine-tuning, additional architectural changes, or external data. Our exploration of self-correction also reveals that, even after several rounds of feedback, strong models like GPT-4V and GPT-4o retain limited capability in leveraging oracle feedback, suggesting promising directions for further research.

Accurate automatic breast ultrasound (BUS) image segmentation is essential for early breast cancer screening and diagnosis. However, it remains challenging owing to (1) breast lesions of various scale and shape, (2) ambiguous boundaries caused by speckle noise and artifacts, and (3) the scarcity of high-quality annotations. Most existing semi-supervised methods employ the mean-teacher architecture, which merely learns semantic information within a single image and heavily relies on the performance of the teacher model. Therefore, we present a novel cross-image semantic correlation semi-supervised framework, named CSC-PA, to improve the performance of BUS image segmentation. CSC-PA is trained based on a single network, which integrates a foreground prototype attention (FPA) and an edge prototype attention (EPA). Specifically, FPA transfers complementary foreground information for more stable and complete lesion segmentation. On the other hand, EPA enhances edge features of lesions by using edge prototype, where an adaptive edge container is proposed to store global edge features and generate the edge prototype. Additionally, we introduce a pixel affinity loss (PAL) to exploit previously ignored contextual correlation in supervision, which further improves performance on edges. Extensive experiments on two benchmark BUS datasets demonstrate that our model outperforms other state-of-the-art methods under different partition protocols. Codes are available at https://github.com/shdkdh/CSC-PA.

The emergence of large multimodal models (LMMs) has brought significant advancements to pathology. Previous research has primarily focused on separately training patch-level and whole-slide image (WSI)-level models, limiting the integration of learned knowledge across patches and WSIs and resulting in redundant models. In this work, we introduce CPath-Omni, the first 15B parameter LMM that unifies patch and WSI analysis, consolidating a variety of tasks at both levels, including classification, visual question answering, captioning, and visual referring prompting. Extensive experiments demonstrate that CPath-Omni achieves state-of-the-art (SOTA) performance across seven diverse tasks on 39 out of 42 datasets, outperforming or matching task-specific models trained for individual tasks. Additionally, we develop a specialized pathology CLIP-based visual processor for CPath-Omni, CPath-CLIP, which, for the first time, integrates different vision models and incorporates a large language model as a text encoder to build a more powerful CLIP model, which achieves SOTA performance on nine zero-shot and four few-shot datasets. Our findings highlight CPath-Omni's ability to unify diverse pathology tasks, demonstrating its potential to streamline and advance the field of foundation model in pathology. The code and model are available at https://github.com/PathFoundation/CPath-Omni.

Current object detectors often suffer significant performance degradation in real-world applications when encountering distributional shifts, posing serious risks in high-stakes domains such as autonomous driving and medical diagnosis. Consequently, the out-of-distribution (OOD) generalization capability of object detectors has garnered increasing attention from researchers. Despite this growing interest, there remains a lack of a large-scale, comprehensive dataset and evaluation benchmark with fine-grained annotations tailored to assess the OOD generalization on more intricate tasks like object detection and grounding. To address this gap, we introduce COUNTS, a large-scale OOD dataset with object-level annotations. COUNTS encompasses 14 natural distributional shifts, over 222K samples, and more than 1,196K labeled bounding boxes. Leveraging COUNTS, we introduce two novel benchmarks: O(OD) and OODG. OODOD is designed to comprehensively evaluate the OOD generalization capabilities of object detectors by utilizing controlled distribution shifts between training and testing data. OODG, on the other hand, aims to assess the OOD generalization of grounding abilities in multimodal large language models (MLLMs). Our findings reveal that, while large models and extensive pre-training data substantially enhance performance in in-distribution (IID) scenarios, significant limitations and opportunities for improvement persist in OOD contexts for both object detectors and MLLMs. In visual grounding tasks, even the advanced GPT-4o and Gemini-1.5 only achieve 56.7% and 28.0% accuracy, respectively. We hope COUNTS facilitates advancements in the development and assessment of robust object detectors and MLLMs capable of maintaining high performance under distributional shifts.

Vision-Language Models (VLMs) trained with contrastive loss have achieved significant advancements in various vision and language tasks. However, the global nature of the contrastive loss makes VLMs focus predominantly on foreground objects, neglecting other crucial information in the image, which limits their effectiveness in downstream tasks. To address these challenges, we propose COSMOS: CrOSs-MOdality Self-distillation for vision-language pre-training that integrates a novel text-cropping strategy and cross-attention module into a self-supervised learning framework. We create global and local views of images and texts (i.e., multi-modal augmentations), which are essential for self-distillation in VLMs. We further introduce a cross-attention module, enabling COSMOS to learn comprehensive cross-modal representations optimized via a cross-modality self-distillation loss. COSMOS consistently outperforms previous strong baselines on various zero-shot downstream tasks, including retrieval, classification, and semantic segmentation. Additionally, it surpasses CLIP-based models trained on larger datasets in visual perception and contextual understanding tasks. Code is available at https://github.com/ExplainableML/cosmos.

Recent vision-language models (VLMs) face significant challenges in test-time adaptation to novel domains. While cache-based methods show promise by leveraging historical information, they struggle with both caching unreliable feature-label pairs and indiscriminately using single-class information during querying, significantly compromising adaptation accuracy. To address these limitations, we propose COSMIC (\underline C lique-\underline O riented \underline S emantic \underline M ulti-space \underline I ntegration for \underline C LIP), a robust test-time adaptation framework that enhances adaptability through multi-granular, cross-modal semantic caching and graph-based querying mechanisms. Our framework introduces two key innovations: Dual Semantics Graph (DSG) and Clique Guided Hyper-class (CGH). The Dual Semantics Graph constructs complementary semantic spaces by incorporating textual features, coarse-grained CLIP features, and fine-grained DINOv2 features to capture rich semantic relationships. Building upon these dual graphs, the Clique Guided Hyper-class component leverages structured class relationships to enhance prediction robustness through correlated class selection. Extensive experiments demonstrate COSMIC's superior performance across multiple benchmarks, achieving significant improvements over state-of-the-art methods: 15.81% gain on out-of-distribution tasks and 5.33% on cross-domain generation with CLIP RN-50.

Fine-grained open-set semi-supervised learning (OSSL) investigates a practical scenario where unlabeled data may contain fine-grained out-of-distribution (OOD) samples. Due to the subtle visual differences among in-distribution (ID) samples, as well as between ID and OOD samples, it is extremely challenging to separate the ID and OOD samples. Due to the subtle visual differences among in-distribution (ID) and OOD samples. Recent Vision-Language Models, such as CLIP, have shown excellent generalization capabilities. However, it tends to focus on general attributes, and thus is insufficient to distinguish the fine-grained details. To tackle the issues, in this paper, we propose a novel CLIP-driven coarse-to-fine semantic-guided framework, named CFSG-CLIP, to progressively focus on the distinctive fine-grained clues. Specifically, CFSG-CLIP comprises a coarse-guidance branch and a fine-guidance branch derived from the pre-trained CLIP model. In the coarse-guidance branch, we design a semantic filtering module to initially filter and highlight local visual features guided by cross-modality features. Then, in the fine-guidance branch, we further design a visual-semantic injection strategy, which embeds category-related visual cues into the visual encoder to further refine the local visual features. By the designed dual-guidance framework, local subtle cues are progressively discovered to distinct the subtle difference between ID and OOD samples. Extensive experiments demonstrate that CFSG-CLIP achieves competitive performance on multiple fine-grained datasets. The source code is available at https://github.com/LxxxxK/CFSG-CLIP.

Multimodal large language models (MLLMs) have garnered widespread attention from researchers due to their remarkable understanding and generation capabilities in visual language tasks (e.g., visual question answering). However, the rapid pace of knowledge updates in the real world makes offline training of MLLMs costly, and when faced with non-stationary data streams, MLLMs suffer from catastrophic forgetting during learning. In this paper, we propose an MLLMs-based dual momentum Mixture-of-Experts (\texttt CL-MoE ) framework for continual visual question answering. We integrate MLLMs with continual learning to utilize the rich commonsense knowledge in LLMs.We introduce a Dual-Router MoE (RMoE) to select the global and local experts using task-level and instance-level routers, to robustly assign weights to the experts most appropriate for the task. Then, we design a dynamic Momentum MoE (MMoE) to update the parameters of experts dynamically based on the relationships between the experts and tasks, so that the model can absorb new knowledge while maintaining existing knowledge. The extensive experimental results indicate that our method achieves state-of-the-art performance on 10 VQA tasks, proving the effectiveness of our approach. The codes and weights will be released on GitHub.

Composed Image Retrieval (CIR) is a multi-modal task that seeks to retrieve target images by harmonizing a reference image with a modified instruction. A key challenge in CIR lies in compositional conflicts between the reference image (e.g., blue, long sleeve) and the modified instruction (e.g., grey, short sleeve). Previous works attempt to mitigate such conflicts through feature-level manipulation, commonly employing learnable masks to obscure conflicting features within the reference image. However, the inherent complexity of feature spaces poses significant challenges in precise conflict neutralization, thereby leading to uncontrollable results. To this end, this paper proposes the Compositional Conflict Identification and Neutralization (CCIN) framework, which sequentially identifies and neutralizes compositional conflicts for effective CIR. Specifically, CCIN comprises two core modules: 1) Compositional Conflict Identification module, which utilizes LLM-based analysis to identify specific conflicting attributes, and 2) Compositional Conflict Neutralization module, which first generates a kept instruction to preserve non-conflicting attributes, then neutralizes conflicts under collaborative guidance of both the kept and modified instructions. Extensive experiments demonstrate the superiority of CCIN over the state-of-the-arts. Code repository: https://github.com/LikaiTian/CCIN.

Transformer-based methods have demonstrated impressive performance in low-level visual tasks such as Image Super-Resolution (SR). However, its computational complexity grows quadratically with the spatial resolution. A series of works attempt to alleviate this problem by dividing Low-Resolution images into local windows, axial stripes, or dilated windows. SR typically leverages the redundancy of images for reconstruction, and this redundancy appears not only in local regions but also in long-range regions. However, these methods limit attention computation to content-agnostic local regions, limiting directly the ability of attention to capture long-range dependency. To address these issues, we propose a lightweight Content-Aware Token Aggregation Network (CATANet). Specifically, we propose an efficient Content-Aware Token Aggregation module for aggregate long-range content-similar tokens, which shares token centers across all image tokens and updates them only during the training phase. Then we utilize intra-group self-attention to enable long-range information interaction. Moreover, we design an inter-group cross-attention to further enhance global information interaction. The experimental results show that, compared with the state-of-the-art cluster-based method SPIN, our method achieves superior performance, with a maximum PSNR improvement of 0.33dB and nearly double the inference speed.

In this work, we propose an extreme compression technique for Large Multimodal Models (LMMs). While previous studies have explored quantization as an efficient post-training compression method for Large Language Models (LLMs), low-bit compression for multimodal models remains under-explored. The redundant nature of inputs in multimodal models results in a highly sparse attention matrix. We theoretically and experimentally demonstrate that the attention matrix's sparsity bounds the compression error of the Query and Key weight matrices. Based on this, we introduce CASP, a model compression technique for LMMs. Our approach performs a data-aware low-rank decomposition on the Query and Key weight matrix, followed by quantization across all layers based on an optimal bit allocation process. CASP is compatible with any quantization technique and enhances state-of-the-art 2-bit quantization methods (AQLM and QuIP#) by an average of 21% on image- and video-language benchmarks. The code is provided in the supplementary materials.

Image registration estimates spatial correspondences between image pairs. These estimates are typically obtained via numerical optimization or regression by a deep network. A desirable property is that a correspondence estimate (e.g., the true oracle correspondence) for an image pair is maintained under deformations of the input images. Formally, the estimator should be equivariant to a desired class of image transformations. In this work, we present careful analyses of equivariance properties in the context of multi-step deep registration networks. Based on these analyses we 1) introduce the notions of [U,U] equivariance (network equivariance to the same deformations of the input images) and [W,U] equivariance (where input images can undergo different deformations); we 2) show that in a suitable multi-step registration setup it is sufficient for overall [W,U] equivariance if the first step has [W,U] equivariance and all others have [U,U] equivariance; we 3) show that common displacement-predicting networks only exhibit [U,U] equivariance to translations instead of the more powerful [W,U] equivariance; and we 4) show how to achieve multi-step [W,U] equivariance via a coordinate-attention mechanism combined with displacement-predicting networks. Our approach obtains excellent practical performance for 3D abdomen, lung, and brain medical image registration. We match or outperform state-of-the-art (SOTA) registration approaches on all the datasets with a particularly strong performance for the challening abdomen registration.

Recent advances in Large Vision-Language Models (LVLMs) have enabled general-purpose vision tasks through visual instruction tuning. While existing LVLMs can generate segmentation masks from text prompts for single images, they struggle with segmentation-grounded reasoning across images, especially at finer granularities such as object parts. In this paper, we introduce the new task of part-focused semantic co-segmentation, which involves identifying and segmenting common objects and their constituent common and unique parts across images. To address this task, we present CALICO, the first LVLM designed for multi-image part-level reasoning segmentation. CALICO features two key components, a novel Correspondence Extraction Module that identifies semantic part-level correspondences, and Correspondence Adaptation Modules that embed this information into the LVLM to facilitate multi-image understanding in a parameter-efficient manner. To support training and evaluation, we curate MixedParts, a large-scale multi-image segmentation dataset containing 2.4M samples across 44K images spanning diverse object and part categories. Experimental results demonstrate that CALICO, with just 0.3% of its parameters finetuned, achieves strong performance on this challenging task.

Long-form video understanding with Large Vision Language Models is challenged by the need to analyze temporally dispersed yet spatially concentrated key moments within limited context windows. In this work, we introduce VideoMindPalace, a new framework inspired by the "Mind Palace", which organizes critical video moments into a topologically structured semantic graph. VideoMindPalace organizes key information through (i) hand-object tracking and interaction, (ii) clustered activity zones representing specific areas of recurring activities, and (iii) environment layout mapping, allowing natural language parsing by LLMs to provide grounded insights on spatio-temporal and 3D context. In addition, we propose the Video MindPalace Benchmark (VMB), to assess human-like reasoning, including spatial localization, temporal reasoning, and layout-aware sequential understanding. Evaluated on VMB and established video QA datasets, including EgoSchema, NExT-QA, IntentQA, and the Active Memories Benchmark, VideoMindPalace demonstrates notable gains in spatio-temporal coherence and human-aligned reasoning, advancing long-form video analysis capabilities in VLMs.

Visual representation models leveraging attention mechanisms are challenged by significant computational overhead, particularly when pursuing large receptive fields. In this study, we aim to mitigate this challenge by introducing the Heat Conduction Operator (HCO) built upon the physical heat conduction principle. HCO conceptualizes image patches as heat sources and models their correlations through adaptive thermal energy diffusion, enabling robust visual representations. HCO enjoys a computational complexity of O(N^1.5), as it can be implemented using discrete cosine transformation (DCT) operations. HCO is plug-and-play, combining with deep learning backbones produces visual representation models (termed vHeat) with global receptive fields. Experiments across vision tasks demonstrate that, beyond the stronger performance, vHeat achieves up to a 3x throughput, 80% less GPU memory allocation, and 35% fewer computational FLOPs compared to the Swin-Transformer.

We present Buffer Anytime, a framework for estimation of depth and normal maps (which we call geometric buffers) from video that eliminates the need for paired video--depth and video--normal training data. Instead of relying on large-scale annotated video datasets, we demonstrate high-quality video buffer estimation by leveraging single-image priors with temporal consistency constraints. Our zero-shot training strategy combines state-of-the-art image estimation models based on optical flow smoothness through a hybrid loss function, implemented via a lightweight temporal attention architecture. Applied to leading image models like Depth Anything V2 and Marigold-E2E-FT, our approach significantly improves temporal consistency while maintaining accuracy. Experiments show that our method not only outperforms image-based approaches but also achieves results comparable to state-of-the-art video models trained on large-scale paired video datasets, despite using no such paired video data.

The development of large-scale image-text pair datasets has significantly advanced self-supervised learning in Vision-Language Processing (VLP). However, directly applying general-domain architectures such as CLIP to medical data presents challenges, particularly in handling negations and addressing the inherent data imbalance of medical datasets. To address these issues, we propose a novel approach that integrates clinically-enhanced dynamic soft labels and medical graphical alignment, thereby improving clinical comprehension and improving the applicability of contrastive loss in medical contexts. Furthermore, we introduce negation-based hard negatives to deepen the model's understanding of the complexities of clinical language. Our approach is easily integrated into medical CLIP training pipeline and achieves state-of-the-art performance across multiple tasks, including zero-shot, fine-tuned classification and report retrieval. To comprehensively evaluate our model's capacity in understanding clinical language, we introduce CXR-Align, a benchmark uniquely designed to evaluate the understanding of negation and clinical information within chest X-ray (CXR) datasets. Experimental results demonstrate that our proposed methods are straightforward to implement and generalize effectively across contrastive learning frameworks, enhancing medical VLP capabilities and advancing clinical language understanding in medical imaging.

Can our brain signals faithfully reflect the original visual stimuli, even including high-frequency details? Although human perceptual and cognitive capacities enable us to process and remember visual information, these abilities are constrained by several factors, such as limited attentional resources and finite capacity of visual memory. When visual stimuli are processed by human visual system into brain signals, some information is inevitably lost, leading to a discrepancy known as the System GAP. Additionally, perceptual and cognitive dynamics, along with technical noise in signal acquisition, degrade the fidelity of brain signals relative to the visual stimuli, known as the Random GAP. When encoded brain representations are directly aligned with the corresponding pretrained image features, the System GAP and Random GAP between paired data challenge the model, requiring it to bridge these gaps. However, due to limited paired data, these gaps are difficult for the model to learn, leading to overfitting and poor generalization to new data. To address these GAPs, we propose a simple yet effective approach called the Uncertainty-aware Blur Prior (UBP). It estimates the uncertainty within the paired data, reflecting the mismatch between brain signals and visual stimuli. Based on uncertainty, UBP dynamically blurs the high-frequency details of the original images, reducing the impact of mismatch and improving alignment. Our method achieves a top-1 accuracy of 50.9% and a top-5 accuracy of 79.7% on the zero-shot brain-to-image retrieval task, surpassing previous state-of-the-art methods. Code is available at https://github.com/HaitaoWuTJU/Uncertainty-aware-Blur-Prior.

Finding semantic correspondences between images is a challenging problem in computer vision, particularly under significant viewpoint changes. Previous methods rely on semantic features from pre-trained 2D models like Stable Diffusion and DINOv2, which often struggle to extract viewpoint-invariant features. To overcome this, we propose a novel approach that integrates geometric and semantic reasoning. Unlike prior methods relying on heuristic geometric enhancements, our framework fine-tunes DUSt3R on synthetic cross-instance data to reconstruct distinct objects in an aligned 3D space. By learning to deform these objects into similar shapes using semantic supervision, we enable efficient KNN-based geometric matching, followed by sparse semantic matching within local KNN candidates. While trained on synthetic data, our method generalizes effectively to real-world images, achieving up to 7.4-point improvements in zero-shot settings on the rigid-body subset of Spair-71K and up to 19.6-point gains under extreme viewpoint variations. Additionally, it accelerates runtime by up to 40 times, demonstrating both its robustness to viewpoint changes and its efficiency for practical applications.