CVPR 2025 论文列表

Existing 3D vision-language (3D-VL) benchmarks fall short in evaluating 3D-VL models, creating a "mist" that obscures rigorous insights into model capabilities and 3D-VL tasks. This mist persists due to three key limitations. First, flawed test data, like ambiguous referential text in the grounding task, can yield incorrect and unreliable test results. Second, oversimplified metrics such as simply averaging accuracy per question answering (QA) pair, cannot reveal true model capability due to their vulnerability to language variations. Third, existing benchmarks isolate the grounding and QA tasks, disregarding the underlying coherence that QA should be based on solid grounding capabilities. To unveil the "mist", we propose Beacon3D, a benchmark for 3D-VL grounding and QA tasks, delivering a perspective shift in the evaluation of 3D-VL understanding. Beacon3D features (i) high-quality test data with precise and natural language, (ii) object-centric evaluation with multiple tests per object to ensure robustness, and (iii) a novel chain-of-analysis paradigm to address language robustness and model performance coherence across grounding and QA. Our evaluation of state-of-the-art 3D-VL models on Beacon3D reveals that (i) object-centric evaluation elicits true model performance and particularly weak generalization in QA; (ii) grounding-QA coherence remains fragile in current 3D-VL models, and (iii) incorporating large language models (LLMs) to 3D-VL models, though as a prevalent practice, hinders grounding capabilities and has yet to elevate QA capabilities. We hope Beacon3D and our comprehensive analysis could benefit the 3D-VL community towards faithful developments.

**M**ultimodal **L**arge **L**anguage **M**odels (MLLMs) have displayed remarkable performance in multimodal tasks, particularly in visual comprehension. However, we reveal that MLLMs often generate incorrect answers even when they understand the visual content. To this end, we manually construct a benchmark with 12 categories and design evaluation metrics that assess the degree of error in MLLM responses even when the visual content is seemingly understood. Based on this benchmark, we test 15 leading MLLMs and analyze the distribution of attention maps and logits of some MLLMs. Our investigation identifies two primary issues: 1) most instruction tuning datasets predominantly feature questions that "directly" relate to the visual content, leading to a bias in MLLMs' responses to other indirect questions, and 2) MLLMs' attention to visual tokens is notably lower than to system and question tokens. We further observe that attention scores between questions and visual tokens as well as the model's confidence in the answers are lower in response to misleading questions than to straightforward ones. To address the first challenge, we introduce a paired positive and negative data construction pipeline to diversify the dataset. For the second challenge, we propose to enhance the model's focus on visual content during decoding by refining the text and visual prompt. For the text prompt, we propose a content-guided refinement strategy that performs preliminary visual content analysis to generate structured information before answering the question. Additionally, we employ a visual attention refinement strategy that highlights question-relevant visual tokens to increase the model's attention to visual content that aligns with the question. Extensive experiments demonstrate that these challenges can be significantly mitigated with our proposed dataset and techniques. The benchmark, training set, and code will be available.

Leveraging the large generative prior of the flow transformer for tuning-free image editing requires authentic inversion to project the image into the model's domain and a flexible invariance control mechanism to preserve non-target contents. However, the prevailing diffusion inversion performs deficiently in flow-based models, and the invariance control cannot reconcile diverse rigid and non-rigid editing tasks. To address these, we systematically analyze the inversion and invariance control based on the flow transformer. Specifically, we unveil that the Euler inversion shares a similar structure to DDIM yet is more susceptible to the approximation error. Thus, we propose a two-stage inversion to first refine the velocity estimation and then compensate for the leftover error, which pivots closely to the model prior and benefits editing. Meanwhile, we propose the invariance control that manipulates the text features within the adaptive layer normalization, connecting the changes in the text prompt to image semantics. This mechanism can simultaneously preserve the non-target contents while allowing rigid and non-rigid manipulation, enabling a wide range of editing types. Experiments on various scenarios demonstrate that our framework achieves flexible and accurate editing, unlocking the potential of the flow transformer for versatile image editing.

Representation learning of pathology whole-slide images (WSIs) has primarily relied on weak supervision with Multiple Instance Learning (MIL). This approach leads to slide representations highly tailored to a specific clinical task. Self-supervised learning (SSL) has been successfully applied to train histopathology foundation models (FMs) for patch embedding generation. However, generating patient or slide level embeddings remains challenging. Existing approaches for slide representation learning extend the principles of SSL from patch level learning to entire slides by aligning different augmentations of the slide or by utilizing multimodal data. By integrating tile embeddings from multiple FMs, we propose a new single modality SSL method in feature space that generates useful slide representations. Our contrastive pretraining strategy, called COBRA, employs multiple FMs and an architecture based on Mamba-2. COBRA exceeds performance of state-of-the-art slide encoders on four different public Clinical Protemic Tumor Analysis Consortium (CPTAC) cohorts on average by at least +4.4% AUC, despite only being pretrained on 3048 WSIs from The Cancer Genome Atlas (TCGA). Additionally, COBRA is readily compatible at inference time with previously unseen feature extractors. Code available at https://github.com/KatherLab/COBRA

Realistic simulation of dynamic scenes requires accurately capturing diverse material properties and modeling complex object interactions grounded in physical principles. However, existing methods are constrained to basic material types with limited predictable parameters, making them insufficient to represent the complexity of real-world materials. We introduce PhysFlow, a novel approach that leverages multi-modal foundation models and video diffusion to achieve enhanced 4D dynamic scene simulation. Our method utilizes multi-modal models to identify material types and initialize material parameters through image queries, while simultaneously inferring 3D Gaussian splats for detailed scene representation. We further refine these material parameters using video diffusion with a differentiable Material Point Method (MPM) and optical flow guidance rather than render loss or Score Distillation Sampling (SDS) loss. This integrated framework enables accurate prediction and realistic simulation of dynamic interactions in real-world scenarios, advancing both accuracy and flexibility in physics-based simulations.

Human pose plays a crucial role in the digital age. While recent works have achieved impressive progress in understanding and generating human poses, they often support only a single modality of control signals and operate in isolation, limiting their application in real-world scenarios. This paper presents UniPose, a framework employing Large Language Models (LLMs) to comprehend, generate, and edit human poses across various modalities, including images, text, and 3D SMPL poses. Specifically, we apply a pose tokenizer to convert 3D poses into discrete pose tokens, enabling seamless integration into the LLM within a unified vocabulary. To further enhance the fine-grained pose perception capabilities, we facilitate UniPose with a mixture of visual encoders, among them a pose-specific visual encoder. Benefiting from a unified learning strategy, UniPose effectively transfers knowledge across different pose-relevant tasks, adapts to unseen tasks, and exhibits extended capabilities. This work serves as the first attempt at building a general-purpose framework for pose comprehension, generation, and editing. Extensive experiments highlight UniPose's competitive and even superior performance across various pose-relevant tasks. The code is available at https://github.com/liyiheng23/UniPose.

This paper presents a unified approach to understanding dynamic scenes from casual videos. Large pretrained vision foundation models, such as vision-language, video depth prediction, motion tracking, and segmentation models, offer promising capabilities. However, training a single model for comprehensive 4D understanding remains challenging. We introduce Uni4D, a multi-stage optimization framework that harnesses multiple pretrained models to advance dynamic 3D modeling, including static/dynamic reconstruction, camera pose estimation, and dense 3D motion tracking. Our results show state-of-the-art performance in dynamic 4D modeling with superior visual quality. Notably, Uni4D requires no retraining or fine-tuning, highlighting the effectiveness of repurposing visual foundation models for 4D understanding. Code and more results are available at: https://davidyao99.github.io/uni4d.

CLIP, a foundational vision-language model, has emerged as a powerful tool for open-vocabulary semantic segmentation. While freezing the text encoder preserves its powerful embeddings, recent studies show that fine-tuning both the text and image encoders jointly significantly enhances segmentation performance, especially for classes from open sets. In this work, we explain this phenomenon from the perspective of hierarchical alignment, since during fine-tuning, the hierarchy level of image embeddings shifts from image-level to pixel-level. We achieve this by leveraging hyperbolic space, which naturally encoders hierarchical structures. Our key observation is that, during fine-tuning, the hyperbolic radius of CLIP's text embeddings decreases, facilitating better alignment with the pixel-level hierarchical structure of visual data. Building on this insight, we propose HyperCLIP, a novel fine-tuning strategy that adjusts the hyperbolic radius of the text embeddings through scaling transformations. By doing so, HyperCLIP equips CLIP with segmentation capability while introducing only a small number of learnable parameters. Our experiments demonstrate that HyperCLIP achieves state-of-the-art performance on open-vocabulary semantic segmentation tasks across three benchmarks, while fine-tuning only approximately 4% of the total parameters of CLIP. More importantly, we observe that after adjustment, CLIP's text embeddings exhibit a relatively fixed hyperbolic radius across datasets, suggesting that the segmentation task has a characteristic level in hyperbolic space.

Model calibration is essential for ensuring that the predictions of deep neural networks accurately reflect true probabilities in real-world classification tasks. However, deep networks often produce over-confident or under-confident predictions, leading to miscalibration. Various methods have been proposed to address this issue by designing effective loss functions for calibration, such as focal loss. In this paper, we analyze its effectiveness and provide a unified loss framework of focal loss and its variants, where we mainly attribute their superiority in model calibration to the loss weighting factor that estimates sample-wise uncertainty. Based on our analysis, existing loss functions fail to achieve optimal calibration performance due to two main issues: including misalignment during optimization and insufficient precision in uncertainty estimation. Specifically, focal loss cannot align sample uncertainty with gradient scaling and the single logit cannot indicate the uncertainty. To address these issues, we reformulate the optimization from the perspective of gradients, which focuses on uncertain samples. Meanwhile, we propose using the Brier Score as the loss weight factor, which provides a more accurate uncertainty estimation via all the logits. Extensive experiments on various models and datasets demonstrate that our method achieves state-of-the-art (SOTA) performance.

Understanding intention and emotion from social media poses unique challenges due to the inherent uncertainty in multimodal data, where posts often contain incomplete or missing modalities. While this uncertainty reflects real-world scenarios, it remains underexplored within the computer vision community, particularly in conjunction with the intrinsic relationship between emotion and intention. To address these challenges, we introduce the Multimodal IntentioN and Emotion Understanding in the Wild (MINE) dataset, comprising over 20,000 topic-specific social media posts with natural modality variations across text, image, video, and audio. MINE is distinctively constructed to capture both the uncertain nature of multimodal data and the implicit correlations between intentions and emotions, providing extensive annotations for both aspects. To tackle these scenarios, we propose the Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework. BEAR consists of two key components: a BEIFormer that leverages emotion-intention correlations, and a Modality Asynchronous Prompt that handles modality uncertainty. Experiments show that BEAR outperforms existing methods in processing uncertain multimodal data while effectively mining emotion-intention relationships for social media content understanding.

Multimodal Large Language Models (MLLMs) have emerged to tackle the challenges of Visual Question Answering (VQA), sparking a new research focus on conducting objective evaluations of these models. Existing evaluation mechanisms face limitations due to the significant human workload required to design Q&A pairs for visual images, which inherently restricts the scale and scope of evaluations. Although automated MLLM-as-judge approaches attempt to reduce the human workload through mutual model evaluations, they often introduce biases. To address these problems, we propose an unsupervised evaluation method -- an Unsupervised Peer review MLLM Evaluation framework. This framework utilizes only image data, allowing models to automatically generate questions and conduct peer review assessments of answers from other models, effectively alleviating the reliance on human workload. Additionally, we introduce the vision-language scoring system to mitigate the bias issues, which focuses on three aspects: (i) response correctness; (ii) the model capability of visual understanding and reasoning; (iii) relevance of text-image matching. Experimental results demonstrate that UPME achieves a Pearson correlation of 0.944 with human evaluations on the MMstar dataset and 0.814 on the ScienceQA dataset, indicating that our UPME framework closely aligns with human-designed QA benchmarks and inherent human preferences.

We present UNIALIGN, a unified model to align an arbitrary number of modalities (\text e.g. , image, text, audio, 3D point cloud, etc.) through one encoder and a single training phase. Existing solutions typically employ distinct encoders for each modality, resulting in increased parameters as the number of modalities grows. In contrast, UNIALIGN proposes a modality-aware adaptation of the powerful mixture-of-experts (MoE) schema and further integrates it with Low-Rank Adaptation (LoRA), efficiently scaling the encoder to accommodate inputs in diverse modalities while maintaining a fixed computational overhead.Moreover, prior work often requires separate training for each extended modality. This leads to task-specific models and further hinders the communication between modalities.To address this, we propose a soft modality binding strategy that aligns all modalities using unpaired data samples across datasets. Two additional training objectives are introduced to distill knowledge from well-aligned anchor modalities and prior multimodal models, elevating UNIALIGN into a high performance multimodal foundation model.Experiments on 11 benchmarks across 6 different modalities demonstrate that UNIALIGN could achieve comparable performance to SOTA approaches, while using merely 7.8M trainable parameters and maintaining an identical model with the same weight across all tasks. Our code shall be released.

Transductive few-shot learning has recently triggered wide attention in computer vision. Yet, current methods introduce key hyper-parameters, which control the pre-diction statistics of the test batches, such as the level of class balance, affecting performances significantly. Such hyper-parameters are empirically grid-searched over validation data, and their configurations may vary substantially with the target dataset and pre-training model, making such empirical searches both sub-optimal and computationally intractable. In this work, we advocate and introduce the unrolling paradigm, also referred to as "learning to optimize", in the context of few-shot learning, thereby learning efficiently and effectively a set of optimized hyperparameters. Specifically, we unroll a generalization of the ubiquitous Expectation-Maximization (EM) optimizer into a neural network architecture, mapping each of its iterates to a layer and learning a set of key hyper-parameters over validation data. Our unrolling approach covers various statistical feature distributions and pre-training paradigms, including recent foundational vision-language models and standard vision-only classifiers. We report comprehensive experiments, which cover a breadth of fine-grained downstream image classification tasks, showing significant gains brought by the proposed unrolled EM algorithm over iterative variants. The achieved improvements reach up to 10% and 7.5% on vision-only and vision-language benchmarks, respectively. The source code and learned parameters are available at https://github.com/ZhouLong0/UNEM-Transductive.

We introduce UHD-Processor, a unified and robust framework for all-in-one image restoration, which is particularly resource-efficient for Ultra-High-Definition (UHD) images. To address the limitations of traditional all-in-one methods that rely on complex restoration backbones, our strategy employs a frequency domain decoupling progressive learning technique, motivated by curriculum learning, to incrementally learn restoration mappings from low to high frequencies. This approach incorporates specialized sub-network modules to effectively tackle different frequency bands in a divide-and-conquer manner, significantly enhancing the learning capability of simpler networks. Moreover, to accommodate the high-resolution characteristics of UHD images, we developed a variational autoencoder (VAE)-based framework that reduces computational complexity by modeling a concise latent space. It integrates task-specific degradation awareness in the encoder and frequency selection in the decoder, enhancing task comprehension and generalization. Our unified model is able to handle various degradations such as denoising, deblurring, dehazing, low-lighting, etc. Experimental evaluations extensively showcase the effectiveness of our dual-strategy approach, significantly improving UHD image restoration and achieving cutting-edge performance across diverse conditions. The code will be available at https://github.com/lyd-2022/UHD-processer.

Cross-Modality Re-Identification (VI-ReID) aims to achieve around-the-clock target matching, benefiting from the strengths of both RGB and infrared (IR) modalities. However, the field is hindered by limited datasets, particularly for vehicle VI-ReID, and by challenges such as modality bias training (MBT), stemming from biased pre-training on ImageNet. To tackle the above issues, this paper introduces an UCM-VeID V2 dataset benchmark for vehicle VI-ReID, and proposes a new self-supervised pre-training method, Cross-Modality Patch-Mixed Self-supervised Learning (PMSL). UCM-VeID V2 dataset features a significant increase in data volume, along with enhancements in multiple aspects. PMSL addresses MBT by learning modality-invariant features through Patch-Mixed Image Reconstruction (PMIR) and Modality Discrimination Adversarial Learning (MDAL), and enhances discriminability with Modality-Augmented Contrasting Cluster (MACC). Comprehensive experiments are carried out to validate the effectiveness of the proposed method.

Deep learning-based computer vision systems adopt complex and large architectures to improve performance, yet they face challenges in deployment on resource-constrained mobile and edge devices. To address this issue, model compression techniques such as pruning, quantization, and matrix factorization have been proposed; however, these compressed models are often highly vulnerable to adversarial attacks. We introduce the Efficient Ensemble Defense (EED) technique, which diversifies the compression of a single base model based on different pruning importance scores and enhances ensemble diversity to achieve high adversarial robustness and resource efficiency. EED dynamically determines the number of necessary sub-models during the inference stage, minimizing unnecessary computations while maintaining high robustness. On the CIFAR-10 and SVHN datasets, EED demonstrated state-of-the-art robustness performance compared to existing adversarial pruning techniques, along with an inference speed improvement of up to 1.86 times. This proves that EED is a powerful defense solution in resource-constrained environments.

Conditional image synthesis is a crucial task with broad applications, such as artistic creation and virtual reality. However, current generative methods are often task-oriented with a narrow scope, handling a restricted condition with constrained applicability. In this paper, we propose a novel approach that treats conditional image synthesis as the modular combination of diverse fundamental condition units. Specifically, we divide conditions into three primary units: text, layout, and drag. To enable effective control over these conditions, we design a dedicated alignment module for each. For the text condition, we introduce a Dense Concept Alignment (DCA) module, which achieves dense visual-text alignment by drawing on diverse textual concepts. For the layout condition, we propose a Dense Geometry Alignment (DGA) module to enforce comprehensive geometric constraints that preserve the spatial configuration. For the drag condition, we introduce a Dense Motion Alignment (DMA) module to apply multi-level motion regularization, ensuring that each pixel follows its desired trajectory without visual artifacts. By flexibly inserting and combining these alignment modules, our framework enhances the model's adaptability to diverse conditional generation tasks and greatly expands its application range. Extensive experiments demonstrate the superior performance of our framework across a variety of conditions, including textual description, segmentation mask (bounding box), drag manipulation, and their combinations. Code is available at https://github.com/ZixuanWang0525/DADG