ICCV 2025 论文列表

Polyp segmentation is vital for early colorectal cancer detection, yet traditional fully supervised methods struggle with morphological variability and domain shifts, requiring frequent retraining. Additionally, reliance on large-scale annotations is a major bottleneck due to the time-consuming and error-prone nature of polyp boundary labeling. Recently, vision foundation models like Segment Anything Model (SAM) have demonstrated strong generalizability and fine-grained boundary detection with sparse prompts, effectively addressing key polyp segmentation challenges. However, SAM's prompt-dependent nature limits automation in medical applications, since manually inputting prompts for each image is labor-intensive and time-consuming. We propose OP-SAM, a One-shot Polyp segmentation framework based on SAM that automatically generates prompts from a single annotated image, ensuring accurate and generalizable segmentation without additional annotation burdens. Our method introduces Correlation-based Prior Generation (CPG) for semantic label transfer and Scale-cascaded Prior Fusion (SPF) to adapt to polyp size variations as well as filter out noisy transfers. Instead of dumping all prompts at once, we devise Euclidean Prompt Evolution (EPE) for iterative prompt refinement, progressively enhancing segmentation quality. Extensive evaluations across five datasets validate OP-SAM's effectiveness. Notably, on Kvasir, it achieves 76.93% IoU, surpassing the state-of-the-art by 11.44%.

Vision-Language Pre-training (VLP) models have exhibited unprecedented capability in many applications by taking full advantage of the learned multimodal alignment. However, previous studies have shown they are vulnerable to maliciously crafted adversarial samples. Despite recent success, these attacks are generally instance-specific and require generating perturbations for each input sample. In this paper, we reveal that VLP models are also susceptible to the instance-agnostic universal adversarial perturbation (UAP). Specifically, we design a novel Contrastive-training Perturbation Generator with Cross-modal conditions (C-PGC). In light that the pivotal multimodal alignment in VLP models is achieved via contrastive learning, we devise to turn this powerful weapon against VLP models themselves. I.e., we employ a malicious version of contrastive learning to train the proposed generator using our carefully crafted positive and negative image-text pairs. Once training is complete, the generator is able to produce universal perturbations that can essentially destroy the established alignment relationship in VLP models. Besides, C-PGC fully utilizes the characteristics of Vision-and-Language (V+L) scenarios by incorporating both unimodal and cross-modal information as effective guidance. Extensive experiments show that C-PGC successfully forces adversarial samples to move away from their original area in the VLP model's feature space, thus fundamentally enhancing attack performance across various victim models and V+L tasks.

Unified vision-language models (VLMs) have recently shown remarkable progress, enabling a single model to flexibly address diverse tasks through different instructions within a shared computational architecture. This instruction-based control mechanism creates unique security challenges, as adversarial inputs must remain effective across multiple task instructions that may be unpredictably applied to process the same malicious content. In this paper, we introduce CrossVLAD, a new benchmark dataset carefully curated from MSCOCO with GPT-4-assisted annotations for systematically evaluating cross-task adversarial attacks on unified VLMs. CrossVLAD centers on the object-change objective--consistently manipulating a target object's classification across four downstream tasks--and proposes a novel success rate metric that measures simultaneous misclassification across all tasks, providing a rigorous evaluation of adversarial transferability. To tackle this challenge, we present CRAFT (Cross-task Region-based Attack Framework with Token-alignment), an efficient region-centric attack method. Extensive experiments on Florence-2 and other popular unified VLMs demonstrate that our method outperforms existing approaches in both overall cross-task attack performance and targeted object-change success rates, highlighting its effectiveness in adversarially influencing unified VLMs across diverse tasks.

Zero-shot depth estimation (DE) models exhibit strong generalization performance as they are trained on large-scale datasets. However, existing models struggle with high-resolution images due to the discrepancy in image resolutions of training (with smaller resolutions) and inference (for high resolutions). Processing them at full resolution leads to decreased estimation accuracy on depth with tremendous memory consumption, while downsampling to the training resolution results in blurred edges in the estimated depth images. Prevailing high-resolution depth estimation methods adopt a patch-based approach, which introduces depth discontinuity issues when reassembling the estimated depth patches, resulting in test-time inefficiency. Additionally, to obtain fine-grained depth details, these methods rely on synthetic datasets due to the real-world sparse ground truth depth, leading to poor generalizability. To tackle these limitations, we propose Patch Refine Once (PRO), an efficient and generalizable tile-based framework. Our PRO consists of two key components: (i) Grouped Patch Consistency Training that enhances test-time efficiency while mitigating the depth discontinuity problem by jointly processing four overlapping patches and enforcing a consistency loss on their overlapping regions within a single backpropagation step, and (ii) Bias Free Masking that prevents the DE models from overfitting to dataset-specific biases, enabling better generalization to real-world datasets even after training on synthetic data. Zero-shot evaluations on Booster, ETH3D, Middlebury 2014, and NuScenes demonstrate that our PRO can be seamlessly integrated into existing depth estimation models. It preserves the performance of original depth estimation models even under grid-based inference on high-resolution images, exhibiting minimal depth discontinuities along patch boundaries. Moreover, our PRO achieves significantly faster inference speed compared to prior patch-based methods.

Representation learning lies at the core of deep reinforcement learning. Although CNNs have traditionally served as the primary models for encoding image observations, modifying the encoder architecture introduces challenges, especially due to the necessity of determining a new set of hyperparameters. To address this problem, we propose a powerful representation learning technique for visual reinforcement learning utilizing Fourier Neural Operators (FNO). Our findings demonstrate that the proposed FNO encoder effectively learns representations from images that encapsulate the underlying differential equations (PDEs) governing the dynamics of the environment in an online model-free RL framework. We demonstrate the applicability of our proposed architecture by replacing the CNN image encoder in PPO, A2C, and Rainbow (a Policy Gradient, Actor-Critic, and Q-Learning RL algorithm, respectively). We achieve state-of-the-art scores (in the model-free RL setting) at both the CARLA Autonomous Driving (from image observations) benchmark and the Atari 100k benchmark. Our proposed FNO encoder is compatible with all model-free reinforcement learning algorithms, enhances both rewards and sample efficiency by implicitly learning the underlying dynamics of the environment, and eliminates the need for additional hyperparameter tuning.

Diffusion Policies have significantly advanced robotic manipulation tasks via imitation learning, but their application on resource-constrained mobile platforms remains challenging due to computational inefficiency and extensive memory footprint. In this paper, we propose LightDP, a novel framework specifically designed to accelerate Diffusion Policies for real-time deployment on mobile devices. LightDP addresses the computational bottleneck through two core strategies: network compression of the denoising modules and reduction of the required sampling steps. We first conduct an extensive computational analysis on existing Diffusion Policy architectures, identifying the denoising network as the primary contributor to latency. To overcome performance degradation typically associated with conventional pruning methods, we introduce a unified pruning and retraining pipeline, optimizing the model's post-pruning recoverability explicitly. Furthermore, we combine pruning techniques with consistency distillation to effectively reduce sampling steps while maintaining action prediction accuracy. Experimental evaluations on the standard datasets, i.e., PushT, Robomimic, CALVIN, and LIBERO, demonstrate that LightDP achieves real-time action prediction on mobile devices with competitive performance, marking an important step toward practical deployment of diffusion-based policies in resource-limited environments. Extensive real-world experiments also show the proposed LightDP can achieve performance comparable to state-of-the-art Diffusion Policies.

Fine-tuning has become the standard practice for adapting pre-trained models to downstream tasks. However, the impact on model robustness is not well understood. In this work, we characterize the robustness-accuracy trade-off in fine-tuning. We evaluate the robustness and accuracy of fine-tuned models over 6 benchmark datasets and 7 different fine-tuning strategies. We observe a consistent trade-off between adversarial robustness and accuracy. Peripheral updates such as BitFit are more effective for simple tasks---over 75% above the average measured by the area under the Pareto frontiers on CIFAR-10 and CIFAR-100. In contrast, fine-tuning information-heavy layers, such as attention layers via Compacter, achieves a better Pareto frontier on more complex tasks---57.5% and 34.6% above the average on Caltech-256 and CUB-200, respectively. Lastly, we observe that the robustness of fine-tuning against out-of-distribution data closely tracks accuracy. These insights emphasize the need for robustness-aware fine-tuning to ensure reliable real-world deployments.

The viewing graph is a compact tool to encode the geometry of multiple views: nodes represent uncalibrated cameras and edges represent fundamental matrices (when available). Most research focuses on theoretical analyses, exploring for which viewing graphs it is possible (in principle) to retrieve cameras from fundamental matrices, in the sense that the problem admits a unique solution for noiseless data. However, the practical task of recovering cameras from noisy fundamental matrices is still open, as available methods are limited to special graphs (such as those covered by triplets). In this paper, we develop the first method that can deal with the recovery of cameras from noisy fundamental matrices in a general viewing graph. Experimental results demonstrate the promise of the proposed approach on a variety of synthetic and real scenarios.

This paper investigates the recently emerged problem of Language-assisted Image Clustering (LaIC), where textual semantics are leveraged to improve the discriminability of visual representations to facilitate image clustering. Due to the unavailability of true class names, one of core challenges of LaIC lies in how to filter positive nouns, i.e., those semantically close to the images of interest, from unlabeled wild corpus data. Existing filtering strategies are predominantly based on the off-the-shelf feature space learned by CLIP; however, despite being intuitive, these strategies lack a rigorous theoretical foundation. To fill this gap, we propose a novel gradient-based framework, termed as GradNorm, which is theoretically guaranteed and shows strong empirical performance. In particular, we measure the positiveness of each noun based on the magnitude of gradients back-propagated from the cross-entropy between the predicted target distribution and the softmax output. Theoretically, we provide a rigorous error bound to quantify the separability of positive nouns by GradNorm and prove that GradNorm naturally subsumes existing filtering strategies as extremely special cases of itself. Empirically, extensive experiments show that GradNorm achieves the state-of-the-art clustering performance on various benchmarks.

Recent advances in Computer Vision have introduced the concept of pretrained representation uncertainty, enabling zero-shot uncertainty estimation. This holds significant potential for Earth Observation (EO), where trustworthiness is critical, yet the complexity of EO data poses challenges to uncertainty-aware methods. In this work, we investigate the generalization of representation uncertainty in EO, considering the domain's unique semantic characteristics. We pretrain uncertainties on large EO datasets and propose an evaluation framework to assess their zero-shot performance in multi-label classification and segmentation EO tasks. Our findings reveal that, unlike uncertainties pretrained on natural images, EO-pretraining exhibits strong generalization across unseen EO domains, geographic locations, and target granularities, while maintaining sensitivity to variations in ground sampling distance. We demonstrate the practical utility of pretrained uncertainties showcasing their alignment with task-specific uncertainties in downstream tasks, their sensitivity to real-world EO image noise, and their ability to generate spatial uncertainty estimates out-of-the-box. Initiating the discussion on representation uncertainty in EO, our study provides insights into its strengths and limitations, paving the way for future research in the field. Code and weights are available at: https://github.com/Orion-AI-Lab/EOUncertaintyGeneralization.

ReLU networks, while prevalent for visual data, have sharp transitions, sometimes relying on individual pixels for predictions, making vanilla gradient-based explanations noisy and difficult to interpret. Existing methods, such as GradCAM, smooth these explanations by producing surrogate models at the cost of faithfulness. We introduce a unifying spectral framework to systematically analyze and quantify smoothness, faithfulness, and their trade-off in explanations.Using this framework, we quantify and regularize the contribution of ReLU networks to high-frequency information, providing a principled approach to identifying this trade-off. Our analysis characterizes how surrogate-based smoothing distorts explanations, leading to an "explanation gap" that we formally define and measure for different post-hoc methods.Finally, we validate our theoretical findings across different design choices, datasets, and ablations.

Traditional image classification requires a predefined list of semantic categories. In contrast, Large Multimodal Models (LMMs) can sidestep this requirement by classifying images directly using natural language (e.g., answering the prompt "What is the main object in the image?"). Despite this remarkable capability, most existing studies on LMM classification performance are surprisingly limited in scope, often assuming a closed-world setting with a predefined set of categories. In this work, we address this gap by thoroughly evaluating LMM classification performance in a truly open-world setting. We first formalize the task and introduce an evaluation protocol, defining various metrics to assess the alignment between predicted and ground truth classes. We then evaluate 13 models across 10 benchmarks, encompassing prototypical, non-prototypical, fine-grained, and very fine-grained classes, demonstrating the challenges LMMs face in this task. Further analyses based on the proposed metrics reveal the types of errors LMMs make, highlighting challenges related to granularity and fine-grained capabilities, showing how tailored prompting and reasoning can alleviate them.

Image-based Virtual Try-On (VTON) techniques rely on either supervised in-shop approaches, which ensure high fidelity but struggle with cross-domain generalization, or unsupervised in-the-wild methods, which improve adaptability but remain constrained by data biases and limited universality. A unified, training-free solution that works across both scenarios remains an open challenge. We propose OmniVTON, the first training-free universal VTON framework that decouples garment and pose conditioning to achieve both texture fidelity and pose consistency across diverse settings. To preserve garment details, we introduce a garment prior generation mechanism that aligns clothing with the body, followed by continuous boundary stitching technique to achieve fine-grained texture retention. For precise pose alignment, we utilize DDIM inversion to capture structural cues while suppressing texture interference, ensuring accurate body alignment independent of the original image textures. By disentangling garment and pose constraints, OmniVTON eliminates the bias inherent in diffusion models when handling multiple conditions simultaneously. Experimental results demonstrate that OmniVTON achieves superior performance across diverse datasets, garment types, and application scenarios. Notably, it is the first framework capable of multi-human VTON, enabling realistic garment transfer across multiple individuals in a single scene. Code is available at https://github.com/Jerome-Young/OmniVTON.

Segment Anything Model 2 (SAM2) has emerged as a strong base model in various pinhole imaging segmentation tasks. However, when applying it to 360^\circ domain, the significant field-of-view (FoV) gap between pinhole (70^\circ x70^\circ) and panoramic images (180^\circ x360^\circ) poses unique challenges. Two major concerns for this application includes 1) inevitable distortion and object deformation brought by the large FoV disparity between domains; 2) the lack of pixel-level semantic understanding that the original SAM2 cannot provide. To address these issues, we propose a novel OmniSAM framework, which makes the first attempt to apply SAM2 for panoramic semantic segmentation. Specifically, to bridge the first gap, OmniSAM first divides the panorama into sequences of patches. These patches are then treated as image sequences in similar manners as in video segmentation tasks. We then leverage the SAM2's memory mechanism to extract cross-patch correspondences that embeds the cross-FoV dependencies, improving feature continuity and the prediction consistency along mask boundaries. For the second gap, OmniSAM fine-tunes the pretrained image encoder and reutilize the mask decoder for semantic prediction. An FoV-based prototypical adaptation module with dynamic pseudo label update mechanism is also introduced to facilitate the alignment of memory and backbone features, thereby improving model generalization ability across different sizes of source models. Extensive experimental results demonstrate that our method outperforms the state-of-the-art methods by large margins, e.g., 79.06% (10.22%\uparrow) on SPin8-to-SPan8, 62.46% (6.58%\uparrow) on CS13-to-DP13.

Diffusion-based generative models have revolutionized object-oriented image editing, yet their deployment in realistic object removal and insertion remains hampered by challenges such as the intricate interplay of physical effects and insufficient paired training data. In this work, we introduce OmniPaint, a unified framework that re-conceptualizes object removal and insertion as interdependent processes rather than isolated tasks. Leveraging a pre-trained diffusion prior along with a progressive training pipeline comprising initial paired sample optimization and subsequent large-scale unpaired refinement via CycleFlow, OmniPaint achieves precise foreground elimination and seamless object insertion while faithfully preserving scene geometry and intrinsic properties. Furthermore, our novel CFD metric offers a robust, reference-free evaluation of context consistency and object hallucination, establishing a new benchmark for high-fidelity image editing.

End-to-end human animation, such as audio-driven talking human generation, has undergone notable advancements in the recent few years. However, existing methods still struggle to scale up as large general video generation models, limiting their potential in real applications. In this paper, we propose OmniHuman, a Diffusion Transformer-based framework that scales up data by mixing motion-related conditions into the training phase. To this end, we introduce two training principles for these mixed conditions, along with the corresponding model architecture and inference strategy. These designs enable OmniHuman to fully leverage data-driven motion generation, ultimately achieving highly realistic human video generation. More importantly, OmniHuman supports various portrait contents (face close-up, portrait, half-body, full-body), supports both talking and singing, handles human-object interactions and challenging body poses, and accommodates different image styles. Compared to existing end-to-end audio-driven methods, OmniHuman not only produces more realistic videos, but also offers greater flexibility in inputs. It also supports multiple driving modalities (audio-driven, video-driven and combined driving signals).

Image Difference Captioning (IDC) aims to generate natural language descriptions of subtle differences between image pairs, requiring both precise visual change localization and coherent semantic expression. Despite recent advancements, existing datasets often lack breadth and depth, limiting their applicability in complex and dynamic environments: (1) from a breadth perspective, current datasets are constrained to limited variations of objects in specific scenes, and (2) from a depth perspective, prior benchmarks often provide overly simplistic descriptions. To address these challenges, we introduce OmniDiff, a comprehensive dataset comprising 324 diverse scenarios--spanning real-world complex environments and 3D synthetic settings--with fine-grained human annotations averaging 60 words in length and covering 12 distinct change types. Building on this foundation, we propose M3Diff, a Multi Modal large language model enhanced by a plug-and-play Multi-scale Differential Perception (MDP) module. This module improves the model's ability to accurately identify and describe inter-image differences while maintaining the foundational model's generalization capabilities. With the addition of the OmniDiff dataset, M3Diff achieves state-of-the-art performance across multiple benchmarks, including Spot-the-Diff, IEdit, CLEVR-Change, CLEVR-DC, and OmniDiff, demonstrating significant improvements in cross-scenario difference recognition accuracy compared to existing methods. The dataset, code, and models will be made publicly available to support further research.

Diffusion models have emerged as a powerful paradigm for generative tasks such as image synthesis and video generation, with Transformer architectures further enhancing performance. However, the high computational cost of diffusion Transformers--stemming from a large number of sampling steps and complex per-step computations--presents significant challenges for real-time deployment. In this paper, we introduce OmniCache, a training-free acceleration method that exploits the global redundancy inherent in the denoising process. Unlike existing methods that determine caching strategies based on inter-step similarities and tend to prioritize reusing later sampling steps, our approach originates from the sampling perspective of DIT models. We systematically analyze the model's sampling trajectories and strategically distribute cache reuse across the entire sampling process. This global perspective enables more effective utilization of cached computations throughout the diffusion trajectory, rather than concentrating reuse within limited segments of the sampling procedure. In addition, during cache reuse, we dynamically estimate the corresponding noise and filter it out to reduce its impact on the sampling direction. Extensive experiments demonstrate that our approach accelerates the sampling process while maintaining competitive generative quality, offering a promising and practical solution for efficient deployment of diffusion-based generative models.