ICCV 2025 论文列表

Parameter-level model merging is an emerging paradigm in multi-task learning with significant promise. Previous research has explored its connections with prediction-level model ensembling--commonly viewed as the upper bound for merging--to reveal the potential of achieving performance consistency between the two. However, this observation relies on certain preconditions, such as being limited to two models, using ViT-based models, and all models are fine-tuned from the same pre-trained checkpoint. To further understand the intrinsic connections between these two paradigms, this paper explores an interesting possibility: If these restrictions are removed, can performance consistency still be achieved between merging and ensembling? To answer this question, we first theoretically establish a performance correlation between merging and ensembling. We find that even when previous restrictions are not met, there is still a way for model merging to attain a near-identical and superior performance similar to that of ensembling. To verify whether our findings are practical, we introduce a validation framework termed \underline Neu ral \underline Lig and (NeuLig). The learning process of NeuLig is meticulously designed with a specialized loss function supported by theoretical foundations. Experimental results demonstrate the robust resilience of NeuLig in terms of both model scale and the number of collaborating models. For instance, for the case involving 5 CLIP-ViT-B/32 models, parameter-level merging achieves the same performance as prediction-level ensembling (merging: 95.44% vs. ensembling: 95.46%).

Stereo images are fundamental to numerous applications, including extended reality (XR) devices, autonomous driving, and robotics. Unfortunately, acquiring high-quality stereo images remains challenging due to the precise calibration requirements of dual-camera setups and the complexity of obtaining accurate, dense disparity maps. Existing stereo image generation methods typically focus on either visual quality for viewing or geometric accuracy for matching, but not both. We introduce GenStereo, a diffusion-based approach, to bridge this gap. The method includes two primary innovations (1) conditioning the diffusion process on a disparity-aware coordinate embedding and a warped input image, allowing for more precise stereo alignment than previous methods, and (2) an adaptive fusion mechanism that intelligently combines the diffusion-generated image with a warped image, improving both realism and disparity consistency. Through extensive training on 11 diverse stereo datasets, GenStereo demonstrates strong generalization ability. GenStereo achieves state-of-the-art performance in both stereo image generation and unsupervised stereo matching tasks. Project page is available at https://qjizhi.github.io/genstereo.

Referring audio-visual segmentation (RAVS) has recently seen significant advancements, yet challenges remain in integrating multimodal information and deeply understanding and reasoning about audiovisual content. To extend the boundaries of RAVS and facilitate future research in this field, we propose Omnimodal Referring Audio-Visual Segmentation (OmniAVS), a new dataset containing 2,104 videos and 61,095 multimodal referring expressions. OmniAVS stands out with three key innovations: (1) 8 types of multimodal expressions that flexibly combine text, speech, sound, and visual cues; (2) an emphasis on understanding audio content beyond just detecting their presence; and (3) the inclusion of complex reasoning and world knowledge in expressions. Furthermore, we introduce Omnimodal Instructed Segmentation Assistant (OISA), to address the challenges of multimodal reasoning and fine-grained understanding of audiovisual content in OmniAVS. OISA uses MLLM to comprehend complex cues and perform reasoning-based segmentation. Extensive experiments show that OISA outperforms existing methods on OmniAVS and achieves competitive results on other related tasks.

Point cloud learning, especially in a self-supervised way without manual labels, has gained growing attention in both vision and learning communities due to its potential utility in a wide range of applications. Most existing generative approaches for point cloud self-supervised learning focus on recovering masked points from visible ones within a single view. Recognizing that a two-view pre-training paradigm inherently introduces greater diversity and variance, it may thus enable more challenging and informative pre-training. Inspired by this, we explore the potential of two-view learning in this domain. In this paper, we propose Point-PQAE, a cross-reconstruction generative paradigm that first generates two decoupled point clouds/views and then reconstructs one from the other. To achieve this goal, we develop a crop mechanism for point cloud view generation for the first time and further propose a novel positional encoding to represent the 3D relative position between the two decoupled views. The cross-reconstruction significantly increases the difficulty of pre-training compared to self-reconstruction, which enables our method to surpass previous single-modal self-reconstruction methods in 3D self-supervised learning. Specifically, it outperforms the self-reconstruction baseline (Point-MAE) by 6.5%, 7.0%, and 6.7% in three variants of ScanObjectNN with the Mlp-Linear evaluation protocol. The code is available at https://github.com/aHapBean/Point-PQAE.

Vision-Language-Action (VLA) is crucial for autonomous decision-making in embodied systems. While current methods have advanced single-skill abilities, their short-horizon capability limits applicability in real-world scenarios. To address this challenge, we innovatively propose MindExplore, a general hierarchical VLA system with cross-skill for long-horizon tasks in highly dynamic sand. The key insight is to iteratively align the knowledge domain of task planning and action execution. Thus, this task-oriented action enables outstanding generalization across a wide range of real-world scenarios. In the reasoning layer, task-specific chains of thought (CoT) are designed for planning long-horizon task sequences and providing meta-action signals. In the acting layer, a simple but powerful Mixture of Policy Experts strategy is built inspired by signals and multimodal inputs for adaptively selecting skill experts and generating closed-loop action sequences. Also, it integrates a lightweight Multimodal Diffusion Policy (MMDP) to enhance spatial perception by fusing multi-visual modality features. Besides, the pioneering memory mechanism establishes feedback between the reasoning and acting layers, facilitating adaptive execution of long-horizon tasks and real-time replanning. Notably, we create SandGo-1k and SandThink-21k, the first expert-level multimodal embodied dataset and CoT dataset tailored for sandy environments. At a high execution frequency of 30 FPS, MindExplore is 3.01 times more successful than existing methods in unstructured and dynamic environments.

Real-time synthesis of physically plausible human interactions remains a critical challenge for immersive VR/AR systems and humanoid robotics. While existing methods demonstrate progress in kinematic motion generation, they often fail to address the fundamental tension between real-time responsiveness, physical feasibility, and safety requirements in dynamic human-machine interactions. We introduce Human-X, a novel framework designed to enable immersive and physically plausible human interactions across diverse entities, including human-avatar, human-humanoid, and human-robot systems. Unlike existing approaches that focus on post-hoc alignment or simplified physics, our method jointly predicts actions and reactions in real-time using an auto-regressive reaction diffusion planner, ensuring seamless synchronization and context-aware responses. To enhance physical realism and safety, we integrate an actor-aware motion tracking policy trained with reinforcement learning, which dynamically adapts to interaction partners' movements while avoiding artifacts like foot sliding and penetration. Extensive experiments on the Inter-X and InterHuman datasets demonstrate significant improvements in motion quality, interaction continuity, and physical plausibility over state-of-the-art methods. Our framework is validated in real-world applications, including virtual reality interface for human-robot interaction, showcasing its potential for advancing human-robot collaboration.

Building autonomous agents that can replicate human behavior in the realistic 3D world is a key step toward artificial general intelligence. This requires agents to be holistic goal achievers and to naturally adapt to environmental dynamics. In this work, we introduce ACTOR, an agent capable of performing high-level, long-horizon, abstract goals in 3D households, guided by its internal value similar to those of humans. ACTOR operates in a perceive-plan-act cycle, extending the ungrounded, scene-agnostic LLM controller with deliberate goal decomposition and decision-making through actively searching the behavior space, generating activity choices based on a hierarchical prior, and evaluating these choices using customizable value functions to determine the subsequent steps. Furthermore, we introduce BehaviorHub, a large-scale human behavior simulation dataset in scene-aware, complicated tasks. Considering the unaffordable acquisition of human-authored 3D human behavior data, we construct BehaviorHub by exploring the commonsense knowledge of LLMs learned from large corpora, and automatically aligning motion resources with 3D scene for knowledgeable generation. Extensive experiments on our established benchmark demonstrate that the proposed architecture leads to effective behavior planning and simulation. BehaviorHub also proves beneficial for downstream task development.

Parameter-efficient fine-tuning (PEFT) has become a standard for adapting large pre-trained models. While low-rank adaptation (LoRA) has achieved notable success, recent studies highlight its limitations when compared to full-rank variants, particularly when scaling to demanding tasks such as vision-language classification or common-sense reasoning.We propose to quantitavely compare full and rank-restricted PEFT methods using a spectrum-controlled matrix approximation benchmark. Our results validate LoRA's rank limitations when approximating matrix presenting highly decorrelated or high frequency features. We further show that full-rank methods can reduce LoRA's approximation error on these matrix types for an equal parameter count.Our evaluation then extends beyond synthetic tasks where we observe that LoRA's restricted work subspace can produce high norm updates, leading to over-fitting and poor out-of-distribution generalization. We address these limits by introducing KRAdapter, a novel PEFT algorithms that uses properties of the Kathri-Rao matrix product to produce weight matrices of higher effective rank and lower norm than related PEFT algorithms.We show the performance improvements of KRAdapter on vision-language models up to 1B parameters and 8B %32Bfor LLMs where we report from 20 to 25 points of accuracy improvements over LoRA when reasoning on commonsense tasks unseen during training. Crucially, KRAdapter maintains the favorable training speed and memory efficiency of LoRA, making it a practical and robust alternative to fine-tune billion-scale parameter models. Code for reproducing toy experiments is available in the supplementary and will be released upon acceptance.

mmWave radars are compact, inexpensive, and durable sensors that are robust to occlusions and work regardless of environmental conditions, such as weather and darkness. However, this comes at the cost of poor angular resolution, especially for inexpensive single-chip radars, which are typically used in automotive and indoor sensing applications. Although many have proposed learning-based methods to mitigate this weakness, no standardized foundational models or large datasets for the mmWave radar have emerged, and practitioners have largely trained task-specific models from scratch using relatively small datasets.In this paper, we collect (to our knowledge) the largest available raw radar dataset with 1M samples (29 hours) and train a foundational model for 4D single-chip radar, which can predict 3D occupancy and semantic segmentation with quality that is typically only possible with much higher resolution sensors. We demonstrate that our Generalizable Radar Transformer (GRT) generalizes across diverse settings, can be fine-tuned for different tasks, and shows logarithmic data scaling of 20% per 10xdata. We also run extensive ablations on common design decisions, and find that using raw radar data significantly outperforms widely-used lossy representations, equivalent to a 10xincrease in training data. Finally, we estimate a total data requirement of ~100M samples (3000 hours) to fully exploit the potential of GRT.

The recent Segment Anything Models (SAMs) have emerged as foundational visual models for general interactive segmentation. Despite demonstrating robust generalization abilities, they still suffer from performance degradations in scenarios that demand accurate masks. Existing methods for high-precision interactive segmentation face a trade-off between perceiving intricate local details and maintaining stable prompting capability, which hinders the applicability and effectiveness of foundational segmentation models. In this paper, we present a SAM2Refiner framework built upon the SAM2 backbone. This architecture allows SAM2 to generate fine-grained segmentation masks for both images and videos while preserving its inherent strengths. Specifically, we design a localization augment module, which incorporates local contextual cues to enhance global features via a cross-attention mechanism, thereby exploiting potential detailed patterns while maintaining semantic information. Moreover, to strengthen the prompting ability toward the enhanced object embeddings, we introduce a prompt retargeting module that renews the embedding with spatially aligned prompt features. In addition, to obtain accurate high resolution segmentation masks, a mask refinement module is devised by employing a multi-scale cascaded structure to fuse mask features with hierarchical representations from the encoder. Extensive experiments demonstrate the effectiveness of our approach, revealing that the proposed method can produce highly precise masks for both images and videos, surpassing state-of-the-art methods.

In this paper, we highlight a critical yet often overlooked factor in most 3D human tasks, namely modeling complicated 3D human with with hand-held objects or loose-fitting clothing. It is known that the parameterized formulation of SMPL is able to fit human skin; while hand-held objects and loose-fitting clothing, are difficult to get modeled within the unified framework, since their movements are usually decoupled with the human body.To enhance the capability of SMPL skeleton in response to this situation, we propose a growth strategy that enables the joint tree of the skeleton to expand adaptively. Specifically, our method, called ToMiE, consists of parent joints localization and external joints optimization. For parent joints localization, we employ a gradient-based approach guided by both LBS blending weights and motion kernels.Once the external joints are obtained, we proceed to optimize their transformations in SE(3) across different frames, enabling rendering and explicit animation.ToMiE manages to outperform other methods across various cases with hand-held objects and loose-fitting clothing, not only in rendering quality but also by offering free animation of grown joints, thereby enhancing the expressive ability of SMPL skeleton for a broader range of applications.

Recent advances in the masked autoencoder (MAE) paradigm have significantly propelled self-supervised skeleton-based action recognition. However, most existing approaches limit reconstruction targets to raw joint coordinates or their simple variants, resulting in computational redundancy and limited semantic representation. To address this, we propose a novel General Feature Prediction framework (GFP) for efficient mask skeleton modeling. Our key innovation is replacing conventional low-level reconstruction with high-level feature prediction that spans from local motion patterns to global semantic representations. Specifically, we introduce a collaborative learning framework where a lightweight target generation network dynamically produces diversified supervision signals across spatial-temporal hierarchies, avoiding reliance on pre-computed offline features. The framework incorporates constrained optimization to ensure feature diversity while preventing model collapse. Experiments on NTU RGB+D 60, NTU RGB+D 120 and PKU-MMD demonstrate the benefits of our approach: Computational efficiency (with 6.2x faster training than standard masked skeleton modeling methods) and superior representation quality, achieving state-of-the-art performance in various downstream tasks.

Large-scale pre-trained foundation models have demonstrated remarkable generalization capabilities across diverse computer vision tasks through fine-tuning. However, existing fine-tuning approaches often encounter challenges in extreme cross-domain few-shot learning scenarios, primarily due to the significant domain shift between pre-training data and target tasks, as well as the scarcity of annotated target samples. To mitigate this issue, we propose a novel absorption adaptation learning framework which meticulously regularizes the fine-tuning procedure of foundation model using an expert model with the same architecture but trained from scratch on the targeted data in two aspects. On one hand, we first design a masked cross-model unidirectional reconstruction scheme, which forces the foundation model to recover the intermediate feature of the expert model in a randomly masked manner. On the other hand, a decision graph association loss is developed to encourage the consistency of token similarity matrix between these two models. By doing these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels are appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation. Sufficient experiments with further observations and analyses underpin our observation and argument. The code is available at https://github.com/NWPUZhoufei/FMA.

Modern retrieval systems often struggle with upgrading to new and more powerful models due to the incompatibility of embeddings between the old and new models. This necessitates a costly process known as backfilling, which involves re-computing the embeddings for a large number of data samples. In vision, Backward-compatible Training (BT) has been proposed to ensure that the new model aligns with the old model's embeddings. This paper extends the concept of vision-only BT to the field of cross-modal retrieval, marking the first attempt to address Cross-modal BT (XBT). Our goal is to achieve backward-compatibility between Vision-Language Pretraining (VLP) models, such as CLIP, for the cross-modal retrieval task. To address XBT challenges, we propose an efficient solution: a projection module that maps the new model's embeddings to those of the old model. This module, pretrained solely with text data, significantly reduces the number of image-text pairs required for XBT learning, and, once it is pretrained, it avoids using the old model during training. Furthermore, we utilize parameter-efficient training strategies that improve efficiency and preserve the off-the-shelf new model's knowledge by avoiding any modifications. Experimental results on cross-modal retrieval datasets demonstrate the effectiveness of XBT and its potential to enable backfill-free upgrades when a new VLP model emerges.

Online education has been widespread in worldwide universities and educational institutions. Lecture slides, a fundamental component of online education, contain a wealth of information, playing a crucial role in learning.However, previous works have not yet paid sufficient attention to understanding lecture slides, including the absence of the large-scale dataset and comprehensive understanding tasks. To facilitate the research about lecture slides understanding, we establish the LecSlides-370K, which consists of 25,542 lectures with 370,078 slides across 15 areas. We also introduce two comprehensive tasks, Lecture Summary and Lecture Question Answering (QA), for providing different perspectives of slides understanding. Furthermore, complex and flexible text relations can hinder the understanding of the internal logic of slides. To address this challenge, we propose a novel method, named SlideParser, which includes an auxiliary branch to predict text relations within slides and enhance attention between related texts, thereby improving slides understanding. With extensive experiments, we show the superiority of our proposed method on both LecSlides-370k and SlideVQA. Dataset and code will be released soon.

Human keypoint detection is fundamental in computer vision, with applications in pose estimation and action recognition. However, existing evaluation metrics (e.g., OKS, PCP, PDJ) rely on human-annotated ground truth, a labor-intensive process that increases costs, limits scalability. To address this, we propose KPAScore (KeyPoint-Answering Score), an annotation-free metric independent of ground truth. It evaluates keypoint detection using a two-stage VLM-based question-answering process: first, the VLM identifies the presence of keypoints within the image, and second, visual prompts are introduced to query the likelihood of each keypoint being accurately localized within a predefined boundary. To validate the rationale behind KPAScore, we propose KPUBench (KeyPoint Understanding Benchmark), which comprehensively evaluates the VLM's ability to determine keypoint presence and localization. Extensive experiments demonstrate KPAScore's effectiveness from three perspectives: consistency to keypoint variation, correlation with traditional metrics, alignment with human perception. We hope KPAScore will reduce reliance on manual annotations, facilitating broader adoption of keypoint detection in real-world applications.

Despite the remarkable progress of deep neural networks (DNNs) in various visual tasks, their vulnerability to adversarial examples raises significant security concerns. Recent adversarial training methods leverage inverse adversarial attacks to generate high-confidence examples, aiming to align adversarial distributions with high-confidence class regions. However, our investigation reveals that under inverse adversarial attacks, high-confidence outputs are influenced by biased feature activations, causing models to rely on background features that lack a causal relationship with the labels. This spurious correlation bias leads to overfitting irrelevant background features during adversarial training, thereby degrading the model's robust performance and generalization capabilities. To address this issue, we propose Debiased High-Confidence Adversarial Training (DHAT), a novel approach that aligns adversarial logits with debiased high-confidence logits and restores proper attention by enhancing foreground logit orthogonality. Extensive experiments demonstrate that DHAT achieves state-of-the-art robustness on both CIFAR and ImageNet-1K benchmarks, while significantly improving generalization by mitigating the feature bias inherent in inverse adversarial training approaches. Code is available at https://github.com/KejiaZhang-Robust/DHAT.

This paper presents Edge-based Mixture of Experts (MoE) Collaborative Computing (EMC2), an optimal computing system designed for autonomous vehicles (AVs) that simultaneously achieves low-latency and high-accuracy 3D object detection. Unlike existing works, the EMC2 introduces a novel scenario-aware MoE architecture optimized for fusing complementary sparse 3D point clouds and dense 2D images to achieve robust multimodal representations for detection. Furthermore, EMC2 integrates an adaptive multimodal data bridge with multi-scale region proposing and scenario-aware routing, dynamically dispatching features to complementary experts based on object visibility and distance. In addition, EMC2 integrates joint hardware-software optimizations, including hardware resource utilization optimization and computational graph simplification, to ensure efficient and real-time inference on resource-constrained edge devices. Experiments on open-source benchmarks clearly show the EMC2 advancements as an end-to-end system. On the KITTI dataset, it achieves an average accuracy improvement of 3.58% and a 159.06% inference speedup compared to 15 baseline methods on Jetson platforms, with similar performance gains on the nuScenes dataset, highlighting its capability to advance reliable, real-time 3D object detection tasks for AVs.

System identification from videos aims to recover object geometry and governing physical laws. Existing methods integrate differentiable rendering with simulation but rely on predefined material priors, limiting their ability to handle unknown ones. We introduce MASIV, the first vision-based framework for material-agnostic system identification. Unlike existing approaches that depend on hand-crafted constitutive laws, MASIV employs learnable neural constitutive models, inferring object dynamics without assuming a scene-specific material prior. However, the absence of full particle state information imposes unique challenges, leading to unstable optimization and physically implausible behaviors. To address this, we introduce dense geometric guidance by reconstructing continuum particle trajectories, providing temporally rich motion constraints beyond sparse visual cues. Comprehensive experiments show that MASIV achieves state-of-the-art performance in geometric accuracy, rendering quality, and generalization ability.