ICCV 2025 论文列表

Transformer-based models have driven significant advancements in Multimodal Large Language Models (MLLMs), yet their computational costs surge drastically when scaling resolution, training data, and model parameters. A key bottleneck stems from the proliferation of visual tokens required for fine-grained image understanding. We propose Skip-Vision, a unified framework addressing both training and inference inefficiencies in vision-language models. On top of conventional token compression approaches, our method introduces two complementary acceleration strategies. For training acceleration, we observe that Feed-Forward Network (FFN) computations on visual tokens induce marginal feature updates. This motivates our Skip-FFN strategy, which bypasses FFN layers for redundant visual tokens. For inference acceleration, we design a selective KV-cache removal mechanism that prunes the skipped key-value pairs during decoding while preserving model performance. Experimental results demonstrate that Skip-Vision reduces training time by up to 35%, inference FLOPs by 75%, and latency by 45%, while achieving comparable or superior performance to existing methods. Our work provides a practical solution for scaling high-performance MLLMs with enhanced efficiency.

Edges are one of the most basic parametric primitives to describe structural information in 3D. In this paper, we study parametric 3D edge reconstruction from calibrated multi-view images. Previous methods usually reconstruct a 3D edge point set from multi-view 2D edge images, and then fit 3D edges to the point set. However, noise in the point set may cause gaps among fitted edges, and the recovered edges may not align with input multi-view images since the edge fitting depends only on the reconstructed 3D point set. To mitigate these problems, we propose SketchSplat, a method to reconstruct accurate, complete, and compact 3D edges via differentiable multi-view sketch splatting. We represent 3D edges as sketches, which are parametric lines and curves defined by attributes including control points, scales, and opacity. During reconstruction, we iteratively sample Gaussian points from a set of sketches and rasterize the Gaussians onto 2D edge images. Then the gradient of the image loss can be back-propagated to optimize the sketch attributes. Our method bridges 2D edge images and 3D edges in a differentiable manner, which ensures that 3D edges align well with 2D images and leads to accurate and complete results. We also propose a series of adaptive topological operations to reduce redundant edges and apply them along with the sketch optimization, yielding a more compact reconstruction. Finally, we contribute an accurate 2D edge detector that improves the performance of both ours and existing methods. Experiments show that our method achieves state-of-the-art accuracy, completeness, and compactness on a benchmark CAD dataset.

Current state-of-the-art methods for skeleton-based temporal action segmentation are predominantly supervised and require annotated data, which is expensive to collect. In contrast, existing unsupervised temporal action segmentation methods have focused primarily on video data, while skeleton sequences remain underexplored, despite their relevance to real-world applications, robustness, and privacy-preserving nature. In this paper, we propose a novel approach for unsupervised skeleton-based temporal action segmentation. Our method utilizes a sequence-to-sequence temporal autoencoder that keeps the information of the different joints disentangled in the embedding space. Latent skeleton sequences are then divided into non-overlapping patches and quantized to obtain distinctive skeleton motion words, driving the discovery of semantically meaningful action clusters. We thoroughly evaluate the proposed approach on three widely used skeleton-based datasets, namely HuGaDB, LARa, and BABEL. The results demonstrate that our model outperforms the current state-of-the-art unsupervised temporal action segmentation methods.

We propose a novel approach for estimating the relative pose between rolling shutter cameras using the intersections of line projections with a single scanline per image. This allows pose estimation without explicitly modeling camera motion. Alternatively, scanlines can be selected within a single image, enabling single-view relative pose estimation for scanlines of rolling shutter cameras. Our approach is designed as a foundational building block for rolling shutter structure-from-motion (SfM), where no motion model is required, and each scanline's pose can be computed independently.We classify minimal solvers for this problem in both generic and specialized settings, including cases with parallel lines and known gravity direction. Furthermore, we develop minimal solvers for the parallel-lines scenario, both with and without gravity priors, by leveraging connections between this problem and the estimation of 2D structure from 1D cameras.Experiments on rolling shutter images from the Fastec dataset demonstrate the feasibility of our approach for initializing rolling shutter SfM, highlighting its potential for further development.The code will be made publicly available.

Event cameras are emerging vision sensors whose noise is challenging to characterize. Existing denoising methods for event cameras are often designed in isolation and thus consider other tasks, such as motion estimation, separately (i.e., sequentially after denoising). However, motion is an intrinsic part of event data, since scene edges cannot be sensed without motion. We propose, to the best of our knowledge, the first method that simultaneously estimates motion in its various forms (e.g., ego-motion, optical flow) and noise. The method is flexible, as it allows replacing the one-step motion estimation of the widely-used Contrast Maximization framework with any other motion estimator, such as deep neural networks. The experiments show that the proposed method achieves state-of-the-art results on the E-MLB denoising benchmark and competitive results on the DND21 benchmark, while demonstrating effectiveness across motion estimation and intensity reconstruction tasks. Our approach advances event-data denoising theory and expands practical denoising use-cases via open-source code. Project page: https://github.com/tub-rip/ESMD

Many studies utilize dual-pixel (DP) sensor phase information for various applications, such as depth estimation and deblurring. However, since DP image features are entirely determined by the camera hardware, DP-depth paired datasets are very scarce, especially when performing depth estimation on customized cameras. To overcome this, studies simulate DP images using ideal optical models. However, these simulations often violate real optical propagation laws, leading to poor generalization to real DP data. To address this, we investigate the domain gap between simulated and real DP data, and propose solutions using the Simulating DP Images from Ray Tracing (Sdirt) scheme. Sdirt generates realistic DP images via ray tracing and integrates them into the depth estimation training pipeline. Experimental results show that models trained with Sdirt-simulated images generalize better to real DP data.

The increasing application of multi-modal large language models (MLLMs) across various sectors has spotlighted the essence of their output reliability and accuracy, particularly their ability to produce content grounded in factual information (e.g. common and domain-specific knowledge). In this work, we introduce SimpleVQA, the first comprehensive multi-modal benchmark to evaluate the factuality ability of MLLMs to answer natural language short questions. SimpleVQA is characterized by 7 key features: it is based on bilingual, it covers multiple tasks and multiple scenarios, ensures high quality and challenging queries, maintains static and timeless reference answers, and is straightforward to evaluate. Our approach involves categorizing visual question-answering items into 9 different tasks around objective events or common knowledge and situating these within 9 scenario domains. Rigorous quality control processes are implemented to guarantee high-quality, concise, and clear answers, facilitating evaluation with minimal variance via an LLM-as-a-judge scoring system. Using SimpleVQA, we perform a comprehensive assessment of leading 18 MLLMs and 8 text-only LLMs, delving into their image comprehension and text generation abilities by identifying and analyzing error cases.

In recent years, it has been found that "grandmother cells" in the primary visual cortex (V1) of macaques can directly recognize visual input with complex shapes. This inspires us to examine the value of these cells in promoting the research of medical image segmentation. In this paper, we design a Similarity Memory Prior Network (Sim-MPNet) for medical image segmentation. Specifically, we propose a Dynamic Memory Weights-Loss Attention (DMW-LA), which matches and remembers the category features of specific lesions or organs in medical images through the similarity memory prior in the prototype memory bank, thus helping the network to learn subtle texture changes between categories. DMW-LA also dynamically updates the similarity memory prior in reverse through Weight-Loss Dynamic (W-LD) update strategy, effectively assisting the network directly extract category features. In addition, we propose the Double-Similarity Global Internal Enhancement Module (DS-GIM) to deeply explore the internal differences in the feature distribution of input data through cosine similarity and euclidean distance. Extensive experiments on four public datasets show that Sim-MPNet has better segmentation performance than other state-of-the-art methods. Our code is available on https://github.com/vpsg-research/Sim-MPNet.

In this paper, we propose SimMLM, a simple yet powerful framework for multimodal learning with missing modalities. Unlike existing approaches that rely on sophisticated network architectures or complex data imputation techniques, SimMLM provides a generic and effective solution that can adapt to various missing modality scenarios with improved accuracy and robustness. Specifically, SimMLM consists of a generic Dynamic Mixture of Modality Experts (DMoME) architecture, featuring a dynamic, learnable gating mechanism that automatically adjusts each modality's contribution in both full and partial modality settings. A key innovation of SimMLM is the proposed More vs. Fewer (MoFe) ranking loss, which ensures that task accuracy improves or remains stable as more modalities are made available. This aligns the model with an intuitive principle: removing one or more modalities should not increase accuracy. We validate SimMLM on multimodal medical image segmentation (BraTS 2018) and multimodal classification (UPMC Food-101, avMNIST) tasks, where it consistently surpasses competitive methods, demonstrating superior accuracy, interpretability, robustness, and reliability across both complete and missing modality scenarios at test time.

Temporal sentence grounding aims to identify exact moments in a video that correspond to a given textual query, typically addressed with detection transformer (DETR) solutions. However, we find that typical strategies designed to enhance DETR do not improve, and may even degrade, its performance in this task. We systematically analyze and identify the root causes of this abnormal behavior: (1) conflicts between queries from similar target moments and (2) internal query conflicts due to the tension between global semantics and local localization. Building on these insights, we propose a simple yet powerful baseline, Sim-DETR, which extends the standard DETR with two minor modifications in the decoder layers: (1) constraining self-attention between queries based on their semantic and positional overlap and (2) adding query-to-frame alignment to bridge the global and local contexts. Experiments demonstrate that Sim-DETR unlocks the full potential of DETR for temporal sentence grounding, offering a strong baseline for future research.

Sign language is a visual language that encompasses all linguistic features of natural languages and serves as the primary communication method for the deaf and hard-of-hearing communities. Although many studies have successfully adapted pretrained language models (LMs) for sign language translation (sign-to-text), the reverse task--sign language generation (text-to-sign)--remains largely unexplored. In this work, we introduce a multilingual sign language model, Signs as Tokens (SOKE), which can generate 3D sign avatars autoregressively from text inputs using a pretrained LM. To align sign language with the LM, we leverage a decoupled tokenizer that discretizes continuous signs into token sequences representing various body parts. During decoding, unlike existing approaches that flatten all part-wise tokens into a single sequence and predict one token at a time, we propose a multi-head decoding method capable of predicting multiple tokens simultaneously. This approach improves inference efficiency while maintaining effective information fusion across different body parts. To further ease the generation process, we propose a retrieval-enhanced SLG approach, which incorporates external sign dictionaries to provide accurate word-level signs as auxiliary conditions, significantly improving the precision of generated signs. Extensive qualitative and quantitative evaluations demonstrate the effectiveness of SOKE.

Sign language representation learning presents unique challenges due to the complex spatio-temporal nature of signs and the scarcity of labeled datasets. Existing methods often rely either on models pre-trained on general visual tasks, that lack sign-specific features, or use complex multimodal and multi-branch architectures. To bridge this gap, we introduce a scalable, self-supervised framework for sign representation learning. We leverage important inductive (sign) priors during the training of our RGB model. To do this, we leverage simple but important cues based on skeletons while pretraining a masked autoencoder. These sign specific priors alongside feature regularization and an adversarial style agnostic loss provide a powerful backbone. Notably, our model does not require skeletal keypoints during inference, avoiding the limitations of keypoint-based models during downstream tasks. When finetuned, we achieve state-of-the-art performance for sign recognition on the WLASL, ASL-Citizen and NMFs-CSL datasets, using a simpler architecture and with only a single-modality. Beyond recognition, our frozen model excels in sign dictionary retrieval and sign translation, surpassing standard MAE pretraining and skeletal-based representations in retrieval. It also reduces computational costs for training existing sign translation models while maintaining strong performance on Phoenix2014T, CSL-Daily and How2Sign.

Federated Learning (FL) enables collaborative machine learning across decentralized clients without sharing raw data, which offers enhanced privacy and improved performance. However, FL is vulnerable to poisoning attacks, compromising model integrity through both untargeted performance degradation and targeted backdoor attacks. Detecting backdoors in FL is challenging due to their stealthy nature and variability in local datasets. Existing defenses struggle against adaptive adversaries and distinguishing between poisoning and genuine dataset anomalies. This paper introduces the Siamese Backdoor Inspector (Sibai), a novel meta-classifier-based poisoning defense for FL. Leveraging the staple few-shot learning technique of Siamese networks, Sibai effectively detects malicious contributions in various scenarios, including settings with strong variations between clients' datasets and encounters with adaptive adversaries. Sibai achieves high detection rates, prevents backdoors, minimizes performance impact, and outperforms eight recent defenses regarding F1 score, poisoning prevention, and consistency across complex scenarios.

We propose SiM3D, the first benchmark considering the integration of multiview and multimodal information for comprehensive 3D anomaly detection and segmentation (ADS) where the task is to produce a voxel-based Anomaly Volume. Moreover, SiM3D focuses on a scenario of high interest in manufacturing: single-instance anomaly detection, where only one object, either real or synthetic, is available for training. In this respect, SiM3D stands out as the first ADS benchmark that addresses the challenge of generalizing from synthetic training data to real test data. SiM3D includes a novel multimodal multiview dataset acquired using top-tier industrial sensors and robots. The dataset features multiview high-resolution images (12 \tt Mpx ) and point clouds (~7M points) for 333 instances of eight types of objects, alongside a CAD model for each type. We also provide manually annotated 3D segmentation GTs for anomalous test samples. To establish reference baselines for the proposed multiview 3D ADS task, we adapt prominent singleview methods and assess their performance using novel metrics that operate on Anomaly Volumes.

Our objective is the automatic generation of Audio Descriptions (ADs) for edited video material, such as movies and TV series. To achieve this, we propose a two-stage framework that leverages "shots" as the fundamental units of video understanding. This includes extending temporal context to neighboring shots and incorporating film grammar devices, such as shot scales and thread structures, to guide AD generation. Our method is compatible with both open-source and proprietary Visual-Language Models (VLMs), integrating expert knowledge from add-on modules without requiring additional training of the VLMs. We achieve state-of-the-art performance among all prior training-free approaches and even surpass fine-tuned methods on several benchmarks. To evaluate the quality of predicted ADs, we introduce a new evaluation measure -- an action score -- specifically targeted to assessing this important aspect of AD. Additionally, we propose a novel evaluation protocol that treats automatic frameworks as AD generation assistants and asks them to generate multiple candidate ADs for selection.

Backpropagation-based approaches aim to align diffusion models with reward functions through end-to-end backpropagation of the reward gradient within the denoising chain, offering a promising perspective. However, due to the computational costs and the risk of gradient explosion associated with the lengthy denoising chain, existing approaches struggle to achieve complete gradient backpropagation, leading to suboptimal results. In this paper, we introduce Shortcut-based Fine-Tuning (ShortFT), an efficient fine-tuning strategy that utilizes the shorter denoising chain. More specifically, we employ the recently researched trajectory-preserving few-step diffusion model, which enables a shortcut over the original denoising chain, and construct a shortcut-based denoising chain of shorter length. The optimization on this chain notably enhances the efficiency and effectiveness of fine-tuning the foundational model. Our method has been rigorously tested and can be effectively applied to various reward functions, significantly improving alignment performance and surpassing state-of-the-art alternatives.

Monocular dynamic reconstruction is a challenging and long-standing vision problem due to the highly ill-posed nature of the task. Existing approaches depend on templates, are effective only in quasi-static scenes, or fail to model 3D motion explicitly. We introduce a method for reconstructing generic dynamic scenes, featuring explicit, persistent 3D motion trajectories in the world coordinate frame, from casually captured monocular videos.We tackle the problem with two key insights: First, we exploit the low-dimensional structure of 3D motion by representing scene motion with a compact set of SE(3) motion bases. Each point's motion is expressed as a linear combination of these bases, facilitating soft decomposition of the scene into multiple rigidly-moving groups. Second, we take advantage of off-the-shelf data-driven priors such as monocular depth maps and long-range 2D tracks, and devise a method to effectively consolidate these noisy supervisory signals, resulting in a globally consistent representation of the dynamic scene. Experiments show that our method achieves state-of-the-art performance for both long-range 3D/2D motion estimation and novel view synthesis on dynamic scenes.