ICCV 2025 Papers

Neural Radiance Fields (NeRF) has emerged as a compelling framework for scene representation and 3D recovery. To improve its performance on real-world data, depth regularizations have proven to be the most effective ones. However, depth estimation models not only require expensive 3D supervision in training, but also suffer from generalization issues. As a result, the depth estimations can be erroneous in practice, especially for outdoor unbounded scenes. In this paper, we propose to employ view-consistent distributions instead of fixed depth value estimations to regularize NeRF training. Specifically, the distribution is computed by utilizing both low-level color features and high-level distilled features from foundation models at the projected 2D pixel-locations from per-ray sampled 3D points. By sampling from the view-consistency distributions, an implicit regularization is imposed on the training of NeRF. We also propose a novel depth-pushing loss that works in conjunction with the sampling technique to jointly provide effective regularizations for eliminating the failure modes. Extensive experiments conducted on various scenes from public datasets demonstrate that our proposed method can generate significantly better novel view synthesis results than state-of-the-art NeRF variants as well as different depth regularization methods.

This paper explains training-time out-of-distribution (OOD) detection from a novel view, i.e., interactions between different input variables of deep neural networks (DNNs). Specifically, we provide a unified understanding of the effectiveness of current training-time OOD detection methods, i.e., DNNs trained with these methods all encode more complex interactions for inference than those trained without training-time methods, which contributes to their superior OOD detection performance. We further conduct thorough empirical analyses and verify that complex interactions play a primary role in OOD detection, by developing a simple-yet-efficient method to force the DNN to learn interactions of specific complexities and evaluate the change of OOD detection performances. Besides, we also use interactions to investigate why near-OOD samples are more difficult to distinguish from in-distribution (ID) samples than far-OOD samples, mainly because compared to far-OOD samples, the distribution of interactions in near-OOD samples is more similar to that of ID samples. Moreover, we discover that training-time OOD detection methods can effectively decrease such similarities.

Incomplete multi-view clustering (IMVC) has gained increasing attention due to its ability to analyze incomplete multi-view data.Despite deep IMVC methods achieved significant progress, they still face two challenges: (I) The method-specific inseparable designs limit their application. (II) Non-independent and identically distributed (Non-IID) missing patterns has not been considered and caused degeneration. To address these issues, we propose a novel unified framework that bridges from deep MVC to deep IMVC, while emphasizing the robustness against Non-IID missing patterns. Our framework has a two-stage process: (I) Multi-view learning on complete data, where our framework is modularly established to be compatible with different multi-view interaction objectives. (II) Transfer learning and clustering on incomplete data, where we propose a multi-view domain adversarial learning method to improve the model robustness to Non-IID missing patterns. Moreover, an intra-view and inter-view imputation strategy is introduced for more reliable clustering.Based on our unified framework, we easily construct multiple IMVC instances and extensive experiments verified their clustering effectiveness.

Recent advancements in large language models (LLMs) have significantly improved their ability to generate natural and contextually appropriate text, enabling more human-like interactions. However, understanding and generating interactive human-like motion, especially involving coordinated interactive motion, remains a challenging problem due to its inherent complexity. To address this, we present MoLaM, the Interactive Motion-LAnguage Model, a unified architecture that jointly processes language and motion modalities for understanding, generating, and controlling interactive motions in multi-turn conversational settings. Unlike prior approaches limited to uni-directional tasks (e.g., text-to-motion or motion-to-text), MoLaM supports a wide range of bi-directional and multi-turn tasks. We also introduce Inter-MT2, a large-scale instruction tuning dataset containing 82.7K multi-turn interactive motion instructions and 153K motion samples, covering tasks such as editing, question answering, and story generation. Leveraging LLMs and motion diffusion models, MoLaM demonstrates strong performance across five interactive motion tasks: motion-to-text, text-to-motion, reaction generation, motion editing, and motion reasoning, outperforming or matching task-specific baselines with a single, unified model.

In recent years, general visual foundation models (VFMs) have witnessed increasing adoption, particularly as image encoders for popular multi-modal large language models (MLLMs). However, without semantically fine-grained supervision, these models still encounter fundamental prediction errors in the context of downstream text-image-related tasks, i.e., perception, understanding and reasoning with images containing small and dense texts. To bridge this gap, we develop TokenFD, the first token-level visual foundation model specifically tailored for text-image-related tasks, designed to support a variety of traditional downstream applications. To facilitate the pretraining of TokenFD, we also devise a high-quality data production pipeline that constructs the first token-level image text dataset, TokenIT, comprising 20 million images and 1.8 billion token-mask pairs. Furthermore, leveraging this foundation with exceptional image-as-text capability, we seamlessly replace previous VFMs with TokenFD to construct a token-level visual-language MLLM, TokenVL, for VQA-based document understanding tasks. Finally, extensive experiments demonstrate the effectiveness of TokenFD and TokenVL. Code, demo, datasets, and weights are available at https://github.com/Token-family/TokenFD.

Long-Tailed Class-Incremental Learning (LT-CIL) remains a fundamental challenge due to biased gradient updates caused by highly imbalanced data distributions and the inherent stability-plasticity dilemma. These factors jointly degrade tail-class performance and exacerbate catastrophic forgetting. To tackle these issues, we propose Geometric Prototype Alignment (GPA), a model-agnostic approach that calibrates classifier learning dynamics via geometric feature-space alignment. GPA initializes classifier weights by projecting frozen class prototypes onto a unit hypersphere, thereby disentangling magnitude imbalance from angular discriminability. During incremental updates, a Dynamic Anchoring mechanism adaptively adjusts classifier weights to preserve geometric consistency, effectively balancing plasticity for new classes with stability for previously acquired knowledge. Integrated into state-of-the-art CIL frameworks such as LUCIR and DualPrompt, GPA yields substantial gains, improving average incremental accuracy by 6.11% and reducing forgetting rates by 6.38% on CIFAR100-LT. Theoretical analysis further demonstrates that GPA accelerates convergence by 2.7X and produces decision boundaries approaching Fisher-optimality. Our implementation is available at https://github.com/laixinyi023/Geometric-Prototype-Alignment.

Recent Mamba-based methods for the pose-lifting task tend to model joint dependencies by 2D-to-1D mapping with diverse scanning strategies. Though effective, they struggle to model intricate joint connections and uniformly process all joint motion trajectories while neglecting the intrinsic differences across motion characteristics. In this work, we propose a structure-aware and motion-adaptive framework to capture spatial joint topology along with diverse motion dynamics independently, named as SAMA. Specifically, SAMA consists of a Structure-aware State Integrator (SSI) and a Motion-adaptive State Modulator (MSM). The Structure-aware State Integrator is tasked with leveraging dynamic joint relationships to fuse information at both the joint feature and state levels in the state space, based on pose topology rather than sequential state transitions. The Motion-adaptive State Modulator is responsible for joint-specific motion characteristics recognition, thus applying tailored adjustments to diverse motion patterns across different joints. Through the above key modules, our algorithm enables structure-aware and motion-adaptive pose lifting. Extensive experiments across multiple benchmarks demonstrate that our algorithm achieves advanced results with fewer computational costs.

Diffusion models have demonstrated powerful capability as a versatilist for dense vision tasks, yet the generalization ability to unseen domains remains rarely explored. This paper presents HarDiff, an efficient frequency learning scheme, so as to advance generalizable paradigms for diffusion based dense prediction. It draws inspiration from a fine-grained analysis of Discrete Hartley Transform, where some low-frequency features activate the broader content of an image, while some high-frequency features maintain sufficient details for dense pixels. Consequently, HarDiff consists of two key components. The low-frequency training process extracts structural priors from the source domain, to enhance understanding of task-related content. The high-frequency sampling process utilizes detail-oriented guidance from the unseen target domain, to infer precise dense predictions with target-related details. Extensive empirical evidence shows that HarDiff can be easily plugged into various dense vision tasks, e.g., semantic segmentation, depth estimation and haze removal, yielding improvements over the state-of-the-art methods in twelve public benchmarks.

We introduce a recipe for generating immersive 3D worlds from a single image by framing the task as an in-context learning problem for 2D inpainting models. This approach requires minimal training and uses existing generative models. Our process involves two steps: generating coherent panoramas using a pre-trained diffusion model and lifting these into 3D with a metric depth estimator. We then fill unobserved regions by conditioning the inpainting model on rendered point clouds, requiring minimal fine-tuning. Tested on both synthetic and real images, our method produces high-quality 3D environments suitable for VR display. By explicitly modeling the 3D structure of the generated environment from the start, our approach consistently outperforms state-of-the-art, video synthesis-based methods along multiple quantitative image quality metrics.

Inverse rendering aims to reconstruct geometry and reflectance from captured images. Display-camera imaging systems offer unique advantages for this task: each pixel can easily function as a programmable point light source, and the polarized light emitted by LCD displays facilitates diffuse-specular separation. Despite these benefits, there is currently no public real-world dataset captured using display-camera systems, unlike other setups such as light stages. This absence hinders the development and evaluation of display-based inverse rendering methods. In this paper, we introduce the first real-world dataset for display-based inverse rendering. To achieve this, we construct and calibrate an imaging system comprising an LCD display and stereo polarization cameras. We then capture a diverse set of objects with diverse geometry and reflectance under one-light-at-a-time (OLAT) display patterns. We also provide high-quality ground-truth geometry. Our dataset enables the synthesis of captured images under arbitrary display patterns and different noise levels. Using this dataset, we evaluate the performance of existing photometric stereo and inverse rendering methods, and provide a simple, yet effective baseline for display inverse rendering, outperforming state-of-the-art inverse rendering methods.Code and dataset are available on our project page at https://michaelcsj.github.io/DIR/.

Whole-body biometric recognition is a challenging multi-modal task that integrates various biometric modalities, including face, gait, and body. This integration is essential for overcoming the limitations of unimodal systems. Traditionally, whole-body recognition involves deploying different models to process multiple modalities, achieving the final outcome by score-fusion (e.g., weighted averaging similarity matrices from each model). However, these conventional methods may overlook the variations in score distributions of individual modalities, making it challenging to improve final performance. In this work, we present Quality-guided Mixture of score-fusion Experts (QME), a novel framework designed for improving whole-body biometric recognition performance through a learnable score-fusion strategy using a Mixture of Experts (MoE). We introduce a novel pseudo quality loss for quality estimation with a modality-specific Quality Estimator (QE), and a score triplet loss to improve the metric performance. Extensive experiments on multiple whole-body biometric datasets demonstrate the effectiveness of our proposed approach, achieving state-of-the-art results across various metrics compared to baseline methods. Our method is effective for multi-modal and multi-model, addressing key challenges such as model misalignment in the similarity score domain and variability in data quality. Code is available at the \href https://github.com/jiezhu23/QME_ICCV25 Project Link

Extracting physically plausible 3D human motion from videos is a critical task. Although existing simulation-based motion imitation methods can enhance the physical quality of daily motions estimated from monocular video capture, extending this capability to high-difficulty motions remains an open challenge. This can be attributed to some flawed motion clips in video-based motion capture results and the inherent complexity in modeling high-difficulty motions. Therefore, sensing the advantage of segmentation in localizing human body, we introduce a mask-based motion correction module (MCM) that leverages motion context and video mask to repair flawed motions; and propose a physics-based motion transfer module (PTM), which employs a prior injected pretrain and adapt approach for motion imitation, improving physical plausibility with the ability to handle in-the-wild and challenging motions. Our approach is designed as a plug-and-play module to physically refine the video motion capture, which also excels in motion generation tasks. Finally, we collected a challenging in-the-wild test set to establish a benchmark, and our method has demonstrated effectiveness on both the new benchmark and existing public datasets. Our project page is : https://physicalmotionrestoration.github.io/

Structure and continuous motion estimation from point correspondences is a fundamental problem in computer vision that has been powered by well-known algorithms such as the familiar 5-point or 8-point algorithm. However, despite their acclaim, these algorithms are limited to processing point correspondences originating from a pair of views each one representing an instantaneous capture of the scene. Yet, in the case of rolling shutter cameras, or more recently, event cameras, this synchronization breaks down. In this work, we present a unified approach for structure and linear motion estimation from 2D point correspondences with arbitrary timestamps, from an arbitrary set of views. By formulating the problem in terms of first-order dynamics and leveraging a constant velocity motion model, we derive a novel, linear point incidence relation allowing for the efficient recovery of both linear velocity and 3D points with predictable degeneracies and solution multiplicities. Owing to its general formulation, it can handle correspondences from a wide range of sensing modalities such as global shutter, rolling shutter, and event cameras, and can even combine correspondences from different collocated sensors. We validate the effectiveness of our solver on both simulated and real-world data, where we show consistent improvement across all modalities when compared to recent approaches. We believe our work opens the door to efficient structure and motion estimation from asynchronous data.

We present a novel framework for training 3D image-conditioned diffusion models using only 2D supervision. Recovering 3D structure from 2D images is inherently ill-posed due to the ambiguity of possible reconstructions, making generative models a natural choice. However, most existing 3D generative models rely on full 3D supervision, which is impractical due to the scarcity of large-scale 3D datasets. To address this, we propose leveraging sparse-view supervision as a scalable alternative. While recent reconstruction models use sparse-view supervision with differentiable rendering to lift 2D images to 3D, they are predominantly deterministic, failing to capture the diverse set of plausible solutions and producing blurry predictions in uncertain regions. A key challenge in training 3D diffusion models with 2D supervision is that the standard training paradigm requires both the denoising process and supervision to be in the same modality. We address this by decoupling the noisy samples being denoised from the supervision signal, allowing the former to remain in 3D while the latter is provided in 2D. Our approach leverages suboptimal predictions from a deterministic image-to-3D model--acting as a "teacher"--to generate noisy 3D inputs, enabling effective 3D diffusion training without requiring full 3D ground truth. We validate our framework on both object-level and scene-level datasets, using two different 3D Gaussian Splat (3DGS) teachers. Our results show that our approach consistently improves upon these deterministic teachers, demonstrating its effectiveness in scalable and high-fidelity 3D generative modeling.

Visual Place Recognition (VPR) enables coarse localization by comparing query images to a reference database of geo-tagged images. Recent breakthroughs in deep learning architectures and training regimes have led to methods with improved robustness to factors like environment appearance change, but with the downside that the required training and/or matching compute scales with the number of distinct environmental conditions encountered. Here, we propose Hyperdimensional One Place Signatures (HOPS) to simultaneously improve the performance, compute and scalability of these state-of-the-art approaches by fusing the descriptors from multiple reference sets captured under different conditions. HOPS scales to any number of environmental conditions by leveraging the Hyperdimensional Computing framework. Extensive evaluations demonstrate that our approach is highly generalizable and consistently improves recall performance across all evaluated VPR methods and datasets by large margins. Arbitrarily fusing reference images without compute penalty enables numerous other useful possibilities, three of which we demonstrate here: improved performance with reduced dimensionality descriptors, stacking synthetic images, and coarse localization to an entire traverse or environmental section.

Generalized Category Discovery (GCD) aims to classify unlabeled data from both known and unknown categories by leveraging knowledge from labeled known categories. While existing methods have made notable progress, they often overlook a hidden stumbling block in GCD: distracted attention. Specifically, when processing unlabeled data, models tend to focus not only on key objects in the image but also on task-irrelevant background regions, leading to suboptimal feature extraction. To remove this stumbling block, we propose Attention Focusing (AF), an adaptive mechanism designed to sharpen the model's focus by pruning non-informative tokens. AF consists of two simple yet effective components: Token Importance Measurement (TIME) and Token Adaptive Pruning (TAP), working in a cascade. TIME quantifies token importance across multiple scales, while TAP prunes non-informative tokens by utilizing the multi-scale importance scores provided by TIME. AF is a lightweight, plug-and-play module that integrates seamlessly into existing GCD methods with minimal computational overhead. When incorporated into one prominent GCD method, SimGCD, AF achieves up to 15.4% performance improvement over the baseline with minimal computational overhead. The implementation code is provided in https://github.com/Afleve/AFGCD.

Knowledge distillation (KD) is an effective method for enhancing a small model, named student, by training it under the supervision of larger teacher models. However, existing studies indicate that a substantial capacity gap between the student and teacher can lead to poor learning for the student model. This capacity gap problem limits the applicability of KD and necessitates careful selection of the teacher's size. Despite its importance, the underlying cause of the capacity gap problem remains underexplored. In this paper, we reveal that a substantial disparity in the output distributions of teacher and student models is a key factor behind this issue. To demonstrate this, we decompose the KD loss into two components: class-wise similarity and intra-class distribution, and analyze the contribution of each term. Our analysis shows that a large distributional mismatch can lead to poor student learning.Inspired by this observation, we propose the Adapted Intra-class Distribution (AID) method, wherein the teacher model is fine-tuned to optimize its intra-class distribution to better align with the student's capacity prior to knowledge distillation. This approach effectively bridges the capacity gap between teacher and student models and consistently achieves state-of-the-art performance across a diverse range of architectures.

Foundation models (FMs) excel in zero-shot tasks but benefit from task-specific adaptation. However, privacy concerns prevent data sharing among multiple data owners, and proprietary restrictions prevent the learning service provider (LSP) from sharing the FM. In this work, we propose BlindFed, a framework enabling collaborative FM adaptation while protecting both parties: data owners do not access the FM or each other's data, and the LSP does not see sensitive task data. BlindFed relies on fully homomorphic encryption (FHE) and consists of three key innovations: (i) FHE-friendly architectural modifications via polynomial approximations and low-rank adapters, (ii) a two-stage split learning approach combining offline knowledge distillation and online encrypted inference for adapter training without backpropagation through the FM, and (iii) a privacy-boosting scheme using sample permutations and stochastic block sampling to mitigate model extraction attacks. Empirical results on four image classification datasets demonstrate the practical feasibility of the BlindFed framework, albeit at a high communication cost and large computational complexity for the LSP. The code can be found at https://github.com/tnurbek/blindfed.