CVPR 2025 Papers

Recent advancements in multimodal large language models (MLLMs) have significantly improved performance in visual question answering. However, they often suffer from hallucinations. In this work, hallucinations are categorized into two main types: initial hallucinations and snowball hallucinations. We argue that adequate contextual information can be extracted directly from the token interaction process. Inspired by causal inference in decoding strategy, we propose to leverage causal masks to establish information propagation between multimodal tokens. The hypothesis is that insufficient interaction between those tokens may lead the model to rely on outlier tokens, overlooking dense and rich contextual cues. Therefore, we propose to intervene in the propagation process by tackling outlier tokens to enhance in-context inference. With this goal, we present FarSight, a versatile plug-and-play decoding strategy to reduce attention interference from outlier tokens merely by optimizing the causal mask. The heart of our method is effective token propagation. We design an attention register structure within the upper triangular matrix of the causal mask, dynamically allocating attention capture attention diverted to outlier tokens. Moreover, a positional awareness encoding method with a diminishing masking rate is proposed, allowing the model to attend to further preceding tokens, especially for video sequence tasks. With extensive experiments, FarSight demonstrates significant hallucination-mitigating performance across different MLLMs on both image and video benchmarks, proving its effectiveness.

Underwater (UW) robotics applications require depth and restored images simultaneously in real-time, irrespective of whether the UW images are captured in good lighting conditions or not. Most of the UW image restoration and depth estimation methods have been devised for images under normal lighting. Consequently, they struggle to perform on poorly lit images. Even though artificial illumination can be used when there is insufficient ambient light, it can introduce non-uniform lighting artifacts in the restored images. Hence, the recovery of depth and restored images directly from Low-Light UW (LLUW) images is a critical requirement in marine applications. While a few works have attempted LLUW image restoration, there are no reported works on joint recovery of depth and clean image from LLUW images. We propose a Self-supervised Low-light Underwater Image and Depth recovery network (SelfLUID-Net) for joint estimation of depth and restored image in real-time from a single LLUW image. We have collected an Underwater Low-light Stereo Video (ULVStereo) dataset which is the first-ever UW dataset with stereo pairs of low-light and normally-lit UW images. For the dual tasks of image and depth recovery from a LLUW image, we effectively utilize the stereo data from ULVStereo that provides cues for both depth and illumination-independent clean image. We harness a combination of the UW image formation process, the Retinex model, and constraints enforced by the scene geometry for our self-supervised training. To handle occlusions, we additionally utilize monocular frames from our video dataset and propose a masking scheme to prevent dynamic transients, that do not respect the underlying scene geometry, from misguiding the learning process. Evaluations on five LLUW datasets demonstrate the superiority and generalization ability of our proposed SelfLUID-Net over existing state-of-the-art methods. The dataset ULVStereo is available at https://github.com/nishavarghese15/ULVStereo.

When discussing the Aerial-Ground Person Re-identification (AGPReID) task, we face the main challenge of the significant appearance variations caused by different viewpoints, making identity matching difficult. To address this issue, previous methods attempt to reduce the differences between viewpoints by critical attributes and decoupling the viewpoints. While these methods can mitigate viewpoint differences to some extent, they still face two main issues: (1) difficulty in handling viewpoint diversity and (2) neglect of the contribution of local features. To effectively address these challenges, we design and implement the Self-Calibrating and Adaptive Prompt (SeCap) method for the AGPReID task. The core of this framework relies on the Prompt Re-calibration Module (PRM), which adaptively re-calibrates prompts based on the input. Combined with the Local Feature Refinement Module (LFRM), SeCap can extract view-invariant features from local features for AGPReID. Meanwhile, given the current scarcity of datasets in the AGPReID field, we further contribute two real-world Large-scale Aerial-Ground Person Re-Identification datasets, LAGPeR and G2APS-ReID. The former is collected and annotated by us independently, covering 4,231 unique identities and containing 63,841 high-quality images; the latter is reconstructed from the person search dataset G2APS. Through extensive experiments on AGPReID datasets, we demonstrate that SeCap is a feasible and effective solution for the AGPReID task.

The rapid growth of dataset scales has been a key driver in advancing deep learning research. However, as dataset scale increases, the training process becomes increasingly inefficient due to the presence of low-value samples, including excessive redundant samples, overly challenging samples, and inefficient easy samples that contribute little to model improvement. To address this challenge, we propose Scale Efficient Training (SeTa) for large datasets, a dynamic sample pruning approach that losslessly reduces training time. To remove low-value samples, SeTa first performs random pruning to eliminate redundant samples, then clusters the remaining samples according to their learning difficulty measured by loss. Building upon this clustering, a sliding window strategy is employed to progressively remove both overly challenging and inefficient easy clusters following an easy-to-hard curriculum. We conduct extensive experiments on large-scale synthetic datasets, including ToCa, SS1M, and ST+MJ, each containing over 3 million samples. SeTa reduces training costs by up to 50% while maintaining or improving performance, with minimal degradation even at 70% cost reduction. Furthermore, experiments on various scale real datasets across various backbones (including CNNs, Transformers, and Mambas) and diverse tasks (instruction tuning, multi-view stereo, geo-localization, composed image retrieval, referring image segmentation) demonstrate the powerful effectiveness and universality of our approach. Code is available at https://github.com/mrazhou/SeTa.

This paper proposes a new autoregressive model as an effective and scalable monocular depth estimator. Our idea is simple: We tackle the monocular depth estimation (MDE) task with an autoregressive prediction paradigm, based on two core designs. First, our depth autoregressive model (DAR) treats the depth map of different resolutions as a set of tokens, and conducts the low-to-high resolution autoregressive objective with a patch-wise casual mask. Second, our DAR recursively discretizes the entire depth range into more compact intervals, and attains the coarse-to-fine granularity autoregressive objective in an ordinal-regression manner. By coupling these two autoregressive objectives, our DAR establishes new state-of-the-art (SOTA) on KITTI and NYU Depth v2 by clear margins. Further, our scalable approach allows us to scale the model up to 2.0B and achieve the best RMSE of 1.799 on the KITTI dataset (5% improvement) compared to 1.896 by the current SOTA (Depth Anything). DAR further showcases zero-shot generalization ability on unseen datasets. These results suggest that DAR yields superior performance with an autoregressive prediction paradigm, providing a promising approach to equip modern autoregressive large models (e.g., GPT-4o) with depth estimation capabilities. Project page: https://depth-ar.github.io/

Semantic segmentation in urban scene analysis has mainly focused on images or point clouds, while textured meshes--offering richer spatial representation--remain underexplored. This paper introduces SUM Parts, the first large-scale dataset for urban textured meshes with part-level semantic labels, covering about 2.5km^2 with 21 classes. The dataset was created using our designed annotation tool, supporting both face and texture-based annotations with efficient interactive selection. We also provide a comprehensive evaluation of 3D semantic segmentation and interactive annotation methods on this dataset.

In this work, we study Weakly Supervised Spatio-Temporal Video Grounding (WSTVG), a challenging task of localizing subjects spatio-temporally in videos using only textual queries and no bounding box supervision. Inspired by recent advances in vision-language foundation models, we investigate their utility for WSTVG, leveraging their zero-shot grounding capabilities. However, we find that a simple adaptation lacks essential spatio-temporal grounding abilities. To bridge this gap, we introduce Tubelet Referral Grounding (TRG), which connects textual queries to tubelets to enable spatio-temporal predictions. Despite its promise, TRG struggles with compositional action understanding and dense scene scenarios. To address these limitations, we propose STPro, a progressive learning framework with two key modules: Sub-Action Temporal Curriculum Learning (SA-TCL), which incrementally builds compositional action understanding, and Congestion-Guided Spatial Curriculum Learning (CG-SCL), which adapts the model to complex scenes by spatially increasing task difficulty. STPro achieves state-of-the-art results on three benchmark datasets, with improvements of 1.0% on VidSTG-Declarative and 3.0% on HCSTVG-v1.

Recent efforts in Gaussian-Splat-based Novel View Synthesis can achieve photorealistic rendering; however, such capability is limited in sparse-view scenarios due to sparse initialization and over-fitting floaters. Recent progress in depth estimation and alignment can provide dense point cloud using few views; however, the resulting pose accuracy is suboptimal. In this work, we present SPARS3R, which combines the advantages of accurate pose estimation from Structure-from-Motion and dense point cloud from depth estimation. To this end, SPARS3R first performs a Global Fusion Alignment process that maps a prior dense point cloud to a sparse point cloud from Structure-from-Motion based on triangulated correspondences. RANSAC is applied during this process to distinguish inliers and outliers. SPARS3R then performs a second, Semantic Outlier Alignment step, which extracts semantically coherent regions around the outliers and performs local alignment in these regions. Along with several improvements in the evaluation process, we demonstrate that SPARS3R can achieve photorealistic rendering with sparse images and significantly outperforms existing approaches.

The emergence of Vision Language Models (VLMs) has brought unprecedented advances in understanding multimodal information. The combination of textual and visual semantics in VLMs is highly complex and diverse, making the safety alignment of these models challenging. Furthermore, due to the limited study on the safety alignment of VLMs, there is a lack of large-scale, high-quality datasets. To address these limitations, we propose a Safety Preference Alignment dataset for Vision Language Models named SPA-VL. In terms of breadth, SPA-VL covers 6 harmfulness domains, 13 categories, and 53 subcategories, and contains 100,788 samples of the quadruple (question, image, chosen response, rejected response). In terms of depth, the responses are collected from 12 open-source (e.g., QwenVL) and closed-source (e.g., Gemini) VLMs to ensure diversity. The construction of preference data is fully automated, and the experimental results indicate that models trained with alignment techniques on the SPA-VL dataset exhibit substantial improvements in harmlessness and helpfulness while maintaining core capabilities. SPA-VL, as a large-scale, high-quality, and diverse dataset, represents a significant milestone in ensuring that VLMs achieve both harmlessness and helpfulness.

Masked video modeling, such as VideoMAE, is an effective paradigm for video self-supervised learning (SSL). However, they are primarily based on reconstructing pixel level details on natural videos which have substantial temporal redundancy, limiting their capability for semantic representation and sufficient encoding of motion dynamics. To address these issues, this paper introduces a novel SSL approach for video representation learning, dubbed as SMILE, by infusing both spatial and motion semantics. In SMILE, we leverage image-language pretrained models, such as CLIP, to guide the learning process with their high-level spatial semantics. We enhance the representation of motion by introducing synthetic motion patterns in the training data, allowing the model to capture more complex and dynamic content. Furthermore, using SMILE, we establish a new self-supervised video learning paradigm capable of learning strong video representations without requiring any natural video data. We have carried out extensive experiments on 7 datasets with various downstream scenarios. SMILE surpasses current state-of-the-art SSL methods, showcasing its effectiveness in learning more discriminative and generalizable video representations.

We introduce SCFlow2, a plug-and-play refinement framework for 6D object pose estimation. Most recent 6D object pose methods rely on refinement to get accurate results. However, most existing refinements either suffer from noises in establishing correspondences, or rely on retraining for novel objects. SCFlow2 is based on the SCFlow model designed for iterative RGB refinement with shape constraint, but formulates the additional depth as a regularization in the iteration via 3D scene flow for RGBD frames. The key design of SCFlow2 is an introduction of geometry constraints into the training of recurrent match network, by combining the rigid-motion embeddings in 3D scene flow and 3D shape prior of the target. We train the refinement network on a combination of dataset Objaverse, GSO and ShapeNet, and demonstrate on BOP datasets with novel objects that, after using our method, the result of most state-of-the-art methods improves significantly, without any retraining or fine-tuning.

Driven by the increasing demand for accurate and efficient representation of 3D data in various domains, point cloud sampling has emerged as a pivotal research topic in 3D computer vision. Recently, learning-to-sample methods have garnered growing interest from the community, particularly for their ability to be jointly trained with downstream tasks. However, previous learning-based sampling methods either lead to unrecognizable sampling patterns by generating a new point cloud or biased sampled results by focusing excessively on sharp edge details. Moreover, they all overlook the natural variations in point distribution across different shapes, applying a similar sampling strategy to all point clouds. In this paper, we propose a Sparse Attention Map and Bin-based Learning method (termed SAMBLE) to learn shape-specific sampling strategies for point cloud shapes. SAMBLE effectively achieves an improved balance between sampling edge points for local details and preserving uniformity in the global shape, resulting in superior performance across multiple common point cloud downstream tasks, even in scenarios with few-point sampling.

Reliable embodied perception from an egocentric perspective is challenging yet essential for autonomous navigation technology of intelligent mobile agents. With the growing demand of social robotics, near-field scene understanding becomes an important research topic in the areas of egocentric perceptual tasks related to navigation in both crowded and unstructured environments. Due to the complexity of environmental conditions and difficulty of surrounding obstacles owing to truncation and occlusion, the perception capability under this circumstance is still inferior. To further enhance the intelligence of mobile robots, in this paper, we setup an egocentric multi-sensor data collection platform based on 3 main types of sensors (Camera, LiDAR and Fisheye), which supports flexible sensor configurations to enable dynamic sight of view from ego-perspective, capturing either near or farther areas. Meanwhile, a large-scale multimodal dataset is constructed, named RoboSense, to facilitate egocentric robot perception. Specifically, RoboSense contains more than 133K synchronized data with 1.4M 3D bounding box and IDs annotated in the full 360^ \circ view, forming 216K trajectories across 7.6K temporal sequences. It has 270xand 18xas many annotations of surrounding obstacles within near ranges as the previous datasets collected for autonomous driving scenarios such as KITTI and nuScenes. Moreover, we define a novel matching criterion for near-field 3D perception and prediction metrics. Based on RoboSense, we formulate 6 popular tasks to facilitate the future research development, where the detailed analysis as well as benchmarks are also provided accordingly. Data desensitization measures have been conducted for privacy protection.

We introduce RoadSocial, a large-scale, diverse VideoQA dataset tailored for generic road event understanding from social media narratives. Unlike existing datasets limited by regional bias, viewpoint bias and expert-driven annotations, RoadSocial captures the global complexity of road events with varied geographies, camera viewpoints (CCTV, handheld, drones) and rich social discourse. Our scalable semi-automatic annotation framework leverages Text LLMs and Video LLMs to generate comprehensive question-answer pairs across 12 challenging QA tasks, pushing the boundaries of road event understanding. RoadSocial is derived from social media videos spanning 14M frames and 414K social comments, resulting in a dataset with 13.2K videos, 674 tags and 260K high-quality QA pairs. We evaluate 18 Video LLMs (open-source and proprietary, driving-specific and general-purpose) on our road event understanding benchmark. We also demonstrate RoadSocial's utility in improving road event understanding capabilities of general-purpose Video LLMs.

Rip currents are strong, localized and narrow currents of water that flow outwards into the sea, causing numerous beach-related injuries and fatalities worldwide. Accurate identification of rip currents remains challenging due to their amorphous nature and the lack of annotated data, which often requires expert knowledge. To address these issues, we present RipVIS, a large-scale video instance segmentation benchmark explicitly designed for rip current segmentation. RipVIS is an order of magnitude larger than previous datasets, featuring 184 videos (212,328 frames), of which 150 videos (163,528 frames) are with rip currents, collected from various sources, including drones, mobile phones, and fixed beach cameras. Our dataset encompasses diverse visual contexts, such as wave-breaking patterns, sediment flows, and water color variations, across multiple global locations, including USA, Mexico, Costa Rica, Portugal, Italy, Greece, Romania, Sri Lanka, Australia and New Zealand. Most videos are annotated at 5 FPS to ensure accuracy in dynamic scenarios, supplemented by an additional 34 videos (48,800 frames) without rip currents. We conduct comprehensive experiments with Mask R-CNN, Cascade Mask R-CNN, SparseInst and YOLO11, fine-tuning these models for the task of rip current segmentation. Results are reported in terms of multiple metrics, with a particular focus on the F2 score to prioritize recall and reduce false negatives. To enhance segmentation performance, we introduce a novel post-processing step based on Temporal Confidence Aggregation (TCA). RipVIS aims to set a new standard for rip current segmentation, contributing towards safer beach environments. We offer a benchmark website to share data, models, and results with the research community, encouraging ongoing collaboration and future contributions, at https://ripvis.ai.

Recently, enhancing image quality in the original RAW domain has garnered significant attention, with denoising and reconstruction emerging as fundamental tasks. Although some works attempt to couple these tasks, they primarily focus on cascade learning while neglecting task associativity within a broader parameter space, leading to suboptimal performance. This work introduces a novel approach by rethinking denoising and reconstruction from a "backbone-head" perspective, leveraging the stronger shared parameter space offered by the backbone, compared to the encoder used in existing works. We derive task-specific heads with fewer parameters to mitigate learning pressure. By incorporating chromaticity-and-noise perception module into the backbone and introducing task-specific supervision during training, we enable simultaneous high-quality results for reconstruction and denoising. Additionally, we design a dual-head interaction module to capture the latent correspondence between the two tasks, significantly enhancing multi-task accuracy. Extensive experiments validate the superiority of the proposed method. Code is available at: https://github.com/csmty/CANS.

The Personalized Aesthetics Assessment (PAA) aims to accurately predict an individual's unique perception of aesthetics. With the surging demand for customization, PAA enables applications to generate personalized outcomes by aligning with individual aesthetic preferences. The prevailing PAA paradigm involves two stages: pre-training and fine-tuning, but it faces three inherent challenges: 1) The model is pre-trained using datasets of the Generic Aesthetics Assessment (GAA), but the collective preferences of GAA lead to conflicts in individualized aesthetic predictions. 2) The scope and stage of personalized surveys are related to both the user and the assessed object; however, the prevailing personalized surveys fail to adequately address assessed objects' characteristics. 3) During application usage, the cumulative multimodal feedback from an individual holds great value that should be considered for improving the PAA model but unfortunately attracts insufficient attention. To address the aforementioned challenges, we introduce a new PAA paradigm called PAA+, which is structured into three distinct stages: pre-training, fine-tuning, and continual learning. Furthermore, to better reflect individual differences, we employ a familiar and intuitive application, physique aesthetics assessment (PhysiqueAA), to validate the PAA+ paradigm. We propose a dataset called PhysiqueAA50K, consisting of over 50,000 annotated physique images. Furthermore, we develop a PhysiqueAA framework (PhysiqueFrame) and conduct a large-scale benchmark, achieving state-of-the-art (SOTA) performance. Our research is expected to provide an innovative roadmap and application for the PAA community. The code and dataset are available in here.

Category-level object pose estimation aims to determine the pose and size of arbitrary objects within given categories. Existing two-stage correspondence-based methods first establish correspondences between camera and object coordinates, and then acquire the object pose using a pose fitting algorithm. In this paper, we conduct a comprehensive analysis of this paradigm and introduce two crucial essentials: 1) shape-sensitive and pose-invariant feature extraction for accurate correspondence prediction, and 2) outlier correspondence removal for robust pose fitting. Based on these insights, we propose a simple yet effective correspondence-based method called SpotPose, which includes two stages. During the correspondence prediction stage, pose-invariant geometric structure of objects is thoroughly exploited to facilitate shape-sensitive holistic interaction among keypoint-wise features. During the pose fitting stage, outlier scores of correspondences are explicitly predicted to facilitate efficient identification and removal of outliers. Experimental results on CAMERA25, REAL275 and HouseCat6D benchmarks demonstrate that the proposed SpotPose outperforms state-of-the-art approaches by a large margin.

Physiological activities can be manifested by the sensitive changes in facial imaging. While they are barely observable to our eyes, computer vision manners can, and the derived remote photoplethysmography (rPPG) has shown considerable promise. However, existing studies mainly rely on spatial skin recognition and temporal rhythmic interactions, so they focus on identifying explicit features under ideal light conditions, but perform poorly in-the-wild with intricate obstacles and extreme illumination exposure. In this paper, we propose an end-to-end video transformer model for rPPG. It strives to eliminate complex and unknown external time-varying interferences, whether they are sufficient to occupy subtle biosignal amplitudes or exist as periodic perturbations that hinder network training. In the specific implementation, we utilize global interference sharing, subject background reference, and self-supervised disentanglement to eliminate interference, and further guide learning based on spatiotemporal filtering, reconstruction guidance, and frequency domain and biological prior constraints to achieve effective rPPG. To the best of our knowledge, this is the first robust rPPG model for real outdoor scenarios based on natural face videos, and is lightweight to deploy. Extensive experiments show the competitiveness and performance of our model in rPPG prediction across datasets and scenes.