CVPR 2025 论文列表

Large-scale pre-trained vision models are becoming increasingly prevalent, offering expressive and generalizable visual representations that benefit various downstream tasks. Recent studies on the emergent properties of these models have revealed their high-level geometric understanding, in particular in the context of depth perception. However, it remains unclear how depth perception arises in these models without explicit depth supervision provided during pre-training. To investigate this, we examine whether the monocular depth cues, similar to those used by the human visual system, emerge in these models. We introduce a new benchmark, DepthCues, designed to evaluate depth cue understanding, and present findings across 20 diverse and representative pre-trained vision models. Our analysis shows that human-like depth cues emerge in more recent larger models. We also explore enhancing depth perception in large vision models by fine-tuning on DepthCues, and find that even without dense depth supervision, this improves depth estimation. To support further research, our benchmark and evaluation code will be made publicly available for studying depth perception in vision models.

Recent advancements in generalizable novel view synthesis have achieved impressive quality through interpolation between nearby views. However, rendering high-resolution images remains computationally intensive due to the need for dense sampling of all rays. Recognizing that natural scenes are typically piecewise smooth and sampling all rays is often redundant, we propose a novel depth-guided bundle sampling strategy to accelerate rendering. By grouping adjacent rays into a bundle and sampling them collectively, a shared representation is generated for decoding all rays within the bundle. To further optimize efficiency, our adaptive sampling strategy dynamically allocates samples based on depth confidence, concentrating more samples in complex regions while reducing them in smoother areas. When applied to ENeRF, our method achieves up to a 1.27 dB PSNR improvement and a 47% increase in FPS on the DTU dataset. Extensive experiments on synthetic and real-world datasets demonstrate state-of-the-art rendering quality and up to 2x faster rendering compared to existing generalizable methods. Code is available at https://github.com/KLMAV-CUC/GDB-NeRF.

While recent depth foundation models exhibit strong zero-shot generalization, achieving accurate metric depth across diverse camera types--particularly those with large fields of view (FoV) such as fisheye and 360-degree cameras--remains a significant challenge. This paper presents Depth Any Camera (DAC), a powerful zero-shot metric depth estimation framework that extends a perspective-trained model to effectively handle cameras with varying FoVs. The framework is designed to ensure that all existing 3D data can be leveraged, regardless of the specific camera types used in new applications. Remarkably, DAC is trained exclusively on perspective images but generalizes seamlessly to fisheye and 360-degree cameras without the need for specialized training data. DAC employs Equi-Rectangular Projection (ERP) as a unified image representation, enabling consistent processing of images with diverse FoVs. Its key components include a pitch-aware Image-to-ERP conversion for efficient online augmentation in ERP space, a FoV alignment operation to support effective training across a wide range of FoVs, and multi-resolution data augmentation to address resolution disparities between training and testing. DAC achieves state-of-the-art zero-shot metric depth estimation, improving delta-1 accuracy by up to 50% on multiple fisheye and 360-degree datasets compared to prior metric depth foundation models, demonstrating robust generalization across camera types.

We present Dense-SfM, a novel Structure from Motion (SfM) framework designed for dense and accurate 3D reconstruction from multi-view images. Sparse keypoint matching, which traditional SfM methods often rely on, limits both accuracy and point density, especially in texture-less areas. Dense-SfM addresses this limitation by integrating dense matching with a Gaussian Splatting (GS) based track extension which gives more consistent, longer feature tracks. To further improve reconstruction accuracy, Dense-SfM is equipped with a multi-view kernelized matching module leveraging transformer and Gaussian Process architectures, for robust track refinement across multi-views. Evaluations on the ETH3D and Texture-Poor SfM datasets show that Dense-SfM offers significant improvements in accuracy and density over state-of-the-art methods. Project page: https://icetea-cv.github.io/densesfm/.

In this paper, we speed up robust two-view relative pose from dense correspondences. Previous work has shown that dense matchers can significantly improve both accuracy and robustness in the resulting pose. However, the large number of matches comes with a significantly increased runtime during robust estimation in RANSAC. To avoid this, we propose an efficient match summarization scheme which provides comparable accuracy to using the full set of dense matches, while having 10-100x faster runtime. We validate our approach on standard benchmark datasets together with multiple state-of-the-art dense matchers.

Hyperspectral 3D imaging captures both depth maps and hyperspectral images, enabling comprehensive geometric and material analysis. Recent methods achieve high spectral and depth accuracy; however, they require long acquisition times--often over several minutes--or rely on large, expensive systems, restricting their use to static scenes. We present Dense Dispersed Structured Light (DDSL), an accurate hyperspectral 3D imaging method for dynamic scenes that utilizes stereo RGB cameras and an RGB projector equipped with an affordable diffraction grating film. We design spectrally multiplexed DDSL patterns that significantly reduce the number of required projector patterns, thereby accelerating acquisition speed. Additionally, we formulate an image formation model and a reconstruction method to estimate a hyperspectral image and depth map from captured stereo images. As the first practical and accurate hyperspectral 3D imaging method for dynamic scenes, we experimentally demonstrate that DDSL achieves a spectral resolution of 15.5 nm full width at half maximum (FWHM), 4 mm depth error, and 6.6 fps.

Estimating correspondences between pairs of deformable shapes remains a challenging problem. Despite substantial progress, existing methods lack broad generalization capabilities and require category-specific training data. To address these limitations, we propose a fundamentally new approach to shape correspondence based on denoising diffusion models. In our method, a diffusion model learns to directly predict the functional map, a low-dimensional representation of a point-wise map between shapes. We use a large dataset of synthetic human meshes for training and employ two steps to reduce the number of functional maps that need to be learned. First, the maps refer to a template rather than shape pairs. Second, the functional map is defined in a basis of eigenvectors of the Laplacian, which is not unique due to sign ambiguity. Therefore, we introduce an unsupervised approach to select a specific basis by correcting the signs of eigenvectors based on surface features. Our model achieves competitive performance on standard human datasets, meshes with anisotropic connectivity, non-isometric humanoid shapes, as well as animals compared to existing descriptor-based and large-scale shape deformation methods.

In recent years, large transformer-based video encoder models have greatly advanced state-of-the-art performance on video classification tasks. However, these large models typically process videos by averaging embedding outputs from multiple clips over time to produce fixed-length representations. This approach fails to account for a variety of time-related features, such as variable video durations, chronological order of events, and temporal variance in feature significance. While methods for temporal modeling do exist, they often require significant architectural changes and expensive retraining, making them impractical for off-the-shelf, fine-tuned large encoders. To overcome these limitations, we propose DejaVid, an encoder-agnostic method that enhances model performance without the need for retraining or altering the architecture. Our framework converts a video into a variable-length temporal sequence of embeddings (TSE). A TSE naturally preserves temporal order and accommodates variable video durations. We then learn per-timestep, per-feature weights over the encoded TSE frames, allowing us to account for variations in feature importance over time. We introduce a new neural network architecture inspired by traditional time series alignment algorithms for this learning task. Our evaluation demonstrates that DejaVid substantially improves the performance of a state-of-the-art large encoder, achieving leading Top-1 accuracy of 77.2% on Something-Something V2, 89.1% on Kinetics-400, and 88.6% on HMDB51, while adding fewer than 1.8% additional learnable parameters and requiring less than 3 hours of training time. Our code is available at https://github.com/darrylho/DejaVid.

All-in-one image restoration aims to recover clear images from various degradation types and levels with a unified model. Nonetheless, the significant variations among degradation types present challenges for training a universal model, often resulting in task interference, where the gradient update directions of different tasks may diverge due to shared parameters. To address this issue, motivated by the routing strategy, we propose DFPIR, a novel all-in-one image restorer that introduces Degradation-aware Feature Perturbations(DFP) to adjust the feature space to align with the unified parameter space. In this paper, the feature perturbations primarily include channel-wise perturbations and attention-wise perturbations. Specifically, channel-wise perturbations are implemented by shuffling the channels in high-dimensional space guided by degradation types, while attention-wise perturbations are achieved through selective masking in the attention space. To achieve these goals, we propose a Degradation-Guided Perturbation Block (DGPB) to implement these two functions, positioned between the encoding and decoding stages of the encoder-decoder architecture. Extensive experimental results demonstrate that DFPIR achieves state-of-the-art performance on several all-in-one image restoration tasks including image denoising, image dehazing, image deraining, motion deblurring, and low-light image enhancement. Our codes are available at https://github.com/TxpHome/DFPIR.

Recently, Gaussian splatting has emerged as a robust technique for representing 3D scenes, enabling real-time rasterization and high-fidelity rendering. However, Gaussians' inherent radial symmetry and smoothness constraints limit their ability to represent complex shapes, often requiring thousands of primitives to approximate detailed geometry. We introduce Deformable Radial Kernel (DRK), which extends Gaussian splatting into a more general and flexible framework. Through learnable radial bases with adjustable angles and scales, DRK efficiently models diverse shape primitives while enabling precise control over edge sharpness and boundary curvature. iven DRK's planar nature, we further develop accurate ray-primitive intersection computation for depth sorting and introduce efficient kernel culling strategies for improved rasterization efficiency. Extensive experiments demonstrate that DRK outperforms existing methods in both representation efficiency and rendering quality, achieving state-of-the-art performance while dramatically reducing primitive count.

In the field of 3D medical imaging, accurately extracting and representing the blood vessels with curvilinear structures holds paramount importance for clinical diagnosis. Previous methods have commonly relied on discrete representation like mask, often resulting in local fractures or scattered fragments due to the inherent limitations of the per-pixel classification paradigm. In this work, we introduce DeformCL, a new continuous representation based on Deformable Centerlines, where centerline points act as nodes connected by edges that capture spatial relationships. Compared with previous representations, DeformCL offers three key advantages: natural connectivity, noise robustness, and interaction facility. We present a comprehensive training pipeline structured in a cascaded manner to fully exploit these favorable properties of DeformCL. Extensive experiments on four 3D vessel segmentation datasets demonstrate the effectiveness and superiority of our method. Furthermore, the visualization of curved planar reformation images validates the clinical significance of the proposed framework.

Recently, state space models (SSM), particularly Mamba, have attracted significant attention from scholars due to their ability to effectively balance computational efficiency and performance. However, most existing visual Mamba methods flatten images into 1D sequences using predefined scan orders, which results the model being less capable of utilizing the spatial structural information of the image during the feature extraction process. To address this issue, we proposed a novel visual foundation model called DefMamba. This model includes a multi-scale backbone structure and deformable mamba (DM) blocks, which dynamically adjust the scanning path to prioritize important information, thus enhancing the capture and processing of relevant input features. By combining a deformable scanning (DS) strategy, this model significantly improves its ability to learn image structures and detects changes in object details. Numerous experiments have shown that DefMamba achieves state-of-the-art performance in various visual tasks, including image classification, object detection, instance segmentation, and semantic segmentation. The code is open source on DefMamba.

Due to the irregular and disordered data structure in 3D point clouds, prior works have focused on designing more sophisticated local representation methods to capture these complex local patterns. However, the recognition performance has saturated over the past few years, indicating that increasingly complex and redundant designs no longer make improvements to local learning. This phenomenon prompts us to diverge from the trend in 3D vision and instead pursue an alternative and successful solution: deeper neural networks. In this paper, we propose DeepLA-Net, a series of very deep networks for point cloud analysis. The key insight of our approach is to exploit a small but mighty local learning block, which reduces 10xfewer FLOPs, enabling the construction of very deep networks. Furthermore, we design a training supervision strategy to ensure smooth gradient backpropagation and optimization in very deep networks. We construct the DeepLA-Net family with a depth of up to 120 blocks --- at least 5xdeeper than recent methods --- trained on a single RTX 3090. An ensemble of the DeepLA-Net achieves state-of-the-art performance on classification and segmentation tasks of S3DIS Area5 (+2.2% mIoU), ScanNet test set (+1.6% mIoU), ScanObjectNN (+2.1% OA), and ShapeNetPart (+0.9% cls.mIoU).

Vision Transformers (ViTs) excel in tackling complex vision tasks, yet their substantial size poses significant challenges for applications on resource-constrained edge devices. The increased size of these models leads to higher overhead (e.g., energy, latency) when transmitting model weights between the edge device and the server. Hence, ViTs are not ideal for edge devices where the entire model may not fit on the device. Current model compression techniques often achieve high compression ratios at the expense of performance degradation, particularly for ViTs. To overcome the limitations of existing works, we rethink model compression strategy for ViTs from first principle approach and develop an orthogonal strategy called \method. The objective of the \method is to encode the model weights to a highly compressed encoded representation using a novel training method, denoted as Unified Compression Training (UCT). Proposed UCT is accompanied by a decoding mechanism during inference, which helps to gain any loss of accuracy due to high compression ratio. We further optimize this decoding step by re-ordering the decoding operation using associative property of matrix multiplication, ensuring that the compressed weights can be decoded during inference without incurring any computational overhead. Our extensive experiments across multiple ViT models on modern edge devices show that \method can successfully compress ViTs at high compression ratios (>14x). \method enables the entire model to be stored on edge device, resulting in unprecedented reductions in energy consumption (>1470x) and latency (>68x) for edge ViT inference. Our code is available at \href https://github.com/ML-Security-Research-LAB/DeepCompress-ViT https://github.com/ML-Security-Research-LAB/DeepCompress-ViT .

In this paper, we consider the conditional generation problem by guiding off-the-shelf unconditional diffusion models with differentiable loss functions in a plug-and-play fashion. While previous research has primarily focused on balancing the unconditional diffusion model and the guided loss through a tuned weight hyperparameter, we propose a novel framework that distinctly decouples these two components. Specifically, we introduce two variables x and z , to represent the generated samples governed by the unconditional generation model and the guidance function, respectively. This decoupling reformulates conditional generation into two manageable subproblems, unified by the constraint x = z . Leveraging this setup, we develop a new algorithm based on the Alternating Direction Method of Multipliers (ADMM) to adaptively balance these components. Additionally, we establish the equivalence between the diffusion reverse step and the proximal operator of ADMM and provide a detailed convergence analysis of our algorithm under certain mild assumptions. Our experiments demonstrate that our proposed method \OurMethod consistently generates high-quality samples while ensuring strong adherence to the conditioning criteria. It outperforms existing methods across a range of conditional generation tasks, including image generation with various guidance and controllable motion synthesis.

Recovering high-quality depth maps from compressed sources has gained significant attention due to the limitations of consumer-grade depth cameras and the bandwidth restrictions during data transmission. However, current methods still suffer from two challenges. First, bit-depth compression produces a uniform depth representation in regions with subtle variations, hindering the recovery of detailed information. Second, densely distributed random noise reduces the accuracy of estimating the global geometric structure of the scene. To address these challenges, we propose a novel framework, termed geometry-decoupled network (GDNet), for compressed depth map super-resolution that decouples the high-quality depth map reconstruction process by handling global and detailed geometric features separately. To be specific, we propose the fine geometry detail encoder (FGDE), which is designed to aggregate fine geometry details in high-resolution low-level image features while simultaneously enriching them with complementary information from low-resolution context-level image features. In addition, we develop the global geometry encoder (GGE) that aims at suppressing noise and extracting global geometric information effectively via constructing compact feature representation in a low-rank space. We conduct experiments on multiple benchmark datasets, demonstrating that our GDNet significantly outperforms current methods in terms of geometric consistency and detail recovery. In the ECCV 2024 AIM Compressed Depth Upsampling Challenge, our solution won the 1st place award. Our codes are available at: https://github.com/Ian0926/GDNet.

We present DecoupledGaussian, a novel system that decouples static objects from their contacted surfaces captured in-the-wild videos, a key prerequisite for realistic Newtonian-based physical simulations. Unlike prior methods focused on synthetic data or elastic jittering along the contact surface, which prevent objects from fully detaching or moving independently, DecoupledGaussian allows for significant positional changes without being constrained by the initial contacted surface. Recognizing the limitations of current 2D inpainting tools for restoring 3D locations, our approach uses joint Poisson fields to repair and expand the Gaussians of both objects and contacted scenes after separation. This is complemented by a multi-carve strategy to refine the object's geometry. Our system enables realistic simulations of decoupling motions, collisions, and fractures driven by user-specified impulses, supporting complex interactions within and across multiple scenes. We validate DecoupledGaussian through a comprehensive user study and quantitative benchmarks. This system enhances digital interaction with objects and scenes in real-world environments, benefiting industries such as VR, robotics, and autonomous driving. Our project page is at: https://wangmiaowei.github.io/DecoupledGaussian.github.io/.

Motion expression video segmentation aims to segment objects based on input motion descriptions. Compared with traditional referring video object segmentation, it focuses on motion and multi-object expressions and is more challenging. Previous works achieved it by simply injecting text information into the video instance segmentation (VIS) model. However, this requires retraining the entire model and optimization is difficult. In this work, we propose DMVS, a simple framework constructed on the existing query-based VIS model, emphasizing decoupling the task into video instance segmentation and motion expression understanding. Firstly, we use a frozen video instance segmenter to extract object-specific contexts and convert them into frame-level and video-level queries. Secondly, we interact two levels of queries with static and motion cues, respectively, to further encode visually enhanced motion expressions. Furthermore, we propose a novel query initialization strategy that uses video queries guided by classification priors to initialize motion queries, greatly reducing the difficulty of optimization. Without bells and whistles, DMVS achieves state-of-the-art performance on the MeViS dataset at a lower training cost. Extensive experiments verify the effectiveness and efficiency of our framework.

Diffusion models are trained by learning a sequence of models that reverse each step of noise corruption. Typically, the model parameters are fully shared across multiple timesteps to enhance training efficiency. However, since the denoising tasks differ at each timestep, the gradients computed at different timesteps may conflict, potentially degrading the overall performance of image generation. To solve this issue, this work proposes a Decouple-then-Merge (DeMe) framework, which begins with a pretrained model and finetunes separate models tailored to specific timesteps. We introduce several improved techniques during the finetuning stage to promote effective knowledge sharing while minimizing training interference across timesteps. Finally, after finetuning, these separate models can be merged into a single model in the parameter space, ensuring efficient and practical inference. Experimental results show significant generation quality improvements upon 6 benchmarks including Stable Diffusion on COCO30K, ImageNet1K, PartiPrompts, and DDPM on LSUN Church, LSUN Bedroom, and CIFAR10. Code is included in the supplementary material and will be released on Github.