CVPR 2025 Papers

Geometric reconstruction of opaque surfaces from images is a longstanding challenge in computer vision, with renewed interest from volumetric view synthesis algorithms using radiance fields. We leverage the geometry field proposed in recent work for stochastic opaque surfaces, which can then be converted to volume densities. We adapt Gaussian kernels or surfels to splat the geometry field rather than the volume, enabling precise reconstruction of opaque solids. Our first contribution is to derive an efficient and almost exact differentiable rendering algorithm for geometry fields parameterized by Gaussian surfels, while removing current approximations involving Taylor series and no self-attenuation. Next, we address the discontinuous loss landscape when surfels cluster near geometry, showing how to guarantee that the rendered color is a continuous function of the colors of the kernels, irrespective of ordering. Finally, we use latent representations with spherical harmonics encoded reflection vectors rather than spherical harmonics encoded colors to better address specular surfaces. We demonstrate significant improvement in the quality of reconstructed 3D surfaces on widely-used datasets.

In the last decade, we have witnessed the introduction of several novel deep neural network (DNN) architectures exhibiting ever-increasing performance across diverse tasks. Explaining the upward trend of their performance, however, remains difficult as different DNN architectures of comparable depth and width -- common factors associated with their expressive power -- may exhibit a drastically different performance even when trained on the same dataset. In this paper, we introduce the concept of the non-linearity signature of DNN, the first theoretically sound solution for approximately measuring the non-linearity of deep neural networks. Built upon a score derived from closed-form optimal transport mappings, this signature provides a better understanding of the inner workings of a wide range of DNN architectures and learning paradigms, with a particular emphasis on the computer vision task. We provide extensive experimental results that highlight the practical usefulness of the proposed non-linearity signature and its potential for long-reaching implications.

Vision Language Models (VLMs) like LLaVA encode images into tokens aligned to the word embedding space of the LLM decoder. Scaling input image resolution is essential for improving performance, especially in text-rich image understanding tasks. However, popular visual encoders such as CLIP-pretrained ViTs become inefficient at high resolutions due to the large number of tokens and high encoding latency caused by stacked self-attention layers. At different operational resolutions, the vision encoder of a VLM can be optimized along two axes: reducing encoding latency and minimizing the number of visual tokens passed to the LLM, thereby lowering overall latency. In this work, we introduce FastVLM, which achieves an optimized trade-off between resolution, latency, and accuracy by incorporating FastViTHD--a new hybrid vision encoder that outputs fewer tokens and significantly reduces encoding time while processing high-resolution images. We provide a comprehensive efficiency analysis of the interplay between image resolution, vision latency, number of visual tokens, and LLM size. In the LLaVA-1.5 setup, we achieve 3.2x improvement in overall time-to-first-token (TTFT) while maintaining similar performance on VLM benchmarks compared to prior works. On text-rich evaluations like TextVQA and DocVQA, FastVLM obtains +8.4% and +12.5% better accuracy than ConvLLaVA at a similar operating point of 144 visual tokens. Compared to LLaVa-OneVision at the highest resolution (1152 x 1152), FastVLM achieves comparable performance on key benchmarks like SeedBench and MMMU, using the same LLM, but with 85x faster TTFT, 3x less vision instruction tuning data, and a vision encoder that is 3.4x smaller.

Multi-view 3D reconstruction remains a core challenge in computer vision, particularly in applications requiring accurate and scalable representations across diverse perspectives. Current leading methods such as DUSt3R employ a fundamentally pairwise approach, processing images in pairs and necessitating costly global alignment procedures to reconstruct from multiple views. In this work, we propose Fast 3D Reconstruction (Fast3R), a novel multi-view generalization to DUSt3R that achieves efficient and scalable 3D reconstruction by processing many views in parallel. Fast3R's Transformer-based architecture forwards N images in a single forward pass, bypassing the need for iterative alignment. Through extensive experiments on camera pose estimation and 3D reconstruction, Fast3R demonstrates state-of-the-art performance, with significant improvements in inference speed and reduced error accumulation. These results establish Fast3R as a robust alternative for multi-view applications, offering enhanced scalability without compromising reconstruction accuracy. Project website: https://fast3r-3d.github.io

Classifiers are important components in many computer vision tasks, serving as the foundational backbone of a wide variety of models employed across diverse applications. However, understanding the decision-making process of classifiers remains a significant challenge. We propose DiffEx, a novel method that leverages the capabilities of text-to-image diffusion models to explain classifier decisions. Unlike traditional GAN-based explainability models, which are limited to simple, single-concept analyses and typically require training a new model for each classifier, our approach can explain classifiers that focus on single concepts (such as faces or animals) as well as those that handle complex scenes involving multiple concepts. DiffEx employs vision-language models to create a hierarchical list of semantics, allowing users to identify not only the overarching semantic influences on classifiers (e.g., the 'beard' semantic in a facial classifier) but also their sub-types, such as 'goatee' or 'Balbo' beard. Our experiments demonstrate that DiffEx is able to cover a significantly broader spectrum of semantics compared to its GAN counterparts, providing a hierarchical tool that delivers a more detailed and fine-grained understanding of classifier decisions.

Driven by the great success of Large Language Models (LLMs) in the 2D image domain, their applications in 3D scene understanding has emerged as a new trend. A key difference between 3D and 2D is that the situation of an egocentric observer in 3D scenes can change, resulting in different descriptions (e.g., "left" or "right"). However, current LLM-based methods overlook the egocentric perspective and simply use datasets from a global viewpoint. To address this issue, we propose a novel approach to automatically generate a situation-aware dataset by leveraging the scanning trajectory during data collection and utilizing Vision-Language Models (VLMs) to produce high-quality captions and question-answer pairs. Furthermore, we introduce a situation grounding module to explicitly predict the position and orientation of observer's viewpoint, thereby enabling LLMs to ground situation description in 3D scenes. We evaluate our approach on several benchmarks, demonstrating that our method effectively enhances the 3D situational awareness of LLMs while significantly expanding existing datasets and reducing manual effort.

Vision Language Models (VLMs) demonstrate significant potential as embodied AI agents for various mobility applications. However, a standardized, closed-loop benchmark for evaluating their spatial reasoning and sequential decision-making capabilities is lacking. To address this, we present MetaVQA: a comprehensive benchmark designed to assess and enhance VLMs' understanding of spatial relationships and scene dynamics through Visual Question Answering (VQA) and closed-loop simulations. MetaVQA leverages Set-of-Mark prompting and top-down view ground-truth annotations from nuScenes and Waymo datasets to automatically generate extensive question-answer pairs based on diverse real-world traffic scenarios, ensuring object-centric and context-rich instructions. Our experiments show that fine-tuning VLMs with the MetaVQA Dataset significantly improves their embodied scene understanding, which is evident not only in improved VQA accuracy but also in emerging safety-aware driving maneuvers. In addition, the learning exhibits strong transferability from simulation to real-world observation. The project webpage is at https://metadriverse.github.io/metavqa.

While multimodal large language models demonstrate strong performance in complex reasoning tasks, they pose significant challenges related to model complexity during deployment, especially for resource-limited devices. In this paper, we propose an automatic pruning method for large vision-language models to enhance the efficiency of multimodal reasoning. Conventional methods rely on the training data of the original model to select the proper pruning ratio for different network components. However, these methods are impractical for large vision-language models due to the unaffordable search costs caused by web-scale training corpus. In contrast, our approach only leverages a small number of samples to search for the desired pruning policy by maximizing its generalization ability on unknown training data while maintaining the model accuracy, which enables the achievement of an optimal trade-off between accuracy and efficiency for large visual language models. Specifically, we formulate the generalization gap of the pruning strategy using the structural risk minimization principle. Based on both task performance and generalization capability, we iteratively search for the optimal pruning policy within a given search space and optimize the vision projector to evolve the search space with higher upper bound of performance. We conduct extensive experiments on the ScienceQA, Vizwiz, MM-vet, and LLaVA-Bench datasets for the task of visual question answering. Using only 64 samples for pruning policy search, EfficientLLaVA achieves an accuracy of 83.05% on ScienceQA, along with a x1.8 speedup compared to the dense LLaVA-v1.5-7B model.

Finding correspondences between 3D shapes is a crucial problem in computer vision and graphics. While most research has focused on finding correspondences in settings where at least one of the shapes is complete, the realm of partial-to-partial shape matching remains under-explored. Yet, it is important since in many applications shapes are only observed partially due to occlusion or scanning. Finding correspondences between partial shapes comes with an additional challenge: We not only want to identify correspondences between points on either shape but also have to determine which points of each shape actually have a partner. To tackle this challenging problem, we present EchoMatch, a novel framework for partial-to-partial shape matching that incorporates the concept of correspondence reflection to enable an overlap prediction within a functional map framework. With this approach, we show that we can outperform current SOTA methods in challenging partial-to-partial shape matching problems. Our code is available at https://echo-match.github.io.

Shape completion, a crucial task in 3D computer vision, involves predicting and filling the missing regions of scanned or partially observed objects. Current methods expect known pose or canonical coordinates and do not perform well under varying rotations, limiting their real-world applicability. We introduce ESCAPE (Equivariant Shape Completion via Anchor Point Encoding), a novel framework designed to achieve rotation-equivariant shape completion. Our approach employs a distinctive encoding strategy by selecting anchor points from a shape and representing all points as a distance to all anchor points. This enables the model to capture a consistent, rotation-equivariant understanding of the object's geometry. ESCAPE leverages a transformer architecture to encode and decode the distance transformations, ensuring that generated shape completions remain accurate and equivariant under rotational transformations. Subsequently, we perform optimization to calculate the predicted shapes from the encodings. Experimental evaluations demonstrate that ESCAPE achieves robust, high-quality reconstructions across arbitrary rotations and translations, showcasing its effectiveness in real-world applications without additional pose estimation modules.

Large multimodal models (LMMs) have achieved impressive progress in vision-language understanding, yet they face limitations in real-world applications requiring complex reasoning over a large number of images. Existing benchmarks for multi-image question-answering are limited in scope, each question is paired with only up to 30 images, which does not fully capture the demands of large-scale retrieval tasks encountered in the real-world usages. To reduce these gaps, we introduce two document haystack benchmarks, dubbed DocHaystack and InfoHaystack, designed to evaluate LMM performance on large-scale visual document retrieval and understanding. Additionally, we propose V-RAG, a novel, vision-centric retrieval-augmented generation (RAG) framework that leverages a suite of multimodal vision encoders, each optimized for specific strengths, and a dedicated question-document relevance module. V-RAG sets a new standard, with a 9% and 11% improvement in Recall@1 on the challenging DocHaystack-1000 and InfoHaystack-1000 benchmarks, respectively, compared to the previous best baseline models. Additionally, integrating V-RAG with LMMs enables them to efficiently operate across thousands of images, yielding significant improvements on our DocHaystack and InfoHaystack benchmarks. Our code and datasets will be made publicly available.

Computer vision foundation models, such as DINO or OpenCLIP, are trained in a self-supervised manner on large image datasets. Analogously, substantial evidence suggests that the human visual system (HVS) is influenced by the statistical distribution of colors and patterns in the natural world, characteristics also present in the training data of foundation models. The question we address in this paper is whether foundation models trained on natural images mimic some of the low-level characteristics of the human visual system, such as contrast detection, contrast masking, and contrast constancy. Specifically, we designed a protocol comprising nine test types to evaluate the image encoders of 45 foundation and generative models. Our results indicate that some foundation models (e.g., DINO, DINOv2, and OpenCLIP), share some of the characteristics of human vision, but other models show little resemblance. Foundation models tend to show smaller sensitivity to low contrast and rather irregular responses to contrast across frequencies. The foundation models show the best agreement with human data in terms of contrast masking. Our findings suggest that human vision and computer vision may take both similar and different paths when learning to interpret images of the real world. Overall, while differences remain, foundation models trained on vision tasks start to align with low-level human vision, with DINOv2 showing the closest resemblance. Our code is available on https://github.com/caiyancheng/VFM_HVS_CVPR2025.

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 .

Recently, Vision Graph Neural Network (ViG) has gained considerable attention in computer vision. Despite its groundbreaking innovation, Vision Graph Neural Network encounters key issues including the quadratic computational complexity caused by its K-Nearest Neighbor (KNN) graph construction and the limitation of pairwise relations of normal graphs. To address the aforementioned challenges, we propose a novel vision architecture, termed Dilated Vision HyperGraph Neural Network (DVHGNN), which is designed to leverage multi-scale hypergraph to efficiently capture high-order correlations among objects. Specifically, the proposed method tailors Clustering and Dilated HyperGraph Construction (DHGC) to adaptively capture multi-scale dependencies among the data samples. Furthermore, a dynamic hypergraph convolution mechanism is proposed to facilitate adaptive feature exchange and fusion at the hypergraph level. Extensive qualitative and quantitative evaluations of the benchmark image datasets demonstrate that the proposed DVHGNN significantly outperforms the state-of-the-art vision backbones. For instance, our DVHGNN-S achieves an impressive top-1 accuracy of 83.1% on ImageNet-1K, surpassing ViG-S by +1.0% and ViHGNN-S by +0.6%.

The Base-New Trade-off (BNT) problem universally exists during the optimization of CLIP-based prompt tuning, where continuous fine-tuning on base (target) classes leads to a simultaneous decrease of generalization ability on new (unseen) classes. Existing approaches attempt to regulate the prompt tuning process to balance BNT by appending constraints. However, imposed on the same target prompt, these constraints fail to fully avert the mutual exclusivity between the optimization directions for base and new. As a novel solution to this challenge, we propose the plug-and-play Dual-Prompt Collaboration (DPC) framework, the first that decoupling the optimization processes of base and new tasks at the prompt level. Specifically, we clone a learnable parallel prompt based on the backbone prompt, and introduce a variable Weighting-Decoupling framework to independently control the optimization directions of dual prompts specific to base or new tasks, thus avoiding the conflict in generalization. Meanwhile, we propose a Dynamic Hard Negative Optimizer, utilizing dual prompts to construct a more challenging optimization task on base classes for enhancement. For interpretability, we prove the feature channel invariance of the prompt vector during the optimization process, providing theoretical support for the Weighting-Decoupling of DPC. Extensive experiments on multiple backbones demonstrate that DPC can significantly improve base performance without introducing any external knowledge beyond the base classes, while maintaining generalization to new classes. Code is available at: https://github.com/JREion/DPC.

In large vision-language models (LVLMs), images serve as inputs that carry a wealth of information. As the idiom "A picture is worth a thousand words" implies, representing a single image in current LVLMs can require hundreds or even thousands of tokens. This results in significant computational costs, which grow quadratically as input image resolution increases, thereby severely impacting the efficiency. Previous approaches have attempted to reduce the number of image tokens either before or within the early layers of LVLMs. However, these strategies inevitably result in the loss of crucial image information. To address this challenge, we conduct an empirical study revealing that all visual tokens are necessary for LVLMs in the shallow layers, and token redundancy progressively increases in the deeper layers.To this end, we propose ViCo, a conical-style visual concentration strategy for LVLMs to boost their efficiency in both training and inference with neglectable performance loss. Specifically, we partition the LVLM into several stages and drop part of the image tokens at the end of each stage with a pre-defined ratio. The dropping is based on a lightweight similarity calculation with a negligible time overhead. Extensive experiments demonstrate that ViCo can achieve over 40% training time reduction and 55% inference FLOPs acceleration on leading LVLMs like LLaVA-NeXT, maintaining comparable multi-modal performance. Besides, ViCo can also serve as a plug-and-play strategy to accelerate inference in a free way, with better performance and lower inference cost than counterparts.

Vision-Language Models (VLMs) have recently witnessed significant progress in visual comprehension. As the permitting length of image context grows, VLMs can now comprehend a broader range of views and spaces. Current benchmarks provide insightful analysis of VLMs in tasks involving complex visual instructions following, multi-image understanding and spatial reasoning. However, they usually focus on spatially irrelevant images or discrete images captured from varied viewpoints. The compositional characteristic of images captured from a static viewpoint remains underestimated. We term this characteristic as Continuous Space Perception. When observing a scene from a static viewpoint while shifting orientations, it produces a series of spatially continuous images, enabling the reconstruction of the entire space. In this paper, we present CoSpace, a multi-image visual understanding benchmark designed to assess the Continuous Space perception ability for VLMs. CoSpace contains 2,918 images and 1,626 question-answer pairs, covering seven types of tasks. We conduct evaluation across 19 proprietary and open-source VLMs. Results reveal that there exist pitfalls on the continuous space perception ability for most of the evaluated models, including proprietary ones. Interestingly, we find that the main discrepancy between open-source and proprietary models lies not in accuracy but in the consistency of responses. We believe that enhancing the ability of continuous space perception is essential for VLMs to perform effectively in real-world tasks and encourage further research to advance this capability.

3D Vision Grounding (3DVG) is a fundamental research area that enables agents to perceive and interact with the 3D world. The challenge of the 3DVG task lies in understanding fine-grained semantics and spatial relationships within both the utterance and 3D scene. To address this challenge, we propose a zero-shot neuro-symbolic framework that utilizes a large language model (LLM) as neuro-symbolic functions to ground the object within the 3D Gaussian Splatting (3DGS) representation. By utilizing 3DGS representation, we can dynamically render high-quality 2D images from various viewpoints to enrich the semantic information. Given the complexity of spatial relationships, we construct a relationship graph and chain of semantics that decouple spatial relationships and facilitate step-by-step reasoning within 3DGS representation. Additionally, we employ a grounded-aware self-check mechanism to enable the LLM to reflect on its responses and mitigate the effects of ambiguity in spatial reasoning. We evaluate our method using two publicly available datasets, Nr3D and Sr3D, achieving accuracies of 60.8% and 91.4%, respectively. Notably, our method surpasses current state-of-the-art zero-shot methods on the Nr3D dataset. In addition, it outperforms the recent supervised models on the Sr3D dataset.

Pre-trained vision-language models (VLMs) have showcased remarkable performance in image and natural language understanding, such as image captioning and response generation. As the practical applications of VLMs become increasingly widespread, their potential safety and robustness issues raise concerns that adversaries may evade the system and cause these models to generate toxic content through malicious attacks. Therefore, evaluating the robustness of open-source VLMs against adversarial attacks has garnered growing attention, with transfer-based attacks as a representative black-box attacking strategy. However, most existing transfer-based attacks neglect the importance of the semantic correlations between vision and text modalities, leading to sub-optimal adversarial example generation and attack performance. To address this issue, we present Chain of Attack (CoA), which iteratively enhances the generation of adversarial examples based on the multi-modal semantic update using a series of intermediate attacking steps, achieving superior adversarial transferability and efficiency. A unified attack success rate computing method is further proposed for automatic evasion evaluation. Extensive experiments conducted under the most realistic and high-stakes scenario, demonstrate that our attacking strategy is able to effectively mislead models to generate targeted responses using only black-box attacks without any knowledge of the victim models. The comprehensive robustness evaluation in our paper provides insight into the vulnerabilities of VLMs and offers a reference for the safety considerations of future model developments.