CVPR 2025 论文列表

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.

Ensuring adherence to traffic sign regulations is essential for both human and autonomous vehicle navigation. While current online mapping solutions often prioritize the construction of the geometric and connectivity layers of HD maps, overlooking the construction of the traffic regulation layer within HD maps. Addressing this gap, we introduce MapDR, a novel dataset designed for the extraction of Driving Rules from traffic signs and their association with vectorized, locally perceived HD Maps. MapDR features over 10,000 annotated video clips that capture the intricate correlation between traffic sign regulations and lanes. Built upon this benchmark and the newly defined task of integrating traffic regulations into online HD maps, we provide modular and end-to-end solutions: VLE-MEE and RuleVLM, offering a strong baseline for advancing autonomous driving technology. It fills a critical gap in the integration of traffic sign rules, contributing to the development of reliable autonomous driving systems.

Multimodal large language models (MLLMs) possess the ability to comprehend visual images or videos, and show impressive reasoning ability thanks to the vast amounts of pretrained knowledge, making them highly suitable for autonomous driving applications. Unlike the previous work, DriveGPT4-V1, which focused on open-loop tasks, this study explores the capabilities of LLMs in enhancing closed-loop autonomous driving. DriveGPT4-V2 processes camera images and vehicle states as input to generate low-level control signals for end-to-end vehicle operation. A multi-view visual tokenizer (MV-VT) is employed enabling DriveGPT4-V2 to perceive the environment with an extensive range while maintaining critical details. The model architecture has been refined to improve decision prediction and inference speed. To further enhance the performance, an additional expert LLM is trained for online imitation learning. The expert LLM, sharing a similar structure with DriveGPT4-V2, can access privileged information about surrounding objects for more robust and reliable predictions. Experimental results show that DriveGPT4-V2 outperforms all baselines on the challenging CARLA Longest6 benchmark. The code and data of DriveGPT4-V2 will be publicly available.

Vision-and-Language Navigation (VLN) agents are tasked with navigating an unseen environment using natural language instructions. In this work, we study if visual representations of sub-goals implied by the instructions can serve as navigational cues and lead to increased navigation performance. To synthesize these visual representations or "imaginations", we leverage a text-to-image diffusion model on landmark references contained in segmented instructions. These imaginations are provided to VLN agents as an added modality to act as landmark cues and an auxiliary loss is added to explicitly encourage relating these with their corresponding referring expressions. Our findings reveal an increase in success rate (SR) of ~1 point and up to ~0.5 points in success scaled by inverse path length (SPL) across agents. These results suggest that the proposed approach reinforces visual understanding compared to relying on language instructions alone.

In this paper, we examine a key limitation in query-based detectors for temporal action detection (TAD), which arises from their direct adaptation of originally designed architectures for object detection. Despite the effectiveness of the existing models, they struggle to fully address the unique challenges of TAD, such as the redundancy in multi-scale features and the limited ability to capture sufficient temporal context. To address these issues, we propose a multi-dilated gated encoder and central-adjacent region integrated decoder for temporal action detection transformer (DiGIT). Our approach replaces the existing encoder that consists of multi-scale deformable attention and feedforward network with our multi-dilated gated encoder. Our proposed encoder reduces the redundant information caused by multi-level features while maintaining the ability to capture fine-grained and long-range temporal information. Furthermore, we introduce a central-adjacent region integrated decoder that leverages a more comprehensive sampling strategy for deformable cross-attention to capture the essential information. Extensive experiments demonstrate that DiGIT achieves state-of-the-art performance on THUMOS14, ActivityNet v1.3, and HACS-Segment.

Hallucinations in Large Vision-Language Models (LVLMs) significantly undermine their reliability, motivating researchers to explore the causes of hallucination. However, most studies primarily focus on the language aspect rather than the visual. In this paper, we address how LVLMs process visual information and whether this process causes hallucination. Firstly, we use the attention lens to identify the stages at which LVLMs handle visual data, discovering that the middle layers are crucial. Moreover, we find that these layers can be further divided into two stages: "visual information enrichment" and "semantic refinement" which respectively propagate visual data to object tokens and interpret it through text. By analyzing attention patterns during the visual information enrichment stage, we find that real tokens consistently receive higher attention weights than hallucinated ones, serving as a strong indicator of hallucination. Further examination of multi-head attention maps reveals that hallucination tokens often result from heads interacting with inconsistent objects. Based on these insights, we propose a simple inference-time method that adjusts visual attention by integrating information across various heads. Extensive experiments demonstrate that this approach effectively mitigates hallucinations in mainstream LVLMs without additional training costs.

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.

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 .

3D Question Answering (3D QA) requires the model to comprehensively understand its situated 3D scene described by the text, then reason about its surrounding environment and answer a question under that situation. However, existing methods usually rely on global scene perception from pure 3D point clouds and overlook the importance of rich local texture details from multi-view images. Moreover, due to the inherent noise in camera poses and complex occlusions, there exists significant feature degradation and reduced feature robustness problems when aligning 3D point cloud with multi-view images. In this paper, we propose a Dual-vision Scene Perception Network (DSPNet), to comprehensively integrate multi-view and point cloud features to improve robustness in 3D QA. Our Text-guided Multi-view Fusion (TGMF) module prioritizes image views that closely match the semantic content of the text. To adaptively fuse back-projected multi-view images with point cloud features, we design the Adaptive Dual-vision Perception (ADVP) module, enhancing 3D scene comprehension. Additionally, our Multimodal Context-guided Reasoning (MCGR) module facilitates robust reasoning by integrating contextual information across visual and linguistic modalities. Experimental results on SQA3D and ScanQA datasets demonstrate the superiority of our DSPNet. Codes will be available at https://github.com/LZ-CH/DSPNet.

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.

Self-supervised visual foundation models produce powerful embeddings that achieve remarkable performance on a wide range of downstream tasks. However, unlike vision-language models such as CLIP, self-supervised visual features are not readily aligned with language, hindering their adoption in open-vocabulary tasks. Our method unlocks this new ability for DINOv2, a widely used self-supervised visual encoder. We build upon the LiT training strategy, which trains a text encoder to align with a frozen vision model but leads to unsatisfactory results on dense tasks. We propose several key ingredients to improve performance on both global and dense tasks, such as concatenating the [CLS] token with the patch average to train the alignment and curating data using both text and image modalities. With these, we successfully train a CLIP-like model with only a fraction of the computational cost compared to CLIP while achieving state-of-the-art results in zero-shot classification and open-vocabulary semantic segmentation.

Vision-Language (VL) alignment across image and text modalities is a challenging task due to the inherent semantic ambiguity of data with multiple possible meanings. Existing methods typically solve it by learning multiple sub-representation spaces to encode each input data as a set of embeddings, and constraining diversity between whole subspaces to capture diverse semantics for accurate VL alignment. Despite their promising outcomes, existing methods suffer two imperfections: 1) actually, specific semantics is mainly expressed by some local dimensions within the subspace. Ignoring this intrinsic property, existing diversity constraints imposed on the whole subspace may impair diverse embedding learning; 2) multiple embeddings are inevitably introduced, sacrificing computational and storage efficiency. In this paper, we propose a simple yet effective Diverse and Hybrid Set-embeddings learning framework (DH-Set), which is distinct from prior work in three aspects. DH-Set 1) devises a novel semantic importance dissecting method to focus on key local dimensions within each subspace; and thereby 2) not only imposes finer-grained diversity constraint to improve the accuracy of diverse embedding learning, 3) but also mixes key dimensions of all subspaces into the single hybrid embedding to boost inference efficiency. Extensive experiments on various benchmarks and model backbones show the superiority of DH-Set over state-of-the-art methods, achieving substantial 2.3%-14.7% rSum improvements while lowering computational and storage complexity.

Visual Prompt Tuning (VPT) has become a promising solution for Parameter-Efficient Fine-Tuning (PEFT) approach for Vision Transformer (ViT) models by partially fine-tuning learnable tokens while keeping most model parameters frozen. Recent research has explored modifying the connection structures of the prompts. However, the fundamental correlation and distribution between the prompts and image tokens remain unexplored. In this paper, we leverage metric learning techniques to investigate how the distribution of prompts affects fine-tuning performance. Specifically, we propose a novel framework, Distribution Aware Visual Prompt Tuning (DA-VPT), to guide the distributions of the prompts by learning the distance metric from their class-related semantic data. Our method demonstrates that the prompts can serve as an effective bridge to share semantic information between image patches and the class token. We extensively evaluated our approach on popular benchmarks in both recognition and segmentation tasks. The results demonstrate that our approach enables more effective and efficient fine-tuning of ViT models by leveraging semantic information to guide the learning of the prompts, leading to improved performance on various downstream vision tasks.

The goal of image cropping is to identify visually appealing crops in an image. Conventional methods are trained on specific datasets and fail to adapt to new requirements. Recent breakthroughs in large vision-language models (VLMs) enable visual in-context learning without explicit training. However, downstream tasks with VLMs remain under explored. In this paper, we propose an effective approach to leverage VLMs for image cropping. First, we propose an efficient prompt retrieval mechanism for image cropping to automate the selection of in-context examples. Second, we introduce an iterative refinement strategy to iteratively enhance the predicted crops. The proposed framework, we refer to as Cropper, is applicable to a wide range of cropping tasks, including free-form cropping, subject-aware cropping, and aspect ratio-aware cropping. Extensive experiments demonstrate that Cropper significantly outperforms state-of-the-art methods across several benchmarks.

Repetitive action counting, which aims to count periodic movements in a video, is valuable for video analysis applications such as fitness monitoring. However, existing methods largely rely on regression networks with limited representational capacity, which hampers their ability to accurately capture variable periodic patterns. Additionally, their supervised learning on narrow, limited training sets leads to overfitting and restricts their ability to generalize across diverse scenarios. To address these challenges, we propose CountLLM, the first large language model (LLM)-based framework that takes video data and periodic text prompts as inputs and outputs the desired counting value. CountLLM leverages the rich clues from explicit textual instructions and the powerful representational capabilities of pre-trained LLMs for repetitive action counting. To effectively guide CountLLM, we develop a periodicity-based structured template for instructions that describes the properties of periodicity and implements a standardized answer format to ensure consistency. Additionally, we propose a progressive multimodal training paradigm to enhance the periodicity-awareness of the LLM. Empirical evaluations on widely recognized benchmarks demonstrate CountLLM's superior performance and generalization, particularly in handling novel and out-of-domain actions that deviate significantly from the training data, offering a promising avenue for repetitive action counting.

Online Action Detection (OAD) detects actions in streaming videos using past observations. State-of-the-art OAD approaches model past observations and their interactions with an anticipated future. The past is encoded using short- and long-term memories to capture immediate and long-range dependencies, while anticipation compensates for missing future context.We identify a training-inference discrepancy in existing OAD methods that hinders learning effectiveness. The training uses varying lengths of short-term memory, while inference relies on a full-length short-term memory. As a remedy, we propose a Context-enhanced Memory-Refined Transformer (CMeRT). CMeRT introduces a context-enhanced encoder to improve frame representations using additional near-past context. It also features a memory-refined decoder to leverage near-future generation to enhance performance. CMeRT achieves state-of-the-art in online detection and anticipation on THUMOS'14, CrossTask, and EPIC-Kitchens-100.

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-action models (VLAs) have shown potential in leveraging pretrained vision-language models and diverse robot demonstrations for learning generalizable sensorimotor control. While this paradigm effectively utilizes large-scale data from both robotic and non-robotic sources, current VLAs primarily focus on direct input--output mappings, lacking the intermediate reasoning steps crucial for complex manipulation tasks. As a result, existing VLAs lack temporal planning or reasoning capabilities. In this paper, we introduce a method that incorporates explicit visual chain-of-thought (CoT) reasoning into vision-language-action models (VLAs) by predicting future image frames autoregressively as visual goals before generating a short action sequence to achieve these goals. We introduce CoT-VLA, a state-of-the-art 7B VLA that can understand and generate visual and action tokens. Our experimental results demonstrate that CoT-VLA achieves strong performance, outperforming the state-of-the-art VLA model by 17% in real-world manipulation tasks and 6% in simulation benchmarks. Videos are available at: https://cot-vla.github.io/.

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.