CVPR 2025 论文列表

Why do gradient-based explanations struggle with Transformers, and how can we improve them? We identify gradient flow imbalances in Transformers that violate FullGrad-completeness, a critical property for attribution faithfulness that CNNs naturally possess. To address this issue, we introduce LibraGrad--a theoretically grounded post-hoc approach that corrects gradient imbalances through pruning and scaling of backward paths, without changing the forward pass or adding computational overhead. We evaluate LibraGrad using three metric families: Faithfulness, which quantifies prediction changes under perturbations of the most and least relevant features; Completeness Error, which measures attribution conservation relative to model outputs; and Segmentation AP, which assesses alignment with human perception. Extensive experiments across 8 architectures, 4 model sizes, and 5 datasets show that LibraGrad universally enhances gradient-based methods, outperforming existing white-box methods--including Transformer-specific approaches--across all metrics. We demonstrate superior qualitative results through two complementary evaluations: precise text-prompted region highlighting on CLIP models and accurate class discrimination between co-occurring animals on ImageNet-finetuned models--two settings on which existing methods often struggle. LibraGrad is effective even on the attention-free MLP-Mixer architecture, indicating potential for extension to other modern architectures. Our code is freely available at https://nightmachinery.github.io/LibraGrad/.

Large Vision-Language Models (LVLMs) have achieved significant progress in recent years. However, the expensive inference cost limits the realistic deployment of LVLMs. Some works find that visual tokens are redundant and compress tokens to reduce the inference cost. These works identify important non-redundant tokens as target tokens, then prune the remaining tokens (non-target tokens) or merge them into target tokens. However, target token identification faces the token importance-redundancy dilemma. Besides, token merging and pruning face a dilemma between disrupting target token information and losing non-target token information. To solve these problems, we propose a novel visual token compression scheme, named Libra-Merging. In target token identification, Libra-Merging selects the most important tokens from spatially discrete intervals, achieving a more robust token importance-redundancy trade-off than relying on a hyper-parameter. In token compression, when non-target tokens are dissimilar to target tokens, Libra-Merging does not merge them into the target tokens, thus avoiding disrupting target token information. Meanwhile, Libra-Merging condenses these non-target tokens into an information compensation token to prevent losing important non-target token information. Our method can serve as a plug-in for diverse LVLMs, and extensive experimental results demonstrate its effectiveness. The code will be publicly available at https://github.com/longrongyang/Libra-Merging.

Spatial reasoning is a fundamental aspect of human cognition, enabling intuitive understanding and manipulation of objects in three-dimensional space. While foundation models demonstrate remarkable performance on some benchmarks, they still struggle with 3D reasoning tasks like arranging objects in space according to open-ended language instructions, particularly in dense and physically constrained environments. We introduce LayoutVLM, a framework and scene layout representation that exploits the semantic knowledge of Vision-Language Models (VLMs) and supports differentiable optimization to ensure physical plausibility. LayoutVLM employs VLMs to generate two mutually reinforcing representations from visually marked images, and a self-consistent decoding process to improve VLMs spatial planning. Our experiments show that LayoutVLM addresses the limitations of existing LLM and constraint-based approaches, producing physically plausible 3D layouts better aligned with the semantic intent of input language instructions. We also demonstrate that fine-tuning VLMs with the proposed scene layout representation extracted from existing scene datasets can improve their reasoning performance.

Current Large Language Vision Models (LLVMs) trained on web videos perform well in general video understanding but struggle with fine-grained details, complex human-object interactions (HOI), and view-invariant representation learning essential for Activities of Daily Living (ADL). This limitation stems from a lack of specialized ADL video instruction-tuning datasets and insufficient modality integration to capture discriminative action representations. To address this, we propose a semi-automated framework for curating ADL datasets, creating ADL-X, a multi-view, multi-modal RGBS instruction-tuning dataset. Additionally, we introduce LLAVIDAL, an LLVM integrating videos, 3D skeletons, and HOIs to model ADL's complex spatiotemporal relationships. For training LLAVIDAL, a simple joint alignment of all modalities yields suboptimal results; thus, we propose a Multimodal Progressive (MMPro) training strategy, incorporating modalities in stages following a curriculum. We also establish ADL MCQ and video description benchmarks to assess LLVM performance in ADL tasks. Trained on ADL-X, LLAVIDAL achieves state-of-the-art performance across ADL benchmarks. Code and data will be made publicly available at https://adl-x.github.io/.

The platonic representation hypothesis suggests that vision and language embeddings become more homogeneous as model and dataset sizes increase. In particular, pairwise distances within each modality become more similar. This suggests that as foundation models mature, it may become possible to match vision and language embeddings in a fully unsupervised fashion, i.e. without parallel data. We present the first feasibility study, and investigate conformity of existing vision and language foundation models in the context of unsupervised, or "blind", matching. First, we formulate unsupervised matching as a quadratic assignment problem and introduce a novel heuristic that outperforms previous solvers. We also develop a technique to find optimal matching problems, for which a non-trivial match is very likely. Second, we conduct an extensive study deploying a range of vision and language models on four datasets. Our analysis reveals that for many problem instances, vision and language representations can be indeed matched without supervision. This finding opens up the exciting possibility of embedding semantic knowledge into other modalities virtually annotation-free. As a proof of concept, we showcase an unsupervised classifier, which achieves non-trivial classification accuracy without any image-text annotation.

In this paper, we observe that the collaboration of various foundation models can perceive semantic and degraded information within images, thereby guiding the low-light enhancement process. Specifically, we propose a self-supervised low-light enhancement framework based on the multiple foundation models collaboration (dubbed FoCo), aimed at improving both the visual quality of enhanced images and the performance in high-level applications. At the feature level, FoCo leverages the rich features from various foundation models to enhance the model's semantic perception during training, thereby reducing the gap between enhanced results and high-quality images from a high-level perspective. At the task level, we exploit the robustness-gap between strong foundation models and weak models, applying high-level task guidance to the low-light enhancement training process. Through the collaboration of multiple foundation models, the proposed framework shows better enhancement performance and adapts better to high-level tasks. Extensive experiments across various enhancement and application benchmarks demonstrate the qualitative and quantitative superiority of the proposed method over numerous state-of-the-art techniques.

The evolution of Large Vision-Language Models (LVLMs) has progressed from single-image understanding to multi-image reasoning. Despite this advancement, our findings indicate that LVLMs struggle to robustly utilize information across multiple images, with predictions significantly affected by the alteration of image positions. To further explore this issue, we introduce Position-wise Question Answering (PQA), a meticulously designed task to quantify reasoning capabilities at each position. Our analysis reveals a pronounced position bias in LVLMs: open-source models excel in reasoning with images positioned later but underperform with those in the middle or at the beginning, while proprietary models like GPT-4o show improved comprehension for images at the beginning and end but struggle with those in the middle. Motivated by these insights, we propose SoFt Attention (SoFA), a simple, training-free approach that mitigates this bias by employing linear interpolation between inter-image causal attention and bidirectional counterparts. Experimental results demonstrate that SoFA effectively reduces position bias and significantly enhances the reasoning performance of existing LVLMs.

Addressing the retrieval of unsafe content from vision-language models such as CLIP is an important step towards real-world integration. Current efforts have relied on unlearning techniques that try to erase the model's knowledge of unsafe concepts. While effective in reducing unwanted outputs, unlearning limits the model's capacity to discern between safe and unsafe content. In this work, we introduce a novel approach that shifts from unlearning to an awareness paradigm by leveraging the inherent hierarchical properties of the hyperbolic space. We propose to encode safe and unsafe content as an entailment hierarchy, where both are placed in different regions of hyperbolic space. Our HySAC, Hyperbolic Safety-Aware CLIP, employs entailment loss functions to model the hierarchical and asymmetrical relations between safe and unsafe image-text pairs. This modelling - ineffective in standard vision-language models due to their reliance on Euclidean embeddings - endows the model with awareness of unsafe content, enabling it to serve as both a multimodal unsafe classifier and a flexible content retriever, with the option to dynamically redirect unsafe queries toward safer alternatives or retain the original output. Extensive experiments show that our approach not only enhances safety recognition, but also establishes a more adaptable and interpretable framework for content moderation in vision-language models. Our source code is available at: https://github.com/aimagelab/HySAC

High-resolution image inputs allow Large Vision-Language Models (LVLMs) to capture finer visual details, improving comprehension. However, the increased training and computational costs associated with such inputs pose significant challenges. A common approach to mitigate these costs involves slicing the input into uniform patches using sliding windows, each aligned with the vision encoder's input size. While efficient, this method fragments the input, disrupting the continuity of context, which negatively impacts cross-patch perception tasks. To address these limitations, we propose HiRes-LLaVA, a novel framework designed to efficiently process high-resolution inputs of any size without altering the original contextual and geometric information. HiRes-LLaVA introduces two key components: (i) a SliceRestore Adapter (SRA) that reconstructs sliced patches into their original form, enabling efficient extraction of both global and local features through down-up-sampling and convolutional layers, and (ii) a Self-Mining Sampler (SMS) that compresses visual tokens based on internal relationships, preserving original context and positional information while reducing training overhead. To assess the ability of handling context fragmentation, we construct a new benchmark, EntityGrid-QA, consisting of edge-related tasks. Extensive experiments demonstrate the superiority of HiRes-LLaVA on both existing public benchmarks and EntityGrid-QA. For example, with SRA, our method achieves a performance improvement of 12% over state-of-the-art LVLMs in addressing fragmentation issues. Additionally, our SMS outperforms other visual token downsamplers, while offering high data efficiency.

Skeleton-based human action recognition has received widespread attention in recent years due to its diverse range of application scenarios. Due to the different sources of human skeletons, skeleton data naturally exhibit heterogeneity. The previous works, however, overlook the heterogeneity of human skeletons and solely construct models tailored for homogeneous skeletons. This work addresses the challenge of heterogeneous skeleton-based action representation learning, specifically focusing on processing skeleton data that varies in joint dimensions and topological structures. The proposed framework comprises two primary components: heterogeneous skeleton processing and unified representation learning. The former first converts two-dimensional skeleton data into three-dimensional skeleton via an auxiliary network, and then constructs a prompted unified skeleton using skeleton-specific prompts. We also design an additional modality named semantic motion encoding to harness the semantic information within skeletons. The latter module learns a unified action representation using a shared backbone network that processes different heterogeneous skeletons. Extensive experiments on the NTU-60, NTU-120, and PKU-MMD II datasets demonstrate the effectiveness of our method in various tasks of action understanding. Our approach can be applied to action recognition in robots with different humanoid structures.

Hallucinations pose a significant challenge to the reliability of large vision-language models, making their detection essential for ensuring accuracy in critical applications. Current detection methods often rely on computationally intensive models, leading to high latency and resource demands. Their definitive outcomes also fail to account for real-world scenarios where the line between hallucinated and truthful information is unclear. To address these issues, we propose HalLoc, a dataset designed for efficient, probabilistic hallucination detection. It features 150K token-level annotated samples, including hallucination types, across Visual Question Answering (VQA), instruction-following, and image captioning tasks. This dataset facilitates the development of models that detect hallucinations with graded confidence, enabling more informed user interactions. Additionally, we introduce a baseline model trained on HalLoc, offering low-overhead, concurrent hallucination detection during generation. The model can be seamlessly integrated into existing VLMs, improving reliability while preserving efficiency. The prospect of a robust plug-and-play hallucination detection module opens new avenues for enhancing the trustworthiness of vision-language models in real-world applications. The HalLoc dataset and code are publicly available at: https://github.com/dbsltm/cvpr25_halloc.

In this paper, we propose H-MoRe, a novel pipeline for learning precise human-centric motion representation. Our approach dynamically preserves relevant human motion while filtering out background movement. Notably, unlike previous methods relying on fully supervised learning from synthetic data, H-MoRe learns directly from real-world scenarios in a self-supervised manner, incorporating both human pose and body shape information. Inspired by kinematics, H-MoRe represents absolute and relative movements of each body point in a matrix format that captures nuanced motion details, termed world-local flows. H-MoRe offers refined insights into human motion, which can be integrated seamlessly into various action-related applications. Experimental results demonstrate that H-MoRe brings substantial improvements across various downstream tasks, including gait recognition(CL@R1: +16.01%), action recognition(Acc@1: +8.92%), and video generation(FVD: -67.07%). Additionally, H-MoRe exhibits high inference efficiency (34 fps), making it suitable for most real-time scenarios. Models and code will be released upon publication.

Event-based cameras asynchronously detect changes in light intensity with high temporal resolution, making them a promising alternative to RGB camera for low-latency and low-power optical flow estimation. However, state-of-the-art convolutional neural network methods create frames from the event stream, therefore losing the opportunity to exploit events for both sparse computations and low-latency prediction. On the other hand, asynchronous event graph methods could leverage both, but at the cost of avoiding any form of time accumulation, which limits the prediction accuracy. In this paper, we propose to break this accuracy-latency trade-off with a novel architecture combining an asynchronous accumulation-free event branch and a periodic aggregation branch. The periodic branch performs feature aggregations on the event graphs of past data to extract global context information, which improves accuracy without introducing any latency.The solution could predict optical flow per event with a latency of tens of microseconds on asynchronous hardware, which represents a gain of three orders of magnitude with respect to state-of-the-art frame-based methods, with 48x less operations per second. We show that the solution can detect rapid motion changes faster than a periodic output. This work proposes, for the first time, an effective solution for ultra low-latency and low-power optical flow prediction from event cameras.

Learning to use tools or objects in common scenes, particularly handling them in various ways as instructed, is a key challenge for developing interactive robots. Training models to generate such manipulation trajectories requires a large and diverse collection of detailed manipulation demonstrations for various objects, which is nearly unfeasible to gather at scale. In this paper, we propose a framework that leverages large-scale ego- and exo-centric video datasets --- constructed globally with substantial effort --- of Exo-Ego4D to extract diverse manipulation trajectories at scale. From these extracted trajectories with the associated textual action description, we develop trajectory generation models based on visual and point cloud-based language models. In the recently proposed egocentric vision-based in-a-quality trajectory dataset of HOT3D, we confirmed that our models successfully generate valid object trajectories, establishing a training dataset and baseline models for the novel task of generating 6DoF manipulation trajectories from action descriptions in egocentric vision.

We present GEM, a Generalizable Ego-vision Multimodal world model that predicts future frames using a reference frame, sparse features, human poses, and ego-trajectories. Hence, our model has precise control over object dynamics, ego-agent motion and human poses. GEM generates paired RGB and depth outputs for richer spatial understanding. We introduce autoregressive noise schedules to enable stable long-horizon generations. Our dataset is comprised of 4000+ hours of multimodal data across domains like autonomous driving, egocentric human activities, and drone flights. Pseudo-labels are used to get depth maps, ego-trajectories, and human poses. We use a comprehensive evaluation framework, including a new Control of Object Manipulation (COM) metric, to assess controllability. Experiments show GEM excels at generating diverse, controllable scenarios and temporal consistency over long generations. Code, models, and datasets are fully open-sourced.

Large Vision-Language Models (LVLMs) have achieved significant progress in combining visual comprehension with language generation.Despite this success, the training data of LVLMs still suffers from Long-Tail (LT) problems, where the data distribution is highly imbalanced.Previous works have mainly focused on traditional VLM architectures, i.e., CLIP or ViT, and specific tasks such as recognition and classification. Nevertheless, the exploration of LVLM (e.g. LLaVA) and more general tasks (e.g. Visual Question Answering and Visual Reasoning) remains under-explored.In this paper, we first conduct an in-depth analysis of the LT issues in LVLMs and identify two core causes: the overrepresentation of head concepts and the underrepresentation of tail concepts.Based on the above observation, we propose an Adaptive Data Refinement Framework (ADR), which consists of two stages: Data Rebalancing (DR) and Data Synthesis (DS).In the DR stage, we adaptively rebalance the redundant data based on entity distributions, while in the DS stage, we leverage Denoising Diffusion Probabilistic Models (DDPMs) and scarce images to supplement underrepresented portions.Through comprehensive evaluations across eleven benchmarks, our proposed ADR effectively mitigates the long-tail problem in the training data, improving the average performance of LLaVA 1.5 relatively by 4.36%, without increasing the training data volume. Our code and data will be publicly released.

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.

Despite their promise to perform complex reasoning, large language models (LLMs) have been shown to have limited effectiveness in end-to-end planning. This has inspired an intriguing question: if these models cannot plan well, can they still contribute to the planning framework as a helpful plan evaluator? In this work, we generalize this question to consider LLMs augmented with visual understanding, i.e., Vision-Language Models (VLMs). We introduce **PathEval**, a novel benchmark evaluating VLMs as plan evaluators in complex path-planning scenarios. Succeeding in the benchmark requires a VLM to be able to abstract traits of optimal paths from the scenario description, demonstrate precise low-level perception on each path, and integrate this information to decide the better path. Our analysis of state-of-the-art VLMs reveals that these models face significant challenges on the benchmark. We observe that the VLMs can precisely abstract given scenarios to identify the desired traits and exhibit mixed performance in integrating the provided information. Yet, their vision component presents a critical bottleneck, with models struggling to perceive low-level details about a path. Our experimental results show that this issue cannot be trivially addressed via end-to-end fine-tuning; rather, task-specific discriminative adaptation of these vision encoders is needed for these VLMs to become effective path evaluators.