CVPR 2025 论文列表

'Pre-training + fine-tuning' has been widely used in various downstream tasks. Parameter-efficient fine-tuning (PEFT) has demonstrated higher efficiency and promising performance comapred to traditional full-tuning. The widely used adapter-based and prompt-based methods in PEFT can be uniformly represented as adding an MLP structure to the pre-trained model. These methods are prone to over-fitting in downstream tasks, due to the difference in data scale and distribution. To address this issue, we propose a new adapter-based PEFT module, i.e., LoKi, which consists of an encoder, a learnable activation layer, and a decoder. To maintain the simplicity of LoKi, we use single-layer linear networks for the encoder and decoder, and for the learnable activation layer, we use a Kolmogorov-Arnold Network (KAN) with the minimal number of layers (only 2 KAN linear layers). With a bottleneck rate much lower than that of Adapter, LoKi is equipped with fewer parameters (only half of Adapter) and eliminates slow training speed and high memory usage of KAN. We conduct extensive experiments on LoKi under image classification and video action recognition across 9 datasets. LoKi demonstrates highly competitive generalization performance compared to other PEFT methods with fewer tunable parameters, ensuring both effectiveness and efficiency.

Recent video large language models (Video LLMs) often depend on costly human annotations or proprietary APIs (e.g., GPT-4o) to produce training data, which limits their training at scale. In this paper, we explore large-scale training for Video LLM with cheap automatic speech recognition (ASR) transcripts. Specifically, we propose a novel streaming training approach that densely interleaves the ASR words and video frames according to their timestamps. Compared to previous studies in vision-language representation with ASR, our method naturally fits the streaming characteristics of ASR, thus enabling the model to learn temporally-aligned, fine-grained vision-language modeling. To support the training algorithm, we introduce a data pipeline for YouTube videos and their closed captions (CC), resulting in \texttt Live-CC-10M pre-training set and \texttt Live-WhisperX-408K high-quality supervised fine-tuning (SFT) set. Remarkably, even without SFT, the pre-trained model \texttt LiveCC-7B demonstrates significant improvements in general video QA and exhibits a new capability in real-time video commentary. To evaluate this, we carefully design a new benchmark \texttt LiveSports-3K , using LLM-as-a-judge to measure the free-form commentary. Experiments show our final model \texttt LiveCC-7B can surpass LLaVA-Video-72B in commentary quality even working in a real-time mode. Meanwhile, it achieves state-of-the-art results at the 7B scale on popular benchmarks such as VideoMME, demonstrating its broad generalizability. All resources of this paper have been released at \href https://showlab.github.io/livecc showlab.github.io/livecc .

Fast 3D clothed human reconstruction from monocular video remains a significant challenge in computer vision, particularly in balancing computational efficiency with reconstruction quality. Current approaches are either focused on static image reconstruction but too computationally intensive, or achieve high quality through per-video optimization that requires minutes to hours of processing, making them unsuitable for real-time applications. To this end, we present TemPoFast3D, a novel method that leverages temporal coherency of human appearance to reduce redundant computation while maintaining reconstruction quality. Our approach is a "plug-and play" solution that uniquely transforms pixel-aligned reconstruction networks to handle continuous video streams by maintaining and refining a canonical appearance representation through efficient coordinate mapping. Extensive experiments demonstrate that TemPoFast3D matches or exceeds state-of-the-art methods across standard metrics while providing high-quality textured reconstruction across diverse pose and appearance, with a maximum speed of 12 FPS.

Text images are unique in their dual nature, encompassing both visual and linguistic information. The visual component encompasses structural and appearance-based features, while the linguistic dimension incorporates contextual and semantic elements. In scenarios with degraded visual quality, linguistic patterns serve as crucial supplements for comprehension, highlighting the necessity of integrating both aspects for robust scene text recognition (STR). Contemporary STR approaches often use language models or semantic reasoning modules to capture linguistic features, typically requiring large-scale annotated datasets. Self-supervised learning, which lacks annotations, presents challenges in disentangling linguistic features related to the global context. Typically, sequence contrastive learning emphasizes the alignment of local features, while masked image modeling (MIM) tends to exploit local structures to reconstruct visual patterns, resulting in limited linguistic knowledge. In this paper, we propose a Linguistics-aware Masked Image Modeling (LMIM) approach, which channels the linguistic information into the decoding process of MIM through a separate branch. Specifically, we design a linguistics alignment module to extract vision-independent features as linguistic guidance using inputs with different visual appearances. As features extend beyond mere visual structures, LMIM must consider the global context to achieve reconstruction. Extensive experiments on various benchmarks quantitatively demonstrate our state-of-the-art performance, and attention visualizations qualitatively show the simultaneous capture of both visual and linguistic information. The code is available at https://github.com/zhangyifei01/LMIM.

Recent years, learned image compression has made tremendous progress to achieve impressive coding efficiency. Its coding gain mainly comes from non-linear neural network-based transform and learnable entropy modeling. However, most studies focus on a strong backbone, and few studies consider a low complexity design. In this paper, we propose LALIC, a linear attention modeling for learned image compression. Specially, we propose to use Bi-RWKV blocks, by utilizing the Spatial Mix and Channel Mix modules to achieve more compact feature extraction, and apply the Conv based Omni-Shift module to adapt to two-dimensional latent representation. Furthermore, we propose a RWKV-based Spatial-Channel ConTeXt model (RWKV-SCCTX), that leverages the Bi-RWKV to modeling the correlation between neighboring features effectively. To our knowledge, our work is the first work to utilize efficient Bi-RWKV models with linear attention for learned image compression. Experimental results demonstrate that our method achieves competitive RD performances by outperforming VTM-9.1 by -15.26%, -15.41%, -17.63% in BD-rate on Kodak, CLIC and Tecnick datasets. The code is available at https://github.com/sjtu-medialab/RwkvCompress .

Multimodal large language models (MLLMs) improve performance on vision-language tasks by integrating visual features from pre-trained vision encoders into large language models (LLMs). However, how MLLMs process and utilize visual information remains unclear. In this paper, a shift in the dominant flow of visual information is uncovered: (1) in shallow layers, strong interactions are observed between image tokens and instruction tokens, where most visual information is injected into instruction tokens to form cross-modal semantic representations; (2) in deeper layers, image tokens primarily interact with each other, aggregating the remaining visual information to optimize semantic representations within the visual modality. Based on these insights, we propose Hierarchical Modality-Aware Pruning (HiMAP), a plug-and-play inference acceleration method that dynamically prunes image tokens at specific layers, reducing computational costs by approximately 65% without sacrificing performance. Our findings offer a new understanding of visual information processing in MLLMs and provide a state-of-the-art solution for efficient inference.

Model editing aims to correct inaccurate knowledge, update outdated information, and incorporate new data into Large Language Models (LLMs) without the need for retraining. This task poses challenges in lifelong scenarios where edits must be continuously applied for real-world applications. While some editors demonstrate strong robustness for lifelong editing in pure LLMs, Vision LLMs (VLLMs), which incorporate an additional vision modality, are not directly adaptable to existing LLM editors. In this paper, we propose LiveEdit, a lifelong vision language model edit to bridge the gap between lifelong LLM editing and VLLMs. We begin by training an editing expert generator to independently produce low-rank experts for each editing instance, with the goal of correcting the relevant responses of the VLLM. A hard filtering mechanism is developed to utilize visual semantic knowledge, thereby coarsely eliminating visually irrelevant experts for input queries during the inference stage of the post-edited model. Finally, to integrate visually relevant experts, we introduce a soft routing mechanism based on textual semantic relevance to achieve multi-expert fusion. For evaluation, we establish a benchmark for lifelong VLLM editing. Extensive experiments demonstrate that LiveEdit offers significant advantages in lifelong VLLM editing scenarios. Further experiments validate the rationality and effectiveness of each module design in LiveEdit.

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.

Deep learning models are known to often learn features that spuriously correlate with the class label during training but are irrelevant to the prediction task. Existing methods typically address this issue by annotating potential spurious attributes, or filtering spurious features based on some empirical assumptions (e.g., simplicity of bias). However, these methods may yield unsatisfying performance due to the intricate and elusive nature of spurious correlations in real-world data. In this paper, we propose a data-oriented approach to mitigate the spurious correlation in deep learning models. We observe that samples that are influenced by spurious features tend to exhibit a dispersed distribution in the learned feature space. This allows us to identify the presence of spurious features. Subsequently, we obtain a bias-invariant representation by neutralizing the spuriousfeatures based on a simple grouping strategy. Then, we learn a feature transformation to eliminate the spuriousfeatures by aligning with this bias-invariant representation. Finally, we update the classifier by incorporating the learned feature transformation and obtain an unbiased model. By integrating the aforementioned identifying, neutralizing, eliminating and updating procedures, we build an effective pipeline for mitigating spurious correlation. Experiments on image and NLP debiasing benchmarks show an improvement in worst group accuracy of more than 20% compared to standard empirical risk minimization (ERM). Codes and checkpoints are available at https://github.com/davelee-uestc/nsf_debiasing.

Hiking on complex trails demands balance, agility, and adaptive decision-making over unpredictable terrain. Current humanoid research remains fragmented and inadequate for hiking: locomotion focuses on motor skills without long-term goals or situational awareness, while semantic navigation overlooks real-world embodiment and local terrain variability. We propose training humanoids to hike on complex trails, fostering integrative skill development across visual perception, decision making, and motor execution. We develop LEGO-H, a learning framework that enables a humanoid with vision to hike complex trails independently. It has two key innovations. (1) A Temporal Vision Transformer anticipates future steps to guide locomotion, unifying local movement and goal-directed navigation. (2) Latent representations of joint movement patterns combined with hierarchical metric learning allow smooth policy transfer from privileged training to real-world training. These techniques enable LEGO-H to handle diverse physical and environmental challenges without relying on predefined motion patterns. Experiments on diverse simulated hiking trails and humanoids with different morphologies demonstrate LEGO-H's robustness and versatility, establishing a strong foundation for future humanoid development.

Composed Image Retrieval (CIR) enables editable image search by integrating a query pair--a reference image ref and a textual modification mod--to retrieve a target image tar that reflects the intended change. While existing CIR methods have shown promising performance using well-annotated triplets \langle ref, mod, tar \rangle, almost all of them implicitly assume these triplets are accurately associated with each other. In practice, however, this assumption is often violated due to the limited knowledge of annotators, inevitably leading to incorrect textual modifications and resulting in a practical yet less-touched problem: noisy triplet correspondence (NTC). To tackle this challenge, we propose a \underline T ask-oriented \underline M odification \underline E nhancement framework (TME) to learn robustly from noisy triplets, which comprises three key modules: Robust Fusion Query (RFQ), Pseudo Text Enhancement (PTE), and Task-Oriented Prompt (TOP). Specifically, to mitigate the adverse impact of noise, RFQ employs a sample selection strategy to divide the training triplets into clean and noisy sets, thus enhancing the reliability of the training data for robust learning. To further leverage the noisy data instead of discarding it, PTE unifies the triplet noise as an adapter mismatch problem, thereby adjusting mod to align with ref and tar in the mismatched triplet. Finally, TOP replaces ref in the clean set with a trainable prompt, which is then concatenated with mod to form a query independent of the visual reference, aiming to mitigate visually irrelevant noise. Extensive experiments on two domain-specific datasets demonstrate the robustness and superiority of TME, particularly in noisy scenarios.

Extracting physical dynamical system parameters from recorded observations is key in natural science. Current methods for automatic parameter estimation from video train supervised deep networks on large datasets. Such datasets require labels, which are difficult to acquire. While some unsupervised techniques--which depend on frame prediction--exist, they suffer from long training times, initialization instabilities, only consider motion-based dynamical systems, and are evaluated mainly on synthetic data. In this work, we propose an unsupervised method to estimate the physical parameters of known, continuous governing equations from single videos suitable for different dynamical systems beyond motion and robust to initialization. Moreover, we remove the need for frame prediction by implementing a KL-divergence-based loss function in the latent space, which avoids convergence to trivial solutions and reduces model size and compute. We first evaluate our model on synthetic data, as commonly done. After which, we take the field closer to reality by recording Delfys75: our own real-world dataset of 75 videos for five different types of dynamical systems to evaluate our method and others. Our method compares favorably to others. Code and data are available online: https://github.com/Alejandro-neuro/Learning_physics_from_video.

Atmospheric turbulence is a major source of image degradation in long-range imaging systems. Although numerous deep learning-based turbulence mitigation (TM) methods have been proposed, many are slow, memory-hungry, and do not generalize well. In the spatial domain, methods based on convolutional operators have a limited receptive field, so they cannot handle a large spatial dependency required by turbulence. In the temporal domain, methods relying on self-attention can, in theory, leverage the lucky effects of turbulence, but their quadratic complexity makes it difficult to scale to many frames. Traditional recurrent aggregation methods face parallelization challenges. In this paper, we present a new TM method based on two concepts: (1) A turbulence mitigation network based on the Selective State Space Model (MambaTM). MambaTM provides a global receptive field in each layer across spatial and temporal dimensions while maintaining linear computational complexity. (2) Learned Latent Phase Distortion (LPD). LPD guides the state space model. Unlike classical Zernike-based representations of phase distortion, the new LPD map uniquely captures the actual effects of turbulence, significantly improving the model's capability to estimate degradation by reducing the ill-posedness. Our proposed method exceeds current state-of-the-art networks on various synthetic and real-world TM benchmarks with significantly faster inference speed. The code will be publicly available.

Single-stream architectures using Vision Transformer (ViT) backbones show great potential for real-time UAV tracking recently. However, frequent occlusions from obstacles like buildings and trees expose a major drawback: these models often lack strategies to handle occlusions effectively. New methods are needed to enhance the occlusion resilience of single-stream ViT models in aerial tracking. In this work, we propose to learn Occlusion-Robust Representations (ORR) based on ViTs for UAV tracking by enforcing an invariance of the feature representation of a target with respect to random masking operations modeled by a spatial Cox process. Hopefully, this random masking approximately simulates target occlusions, thereby enabling us to learn ViTs that are robust to target occlusion for UAV tracking. This framework is termed ORTrack. Additionally, to facilitate real-time applications, we propose an Adaptive Feature-Based Knowledge Distillation (AFKD) method to create a more compact tracker, which adaptively mimics the behavior of the teacher model ORTrack according to the task's difficulty. This student model, dubbed ORTrack-D, retains much of ORTrack's performance while offering higher efficiency. Extensive experiments on multiple benchmarks validate the effectiveness of our method, demonstrating its state-of-the-art performance. Codes is available at https://github.com/wuyou3474/ORTrack.

Video-based high-density crowd analysis and prediction has been a long-standing topic in computer vision. It is notoriously difficult due to, but not limited to, the lack of high-quality data and complex crowd dynamics. Consequently, it has been relatively under studied. In this paper, we propose a new approach that aims to learn from in-the-wild videos, often with low quality where it is difficult to track individuals or count heads. The key novelty is a new physics prior to model crowd dynamics. We model high-density crowds as active matter, a continumm with active particles subject to stochastic forces, named `crowd material'. Our physics model is combined with neural networks, resulting in a neural stochastic differential equation system which can mimic the complex crowd dynamics. Due to the lack of similar research, we adapt a range of existing methods which are close to ours, and other possible alternatives for comparison. Through exhaustive evaluation, we show our model outperforms existing methods in analyzing and forecasting extremely high-density crowds. Furthermore, since our model is a continuous-time physics model, it can be used for simulation and analysis, providing strong interpretability. This is categorically different from most deep learning methods, which are discrete-time models and black-boxes.

Video-text retrieval, the task of retrieving videos based on a textual query or vice versa, is of paramount importance for video understanding and multimodal information retrieval. Recent methods in this area rely primarily on visual and textual features and often ignore audio, although it helps enhance overall comprehension of video content.Moreover, traditional models that incorporate audio blindly utilize the audio input regardless of whether it is useful or not, resulting in suboptimal video representation. To address these limitations, we propose a novel video-text retrieval framework, Audio-guided VIdeo representation learning with GATEd attention (AVIGATE), that effectively leverages audio cues through a gated attention mechanism that selectively filters out uninformative audio signals.In addition, we propose an adaptive margin-based contrastive loss to deal with the inherently unclear positive-negative relationship between video and text, which facilitates learning better video-text alignment.Our extensive experiments demonstrate that AVIGATE achieves state-of-the-art performance on all the public benchmarks.

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.

Salient Object Ranking (SOR) aims to study human attention shifts across different objects in the scene. It is a challenging task, as it requires comprehension of the relations among the salient objects in the scene. However, existing works often overlook such relations or model them implicitly. In this work, we observe that when Large Vision-Language Models (LVLMs) describe a scene, they usually focus on the most salient object first, and then discuss the relations as they move on to the next (less salient) one. Based on this observation, we propose a novel Language-Guided Salient Object Ranking approach (named LG-SOR), which utilizes the internal knowledge within the LVLM-generated language descriptions, i.e., semantic relation cues and the implicit entity order cues, to facilitate saliency ranking. Specifically, we first propose a novel Text-Guided Visual Modulation (TGVM) module to incorporate semantic information in the description for saliency ranking. TGVM controls the flow of linguistic information to the visual features, suppresses noisy background image features, and enables propagation of useful textual features. We then propose a novel Text-Aware Visual Reasoning (TAVR) module to enhance model reasoning in object ranking, by explicitly learning a multimodal graph based on the entity and relation cues derived from the description. Extensive experiments demonstrate superior performances of our model on two SOR benchmarks.