CVPR 2025 Papers

Despite a big leap forward in capability, multimodal large language models (MLLMs) tend to behave like a sloth in practical use, i.e., slow response and large latency. Recent efforts are devoted to building tiny MLLMs for better efficiency, but the plethora of visual tokens still used limit their actual speedup. In this paper, we propose a powerful and fast tiny MLLM called FlashSloth. Different from previous efforts, FlashSloth focuses on improving the descriptive power of visual tokens in the process of compressing their redundant semantics. In particular, FlashSloth introduces embedded visual compression designs to capture both visually salient and instruction-related image information, so as to achieving superior multimodal performance with fewer visual tokens. Extensive experiments are conducted to validate the proposed FlashSloth, and a bunch of tiny but strong MLLMs are also comprehensively compared, e.g., InternVL-2, MiniCPM-V2 and Qwen2-VL. The experimental results show that compared with these advanced tiny MLLMs, our FlashSloth can greatly reduce the number of visual tokens, training memory and computation complexity while retaining high performance on various VL tasks. Our code is released at: https://github.com/codefanw/FlashSloth.

Class activation map (CAM) has been widely used to highlight image regions that contribute to class predictions. Despite its simplicity and computational efficiency, CAM often struggles to identify discriminative regions that distinguish visually similar fine-grained classes. Prior efforts address this limitation by introducing more sophisticated explanation processes, but at the cost of extra complexity. In this paper, we propose Finer-CAM, a method that retains CAM's efficiency while achieving precise localization of discriminative regions. Our key insight is that the deficiency of CAM lies not in "how" it explains, but in "what" it explains. Specifically, previous methods attempt to identify all cues contributing to the target class's logit value, which inadvertently also activates regions predictive of visually similar classes. By explicitly comparing the target class with similar classes and spotting their differences, Finer-CAM suppresses features shared with other classes and emphasizes the unique, discriminative details of the target class. Finer-CAM is easy to implement, compatible with various CAM methods, and can be extended to multi-modal models for accurate localization of specific concepts. Additionally, Finer-CAM allows adjustable comparison strength, enabling users to selectively highlight coarse object contours or fine discriminative details. Quantitatively, we show that masking out the top 5% of activated pixels by Finer-CAM results in a larger relative confidence drop compared to baselines. The source code and demo are available at https://github.com/Imageomics/Finer-CAM.

Visual instruction tuning (VIT) for large vision-language models (LVLMs) requires training on expansive datasets of image-instruction pairs, which can be costly. Recent efforts in VIT data selection aim to select a small subset of high-quality image-instruction pairs, reducing VIT runtime while maintaining performance comparable to full-scale training. However, a major challenge often overlooked is that generating instructions from unlabeled images for VIT is highly expensive. Most existing VIT datasets rely heavily on human annotations or paid services like the GPT API, which limits users with constrained resources from creating VIT datasets for custom applications. To address this, we introduce Pre-Instruction Data Selection (PreSel), a more practical data selection paradigm that directly selects the most beneficial unlabeled images and generates instructions only for the selected images. PreSel first estimates the relative importance of each vision task within VIT datasets to derive task-wise sampling budgets. It then clusters image features within each task, selecting the most representative images with the budget. This approach reduces computational overhead for both instruction generation during VIT data formation and LVLM fine-tuning. By generating instructions for only 15% of the images, PreSel achieves performance comparable to full-data VIT on the LLaVA-1.5 and Vision-Flan datasets. The link to our project page: https://bardisafa.github.io/PreSel

Combining the advantages of conventional and event cameras for robust visual tracing has drawn extensive interest. However, existing tracking approaches heavily engage in complex cross-modal fusion modules, leading to higher computational complexity and training challenges. Besides, these methods generally ignore the effective integration of historical information, which is crucial to grasping the change in the target's appearance and motion trends. Given the recent advancements in Mamba's long-range modeling and linear complexity, we explore its potential in addressing the above issues in RGBE tracking tasks. Specifically, we first propose an efficient fusion module based on Mamba, which utilizes a simple gate-based interaction scheme to achieve effective modality-selective fusion. This module can be seamlessly integrated into the encoding layer of prevalent Transformer-based backbones. Moreover, we further present a novel historical decoder that leverages Mamba's advanced long sequence modeling to effectively capture the target appearance changes with autoregressive queries. Extensive experiments show that our proposed approach achieves state-of-the-art performance on multiple challenging short-term and long-term RGBE benchmarks. Besides, the effectiveness of each key Mamba-based component of our approach is evidenced by our thorough ablation study.

Pedestrian attribute recognition (PAR) seeks to predict multiple semantic attributes associated with a specific pedestrian. There are two types of approaches for PAR: unimodal framework and bimodal framework. The former one is to seek a robust visual feature. However, the lack of exploiting semantic feature of linguistic modality is the main concern. The latter one utilizes prompt learning techniques to integrate linguistic data. However, static prompt templates and simple bimodal concatenation cannot to capture the extensive intra-class attribute variability and support active modalities collaboration. In this paper, we propose an Enhanced Visual-Semantic Interaction with Tailored Prompts (EVSITP) framework for PAR. We present an Image-Conditional Dual-Prompt Initialization Module (IDIM) to adaptively generate context-sensitive prompts from visual inputs. Subsequently, a Prompt Enhanced and Regularization Module (PERM) is proposed to strengthen linguistic information from IDIM. We further design a Bimodal Mutual Interaction Module (BMIM) to ensure bidirectional modalities communication. In addition, existing PAR datasets are collected over a short period in limited scenarios, which do not align with real-world scenarios. Therefore, we annotate a long-term person re-identification dataset to create a new PAR dataset, Celeb-PAR. Experiments on several challenging PAR datasets show that our method outperforms state-of-the-art approaches.

Visual In-Context Learning (VICL) enables adaptively solving vision tasks by leveraging pixel demonstrations, mimicking human-like task completion through analogy. Prompt selection is critical in VICL, but current methods assume the existence of a single "ideal" prompt in a pool of candidates, which in practice may not hold true. Multiple suitable prompts may exist, but individually they often fall short, leading to difficulties in selection and the exclusion of useful context. To address this, we propose a new perspective: prompt condensation. Rather than relying on a single prompt, candidate prompts collaborate to efficiently integrate informative contexts without sacrificing resolution. We devise Condenser, a lightweight external plugin that compresses relevant fine-grained context across multiple prompts. Optimized end-to-end with the backbone, Condenser ensures accurate integration of contextual cues. Experiments demonstrate Condenser outperforms state-of-the-arts across benchmark tasks, showing superior context compression, scalability with more prompts, and enhanced computational efficiency compared to ensemble methods, positioning it as a highly competitive solution for VICL. Code is open-sourced at https://github.com/gimpong/CVPR25-Condenser.

Traffic Anomaly Understanding (TAU) is essential for improving public safety and transportation efficiency by enabling timely detection and response to incidents. Beyond existing methods, which rely largely on visual data, we propose to consider audio cues, a valuable source that offers strong hints to anomaly scenarios such as crashes and honking. Our contributions are twofold. First, we compile AV-TAU, the first large-scale audio-visual dataset for TAU, providing 29,865 traffic anomaly videos and 149,325 Q&A pairs, while supporting five essential TAU tasks. Second, we develop EchoTraffic, a multimodal LLM that integrates audio and visual data for TAU, through our audio-insight frame selector and dynamic connector to effectively extract crucial audio cues for anomaly understanding with a two-phase training framework. Experimental results on AV-TAU manifest that EchoTraffic sets a new SOTA performance in TAU, outperforming the existing multimodal LLMs. Our contributions, including AV-TAU and EchoTraffic, pave a new direction for multimodal TAU.

Sound-guided object segmentation has drawn considerable attention for its potential to enhance multimodal perception. Previous methods primarily focus on developing advanced architectures to facilitate effective audio-visual interactions, without fully addressing the inherent challenges posed by audio natures, i.e., (1) feature confusion due to the overlapping nature of audio signals, and (2) audio-visual matching difficulty from the varied sounds produced by the same object. To address these challenges, we propose Dynamic Derivation and Elimination (DDESeg): a novel audio-visual segmentation framework. Specifically, to mitigate feature confusion, DDESeg reconstructs the semantic content of the mixed audio signal by enriching the distinct semantic information of each individual source, deriving representations that preserve the unique characteristics of each sound. To reduce the matching difficulty, we introduce a discriminative feature learning module, which enhances the semantic distinctiveness of generated audio representations. Considering that not all derived audio representations directly correspond to visual features (e.g., off-screen sounds), we propose a dynamic elimination module to filter out non-matching elements. This module facilitates targeted interaction between sounding regions and relevant audio semantics. By scoring the interacted features, we identify and filter out irrelevant audio information, ensuring accurate audio-visual alignment. Comprehensive experiments demonstrate that our framework achieves superior performance in AVS datasets. Our code will be publicly available.

Aiming to achieve class-agnostic perception in visual object tracking, current trackers commonly formulate tracking as a one-shot detection problem with the template-matching architecture. Despite the success, severe environmental variations in long-term tracking raise challenges to generalizing the tracker in novel situations. Temporal trackers try to fix it by preserving the time-validity of target information with historical predictions, e.g., updating the template. However, solely transmitting the previous observations instead of learning from them leads to an inferior capability of understanding the tracking scenario from past experience, which is critical for the generalization in new frames. To address this issue, we reformulate temporal learning in visual tracking as a History-to-Future process and propose a novel tracking framework DreamTrack. Our DreamTrack learns the temporal dynamics from past observations to dream the future variations of the environment, which boosts the generalization with the extended future information from history. Considering the uncertainty of future variation, multimodal prediction is designed to infer the target trajectory of each possible future situation. The experiments demonstrate that our DreamTrack achieves leading performance with real-time inference speed. In particular, DreamTrack obtains SUC scores of 76.6%/87.9% on LaSOT/TrackingNet, surpassing all recent SOTA trackers.

Accurate 3D reconstruction is frequently hindered by visual aliasing, where visually similar but distinct surfaces (aka, doppelgangers), are incorrectly matched. These spurious matches distort the structure-from-motion (SfM) process, leading to misplaced model elements and reduced accuracy. Prior efforts addressed this with CNN classifiers trained on curated datasets, but these approaches struggle to generalize across diverse real-world scenes and can require extensive parameter tuning. In this work, we present Doppelgangers++, a method to enhance doppelganger detection and improve 3D reconstruction accuracy. Our contributions include a diversified training dataset that incorporates geo-tagged images from everyday scenes to expand robustness beyond landmark-based datasets. We further propose a Transformer-based classifier that leverages 3D-aware features from the MASt3R model, achieving superior precision and recall across both in-domain and out-of-domain tests. Doppelgangers++ integrates seamlessly into standard SfM and MASt3R-SfM pipelines, offering efficiency and adaptability across varied scenes. To evaluate SfM accuracy, we introduce an automated, geotag-based method for validating reconstructed models, eliminating the need for manual inspection. Through extensive experiments, we demonstrate that Doppelgangers++ significantly enhances pairwise visual disambiguation and improves 3D reconstruction quality in complex and diverse scenarios.

Large Multimodal Models (LMMs) have emerged as powerful models capable of understanding various data modalities, including text, images, and videos. LMMs encode both text and visual data into tokens that are then combined and processed by an integrated Large Language Model (LLM). Including visual tokens substantially increases the total token count, often by thousands. The increased input length for LLM significantly raises the complexity of inference, resulting in high latency in LMMs. To address this issue, token pruning methods, which remove part of the visual tokens, are proposed. The existing token pruning methods either require extensive calibration and fine-tuning or rely on suboptimal importance metrics which results in increased redundancy among the retained tokens. In this paper, we first formulate token pruning as Max-Min Diversity Problem (MMDP) where the goal is to select a subset such that the diversity among the selected tokens is maximized. Then, we solve the MMDP to obtain the selected subset and prune the rest. The proposed method, DivPrune, reduces redundancy and achieves the highest diversity of the selected tokens. By ensuring high diversity, the selected tokens better represent the original tokens, enabling effective performance even at high pruning ratios without requiring fine-tuning. Extensive experiments with various LMMs show that DivPrune achieves state-of-the-art accuracy over 16 image- and video-language datasets. Additionally, DivPrune reduces both the end-to-end latency and GPU memory usage for the tested models. The code is available here.

Infants develop complex visual understanding rapidly, even preceding of the acquisition of linguistic skills. As computer vision seeks to replicate the human vision system, understanding infant visual development may offer valuable insights. In this paper, we present an interdisciplinary study exploring this question: can a computational model that imitates the infant learning process develop broader visual concepts that extend beyond the vocabulary it has heard, similar to how infants naturally learn? To investigate this, we analyze a recently published model in Science by Vong et al., which is trained on longitudinal, egocentric images of a single child paired with transcribed parental speech. We perform neuron labeling to identify visual concept neurons hidden in the model's internal representations. We then demonstrate that these neurons can recognize objects beyond the model's original vocabulary. Furthermore, we compare the differences in representation between infant models and those in modern computer vision models, such as CLIP and ImageNet pre-trained model. Ultimately, our work bridges cognitive science and computer vision by analyzing the internal representations of a computational model trained on an infant visual and linguistic inputs. Our code is available at https://github.com/Kexueyi/discover_infant_vis.

Concept Bottleneck Models (CBMs) try to make the decision-making process transparent by exploring an intermediate concept space between the input image and the output prediction. Existing CBMs just learn coarse-grained relations between the whole image and the concepts, less considering local image information, leading to two main drawbacks: i) they often produce spurious visual-concept relations, hence decreasing model reliability; and ii) though CBMs could explain the importance of every concept to the final prediction, it is still challenging to tell which visual region produces the prediction. To solve these problems, this paper proposes a Disentangled Optimal Transport CBM (DOT-CBM) framework to explore fine-grained visual-concept relations between local image patches and concepts. Specifically, we model the concept prediction process as a transportation problem between the patches and concepts, thereby achieving explicit fine-grained feature alignment. We also incorporate orthogonal projection losses within the modality to enhance local feature disentanglement. To further address the shortcut issues caused by statistical biases in the data, we utilize the visual saliency map and concept label statistics as transportation priors. Thus, DOT-CBM can visualize inversion heatmaps, provide more reliable concept predictions, and produce more accurate class predictions. Comprehensive experiments demonstrate that our proposed DOT-CBM achieves SOTA performance on several tasks, including image classification, local part detection and out-of-distribution generalization.

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

In this paper, we present DV-Matcher, a novel learning-based framework for estimating dense correspondences between non-rigidly deformable point clouds. Learning directly from unstructured point clouds without meshing or manual labelling, our framework delivers high-quality dense correspondences, which is of significant practical utility in point cloud processing. Our key contributions are two-fold: First, we propose a scheme to inject prior knowledge from pre-trained vision models into geometric feature learning, which effectively complements the local nature of geometric features with global and semantic information; Second, we propose a novel deformation-based module to promote the extrinsic alignment induced by the learned correspondences, which effectively enhances the feature learning. Experimental results show that our method achieves state-of-the-art results in matching non-rigid point clouds in both near-isometric and heterogeneous shape collection as well as more realistic partial and noisy data. Our code is available at https://github.com/rqhuang88/DV-Matcher.

In recent years, numerous tasks have been proposed to encourage model to develop specified capability in understanding audio-visual scene, primarily categorized into temporal localization, spatial localization, spatio-temporal reasoning, and pixel-level understanding. Instead, human possesses a unified understanding ability for diversified tasks. Therefore, designing an audio-visual model with general capability to unify these tasks is of great value. However, simply joint training for all tasks can lead to interference due to the heterogeneity of audiovisual data and complex relationship among tasks. We argue that this problem can be solved through explicit cooperation among tasks. To achieve this goal, we propose a unified learning method which achieves explicit inter-task cooperation from both the perspectives of data and model thoroughly. Specifically, considering the labels of existing datasets are simple words, we carefully refine these datasets and construct an Audio-Visual Unified Instruction-tuning dataset with Explicit reasoning process (AV-UIE), which clarifies the cooperative relationship among tasks. Subsequently, to facilitate concrete cooperation in learning stage, an interaction-aware LoRA structure with multiple LoRA heads is designed to learn different aspects of audiovisual data interaction. By unifying the explicit cooperation across the data and model aspect, our method not only surpasses existing unified audio-visual model on multiple tasks, but also outperforms most specialized models for certain tasks. Furthermore, we also visualize the process of explicit cooperation and surprisingly find that each LoRA head has certain audio-visual understanding ability. Code and dataset: https://github.com/GeWu-Lab/Crab

In the rapidly advancing field of image generation, *Visual Auto-Regressive* (VAR) modeling has garnered considerable attention for its innovative next-scale prediction approach. This paradigm offers substantial improvements in efficiency, scalability, and zero-shot generalization. Yet, the inherently coarse-to-fine nature of VAR introduces a prolonged token sequence, leading to prohibitive memory consumption and computational redundancies. To overcome these bottlenecks, we propose *Collaborative Decoding* (CoDe), a novel decoding strategy tailored to the VAR framework. CoDe capitalizes on two critical observations: the substantially reduced parameter demands at larger scales and the exclusive generation patterns across different scales. Based on these insights, we partition the multi-scale inference process into a seamless collaboration between a large model and a small model. The large model serves as the 'drafter', specializing in generating low-frequency content at smaller scales, while the smaller model serves as the 'refiner', solely focusing on predicting high-frequency details at larger scales. This collaboration yields remarkable efficiency with minimal impact on quality: CoDe achieves a 1.7x speedup, slashes memory usage by around 50%, and preserves image quality with only a negligible FID increase from 1.95 to 1.98. When drafting steps are further decreased, CoDe can achieve an impressive 2.9x acceleration, reaching over 41 images/s at 256x256 resolution on a single NVIDIA 4090 GPU, while preserving a commendable FID of 2.27.

A person downloading a pre-trained model from the web should be aware of its biases. Existing approaches for bias identification rely on datasets containing labels for the task of interest, something that a non-expert may not have access to, or may not have the necessary resources to collect: this greatly limits the number of tasks where model biases can be identified. In this work, we present Classifier-to-Bias (C2B), the first bias discovery framework that works without access to any labeled data: it only relies on a textual description of the classification task to identify biases in the target classification model. This description is fed to a large language model to generate bias proposals and corresponding captions depicting biases together with task-specific target labels. A retrieval model collects images for those captions, which are then used to assess the accuracy of the model w.r.t. the given biases. C2B is training-free, does not require any annotations, has no constraints on the list of biases, and can be applied to any pre-trained model on any classification task. Experiments on two publicly available datasets show that C2B discovers biases beyond those of the original datasets and outperforms a recent state-of-the-art bias detection baseline that relies on task-specific annotations, being a promising first step toward addressing task-agnostic unsupervised bias detection.