CVPR 2025 Papers

Large pre-trained Vision-Language Models (VLMs) such as CLIP have demonstrated excellent zero-shot generalizability across various downstream tasks. However, recent studies have shown that the inference performance of CLIP can be greatly degraded by small adversarial perturbations, especially its visual modality, posing significant safety threats. To mitigate this vulnerability, in this paper, we propose a novel defense method called Test-Time Adversarial Prompt Tuning (TAPT) to enhance the inference robustness of CLIP against visual adversarial attacks. TAPT is a test-time defense method that learns defensive bimodal (textual and visual) prompts to robustify the inference process of CLIP. Specifically, it is an unsupervised method that optimizes the defensive prompts for each test sample by minimizing a multi-view entropy and aligning adversarial-clean distributions. We evaluate the effectiveness of TAPT on 11 benchmark datasets, including ImageNet and 10 other zero-shot datasets, demonstrating that it enhances the zero-shot adversarial robustness of the original CLIP by at least 48.9% against AutoAttack (AA), while largely maintaining performance on clean examples. Moreover, TAPT outperforms existing adversarial prompt tuning methods across various backbones, achieving an average robustness improvement of at least 36.6%. Code is available at https://github.com/xinwong/TAPT.

Action-driven stochastic human motion prediction aims to generate future motion sequences of a pre-defined target action based on given past observed sequences performing non-target actions. This task primarily presents two challenges. Firstly, generating smooth transition motions is hard due to the varying transition speeds of different actions. Secondly, the action characteristic is difficult to be learned because of the similarity of some actions. These issues cause the predicted results to be unreasonable and inconsistent. As a result, we propose two memory banks, the Soft-transition Action Bank (STAB) and Action Characteristic Bank (ACB), to tackle the problems above. The STAB stores the action transition information. It is equipped with the novel soft searching approach, which encourages the model to focus on multiple possible action categories of observed motions. The ACB records action characteristic, which produces more prior information for predicting certain actions. To fuse the features retrieved from the two banks better, we further propose the Adaptive Attention Adjustment (AAA) strategy. Extensive experiments on four motion prediction datasets demonstrate that our approach consistently outperforms the previous state-of-the-art. The demo and code are available at https://hyqlat.github.io/STABACB.github.io/.

Vision Language Models (VLMs) can produce unintended and harmful content when exposed to adversarial attacks, particularly because their vision capabilities create new vulnerabilities. Existing defenses, such as input preprocessing, adversarial training, and response evaluation-based methods, are often impractical for real-world deployment due to their high costs. To address this challenge, we propose ASTRA, an efficient and effective defense by adaptively steering models away from adversarial feature directions to resist VLM attacks. Our key procedures involve finding transferable steering vectors representing the direction of harmful response and applying adaptive activation steering to remove these directions at inference time. To create effective steering vectors, we randomly ablate the visual tokens from the adversarial images and identify those most strongly associated with jailbreaks. These tokens are then used to construct steering vectors. During inference, we perform the adaptive steering method that involves the projection between the steering vectors and calibrated activation, resulting in little performance drops on benign inputs while strongly avoiding harmful outputs under adversarial inputs. Extensive experiments across multiple models and baselines demonstrate our state-of-the-art performance and high efficiency in mitigating jailbreak risks. Additionally, ASTRA exhibits good transferability, defending against unseen attacks (i.e., structured-based attack, perturbation-based attack with project gradient descent variants, and text-only attack).

The task of occupancy forecasting (OCF) involves utilizing past and present perception data to predict future occupancy states of autonomous vehicle surrounding environments, which is critical for downstream tasks such as obstacle avoidance and path planning. Existing 3D OCF approaches struggle to predict plausible spatial details for movable objects and suffer from slow inference speeds due to neglecting the bias and uneven distribution of changing occupancy states in both space and time. In this paper, we propose a novel spatiotemporal decoupling vision-based paradigm to explicitly tackle the bias and achieve both effective and efficient 3D OCF. To tackle spatial bias in empty areas, we introduce a novel spatial representation that decouples the conventional dense 3D format into 2D bird's-eye view (BEV) occupancy with corresponding height values, enabling 3D OCF derived only from 2D predictions thus enhancing efficiency. To reduce temporal bias on static voxels, we design temporal decoupling to improve end-to-end OCF by temporally associating instances via predicted flows. We develop an efficient multi-head network EfficientOCF to achieve 3D OCF with our devised spatiotemporally decoupled representation. A new metric, conditional IoU (C-IoU), is also introduced to provide a robust 3D OCF performance assessment, especially in datasets with missing or incomplete annotations. The experimental results demonstrate that EfficientOCF surpasses existing baseline methods on accuracy and efficiency, achieving state-of-the-art performance with a fast inference time of 82.33ms with a single GPU. Our code is released at: https://github.com/BIT-XJY/EfficientOCF.

Contrastive Language-Image Pre-training (CLIP) \citep radford2021learning has emerged as a pivotal model in computer vision and multimodal learning, achieving state-of-the-art performance at aligning visual and textual representations through contrastive learning. However, CLIP struggles with potential information misalignment in many image-text datasets and suffers from entangled representation. On the one hand, short captions for a single image in datasets like MSCOCO may describe disjoint regions in the image, leaving the model uncertain about which visual features to retain or disregard. On the other hand, directly aligning long captions with images can lead to the retention of entangled details, preventing the model from learning disentangled, atomic concepts -- ultimately limiting its generalization on certain downstream tasks involving short prompts. In this paper, we establish theoretical conditions that enable flexible alignment between textual and visual representations across varying levels of granularity. Specifically, our framework ensures that a model can not only preserve cross-modal semantic information in its entirety but also disentangle visual representations to capture fine-grained textual concepts. Building on this foundation, we introduce \ours, a novel approach that identifies and aligns the most relevant visual and textual representations in a modular manner. Superior performance across various tasks demonstrates its capability to handle information misalignment and supports our identification theory. The code is available at https://github.com/Mid-Push/SmartCLIP.

Open-world interpretation aims to accurately localize and recognize all objects within images by vision-language models (VLMs). While substantial progress has been made in this task for natural images, the advancements for remote sensing (RS) images still remain limited, primarily due to these two challenges. 1) Existing RS semantic categories are limited, particularly for pixel-level interpretation datasets. 2) Distinguishing among diverse RS spatial regions solely by language space is challenging due to the dense and intricate spatial distribution in open-world RS imagery. To address the first issue, we develop a fine-grained RS interpretation dataset, Sky-SA, which contains 183,375 high-quality local image-text pairs with full-pixel manual annotations, covering 1,763 category labels, exhibiting richer semantics and higher density than previous datasets. Afterwards, to solve the second issue, we introduce the vision-centric principle for vision-language modeling. Specifically, in the pre-training stage, the visual self-supervised paradigm is incorporated into image-text alignment, reducing the degradation of general visual representation capabilities of existing paradigms. Then, we construct a visual-relevance knowledge graph across open-category texts and further develop a novel vision-centric image-text contrastive loss for fine-tuning with text prompts. This new model, denoted as SkySense-O, demonstrates impressive zero-shot capabilities on a thorough evaluation encompassing 14 datasets over 4 tasks, from recognizing to reasoning and classification to localization. Specifically, it outperforms the latest models such as SegEarth-OV, GeoRSCLIP, and VHM by a large margin, i.e., 11.95%, 8.04% and 3.55% on average respectively.

Prompt tuning (PT) has long been recognized as an effective and efficient paradigm for transferring large pre-trained vision-language models (VLMs) to downstream tasks by learning a tiny set of context vectors. Nevertheless, in this work, we reveal that freezing the parameters of VLMs during learning the context vectors neither facilitates the transferability of pre-trained knowledge nor improves the memory and time efficiency significantly. Upon further investigation, we find that reducing both the length and width of the feature-gradient propagation flows of the full fine-tuning (FT) baseline is key to achieving effective and efficient knowledge transfer. Motivated by this, we propose Skip Tuning, a novel paradigm for adapting VLMs to downstream tasks. Unlike existing PT or adapter-based methods, Skip Tuning applies Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip) upon the FT baseline without introducing extra context vectors or adapter modules. Extensive experiments across a wide spectrum of benchmarks demonstrate the superior effectiveness and efficiency of our Skip Tuning over both PT and adapter-based methods. Code: https://github.com/anonymity-007/SkipT.

Vision transformers (ViTs) have emerged as a popular backbone for visual tracking. However, complete ViT architectures are too cumbersome to deploy for unmanned aerial vehicle (UAV) tracking which extremely emphasizes efficiency. In this study, we discover that many layers within lightweight ViT-based trackers tend to learn relatively redundant and repetitive target representations. Based on this observation, we propose a similarity-guided layer adaptation approach to optimize the structure of ViTs. Our approach dynamically disables a large number of representation-similar layers and selectively retains only a single optimal layer among them, aiming to achieve a better accuracy-speed trade-off. By incorporating this approach into existing ViTs, we tailor previously complete ViT architectures into an efficient similarity-guided layer-adaptive framework, namely SGLATrack, for real-time UAV tracking. Extensive experiments on six tracking benchmarks verify the effectiveness of the proposed approach, and show that our SGLATrack achieves a state-of-the-art real-time speed while maintaining competitive tracking precision. Codes and models are available at https://github.com/GXNU-ZhongLab/SGLATrack.

Integrating large language models (LLMs) into autonomous driving has attracted significant attention with the hope of improving generalization and explainability. However, existing methods often focus on either driving or vision-language understanding but achieving both high driving performance and extensive language understanding remains challenging. In addition, the dominant approach to tackle vision-language understanding is using visual question answering. However, for autonomous driving, this is only useful if it is aligned with the action space. Otherwise, the model's answers could be inconsistent with its behavior. Therefore, we propose a model that can handle three different tasks: (1) closed-loop driving, (2) vision-language understanding, and (3) language-action alignment. Our model SimLingo is based on a vision language model (VLM) and works using only camera, excluding expensive sensors like LiDAR. SimLingo obtains state-of-the-art performance on the widely used CARLA simulator on the Bench2Drive benchmark and is the winning entry at the CARLA challenge 2024. Additionally, we achieve strong results in a wide variety of language-related tasks while maintaining high driving performance.

Building Graphical User Interface (GUI) assistants holds significant promise for enhancing human workflow productivity. While most agents are language-based, relying on closed-source API with text-rich meta-information (e.g., HTML or accessibility tree), they show limitations in perceiving UI visuals as humans do, highlighting the need for GUI visual agents.In this work, we develop a vision-language-action model in the digital world, namely "Our Model," which features the following innovations:1. **UI-Guided Visual Token Selection** to reduce computational costs by formulating screenshots as a UI-connected graph, adaptively identifying their redundant relationships and serving as the criteria for token selection during self-attention blocks. 2. **Interleaved Vision-Language-Action Streaming** that flexibly unifies diverse needs within GUI tasks, enabling effective management of visual-action history in navigation or pairing multi-turn query-action sequences per screenshot to enhance training efficiency. 3. **Small-Scale High-Quality GUI Instruction-Following Datasets** by careful data curation and employing a resampling strategy to address significant data type imbalances. With the above components, our model, a lightweight 2B model using 256K data, achieves a strong 75.1% accuracy in zero-shot screenshot grounding. Its UI-guided token selection further reduces 33% of redundant visual tokens during training and speeds up performance by 1.4x. Navigation experiments across web, mobile, and online environments further underscore the effectiveness and potential of our model in advancing GUI visual agents.

Skeleton-based human action recognition is promising due to its privacy preservation, robustness to visual challenges, and computational efficiency. Especially, the practical necessity to recognize unseen actions has led to increased interest in zero-shot skeleton-based action recognition (ZSSAR). Existing ZSSAR approaches often rely on manually crafted action descriptions or visual assumptions to enhance knowledge transfer, which is limited in flexibility and prone to inaccuracies and noise. To overcome this, we introduce Semantic-guided Cross-Modal Prompt Learning (SCoPLe), a novel framework that replaces manual guidance with data-driven prompt learning for refinement and alignment of skeletal and textual features. Specifically, we introduce a dual-stream language prompting module that preserves the original semantic context from the pre-trained text encoder while still effectively tuning its ouput for ZSSAR task adaptation. We also introduce a joint-shaped prompting module that learns tuning for skeleton features and incorporate an adaptive visual representation sampler that leverages text semantics to strengthen the cross-modal prompting interactions during skeleton-to-text embedding projection. Experimental results on the NTU-RGB+D and PKU-MMD datasets demonstrate the state-of-the-art performance of our method in both ZSSAR and generalized ZSSAR scenarios.

Natural language goes beyond dryly describing visual content. It contains rich abstract concepts to express feeling, creativity and properties that cannot be directly perceived. Yet, current research in Vision Language Models (VLMs) has not shed light on abstract-oriented language.Our research breaks new ground by uncovering its wide presence and under-estimated value, with extensive analysis.Particularly, we focus our investigation on the fashion domain, a highly-representative field with abstract expressions. By analyzing recent large-scale multimodal fashion datasets, we find that abstract terms have a dominant presence, rivaling the concrete ones, providing novel information, and being useful in the retrieval task. However, a critical challenge emerges: current general-purpose or fashion-specific VLMs are pre-trained with databases that lack sufficient abstract words in their text corpora, thus hindering their ability to effectively represent abstract-oriented language. We propose a training-free and model-agnostic method, Abstract-to-Concrete Translator (ACT), to shift abstract representations towards well-represented concrete ones in the VLM latent space, using pre-trained models and existing multimodal databases.On the text-to-image retrieval task, despite being training-free, ACT outperforms the fine-tuned VLMs in both same- and cross-dataset settings, exhibiting its effectiveness with a strong generalization capability. Moreover, the improvement introduced by ACT is consistent with various VLMs, making it a plug-and-play solution.

Large vision-language models (LVLMs) have shown remarkable capabilities in interpreting visual content. While existing works demonstrate these models' vulnerability to deliberately placed adversarial texts, such texts are often easily identifiable as anomalous. In this paper, we present the first approach to generate scene-coherent typographic adversarial attacks that mislead advanced LVLMs while maintaining visual naturalness through the capability of the LLM-based agent. Our approach addresses three critical questions: what adversarial text to generate, where to place it within the scene, and how to integrate it seamlessly. We propose a training-free, multi-modal LLM-driven scene-coherent typographic adversarial planning (SceneTAP) that employs a three-stage process: scene understanding, adversarial planning, and seamless integration. The SceneTAP utilizes chain-of-thought reasoning to comprehend the scene, formulate effective adversarial text, strategically plan its placement, and provide detailed instructions for natural integration within the image. This is followed by a scene-coherent TextDiffuser that executes the attack using a local diffusion mechanism. We extend our method to real-world scenarios by printing and placing generated patches in physical environments, demonstrating its practical implications. Extensive experiments show that our scene-coherent adversarial text successfully misleads state-of-the-art LVLMs, including ChatGPT-4o, even after capturing new images of physical setups. Our evaluations demonstrate a significant increase in attack success rates while maintaining visual naturalness and contextual appropriateness. This work highlights vulnerabilities in current vision-language models to sophisticated, scene-coherent adversarial attacks and provides insights into potential defense mechanisms. We release our code at https://github.com/tsingqguo/scenetap.

High-resolution perception of visual details is crucial for daily tasks. Current vision pre-training, however, is still limited to low resolutions (e.g., 378 x 378 pixels) due to the quadratic cost of processing larger images. We introduce PS3 that scales CLIP-style vision pre-training to 4K resolution with a near-constant cost. Instead of contrastive learning on global image representation, PS3 is pre-trained by selectively processing local regions and contrasting them with local detailed captions, enabling high-resolution representation learning with greatly reduced computational overhead. The pre-trained PS3 is able to both encode the global image at low resolution and selectively process local high-resolution regions based on their saliency or relevance to a text prompt. When applying PS3 to multi-modal LLM (MLLM), the resulting model, named VILA-HD, significantly improves high-resolution visual perception compared to baselines without high-resolution vision pre-training such as AnyRes and S^2 while using up to 4.3x fewer tokens. PS3 also unlocks appealing scaling properties of VILA-HD, including scaling up resolution for free and scaling up test-time compute for better performance. Compared to state of the arts, VILA-HD outperforms previous MLLMs such as NVILA and Qwen2-VL across multiple benchmarks and achieves better efficiency than latest token pruning approaches. Finally, we find current benchmarks do not require 4K-resolution perception, which motivates us to propose 4KPro, a new benchmark of image QA at 4K resolution, on which VILA-HD outperforms all previous MLLMs, including a 14.5% improvement over GPT-4o, and a 3.2% improvement and 2.96x speedup over Qwen2-VL.

Vision-language temporal alignment is a crucial capability for human dynamic recognition and cognition in real-world scenarios. While existing research focuses on capturing vision-language relevance, it faces limitations due to biased temporal distributions, imprecise annotations, and insufficient compositionally. To achieve fair evaluation and comprehensive exploration, our objective is to investigate and evaluate the ability of models to achieve alignment from a temporal perspective, specifically focusing on their capacity to synchronize visual scenarios with linguistic context in a temporally coherent manner. As a preliminary step, we present the statistical analysis of existing benchmarks and reveal the existing challenges from a decomposed perspective. To this end, we introduce SVLTA, the synthetic vision-language temporal alignment derived via a well-designed and feasible control generation method within a simulation environment. The approach considers commonsense knowledge, manipulable action, and constrained filtering, which generates reasonable, diverse, and balanced data distributions for diagnostic evaluations. Our experiments reveal diagnostic insights through the evaluations in temporal question answering, distributional shift sensitiveness, and temporal alignment adaptation.

Zero-shot multi-label recognition (MLR) with Vision-Language Models (VLMs) faces significant challenges without training data, model tuning, or architectural modifications. Existing approaches require prompt tuning or architectural adaptations, limiting zero-shot applicability. Our work proposes a novel solution treating VLMs as black boxes, leveraging scores without training data or ground truth. We make two contributions. First, we find that VLM scores suffer from image- and prompt-specific biases, and that simple standardization is surprisingly effective at removing these and boosting MLR performance. And second, we introduce compound prompts grounded in realistic object combinations. Our analysis reveals "AND"/"OR" signal ambiguities that cause maximum compound scores to be surprisingly suboptimal compared to second-highest scores. We introduce an adaptive fusion method to address this issue. Our method enhances other zero-shot approaches, consistently improving their results. Experiments show superior mean Average Precision (mAP) compared to methods requiring training data, achieved through refined object ranking for robust zero-shot MLR. Code can be found at https://github.com/kjmillerCURIS/SPARC.

The integration of Vision-Language Models (VLMs) into autonomous driving systems has shown promise in addressing key challenges such as learning complexity, interpretability, and common-sense reasoning. However, existing approaches often struggle with efficient integration and real-time decision-making due to computational demands. In this paper, we introduce SOLVE, an innovative framework that synergizes VLMs with end-to-end (E2E) models to enhance autonomous vehicle planning. Our approach emphasizes knowledge sharing at the feature level through a shared visual encoder, enabling comprehensive interaction between VLM and E2E components. We propose a Trajectory Chain-of-Thought (T-CoT) paradigm, which progressively refines trajectory predictions, reducing uncertainty and improving accuracy. By employing a temporal decoupling strategy, SOLVE achieves efficient asynchronous cooperation, aligning high-quality VLM outputs with E2E real-time performance. Evaluated on the nuScenes dataset, our method demonstrates significant improvements in trajectory prediction accuracy, paving the way for more robust and reliable autonomous driving systems.

Human beings are social animals. How to equip 3D autonomous characters with similar social intelligence that can perceive, understand and interact with humans remains an open yet foundamental problem. In this paper, we introduce SOLAMI, the first end-to-end Social vision-Language-Action (VLA) Modeling framework for Immersive interaction with 3D autonomous characters. Specifically, SOLAMI builds 3D autonomous characters from three aspects: 1) Social VLA Architecture: We propose a unified social VLA framework to generate multimodal response (speech and motion) based on the user's multimodal input to drive the character for social interaction. 2) Interactive Multimodal Data: We present SynMSI, a synthetic multimodal social interaction dataset generated by an automatic pipeline using only existing motion datasets to address the issue of data scarcity. 3) Immersive VR Interface: We develop a VR interface that enables users to immersively interact with these characters driven by various architectures. Extensive quantitative experiments and user studies demonstrate that our framework leads to more precise and natural character responses (in both speech and motion) that align with user expectations with lower latency.

Existing view transformations in vision-centric 3D Semantic Scene Completion (SSC) inevitably experience erroneous feature duplication in the reconstructed voxel space due to occlusions, leading to a dilution of informative contexts. Furthermore, semantic classes exhibit high variability in their appearance in real-world driving scenarios. To address these issues, we introduce a novel 3D SSC method, called SOAP, including two key components: an occluded region-aware view projection and a scene-adaptive decoder. The occluded region-aware view projection effectively converts 2D image features into voxel space, refining the duplicated features of occluded regions using information gathered from previous observations. The scene-adaptive decoder guides query embeddings to learn diverse driving environments based on a comprehensive semantic repository. Extensive experiments validate that the proposed SOAP significantly outperforms existing methods for the vision-centric 3D SSC on automated driving datasets, SemanticKITTI and SSCBench. Code is available at https://github.com/gywns6287/SOAP.