CVPR 2025 Papers

We present PAVE, a framework for adapting pre-trained video large language models (Video-LLMs) to downstream tasks that incorporate side-channel signals, such as audio, camera pose, or high frame rate videos. PAVE introduces a lightweight adaptation strategy called "patching", which adds a small number of parameters and operations to the base model without modifying its architecture or pre-trained weights. We demonstrate that PAVE effectively enhances pre-trained Video-LLMs with the cost of adding <1% additional FLOPs and parameters for diverse tasks, including audio-visual understanding, 3D reasoning, and multi-view video understanding, surpassing state-of-the-art task-specific models. Moreover, when applied to high frame rate videos, PAVE further improves video understanding, enhancing the performance of strong base models. Finally, our experiments show that our framework generalizes well across different Video-LLMs.

Vision language models (VLMs) respond to user-crafted text prompts and visual inputs, and are applied to numerous real-world problems. VLMs integrate visual modalities with large language models (LLMs), which are well known to be prompt-sensitive. Hence, it is crucial to determine whether VLMs inherit this instability to varying prompts. We therefore investigate which prompt variations VLMs are most sensitive to and which VLMs are most agnostic to prompt variations. To this end, we introduce PARC (Prompt Analysis via Reliability and Calibration), a VLM prompt sensitivity analysis framework built on three pillars: (1) plausible prompt variations in both the language and vision domain, (2) a novel model reliability score with built-in guarantees, and (3) a calibration step that enables dataset- and prompt-spanning prompt variation analysis. Regarding prompt variations, PARC's evaluation shows that VLMs mirror LLM language prompt sensitivity in the vision domain, and most destructive variations change the expected answer. Regarding models, outstandingly robust VLMs among 22 evaluated models come from the InternVL2 family. We further find indications that prompt sensitivity is linked to training data. https://github.com/NVlabs/PARC

Visual Language Models require substantial computational resources for inference due to the additional input tokens needed to represent visual information. However, these visual tokens often contain redundant and unimportant information, resulting in an unnecessarily high number of tokens. To address this, we introduce PACT, a method that reduces inference time and memory usage by pruning irrelevant tokens and merging visually redundant ones at an early layer of the language model. Our approach uses a novel importance metric to identify unimportant tokens without relying on attention scores, making it compatible with FlashAttention. We also propose a novel clustering algorithm, called Distance Bounded Density Peak Clustering, which efficiently clusters visual tokens while constraining the distances between elements within a cluster by a predefined threshold. We demonstrate the effectiveness of PACT through extensive experiments.

Vision-language models (VLMs), such as CLIP, have shown remarkable capabilities in downstream tasks. However, the coupling of semantic information between the foreground and the background in images leads to significant shortcut issues that adversely affect out-of-distribution (OOD) detection abilities. When confronted with a background OOD sample, VLMs are prone to misidentifying it as in-distribution (ID) data. In this paper, we analyze the OOD problem from the perspective of shortcuts in VLMs and propose OSPCoOp which includes background decoupling and mask-guided region regularization. We first decouple images into ID-relevant and ID-irrelevant regions and utilize the latter to generate a large number of augmented OOD background samples as pseudo-OOD supervision. We then use the masks from background decoupling to adjust the model's attention, minimizing its focus on ID-irrelevant regions. To assess the model's robustness against background interference, we introduce a new OOD evaluation dataset, ImageNet-Bg, which solely consists of background images with all ID-relevant regions removed. Our method demonstrates exceptional performance in few-shot scenarios, achieving strong results even in one-shot setting, and outperforms existing methods. The code and proposed ImageNet-Bg are available at https://github.com/HAIV-Lab/OSPCoOp_Imagenet-bg.

Top-down attention plays a crucial role in the human vision system, wherein the brain initially obtains a rough overview of a scene to discover salient cues (i.e., overview first), followed by a more careful finer-grained examination (i.e., look closely next). However, modern ConvNets remain confined to a pyramid structure that successively downsamples the feature map for receptive field expansion, neglecting this crucial biomimetic principle. We present OverLoCK, the first pure ConvNet backbone architecture that explicitly incorporates a top-down attention mechanism. Unlike pyramid backbone networks, our design features a branched architecture with three synergistic sub-networks: 1) a Base-Net that encodes low/mid-level features; 2) a lightweight Overview-Net that generates dynamic top-down attention through coarse global context modeling (i.e., overview first); and 3) a robust Focus-Net that performs finer-grained perception guided by top-down attention (i.e., look closely next). To fully unleash the power of top-down attention, we further propose a novel context-mixing dynamic convolution (ContMix) that effectively models long-range dependencies while preserving inherent local inductive biases even when the input resolution increases, addressing critical limitations in existing convolutions. Our OverLoCK exhibits a notable performance improvement over existing methods. For instance, OverLoCK-T achieves a Top-1 accuracy of 84.2%, significantly surpassing ConvNeXt-B while using only around one-third of the FLOPs/parameters. On object detection, our OverLoCK-S clearly surpasses MogaNet-B by 1% in AP^b. On semantic segmentation, our OverLoCK-T remarkably improves UniRepLKNet-T by 1.7% in mIoU. Code is publicly available at https://rb.gy/wit4jh.

Existing image-to-3D creation methods typically split the task into two individual stage: multi-view image generation and 3D reconstruction, leading to two main limitations: (1) In multi-view generation stage, the multi-view generated images present a challenge to preserving 3D consistency;; (2) In 3D reconstruction stage, there is a domain gap between real training data and generated multi-view input during inference. To address these issues, we propose Ouroboros3D, end-to-end trainable framework that integrates multi-view generation and 3D reconstruction into a recursive diffusion process through feedback mechanism.Our framework operates through iterative cycles where each cycle consists of a denoising process and a reconstruction step.By incorporating a 3D-aware feedback mechanism, our multi-view generative model leverages the explicit 3D geometric information (e.g. texture, position) from the feedback of reconstruction results of the previous process as conditions, thus modeling consistency at the 3D geometric level. Furthermore, through joint training of both the multi-view generative and reconstruction models, we alleviate reconstruction stage domain gap and enable mutual enhancement within the recursive process. Experimental results demonstrate that Ouroboros3D outperforms methods that treat these stages separately and those that combine them only during inference, achieving superior multi-view consistency and producing 3D models with higher geometric realism.

Class Incremental Learning (CIL) aims to enable models to learn new classes sequentially while retaining knowledge of previous ones. Although current methods have alleviated catastrophic forgetting (CF), recent studies highlight that the performance of CIL models is highly sensitive to the order of class arrival, particularly when sequentially introduced classes exhibit high inter-class similarity. To address this critical yet understudied challenge of class order sensitivity, we first extend existing CIL frameworks through theoretical analysis, proving that grouping classes with lower pairwise similarity during incremental phases significantly improves model robustness to order variations. Building on this insight, we propose Graph-Driven Dynamic Similarity Grouping (GDDSG), a novel method that employs graph coloring algorithms to dynamically partition classes into similarity-constrained groups. Each group trains an isolated CIL sub-model and constructs meta-features for class group identification. Experimental results demonstrate that our method effectively addresses the issue of class order sensitivity while achieving optimal performance in both model accuracy and anti-forgetting capability. Our code is available at https://github.com/AIGNLAI/GDDSG.

Rolling shutter (RS) cameras dominate consumer and smartphone markets. Several methods for computing the absolute pose of RS cameras have appeared in the last 20 years, but the relative pose problem has not been fully solved yet. We provide a unified theory for the important class of order-one rolling shutter (RS_1) cameras. These cameras generalize the perspective projection to RS cameras, projecting a generic space point to exactly one image point via a rational map. We introduce a new back-projection RS camera model, characterize RS_1 cameras, construct explicit parameterizations of such cameras, and determine the image of a space line. We classify all minimal problems for solving the relative camera pose problem with linear RS_1 cameras and discover new practical cases. Finally, we show how the theory can be used to explain RS models previously used for absolute pose computation.

In clinical practice, panoramic dental radiography is a widely employed imaging technique that can provide a detailed and comprehensive view of dental structures and surrounding tissues for identifying various oral anomalies. However, due to the complexity of oral anomalies and the scarcity of available data, existing research still suffers from substantial challenges in automated oral anomaly detection. To this end, this paper presents a new hospital-scale panoramic X-ray benchmark, namely "OralXrays-9", which consists of 12,688 panoramic X-ray images with 84,113 meticulously annotated instances across nine common oral anomalies. Correspondingly, we propose a personalized Multi-Object Query-Aware Mining (MOQAM) paradigm, which jointly incorporates the Distribution-IoU Region Proposal Network (DI-RPN) and Class-Balanced Spherical Contrastive Regularization (CB-SCR) mechanisms to address the challenges posed by multi-scale variations and class-imbalanced distributions. To the best of our knowledge, this is the first attempt to develop AI-driven diagnostic systems specifically designed for multi-object oral anomaly detection, utilizing publicly available data resources. Extensive experiments on the newly-published OralXrays-9 dataset and real-world nature scenarios consistently demonstrate the superiority of our MOQAM in revolutionizing oral healthcare practices.

Developing a face anti-spoofing model that meets the security requirements of clients worldwide is challenging due to the domain gap between training datasets and the diverse end-user test data. Moreover, for security and privacy reasons, it is undesirable for clients to share a large amount of their face data with service providers. In this work, we introduce a novel method in which the face anti-spoofing model can be adapted by the client itself to a target domain at test time using only a small sample of data while keeping model parameters and training data inaccessible to the client. Specifically, we develop a prototype-based base model and an optimal transport-guided adaptor that enables adaptation in either a lightweight training or training-free fashion, without updating the base model's parameters. Furthermore, we propose geodesic mixup, an optimal transport-based synthesis method that generates augmented training data along the geodesic path between source prototypes and the target data distribution. This allows training a lightweight classifier to effectively adapt to target-specific characteristics while retaining essential knowledge learned from the source domain. In cross-domain and cross-attack settings, compared with recent methods, our method achieves average relative improvements of 19.17% in HTER and 8.58% in AUC, respectively.

Optical imaging capable of resolving nanoscale features would revolutionize scientific research and engineering applications across biomedicine, smart manufacturing, and semiconductor quality control. However, due to the physical phenomenon of diffraction, the optical resolution is limited to approximately half the wavelength of light, which impedes the observation of subwavelength objects such as the native state coronavirus, typically smaller than 200 nm. Fortunately, deep learning methods have shown remarkable potential in uncovering underlying patterns within data, promising to overcome the diffraction limit by revealing the mapping pattern between diffraction images and their corresponding ground truth object images. However, the absence of suitable datasets has hindered progress in this field--collecting high-quality optical data of subwavelength objects is highly difficult as these objects are inherently invisible under conventional microscopy, making it impossible to perform standard visual calibration and drift correction. Therefore, we provide the first general optical imaging dataset based on the "building block" concept for challenging the diffraction limit. Drawing an analogy to modular construction principles, we construct a comprehensive optical imaging dataset comprising subwavelength fundamental elements, i.e., small square units that can be assembled into larger and more complex objects. We then frame the task as an image-to-image translation task and evaluate various vision methods. Experimental results validate our "building block" concept, demonstrating that models trained on basic square units can effectively generalize to realistic, more complex unseen objects. Most importantly, by highlighting this underexplored AI-for-science area and its potential, we aspire to advance optical science by fostering collaboration with the vision and machine learning communities.

While text-to-video diffusion models have made significant strides, many still face challenges in generating videos with temporal consistency. Within diffusion frameworks, guidance techniques have proven effective in enhancing output quality during inference; however, applying these methods to video diffusion models introduces additional complexity of handling computations across entire sequences. To address this, we propose a novel framework called MotionPrompt that guides the video generation process via optical flow. Specifically, we train a discriminator to distinguish optical flow between random pairs of frames from real videos and generated ones. Given that prompts can influence the entire video, we optimize learnable token embeddings during reverse sampling steps by using gradients from a trained discriminator applied to random frame pairs. This approach allows our method to generate visually coherent video sequences that closely reflect natural motion dynamics, without compromising the fidelity of the generated content. We demonstrate the effectiveness of our approach across various models. Project Page: https://motionprompt.github.io

Scattered light from pulsed lasers is increasingly part of our ambient illumination, as many devices rely on them for active 3D sensing. In this work, we ask: can these "ambient" light signals be detected and leveraged for passive 3D vision? We show that pulsed lasers, despite being weak and fluctuating at MHz to GHz frequencies, leave a distinctive sinc comb pattern in the temporal frequency domain of incident flux that is specific to each laser and invariant to the scene. This enables their passive detection and analysis with a free-running SPAD camera, even when they are unknown, asynchronous, out of sight, and emitting concurrently. We show how to synchronize with such lasers computationally, characterize their pulse emissions, separate their contributions, and--if many are present--localize them in 3D and recover a depth map of the camera's field of view. We use our camera prototype to demonstrate (1) a first-of-its-kind visualization of asynchronously propagating light pulses from multiple lasers through the same scene, (2) passive estimation of a laser's MHz-scale pulse repetition frequency with mHz precision, and (3) mm-scale 3D imaging over room-scale distances by passively harvesting photons from two or more out-of-view lasers.

The growing reliance on Artificial Intelligence (AI) in critical domains such as healthcare demands robust mechanisms to ensure the trustworthiness of these systems, especially when faced with unexpected or anomalous inputs. This paper introduces the Open Medical Imaging Benchmarks for Out-Of-Distribution Detection (OpenMIBOOD), a comprehensive framework for evaluating out-of-distribution (OOD) detection methods specifically in medical imaging contexts. OpenMIBOOD includes three benchmarks from diverse medical domains, encompassing 14 datasets divided into covariate-shifted in-distribution, near-OOD, and far-OOD categories. We evaluate 24 post-hoc methods across these benchmarks, providing a standardized reference to advance the development and fair comparison of OOD detection methods. Results reveal that findings from broad-scale OOD benchmarks in natural image domains do not translate to medical applications, underscoring the critical need for such benchmarks in the medical field. By mitigating the risk of exposing AI models to inputs outside their training distribution, OpenMIBOOD aims to support the advancement of reliable and trustworthy AI systems in healthcare. The repository is available at https://github.com/remic-othr/OpenMIBOOD.

Multimodal Large Language Models (MLLMs) have made significant strides in visual understanding and generation tasks. However, generating interleaved image-text content remains a challenge, which requires integrated multimodal understanding and generation abilities. While the progress in unified models offers new solutions, existing benchmarks are insufficient for evaluating these methods due to data size and diversity limitations. To bridge this gap, we introduce OpenING, a comprehensive benchmark comprising 5,400 high-quality human-annotated instances across 56 real-world tasks. OpenING covers diverse daily scenarios such as travel guide, design, and brainstorming, offering a robust platform for challenging interleaved generation methods. In addition, we present IntJudge, a judge model for evaluating open-ended multimodal generation methods. Trained with a novel data pipeline, our IntJudge achieves an agreement rate of 82.42% with human judgments, outperforming GPT-based evaluators by 11.34%. Extensive experiments on OpenING reveal that current interleaved generation methods still have substantial room for improvement. Key findings on interleaved image-text generation are further presented to guide the development of next-generation models.

Recent advancements in visual generation technologies have markedly increased the scale and availability of video datasets, which are crucial for training effective video generation models. However, a significant lack of high-quality, human-centric video datasets presents a challenge to progress in this field. To bridge this gap, we introduce OpenHumanVid, a large-scale and high-quality human-centric video dataset characterized by precise and detailed captions that encompass both human appearance and motion states, along with supplementary human motion conditions, including skeleton sequences and speech audio. To validate the efficacy of this dataset and the associated training strategies, we propose an extension of existing classical diffusion transformer architectures and conduct further pretraining of our models on the proposed dataset. Our findings yield two critical insights: First, the incorporation of a large-scale, high-quality dataset substantially enhances evaluation metrics for generated human videos while preserving performance in general video generation tasks. Second, the effective alignment of text with human appearance, human motion, and facial motion is essential for producing high-quality video outputs. Based on these insights and corresponding methodologies, the straightforward extended network trained on the proposed dataset demonstrates an obvious improvement in the generation of human-centric videos.

The challenge in open-world object detection, similarly to few- and zero-shot learning, is to generalize beyond the class distribution of the training data. In this paper, we propose a general class-agnostic objectness measure to limit bias toward labeled samples. One issue in open-world detection is that previously unseen objects are often misclassified as known categories or filtered as background by classifiers. To prevent this, we explicitly model the joint distribution of objectness and category labels using variational approximation. However, without sufficient labeled data, minimizing the KL divergence between the estimated posterior and a static normal prior fails to converge. Our theoretical analysis identifies the root cause of this failure and motivates adopting a Gaussian prior with variance dynamically adapted to the estimated posterior as a surrogate. To further reduce misclassification, we introduce an energy-based margin loss that encourages unknown objects to move toward high-density regions of the distribution, thus reducing the uncertainty of unknown detections. Our Open-World OBJectness modeling (OWOBJ) boosts novel object detection, especially in low-data regimes. OWOBJ is a flexible plugin that outperforms baselines in Open-World, Few-Shot, and zero-shot Open-Vocabulary Object Detection.