CVPR 2025 Papers

Autofocus is a crucial component of modern digital cameras. While recent learning-based methods achieve state-of-the-art in focus prediction accuracy, they unfortunately ignore the potential focus hunting phenomenon of back-and-forth lens movement in the multi-step focusing procedure. To address this, in this paper, we propose an expert regularized deep reinforcement learning (DRL)-based approach for autofocus, which can utilize the sequential information of lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting. Our method generally follows an actor-critic framework. To accelerate the DRL's training with a higher sample efficiency, we initialize the policy with a pre-trained single-step prediction network, where the network is further improved by modifying the output of absolute in-focus position distribution to the relative lens movement distribution to establish a better mapping between input images and lens movement. To further stabilize DRL's training with a lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset of expert trajectories to regularize the actor network's training. Empirical evaluations show that our method outperforms those learning-based methods on public benchmarks, with higher single- and multi-step prediction accuracy, and a significant reduction of focus hunting rate.

To operate safely, autonomous vehicles (AVs) need to detect and handle unexpected objects or anomalies on the road. While significant research exists for anomaly detection and segmentation in 2D, research progress in 3D is underexplored. Existing datasets lack high-quality multimodal data that are typically found in AVs. This paper presents a novel dataset for anomaly segmentation in driving scenarios. To the best of our knowledge, it is the first publicly available dataset focused on road anomaly segmentation with dense 3D semantic labeling, incorporating both LiDAR and camera data, as well as sequential information to enable anomaly detection across various ranges. This capability is critical for the safe navigation of autonomous vehicles. We adapted and evaluated several baseline models for 3D segmentation, highlighting the challenges of 3D anomaly detection in driving environments. Our dataset and evaluation code will be openly available, facilitating the testing and performance comparison of different approaches.

Spike-based Transformer presents a compelling and energy-efficient alternative to traditional Artificial Neural Network (ANN)-based Transformers, achieving impressive results through sparse binary computations. However, existing spike-based transformers predominantly focus on spatial attention while neglecting crucial temporal dependencies inherent in spike-based processing, leading to suboptimal feature representation and limited performance. To address this limitation, we propose Spiking Transformer with Spatial-Temporal Attention (STAtten), a simple and straightforward architecture that efficiently integrates both spatial and temporal information in the self-attention mechanism. STAtten introduces a block-wise computation strategy that processes information in spatial-temporal chunks, enabling comprehensive feature capture while maintaining the same computational complexity as previous spatial-only approaches. Our method can be seamlessly integrated into existing spike-based transformers without architectural overhaul. Extensive experiments demonstrate that STAtten significantly improves the performance of existing spike-based transformers across both static and neuromorphic datasets, including CIFAR10/100, ImageNet, CIFAR10-DVS, and N-Caltech101.

3D Gaussian Splatting (3D-GS) is a recent 3D scene reconstruction technique that enables real-time rendering of novel views by modeling scenes as parametric point clouds of differentiable 3D Gaussians. However, its rendering speed and model size still present bottlenecks, especially in resource-constrained settings. In this paper, we identify and address two key inefficiencies in 3D-GS to substantially improve rendering speed.These improvements also yield the ancillary benefits of reduced model size and training time. First, we optimize the rendering pipeline to precisely localize Gaussians in the scene, boosting rendering speed without altering visual fidelity. Second, we introduce a novel pruning technique and integrate it into the training pipeline, significantly reducing model size and training time while further raising rendering speed. Our Speedy-Splat approach combines these techniques to accelerate average rendering speed by a drastic 6.71x across scenes from the Mip-NeRF 360, Tanks&Temples, and Deep Blending datasets.

Current learning-based methods predict NeRF or 3D Gaussians from point clouds to achieve photo-realistic rendering but still depend on categorical priors, dense point clouds, or additional refinements. Hence, we introduce a novel point cloud rendering method by predicting 2D Gaussians from point clouds. Our method incorporates two identical modules with an entire-patch architecture enabling the network to be generalized to multiple datasets. The module normalizes and initializes the Gaussians utilizing the point cloud information including normals, colors and distances. Then, splitting decoders are employed to refine the initial Gaussians by duplicating them and predicting more accurate results, making our methodology effectively accommodate sparse point clouds as well. Once trained, our approach exhibits direct generalization to point clouds across different categories. The predicted Gaussians are employed directly for rendering without additional refinement on the rendered images, retaining the benefits of 2D Gaussians. We conduct extensive experiments on various datasets, and the results demonstrate the superiority and generalization of our method, which achieves SOTA performance.

Object removal has so far been dominated by the mask-and-inpaint paradigm, where the masked region is excluded from the input, leaving models relying on unmasked areas to inpaint the missing region. However, this approach lacks contextual information for the masked area, often resulting in unstable performance. In this work, we introduce SmartEraser, built with a new removing paradigm called Masked-Region Guidance. This paradigm retains the masked region in the input, using it as guidance for the removal process. It offers several distinct advantages: (a) it guides the model to accurately identify the object to be removed, preventing its regeneration in the output; (b) since the user mask often extends beyond the object itself, it aids in preserving the surrounding context in the final result. Leveraging this new paradigm, we present Syn4Removal, a large-scale object removal dataset, where instance segmentation data is used to copy and paste objects onto images as removal targets, with the original images serving as ground truths. Experimental results demonstrate that SmartEraser significantly outperforms existing methods, achieving superior performance in object removal, especially in complex scenes with intricate compositions.

Traditional reinforcement learning methods for human-object interaction (HOI) rely on labor-intensive, manually designed skill rewards that do not generalize well across different interactions. We introduce SkillMimic, a unified data-driven framework that fundamentally changes how agents learn interaction skills by eliminating the need for skill-specific rewards. Our key insight is that a unified HOI imitation reward can effectively capture the essence of diverse interaction patterns from HOI datasets. This enables SkillMimic to learn a single policy that not only masters multiple interaction skills but also facilitates skill transitions, with both diversity and generalization improving as the HOI dataset grows. For evaluation, we collect and introduce two basketball datasets containing approximately 35 minutes of diverse basketball skills. Extensive experiments show that SkillMimic successfully masters a wide range of basketball skills including stylistic variations in dribbling, layup, and shooting. Moreover, these learned skills can be effectively composed by a high-level controller to accomplish complex and long-horizon tasks such as consecutive scoring, opening new possibilities for scalable and generalizable interaction skill learning. Project page: https://ingrid789.github.io/SkillMimic/

We introduce Sketchtopia, a large-scale dataset and AI framework designed to explore goal-driven, multimodal communication through asynchronous interactions in a Pictionary-inspired setup. Sketchtopia captures natural human interactions, including freehand sketches, open-ended guesses, and iconic feedback gestures, showcasing the complex dynamics of cooperative communication under constraints. It features over 20K gameplay sessions from 916 players, capturing 263K sketches, 10K erases, 56K guesses and 19.4K iconic feedbacks. We introduce multimodal foundational agents with capabilities for generative sketching, guess generation and asynchronous communication. Our dataset also includes 800 human-agent sessions for benchmarking the agents. We introduce novel metrics to characterize collaborative success, responsiveness to feedback and inter-agent asynchronous communication. Sketchtopia pushes the boundaries of multimodal AI, establishing a new benchmark for studying asynchronous, goal-oriented interactions between humans and AI agents. The dataset can be found at https://sketchtopia25.github.io/

While modern visual recognition systems have made significant advancements, many continue to struggle with the open problem of learning from few exemplars. This paper focuses on the task of object detection in the setting where object classes follow a natural long-tailed distribution. Existing methods for long-tailed detection resort to external ImageNet labels to augment the low-shot training instances. However, such dependency on a large labeled database has limited utility in practical scenarios. We propose a versatile and scalable approach to leverage optional unlabeled images, which are easy to collect without the burden of human annotations. Our SimLTD framework is straightforward and intuitive, and consists of three simple steps: (1) pre-training on abundant head classes; (2) transfer learning on scarce tail classes; and (3) fine-tuning on a sampled set of both head and tail classes. Our approach can be viewed as an improved head-to-tail model transfer paradigm without the added complexities of meta-learning or knowledge distillation, as was required in past research. By harnessing supplementary unlabeled images, without extra image labels, SimLTD establishes new record results on the challenging LVIS v1 benchmark across both supervised and semi-supervised settings.

Reconstructing dynamic radiance fields from video clips is challenging, especially when entertainment videos like TV shows are given. Many challenges make the reconstruction difficult due to (1) actors occluding with each other and having diverse facial expressions, (2) cluttered stages, and (3) small baseline views or sudden shot changes. To address these issues, we present ShowMak3r, a comprehensive reconstruction pipeline that allows the editing of scenes like how video clips are made in a production control room. In ShowMak3r, a 3DLocator module locates recovered actors on the stage using depth prior, and estimates unseen human poses via interpolation. The proposed ShotMatcher module then tracks the actors under shot changes. Furthermore, ShowMak3r introduces a face-fitting network that dynamically recovers the actors' expressions. Experiments on Sitcoms3D dataset show that our pipeline can reassemble TV show scenes with new cameras at different timestamps. We also demonstrate that ShowMak3r enables interesting applications such as synthetic shot-making, actor relocation, insertion, deletion, and pose manipulation. Project page : https://nstar1125.github.io/showmak3r

Medical image segmentation remains challenging due to the vast diversity of anatomical structures, imaging modalities, and segmentation tasks. While deep learning has made significant advances, current approaches struggle to generalize as they require task-specific training or fine-tuning on unseen classes. We present Iris, a novel In-context Reference Image guided Segmentation framework that enables flexible adaptation to novel tasks through the use of reference examples without fine-tuning. At its core, Iris features a lightweight context task encoding module that distills task-specific information from reference context image-label pairs. This rich context embedding information is used to guide the segmentation of target objects. Given a decoupled architecture on 3D data processing, Iris supports diverse inference strategies including one-shot inference, context example ensemble, object-level context example retrieval, and in-context tuning. Through comprehensive evaluation across twelve datasets, we demonstrate that Iris performs strongly compared to specialized supervised models on in-distribution tasks. On seven held-out dataset, Iris shows superior generalization to out-of-distribution data and unseen classes. Further, Iris's task encoding module can automatically discover anatomical relationships across datasets and modalities, offering insights into cross-modality medical objects without explicit anatomical supervision.

Recent methods have made significant progress in optical flow estimation. However, the evaluation of these methods focuses mainly on improved accuracy in benchmarks and often overlooks the analysis of network robustness, which may be important in safety-critical scenarios such as autonomous driving. In this paper, we propose a novel method for robustness evaluation by modifying data from original benchmarks. Unlike previous benchmarks that focus on complex scenes, we propose to modify shape and texture of objects from the original images in order to analyze the sensitivity to these changes observed in the output. Our aim is to identify common failure cases of state-of-the-art (SOTA) methods to evaluate their robustness and understand their behaviors. We show that: Optical flow methods are more sensitive to shape changes than to texture changes; and optical flow methods tend to "remember" objects seen during training and may "ignore" the motion of unseen objects. Our experimental results and findings provide a more in-depth understanding of the behavior of recent optical flow methods.

Accurate depth estimation from monocular videos remains challenging due to ambiguities inherent in single-view geometry, as crucial depth cues like stereopsis are absent. However, humans often perceive relative depth intuitively by observing variations in the size and spacing of objects as they move. Inspired by this, we propose a novel method that infers relative depth by examining the spatial relationships and temporal evolution of a set of tracked 2D trajectories. Specifically, we use off-the-shelf point tracking models to capture 2D trajectories. Then, our approach employs spatial and temporal transformers to process these trajectories and directly infer depth changes over time. Evaluated on the TAPVid-3D benchmark, our method demonstrates robust zero-shot performance, generalizing effectively from synthetic to real-world datasets. Results indicate that our approach achieves temporally smooth, high-accuracy depth predictions across diverse domains.

Knowledge-based visual question answering (KBVQA) separates image interpretation and knowledge retrieval into separate processes, motivated in part by the fact that they are very different tasks. In this paper, we transform the KBVQA into linguistic question-answering tasks so that we can leverage the rich world knowledge and strong reasoning abilities of Large Language Models (LLMs). The caption-then-question approach to KBVQA has been effective but relies on the captioning method to describe the detail required to answer every possible question. We propose instead a Question-Aware Captioner (QACap), which uses the question as guidance to extract correlated visual information from the image and generate a question-related caption. To train such a model, we utilize GPT-4 to build a corresponding high-quality question-aware caption dataset on top of existing KBVQA datasets. Extensive experiments demonstrate that our QACap model and dataset significantly improve KBVQA performance. Our method, QACap, achieves 68.2% accuracy on the OKVQA validation set, 73.4% on the direct-answer part of the A-OKVQA validation set, and 74.8% on the multiple-choice part, all setting new SOTA benchmarks.

Most previous scene text spotting methods rely on high-quality manual annotations to achieve promising performance. To reduce their expensive costs, we study semi-supervised text spotting (SSTS) to exploit useful information from unlabeled images. However, directly applying existing semi-supervised methods of general scenes to SSTS will face new challenges: 1) inconsistent pseudo labels between detection and recognition tasks, and 2) sub-optimal supervisions caused by inconsistency between teacher/student. Thus, we propose a new Semi-supervised framework for End-to-end Text Spotting, namely SemiETS that leverages the complementarity of text detection and recognition. Specifically, it gradually generates reliable hierarchical pseudo labels for each task, thereby reducing noisy labels. Meanwhile, it extracts important information in locations and transcriptions from bidirectional flows to improve consistency. Extensive experiments on three datasets under various settings demonstrate the effectiveness of SemiETS on arbitrary-shaped text. For example, it outperforms previous state-of-the-art SSL methods by a large margin on end-to-end spotting (+8.7%, +5.6%, and +2.6% H-mean under 0.5%, 1%, and 2% labeled data settings on Total-Text, respectively). More importantly, it still improves upon a strongly supervised text spotter trained with plenty of labeled data by 2.0%. Compelling domain adaptation ability shows practical potential. Moreover, our method demonstrates consistent improvement on different text spotters. Moreover, our method demonstrates consistent improvement on different text spotters. Code will be available at https://github.com/DrLuo/SemiETS.

Most vision-language models today are primarily trained on English image-text pairs, with non-English pairs often filtered out. Evidence from cross-cultural psychology suggests that this approach will bias models against perceptual modes exhibited by people who speak other (non-English) languages. We investigate semantic and expressive variation in image captions across different languages; we analyze both human-annotated datasets and model-produced captions. By analyzing captions across seven languages (English, French, German, Russian, Chinese, Japanese, Korean) in high-quality image captioning datasets (Crossmodal and Visual Genome), we find that multilingual caption sets tend to provide richer visual descriptions than monolingual (including English-only) ones; multilingual sets contain 46.0% more objects, 66.1% more relationships, and 66.8% more attributes. We observe the same results with multilingual captions produced by LLaVA and the Google Vertex API: for example, compared to monolingual captions, they cover 21.9% more objects,18.8% more relations, and 20.1% more attributes. These suggest that, across a large number of samples, different languages bias people and models to focus on different visual concepts. Finally, we show that models trained on image-text data in one language perform distinctly better on that language's test set. Our work points towards the potential value of training vision models on multilingual data sources to widen the range/variation of descriptive information those models are exposed to.

We propose SelfSplat, a novel 3D Gaussian Splatting model designed to perform pose-free and 3D prior-free generalizable 3D reconstruction from unposed multi-view images. These settings are inherently ill-posed due to the lack of ground-truth data, learned geometric information, and the need to achieve accurate 3D reconstruction without finetuning, making it difficult for conventional methods to achieve high-quality results. Our model addresses these challenges by effectively integrating explicit 3D representations with self-supervised depth and pose estimation techniques, resulting in reciprocal improvements in both pose accuracy and 3D reconstruction quality. Furthermore, we incorporate a matching-aware pose estimation network and a depth refinement module to enhance geometry consistency across views, ensuring more accurate and stable 3D reconstructions. To present the performance of our method, we evaluated it on large-scale real-world datasets, including RealEstate10K, ACID, and DL3DV. SelfSplat achieves superior results over previous state-of-the-art methods in both appearance and geometry quality, also demonstrates strong cross-dataset generalization capabilities. Extensive ablation studies and analysis also validate the effectiveness of our proposed methods.

We present a method for finding cross-modal space-time correspondences. Given two images from different visual modalities, such as an RGB image and a depth map, our model identifies which pairs of pixels correspond to the same physical points in the scene. To solve this problem, we extend the contrastive random walk framework to simultaneously learn cycle-consistent feature representations for both cross-modal and intra-modal matching. The resulting model is simple and has no explicit photo-consistency assumptions. It can be trained entirely using unlabeled data, without the need for any spatially aligned multimodal image pairs. We evaluate our method on both geometric and semantic correspondence tasks. For geometric matching, we consider challenging tasks such as RGB-to-depth and RGB-to-thermal matching (and vice versa); for semantic matching, we evaluate on photo-sketch and cross-style image alignment. Our method achieves strong performance across all benchmarks.

Cone-beam computed tomography (CBCT) is a standard imaging modality in orofacial and dental practices, providing essential 3D volumetric imaging of anatomical structures, including jawbones, teeth, sinuses, and neurovascular canals. Accurately segmenting these structures is fundamental to numerous clinical applications, such as surgical planning and implant placement. However, manual segmentation of CBCT scans is time-intensive and requires expert input, creating a demand for automated solutions through deep learning. Effective development of such algorithms relies on access to large, well-annotated datasets, yet current datasets are often privately stored or limited in scope and considered structures, especially concerning 3D annotations. This paper proposes ToothFairy2, a comprehensive, publicly accessible CBCT dataset with voxel-level 3D annotations of 42 distinct classes corresponding to maxillofacial structures. We validate the dataset by benchmarking state-of-the-art neural network models, including convolutional, transformer-based, and hybrid Mamba-based architectures, to evaluate segmentation performance across complex anatomical regions. Our work also explores adaptations to the nnU-Net framework to optimize multi-class segmentation for maxillofacial anatomy. The proposed dataset provides a fundamental resource for advancing maxillofacial segmentation and supports future research in automated 3D image analysis in digital dentistry.