CVPR 2025 Papers

Hybrid Mamba-Transformer networks have recently garnered broad attention. These networks can leverage the scalability of Transformers while capitalizing on Mamba's strengths in long-context modeling and computational efficiency. However, the challenge of effectively pretraining such hybrid networks remains an open question. Existing methods, such as Masked Autoencoders (MAE) or autoregressive (AR) pretraining, primarily focus on single-type network architectures. In contrast, pretraining strategies for hybrid architectures must be effective for both Mamba and Transformer components. Based on this, we propose Masked Autoregressive Pretraining (MAP) to pretrain a hybrid Mamba-Transformer vision backbone network. This strategy combines the strengths of both MAE and Autoregressive pretraining, improving the performance of Mamba and Transformer modules within a unified paradigm. Experimental results show that the hybrid Mamba-Transformer vision backbone network pretrained with MAP significantly outperforms other pretraining strategies, achieving state-of-the-art performance. We validate the method's effectiveness on both 2D and 3D datasets and provide detailed ablation studies to support the design choices for each component.

Long-term dense action anticipation is very challenging since it requires predicting actions and their durations several minutes into the future based on provided video observations. To model the uncertainty of future outcomes, stochastic models predict several potential future action sequences for the same observation. Recent work has further proposed to incorporate uncertainty modelling for observed frames by simultaneously predicting per-frame past and future actions in a unified manner. While such joint modelling of actions is beneficial, it requires long-range temporal capabilities to connect events across distant past and future time points. However, the previous work struggles to achieve such a long-range understanding due to its limited and/or sparse receptive field. To alleviate this issue, we propose a novel MANTA (MAmba for ANTicipation) network. Our model enables effective long-term temporal modelling even for very long sequences while maintaining linear complexity in sequence length. We demonstrate that our approach achieves state-of-the-art results on three datasets - Breakfast, 50Salads, and Assembly101 - while also significantly improving computational and memory efficiency. Our code is available at https://github.com/olga-zats/DIFF_MANTA.

We present MANTA, a visual-text anomaly detection dataset for tiny objects. The visual component comprises over 137.3K images across 38 object categories spanning five typical domains, of which 8.6K images are labeled as anomalous with pixel-level annotations. Each image is captured from five distinct viewpoints to ensure comprehensive object coverage. The text component consists of two subsets: Declarative Knowledge, including 875 words that describe common anomalies across various domains and specific categories, with detailed explanations for \left< what, why, how \right>, including causes and visual characteristics; and Constructivist Learning, providing 2K multiple-choice questions with varying levels of difficulty, each paired with images and corresponded answer explanations. We also propose a baseline for visual-text tasks and conduct extensive benchmarking experiments to evaluate advanced methods across different settings, highlighting the challenges and efficacy of our dataset.

Simultaneous localization and mapping (SLAM) systems with novel view synthesis capabilities are widely used in computer vision, with applications in augmented reality, robotics, and autonomous driving. However, existing approaches are limited to single-agent operation. Recent work has addressed this problem using a distributed neural scene representation. Unfortunately, existing methods are slow, cannot accurately render real-world data, are restricted to two agents, and have limited tracking accuracy. In contrast, we propose a rigidly deformable 3D Gaussian-based scene representation that dramatically speeds up the system. However, improving tracking accuracy and reconstructing a globally consistent map from multiple agents remains challenging due to trajectory drift and discrepancies across agents' observations. Therefore, we propose new tracking and map-merging mechanisms and integrate loop closure in the Gaussian-based SLAM pipeline. We evaluate \ours on synthetic and real-world datasets and find it more accurate and faster than the state of the art.

With advances in deep learning models and the availability of large-scale 3D datasets, we have recently witnessed significant progress in single-view 3D reconstruction. However, existing methods often fail to reconstruct physically based material properties given a single image, limiting their applicability in complicated scenarios. This paper presents a novel approach (named MAGE) for generating 3D geometry with realistic decomposed material properties given a single image as input. Our method leverages inspiration from traditional computer graphics deferred rendering pipelines to introduce a multi-view G-buffer estimation model. The proposed model estimates G-buffers for various views as multi-domain images, including XYZ coordinates, normals, albedo, roughness, and metallic properties from a single-view RGB image. To address the inherent ambiguity and inconsistency in generating G-buffers simultaneously, we also formulate a deterministic network from the pretrained diffusion models and propose a lighting response loss that enforces consistency across these domains using PBR principles. Finally, we propose a large-scale synthetic dataset rich in material diversity for our model training. Experimental results demonstrate the effectiveness of our method in producing high-quality 3D meshes with rich material properties. Our code and dataset can be found at https://www.whyy.site/paper/mage.

To perceive, humans use memory to fill in gaps caused by our limited visibility, whether due to occlusion or our narrow field of view. However, most 3D object detectors are limited to using sensor evidence from a short temporal window (0.1s-0.3s). In this work, we present a simple and effective add-on for enhancing any existing 3D object detector with long-term memory regardless of its sensor modality (e.g., LiDAR, camera) and network architecture. We propose a model to effectively align and fuse object proposals from a detector with object proposals from a memory bank of past predictions, exploiting trajectory forecasts to align proposals across time. We propose a novel schedule to train our model on temporal data that balances data diversity and the gap between training and inference. By applying our method to existing LiDAR and camera-based detectors on the Waymo Open Dataset (WOD) and Argoverse 2 Sensor (AV2) dataset, we demonstrate significant improvements in detection performance (+2.5 to +7.6 AP points). Our method attains the best performance on the WOD 3D detection leaderboard among online methods (excluding ensembles or test-time augmentation).

Real-time multi-agent collaboration for ego-motion estimation and high-fidelity 3D reconstruction is vital for scalable spatial intelligence. However, traditional methods produce sparse, low-detail maps, while recent dense mapping approaches struggle with high latency.To overcome these challenges, we present MAC-Ego3D, a novel framework for real-time collaborative photorealistic 3D reconstruction via Multi-Agent Gaussian Consensus. MAC-Ego3D enables agents to independently construct, align, and iteratively refine local maps using a unified Gaussian splat representation. Through Intra-Agent Gaussian Consensus, it enforces spatial coherence among neighboring Gaussian splats within an agent. For global alignment, parallelized Inter-Agent Gaussian Consensus, which asynchronously aligns and optimizes local maps by regularizing multi-agent Gaussian splats, seamlessly integrates them into a high-fidelity 3D model. Leveraging Gaussian primitives, MAC-Ego3D supports efficient RGB-D rendering, enabling rapid inter-agent Gaussian association and alignment.MAC-Ego3D bridges local precision and global coherence, delivering higher efficiency, largely reducing localization error, and improving mapping fidelity. It establishes a new SOTA on synthetic and real-world benchmarks, achieving a 15xincrease in inference speed, order-of-magnitude reductions in ego-motion estimation error for partial cases, and RGB PSNR gains of 4 to 10 dB. Our code will be made publicly available.

Multi-instance learning (MIL) has demonstrated impressive performance in whole slide image (WSI) analysis. However, existing approaches struggle with undesirable results and unbearable computational overhead due to the quadratic complexity of Transformers. Recently, Mamba has offered a feasible solution for modeling long-range dependencies with linear complexity. However, vanilla Mamba inherently suffers from contextual forgetting issues, making it ill-suited for capturing global dependencies across instances in large-scale WSIs. To address this, we propose a memory-driven Mamba network, dubbed M3amba, to fully explore the global latent relations among instances. Specifically, M3amba retains and iteratively updates historical information with a dynamic memory bank (DMB), thus overcoming the catastrophic forgetting defects of Mamba for long-term context representation. For better feature representation, M3amba involves an intra-group bidirectional Mamba (BiMamba) block to refine local interactions within groups. Meanwhile, we additionally perform cross-attention fusion to incorporate relevant historical information across groups, facilitating richer inter-group connections. The joint learning of inter- and intra-group representations with memory merits enables M3amba with a more powerful capability for achieving accurate and comprehensive WSI representation. Extensive experiments on four datasets demonstrate that M3amba outperforms the state-of-the-art by 6.2% and 7.0% in accuracy on the TCGA BRAC and TCGA Lung datasets while maintaining low computational costs.

Human pose estimation is a critical task in computer vision for applications in sports analysis, healthcare monitoring, and human-computer interaction. However, existing human pose datasets are collected either from custom-configured laboratories with complex devices or they only include data on single individuals, and both types typically capture daily activities. In this paper, we introduce the M3GYM dataset, a large-scale multimodal, multi-view, and multi-person pose dataset collected from a real gym to address the limitations of existing datasets.Specifically, we collect videos for 82 sessions from the gym, each session lasting between 40 to 60 minutes. These videos are gathered by 8 cameras, including over 50 subjects and 47 million frames. These sessions include 51 Normal fitness exercise sessions as well as 17 Pilates and 14 Yoga sessions. The exercises cover a wide range of poses and typical fitness activities, particularly in Yoga and Pilates, featuring poses with stretches, bends, and twists, e.g., humble warrior, fire hydrants and knee hover side twists. Each session involves multiple subjects, leading to significant self-occlusion and mutual occlusion in single views.Moreover, the gym has two symmetric floor mirrors, a feature not seen in previous datasets, and seven lighting conditions. We provide frame-level multimodal annotations, including 2D&3D keypoints, subject IDs, and meshes. Additionally, M3GYM uniquely offers labels for over 500 actions along with corresponding assessments from sports experts.We benchmark a variety of state-of-the-art methods for several tasks, i.e., 2D human pose estimation, single-view and multi-view 3D human pose estimation, and human mesh recovery. To simulate real-world applications, we also conduct cross-domain experiments across Normal, Yoga, and Pilates sessions. The results show that M3GYM significantly improves model generalization in complex real-world settings.

Recent advances in Multi-Modal Large Language Models (M-LLMs) show promising results in video reasoning. Popular Multi-Modal Large Language Model (M-LLM) frameworks usually apply naive uniform sampling to reduce the number of video frames that are fed into an M-LLM, particularly for long context videos. However, it could lose crucial context in certain periods of a video, so that the downstream M-LLM may not have sufficient visual information to answer a question. To attack this pain point, we propose a light-weight M-LLM -based frame selection method that adaptively select frames that are more relevant to users' queries. In order to train the proposed frame selector, we introduce two supervision signals (i) Spatial signal, where single frame importance score by prompting an M-LLM; (ii) Temporal signal, in which multiple frames selection by prompting Large Language Model ( LLM ) using the captions of all frame candidates. The selected frames are then digested by a frozen downstream video M-LLM for visual reasoning and question answering. Empirical results show that the proposed M-LLM video frame selector improves the performances various downstream video Large Language Model (video-LLM) across medium (ActivityNet, NExT-QA) and long (EgoSchema, LongVideoBench) context video question answering benchmarks.

Video portrait relighting remains challenging because the results need to be both photorealistic and temporally stable.This typically requires a strong model design that can capture complex facial reflections as well as intensive training on a high-quality paired video dataset, such as dynamic one-light-at-a-time (OLAT). In this work, we introduce Lux Post Facto, a novel portrait video relighting method that produces both photorealistic and temporally consistent lighting effects. From the model side, we design a new conditional video diffusion model built upon state-of-the-art pre-trained video diffusion model, alongside a new lighting injection mechanism to enable precise control. This way we leverage strong spatial and temporal generative capability to generate plausible solutions to the ill-posed relighting problem. Our technique uses a hybrid dataset consisting of static expression OLAT data and in-the-wild portrait performance videos to jointly learn relighting and temporal modeling. This avoids the need to acquire paired video data in different lighting conditions. Our extensive experiments show that our model produces state-of-the-art results both in terms of photorealism and temporal consistency.

Capturing high-quality photographs under diverse real-world lighting conditions is challenging, as both natural lighting (e.g., low-light) and camera exposure settings (e.g., exposure time) significantly impact image quality. This challenge becomes more pronounced in multi-view scenarios, where variations in lighting and image signal processor (ISP) settings across viewpoints introduce photometric inconsistencies. Such lighting degradations and view-dependent variations pose substantial challenges to novel view synthesis (NVS) frameworks based on Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS). To address this, we introduce Luminance-GS, a novel approach to achieving high-quality novel view synthesis results under diverse challenging lighting conditions using 3DGS. By adopting per-view color matrix mapping and view adaptive curve adjustments, Luminance-GS achieves state-of-the-art (SOTA) results across various lighting conditions--including low-light, overexposure, and varying exposure--while not altering the original 3DGS explicit representation. Compared to previous NeRF- and 3DGS-based baselines, Luminance-GS provides real-time rendering speed with improved reconstruction quality. The source code is available at https://github.com/cuiziteng/Luminance-GS.

We introduce LumiNet, a novel architecture that leverages generative models and latent intrinsic representations for transferring lighting from one image to another. Given a source image and a target lighting image, LumiNet generates a relit version of the source scene that captures the target's lighting. Our approach makes two key contributions: a data curation strategy from the StyleGAN-based relighting model for our training, and a modified diffusion-based ControlNet that processes both latent intrinsic properties from the source image and latent extrinsic properties from the target image. We further improve lighting transfer through a learned adaptor that injects the target's latent extrinsic properties via cross-attention and light-weight fine-tuning. Unlike traditional ControlNet, which generates images with conditional maps from a single scene, LumiNet processes latent representations from two different images - preserving geometry and albedo from the source while transferring lighting characteristics from the target. Experiments demonstrate that our method successfully transfers complex lighting phenomena including specular highlights and indirect illumination across scenes with varying spatial layouts and materials, outperforming existing approaches on challenging indoor scenes using only images as input.

In security-sensitive fields, data should be encrypted to protect against unauthorized access and maintain confidentiality throughout processing. However, traditional networks like ViTs and CNNs return different results when processing original data versus its encrypted form, meaning that they require data to be decrypted, posing a security risk by exposing sensitive information. One solution for this issue is using polynomial networks, including state-of-the-art Multilinear Operator Networks, which return the same outputs given the real data and their encrypted forms under Leveled Fully Homomorphic Encryption. Nevertheless, these models are susceptible to catastrophic forgetting in incremental learning settings. Thus, this paper will present a new low-rank adaptation method combined with the Gradient Projection Memory mechanism to minimize the issue. Our proposal is compatible with Leveled Fully Homomorphic Encryption while achieving a sharp improvement in performance compared to existing models.

Pre-training backbone networks on a general annotated dataset (e.g., ImageNet) that comprises numerous manually collected images with category annotations has proven to be indispensable for enhancing the generalization capacity of downstream visual tasks. However, those manually collected images often exhibit bias, which is non-transferable across either categories or domains, thus causing the model's generalization capacity degeneration. To mitigate this problem, we present a low-biased general annotated dataset generation framework (lbGen). Instead of expensive manual collection, we aim at directly generating low-biased images with category annotations. To achieve this goal, we propose to leverage the advantage of a multimodal foundation model (e.g., CLIP), in terms of aligning images in a low-biased semantic space defined by language. Specifically, we develop a bi-level semantic alignment loss, which not only forces all generated images to be consistent with the semantic distribution of all categories belonging to the target dataset in an adversarial learning manner, but also requires each generated image to match the semantic description of its category name. In addition, we further cast an existing image quality scoring model into a quality assurance loss to preserve the quality of the generated image. By leveraging these two loss functions, we can obtain a low-biased image generation model by simply fine-tuning a pre-trained diffusion model using only all category names in the target dataset as input. Experimental results confirm that, compared with the manually labeled dataset or other synthetic datasets, the utilization of our generated low-biased dataset leads to stable generalization capacity enhancement of different backbone networks across various tasks, especially in tasks where the manually labeled samples are scarce.

Approximate nearest neighbor search (ANNS) is an essential building block for applications like RAG but can sometimes yield results that are overly similar to each other. In certain scenarios, search results should be similar to the query and yet diverse. We propose LotusFilter, a post-processing module to diversify ANNS results. We precompute a cutoff table summarizing vectors that are close to each other. During the filtering, LotusFilter greedily looks up the table to delete redundant vectors from the candidates. We demonstrated that the LotusFilter operates fast (0.02 [ms/query]) in settings resembling real-world RAG applications, utilizing features such as OpenAI embeddings. Our code is publicly available at https://github.com/matsui528/lotf.

Our objective is to translate continuous sign language into spoken language text. Inspired by the way human interpreters rely on context for accurate translation, we incorporate additional contextual cues together with the signing video, into a new translation framework. Specifically, besides visual sign recognition features that encode the input video, we integrate complementary textual information from (i) captions describing the background show, (ii) translation of previous sentences, as well as (iii) pseudo-glosses transcribing the signing. These are automatically extracted and inputted along with the visual features to a pre-trained large language model (LLM), which we fine-tune to generate spoken language translations in text form. Through extensive ablation studies, we show the positive contribution of each input cue to the translation performance. We train and evaluate our approach on BOBSL -- the largest British Sign Language dataset currently available. We show that our contextual approach significantly enhances the quality of the translations compared to previously reported results on BOBSL, and also to state-of-the-art methods that we implement as baselines. Furthermore, we demonstrate the generality of our approach by applying it also to How2Sign, an American Sign Language dataset, and achieve competitive results.

Anamorphosis refers to a category of images that are intentionally distorted, making them unrecognizable when viewed directly. Their true form only reveals itself when seen from a specific viewpoint, which can be through some catadioptric device like a mirror or a lens. While the construction of these mathematical devices can be traced back to as early as the 17th century, they are only interpretable when viewed from a specific vantage point and tend to lose meaning when seen normally. In this paper, we revisit these famous optical illusions with a generative twist. With the help of latent rectified flow models, we propose a method to create anamorphic images that still retain a valid interpretation when viewed directly. To this end, we introduce Laplacian Pyramid Warping, a frequency-aware image warping technique key to generating high-quality visuals. Our work extends Visual Anagrams (Geng et al. 2024) to latent space models and to a wider range of spatial transforms, enabling the creation of novel generative perceptual illusions.

Humans perceive and comprehend their surroundings through information spanning multiple frequencies. In immersive scenes, people naturally scan their environment to grasp its overall structure while examining fine details of objects that capture their attention. However, current NeRF frameworks primarily focus on modeling either high-frequency local views or the broad structure of scenes with low-frequency information, limited to balance both. We introduce FA-NeRF, a novel frequency-aware framework for view synthesis that simultaneously captures the overall scene structure and high-definition details within a single NeRF model. To achieve this, we propose a 3D frequency quantification method that analyzes the scene's frequency distribution, enabling frequency-aware rendering. Our framework incorporates a frequency grid for fast convergence and querying, a frequency-aware feature re-weighting strategy to balance features across different frequency contents. Extensive experiments show that our method significantly outperforms existing approaches in modeling entire scenes while preserving fine details.