CVPR 2026 论文列表

Multi-camera 3D object detection (MC3D) has attracted increasing attention with the growing deployment of multi-sensor physical agents, such as robots and autonomous vehicles. However, MC3D models still struggle to generalize to unseen platforms with new multi-camera configurations. Current solutions simply employ a meta-camera for unified representation but lack comprehensive consideration. In this paper, we revisit this issue and identify that the devil lies in spatial prior discrepancies across source and target configurations, including different intrinsics, extrinsics, and array layouts. To address this, we propose CoIn3D, a generalizable MC3D framework that enables strong transferability from source configurations to unseen target ones. CoIn3D explicitly incorporates all identified spatial priors into both feature embedding and image observation through spatial-aware feature modulation (SFM) and camera-aware data augmentation (CDA), respectively. SFM enriches feature space by integrating four spatial representations, such as focal length, ground depth, ground gradient, and Plucker coordinate. CDA improves observation diversity under various configurations via a training-free dynamic novel-view image synthesis scheme. Extensive experiments demonstrate that CoIn3D achieves strong cross-configuration performance on landmark datasets such as NuScenes, Waymo, and Lyft, under three dominant MC3D paradigms represented by BEVDepth, BEVFormer, and PETR.

This paper addresses the challenge of training a single network to jointly perform multiple dense prediction tasks, such as segmentation and depth estimation, i.e., multi-task learning (MTL). Current approaches mainly capture cross-task relations in the 2D image space, often leading to unstructured features lacking 3D-awareness. We argue that 3D-awareness is vital for modeling cross-task correlations essential for comprehensive scene understanding. We propose to address this problem by integrating correlations across views, i.e., cost volume, as geometric consistency in the MTL network. Specifically, we introduce a lightweight Cross-view Module (CvM), shared across tasks, to exchange information across views and capture cross-view correlations, which are integrated with MTL encoder features for multi-task prediction. This module is architecture-agnostic and can be applied to both single- and multi-view data. Extensive results on NYUv2 and PASCAL-Context demonstrate that our method effectively injects geometric consistency into existing MTL methods to improve performance.

Out-of-distribution (OOD) detection aims to identify samples that deviate from in-distribution (ID). One popular pipeline addresses this by introducing negative labels distant from ID classes and detecting OOD based on their distance to these labels.However, such labels may present poor activation on OOD samples, failing to capture the OOD characteristics. To address this, we propose \underline T est-time \underline A ctivated \underline N egative \underline L abels (TANL) by dynamically evaluating activation levels across the corpus dataset and mining candidate labels with high activation responses in the testing process. Specifically, TANL identifies high-confidence test images online and accumulates their assignment probabilities over the corpus to construct a label activation metric.Such a metric leverages historical test samples to adaptively align with the test distribution, enabling the selection of distribution-adaptive activated negative labels. By further exploring the activation information within the current testing batch, we introduce a more fine-grained, batch-adaptive variant. To fully utilize label activation knowledge, we propose an activation-aware score function that emphasizes negative labels with stronger activations, boosting performance and enhancing its robustness to the label number.Our TANL is training-free, test-efficient, and grounded in theoretical justification. Experiments on diverse backbones and wide task settings validate its effectiveness.Notably, on the large-scale ImageNet benchmark, TANL significantly reduces the FPR95 from 17.5% to 9.8%.Codes will be released.

Although 4D reconstruction based on Gaussian Splatting has achieved many impressive results, reconstructing real-world images captured by a casual monocular camera remains a significant challenge. In dynamic scenes, as the camera and objects move during the exposure time, these input images inevitably contain a considerable amount of motion blur, which severely compromises the quality of reconstruction and new viewpoint synthesis. The existing deblurring 3D Gaussian models still cannot handle motion blur issues in real dynamic scenes. To address these challenges, we propose MSCD-GS--a novel method for motion-separated collaborative deblurring 4D reconstruction via Gaussian Splatting, capable of effectively handling motion-blurred inputs. Specifically, due to the distinct motion characteristics of static and dynamic Gaussians, we perform separate motion models for dynamic scene reconstruction. To predict Gaussian changes over exposure time, we designed motion-aware networks for static and dynamic Gaussians, thereby synthesizing virtual blurred images. Finally, we utilize the results from the deblurring network and the synthesized images to supervise 4D reconstruction collaboratively. Extensive experiments demonstrate that MSCD-GS can effectively reconstruct high-quality dynamic scenes from blurred image inputs, with performance surpassing existing methods.

Causal graphs play a crucial role in AI research as they reveal the data generation processes underlying real-world machine learning and computer vision tasks. Recent studies have leveraged causal graphs to develop more robust and interpretable models. However, limited or biased data often lead to inaccurate causal graph estimation, reducing a model's transferability to unseen domains. To address this challenge, we propose a novel framework that performs Bayesian inference over causal graphs to capture potential underlying causal relations and identify invariant causal features for DG prediction. The key advantage of our framework lies in its ability to quantify causal graph uncertainty in the context of prediction tasks and incorporate it into the prediction process. Our proposed uncertainty provides valuable insights into (i) the reliability of our method on specific datasets, (ii) the alignment between learned causal graphs and unseen test domains, and (iii) the confidence of our predictions. In particular, we go beyond merely quantifying uncertainty and leverage it as weighting factors in a weighted Bayesian inference scheme. Empirical results on multiple benchmark distribution-shift datasets show that our algorithm, **C**ausal **G**raph **U**ncertainty-guided **Bayes**ian Inference (CGU-Bayes), outperforms existing DG methods on challenging datasets and achieves state-of-the-art performance overall.

Unsupervised domain adaptation (UDA) for pose estimation promises transfer from synthetic to real domains but often suffers instability under domain shift. Prior work attributes this deterioration to gradient interference between source supervision and target consistency. This conflict is distinct in pose estimation, where sparse and heterogeneous supervision signals cause gradients to be highly sensitive to small localization errors and lead to unstable updates. To address these challenges, we propose HamiPose, a Hamiltonian optimization framework that transports decoupled and confidence-calibrated gradients within a unified geometry to mitigate instability. HamiPose first refines gradient interaction through keypointwise geometry decomposition, orthogonally projecting target gradients to preserve nonconflicting component. Channelwise gated alignment then calibrates the parallel component with confidence and alignment, producing decoupled, confidence-calibrated gradients. These gradients are advanced by a Hamiltonian optimizer with a symplectic integrator, providing controlled momentum that stabilizes updates. Extensive experiments demonstrate that HamiPose achieves state-of-the-art performance in UDA pose estimation while maintains strong performance under domain generalization settings.

Vision-Language Models (VLMs) have achieved remarkable progress in multimodal understanding, yet their positional encoding mechanisms remain suboptimal. Existing approaches uniformly assign positional indices to all tokens, overlooking variations in information density within and across modalities, which leads to inefficient attention allocation where redundant visual regions dominate while informative content is underrepresented. We identify positional granularity as an implicit resource and propose MODIX (Multimodal Information-Driven Positional IndeX Scaling), a training-free framework that dynamically adapts positional strides based on modality-specific contributions. MODIX jointly models intra-modal density via covariance-based entropy and inter-modal interaction via cross-modal alignment to derive unified scores, which rescale positional indices to allocate finer granularity to informative modalities while compressing redundant ones, without requiring any modification to model parameters or architecture. Experiments across diverse architectures and benchmarks demonstrate that MODIX consistently improves multimodal reasoning and adaptively reallocates attention according to task-dependent information distributions, suggesting that positional encoding should be treated as an adaptive resource in Transformers for multimodal sequence modeling.

Vision-language models (VLMs) like CLIP have shown impressive generalization capabilities, yet their potential for Cross-Domain Few-Shot Learning (CDFSL) remains underexplored, where the model needs to transfer source-domain information to target domains with scarce training data. While the attention sink phenomenon has been observed in VLMs for certain tasks, its role in CDFSL scenarios has not been studied. In this paper, we uncover a critical issue overlooked by prior works: standard target-domain few-shot fine-tuning in CDFSL significantly exacerbates the attention sink problem, leading to poor discriminability across classes. To understand this phenomenon, through extensive experiments, we interpret it as the model's shortcut learning for domain adaptation: to overcome the huge domain gap between the source and target domains, the model shows a high tendency to push tokens that are initially closer to target-domain classes (i.e., simple tokens) to be even closer to these classes, exacerbating the attention sink and wasting the capability of learning other discriminative but initially further tokens (i.e., hard tokens). To address this, we propose a novel approach to dynamically re-weight tokens according to their relevance with target-domain classes during the target-domain finetuning, which explicitly suppresses the model's reliance on these simple tokens and enhances the learning of hard tokens, reducing sink tokens and enhancing discriminability. Extensive experiments on four benchmark datasets validate the rationale of our method, demonstrating new state-of-the-art performance. Our codes are available at https://github.com/shuaiyi308/TIR.

Autoregressive (AR) models have demonstrated remarkable performance in generating high-fidelity images. However, their inherently sequential next-token prediction leads to significantly slows inference. Recent studies have introduced Jacobi-style decoding to accelerate autoregressive image generation. Extending the draft sequence initially improves efficiency, yet the acceleration quickly saturates as error propagation in the one-dimensional sequence hinders convergence.Observing that images exhibit strong local spatial correlations, we propose Parallel Jacobi Decoding (PJD), a training-free decoding approach that expands draft tokens in the two-dimensional spatial domain to enable efficient spatially parallel refinement.PJD adjusts the attention mask to mitigate error accumulation and improve convergence stability. Extensive experiments on diverse datasets show that PJD achieves **4.8x-6.4x** acceleration across multiple autoregressive image generation models while maintaining competitive generation quality.

Video Diffusion Transformers (DiTs) have been synthesizing high-quality video with high fidelity from given text descriptions involving motion. However, understanding how Video DiTs convert motion words into video remains insufficient. Furthermore, while prior studies on interpretable saliency maps primarily target objects, motion-related behavior in Video DiTs remains largely unexplored. In this paper, we investigate concrete motion features that specify when and which object moves for a given motion concept. First, to spatially localize, we introduce GramCol, which adaptively produces per-frame saliency maps for any text concept, including both motion and non-motion. Second, we propose a motion-feature selection algorithm to obtain an Interpretable Motion-Attentive Map (IMAP) that localizes motion spatially and temporally. Our method discovers concept saliency maps without the need for any gradient calculation or parameter update. Experimentally, our method shows outstanding localization capability on the motion localization task and zero-shot video semantic segmentation, providing interpretable and clearer saliency maps for both motion and non-motion concepts.

Human mesh recovery (HMR) models 3D human body from monocular videos, with recent works extending it to world-coordinate human trajectory and motion reconstruction. However, most existing methods remain offline, relying on future frames or global optimization, which limits their applicability in interactive feedback and perception-action loop scenarios such as AR/VR and telepresence. To address this, we propose OnlineHMR, a fully online framework that jointly satisfies four essential criteria of online processing, including system-level causality, faithfulness, temporal consistency, and efficiency. Built upon a two-branch architecture, OnlineHMR enables streaming inference via a causal key-value cache design and a curated sliding-window learning strategy. Meanwhile, a human-centric incremental SLAM provides online world-grounded alignment under physically plausible trajectory correction. Experimental results show that our method achieves performance comparable to existing chunk-based approaches on the standard EMDB benchmark and highly dynamic custom videos, while uniquely supporting online processing. Page and code are available at https://tsukasane.github.io/Video-OnlineHMR/.

Deep learning models for medical image analysis often act as "black boxes," seldom aligning with clinical guidelines or explicitly linking decisions to supporting evidence. This is especially critical in Alzheimer's disease (AD), where predictions should be grounded in both anatomical and clinical findings. We present EMAD, a vision-language framework that generates structured AD diagnostic reports with each claim explicitly grounded in multimodal evidence. EMAD uses a hierarchical Sentence-Evidence-Anatomy (SEA) Grounding mechanism: (i) sentence-to-evidence grounding links generated sentences to clinical evidence phrases, and (ii) evidence-to-anatomy grounding localizes corresponding structures on 3D brain MRI. To reduce dense annotation requirements, we propose GTX-Distill, which transfers grounding behavior from a teacher trained with limited supervision to a student operating on model-generated reports. We further introduce Executable-Rule GRPO, a reinforcement fine-tuning scheme with verifiable rewards that enforces clinical consistency, protocol adherence, and reasoning-diagnosis coherence. On the AD-MultiSense dataset, EMAD achieves state-of-the-art diagnostic accuracy and produces more transparent, anatomically faithful reports than existing methods; we will release code and grounding annotations to support future research in trustworthy medical vision-language models.

Controllable driving scene generation is critical for realistic and scalable autonomous driving simulation, yet existing approaches struggle to jointly achieve photorealism and precise control. We introduce HorizonForge, a unified framework that reconstructs scenes as editable Gaussian Splats and Meshes, enabling fine-grained 3D manipulation and language-driven vehicle insertion. Edits are rendered through a noise-aware video diffusion process that enforces spatial and temporal consistency, producing diverse scene variations in a single feed-forward pass without per-trajectory optimization. To standardize evaluation, we further propose HorizonSuite, a comprehensive benchmark spanning ego- and agent-level editing tasks such as trajectory modifications and object manipulation. Extensive experiments show that Gaussian Splatting delivers substantially higher fidelity than alternative 3D representations, and that temporal priors from video diffusion are essential for coherent synthesis. Combining these finding, HorizonSuite establishes a simple yet powerful paradigm for photorealistic, controllable driving simulation. achieving an 83.4% user-preference gain and a 25.19% FID improvement over the second best state-of-the-art method.

Generative models, with their success in image and video generation, have recently been explored for synthesizing effective neural network weights. These approaches take trained neural network checkpoints as training data, and aim to generate high-performing neural network weights during inference. In this work, we examine four representative, well-known methods in this emerging area on their ability to generate novel model weights, i.e., weights that are different from the checkpoints seen during training. Contrary to claims in prior work, we find that these methods synthesize weights largely by memorization: they produce either replicas, or at best simple interpolations, of the training checkpoints. Current methods fail to outperform simple baselines, such as adding noise to the weights or taking a simple weight ensemble, in obtaining different and simultaneously high-performing models. Our further results suggest that the memorization potentially resulted from limited data, overparameterized models, and the underuse of structural priors specific to weight data. Our findings highlight the need for more careful design and evaluation of generative models in new domains.

We present Motion 3-to-4, a feed-forward framework for synthesising high-quality 4D dynamic objects from a single monocular video and an optional 3D reference mesh. While recent advances have significantly improved 2D, video, and 3D content generation, 4D synthesis remains difficult due to limited training data and the inherent ambiguity of recovering geometry and motion from a monocular viewpoint. Motion 3-to-4 addresses these challenges by decomposing 4D synthesis into static 3D shape generation and motion reconstruction. Using a canonical reference mesh, our model learns a compact motion latent representation and predicts per-frame vertex trajectories to recover complete, temporally coherent geometry. A scalable frame-wise transformer further enables robustness to varying sequence lengths. Evaluations on both standard benchmarks and a new dataset with accurate ground-truth geometry show that Motion 3-to-4 delivers superior fidelity and spatial consistency compared to prior work. Project page is available at https://motion3-to-4.github.io/.

Video object detection has gained notable progress with the advent of transformers. While transformers excel at modeling long-range contextual dependencies, the quadratic complexity limits their efficiency in long-sequence processing. In contrast, Mamba offers greater efficiency in modeling long sequences but tends to exhibit relatively limited contextual learning capability compared with transformers, and its application to video object detection remains unexplored. To harness the complementary strengths of transformers and Mamba, we propose a hybrid Transformer-Mamba network for video object detection (TMambaDet), a pioneering framework in this domain that combines the long-range modeling power of transformers with the efficient long-sequence processing capability of Mamba. Our TMambaDet is characterized by three core components: 1) a spatial adaptive deformable transformer encoder to effectively model the long-range dependencies within each frame, enabling intra-frame feature aggregation that substantially improves the spatial feature representations of objects; 2) a temporal cascaded bidirectional Mamba encoder to efficiently capture the long-range dependencies across frames in video sequences with linear complexity, enabling inter-frame feature aggregation that effectively enhances the temporal feature representations of objects; 3) a Mamba entangled transformer decoder to fully explore the interactions between object queries and spatial-temporal features, enabling fine-grained query-feature alignment that effectively enriches the instance-level representations of object queries. We conduct experiments on the ImageNet VID and EPIC-KITCHENS-55 datasets, showing that TMambaDet achieves state-of-the-art results.

Material classification has emerged as a critical task in computer vision and graphics, supporting the assignment of accurate material properties to a wide range of digital and real-world applications. While traditionally framed as an image classification task, this domain faces significant challenges due to the scarcity of annotated data, limiting the accuracy and generalizability of trained models. Recent advances in vision-language foundation models (VLMs) offer promising avenues to address these issues, yet existing solutions leveraging these models still exhibit unsatisfying results in material recognition tasks. In this work, we propose a novel framework that effectively harnesses foundation models to overcome data limitations and enhance classification accuracy. Our method integrates two key innovations:(a) a robust image generation and auto-labeling pipeline that creates a diverse and high-quality training dataset with material-centric images, and automatically assigns labels by fusing object semantics and material attributes in text prompts; (b) a prior incorporation strategy to distill information from VLMs, combined with a joint fine-tuning method that optimizes a pre-trained vision foundation model alongside VLM-derived priors, preserving broad generalizability while adapting to material-specific features.Extensive experiments demonstrate significant improvements on multiple datasets. We show that our synthetic dataset effectively captures the characteristics of real world materials, and the integration of priors from vision-language models significantly enhances the final performance.

Reinforcement learning (RL) has become a powerful tool for post-training visual generative models, with Group Relative Policy Optimization (GRPO) increasingly used to align generators with human preferences. However, existing GRPO pipelines rely on a single scalar reward per sample, treating each image or video as a holistic entity and ignoring the rich spatial and temporal structure of visual content. This coarse supervision hinders the correction of localized artifacts and the modeling of fine-grained perceptual cues. We introduce Visual Preference Policy Optimization (ViPO), a GRPO variant that lifts scalar feedback into structured, pixel-level advantages. ViPO employs a Perceptual Structuring Module that uses pretrained vision backbones to construct spatially and temporally aware advantage maps, redistributing optimization pressure toward perceptually important regions while preserving the stability of standard GRPO. Across both image and video benchmarks, ViPO consistently outperforms vanilla GRPO, improving in-domain alignment with human-preference rewards and enhancing generalization on out-of-domain evaluations. The method is architecture-agnostic, lightweight, and fully compatible with existing GRPO training pipelines, providing a more expressive and informative learning signal for visual generation.

Talking face generation has gained significant attention as a core application of generative models.To enhance the expressiveness and realism of synthesized videos, emotion editing in talking face video plays a crucial role.However, existing approaches often limit expressive flexibility and struggle to generate extended emotions.Label-based methods represent emotions with discrete categories, which fail to capture a wide range of emotions.Audio-based methods can leverage emotionally rich speech signals--and even benefit from expressive text-to-speech (TTS) synthesis--but they fail to express the target emotions because emotions and linguistic contents are entangled in emotional speeches.Images-based methods, on the other hand, rely on target reference images to guide emotion transfer, yet they require high-quality frontal views and face challenges in acquiring reference data for extended emotions (e.g., sarcasm).To address these limitations, we propose Cross-Modal Emotion Transfer (C-MET), a novel approach that generates facial expressions based on speeches by modeling emotion semantic vectors between speech and visual feature spaces.C-MET leverages a large-scale pretrained audio encoder and a disentangled facial expression encoder to learn emotion semantic vectors that represent the difference between two different emotional embeddings across modalities.Extensive experiments on the MEAD and CREMA-D datasets demonstrate that our method improves emotion accuracy by 14% over state-of-the-art methods, while generating expressive talking face videos--even for unseen extended emotions. Code, checkpoint, and demo are available at https://chanhyeok-choi.github.io/C-MET/.