AAAI 2026 论文列表

Collaborative perception leveraging intermediate feature fusion has emerged as a leading paradigm to significantly enhance the environmental perception capabilities of autonomous driving systems. However, existing methods typically rely on discriminative supervision guided by downstream tasks. This paradigm compels models to learn minimal, task-specific representations, which conflicts with the goal of cooperative perception to capture comprehensive information, thereby limiting generalization. To address this issue, we propose DiGS-CP, a novel two-stage generative supervised collaborative perception framework. Specifically, we introduce a diffusion-based generative task that conditions on fused object-level features to generate representations of object-level point clouds. The proposed generative supervision provides fine-grained, task-agnostic signals that encourages the fusion module to learn comprehensive representations beyond task-specific requirements. By preserving and integrating complementary information from collaborative agents, our approach overcomes the limitations of task-specific learning and enhances the generalizability of the learned features. Furthermore, our two-stage architecture requires agents to transmit only object-level features, significantly reducing communication overhead. Extensive experiments on three benchmark datasets demonstrate that DiGS-CP achieves state-of-the-art performance in 3D object detection, while maintaining low bandwidth requirements and exhibiting excellent generalization ability.

Scene-aware motion synthesis has been widely researched recently due to its numerous applications. Prevailing methods rely heavily on paired motion-scene data, while it is difficult to generalize to diverse scenes when trained only on a few specific ones. Thus, we propose a unified framework, termed Diffusion Implicit Policy (DIP), for scene-aware motion synthesis, where paired motion-scene data are no longer necessary. In this paper, we disentangle human-scene interaction from motion synthesis during training, and then introduce an interaction-based implicit policy into motion diffusion during inference. Synthesized motion can be derived through iterative diffusion denoising and implicit policy optimization, thus motion naturalness and interaction plausibility can be maintained simultaneously. For long-term motion synthesis, we introduce motion blending in joint rotation power space. The proposed method is evaluated on synthesized scenes with ShapeNet furniture, and real scenes from PROX and Replica. Results show that our framework presents better motion naturalness and interaction plausibility than cutting-edge methods. This also indicates the feasibility of utilizing the DIP for motion synthesis in more general tasks and versatile scenes.

Human motion synthesis in 3D scenes relies heavily on scene comprehension, while current methods focus mainly on scene structure but ignore the semantic understanding. In this paper, we propose a human motion synthesis framework that take an unified Scene Semantic Occupancy (SSO) for scene representation, termed SSOMotion. We design a bi-directional tri-plane decomposition to derive a compact version of the SSO, and scene semantics are mapped to an unified feature space via CLIP encoding and shared linear dimensionality reduction. Such strategy can derive the fine-grained scene semantic structures while significantly reduce redundant computations. We further take these scene hints and movement direction derived from instructions for motion control via frame-wise scene query. Extensive experiments and ablation studies conducted on cluttered scenes using ShapeNet furniture, as well as scanned scenes from PROX and Replica datasets, demonstrate its cutting-edge performance while validating its effectiveness and generalization ability.

Text-guided image inpainting endeavors to generate new content within specified regions of images using textual prompts from users. The primary challenge is to accurately align the inpainted areas with the user-provided prompts while maintaining a high degree of visual fidelity. While existing inpainting methods have produced visually convincing results by leveraging the pre-trained text-to-image diffusion models, they still struggle to uphold both prompt alignment and visual rationality simultaneously. In this work, we introduce FreeInpaint, a plug-and-play tuning-free approach that directly optimizes the diffusion latents on the fly during inference to improve the faithfulness of the generated images. Technically, we introduce a prior-guided noise optimization method that steers model attention towards valid inpainting regions by optimizing the initial noise. Furthermore, we meticulously design a composite guidance objective tailored specifically for the inpainting task. This objective efficiently directs the denoising process, enhancing prompt alignment and visual rationality by optimizing intermediate latents at each step. Through extensive experiments involving various inpainting diffusion models and evaluation metrics, we demonstrate the effectiveness and robustness of our proposed FreeInpaint.

This paper tackles the challenging task of achieving storage-efficient yet high-fidelity motion representation in large-scale dynamic 3D Gaussian Splatting. Our motivation stems from the truth that existing urban-scale methods, which rely on massive and unstructured individual Gaussians for scene modeling, face a critical scalability bottleneck. Inspired by recent advances in the 3DGS-based compression beyond autonomous driving, we address this challenge by leveraging the compression capability of anchor-driven methods. However, this is non-trivial as our exploratory experiments reveal that the direct application of this paradigm to dynamic, large-scale urban scenes results in performance degradation. We attribute this phenomenon to the hierarchical anchor design that severely loses dynamic information. To this end, we propose Hierarchical Dynamic Gaussian Splatting (HDGS), a novel framework designed to adapt the anchor-based Gaussian paradigm to 4D urban environments. We first establish a local support network to reinforce inter-anchor consistency, mitigating geometric and appearance fractures caused by supervision attenuation in deep hierarchies. Then, we handle heterogeneous object motion via coarse-to-fine decomposition, where high-level anchors model coarse dynamics and low-level anchors refine them with residual deformations. Third, we introduce a hybrid supervision scheme that fuses global geometric constraints and local pixel-level cues to alleviate geometrically inconsistent reconstruction under sparse LiDAR. Extensive experiments show that HDGS reduces storage by 69.0% while maintaining or even improving rendering fidelity compared to state-of-the-art methods.

Semantic understanding of large-scale aerial scenes represents a critical challenge in 3D computer vision, hindered by the prohibitive cost of dense annotation. This paper introduces EvoPropGS, a novel approach for the semantic segmentation of 3D Gaussian Splatting models that requires only minimal supervision. Our core insight is to leverage the inherent structural repetitions within aerial environments to propagate semantic information from a sparse set of annotations across the entire 3D scene. Our approach constructs a prompt library by pairing SAM-generated mask candidates with DINOv2 feature embeddings from annotated views. For unannotated regions, we generate pseudo-labels by matching region proposals with these featured prompts via cosine similarity. We then formulate optimal prompt selection as a discrete optimization problem solved via evolutionary search, guided by our novel fitness function that evaluates both 3D consistency and 2D semantic coherence. Extensive experiments demonstrate that EvoPropGS achieves accurate segmentation with only 2 percent annotated pixels.

In this work, we introduce a novel high-fidelity full-head 3D avatar generation method from a single image, regardless of perspective, style, expression, or accessories. Prior works often fail to preserve consistent head geometry and facial details, primarily due to their limited capacity in modeling fine-grained facial textures and maintaining identity information. To address these challenges, we construct a new high-quality dataset containing 227 sequences of digital human portraits captured from 96 different perspectives, totalling 21,792 frames, featuring high-quality facial texture details. To further improve performance, we propose a novel multi-view diffusion named ID-TS diffusion model, which integrate identity and expression information into the two-stage multi-view diffusion process. The low-resolution stage ensures structural consistency of heads across multiple views, while the high-resolution stage preserves facial detail fidelity and coherence. Finally, we propose an enhanced feed-forward Gaussian avatar reconstruction method that optimizes the network on multi-view images of each single subject, significantly improving 3D facial texture details. Extensive experiments show that our method demonstrates robust performance across challenging scenarios, while showcasing broad applicability across numerous downstream tasks.

Facial Expression Recognition (FER) seeks to classify affective states from facial images, which remains a challenging problem due to variations in real-world conditions. FER task becomes particularly complex when handling unconstrained environments characterized by partial occlusions, different head poses, and so on. To address the above problems, current approaches rely on extensive learnable parameters and complex model architectures, which inevitably lead to overfitting and cause the FER model to focus on non-discriminative facial regions. In this work, we propose an HKAFER model that can adaptively enhance visual expression representations through efficiently fine-tuning the image encoder in large Visual Foundation Models (VFMs) and Vision-Language Models (VLMs). Specifically, we establish Heterogeneous Kronecker Adaptation (HeKA), which consists of multi-scale adapters based on Kronecker product in a parallel manner, offering significantly diverse subspaces to learn the incremental matrices. Besides, we also propose Dual-Branch Interactive Router (DBIR) to dynamically assign the weights of adapters, which promotes collaboration and information flow among them. In this way, our HKAFER can effectively capture robust spatial features and the regional associations. Experimental results demonstrate that our proposed model not only outperforms state-of-the-art methods on several FER benchmarks but also uses significantly fewer trainable parameters.

In recent years, Gaussian scene representations have achieved a series of promising results in 3D reconstruction. Compared to the previous 3DGS paradigm, the latest reconstruction approach 2DGS can achieve more accurate geometric representation using fewer Gaussian points. Accordingly, developing a panoramic segmentation algorithm suitable for 2DGS-reconstructed scenes is of significant importance. However, existing segmentation methods are primarily designed for 3DGS. They either fail to account for all objects in complex segmentation scenes or suffer from significant performance degradation when applied to 2D Gaussian scenes. Moreover, these methods consistently exhibit poor cross-dataset generalization. To address these issues, we propose IQGS, a segmentation framework applicable to 2DGS representations. Specifically, IQGS employs per-instance query and relaxed object-level supervision instead of strict pixel-level ID supervision, effectively mitigating the segmentation performance degradation that occurs when applied to 2DGS. At the same time, by learning features independent of specific object ID assignments, IQGS enhances its ability to generalize across diverse datasets. Our method achieves impressive panoramic segmentation results across multiple datasets, with an average mIoU of 66.6%, surpassing the state-of-the-art method Gaussian Grouping, which achieves 57.17%.

Self-supervised monocular depth estimation methods severely compromise accuracy in dynamic objects due to their static scene assumption. Existing approaches for dynamic scenes suffer from two critical shortcomings: 1) reliance on supervised segmentation models (requiring costly annotations) or computationally intensive multi-branch models to isolate moving objects, and 2) simple integration of 2D/3D motion flow without reliable supervision for dynamic objects. We propose AdaDepth, a two‑stage framework that jointly performs unsupervised scene decomposition and dynamic-aware depth learning. In the initial structural stage, our geometry-motion joint scene decomposition (GMoDecomp) module ensures the robust generation of a depth prior and simultaneously partitions the scene into multiple regions through the fusion of geometric and motion cues. In the region-adaptive refinement stage, we exploit the depth prior and decomposed regions to introduce motion-aware and geometry-consistent constraints, effectively improving depth estimation in dynamic scenes. AdaDepth achieves accurate depth prediction in highly dynamic scenes without relying on external labels or specialized segmentation models. Extensive experiments on KITTI, Cityscapes, and Waymo Open demonstrate its superiority over state-of-the-art approaches.

Previous studies leveraging artificial neural networks have been used to investigate the semantic coding within human visual cortex. However, building an interpretable label-free framework that can effectively map brain responses to multiple coexisting semantic concepts remains largely unexplored. Here, we propose BrainLMM, a label-free framework for multi-semantic mapping of voxel responses by combining diverse vision encoders with the Describe-and-Dissect strategy, enabling a hypothesis-free analysis of the human high-level visual cortex. First, we construct voxel-wise encoding models leveraging diverse vision encoders to predict visual cortical responses to natural scene images. Then, we use BrainLMM to map individual brain voxels to multiple semantics without requiring any predefined labels. To evaluate the effectiveness of our method, we compute Pearson correlation coefficients to compare the multi-semantic mappings produced by BrainLMM and CLIP-MSM with ground-truth voxel responses within selective cortical areas. Our findings indicate that BrainLMM achieves more accurate predictions of visual responses compared to CLIP-MSM. Finally, to demonstrate the multi-semantic mapping capability of our method, we project multiple representative semantic concepts onto the cortical surface for visualization. Our method enables the discovery of voxels that exhibit strong activation in response to previously undefined semantic concepts across two independent datasets: the Natural Scenes Dataset (NSD) and the Natural Object Dataset (NOD).

For video anomaly detection, it's both important to detect when the event happens and what the event is. The tasks of temporal grounding and semantic understanding can benefit from joint learning, but no existing work support it. To address this problem, we introduce VAGU (Video Anomaly Grounding and Understanding), the first benchmark designed to jointly evaluate semantic understanding and precise temporal grounding of anomalies, with comprehensive annotations and objective multiple-choice Video QA. Besides, we propose Glance then Scrutinize (GtS), the first training-free framework that achieves the best balance performance in both accuracy and efficiency. GtS uniquely balances high temporal precision and semantic interpretability while meeting practical speed requirements, outperforming previous methods in real-world scenarios. Furthermore, we introduce the JeAUG metric for holistic evaluation of both speed and accuracy. Extensive experiments demonstrate the superior effectiveness and practicality of our benchmark, framework, and metric.

Multimodal remote sensing image joint classification has achieved significant progress. However, existing methods primarily focus on designing modality-specific networks, lacking adaptive generalization capabilities in diverse and dynamic modality combinations encountered in real-world scenarios. Inspired by the generalization capabilities of visual foundation model in downstream tasks, we propose a unified Text-guided Arbitrary Modalitiy Prompting (T-APT) framework, which leverages complementary fused features to drive the foundation model and employs text-guided modality-specific prior knowledge as cross-modal prompts to fine-tune a pretrained Vision Transformer (ViT) model. Specifically, a Mamba-Based Arbitrary Modal-Focused Feature Capture (MAMF-FC) module is designed to extract complementary joint features and modality-specific prior knowledge from arbitrary modalities through a shared-specific scanning encoder-decoder architecture. Subsequently, a Text-Guided Modality-Aware Prompt Tuning (TMPT) module is proposed to support the adaptation of fused features to the foundation model, enabling our arbitrary remote sensing image classification task. Extensive experiments on public datasets spanning multispectral (MS), hyperspectral (HS), light detection and ranging (LiDAR), and synthetic aperture radar (SAR) modalities demonstrate that our T-APT achieves classification performance comparable to specialized networks across arbitrary modal combinations.

We present VisAssist, the first large-scale video question-answering dataset with 13,413 real-world videos captured by visually impaired users, addressing a critical gap in assistive vision research. Unlike existing benchmarks relying on third-person footage, VisAssist provides authentic first-person perspectives that uniquely capture challenges in blind photography—including unconventional framing, motion artifacts, and frequent information omission. Benchmark evaluations of SOTA multimodal models reveal systematic limitations: severe deficiencies in spatial reasoning when processing dynamic first-person viewpoints, an inability to distinguish missing information from poor capture quality leading to hazardous hallucinations, and fragile text understanding especially for non-Latin scripts under suboptimal conditions. This work establishes a vital real-world benchmark and underscores the need for specialized architectures in visual assistance systems.

Large Vision-Language Models (LVLMs) have achieved impressive performance across a wide range of multimodal tasks. However, they still face critical challenges in modeling long-range dependencies under the usage of Rotary Positional Encoding (ROPE). Although it can facilitate precise modeling of token positions, it induces progressive attention decay as token distance increases, especially with progressive attention decay over distant token pairs, which severely impairs the model's ability to remember global context. To alleviate this issue, we propose inference-only Three-step Decay Resilience Strategies (T-DRS), comprising (1) Semantic-Driven DRS (SD-DRS), amplifying semantically meaningful but distant signals via content-aware residuals, (2) Distance-aware Control DRS (DC-DRS), which can purify attention by smoothly modulating weights based on positional distances, suppressing noise while preserving locality, and (3) re-Reinforce Distant DRS (reRD-DRS), consolidating the remaining informative remote dependencies to maintain global coherence. Together, the T-DRS recover suppressed long-range token pairs without harming local inductive biases. Extensive experiments on Vision Question Answering (VQA) benchmarks demonstrate that T-DRS can consistently improve performance in an inference-only manner.

Current text-to-image models face challenges in visual text rendering: text encoders like CLIP and T5 lack glyph-level understanding and often struggle to distinguish between the specific words to be rendered and their intended semantic meaning within prompts. In addition, inconsistencies between the base model and its plugins further compromise the quality of synthesized images. In this paper, we enhance the existing text-to-image method by addressing the following aspects: (1) Text-Glyph Alignmentin a Visual Question Answering (VQA) manner to enable glyph understanding for the text encoder. This involves establishing an explicit alignment between the representations of the glyphs and their detailed attribute descriptions, which boosts the model's ability to capture fine-grained visual features of the text. (2) Accurate and harmony visual text rendering: integrating pre-aligned glyph-visual embeddings with semantic text tokens through the Multimodal Diffusion Transformer(MMDiT) synchronously, ensuring coherent feature alignment and enhancing both the robustness and fidelity of visual text rendering. (3) Image Aesthetic Refinement: leveraging a multisource data training strategy that incorporates diverse, high-quality image-text pairs from various domains, exposing the model to extensive linguistic and visual diversity while maintaining superior aesthetic quality throughout training. Our experiments demonstrate that the proposed approach significantly outperforms the existing state-of-the-art method.

Recent advances in deep learning-based 3D representation have achieved remarkable success, particularly in modeling static high-fidelity geometries. However, the extension of these techniques to dynamic 3D scenes introduces a critical challenge of effectively representing spatio-temporal dependencies, i.e., jointly modeling detailed spatial structures within frames and temporal dynamics across frames. To address this challenge, this paper proposes that the temporal evolution observed in dynamic 3D scenes is fundamentally attributable to the deformation of underlying spatial structures. To capture this relationship, we introduce a unified continuous 4D latent space representation incorporating a structure-equivalence prior, named SEP-4D. The core of SEP-4D is an efficient 4D tensor decomposition-fusion approach. This method fuses decomposed learnable 2D feature planes via a plane-wise spatio-temporal fusion mechanism of planar distributions, explicitly enforcing the principle that temporal evolution originates from geometric deformations of the 3D structure. To mitigate the associated computational demands, we sample the 3D probability volumes generated by VAE-based fusion into a spatio-temporally consistent 4D latent representation. The efficacy of our approach is validated through experiments on the fundamental task of 4D occupancy reconstruction. Extensive results demonstrate that, by leveraging the inherent equivalence of temporal dynamics and structural deformation, our method achieves high-quality reconstruction across various sequence lengths. Notably, for 4-frame scenes, we attain an impressive 91.68% mIoU, significantly outperforming state-of-the-art baselines on standard benchmarks.

Current multimodal large language models (MLLMs) struggle with hour-level video understanding, facing significant challenges not only in modeling the substantial information volume of long videos but also in overcoming the memory wall and resource constraints during both training and inference. Although recent training-free approaches have alleviated resource demands by compressing visual features, their reliance on incomplete visual information limits the performance potential. To address these limitations, we propose Adaptive Pivot Visual information Retrieval (APVR), a training-free framework that hierarchically retrieves and retains sufficient and important visual information. It breakthroughs the memory wall limitation via two complementary components: Pivot Frame Retrieval employs query expansion and iterative spatio-semantic confidence scoring to identify relevant video frames, and Pivot Token Retrieval performs query-aware attention-driven token selection within up to 1024 pivot frames. This dual granularity approach enables the processing of hour-long videos while maintaining semantic fidelity. Experimental validations on three different baseline MLLMs demonstrate significant performance improvements up to 9.5%, 4.6% and 9.7% on LongVideoBench, VideoMME and MLVU, respectively. APVR achieves state-of-the-art results for both training-free and training-based approaches.

Accurately forecasting the spatiotemporal dynamics of biological systems, such as human pluripotent stem cell (hPSC)-derived cardiac organoids, from microscopy time-series is a critical challenge in biomedicine with profound implications for drug discovery. Existing generative models often fail to capture the intricate dynamics of organoid development, struggling with their irregular morphology, indistinct boundaries, and complex spatiotemporal patterns. To overcome these limitations, we introduce OrgaCast, a novel multimodal conditional diffusion model for high-fidelity organoid forecasting. OrgaCast uniquely conditions the generative process on three synergistic modalities: (i) historical image sequences, captured by a dedicated spatiotemporal control module; (ii) structured numerical metadata defining experimental conditions; and (iii) descriptive text captions summarizing the biological context. This comprehensive conditioning enables the generation of forecasts with high visual accuracy and biological plausibility. Furthermore, to enhance the model's utility in critical research settings, we introduce a post-hoc uncertainty quantification method that produces intuitive confidence maps, bolstering the interpretability and trustworthiness of predictions. Extensive experiments on a challenging cardiac organoid dataset demonstrate that OrgaCast outperforms baselines in metrics such as SSIM, PSNR, and LPIPS. Our framework presents a robust solution for biological forecasting, promising to accelerate research discovery while minimizing experimental costs and manual effort.