AAAI 2026 Papers

Modeling dynamic scenes through 4D Gaussians offers high visual fidelity and fast rendering speeds, but comes with significant storage overhead. Recent approaches mitigate this cost by aggressively reducing the number of Gaussians. However, this inevitably removes Gaussians essential for high-quality rendering, leading to severe degradation in dynamic regions. In this paper, we introduce a novel 4D anchor-based framework that tackles the storage cost in different perspective. Rather than reducing the number of Gaussians, our method retains a sufficient quantity to accurately model dynamic contents, while compressing them into compact, grid-aligned 4D anchor features. Each anchor is processed by an MLP to spawn a set of neural 4D Gaussians, which represent a local spatiotemporal region. We design these neural 4D Gaussians to capture temporal changes with minimal parameters, making them well-suited for the MLP-based spawning. Moreover, we introduce a dynamic-aware anchor growing strategy to effectively assign additional anchors to under-reconstructed dynamic regions. Our method adjusts the accumulated gradients with Gaussians' temporal coverage, significantly improving reconstruction quality in dynamic regions. Experimental results highlight that our method achieves state-of-the-art visual quality in dynamic regions, outperforming all baselines by a large margin with practical storage costs.

Massive multi-modality datasets are fundamental to the success of large video-language models. However, existing datasets often focus on providing textual descriptions for visual content, treating audio, particularly music, as weakly related information. This overlooks the inherent semantic correlation between visual narratives and musical scores, limiting the development of models for fine-grained cross-modal understanding and generation. To address this gap, we introduce VMChill, a large-scale, fine-grained multimodal video dataset. We leverage trailers as our data source, as they are professionally edited to create a strong synergy between visual pacing, scene transitions, and background music for narrative and emotional impact. Our dataset comprises over 20 million video clips derived from more than 27.1k hours of high-resolution trailer videos. To annotate this data, we propose a systematic multimodal captioning framework. This framework first employs specialized unimodal models to extract descriptive features from multiple perspectives, including visual content, motion dynamics, and musical attributes (e.g., genre, instruments, mood). Subsequently, a large language model (LLM) is utilized to adaptively fuse these diverse descriptions into a single, coherent, and rich multimodal caption. This process yields VMChill-2M, a high-quality subset of 2 million clips with detailed multimodal annotations, and VMChill-Test, a manually refined test set for evaluation. We conduct extensive experiments on downstream tasks, including video understanding and generation, to establish benchmarks and demonstrate the dataset's quality. The results validate that VMChill effectively enhances model performance, highlighting its potential to facilitate future research in fine-grained multimodal learning. We will release the dataset, annotation codebase, and processing pipelines to support community research.

Existing methods for human parsing into body parts and clothing often use fixed mask categories with broad labels that obscure fine-grained clothing types. Recent open-vocabulary segmentation approaches leverage pretrained text-to-image (T2I) diffusion model features for strong zero-shot transfer, but typically group entire humans into a single person category, failing to distinguish diverse clothing or detailed body parts. To address this, we propose Spectrum, a unified network for part-level pixel parsing (body parts and clothing) and instance-level grouping. While diffusion-based open-vocabulary models generalize well across tasks, their internal representations are not specialized for detailed human parsing. We observe that, unlike diffusion models with broad representations, image-driven 3D texture generators maintain faithful correspondence to input images, enabling stronger representations for parsing diverse clothing and body parts. Spectrum introduces a novel repurposing of an Image-to-Texture (I2Tx) diffusion model—obtained by fine-tuning a T2I model on 3D human texture maps—for improved alignment with body parts and clothing. From an input image, we extract human-part internal features via the I2Tx diffusion model and generate semantically valid masks aligned to diverse clothing categories through prompt-guided grounding. Once trained, Spectrum produces semantic segmentation maps for every visible body part and clothing category, ignoring standalone garments or irrelevant objects, for any number of humans in the scene. We conduct extensive cross-dataset experiments—separately assessing body parts, clothing parts, unseen clothing categories, and full-body masks—and demonstrate that Spectrum consistently outperforms baseline methods in prompt-based segmentation.

Typical detection-free methods for image-to-point cloud registration leverage transformer-based architectures to aggregate cross-modal features and establish correspondences. However, they often struggle under challenging conditions, where noise disrupts similarity computation and leads to incorrect correspondences. Moreover, without dedicated designs, it remains difficult to effectively select informative and correlated representations across modalities, thereby limiting the robustness and accuracy of registration. To address these challenges, we propose a novel cross-modal registration framework composed of two key modules: the Iterative Agents Selection (IAS) module and the Reliable Agents Interaction (RAI) module. IAS enhances structural feature awareness with phase maps and employs reinforcement learning principles to efficiently select reliable agents. RAI then leverages these selected agents to guide cross-modal interactions, effectively reducing mismatches and improving overall robustness. Extensive experiments on the RGB-D Scenes v2 and 7-Scenes benchmarks demonstrate that our method consistently achieves state-of-the-art performance.

In industrial point cloud analysis, detecting subtle anomalies demands high-resolution spatial data, yet prevailing benchmarks emphasize low-resolution inputs. To address this disparity, we propose a scalable pipeline for generating realistic and subtle 3D anomalies. Employing this pipeline, we developed MiniShift, the inaugural high-resolution 3D anomaly detection dataset, encompassing 2,577 point clouds, each with 500,000 points and anomalies occupying less than 1% of the total. We further introduce Simple3D, an efficient framework integrating Multi-scale Neighborhood Descriptors (MSND) and Local Feature Spatial Aggregation (LFSA) to capture intricate geometric details with minimal computational overhead, achieving real-time inference exceeding 20 fps. Extensive evaluations on MiniShift and established benchmarks demonstrate that Simple3D surpasses state-of-the-art methods in both accuracy and speed, highlighting the pivotal role of high-resolution data and effective feature aggregation in advancing practical 3D anomaly detection.

With the rapid development of generative AI, image steganography has garnered widespread attention due to its unique concealment. Recent studies have demonstrated the practical advantages of Fixed Neural Network Steganography (FNNS), notably its ability to achieve stable information embedding and extraction without any additional network training. However, the stego images generated by FNNS still exhibit noticeable distortion and limited robustness. These drawbacks compromise the security of the embedded information and restrict the practical applicability of the method. To address these limitations, we propose Robust Fixed Neural Network Steganography (RFNNS). Specifically, a texture-aware localization technique selectively embeds perturbations carrying secret information into regions of complex textures, effectively preserving visual quality. Additionally, a robust steganographic perturbation generation (RSPG) strategy is designed to enhance the decoding accuracy, even under common and unknown attacks. These robust perturbations are combined with AI-generated cover images to produce stego images. Experimental results demonstrate that RFNNS significantly improves robustness compared to state-of-the-art FNNS methods, achieving an average increase in SSIM of 23% for recovered secret images under common attacks. Furthermore, the LPIPS value of recovered secrets images against previously unknown attacks achieved by RFNNS was reduced to 39% of the SOTA method, underscoring its practical value for covert communication.

Continual learning for action recognition is a critical capability for next-generation Extended Reality (XR) systems. Yet it faces a severe real-world challenge: strict user privacy that prohibits data rehearsal. While recent prompt-based continual learning methods show promise, we argue their core 'flat,' single-granularity design fundamentally misaligns with the complexity of human actions. This monolithic architecture fails to model the inherent hierarchical structure and overlooks standard action primitives shared across tasks, resulting in suboptimal performance and hindered knowledge transfer. To overcome this limitation, we propose DPCA, a novel spatio-temporal continual learning framework with multi-granularity adaptive prompting. DPCA learns three synergistic components to resolve this mismatch. First, the task-specific prompter employs a multi-granularity query system to capture the unique, compositional semantics of each action. Second, the task-agnostic prompter learns a globally shared vocabulary of ``action primitives," providing a stable and generalizable knowledge base to mitigate catastrophic forgetting. Finally, we introduce a Dissimilarity Attention Rectification at each granularity level, leveraging a reverse attention mechanism to model class-agnostic background information and effectively alleviating overfitting. The synergy between these components enables robust model adaptation without requiring access to past data. Rigorous experiments on multiple large-scale benchmarks (including NTU RGB+D), under a strict rehearsal-free, few-shot protocol, confirm that DPCA establishes a new state-of-the-art. This advance paves the way for the realization of truly adaptive and privacy-respecting XR systems.

We present 360Explorer, a novel approach for generating 4D controllable panoramic videos conditioned on user-provided 3D instructions for exploring and manipulating dynamic worlds. Compared to existing perspective-based methods struggle to address spatial consistency during camera rotation in place, we introduce the panoramic view in controllable video generation models to inherently maintain the view recall consistency. By introducing dynamic point clouds as the 4D scene representations, 360Explorer unifies the modeling of camera transformations and object movements as incomplete renders to describe precise control instructions in 3D worlds. To tackle the data limitation in acquiring multi-viewpoint panoramic videos, we further propose a reverse warping strategy to construct the training dataset on easily accessible monocular panoramic videos. Extensive experiments demonstrate that 360Explorer achieves superior performance in creating 4D controllable panoramic videos with camera transformation and object movements aligned with diverse provided instructions.

3D hand pose estimation that involves accurate estimation of 3D human hand keypoint locations is crucial for many human-computer interaction applications such as augmented reality. However, this task poses significant challenges due to self-occlusion of the hands and occlusions caused by interactions with objects. In this paper, we propose HandMCM to address these challenges. Our HandMCM is a novel method based on the powerful state space model (Mamba). By incorporating modules for local information injection/filtering and correspondence modeling, the proposed correspondence Mamba effectively learns the highly dynamic kinematic topology of keypoints across various occlusion scenarios. Moreover, by integrating multi-modal image features, we enhance the robustness and representational capacity of the input, leading to more accurate hand pose estimation. Empirical evaluations on three benchmark datasets demonstrate that our model significantly outperforms current state-of-the-art methods, particularly in challenging scenarios involving severe occlusions. These results highlight the potential of our approach to advance the accuracy and reliability of 3D hand pose estimation in practical applications.

Accurate 3D scene motion perception significantly enhances the safety and reliability of an autonomous driving system. Benefiting from its all-weather operational capability and unique perceptual properties, 4D mmWave radar has emerged as an essential component in advanced autonomous driving. However, sparse and noisy radar points often lead to imprecise motion perception, leaving autonomous vehicles with limited sensing capabilities when optical sensors degrade under adverse weather conditions. In this paper, we propose RadarMP, a novel method for precise 3D scene motion perception using low-level radar echo signals from two consecutive frames. Unlike existing methods that separate radar target detection and motion estimation, RadarMP jointly models both tasks in a unified architecture, enabling consistent radar point cloud generation and pointwise 3D scene flow prediction. Tailored to radar characteristics, we design specialized self-supervised loss functions guided by Doppler shifts and echo intensity, effectively supervising spatial and motion consistency without explicit annotations. Extensive experiments on the public dataset demonstrate that RadarMP achieves reliable motion perception across diverse weather and illumination conditions, outperforming radar-based decoupled motion perception pipelines and enhancing perception capabilities for full-scenario autonomous driving systems.

Glass surfaces challenge object detection models as they mix the transmitted background with the reflected surrounding, creating confusing visual patterns. Previous methods relying on low-level cues (e.g., reflections and boundaries) or surrounding semantics are often unreliable in complex real-world scenarios. A glass image inherently comprises three distinct semantic components: semantics of the transmitted content, semantics of the reflected content, and semantics of the surrounding content. In this work, we observe that there is a relationship among these three types of semantics, where reflection semantics closely resembles surrounding semantics, while these two types of semantics tend to be different from the transmission semantics. For example, when on a street, we may see into a cafeteria through a glass wall, intermixed with reflection of the street, while the glass is surrounded by other street contents like shops and pedestrians, thereby creating a unique multi-semantic signature. Based on this observation, we propose the Multi-Semantic Net, MSNet, which identifies transmission, reflection, and surrounding semantics from glass images and exploits their relationships for glass surface detection. MSNet consists of two novel modules: (1) A Semantic Decomposition Module (SDM) containing Dual-Semantics Extraction Block to extract original image and reflection semantics and Semantic Elimination Block to progressively derive transmission and surrounding semantics, and (2) An Adaptive Semantic Fusion Module (ASFM) to fuse these semantic components and adaptively learn their relationships to handle varying reflection conditions. Extensive experiments demonstrate that MSNet surpasses SOTA methods on public glass detection benchmarks.

This paper addresses cross-view geo-localization in real-world scenarios, where the field-of-view (FoV) is restricted and the orientation is unknown for ground-view images. This task is extremely challenging due to the huge domain gap. Existing methods typically treat tasks with different FoVs as independent tasks. These approaches not only require separate retraining for each FoV, but also neglect the strong correlations between different FoVs, leading to poor performance under extremely limited FoV. To overcome these limitations, we propose HCL-Geo, a framework follows human-like continual learning paradigm of "first learn, then review" for geo-localization: in the first "learn" stage, tasks are presented to the model in an easy-to-hard sequence to enable gradual learning and knowledge retention, so that their natural correlations could be exploited to facilitate knowledge transfer. In the second "review" stage, expert modules are incorporated to efficiently handle tasks with varying FoVs. This approach eliminates the need for retraining separate models and demonstrates state-of-the-art performance across different FoVs with strong generalization capabilities. Remarkably, the recall rate@top-1 improves from 49.1% to 68.3% and from 24.6% to 34.3% respectively on CVUSA and CVACT benchmarks with 70° FoV.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in multimodal understanding and generation by integrating visual and textual data. However, these models frequently exhibit object hallucination problems: generating outputs that are inconsistent with the input image. Existing improved methods for mitigating hallucinations still suffer from two key limitations: dynamic approaches based on logits or attention mechanisms risk suppressing valuable linguistic priors, whereas static methods that employ fixed intervention vectors lack the flexibility to adapt to diverse images and questions. To address these issues, we propose RFI (Rectified Flow Intervention), a novel approach that harnesses the linear trajectory design of rectified flow for input-specific adaptation and employs gradient correction to ensure coherent generation, effectively combining the adaptability of dynamic methods with the stability of static ones. RFI dynamically predicts latent-space intervention vectors while requiring only a single forward pass in LVLMs per question, achieving computational efficiency (1.09x latency overhead for 100 new tokens). Extensive experiments show RFI significantly reduces hallucinations, achieving superior performance compared to existing advanced methods, highlighting its effectiveness as a lightweight plug-and-play method for reducing LVLM's hallucination in practical applications.

Color temperature, as a crucial attribute influencing image color, plays a critical role in Image Aesthetics Assessment (IAA). Yet, within the existing IAA field, little light has been shed on assessing the aesthetic quality of image color temperature. To bridge this gap, we introduce a new task: Image Color Temperature Aesthetics Assessment (ICTAA). However, this task poses the following challenges: 1) Perceptual Sensitivity: humans exhibit high sensitivity to subtle shifts in color temperature, necessitating a model to enable fine-grained discrimination; 2) Spectral Continuity: The theoretical modeling of color temperature aesthetics requires continuous labels; however, the just-noticeable-difference property of human perception makes continuous labeling infeasible, necessitating a well-designed labeling strategy. To address the aforementioned challenges, we make the following efforts. First, we propose a multi-modal contrastive learning framework, ICTA2Net, that models color temperature differences between image pairs while strictly controlling other visual attributes. Second, leveraging color temperature transitivity, we design a weakly supervised strategy that discretely samples images based on anchor images and human perception to build contrastive relations across color temperatures, enabling learning from discrete labels. Thirdly, we construct a color temperature aesthetics dataset, ICTAA240K, and a benchmark for validation. Additionally, we propose a new metric, Information Entropy-weighted Accuracy (IEA), which weights accuracy by the degree of annotation disagreement to reflect model performance across varying sample difficulties, complementing existing evaluation metrics. Experiments show our method outperforms existing state-of-the-art IAA methods on ICTAA240K, thereby setting an effective roadmap for ICTAA.

Video diffusion generation suffers from critical sampling efficiency bottlenecks, particularly for large-scale models and long contexts. Existing video acceleration methods, adapted from image-based techniques, lack a single-step distillation ability for large-scale video models and task generalization for conditional downstream tasks. To bridge this gap, we propose the Video Phased Adversarial Equilibrium (V-PAE), a distillation framework that enables high-quality, single-step video generation from large-scale video models. Our approach employs a two-phase process. (i) Stability priming is a warm-up process to align the distributions of real and generated videos. It improves the stability of single-step adversarial distillation in the following process. (ii) Unified adversarial equilibrium is a flexible self-adversarial process that reuses generator parameters for the discriminator backbone. It achieves a co-evolutionary adversarial equilibrium in the Gaussian noise space. For the conditional tasks, we primarily preserve video-image subject consistency, which is caused by semantic degradation and conditional frame collapse during the distillation training in image-to-video (I2V) generation. Comprehensive experiments on VBench-I2V demonstrate that V-PAE outperforms existing acceleration methods by an average of 5.8% in the overall quality score, including semantic alignment, temporal coherence, and frame quality. In addition, our approach reduces the diffusion latency of the large-scale video model (e.g., Wan2.1-I2V-14B) by 100 times, while preserving competitive performance.

Medical images exhibit latent anatomical groupings, such as organs, tissues, and pathological regions, that standard Vision Transformers (ViTs) fail to exploit. While recent work like SBM-Transformer attempts to incorporate such structures through stochastic binary masking, they suffer from non-differentiability, training instability, and the inability to model complex community structure. We present DCMM-Transformer, a novel ViT architecture for medical image analysis that incorporates a Degree-Corrected Mixed-Membership (DCMM) model as an additive bias in self-attention. Unlike prior approaches that rely on multiplicative masking and binary sampling, our method introduces community structure and degree heterogeneity in a fully differentiable and interpretable manner. Comprehensive experiments across diverse medical imaging datasets, including brain, chest, breast, and ocular modalities, demonstrate the superior performance and generalizability of the proposed approach. Furthermore, the learned group structure and structured attention modulation substantially enhance interpretability by yielding attention maps that are anatomically meaningful and semantically coherent.

Contact-based grasp generation plays a crucial role in various applications. Recent methods typically focus on the geometric structure of objects, producing grasps with diverse hand poses and plausible contact points. However, these approaches often overlook the physical attributes of the grasp, specifically the contact force, leading to reduced stability of the grasp. In this paper, we focus on stable grasp generation using explicit contact force predictions. First, we define a force-aware contact representation by transforming the normal force value into discrete levels and encoding it using a one-hot vector. Next, we introduce force-aware stability constraints. We define the stability problem as an acceleration minimization task and explicitly relate stability with contact geometry by formulating the underlying physical constraints. Finally, we present a pose optimizer that systematically integrates our contact representation and stability constraints to enable stable grasp generation. We show that these constraints can help identify key contact points for stability which provide effective initialization and guidance for optimization towards a stable grasp. Experiments are carried out on two public benchmarks, showing that our method brings about 20% improvement in stability metrics and adapts well to novel objects.

Curvilinear structure segmentation (CSS) plays a vital role in industrial applications, including medical imaging and structural health monitoring. Recently, the strong capacity of the Segment Anything Model (SAM) has inspired its downstream application in CSS tasks. To adapt SAM to CSS tasks, previous methods heavily rely on a certain number of samples and costly pixel-level annotation, which are hard to access for a new scenario. Considering this, the goal of our work is to adapt SAM in a very cost-effective setting where only a single unlabeled image is given. This is far more challenging than the typical supervised, unsupervised, or self-supervised learning manner that needs a large number of training samples. To tackle this problem, we propose a finetuning-free SAM for curvilinear structure segmentation, called curvilinear-aware prompt learning (CaPro), which aims to automatically learn visual prompts via a single unlabeled image. In the first stage, we generate extensive curvilinear structures and oriented sub-curvilinear box annotations. To increase the realism of generated curvilinear structures, we adapt these structures into real image domains via the Fourier Transform using a single real-world unlabeled image. Now, these adapted images can be used to train our oriented sub-curvilinear detector. In the second stage, we propose the curvilinear-aware discrete representation matching to filter those unreliable detection results. Afterward, these reliable detection results can be converted into informative prompts, contributing to the cost-effective SAM adaptation to CSS tasks. Experiments demonstrate the effectiveness of CaPro on medical image and crack segmentation tasks.

Underwater Instance Segmentation (UIS), integrating pixel-level understanding and instance-level discrimination, is a pivotal technology in marine resource exploration and ecological protection. In recent years, large-scale pretrained visual foundation models, exemplified by DINO, have advanced rapidly and demonstrated remarkable performance on complex downstream tasks. In this paper, we demonstrate that DINO can serve as an effective feature learner for UIS, and we introduce DiveSeg, a novel framework built upon two insightful components: (1) The AquaStyle Aligner, designed to embed underwater color style features into the DINO fine-tuning process, facilitating better adaptation to the underwater domain. (2) The ObjectPrior Prompter, which incorporates binary segmentation-based prompts to deliver object-level priors, provides essential guidance for instance segmentation task that requires both object- and instance-level reasoning. We conduct thorough experiments on the popular UIIS and USIS10K datasets, and the results show that DiveSeg achieves the state-of-the-art performance.