IJCAI 2026 论文列表

Most text-to-motion models treat motion duration as a fixed hyperparameter rather than a variable inferred from semantics, inducing two systematic failure modes: duration trailing on short prompts and stage-wise semantic collapse on long, multi-stage prompts. To address this issue, we propose TPPMG (Temporal Planning-driven Progressive Motion Generation), which factorizes generation as semantics -> temporal structure -> motion realization. TPPMG comprises two components: (1) a Temporal Planner that maps text to an explicit stage-duration script via an LLM with a lightweight duration calibrator; and (2) a Progressive Diffusion Generator that bridges fixed-length diffusion training and variable-length inference through heterogeneous noise scheduling and a sliding-window mechanism, enabling budget-driven variable-length motion synthesis. To evaluate both failure modes with verifiable ground truth, we construct three benchmarks: HumanML3D-Short, HumanML3D-Concat, and BABEL-Concat, along with a comprehensive evaluation suite. Experiments show that TPPMG suppresses duration trailing (UR: 0.52 -> 0.85), substantially improves multi-stage structure recovery (PhaseRecall: 0.56 -> 0.90, BoundaryF1: 0.28 -> 0.56), and reduces boundary discontinuity at stage transitions.

In the field of skeleton-based human action recognition, Graph Convolutional Networks (GCNs) have become a dominant framework. However, existing GCN-based approaches often treat the sequences of two-person interaction as separate entities, ignoring the inherent semantic dependencies and spatial correlations between interacting subjects. Furthermore, high-order skeleton representations naturally exhibit non-Euclidean structures, where Euclidean deep learning models are inherently limited in explicitly capturing and preserving such geometric information. As a countermeasure, we propose a Riemannian Graph Convolutional Network (RGCN) that operates on the Symmetric Positive Definite (SPD) manifolds. Specifically, we model high-order skeletal statistics via Gaussian embedding and propose a Riemannian network to capture inter-subject interactions and global correlations. The proposed RGCN is instantiated under three SPD geometries, and its effectiveness is validated through extensive experiments on three interaction benchmarks. Extensive experimental results show that RGCN provides a competitive and geometrically grounded alternative for skeleton-based interaction recognition.

Cooperative perception allows agents to extend perceptual capabilities through inter-agent communication. However, most existing methods still adopt a frame-wise paradigm, which limits the exploitation of spatio-temporal consistency in dynamic scenes. Therefore, when observations are partially occluded or degraded, these methods are unable to leverage historical context for compensation, leading to unstable perception and reduced detection accuracy. To address these challenges, we propose World4V2X, the first world model framework tailored for V2X cooperative perception. The proposed method first introduces a spatial observability modeling module that defines spatial consistency boundaries to distinguish reliable regions from uncertain ones, thereby enabling spatial consistency modeling over multi-agent heterogeneous observations. Building upon this, we construct a consistency-guided world modeling module. Specifically, a temporal consistency evolution mechanism leverages features from historical and current frames to model the state evolution of the scene, capturing environmental dynamics and producing temporal consistency beliefs. Meanwhile, a consistency-guided deterministic reconstruction mechanism exploits spatial boundaries and temporal beliefs to perform diffusion-based refinement for robust cooperative perception. Extensive experiments on the OPV2V and V2XSet datasets demonstrate that World4V2X achieves state-of-the-art perception performance across diverse V2X scenarios.

While class-agnostic object counting has recently evolved from image-exemplar to language-guided paradigms, existing methods are limited by text polysemy and the lack of prompts in audio-sensing scenarios. To overcome these challenges, we introduce a sound-centric counting paradigm, enabling models to "listen and count" using the vivid and discriminative signatures of sound cues. We propose S2ICount, which achieves seamless, fine-grained audio-visual alignment for sound-guided counting. At its core, SoundMamba leverages linear-complexity long-range semantic modelling to align cross-modal features, while the Sound Calibration Module designs a circular scanning mechanism to progressively reconcile modality discrepancies using multi-scale cues from a Stable Diffusion backbone. To enhance saliency awareness, a Multi-tier Hybrid Optimization strategy enforces consistency across magnitude, semantics, and distribution. In addition, we introduce SoundCount, the first large-scale sound-oriented counting dataset, comprising over 2,000 scenes and 1,650 audio sources across 66 object categories. Extensive experiments demonstrate that S2ICount achieves state-of-the-art performance, validating the effectiveness of sound as a robust guidance modality for counting. The dataset, demo and implementation code are made publicly accessible at https://github.com/ZSTU-CV-Lab/S2ICount.

Learning reliable two-view correspondences is essential for geometric computer vision applications. Existing graph-based pruning methods typically aggregate information unidirectionally, capturing only whether a correspondence receives neighborhood support while ignoring its structural contribution during aggregation. This makes it difficult to distinguish locally clustered outliers from globally critical inliers. To overcome this limitation, we propose a Bidirectional Aggregation Graph Network (BAG-Net), which introduces a bidirectional aggregation perspective for correspondence pruning. Specifically, we design Bidirectional Dynamic Graph Convolution (BiDGC) that jointly models forward consistency aggregation and reverse structural dependency, enabling the network to identify key inliers that are widely referenced during feature propagation. To further enhance robustness, we design a Stage-wise Graph Expansion-Contraction (SGEC) architecture that combines multi-scale receptive field evolution with Stage-wise Squeeze-and-Excitation (SSE) and Deep-Guided Modulation (DGM) to stabilize feature propagation across network stages. Extensive experiments on the YFCC100M and SUN3D datasets demonstrate that BAG-Net significantly outperforms existing SOTA methods in both outlier rejection and camera pose estimation, achieving 66.10% and 78.88% mAP5^o% on YFCC100M in known and unknown scenes, respectively. Source code: https://github.com/LeyiWang13/BAG-Net.

Open-vocabulary 3D understanding aims to align 3D representations with a unified vision-language semantic space. However, existing methods suffer from the challenge of structural asymmetry caused by sparse observations and holistic geometries. Additionally, the inherent semantic chasm between discrete coordinates and abstract natural language hinders cross-modal mapping. This work introduces SpecBridge, a 3D-2D-Text pre-training framework that leverages CLIP priors as a foundational bridge to connect three modalities by synergizing spectral graph theory with transitive semantic learning. The method consists of two main sub-modules. First, Spectral Eigen-Modal Alignment is presented to construct cross-modal correlations within the intrinsic eigen-spectral space via Laplacian eigendecomposition. By aligning low-frequency geometric harmonics with CLIP-guided observation inference, it maintains structural consistency between 3D geometries and 2D projections. Second, a Transitive Spectral-Semantic Alignment method is developed to establish a 3D→2D→Text propagation chain across the CLIP priors bridge. It distills dense CLIP priors into 3D representations through transitive distillation, effectively mitigating the semantic chasm between geometry and language. Extensive experiments confirm that the proposed SpecBridge demonstrates state-of-the-art performance by overcoming challenges triggered by modality discrepancies.

Recently, diffusion models have been applied to reformulate the regressive portrait matting task as a gradual denoising-driven generative process, alleviating the inherent limitations of discriminative models via iterative error correction. However, due to the lack of high-quality ground-truth mattes, existing generative methods for portrait matting still suffer from severe detail degradation, especially in structurally delicate hair strands and pixels with ambiguous colors and context. To address these challenges, we propose OrienDiffusion, a diffusion-based framework tailored for fine-grained portrait matting. Specifically, we construct a Hair Orientation Field to conditionally guide the denoising process, constraining the generation toward authentic structural consistency in hair regions. Furthermore, instead of relying on overly encoded deep features or ambiguous original color cues as existing methods do, we develop a Pixel Color-anchored Local Structure Embedding approach, which models alpha matte transitions as local structural variations, achieving accurate alpha estimation in pixels with ambiguous colors or context. Extensive experiments on multiple portrait matting datasets demonstrate that OrienDiffusion achieves state-of-the-art performance in these matting challenging regions. Code is available at https://github.com/xtz0001/OrienDiffusion.

Weakly supervised Referring Expression Comprehension (WREC) aims to localize referred objects using only image-text pairs without box-level annotations. Existing one-stage methods predominantly rely on anchor-level alignment, which suffers from two fundamental limitations: (1) anchors represent local visual patches rather than holistic objects, and (2) they lack the capability to model inter-object relations. To address these issues, we propose OCAAD, an Object-Centric Alignment and Anchor Distillation framework. Our key insight is that different self-attention heads in DINOv2 naturally attend to distinct semantic regions, effectively capturing object-level information. Building on this, OCAAD introduces two complementary modules: (1) Anchor-Object Distillation Module (AODM), which bridges the semantic gap between anchors and objects by transferring object-level knowledge from attention heads to anchors via overlap-aware contrastive learning; and (2) Intra-Modal Relation Consistency (IRC), which explicitly models inter-object relations by enforcing the relational structure among linguistic entities to match that of their visual counterparts. Extensive experiments on RefCOCO, RefCOCO+, and RefCOCOg demonstrate that OCAAD achieves new state-of-the-art performance, validating the effectiveness of objectcentric alignment for WREC. Code is available at https://github.com/VILAN-Lab/OCAAD.

Federated learning (FL) enables privacy-preserving video human activity recognition without centralizing data. Despite its potential, FL remains fundamentally constrained by the high communication overhead incurred during training. Existing methods, such as model quantization, gradient sparsification and low-rank approximation, aim to mitigate this overhead but often at the cost of model expressivity, leading to reduced classification accuracy. In this work, we take a representation-centric perspective and show that for video-base activity recognition using pretrained embeddings, class-discriminative information lies in a very low-dimensional subspace, that preserves linear seperability between activity classes. Motivated by this observation, we propose Federated discriminative subspace discovery, a framework that collaboratively identify and operate within such a subspace. Instead of communicating full classifier updates, each client first projects its local video embeddings using a shared random projection matrix. It then computes the covariance of the projected embeddings and transmits this statistic to the server for aggregation. The server reconstructs a global covariance matrix, performs singular value decomposition, and selects the top-k eigenvectors that define the global class-discriminative subspace. Subsequent communication occurs only in this compact subspace. Extensive experiments on UCF101, HMDB51, and Toyota-SmartHome demonstrate that, learning and sharing the classifier in the discovered subspace, reduces communication costs by up to 61%, with a moderate drop in activity recognition accuracy compared to the full model.

Unlocking the spatial intelligence of multimodal large language models (MLLMs) is crucial for understanding and interacting with the 3D world. Prevailing approaches typically inject spatial priors via external tools, which impose significant inference overhead, or rely on latent feature distillation, which remains uninterpretable and lacks fine-grained geometric constraints. To address these issues, we propose SpatialSV, a framework designed to internalize robust 3D spatial awareness within MLLMs while simultaneously offering inherent interpretability. Deviating from passive feature imitation, SpatialSV employs task-oriented visual supervision, compelling the model to actively lift its 2D visual features into explicit 3D representations, including depth maps, camera poses, and point clouds. Crucially, this 2D-to-3D lifting process provides a transparent window into the model’s representations: the resulting 3D reconstructions serve as an intuitive proxy for visualizing and diagnosing the quality of the model’s intrinsic spatial knowledge. Extensive experiments across multiple benchmarks and models demonstrate the effectiveness of SpatialSV in enhancing and interpreting MLLMs’ spatial intelligence. Furthermore, the framework exhibits strong generalization in semi-supervised settings, validating its potential to leverage unlabeled visual data for scalable, interpretable spatial representation learning.

Unauthorized personalization based on diffusion models pose a severe and growing threat to digital privacy by enabling the unauthorized replication and exploitation of individual identities. Existing disrupting-based defenses primarily add invisible perturbations arbitrarily across the entire image space to disrupt the generation process. However, we reveal that these methods fundamentally overlook the spatio-temporal dynamics of the personalization process, resulting in inefficient optimization that fails to sufficiently disrupt the core identity encoding mechanism. To mitigate these limitations, we propose IdentityMask, a robust protection framework that shifts the paradigm from arbitrary confusion to precise, targeted feature corruption. By anchoring the perturbation on subject-specific semantics and prioritizing the most critical diffusion timesteps, our framework ensures the disruption is maximized precisely where the identity is encoded. Additionally, a novel manifold projection strategy is introduced to embed the adversarial signals into the intrinsic structure of the image, rendering the protection resilient against state-of-the-art purification. Extensive experiments across diverse datasets, personalization techniques, and defense settings demonstrate that IdentityMask consistently outperforms prior state-of-the-art approaches in both protection efficacy and robustness.

Estimating the 6D pose of novel objects without CAD models or video sequences remains a challenging problem. Recent works explore text-driven approaches to address this challenge in an open-vocabulary manner. However, these methods typically treat text embeddings as static priors, which lack the flexibility to adapt to specific visual scenes, thereby limiting reliable cross-view correspondences. In this paper, we propose a framework that modulates static textual embeddings into adaptable semantic guidance for RGB-D open-vocabulary 6D pose estimation. Specifically, we introduce a Text Feature Modulation (TFM) module that uses learnable queries to transform fixed text embeddings into adaptive semantic representations. These representations are explicitly calibrated by a Foreground-Grounded Semantic Alignment (FGSA) strategy, which suppresses background distractions via contrastive learning. Furthermore, we design a Semantic-Anchored Cross-view Fusion (SACF) module that aggregates image features from the reference and query views using modulated semantics as an invariant anchor. Finally, we predict object segmentation and establish semantic-aware correspondences, leveraging modulated textual semantics to filter mismatches for robust 6D pose estimation. Extensive experiments on standard benchmarks demonstrate that our method achieves superior performance.

Zero-shot multimodal anomaly detection is critical for identifying structural defects that are often invisible to traditional RGB images, particularly in scenarios lacking target domain samples. However, existing methodologies face two significant impediments: the prohibitive computational overhead caused by relying on multi-view rendering for depth processing, and the inability of static text prompts to adapt to fine-grained local anomalies. To address these challenges, a unified zero-shot multimodal anomaly detection framework GRASP is proposed. Firstly, a Frequency Domain Enhancement module is introduced to replace costly rendering with spectral transformations, directly synthesizing high-fidelity depth images for efficient utilization. Secondly, a Prompt-Conditioned Variational module is designed to bridge the semantic gap by grounding global textual descriptions into local visual nuances. Finally, a Dual Cross-Injection Alignment module is proposed to enable robust feature fusion for enhancing anomaly classification performance, while a Pyramid Anomaly Map Recalibration module further refines anomaly localization across multiple scales. Extensive experiments on MVTec 3D-AD and Eyecandies demonstrate that GRASP establishes a new state-of-the-art, yielding substantial improvements of 3.0 points in I-AUROC and 2.7 points in AUPRO compared to the SOTA method.

Point cloud completion seeks to recover geometrically consistent shapes from partial or sparse 3D observations. Although recent methods have achieved reasonable global shape reconstruction, they often rely on Euclidean proximity and overlook the intrinsic nonlinear geometric structure of point clouds, resulting in suboptimal geometric consistency and semantic ambiguity. In this paper, we present a manifold-aware point cloud completion framework that explicitly incorporates nonlinear geometry information throughout the feature learning pipeline. Our approach introduces two key modules: a Geodesic Distance Approximator (GDA), which estimates geodesic distances between points to capture the latent manifold topology, and a Manifold-Aware Feature Extractor (MAFE), which utilizes geodesic-based k-NN groupings and a geodesic-relational attention mechanism to guide the hierarchical feature extraction process. By integrating geodesic-aware relational attention, our method promotes semantic coherence and structural fidelity in the reconstructed point clouds. Extensive experiments on benchmark datasets demonstrate that our approach consistently outperforms state-of-the-art methods in reconstruction quality. The code has been available online https://anonymous.4open.science/r/Manifold-Aware-Point-Cloud-Completion--F380/README.md.

Large visual-language models demonstrate exceptional transferability in biomedical image analysis, yet their robustness remains challenging in generalization scenarios such as few-shot learning. This limitation stems from an overly entangled prompt space, causing instability in the conditional distribution, coupled with saliency bias that drives attention to dominate key regions, thereby restricting the coverage of visual representations. Furthermore, cross-modal fusion often struggles to extract critical information signals from redundant alignments. To address these structural bottlenecks, we propose TriCLIP, a decomposition-based framework designed to reconstruct the entire information flow across prompt conditioning, visual evidence extraction, and cross-modal fusion. Prompt Distribution Regularization enhances robustness to linguistic variations by optimizing class-conditional random context perturbations for continuous prompts. Feedback-driven Counterfactual Masking constructs a feedback-driven reverse saliency view that suppresses dominant evidence to promote complementary cue learning and mitigate attention bias. Finally, Selective Regulation Fusion regulates fusion selectivity through paired reinforcement-suppression principles, amplifying information correspondence while suppressing redundant channels. Extensive experiments on diverse biomedical benchmarks indicate that the proposed TriCLIP consistently outperforms existing methods.

Knowledge-based Visual Question Answering (KB-VQA) requires models to answer visual questions by reasoning over external knowledge beyond the given image. Existing approaches suffer from two main limitations. First, candidate knowledge is often retrieved in a single modality, either textual or visual, which prevents effective use of heterogeneous and complementary evidence. Second, current approaches typically employ fixed fusion weights, ignoring the varying importance of modalities for different queries. Consequently, irrelevant evidence is introduced while critical knowledge may be overlooked. To address these issues, we propose Dynamic Multi-Path Retrieval for KB-VQA (DMRAG). Our framework retrieves candidates through multiple retrieval paths that capture complementary visual and semantic cues. It then performs Question-Adaptive Gated Fusion (QGF) to balance contributions from different modalities according to the query’s information need. The fused candidates are further refined via multimodal rearrangement to support accurate answer generation. Experiments on the E-VQA and InfoSeek datasets show that DMRAG improves both retrieval recall and answer accuracy over prior methods, demonstrating its effectiveness for KB-VQA. Our code is available at https://github.com/qwqq335/DMRAG.

Domain Generalization (DG) for medical image segmentation is both highly challenging and critically important. However, existing medical DG methods largely overlook the issue of Catastrophic Forgetting (CF): Models often sacrifice their ability to retain source-domain knowledge while pursuing cross-domain robustness. This can directly threaten diagnostic safety in already-deployed clinical scenarios. To address this, we investigate data augmentation strategies and catastrophic forgetting for medical image DG segmentation. First, we propose a structure-guided style diffusion augmentation method. Constrained by anatomical structure consistency in the frequency domain, this method performs cross-domain diffusion on the amplitude spectrum, generating samples with more diverse and broader style coverage to better support domain generalization. Then, we design a collaborative learning network with a dual-branch interactive architecture (CoDG-Net), together with a novel learning bias-guided strategy that adaptively regulates knowledge transfer at both the layer level and the task level, thereby effectively mitigating catastrophic forgetting on the source domain. Experiments and ablation studies on single-source and multi-source medical DG benchmark datasets demonstrate that CoDG-Net not only outperforms existing state-of-the-art methods in target-domain segmentation performance, but also achieves a lower forgetting rate on the source-domain data. The code is available at: https://github.com/wangprocess/CoDG-Net.

Open‑vocabulary video instance segmentation (OV‑VIS) couples spatial‑temporal reasoning with language grounding, yet its adversarial robustness has remained unexplored. We present the Dual-Objective Triggers (DOT), the first transferable attack on OV-VIS that simultaneously exploits the vision–language coupling and temporal coherence. DOT deploys a Dual Semantic Perturbation Module that overlays two complementary triggers: a Semantic Suppression Trigger erases the alignment between the true object and the query, while a Plausible Replacement Trigger steers the tracker toward a phantom trajectory that is visually plausible and text‑consistent. To amplify cross‑model transferability without sacrificing perceptual fidelity, we introduce Phase‑Guided Adversarial Training, which injects perturbations primarily in the phase spectrum while blending amplitudes with clean references. Extensive experiments on four state‑of‑the‑art OV‑VIS implementations demonstrate that DOT reduces mAP by up to 69.3% and raises attack success rate by up to 98%, outperforming the strongest baselines by a factor of 1.6× on average, while maintaining a PSNR of 52.49 dB, thus exposing critical security vulnerabilities and laying a foundation for future research on robust and trustworthy vision–language systems.

We introduce the Visual Implicit Geometry Transformer (ViGT), an autonomous driving geometric model that estimates continuous 3D occupancy fields from surround-view camera rigs. ViGT represents a step towards foundational geometric models for autonomous driving, prioritizing scalability, architectural simplicity, and generalization across diverse sensor configurations. Our approach achieves this through a calibration-free architecture, enabling a single model to adapt to different sensor setups. Unlike general-purpose geometric foundational models that focus on pixel-aligned predictions, ViGT estimates a continuous 3D occupancy field in a bird’s-eye-view (BEV) addressing domain-specific requirements. ViGT naturally infers geometry from multiple camera views into a single metric coordinate frame, providing a common representation for multiple geometric tasks. Unlike most existing occupancy models, we adopt a self-supervised training procedure that leverages synchronized image-LiDAR pairs, eliminating the need for costly manual annotations. We validate the scalability and generalizability of our approach by training our model on a mixture of five large-scale autonomous driving datasets (NuScenes, Waymo, NuPlan, ONCE, and Argoverse) and achieving state-of-the-art performance on the pointmap estimation task, with the best average rank across all evaluated baselines. We further evaluate ViGT on the Occ3D-nuScenes benchmark, where ViGT achieves comparable performance with supervised methods.