Video Anomaly Detection aims to identify video segments containing abnormal events. However, detecting anomalies relies more heavily on temporal modeling, particularly when anomalies exhibit only subtle deviations from normal events. However, most existing methods inadequately model the heterogeneity in spatiotemporal relationships, especially the dynamic interactions between human pose and video appearance. To address this, we propose GSMTNet, a dual-stream heterogeneous unsupervised network integrating gated spatio-temporal graph convolution and multi-scale temporal learning. First, we introduce a dynamic graph structure learning module, which leverages gated spatio-temporal graph convolutions with manifold transformations to model latent spatial relationships via human pose graphs. This is coupled with a normalizing flow-based density estimation module to model the probability distribution of normal samples in a latent space. Second, we design a hybrid dilated temporal module that employs multi-scale temporal feature learning to simultaneously capture long- and short-term dependencies, thereby enhancing the separability between normal patterns and potential deviations. Finally, we propose a dual-stream fusion module to hierarchically integrate features learned from pose graphs and raw video sequences, followed by a prediction head that computes anomaly scores from the fused features. Extensive experiments demonstrate state-of-the-art performance, achieving 86.81% AUC on ShanghaiTech and 70.43% on UBnormal, outperforming existing methods in rare anomaly scenarios.

Unsupervised video anomaly detection (VAD) aims to identify unusual events by learning from unlabeled videos. However, many current methods overlook the fine-grained spatiotemporal dynamics of human poses, which are crucial for detecting localized anomalies like falls or assaults. Prevailing methods that rely on raw RGB frames are often susceptible to variations in lighting and background and struggle to capture the precise structural relationships of human bodies over time. To bridge this gap, we propose UHPose-VAD, a novel unsupervised framework that integrates human pose dynamics with normalizing flow within a graph-based probabilistic model to capture anomalies through spatiotemporal Gaussian distributions. Our framework first extracts human pose keypoints and normalizing flow features. These are then modeled by a graph convolutional network that adaptively learns the graph connectivity, effectively mapping the data to a latent space. This approach allows the model to explicitly reason about the spatiotemporal relationships between body joints, making it inherently more robust and interpretable for human-centric anomaly detection. Finally, a Gaussian Mixture Model fits the latent features of normal training data, learning the intrinsic manifold of regular motion patterns. Extensive experiments on ShanghaiTech and UBnormal datasets show that UHPose-VAD achieves state-of-the-art performance among unsupervised methods, with AUC scores of 86.1% and 69.4%, respectively.

Multimodal Emotion Recognition in Conversation (MERC) aims to infer speakers’ affective states from heterogeneous cues within evolving dialogue contexts. Although graph-based methods have shown strong potential in modeling conversational structures, they often introduce history through dense or predefined connections and obtain unified multimodal representations before prediction. This paradigm may propagate irrelevant contextual information, overlook the different roles of same-speaker and cross-speaker dependencies, and underexploit useful corrective cues carried by inconsistent modalities. To address these limitations, we propose HiRoC, a multimodal emotion recognition framework that selectively routes emotion-relevant history and calibrates predictions using cross-modal disagreement. Specifically, a Historical Relevance Filter (HRF) scores how relevant each past utterance is to the current one and suppresses weakly related context before propagation. Guided by the resulting dependency prior, a Speaker-Typed Routing Graph (STRG) keeps intra- and inter-speaker dependencies apart and propagates the retained history along sparse, temporally directed paths. In parallel, the filtered modality-specific representations feed a hypergraph backbone for high-order utterance- and modality-level interactions, which the routing branch complements with speaker-aware temporal dependencies. A Cross-Modal Residual Correction (CMRC) module then uses modal disagreement to adjust the main prediction at the decision level, so that conflicting cues are not averaged away. On IEMOCAP and MELD, HiRoC outperforms representative state-of-the-art methods.