HOOI Detection: Cascade-Clue Integrated Modeling over Multiple Temporal Segments

作者：Mingxuan Zhang, Qi He, Zhaoquan Yuan, Tingquan He, Rong Li · 年份：2025 · DOI：10.1145/3731715.3733360 · 被引用次数：1 · 研究领域：Anomaly Detection Techniques and Applications、Time Series Analysis and Forecasting、Machine Learning and Data Classification

To fully comprehend a visual scene, recognizing and localizing interaction actions are essential components. Recently, significant advances have been made in detecting human-object interaction actions, which aim to capture pairwise relations between entities in the scene. Although these methods have made significant progress, they ignore the human-object-object interaction (HOOI) actions that frequently occur between a human and two objects in the real world. To advance related research, a new task named HOOI detection is introduced. It aims to accurately localize the humans in each video frame and identify the HOOI actions they perform. For this purpose, two novel HOOI datasets oriented to industrial production and daily life are constructed. These new datasets provide essential data support for in-depth research of HOOI detection. Furthermore, a cutting-edge method named Cascade-Clue Integrated Modeling over Multiple Temporal Segments (C2TS) is proposed for effectively detecting HOOI actions. Specifically, considering the phased characteristics of the action, C2TS comprehensively considers the HOOI information in the preceding, neighborhood, and subsequent temporal segments. For each temporal segment, the Cascaded Modeling and Clue Augmentation methods are applied to extract the corresponding HOOI features. The final detection result is obtained by classifying the aggregated HOOI features from the three temporal segments. Experiments conducted on the two proposed HOOI-relat...