01644nas a2200265   4500000000100000000000100001008004100002260001200043653002000055653001800075653003100093653000800124653002500132100001200157700001400169700001500183700001700198700001300215245007500228856008100303300000900384490001300393520095800406022001401364       9998                            d  c06/202310aComputer vision10aDeep Learning10aGated Graph Neural Network10aHOI10aImage Classification1 aZhan Su1 aRuiyun Yu1 aShihao Zou1 aBingyang Guo1 aLi Cheng00aSpatial-Aware Multi-Level Parsing Network for Human-Object Interaction  uhttps://www.ijimai.org/journal/sites/default/files/2023-06/ip2023_06_004.pdf  a1-100 vIn Press3 aHuman-Object Interaction (HOI) detection focuses on human-centered visual relationship detection, which is a challenging task due to the complexity and diversity of image content. Unlike most recent HOI detection works that only rely on paired instance-level information in the union range, our proposed Spatial-aware Multilevel Parsing Network (SMPNet) uses a multi-level information detection strategy, including instance-level visual features of detected human-object pair, part-level related features of the human body, and scene-level features extracted by the graph neural network. After fusing the three levels of features, the HOI relationship is predicted. We validate our method on two public datasets, V-COCO and HICO-DET. Compared with prior works, our proposed method achieves the state-of-the-art results on both datasets in terms of mAProle, which demonstrates the effectiveness of our proposed multi-level information detection strategy.  a1989-1660