Effective actor-centric human-object interaction detection

While Human-Object Interaction(HOI) Detection has achieved tremendous advances in recent, it still remains challenging due to complex interactions with multiple humans and objects occurring images, which would inevitably lead ambiguities. Most existing methods either generate all human-object pair candidates infer their relationships by cropped local features successively a two-stage manner, or directly predict interaction points one-stage procedure. However, the lack of spatial configurations reasoning steps two- one- stage respectively limits performance such scenes. To avoid this ambiguity, we propose novel actor-centric framework. The main ideas are that when inferring interactions: 1) non-local entire image guided actor position obtained model relationship between context, then 2) use an object branch pixel-wise area prediction, where denotes central area. Moreover, also get prediction composition strategy based on center-point indexing final HOI prediction. Thanks usage partly-coupled property human-objects strategy, our proposed framework can detect more accurately especially for images. Extensive experimental results show method achieves state-of-the-art V-COCO HICO-DET benchmarks is robust persons and/or