人の骨格情報と物体情報を用いたインタラクション認識手法の提案

抄録

To understand human-centric images and videos, detecting human-object interactions is an important probrem. Motion recognition based on skeleton information has been studied, but it is not possible to accurately estimate target object using only skeleton information. Therefore, we propose a method for detecting human-object interactions using the skeleton information in conjuction with the object recognition information. We define object-specific interaction by this method. Our hypothesis is that by training human-object interactions using the same recognition method as a human, sufficient recognition accuracy can be obtained even with a small dataset. Since we use only coordinates for training, robust recognition is possible in various situations.