Using deep multiple instance learning for action recognition in still images
Cagdas Bas, Cemil Zalluhoğlu, Nazlı İkizler-Cinbiş · 2017
Recognizing actions from still images just by analyzing human appearance is a changeling computer vision problem. In this problem, objects in the scene and scene itself can help to improve recognition. In this paper we propose to extract object like windows from still images and use this possible object areas with Multiple Instance Learning framework to learn automated object recognition task. It is important to incorporate Multiple Instance Learning with Deep learning methods which are very successful at classifying images, for leveraging semi (weakly) supervised human action recognition task. Proposed method is evaluated with Stanford 40, the important benchmark in still image action dataset, and the result showed that multiple instance learning method improved the accuracy of deep learning methods.