HUMAN ACTION RECOGNITION AND DETECTION IN HUMAN-CENTRIC SCENES

Abstract

This thesis studies the problem of human action recognition and detection in single images. The goal is to recognize human actions, and localize the actions between the interactive humans and objects in the visual scenes. In the first work, we introduce a human-mask loss to automatically guide the activations of the feature maps to the target human who is performing the action, and hence suppress the activations of misleading contexts. In the second work, we propose to use a spatial enhancement approach and exclusive object prior for Human-Object Interaction (HOI) detection. In the following parts, we address the problem of learning with less supervision. In the third work, we propose a certainty-driven consistency loss for the general semi-supervised image classification task. In the fourth work, we propose a semi-supervised approach for HOI detection using data distilled pseudo labels and consistency regularization. Extensive experiments demonstrate the effectiveness of our proposed approaches.