Reinforcing Pedestrian Parsing on Small Scale Dataset

Abstract

In this paper we address the problem of automatic pedestrian parsing in surveillance video with only a small number of training samples. Although human parsing has achieved great success with high-capacity models, it is still quite challenging to parse pedestrians in practical surveillance conditions because complicated environmental interferences need more pixel-level training samples to fit. But creating large datasets with pixel-level labels has been extremely costly due to the vast amount of human effort required. Our method is developed to capture the pedestrian information from the non-labeled datasets to update the trained model by reinforcement learning, which achieves elegant performance with only much fewer pixel-level labeled samples. Both quantitative and qualitative experiments conducted on practical surveillance datasets have shown the effectiveness of the proposed method.