A NOVEL SEMI-SUPERVISED DETECTION APPROACH WITH WEAK ANNOTATION

Abstract

In this work we propose a semi-supervised learning approach for object detection where we use detections from a preexisting detector to train a new detector. We differ from previous works by coming up with a relative quality metric which involves simpler labeling and by proposing a full framework of automatic generation of improved detectors. To validate our method, we collected a comprehensive dataset of more than two thousand hours of streaming from public traffic cameras that contemplates variations in time, location and weather. We used these data to generate and assess with weak labeling a car detector that outperforms popular detectors on hard situations such as rainy weather and low resolution images. Experimental results are reported, thus corroborating the relevance of the proposed approach.