HTS: High-Quality Training Set Selection for Pedestrian Detection
Junjie Li, Kai Shuang, Wentao Zhang · 2021
Training a pedestrian body detector generally involves the supervision of body bounding-boxes. We find in some scenarios the supervision is not clear due to mixing background or overlapping persons. Such informative but noisy supervision imposes a challenge to learning a well-performed detector. In this work, we resolve this problem from the perspective of data selection and propose a learning-based high-quality training set selection strategy (HTS). Our strategy aims at selecting (preprocessing) a training set with best trade-off between useful information and noise from the original training set. Our strategy is composed of three steps. First we design a metric trade-off quality to evaluate the trade-off in a training set. Then we search a training set that achieves best trade-off quality based on parameterizing the training set. Finally the learned parameters are applied to the original training set by a novel re-weighting scheme. Due to improved handling of the noisy data, our proposed strategy shows consistent superiority over the typical data selection strategy for different pedestrian detectors on different datasets (Caltech and CityPersons datasets) without any specifically designed network architecture or bells and whistles.