SynPoses: Generating Virtual Dataset for Pedestrian Detection in Corner Cases

Abstract

Pedestrian detection based on deep learning methods makes a big hit during these days. The key to achieve excellent results for deep learning-based pedestrian detection methods is a high-quality dataset. There are lots of outstanding datasets for pedestrian detection like EuroCity Persons, CityPersons, Caltech, etc. However, their efforts are not dedicated to the traffic scenarios with pedestrians in various poses. When the state-of-the-art detectors trained on datasets described above encounter the scene with pedestrians in variously complicated poses, the results are mostly worse than expected. In order to alleviate this problem, we propose a framework called SynPoses to create synthetic human with complicated poses in high quality. Experimental results show that, when facing scenarios with human in diverse poses, the performance of detectors trained on augmented dataset outperforms those trained on original dataset.