Pedestrian Detection for Autonomous Vehicles Using Virtual-to-Real Augmentation

Annotated data are essential to the success of training deep neural models for autonomous driving. Practically, it is both expensive and time consuming to collect and annotate plenty of data driving cross the city. It is even harder to cover the corner cases involving rare pedestrian poses with occlusions. To tackle this problem, we propose a data augmentation method for pedestrian detection. First, we take a SMPL model to generate pedestrians with a variety of poses and textures. Second, we insert the rendered pedestrians into images with real traffic background according to vanishing point. Occlusions can also be set easily by taking use of segmentation annotations of original real dataset. We evaluated the effectiveness of the generated synthetic dataset using multiple detectors. Results of three experiments shown that detectors trained with mixture of synthetic and real data outperform those trained on only real datasets.