Advanced Pedestrian Dataset Augmentation for Autonomous Driving

Abstract

Having the ability of generating people images in arbitrary, yet admissible, pose is a crucial prerequisite for Autonomous Driving applications. Firstly, because the existing datasets are quite limited in the human pose variation and appearance. Secondly, because the strict safety requirements call for the ability of validation on rare situations. Generating realistically looking people images is very challenging problem due to various transformations of individual body parts [2,6] self occlusions etc. We propose a novel approach for person image generation. Our approach allows generating people images in a required pose, indicated by specific pose keypoints and deals with occlusions. We build on top of the recent prevailing success of Generative Adversarial Networks [10]. Our contributions comprise of the networks architecture, as well as the novel loss terms specifically designed to generate visually appealing pedestrians fitting the surrounding environment well.