GAN Based Method for Labeled Image Augmentation in Autonomous Driving

Abstract

Deep learning models in Autonomous Driving perception tasks commonly use supervised learning methods and thus highly depend on labeled data. Training with more labeled data tends to bring better results, which highlights the meaning of data augmentation. Currently there are two difficulties when doing data augmentation. Firstly, it is time consuming to manually label the collected raw data. The second issue is that the diversity of a dataset is limited by the collection environment and time. In this paper, we proposed to use the current state of the art Multimodal Unsupervised Image-to-Image Translation (MUNIT) to generate synthesized images from labeled data. One of the benefits is that the generated data are automated labeled since they share the same ground truth with the raw data. Then we used the augmentation dataset to do different tasks including drivable area detection and object detection to prove that the data could be used to improve the performance of convolution neural networks (CNNs). We also designed an auto labelling tool that people could do labelling with the help of the improved CNN. The whole process is like a close loop that finishes labelling tasks while making progresses by itself. Generally speaking, our approach introduces an auto labelling pipeline based on unsupervised image-to-image translation to increase the amount and diversity of labeled data.