Image Synthesis Techniques for Improving Trash Object Detection Performance of Self-Driving Road Cleaning Vehicles

Object detection using deep learning has been actively studied in the computer vision task and is used in various tasks. In order to utilize the deep learning object detection model in various tasks, not only the quantity and quality of training data but also diversity has become an important factor. However, if the data is rare because the task which utilizes the object detection model is special, it may not be possible to collect the large amount of data required for the model’s training. In this paper, we study a data augmentation technique that synthesizes images to learn an object recognition model in a special environment called an autonomous road cleaner that recognizes and cleans trash on the road. Furthermore, we propose a method of applying perspective transformations to synthesize more realistic data and analyze the impact on object detection model through experiments.