Application of Synthetic Data on Object Detection Tasks

摘要

Object detection is a computer vision task that identifies and locates one or more effective targets from image or video data. The accuracy of object detection heavily depends on the size and the diversity of the utilized dataset. However, preparing and labeling an adequate dataset to guarantee a high level of reliability can be time-consuming and labor-intensive, because the process of building data requires manually setting up the environment and capturing the dataset while keeping its variety in scenarios. There have been several efforts on object detection that take a long time to prepare the input data for training the models. To deal with this problem, synthetic data have emerged as a potential for the replacement of real-world data in data preparation for model training. In this paper, we provide a technique that can generate an enormous synthetic dataset with little human labor. Concretely, we have simulated the environment by applying the pyBullet library and capturing various types of input images. In order to examine its performance on the training model, we integrated a YOLOv5 object detection model to investigate the dataset. The output of the conducted model was deployed in a simulation robot system to examine its potential. YOLOv5 can reach a high accuracy of object detection at 93.1% mAP when solely training on our generated data. Our research provides a novelistic method to facilitate the understanding of data generation process in preparing datasets for deep learning models.