A Generative-Adversarial Network-Based Method for Image Synthesis of Diverse Pedestrian
Bo Li, Zhenyuan Liu, Xingyu Xing, Tong Jia, Yu-Xiao Lu, Junyi Chen · 2023
The training and testing of visual recognition algorithms require a large number of target images, but diverse targets in the existing data set are insufficient. Besides, edge cases, such as pedestrian-obscured scenarios, are scarce. In this paper, we propose a method of diverse target generation based on generative adversarial network. The input data are processed by the location module and the shape module, which realize the generation of targets with diverse, realistic shapes at suitable new locations and the generation of obscured targets. By designing two parallel paths, the unsupervised path and the supervised path, the collapse problem is effectively solved. YOLO is selected to test the performance of the generated targets. The results show that the synthesized images are close to the original images and can meet the requirements of augmenting the data set.