Pedestrian Image Generation with Target Pose Based on the Improved Cyclegan

Person image generation mainly contains two parts: The first part is to learn people's appearance information, and the second part is to learn people's posture information. this paper is to generate pedestrian images by using the improved Cycle-Consistent Adversarial Networks (CycleGAN[2]). The network uses a loop structure consisting of two generators G1, G2 and two discriminators D1, D2. In order to preserve the appearance and target pose information of the original pedestrian, we used paired images of the same pedestrian during training and a jump connection is used in the generators G1, G2. The experimental results show that the network structure can generate pedestrian images with original appearance feature information and target pose information.