Pose and Color-Gamut Guided Generative Adversarial Network for Pedestrian Image Synthesis.

Tremendous transfer requirements in pedestrian reidentification (Re-ID) tasks have greatly promoted the remarkable success in pedestrian image synthesis, to relieve the inconsistency in poses and lighting. However, existing approaches are confined to transferring in a particular domain and are difficult to combine, since pose and color variables locate in two independent domains. To facilitate the research toward conquering this issue, we propose a pose and color-gamut guided generative adversarial network (PC-GAN) that performs joint-domain pedestrian image synthesis conditioned on certain pose and color-gamut through a delicate supervision design. The generator of the network comprises a sequence of cross-domain conversion subnets, where the local displacement estimator, color-gamut transformer, and pose transporter coordinate their learning pace to progressively synthesize images in desired pose and color-gamut. Ablation studies have demonstrated the efficacy and efficiency of the proposed network both qualitatively and quantitatively on Market-1501 and DukeMTMC. Furthermore, the proposed architecture can generate training images for person Re-ID, alleviating the data insufficiency problem.