Semantic-aware multi-view person image generation for re-identification

Synthesizing realistic person images that preserve specific characteristics is an effective and promising approach to enhance input diversity for the re-identification (ReID) task. However, existing person image generative methods have not adequately taken into account the distribution variations of different parts of the human body across different viewpoints, resulting in the generation of unrealistic images with considerably incorrect intra-class variation. To this end, we propose a semantic-aware generative adversarial framework for high-fidelity multi-view person image generation, which provides region-level fine-grained information for person ReID. The proposed framework involves an innovative structure encoder integrating 3D information and regional parts of the person in the input sample, a group of semantic-aware generators creating intrinsic features and adaptive region weights to guide rendering, and a dual-branch generative adversarial module that synthesizes segmentation masks as supplementary supervisory information for local details of each region. Experimental results on several person ReID datasets demonstrate that the proposed method generates more realistic images with superior diversity and appearance consistency compared to existing generative models. Moreover, the proposed method significantly improves the performance of ReID methods, contributing to state-of-the-art performance.