Color Transfer to Anonymized Gait Images While Maintaining Anonymization

Gait anonymization helps prevent the identification of people by gait recognition systems using videos uploaded to social media. Our current gait anonymization approach is to first modify the silhouette of the gait sequence and then transfer the colors of the skin, hair, clothing, etc. in the original gait images to the modified gait images to produce a final RGB anonymized gait image sequence. Since users typically care about the quality of the generated videos as they want to share them with family and friends in addition to caring about privacy, the generated videos should contain color images that are sharp and finely textured. Existing anonymization models are unable to produce such images. In this paper, we focus on color transfer while maintaining anonymization. This is challenging because the original gait images may consist of multiple colors, the modified gait differs from the original one, there is no real ground truth anonymized gait, and it may be difficult to exactly separate the foreground colors from the background colors in the original gait images. To overcome this problem, we propose transferring the colors without using ground truth and without extracting the colors in the original gait images. In this model, the overlapping region between the two gaits is first located, and the colors in that region in the original images are transferred to the modified images. The colors in the remaining region are interpolated from the color in the overlapping region, so the colors in the overlapping and non-overlapping regions are coherent. Quantitative and qualitative experiments demonstrated that the proposed model is more effective than our previous models with no reduction in anonymization.