A novel view synthesis approach based on view space covering for gait recognition.

Gait recognition has proven to be effective for long-distance human recognition. View angle, one form of the gait variations, can change the human appearance greatly and reduce its performance. For most existing gait datasets, the angle interval between the two nearest views is large. This means that the angle does not cover the entire view space and prevent better view-invariant feature extraction for CNN. Additionally, the angles between cameras and people vary widely in typical camera deployments for monitoring people. In this paper, we, therefore, propose a novel view synthesis approach based on view space covering to deal with the challenge of large-angle interval. Specifically, a Dense-View GEIs Set (DV-GEIs) is introduced to expand this view approach, from 0° to 180° with 1° interval. GEI is a popular feature representation for gait, which can be obtained by aligning human silhouettes and averaging them in a gait cycle. In order to synthesize DV-GEIs set, Dense-View GAN (DV-GAN) is proposed to model the gait attribute distribution and generate new GEIs with various views. DV-GAN consists of a generator, discriminator, and monitor, where the monitor is designed to preserve human identification and view information. Compared with our previous work DV-GAN-pre, we add a center for each object in the monitor to improve the discriminative capability of synthesized images during the modeling process. The proposed method is evaluated on the CASIA-B and OU-ISIR dataset. The experimental results show that view space covering is an effective way to light the burden of view-invariant feature extraction for CNN and make the feature more discriminative. We believe the idea of view space covering will further improve the development of gait recognition.