S2Gait: RGB-based Gait Recognition with Style Feature Sampling Data Augmentation

Gait is unique to individuals and can be acquired from a distance, making it difficult to disguise. Gait videos also contain many elements unrelated to gait, which make gait recognition challenging. Departing from common approaches that use preprocessing such as silhouette extraction, the RGB-based method extracts gait features directly from RGB gait videos. RGB-based methods leverage the difference between two inputs with different attributes to separate gait-related/unrelated features, but their separation performance depends on the diversity of the dataset. To increase the amount and diversity of training data, we focus on the latent space of gait-independent features (style features), which are usually not needed for gait recognition. In this paper, we propose S2Gait (Style feature Sampling Gait), which augments the training data online with images generated from gait-dependent features of the input images and sampled style features. Experiments demonstrate that the proposed method surpasses existing RGB-based methods on almost all metrics for both generated image quality and identification accuracy. We also explore the relationship between the amount of data augmentation and performance taking advantage of our method’s flexibility to generate a wide variety of gait images.