Attention-Based Network for Cross-View Gait Recognition

摘要

Existing gait recognition approaches based on CNN (Convolutional Neural Network) extract features from different human parts indiscriminately, without consideration of spatial heterogeneity. This may cause a loss of discriminative information for gait recognition, since different human parts vary in shape, movement constraints and so on. In this work, we devise an attention-based embedding network to address this problem. The attention module incorporated in our network assigns different saliency weights to different parts in feature maps at pixel level. The embedding network strives to embed gait features into low-dimensional latent space such that similarities can be simply measured by Euclidian distance. To achieve this goal, a combination of contrastive loss and triplet loss is utilized for training. Experiments demonstrate that our proposed network prevails over the state-of-the-art works on both OULP and MVLP dataset under cross-view conditions. Notably, we achieve 6.4% rank-1 recognition accuracy improvement under 90° angular difference on MVLP and 3.6% under 30° angular difference on OULP.