Multi-scale multi-patch person re-identification with exclusivity regularized softmax

Discriminative feature learning is critical for person re-identification. To obtain abundant visual information from the input person image, we first propose a novel network that extracts multi-scale patch-level deep features. Then, we propose an improved softmax loss function for learning more compact and more discriminative feature vectors. Specifically, we integrate feature pyramid blocks and region-level global average pooling functions into the feature extraction network, introduce the well-established normalization techniques in face recognition algorithms into person re-ID, and penalize the redundancy in feature vectors by minimizing the l1,2 norm of the weight matrix in the softmax layer. Experiments on three large-scale datasets under the standard settings show the effectiveness of the proposed method. Moreover, we report our cross-domain re-ID results by training re-ID models on source datasets and testing them on other target datasets.