Balancing People Re-Identification Data for Deep Parts Similarity Learning

The learning of a distance metric, to measure the degree of appearance similarity between a pair of person images, becomes a remarkable challenge in person re-identification due to the unbalanced nature of the datasets, where the number of instances of a query identity is very limited against the vast quantity of different people representations, among which it must be identified. This article presents two network models, a Siamese and a Triplet one, which exploit the multiple possible combinations of training samples in pairs and triplets, respectively. Both models have been used to learn features for nine different body parts, which has been simultaneously analyzed to embed the inter-view variations in a Mahalanobis distance. The influence of the model and training data in the features learning has been evaluated through several tests over the challenging PRID2011 dataset, besides of the proposed system re-identification performance in comparison with other state-of-the-art approaches.

Balancing People Re-Identification Data for Deep Parts Similarity Learning Articles