Authors:
María José Gómez-Silva
;
José María Armingol
and
Arturo de la Escalera
Affiliation:
Universidad Carlos III de Madrid, Spain
Keyword(s):
Re-identification, Deep Learning, Siamese Network, Contrastive Loss Function.
Related
Ontology
Subjects/Areas/Topics:
Computer Vision, Visualization and Computer Graphics
;
Motion, Tracking and Stereo Vision
;
Video Surveillance and Event Detection
Abstract:
The selection of discriminative features that properly define a person appearance is one of the current challenges for person re-identification. This paper presents a three-dimensional representation to compare person images, which is based on the similarity, independently measured for the head, upper body, and legs from two images. Three deep Siamese neural networks have been implemented to automatically find salient features for each body part. One of the main problems in the learning of features for re-identification is the presence of intra-class variations and inter-class ambiguities. This paper proposes a novel normalized double-margin-based contrastive loss function for the training of Siamese networks, which not only improves the robustness of the learned features against the mentioned problems but also reduce the training time. A comparative evaluation over the challenging PRID 2011 dataset has been conducted, resulting in a remarkable enhancement of the single-shot re-ident
ification performance thanks to the use of our descriptor based on deeply learned features in comparison with the employment of low-level features. The obtained results also show the improvements generated by our normalized double-margin-based function with respect to the traditional contrastive loss function.
(More)