A comparative analysis of deep Re-Identification models for matching pairs of identities

Details

Supervisors

De Vleeschouwer, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées

Abstract

This thesis offers a comparative analysis of deep learning models for Pair Re-Identification. Pair Re-Identification is the task of answering the question "Is this pair of images, each containing a person, from the same person identity ?" This problem differs from usual problems in person Re-Identification where one deals with "finding, in a large gallery of images, the n closest images to a query image". Models for Re-Identification (Re-ID) handle their task by mapping images into a feature space where each image is represented by an embedding. The distances between the embeddings can then be used as a measure of similarity for Pair Re-ID or to sort the gallery for the usual problem. The deep learning models implemented for the purposes of this thesis are Convolutional Neural Networks. They draw from existing state of the art models for Re-ID which differ mainly in the supervision with which they are trained. Our models are supervised either by classification or by metric learning or by a combination of both either in succession or simultaneously. Hence the comparative analysis focuses on comparing and contrasting the effects of the supervision on the feature space learned by the models. More precisely on the distribution of distances between embeddings of the same identity and on the distribution of distances between embeddings of different identities. This type of analysis can not be found in the literature.