Deep Visual Feature Learning for Person Re-identification

Résumé

In recent years the use of surveillance cameras in the city increased dramatically. There is thus a big demand for person re-identification (Re-ID) algorithms. The goal of Person re-identification is to find the target person in other non-overlapping camera views, which represents a real challenge in practical applications. In this thesis, we present a research carried out on person re-identification which aims to propose new models that are able to learn robust and complementary visual features. To this end, we focused on exploiting complementary information learned from local and global features. The first proposed approach is based on a multi-stream model combining the original image and multiple person body parts to capture a complementary information. Since the first approach required many processing times such as segmentation and individual training for each stream, we proposed a second model working in an end to end fashion. This model is based on the attention mechanism which allow the model to selectively focus on the most relevant parts of the input image. In addition, we proposed the use of a feature dropping technique that push the model to learn from some less relevant regions which can improve the final prediction. The experimental results conducted on several public datasets show that the proposed approaches outperform the sate-of-the-art methods which proves their effectiveness.

Citer ce document

Accès au document

Ce lien n'est plus accessible actuellement. Contactez l'institution d'origine.

Auteur(s)

Statistiques

Consultations : 2

Téléchargements : 0