Neural Architectures for Feature Embedding in Person Re-Identification: a comparative view

Abstract

Solving Person Re-Identification (Re-Id) through Deep Convolutional Neural Networks is a daunting challenge, due to the small size and variety of the training data, especially in Single-Shot Re-Id, where only two images per person are available. The lack of training data causes the overfitting of the deep neural models, leading to degenerated performance. This article explores a wide assortment of neural architectures that have been commonly used for object classification and analyses their suitability in a Re-Id model. These architectures have been trained through a Triplet Model, and evaluated over two challenging Single-Shot Re-Id datasets, PRID2011 and CUHK. This comparative study is aimed at obtaining the best-performing architectures, and some concluding guidance to optimize the features embedding for the Re-Identification task. The obtained results present Inception-Resnet and DenseNet as potentially useful models, especially when compared with other methods, specifically designed for solving Re-Id.