Transformer-based domain generalization of person re-identification

Abstract

Person re-identification (Re-ID), is a technology that uses computer vision technology to determine whether there is a specific pedestrian in an image or video sequence. The purpose of domain generalizable (DG) person Re-ID is to train a robust person Re-ID model with great generalizability that can achieve relatively high accuracy on unseen datasets. Although some CNN-based models achieve high accuracy on cross-domain evaluations, there is still a lot of room for improvement.

The original Transformer [11] has been widely used in natural language processing area since 2017. It uses self-attention mechanism to update the embedding. In computer vision, some methods using Transformer are proposed to solve the long-range correlation extraction problem. However, the matching ability of transformer-based DG Re-ID has not been studied yet.

This dissertation proposed a pipeline with a CNN-based backbone feature extractor and a Transformer-based encoder-decoder module to solve the domain generalization problem of person re-identification. Some pre-processing and post-processing techniques are used to achieve higher accuracy such as reranking, BNNeck and temporal lift. The ablation studies of parameters in these modules are employed. The result analysis and future prospects are discussed.