Investigating the Impact of Illumination and Viewpoint Variations on Transformer-based Person Re-Identification

Abstract

Variations in visual factors such as illumination, viewpoint, resolution, background, pose, and so on are commonly regarded as significant issues in object re-identification (re-ID). Despite widespread recognition of their importance in determining the performance of an object re-ID model, not enough attention is paid to how these factors affect re-ID systems. One of the major impediments to investigating how these factors affect the performance of re-ID models is the lack of datasets with unbiased distribution of these difficult visual conditions. To make up for the lack of large-scale datasets with a balanced distribution of such photometric and geometric transforms, recent studies suggest using game engines to generate synthetic datasets. This study proposes a quantitative investigation of the impact of two critical visual factors: illumination and Tranfomer-based re-ID models on synthetic dataset.