Long-Term Person Re-identification Model with a Strong Feature Extractor
Adham Gamal, Nadeen Shoukry, Mohammed A.‐M. Salem · 2021
Person Re-Identification is a challenging task in the field of computer vision, it has gained a huge space due to its importance in our life. Person Re-ID is mainly utilized in security applications, it is used everywhere like airports, metro stations, banks, etc. Due to its crucial importance many approaches were proposed to perform such task, the explosion of deep learning and its ability to outperform most of the state of-the-art approaches motivated researchers to use it in person re-identification, moreover researchers introduced benchmark person re-id datasets to train these models, which were able to provide remarkable results and provide a solid benchmark for comparison. The only limitation to all these models and datasets is that they were built assuming no pedestrian clothes variations. In our work we aim to address the problem of Long Term Person Re-Identification where the clothes variation exists. We start by experimenting and comparing the most recent state-of-the-art deep learning models that were developed for that task such as OSNet and PCB. After that we introduce our proposal to replace the backbone network of PCB with OSNet to act as a feature extractor. All of these models were trained and tested using NKUP dataset and Market-1501. The benchmark dataset Market-1501 was used to compare our model’s performance with several standard models. Our model was able to outperform the" PCB" model on Market-1501 and improve the accuracy from 87.71% to 91.3%. On the NKUP dataset our model improved the accuracy of" PCB" model from 16.1% to 17.6%.