This repository contains a novel self-attention based person re-identification architecture. The proposed architecture outperformed CNN based re-id architectures when evaluated on three public re-id benchmarks i.e. Market1501, DukeMTMC-ReID and MSMT17. The complete code along with pretrained models is shared via this repository.
Install torchreid. Person Re-id datasets can be downloaded from the links given below. Extract the datasets and place in the respective sub-folders in the "reid-data" folder.
Market 1501 dataset is an open-access dataset and can be downloaded from Openlink
DukeMTMC-reID dataset is an open-access dataset, the dataset can be downloaded via GoogleDriver and BaiduYun, the download links are shared by the authors of DukeMTMC-ReID dataset via their github repository. The authors of the dataset are very generous to handle the queries regarding dataset download links as mentioned here.
MSMT17 dataset can be downloaded by following the detailed instructions available here
Performer(base) and Performer(SCM) training models can be downloaded from GoogleDriver
Place the trained weights in the trained-models folder and update the paths in the evaluation script.
Evaluate the models using command: python eval_performer.py
If you find this code useful to your research, please cite the following paper.
@article {PER-FORMER: Rethinking Person Re-identification Using Transformer Augmented with Self-Attention and Contextual Mapping, Under Review in TVCJ}