Structure-Regularized Attention
Code for paper "Structure-Regularized Attention for Deformable Object Representation".
The code contains the ResNet50-StRA network for person re-identification task and the corresponding configurations to reproduce our results.
Requirements
To install requirements:
pip install -r requirements.txt
Training
The reported results are trained and evaluated based on the existing person ReID framework: https://github.com/KaiyangZhou/deep-person-reid.
To reproduce the reported results, please pull from the above deep-person-reid repository (version 0.9.1)
and download the datasets following the instructions. Then integrate with our network code:
mv models ./deep-person-reid/torchreid
To train our StRA-ResNet50 on Market-1501 dataset:
python train.py
To train on different datasets, change different dataset sources in train.py.
Evaluation
Simply running the following code will give the reported results on Market-1501 dataset:
python eval.py
Results
By downloading our pre-trained model and running the evaluation code, the followng results will be obtained. And the result reported in the paper is the mean of 3 runs.
| Dataset | mAP | rank1 | rank5 | rank10 |
|---|---|---|---|---|
| Market1501 | 84.2% | 94.0% | 97.6% | 98.5% |
Pre-trained Model
The checkpoint model on Market-1501 dataset can be found at https://drive.google.com/file/d/1oXLY60iX8Vkbp-iTlrrzzkrFjME1Esun/view?usp=sharing.