PFD：Pose-guided Feature Disentangling for Occluded Person Re-identification based on Transformer
This repo is the official implementation of "Pose-guided Feature Disentangling for Occluded Person Re-identification based on Transformer(PFD), Tao Wang, Hong Liu, Pinghao Song, Tianyu Guo& Wei Shi" in PyTorch.
Pipeline
Dependencies
-
timm==0.3.2
-
torch==1.6.0
-
numpy==1.20.2
-
yacs==0.1.8
-
opencv_python==4.5.2.54
-
torchvision==0.7.0
-
Pillow==8.4.0
Installation
pip install -r requirements.txt
If you find some packages are missing, please install them manually.
Prepare Datasets
mkdir data
Please download the dataset, and then rename and unzip them under the data
data
|--market1501
|
|--Occluded_Duke
|
|--Occluded_REID
|
|--MSMT17
|
|--dukemtmcreid
Prepare ViT Pre-trained and HRNet Pre-trained Models
mkdir data
The ViT Pre-trained model can be found in ViT_Base, The HRNet Pre-trained model can be found in HRNet, please download it and put in the './weights' dictory.
Training
We use One GeForce GTX 1080Ti GPU for Training Before train the model, please modify the parameters in config file, please refer to Arguments in TransReID
python occ_train.py --config_file {config_file path}
#example
python occ_train.py --config_file 'configs/OCC_Duke/skeleton_pfd.yml'
Test the model
First download the Occluded-Duke model:Occluded-Duke
To test on pretrained model on Occ-Duke: Modify the pre-trained model path (PRETRAIN_PATH:ViT_Base, POSE_WEIGHT:HRNet, WEIGHT:Occluded-Duke) in yml, and then run:
## OccDuke for example
python test.py --config_file 'configs/OCC_Duke/skeleton_pfd.yml'
Occluded-Duke Results
| Model | Image Size | Rank-1 | mAP |
|---|---|---|---|
| HOReID | 256*128 | 55.1 | 43.8 |
| PAT | 256*128 | 64.5 | 53.6 |
| TransReID | 256*128 | 64.2 | 55.7 |
| PFD | 256*128 | 67.7 | 60.1 |
| TransReID* | 256*128 | 66.4 | 59.2 |
| PFD* | 256*128 | 69.5 | 61.8 |
$*$means the encoder is with a small step sliding-window setting
Occluded-REID Results
| Model | Image Size | Rank-1 | mAP |
|---|---|---|---|
| HOReID | 256*128 | 80.3 | 70.2 |
| PAT | 256*128 | 81.6 | 72.1 |
| PFD | 256*128 | 79.8 | 81.3 |
Market-1501 Results
| Model | Image Size | Rank-1 | mAP |
|---|---|---|---|
| HOReID | 256*128 | 80.3 | 70.2 |
| PAT | 256*128 | 95.4 | 88.0 |
| TransReID | 256*128 | 95.4 | 88.0 |
| PFD | 256*128 | 95.5 | 89.6 |
Citation
If you find our work useful in your research, please consider citing this paper! (preprint version will be available soon)
@inproceedings{wang2022pfd,
Title= {Pose-guided Feature Disentangling for Occluded Person Re-identification based on Transformer},
Author= {Tao Wang, Hong Liu, Pinhao Song, Tianyu Guo and Wei Shi},
Booktitle= {AAAI},
Year= {2022}
}
Acknowledgement
Our code is extended from the following repositories. We thank the authors for releasing the codes.
License
This project is licensed under the terms of the MIT license.
111 Dec 31, 2022
28 Nov 16, 2022
35 Dec 6, 2022
127 Dec 28, 2022
87 Dec 30, 2022
59 Dec 17, 2022
121 Dec 25, 2022
10.1k Dec 28, 2022
11.3k Dec 30, 2022
跟论文中的结果大约相差2%，我使用的时pytorch1.7.1, cuda10.2, python3.7.13;所以我想知道这是什么原因造成的呢？
期待您的回复。
287 Jan 04, 2023
2 Nov 01, 2022
181 Nov 18, 2022
5 Apr 05, 2022
228 Dec 29, 2022
181 Dec 19, 2022
36 Dec 29, 2022
4 Sep 19, 2022
62 Nov 28, 2022
111 Jan 05, 2023
98 Dec 07, 2022
4 Aug 27, 2022
2.2k Jan 08, 2023
44 Jan 04, 2023
57 Dec 28, 2022
25 Sep 21, 2022
75 Dec 23, 2022
9 Apr 04, 2022
50 Dec 16, 2022
223 Jan 04, 2023