MutipleFeature-for-PRID

MutipleFeature-for-PRID is implemented in PyTorch

Prerequisites

Pytorch 1.0
cuda 9.0
python 3.6
1GPUs Memory>24G We Recommend Tesla V100

Datasets

To use our code, firstly you should download ReID dataset (Market1501,DukeMTMC-reID,CUHK03-NP and MSMT17) from Here(saqs).

Here we use the CUHK03 dataset as an example for description.

"detected" means the bounding boxes are estimated by pedestrian detector

"labeled" means the bounding boxes are labeled by human

CUHK03_np
│ 
└───Labeled
│   └───bounding_box_test
│   │   │   0003_c1_21.jpg
│   │   │   0003_c1_23.jpg
│   │   │   ...
│   └───bounding_box_train
│   │   │   0001_c1_1.png
│   │   │   0001_c1_2.png
│   │   │   ...
│   └───query
│   │   │   0003_c1_22.png
│   │   │   0003_c2_27.png
│   │   │   ...
└───detected
│   └───bounding_box_test
│   │   │   0003_c1_21.jpg
│   │   │   0003_c1_23.jpg
│   │   │   ...
│   └───bounding_box_train
│   │   │   0001_c1_1.png
│   │   │   0001_c1_2.png
│   │   │   ...
│   └───query
│   │   │   0003_c1_22.png
│   │   │   0003_c2_27.png
│   │   │   ...

Train

First ,you must run the perpare.py to get 'pytorch' folder.

You need prepare your dataset-dir in the 'prepare.py' flie

download_path = '.../dataset/CUHK'

In our train.py,we give you some options,as follows:

parser = argparse.ArgumentParser(description='Training')
#you can choose the gpu to run the trainmodel.
parser.add_argument('--gpu_ids',default='1', type=str,help='gpu_ids: e.g. 0  0,1,2  0,2')
#the model name you want to save.
parser.add_argument('--name',default='...', type=str, help='output model name')
#the dataset direction.
parser.add_argument('--data_dir',default='.../cuhk03-np/labeled/pytorch',type=str, help='training dir path')
#the batchsize you choose in train process,we recommend 64,and you also choose 128 batchsize.
parser.add_argument('--batchsize', default=64, type=int, help='batchsize')
#REA p=0.5,you can use other number [0,1].
parser.add_argument('--erasing_p', default=0.5, type=float, help='Random Erasing probability, in [0,1]')
#warm up epoch.
parser.add_argument('--warm_epoch', default=10, type=int, help='the first K epoch that needs warm up')
#all epoch you should use in the train process.
parser.add_argument('--epochnum', default=150, type=int, help='please to select the epoch num')
#LR.
parser.add_argument('--base_lr', default=0.01, type=float, help='the base_learning rate')
#triplet loss margin.
parser.add_argument('--tripletmargin', default=1.0, type=float, help='the tripletmargin')
#warm up LR.
parser.add_argument('--warmup_begin_lr', default=3.5e-5, type=float, help='warmup learning rate')
#LR decray.
parser.add_argument('--factor', default=0.1, type=float, help='the learning rate decracy')
#using the color jitter.
parser.add_argument('--color_jitter', action='store_true', help='use color jitter in training' )
#using the attention models.
parser.add_argument('--attentionmodel', action='store_true', help='use the attention model')
#change to the test process.
parser.add_argument('--testing', action='store_true', help='import testing features')

Usage

python3 train.py --gpu_ids .. --name .. --data_dir ../cuhk03-np/labeled/pytorch --batchsize 64 --erasing_p 0.5 --warm_epoch 10 --epochnum 150 --base_lr 0.01 --tripletmargin 1.0 --warmup_begin_lr 3e-4 --factor 0.5  --attentionmodel

python3 train.py --gpu_ids .. --name .. --data_dir ../cuhk03-np/detected/pytorch --batchsize 64 --erasing_p 0.5 --warm_epoch 10 --epochnum 150 --base_lr 0.01 --tripletmargin 1.0 --warmup_begin_lr 3e-4 --factor 0.5  --attentionmodel

Test

parser = argparse.ArgumentParser(description='testing')
parser.add_argument('--gpu_ids',default='1', type=str,help='gpu_ids: e.g. 0  0,1,2  0,2')
parser.add_argument('--name',default='duke', type=str, help='output model name')
parser.add_argument('--test_dir',default='.../cuhk03-np/labeled/pytorch',type=str, help='training dir path')
parser.add_argument('--train_all', action='store_true', help='use all training data' )
parser.add_argument('--batchsize', default=64, type=int, help='batchsize')
parser.add_argument('--epochnum', default='last', type=str, help='please to select the epoch num')
parser.add_argument('--testing', action='store_true', help='import testing features')
parser.add_argument('--attentionmodel', action='store_true', help='use the attention model')
opt = parser.parse_args()

python3 test.py --gpu_ids ... --data_dir ../cuhk03-np/labeled/pytorch --name ... --batchsize 64 --epochnum last --train_all --attentionmodel --testing
python3 test.py --gpu_ids ... --data_dir ../cuhk03-np/detected/pytorch  --name ... --batchsize 64 --epochnum last --train_all --attentionmodel --testing

Evaluation

# Using CPU
python3 evaluate.py
# If you want to attain results quickly you can do:
python3 evaluate_gpu.py
# Also, you can use the reranking
python3 evaluate_rerank.py

Datasets	TOP@1	TOP@5	TOP@10	mAP
CUHK_Detected	0.775000	0.894286	0.932857	0.724668
CUHK_Labeled	0.802857	0.912143	0.954286	0.761700
DukeMTMC-reID	0.901729	0.945242	0.960952	0.801553

Visualization

#you should choose the dateset dir
parser.add_argument('--test_dir',default='...\DukeMTMC-reID\pytorch',type=str, help='./test_data')

python3 demo.py --test_dir ../cuhk03-np/labeled/pytorch 
python3 demo.py --test_dir ../cuhk03-np/detected/pytorch

victor-gui/MutipleFeature-for-PRID