The architecture follows the NTU ROSE ReID Project guide line Person_ReID_Baseline. Some of codes are copy from L1aoXingyu's reid_baseline.
ResNet50 Last Stride 1
from huanghoujing's triplet baselineWarmupMultiStepLR
: is from FAIR's paper: 'Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour'
- python 3
- pytorch 1.0 + torchvision
- yacs Yet Another Configuration System
- fire Automatically generating command line interfaces (CLIs)
Install all dependences libraries
pip3 install -r requirements.txt
Use different yaml config files for different experiment settings. All the config files are store in folder config
. Please use different OUTPUT_DIR
names for different experiments to avoid conflit and accidentally files overwritten.
This code support CUHK03, Market1501, DukeMTMC and MSMT17 datasets. All these dataset should be defined in the DATASETS.NAMES
of the config file, our code will be download the corresponding dataset automatically (into the datasets
folder). As this fuction require access to Google Drive, it will not work in China.
Currently support:
- As the last feature maps increase. Batch Size 128 will use more than 12G
- Batch Size 64 use around 10G GPU memory
python train.py ./config/market_softmax.yaml
### Change GPU
python train.py ./config/market_softmax.yaml --DEVICE=cuda:5
### No Re-Ranking
python test.py ./config/market_softmax.yaml
### Change GPU
python test.py ./config/market_softmax.yaml --DEVICE=cuda:5
### With Re-Ranking
python test.py ./config/market_softmax.yaml --RE_RANKING=True
### Market1501 -> DukeMTMC
python test_cross_dataset.py ./config/market_softmax.yaml DukeMTMC
Results Compare with Person ReID Baseline
Dataset | Softmax | Strong Softmax |
---|---|---|
CUHK03 | 56.1 (52.4) | 61.1 (56.2 |
Market1501 | 91.6 (78.7) | 92.5 (80.2) |
DukeMTMC | 83.4 (66.6) | 84.8 (68.3) |
MSMT17 | 69.0 (40.1) | 71.4 (42.5) |
Softmax+Triplet | Strong Softmax+Triplet | |
---|---|---|
CUHK03 | 65.6 (61.8) | 66.3 (61.8) |
Market1501 | 93.2 (82.0) | 93.4 (83.1) |
DukeMTMC | 86.4 (72.4) | 86.2 (72.5) |
MSMT17 | 73.9 (46.4) | 74.6 (47.3) |