retrieval.pytorch

dependence

• python3
• pytorch
• numpy
• scikit-learn
• tqdm
• yacs

You can install yacs by pip. Other dependencies can be installed by 'conda'.

prepare dataset

first, you need to download the dataset from here. Then, you can move them into the directory $DATASET and decompress them by

unzip train_data.zip
unzip test_data_A.zip
unzip test_data_B.zip

Then remove the empty directory in train_data:

cd train_data
rm -rf DIGIX_001453
rm -rf DIGIX_001639
rm -rf DIGIX_002284

Finally, you need to edit the file src/dataset/datasets.py and set the correct values for traindir, test_A_dir, test_B_dir.

traindir = '$DATASET/train_data'
test_A_dir = '$DATASET/test_data_A'
test_B_dir = '$DATASET/test_data_B'

Train the network to extract feature

You can train dla102x and resnet101 by the below comands.

python experiments/DIGIX/dla102x/cgd_margin_loss.py
python experiments/DIGIX/resnet101/cgd_margin_loss.py

To train fishnet99, hrnet_w18 and hrnet_w30, you need to download their imagenet pretrained weights from here. Specifically, download fishnet99_ckpt.tar for fishnet99, download hrnetv2_w18_imagenet_pretrained.pth for hrnet_w18, download hrnetv2_w30_imagenet_pretrained.pth for hrnet_w30. Then you need to move these weights to ~/.cache/torch/hub/checkpoints to make sure torch.hub.load_state_dict_from_url can find them.

Then, you can train fishnet99, hrnet_w18, hrnet_w30 by

python experiments/DIGIX/fishnet99/cgd_margin_loss.py
python experiments/DIGIX/hrnet_w18/cgd_margin_loss.py
python experiments/DIGIX/hrnet_w30/cgd_margin_loss.py

After Training, the model weights can be found in results/DIGIX/{model}/cgd_margin_loss/{time}/transient/checkpoint.final.ckpt. We also provide these weights file.

extract features for retrieval

You can download the pretrained model from here and move them to pretrained directory.

Then, run the below comands.

python experiments/DIGIX_test_B/dla102x/cgd_margin_loss_test_B.py
python experiments/DIGIX_test_B/resnet101/cgd_margin_loss_test_B.py
python experiments/DIGIX_test_B/fishnet99/cgd_margin_loss_test_B.py
python experiments/DIGIX_test_B/hrnet_w18/cgd_margin_loss_test_B.py
python experiments/DIGIX_test_B/hrnet_w30/cgd_margin_loss_test_B.py

When finished, the query feature for test_data_B can be found in results/DIGIX_test_B/{model}/cgd_margin_loss_test_B/{time}/query_feat. And the gallery feature can be found in results/DIGIX_test_B/{model}/cgd_margin_loss_test_B/{time}/gallery_feat.

Post process

You can download features from here. Then, you can put it into the directory features and decompress the files by

tar -xvf DIGIX_test_B_dla102x_5088.tar
tar -xvf DIGIX_test_B_fishnet99_5153.tar
tar -xvf DIGIX_test_B_hrnet_w18_5253.tar
tar -xvf DIGIX_test_B_hrnet_w30_5308.tar
tar -xvf DIGIX_test_B_resnet101_5059.tar

Then the features directory will be organized like this:

|-- DIGIX_test_B_dla102x_5088.tar  
|-- DIGIX_test_B_fishnet99_5153.tar  
|-- DIGIX_test_B_hrnet_w18_5253.tar  
|-- DIGIX_test_B_hrnet_w30_5308.tar  
|-- DIGIX_test_B_resnet101_5059.tar 
|-- DIGIX_test_B_dla102x_5088  
| |-- gallery_feat  
| |-- query_feat  
|-- DIGIX_test_B_fishnet99_5153  
| |-- gallery_feat  
| |-- query_feat  
|-- DIGIX_test_B_hrnet_w18_5253  
| |-- gallery_feat  
| |-- query_feat  
|-- DIGIX_test_B_hrnet_w30_5308  
| |-- gallery_feat  
| |-- query_feat  
|-- DIGIX_test_B_resnet101_5059  
| |-- gallery_feat  
| |-- query_feat  

Now, post process can be executed by

python post_process/rank.py --gpu 0 features/DIGIX_test_B_fishnet99_5153 features/DIGIX_test_B_dla102x_5088 features/DIGIX_test_B_hrnet_w18_5253 features/DIGIX_test_B_hrnet_w30_5308 features/DIGIX_test_B_resnet101_5059