• Extract SIFT descriptor

• Extract popular CNN features: inceptionresnetv2/inceptionv4/resnet/resnext/xception/senet/polynet etc. see pretrainedmodels/models for more details.

• Feature Postprocess: PCA + whiten

• Visualize top k image retrieval results.

pytorch >=1.0
torchvision
PIL
h5py
numpy
tqdm
munch
matplotlib

python gen_datalist.py img_dir image_list 

example:

python gen_datalist.py  dataset/paris6k/train train.lst

Same to val.lst

python extract.py model_name image_list save_file gpu_id batch_size

example:

python extract.py resnet101 train.lst paris-train.hdf5 0 128

Same to val.lst

Note: you can modify utils/parser.py to fit your need

python evaluate.py --datasets=paris6k --features=paris-val.hdf5

cd eval/

# calculate distance
python hnsw.py train_db test_db top_k

example:

python hnsw.py paris-train.hdf5 paris-val.hdf5 5

A file named paris_rank is generated, format:

val_image_id train_image_id1 train_image_id2 ... train_image_idk

Use mAP as evaluation metric:

python map.py pred_file

example:

python map.py paris_rank

I tested this code on ThePerfect500k dataset, please verify the ground truth file in map.py which is a csv file in format:

val_image_id, train_image_id1 ... train_image_idn

• PCA+Whiten for CNN features
• Feature fusion

Update by Skye

06/22/2019