Sub-region localized hashing for fine-grained image retrieval (sRLH)

Dataset Preparation

CUB-200-2011: link：https://pan.baidu.com/s/1cWVu7JHSQV9Pvw-dLlkDQw, extraction code：zzlz.

Details

|--CUB_200_2011 
  |--images 
       |--001...
       |--002... 
       ... 
  |--classes.txt 
  |--image_class_labels.txt 
  |--image.txt 
  |--train_test_split.txt

FGVC-Aircraft: link：https://pan.baidu.com/s/1MEiwAJbBGmsCbpZ5u19x8Q, extraction code：91su.

Details

|--FGVC-aircraft
  |--data
    |--images
       |--...
    |--test.txt
    |--train.txt

Stanford Cars: link：https://pan.baidu.com/s/1c6mivvIXXEjERP2ilDtHNg, extraction code：o96t.

Details

|--Stanford_Cars
  |--cars_test
    |--...
  |--cars_train
    |--...
  |--test.txt
  |--train.txt

Stanford Dogs: link：https://pan.baidu.com/s/1mBDOOVwgT0RAzjIITlwbgg, extraction code：ivsu.

Details

|--dogs
  |--images
    |--Images
      |--file
      |--file
      ...
  |--lists
    |--file
    |--file
    ...
  |--test_data.mat
  |--train_data.mat

Finetune

(1) Put the parameters of Resnet18 into the path .models/petrained. This parameters can be download at link：https://pan.baidu.com/s/1uGfo2JCiX4GmqkGE2waG7A, extraction code：7bu5.
(2) Finetune the network with the cross-entropy loss for classification, such as: python finetune_cub.py.
(3) Choose the network with minimum loss as the finetuned network.
You can also use our pretrained models. The pretrained models can be download at link：https://pan.baidu.com/s/15FlAAZD9NZtW9MVKdwy7RA, extraction code：fith.

Train

(1) Put the finetuned network into the path .checkpoint.
(2) Train the network, such as: python cub_train.py

Citation

@article{xiang2021sub,
title={Sub-region localized hashing for fine-grained image retrieval},
author={Xiang, Xinguang and Zhang, Yajie and Jin, Lu and Li, Zechao and Tang, Jinhui},
journal={IEEE Transactions on Image Processing},
volume={31},
pages={314--326},
year={2021},
publisher={IEEE}
}