dangkhoasdc / sah

The simultaneous embedding and hashing framework

Introduction

This is sah, the hashing framework which simultaneously learn embedding image representation and hashing code. It also implements relaxed-ba, an improved of Binary Autoencoder. We also provide pre-trained hashing model for following datasets:

• Oxford5k.
• Holidays

Citation

The work will be presented in CVPR 2017.

If you use the code in your paper, then please cite it as:

@InProceedings{Do_2017_CVPR,
    author = {Do, Thanh-Toan and Le Tan, Dang-Khoa and Pham, Trung T. and Cheung, Ngai-Man},
    title = {Simultaneous Feature Aggregating and Hashing for Large-Scale Image Search},
    booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {July},
    year = {2017}
}

Requirement & Data

The code requires Matlab version 2014 or higher.

Data

All training data in the code are extracted from conv5 layer in the network VGG-16. All data should be put into the data folder. The models have to be put on workdir. We provide following data for training/testing:

• Training data/ Dataset/ Query and groundtruth of Oxford5k
  • paris6k_conv5_vgg16
  • oxford5k_conv5_vgg16
  • oxfordq_conv5_vgg16
  • groundtruth
• Training data/ Dataset/ Query and groundtruth of Holidays
  • flickr10k_conv5_vgg16
  • holidays_conv5_vgg16
  • holidaysq_conv5_vgg16
  • groundtruth

The pre-trained models are provided as the following format:

sah_{dataset-name}_c{code-length}

Testing models

In test_retrieval.m, please change dataset to holidays/oxford5k to test on Holidays and Oxford5k , respectively. You also need to change code_length to 8/16/24/32 in order to achieve the result as which are reported in the paper. By default, please put datasets on the data folder and pre-trained sah models into the workdir folder. Finally, run test_retrieval.

The table shows the mAP results in Oxford5k and Holidays datasets:

Dataset	8	16	24	32
Oxford5k	8.44	12.70	15.26	18.05
Holidays	7.52	21.15	33.02	39.18

Training models

The training function is implemented on train.m. Similar to test_retrieval, please change the dataset name and code length in dataset and code_length, respectively. To finetune parameters, you can consider these following variables:

• lambda.
• beta.
• gamma.
• mu

For more detail, please check the paper. After training, the model is saved in the workdir folder as the format I have mentioned before. Please noting that the model is only saved when it achieves higher result than max_map.