REPRODUCTION.md

Reproduction

Prerequisite

• Firstly make sure, you have followed installation guide. and downloaded datasets and pre-trained models.

• Set correct paths in localpaths.m

  • datasets_directory (Main directory of all the datasets).

  • paths.m_directory (Pre-computed MAQBOOL data). Create a directory structure shown below.

    maqbool-data
    ├── models
    ├── post_computed_files #fast verify results
    └── pre_computed_files  #fast results
    

    On our project page, you can download pre_computed and post_computed files. If you just want to test the results.

• Dataset and Pre-trained Models

  • Please download pre-trained models and datasets (Pittsburgh, Tokyo247 and ToykoTM) from NetVLAD project website.
  • We have used VGG-16 + NetVLAD + whitening models only as it has top NetVLAD performance.

• Rename localPaths.setup->localPaths.m in the maqbool home directory. Update paths of datasets folders of datasets_directory(datasets path) and m_directory(to store computed data / checkpoints).

Training and Testing

• To run the main file, open MATLAB and run run main_wsd.m.

Training MAQBOOL Layer

MAQBOOL PDL layer is trained using TOKYOTM dataset. First 250 test samples are taken.

If you didn't download/use from our project page, it will start train and create a model based on first 250 samples of TOKYOTM dataset.

Testing on different datasets

Add configuration in the config_wsd.m file before testing. For instance, if you want to use model trained on pittsburgh datasets and test on Tokyo datasets at feature dimension 512. So the configuration will be

net_dataset = 'pitts30k'; % use tokyoTM if you want to use the tokyo based pre-trained model. job_datasets = 'tokyo247'; % use 'pitts30k' if you want to test on pitts30k. f_dimension = 512; % use '4096', if you want to test on 4k feature dimension.