Recurrent Attention Models for Depth-Based Person Identification [website] [arxiv] [pdf]
Albert Haque, Alexandre Alahi, Li Fei-Fei
CVPR 2016

1. Clone the repo: git clone https://github.com/ahaque/ram_person_id.git

2. (Optional) If you are using GPU/CUDA:

    luarocks install cutorch
    luarocks install cunn
    luarocks install cunnx
   

3. Install HDF5: sudo apt install libhdf5-serial-dev hdf5-tools

4. Install more Lua packages (nn, dpnn, image, etc.): ./install_custom_rocks.sh

5. Confirm that the custom rocks were installed correctly by running torch and checking if dp.Dpit exists. From the bash/command line:

    $ th
    th> require 'dp'
    th> dp.Dpit
    table: 0x41d27e80       [0.0001s]
   

   If you see an error, you may need to refresh the lua package cache. Run: package.loaded from the torch command line.

The datasets can be downloaded from the publishers' websites:

• DPI-T: Depth-Based Person Identification from Top View [website]
• BIWI: BIWI RGBD-ID Dataset [website]
• IAS: IAS-Lab RGBD-ID [website]
• IIT PAVIS: RGB-D person Re-Identification Dataset [website]

To automatically download the DPI-T dataset, run: ./download_datasets.sh.

1. Navigate to src/encoder.

2. Train the encoder:

    th train.lua
    th train.lua -gpuid 0
   

3. The model will be saved to opt.dir every opt.save_interval epochs.

Note: You must have a saved encoder model before training the recurrent attention model.

1. Navigate to the src folder. The file opts.lua contains the training options, model architecture definition, and optimization settings.

    th train.lua
    th train.lua -gpuid 0
   

2. Done.

Haque, A., Alahi, A., Fei-Fei, L.: Recurrent attention models for depth-based person identification. CVPR, 2016.

Bibtex:

@inproceedings{haque2016cvpr,
    author = {Haque, Albert and Alahi, Alexandre and Fei-Fei, Li},
    title = {Recurrent Attention Models for Depth-Based Person Identification},
    booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2016}
}