InnovArul / rgb_IR_personreid

An independent implementation to reproduce the paper "RGB-Infrared Cross-Modality Person Re-Identification" from ICCV2017

rgb_IR_personreid

RGB-D ICCV-2017 paper

Author's webpage for ICCV2017 paper "RGB-Infrared Cross-Modality Person Re-Identification": http://isee.sysu.edu.cn/project/RGBIRReID.htm

Dataset

Download the dataset from the following link and extract contents to the folder datasets/sysu_mm01.

Dropbox (from ICCV-2017)

The folder contents of datasets folder will look like:

./datasets
    - sysu_mm01
        -- cam1
        -- cam2
        -- cam3
        -- cam4
        -- cam5
        -- cam6
        -- exp

Training

To train the model (ResNet6 from model.py), go to src_py folder in command prompt and execute:

python doall.py

The training logs and models will be saved under <root>/scratch folder.

Testing

To test the model, go to src_py folder and edit the file doall_test.py file to place the path of pretrained model:

pretrained_model = "../scratch/..."

Then execute the command below to test the model:

python doall_test.py

The features of all the images from different cameras will be stored with file names suffixed _camX where X=cam number under the corresponding scratch log folder. These features will be used by Matlab evaluation code to calculate the relevant metrics.

Metrics calculation

To evaluate and get the metrics, the authors have release a Matlab evaluation script in the github repository: https://github.com/wuancong/SYSU-MM01/blob/master/evaluation.

Given the features, the evaluation code in their repository calculates the rank-1, mAP metrics. Open the demo.m file inside <root>/src_py/SYSU-MM01/evaluation/ directory and give the relevant information regarding feature path, result folder, prefix of the model and execute the demo.m file.

Alternatively, you can use python version of evaluation script from <root>/src_py/SYSU_MM01_pythoneval/. Refer to the README file of SYSU_MM01_pythoneval project for more information on how to use the evaluation routine. It is as similar to Matlab evaluation, but implemented in python to have seemless interface with python scripts.