• Python2.7
• Tensorflow 1.1.0
• Numpy
• Scipy

• Download the dataset from https://vision.soe.ucsc.edu/node/178

• Extract VIPeR.v1.0.zip file. You should get a VIPeR directory containing cam_a and cam_b folders.

• Run viper/imgdump.py script with VIPeR directory as the first CLI arg, and an output directory for storing the data batch files. Example:

  $ cd viper
  $ mkdir viper-data
  $ ./imgdump.py ~/Downloads/VIPeR.v1.0/VIPeR viper-data
  

  We'll be using viper-data for training/testing our models.

• Download the dataset from http://www.ee.cuhk.edu.hk/~xgwang/CUHK_identification.html

• Extract CUHK01.zip file. You should get a campus directory containing dataset image files.

• Run cuhk01/imgdump.py script with campus directory as the first CLI arg, and an output directory for storing the data batch files. Example:

  $ cd cuhk01
  $ mkdir cuhk01-data
  $ ./imgdump.py ~/Downloads/CUHK01/campus cuhk01-data
  

  We'll be using cuhk01-data for training/testing our models.

Both the models have train.py & test.py files for training & testing respectively. Both the files need two command-line args (in order):

• data dir: dataset dir path
• out dir: results dir path Used for storing model checkpoints and model statistics.

test.py supports an additional third CLI arg: --run-once; which evaluates the model accuracy on the test dataset once and exits. If --run-once is not specified, the script periodically (every 60s) evaluates the model accuracy on the test dataset. This behavior is helpful in observing the test accuracy trend against training steps.

Examples:

1. The following commands train the model and evaluate the trained model's performance on the test dataset:

      $ cd cuhk01
      $ ./train.py cuhk01-data out
      $ ./test.py cuhk01-data out --run-once
   

2. The following commands train and test the model simultaneously. Training in-progress model's performance on the test dataset is evaluated every 60s:

      $ cd viper
      $ ./train.py viper-data out &
      $ ./test.py viper-data out
   

You can visualize all the parameters and results using TensorBoard during/after training. You can navigate to http://localhost:6006 after running the following command:

  $ python -m tensorflow.tensorboard --logdir=out

out is the directory path passed as the second CLI arg to train.py and test.py scripts.

• VIPeR takes ~1 hour to converge on CPU based training. We haven't tried training on GPU.
• CUHK01 takes ~4.5 hours to converge on CPU based training. We haven't tried training on GPU.