siamese_person_re_id

Siamese Network for Person Re-Identification in TensorFlow

Usage

For academic research purposes only. Please cite this GitHub repository.

Dependencies

TensorFlow version >= 1.5
NumPy 1.14.0
SciPy 1.0.0
PIL 5.0.0
IPython 6.2.1
jupyter 1.0.0

Dataset

Download the MARS dataset from here:

• MARS - The dataset used for training

Next put the data into a directory with the following structure:

./mars/
  categories/
    0001/
      ....jpg
      ....jpg
      ....jpg
    0002/
      ....jpg
    0003/
      ....jpg
    0004/
      ....jpg
    ...

You can creade a TFRecord file for TensorFlow consumption. This will create a training and validation set for your dataset:

$ python3 create_tf_record.py --tfrecord_filename=mars --dataset_dir=/path/to/dataset/

Training

The training script creates the TFRecord file and then trains the model on the generated TFRecord file.

$ python3 train_siamese_network.py --data /path/to/dataset/

The updates are stored in ./train.log/ which can be seen in TensorBoard.

Inference

To test the model you can run the test script with two test images as such:

$ python3 test_siamese_network.py --img1 /path/to/image1/ --img2 /path/to/image2

Alternatively you can use the two provided IPython notebooks to test the network. The following notebook runs the test on 20 random validation set image pairs and displays the results.

$ jupyter notebook test_siamese_network_jupyter_validation.ipynb

The following notebook runs the test on the 7 test images provided in the main directory and displays the results.

$ jupyter notebook test_siamese_network_jupyter_images.ipynb

To improve the results

• Use data augmentation [AutoAugment] (https://arxiv.org/abs/1805.09501)
• Transfer learning; use a pre-trained model such as VGG16/19 or ResNet instead of training from scratch
• Add more layers; increase model size; train for longer
• Experiment with the architecture and hyper-parameters