Instructions to setup training infrastructure

1. Install required python packages from requirements.txt file.
2. Place the input data images and json file as per samples in the repository.
3. Path for Json file is defined in config_2.py file and dataset_500.json contains relative paths for the input images.
4. Surrounding stream input size can be modified in config_2.py file through variables image_size_w_p & image_size_h_p, default values are 96,48 respectively.
5. Current data is splitted in 5 sets beforehand and Train:Validation:Test split is 2:1:2.
6. Data pre-processing parallelization can be controlled through max_workers parameter in siamese_two_stream.py, default = 10.
7. suffix and model_name parameters from siamese_two_stream.py can be modified to change the saved model names.

Commands

Train

python siamese_two_stream.py train

Test

python siamese_two_stream.py test models-car-96-96/

Test resuls on Vehicle Reid dataset

• Combined Results -
  • Precision:94.56 | Recall:91.08 | F-score:92.62

Model Architecture and input example