Darknet

• Training Configuration
• Tricky Details
• Convert Data to darknet format
  • Roboflow
  • GTSDB

Training Configuration

See Jupyter notebook.

Tricky Details

You cannot directly use the pre-trained weights from pjreddie.

• You must extract the weights of the first 15 conv layers by running ./darknet partial cfg/yolov3-tiny.cfg yolov3-tiny.weights yolov3-tiny.conv.15 15 source

Convert Data to darknet format

Roboflow

Use the class label remapping feature.

GTSDB

• AlexeyAB recommended parser
  • in common_config.py configure:

    CLASS_NUMBER = 3
    classes_names = ["PROHIBITORY", "DANGER", "MANDATORY"]

  • in datasets_parsers/gtsdb_parser.py:

    traffic_sign_classes["0-prohibitory"] = [0, 1, 2, 3, 4, 5, 7, 8, 9, 10, 15, 16]
    traffic_sign_classes["1-danger"] = [11, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]
    traffic_sign_classes["2-mandatory"] = [33, 34, 35, 36, 37, 38, 39, 40]
    traffic_sign_classes[str(OTHER_CLASS) + "-" + OTHER_CLASS_NAME] = [14, 13, 17, 6, 12, 32, 41, 42]