Skip to main content

Classification

Overview​

  • A ResNet will be used on top of the YOLO object detection network in order to infer the final sign class (whereas YOLO only predicts the preliminary classes MANDATORY, DANGER, PROHIBITORY and OTHER).
  • The bounding boxes predicted by the YOLO network will be the inputs of the ResNet.

Training​

  • Classification
    • Colab/: training notebook for ResNet using Google Colab
    • TAO/: training notebooks for ResNet50 using Nvidia TAO Toolkit
    • Note: We will not use the Google Colab ResNet model because it did not work (yet) in the Deepstream pipeline (TODO: figure out why). Instead we will use a ResNet trained with Nvidia's TAO toolkit (see TAO/ folder).