Prepare dataset from public and custom sources to retrain a less class object detection model

My purpose is to retrain an object detection model that can detect classes:

  • bicycle
  • motocycle
  • people
  • custom doorsign

I was told to do it on a pre-trained unpruned NGC model, like trafficcamnet.

The question is, I can prepare the dataset only for custom doorsign, the other 3 common classes will be from public dataset like PASCAL VOC or MS COCO, I would like to know how to only extract targeting sub classes(that 3 common classes: bicycle, motocycle, people) from the public dataset? Does TAO support train for sub-classes in a dataset?

Yes, it can. Just set the classes you want to train. For example, if you only want to train “car” class. Then set as below.
target_class_mapping {
key: “car”
value: “car”
}