Object detection with maskrcnn-benchmark on custom dataset

I have a dataset of 13k classes with 6 labels (none of which are in COCO) and I am trying to use the Fast RCNN model to train an object detector. Has anyone had success using this https://github.com/facebookresearch/maskrcnn-benchmark/issues/521?

Currently, the blockers are:

  • Formatting the dataset in COCO-style.
  • Implementing the Dataset class as shown here.