I am using tensorflow 2 object detection api with a custom dataset to train object detection networks. Some of the labels in the dataset are too big and cover regions that they should not cover. Is there a way to discard specific labels (based on the size of their bounding box) from the dataset during training, with the pipeline config file?
Does the input reader protos allow filtering labels by size or should I filter the labels in the xml before turning them into tfrecords?
So far i have read some of the protocol buffers from the tf2 object detection api GitHub but have not found what I need.
