Add Multilabel Classification Support

🚀 Feature

Provide support for multilabel classification, e.g. Vehicle -> Sedan.

Motivation

There are several datasets that provide multiple labels for a class. Also, this appears to have a positive impact on transfer learning, new label introduction.

Pitch

1. Provide an option to make use of multilabel categories.
2. Pass through the supercategory attribute from the json annotations.json file. It’d be nice if this was abstracted to enable any number of labels per box but that may require a custom dataset
3. Increase the number of classes to N_classes + N_superclasses or N_nodes in the case of a 3 or more level class hierarchy.
4. Switch to BCE loss if there are multiple labels per box.
5. Generate the truth vector for both precise class and any parent or supercategory(s) of the box.

Alternatives

YOLOv3 implements it in this way, I believe. RFCN takes a different a more divergent approach.

[steve-goley]

• Open Images put this out: https://storage.googleapis.com/openimages/2018_04/bbox_labels_600_hierarchy_visualizer/circle.html
• COCO has supercategories as well, though I’m not sure they are frequently used.
• A recent challenge, xView, had some coarse categorical nature to it as well

[SaiSarath01]

Hi did any one find way for multi label mask rcnn?