Attributes for Object Labels - How to use in model training?

[sekh77]

Hi -

I'm using CVAT tool to create annotations for Safety Data Sheets (SDS) Hazard Pictograms. There are nine pictograms - all have the shape of a Rombus. But each of these 9 images has different names (in other words attributes) - [example: exclamation, flammable, acid, bottle etc.]. CVAT has a mechanism to add attributes to a label. I have only one label (class) which is "pictogram". And I'm adding 9 attributes to this label so that during annotation I can clearly specify the attribute for that picture.

I'm using mmdetection v2.5.0 for training Cascade Mask R-CNN HRNet model to detect the pictograms and its attributes. My annotation file takes the COCO JSON format.

Question:
How do I specify to the training about the label attributes? This is the config file I'm using for training - "cascade_mask_rcnn_hrnetv2p_w40_20e.py", and I have mentioned "classes = [pictogram]" in this config file. I'm not sure how to also specify the attributes so that the model can learn and report the label as well as its attribute when an image is given. Hope you can guide me.

Thanks,
Sekhar H.