Mask RCNN model not properly generating circular segmentation masks

[abhinavchadaga]

Hello. I am training the mask_rcnn_R_50_FPN_3x.yaml model on a custom dataset of elevator panels to perform instance segmentation and recognize each label and button on an elevator panel. When running inference, I am noticing that while the bounding box accuracy is high for both the label and button class, the circular segmentation masks are either not being properly displayed or generated.

I annotated the dataset using the VGG annotator and the circular features were annotated using the circle/ellipse tool.

I've attatched an example below showing what is being detected on inference. All bounding boxes are correct and instances are properly categorized, but the circular "button" class segmentation masks do not appear to be showing up, or aren't generated.