Mask RCNN augmentation and class weight config


I am working on training a Mask RCNN model using custom data and the TLT 3.0 toolkit [1].

I can see that the Faster RCNN detection model [2] has many more config options (e.g. customize augmentation, specifiy class vs. box weights etc), while the Mask RCNN config seems more scarce, although it would expect the models to be of similar structure except for the added mask head for MaskRCNN.

Is it possible to further configure the Mask RCNN in the TLT 3.0 beside what is documented? Specifically I am looking to configure augmentation and specify different weights for different classes, i.e. certain classes are more important to get right.

[1] Instance Segmentation — Transfer Learning Toolkit 3.0 documentation
[2] FasterRCNN — Transfer Learning Toolkit 3.0 documentation

Best regards,

I will sync internally about your request.