Hi, I want to train a mask rcnn model on a custom dataset.
Will be using the Detectron repo for this.
Some objects in the dataset are very small
by up scaling all images and by using the following RPN
settings,
RPN_MAX_LEVEL: 6
RPN_MIN_LEVEL: 2
RPN_ASPECT_RATIOS: (0.25, 0.5, 1.0, 2.0, 4.0)
RPN_ANCHOR_START_SIZE: 8
Hopefully the small objects can be detected
The question is, should I edit the params for ROI pooling?
I would be referring to
ROI_CANONICAL_SCALE
ROI_CANONICAL_LEVEL
ROI_MAX_LEVEL
ROI_MIN_LEVEL
Thanks for your advice!