Num classes in BBox head maskRCNN

[lauriebax]

Hi,
When I tried to train my Mask RCNN on my custom data set I tried to set the class_weights for the CE-loss. It complained that it needed 8 weights, instead of the 7 classes I defined in my data set. Digging through the code I found this file ConvFCBBoxHead where the head is made with self.num_classes + 1.

Can anyone explain to me why this is done and how I should interpret the additional class?

Thanks in advance

[zachcoleman]

It is a background class (i.e. your bbox head classifying it as BG/not-one-of-target-classes). See this code from base class:

mmdetection/mmdet/models/roi_heads/bbox_heads/bbox_head.py

Lines 286 to 287 in 9d3e162

	bg_class_ind = self.num_classes
	# 0~self.num_classes-1 are FG, self.num_classes is BG