Only objects described in terms of polygons are included in training

[realwecan]

Hi Detectron,

According to json_dataset.py, it seems that all crowd objects are excluded from training:

Detectron/lib/datasets/json_dataset.py

Lines 227 to 228 in 1a3ae0c

	# Set overlap to -1 for all classes for crowd objects
	# so they will be excluded during training

Therefore all objects (used in training) are described in terms of polygons, i.e., one object, one polygon. All crowd objects (i.e., one object, more than one polygons) will be ignored. Am I correct?

Thanks!

[ambigus9]

It seems that you are correct, maybe it's for saving memory in the training process.

[Zoalfekar536]

C bc h.fig

[rbgirshick]

Crowd "objects" are cases where there are too many object instances to accurately annotate. E.g., imagine a sports stadium with hundreds of tiny people in the seats -- the whole mass of people will be marked as a crowd. These should be excluded from training because they are not individual object instances.