Allow empty object bounding box for SSD training

[jo7ueb]

Hi,

Thank you for your great library.

I'm training SSD model based on your example code (https://github.com/chainer/chainercv/blob/master/examples/ssd/train.py) .

This example code requires at least one object annotated for each data in dataset. I want this requirement to be exempted because some of data in my dataset has no object annotated.

Do you think my idea is good for chainercv? If yes, I'll try to modify chainercv

Thanks.