avoid zero division in bbox2loc

[knorth55]

No description provided.

[yuyu2172]

How about np.minimum(..., 1e-14)?

[Hakuyume]

How about using eps which can get from numpy.finfo(https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.finfo.html) instead of magic value 1e-14?

[Hakuyume]

This is just a question. Are there any cases that height or width becomes 0? According to assert_is_bbox, both height and width should be a positive number (https://github.com/chainer/chainercv/blob/master/chainercv/utils/testing/assertions/assert_is_bbox.py#L25-L28).

[knorth55]

rpn output rois can be height =0 and width = 0 at initial state in training.
and this rois can be used proposal_target_creator

[knorth55]

thank you for review.
I updated

[Hakuyume]

rpn output rois can be height =0 and width = 0 at initial state in training.
and this rois can be used proposal_target_creator

I see. That makes sense.

[Hakuyume]

LGTM. Thank you!