Format of bbox about chainercv HOT 11 CLOSED

label_id is always integer. But coordinates sometimes take float values.

I understand that this is a problem.
Currently, this is handled by making a bounding box to be float type (e.g. https://github.com/pfnet/chainercv/blob/master/chainercv/datasets/pascal_voc/voc_detection_dataset.py#L100), and asking user to convert it to integer type whenever a user needs integer data.

Personally, I do not like to add additional type (bbox_dtype = [('rect', 'f8', 4), ('label', 'u4')]), as it would be cumbersome for users to import or copy this dtype every time to their own projects. I think this is not really Pythonic. Also, I am not sure if this dtype is a standard convention in the community or not.

How about changing type of tuple returned by Dataset?
Currently, it returns

img (CHW) and bbox (R, 5)

Instead,

img (CHW), bbox_no_label (R, 4), and Label (R,)

from chainercv.

Also, it is important that DetectionVisReport is easy enough for users to use.
http://chainercv.readthedocs.io/en/latest/reference/extensions.html#chainercv.extensions.DetectionVisReport

from chainercv.

I understand why you do not want to use a structured array. I agree that an additional dtype is a little cumbersome.
As you suggested, img (CHW), bbox_no_label (R, 4), and Label (R,) seems a better solution.

from chainercv.

By defining bbox as a set of four coordinates, we can replace the name bbox_no_label with bbox.
This definition is natural because bbox of Pascal VOC does not contain class information.

from chainercv.

Does img (CHW) float32, bbox (R, 4) float32, label (R,) int32 sound good to you?

from chainercv.

Yes. It looks good to me.

from chainercv.

OK.

I will change specification of bbox related functionalities.

from chainercv.

@furushchev
What do you think about this proposal?

from chainercv.

LGTM to change to (R, 4) because:

• For resizing bounding box according to image size, (R, 4) would be better.
• labels are integer, values of bbox may be float.

from chainercv.

I find the same issue in keypoint.
https://github.com/pfnet/chainercv/blob/master/chainercv/transforms/keypoint/resize_keypoint.py#L4

vaild is always bool while x and y may be float.

from chainercv.

Resolved #47.

from chainercv.