@alalek, perhaps, we will need to update opencv_face_detector.prototxt on CI. Or it's better to replace tests to use an origin deploy.prototxt?

Update file hash here too: https://github.com/opencv/opencv_extra/blob/3.4/testdata/dnn/download_models.py#L227-L229

And probably we need to change target branch to 3.4 in URL (I'll change that to master during 3.4 merging)

In case of breaking of existed tests - add and use updated file under a new name, like opencv_face_detector_20190722.prototxt.

@alalek, please let's wait for @lubagov feedback before merge.

Red box, with original padding (pad: 0, on both), blu box, from this commit.

My opinion, Conv_9 * does not work at all, although Conv_8 * works. I think, if the resolution is less than 300x300, then the image needs to be scaled to 300x300. Or use protopxt, without the last one, or two SSD blocks. Depending on how much less resolution.

@lubagov, Thanks for quick response! May I ask you to share an origin image as well?

Original image:

Here normal output with original padding:
, face position [24, -4, 239, 301], face size 215x305,
and here work conv9_2_mbox_priorbox (min_size: 264.0, max_size: 315.0).

Here with suggest padding:
, face position [59, 22, 166, 248], face size
107x226 it is only fit in conv8_2_mbox_priorbox (min_size: 213.0, max_size: 264.0)

Please check, train.prototxt file also. Because it also has wrong padding, and wrong train network.

@lubagov, Will be fixed by #14938. Thanks!