mahyarnajibi / SNIPER

SNIPER / AutoFocus is an efficient multi-scale object detection training / inference algorithm
Other
2.69k stars 449 forks source link

SSH face detector on MXNet. #26

Open nttstar opened 6 years ago

nttstar commented 6 years ago

Hi authors, Great work! And I know that you're migrating caffe SSH to this repo which will be re-write by MXnet. I have done such work these days but can not achieve the same accuracy as it was by Caffe. For example, single scale 79.1% mAP on widerface-hard, vs reported 81.4% from caffe SSH. Do you have any tips on it?

Thanks.

mahyarnajibi commented 6 years ago

Hi Jia, Unfortunately we didn't have time to complete the port yet. So we are still not sure what might be the problem in your case. Maybe safer to use the SSH repo for now till we reproduce the results...

nttstar commented 6 years ago

Hi, is SNIPER good for SSD-liked detectors? Like your SSH approach. This question may extend to another one: is SNIPER good for FPN-liked structure? From my knowledge they(SNIPER vs SSD/FPN) may have some inconsistent objectives.

nttstar commented 6 years ago

I have made one MXNet re-implementation here

nttstar commented 6 years ago

But sniper was not used. :)

chinakook commented 6 years ago

@nttstar At the end of https://github.com/msracver/Deformable-ConvNets/issues/99 , bharatsingh430 described the differences of snip and fpn.