leetenki/YOLOv2

Hove you tried training with VOC dataset?

wkentaro opened this issue · 4 comments

wkentaro commented

I'd like to know if the accuracy was also reproduced.
Have you tried this?

naoto0804 commented

I've modified the code and tried testing performance on VOC2007-test, but the mAP was 74.2 on 416x416 input, 0.45 nms thresh, 0.005 detection thresh (all parameters are written in the original paper. Original mAP was 76.8).
I can't figure out what's wrong, do you have any suggestions?
https://github.com/naoto0804/YOLOv2/blob/voc/test.py

wkentaro commented

@naoto0804 Thank you for sharing the info.
Currently I'm wondering why the pad parameters are different among Chainer/darknet implementations.
In Chainer one it has pad=0: https://github.com/leetenki/YOLOv2/blob/master/yolov2.py#L28, but darknet has all pad=1.

Breakend commented

@naoto0804 as in you tried to train using this code or simply load in the weights from darknet?

naoto0804 commented

@Breakend Sorry for my extremely late reply! I tried to simply load the weights from darknet and tested.

  • 👍1