search

biubug6 / Pytorch_Retinaface

Retinaface get 80.99% in widerface hard val using mobilenet0.25.
MIT License
2.63k stars 774 forks source link

an error in data augment? #18

Closed Keysmis closed 5 years ago

Keysmis commented 5 years ago

I think you have a mistake in data_augment.py preproc ruction.

image

maybe you should multiply the resize_size??

hegc commented 5 years ago

boxes_t[:, 0::2] = boxes_t[:, 0::2] * (self.img_dim / width) / self.img_dim, so the self.img_dim is reduced.

SnowRipple commented 5 years ago

@biubug6 A little of-topic.. Can you described the meaning behind "min_sizes" and "steps" in config please? What would be the reason for modifying them?

biubug6 commented 5 years ago

"min_sizes" is anchor size on different scales, "steps" is downsampling rate of feature map. I don't modify them?

SnowRipple commented 5 years ago

@hegc Sorry. maybe I'm missing something but both boxes and landmarks are normalized with width and height of the image before resizing, why? Since height and width change after resizing shouldn't they be normalized with width and height after resizing?