tinaface_r50_fpn_widerface.pth mAP仅0.628

[dan520520520]

tinaface_r50_fpn_widerface.pth测试，按照Issues #25的过程做了一遍，还是下面的结果，做了filter_widerface_val.py以后感觉xml没什么变化? +-------+-------+---------+--------+-------+ | class | gts | dets | recall | ap | +-------+-------+---------+--------+-------+ | face | 31957 | 8420592 | 0.837 | 0.628 | +-------+-------+---------+--------+-------+ | mAP | | | | 0.628 | +-------+-------+---------+--------+-------+

[mike112223]

Please follow the latest official instructions (this) to do the testing. If you still get wrong results, please provide the complete testing log.

[dan520520520]

@mike112223 filter_widerface_val.py时会遇到中断，把assert np.sum(bbox == gt_box) == 4注释掉了，有影响吗？

[mike112223]

Have you solved the problem? I followed the instructions again, and i did not stuck at running filter_widerface_val.py.

[dan520520520]

Have you solved the problem? I followed the instructions again, and i did not stuck at running filter_widerface_val.py. 解决了，dict(typename='Resize', keep_ratio=True)，我之前keep_ratio=False关掉了，打开就可以了。我是按照步骤做，达到了0.923.

[dan520520520]

不运行filter_widerface_val.py，可以达到0.923吗？

[mike112223]

Check #24. If you skip this step, some extremely hard faces will be taken into account in evaluation, which will definitely make the score low.

[twmht]

@mike112223

is filtering bounding box necessary?

I did not see the same thing in retinaface (https://github.com/biubug6/Pytorch_Retinaface/blob/master/test_widerface.py)

[mileistone]

@mike112223

is filtering bounding box necessary?

I did not see the same thing in retinaface (https://github.com/biubug6/Pytorch_Retinaface/blob/master/test_widerface.py)

It's necessary.

[twmht]

@mileistone

can you provide the reference?

[mileistone]

It's a postprocessing method proposed by us, there is no reference.

[twmht]

@mileistone

I change the backbone to mobilnet0.25 in order to reproduce a better result than https://github.com/biubug6/Pytorch_Retinaface.

However, the result is inferior, for example, the easy set is 84.8 vs 88.67. they use ohem which is inferior than focal loss in theory. Do you have any idea？