AIZOOTech/FaceMaskDetection

Testset PR curve

Opened this issue · 4 comments

Would you be so kind and tell me how you achieved the results: face: 0.896 and face_mask: 0.919?

Are these mAP?
Are these the results for model: face_mask_detection.pb?
What value did you take for the conf_thresh?
and what about IoU_thresh in eval process?

Regards.

Yes, this is the ap for face and face with mask.
face_mask_detection.pb and other models have same results, espically for the model360.pth, model360 has bigger input size with 360, the mAP is higher than others.
conf_thresh=0.02, IoU_thresh=0.5

Thank you for your quick response.

I noticed that the face_mask_detection.pth model also accepts 360x360 input data. Is that OK?

To avoid ambiguity: with IoU_thresh I meant IoU when calculating mAP not for NMS.

Which code did you use to calculate the mAP? I use evaluation code and validation for WiderFace (and equivalent: https://github.com/wondervictor/WiderFace-Evaluation) and got different results.

Regards.

If you use input size with 360, you'd better use model360.pth. face_mask_detection.pth can be loaded, however, the result may have some problem. I will remove the model360.pth when I have time. All the other models of tf, mxnet, caffe have exactly same result.
You can use any codes to calculate mAP as long as their codes is correct.
I write the evaluation codes by myself.

You can use any codes to calculate mAP as long as their codes is correct.
I write the evaluation codes by myself.

I try different codes for calculation mAP but couldn't get your result. Could you share your code, please.
Thank you.

Regards.