search

zyainfal / Joint-Holistic-and-Masked-Face-Recognition

4 stars 2 forks source link

Questions about the performance of iResNet-100 model on SMFRD in Table 5 of the paper #5

Open ClimberXIE opened 3 months ago

ClimberXIE commented 3 months ago

Hi, I tested the IResNet100 model you released on SMFRD dataset, but the accuracy I get is only about 74%. I am not sure whether it is due to different test pairs. Could you please provide the test pair file of SMFRD? Looking forward to your reply. Best wishes.

zyainfal commented 3 months ago

Yes, the original SMFRD data is not well aligned and we re-aligned faces by RetinaFace, which is noted in the experiment part of the paper.

As to the data we used for experiments, we zipped and uploaded the file on Google Drive and Baidu Pan (please check readme for download)

ClimberXIE commented 3 months ago

Thank you for your prompt reply. I've checked the data. Our data is also aligned using Retinaface, and it looks the same as the data you provided, so I think it's due to different test pairs. But it seems you don't provide the test pairs file of SMFRD in the decompressed data?

zyainfal commented 3 months ago

Hi,

we used pairs for LFW, CFP-FP, and AgeDB-30.

The performance of SMFRD is the average of accuracy4, accuracy5, and accuracy6 as in Line 463 to 467 in trainer.py, https://github.com/zyainfal/Joint-Holistic-and-Masked-Face-Recognition/blob/45394db77ba26afc4286095f5f7bd361770dbc15/code/trainer.py#L463-L467

therefore showing the impact of masked faces.

ClimberXIE commented 3 months ago

Thanks again for your reply, I have solved my problem now！