search

TreB1eN / InsightFace_Pytorch

Pytorch0.4.1 codes for InsightFace
MIT License
1.74k stars 423 forks source link

How to get 10-fold split lists for CALFW and CPLFW? #108

Open wandering007 opened 4 years ago

wandering007 commented 4 years ago

Hi, thanks for providing this awesome project.

After I downloaded CALFW and CPLFW datasets from the official websites, I found that there are no off-the-shelf 10-fold cross-validation lists among the files.

I want to ask that how the project process and evaluate CALFW and CPLFW datasets. It would be appreciated if you could provide me the 10-fold split list files. Thanks a lot!

hjy1312 commented 4 years ago

For CPLFW,“Based on it, our CPLFW dataset has been divided into 10 separate folds using the same identities contained in the LFW 10 folds.”

wandering007 commented 4 years ago

@hjy1312 Thanks! I'll try it.

wandering007 commented 4 years ago

@hjy1312 Hi, I have split the 10 folds correctly after matching identities with LFW dataset. But the results on CALFW and CPLFW are quite low (LFW: 99.7%, CALFW: ~70%, CPLFW: ~65%). Have you been ever faced with this issue before? Thanks!

hjy1312 commented 4 years ago

@wandering007 Compared with the LFW pair list file, I suppose you can directly split the pair list into 10 parts in sequence. But I'm not sure whether the data should be divided in this way.

wandering007 commented 4 years ago

@wandering007 Compared with the LFW pair list file, I suppose you can directly split the pair list into 10 parts in sequence. But I'm not sure whether the data should be divided in this way.

Yes, quite like this. I am quite sure that the 10-fold split problem is solved. But I still get low results, which makes me very confused.

hjy1312 commented 4 years ago

@wandering007 Maybe there're other factors influencing the experimental results. I test my model with the directly split 10 folds data, the accuracy is about 82.6%. And when I test the model with randomly selected 10 folds according to , the accuracy is 80.667%.