search

tks1998 / Pytorch-Face-recongition-state-of-the-art-Qmul-surveface-

Implementation state of the art for face recognition algorithm.
MIT License
0 stars 0 forks source link

Questions about testing on QMUL-SurvFace #1

Open kidwhh opened 3 years ago

kidwhh commented 3 years ago

Thank you for your reply. I have a few questions for you First, did you train with the training set of QMUL-Surveillance Second, of the 85% you measured, what kind of test is this? Did you run the Face Identification test? Third, can you provide the code to use the model to generate the feature file? It can be directly used for Matlab file testing. I would like to refer to your findings to find my mistakes.

The lack of good results on this data set has bothered me for a long time

My email is whh39686@mail.ustc.edu.cn, we can also send email to discuss, I really need help

tks1998 commented 3 years ago

I just updated my code with a test-verification notebook and adaptive margin algorithms. I train many models, And I lose my best model but I provide model ~83% auc score at drive https://drive.google.com/file/d/13EB-kDFSC54LSaGycbNRYl1CDRP5U2-D/view?usp=sharing. My model use backbone efficient net b3 + adaptive margin for arc-face, In my repos you can run to generate model, And In my experience, you should early stop to get the best model with my repos. In the feature, I'll be update the code with an early stop to avoid overfitting. And We can discuss on my repos to develop algorithms.

kidwhh commented 3 years ago

@tks1998 You may have misunderstood me. I want to use arcface model (https://github.com/TreB1eN/InsightFace_Pytorch) or the adaptive arcface reproduce the QMUL - SurvFace results (https://qmul-survface.github.io/index.html). I don't know if your 83% result is face verification or face recognition. Is the TAR @ FAR =?% Are you using the training set given by QMUL - Survface to train model ?
The other one is, when I trained a model, how can I use this model to generate feature files gallery. mat ,mated_probe. mat and unmated_probe. mat according to the requirements of QMUL - Survface dataset? Finally, the MATLAB function is used to get the test results. Can you provide this part of the code? image