search

suzhigangssz / AVIH

Code for Visual Information Hiding Based on Obfuscating Adversarial Perturbations
12 stars 2 forks source link

lfw evaluation #4

Closed Baicheng42 closed 8 months ago

Baicheng42 commented 8 months ago

Hi, Thanks for your great work!

I met some questions about evaluations of LFW dataset. I used your Arcface model here and the code of Dataloader and Test protocol from https://github.com/QuasarLight/Pytorch_Face_Recognition-, but I failed to get a high accuracy. The acc is only 67.33 even on the unprotected data i.e. original face images.

Could you please provide me a link about how to prepare the dataloader and test code. Thank you very much!

Best regard.

suzhigangssz commented 8 months ago

Hi, Thanks for your great work!

I met some questions about evaluations of LFW dataset. I used your Arcface model here and the code of Dataloader and Test protocol from https://github.com/QuasarLight/Pytorch_Face_Recognition-, but I failed to get a high accuracy. The acc is only 67.33 even on the unprotected data i.e. original face images.

Could you please provide me a link about how to prepare the dataloader and test code. Thank you very much!

Best regard.

The input range of the ArcFace we referenced is [0,255], so don't forget to adjust the samples before feeding them into the model. Also we fixed a small bug in line 30 of the avih.py file. Note that the testing method in our paper is different from the common way of testing lfw (see Section 4.1 Evaluation metrics in the paper for more details), but this does not affect the effectiveness of the method, and you can choose this method for evaluation.

Baicheng42 commented 8 months ago

Thank you again! I ignored that since I downloaded the previous version of this code, and now I have fixed this.