search

ZhaoJ9014 / face.evoLVe

🔥🔥High-Performance Face Recognition Library on PaddlePaddle & PyTorch🔥🔥
MIT License
3.44k stars 760 forks source link

#Accuracy on my dataset is 82% #72

Closed pp3993 closed 5 years ago

pp3993 commented 5 years ago

Hi I evaluated the performances of pre-trained model on my private Indian data-sets, where accuracy is 82% only. what I need to do to improve performance on my datasets. Thank-you

ZhaoJ9014 commented 5 years ago

Hi I evaluated the performances of pre-trained model on my private Indian data-sets, where accuracy is 82% only. what I need to do to improve performance on my datasets. Thank-you

Hi, thank you for your question. Since our model is pre-train on MS-Celeb-1M, whose race source is global based, it may perform not that well on your own dataset containing Indian faces mostly (significantly different data domain). To improve the accuracy on your data, we recommend you fine-tune our released model on your own training data and perform the evaluation again. Hope this helps:)

pp3993 commented 5 years ago

Do i need to train freshly this model on indian dataset?

pp3993 commented 5 years ago

here, fine-tune means creation of million of indian dataset or change of hyperparameter on pretrained model?