This is a sub-field of liyana face analysis project. Under this title, i intend to reach state-of-art face recognition accuracies and use my methods on real life applications.
Model can be trained with both ArcFace head and Softmax head. Parameters supported with comments in Python file.
Get evaluation scores for LFW, AgeDB and CFP.
| Model | Architecture | Epochs | LFW Acc | AgeDB Acc | CFP Acc |
|---|---|---|---|---|---|
| A | InceptionResNetV1 | 9 | %99.53 | %95.11 | %93.97 |
| B | ResNet50V2 | 11 | %99.51 | %94.53 | %93.60 |
| C | L_Resnet50_E_IR | 7 | %99.70 | %96.75 | %97.34 |
PS: I train models on Google Colab
-
Train with ResNet50V2
-
Train with ResNet101V2(Results are even worse than ResNet50V2 Model, not gonna share this one)
-
Train with L_Resnet50_E_IR(I could try to train more epochs but i will focus on Resnet100 for now)
-
Train with L_Resnet100_E_IR
-
Train with VarGFaceNet
-
Re-train with lower weight decay
-
Train with Dataset V4
First download data to"dataset" folder, then use this script to turn it into tfrecord.
Run this script. Parameters supported in Python file. Script will test on LFW at every 10k step.
PS: Default model architecture set to InceptionResNetV1, check this file for other architectures such ResNet.
Go to script. Parameters supported with comments in Python file.
color is green because faces belong to same person
color is red because faces belong to different persons