This project is performed using Pytorch and ResNet-34 is used as backbone. The Datasets used were UTKFace and MORPH-2. The face images were cropped and preprocessed. It is split into Train-Validation-Test sets(80-10-10). The Images are labeled in .csv files. The Batch size used is 256.
This is the Cross Entropy Architecture.The Soft max layer is applied in the final layer.The Optimizer used was ADAM and the learning Rate is 0.0005. The Accuracy of Gender is 83.78%(UTKFace) and 99.18%(MORPH-2) and Mean Absolute Error for Age is 7.13(UTKFace) and 3.27(MORPH-2).
This is Consistent Ranked Logits(CORAL) Architecture.The n-1 indvidual classifiers are applied in the final layer.The Optimizer used was ADAM and the loss function is Weighted Cross Entropy Loss Function.The learning rate is 0.0005. The Accuracy of Gender is 89.20%(UTKFace) and 99.38%(MORPH-2) and Mean Absolute Error for Age is 5.58(UTKFace) and 2.57(MORPH-2).
The Results are:-
MAE: Mean Absolute Error and RMSE: Root Mean Squared Error.
Conclusion: We have observed that "CORAL outperformed both Ordinal and Cross Entropy approaches for Age and Gender Detection".