SVC (kernel = rbf) 97.13
94.64 Table 11. Untuned Baseline for Age estimation
Method
Train (MAE)
Test (MAE)
Decision Tree
0.05
9.86
Gradient Boosted Trees
4.97
6.17
XGBoost
5.00
5.89
Random Forest
1.91
5.75
Linear Regression
4.93
5.61
Linear SVR
4.85
5.58
SVR (kernel = rbf) 4.85
5.49 Clearly, even simple linear regression outperformed training our custom CNN model
for age estimation and logistic regression came remarkably close to the custom CNN
architecture for gender classification on the features extracted using SENet50_f.
As expected, our model performs relatively poorly while predicting ages for people
above 70 years of age. This is quite evident from Table 2. where it can be seen that
there are only 5.78 % images in the dataset belonging to people above 70 (albeit the
dataset is quite evenly balanced when it comes to gender). We believe much better
results can be attained using a more balanced and larger dataset.
6 Conclusion Inspired by the recent developments in this field, in this paper we proposed two ways
to deal with the problem of age estimation, age and gender classification - a custom
CNN architecture and transfer learning based pre-trained models. These pre-trained
models helped us combat overfitting to a large extent. It was found that our models
generalized very well with minimal overfitting, when tested on real-life images.
We plan to extend our work on a larger and more balanced dataset with which we
can study biases and experiment with more things in order to improve the
12
generalizability of our models. In future research, we hope to use this work of ours as
a platform to improvise and innovate further and contribute to the deep learning com-
munity.
