Shahar Rotem
Afik Bar
The dataset contains 24,345 facial image, splitted to train and test sets.
As we can see from this random samples, Images seems to cover a variety of ages (infants and seniors), and a multiple races.
Furthermore, we can notice that the masks "type" are not fixed to certain color or shape.
Both of the sets are balanced, as we can see from the following plots:
To obtain better generalization, we’ve used several transformations:
-
Random Cropping
-
Random Horizontal Flipping
-
Normalizing each color channel.
Our neural network consists of 7 Convolutional layers with batch normalization & ReLU as activation function, 1 average pool, and 1 fully connected layers.
We found out that Dropout consistently preformed worse (Possibly due to Batch normalization).
Our optimizer utilizes Weight decay, which limits the learned change in weights for each batch, we’ve used the default value of 0.01.
We were able to reach a F1-Score of 96.1 on the test-set.
To obtain better generalization, we’ve used several transformations:
-
Random Cropping
-
Random Horizontal Flipping
-
Normalizing each color channel.
We’ve used MobileNet V3.
We found out that Dropout didn’t affect results at all (Possibly due to Batch normalization).
Our optimizer utilizes Weight decay, which limits the learned change in weights for each batch, we’ve used the default value of 0.01.
We were able to reach an accuracy of 97.95 on the test set.
We’ve tested multiple published models, however, MobileNet intrigued us the most, since it is intended to by used by low budget processing units (In smartphones), which we believe fit the task (: Side note - PyTorch has an implementation of MobileNetV2, but we’ve found out that V3 performed just slightly better.