Challenges Faced
Class Imbalance
We faced around 4:1 class imbalance when detecting faceshield and goggles
Here are the strategies that we used to mitigate class-imbalance:
Image Augmentation
Using different loss function, where model faces more loss when it mis-classifies a minority class.
Image Augmentation
Image augmentation is a technique of altering the existing images to create some more data for the model training process.
The picture shows the common image level augmentations that are applied:
An image in the dataset obtained after applying :
Flip operation
Shear operation
Results
Results before augmentation
Results after augmentation. The mean average precision score of goggles class increased.
Focal Loss
Focal Loss function downweighs the easy to classify examples and focuses on hard examples. As the modulating factor (γ) increases, the criteria of well-classified examples decreases.
Results
Results before using focal loss
Results after using focal loss. The mean average precision of the goggles class increased.