Based on our current labeling information, we have the following distribution:

1. Individual student
2. Board / materials
3. Whole class
4. Other

It's evident from the distribution that our dataset is imbalanced. To address this:

• Weighted Loss Function: We can assign different weights to classes, making the model more sensitive to underrepresented classes.
• SMOTE (Synthetic Minority Over-sampling Technique): This technique can generate synthetic samples for minority classes, improving the balance.
• Oversampling: By randomly duplicating samples from the minority classes, we can increase their representation in the dataset.

While we have access to some tabular data like skeleton coordinates for training, I understand that in a real production environment, we might not always have the luxury of labeling this data. Consequently, I'm prioritizing the following approaches:

• Image Augmentation: Enhancing data diversity through methods like grayscale conversion, rescaling, colorization, etc.
• Incorporating Depth Images: Adding depth images as prior information can provide additional context, potentially enhancing the model's accuracy.