I am training an object detection model using the YOLOv5 architecture. I have the following classes and counts.
| class | n_instances |
| ----- | ----------- |
| dog | 1000 |
| cat | 4000 |
| fox | 8000 |
My training dataset is imbalanced. The difference between the largest and the smallest group is (8000/(8000 + 1000)) - (1000/(8000 + 1000)) = 77%. What should I do?
- Option 1: Just ignore it and train the model with this imbalanced dataset.
- Option 2: Balance the dataset by removing 7000 foxes, 3000 cats and train the model with a balanced dataset of 1000 instances of each class.
I can imagine that option 2 is better suited for very imbalanced datasets, and option 1 for only slightly imbalanced datasets. If this is true, what would define 'very' and 'slightly'? In other words, at which percentage would be the threshold?
