We ran into several challenges:

1. Low quality data
2. Manual removal of noise
3. Manual labelling for classification
4. Insufficient amount of data, We needed 20,000 images to accurately classify the validation set, We only had about 400 good quality images.

The first and the most time consuming step was to prune the data manually! We dedicated roughly 2 hrs to make sure that the data feeded into the model is proper and in the correct format.

We tackled these challenges using following techniques:

1. Data augmentation to tackle amount of data, We aumented the pruned data generting about 900 images.
2. Used tools like Roboflow to label the existing data