How to handle irregularities in dataset for classification

I have custom dataset for image classification in which one class has more than 1000 images but another class has about only 250 images. How can I handle such irregularities in the dataset.

Hi @Kalaria_Pinkesh,
You can either use some augmentation methods to create more data for smaller classes, or use some sampling methods for bigger ones to level up the bias during the training process.

One possible approach for unbalanced dataset, can be to just give more weight to class with less instances. For e.g in your case, class with only 250 images can have a weight of 4 (while class with 1000 images be given a weight of 1)

@Ishvinder, can you describe this a bit more? Does it mean, collect a sample out of dataset such that sample is balanced in terms of classes?

Sure, I meant to say that sampling out a set of 250 images from the large class is also an option.

2 Likes