Voting Classifier (Multi-View Classifier)

fastai
1

Hello,

I’m new to fastai and have a image classification problem where there is one item per image and possible 5 classes.
With the help of the fastai course and this forum I was able to make a model with 85% accuracy.
Now to improve the accuracy I’d like to make use of 4 different perspectives of the item. This means I have in total 4 images of the same item with different perspectives and I’d like to make some kind of voting classifier, which looks at all the 4 images instead of just one.
I came acros the word Multi-View Classifier.
Is this some how possible with fastai? Or do I have to code this from scratch?

Thanks for your support.
Rafa

2

Would data augmentation during training and test time augmentation during inference address your goal?

Jeremy introduces data augmentation during training in 02_production.ipynb and test time augmentation in 07_sizing_and_tta.ipynb. He also goes over test time augmentation in one of his live coding videos. Augmentation can involve warping the images (which seems similar to having images with different perspectives of an object).

I am not familiar with voting classifiers or Multi-View classifiers, so I can’t speak to how data augmentation and test time augmentation compare to them, but I’m sharing these concepts in case you haven’t come across them yet in fastai.

3

Thanks vbakshi for your reply.
I know data augmentation during training. I came across tta briefly and have to take a look at it for this specific task.

1 Like