TransWikia.com

Few shot learning in object detection

Data Science Asked by Jan Pisl on January 2, 2021

I am working on a project to detect buildings from satellite imagery in Tanzania using convolutional neural nets. I use a pre-trained model which I further train on a selected area in Tanzania.

The model performs well (f1 score ≈ 0.85) when tested on an area near where it was trained but much worse (f1 score ≈ 0.3) when tested in a region further away. To tackle this issue, I thought I could use few-shot learning – I would first train the network on a large dataset in one region and when I want to use the model in a different region, I would fine-tune it with a few examples from that region.

However, when I read about few-shot learning, for example in this article or in this paper, they describe it as a method for learning new classes/categories from few examples. But I don’t want the model to learn new classes. I want it to learn to recognize the same class, only on images looking slightly different than those used for training.

Question: Is few-shot learning suitable for what I want to achieve? If not, what could I use instead?

Add your own answers!

Ask a Question

Get help from others!

© 2024 TransWikia.com. All rights reserved. Sites we Love: PCI Database, UKBizDB, Menu Kuliner, Sharing RPP