Semantic segmentation in high-resolution images with high variance - cannot avoid underfitting

Question

I am working on a dataset of 2K images for a semantic segmentation problem. I want to detect and localize small objects, with the smallest mask to be 5x5 pixels. The images include 5 different textures, quite different from each other.

I am using Unet and EfficientNetb0-3 as backbones, but whatever I do, I get underfitting. I've reached the conclusion that it is probably due to the fact that the resolution is too high and the variance as well.

Do you have any advice on how I could overcome this underfitting problem?

Answer 1

Given your mask is very small, you should look at reducing your convolutions to 2x2 since that will help aggregate more information from these smaller masks. EfficientNet has 3x3 and 5x5 convolutions which may not be suited for your purpose.

It is a better idea for you to train from scratch using smaller convolutions (2x2). Also, since you will lose the edge of transfer learning, look at self-training, if it is possible for your network.