Image Segmentation Overfitting


I have a question regarding the Image Segmentation with U-Net Assignment. The assignment shows high accuracy on the training set but it does not show model performance on any validation/test dataset (it does not give the chance either to upload own images as in other assignments).

I tried to run the model with my own images from streets in the country where I live (Colombia) but found poor quality output. That gives me the impression that the model from the notebook is overfitting to the training set. I increased the training dataset size by downloading the entire 5,000-image Carla dataset and run the model on it, but the result quality did not improve much. Do you have any advice on how to improve model performance with new test images?


Boris M.

In general, fine tuning a model to your custom dataset yields good results. That said, here are a few things to consider:

  1. The model used in the assignment is much smaller than the one described in the paper. A bigger model is one of the options to consider.
  2. Here’s a link that shows how to employ transfer learning for training only the upsampling block of unet. See if this helps as well.