Low Segmentation Accuracy

Hi All! First attempt at using a semantic segmentation model and getting ~10% accuracy (goes down with each epoch).

Public model with my data here:



  1. Segmented 20 photos of burgers and hotdogs (10 each). Created a mask for each image and the codes are: void = 0, burger = 1, and hotdog = 2
  2. Uploaded the data and ran a unet_learner

Was following this tutorial https://walkwithfastai.com/Segmentation on segmentation from a fast.ai alumni except I need to do it with a custom dataset so I started testing on burger and hotdog photos.

I’m making a mistake somewhere (too little data, wrong pixel segmentation, error loss function etc.) but having trouble debugging the mistake.

Any help spotting the error / improving the model would be greatly appreciated!


Hello! It seems like you’ve been able to make some progress since then, with the loss steadily decreasing. It seems like with more epochs of training you’ll have an okay result. What worked to solve your problem?

I also trained a segmentation model getting only 2% accuracy but the results are pretty great.

Even though the data is not perfect, model is able to capture that information. Here in target, you can see that some humans are not properly segmented. So, I would recommend you train and then maybe use segmentation interpretation and visualize the results to get how your model is performing