Day 14

You are given a task of image segmentation. You decide to implement UNet to solve the problem. You observe that the segmentation masks generated by the model after training aren't very accurate ( IoU score or Dice). Which one of the following is a correct inference?

Maybe the architecture isn't well suited for the task. Try out a modern segmentation model like UNet++ or DeepLabV3+

Maybe the contrast isn't sharp enough in the dataset images for the UNet model to train on.

Maybe the pixel resolution isn't high enough for the detailed segmentation required resulting in imperfect outputs.

All of the above

Avenir Light is a clean and stylish font favored by designers. It's easy on the eyes and a great go to font for titles, paragraphs & more.

All of the given alternatives seem a possible cause of faulty segmentation output.

There might be a possibility that the model architecture is naive for the complexity of the task at hand or the dataset has some inherent issues which need to be addressed. It's always either the dataset or the model. If such a situation arises, try out 2 latest models as an alternative. If that doesn't give a performance boost, there is definitely some problem with the dataset. Fix that.

Join us on Slack

Day 14

Credits :

Day 14

Credits :

​