Week3, Assignment 2, Questions about 'true mask' and 'out-of-sample' prediction


I finished this assignment without issues but I have a couple of questions:

  1. Where did the true mask for each image come from? From manual labeling (unlikely…) or some other type of clustering algorithm? If it is the latter, then why do we want to train a U-net to do the clustering if an existing algorithm is already doing so well by giving us the ‘true’ mask? Is it because U-net is computationally efficient?

  2. As the true masks for the images are used for training the U-net, the accuracy computed in this assignment thus reflects an ‘in-sample’ error of the trained U-net. I would be interested in testing the ‘out-of-sample’ accuracy by using the trained U-net for prediction on a test image. Is there a way I can upload some test images of my own and use U-net for prediction? What would the prediction in tensorflow language be like?


1 Like

Hi @yezhang.hydro ,

Creating masked images can be anywhere from an easy to a mighty task. You can search online and you can find many ways to do this task.

Let me tell you how we did it for the assignment:

We used the CARLA simulator for this assignment. Some self driving cars in the city were created, added a car with a RGB camera and a semantic segmentation camera. Recording was started with those two synced and repeated several times in different cities and configurations.

Here’s some material which can help you:

Hope this helps,