There are 23 class in the week 3 image segmentation exercise. I have two questions:
- What are the classes?
- How was the training data created? Did someone have to label all the pixels on the image into the relevant categories?
Good questions and the instructions in the notebook don’t provide any answers. The only clue is that the dataset is from the CARLA project. A little googling finds their website here. In 5 or 10 minutes of digging around there, I have yet to find any documentation about the dataset.
You can also find CARLA referenced in lots of papers. The creation of the dataset looks incredibly tedious. Googling “how to create semantic labels for images” gets a bunch of ads for software packages that sound interesting followed by this medium article.
Disclaimer: obviously I do not know the answers. The U-Net material is all brand new as of the big rewrite of DLS published in April 2021. I have not had time to explore any further than what Prof Ng says in the lectures and what’s in the notebook. Well, I did mess with the dataset a little. They are PNG files with 4 channels instead of 3. I got as far as figuring out that they don’t really use the alpha channel: the alpha values are 255 in all the images. And the labels are on channel 0 of the mask files. That’s as far as I got. Sorry!
1 Like