Label issue

[epicjung]

Maybe this is not the issue of your code but SemanticKitti label itself, but I just want to make sure before I dive too deep into training your network.

I've run your code with SemanticKitti dataset and the result of the predicted range image is shown in the below link.

https://www.notion.so/Result-f3c2e9ea10b248a289d3dee2e49a4907

If you see the image, the range projected image on the top and the predicted range image in the middle have some points colored at the far end of the road, yet the ground truth image does not have the colored points. Does this mean that those points are actually unlabelled (black), while the prediction predicts a class? This frequently happens in some images.

I am just wondering if this is the ground-truth problem in SemanticKitti or anything wrong in the code? Did you have any problem just like this while training?

Thank you so much in advance!

[chenfengxu714]

Good observation! We didn't notice this before. I think it is the dataset that doesn't provide labels on these kinds of scenes, because they usually represents very far points where it is very hard to recognize what class they are. If it appears very often in many cases, we do believe it will cause some problems, yet we don't notice any currently and will appreciate it if you could explore this further.