Question about the semantic_prob data about overlapnet

[Chen-Xieyuanli]

Hey @CK-Sung,

1. The semantic_probs are the results of the second last layer of RangeNet++ which contains the probabilities over all the classes.

2. As far as I know, the SemanticKITTI dataset only provides the ground truth labels but not the probabilities over all the classes. You could try directly training OverlapNet with label values. However, since the labels are discrete, the results may not very good. You could also generate "probabilities" like dataset for training OverlapNet using SemanticKITTI by assigning the real class channel as 1 and other classes as 0.

3. As I answered here, you could use rangenetlib to generate the semantic estimates with probabilities over all classes.

I hope it helps and thank you for following our work!

[CK-Sung]

Thank you for answer and great works!!