yanx27 / 2DPASS

2DPASS: 2D Priors Assisted Semantic Segmentation on LiDAR Point Clouds (ECCV 2022) :fire:
MIT License
397 stars 51 forks source link

two classes have not been trained on SemanticKitti #54

Closed potaosking closed 1 year ago

potaosking commented 1 year ago

image Hello,Thank you for your great work!When I train on SemanticKitti,the motorcyclist and other-ground classes get very low score,

brahami14 commented 1 year ago

image i have the same issue

Montyro commented 1 year ago

The validation sequence of semantic-kitti has only 5 points labeled as motorcyclist, therefore it cannot distinguish them as a motorcyclist, obtaining a 0% IoU score. The same issue happens with the other-ground class.

If you run in on test and upload results to the benchmark server, you will see it indeed learns those classes.

potaosking commented 1 year ago

The validation sequence of semantic-kitti has only 5 points labeled as motorcyclist, therefore it cannot distinguish them as a motorcyclist, obtaining a 0% IoU score. The same issue happens with the other-ground class.

If you run in on test and upload results to the benchmark server, you will see it indeed learns those classes.

thanks