search

PRBonn / semantic-kitti-api

SemanticKITTI API for visualizing dataset, processing data, and evaluating results.
http://semantic-kitti.org
MIT License
783 stars 187 forks source link

What is different between outlier and unlabelled points ? #17

Closed hagianga21 closed 5 years ago

hagianga21 commented 5 years ago

Hi, What is different between outliers and unlabelled points?

jbehley commented 5 years ago

For the evaluation both are equally ignored.

However, while labeling outlier correspond mainly to reflections or otherwise distorted measurement.

In the dataset the main cause for outliers are windshields of cars that cause false measurements below the street. But I also so many cases with windows reflecting cars or buildings from the other side of the street.

hagianga21 commented 5 years ago

Yeah, thanks for that. I am just a little curious. How about the reason for the unlabelled points? When I visualize the dataset. I also see unlabelled points class beside the outlier class.

jbehley commented 5 years ago

We labeled up to a range of 50 m each scan (but we used always the aggregated point clouds for labeling); sometimes it was also not possible within this range to definitively assign a class or it was unclear which class to assign. We then just left it unlabeled.

We choose 50 m, since it seemed reasonable at this maximal distance to decide which class is there with a single scan. But the decision is also debatable, of course.

Hope that explains your impression.

hagianga21 commented 5 years ago

Thanks so much for your explanation.