search

PRBonn / lidar-bonnetal

Semantic and Instance Segmentation of LiDAR point clouds for autonomous driving
http://semantic-kitti.org
MIT License
915 stars 204 forks source link

labels/semantic-kitti.yaml #75

Closed akouri-dd closed 3 years ago

akouri-dd commented 3 years ago

I'm a little confused by some of the values in the labels .yamls.

  1. What method is used to generate the values for content -- I'm assuming it's the number of points in the dataset for each class divided by the total number of points in the dataset?
  2. How is learning_map supposed to work? I have tried creating my own mapping, but it causes the IoU scores for the remaining classes to go down significantly.
  3. What is the purpose of learning_map_inv? Why not just dynamically calculate this, since there is no new information provided?
  4. What does learning_ignore do?
jbehley commented 3 years ago

Sorry for the late reply:

  1. content is the proportion of points per class uses for weighting in the weighted cross entropy.
  2. learning_map maps our semanticKITTI annotation labels (like 10 for car, 30 for person) to a range that is used for learning, such as 0-19.
  3. learning_map_inv is the opposite direction, which is also the way you have to submit the results to the codalab evaluation.
  4. learning_ignore are the classes that are ignored in the training. (I think.)
jbehley commented 3 years ago

If you have further questions regarding this specific topic feel free to re-open this issue or open another issue for other questions.

xdtzzz commented 2 years ago

hi, bro. @akouri-dd I think you also use the code to train your own dataset. I wonder have you succeed? How and what the content in labels/semantic-kitti.yaml should be modified? thanks!