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

Are the semantic segementation dataset the same as the semantic completion? #82

Closed TintinWuNTUEE closed 2 years ago

TintinWuNTUEE commented 3 years ago

Since we want to use the panoptic segementation label for semantic completion trainning, are the data the same?Are they just different representations with different label or are they totally different data? Thanks a lot !!

jbehley commented 3 years ago

The original data is the same from which the different representations are generated.

The completion data is just a different representation (voxels) instead of points, but the filenames 00000.bin of the voxels directory correspond to the same point cloud in the velodyne directory or the labels directory. The original data provided the voxels only for every 5th scan. However, we also generate all voxel volumes for each scan, which you can download on our webpage.

As described in our papers, the voxels are generated from aggregating multiple scans before the actual scan, i.e., for generating 000100.bin in the voxels directory of a sequence, we used the 100 scans before (0, 1, 2, ..., 100) and the corresponding poses to generate a aggregated point cloud that is then cast into a voxel grid, where each voxel is labeled. Find more information on the data format in the README in the zip file.

The labels are also exactly the same.

TintinWuNTUEE commented 3 years ago

Thanks for answering !!

TimChou-ntu commented 3 years ago

Hi, I'm wondering since the point cloud label provides instance label with segmentation label, whether the completion voxel label also provides instance label with segmentation label? or how could I get it? Thanks a lot!

jbehley commented 3 years ago

Currently, we just account for semantic labels in the generation process. Currently, we have no further plans to also provide instance labels for a separate benchmark. However, you can modify the voxelizer to also provide instance labels.