Clarifying nuScenes results

ldkong1205 / LaserMix

[CVPR 2023 Highlight] LaserMix for Semi-Supervised LiDAR Semantic Segmentation

https://ldkong.com/LaserMix

Apache License 2.0

272 stars 17 forks source link

Clarifying nuScenes results #8

Closed cavitcakir closed 10 months ago

cavitcakir commented 10 months ago

Hi, thank you for the codebase. I have questions for you;

Could you clarify whether the results for nuScenes are obtained with Cylinder3D or SPVCNN?
If it's Cylinder3D, could you release the config file or the training hyperparameter details (num epochs, voxel size, etc.) used to obtain the supOnly results for nuScenes with voxel backbone?

Thank you.

ldkong1205 commented 10 months ago

Hi @cavitcakir, thanks for your interest in this work!

The results reported in the paper were from Cylinder3D, with the spconv backend. We plan to support SPVCNN in our next update.
Sure, we will provide the config file in our next update. Please allow us a bit more time for this. Meanwhile, you can find some details in our supplementary file.
Our next update should be ready soon. We will keep you updated!

ldkong1205 commented 10 months ago

Hi @cavitcakir, we have migrated the old codebase to MMDetection3D. A pre-release version is available in this repository, including the training, evaluation, and visualization code. Feel free to explore and let us know if you encounter any issues.

The training log and checkpoints of LaserMix on SemanticKITTI will be available in the next few days. The results from our new implementation are better than the ones reported in the paper.

The support for nuScenes, ScribbleKITTI, and Waymo Open will be available next month. We will also support the semi-supervised learning version of the LiDAR panoptic segmentation task.

ldkong1205 commented 10 months ago

Closing this issue now. Feel free to open a new one if you have any other questions!