search

songw-zju / Meta-RangeSeg

The official implementation of "Meta-RangeSeg: LiDAR Sequence Semantic Segmentation Using Multiple Feature Aggregation" (RA-L with IROS 2022)
MIT License
45 stars 6 forks source link

Good work! Looking forward to the code. #1

Closed huixiancheng closed 2 years ago

huixiancheng commented 2 years ago

In fact, I did try to incorporate BD Loss. However, since Lite-HDSeg is not open source, I refer to this repo for the implementation.

However, in my code, it is not as good as described in Lite-HDSeg and your ablation experiments.

Would you like to share this part of the code and your (open3d) visualization code?

songw-zju commented 2 years ago

Hi @huixiancheng , thanks for your interest. I incorpotated BD Loss in a similar way to yours. I think there are two reasons for the difference in performance:

  1. Due to the limitation of computing resources, I performed ablation experiments on a subset of semantickitti datasets, which may be slightly different from the full datasets.
  2. The backbone I adopted in the framework is weaker than yours, and the effect is more obvious after using the BD Loss.

The code for this project is still being worked out. The visualization part is modified from Open3D-Semantic-KITTI-Vis(based on open3d) and spvnas(based on mayavi), you can refer to it first.

huixiancheng commented 2 years ago

thx for your advice~

huixiancheng commented 2 years ago

It should be clarified that recently I run a strict ablation experiment again and the results show that BD loss actually brings a 2.0 miou improvement. Maybe I made a mistake before. Thank you again for your advice and apologize for any possible misleading information.