search

hustvl / Symphonies

[CVPR 2024] Symphonies (Scene-from-Insts): Symphonize 3D Semantic Scene Completion with Contextual Instance Queries
https://arxiv.org/abs/2306.15670
MIT License
158 stars 7 forks source link

Kitti360 ground truth labels #14

Closed Iana-Zhura closed 6 months ago

Iana-Zhura commented 6 months ago

Hello! Thank you for your interesting work! I have a question regarding kitti360 dataset. Did you use ground truth labels provided by SSCbench (from here https://github.com/ai4ce/SSCBench/tree/main/dataset/KITTI-360)? Did you encounter a problem that there are unlabeled areas in ground truth? Can you share please what preprocessing steps you have done with kitti360 labels before training. Thank you in advance!

npurson commented 6 months ago

We indeed encountered the same issue where a considerable portion of the areas were labeled as "ignored". We did not perform any preprocessing and instead trained and evaluated the model with the ignored voxels masked, as is typically done.

But for visualization, we found it necessary to replace the predictions of gt ignored voxels with the "empty" class to generate interpretable visualizations. Below, we present our visualization results on SSCBench-KITTI-360:

Prediction: 000190_pred

Ground Truth: 000190_gt

I hope this helps!

Iana-Zhura commented 6 months ago

Thank you for your quick reply! Yes, it helps:)