Clustering is back: Reaching state-of-the-art LiDAR instance segmentation without training

Official implementation of the method ALPINE. Check out the article for more details!

Clustering is back: Reaching state-of-the-art LiDAR instance segmentation without training, by Corentin Sautier, Gilles Puy, Alexandre Boulch, Renaud Marlet, and Vincent Lepetit

ALPINE takes semantic LiDAR predictions as input, and predicts instance masks, evaluated on LiDAR panoptic benchmarks.

If you use ALPINE in your research, please cite:

@article{sautier2025alpine,
  title = {Clustering is back: Reaching state-of-the-art {LiDAR} instance segmentation without training},
  author = {Corentin Sautier and Gilles Puy and Alexandre Boulch and Renaud Marlet and Vincent Lepetit},
  journal={arxiv},
  year = {2025}
}

Using Alpine in your own codebase

First install Alpine using:

pip install .

Alpine's usage is similar to other scikit-learn clustering:

from alpine import Alpine
# Provide your own point cloud X and semantic labels y
X = np.array([[0, 0], [1, 1], [2, 2], [3, 3]])
y = np.array([0, 0, 1, 1])
# Provide a list of 'thing' classes matching separable objects in your semantic labels
thing_indexes = [0, 1]
# Provide a dict of standard size of object for each 'thing' class
thing_bboxes = {0: [2., 2.], 1: [2., 2.]}
alpine = Alpine(thing_indexes, thing_bboxes, split=True)
instance_labels = alpine.fit_predict(X, y)

Other parameters include whether or not to activate the box splitting (default is false), the number of neighbors k and the margin of the box splitting.

Evaluation Alpine in our codebase

Datasets

This project has been adapted to SemanticKITTI, nuScenes and SemanticPOSS. Please use the following convention (or provide the --path_dataset argmument):

UNIT
├── datasets
│   ├── nuscenes
│   │   │── v1.0-trainval
│   │   │── samples
│   │   │── sweeps
│   ├── semantic_kitti
│   │   │── sequences
│   ├── semantic_poss
│   │   │── sequences

Use your own semantic masks

For SemanticKITTI and SemanticPOSS, you should preferably use the official prediction formatting requested by SemanticKITTI.

For nuScenes you can either use the semantic segmentation formatting or the panoptic segmentation formatting.

Running ALPINE

ALPINE can be ran with the command:

python alpine_semantickitti.py --path_to_files [path] --split

Other arguments can disable the box splitting for faster runtime at some performance cost, or save the predictions to disk.

Running ALPINE with WaffleIron predictions

To obtain semantic predictions on SemanticKITTI using WaffleIron, first install WaffleIron and its dependencies, then you can run the following code:

cd waffleiron

wget https://github.com/valeoai/WaffleIron/releases/download/v0.2.0/waffleiron_kitti.tar.gz
tar -xvzf waffleiron_kitti.tar.gz

python predict_semantickitti.py \
--path_dataset ../semantic_kitti/ \
--ckpt ./pretrained_models/WaffleIron-48-256__kitti/ckpt_last.pth \
--config WaffleIron-48-256__kitti.yaml \
--result_folder ./predictions_kitti \
--phase val \
--num_votes 12 \
--num_workers 6

Predicted semantic segments will be generated in waffleiron/predictions_kitti and can be evaluated running:

python alpine_semantickitti.py --path_to_files waffleiron/predictions_kitti --split

Ensembling

In the article, we ran ensembling by summing predicted semantic probabilities together. To obtain this behavior, you can provide multiple path_to_files with output probabilities. This ensembling method might not be effective for any combination of predictions.

Acknowledgement

We want to acknowledge the authors of MODEST from which we reuse a box fitting algorithm.

We incorporated a minimal working example of WaffleIron into this repository to provide working semantic masks.