roboticslaburjc-org/lidar-segmentation-beyond-road

LiDAR Segmentation Beyond the Road: Performance and Robustness in Off-Road Settings

David Pascual-Hernández, Roberto Calvo-Palomino, Inmaculada Mora-Jiménez, Jose María Cañas-Plaza

This repository provides model weights and evaluation dataset for the paper LiDAR Segmentation Beyond the Road: Performance and Robustness in Off-Road Settings:

• dataset/: labeled LiDAR scans from proprietary sensor data provided by Celestia|TST.
• weights/: trained model weights for all experiments in the paper.

Weights naming

<dataset>-<model>-<input_features>-<data_augmentation>

• Dataset: training dataset used.
• Model: KPConv, RandLANet, MinkUNet, SPVCNN, Cylinder3D, SphereFormer, or LSK3DNet.
• Input features: point features used as input (xyzor xyzi if intensity is included).
• Data augmentation: none, light or strong.

Models have been finetuned from the pretrained weights provided by:

• Open3D-ML → KPConv, RandLANet
• mmdetection3d → MinkUNet, SPVCNN, Cylinder3D
• Official repositories → SphereFormer, LSK3DNet