Inofficial PyTorch implementation of Boosting LiDAR-based Semantic Labeling by Cross-Modal Training Data Generation (Piewak et al., 2018).
The Autolabeling process is currently not used, instead the converted KITTI data from SqueezeSeg is used. For better convergence we add batch normalization after each convolutional layer.
| Car | Pedestrian | Cyclist | mIoU | |
|---|---|---|---|---|
| SqueezeSeg | 64.6 | 21.8 | 25.1 | 37.2 |
| SqueezeSegV2 | 73.2 | 27.8 | 33.6 | 44.9 |
| LiLaNet | 67.6 | 36.9 | 31.9 | 45.5 |
- Install PyTorch (pytorch.org)
pip install -r requirements.txt- Download the KITTI Lidar dataset
Train model:
Important: The dataset-dir must contain the lidar_2d and the ImageSet folder.
python train_kitti.py --dataset-dir 'data/kitti'