HPGNN - A Hierarchical Graph Neural Network for Semantic Segmentation of Large Scale Outdoor Point Clouds

Official implementation of the paper 'HPGNN - A Hierarchical Graph Neural Network for Semantic Segmentation of Large Scale Outdoor Point Clouds' (ICPR 2022).

❗ It is not possible to retrain this model using this repo with later versions of tensorflow due to data loader differences. The inference can be used with provided checkpoints. We're expecting to port the code to a later version of tensorflow. Sorry for any inconvenience caused.

Requirements:

- tensorflow-gpu 2.5.0
- scikit-learn 0.24.2
- matplotlib 3.4.1
- tensorboard

Dataset folder structure

|-dataset
	|-sequences
        |-00
            |-velodyne
                |-000000.bin
                |-...
            |labels
                |-000000.labels
                |-...
        |-... 
	|-nuscenes
		|-velo
			|-n008-2018-05-21-11-06-59-0400__LIDAR_TOP__1526915243047392.pcd.bin
			|-...
		|-lidarseg
			|-0a0c9ff1674645fdab2cf6d7308b9269_lidarseg.bin
			|-...

Combine all 'pcd.bin' files in 'samples' folders of NuScenes dataset into the 'velo' folder. Training, validation, and testing without overlap is handled by the split files.

Training

Initialize the environment

conda create -n hpgnn python=3.8
conda activate hpgnn
pip install tensorflow-gpu==2.5.0 scikit-learn==0.24.2 matplotlib==3.4.1 tensorboard

KITTI dataset

python train_kitti.py --config hpgnn_kitti

NuScenes dataset

python train_nuscenes.py --config hpgnn_nuscenes

Inference

KITTI dataset

python inference_kitti.py --config hpgnn_kitti

NuScenes dataset

python inference_nuscenes.py --config hpgnn_nuscenes