Semantic segmentation of AHN3 point clouds using a PyTorch implementation of Dynamic Graph CNN for Learning on Point Clouds (DGCNN)
Compared to the original PyTorch implementation, several changes are made in this repository to enable training and testing DGCNN on AHN3 point cloud data:
- prepare_data/extract_ahn3_annotations.py: convert AHN3 point clouds into S3DIS data format (.npy). The AHN3 dataset is already converted from .las to .txt in CloudCompare.
- data.py: add S3DISDataset class to load points block by block, instead of using large hdf5 files.
- main_semseg.py: add code to output segmented point cloud with labels (.txt), which can be used for later visualization.
- postprocess_data: add code for the multi-scale combination, by using results generated with different block sizes and k values in DGCNN.
| block size (m) | k | overall accuracy (%) | average per-class accuracy (%) | mean IoU (%) |
|---|---|---|---|---|
| 30 | 20 | 91.72 | 81.53 | 74.94 |
| 50 | 20 | 93.28 | 89.39 | 81.73 |
| 50 | 15 | 92.38 | 88.51 | 79.98 |
| 30 & 50 | 20 | 93.51 | 91.60 | 82.34 |
| 50 | 15&20 | 93.37 | 90.48 | 82.46 |
More Detailed description and results can be found in the project report.