This project is developed based on the following three projects:
- https://github.com/facebookresearch/maskrcnn-benchmark.git
- https://github.com/facebookresearch/SparseConvNet.git
- https://github.com/traveller59/second.pytorch.git
Install required softwares according to the guide "./docs/Install.md".
Train or test by modifying and running the script "./run.sh".
- Input: point cloud of indoor building.
- Five objects: wall, window, door, ceiling, floor
- Output: 3D bounding boxes of objectis
The dataset is modified from SUNCG by the processing functions in: ./data3d/suncg_utils/suncg_preprocess.py
Some data samples can be downloaded in https://unsw-my.sharepoint.com/:f:/g/personal/z5105843_ad_unsw_edu_au/EtCliIZlMbFAotFQruB_bqYBsdAKxcJmk2TGJDOdNHGjNQ?e=PXjKmG
The training model is available in https://unsw-my.sharepoint.com/:f:/g/personal/z5105843_ad_unsw_edu_au/EqihMsW90rZKnWLM8fT-SI4B_3ueak_0TP66bLgBU_mm2Q?e=owjPb2
- Shortest wall instance: Long wall pieces are croped by intersected ones to generate short instances.
- time per building: 4.75 s
| Wall | Window | Door | Floor | Ceiling | Classes Mean | |
|---|---|---|---|---|---|---|
| AP(%) | 89.12 | 76.42 | 90.91 | 83.57 | 88.91 | 85.79 |
| AIoU(%) | 84.09 | 67.92 | 80.20 | 78.53 | 84.42 | 79.03 |
- Looking for cooperation to improve the quality of the dataset and rich it with real scanning data.
Each intance is colorized by a random color, except blue denotes incorrect detection or missed ground truth.
| Synthetic mesh | Point Cloud |
|---|---|
 |
 |
| Ground truth | Detection |
 |
 |
| Synthetic mesh | Point Cloud |
|---|---|
 |
 |
| Ground truth | Detection |
 |
 |
| Synthetic mesh | Point Cloud |
|---|---|
 |
 |
| Ground truth | Detection |
 |
 |
| Synthetic mesh | Point Cloud |
|---|---|
 |
 |
| Ground truth | Detection |
 |
 |