- Python >= 3.7
- PyTorch >= 1.2
- CUDA >= 10.0
- Package: glob, h5py, sklearn, plyfile,tensorboardX
The network architecture of GTNet is shown below:
- GTNet is mainly composed of Graph Transformer, which is mainly divided into Local Transformer and Global Transformer:
- Dynamic graph update process:
python main_cls.py
python main_cls.py --exp_name=cls_eval --eval=True --model_path='xxx'
ModelNet40 dataset
| Mean Class Acc | Overall Acc | |
|---|---|---|
| This repo (2048 points) | 91.2 | 93.6 |
Note: The training modes 'full dataset' and 'with class choice' are different.
python main_partseg.py --exp_name=partseg
- With class choice, for example Earphone python main_partseg.py --exp_name=partseg_airplane --class_choice=Earphone
python main_partseg.py --exp_name=partseg_eval --eval=True --model_path='xxx'
ShapeNet part dataset
| Mean IoU | Airplane | Bag | Cap | Car | Chair | Earphone | Guitar | Knife | Lamp | Laptop | Motor | Mug | Pistol | Rocket | Skateboard | Table | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| This repo | 85.1 | 84.1 | 77.7 | 82.7 | 77.4 | 91.0 | 76.3 | 91.8 | 86.5 | 83.5 | 96.1 | 58.5 | 92.4 | 81.9 | 53.5 | 76.6 | 82.9 |
Both .txt and .ply file can be loaded into MeshLab for visualization. For the usage of MeshLab on .txt file. The .ply file can be directly loaded into MeshLab by dragging. python main_partseg.py --exp_name=partseg_eval --eval=True --model_path='xxx' --visu=Earphone_0 --visu_format=ply
The visualization result:
You have to download Stanford3dDataset_v1.2_Aligned_Version.zip manually from https://goo.gl/forms/4SoGp4KtH1jfRqEj2 and place it under data/
This task uses 6-fold training, such that 6 models are trained leaving 1 of 6 areas as the testing area for each model. python main_semseg.py --test_area=1 python main_semseg.py --test_area=2 python main_semseg.py --test_area=3 ... python main_semseg.py --test_area=6
- Evaluate in all areas after 6 models are trained python main_semseg.py --exp_name=semseg_eval --test_area=all --eval=True --model_root='xxx'
Stanford Large-Scale 3D Indoor Spaces Dataset (S3DIS) dataset
| Mean IoU | Overall Acc | |
|---|---|---|
| This repo | 64.3 | 86.6 |
Both .txt and .ply file can be loaded into MeshLab for visualization. For the usage of MeshLab on .txt file. The .ply file can be directly loaded into MeshLab by dragging. python main_semseg_s3dis.py --exp_name=semseg_eval --test_area=all --eval=True --model_root='xxx' --visu_format=ply
The visualization result:
Color map:
GTNet code refers to the following codes: