This repo is implementation for PointNet in pytorch. The model is in pointnet/model.py.
The project is forked from point.pytorch
git clone https://github.com/nicolgo/pointnet.pytorch.git
cd pointnet.pytorch
pip install -e .
Download and build visualization tool (Linux)
cd script
bash build.sh #build C++ code for visualization
bash download.sh #download dataset
Build dynamic library on Windows
cd utils
mkdir build & cd build
cmake -DCMAKE_GENERATOR_PLATFORM=x64 .. or cmake -A x64 ..
cmake --build .
Training(for modelnet40, dataset path is useless since we rewrite the code getting modelnet40 dataset)
cd utils
python train_classification.py --dataset <dataset path> --nepoch=<number epochs> --dataset_type <modelnet40 | shapenet>
python train_segmentation.py --dataset <dataset path> --nepoch=<number epochs>
Use --feature_transform to use feature transform.
change the network with three symmetric functions(max/min/mean), just disable the comment of the code below in model.py
# self.use_more_features = True
if self.use_more_features:
On ModelNet40:
| Overall Acc | |
|---|---|
| Original implementation | 89.2 |
| this implementation(w/o feature transform) | 89.99 |
| this implementation(w/ feature transform) | 89.59 |
| Overall Acc | |
|---|---|
| Original implementation | N/A |
| this implementation(w/o feature transform) | 97.59 |
| this implementation(w/ feature transform) | 96.97 |
Segmentation on A subset of shapenet.
| Class(mIOU) | Airplane | Bag | Cap | Car | Chair | Earphone | Guitar | Knife | Lamp | Laptop | Motorbike | Mug | Pistol | Rocket | Skateboard | Table |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Original implementation | 83.4 | 78.7 | 82.5 | 74.9 | 89.6 | 73.0 | 91.5 | 85.9 | 80.8 | 95.3 | 65.2 | 93.0 | 81.2 | 57.9 | 72.8 | 80.6 |
| this implementation(w/o feature transform) | 73.5 | 71.3 | 64.3 | 61.1 | 87.2 | 69.5 | 86.1 | 81.6 | 77.4 | 92.7 | 41.3 | 86.5 | 78.2 | 41.2 | 61.0 | 81.1 |
| this implementation(w/ feature transform) | 87.6 | 81.0 |
Note that this implementation trains each class separately, so classes with fewer data will have slightly lower performance than reference implementation.
Sample segmentation result:
