This is a PyTorch implementation of PointNet (CVPR 2017), with comprehensive experiments.
Refer to requirements.txt for common dependencies which are fairly easy to install via conda or pip. You may also need to install PyMesh. See here for instructions to install.
The code supports Python3 and PyTorch 0.4.1+.
This code implements object classification on ModelNet, shape part segmentation on ShapeNet and indoor scene semantic segmentation on the Stanford 3D dataset.
Download the ModelNet10 dataset from here or the ModelNet40 dataset from here. Unzip and run
python train_cls.py -dset modelnet40 -r modelnet_root_dir -np number_of_points_to_sample
Download the ShapeNet dataset from here. Unzip and run
python train_seg.py -dset shapenet16 -r shapenet_root_dir -np number_of_points_to_sample
Download the S3DIS dataset from here (you need to submit a request). Unzip and do
cd Stanford3dDataset_v1.2
mkdir train test
mv Area_1 Area_2 Area_3 Area_4 Area_6 train
mv Area_5 test
to create train/test split. Then set gen_labels=True in the class S3dDataset in datasets.py and do
python datasets.py
to generate labels for the train and test set respectively. After that always set gen_labels=False. With labels generated do
python train_seg.py -dset s3dis -r s3dis_root_dir -np number_of_points_to_sample
to start training.
First do sh build.sh, then use show_seg.py to visualize segmented object parts. Below are some example results.
For S3DIS, you have to combine scene components along with their labels into one text file (cat and paste seems to be an easy way to do this) and then pass it to show_seg_s3dis.py. Below are some example results (removed some clutter classes for better visualization).
Certain design choices in the original paper are not implemented here for simplicity. There is some performance gap on ModelNet classification, for ShapeNet and S3DIS seems to be on par with the original paper.
| accuracy | |
|---|---|
| ModelNet10 | 87.2% |
| ModelNet40 | 85.4% |
| accuracy | class avg IoU | |
|---|---|---|
| ShapeNet | - | 82.9% |
| accuracy | class avg IoU | |
|---|---|---|
| S3DIS | 72.1% | 50.6% |
pointnet.pytorch (many thanks)