megayeye / triangle-net

(paper code) Triangle-Net: Towards Robustness in Point Cloud Learning

Classification

1. Prepare data

   For ModelNet 40, download dataset from https://shapenet.cs.stanford.edu/media/modelnet40_ply_hdf5_2048.zip

   Next, extract it to <code folder>/data/modelnet40_ply_hdf5_2048 folder.

   For ScanObjectNN, please get the download link according to the instruction from this link (https://hkust-vgd.github.io/scanobjectnn/). Then, extract training_objectdataset_augmentedrot_scale75.h5 and test_objectdataset_augmentedrot_scale75.h5 to <code folder>/data/ScanObjectNN_nobg.

2. training

   for training on ModelNet40 with reconstruction network:

   python train_recon.py
   

   for training on ModelNet40 without reconstruction network, of which the training is faster at a cost of minor accuracy drop:

   python train_wo_recon.py
   

   For both of the training configuration, --n_points can specify the number of points.

   For training on ScanObjectNN:

   python train_scanobjects.py
   

Segmentation

1. Prepare data

   Download dataset from: https://shapenet.cs.stanford.edu/media/shapenetcore_partanno_segmentation_benchmark_v0_normal.zip

   Next, extract it to <code folder>/data/shapenetcore_partanno_segmentation_benchmark_v0_normal folder

2. Preprocessing

   To accelerate disk IO, we save the dataset as npy files:

   python segment_data_preprocess.py
   

3. Training

   python train_partseg.py
   

Comparison experiment

We refer the following code for comparison experiments

[PointNet & PointNet++] https://github.com/yanx27/Pointnet_Pointnet2_pytorch

[DGCNN] https://github.com/WangYueFt/dgcnn

[RI-CONV] https://github.com/hkust-vgd/riconv

[3DmFV] https://github.com/sitzikbs/3DmFV-Net