A VoxelNet like approach for Shape Classification that works directly on PointClouds. PoxelNet is a combination of Point-based and Volumetric approaches that utilizes the advantages of point-processing and aggregating local information. The Volumetric approach is a Sparse Tensor Network that utilizes Minkowski Engine for efficient sparse operations.

The user needs to have GPU enabled machine with Linux Environment in order to run MinkowskiEngine effectively.

• Python version >=3.6
• Linux Machine (requirement of Minkowski Engine)
• GPU enabled device for inference

• Pytorch (>= 1.7)
• MinkowskiEngine
• numpy
• scikit-learn

Optional Requirements for Visualization and creating PointCloud dataset

• trimesh
• matplotlib

The dataset is in the form of 3D CAD models.

Download the dataset and put it into the dataset/ folder
Links for downloading -

• ModelNet 40: http://modelnet.cs.princeton.edu/ModelNet40.zip

The files can be viewed using -

• Any text editor: For seeing vertex and face information
• MeshLab: For seeing the CAD model
• We have also provided helper functions in utils.py that allows the user to inspect the meshes using trimesh.

PoxelNet works on 3D point clouds and not meshes. Hence, random sampling of the points is required over the 3D meshes. create-point_cloud.py file does the work of sampling. In order to get the PointClouds, the user just needs to run the following command -

python create_point_cloud.py --path=<Path_to_Original_ModelNet40_dataset> --N=<Number_of_points>

The default value of N is set to 4096

In order to directly get the points, we have provided links to Point Clouds created using Uniform Sampling from the ModelNet40 dataset.

• Training data with 4096 points
• Testing data with 4096 points
• Training data with 2048 points
• Testing data with 2048 points

Note: In order to do training/testing, the code requires you to provide just the name of the directory. Ensure that "train.npy" and "test.npy" files exist within this directory with exactly the same name. (Rename files with 2048 points accordingly.)

In order to train the model from scratch, the user needs to run the following command -

python train.py --path_to_dataset=<path_to_point_cloud_dataset> --store_weights=<path_where_training_weights_will_be_stored>

In order to resume training of the model from a previous checkpoint, the user needs to run -
python train.py --path_to_dataset=<path_to_dataset> --store_weights=<path_for_storing_weights> --load_pretrained --store_weights=<path_to_previous_checkpoint>

More optional arguments and their descriptions that can be passed to the functions are included in the file. You can see them using the --help flag

The testing of the model can be performed via the test.py file script. The user needs to have pretrained model stored in order to make the predictions directly without training
python test.py --path_to_dataset=<path_to_dataset> --path_weights=<path_to_the_pretrained_model>

Note: The path to dataset should include the path to the directory. The directory then should contain the train.npy and test.npy files which are provided above. The files include the normalized pointcloud for every Modelnet40 object along with corresponding labels

Note: To test the provided poxelnet_2048.pth model (Model trained with 2048 points), make sure to use the dataset with 2048 points as train.npy and test.npy. Similarly, to test the provided poxelnet_4096.pth model (Model trained with 4096 points), make sure to use the dataset with 4096 points as train.npy and test.npy.