PointCNN using Pytorch

Dataset: ModelNet40, The UoY 3D Face Dataset

Reference

This work is based on hxdengBerkeley's PointCNN.Pytorch. A Pytorch implementation of PointCNN.

The UoY 3D Face Dataset was used to test the PointCNN's performance of matching facial masks to faces. The data in this dataset has been manually preprocessed and labelled. Preprocessing includes aligning nose area of each face to the same location. This facilitates the later automatic labelling.

Improvements

Many improvements have been done to increase the GPU efficiency (was 95% of the time at 0% on GPU but CPU was consistantly busy) and significantly speed up the training process. Most of the calculations in model have been move to GPU but they can still be switched back to CPU. When training on ModelNet40 Dataset, a minimum of 3.5Gb GPU MEM are expected. While there are some CPU compulsory calculations, the overall GPU efficiency is around 85%.

• Move a majority of computational workload to GPU from CPU
• Rewrite the self-defined data loader to torch.utils.data.Dataset and torch.utils.data.Dataloader
• Add adam as an alternative optimizer
• Add generating and saving plot of train/test accuracy and train loss
• Add UoY
• Adjustable train set size (randomly select data but considering distribution over classes)(Under implementation)

Environment

This project relies on a variety of packages. To make the environment building-up process more straightforward, a configuration file, xconv_conda_env_setup.yml, is provided. It is recommended to use Anaconda to install the environment.

For getting and installing Anaconda, Please visit Anaconda's official website to download and install the Anaconda on your devices.

Setting up ENV and Training

If you already have the ENV set up, simply jump over step 2.

1. cd to the project directory (where xconv_conda_env_setup.yml is located at)
2. In your terminal, use conda create --name env_name --file xconv_conda_env_setup.yml. Replace env_name to your desired environment name.
3. In your terminal, use conda activate env_name to switch to this environment.
4. Prepare data according to yangyanli/PointCNN (Modelnet40)
5. In your terminal, use python train_pytorch_modelnet40.py to start training. There are multiple commandline arguments acceptable. (run python train_pytorch_dataloader.py for preprocessed UoY dataset)
6. Have fun

Result

After preprocessing on UoY dataset, there are 250 faces remain in the dataset. 10 of them were randomly picked up as the standards for generating 10 labels. The result is shown in the graph below.