Improved Training for 3D Point Cloud Classification

Structure

experiments/  *** includes all modified experiments ***
criterions/   *** all the loss functions ***
datasets/     *** all the datasets ***
main.py
readme.md

Datasets

Two datasets are used in this project:

Dataset Name	Download Link
ModelNet10	Link
ModelNet40	Link

Please download the dataset and unzip it in the datasets folder with following structure:

datasets/
    ModelNet10/
        bathtub/
            train/
                <.off files>
            test/
                <.off files>
        bed/
    ModelNet40/
        ....

How to run the code

To run the final best result of the mehtod, use the following script. This will give an accuracy of 91.52%.

python main.py

For different ablation, sensitivity analysis on hyper-parameters, and other experiments with model architecture, use the scripts in the experiments folder.

python experiments/lr_0.1.py

The experiment directory contains files for different experiments.

batch_size_8.py
batch_size_16.py
batch_size_32.py

.....

ALl the experiments' outputs are saved in the checkpoints folder.

batch_size_8.txt
batch_size_16.txt
batch_size_32.txt

.....

Most of the file names are intuitive and indicative of the experiments. The files that require description are:

main_exp_1.py to main_exp_10.py --> experiments with different version of 1D CNNs. Description is included in each file.

Run time

The code requires a Colab GPU machine around 10 hours to run the default 15 epochs settings.

Acknowledgements

This project was build using the help from the following sources and repositories:
https://github.com/fxia22/pointnet.pytorch --> PointNet Implementation
https://github.com/nikitakaraevv/pointnet --> PointNet Implementation
https://discuss.pytorch.org/t/is-this-a-correct-implementation-for-focal-loss-in-pytorch/43327/22 --> Focul loss