This repository contains a PyTorch implementation of historgram and RANSAC Pooling layers for robust point cloud classification, segmentation, and normal vector estimation. Both layers were tested with PointNet and DGCNN. More details can be found in this paper. This code is mostly based on Deep Declarative Netwroks. For requirements, please refere to their repository.
If you find our work useful in your research, please consider citing:
@article{mukhaimar2023robust,
title={Robust pooling through the data mode},
author={Mukhaimar, Ayman and Tennakoon, Ruwan and Hoseinnezhad, Reza and Lai, Chow Yin and Bab-Hadiashar, Alireza},
journal={Intelligent Systems with Applications},
volume={17},
pages={200162},
year={2023},
publisher={Elsevier}
- Install PyTorch (pytorch.org)
pip install -r requirements.txt
coming soon...
To train a model, run main.py:
python3 main.py --outlier_fraction [OUTLIER_FRACTION] --noise_level [NOISE_LEVEL] --miss_fraction [MISS_FRACTION] --robust_type [ROBUST_TYPE]The ROBUST_TYPE can be RANSAC or Histogram {'RN', 'HS'} or none for max-pooling. Other pooling types introduced in DDN {'Q', 'PH', 'H', 'W', 'TQ'} can be used but the DDN library path needs to be added first.
Default OUTLIER_FRACTION, NOISE_LEVEL, and MISS_FRACTION values are zero. Outliers are normally distributed in the unit cude.