GT-MilliNoise: Graph Transformer for Point-wise Denoising of Indoor Millimeter-Wave Point Clouds

(Paper under submission) [Project] [Paper]

This GitHub contains the code to implement GT-MilliNoise architecture for millimetre-wave point cloud denoising using the MilliNoise dataset.
This Github also provides benchmarking on state-of-art architecture (PointNet, PointNet++, DGCNN, Transformer) on MilliNoise point cloud denoising.

MillNoise Dataset: [Dataset] [Paper]

Overview

Citation

Please cite this paper if you want to use it in your work,

@article{coming soon,
  }

Installation

Install TensorFlow. The code has been tested with Python 3.6, TensorFlow 1.10.0, CUDA 9.0 and cuDNN 7.3

Compile the code. You must select the correct CUDA version and install Tensorflow on your computer. For that edit the Makefiles to the paths of your Cuda and Tensorflow directories. The Makefiles to compile the code are in modules/tf_ops

Directory paths

You must specific the following directory (parse as input to train.py and test.py)

1. --data-dir: Path were the dataset is stored
2. --log-dir: Path were the model outputs should be saved.

Usage

To train a model for millimeter-wave point cloud denoising:

python train.py --version <name_of_model> --data-split <#split> --model <architecture_name>  --seq-length <input_frames>

For example:

python train.py --version v0 --data-split 4 --model GT  --seq-length 12 

Trains a GT-Millinoise model using dataset split #4 (Fold-4)

To evaluate the model

python test.py --version v0 --data-split 4 --model GT --seq-length 12 --manual-restore 2 

--manual-restore 2: loads best model in validation, --manual-restore 1 allows to choose a specific checkpoint

Splits:

#Split:	K-Fold:	Training Scenarios	Test-Scenarios
--data-split 11	Fold 1-6	All [1-6]	All [1-6]
--data-split 17	Fold 1,2,3	[4,5,6]	[1,2,3]
--data-split 4	Fold 4	[1,2,3,5,6]	[4]
--data-split 16	Fold 5	[1,2,3,4,6]	[5]

Available Models:

Model	Name	Description
--model GT	GT-Millinoise	Standart GT-Millinoise architecture
--model PointNet	PoinNet
--model PointNet_2	PointNet++
--model DGCNN	Dynamci Graph CNN (DGCNN)
--model Transformer	Vannila-Transformer
--model GT_intensity	GT-Millinoise	GT-Millinoise with intensity as input
--model GT_velocity	GT-Millinoise	GT-Millinoise with velocity as input
--model GT_noTC	GT-Millinoise	GT-Millinoise with without temporal block

Datasets

The models were evaluated with the following datasets:

1. Complete Raw Data (JSON)   2. Pre-processed Data (Numpy)  

We provide the dataset in two formats.

1. The original JSON data with complete information (intensity, velocity, roto-translation coordinates).
2. Pre-processed data converted to numpy, used in our experimentation

Acknowledgement

The parts of this codebase are borrowed from Related Repos:

Related Repos

1. PointRNN TensorFlow implementation: https://github.com/hehefan/PointRNN
2. PointNet++ TensorFlow implementation: https://github.com/charlesq34/pointnet2
3. Dynamic Graph CNN for Learning on Point Clouds https://github.com/WangYueFt/dgcnn
4. Millinoise https://github.com/c3lab/MilliNoise