3D_pointclouds_super-resolution

Copyright (C) 2018

Rémi Petitpierre, Anna Fredriksson Häägg, Martin Sigmund

Ecole polytechnique fédérale de Lausanne – EPFL

GNU General Public License v 3.0

https://github.com/RPetitpierre/3D_pointclouds_super-resolution

Jupyter notebook which will run the whole model by calling the various different functions. This notebook also directly contains all the machine learning part of the project, including the neural network as well as the training and testing algorithms, that we implemented

The functions contained in this python file help to cut the massive dataset in individual raster files (.ras format, composed of zeros and ones). They were directly furnished by the hosting lab (see reference below).

Functions :

• ml_raster_export(ml_data, ml_path)
• ml_raster_import(ml_path, ml_width)
• ml_dataset_import(ml_args_dataset, ml_args_raster, ml_args_width)

Taken from : turing-project, Nils Hamel Copyright (C) 2016-2018 DHLAB, EPFL GNU General Public License v 3.0 https://github.com/nils-hamel/turing-project.git

These functions will help to import rasters contained in .ras files and load them in the form of numpy arrays. They were directly furnished by the hosting lab (see reference below).

Functions :

• ml_raster_import(ml_path)
• get_ml_data(ml_args_raster)

Adapted from : turing-project, Nils Hamel Copyright (C) 2016-2018 DHLAB, EPFL GNU General Public License v 3.0 https://github.com/nils-hamel/turing-project.git

The function it contains helps to convert rasters (composed of zeros and ones) into points coordinates contained in a 3D numpy array. This function was directly furnished by the hosting lab (see reference below).

Functions :

• ml_raster_convert(ml_raster)

Taken from : turing-project, Nils Hamel Copyright (C) 2016-2018 DHLAB, EPFL GNU General Public License v 3.0 https://github.com/nils-hamel/turing-project.git

Contains various functions used for post-processing and displaying outputs.

Functions:

• find_ground(flatten, position, size)
• flatten_ground(size, raster)
• complete_ground(raster)
• print_3D(y, n, m, mode)
• print_3D_pred(y, x, n = 0, advanced_mode = True)

This python file contains the arguments (global variables) that are used by the other functions dealing with the raster files.

Functions : init()

Will automatically be created by Python when running the code.

This folder is the supposed location of the full dataset.

This folder contains the model file which is created during the training process.

This folder is subdivided between the high_res and low_res folders which are meant to contain the individual rasters once they have been extracted (see extract_dataset.py).

Just contains the beautiful cnn_model image displayed in the Jupyter notebook.

For this project 3D LIDAR data of the city of Geneva in the year 2005 was used. The dataset belongs to the EPFL DHLAB, this is the reason why it is not furnished there. It is composed of 13597 rasters (volumetric elements), each encoding one part of the Geneva region. For each raster there are both a high- and low-resolution version. 32x32x32 binary pixels (high resolution) and 16x16x16 binary pixels (low resolution). Each raster covers roughly 153 m in width, length and height with a precision of roughly 4,8 m for the high resolution data and 9,6 m for the low-resolution data. Additional informations, including on the rest of the dataset (which was not used in this project) can be found at the following address : https://github.com/nils-hamel/turing-project/blob/master/README.md

In order to run the program, you will need to have the following libraries installed :

Python core libraries :

• random
• os
• sys

External libraries :

• torch "Pytorch" Deep convolutional neural network implementation and evaluation https://pytorch.org/docs/stable/index.html

• sklearn "Scikit learn" Splitting test and train set https://scikit-learn.org/stable/documentation.html

• numpy 3D data handling and regular computation https://docs.scipy.org/doc/

• matplotlib Visualization of 3D data https://matplotlib.org/contents.html

• IPython.display Image display in the Jupyter notebook https://ipython.org/ipython-doc/3/api/generated/IPython.display.html

The project runs on Mac OSX. Please consider that the pytorch library is still unstable on Windows and the code may therefore not run on certain PCs.