fayolle / csg_recovery_py

CSG recovery from a 3D point-cloud (Python).

Point-cloud to CSG

Simple implementation of the genetic programming code used in this paper to generate a CSG expression from a 3D point-cloud. This is a pure Python version.

How to install / compile

Compile the segmentation program (based on Ransac) in the subdirectory 'segmentation'. It relies on the library boost. This code is a modified version of the code from the efficient RANSAC paper.

The CSG tree search program (based on genetic programming) in the subdirectory 'gp' is written in Python. It depends on NumPy only (see also the 'requirements.txt' file).

How to run

Assuming that the compiled programs for the segmentation is copied in the subdirectory 'bin'.

Step 1: Segmentation

bin/ransac_command data/test.xyzn out/test-segmented.segps [data/test.conf] > out/log.txt

The program 'ransac_command' has three arguments: the name of the input 3D point-cloud (first argument), the name of the output point-cloud with segmentation information (second argument), a config file (optional).

Step 2: Recovery step

Produce an expression for the CSG tree

python gp/gp.py out/test-segmented.fit data/test.xyzn out/test_expression.txt out/test_primitives.txt 150 3000 0.3 0.3

The program 'gp' has the following arguments:

• The name of the file with the list of primitives (computed by 'ransac_command')
• The name of the input 3D point-cloud
• The name of the file where the CSG expression will be written
• The name of the file where the list of primitives used in the CSG expression will be written
• The number of creatures per generation
• The maximumn number of generations
• The mutation rate
• The crossover rate

Step 3: C/C++ source file generation and evaluation

python utils/create_eval_source.py out/test-segmented.fit out/test_expression.txt out/test.cpp data/test.xyzn 

The C++ source code with the CSG expression can be found in the file 'out/test.cpp'. The function defined in the C++ file is an implicit surface (approximate SDF) that can be rendered by ray-tracing or after meshing with the Marching Cubes algorithm.

Note: It is possible to generate a tree (using Graphviz format) corresponding to the CSG expression.

python utils/tree_from_expression.py out/test_expression.txt out/test_primitives.txt out/test_tree.dot 

The output file is in 'out/test_tree.dot'. It can be processed with the 'dot' command of GraphViz.

Additional notes

Segmentation / Fitting

Obviously, the results depend a lot on the list of primitives passed to the CSG tree search. The version provided here uses RANSAC. There are several ways to improve the output of this segmentation/fitting step such as, for example: Section 4.1 of this paper, this paper or this one.

Improved CSG recovery methods

Several methods have been devised for improving and accelerating the CSG tree search: This work and this one still use genetic programming and propose several improvements. This work propose a different approach based on program synthesis.

Implementations can be found below:

• https://github.com/mafried/csg_playground/tree/wscg18
• https://github.com/mafried/csg_playground/tree/pointselection
• https://github.com/yijiangh/InverseCSG

CSG expression optimization

The CSG expression generated by the method above can be further optimized by considering methods such as the ones proposed in this paper.

An implementation can be found here:

• https://github.com/mafried/csg_playground/tree/optimization

Reference

Link to the paper where the approach is described and the corresponding bibtex entry

@article{pc2csg2016,
title = {An evolutionary approach to the extraction of object construction trees from 3D point clouds},
journal = {Computer-Aided Design},
volume = {74},
pages = {1-17},
year = {2016},
issn = {0010-4485},
author = {Fayolle, Pierre-Alain and Pasko, Alexander},
}