geometry clustering of shapes based on trained features
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
Dependencies
Documentation
Final
Working folder moving, adding treesloth as additional dependency Oct 23, 2017
.gitignore
LICENSE
README.md

README.md

metashape

geometry clustering of shapes based on trained features

The goal is to get a set of geometries and cluster them to see patterns in geometric shapes.

Initial Primitives

As a prototype, we developed a sample set of 16 elements that were categorized into cones, cylinders, spheres, and boxes.

Initial Clustering

This set was trained in scikit-learn and Accord.NET through dodo, lunchbox, owl. Owl fit our needs in predictive modeling most effectively.

Shape Analysis

Shape grammar has been notoriously difficult to quantify over the years manually, so machine learning is a great application of this

Consistent Geometry Analysis

First, we take a dodecahedron to reduce polar biases and perform raycasting to determine vertex distances.

We also create a bounding box and then slice the building into upper, middle, and low slices as well as into levels describing terraces.

Additionally, we develop a contour and create a mesh to consistently calculate volume and area of the geometry.

Geometric Properties

The general properties we have focused on:

  • volume
  • surface area
  • centroids
  • orientation
  • ratios of properties

Machine Learning

To categorize the building shapes, we used a clustering algorithm that would automatically group buildings based on their geometric properties.

For clustering algorithms, we initially tried a regression classification to test plugins, but decided that Owl was the best solution for our use case. We used the following features:

  • Area
  • Aspect Ratios
  • Centroid
  • Eccentricity
  • Orientation
  • Slice Area
  • Surface to Volume Ratio
  • Terraces
  • Vertex Mean Distance
  • Vertex Farthest Distance
  • Raycasting
  • Volume to Bounding Box Ratio
  • Volume

These provided us with enough features to distinguish differences between surrounding buildings.

Future Work

We hope to include further feature analysis for our machine learning algorithms.