DID

Diagrammatic Image Dataset for Architectural Machine Learning

Carnegie Mellon University

Pittsburgh, PA, USA

2019

This research proposes a system and method for generating image datasets for identifying design patterns. Diagrammatic image dataset is a collection of abstracted images synthesized with specific relationships that the user intends to train the neural network to identify patterns in a design. The diagrammatic images are derived from digital geospatial data to abstract information in terms of shapes, colors, and line types. The images are in turn output as rasterized two-dimensional data format that can be used for training neural networks. Generating diagrammatic dataset for identifying design patterns can overcome the challenges of excess noise in the data due to the variegated nature of readily available geospatial information. In addition, this method can allow the users to customize the relationships he/she is interested in training the neural networks with.

Collaborators:
Jinmo Rhee, Pedro Veloso