MICHAEL HASEY // AI & ARCHITECTURE

The fastest and easiest way to make custom, large, and robust architectural imagery datasets for AI deep learning projects

Create datasets for any architect, any building, any style, any location

year: 2020

type: database software app

GitHub: click for source code

ARCHI_BASE is a soon to be released online app that autonomously creates very large custom datasets of labelled and sorted architectural imagery to train deep learning neural networks.

With artificial intelligence changing the way we design and build in unprecedented ways, the accumulation of good and reliable data has never been so important.

Today’s Problem

The current method of creating large databases of architectural image data is inadequate and slow.

To create a custom image database today, a user must manually search through multiple and scattered image sources, filter through thousands of un-labelled and non-relevant images, download the correct images one by one, and then painstakingly label and sort each image into their corresponding folders to prepare for training.

“With todays current method, it would take around 30 hours of manual labour to create a small database of just 10,000 images”

The ARCHI_BASE solution

ARCHI_BASE solves this problem by providing an online tool that autonomously finds, labels, and sorts architectural images into custom datasets of any size in a fraction of the time.

Instead of taking 30 hours to manually create a 10,000 image dataset, ARCHI_BASE can complete this task in an hour with zero human interaction required.

Create datasets for any architect, any building, any style, any location.

archi_base diagram.jpg

ARCHI_BASE IS…

18.75 X FASTER AUTONOMOUS ACCURATE

How Does ARCHI_BASE work?

Overview

ARCHI_BASE employs a 3-step pipeline that leverages sophisticated AI & ML computer vision technology.

Step 1: Train a ResNet Image classification model to identify specific architectural image types.

Step 2: User identifies what to search for and how many images required, and our system autonomously scrapes multiple online image databases.

Step 3: The ResNet-32 model classifies each scraped image, labels it and then places it into its corresponding folder.

Your Image database is now ready to train deep neural networks!

“While a robust 100,000 image dataset would take 330 hours to create manually, Archi_base completes this task in just 16 hours.”

Deep Learning Neural Network Algorithms

“Neural networks are a set of algorithms, modelled loosely after the human brain, that are designed to recognize patterns. They interpret sensory data through a kind of machine perception, labeling or clustering raw input. The patterns they recognize are numerical, contained in vectors, into which all real-world data, be it images, sound, text or time series, must be translated.” - A.I. Wiki

ARCHI_BASE uses a ResNet-32 Image Classification model due to its superior image classification abilities over standard, “vanilla” neural networks. In other words, ResNet models can easily learn patterns within images and then associate them to a class such as “dog” or “cat”. Once trained, ResNet have the ability to determine whether new and unseen images belong to that class or not. Such computational autonomy to “understand” what is being shown has only recently been developed, and has forced us to reconsider our views of computer-human dependency.

For ARCHI_BASE, we taught our ResNet model to learn the patterns inherent within architectural images. Learned image patterns include those associated with building interiors, building exteriors, building aerial views, buildings with people in the frame, parts of buildings, and building textures. Once learned, our model can classify whether new images belong to any of these class categories and then label and sort them into folders accordingly.

Vanilla Convolutional Neural Networks

Limited Performance

Vanilla convolutional neural networks are some of the earliest image classification models used for deep learning. In order for these networks to perform well, image data is sent through a series of convolutional layers to extract exceedingly intricate patterns of information. As more patterns are extracted, the better the network can classify images. However, vanilla networks hit a road block when they failed to increase performance despite more layers being added. In fact, as more layers were added, the original data quickly degraded and led to increased error rates. As a response, the ResNet Image Classification model was created as a means to circumnavigate this problem and boost performance.

ResNet Image Classifiers

Superior Abilities

To overcome this problem, Microsoft Research developed the Residual Neural Network, or ResNet for short. By introducing shortcut connections, they allowed data flowing through the network to skip one or more layers therefore preserving the original data throughout the entire algorithmic process. As a result, the increasing error loss issue was solved and additional layers could be added into the network to extract deeper patterns of information.