Fill the form and we’ll contact you soon

Your information has been received.
Looks like we're having trouble

Fill the form and we’ll contact you soon

Your information has been received.
Looks like we're having trouble

Case Study

Roof Detection Algorithm

We’re not a regular consulting firm. More than Data & AI suppliers, we are reliable partners. Both transparency and independence are at the core of how we work and guide the development of strong and reliable partnerships with our clients.

Contact Us

Overview

One of our coolest projects focused on targeting roofs on satellite imaging for a utilities customer. We were not only targeting roofs but also looking at any obstruction that might prevent the setup of solar panels or other devices - we achieved an IOU of 0.7 using a U-NET architecture, producing outstanding results in less than one month.

Challenge

While having few tagged data was an issue in the past, with most state-of-art computer vision models, that is not the case.

Problem

Vision model with generalization power. Our POC ended up opening up a ton of new possibilities for our customer with few risk and investment.

Approach

Just with 60 tagged images, we were able more models: obstruction detection and to produce a really performant computer roof facing. Vision model with generalization power. Our POC ended up opening up a ton of new possibilities for our customer with few risk and investment.

Computer Vision

Our model is based on a U-Net While having few tagged data was an convolutional architecture. With just a issue in the past, with most state-of-art couple of weeks we were able to produce a computer vision models that is not the really decent baseline for roof detection. case. During the project, we also developed two Just with 60 tagged images, we were able more models: obstruction detection and to produce a really performant computer roof facing.

Machine Learning Model

Time Line

01

arrowelipse

Image tags and Model Training.

02

arrowelipse

Checkpoint

03

arrowelipse

Complementary model creation, north roof and obstruction.

04

arrowelipse

Roof detection model delivered

arrowtimeline

Jan

Feb

Mar

Apr

Key Insights

Our POC opened up several new possibilities for our customer.

With a U-Net architecture, we were able to train different computer vision models for different targets really quickly,

60 tagged images were Enough for a satisfactory POC