The Core of our Tech

Advanced Computer Vision

Our app utilizes computer vision to accurately identify windows in real-time, enhancing measurement accuracy and speed by instantly analyzing visual data from your smartphone's camera.

Machine Learning

Leverage machine learning algorithms, our system continuously improves, learning from thousands of window measurements to increase precision and efficiency predict dimension, even in complex environments.

LiDAR / Sensors

LiDAR uses laser pulses for details surfaces mapping, prioritizing precision over colour vision.  Our adaptable technology ensures remarkable accuracy, using your phone's cameras and motion sensors even on devices without LiDAR.

Learn about our Algorithm:
Beam Sight

Our algorithm isn't just special - it's at the forefront of our window measurement technology, crafted to ensure reliability, adaptability, and accuracy in every scan.


Through a blend of tech-savviness and meticulous data curation, our algorithm redefines whats possible in window measurement technology.  It's not just special; its a game-changer.

Photo%20gallery.svg
Extensive Learning from Over 17k Images
Brightness.svg
Embracing Diversity for Real-World Accuracy
Balance.svg
Ensuring Fairness with Balanced Classes
Note.svg
Accurate Annotations for Super Learning
Group.svg
Prioritizing High-Quality Data
Group%20-1-.svg
Adapting to Variability for Robust Recognition
Icon.svg

Help Beam Sight Learn

Share your feedback on the app's performance to help us enhance and refine our data models.  Each piece of feedback continue to train our algorithms, making measurements even more accurate and user-friendly for everyone.

Perfecting Precision:
​Training Our Data Model

Rendell Bernardes

Chief Technology Officer 

Overview
Training Data
Overview

Training a model requires extensive data and meticulous annotation curation. Our dataset includes over five thousand images covering a wide variety of window types, enabling us to achieve over 75% accuracy in detection. During the process, we underwent multiple fine-tuning phases, iterating on the parameters to attain the best possible detection results.

Training Data
Using high-quality tools to support our annotation process has allowed us to achieve excellent training results. Our goal is to enable the computer to discover optimal parameters and patterns and make data-driven decisions. But how do we ensure the machine makes accurate predictions on images? We review the tests to see if the results are acceptable; if not, we need to check the data input into the model. For example, if you want to detect windows in Canadian houses, it’s not advisable to include images of houses from Brazil.

In such scenarios, our engineers step in, dividing the data into three parts: Training, Validation, and Testing. To achieve superior performance, we employ deep neural network algorithms. These are particularly effective when is challenging to identify the right metrics, especially when these exhibit chaotic behaviour.