IRISBOND – Mamu, Low Resolution Eye-Tracking Software
IRISBOND
Sector: Health
Business Case
The proposal involves the development of a computer vision-based eye tracking algorithm for mobile devices. Taking the capabilities of high-resolution eye tracking devices (Hiru) as a reference to try to replicate such functionalities by lowering the costs of having a dedicated hardware.
Objectives
The use case objective is to have a low-cost AI-based eye-tracking algorithm that allows the integration into any device (mobile or desktop) containing a webcam. This has to be, on the one hand, adaptable to the circumstances of the user (variability of lighting and facial features) and, on the other hand, light enough to be able to function in real time.
Use case
Since the solution has many modules that make use of AI, the effort was focused on the development of the neural network, adjusting it to the needs of our problem. For the less essential modules, ways were sought to obtain more generic open-source algorithms and off-the-shelf products.
Infrastructure
On Premise
Technology
Machine learning and deep learning Computer vision
Data
1,000,000 synthetic images with their gaze angle label (public generator). 10,000 real labelled images (angle).
Resources
At the organisational level, 1 project manager, 3 SW technicians (computer vision and microservices) were required. Technical development mainly by Vicomtech and Irisbond (algorithm), and technical validation, requirements, feasibility and integration with an Android app by Irisbond only.
Difficulties and learning
Constraints in obtaining real images, leading to the exploration of alternative data sources. Experimenting with third-party resources and algorithms and customising them to suit our specific problem. Overcoming the integration challenges of a computer vision (deep learning) algorithm on mobile devices with limited computational capabilities. A high level of model sensitivity to environmental changes, including variations in lighting and user positions. Changes in light and user position.
KPIs (business impact and metrics of the model)
Precisión de predicción: Sin calibración: Error en ángulos 6º, Error en cm a 40 cm 4,18 cm Con calibración: Error en ángulos 3º, Error en cm a 40 cm 2,09 cm
Funding
Hazitek, Samsung
Collaborators, Partners
Samsung, Vicomtech
