IRISBOND – Hiru image labelling support tool
IRISBOND
Sector: ICT
Business Case
It is proposed to develop an image labelling support tool for machine vision hiru. Irisbond is working on generating an extensive database of images captured by the hiru on users. Currently, the process of labelling/validation of different visual features of the images was done manually, making the process very inefficient. This allows us to evaluate different versions of our pupil detection algorithm objectively.
Objectives
The objective of this internal tool is to facilitate the image labelling process, reducing the total number of images that have to be supervised by a human being. If the previously saved labelling differs greatly from the information returned by the AI algorithm, a flag would be raised for the human to supervise it more closely, without the need for human supervision of the total samples.
Use case
It was possible to reduce the number of images that need to be monitored by humans to 15% of the total. Additionally, the arduous labelling/supervision process was greatly streamlined, without introducing any false positives to date.
Infrastructure
On Premise
Technology
Image recognition/processing Automatic or Deep Learning
Data
It is a database of tens of thousands of IR images of users using our device. These images have labels associated with different facial features such as eye position, glints and eye bounding box. This database is internal and private due to the confidentiality of the users.
Resources
Vicomtech
Difficulties and learning
Technology bonus
KPIs (business impact and metrics of the model)
At the organisational level, 1 project manager, 2 SW technicians (computer vision and micro services) were required. Model development carried out mainly by Vicomtech, and technical validation, requirements and feasibility by Irisbond.
