What can a small device do without a large GPU?
Training a vision model and using it are different jobs. The camera-side task was inference: applying an already-trained model to new images, with limited power and processing capacity.
Cortexica · 2014–2019
AI cameras and applied vision.
At Cortexica, I brought computer vision onto small edge devices, from camera prototypes to checks for protective clothing. The question was practical: could a camera system recognise what mattered, close to where it was happening?
Watch CortexiCAM
Question and contribution
Put the trained model beside the camera.
Software engineering, then innovation leadership
I joined Cortexica as a GPU Engineer in February 2014 and became Head of Innovation in June 2016, staying until September 2019. I evaluated emerging hardware and developed prototypes including CortexiCAM and the PPE demonstrations shown here.
Cortexica software and partner hardware
Within Cortexica’s Innovation Department, I worked on connecting our vision software to edge hardware. We developed the PPE kit with UP, and demonstrated protective-equipment checks at the Axis Experience Centre in Madrid.
Cortexica’s technical documentationImplementation
From object detection to a visible PPE check.
UP Squared and Myriad 2
For the experiment I described on 5 April 2018, I put one AI Core module into an UP Squared computer running Ubuntu. I ran three MobileNet-SSD models in sequence to locate objects in images, at around 10 frames per second. The models were still unoptimised and not well trained; the experiment showed what that small setup could do.
Read the setup and reported resultCortexiCAM
With CortexiCAM, I brought inference into a camera prototype based on Myriad-X. This was a later vision processor than the Myriad 2 used in my UP Board experiment. The film shows the prototype I developed during the Cortexica period.
Watch CortexiCAM on YouTubeRecognising helmets and high-visibility clothing
PPE means personal protective equipment. Our system used a C++ service to analyse images and return object labels, boxes and inference times to a display application. In the Axis demonstration, we applied that to helmets and high-visibility jackets, turning the detections into checks a person could see.
Watch the Axis Experience Centre PPE demoWhat worked
Vision running on small edge hardware.
The result was a set of working prototypes: object detection on the UP Squared setup, inference in CortexiCAM, and protective-equipment checks with a visible interface. The demonstrations made the software’s output tangible, on the hardware it was intended to run on.
The frame-rate figure belongs to the 2018 Myriad 2 experiment; the PPE development kit was for evaluation and proof-of-concept work.
Public artifacts
The demonstrations.
Watch Generator
Later work · Oosto era
Testing perception in a known world.
Later, during the Oosto era, I explored controlled synthetic scenes with Generator. This opened another way to test perception: start with a world where the scene and camera geometry are known. I shared the v1.0 video in January 2024.