Detecting health risks in a non-invasive way, a Viapontica AI and Cisco Inc collaboration
21 April 2020
Concept to prototype: Open data with smart cost-effective air quality sensors
3 September 2021Quick 3D design & prototyping to support ML Ops
Often training data is not available immediately at project kick off. Here is how we cobbled together some parts and 3D printed others to build a waterproof casing to collect initial training data and avoid delays on a recent project.
We specialize in building bespoke AI models. Тhe success and performance of the AI models which we build for our clients depends on the availability, quality and size of training data. Often data is not available immediately at the onset of project – data may not yet be collected and stored in a suitable format, confidentiality and privacy terms may restrict the use of data, or data from a particular training distribution may not be available at all.
When training data is not immediately available, we get creative – sometimes we generate synthetic data, other times we go out in the world to collect real-world data, as we did on this occasion.
On a current project, Smart Roads AI, we knew that DPIA, Data Protection Impact Assessments, approvals would take time to be reviewed and granted. But even when client training data would become available, we suspected that the available data would be skewed and it would not present enough samples from some distributions (use cases) which were important for our client, for example road data showing:
- adverse weather conditions, such as heavy rain, or direct sunlight
- presence of debris or dirt, such as dried mud, on the road surface
- unusual mounting angles of the camera equipment, such as highly obtuse or acute mounting angles
- variety of road surfaces, such as concrete and cobbled stone, in addition to asphalt
- option for external mounting
The lack of training data increases the risk of knock-on delays. We knew we would have to collect data ourselves. We also quickly discovered that none of the off-the-shelf mounting kits met our mounting requirements. We wanted to be able to mount recording equipment externally to vehicles in order to assess the feasibility of using infra red sensors for depth sensing and remote measurement capabilities. However, we did not find any products that met the identified requirements in terms of ruggedness, IP (weatherproof) rating, theft deterrence, cooling, and space options for storage of auxiliary hardware such as external battery, which was necessary for continuous 16 hour recording.
We set on build a custom enclosure to collected data. Here is how we did it.
Please note the "live photos" below are embedded as videos which will not play correctly on some mobile devices.