Efficient edge perception
Reliable perception on inexpensive, CPU-only hardware — not GPUs in the cloud. Doing more with fewer inferences is a product advantage, not just an optimisation.
The interesting question is what's enough
Perception research is usually framed around accuracy — bigger models, more compute, better benchmarks. We're interested in a different axis: sufficiency. How little detection does a reliable operational decision actually require? Often the honest answer is: far less than the field assumes. We build the product first, then let real deployments ask the hard questions.
What we're studying
Applied research, driven by what deployment demands — not the other way around.
Operational trust
The gap between "the model detected something" and "the operator can act on the alert." We study the temporal and decision layers that make an alert trustworthy in the real world.
Evidence over effort
Pre-registered, falsifiable gates and held-out field validation. We trust a result only when the data earns it — and we record what the data doesn't justify.
Open questions we're chasing
Live threads, still being tested in the field. We publish what the evidence earns, and hold back what isn't proven yet.
How little is enough?
How sparse can semantic detection be before an operational decision — presence, dwell, intrusion — starts to degrade?
The cheapest sufficient abstraction
What is the simplest state a system can maintain that still supports a correct decision — and what can it stop paying for?
Simpler as better
When does a leaner architecture make a better product — more robust and cheaper to run — rather than merely a smaller one?
How we work
Field-first and long-horizon. We freeze a hypothesis, deploy across varied real sites, hold out locations we never tune on, and let months of live evidence — not a curated clip — decide what's true. It's slower, and it's the point.
Work with us
For ongoing research, please contact us. We welcome research collaboration, field-test and pilot partners, edge-hardware partners, and academic enquiries into efficient, on-device perception.
Get in touch
Prefer email? hello@perceptionorigin.com