Above-SOTA computer vision on your existing CCTV
No new cameras, no server room. Cassette runs TAPe + ML — the engine from our technology partner comexp.net — on the hardware you already own, and reacts the moment something happens.
Inside TAPe + ML
Instead of slicing each frame into patches or pushing raw pixels through a giant network, TAPe pre-encodes the image into a small set of structured building blocks and recognises objects directly, in a single pass.
The payoff is SOTA-level accuracy at roughly a thousandth of the size: full object detection in under 100,000 parameters — against the 100M+ of mainstream models — running on ordinary CPUs, no GPU. Lighter to deploy on the cameras you already have, and cheap to run at scale. Built by our technology partner comexp.net.
How it stacks up
Published benchmarks from comexp, on standard datasets.
COCO detection · mAP@50
Level with RF-DETR-2XL and well above YOLO — at under 100K parameters instead of 127M. Source.
Industrial pilot · TAPe + ML vs YOLO 12s
Same small pilot dataset (≤50 images/class). TAPe + ML leads on both metrics, most on the strict mAP@50-95. Source.
From frame to action, threaded end to end
Camera
Your existing CCTV feed, exactly as it is today.
Edge inference
TAPe + ML runs the detection scenario on-site, in milliseconds.
Event
A mismatch, empty shelf or concealment is recognised.
Alert / action
Supervisor alerted, self-checkout paused, clip attached.
A pilot in a few weeks, on your own footage
Pilot on 2-5 stores and measure shrinkage before and after. One-time setup from ~AED 5,500.
Start a pilot