What Calyston's AI actually detects — the object list, explained

People ask us a fair question: "What can the AI actually recognize?" It's worth answering precisely, because the honest answer has two parts — and knowing the difference is the key to getting good results.

Calyston runs two different kinds of AI, and they do very different jobs.

Layer 1 — the object detector (a fixed, reliable vocabulary)

Every frame Calyston decides to analyze runs through a fast object detector. This model was trained on a fixed list of object types — the well-known COCO set of 80 everyday objects — and it can only ever report things from that list. You can't teach it a new word by typing one; it physically only outputs what it was trained on. The upside: for the things it does know, it's fast, runs on a normal CPU, and draws an exact box around each one.

Calyston groups the detector's output into three categories that turn into events you can filter and alert on:

🧍 People

• person

That's the whole category — and it's the one that matters most for security. "Was someone there?" is the question almost every alert is really asking.

🚗 Vehicles

• car
• truck
• bus
• motorcycle
• bicycle
• train
• boat
• airplane

🐾 Animals

• dog
• cat
• bird
• horse
• cow
• sheep
• bear
• elephant
• zebra
• giraffe

(Yes, the list includes zebras and giraffes — that's just the public dataset the model was trained on. Your driveway camera will happily ignore them.)

The detector technically recognizes ~50 other COCO objects too — things like backpack, chair, bottle, laptop, cell phone. Calyston deliberately does not turn those into events. A security system that alarms on a chair is noise, not signal. We track people, vehicles and animals because those are the things that move, arrive, and matter on a property.

Layer 2 — natural-language search (describe anything)

Here's where it gets interesting. Calyston also runs a vision-language model that lets you search your footage by describing it in plain words — even for things that aren't in the fixed list above.

Type a phrase and Calyston scores every saved detection by how well it matches:

• "a red car"
• "person carrying a backpack"
• "someone in a hi-vis jacket"
• "a delivery van"

This is similarity matching, not a fixed classifier — so it's flexible and open-ended, but also fuzzier than the object detector. Concrete, visual things (colors, clear objects, distinctive scenes) work best. Subtle attributes — guessing a person's gender, or recognizing an action like "smoking" from a single frame — are far less reliable, especially on typical CCTV resolution. We tell you this plainly because vendors who demo these queries usually do it on expensive cameras with dedicated AI chips and staged conditions. Calyston runs on the cameras you already own, on a box you control, with nothing sent to a cloud — and we'd rather set honest expectations than oversell.

The short version

• Reliable, boxed, alertable: people, vehicles, animals (the object detector).
• Flexible, describe-anything search: the natural-language layer, best on clear visual descriptions.
• Everything stays on your machine. No frame ever leaves your network to be analyzed.

Natural-language search is included on every paid plan, and the object detector with smart events comes with AI-enabled plans. The free Community tier includes motion detection and events — the foundation everything above builds on.