/ use case 001 · waste reduction

Throw away less.
Pay for fewer collections.

What you throw away costs you twice: once in fines and contamination penalties, again in unnecessary collections. Darkfield watches every bin on the cameras you already have — fill levels, contamination, mis-routed items — and calls the manager when something needs attention. Fewer collections, cleaner streams, a continuous audit you can hand to a regulator.

deployed in · automotive workshops · commercial waste streams · retail loading docks
/ what we detect

Three things, watched continuously.

a / bin fill level

How full is each bin, right now?

Quantitative percentage updated on every keyframe. Triggers a collection request before overflow; suppresses unnecessary collections of nearly-empty bins.

b / contamination

Is there an item in the wrong stream?

Detects oil-soaked rags in recycling, food in dry waste, hazardous in general — and flags it the moment it lands. Direct phone call to the on-call manager.

c / material classification

What's actually in this bin?

Per-item segmentation across 7+ categories: cardboard, mixed plastic, food plastic container, metal, motor-oil bottle, organic, unknown (flagged for review).

d / collection event

Was the bin actually emptied?

Multi-signal verification — bin position movement, sharp drop in fill percentage, scheduled-time match — plus a vision-language model auditing the surrounding clip.

/ how the detection works

Several methods.
Picked per camera.

No two bin compounds behave the same. Outdoor sites move bins after every collection. Indoor bins near service bays may not be present at all between shifts. Some are colour-coded, some are numbered, some are unmarked. We pick the right method per camera during onboarding, based on what works on your scene.

Fill-level estimation

four complementary approaches

methodbest forhow it workssetup
1 · VLM directday-one resultsVLM estimates fill % from the framenone
2 · segmentationquantitative measurementSegments waste vs. bin, computes area rationone
3 · trained modelhighest accuracyPer-camera model auto-trained on 1 & 2's outputs~3 days
4 · ArUco markerstricky angles / lightingQR-style stickers inside the bin at fixed heightsstickers

Bin identification

five ways to know which bin is which

methodbest forhow it workssetup
1 · zonelarge stable sitesPredefined fixed regions per binmanual
2 · colourcolour-coded streamsGeneric detection by colour signatureautomatic
3 · VLM labelany labelled binReads stickers and painted markingsautomatic
4 · OCRprinted bin numbersOCR + bbox handoff to segmentationautomatic
5 · trainedcomplex compoundsPer-bin detection model via auto-train~3 days

Contamination detection

two approaches, picked by site

methodbest forhow it workstrade-off
a · class-basedlimited material varietyGreasy box vs. clean box are separate classes — single-pass detectionclasses × states grows fast
b · attribute-layeredmulti-material sitesDetect material first, then detect grease / oil / staining as attributesneeds labelled stains

// all methods validated at TRL 5 · selection happens during the onboarding survey, no partner involvement

/ when something matters

A real call, recorded.

When the AI detects contamination, it doesn't email a dashboard — it picks up the phone. The recipient asks questions. The AI answers, logs the response, and follows up. Below is an actual recording of one such call, with the audio you'd hear at the supervisor's end.

AI Hello, this is your waste monitoring AI. I'm calling to inform you that dry recyclable items have been detected in a general waste bin.
Human I see. What items are we talking about?
AI The items detected were a cardboard box and metal cans. This occurred at 12:40 in general waste bin number two on the right side.
Human I see. Can you send me a text message with these details?
AI I will send a text message with the photo and details to your mobile device immediately. Do you require any further information regarding this incident?
Human No, that's fine. Thank you.
AI You are welcome. I have sent the details to your mobile. Have a productive day.

Detection, natural-language notification, supervisor confirmation, SMS follow-up, and remediation logging — all in a single sub-one-minute loop, with no human on the platform side.

how alerts work →
/ the data table

What you receive, row by row.

timebineventseverity
14:14:08bin-04 · recyclingoil-soaked rag detected (×2)P1 · call
14:08:31bin-02 · generalfill 87% · approaching thresholdP3 · log
12:42:11bin-09 · scrap-metalaerosol can · poked · routed correctlyok · log
11:02:48bin-07 · hazardousmisroute attempted · paint tin · routed backP2 · sms
08:15:01bin-04 · recyclingcollection event verifiedok · log
/ try it on your bins

First detection on day one. No financial commitment for two weeks.

We typically have bin fill, contamination, and collection events flowing within 48 hours of connecting your first camera. If we can't reach the accuracy you agreed in two weeks, you pay nothing — and we remove anything we installed.

other use cases
typical sites
1–22 cameras
training time
~3 days · per-camera
compliance
UK GDPR · anonymised in RAM