Industry field guide
Recycling sortation data for AI material recognition
A recycling model has to recognize materials when packaging is dirty, crushed or partly hidden.
- 2022 NAICS
- 562920
- Practice pack
- Example and answer key (JSON)
- Updated
What this guide tests
Which streams follow from the verified labels, what happens to C, and why does the proposed split weakly test new-object generalization?
See the specimen and fields · Work through the fictional records
Task and source context
A recycling model has to recognize materials when packaging is dirty, crushed or partly hidden. AMP’s product documentation shows a commercial use of computer vision in waste sorting. It does not show that AMP is buying outside footage. A facility considering a data project can propose a bench collection around a defined classification or error. AMP sorting technology
Give the model a real ambiguity
Use approved clean material samples on a stationary table. Stage a transparent container under glare, a crumpled item with a label removed and two visually similar items that should enter different local streams. Record the verified material class and the local rule that governs the routing. Annotate overlaps and conditions where the human reviewer cannot identify an item. Keep examples of uncertain labels separate.
Hold back complete object instances rather than random frames from the same video. If the buyer needs conveyor footage, agree separately on camera position, belt speed, lighting and whether an item’s material is independently confirmed.
Stay outside active sorting equipment
Use authorized training materials and a controlled location. Keep workers, customer addresses on packaging, proprietary layout and operational telemetry outside the initial sample. Do not create unsafe exposure to hazardous or contaminated waste for a recording.
Describe the stream and decision. Include which material classes you can verify, what equipment is out of scope and what a buyer can measure from the sample. Describe collection capabilities
Industry research
The U.S. Census Bureau’s industry definitions establish which activities these codes cover. The workflow sources in this guide describe the task or record type. Fiund’s pilot below uses fictional material to test a specific decision within that scope.
562920 · Materials Recovery Facilities
This industry comprises establishments primarily engaged in (1) operating facilities for separating and sorting recyclable materials from nonhazardous waste streams (i.e., garbage) and/or (2) operating facilities where commingled recyclable materials, such as paper, plastics, used beverage cans, and metals, are sorted into distinct categories.
Build a pilot sample
Set up this pilot
Stage clean inert objects under an original sorting rule. Hold out all views of one object and grade verified material, selected stream and uncertain identity separately.
Prepare the test packets
- Start with the worked case. Use the fictional record types below as a guide to the source files you need. Give each packet a stable ID and save its source versions with it.
- Change one fact. Make one packet with enough evidence to answer the question and one where the evidence points to a different answer. Keep the other fields fixed so a reviewer can tell what changed.
- Remove the decisive evidence. Give the reviewer a third packet with a plausible near-match in its place. Have them write what can still be established and what evidence would resolve the rest.
Fields the reviewer needs
- Object ID and verified material source
- Defines ground truth.
- Local rule, stream and uncertainty
- Supports routing.
- Occlusion and instance-level split
- Preserves capture and evaluation limits.
Use Build an AI evaluation sample for service handoffs to turn the specimen into a brief with a defined task, source and review owner.
Fictional exercise
Verified material and a local stream rule can disagree with a visual shortcut
Owned clean training objects on a stationary table use original annotations and an invented routing rule. There is no conveyor, hazardous waste, actual facility rule or claim of buyer demand.
| Record ID | Source | What it records |
|---|---|---|
| RC1 | Local practice rule | Verified material PLASTIC goes to stream P; verified GLASS to G. Unidentified objects route to REVIEW. Rule is original to this fixture. |
| RC2 | Objects A and B | Transparent A has verified plastic provenance, glare partly obscures its shape. Similar transparent B has verified glass provenance. Distinct full object IDs. |
| RC3 | Object C and split draft | C is crumpled with removed label; reviewer cannot verify material. Draft training/test split uses adjacent frames of A in both sets. |
Question for the model. Which streams follow from the verified labels, what happens to C, and why does the proposed split weakly test new-object generalization?
Show the reference answer
RC1 and RC2 route verified plastic A to P and glass B to G despite their similar transparent appearance. RC3's C lacks a verified material label and belongs in REVIEW rather than a forced class. Adjacent frames of A in both training and test reuse the same object and are a weak test of transfer to unseen instances. Hold out complete separately verified objects and preserve uncertainty instead of using transparency as a universal material shortcut.
How to check the result
Ask for the answer, the record IDs that support it and any fact the evidence leaves unresolved. Have a reviewer compare the response against the reference before deciding whether the sample is ready to expand.
Errors to flag in this case
- Inferring material solely from transparency or a familiar shape.
- Forcing an uncertain class or claiming new-object transfer from adjacent frames of the same object.
Change the inputs and run it again
Use an original opaque plastic object and a clear nonmatching prop with independently reviewed labels. Test material evidence rather than the initial appearance pattern.
Score the decision and the evidence separately
Save the reviewer’s expected answer, supporting record IDs and unresolved fields in a separate answer key. Score the model on the decision, its citations and its handling of the deliberate gap. A correct answer citing a different case counts as an evidence error.
Before expanding collection, have a second reviewer run the same checks on the held-out packet. Keep related versions, views and exports of the same case together so they stay on one side of the training and evaluation split.
Check permission and scope
Use approved clean objects away from operating equipment. Exclude workers, addresses, layouts and telemetry; no contaminated material is needed.
Exclude hazardous operations, real site coordinates and restricted records. Use inert objects in a visual exercise.
Describe an authorized archive through Fiund’s existing-data assessment, or describe a new capture capability through the collection brief. The contributor agreement covers the terms for participation.
Sources and further reading
Resources for this workflow
Company and program reviews
Use these reviews to compare sourcing, capture or evaluation approaches. Confirm the applicable program, accepted sources and terms with the provider.
Compare the approach
micro1 review for experts, trade footage and company data
Compare the prerequisites for an approved recording project before planning a contribution.
Compare the approach
Sama review for computer-vision labeling and model evaluation
Review labeling scope for visual state and exception checks in staged footage.
Related industry examples
Industry guide · NAICS 562998
Catch basin cleaning and inspection asset records for AI
A cleaning ticket and later inspection may refer to different assets with similar identifiers.
Industry guide · NAICS 562910
Environmental remediation sample custody and clearance records for AI
A remediation project may have sampling, lab receipt, results and a separate decision about clearance.
Industry guide · NAICS 562991
Septic pump-out and maintenance visit records for AI
A homeowner’s bill may list a visit without containing the evidence for every service performed.