Start with the decision, not the model

Czech company 3iD develops machines for returning reusable cups. Its operational question is very specific: can this particular cup be used again? The bottom, wall and holder all need inspection. According to the case published by the European Digital Innovation Hubs Network, a human assessed every returned cup when the company sought assistance. Such a process is hard to scale as the machine fleet grows.

3iD did not start with a large multimodal model. Together with EDIH-DIGIMAT and INTEMAC's testbed, it tested a narrower hypothesis: would an inexpensive camera, stable lighting and comprehensible image-processing algorithms suffice? This is useful for small businesses because it separates three things often conflated: whether a prototype works technically, whether a future machine will be reliable, and whether the entire network will generate an economic return.

The pilot should not be retold as a completed industrial rollout. The source describes a testbed study and working code; the first functional machine prototype still had to be designed. Round-the-clock operation and the elimination of manual sorting are possibilities for a subsequent stage, not measured outcomes from an operating fleet.

What the team actually tested

The test configuration used a Raspberry Pi 5 with 8 GB of memory, a 12-megapixel Camera Module 3 and a directed light. The case prices these listed components at CZK 2,748, about €110. This is the cost of the specified feasibility-test kit, not a finished machine or total cost of ownership: enclosure, mechanics, installation, operation and maintenance are not included.

Three checks were implemented in Python and OpenCV:

  • For the wall, the system detected defect edges with Canny on both the normal and inverted images. It compared the edge-pixel count with an adjustable threshold.
  • For the bottom, it restricted analysis to a circular region of interest, suppressing background noise, and checked the outlines of cracks and deformations.
  • For the blue holder, it selected an HSV colour range, found the contour and measured its angular deviation from tolerance.

An interface displayed the live camera feed, processed output and counters for acceptable and rejected cups. According to the case authors, all three algorithms ran in real time on this hardware without a GPU or cloud service. Here, “local AI” means applied computer vision next to the camera, not necessarily a generative model. OpenCV documents both Canny edge detection and HSV range selection; their behaviour on specific defects is easier to verify than a language model's free-form answer.

Where the pilot's certainty ends

The authors report reliable detection of typical damage in their test, but do not publish a confusion matrix, evaluation-set size or precise sensitivity and false-rejection rates. “High accuracy” therefore cannot be converted into a quality guarantee for another batch of cups. Very fine cracks with insufficient contrast could be missed. A single camera position cannot guarantee a view of the entire wall circumference; the source recommends mechanical rotation for full coverage.

A practical evaluation for a Russian operator of a similar process should include different batches, dirt levels, scratches, moisture, glare and lighting changes. It needs both acceptable and defective samples labelled by a responsible specialist. Measure the two errors separately: a defective item returned to circulation, and a usable item discarded. Their costs differ for food-contact packaging; acceptance criteria should be agreed before thresholds are tuned.

When capture conditions vary, stabilise lighting and geometry before trying to “fix” the algorithm. The case identifies controlled illumination as a key condition. The camera should not be asked to guess from reflections: its job is to observe a reproducible feature in a reproducible image.

How a small business can adapt the idea

Begin with a decision map rather than a model shortlist. Which defects prohibit reuse? Which are acceptable? What happens to a borderline item? Who may change a threshold, and who owns the final sorting decision? The same questions apply to cups, packaging, parts, textiles and labels.

A minimal pilot architecture is: fixed imaging area and lighting → local frame processing → separate checks for different item regions → “accept / reject / manual review” decision → a log of the image, rule version and operator decision. For potentially costly uncertainty, the third outcome matters more than flashy automation. Humans should receive disputed cases rather than repeat the camera's work for the entire stream.

Local processing helps when latency matters, connectivity is unreliable, or images and process settings should not leave the site. But avoiding the cloud does not remove the need for code versioning, image-access controls, a fallback for camera failure or personal-data handling if people appear in a frame. The published pilot establishes technical feasibility of the algorithms; these safeguards are editorial requirements for a future deployment, not claimed features of 3iD's prototype.

Economics without a magic “remove the operator” line

Estimate the whole workflow, not just the board and camera. The budget may include item positioning, lighting, lens protection, rotation where needed, integration with a sorter, maintenance, labelling test samples, quality monitoring and handling disputed cases. In an unattended machine, downtime and repairs may cost more than computation.

One illustrative calculation for your own operation: multiply monthly item volume by actual manual inspection time and the fully loaded cost of that time. Compare the result not with the board's purchase price, but with monthly depreciation of the complete installation, maintenance and residual human review. The assumptions are a stable flow, a known share of disputed items and the same quality requirement before and after deployment. A small flow or highly variable items may never justify the investment.

A useful first step is to collect 100–200 representative images from your own conditions, including known difficult examples, and run a blind comparison against a specialist's decisions. That is a pilot collection size, not a statistical reliability guarantee; production acceptance requires a sample and criteria sized to the cost of errors. If thresholds and geometry solve the problem, a generative model need not be interviewed for the job. If defects are diverse and rules fail, compare learned vision with the simple baseline then.

The 3iD case does not prove that every cup is easy to inspect with a camera. It shows how to reduce capital risk: test optics, algorithms and edge cases before building the complete machine. For a business owner, that is a more valuable pilot result than a polished demonstration on one perfect cup.