Visual measurement, counting & volume estimation

Counting items and measuring size or volume is still done with tape measures, tally sheets, scales and manual survey walks. Vision models detect and segment the objects, and calibrated geometry (reference markers, depth sensors, stereo or photogrammetry) converts pixels into millimetres, cubic metres or piece counts. Typical result: faster, repeatable and documented measurements, with a human check wherever the number has contractual or legal weight.

Problem & value

• Today: goods receipt counts, timber stack volumes, stockpile surveys, parcel dimensioning, tolerance checks on parts and crop yield counts rely on manual work, hand tools or expensive survey crews. The results are slow, subjective and rarely backed by evidence images.
• AI value: one photo or short video yields a count or dimension with a stored audit image; drones and phones replace dangerous or time-consuming access (e.g. climbing stockpiles); counting scales to thousands of small items that humans miscount.
• Benefit figures: vendor claims vary strongly by setup; verify on your own scenes [3][5].

Variants

• Discrete counting: parts on trays, pipes/bars/logs on trucks, pallets, bags, seedlings, fruit, fish, cells in a lab (dense-object counting, class-agnostic "count anything" models).
• 2D dimensional measurement: length, diameter, area, angle, gap/flush from a calibrated camera (telecentric optics for precision parts, reference marker for phone-based use).
• 3D volume: bulk material and stockpiles (drone photogrammetry or LiDAR), timber stacks (solid volume from stack photo), parcel and pallet dimensioning (ToF/stereo), excavation and fill volumes on construction sites (see VIS-040, VIS-021).
• Agriculture: plant/fruit counting, canopy and biomass estimates (see VIS-005).

How it works

1. capture (fixed camera / phone / drone / depth sensor)
2. calibration (intrinsics, reference marker or scale bar, ground control points)
3. detection / instance segmentation (YOLO-class, SAM-class, open-vocabulary)
4. geometry (pixel
5. metric via homography, depth map, point cloud or mesh)
6. counting / measuring / volume integration (+ density
7. weight)
8. plausibility & uncertainty checks (tolerance, occlusion, double-count)
9. human review of low-confidence cases
10. record in ERP / WMS / QMS

• Counting: a detector or segmenter proposes instances; tracking avoids double counts in video; density-map or point-prompt models handle crowded scenes with overlap.
• Measuring: instance masks give silhouettes; with a known scale (marker, fixed camera distance, depth) the mask becomes millimetres. Monocular metric depth models help where no sensor exists but are less accurate than real depth hardware.
• Volume: integrate a surface model above a base plane (photogrammetry DSM, LiDAR point cloud, or depth-camera cloud); for timber, estimate stack face area and depth, or measure individual log diameters and lengths.
• Human-in-the-loop: review of low-confidence frames, sampled audits against manual counts, and signed-off values for anything used in billing or legal trade.

Data preparation

The data are images, video or point clouds from fixed cameras, phones, drones or depth sensors, plus reference measurements (weighbridge tickets, inventory counts, CAD or BIM data) that typically live in the ERP, WMS, quality system or in Excel sheets. Typical gaps: inconsistent lighting and camera angles, no reference scale or calibration marker, no ground truth to compare against, photos scattered on phones and file shares, and unclear rights where staff or customers appear in the picture.

Preparation steps: fix the capture setup first (camera position, lighting, scale markers, mounting), which matters more than the model. Collect 100 to 300 representative images per scenario and pair them with manually verified counts or measurements as a test set (PREP-018). Harvest existing labels from weighbridge, delivery and inventory records (PREP-020). Export ERP reference data via API or file (PREP-005) and tidy spreadsheet counts (PREP-011). Annotate with open tools such as CVAT or Label Studio only when a custom model is needed. Check whether people are visible and mask or avoid them (PREP-022).

Rough share of effort: 30 to 40 percent, dominated by capture setup and ground-truth collection. Buying a product reduces but does not remove this work, because calibration still needs reference measurements.

Implementation paths

Which path for whom

• smb (A): Mature counting and volumetric products (camera- or drone-based) exist at SMB prices and the task, such as counting pallets or measuring stockpiles, is standard.
• enterprise (B): Measurements must flow into ERP, WMS or quality systems, so cloud vision APIs or custom models integrated with own systems fit better than a boxed product.
• developer (C): Open-source detection and depth models (YOLO-family, SAM, depth estimators) run on HW1 and give full control over calibration.
• public (C): Surveying, construction or environment agencies keep imagery on own or sovereign EU infrastructure and avoid vendor lock-in.

Path A — Buy (off-the-shelf)

• Parcel/pallet dimensioning (fixed devices): Cubiscan, Mettler-Toledo (incl. certified dimensioners), Zebra, Cognex, Keyence (3D profilers and dimensioners). Some are approved for legal-for-trade use.
• Industrial 2D/3D measurement & counting: Cognex, Keyence, Zebra Aurora, MVTec HALCON (DE), Basler (DE) with Zeiss/Isra-type metrology.
• Stockpile and site volumes: DJI Terra/Enterprise drones, Pix4D (CH), Propeller Aero, DroneDeploy, Trimble, Leica/Hexagon; Teledyne FLIR-style laser scanners for silos. Check EU data location.
• Timber: photo-based log and stack measurement apps (e.g. Trestima (FI), Haglöf (SE)); verify national acceptance for trade measurement before use.
• Choose A when: the object is standard (parcels, pallets, logs, stockpiles) and measurement must be certified or quickly operational.

Path B — Build on cloud APIs

• Detection/segmentation: managed services such as Roboflow (hosted), Azure AI Vision (Custom Vision is scheduled for retirement on 25 Sept 2028; plan migration), Google Vertex AI, AWS Rekognition Custom Labels (EU regions available).
• Multimodal models: Gemini and Claude vision can count and roughly estimate size in single images; suitable for triage, not for tolerance-level measurement.
• Photogrammetry as a service: Pix4D (CH) / Bentley iTwin Capture / DroneDeploy cloud for volume reports.
• Orchestration: edge capture + cloud inference; results into ERP/WMS via REST, n8n or Power Automate.
• Choose B when: you need custom classes or workflows but no on-prem constraint.

Path C — Self-host (open source)

• Detectors/segmenters: Ultralytics YOLO (AGPL, check licensing), RT-DETR, RF-DETR, YOLOX (Apache); Segment Anything 2 (Meta, Apache) and Grounding DINO / open-vocabulary detectors for no-label starts.
• Depth/3D: Depth Anything V2 (relative; metric variants exist), Open3D, COLMAP, OpenDroneMap (orthophoto, DSM and volume tools), PDAL for point clouds.
• Runtime: OpenCV for calibration (ArUco/ChArUco markers), ONNX Runtime/TensorRT, NVIDIA DeepStream.
• Hardware: HW1 (Jetson Orin / NPU) for fixed cameras; HW2 for photogrammetry and batch processing.

Path D — Custom training

Worth it for unusual objects (specific parts, bundles, seedlings) or tight tolerances. A few hundred to a few thousand annotated images with varied lighting and occlusion; for dimensional accuracy collect reference measurements from calibrated tools as ground truth, not just labels.

State of the art (2026)

• Promptable segmentation and open-vocabulary detection (SAM 2, Grounding DINO-style models) make first counts possible without training data; fine-tuned small detectors still win on speed and accuracy for fixed production scenes.
• Monocular metric depth has improved substantially (Depth Anything V2 and successors, Metric3D, UniDepth), but accuracy is scene-dependent. For volumes and tolerances, real depth sensors, stereo or photogrammetry with reference points remain more reliable.
• Phone LiDAR/ToF and gaussian-splat/NeRF reconstructions make capture cheap, yet scale errors, reflective or textureless surfaces and moving objects degrade results.
• Limits and failure modes: occlusion and stacked/overlapping items (undercounting), double counting in video, perspective and lens distortion, shadows, wet or glossy materials, hidden voids inside stacks or piles, wrong density assumptions when converting volume to weight, and drift in uncalibrated setups. Always report an uncertainty and validate against manual reference on a sample.

Shared resources & combinations

• Same stack (object-detection + segmentation + tracking + edge-runtime): VIS-001 defect detection, VIS-008 people counting, VIS-019 warehouse inventory counting, VIS-020 waste sorting.
• 3D/photogrammetry pipeline: VIS-003 drone infrastructure inspection, VIS-021 aerial change detection, VIS-040 scan-to-BIM, VIS-023 damage assessment from photos.
• Domain overlaps: VIS-005 crop counting, VIS-033 food-waste weighing/volume estimation.

Effort, cost & KPIs

• Path A: days to weeks; hardware dimensioners cost from thousands to tens of thousands of euros; software volume tools are usually per-seat or per-flight.
• Path B: 1–3 months incl. calibration, integration and validation.
• Path C: 2–4 months; edge devices or one GPU plus capture hardware and annotation effort.
• Cost drivers: calibration and capture discipline, annotation, ground-truth measurement, hardware.
• KPIs: count error (MAPE vs manual), mean absolute error in mm / % volume error, bias, repeatability (same object, repeated capture), share of items needing manual review, time per measurement, dispute/claim rate.

People & change

Involve the department lead (warehouse, production, site management) first, as they own the process and the tolerance for error. End users such as warehouse staff or site foremen help define capture routines and test the first results; IT and data owners connect ERP or WMS and approve cameras or drones on the network, with IT security reviewing remote access. Management sets the accuracy threshold that justifies replacing manual counts.

Works council: depends. If cameras only observe goods and stockpiles the case is mild, but fixed cameras in workplaces or workflows that log who counted what can fall under §87(1) no. 6 BetrVG, since objective suitability for monitoring is enough. §90(1) no. 3 requires timely information about AI-planned work processes. Agree that images are not used for performance evaluation and write this into a short works agreement. Drone flights may need separate permissions and data-protection checks if people are filmed; involve the DPO and run a DPIA if workplaces are covered continuously.

AI literacy under Art. 4: explain what accuracy to expect, typical failure modes (occlusion, shadows, unusual packaging) and when to recount manually. Start with a pilot at one site or one product group, run manual and AI counts in parallel for four weeks, and report differences openly. Address fears of control and job loss by showing that the tool takes over tedious counting while staff confirm exceptions and keep the final say. See GUIDE-009, GUIDE-004 and GUIDE-008.

Risks, regulation & compliance

• EU AI Act: typically minimal risk. It becomes high-risk where the system is a safety component of regulated products (Annex I, e.g. under the Machinery Regulation) or used in critical infrastructure safety; AI literacy (Art. 4) applies. Body measurement for people profiling or emotion/biometric categorisation would need separate assessment (Art. 5 prohibitions) and is out of scope here.
• Metrology: values used for trade or billing may fall under the Measuring Instruments Directive 2014/32/EU and national legal metrology (in Germany, MessEG/MessEV with PTB and calibration authorities); uncertified AI measurement is a decision aid unless approved [2].
• GDPR: counting or measuring objects is generally no personal data; images of workers, yards with people, number plates on trucks, or drone footage of neighbouring land can be. Use masking, signage, data minimisation and short retention; drone operation needs EASA rules (Regulation (EU) 2019/947) [4].
• Mitigations: show uncertainty and evidence images, check calibration with reference artefacts, audit samples, keep original images.

Getting started

1. Define the decision the number drives and the required accuracy (count tolerance, mm, % volume).
2. Fix capture: lighting, camera position, reference marker or ground control points; collect 100–300 representative images with manual reference values.
3. Try a pre-trained or open-vocabulary model (SAM 2 + detector) and measure the error against reference.
4. Add calibration and uncertainty reporting; decide which cases go to human review.
5. Pilot in parallel with the manual process for a few weeks before switching, then monitor drift.

Sources

1. Meta AI — Segment Anything 2 — https://github.com/facebookresearch/sam2
2. Directive 2014/32/EU (Measuring Instruments Directive) — https://eur-lex.europa.eu/eli/dir/2014/32/oj
3. OpenDroneMap documentation (orthophoto, DSM, volume) — https://docs.opendronemap.org/
4. Commission Implementing Regulation (EU) 2019/947 (drones) — https://eur-lex.europa.eu/eli/reg_impl/2019/947/oj
5. Depth Anything V2 — https://github.com/DepthAnything/Depth-Anything-V2
6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. Open3D — https://www.open3d.org/
8. Ultralytics YOLO documentation — https://docs.ultralytics.com/