This is a premium annotated dataset. By requesting access, you agree to credit UtilityMeters.ai as the data source in any research, product, or publication that uses it, and confirm you will not resell or redistribute the raw files.

DS_mech_scor_ser_labeled_v1_2026

Mechanical Meter Dials + Serial Numbers — Annotated Vision Dataset

9,299 labeled image kits · 35 classes · YOLO + PASCAL VOC · dual-target annotations

Dataset Summary

This dataset trains a model to read the consumption value and identify the physical device in a single forward pass. Most meter-reading datasets stop at digits — this one goes further: every image carries dual-target annotations, so a network learns to extract the billing number from the analog reels and localize and decode the manufacturer's serial number in the same frame. That's the link between "we can read a photo" and "we can automatically match a reading to a customer account."

Bonus: every purchase of the full dataset ships with a proprietary neural network trained directly on this 35-class taxonomy, with a measured mAP50-95 of 94%, provided free of charge under AGPL-3.0.

👉 Get the full dataset + free model weights: https://utilitymeters.ai/datasets-utility-meters/dataset-mechanical-counters The card below describes the structure and sample available on this page; the complete 9,299-kit dataset and pre-trained weights are hosted on the UtilityMeters.ai platform.

Dataset Structure

DS_mech_scor_ser_labeled_v1_2026/
├── Images/                 # 9,299 utility meter photos (JPG)
├── Labels_YOLO/            # normalized relative coordinates (.txt)
├── Labels_PASCAL_VOC/      # absolute pixel coordinates (.xml)
├── classes.txt             # 35-class master list
└── README_Documentation.txt

Directory layout follows Ultralytics conventions, so training starts with a config file instead of a data-engineering project.

Data Instances

Each image kit consists of one JPG photograph paired with two label files (YOLO .txt and PASCAL VOC .xml) describing the same bounding boxes in different coordinate formats.


Valid kits (image + .txt + .xml)	9,299
Data type	Real photographs of utility meters (JPG)
Annotations	YOLO (normalized `.txt`) and PASCAL VOC (absolute-pixel `.xml`)
Classes	35 (dual-target taxonomy)
Resolution	Longest side pre-scaled to exactly 640 px
Extras	`classes.txt` master list + README documentation

Data Fields — the 35-class dual-target taxonomy

The label space is deliberately split into logical segments so the model never confuses a billing metric with a hardware label:

Metrics block (IDs 0–9) — clean, fully settled digits on the consumption register.
Mid-rotations (IDs 10–19) — transition states such as 0_1 … 9_0.
Serial verification (IDs 20–31) — a separate character space (s0…s9, s/, s_) for serial numbers.
Target zones (IDs 32–34) — comma, scoreboard (the register window), and sa (the serial box), for region-first, decode-second pipelines.

Why This Data Trains Better Models

Solves the half-digit problem. Generic OCR engines collapse the moment a mechanical wheel sits halfway between two numbers. Dedicated rolling-wheel classes teach the network to resolve those intermediate states explicitly instead of guessing.
Genuinely "in the wild." Faded and stamped serials on metal, harsh flash glare on protective glass, skewed shooting angles, dust, grime, and overpaint — annotations were produced on real field imagery, not studio shots.
Dual formats, zero conversion scripts. Both YOLO and PASCAL VOC labels ship side by side, whatever framework your stack prefers.
VRAM-friendly by design. 640 px normalization keeps enough pixel density to resolve tiny weathered serial characters while keeping batch sizes (and GPU bills) sane.

Intended Uses

Automated Meter Reading (AMR) and remote-reading apps, utility field-inspection and anti-fraud tooling, smart-city and IoT edge devices, self-service customer photo submission flows, and asset-tracking systems that need to bind a reading to a specific physical device.

Licensing Information

Images and annotations come with full commercial rights for training proprietary ML/AI applications. The bonus pre-trained model is provided free of charge under GNU AGPL-3.0. Redistribution or resale of the raw files is not permitted. Full terms at the link above.

Related Datasets from UtilityMeters.ai

This dataset is one module of a larger catalogue of annotated datasets and raw image collections for water, gas, electricity, and heat meters from 100+ European manufacturers:

Annotated datasets — mechanical dials only (DS_mech_dials_labeled_v1_2026), digital LCD/LED screens (DS_lcd_screens_labeled_v1_2026), a hybrid mechanical + digital taxonomy for one universal reader (DS_hybrid_dials_labeled_v1_2026), LCD + serial numbers (DS_lcd_serial_labeled_v1_2026), and the 35-class Ultimate Mix covering dials, screens, and serials together (DS_mixed_scor_ser_labeled_v1_2026). Every complete dataset ships with matching pre-trained weights as a bonus.
Raw, unlabeled data — 20,000+ real meter photographs across varied resolutions, orientations, lighting conditions, and edge cases, pre-sorted into mechanical and electronic sets.
Pre-trained models — production-ready recognition weights reaching up to 99% mAP50 / 87% mAP50-95, ready to drop into a mobile app, REST API, or IoT SDK.

Browse everything: https://utilitymeters.ai

Downloads last month: 11

Total file size:

6.17 MB