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.

Utility Meter Digits — Hybrid Mix (Analog + Digital) · 50-Image Preview

This is a free 50-image preview of DS_mixed_scor_labeled_v1_2026 — 21,914 fully annotated real-world utility meter images.

We published these 50 photos for one reason: to show exactly what a serious meter-reading dataset should look like before you spend a single dollar or a single week of labeling time. Every sample here is a real field capture with real annotations — no synthetic renders, no clean studio shots, no cropped-to-perfection displays.

What you're looking at

Open any image in the preview and you'll see the same annotation discipline applied across the full 21,914-kit release:

One image → three files. Every photo ships with a YOLO .txt (normalized coordinates) and a PASCAL VOC .xml (absolute pixels). Drop it straight into Ultralytics, SSD, or Faster R-CNN — no converters, no glue scripts.
A 22-class unified taxonomy that finally solves the rolling-drum problem: digits 0–9 (IDs 0–9), ten dedicated mid-rotation classes 0_1 … 9_0 (IDs 10–19) for analog drums caught between values, plus two structural anchors — comma and scoreboard (the reading zone itself).
Mechanical and digital in one label space. Analog dials use the full rolling spectrum, LCD/LED displays use standard digits. One model. Zero device-routing logic.
640 px longest side, standardized. Enough pixel density for tiny distant displays, gentle enough on VRAM for large-scale runs.

Why the ugly photos are the point

These are not the images a vendor picks for a landing page. In the preview and across the full set you'll find heavy glass glare, washed-out low-contrast LCD matrices, cracked covers, dust and grime, deep basement shadows, skewed angles, and even inverted frames. Field conditions, labeled by hand.

The full dataset spans water, gas, electricity, and heat meters from 100+ European manufacturers. That diversity is the anti-overfitting safeguard: your network stops memorizing case colors, bezel shapes, and brand logos, and starts actually decoding characters.

The full dataset — and a production model on top

Purchase the complete dataset and receive our proprietary, production-ready neural network trained explicitly on this 22-class universal taxonomy. The model has achieved a proven mAP50-95 of 94%, allowing you to bypass months of R&D and deploy instantly.

→ Full annotated dataset: https://utilitymeters.ai/datasets-utility-meters/dataset-mix-meters

Licensing in plain terms: datasets (images + annotations) come with full commercial rights to train your own proprietary ML/AI products. The bonus pre-trained weights are provided free under GNU AGPL-3.0. Redistribution or resale of raw files is not permitted.

Also worth a look

Raw, unlabeled images — 20,000+ in-the-wild captures across water, gas, electricity, and heat. Available as dials only or dials + serial numbers, and as mechanical-only, LCD-only, or hybrid subsets. From 1,000 images up, starting at $0.35/image: https://utilitymeters.ai/images-utility-meters
Pre-trained YOLO models — six networks from a 12-class LCD specialist (mAP 98.5%) to our 35-class flagship all-in-one engine that reads consumption and serial numbers in a single pass, shipped in PyTorch, ONNX, and TFLite with knowledge-distillation docs for mobile: https://utilitymeters.ai/ai-models-utility-meters

Come say hi

Building AMR retrofits, a mobile meter-reading app, or smart-city asset tracking? Grab the 50 samples, run them through your current pipeline, and see where it breaks. Then talk to us — we'll tell you honestly which subset you actually need.

UtilityMeters.ai — real meters, real conditions, honest labels.

Downloads last month: 66

Total file size:

3.23 MB