rwanda-water-meter-reading-yolov8n

Rwanda Water Meter Reading YOLOv8n

This repository contains the selected model for the Rwanda Water Meter Reading System MVP. It detects the complete meter, the reading window, and digits 0-9. The application keeps digits inside the best window and orders them from left to right to construct a reading.

Live demonstration

Test the model without installing anything:

https://huggingface.co/spaces/Beni4/rwanda-water-meter-reading-demo

Validation performance

Validation was reproduced with Ultralytics 8.4.49 at an image size of 640 on 249 validation images containing 1,589 labeled objects. Class-by-class results are available in per_class_metrics.csv, and validation_metrics.png provides assessment evidence.

Classes

The class order is also stored in classes.yaml.

Unclear digit policy

u is not a trained object class. During post-processing, a detected digit with confidence below 0.40 is returned as u. The same result can be inserted when digit spacing indicates a likely missing position. This prevents uncertain digits from being presented as reliable readings.

Files

• best.pt: original evaluated Ultralytics YOLOv8n checkpoint.
• best.onnx: browser and cross-platform export of best.pt.
• reading_postprocess.py: window filtering and reading reconstruction.
• classes.yaml: authoritative class order and u policy.
• training_args.yaml: parameters from the selected training experiment.
• model_metadata.yaml: dataset, training, and validation provenance.
• metrics.json: overall validation metrics.
• per_class_metrics.csv: class-by-class validation metrics.
• validation_metrics.png: validation evidence.
• CHECKSUMS.sha256: SHA-256 identities for both model artifacts.

Python usage

Install the same Ultralytics release used for validation:

pip install ultralytics==8.4.49

Run object detection:

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict(
    source="water_meter.jpg",
    imgsz=640,
    conf=0.03,
)

for result in results:
    print(result.boxes)

Candidate confidence 0.03 retains weak digit candidates for reading post-processing. It is intentionally different from the 0.40 threshold used to convert an uncertain selected digit to u.

Training data

The selected run used 1,248 images: 329 teacher-provided meter images and a limited 919-image Roboflow digit subset. The private training images are not redistributed in this model repository. Users must verify ownership and source licenses before reusing or publishing the original datasets.

Limitations

• Performance can decrease with blur, glare, severe perspective, occlusion, or very small digits.
• Inverted images can change the semantic meaning of digits such as 6 and 9.
• The ONNX export can move borderline confidence values across the u threshold; the submitted assessment checkpoint is best.pt.
• A predicted reading should be reviewed whenever it contains u.