Modelpublic

ntarekp/watermeter-yolov8-obb

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes

Model Card

Water Meter Reading System (WMRS) — YOLOv8-OBB

Two-stage oriented-bounding-box pipeline for reading analog water meter displays.

Files

—stage1_meter_window_digits.pt — detects meter body, digit window, and digits in the full photo.
—stage2_digits.pt — digit-only detector, trained on cropped/upscaled window regions for tighter localization.

Stage-1 test set results (mAP50 / mAP50-95)

Class	mAP50	mAP50-95
meter	0.995	0.917
window	0.985	0.802
0	0.990	0.766
1	0.903	0.711
2	0.956	0.734
3	0.925	0.736
4	0.874	0.677
5	0.926	0.712
6	0.915	0.684
7	0.945	0.754
8	0.941	0.737
9	0.958	0.722
u	0.767	0.545

Stage-2 test set results (cropped digit-only model)

Class	mAP50	mAP50-95
0	0.994	0.780
1	0.917	0.705
2	0.989	0.769
3	0.960	0.740
4	0.994	0.773
5	0.990	0.717
6	0.978	0.741
7	0.983	0.739
8	0.954	0.725
9	0.952	0.687
u	0.641	0.422

Stage-2 substantially improved digit recall (avg +9 pts) but u regressed — likely due to crop clipping or augmentation on this rare, low-instance class. Noted as a known limitation.

Dataset

510 labeled water meter images, 70/20/10 train/val/test split, 13-class OBB annotations (meter, window, digits 0-9, u = unclear).

Usage

See the demo Space: https://huggingface.co/spaces/ntarekp/watermeter-reader-demo