manhole-detection-yolo26

Manhole and drainage asset detection (YOLO26)

Models trained by WeScan Solutions to find manholes, inspection covers and drainage grates in survey orthophotos, as part of the internal geo-object-detection R&D platform.

They return masks as well as boxes, so each detection becomes a polygon that can be written straight to a GeoPackage in the orthophoto's own CRS.

Files

Class names are Italian: chiusino = manhole cover, caditoia = drainage grate, rettangolare = rectangular, circolare = circular.

Which segmentation model to use — this matters

The two segmentation models are trained for different object scales and are not interchangeable:

Running either at the wrong scale gives confident nonsense rather than an obvious failure. Measured on a 640 px tile whose ground truth is a single manhole about 180 px across:

• TOMBINI4 → 1 detection, 205 px wide, confidence 0.60 — correct
• TOMBINI-HIGHRES → 6 detections, 13–19 px wide, confidence up to 0.68 — all spurious

Both look plausible in a results table. Check one tile visually before a full run.

Metrics

Detection training run, 50 epochs, yolo26m.pt base, 640 px, batch 24:

Usage

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

weights = hf_hub_download("P-khosravi/manhole-detection-yolo26", "TOMBINI4.pt")
model = YOLO(weights)

result = model.predict("orthophoto_tile.png", conf=0.30, imgsz=640)[0]
print(result.boxes, result.masks)

For a full orthophoto, use the tiled pipeline from the project repository, which handles overlapping tiles, seam merging and georeferenced export.

Demo

https://huggingface.co/spaces/P-khosravi/geo-object-detection

Licence

AGPL-3.0, following Ultralytics YOLO. Trained on WeScan survey data; the training imagery itself is client property and is not distributed.