drywall-yoloe

YOLOE (open-vocab YOLO11 + text-prompted seg head) fine-tuned via Ultralytics YOLOEPESegTrainer. yoloe-11s-seg backbone, imgsz 640, 30 epochs, prompt augmentation (paraphrase per batch).

Test metrics

Task	Dice	mIoU	Precision	Recall	ms/img
Crack	0.681	0.547	0.690	0.757	~30
Taping	0.859	0.774	0.807	0.956	~30

Best of all three pipelines on Dice + latency.

Load + predict

from huggingface_hub import hf_hub_download
from ultralytics import YOLOE

ckpt = hf_hub_download(repo_id="ravindrakapse/drywall-yoloe", filename="best.pt")
model = YOLOE(ckpt)
results = model.predict("image.jpg", imgsz=640, conf=0.05)

See load_models.py.