pipevision-yolo26m-seg

Unified YOLO26m instance-segmentation model for sewage / plumbing pipe defect detection. Trained as part of the Plumbing Defect Detection & Pipe Inspection System (UTS CNN Assignment 3, Team Plumbers of UTS) and shipped as a single ONNX file for in-browser inference via the pipevision-ai web client.

• Architecture: YOLO26m-seg (P3/P4/P5 detection + mask prototypes head)
• Input: float32 [1, 3, 640, 640], NCHW, RGB, 0..1, letterboxed
• Outputs: detections [1, 300, 38] + prototypes [1, 32, 160, 160]
• Classes (7): Buckling, Crack, Debris, Hole, Joint offset, Obstacle, Utility intrusion
• Precision: FP16 (ONNX, opset 17)
• SHA-256 (yolo26m-seg-fp16.onnx): 3015a5cca1cce704912aebc01c24d2287af4e07514f279cf81c6cbcc63b4b922

Files

File	Purpose
yolo26m-seg-fp16.onnx	Web / edge inference (47 MB, FP16, opset 17)
best.pt	Ultralytics native checkpoint for further training
metadata.yaml	I/O contract, decode rules, thresholds, metrics
classes.json	Class ID → name map
per_class_metrics.csv	Per-class P / R / AP50 / AP for box + mask
results.csv, results.png	Training curves
confusion_matrix*.png, test_*.png	Eval visualisations
qualitative/	Sample predictions on the test split

Test Metrics (Roboflow v1 test split, 98 images / 229 instances)

Metric	Box	Mask
mAP@0.5	0.534	0.475
mAP@0.5:0.95	0.302	0.271

Per-class (test split)

Class	AP50 (box)	AP (box)	AP50 (mask)	AP (mask)
Utility intrusion	0.901	0.594	0.920	0.628
Hole	0.832	0.555	0.508	0.328
Obstacle	0.704	0.347	0.683	0.376
Debris	0.597	0.260	0.604	0.313
Crack	0.397	0.208	0.266	0.109
Joint offset	0.225	0.104	0.296	0.124
Buckling	0.080	0.044	0.048	0.016

Buckling regression is a known training artefact (cls_weights was silently dropped in Ultralytics 8.4.37); a short fine-tune from best.pt with the 8.4.51 trainer and the original Crack 2.6 / Hole 2.4 / Utility 2.8 weights is the planned fix.

Mask Decode (TS / Python)

1. detections = output0[:, :, :].T                # [N, 38]
2. boxes, scores, class_ids, coeffs               # split per spec in metadata.yaml
3. Apply confidence threshold (default 0.25) and class-wise NMS (IoU 0.45)
4. proto = output1[0]                              # [32, 160, 160]
5. mask  = sigmoid(coeffs @ proto.reshape(32, 25600)).reshape(160, 160)
6. crop mask to letterboxed bbox, threshold > 0.5, resize to original frame

Full client implementation: pipevision-ai/src/lib/postprocess.ts.

Quick Start — Ultralytics

from ultralytics import YOLO
model = YOLO('best.pt')                            # or yolo26m-seg-fp16.onnx
results = model.predict('pipe_frame.jpg', conf=0.25, iou=0.45)
for r in results:
    boxes = r.boxes.xyxy.cpu().numpy()
    masks = r.masks.data.cpu().numpy()             # [N, H, W]

Quick Start — ONNX Runtime

import numpy as np, onnxruntime as ort
sess = ort.InferenceSession('yolo26m-seg-fp16.onnx')
x = np.random.rand(1, 3, 640, 640).astype(np.float32)
det, proto = sess.run(None, {'images': x})
print(det.shape, proto.shape)   # (1, 300, 38) (1, 32, 160, 160)

Training Recipe

Knob	Value
Base	yolo26m-seg.pt (COCO segmentation pretrained)
Hardware	SageMaker GPU (ml.g5/g4dn class)
Optimizer	MuSGD, cosine LR, AMP on
imgsz / batch / epochs	640 / 8 / 200 (patience 50)
Augmentation	mosaic 1.0, close_mosaic 20, mixup 0.10, copy_paste 0.30, scale 0.6, degrees 10, hsv_v 0.30, hsv_s 0.60, fliplr 0.5
Mask labels	SAM 2.1_b pseudo-polygons from YOLO bboxes (see src/generate_seg_labels.py)

Dataset

• Source: Roboflow Universe — Sewage Defect Detection v1
• Splits: 686 train / 196 valid / 98 test
• Original labels: bounding boxes; polygon masks generated with SAM 2.1_b
• Image dimensions: 640 × 480 (CCTV frames)

Limitations

• Trained on a small (~1 k images) labelled set; expect drift on cameras with very different colour balance or non-CCTV optics (e.g. drone or borescope).
• Buckling AP50 is 0.08 — do not deploy this class without retraining with proper class weighting.
• Native input scale is 640 px. SAHI sliced inference is not recommended; it hurt mAP on the source dataset (see project report §7.5).
• FP16 ONNX: verified box-count parity with FP32 on the test set; expect ≤ 1 pp mask mAP delta on borderline thin Cracks.

Citation

@misc{plumbersofuts2026pipevision,
  title   = {pipevision-yolo26m-seg: instance segmentation for sewage pipe defects},
  author  = {Plumbers of UTS},
  year    = {2026},
  howpublished = {\url{https://huggingface.co/gracefullight/pipevision-yolo26m-seg}}
}

License

AGPL-3.0 (inherited from Ultralytics YOLO26 base weights).