pipevision-yolo26m-seg

PipeVision YOLO26m-seg — Pipeline Defect Segmentation (ONNX FP16)

Model: yolo26m-seg-pipevision-fp16 Format: ONNX FP16 (opset 17) Input: images [1, 3, 640, 640] — NCHW, float16, RGB normalized [0, 1], letterboxed (pad 114) Outputs:

• output0 [1, 100, 38] float32 — end-to-end (NMS-included): x1, y1, x2, y2, conf, classId + 32 mask coefficients per detection
• output1 [1, 32, 160, 160] float16 — mask prototypes

Team: Plumbers of UTS (Bo Zhao, Jadyn Braganza, Eunkwang Shin)

---

Model Description

This model performs instance segmentation of structural defects in pipeline CCTV inspection images. It is exported from a YOLO26m-seg PyTorch checkpoint (best.pt) trained on a 6-class pipeline-defect segmentation dataset.

The model is exported end-to-end with NMS included in the ONNX graph (output0 already carries post-NMS detections capped at 100). Client code decodes the fixed [1, 100, 38] tensor directly — no separate NMS pass is required, though PipeVision still runs a light client-side IoU dedup as a guard.

Each detection's instance mask is reconstructed from the 32 mask coefficients and the shared output1 prototypes: sigmoid(coeffs · prototypes), thresholded at 0.5, cropped to the box and resized to image space.

---

Intended Use

• Browser-based pipeline defect visualization (client-side ONNX Runtime Web)
• Academic research and portfolio demonstration
• Offline inspection tool prototype

Not suitable for: Safety-critical production deployment, real-time video streams, mobile-only inference (model is ~45 MB).

---

Training Data

Dataset: Pipeline-defect instance-segmentation set (custom, YOLO polygon format)

Classes (6 total):

The class set follows a long-tail distribution — Deposition and Disconnect are the rarest categories in the test split.

---

Training Configuration

---

Metrics

Validation-set metrics at the final training epoch (results.csv, epoch 150):

Per-class (validation, per_class_metrics.csv)

---

Inference Defaults

conf_threshold: 0.25
iou_threshold: 0.45
max_detections: 100

ONNX Runtime Web Example (TypeScript)

import * as ort from 'onnxruntime-web';

const session = await ort.InferenceSession.create(MODEL_URL, {
  executionProviders: ['webgpu', 'wasm'],
});

// Input: preprocessed Float16Array [1, 3, 640, 640] (letterboxed, RGB/255)
const feeds = { images: new ort.Tensor('float16', data, [1, 3, 640, 640]) };
const results = await session.run(feeds);

// output0 [1, 100, 38]: rows of [x1,y1,x2,y2,conf,classId, ...32 mask coeffs]
//   — already NMS'd; drop rows with conf < threshold.
// output1 [1, 32, 160, 160]: mask prototypes for mask reconstruction.
const detections = results['output0'];
const prototypes = results['output1'];

---

Limitations

1. Long-tail dataset — rare classes (Deposition, Disconnect) have fewer training instances; mask [email protected]:0.95 trails the box metric on these.
2. Single domain — trained on one dataset source; performance on other pipe inspection cameras may differ.
3. FP16 precision — minor numerical differences vs the FP32 baseline.
4. No temporal context — designed for single-image inference, not video.
5. Validation metrics — numbers above are validation-set (no held-out test split was evaluated separately).

---

License

MIT — Free to use for research and non-commercial projects.