SharonMelhi/BluEcho-SSS-Pipeline

BluEcho · Side-scan pipeline detector

Created for Smart India Hackathon (SIH) 2026 by Khushi Mhamane, Sharon Melhi, Kirti Rajput, Peeyush Rampal, Aditya Banerjee, and Aditya SS Varma.

This repository contains the existing sonar-trained YOLO11n pipeline checkpoint used by BluEcho, its ONNX export, and the original frozen inference and evaluation manifests. This publication does not retrain or improve the detector.

Intended input and output

• Input: low-frequency side-scan sonar images compatible with the SubPipe preparation (SSS_LF).
• Class: Pipeline (ID 0), bounding-box detection.
• Preserve uint8 RGB intensity, tile at 640 pixels with stride 480, and pad with RGB 114. Do not squash a long sonar strip into a square.
• Map tile predictions into source-image coordinates and use class-correct NMS at IoU 0.5. Adjacent segments are not reconstructed into a complete physical pipeline.
• Frozen operational score threshold: 0.6818633675575256. Scores are uncalibrated.

Recorded evidence

The frozen manifest reports 58.12% precision, 49.04% recall, and 0.53195 F1 at the validation-selected operating point, with one-to-one class-correct matching at IoU 0.50. The manifest identifies 674 validation originals and 704 internal benchmark originals. These are correlated observations from one survey, not independent physical pipelines. The historical internal benchmark was reused; it is not a fresh holdout or an independent-site test. The ONNX parity report records the checked validation comparison; it does not establish new field accuracy.

No performance is claimed here for shipwrecks, cylinders, entangled nets, plastic composition, new survey sites, or onboard real-time operation. Geotagging requires separately verified source metadata; the image detector alone does not provide latitude or longitude.

Download and run

from huggingface_hub import snapshot_download
snapshot_download("SharonMelhi/BluEcho-SSS-Pipeline", local_dir="pipeline-model")

Install bluecho-sonar with the documented inference dependencies, then import the native checkpoint while keeping its original manifest alongside it:

bluecho models import --model sss-pipeline-v3 --registry ./models --local ./pipeline-model/native.pt
bluecho inspect ./sonar.pbm --modality SSS_LF --model sss-pipeline-v3 --registry ./models --output ./inspection

The native runtime was verified with PyTorch 2.4.1 and Ultralytics 8.4.150. ONNX is FP32, opset 17, input [1, 3, 640, 640]; see manifest.json for hashes and onnx-parity-policy.json for the comparison policy. Raw direct YOLO prediction does not reproduce BluEcho's full tiling and source-coordinate merging workflow.

Files, source and licences

• native.pt: sonar-trained native checkpoint.
• native.onnx: parity-checked ONNX export.
• manifest.json, preprocessing.json, classes.json, evaluation.json: original immutable handoff metadata.
• onnx-parity.json, onnx-parity-policy.json: original comparison evidence.
• training_source/: the available Phase 2B Python preparation, training and evaluation source; paths are supplied through configuration. Raw training data is excluded.
• BluEcho application source and package documentation.

Model/software redistribution preserves inherited Ultralytics AGPL-3.0 obligations; the accompanying project source is AGPL-3.0-or-later. Training data originates from SubPipe, released under CC BY 4.0, with attribution to OceanScan-MST / REMARO. See the SubPipe source documentation and Zenodo record 12666132. No dataset images or annotations are uploaded in this model repository. Original manifests preserve historical statements such as “no public release performed”; this model card records the subsequent authorized publication.