drishti-detector

DRISHTI — marine-debris detector for side-scan sonar

YOLOv8s fine-tuned to detect man-made seabed hazards in side-scan sonar (SSS) imagery. Built for Smart India Hackathon 2026, Problem Statement 26057 (Ministry of Earth Sciences / NIOT).

Ships as PyTorch and as ONNX — the ONNX path runs the full pipeline with onnxruntime + numpy + opencv and no PyTorch import at all (~50 MB runtime vs ~1.5 GB), which is what makes it deployable on an AUV.

• Code: https://github.com/Rehan9599/Sonar-Drishti
• Full technical record: docs/PROJECT_RECORD.html in that repo

Classes

5 trained, 4 shipped. crab_pot is trained as a hard-negative-ish class and filtered downstream — it never reached a usable AP (see Limitations).

Files

Results — held-out test set (850 tiles)

Per class (AP@50):

On the shipwreck number. It is measured on a deliberately hard test set (a 50 %-overlap re-tile tripled the shipwreck instances with partial and near-duplicate tiles). The 2026 literature on the same dataset (AI4Shipwrecks) reports a vanilla YOLOv8 detection baseline of mAP50 0.716 and a best-in-class 0.755 (DFSE-YOLO), while human inter-annotator agreement on SSS wrecks is only 50–60 % (SW-Net). Our number is that ceiling minus a harder split and multi-class dilution — not a training failure.

Confidence calibration

Raw detector scores are not probabilities. calibrator.pkl holds per-class Platt-scaling models; a single global calibrator made things worse.

Usage

Torch-free (recommended for deployment)

import cv2, numpy as np, onnxruntime as ort

sess = ort.InferenceSession("best_detector.onnx", providers=["CPUExecutionProvider"])
img = cv2.resize(cv2.imread("tile.png"), (640, 640))
x = np.ascontiguousarray((img[:, :, ::-1] / 255.0).astype(np.float32).transpose(2, 0, 1)[None])
out = sess.run(None, {sess.get_inputs()[0].name: x})[0]   # [1, 4+nc, 8400]

Full decode + NMS + calibration: edge/edge_infer.py in the GitHub repo.

Ultralytics

from ultralytics import YOLO
model = YOLO("best_detector.pt")
results = model.predict("tile.png", conf=0.10)

Preprocess your input. This model was trained on Lee speckle filter + CLAHE tiles. Apply despeckle_clahe() (ml/scripts/preprocess_sonar.py) before inference, or accuracy degrades silently. Use conf=0.10 — the per-class gate in Module 2 does the real cut.

Training

Exact hyper-parameters in args.yaml; the epoch-by-epoch log in results.csv.

Training data

~4,775 training tiles assembled from six real SSS surveys plus procedural synthetic data. ghost_net is 100 % synthetic — no public real ghost-net-in-SSS dataset exists (a Microsoft AI for Good / WWF effort had 412 real segments total and called it a feasibility study).

The dataset is not redistributed here; source provenance and the rebuild scripts are in the GitHub repo (docs/PROJECT_RECORD.html §15, ml/scripts/).

Limitations

• Does not transfer across surveys without fine-tuning. On an unseen instrument/seabed (AURORA) it produced near-zero detections. This is documented literature behaviour — GhostNetZero reports IoU 0.740 → 0.547 across regions — and the remedy is the same: fine-tune on a small labelled sample of the target sonar.
• ghost_net metrics are synthetic-on-synthetic. Treat as proof-of-capability, not field performance.
• crab_pot is not usable (AP@50 0.193 after two runs and a data doubling). It is a separability problem, not a data-volume one. Filtered from the product.
• INT8 is broken for this task. Static PTQ collapsed accuracy to 0.00 precision / 0.00 recall on every class, and was slower than FP32 on CPU (164 ms vs 90 ms) without VNNI/AVX-512. The file is published only so the result is reproducible. Use FP32 or FP16.
• Positional accuracy is a search area, not a survey fix. The tow-fish position recovers to < 1 m, but detection position also depends on across-track scale; two navigation paths can disagree by ~100 m on the same target.

Benchmarks (laptop CPU, onnxruntime 1.29, no CUDA EP)

Citation

@software{drishti2026,
  title  = {DRISHTI: AI-Powered Marine Debris Detection from Side-Scan Sonar},
  author = {Fazal, Rehan and others},
  year   = {2026},
  note   = {Smart India Hackathon 2026, Problem Statement 26057},
  url    = {https://github.com/Rehan9599/Sonar-Drishti}
}