rehan9599/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).

id	class	shipped
0	crab_pot	✗ filtered
1	submarine_pipeline	✓
2	shipwreck	✓
3	ghost_net	✓ (synthetic training data)
4	mine_cylinder	✓

Files

File	Size	Notes
best_detector.pt	22.5 MB	shipped model — Run 3, trained on Lee+CLAHE preprocessed tiles
best_detector.onnx	44.8 MB	FP32 ONNX — the edge deployment model, same weights
best_detector_fp16.onnx	21.4 MB	FP16 — Jetson GPU target
best_detector_int8.onnx	11.0 MB	INT8 — excluded, kept for the record (see Limitations)
best_detector_prep.pt / best_detector_raw.pt	22 MB ea.	Run 3 / Run 2 backups
calibrator.pkl	2 KB	per-class Platt-scaling calibrators (Module 2)
results.csv, args.yaml, *.png	~1 MB	training curves, confusion matrix, PR curves, exact hyper-parameters

Results — held-out test set (850 tiles)

Metric	Value
mAP@50	0.580
mAP@50-95	0.434
precision	0.734
recall	0.629
false-positive rate	0.266

Per class (AP@50):

Class	AP@50	Note
ghost_net	0.995	synthetic data — not a field number
submarine_pipeline	0.984	production-grade
mine_cylinder	0.424	real-data ceiling
shipwreck	0.302	see below
crab_pot	0.193	filtered from the product

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.

	Expected Calibration Error (test)
raw scores	0.052
per-class calibrated	0.037

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

base	yolov8s.pt (COCO-pretrained)
image size	640 × 640
batch	16
epochs	120 requested, early-stopped at 96 (patience 25)
hardware	RTX 4050 laptop, ~2.1 h
preprocessing	Lee speckle filter + CLAHE, applied identically at train and serve
augmentation	SSS-tuned: colour off (grayscale), vertical flip off (preserves highlight→shadow polarity), mosaic 0.8, mixup 0.1, erasing 0.4, rotate ±10°, shear 2°
loss	CIoU + BCE + DFL

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)

Model	Size	Latency	Accuracy
FP32 ONNX	44.8 MB	~90 ms/tile (11 FPS)	matches PyTorch
FP16 ONNX	21.4 MB	n/a on CPU	Jetson GPU target
INT8 ONNX	11.0 MB	164 ms/tile	collapsed

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}
}