rehan9599/drishti-sss

The preview viewer is off by design. This is a YOLO-format detection set ({split}/images/*.jpg + {split}/labels/*.txt), meant to be pulled with snapshot_download and trained with Ultralytics — it is not a load_dataset() dataset, and HF's auto-parquet converter cannot parse the paired label files.

DRISHTI — side-scan sonar training splits

The assembled, preprocessed train / val / test tiles behind the DRISHTI detector — an SIH 2026 (PS 26057) marine-debris and anomaly detector for side-scan sonar. YOLO format, 640 px tiles.

Licence — CC-BY-SA-4.0

This release is Creative Commons Attribution-ShareAlike 4.0 International. You may share and adapt it, including commercially, provided you (a) credit the upstream sources below and this repo, and (b) release any derivative dataset under CC-BY-SA-4.0 as well.

The ShareAlike term is inherited: two of the source datasets (mine, and the crab_pot set that is excluded here) are themselves CC-BY-SA-4.0, so the assembled whole must be too.

Attribution (required)

Portion	Source	Licence
pipe, bg	SubPipe / SubPipeMini2 — Álvarez-Tuñón et al., OceanScan-MST	CC-BY-4.0
wreckA	AI4Shipwrecks — Sethuraman et al., UM Field Robotics / NOAA Thunder Bay	CC-BY-4.0
wreckR	Side Scan Sonar (Ship, Plane) — Dae Hyeok Lee, Roboflow Universe	CC-BY-4.0
mine	Sonar Imaging Mine Detection — MILCO contacts	CC-BY-SA-4.0
synth	procedural acoustic generator (ml/scripts/build_synthetic_data.py)	original work, CC-BY-SA-4.0 with the rest

The synth background canvases are drawn from the Roboflow SSS set (CC-BY-4.0), then composited with procedurally modelled objects and filtered — the source imagery is not recognisable in the output.

What is not in this release

• crab_pot (class id 0) — 1,200 tiles were used to train the shipped model as a hard negative, from a HuggingFace set that is now access-gated. They are omitted here. Class 0 therefore has zero examples in this repo. To reproduce the exact training set, rebuild from the GitHub ml/scripts/ recipe.
• KLSG / SeabedObjects (Ship & Airplane) — never contributed a tile (KLSG_TRAIN_CAP = 0 in build_dataset.py); its weak full-frame boxes hurt shipwreck precision. Listed only so the exclusion is on the record.

Contents

Split	Images	Labels
train	3,875	3,875
val	630	630
test	700	700
total	5,205	5,205

~1.8 GB. Layout:

train/images/*.jpg   train/labels/*.txt
val/images/*.jpg     val/labels/*.txt
test/images/*.jpg    test/labels/*.txt
drishti.yaml         # Ultralytics data config

Labels are YOLO boxes: class_id x_center y_center width height (normalised).

Classes

id	class	in this release
0	crab_pot	— (excluded, see above)
1	submarine_pipeline	✓
2	shipwreck	✓
3	ghost_net	✓ (100 % synthetic)
4	mine_cylinder	✓

The shipped product uses classes 1–4. ghost_net is fully 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.

Preprocessing — already applied

Every tile has been through Lee speckle filter + CLAHE (despeckle_clahe() in ml/scripts/preprocess_sonar.py, clip 3.0, 8×8 grid, 7×7 Lee kernel). Do not apply it again. If you train on these tiles, apply the same filter to your inference inputs.

Speckle in sonar is multiplicative (I_obs = I_true · n), so a plain blur destroys the object and shadow edges that carry the signal. The Lee filter is a local MMSE estimate that smooths flat seabed and preserves edges; CLAHE lifts faint contrast with a clip limit so flat-sand speckle is not amplified.

Note: an ablation found this preprocessing gave no accuracy gain over raw tiles once training-time augmentation was strong — it is retained because CLAHE'd input makes acoustic shadows more detectable for the downstream geometry check.

Provenance — per prefix (train split)

Filenames carry their source.

Prefix	Count	Source	Nature
synth	1,250	procedural generator, modelled acoustic physics	original work
pipe	1,000	SubPipeMini2 survey strips, tiled	real
wreckA	546	AI4Shipwrecks transects (pixel masks → boxes)	real
bg	500	object-free SubPipeMini2 tiles (hard negatives)	real
wreckR	354	Roboflow side-scan-sonar (Ship + Plane)	real
mine	225	Kaggle sonar-mine (MILCO contacts)	real

Full source detail: docs/PROJECT_RECORD.html §03 and §15 in the GitHub repo.

Known caveats

• Shipwreck split not yet audited site-disjoint. An audit to guarantee no wreck site appears in both train and test is pending; treat shipwreck metrics as optimistic until it lands.
• The test set is deliberately hard. A 50 %-overlap re-tile tripled shipwreck test instances with partial and near-duplicate tiles. Numbers on this split are not comparable to papers using a non-overlapping tiling of the same source.
• ghost_net evaluation is synthetic-on-synthetic. Not a field number.
• Class imbalance is deliberate. Per-class caps per split; a dominant class suppresses accuracy on feature-dissimilar classes.

Usage

from huggingface_hub import snapshot_download
snapshot_download("rehan9599/drishti-sss", repo_type="dataset",
                  local_dir="ml/data/splits")

yolo detect train data=drishti.yaml model=yolov8s.pt imgsz=640 epochs=120 batch=16

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}
}

Please also cite the upstream datasets listed under Attribution.