idg101/Aperture_Lab_Synthetic_Side_Scan_Sonar_v1

ApertureLab Synthetic Side-Scan Sonar Dataset v1.0 (EdgeTech 4125i)

Side-scan sonar is a standard instrument for surveying the seabed, and models that find objects in its imagery need labelled examples of those objects on many kinds of seabed. Real surveys rarely provide them: the objects are sparse, their positions are seldom known exactly, and each image must be labelled by hand. This dataset supplies 250 simulated side-scan swaths with exact labels for seven classes of seabed object, each with the raw data it was imaged from.

Every scene was produced by ApertureLab's point-scatterer sonar simulator and side-scan image formation. The sonar is modelled on the EdgeTech 4125i's 400 kHz channel, imaging a 75 m starboard swath from a towfish flying 6 to 10 m above the bed.

Classes

id	class	what it is	instances
1	ghost_net	a net mound with two to four trailing net lines	194
2	rope_line	a polyethylene line lying on the bed	144
3	container	a 20 ft shipping container	103
4	wreck	a small boat hull, 9 to 13 m long	91
5	tire	a vehicle tyre	392
6	pipe	a steel pipeline 0.2 to 0.5 m across, up to half buried	118
7	natural	a boulder or a rock pile	512

A ghost net is one instance: its mound and trailing net lines share one polygon. Rocks are labelled only as placed natural objects; the seabeds carry no unlabelled boulder fields.

Files

path	contents
images/sss_XXXX_drc.png	the displayed image (dynamic-range compressed), slant range to the right, along track up
xtf/sss_XXXX.xtf	the raw side-scan data for the same swath, in XTF
labels/sss_XXXX_coco.json	COCO labels for that image: polygons, boxes, and per-object attributes
scenes/sss_XXXX.yaml	the complete scene description the simulator ran
coco_all.json	every image and annotation in one COCO file
manifest.csv	one row per scene: split, seed, altitude, sound speed, seabed family, bathymetry, and instance counts per class

Labels are in the image's own pixel frame, which is slant range. Each COCO image record carries its pixel size, the along-track and range extents, the altitude and the sound speed; each annotation carries the object's height, burial and acoustic-shadow polygon where they apply.

Splits

split	scenes
train	200
val	25
test	25

How the scenes vary

• Seabed: sand, silt, mud or gravel, in uniform, patchwork, rippled or pockmarked arrangements, with ripple fields, material patches and gentle relief (rolling ground, tilt, sand waves), drawn from the recipes of the ApertureLab synthetic SAS dataset v1.1 with relief scaled for the lower towfish.
• Acquisition: altitude 6 to 10 m, sound speed 1480 to 1520 m/s, and small sway, heave, roll, pitch and yaw of the towfish.
• Objects: two to seven classes per scene at random positions, headings and sizes, with no overlaps and no crossing lines or pipes; pipes are partly buried.

Limitations

The data are simulated. The images carry no electronic noise, and the objects are generic meshes and primitives rather than surveyed examples. Models trained on these data should be evaluated on real survey data before they are relied on.

License and attribution

Released under the ApertureLab Synthetic Sonar Dataset License (ALSD-1.0), side-scan edition; see LICENSE.md, including the exclusions in Section 3. Credit "ApertureLab Synthetic Side-Scan Sonar Dataset, Isaac Gerg" with a link to this page. More about ApertureLab: https://gergltd.com/aperturelab/