Ground-truth labels from the scene file

Labeled sonar imagery is scarce because labeling it is expensive: an analyst inspects a gray, speckled scene and decides where each object ends and the seafloor begins. ApertureLab writes six categories of annotation for every simulated image, in COCO format, from the same scene file that the simulation reads. The position, footprint and heading of every object are known when the image is formed, so the labels are computed from the scene description instead of inferred from the image.

Six annotation categories in world coordinates

Each simulated image is accompanied by annotations for objects, rocks, biota, pockmarks, trawl scars and seafloor types. They are generated in world coordinates and projected into the image, so they remain registered with the image when the acquisition geometry changes.

Objects are labeled with an oriented rectangle taken from the footprint and heading of the object in the scene file, instead of an axis-aligned box fitted to the image afterward.

Cables and pipelines are labeled with one buffered polyline per run, split at gaps and at heading breaks, so a pipeline that changes direction is one annotation instead of a chain of fragments.

Clutter fields (rocks, shells, biota) are labeled with one bounding region instead of a box per pebble.

Seafloor types are labeled with a traced, Chaikin-smoothed outline per zone, taken from the zone raster that the simulator rendered.

Burial is recorded as a per-object property, so a partially buried target is distinguishable from an exposed one at training time.

COCO export with the imaging geometry

Annotations are exported as COCO detection JSON with polygon segmentation, the format that off-the-shelf detectors read without a custom loader. Each image record also carries pixel_size_m, along_extent_m, and range_extent_m, so a pixel box can be converted back into meters, and a georeference attached later, without re-deriving the imaging geometry.

Per-image files are concatenated into one dataset with merge_coco(). Because each scene is a file, a dataset is generated as a parameter sweep over scene files, and its cost scales linearly with the number of images and requires no survey vessel.

A simulated synthetic aperture sonar scene beside the same scene overlaid with green object bounding boxes, object names and positions, and yellow traced seafloor-zone boundaries

A simulated scene is shown beside the same scene with its labels. Oriented object footprints are drawn in green with class, position and heading, and the traced boundary of each seafloor type in yellow. Both label sets are written from the scene file at simulation time.

Scene generators and their outputs

Five scene generators produce the collections below: three yield labeled imagery, and two yield scene configurations from which imagery can be simulated.

The public dataset contains 1000 HISAS-class swaths with 2000 labeled targets, 2479 labeled clutter objects, COCO polygons and every scene file, and is published on Hugging Face. It is described on the dataset page.

A target-on-background grid places 60 instances of one object across ten seafloor backgrounds so that a 256 by 256 chip lies entirely inside one homogeneous background, and lists each instance in a CSV manifest.

Change-detection pairs are scene configurations for the same patch before and after a change, pixel-registered by construction, with height-map previews; no imagery has been simulated from them.

The linear-feature scenes are 20 simulated scenes of pipes and cables with varying burial, constrained so that no two runs intersect.

A 151-scene ocean archetype library holds scene configurations without rendered imagery, covering wrecks, pipelines, debris fields, trawl scars and pockmarks; each scene is parametric and can be reseeded.

A grid of identical targets rendered against ten different seafloor types including fine ripple, coarse ripple, rocky, mud, and gravel

One object is shown on ten seafloor types. The domain randomization comes from simulating each background as a different seafloor instead of from post-processing the image.

Each image in these collections and its labels come from one simulation of one scene file; no image was labeled by hand.