sauvc-v1

SAUVC synthetic underwater vision dataset

jack-www/sauvc-v1 is the first public Hugging Face publication of the accepted internal capture release v2, generated on 2026-10-10 with rovnext, Isaac Sim 5.1, and OceanSim. The repository name v1 identifies the public publication; it does not refer to the rejected internal v1 capture. All canonical release files retain their original paths and bytes. This card and metadata.jsonl are additive conveniences outside the canonical release manifest.

Browse saved images, ground truth, and both legacy model predictions. This separately hosted viewer may be stopped independently of the Hugging Face dataset.

Contents and splits

There are 4,500 RGB images, 19,647 visible-instance annotations, 50 layouts, and 900 explicit negative images. Each layout has 10 planned views/state captures and 9 lighting/visibility variants. Every view, state, and variant of a layout stays in one split; do not randomly split individual images.

Canonical category IDs and visible annotation counts: 1 gate (2,736), 2 drum (10,278), 3 marker (387), 4 flare (3,069), 5 golf_ball (2,997), 6 pinger (180). Colours and physical object states remain in the per-image/world-object metadata rather than creating extra category IDs.

Annotation distance_m is the physical Euclidean distance from camera eye to the aligned world bounding-box centre of the physical object. It is not apparent reflected range, minimum surface distance, or optical path length. Boxes span all visible pixels of the labelled physical semantic object, including supports/frame members assigned to that object; they are not outlines of only a bucket opening, coloured cloth, or scoring zone. Occluded/offscreen world instances are recorded separately and do not receive invented visible positives.

Load the RGB view

The default rgb configuration explicitly selects only images/ PNGs ; instance-ID and diagnostic PNGs are excluded. drop_metadata: true keeps the all-image convenience metadata separate, so it cannot accidentally duplicate rows across splits. It preserves the original split names train, val, and test and adds no directory-derived classification labels.

from datasets import load_dataset

data = load_dataset("jack-www/sauvc-v1", "rgb")
example = data["train"][0]
image = example["image"]
# The rgb view has only image; use annotations.json for full labels.
# metadata.jsonl is a compact, separate lookup keyed by file_name/sample_id.

Download metadata.jsonl separately and join by its relative RGB file_name or stable sample_id; it is not automatically attached to the RGB-only loading configuration. This intentionally avoids loading the full metadata table into each split.

metadata.jsonl is a compact projection of the original image/annotation records: IDs, split/layout/geometry/variant, dimensions, explicit negative flag, and parallel per-object arrays of IDs, canonical categories, visible boxes/area, and physical range. It does not replace the canonical RLE masks or richer per-image JSON. Empty object arrays mean a genuinely zero-annotation frame. Category IDs are 1-based as above.

For exact canonical files or custom loaders, download the repository snapshot and read annotations.json and the referenced assets. Check each canonical file against manifest.json; the additive Hugging Face card/metadata are not part of that original manifest.

Label scope and limitations

Positive labels cover directly visible submerged physical target parts. Reflected target appearances and water-interface imagery remain background, with no extra positive or ignore annotations. Native instance AOVs cannot identify above-water parts transmitted through the surface, so those parts are excluded by this release policy. This means the dataset is not a general annotation set for every optically visible part of a partly emergent target.

The broad water-interface diagnostic also includes transmitted/background water pixels. Its overlap statistic can identify detections on that region, but cannot establish that every overlapping prediction is a reflection false positive.

Golf-ball states describe actual simulated physical states and poses. They are not a human referee judgement or an independent drop-success ground-truth oracle. Paired flare/ball views are grouped by their layout and geometry facts.

Three lighting settings are crossed with nominal visibility settings 80 m, 50 m, and 15 m. The rendering coefficients are saved, but these visibility settings are not calibrated to a competition venue; the 15 m setting is a stress condition. The imagery is synthetic and should not be treated as evidence of real-world performance without real camera data.

Existing model evaluation

Both legacy all_combined_11s and all_combined_26s models were run on every one of the 4,500 images. A single confidence threshold per model was selected only on the 900 validation images and then frozen across test, train, lighting, and haze. The held-out headline is the 900-image/10-layout test split; train and all-image operating-point summaries are diagnostic, and the all-image summary includes the validation threshold-selection images.

The legacy task has four canonical categories: gate, drum, marker, flare. Raw cbuckets and tubs predictions map to drum, and raw flag maps to marker; raw predictions are preserved. Golf-ball and pinger ground truth remain in this dataset but are excluded from those legacy models’ metrics. AP uses 101-point interpolation at IoU .50:.95; the reports are not an official COCO summary with COCO area/crowd/maxDet subreports.

Integrity and provenance

Accepted canonical manifest SHA-256:

983c4ea62ac12d73b1353214e4aa7c33b87618445e00674a5f59fdcd62ec4b0f

Frozen source archive SHA-256:

9b510c16ec7c66b94a190c0635fbd306edd88ca454e4b985e564606a5677cd41

Validation established exact native instance, direct-label mask, metric depth, and broad water-interface consistency across every 9-variant geometry group; positive float32 subnormal depths were rejected. Every canonical copied file was independently checked against original worker SHA receipts, followed by a complete sealed-file reread. The earlier internal v1 capture was rejected after an AOV-buffer aliasing defect and is not included here.

This Hugging Face publication keeps the repository’s existing MIT licence designation. No model checkpoint or inference runtime is required to read the dataset.

ImageFolder/configuration references: Hugging Face image datasets and manual split configuration.