jack-www/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.

Annotation versions

V3 revises the ground truth on the same 4,500 RGB images. It labels the coloured flare poles, excluding their weighted bases and sockets, and the gate uprights and top bar, excluding horizontal floor supports. Vertical bottom-post segments remain part of the gate frame. Drum, marker, golf-ball and pinger labels are unchanged. Reflection appearances remain background. The existing limitation on above-water refracted parts still applies.

Label release	Annotation master	Scope
v2	annotations.json	Original whole objects, including flare bases and gate supports
v3	annotations/v3/annotations.json	Flare poles and gate frames, with the supports above excluded

Tags v2 and v3 pin their publication snapshots. The root annotations.json and metadata.jsonl continue to describe v2. For v3, select the annotation master under annotations/v3/, or build the resolved v3 view below. The rgb loading configuration is the same for both versions. Image IDs, layout splits, RGB, depth, physical object identities, poses, and physical distance_m values are unchanged; distance still uses the original whole-object world bounding-box centre.

V3 contains 19,575 annotations and 900 empty-label images: gate 2,736, drum 10,278, marker 387, flare 2,997, golf_ball 2,997, pinger 180. An object with only an excluded base/support visible has no positive v3 annotation. Its original physical visibility is retained in metadata separately from selected-part visibility.

The small overlay owns revised COCO annotations, 500 shared geometry instance maps, per-image metadata/records, part-replay proof, and a separate manifest binding all shared assets to v2. RGB and depth are stored once. The two legacy models reuse their exact saved predictions; each label release selects its own threshold on the same validation split and reports the same held-out test layouts.

Download on your training machine and make a complete v3 filesystem view (Python standard library only; shared assets are symlinked):

hf download jack-www/sauvc-v1 --type dataset --revision v3 --local-dir sauvc
python sauvc/annotations/v3/materialize.py sauvc/annotations/v3 sauvc sauvc-v3

If you already have the original repository files, download only annotations/v3/** with the same hf download command's --include option. The view builder checks the annotation overlay and source-manifest binding; add --verify-shared to rehash every shared RGB/depth/diagnostic file. Keep the source and overlay directories alongside the view because its symlinks refer to them.

sauvc-v3/annotations.json then names the revised labels and all its asset paths resolve. In the rovnext repository, the existing rovnext-cv export-yolo command accepts this resolved root and writes a separate YOLO export. No training is performed by creating the view.

V3 annotation manifest SHA-256:

0fe7c12b82804559cbd4146ea6ca368b11588593dd1494680f4618f59e80cffb

Original v2 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.

Original split	Images	Layouts	Purpose
train	2,700	30	Training; current legacy predictions are diagnostic
val	900	10	Validation and legacy confidence-threshold selection
test	900	10	Held-out test

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.

Path	Format and meaning
images/*.png	Native RGB, front 1920×1200 or down 1280×800
annotations.json	COCO-compatible images/categories, tight visible-mask boxes [x,y,width,height] in native pixels, visible area, and uncompressed column-major mask RLE
instances/*.png	Unsigned per-pixel physical instance IDs; 0 is background; join with each annotation instance_id
depth/*.npz	Float32 range_m (native geometric distance to camera) and axial_depth_m; these are geometric depths, not wet optical-path lengths
metadata/*.json	Camera intrinsics/pose, scene layout, lighting, actual haze coefficients, object poses/colours/physical states, visibility facts and source hashes
diagnostics/*.reflections.png	Binary broad water-interface region for diagnostic overlap; it is not an exact reflected-object mask
records/*.json	Per-capture canonical record used to build the merged annotations
provenance/	Frozen source archive, plans, render recipe and capture provenance
manifest.json, manifest.sha256	Immutable canonical file inventory, byte sizes and SHA-256 checksums

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.

Original v2 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.