VR180: A Real-Captured Stereoscopic Dataset

VR180 is a collection of stereoscopic 180-degree video clips and stills recorded with two Blackmagic URSA Cine Immersive cameras. Every sample is provided as camera-native Blackmagic RAW (BRAW) and as two separate-eye HEVC representations of the same views, native fisheye and half-equirectangular, with the camera's factory lens calibration and scene annotations.

Version 1.0, released 6 October 2026.

Contents

Property	Value
Samples	1,211 (636 video clips, 575 stereo stills)
Total video duration	2,220.8 s
Clip duration	0.4–11.8 s, median 5 s
Frame rate	90 fps (632 clips), 30 fps (4 clips)
Native fisheye	8,160 × 7,200 per eye
Half-equirectangular	7,200 × 7,200 per eye, 180° × 180°
BRAW	5.411 TB
Fisheye HEVC, both eyes	314.779 GB
Half-equirectangular HEVC, both eyes	278.018 GB

Sizes are decimal bytes of the media files.

Files

braw/<sample_id>/<sample_id>.braw
braw/<sample_id>/metadata.json
h265_fisheye/<sample_id>/LEFT.mov
h265_fisheye/<sample_id>/RIGHT.mov
h265_fisheye/<sample_id>/metadata.json
h265_eq/<sample_id>/LEFT.mov
h265_eq/<sample_id>/RIGHT.mov
h265_eq/<sample_id>/metadata.json

File	Content
manifest.json	Every sample with its type, recording date, calibration ID, annotations and file paths
SHA256SUMS	SHA-256 of every media and metadata file
split-groups.json	Groups for leakage-aware train/test splits
metadata.schema.json	JSON Schema of the per-sample metadata
docs/DATA_FORMAT.md	Metadata fields
docs/ANNOTATIONS.md	Label definitions
docs/DATASHEET.md	Datasheet

Download

Access is gated with automatic approval: sign in, request access on this page, then authenticate with hf auth login.

# Index files
hf download lulinxuan/VR180 manifest.json SHA256SUMS split-groups.json \
  --repo-type dataset --local-dir ./VR180

# One sample in half-equirectangular form; replace <sample_id> with an ID from the manifest
hf download lulinxuan/VR180 --repo-type dataset \
  --include 'h265_eq/<sample_id>/*' --local-dir ./VR180

Use braw/*, h265_fisheye/* or h265_eq/* patterns to select representations, and */*/metadata.json for metadata only. The full dataset is about 6 TB. BRAW files need a Blackmagic RAW decoder, such as DaVinci Resolve or the Blackmagic RAW SDK.

Capture and processing

All footage was recorded by Linxuan Lu or by camera operators hired by Linxuan Lu. Clips were kept short to limit data volume and were chosen to cover diverse scenes. Video clips were cut from camera clips with the Blackmagic RAW SDK trim operation, which copies the compressed frames without re-encoding; stills are original camera clips. The HEVC representations (QuickTime, Main 10, 4:2:0, Rec.709 with gamma 2.4) were rendered in DaVinci Resolve. The half-equirectangular files cover 180° × 180° per eye; their projection was produced by Resolve and may differ slightly in orientation from other conversion tools. The camera sensor crops the lens image at the top and bottom, so about 11% of each half-equirectangular frame, mostly near the zenith and nadir, is black, while the fisheye files also contain image content beyond 90° at the sides.

Annotations

Each sample has a scene description, one of 41 scene types, an optional subtype, camera and object motion (videos), and eight visual challenges, each marked present, absent, uncertain or not applicable with a short reason. Labels were generated with the DeepSeek V4 API from sparse preview frames and reviewed with GPT-6 Astra; the six event scene types were checked by Claude (Anthropic). They were not verified by humans; use them to browse, filter and stratify the data rather than as ground truth. See docs/ANNOTATIONS.md.

Splits

No train/test split is fixed. Several samples can come from one camera clip, so split by group rather than by sample: split-groups.json assigns each sample a recording_group (same camera, recording date and clip number; 712 groups) and an overlap_component (recording groups joined by visual-similarity matches; 662 components). Put each overlap_component in a single split.

Validation

• All 1,211 samples have their BRAW file, four HEVC files and three metadata files.
• For all 4,844 HEVC files, dimensions, frame counts and frame rates match the BRAW source, and left and right eyes are paired.
• All 2,422 half-equirectangular files (399,484 eye-frames) decode completely without errors.
• For every sample, the tone statistics of the fisheye and half-equirectangular files agree (frame 0, mean saturation ratio 0.85–1.41, median 1.02).
• All metadata files validate against metadata.schema.json.

People and privacy

The footage was recorded mostly in public places such as event venues, streets, gardens, museums and zoos. Bystanders, audiences and performers may appear and some may be identifiable. Individual consent was not obtained from people incidentally captured, and faces are not blurred, because the dataset provides unmodified camera-native imagery. The annotations contain no information about the identity of any person. Identifying, tracking or profiling individuals is not an intended use of this dataset.

If you appear in the dataset and want samples removed, open a Discussion with the sample ID or a description of the scene and date. Removed samples are excluded from later versions.

Code

A converter between the native fisheye and half-equirectangular representations that uses the factory calibration, and the code of the accompanying analyses, are at https://github.com/lulinxuan/vr180-dataset-tools (MIT License).

License and citation

The media and annotations are licensed under CC BY-NC 4.0. For commercial use, contact the author through Discussions.

@misc{vr180_2026,
  author  = {Lu, Linxuan},
  title   = {{VR180}: A Real-Captured Stereoscopic Dataset},
  year    = {2026},
  version = {1.0},
  url     = {https://huggingface.co/datasets/lulinxuan/VR180}
}