Researchers release VR180 dataset shot on Blackmagic cameras: 1211 stereo samples

Researchers have published a VR180 dataset captured directly on two professional Blackmagic URSA Cine Immersive cameras. It contains 1211 samples: 636 stereo video clips with a total duration of 2220.8 seconds (most shot at 90 frames per second) and 575 stereo photos. The main result of the accompanying experiment: encoding native fisheye requires a median of 38% more bitrate than half-equirectangular at the same quality in the viewport.

How does this dataset differ from

YouTube-sourced collections?

Most public VR180 resources are collected from online platforms such as YouTube. The videos there have already been stitched, projected and compressed by unknown pipelines, and no lens parameters are attached. For compression, quality assessment or scene reconstruction work, this is a serious problem: artifacts from the original processing get mixed in with whatever the researcher wants to measure.

The authors took the opposite approach and put everything the camera produces into the repository. Each sample is available in three representations, and alongside them sit the factory lens calibration and portable utilities for converting between fisheye and half-equirectangular.

• Size: 1211 samples, of which 636 are videos (about 3.5 seconds per clip on average) and 575 are photos.
• Cameras: two Blackmagic URSA Cine Immersive units.
• Formats: Blackmagic RAW, separate per-eye fisheye HEVC (8160×7200 per eye) and half-equirectangular HEVC (7200×7200 per eye).
• Frame rate: mostly 90 fps.
• Extras: factory lens calibration, converters, and AI-generated descriptions of scenes and visual challenges.
• Access: the dataset is on Hugging Face, the tools are on GitHub.

The scale of the frames is notable. A single fisheye eye is about 58.8 megapixels, a stereo pair exceeds 117 megapixels, and at 90 fps that comes to roughly 10.6 billion pixels per second before compression. Working with streams like this calls for real source footage, not recompressed copies.

What did the encoding test show?

The authors tested how the two representations behave under compression. They re-encoded the fisheye and half-equirectangular renders with the x265 encoder on 24 clips, for both eyes, at four bitrate levels and for nine viewing directions. Quality was compared in the viewport, that is, in what a user sees in the headset.

The result is unambiguous: on all 24 clips, native fisheye needed more bitrate than half-equirectangular at equal quality. The median difference was +38%. It held across the entire field of view, not just at the periphery, where fisheye distortion is expected to be stronger. The abstract does not explain why, and that remains a question for future work.

The limitations are also clear. The experiment covers one codec (x265) and 24 of the 636 clips. It does not confirm conclusions about other codecs, including AV1 and neural ones. The scene annotations are AI-generated rather than labeled by hand, so annotation quality should be spot-checked.

Who should use it, and how?

For teams building VR180 streaming, the takeaway is practical. If your pipeline stores or transmits native fisheye, budget roughly a third more bitrate than for half-equirectangular, or convert before encoding. But conversion introduces its own losses, and now they can be measured, because the release includes both the calibration and the tools.

For researchers working on neural compression, novel view synthesis and quality assessment, the dataset offers a clean starting point. Raw Blackmagic RAW lets you choose your own color processing and compression, and the factory calibration removes the need to estimate lens parameters yourself. Before using it, check the license on the Hugging Face page and plan disk space in advance: 8K stereo in RAW takes up a lot.