A Camera-Native Stereo VR180 Dataset for Immersive Video Research

Authors: Linxuan Lu

Submitted: 7 October 2026

arXiv: 2610.10607 [cs.CV]

Subjects: Computer Vision and Pattern Recognition (cs.CV); Multimedia (cs.MM)

ACM Classes: H.5.1; I.4.2

Abstract

Immersive VR180 video is increasingly produced with professional stereo fisheye cameras, yet public VR180 research resources are mostly collected from online platforms such as YouTube—already stitched, projected, and compressed by unknown pipelines, and without lens calibration. This limits reproducibility and makes it difficult to study the effects of camera-native processing on downstream tasks.

We present a firsthand-captured stereo VR180 dataset recorded with two Blackmagic URSA Cine Immersive cameras. The dataset contains 1,211 samples:

• 636 stereo video clips (2,220.8 s, mostly 90 fps)
• 575 stereo stills

Each sample is released in three complementary forms:

1. Camera-native Blackmagic RAW
2. Separate-eye native fisheye HEVC (8160×7200 per eye)
3. Half-equirectangular HEVC (7200×7200 per eye)



The release also includes the factory lens calibration, portable fisheye/half-equirectangular conversion tools, and AI-generated scene and visual-challenge annotations.




Key Finding: Fisheye vs. Half-Equirectangular Coding Efficiency




To characterize the coding behavior of the released formats, we re-encoded the fisheye and half-equirectangular renders with x265 across 24 clips, both eyes, four rate points, and nine viewing directions. At equal viewport quality, native-fisheye coding required more bitrate than half-equirectangular coding for all 24 clips (median +38%), and this gap held in every part of the field of view. This result highlights a practical trade-off between preserving native sensor geometry and minimizing bitrate for viewport-adaptive streaming.




Contributions



• A camera-native stereo VR180 dataset with both video and still content captured under controlled conditions
• Multiple representation formats (RAW, native fisheye HEVC, half-equirectangular HEVC) to support a range of research pipelines
• Factory lens calibration and conversion tools for reproducible processing
• AI-generated scene and visual-challenge annotations for benchmarking
• A systematic rate–quality comparison of fisheye versus half-equirectangular coding under viewport-quality constraints



Resources



• Dataset: https://huggingface.co/datasets/lulinxuan/VR180
• Code: https://github.com/lulinxuan/vr180-dataset-tools



Citation




@misc{lu2026vr180,
  title={A Camera-Native Stereo VR180 Dataset},
  author={Linxuan Lu},
  year={2026},
  eprint={2610.10607},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  doi={10.48550/arXiv.2610.10607}
}




Comments: 6 pages, 5 figures, 6 tables.

via ArXiv CV