PaintUnicorn: Multi-view video datset

PaintUnicorn: Real-World Multiview Dynamic Scene Dataset

**PaintUnicorn** is a real-world multiview video dataset captured with 28 synchronized rolling shutter cameras arranged on a semicircular rig with 4 vertical layers. It is designed for benchmarking neural rendering, dynamic view synthesis, and 4D reconstruction methods under real-world conditions.

---

 License

Creative Commons Attribution 4.0 International (CC BY 4.0)

---

Terms of Use

If you use this dataset in any publication or project, please cite the following:

@misc{PaintUnicorn_2025,

  title     = {PaintUnicorn: Multi-view video datset},

  author    = {Eva Dubar and Daniele Bonatto and Sarah Dury and Gauthier Lafruit},

  year      = {2025},

  doi       = {10.5281/zenodo.16312255},

  publisher = {Zenodo}

}

Production:

Laboratory of Image Synthesis and Analysis, LISA department, Ecole Polytechnique de Bruxelles, Universite Libre de Bruxelles, Belgium.

Dataset Overview

This dataset captures dynamic scenes with a wide range of real-world artifacts:

- Non-Lambertian surfaces and specular highlights

- Moving human subject, hair, clothes, transparent and reflective objects

- Depth maps with geometric noise due to challenging materials

These characteristics make PaintUnicorn a robust benchmark for methods that aim to handle realistic scenarios, moving beyond controlled synthetic settings.

Content:

In addition to the images and their depth maps, an accurate camera calibration file is provided following the format of [2].

The dataset contains:

- a `camera.json` Camera intrinsics and extrinsics in OMAF coordinates (X: forward, Y: left, Z: up) [3]

- a `yuv_unprocessed.zip` Undistorted and color-corrected frames, cropped for spatial consistency

- a `yuv_processed.zip` Raw frames from all 28 cameras (distorted, uncorrected)

- `depth.zip` Per-frame depth maps (hybrid from COLMAP and DepthPro refinement) in RVS format [1]

- `cameras.txt`, `images.txt` from COLMAP reconstruction

- `poses_bounds.np` LLFF-style camera poses and scene bounds

 References and links:

[1] D. Bonatto, S. Fachada, S. Rogge, A. Munteanu and G. Lafruit, "Real-Time Depth Video-Based Rendering for 6-DoF HMD Navigation and Light Field Displays," in IEEE Access, vol. 9, pp. 146868-146887, 2021, doi: 10.1109/ACCESS.2021.3123529.

[2] S. Fachada, B. Kroon, D. Bonatto, B. Sonneveldt, et G. Lafruit, "Reference View Synthesizer (RVS) 2.0 manual, [N17759]", july. 2018.

[3] S. Fachada, D. Bonatto, M. Teratani, and G. Lafruit, "Intechopen - View Synthesis tool for VR Immersive Video", 2022.