Deblur-NeRF-SAI
Authors/Creators
Description
Modified version of the Deblur-NeRF dataset, used in under the CC BY 4.0 license. The original dataset, created by Ma Li, Li Xiaoyu, Liao Jing, Zhang Qi, Wang Xuan, Wang Jue, and Pedro V. Sander , is described in https://doi.org/10.1109/CVPR52688.2022.01252 and avaiable at https://limacv.github.io/deblurnerf/.
The data was extended to include a variant with artificial rolling shutter artefacts. In addition, we regenerated the motion blurred images using the Blender files, which we modified by fixing the random seed, and also included velocity information in our outputs. Furthermore, we switched from the custom 10-sample blur implementation to Blender's built-in motion blur and rolling-shutter effects, which are less prone to sampling artefacts. Finally, we adjusted the caustics rendering settings in the Blender scenes to reduce non-deterministic raytracing sampling noise in the data with reasonable sample counts.
Files
blend_files.zip
Additional details
Related works
- Is derived from
- Publication: 10.1109/CVPR52688.2022.01252 (DOI)
- Is described by
- Preprint: arXiv:2403.13327 (arXiv)
Software
- Repository URL
- https://github.com/SpectacularAI/3dgs-deblur