Light Field Dataset LF@FIT

This dataset contains 4K synthetic light field scenes with 15x15 views. All views are positioned in a uniform grid with constant spaces.

Related research

The dataset was released as a part of the paper Lightweight all-focused light field rendering. The paper evaluates previously proposed ways of disparity/focus map estimation and proposed novel algorithms to improve the resulting quality of the map and the computational performance. Such maps are used to synthetize all-focused novel views in light field rendering.

The following video quickly introduces the problem and describes the solution:

The following demo demonstrates the simplified mechanism of focus map generation and relevant novel view synthesis from light field.

The green cameras mark the 4 views that are used in the focus map generation.

Current position: 0.5 0.5

Click into the grid to select a desired view..

Loading resources. This might take a few minutes. Downloading 40 MB of light field images. The demo will be working once the downloading is finished.

The slider splits the result to the synthetic view and the focus map.

The view position can also be changed by pressing and dragging left mouse button inside the scene.

Asset sources

All assets used in the dataset scenes were available under CC0 licence. They were obtained from the following sources:

Textures: ambientcg.com

Stanford Bunny: graphics.stanford.edu

Public domain 3D models: sketchfab.com

Blender demo files: blender.org

Citation

If you find our dataset or method useful, please cite our work:

@article{chlubna2024light,
title = {Lightweight all-focused light field rendering},
journal = {Computer Vision and Image Understanding},
pages = {104031},
year = {2024},
issn = {1077-3142},
doi = {10.1016/j.cviu.2024.104031},
url = {https://www.sciencedirect.com/science/article/pii/S1077314224001127},
author = {Tomáš Chlubna and Tomáš Milet and Pavel Zemčík}}

How to use

The dataset was rendered with Cycles rendering engine using Blender 3D editing software. The .blend project files are available and contain the same settings as were used in the rendering of the views. The light field views were rendered with a custom Light Field Data Renderer & Capturing Addon and the parameters used in its GUI are listed at eachs cene below.

Two sparser datasets (8x8 as rendering input and the rest between the input views as reference) used to evaluate the synthesis in between the frames can be extracted using this script.

The views were rendered as PNG images and compressed using lossless AV1 video encoding with FFmpeg:

ffmpeg -i data/%05d.png -c:v libaom-av1 -crf 0 -aom-params lossless=1 data.mkv

The views were can be extracted with ffmpeg:

ffmpeg -i data.mkv data/%05d.png

The processing script is also available.

Scenes

The scenes were obtained and designed to represent classic candidates for light field rendering and photography in general. Some of the scenes are examples of easy/hard cases for image-based rendering methods and serve for benchmarking and testing purposes. Light field video examples are also present in a form of short seamless loop animations. The parameters can be used to reproduce the render process using the published Blender addon or to render a novel view with the published light field focusing and rendering code. Center and four corner views are shown in the pictures.

Blender BMW

Two reflective objects in neutral simple setting. Example of possible product advertisment scenario.

Blender Cat

Furry animal in grassy environment with depth-of-field effect. Artistic scene like from an animated movie.

Blender Classroom

Classic example of indoors scene - room with a lot of chairs and desks with unique objects in the back.

Blender Diorama

Indoors scene with a few objects of interest.

Buildings

Outdoors scene with buildings and one object of interest in the front.

Colorful

The whole color spectra in one scene.

Colorless

The whole scene is desaturated, mostly consisting of shades of gray.

Cornell Box

Standard testing scene for 3D rendering methods.

Diffuse

Simple 3D objects containing only diffuse materials.

Face

Human face portrait containing depth of field and subsurface scattering materials.

Green Screen

Various objects placed in front of a green screen setting which is commonly used in cinematography to capture objects without background.

High Frequency

Natural scene containing a lof of high-frequency features.

Large Depth

Outdoors environment with very large depth range and objects distributed along the whole range.

Low Depth

Objects lying on a table with very small depth range.

Low Frequency

Outdoors scene containing a lot of low frequency features and color gradients.

Macro

Example of close macro photography.

Reflective

Ultra reflective scene with caustics and tranparency.

Simple Setting

Simple black-screen scene with lit objects of interest in the middle.

Single Object

One single object on black background.

Stanford Bunny

Variation of Stanford Bunny - famous testing 3D model. Contains basic scene elements: object of interest, flat ground and static sky.

Text

Objects containing text in 3D environment.

Volumetric

Volumetric effects on black background. The scene contains fire and smoke simulation with sparks simulated by particles.

Bonfire Animation

Animated light field video of bright and dark neutrally colored areas with volumetric effects.

Street Animation

Animated light field video capturing a street with passing reflective car.

Timelapse Animation

Animated light field video with a wast landscape and changing sun elevation.

Keying Animation

Moving character in front of green screen setup.