SyntheticFur Dataset Description

• Paper: https://arxiv.org/abs/2105.06409
• Video: https://youtu.be/5uraQu_5Tyg

Overview

Collecting and generating high quality fur images is an expensive and difficult process that requires content specialists to generate. By releasing this unique dataset with high quality lighting simulation via ray tracing, this can save time for researchers seeking to advance studies of fur rendering and simulation, without having to recreate this laborious process.

The dataset was used for neural rendering research at Google that takes advantage of rasterized image buffers and converts them into high quality raytraced fur renders. We believe that this dataset can contribute to the computer graphics and machine learning community to develop more advanced techniques with fur rendering.

Download

The dataset are divded into png images (~20GB), and Alembic simulation files (> 1TB).

We recommend to download the images separately from the Alembic simulation files since they can take up a lot of your local storage. The simulation files are entirely optional, and only needed if you want to experiment on building a neural physics model.

Individual zip files containing portions of the dataset can be separately downloaded:

• All images
• Alembic files (bunny)
• Alembic files (rubbertoy)
• Alembic files (sphere)
• Alembic files (torus)
• Alembic files (tube)

What's in this?

The dataset contains ~140,000 images and 15 Alembic files generated from scratch using Houdini and Zync.

The dataset has the following definitions:

• Scene: a sequence of frames that together represents a continuous motion of fur. Each scene contains frames of images and optionally 1 corresponding Alembic file for simulation data. See Table 1 - 5 below for details about each scene.
• Frame: each frame is a set of conditional images that represents the ground truth and the input images.

Data Structure

Images

• Resolution: 1024x1024
• Format: png
• Directory naming convention: <scene_name>/<buffer_name><4 digit frame count>.png
• Frames / scene: 720
• Buffer names: HighQualityRender (ground truth), Rasterized, SceneDepth, LitPrimitive, GuideColored, WorldNormal, Mask

Example

HighQualityRender (ground truth)

GuideColored	LitPrimitive	SceneDepth	WorldNormal	Mask	Rasterized

Note:

• The GuideColored buffer has random colors for white fur scenes, and brown color for brown fur scenes.
• The Mask buffer can be used for sampling specific regions of interest if trained the images as crops.
• We find that the Rasterized buffer type takes too long and costly to generate and does not fit our description of inexpensive inputs, therefore excluding it during training.

Simulations

Alembic is an open computer graphics interchange framework. The Alembic files capture the fur strand positions per frame.

A groom consists of hair strands. Each hair strand is a series of connected line segments. The following information can be extracted from the Alembic files:

• Number of hair strands
• Number of segments per hair
• The hair point’s position per frame
• The hair point’s velocity per frame
• etc.

Note: Each Alembic file only contains the definition of the groom, and does not include the skin primitives nor the guide curves.

Scene Description

See here.

Houdini Scene Setup Guide

See here.

Citation

@misc{le2021syntheticfur, title={SyntheticFur dataset for neural rendering}, author={Trung Le and Ryan Poplin and Fred Bertsch and Andeep Singh Toor and Margaret L. Oh}, year={2021}, eprint={2105.06409}, archivePrefix={arXiv}, primaryClass={cs.LG} }