RenderMe-360
Large Digital Asset Library and Benchmark Towards High-fidelity Head Avatars

Get Started

Overview

High Fidelity

   2K Logo   60 view   360 Logo

High Diversity

Different eras of ages✓
  Distinctive-Styles Materials ✓
  Dynamic Motions ✓


Rich Annotations

*Cameras' Calibration* *Background Matting*
*SCAN* *2D-3D Key Points*
 *FLAME Fitting* *Text Annotation*

Synthesizing high-fidelity head avatars is a central problem for many applications on AR, VR, and Metaverse. Compared to fast progress on the head avatar synthesizing algorithm, current public datasets can only support researchers to explore high-fidelity head avatars in one or two aspects, such as viewpoint, pose, hairstyle, or motion. Moreover, existing datasets usually contain portrait images with limited amounts and narrow distribution over different attributes, such as expressions and ages.

In this paper, we present RenderMe-360, a comprehensive 4D human head dataset to drive advance in head avatar algorithms across different scenarios. RenderMe-360 contains over 243 million complete head frames, diverse appearances, and over 80 million video sequences from 500 different identities captured by 60 synchronized HD cameras at 30 fps.

What makes RenderMe-360 different?

ID

Views

Expressions

Sentences

Wigs

Frames

Scales

  • RenderMe-360 is a large scale dataset with 500 IDs and 243M frames in total, far exceeds other datasets.
  • A wide diversity including era, ethnicity, accessory and makeup.
  • Each subject capture about 20-30 performance parts in cluding expression, hair, and speech

Realism

  • Build a multi-video camera capture cylinder called POLICY to capture synchronized multi-view videos
  • 60 instructive cameras / 2448 × 2048 / 30FPS for video capture

Granularity

Rich and multimodal annotation far beyond other datasets

  • face landmark 2d & 3d
  • front-back matting
  • FLAME parameters
  • scan mesh
  • uv map
  • Action Units
  • appearance annotation
  • text description



Statistics

Identity Statistic

  • Ages between 5 and 80. Teenagers and adults form the major part.
  • Height-weight distribution map. It indicates a large part of the models is located in height between 155cm and 185cm, and weight between 50kg and 90kg.
  • Gender-balanced with 4 ethnicities (217 Asian, 140 White, 88 Black, and 55 Brown)

Annotation Statistic

We obtain a dataset with more than 243M frames which are fine-grained annotated.

  • RenderMe-360: Expression-Part, Wig-Part, and Speech-Part
  • 71M frames with 2D landmarks and matting
  • 4.8M frames with 3D landmarks
  • 0.6M FLAME Fitting

Camera View Distribution

  • Our capture system has 60 cameras which forms 3 and a half circles.
  • The cameras are split into two parts with an angle divider from 90 degree to 270 degree
  • The frontal views take up 65%.

Accessory Statistic

  • Parts of Asians (more than 40%) are captured with special clothing and head accessories
  • Distributions of head accessories are only calculated among Asians
  • More than 50% of subjects with accessories wear hats with or without a brim
  • Less than 10% of subjects choose headband or head veil.

Hair Style

  • 7 styles for wigs, 2 with men’s style, and 5 with women’s style
  • We randomly sampled 8 to 12 wigs for captured subjects, wig styles are not specified for gender.
  • Six colors are not evenly distributed among each wig.
  • subjects captured with black and brown are the majority
  • yellow color has the least portion

Word Frequency Chinese

We calculated word frequency for Chinese. From the cloud visualization, word frequency is indicated by the size of each character. The most frequent word ``Hai Pa'' in Chinese appears nearly 450 times among all sentences, while the least frequent one ``Ji Jiu'' is less than 50. We only summarize the phrased in Chinese, but not single characters like ``de'', ``shi'', ``wo'' and etc., since there have not specific implications.

Word Frequency English

We calculated word frequency for English. The most frequent word, ``Drawing'', occurs more than 600 times, while the least frequent one ``Ambitious'' is close to 0.