Researchers from Google AI and the University of Central Florida Released the Open-Source Virtual Avatar Library for Inclusion and Diversity (VALID)

The research team from Google AR & VR, in collaboration with the University of Central Florida, conducted a comprehensive study to validate a virtual avatar library called VALID, consisting of 210 fully rigged avatars representing seven diverse races. The selection of seven races was under the guidance of the US Census Bureau. They utilized data-driven facial averages and collaborated with volunteer representatives for each ethnicity to create 42 base avatars(7 races X 2 genders X 3 individuals). The study involved a global participant pool to obtain validated labels and metadata for the perceived race and gender of each avatar.

The validation process employed Principle Component Analysis (PCA) and K-means clustering to understand how participants perceived the avatars’ races. To ensure a diverse representation of perspectives by balancing participants by race and gender, a total of 132 participants from 33 different countries worldwide were chosen for the study.

Results revealed consistent recognition of Asian, Black, and White avatars by participants of various races. However, avatars representing American Indian and Alaska Native (AIAN), Hispanic, Middle Eastern, and North African (MENA), and Native Hawaiian and Pacific Islander (NHPI) races showed more ambiguity, with differences in perception based on participant race. An avatar is named after the race if the avatar was identified as its intended race by corresponding same-race participants.

In the discussion, the researchers highlighted the successful identification of Asian, Black, and White avatars with more than a 95% agreement rate across all participants, challenging the notion of lower accuracy of around 65-80% in identifying faces of races different from one’s own. They attributed this to perceptual expertise or familiarity with diverse racial groups, possibly influenced by global media representation.

Own-race bias effects were observed, with some avatars being correctly identified primarily by participants of the same race. For instance, Hispanic avatars received mixed ratings across participants but were more accurately perceived by Hispanic-only participants. The study emphasized the importance of considering participants’ race in virtual avatar research to ensure accurate representation.

Certain avatars were labeled ambiguous due to unclear identification, with factors like hairstyle influencing perception. The validation of Native Hawaiian and Pacific Islander avatars faced limitations, highlighting the challenges of representation and the need for broader recruitment efforts.

The research team discussed implications for virtual avatar applications, emphasizing the potential for in-group and out-group categorization leading to stereotyping and social judgments. They suggested incorporating regulations to improve interracial interactions in virtual reality.

As a contribution to the research community, the team provided open access to the VALID avatar library, offering diverse avatars suitable for various scenarios. The library includes avatars with 65 facial blend shapes for dynamic expressions and is compatible with popular game engines like Unity and Unreal. The researchers acknowledged limitations, such as the focus on young and fit adults. They outlined plans to expand diversity by introducing different regional categories, body types, ages, and genders in future updates.

In conclusion, the research team successfully created and validated a diverse virtual avatar library, challenging stereotypes and promoting inclusion. The study highlighted the impact of own-race bias on avatar perception and provided valuable insights for developing and applying virtual avatars in various fields. The open-access VALID library is positioned as a valuable resource for researchers and developers seeking diverse and inclusive avatars for their studies and applications.

Check out the Paper, GitHub, and Google Blog. All credit for this research goes to the researchers of this project. Also, don’t forget to join our 33k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

If you like our work, you will love our newsletter..