Avatar SDK generates recognizable avatars from selfies. Our ultimate goal is for you to recognize your friends’ avatars within a 3D experience. Traditionally, we’ve assessed the likeness of our avatars using focus groups. Yet, these focus groups are costly, and we are limited to testing only major releases on them. Recently, an external event piqued our interest in employing facial recognition technology. We’ll discuss this event later, but for now, let’s delve into how we’re utilizing a facial recognition system to test avatar likeness.

Typically, a facial recognition system calculates a ‘distance’ between two faces. The smaller this distance, the greater the likeness between these faces. We generate avatars for each image in our test dataset and measure the likeness to the original image. This provides us insight into how recognizable the avatars are. This method enables us to test every minor release, note subtle changes in likeness, and further optimize it.

Additionally, upon adopting the facial recognition system, we discovered that our avatars already surpass some impressively stringent tests. For instance, we measure distances between each avatar and all test images, selecting the closest image. Intriguingly, for each avatar, the closest image happens to be the very image used to create that avatar! This result is consistent for 100% of the images in our test dataset, which encompasses all ethnicities and ages!

For more information, please read the white paper https://docs.avatarsdk.com/pub/White_paper_-_Avatars_you_can_recognize.pdf. Test our new MetaPerson avatars at https://metaperson.avatarsdk.com!