Expressions in Virtual Reality

Recently Google Machine Perception researchers, in collaboration with Daydream Labs and YouTube Spaces, presented a solution for virtual headset ‘removal’ for mixed reality in order to create a more rich and engaging VR experience. While that work could infer eye-gaze directions and blinks, enabled by a headset modified with eye-tracking technology, a richer set of facial expressions — which are key to understanding a person's experience in VR, as well as conveying important social engagement cues — were missing.

Today we present an approach to infer select facial action units and expressions entirely by analyzing a small part of the face while the user is engaged in a virtual reality experience. Specifically, we show that images of the user’s eyes captured from an infrared (IR) gaze-tracking camera within a VR headset are sufficient to infer at least a subset of facial expressions without the use of any external cameras or additional sensors.

Left: A user wearing a VR HMD modified with eye-tracking used for expression classification (Note that no external camera is used in our method; this is just for visualization). Right: inferred expression from eye images using our model. A video demonstrating the work can be seen here.

We use supervised deep learning to classify facial expressions from images of the eyes and surrounding areas, which typically contain the iris, sclera, eyelids and may include parts of the eyebrows and top of cheeks. Obtaining large scale annotated data from such novel sensors is a challenging task, hence we collected training data by capturing 46 subjects while performing a set of facial expressions.

To perform expression classification, we fine-tuned a variant of the widespread Inception architecture with TensorFlow using weights from a model trained to convergence on Imagenet. We attempted to partially remove variance due to differences in participant appearance (i.e., individual differences that do not depend on expression), inspired by the standard practice of mean image subtraction. Since this variance removal occurs within-subject, it is effectively personalization. Further details, along with examples of eye-images, and results are presented in our accompanying paper.

Results and Extensions
We demonstrate that the information required to classify a variety of facial expressions is reliably present in IR eye images captured by a commercial HMD sensor, and that this information can be decoded using a CNN-based method, even though classifying facial expressions from eye-images alone is non-trivial even for humans. Our model inference can be performed in real-time, and we show this can be used to generate expressive avatars in real-time, which can function as an expressive surrogate for users engaged in VR. This interaction mechanism also yields a more intuitive interface for sharing expression in VR as opposed to gestures or keyboard inputs.

The ability to capture a user’s facial expressions using existing eye-tracking cameras enables a fully mobile solution to facial performance capture in VR, without additional external cameras. This technology extends beyond animating cartoon avatars; it could be used to provide a richer headset removal experience, enhancing communication and social interaction in VR by transmitting far more authentic and emotionally textured information.

Acknowledgements
The research described in this post was performed by Steven Hickson (as an intern), Nick Dufour, Avneesh Sud, Vivek Kwatra and Irfan Essa. We also thank Hayes Raffle and Alex Wong from Daydream, and Chris Bregler, Sergey Ioffe and authors of TF-Slim from Google Research for their guidance and suggestions.

This technology, along with headset removal, will be demonstrated at Siggraph 2017 Emerging Technologies.