Hi everyone,
We are currently exploring a native Apple Vision Pro (visionOS) client for a psychological screening tool that relies on Py-Feat.
Due to Apple's strict privacy restrictions on visionOS, third-party apps cannot access the raw inward/outward camera feeds, nor do they expose the underlying ARKit numerical blend shapes. Instead, when an app requests camera access, visionOS returns a 2D video feed of the user's Persona (a real-time, photorealistic 3D CGI avatar that mimics the user's facial muscles).
This means the video payload we would send to our backend for Py-Feat to analyze is a video of a synthetic avatar, rather than a real human face.
My questions for the community/maintainers:
Has anyone tested Py-Feat's AU or emotion extraction accuracy against video feeds of Apple Personas (or similar high-fidelity avatars like Meta Avatars)?
Do the Py-Feat ML models suffer significant degradation in accuracy due to the lack of natural skin texture, human lighting, or micro-wrinkles on the Persona?
Hi everyone,
We are currently exploring a native Apple Vision Pro (visionOS) client for a psychological screening tool that relies on Py-Feat.
Due to Apple's strict privacy restrictions on visionOS, third-party apps cannot access the raw inward/outward camera feeds, nor do they expose the underlying ARKit numerical blend shapes. Instead, when an app requests camera access, visionOS returns a 2D video feed of the user's Persona (a real-time, photorealistic 3D CGI avatar that mimics the user's facial muscles).
This means the video payload we would send to our backend for Py-Feat to analyze is a video of a synthetic avatar, rather than a real human face.
My questions for the community/maintainers:
Has anyone tested Py-Feat's AU or emotion extraction accuracy against video feeds of Apple Personas (or similar high-fidelity avatars like Meta Avatars)?
Do the Py-Feat ML models suffer significant degradation in accuracy due to the lack of natural skin texture, human lighting, or micro-wrinkles on the Persona?