search

google-ai-edge / mediapipe

Cross-platform, customizable ML solutions for live and streaming media.
https://ai.google.dev/edge/mediapipe
Apache License 2.0
27.39k stars 5.15k forks source link

eyeWidenLeft/eyeWidenRight/noseSneerLeft/noseSneerRight blendshapes are always near 0 #4450

Open TigerHix opened 1 year ago

TigerHix commented 1 year ago

Have I written custom code (as opposed to using a stock example script provided in MediaPipe)

None

OS Platform and Distribution

Python 3.9 (Windows)

Mobile device if the issue happens on mobile device

No response

Browser and version if the issue happens on browser

No response

Programming Language and version

Python

MediaPipe version

No response

Bazel version

No response

Solution

Face Landmarks

Android Studio, NDK, SDK versions (if issue is related to building in Android environment)

No response

Xcode & Tulsi version (if issue is related to building for iOS)

No response

Describe the actual behavior

Similar to #4436, I was not able to make any of those blendshapes value greater than 0.05.

Describe the expected behaviour

These blendshapes should have values greater than 0.05 when I sneer my nose / raise my eyebrows.

Standalone code/steps you may have used to try to get what you need

Official Python example code can reproduce this issue.

Other info / Complete Logs

This could be a problem with the diversity of the model's training data (I am an Asian male).
Shrimperator commented 8 months ago

Unfortunately this is still a major issue that noticeably decreases the quality of blendshapes face landmarker returns. The landmarks are great, but a large amount of blendshapes are inaccurate or barely go above 0 :/