Replies: 4 comments
-
Standard FaceMesh is 468 points, then its up to post processing to
new FaceMesh-With-Attention is simmilar except it does that inside the model
Size wise, its double the size of normal FaceMesh, but still within reason But...MediaPipe releases only TFLite models and converting this is not going to be easy as multiple models are linked with a custom op And I just noticed there are now two variations of BlazeFace detector full variation as it has been Btw, there is also new FaceGeometry model which runs as calculator that takes mesh and performs Intention is more precise AR usage such as wearing fake glasses and what not |
Beta Was this translation helpful? Give feedback.
-
Just closing the loop...
Don't know why MediaPipe is going in that direction instead of contributing to TF, but as it is, it might take someone to reimplement the model from scratch (it is documented in sources) as an easier option than trying to convert it from their TFLite binary |
Beta Was this translation helpful? Give feedback.
-
Did some testings and it doesn't seem noticeably better than the current solution in terms of accuracy. A unified model has its advantage though, especially for the Mediapipe JS solutions like It will be interesting to see if the official TFJS side will do a conversion or not. |
Beta Was this translation helpful? Give feedback.
-
TFJS side stopped doing MediaPipe conversions last year, all of contributions since then are directly from MediaPipe team |
Beta Was this translation helpful? Give feedback.
-
https://github.com/google/mediapipe/releases/tag/v0.8.8
At the first glance, it seems they have (finally) integrated the iris/blink detection into the face model. However, the model description seems to suggest that it's actually a new model using a new approach. More accurate lips and general 3D coordinates perhaps?
https://google.github.io/mediapipe/solutions/face_mesh.html#attention-mesh-model
Beta Was this translation helpful? Give feedback.
All reactions