Closed tri-bao closed 1 year ago
@gkarpiak do you have any advice?
Meanwhile, I tried both and the accuracy of lips&eyes landmarks produce by the landmark_with_attention
model is less than the combination of the model face_landmark
(without attention) and iris_landmark
. By "accuracy", I mean the sensitive of landmarks position to the movement of lips and eyes.
I thought that landmark_with_attention
alone will perform faster than face_landmark
+ iris_landmark
with the trade off of accuracy and I used it. Everything works fine in 4 different Android phone models until I tried it on Samsung A12. On that phone, the predicted landmarks fluctuate a lot between frames. If we draw a facial bounding box, we can clearly see the fluctuation. If we calculate the overlap of the facial bounding box in 2 consecutive frames, they don't overlap each other although the face and the phone (hence the camera) stand still.
Samsung A12 phone has:
MediaTek Helio P35
IMG PowerVR GE8320
For more information, Android phone models that the landmark_with_attention
works well are:
Hello @tri-bao,
Are you still looking for resolution on this Or the issue has been resolved? Thank you!!
hello @kuaashish
It hasn't been resolved but I give up looking for the resolution. I don't use face_landmark_with_attention
instead I use face_landmark
+ iris_landmark
In the
iris_tracking
graph,face_landmark.tflite
is firstly used to have 468 facial landmarks and theniris_landmark.tflite
is used to produce 10 more landmarks for irises. However, I notice thatface_landmark_with_attention.tflite
already produces 478 facial landmarks including iris ones. Are they same? In another word, can I just modify theiris_tracking
graph to only useface_landmark_with_attention.tflite
and still have the same result with the same accuracy?Thanks