Closed sthalik closed 9 years ago
Yaw angle with big deviation from near-frontal can be handled by learning new detector, e.g. for full profile and then switching between detector's dependent on the yaw angle estimation.
Solution I chose is to ignore poses where the model and the flandmark detector's rotation differ by more than a threshold amount.
Critical yaw angle value is problematic to estimate as it depends on pitch too.
High face yaw causes false positive detections which don't correspond to what's in the image.
Trying to guard against it by guessing yaw and skipping flandmark_detect invocation but it's kludgy and doesn't work anyway.