Inaccuracy when yaw is over around 80 degrees

[dl-maxwang]

The pose estimation seems to work fine in -45-45 degree, however it seems accuracy is keep going down when yaw is over 80 degrees, when yaw is 90 degrees, the estimated yaw is 0...

is this caused by inaccuracy of facial landmark?(I spotted a lot of misplaced landmark in my own test image sets)

[yinguobing]

I'm afraid so. The pose is derived from the landmarks. Extreme pose makes it very difficulty to obtain accuracy landmarks.

[pirate-lofy]

Hi, @dl-maxwang, I want to know how u can calculate yaw and pitch in this code. thanks

[dl-maxwang]

Hi, @dl-maxwang, I want to know how u can calculate yaw and pitch in this code. thanks

You can find these code in estimate_head_pose.py

which use cv2.solvePnP to calculate rotation vector and transaltion vector. the parameter is:

object_points = all model points # all model points is save in assets/model.txt
image_points = detected face points
camera_matrix = 
             [
               [img_size, 0, camera_center[0]],
               [0, img_size, camera_center[1]],
               [0, 0, 1]
             ]
distCoeffs = np.zeros((4, 1)) #assume no distortion

[pirate-lofy]

sorry, but I'm not that good at these stuff, can u explain more