erkil1452/gaze360

test error

Closed this issue · 12 comments

Using the author's code,the mean angle error of the model is 12.8 degrees.There is a certain gap with the results in the essay.What could be the reason?

Hmm .. I have reproduced the authors code many times and I can confirm that I can reproduce 13.5.
I recommend first downloading the weights provided by the authors and confirm if you can reproduce 13.5. If not, then it indicates some data files might be missing.

Thank you very much.Inspired by you,I reconfirmed and the result is 13.5.

image
Why are pitch and yaw defined in this way?

The tanh should not be interpreted geometrically. It is a common nonlinearity used in machine learning and we use it to squeeze the output to the range (-1,1). This is then easy to rescale to the range required for yaw (-pi, pi) and pitch (-pi/2, pi/2) angles as seen above.

The output is yaw and pitch in polar coordinates which can be converted into 3D direction using

x = sin(yaw) * cos(pitch)
y = sin(pitch)
z = cos(yaw) * cos(pitch)

Does this 3D direction refer to the coordinates (x, y, z) of the vector in the Cartesian coordinate system?

Yes, it is a 3D direction in the Cartesian coordinate system of the camera.

The gaze vector components (x, y and z).