Confusion about yaw and pitch.

[shaform]

In your model, the forward function returns pre_yaw_gaze and pre_pitch_gaze:

L2CS-Net/l2cs/model.py

Line 70 in a4d8f7f

return pre_yaw_gaze, pre_pitch_gaze

However, in the pipeline, the two variables are assigned as gaze_pitch and gaze_yaw:

L2CS-Net/l2cs/pipeline.py

Line 122 in a4d8f7f

gaze_pitch, gaze_yaw = self.model(img)

It seems yaw and pitch are reversed. Why would this be the case?

[tiamo405]

Have you found the reason why? If left as is, can it be trained?

[tiamo405]

line 207 train.py: pitch, yaw = model(images_gaze)
line 22,23:
self.fc_yaw_gaze = nn.Linear(512 * block.expansion, num_bins)
self.fc_pitch_gaze = nn.Linear(512 * block.expansion, num_bins)
line 68 69 70:
pre_yaw_gaze = self.fc_yaw_gaze(x)
pre_pitch_gaze = self.fc_pitch_gaze(x)
return pre_yaw_gaze, pre_pitch_gaze
Because both have the same calculation structure, I think changing the return part is fine, the result when training will not affect anything.

[shaform]

It turns out the pipeline was not written by the authors, but it was created in #18. So perhaps the PR author would know the answer.

[Reblexis]

I'm pretty certain that it's flipped accidently here and then also in the draw_gaze function (vis.py). So the correct order should begaze_yaw, gaze_pitch = self.model(img)and then in the draw_gaze function flip pitchyaw[0] and pitchyaw[1]uses.