search

ZhenglinZhou / STAR

[CVPR 2023] STAR Loss: Reducing Semantic Ambiguity in Facial Landmark Detection
157 stars 17 forks source link

Not properly working in case of occlusion and tilted head pose. #26

Closed Jourtics closed 4 months ago

Jourtics commented 6 months ago

Hey @ZhenglinZhou The STAR model gives extrapolated landmarks in case of occlusion and tilted head pose like this-- Star_occlusion_3 Is the model not taking care of the above scenarios? Ideally, the landmarks along the lip and facial contour should not be detected. Most of the facial landmark detection models have this issue.

TimSC commented 6 months ago

Most shape predictors just do their best in the case of occlusion. You are right, most methods don't tell you if a point is occluded.

ZhenglinZhou commented 5 months ago

Hi @Jourtics, thanks for your interest!

You are right, the model is encouraged to output landmarks with structural constraint, which is seen as robustness in occlusion. To detect the visibility of landmarks, you could consider LUVLi.

Also thanks @TimSC for the kind reply.

If you have other questions, feel free to leave a comment.