search

emilianavt / OpenSeeFace

Robust realtime face and facial landmark tracking on CPU with Unity integration
BSD 2-Clause "Simplified" License
1.46k stars 152 forks source link

landmark losss problem #24

Closed yutaka329 closed 3 years ago

yutaka329 commented 3 years ago

nice work, I dont comprehend heatmap offset format. such as x_offset= 223 * log(p/(1-p)) / 16. can you provide some explanation? thank you.

emilianavt commented 3 years ago

Hi!

Offset values are in the range [0, 1), normalized with respect to the full resolution of the image. However, during training, the model was not able to learn offsets very well, because the differences from pixel to pixel were very slight. To make the differences stronger and offsets easier to learn, when creating the offset maps for training, I multiply offsets by 16 and take the sigmoid to bring them back into the range of [0, 1). log(p/(1-p)) / 16 is the inverse of multiplying by 16 and taking the sigmoid.

I hope that helps a bit. If you have any further questions, please let me know.

yutaka329 commented 3 years ago

got it , thank you a lot for your reply