Trained nets

[marcinkeviciusp]

Hi, any chance we can get the trained networks / weights? Seems a bit pointless having everyone retrain it, especially since I saw an issue about the database no longer being available.

Thanks, Povilas

[christopher-beckham]

Hi Marcin,

Pre-trained checkpoints can be found at the link that is here: https://github.com/joelmoniz/DepthNets/tree/master/depthnet-pytorch#experiments

Unless you mean wanting a pretrained network that maps images of faces directly to keypoints. We had one that we used internally, but it was trained a very long time ago under the Theano framework (@SinaHonari would have more details about this). There should be some off-the-shelf image-to-keypoint regression networks lying around on the internet that you can use. While some may generate 68 keypoints instead of 66, you could roughly figure out a heuristic to 'convert' to 66 keypoints through the use of this reference image. (We actually had to do this at one point for the opposite direction: converting from 66 to 68 keypoints, as computed in this method: https://github.com/joelmoniz/DepthNets/blob/master/depthnet-pytorch/util.py#L137-L148)

Let me know if there are any other concerns.

Thanks,
Chris

[christopher-beckham]

This looks like a viable alternative to 3DFAW, and they also use 68 keypoints:

https://github.com/jiankangdeng/MenpoBenchmark