Other models available other than 4-stacked?

[Yozey]

Hello Adrian,

Thanks for sharing your code in pytorch. I noticed that you have probably tested several models with less stacked FAN here.

However, only 4-stacked pretrained model is available to download from your website. I would like to know that if it is possible to share smaller models even though the performance might be a little bit worse.

Thank you very much in advance.

[1adrianb]

Hi @Yozey ,

Yes, I was in the process of converting/retraining the smaller models but most of them are not fully retrained. I will resume this at some point during the next week and add them here. Sorry for the delay.

[ggsonic]

Hi @1adrianb , will the 2dfan-2.pth.tar be available? i want to have a test. or will you please send me a 2-stacked 2dfan model to my gmail ggsonic@gmail.com. Thanks!

[korabelnikov]

Any changes in subject? Does not-stacked model available now?

[nightmaredimple]

I found that the face-alignment process cost 3.5s in 2D alignment and 2.0s in 2D-to-3D alignment.So the 3D alignment process cost almost 5.5s in total.(with Tesla K40c) However,in CPU mode, it cost 2.5s in 2D alignment and 1.4s in 2D-to-3D alignment. Is that normal? PS: if it is possible to share smaller models

[1adrianb]

Hi @nightmaredimple , The entire process is dominated by dlib face detector. In CPU mode dlib uses basically a V-J face detector while in cuda mode a cnn-based detector.The 2D/2.5D detector alone (ignoring the face detection) should be able to run at 20-30fps. A simple way to speedup the process is to use a faster face detector. Another method of speeding this up is using a batch>1.

[nightmaredimple]

@1adrianb Thanks for your reply！I‘ve done a simple test(detect_landmarks_in_image.py),the results as below: In GPU mode: face detection:0.46s 2D alignment:3.22s 2D-to-3D alignment:2.04s Total time:5.8s then the same detected face for the second alignment: 2D alignment:0.08s 2D-to-3D alignment:0.1s so make the alignment twice can speed up!

In CPU mode: face detection:0.15s 2D alignment:2.22s 2D-to-3D alignment:1.10s Total time:3.4s then the same detected face for the second alignment: 2D alignment:2.34s 2D-to-3D alignment:1.355s so the GPU can speed up the alignment process

emmm, maybe my GPU (Tesla K40c) can't make it real-time，but I still appreciate for your help.By the way,is there any more models can be shared except for 2DFAN-4 and 3DFAN-4?

[1adrianb]

Hi @nightmaredimple, that makes sense, the first pass is taking longer due to all initialisation and copying stuff from cpu to cuda. The 20-30 was achieved usign a Titan X Pascal, however a simple way to speed up is to use a 2 stack network (basically 2x improvment with almost no loss in performance), however this will require retraining the network for best performance. Another suggestion is to try to export the network to a static graph (via onnx) wich should hopefully fuse some operations and improve speed. I will probaly push this change at a later point.

Not at this time unfortuantelly.

[njho]

Hi @1adrianb I'd also be interested in seeing the 2 stack network. Any update? That being said, great work!

[moncio]

Hi @1adrianb , any kind of advance in this task? thank you so much

[moncio]

Sorry for being heavy @1adrianb again, but I am very interested to know about this topic if there is any progress. Thanks in advance.

[BigTom007]

Hi @1adrianb , I'm also very interested about 2 stack network, since 2 stack Hourglass was used in your other paper.