Combine SAN with SBR

[d-x-y]

Combine SAN with SBR

[D-X-Y]

Not yet. SAN costs maybe 20 hours to train a small network (CPM) and SBR is much slower than SAN due to the implementation of differentiable registration layer. Their combination will further slow down the training procedure, and thus we did not try this.

[keyxuliang]

I have a facial landmark detector network by myself. I want to combine the network with SBR to improve it's performance. As the SBR paper mentioned: you have tried regression network+SBR. Would you mind sharing a simple demo for that?
Thanks!

[D-X-Y]

I will discuss with my mentor about this, and let you know ASAP.

[keyxuliang]

Thank you very much.
What's the meaning of 'ASAP'?

[D-X-Y]

ASAP = as soon as possible

[keyxuliang]

Thanks! Looking forward to your sharing.

[D-X-Y]

Hi, @keyxuliang , I have discussed with my mentor. We can not release the regression model at this moment because our extension version is still under revision. If you want to use a regression-based detector, the simplest way is to use MobileNet backbone (https://github.com/tonylins/pytorch-mobilenet-v2/blob/master/MobileNetV2.py), and replace its classifier by a (number-of-points*2)-dimensional FC layer.

[keyxuliang]

Ok，I will try it.

[D-X-Y]

Please feel free to re-open it if need.