search

AaronJackson / vrn

:man: Code for "Large Pose 3D Face Reconstruction from a Single Image via Direct Volumetric CNN Regression"
http://aaronsplace.co.uk/papers/jackson2017recon/
MIT License
4.52k stars 747 forks source link

What does the NME in you paper stand for? #130

Closed crazySyaoran closed 4 years ago

AaronJackson commented 4 years ago

It is explained under the section "Error Metric" - but it's Normalised Mean Error.

crazySyaoran commented 4 years ago

Yes I'm sorry I found it and closed the issue. But I didn't find how do you change the hard (binary) result into a soft (real) predictions. Did you give some detail in your paper? Thanks

AaronJackson commented 4 years ago

During training, binary predictions are used (via Sigmoid activation function). Since it is Sigmoid, the values will not always be 1 or 0, they will be close to either of these numbers. If you have a surface extraction function which can uses non binary values, it can be used to improve detail (on better trained models) and result in a less blocky model. MATLAB's isosurface function supports these non binary values and hence the MATLAB version in this repo works better than the Python version.

crazySyaoran commented 4 years ago

Cool. Thank you.