search

anhttran / 3dmm_cnn

Regressing Robust and Discriminative 3D Morphable Models with a very Deep Neural Network
Other
857 stars 254 forks source link

Could be 3d mode trimed as float type，it will be smaller。and Could DNN model more smaller #8

Closed sunjunlishi closed 7 years ago

sunjunlishi commented 7 years ago

Could be 3d mode trimed as float type，it will be smaller。and Could DNN model more smaller 。 it will more smart using

TalHassner commented 7 years ago

@sunjunlishi we really don't understand what you mean. In the future, please try to be clearer with your questions (contributions?) and we will be happy to try and help.

sunjunlishi commented 7 years ago

firstly,BaselFaceModel_mod.mat is 116M,it's data has double type,change to float ,could be more smaller secondly: 3dmm_cnn_resnet_101.caffemodel is 172M,could be more smaller,to 10M? so that,so that many many android app using based on your method

sunjunlishi commented 7 years ago

I have a note, multiple photos from different angles, using your method to calculate the results, weighted addition, is not the final result will be better?. And if the texture is added, is the effect better?. Your work is great

TalHassner commented 7 years ago

@sunjunlishi RE smaller size files, you are correct. BFM, however, is not ours to distribute. Reducing the size of a deep network is something we are considering, but I think it may require more than just using smaller variable types to to this right without changing the quality of the output.

RE several photos from different angles, yes, in our experience that seemed to help a lot. Also using texture improved the results too.

sunjunlishi commented 7 years ago

could The ave pooling replaces the full join layer ,to that reduce mode's size