Full head model

[samsgates]

how to create full head model for training dataset( same like BFM, but full head) output will be full head 3d model only mesh, no color or texture

[anhttran]

Hi, To use the full head, you can change trimBaselFace.py:

1. Do not use vertex_indices3, which is for mesh trimming
2. Change this line of code: faces = mod_struct.faces to

   faces = model['tl']
   faces[:,[1, 2]] = faces[:,[2, 1]]

Note that the landmark indices will be outdated. Though, we do not use these landmarks in this release.

[samsgates]

i didn't seen the below line in code

faces = mod_struct.faces

any way i have changed "trimBaselFace.py" code as your instruction. but full head model not generated in output, i got same result like half head model.

I have attached edited file as "trim.py", please update and send me back with changes required for generate full head

trim.py.zip

[dingtiandu]

Hi Your result 3d ALPHA only has 99 numbers.Can I get a ALPHA of 199 numbers, just like the old version of BFM dataset? Hoping for replying.

[anhttran]

@samsgates My bad. I looked into a wrong code. Please find the updated code for full head model below.

##MATLAB: BFM.shapeMU = model.shapeMU/1000;
shapeMU = model['shapeMU']/1000;
##MATLAB: BFM.shapeMU(1:3:end) = BFM.shapeMU(1:3:end) + mod_struct.tx;
shapeMU[0:-1:3] = shapeMU[0:-1:3] +  mod_struct.tx
##MATLAB: BFM.shapeEV = model.shapeEV(1:99)/1000;
shapeEV = model['shapeEV'][0:99]/1000.
##MATLAB BFM.shapePC = model.shapePC(:,1:99);
shapePC = model['shapePC'][:,0:99]
##MATLAB  BFM.texMU = model.texMU(:,:);
texMU = model['texMU']
##MATLAB  BFM.texPC = model.texPC(:,1:99);
texPC = model['texPC'][:,0:99];
##MATLAB  BFM.texEV = model.texEV(1:99);
texEV = model['texEV'][0:99];
innerLandmarkIndex = mod_struct.innerLandmarkIndex  # Outdated
outerLandmarkIndex = mod_struct.outerLandmarkIndex # Outdated

#% Faces
faces_tmp = model['tl']
#faces = faces_tmp,[2 1 3])
faces = np.dstack([faces_tmp[:,1], faces_tmp[:,0], faces_tmp[:,2]])
faces = faces[0]

@dingtiandu We only use 99 parameters for shape & texture since they are good enough for face recognition purpose according to Blanz's paper: http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=1227983

[samsgates]

Thanks for your response