tranluan / Nonlinear_Face_3DMM

Source code for "Nonlinear 3D Face Morphable Model"
http://cvlab.cse.msu.edu/project-nonlinear-3dmm.html
Apache License 2.0
672 stars 123 forks source link

After training your model, the Rendered Image has no texture on it! #56

Open HOMGH opened 3 years ago

HOMGH commented 3 years ago

Hi @tranluan I could train your model by running model_non_linear_3DMM.py file. The rendered output has the correct shape and projection matrix, but it doesn't have any texture on it. The rendered output is same as follows:

rendered_image_05_0012 rendered_image_01_0011

I checked the warp_texture function and it is working, but it seems its input texture is not correct! That's why the output doesn't have any texture! The function used for texture generation is : generator_decoder_albedo(tex_fx_300W, is_reuse=False, is_training=True). Is it the right function you used as your albedo decoder?

chaoshiedwin commented 3 years ago

I believe it is because 3DMM_tri_2d_barycoord.dat file is not correct, check out this issue: https://github.com/tranluan/Nonlinear_Face_3DMM/issues/27

cassiePython commented 3 years ago

Have you successfully corrected this issue? @chaoshiedwin @HOMGH

HOMGH commented 3 years ago

Have you successfully corrected this issue? @chaoshiedwin @HOMGH

Yes, I just run the ....proxy.py file and it works. But still does not converge! (but at least results in something!)

chaoshiedwin commented 3 years ago

Have you successfully corrected this issue? @chaoshiedwin @HOMGH

use 3DMM_tri_2d_barycoord_reduce.dat instead of 3DMM_tri_2d_barycoord.dat works for me.

HOMGH commented 3 years ago

Have you successfully corrected this issue? @chaoshiedwin @HOMGH

use 3DMM_tri_2d_barycoord_reduce.dat instead of 3DMM_tri_2d_barycoord.dat works for me.

Hi @chaoshiedwin , I actually gave up with that code and tried to run the model_non_linear_3DMM_proxy.py. It is on 11th epoch on 124k images but I think the results are not good . (attached images). Could you train the model? image image

huyanfei-cqupt commented 3 years ago

use 3DMM_tri_2d_barycoord_reduce.dat instead of 3DMM_tri_2d_barycoord.dat works for me.

Hi,@chaoshiedwin Did you get through this code? Has the L0 loss converged in the pre-training phase?

huyanfei-cqupt commented 3 years ago

Have you successfully corrected this issue? @chaoshiedwin @HOMGH

use 3DMM_tri_2d_barycoord_reduce.dat instead of 3DMM_tri_2d_barycoord.dat works for me.

Hi,@chaoshiedwin, can you run the model_non_linear_3DMM.py successfully? the G_loss_frecon how to write it? any advices would be highly appreciated.