hiroharu-kato / view_prior_learning Goto Github PK

Hi, could you shed light on the usage of obj_scale, tanh_scale, scaling variables in FCShapeDecoder respectively?

• Why you multiply the intermediate value with the above scaling factor to get the final vertex coordinates? Is that because to reduce the destruction of noise?
• How to choose the value of those variables? Is that empirically found useful?
  

  view_prior_learning/mesh_reconstruction/decoders.py

  Lines 382 to 390 in 6f1afa8

  	class FCShapeDecoder(chainer.Chain):
  	def __init__(
  	self, dim_in=512, scaling=1.0, use_bn=False, fc_connection=False, symmetric=False):
  	super(FCShapeDecoder, self).__init__()
  	self.grid_size = 16
  	self.obj_scale = 0.5
  	self.tanh_scale = 1.2
  	self.scaling = scaling

I also notice that the different method is used to reconstruct 3D mesh in your another repo mesh_reconstruction. What makes you choose different reconstruction algorithms?

Thank you in advance!

Hi, Could you please share the requirements.txt of this fantastic project? And how to use the neural renderer, is it the same as instructed in neural mesh render?

Thank you in advance!

Hi authors,

Would like to know if you can provide a link to download the preprocessed dataset? Thanks!

https://github.com/hiroharu-kato/view_prior_learning/blob/master/mesh_reconstruction/dataset_shapenet.py#L51

Hi @hiroharu-kato

Really enjoyed reading the VPL paper, thanks for making the code public!
I was trying to port the code to python3 and CUDA. I am a bit doubtful about how the voxelize_sub1 should be done. For example, it is not clear to me that how on ine 14 in voxelization.py, i is used, whereas input to the function is j. Also, when it is written voxel=1 like here does it mean that i'th element of voxel array is set to 1?

Any help in this regard is highly appreciated!

Hi @hiroharu-kato Thanks for the code. I have stuck up with some questions and it expecting answers. I have understood the reconstructor part which generate mesh using encoder-decoder architecture.
1.) For single view training, does discriminator takes only one viewpoint (out of 20) as correct views and remaining 19 as incorrect?
2.) What will be partitions of dataset for single view training of discriminator?
3.) What does real views mean In view-based training, reconstructed views from
observed viewpoints converge to the real views in a training
dataset by minimizing the reconstruction loss (single view)?