search

optas / latent_3d_points

Auto-encoding & Generating 3D Point-Clouds.
Other
506 stars 109 forks source link

'approx_match' is not defined? #12

Closed VikVelev closed 5 years ago

VikVelev commented 5 years ago

Hello, First of all, great paper and great results. I have a few questions regarding the architecture and compiling. I managed to compile everything, using python2.7 and all the tested versions of the libraries (except CUDA, but the compilation was successful (i think?)so I ignored it), but as soon as I run the notebook I get stomped by this message: 

NameError: global name 'approx_match' is not defined

My second question is how does 1D Convolution with Stride 1 and Kernel 1 actually work in your case with point clouds? Isn't that just multiplication?

VikVelev commented 5 years ago

Fixed the first issue by installing CUDA 8.0, but my question about the 1D convolutions still stands :) thanks for your time.

optas commented 5 years ago

Hi,

I am glad you figured out the CUDA stuff.

1D convolution happens wrt. to each point. That is indeed a multiplication (dot-product) between the three coordinates (x,y,z) of each point and the learned kernel.

Best,