Abstract:

Neural networks are a set of algorithms whose operation is inspired by biological neurons, these networks have been developed to solve problems: control, recognition of shapes or words, decision, and memorization. In this work, we tried to make an implementation that combines the advantages of the compact point cloud representation but uses the traditional 2D ConvNet to learn the prior knowledge about the shapes. And by combining the 3 modules together, the convolution structure generator 2D and the merge and pseudo-rendering modules, we have obtained an end-to-end model that learns to generate a compact point cloud representation from a single 2D image, using only a convolution structure generator 2D. And at the end we got as final result: from a single RBG image → 3D point cloud

Publisher's Version