Colorization of output

[rauldds]

Autoencoder decoder

• Input: Marching Cubes mesh generated from TSDF (output of GRNET or even already generated target meshes). We apply it the texture of incomplete color input mesh. Shape of Input (Batch Size, N Vertices, XYZ+UVW). Where XYZ+UVW are just coordinates and color texture (6 values).
• Target: Marching Cubes mesh generated from TSDF (output of GRNET or even already generated target meshes). We apply it the texture of the target color mesh. Shape of Target (Batch Size, N Vertices, XYZ+UVW)

Why do the thing above? (to keep the number of vertices constant, otherwise, we have different numbers of vertices and faces).

[rauldds]

Tested approaches were:

1. Test 1

• Input [Batch, XYZ+RGB, Vertices]
• Output [Batch, XYZ+RGB, Vertices]

2. Test 2

• Input [Batch, RGB, Vertices]
• Output [Batch, RGB, Vertices]

Maybe would be nice to test (should be modified in the dataset class):

• Input [Batch, XYZ+RGB, Vertices]
• Output [Batch, RGB, Vertices]

[rauldds]

Also could try different losses (MSEloss, L2 loss, smoothL1 loss)