Isabella98Liu
commented
1 month ago Hi, thanks for your question! Yes, we need to query the MLP to obtain the color of the mesh. The color MLP takes the location in the canonical space and outputs the color. In our paper, we deformed the mesh back to the canonical space and queried its vertices' color. However, you can also obtain a canonical mesh and use the forward deformation network. This way, you only need to query the color MLP once, but performance may drop, and it cannot handle topology changes during deformation.
When comparing with other methods, we extract their mesh from the density field and obtain their vertices' color from the radiance of the surface points. We also tried volume rendering (which involves massive querying of the point's radiance and integration along the ray) to obtain the vertex color in our baselines, but the results were not better than ours. We believe this is because other methods do not provide accurate geometry, leading to inaccurate radiance querying during rendering.
Thanks for your interesting work and open-source code. I noticed that the extracted mesh has no color. When rendering the image, I need to query the MLP to get the color of the mesh. I have two questions. 1. Do I need to query the color through MLP every time I want to use this mesh? 2. The paper compares the rendered images of meshes extracted by other methods. Those methods do not use MLP to model the color of the mesh. Is it fair to compare the images rendered from their meshes?