about the metrics on scannet
chensjtu opened this issue · comments
hey! I just found that in the latest paper, there don't exist any quantitative results of Neural RGB-D surface recon on SCANNET dataset. Can you share the results so that I won't need to re-eval. Many thanks for your wonderful work!
The quantitative results don't exist because there is no ground truth data on which to evaluate.
Since MonoSDF and manhattan SDF both use SCANNET's mesh as ground truth to cal metrics, maybe you can also use provided mesh to cal the recon metric?
This wouldn't make sense because the proposed methods fixes some issues of the Scannet meshes (like misaligned walls), so the metrics might end up being lower in spite of a better reconstruction.