Inside the Supplementary Material, the definition of Chamfer -L1 has a term: the surface area of the mesh.
But in the evaluation code, the surface area is implemented as the summation of the points number.
Atals/Pix2mesh/Onet use 100k sampled points from predicted mesh and 100k sampled gt points to calculate the Chamfer-L1 loss.
But how about PSGN? PSGN only predict 1024 points, is the metric reported in the paper are calculated on 1024 points instead of 100k points? Is this a fair comparison?
Really confused. Hope someone can give me some insight? Thank you very much.
Inside the Supplementary Material, the definition of Chamfer -L1 has a term: the surface area of the mesh. But in the evaluation code, the surface area is implemented as the summation of the points number. Atals/Pix2mesh/Onet use 100k sampled points from predicted mesh and 100k sampled gt points to calculate the Chamfer-L1 loss. But how about PSGN? PSGN only predict 1024 points, is the metric reported in the paper are calculated on 1024 points instead of 100k points? Is this a fair comparison?
Really confused. Hope someone can give me some insight? Thank you very much.