Closed chenhsuanlin closed 4 years ago
Hi @chenhsuanlin , thanks for your interest in the project!
I hope I understood you correctly that you are just interested in the final result tables? If so, you can find the tables you asked for on page 10 of our supplementary material.
Regarding the Chamfer distance, you are right that we use the Euclidean distance for the distance between two points, and then add accuracy and completeness without squaring them. the L1 / L2 is rather referring to the "loss" at the end, so if you take the "l2 loss" (x^2) or the "l1 loss" (abs(x)), and we are taking the l1 loss.
I hope this helps - good luck with your research!
Yes this is what I was looking for, I didn't realize it was in the supp material. Thanks!
Hi @m-niemeyer, I have a quick follow-up question: are the reported chamfer distance numbers scaled? As I understand, the shapes are normalized to fit unit cubes (between [-0.5,0.5] in xyz), so errors in the order of 0.2 or 0.3 would be extremely large. (I'm guessing it's scaled by 10?) Thanks!
Hi @chenhsuanlin , yes, you are right - they multiplied by a factor of 10!
We follow previous works here to use "use 1/10 times the maximal edge length of the current object’s bounding box as unit 1" (see Section 4 - Metrics).
Thanks for the clarification!
Hi! I'd like to ask if it'd be possible to provide the Chamfer distance results in Table 1 (2D supervision section) in the uncombined form, i.e. as the accuracy and completeness metrics separately (described in the Occupancy Networks paper)?
Also, a side question, I'm curious why it's called Chamfer-L1, as to my understanding it's not really measured as an L1 norm but rather an unsquared L2 norm.
Thanks in advance!