Question on comparison on Table 1 in the paper.

[Youngju-Na]

Hi! Thanks for the insightful and fantastic work. I really enjoyed the paper :)

I have a question about the comparison in Table 1. As I can see that there are Novel view synthesis models such as NeRF and PointNeRF, Is the volume density field changed to the SDF field during the test as illustrated in the Preliminary section (3)?

I was interested in the results of PointNeRF because you only tested with neural points but still shows good results in some scenes. Thanks again, fantastic work.

[wutong16]

Hi! For Novel view synthesis models (e.g., DVGO), we apply a threshold on the density field to extract the mesh, as described in the appendix Sec. B.3-Threshold selection.

For PointNeRF, we tried two evaluation modes:

• The first mode is to directly calculate the CD between the neural points and the ground truth mesh (the common evaluation code will sample a dense point cloud from it), and its results are indeed quite good in some scenes (e.g., scene 83, 97, and, 110). Based on my observation, 97 and 110 are typical cases with obvious specular reflection, which cause a flaw in mesh for NeuS, VolSDF, and our method due to the design of conditioning the color on geometry cues. As for 83, I'm not sure the specific reason why, but I suspect that the shaggy beard of the doll in scene 83 may not be friendly to surface-based methods.
• The second mode is to convert the neural point to a water-tight mesh for evaluation, as described in appendix Sec. G, while we find the results become worse due to the error accumulation in the point2mesh process, which is hardly avoidable.