I was testing the models and I noticed some distortions in my point clouds caused (probably) by the interpolation back to the image shape. It happens in sharp corners, where the interpolation sometimes places points between two (very distant) depth levels, instead of leaving it as a "sharp edge".
Mainly, I wanted to know why the authors decided on the bilinear interpolation, and if there's a better solution then just using a simpler interpolation.
My temporary fix was just to modify this line to get a sharper depth image on the original image resolution.
Hi @felipecadar, thank you for raising this! We agree that the 'nearest' interpolation should be a better practice. We will use this interpolation probably in our future versions.
Hi! Congrats on the work!
I was testing the models and I noticed some distortions in my point clouds caused (probably) by the interpolation back to the image shape. It happens in sharp corners, where the interpolation sometimes places points between two (very distant) depth levels, instead of leaving it as a "sharp edge".
Mainly, I wanted to know why the authors decided on the
bilinear
interpolation, and if there's a better solution then just using a simpler interpolation.My temporary fix was just to modify this line to get a sharper depth image on the original image resolution.
from
to