How to get the 2D location of each vertex in each render? about pytorch3d HOT 8 CLOSED

AH! I think the problem was related with the issue #78
Doing

ndc_points [:,1] *= -1

solved

Thanks for the help! And congratulations for this amazing work.

from pytorch3d.

@filipetrocadoferreira Do you mean for each pixel in the rendered image you want to know the position of the vertices?

from pytorch3d.

Basically, is applying the same transformations as the render to the 3D vertex and get their 2D location in the image.

from pytorch3d.

using cameras.transform_points() i get the points in 'screen' coordinate system, but to go from this to 2D pixel coordinates I'm having a bit of a trouble

from pytorch3d.

By 2D pixel coordinates do you mean NDC coordinates in the range [-1, 1] or coordinates in the range x = [0, W], y = [0, H] where the image is of size (H, W)?

If you have a code snippet you can provide of what you are trying to achieve that would also be helpful.

from pytorch3d.

range x = [0, W], y = [0, H] where the image is of size (H, W)

This one.
I'm having trouble to go from NDC to image coordinates.

from pytorch3d.

Ok so the output of cameras.transform_points() will be in NDC coordinates in the range [-1, 1]. If i understand correctly, from this you want to go to pixel coordinates in the range y: [0, H] and x: [0, W].

To do this you just need to do the inverse of this PixToNdc function:

i.e. In C++

float NdcToPix(int i, int S) {
  // where S is the size of the dimension e.g. H or W
  return ((i + 1.0f) * S - 1.0f)/2.0;
}

from pytorch3d.

@filipetrocadoferreira oh ok makes sense that the rendering problem was affecting your calculations! I will close this issue - please monitor #78 for updates on the rendering fix.

from pytorch3d.