Axes convention about neo-360 HOT 3 CLOSED

Our axis convention is unique and unfortunately doesn't share the classical opencv/openGL conventions. You could however utilize the following helper function to convert it to NeRF i.e. OpenGL format:

def convert_pose_PD_to_NeRF(C2W):

    flip_axes = np.array([[1,0,0,0],
                         [0,0,-1,0],
                         [0,1,0,0],
                         [0,0,0,1]])
    C2W = np.matmul(C2W, flip_axes)
    return C2W

Optionally before converting it to OpenGL, you could consider canonicalizing the poses so they are centered at origin
c2w[:3, 3] = c2w[:3, 3] - obj_location

Thanks for your patience and we will release all of these details with our codebase soon!

from neo-360.

Correct, it is the center of the hemisphere used to generate camera poses. averaging the x/y axes values should also work the same way. For overfitting it probably shouldn't matter as much but for generalizability this should remove any biases in camera locations in world coordinates.

Other than this, it is a good practice to do pca on the poses which should give you similar results.

from neo-360.

Ok I see, that's what I suspected. Is there a reason for this choice?

Optionally before converting it to OpenGL, you could consider canonicalizing the poses so they are centered at origin

What is obj_location, I saw it in the pose.json file, the center of the hemisphere used to generate the camera views ?

I have been setting the world origin as the average of the cameras position on the x and y axes and keeping the z-axis value as 0.

Thank you for your availability :)

from neo-360.