Skip to content

Latest commit

 

History

History
53 lines (36 loc) · 2.19 KB

coordinate_conventions.md

File metadata and controls

53 lines (36 loc) · 2.19 KB

Coordinate Conventions in pytorch-dense-correspondence

Keeping coordinate systems consistent is not a spontaneous event.

There are a few reasons this is not always straightforward:

  • Typical computer vision "x,y,z" convention in 3D space is "right-down-forward"
  • Associated with this right-down-forward convention, it's easiest to think about pixel positions as being associated with this right, down convention. This gives us the "(u, v)" convention for defining pixel coordinates.
  • On the other hand, images that are arrays/tensors in numpy or torch are accessed as row, column, in that order. That isn't consisent with the (u, v) convention
  • Also on the other hand, typical robot convention is to think of "x" as being forward, away from the robot, and "y" to the left. This gives "forward-left-up" for "x,y,z"

Conventions

In this repo we would like to follow this simple convention for cartesian 3D frames:

  • All cartesian frames are specified in right-down-forward coordinates.

But we currently have a more hybrid / mixed approach with accessing pixel data:

  • When doing 3D vision, we use the OpenCV "(u,v)" convention for pixel coordinates. This aligns with our cartesian coordinate system as follows in this diagram:

  • All data stored as numpy arrays or torch tensors should be stored in row, column format. This means that in order to access the correct pixel in this data, if using (u,v) coordinates, you should access the data as:

    pixel_at_u_v = data_tensor[v,u]
  • To convert between (u,v) and single index the formulas are n = u + image_width * v. Similarly (u,v) = (n % image_width, n / image_width).

Frames

This screenshot of rviz shows the primary frames in use:

  • The frame at the bottom of the robot arm (it says "base" but is hidden) is considered "world" frame.
  • The RGB optical frame
  • The depth optical frame * (but note that all data by the time it hits this repo is instead in the RGB optical frame after registering the depth image)

Note that in the rviz screenshot, red=x, green=y, blue=z.