Notes for An Invitation to 3-D Vision.

A geometric model of image formation

Goal: project points in 3-D space onto images in a 2-D image plane.

Process:

  1. coordinate transformations between the camera frame and the world frame
  2. projection of 3-D coordinates onto 2-D image coordinates
  3. coordinate transformations between possible choices of image coordinate frame

Inversed chain - camera calibration

An ideal perspective camera

Recall the coordinates $X=[X,Y,Z]^T$ of the same point $p$ relative to the camera frame are given by a rigid-body transformation $g=(R,T)$ of $X_0$:

Adopting the frontal pinhole camera model, we see that the point $X$ is projected onto the image plane at point

$Z$: the depth of the point $p$.

In homogeneous coordinates, this relationship can be written as

which is equivalent as

The coordinate $Z$ (or the depth of the point $p$) is usually unknown, so we simply write it as an arbitrary positive scalar $\lambda \in \Bbb R_+$.

Define two matrices

$\Pi_0$: standard (or canonical) projection matrix.

The overall geometric model for an ideal camera can be describe as

or in matrix form,

Camera with intrinsic parameters

Goal: specify the relationship between the retinal plane coordinate frame and the pixel array.

Actual image coordinates

$(o_x,o_y)$: coordinates (in pixel) of the principal point (where the z-axis intersects the image plane) relative to the image reference frame.

$s_x, s_y$: scale factors. When $s_x=s_y$, each pixel is square.

$s_\theta$: skew factor, proportional to $cot(\theta)$, where $\theta$ is the angle between the image axes $x_s,y_s$.

The transformation matrix

An intrinsic parameter matrix / calibration matrix refers to

3DV

Representation of a Three-Dimensional Moving Scene «
Prev «
» Image Primitives and Correspondence
» Next