Decomposing the affine transformation matrix with the SVD

314 Views Asked by At

In the textbook I am studying Multiple View Geometry in Computer Vision, Second Edition. Richard Hartley, Andrew Zisserman. pg 40, it is stated that

The affine matrix $A$ can always be decomposed as $$R(\theta)R(-\phi)DR(\phi)$$ where $R(-\phi)$ and $R(\phi)$ are rotations by $\theta$ and $\phi$ respectively and $D$ is the diagonal matrix. This decomposition follows directly from the SVD: writing $A = UDV^T = (UV^T)(VDV^T) = R(\theta)(R(-\phi)DR(\phi))$ since $U$ and $V$ are orthogonal matrices.

So I have a couple questions.

  1. I thought orthonormal matrices such as $V$ are not just rotational matrices but also can be reflection matrices? If so, why does the decomposition above state that they are just rotation matrices?
  2. What is the point in decomposing it into 3 rotational matrices? Couldn't it just be decomposed as $A = UDV^T = R(\theta)DR(\phi)$?