cool new equivariance result

let $\Phi$ be an U, U equivariant registration network.

\begin{aligned} \Psi_0 := \Phi \\ \Psi_1 := TwoStep\{\Psi_0, \Phi\} \\ \Psi_2 := TwoStep\{\Psi_1, \Phi\} \\ \Psi_3 := TwoStep\{\Psi_2, \Phi\} \\ ... \end{aligned}

If the sequence $\Psi_i$ is convergent, call its limit $\Psi$

using usual fixed point argument,

\Psi = TwoStep\{\Psi, \Phi\}

Apply to images A, B

\Psi[A, B] = \Psi[A, B] \circ \Phi[A \circ \Psi[A, B], B]

\Phi[A \circ \Psi[A, B], B] = id

Substitute $A = M \circ W, B = F \circ U$

\Phi[M \circ W \circ \Psi[M \circ W, F \circ U], F \circ U] = id

Use that $\Phi$ is U, U equivariant

U^{-1} \circ \Phi[M \circ W \circ \Psi[M \circ W, F \circ U] \circ U^{-1}, F] \circ U = id

\Phi[M \circ W \circ \Psi[M \circ W, F \circ U] \circ U^{-1}, F] = id

Substitute A = M, B = F, use that both sides are identity

\Phi[M \circ W \circ \Psi[M \circ W, F \circ U] \circ U^{-1}, F] = \Phi[M \circ \Psi[M , F], F]

This is not quite the same thing as W, U equivariance, which can be stated as

W \circ \Psi[M \circ W, F \circ U] \circ U^{-1} = \Psi[M, F]

But it's really close. Informally, we showed, "If Psi is not W, U equivariant, it has to be either in a way that the resulting transform warps M identically, or in a way that the differently warped M causes Phi to have the same output"

Back to Reports

subdavis.com forrestli.com