Gradient Notation in Level Set Methods

98 Views Asked by At

I'm trying to get my head around level set based segmentation. I've considered many sources on that topic, but all of them use a formulation that confuses me. Let me describe briefly by quoting exemplarily from Cremers (2006), where we write $\Omega$ to denote the image domain:

Countours are represented as the zero level line of some emedding function $\varphi : \Omega \to \mathbb R$: $$\mathcal C = \left\{ x \in \Omega \middle| \varphi\left(x\right) = 0 \right\}$$ [...] A contour is propagated by evolving a time-dependent embedding function $\varphi\left(x, t\right)$ according to an appropriate partial differential equation. [...] Since $\varphi\left(\mathcal C\left(t\right), t\right) = 0$ at all times, the total time derivative of $\varphi$ at locations of the contour must vanish: $$\frac{\operatorname d}{\operatorname dt} \varphi\left(\mathcal C\left(t\right), t\right) \stackrel{(\ast)}{=} \nabla \varphi \frac{\partial \mathcal C}{\partial t} + \frac{\partial\varphi}{\partial t} = \text{[...]} = 0$$

Why does the equation $(\ast)$ hold? Obviously, $\varphi$ was extended to be a $\Omega \times \mathbb R \to \mathbb R$ function. So as far as I can tell from e.g. Wikipedia, the total time derivative of $\varphi$ should rather read $$\frac{\operatorname d}{\operatorname dt} \varphi\left(\mathcal C\left(t\right), t\right) = \frac{\partial \varphi}{\partial x} \frac{\partial \mathcal C}{\partial t} + \frac{\partial\varphi}{\partial t}\text,$$ but everybody seems to rely on the gradient $\nabla \varphi$ in place of $\partial \varphi / \partial x$. Why is that? Isn't it $$\nabla \varphi = \begin{pmatrix}\partial \varphi / \partial x \\ \partial \varphi / \partial t\end{pmatrix}\text?$$ Could it be, that $\nabla \varphi$ for some (confusing) reason is commonly used to denote $\partial \varphi / \partial x$?

1

There are 1 best solutions below

0
On BEST ANSWER

In case somebody else will be wondering about this in the future:

  • Yes, $\nabla\varphi$ denotes the spatial derivative of $\varphi$, that is $\partial \varphi / \partial x$ in the given context.
  • IMHO, $\varphi_t$ would be a better notation than $\varphi\left(x, t\right)$, because this would render $\nabla \varphi$ unambiguous.