I'm trying to implement SIFT algorithm. After I found the min/max points in DOG images I need to find the real min/max subpixels by taylor exapnsion:
$$ \begin{equation} D(\mathbf{x}) = D + \frac{\partial D^T}{\partial \mathbf{x}}\mathbf{x} + \frac{1}{2}\mathbf{x}^T\frac{\partial^2D}{\partial \mathbf{x}^2}\mathbf{x} \end{equation} $$
http://en.wikipedia.org/wiki/Scale-invariant_feature_transform
I'm not sure how I do it. I understand the idea behind using taylor expansion.
Consider one-dimensional function $f(x)$.
The first order taylor expansion is $f(x_0+h) \approx f(x_0) + f'(x_0)h$
The second order taylor exapnsion is $f(x_0+h) \approx f(x_0) + f'(x_0)h + \frac 1 2 f''(x_0)h^2$
Now we expand three-dimensional function. $$ D(\mathbf{x_0}+\mathbf h) \approx D(\mathbf{X_0}) + \bigg(\frac{ \partial D}{\partial \mathbf x}\bigg)^T\bigg|_{\mathbf x=\mathbf x_0}\mathbf h + \frac 1 2 \mathbf h^TH(\mathbf x)\mathbf h $$ , where $$ \frac {\partial D}{\partial \mathbf x} = \begin{bmatrix} \frac {\partial D}{\partial x} \\ \frac {\partial D}{\partial y} \\ \frac {\partial D}{\partial \sigma}\end{bmatrix} = \begin{bmatrix} \frac {D(x+1,y,\sigma) - D(x-1,y,\sigma)}{2} \\ \frac {D(x,y+1,\sigma)-D(x,y-1,\sigma}{2} \\ \frac {D(x,y,\sigma+1) - D(x,y,\sigma-1)}{2} \end{bmatrix} $$ $$ H(\mathbf x) = \begin{bmatrix} D_{xx} & D_{xy} & D_{x\sigma} \\ D_{yx} & D_{yy} & D_{y\sigma} \\ D_{\sigma x} & D_{\sigma y} & D_{\sigma\sigma} \end{bmatrix} $$
Additionally, $D_{xy}$ $$= \frac { \frac{D(x+1,y+1,s)-D(x-1,y+1,s)}{2} - \frac{D(x+1,y-1,s)-D(x-1,y-1,s)}{2} } 2 $$
Let's go back. $$D' = \bigg ( \frac{ \partial D}{ \partial \mathbf x} \bigg ) ^T + H(\mathbf x) \mathbf h = 0 $$ $$\mathbf h = - H^{-1}(\mathbf x) \bigg ( \frac{ \partial D }{\partial \mathbf x} \bigg ) ^T $$
Assume that $\mathbf{x_0}$ is discrete feature point. Iteratively we can get $\mathbf h$.
If $\mathbf{x_0}$ is the closest point to $\mathbf h$, then we select $\mathbf{x_0}+\mathbf h$. (Imagine that 3-dimensional space)
My english ability is poor and there may be some errors.
