Does ippiMatchTemplate able to get subpixel accuracy?

[shenganx]

Hi there,
I wonder that whether ippiMatchTemplate could reach subpixel accuracy?
Thx

[seiji-torigoe]

When it is the easiest
1.I use ippiMatchTemplate_CoeffNormed_8u32f_C1R.
2.I use the maximum and the four pixels in neighborhood of the result.
3.I do curve fitting by using it.
4.I solve the simultaneous equations every x and y.

simultaneous equations
MacthValue1 := a * x1^2 + b * x1 + c;
MacthValue2 := a * x2^2 + b * x2 + c;
MacthValue3 := a * x3^2 + b * x3 + c;

differential
0 := 2 * a * x + b;

SubPixel := Peak
x := - b / a / 2;

y is also similar.

[shenganx]

Hi, thank you for your reply.
I did use this simple approach to intropolate the subpixel accuracy. but I do not think there is lot of proof of that. Why use polynomal why not Gaussian?
Thx

[seiji-torigoe]

When I use the value of three pixels to calculate,
the Gaussian function obtains the same result.
The Gaussian function can be transformed into the same type.
ln(MatchValue) := a * x^2 + b * x + c;

[shenganx]

Hi thank you very much.
I can not post plot here, the problem I met is the correlation result jittering beteween two pixels, so I assume the correct correlation result should be in the middle and that is why I want subpixel accuracy. The polynomial interpertation approach I tried. If you looked at my matlab plot, the interpertation results also jittering between two values and the shape of the subpixel result is identical to the orginal one. they just have a small offset. SO that is why I wander whether this approach is good enough.
Thank you again for you kind help.

[seiji-torigoe]

I think,
When the image is symmetry, there is no offset.