Sum of absolute differences (SAD) diverse results

Saira

Hi,,

I would like to use Sum of Absolute Differences(SAD) for the purpose of tracking an object on DM6437. SAD generates values in a diverse range. How to normalize SAD for acceptable end results ??

and what will be the threshold of SAD values on which to decide about the template updation ,,,, like for normalized cross correlation the suitable threshold is 0.8. ????

Regards

Saira

over 13 years ago

Senthil Kumar Yogamani over 13 years ago

TI__Intellectual 1450 points

Hello Saira,

There is no standard way to normalize SAD.

Typically it is not normalized because it would take more computations and a heuristic threshold is chosen for the SAD score by trial and error method.

Here is one of the ways which would bound it within 0 to 1.

Think of the 2D blocks to be matched as 1D vectors x and y.

SAD(x,y) = | x - y |₁ (L1 norm)

NSAD(x,y) = | x - y |₁/ ( | x |_{1 +}| y |₁) (Normalized SAD score)

Basically you take the sum of all values in x and y, add them and use it to divide the SAD score.

This normalized metric will give 0 when x and y are the same and will 1 when one of the value is 0.

This normalized metric is also incremental wrt changes in x and y.

Choosing the threshold is often by trial and error. 0.8 might be a good threshold for NSAD too.

Regards,
Senthil

Saira over 13 years ago in reply to Senthil Kumar Yogamani

Intellectual 565 points

Thanx alot...

I vll try it

Regards,

Saira

neo marker over 13 years ago in reply to Senthil Kumar Yogamani

Prodigy 35 points

Hi Senthil

Can you please help me out with computer vision and openCV, I am working on AR system and want to write a library module to recognize some patterns.

regards

neo

neo marker over 13 years ago in reply to neo marker

Prodigy 35 points

hi senthil

i need some help on opencv ,

regards

neo

Processors

Processors forum

Sum of absolute differences (SAD) diverse results