Stitching images

[magnussq]

Hi there!

I am using Ipp 5.1 to get several overlapping 24bit rgb images to be merged into a larger image. For this I would like to use a proximity measure function to find the best match between pairs of images, the ippiSqrDistanceFull or ippiCrossCorrFull functions.

The problem I face is when the overlap is small, then these functions give very unreliable results. The reason is that the images gets a zero-padding. This makes SqrDistance useless of course, while cross correlation functions very rarely finds a good match.

So I am wondering if I am on the wrong track with using these functions for stitching images? In principle the sqr distance method seems good, but it needs to be on a per-pixel average and disregard any padding, something which IPP cant do. I am working with SqrDistanceValid now instead but overlapping equal sized images this mean many smaller matches instead of one big match.

Regards
Magnus

[Intel_C_Intel]

Hi, Magnus!

To use cross corr for your purpose you need compapatively big areas in images to stitch. It can be insufficients if image brightness or other characteristics differ.

Generally, image stitching (panorame image creation) is done via key points search in images. The most popular are SIFT (scale-invariant...) features. We plan to include functions to cover the most time-consuming part of SIFT algorithm to IPP and to provide the corresponding sample.

Alexander