We can go far by analysing various synthetic images – but real-world images are different and come with problems we have to solve and adapt to. I suggest we use this blog category to add things we have found. Now and then I will collect submitted items onto a Master List Of Real World Problems (MLORWP!).
This is the MLORWP I can think of right now:
1) Real images are not centred in the image frame – take this into account when using synthetic images for analysis of some sort – the synthetic images are all right in the middle of the frame.
2) Real images have slight variations in scale and rotation due to slippage etc in the hardware. While we can always measure what the problem is we need to make sure that e.g. synthetic images are generated for the same conditions (e.g. image rotation – the CCD is mor eor less free to rotate!).
3) Use of FFT methods to convolve images carry some consequences – one Chris put his finger on is the centering of objects by the very act of folding: We need to make sure that when we model real images the resulting synthetic image is offset by the right amount. Henriette’s Python project with Kalle Åström at Lund U could come in handy here.
4) The noise in real images is probably higher (never lower) than that given by Poisson’s distribution.