A Resample Frequency below 0.0014 only spits out the color source twice at the origin, one horizontally flipped and rotated by -90° in the background, values above 0.0014 process normal. Using PAL D1 that is. I notice when using bigger formats the processing starts earlier.
Probably any value where Width/Frequency is <= 1.
- Chad
Chad was right. The issue is fixed and will be in the next release.
Brian!
Using the May 8 2008 build with Fusion x64 5.3 build 60…
If you increase the sample frequency, the “Make White…” option doesn’t seem to work well. I assume it’s sampling on both sides of a white region, and interpolating inbetween, but in the case of the samples on the doc page, it’s a bit of a problem.
If the resample frequency is low, you get the same line inserted over and over, resulting in a bad smear. But if you increase it, the white ball gets stretched. If you try to resize to say, 850 pixels, there’s no way to get both, no smearing, and no insertion in the white regions.
- Chad
Yeah, eventually the resize will need to cut through white areas. One issue is that the sobel detector in fusion seems to normalize the result, so some of the pixels in the image are assigned infinite energy when they should not be. The best way of dealing with this is to turn off the “Make White…” button, and paint on the sobel image with a colour value greater that 1.0.
Brian
On my tests, the built in Sobel is not normalizing, but it’s definitely not right, either. Much more blown out than a homemade Sobel makes, ala (sqrt(gx^2+gy^2)). But if you normalize both the built in and the homemade 3x3, you get identical results.
I’m definitely running into images where there is no good compromise between “sampling small enough to respect the high energy” and “sampling low enough to prevent smearing”. Is there a way to do per-pixel sample frequency? Or is that nonsense?
- Chad
The resample frequency is the percentage of the width of the image, so its defined on a per-pixel bases that scales with the auto proxy. I could potentially add an option to never cut through high energy regions, then the resampling frequency would max out on the maximum number of cuts that are possible. Would that work?
Brian!
Not sure exactly. I’m having such a bugger of a time getting this to work right, I’m trying all sorts of custom kernels to make the energy map, trying different resample frequencies, etc… Don’t sweat it, as I’m still trying to determine what would actually help.