THe second one is a little taller, dunno which is the correct ratio tho. Whta I can assure is that the second one also has much less artifacts. Hope that helps.
Sorry for the late reply. The second one is the automatically stitched one, I have tried to utilize every frame in a scene to reduce the amount of compression artifacts from the encode. One of the main reason for making a program was to experiment with stuff I simply can't do in Gimp.
The second one because a little bit taller because one of the images was misaligned, thanks for noticing it, I fixed it now after quite a bit of fiddling with the algorithm.
But it is one of the areas I think will be difficult to get completely correct even with manual stitching, because I often see that the frames doesn't align properly on the pixel boundaries. Sometimes the frame is like 0.5 pixels off. I'm wondering if that could be an advantage, if I could align them with sub-pixel precision? But I think that is beyond my abilities...
I'm having a bit trouble with the noise filtering and automatic watermark (tv logo, credits. etc.) removal, so if anyone have some experience in that I would love to hear from you.
Here is how it looks now: (not a blind-test this time)
http://screenshotcomparison.com/comparison/2426
The second one because a little bit taller because one of the images was misaligned, thanks for noticing it, I fixed it now after quite a bit of fiddling with the algorithm.
But it is one of the areas I think will be difficult to get completely correct even with manual stitching, because I often see that the frames doesn't align properly on the pixel boundaries. Sometimes the frame is like 0.5 pixels off. I'm wondering if that could be an advantage, if I could align them with sub-pixel precision? But I think that is beyond my abilities...
I'm having a bit trouble with the noise filtering and automatic watermark (tv logo, credits. etc.) removal, so if anyone have some experience in that I would love to hear from you.
Here is how it looks now: (not a blind-test this time)
http://screenshotcomparison.com/comparison/2426
The thing is that the screenshot is already so blurry, not sure of any conventional method will clean it and leave it without much more blur that what already exist on it. Any method you use you'll basically want to leave any kind of guassian blur filter on it and just use a high chromatic value.
I cry whenever i see any sort of screenshot post on moe.
I cry whenever i see any sort of screenshot post on moe.
If it wasn't obvious I'm not working with that screenshot stitch in particular, I'm trying to make a generic method to apply to any (simple) screenshot stitch. For the example I specifically took a poor TV encode so that the issues would be more noticeable.
I'm developing my own program because I want to try out unconventional methods, there is no point in competing with Gimp. I can pull the image into Gimp afterwards to do that kind stuff.
Specifically, right now I'm trying to develop a filter which can take several overlapping layers and combine them, while extracting as much information from each layer as possible. Furthermore, it must detect irregularities and discard those.
I have now tried experimenting with sub-pixel alignment in Gimp by simply scaling two images up 8x, aligning, merging and scaling down again, and it seems like there might be a slight improvement. The lines seems a little bit sharper than in the resulting image than in the two source images.
However that is for two images, I have yet to try it out on a large scale, but I will have to implement it properly then. Scaling up and down is simply too CPU and RAM intensive...
I'm developing my own program because I want to try out unconventional methods, there is no point in competing with Gimp. I can pull the image into Gimp afterwards to do that kind stuff.
Specifically, right now I'm trying to develop a filter which can take several overlapping layers and combine them, while extracting as much information from each layer as possible. Furthermore, it must detect irregularities and discard those.
I have now tried experimenting with sub-pixel alignment in Gimp by simply scaling two images up 8x, aligning, merging and scaling down again, and it seems like there might be a slight improvement. The lines seems a little bit sharper than in the resulting image than in the two source images.
However that is for two images, I have yet to try it out on a large scale, but I will have to implement it properly then. Scaling up and down is simply too CPU and RAM intensive...
Unconventional method discovered ; )
I was just messing around with sharpening, and I found something that appears to work rather well. It is basically a Laplace edge-detect based sharpener.
Using a DVD re-encode:
http://screenshotcomparison.com/comparison/39630
Using a HDTV re-encode:
http://screenshotcomparison.com/comparison/39632
(Ignore the mouths, I have yet to add animation.)
This is just some early experimentation in GIMP, final results should be higher quality. I expect the posterization which can slightly be seen in the first image to disappear once I implement it in my 16-bit rendering pipeline.
The more real issue is with colored edges (seen the shadows), where it easily overshoots. Maybe that will be reduced when I only do the sharpening in the lightness channel, but I do not expect miracles... (Can't test it in GIMP because of posterization hell.)
Also the filter is likely highly sensitive to noise/compression artifacts. Not an issue for me though, usually these slides contains ~80-120 images which removes that nearly completely.
So what do you think? I'm not completely happy, but it certainly seems better than the Gaussian based unsharp mask.
I was just messing around with sharpening, and I found something that appears to work rather well. It is basically a Laplace edge-detect based sharpener.
Using a DVD re-encode:
http://screenshotcomparison.com/comparison/39630
Using a HDTV re-encode:
http://screenshotcomparison.com/comparison/39632
(Ignore the mouths, I have yet to add animation.)
This is just some early experimentation in GIMP, final results should be higher quality. I expect the posterization which can slightly be seen in the first image to disappear once I implement it in my 16-bit rendering pipeline.
The more real issue is with colored edges (seen the shadows), where it easily overshoots. Maybe that will be reduced when I only do the sharpening in the lightness channel, but I do not expect miracles... (Can't test it in GIMP because of posterization hell.)
Also the filter is likely highly sensitive to noise/compression artifacts. Not an issue for me though, usually these slides contains ~80-120 images which removes that nearly completely.
So what do you think? I'm not completely happy, but it certainly seems better than the Gaussian based unsharp mask.
Spiller
Automatic screenshot stitching
I spend the day yesterday on it and I would like some feedback on the quality. The following link contains a comparison of two stitches, one of them done by my program and the other one done manually in Gimp:
http://screenshotcomparison.com/comparison/165979
Which issues are there and what do you think I could improve? (both for manual and automatic stitching)
Btw, do you have any tips or guides on how to improve quality of this kind of stitching?
Nowadays all anime comes as Hi10p, do you think there would be a significant improvement to be gained from using 30 bit screenshots instead of 24 bit? (While I am a quality freak, I'm not downloading raws just for this.)