[Kdenlive-devel] Blender cannot perform "flexible" resampling

[Zachary Drew]

The MLT framework's architecture should allow for this effect to be  
implemented without much difficulty in kdenlive.

You will need a custom mlt producer. This is because a producer is  
required to supply a frame with the output resolution. This producer  
would use a library to open the image and apply the pan and zoom  
effect. Normally, this would be a filter but not here because of the  
native image resolution requirement. You can probably use built-in MLT  
transitions to dissolve between images. This assumes it is ok to apply  
the fade at output resolution and not native image resolution.

You can modify producer pixbuf supplied in the MLT framework as a  
starting point. You will probably want to use the inigo command line  
tool for development.

Building an interface in kdenlive is probably going to be more work  
and you'll need guidance from someone else on that.

Hopefully that helps!

Zach


On Dec 29, 2007, at 6:03 PM, Gerhard Hübner <Gerhard.Huebner at huebner-technologie.d 
e> wrote:

>
>>> The bad news: Blender downsamples every imported image to the  
>>> selected
>>> video resolution. When I zoom into an image using the transform  
>>> effect,
>>> Blender does NOT used the original image quality but just inflates  
>>> the
>>> pixels of the previously downsampled image!!! This ruins the whole
>>> function and makes it practically useless. My big multi-megapixel- 
>>> images
>>> just get squeezed to nothing and then inflated into an ugly  
>>> pixelated
>>> mess.
>>
>> AFAIK this doesn't happens in blender, i.e. it doesn't downsample.
>> You have problably set the render resolution to a low value.  
>> Furthermore
>> if you have a low res image and you set resolution to HD, it can  
>> even do upsampling
>
> I would not recommend to increase the video resolution as this results
> in HUGE output files which have to be downsampled again to be burnt on
> DVD or similar media. It could be done in a much much more efficient
> fashion. Although the mathematics is really simple, it seems to  
> require
> an explanation:
>
> 1. The image resolution and the output resolution of the video are TWO
> DIFFERENT THINGS!!!
>
> 2. Image resolution is usually greater than output resolution!
>
> 3. Therefore, when zooming into an image while watching the DVD  
> video, I
> want to see MORE DETAILS, not inflated pixels!!!!!
>
> Example:
>
> Let's say we have a picture with 2000x3000 pixels. I want to perform a
> continuous zoom from factor 1 to factor 2 over 10 seconds. This means,
> at factor 2, there are still 1000x1500 pixesl available, still much  
> much
> more than the DVD resolution!!! So, at zoom factor 1, 2000x3000 pixels
> get downsampled to DVD resolution. While the zoom continues, a smaller
> part of the image is selected and downsampled for each consecutive
> frame. At the final frame of the zoom effect, 1000x1500 pixels are
> downsampled to PAL. Therefore, throughout the whole zoom range from  
> 1 to
> 2, the image is ALWAYS sharp and there is absolutely no need to create
> huge output formats.
>
> These would be the keyframe parameters during the 10-second zoom from
> factor 1 to factor 2:
>
> 00 seconds: downsample from 2000x3000 to DVD
> 01 seconds: downsample from 1900x2850 to DVD
> 02 seconds: downsample from 1800x2700 to DVD
> 03 seconds: downsample from 1700x2550 to DVD
> 04 seconds: downsample from 1600x2400 to DVD
> 05 seconds: downsample from 1500x2250 to DVD
> 06 seconds: downsample from 1400x2100 to DVD
> 07 seconds: downsample from 1300x1950 to DVD
> 08 seconds: downsample from 1200x1800 to DVD
> 09 seconds: downsample from 1100x1650 to DVD
> 10 seconds: downsample from 1000x1500 to DVD
>
> This isn't rocket science, is it?
>
> Unfortunately, Blender is unable to do this simple kind of math.  
> Blender
> would depend on the output resolution being set to 1000x1500 and the
> final video being downsampled again to DVD.
>
> By the way, after fiddling around with Cinelerra, I found out that it
> handles this kind of transformation smoothly without any problems.  
> Seems
> like it isn't as horrible as it felt initially...
>
>
>
> --- 
> ----------------------------------------------------------------------
> This SF.net email is sponsored by: Microsoft
> Defy all challenges. Microsoft(R) Visual Studio 2005.
> http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
> _______________________________________________
> Kdenlive-devel mailing list
> Kdenlive-devel at lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/kdenlive-devel