Quote:
Originally Posted by elguaxo
|
Indeed, a good post.
More technically, there are many issues with the bpp approach.
When you downscale a film by, say, a factor of 2 in each direction in order to encode it, you're
effectively doing three things (this is in MPEG-4 ASP terms):
1. Increasing the macroblock size from 16x16 to 32x32, reducing overhead but also reducing the fine-grainedness of mode decisions and partition sizes.
2. Increasing the transform size from 8x8 to 16x16, compressing more efficiently in flat areas but much less so on sharp edges.
3.
Lowpassing the bottom, right, and bottom right quadrants of the "virtual" 16x16 transform. This is where the bits are saved.
4. Reducing motion vector precision from halfpel to fullpel.
The sacrifices here, especially 4), are really going to hurt compression.
Now, not to say that there isn't a point at which downscaling is worth it; this point is just a lot higher quantizer than you think it is, probably around the time that a large portion of bits are spent on things that are easily encoded on a larger scale: macroblock modes, large partitions and their motion vectors, and DC coefficients.