Hacker News Reader

Comment by Xcelerate | Hacker News Reader

I don't think this is necessarily a good idea. For one thing, automatically cutting off the edges of a frame is undesirable, exactly for the reason mentioned (in a conference call, it cuts the people on the edges off).

Secondly, zooming is not as effective as real compression. What I mean by this is: assume you have a high resolution image. One way to save bandwidth is to use a well-designed compression algorithm optimized for the human visual system. The second way is to just shrink the image, and then "stretch" it back to the original size, which -- in a sense -- is what Skype is doing here. Which is going to be more effective?

View on HN

One way to save bandwidth is to use a well-designed compression algorithm optimized for the human visual system. The second way is to just shrink the image, and then "stretch" it back to the original size, which -- in a sense -- is what Skype is doing here. Which is going to be more effective?

It really depends on how much data you need to shave off. Beyond a certain point, you'll get better visual results by reducing image size rather than increasing compression. As an example, an SD movie encoded with H.264 to a file size of 300MB would look a lot better than an HD movie encoded to the same file size, even when played back at the same on-screen dimensions.

This made me think... Is that a misfeature or bug in the compression algorithm? Why wouldn't the video codec adapt appropriately? If lower res causes a more even and smooth image, why shouldn't it scale down?

I had typed a lengthy introduction to an explanation, but realized I don't know exactly where all the bits are going in a high-res low-bitrate image. All I can say with confidence is that if you squeeze a high-resolution image into the same size as a low resolution image, the low resolution image will look better. The high resolution image will basically be reduced to storing DC coefficients, so you'll see these giant 16x16/8x8/4x4/etc. pixel blocks of solid color, with maybe a bit of pattern on them that barely correlates with what the original looked like.

Maybe it comes down to the per-macroblock overhead; maybe fewer blocks total in the lower resolution image allows more bits to be allocated to frequency coefficients instead of DC offsets.

Yeah, I've observed the effect, but it never occured to me that it doesn't (or shouldn't) have to be that way :) We should be able to demand codecs that are smarter and more adaptive :)

It's hard to imagine why anyone would want a blocky highres lowbitrate stream if the lowres variants are always better, for a given timespan of video

Though this isn't implemented by the reference encoder yet, the VP9 bitstream supports this (encoding a downscaled frame and signaling to the decoder to upscale it before displaying it).

Reply on news.ycombinator.com