As a small follow-up to my 2009 post ["Why You Shouldn't Rotate Your Images In Windows"] detailing image degradation from repeated jpeg compression, I have done some experiments with video and the effects repeated saves using the XVID [wiki] implementation of the MP4 standard will bring.
I won't be going into the same detail as the other post, but I may return to this again at some point in the future. I've only used one example video, which is Italian make-up artist Giuliana (Makeup Delight) [youtube]. The video choice was somewhat arbitary, but there is good contrast in image content. Lots of movement in the foreground with facial expressions, while the background remains relatively still.
As you will see, the final videos get fairly mangled by Youtube's own compression algorithm, which is why I put off trying video experiments for so long. You can increase the resolution to 480, which helps, and I have included some still frames from the videos below too.
I dialled everything back on the XVID codec settings, to really get the compression process working. When I initially started experimenting I found that the degradation process was taking too long, and while I probably could have implemented an automated process fairly easily, I just wanted to try this on-the-fly. So to increase entropy in the image, I applied two types of filters. With the video above I have added a simple (standard VirtualDub) sharpening filter, at about 50%. This effectively magnified the effects of each iterative save, by increasing the detail in the changed edges. I only went as far as 52 iterations, but you can clearly see the effects it had on the video. The image becomes infested with squares, with multitude of apparent geometric patterns. More on that later.
N.B. no audio track
For the second video, I used a error diffusion dithering filter [http://whitewing.co.uk/vdub_errdiff.html] which gave the effect of increasing the relative noise in the image. The effect here, looks more like the results garnered when experimenting with jpeg compression. It's not that clear in the Youtube version, but the background really begins to move about even after just one save. Most apparent here is the sudden flashes of noise which gradually fade back. I imagine these are based on the key-frame points in the original video.
I'm not going to pretend I know exactly what's going on here, but I would assume that the shapes represent the xvid quantize matrices, which due to the extreme measures of compression, have reduced them themselves down to binary values (either 0 or 255). I'm also going to guess that the reason they keep changing between frames, is because of a cascade effect from surrounding matrix squares. Someone reading this can probably explain it better, and it would be great if they offered an explanation in a comment :)
A 100% crop of a frame from the dithered version
That's as much as I want to discuss for now. Be sure to check out Makeup Delight's blog [ http://www.makeupdelight.com/ ] and Youtube account [ makeupdelight2009 ]. I've been following her videos for some years now and she has produced many videos detailing fine examples of her makeup artistry. Ciao.
Posted at 20:58