Ben Hur retargeting challenge

Alec Jacobson

June 29, 2011

weblog/

Recently I got the chance to see a friend's new image retargeting algorithm at work. It was very impressive and the best I've seen so far. I particularly liked that uniform scaling was within the set of possible solutions and actually occurred frequently. I like this because often uniform scaling looks way better than fancy retargeting techniques which sprinkle artifacts all over the image. See the results reminded me again of Ben Hur. This movie was shot at a whopping 2.76:1 aspect ratio. Retargeting it to say, 4:3, would be a mess. All the times when I remember seeing it shown on TV (before widescreens became popular) it was shown with "letter boxes" rather than "pan and scan" or any other attempt to retarget the footage. Not even considering the fact that the original aspect ratio was intentional and that by retargeting it you'd be trivializing the directors' choices (would a museum crop or stretch a painting to fit the gallery walls better?). Not even considering this, Ben Hur poses a striking challenge for retargeting techniques. To me it represents a real stress test. Much harder than most of the images in the RetargetMe dataset. Here are some stills I found flipping through the film that I thought would be particularly difficult to retarget: ben-hur-arab-2 ben-hur-arab-3 ben-hur-arab ben-hur-drummers ben-hur-march ben-hur-parade-close ben-hur-parade ben-hur-senate ben-hur-ship-rowers ben-hur-trumpets ben-hur-chariot I ran some 4 of the easiest methods for retargeting on the last image.

Letter boxing

ben-hur-chariot-letter-box

Uniform scale

ben-hur-chariot-uniform-scale

1/2 Uniform scale, 1/2 letter box

ben-hur-chariot-uniform-scale-and-letter-box

Seam carving

ben-hur-chariot-seam-carving Personally none of the above options besides Letter boxing are acceptable. Not even close. I would really like to see a retargeting algorithm that not only includes uniform scaling and cropping in the set of possible solutions, but also letter boxing. Do you know of one? It goes without saying that the problem is much harder when considering not just a single frame but the entire scene or film's worth of video. Letter boxing may be harder to include as an option when considering temporal coherence, but I still think it must be a possible solution. For films like Ben Hur, which purposefully utilize their entire aspect ratio, squishing a scene or cropping out elements would be far worse than the price of two black bars.

Comments

June 29, 2011, Colin
Knowing very little about the subject, I would think that an algorithm could simplify to rescaling (as a special case), but I can't see how an algorithm could simplify to letter box without having some kind of thresholding on a confidence value. I'm not sure if confidence is already defined but it would seem to me that it would be hard for an algorithm to know, for example, that the horse's faces got all messed up and that that's a bad thing. Also, I want to see a video of seam-carving frame-by-frame on a video. I think that would be a really interesting effect.
June 30, 2011, Colin
I would think it would take an obscene amount of special cases to get the saliency map working well enough. "Another cool effect I’ve been thinking about is using seam-carving to shrink and image then starting the algorithm over on the small image to stretch it back to the original size." Let's see it! The new datamoshing!