Stalker
There’s this outgrowth of music retrieval types that love to do summarization — they take in a four minute pop song and find the “optimal” summary– maybe it’s 10 seconds long. Of course, with nothing but the signal to lead you, all the summarization algorithms read like pulpy redactions from a Pierce/Shannon appendix: the maximal highest aggregate entropy sequence or degenerate KL divergence from the HMM state path (did you like those? I just made them up. I should always keep Matlab open.) I guess these are fancy ways of saying “chorus.” I’d rather listen to the bridge for a good introduction to a track. If the bridge is good, the track’ll be good. But is the bridge only good in relation to the verse? Deep stuff- think Christgau has a piece on this?
I bring this up because I’ve been fascinated with some computational way to do the same thing on movies. Same thing, right? Just a signal. So I used my human brain to pull a ground truth™ frame on the classic Stalker. Obviously, from that one frame you’ve got the gestalt of the whole movie right there. There’s some guys, they’re in a room that’s falling apart, someone just turned on a light. That’s Tarkovsky’s entire oeuvre right there. So now I’ve got to spend the next few months on my peeling paint SVM and my HTK chatty Russian state transition matrix.