verne.freenode.net changed the topic of #mlpack to: http://www.mlpack.org/ -- We don't respond instantly... but we will respond. Give it a few minutes. Or hours. -- Channel logs: http://www.mlpack.org/irc/
witness has quit [Quit: Connection closed for inactivity]
pgunnars has joined #mlpack
< pgunnars>
hey
< naywhayare>
pgunnars: hello
< naywhayare>
there was a message some time back I meant to respond to but you didn't come back... let me dig it up
< naywhayare>
think it had to do with boost
< naywhayare>
ah, yeah, it was boost. you should be able to do that by specifying -DBOOST_ROOT when you call cmake, but it's been several weeks so I wouldn't be surprised if you've got it worked out by now :)
< pgunnars>
whuuuu
< pgunnars>
i cant even remember hahah
< pgunnars>
oh wait yeah
< pgunnars>
thats right
< pgunnars>
was having trouble installing
< pgunnars>
anyway, ive got a general ml question
< pgunnars>
If I have a model which detects concepts in images/frames, whats a good strategy of aggregating the results to video concept detection
< naywhayare>
hmm, so I don't really do CV applications, so my first guess is going to be something that definitely isn't the best thing to do...
< naywhayare>
certainly if you're trying to detect a video concept, some notion of continuity between images/frames is necessary
< pgunnars>
yeah
< pgunnars>
ofc
< naywhayare>
hm, hang on
< pgunnars>
but different scenes can have different concepts
< pgunnars>
so you cant just average it out somehow
< pgunnars>
a video definitely contains a concept even if its only for a second
< naywhayare>
right
< naywhayare>
so you're already detecting concepts in individual frames
< pgunnars>
yeah
< naywhayare>
why not compare concepts from one frame to the next, and if they are sufficiently close/similar, you mark them as being part of the same video concept?
< naywhayare>
"sufficiently close" might be some comparison of how nearby the concepts are (in the image) and the shape of the concept
< naywhayare>
I'd be willing to bet any one of the opencv guys would have way better suggestions than I do :)
< pgunnars>
not sure that will be possible
< pgunnars>
yeah true sorry
< pgunnars>
shouldve tried that first
< pgunnars>
duh
< pgunnars>
thanks tho
< naywhayare>
:)
< naywhayare>
CV is a fascinating field, I've just never had time to get into it much
< naywhayare>
lots of heuristics
< pgunnars>
i think im going to have to limit myself to detecing concepts in scenes
< pgunnars>
change scenes when descriptors drastically change
< pgunnars>
use sequential probability ratio test for frames within scene?