trying to understand what is feasable in terms of fast video-tracking performance:
1 is the understanding right that for best results in vvvv, one needs custom per-camera opencv nodes written?
2 are there examples of tested hi-res/small-delay/high-freq cameras ?
3 what could be the research grade specs in the moving-mirror-demo? custom os etc
halso, gpu tracking, could it make sense ? http://docs.opencv.org/modules/gpu/doc/introduction.html
-
“best results” depends on optimizing for specific use cases, so no, there are more options than opencv in vvvv, even though it is quite mighty. Straight copy to GPU for further parallel processing is good too.
-
@tmp)) made some impressive progress with hybrid gpu/cpu tracking with kinects, completely without opencv. it is a viable option, because analysing and filtering can be done on gpu even better. For interaction you need live ReadBack to the cpu, which is possible with dx11 in realtime. There is this old ((forum:pipet-vs-compute-pipet discussing different approaches and even hinting at a performance competition.
all current approaches are mighty expensive because all attempt to order calculation and actual readback within the same frame instead of deferring it by two frames.
It’s off topic, but imo the challenge has not been fulfilled yet. -
I don’t have any xp with high speed cameras (above 120fps anyway). As for the mirrors, that should be possible too with some tinkering punk, because really it is the same tech as for lasers LD200 (Devices) Lumax (Devices) SourceBuffer (DShow9) after the cam is focused like that.