Hi,
can you talk a little bit about the feasibility of some simple SIFT implementation (maybe only hundres of features instead of thousands)? What can we expect to work and what is simply beyond the processing power at hand?
Best,
Basil
I can’t really give you a concrete answer here as a lot of these kind of details depend on implementation. SIFT is definitely do-able on Pixy (and has been on my wish-list of things to add for some time), but to prevent degraded performance you may need to take some shortcuts, which is typical for these kind of algorithms on embedded devices.
Given the processor speed (208MHz), frame size (640x400 at 50 fps), and steps of the algorithm we can come up with a rough estimation as to how long each frame would take to process (sorry, but I’ll leave the rest up to you).
I’m sure this is of interest to others out there, so hopefully we get some early contributors to the project, otherwise we’ll do what we can to get a decent implementation of SIFT on Pixy within the next few months or so.
Scott