At Edge Impulse both sound and acceleration are time based inputs lasting a second or more. Computer Camera vision at Edge Impulse is for static frames. Any chance in the future for a really small resolution, say 32 x 32 and short duration say 100 milliseconds of video input?
Just enough to detect direction of motion or short duration complex motions etc.
Video input would be a very impressive next step.
Hi @Rocksetta, yeah absolutely an interesting thought. We’re actually adding something like this to deal with special types of cameras in the near(ish) future, so a natural progression would be to do something like this for video too. @dansitu or @matkelcey can probably comment better on the on-device performance.
I’ll leave it to @dansitu to comment re: device performance but, yes, video is a natural progression and we have a number of options to explore; ranging from full blown 3D convolutions to just simply stacking the time dimension into channels (e.g. a (32, 32, 3) image over a burst of 5 frames becomes (32, 32, 15)). I’m particularly excited about the special camera support @janjongboom mentioned where we can really support super low power setups. If you have any more specific info about a use case we can help formulate a first version.
Yeah the stacking of things is most interesting to me, as it also applies to other sensors (one of our customers is actually doing this), and we have a tracking bug for this open for >1 year
If you get video working on the Arduino Portenta with either of the 320*320 GRAYSCALE vision shields, you know I will fully test it out and point out any issues.
I have a few Portenta’s
Performance-wise it should definitely be doable on the Portenta, which has a fast processor and a bunch of RAM!