YOLO or SSD implementation

[ghostjat]

Is your feature request related to a problem? Please describe. A clear and concise description of what the problem is. "Object detection & recognition in image frame"

Describe the solution you'd like i did some google and found yolo & ssd is a good way for it.

Describe alternatives you've considered I tried darknet using ffi to try yolo

[andrewdalpino]

Hey @ghostjat thanks for the great suggestion!

I'm currently unfamiliar with YOLO and SSD but will do some research to see if it's something we can implement in Rubix. We may need to solve another problem first beforehand - namely, the one of enabling efficient storage and parallel computing on matrices. We're currently working on this particular problem as it will enable large-scale deep learning in PHP. Computer vision systems tend to require alot of compute. I'll keep you updated in this thread.

[andrewdalpino]

I forgot to ask @ghostjat, how was your experience using FFI to interface with Darknet?

We're considering moving some code from PHP to C if FFI is a viable solution in the future

Thanks again

[ghostjat]

Hey @andrewdalpino My experience is good with ffi but with darknet i faced callback issue. FFI performance is slow when working with c struct but consume less memory. There some limitations with ffi, it doesn't support macros func From the point of view of functionality awsm i created interface for SOD, CCV & vlfeat. So in ffi u will get less memory consumption but slow.

[andrewdalpino]

Were you able to get Darknet to work still with the callback issue @ghostjat? If so, how slow are we talking? Comparing Darknet with and without FFI.

[ghostjat]

I failed to load image in Darknet. therefore i can't say about ffi performance with Darknet. as per my previous working experience accessing c struct in ffi is almost 2x slower than php array.