Make a new version of the Movidius Compute stick which has the 415 or 435 cameras plug straight into the stick to eliminate
throughput issues on slower micro PC devices. The Stick itself still plugs into the PC USB port.
If you dont want to use the 415 or 435 cameras Intel could also sell the hardware that would allow any camera to interface to the stick directly.
Or maybe just offer the hardware to do this and the port on the Compute stick.
From this wonderful marriage you can develop the nirvana of an SDK which can deliver the BOXES names and the percentages,
Optionally you can deliver the actual images, and the distances and or heat maps via the USB to applications.
Using just the info with no pictures greatly reduces the data size.
This would make it NIRVANA .
You can concentrate on developing the business logic and neural network aka deep learning code on a more powerful box
and when complete stick it onto a small micro device that needs just the business logic and not the image information.
Taking a huge burden off the IOT device. Think of simple things like facial recognition etc.
If someone doesn't want to use this option no problem have the stick work as it has but PLEASE have it directly support and optimize
Tiny Yolo. Most Demos I have seen show this off now.
Hopefully getting the low cost disconnected Deep Learning device of my dreams.
Thanks for posting your opinion in the RealSense community. The community will analyze your idea which sounds really interesting.