It looks like you're new here. If you want to get involved, click one of these buttons!
i wonder if its possible to do stereo vision processing on a raspberry with this stick for example depth measuring. Did someone try something like that? What framerate can be expected?
Some examples would be fantastic.
Were you able to do stereo vision processing on the pi?
in my minds this would be a reason to have 2 units. One to create regional information via stereo processing the other to evaluate the images leaving you one frame behind in the realtime evaluation.
@chicagobob123 depth based image processing is an interesting and challenging topic. You probably already know this - depending on your application requirements, you can choose to go with structured light based depth cameras (like RealSense), LiDAR (RP LiDAR for cheaper alternatives), stereo/binocular vision or many other depth solutions out there. If you are specifically interested in doing DNN accelerated depth map predictions, here's a good resource - https://papers.nips.cc/paper/5539-depth-map-prediction-from-a-single-image-using-a-multi-scale-deep-network.pdf. Here's a high-level explanation of the threory behind stereo vision - https://simonl02.users.greyc.fr/files/Documents/Teaching/VisionLabs/lab7.pdf. The hardware part of this problem is equally challenging since the displacement between the 2 cameras plays a major factor in your disparity calculations. Here's a good reference to tackle the HW problem - https://www.raspberrypi.org/blog/real-time-depth-perception-with-the-compute-module/.
That being said, how does any of the above references relate to NCS? The short answer to your question is: No, we do not have out of box sample codes to do stereo vision on RPI + NCS. This is more of an application level problem, which would involve piecing together multiple hardware and software components as mentioned in the above references. As the community grows, and we start rolling out other functionalities on NCS, somebody might be able to piece together these pieces to create a functional DNN based depth perception on NCS.
This begs the question how do you get any performance for something like YOLO or R-CNN?
@chicagobob123 we don't yet support Yolo and R-CNN. Please see release notes associated with each software version for more details on supported layers. https://ncsforum.movidius.com/categories/downloads > Pick the latest software release > NC SDK Release Notes