I was curious as to if anyone has used the RPi3 for real-time image processing? If so, was there any computational delay while doing it real-time?
I am planning to use an external GPU to train a CNN that can classify weeds and crops. I want to make sure the RPi3 can handle image processing features like segmentation and be able to classify the weed and crops in real-time using pixel-wise classification.
Hi Kannan. I’m not sure if this question is best answered by our community. You seem to ask if the Raspberry Pi 3 is powerful enough to perform a task that isn’t related to the FarmBot. While FarmBot will support weed detection, you are about to implement a custom solution.
I personally think your question is best answered on Stack Exchange, but it’s possible someone here is trying to achieve the same thing, and they will weigh in with their considerations.
A Pi3 has a vector processing unit and a fair amount of compute power. However, its not really up to the task of doing intensive image processing, just light weight stuff like affine transforms or transcoding and the like. Heavy weight/real time image processing takes openCL working against a fair number of GPU cores.
You could “stream” images off of a Pi and then do image processing somewhere else.
One other limitation of the Pi architecture is that all the “network” interfaces are basically hanging off of a single USB hub. That means your USB camera, ethernet and even the on board WiFi share the same bandwidth on the USB hub Sadly, this greatly limits the video + networking performance of a Pi. BTW, this is the same architecture as the Pi2. The 3 really didn’t change that much other than the Arm core and speed.