Hello, I have 2 questions:
I am using the NCS module on a PI3B+ I am getting around 8FPS with inference time 0.80ms which is decent.
The PI resources are from 50-80% by doing this, I am using the 20 classes Graph (person,car,tvmonitor,etc…)
I wonder if adding a new NCS will reduce the inference time?
Also if i retrain the model to use only 1 class would that make the inference time faster with one NCS module?
@chop It wouldn't really reduce the individual inference times as no information is actually shared between the devices. When adding another device, it pretty much lets you process another frame at the same time. So 2 devices = processing 2 frames at the same time. This means you will be able to process a video faster because you have more devices to process the frames from the video.
Thanks for the response, what about the a retrain of the model MobileNet SSD from 20 classes to 1 class would that make it faster?
I am just trying to understand as some graphs process faster than other, the birds graph is way faster but doesnt have person class so i cant use it.
I just cant understand how to get the fastest from that, i just want resources available to do other opencv operations.
Do we have some benchmarks with the PI? like what would be the fastest framerate archieved?
@chop I think your retrained model would depend on the accuracy you're trying to achieve. You might be able to just fine tune the last layers of the network and that may be good enough. I can't provide any Raspberry Pi benchmarks but I think some of our community members have posted some of their findings on the forum.