Back on track! I have been trying to do some video webcam object detection benchmark between CPU only (i.e. IMX8PLUS SoC sans NPU) and with a Google Coral M.2 Edge TPU card.
My main sample code is from the Coral.ai camera support github and I modded it a bit to add FPS for comparison. There are quite a few pretrained models (already converted to Tensorflow Lite) that is available …
inception_v1/2/3/4_224/229
mobilenet_ssd_v1/2_coco/face
mobilenet_v1/2_1.0_224
So many to test! But for today, we want to just compare the performance between CPU vs with TPU accelerator. Our setup ..
Python 3.92 OpenCV 4.5.1 Tensorflow 2.12.0
mobilenet_ssd_v2_coco
IMX8PLUS 2GB RAM
Microsoft USB HD LifeCam
I will skip all the installation as pretty straight-forward and also platform dependant. I did have to install GStreamer as required by the examples-camera/opencv/detect.py script.
CPU Only
With M.2 Edge TPU Accelerator
I am getting 3X FPS running image recognition with the 4TOPS Coral TPU vs just the IMX8 SoC (which is also maxed out on one of the ARM CortexA53 core).
Look out for my further testing between the different models next …
Related Posts
Machine Learning for IoT .. an old school EEE rite of passage. Learn, Test, Integrate, Test and Deploy Solutions 