Thank you for the experiment. There is one more paramters that would have been very interesting to measure - the overall CPU load. Running Tensorflow on the CPU is incredibly expensive (object detection on a single camera uses about 73% CPU for a single HD camera at 5 FPS in a RPi 4B 4Gb). A Coral USB is bound to reduce that dramatically, which also includes the overall CPU temperature long term. It would be really interesting to see a part 2 where you evaluate that aspect as well. Thank you for the time you dedicated to this - it is really useful.