Hi tls,
Back on the farm, we used to count the legs and divide by four.
This is a proof-of-concept arrangement, so not optimized for commercial exploitation. It is at this stage an algorithm running in Raspberry Pi.
The heat sink indicates the processor is generating a bit of heat. I'm not sure if they could dispense with the heat sink if they had Akida, but they could use a smaller one, which would lighten the load for the drone, and, as we know, Akida would extend the battery charge.
Apparently Edge impulse does a lot of this proof-of-concept modelling:
https://www.edgeimpulse.com/blog/getting-more-cycles-per-second-with-fomo
Note that this system does not appear to have one-shot learning, and needs a significant training database.
Basically they are using Edge Impulse's FOMO (Faster Objects, More Objects) algorithm:
https://www.edgeimpulse.com/blog/announcing-fomo-faster-objects-more-objects
It appears that there are synergies in using FOMO and Akida.
I hadn't previously looked into EI's FOMO, but the above link provides an excellent introduction, and the synergies become apparent.
FOMO uses a form of CNN and has significant limitations.