At this point, our primary risk relates to the ML, and our ability to run that. Since this will be the most computationally intensive part of our project, making sure that we have enough power to run that algorithm has been our main concern, so that we achieve the time requirements we set out for ourselves. At this point, we are experimenting with/considering 3 different methods to run our ML: 1. Running identification on the pi. This seems the riskiest, and is something we hope to firmly rule in/out by the end of this weekend or beginning of next week. 2. Running identification on the web server, i.e. completely off of the hardware. This is generally the largest unknown, and so it is something we will be doing much more research on in the next few days, but it would have the advantage of significantly reducing the complexity of what’s running on our hardware (and allow us to drive the theoretical consumer cost down even further). 3. Running identification on a Jetson Nano. This is the ‘safest option’, and so it is something that we intend to be prepared for by getting a Jetson from inventory ASAP and trying to learn its setup. However it also adds a bit of cost to the project, as well as a significant amount of theoretical consumer cost, which is why we still intend to look into options 1 and 2. We hope that looking into multiple options will allow us to mitigate the potential risk that one or several of the options proves infeasible. The primary work required to evaluate each involves setting up the platforms and having an operational CNN.
We are considering where to run the ML, as detailed above. If we find that we can run it on the Pi, then this wouldn’t be a change. If we end up running it on the server, we will need to change the general structure of how the pi communicates with the web app, but otherwise this shouldn’t be a large change to the project as a whole. If we find that our only choice is to run it on a Jetson, this will be the largest change to the project – we can get a Nano from inventory without needing to use our budget, but would probably also have to acquire a wifi adapter of some sort, which would require a bit of budget. The current impact of the changes we are considering would mostly affect communication (if we intend to host an instance of the ML separate from the hardware) and obviously our hardware.
Trying to get the CV/ML code to run on hardware was always scheduled for this week, so this should stay as intended. There will be additional research needed at the beginning of this week to hopefully narrow down the feasibility. We also intend to acquire a Nano to experiment with ASAP.
