2019-07-25, 03:06 AM
After racing shorthanded in the Labrador Current in the fog, I spent a while contemplating icebergs. They don't show up on radar. Apparently they're more easily seen in IR than visible light, and especially in the fog. Same might possibly be true of buoys---as well as other boats---since they probably won't be quite at sea temperature. And if fog truly scatters IR less than visible, this could be fun! (Yes, I know that putting an IR camera onboard isn't the same as hooking your autopilot to it, but work with me for a moment...
Training a deep network to distinguish obstacle type would take a lot of CPU, but just running it forward at a low frame rate (1 Hz should be ample) shouldn't require a huge computer, especially considering that a very powerful network is likely unnecessary: the cost of misidentifying a kayak as a crab pot buoy should be low, since we probably take similar actions in each case. I wonder if an RPi might be enough for a first pass...
Training a deep network to distinguish obstacle type would take a lot of CPU, but just running it forward at a low frame rate (1 Hz should be ample) shouldn't require a huge computer, especially considering that a very powerful network is likely unnecessary: the cost of misidentifying a kayak as a crab pot buoy should be low, since we probably take similar actions in each case. I wonder if an RPi might be enough for a first pass...