This is a good example of a hardware requirement for the MV1 that is not flexible to us as developers. We don’t plan to add a second camera, NIIR cameras, or adjustable height features anytime soon. I’m worried that we’re jumping to the ‘cheats’ (which rely upon changes to the hardware) before fully understanding the requirements and the nature of the platform. Just as one quick example, here’s another thing to consider about our dataset being timeseries:
They will always be sequential - this is a timeseries dataset. Depending on our rate of capture (currently once an hour, but could be more if there was a use case) we can easily reference an image that should be very similar to the one being taken.
I’m concerned about your approach @hmw not being applicable to anything beyond one set of images. Assuming step 1 is completed correctly, using cards to identify the ROI, wouldn’t any movement of the reservoir itself throw that ‘calibration’ off?
In a school environment, students will inevitably be removing the reservoir to fill it with water, at which time they may also remove the pots to measure them or thin them as well. I would like for any solution to be able to work for MV1 users with 3" pots but also using 2" net cups, net pots, or hydroton balls as well. Unless we first explore that route and determine it to be impossible - I’d like to maintain that goal. Photos in this thread are good examples: MV1 Summer Trials
I know this is not the same as an algorithm to accurately measure plants. That’s not what we want. We want to know if Suzies tomato plant is dying (getting smaller) or if Billies radishes have germinated (any green dots) yet - that’s it.
I would be curious for someone on @ajc3xc’s team to read this paper titled ‘Segmentation of Overlapping Plants in Multi-plant Image Time Series’ by Dr. Malia Gehan at Danforth.