Here’s the first “in the wild” results from a model I’ve been working on for a while now, having received help from the Roboflow team, specifically Jacob and Mohamed.
This is a classifier project to detect plants on some property in rural Colorado.
For now, it is separated into two separate models, one detects a type of weed called “mullein”, which has 2 growth stages. One is a ground-cover plant, which spreads over whatever ground it can find. The next is to grow up to 1M(3 feet) high, very quickly, then it dries out and spreads its seeds everywhere. Its considered a weed and a fire hazard there.
The next model detects prickly pear cactus, which is attractive to wildlife and can stabilize hillsides(among other attributes). So, one worth having around, one to keep under control.
Both are based on 200 or so images taken “in the wild” on a DSLR. I outsourced the labeling to a 3rd party then uploaded the datasets to Roboflow with pre-processing and augments provided by the Starter bundle . That kicked each dataset up to about 750 images, split about 50%/25%/25% train/validate/test.
It took about 12 training credits to get the models trained to this point, including a number of missteps on my part.
I started working on this during the Pandemic lockdown period as a skills-building project to enhance the offerings of my consulting business and finally got back out to the field to test it this weekend. The model was deployed to an OAK-D Lite
attached to my laptop, running a Python script I wrote implementing the roboflowoak SDK.
Still to be done:
- Continue working on the cactus model. It still needs help.
- Deploy them onto a rover or drone that can mark the GPS locations of the plants for either tending or elimination.
- Further testing to figure specific cases where the model(s) succeed and fail and roll that into the next edition.
