This is something a little different to open source GPS, but Brian reached out over Twitter and suggested you might be group interested in a new open source ag initiative: Weed-AI
It’s a new platform we developed for the upload/download/standardisation of annotated weeds imagery for the development of machine learning based weed recognition models. It was developed out of a fairly severe bottleneck in the availability of publicly available annotated images to train up these algorithms for anyone to develop their own spot sprayers (or whatever site-specific weed control). We’re hoping to build up a community of people (very much looking at the successful community here as a great example!) that would contribute images and also site improvements.
The Github repository for the site has all the code and is another opportunity for involvement, though less of a priority currently than annotated images.
If you’d like to contribute imagery that would be fantastic and it can be as small/large a dataset as you are happy to do. The most important points are:
- images must be collected in fairly consistent and reportable format
- images must be annotated (acceptable format is COCO, but we have converters for PASCAL VOC/and mask annotations)
- the AgContext and Metadata files need to be completed
That’s about it! There’s more information about the site and specifics about acceptable formats on the About and WeedCOCO pages of Weed-AI if you’re after that. But would love to hear any questions, suggestions or feedback on the process. We also have an OpenWeedLocator (OWL) that is very close to being ready to launch but we’re just finalising an associated paper, so stayed tuned for another open source hardware/software implementation for site specific weed control in the very near future.