RLS volunteers are all too familiar with the experience of spending hours after a dive to ensure that all the fish and invertebrate IDs are accurate to species level. This human dedication to getting IDs right contributes to the high quality of RLS data, but with advancements in computer vision, the task of fish identification can be partly automated: In theory, we should be able to upload our photos, note the site we surveyed, and get a list of probable IDs.
Well, I built a tool to do just that, and it is time to share it with the world: "Ichthy-what" is now available for initial testing on the RLS website. Using it should be self-explanatory, but here’s a quick demo video of its initial version:
https://drive.google.com/file/d/1hpfRY26ZQXHzhYpAIP-3ShsqpuCSXfth/view
At this point, the tool is in an early experimental phase. This is partly because I’m not entirely confident about the stability of the software package behind the tool. In addition, don’t expect it to be perfectly accurate: In my tests on RLS fish images, the correct answer was one of the top-ten matches in over 70% of cases. From anecdotal evidence on my own images, it often gets the right answer when the image is cropped to show a single fish. In any case, the current accuracy level should be good enough for the tool to be useful.
Feedback is very welcome and can be sent via the RLS contact form or directly to me. I’m especially interested in hearing whether there are any glitches other than inaccurate matches, but any other ideas for improvement would also be great. Some ideas that are already in the pipeline are: extending the tool to support invertebrates, choosing an area on the map rather than specifying coordinates, and making it easier to narrow down results by family and other classification levels. However, there are no guarantees when these will be implemented, as it is a hobby project. Still, I hope that RLS volunteers and other fish lovers will find the tool useful!
