I am not sure if this is the right place to ask, so it’d be great if someone can point me to a more dedicated forum in case I am off topic here.
I am a professional software developer (C, Python) and while I did try out a few things with PyTorch and Diffusers - I am not an ML engineer, so I am looking for someone with ML expertise who’d be interested to team up for a non commercial open source project. I can do quite a lot around application development, but I clearly lack the required ML knowledge. Perhaps researches who are focused on AI may find this idea interesting for some university work, then you could see me as a research software engineer supporting you.
So, here’s what I have in mind: I would like to create an application which would be able to identify and detect individual fallow deer in a video. To be more precise, each individual fallow deer have their own distinct fur pattern, as far as I can tell from real life observations, they keep it throughout their entire life. So we’re not talking about a classification “this is a deer”, but rather - like human face recognition mapping to a person: this is “Doe 1”, “Doe 2” and so on.
Caveat: patterns on the left and right side of the animal are not equal. Also, once they get their winter fur the patterns become invisible, but they do reemerge when they change back to the summer fur. For simplicity, I would ignore the winter case and only focus on the summer footage where the fur patterns are clearly visible.
Each new pattern that is seen for the first time needs to go to a database, so that later, in another video, the same deer could be identified again. The application would then output a list of unique deer individuals which appeared in the footage.
On a side note: I have gigabytes and countless hours worth of fallow deer videos, which could be processed and used for training.
Just thinking out loud, my approach would be to use something like SegmentAnything to first fish out the generic “deer” masks from the video and then as a second step look at the pattern. I am however not quite sure how to approach the pattern task.
The forum won’t let me put two images in the post, so I am uploading the one where I marked the most distinct part in the fur of this particular doe.
Of course the algorithm would take more into account, I highlighted only the part which I used myself to “manually” identify this particular doe.
If we think broader, I would assume that models could be trained to identify individuals for different kinds of animals, provided that they do have distinct fur patterns. Practical use could include counting the population of wildlife, so perhaps the result of this work could be useful to others, for instance wildlife scientists.
I am not sure about the scope and effort, but my feeling tells me, that this would be a complex and challenging project.
If anyone finds this idea interesting and would like to collaborate - please reach out!