Great video! I've been wanting to get into AI and needed an idiots guide to create a dataset. At the moment I'm using Frigate for the detection and "whosatmyfeeder" for the classification, so being able to create my own dataset will be invaluable. Do you know where I can get a test lite dataset of UK birds before I start to create my own? I have a feeling this is going to be a big rabbit hole that I'm just falling into :-)
Great that you're enjoying the project!!! I tried an off the shelf dataset in part 1 and it really didn't work too well. The big problem is simply that the bird forms such a small portion of the overall image so I found it finding lots of false positives like shadows being identified as crows. So I'd suggest you take the leap and make your own. Even with a very modest dataset of 100-200 images, you can get some really good results. It's laborious and time consuming but you get a far better end result. Let me know how you get on!
@@jeffs_piinthesky Still haven't got past getting pyenv installed and working (after many hours) but it's sometime the journey thats of most value and not the results :-) Hopefully the results will be good!
Seeing the AI identify even a few birds in less than perfect conditions was very cool
Indeed! I was pretty thrilled with the outcome!!! Should be really good when I have the model refined with more sample data!!
Great video! I've been wanting to get into AI and needed an idiots guide to create a dataset. At the moment I'm using Frigate for the detection and "whosatmyfeeder" for the classification, so being able to create my own dataset will be invaluable. Do you know where I can get a test lite dataset of UK birds before I start to create my own? I have a feeling this is going to be a big rabbit hole that I'm just falling into :-)
Great that you're enjoying the project!!! I tried an off the shelf dataset in part 1 and it really didn't work too well. The big problem is simply that the bird forms such a small portion of the overall image so I found it finding lots of false positives like shadows being identified as crows. So I'd suggest you take the leap and make your own. Even with a very modest dataset of 100-200 images, you can get some really good results. It's laborious and time consuming but you get a far better end result. Let me know how you get on!
@@jeffs_piinthesky Still haven't got past getting pyenv installed and working (after many hours) but it's sometime the journey thats of most value and not the results :-) Hopefully the results will be good!