Training Dataset?

[Erotemic]

I'm wondering if the data used to train your model is available or if you are interested in making it available?

I'm working on a similar project where I am attempting to detect dog poop from phone images. Details are in the README of https://github.com/Erotemic/shitspotter

I'm wondering if images of poop itself are visible in your images? If they are I'm wondering if my current models can find them.

[calebolson123]

Hey @Erotemic, cool project. I might be able to dig up some images I used to train the model but:

1. there's very few needed for the type of model used, maybe 30-50
2. I would bet maybe 1/4 of them have poop visible in frame. The model is not finding poop, rather a dog pooping. And assumes the dog isn't fake pooping lol

[Erotemic]

That's interesting you were able to make it work with so few examples. My guess is that because the domain is fixed to your backyard the model was able to take some shortcuts, but it also helps to start with a good pretrained model. I need to try a transfer learning experiment on my data to see how well that works.

I'd be more than happy to dig through 50 images to find the 10-or-so that I could test my algorithms on. I would also be happy to host them on my IPFS server if you wanted to make the data more widely available.

Does your dog fake poop? I've never seen my dogs "assume the position" unless they were trying to drop one. There are cases where they will try, but are constipated or something, so there's no poop afterwords, but if you are training a pose-recognition model, I still think it's correct to detect that.

Not sure what the false positive rate of your system is, but I imagine you could ensemble the pose detector with a poop detector to get a cleaner signal.

[zoltanszalontay]

Hi Caleb, great project! Have you created augmented images to increase the training set?