Over the past 3 years I’ve made significant progress on ShitSpotter. I’ve collected over 5000 images (31 GB). Recently I’ve started annotating images and training models.
I annotated 400 images before I started training my first models. It was subpar. I annotated up to 900 images, and started seeing loss curves like this:
With corresponding training curve being less well behaved (900 annotated images is still probably not enough).
With my old model I had to cherry pick to find results that looked semi-ok. With the new model, good results are far more frequent.
However, the larger problem is still unsolved:
But the network is still learning, and separating leafs from poop will be harder. However, the easy cases are doing well, up to some interesting cases:
With more data, and more compute I have no doubt this will start working well.
Erotemic's Blog 