Tracking animal pose (that is, identifying the positions foo their major joints) is a major frontier in neuroscience. When combined with neural recordings, pose tracking allows for identifying the relationship between neural activity and movement, and decision-making inferred from movement. OpenMonkeyStudio is a system designed to allow tracking of rhesus macaques in large freely moving environments.
Tracking monkeys is at least an order of magnitude more difficult than tracking mice, flies, and worms. Monkeys are, basically, large furry blobs; they don’t have clear body segmentations. And their movements are much richer and more complex. For these reasons, out of the box systems don’t work with monkeys.
The major innovation of our OpenMonkeyStudio is how it tackles the annotation problem. Deep learning systems aren’t very good at generalization. They can replicate things they have seen before or things that are kind fo similar to what they have seen. So the important thing is giving them a sufficiently large training set. We ideally want to have about a million annotated images. That would cost about $10 million and we don’t have that kind of money. So we use several cool tricks, which we describe in our paper, to augment a small dataset and turn it into a large one. Doing that works very well, and results in a system that can track one or even two interacting monkeys.