It’s a general solution to the problem of pose tracking macaques.
-
-
Prikaži ovu nit
-
It turns out macaque tracking is a LOT harder than tracking worms, flies, or birds. They are big furry monochrome blobs with very little articulation, they display a lot of poses, and they are basically acrobats.pic.twitter.com/U0DhHwZUY0
Prikaži ovu nit -
Unlike humans, where there exist few large (million plus) annotated databases, nothing comparable works for monkeys. And the human networks don’t generalize - we tried! In our view, the annotation problem is THE major barrier to macaque tracking.pic.twitter.com/RtYrTiBvqU
Prikaži ovu nit -
So we built a large pen in which rhesus macaques could move freely. We placed 62 machine vision cameras around the pen at precise positions and angles. The walls are chroma key green to aid in segmentation.pic.twitter.com/5kS5QCuG3i
Prikaži ovu nit -
We filmed a large number of videos and worked with Hive AI to annotate about 10,000 images.
Prikaži ovu nit -
The paper goes over the cool tricks we developed to overcome the annotation problem. The result is a library of effectively 200,000 images, which we will share with the public when the paper is accepted.pic.twitter.com/fkaBT9vUzK
Prikaži ovu nit -
We will also share the model itself. It can do robust low error tracking of 13 joints at 30 Hz.pic.twitter.com/2zw1i6F6Ks
Prikaži ovu nit -
-
With this we can use clustering algorithms (here, UMAP), to automatically segment behavior into meaningful actions like sitting, walking, and climbing.pic.twitter.com/ZxAQu6D2qe
Prikaži ovu nit -
We have started doing behavioral training with our subjects in this cage, and we can get them to do some basic foraging tasks.https://www.nature.com/articles/s41598-019-51442-z …
Prikaži ovu nit -
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.