320 — Beef Cattle Instance Segmentation Using Fully Convolutional Neural Network

I guess I read about cows a lot now?

Cattle video segmentation is hard: In this case, ten beef heifers were recorded via CCTV over a two week stay in cattle pens. But cows change their posture and shape when going from standing to sitting position; cattle all look… well… not to be that guy, but… pretty much all the same. So that’s hard too. And they walk in front of each other and they walk behind things like structural supports: Hard. And because the animals are in a farm enclosure, the roof has large gaps in it to allow natural light in. But that means that animals are very unevenly lit, since the artificial lighting is, in the authors’ words, “quite bad.”

To address these problems, the authors designed a simple framework dubbed “MaskSplitter” which avoids the generation of bounding boxes and instead directly trains upon determining the difference between “good” and “bad” masks, given a large training set corpus. This means that the net can run quickly and with little overhead, which is perfect for, say, an AWS DeepLens sitting on a shelf in a barn.