I've been using DeepLabCut for physiology research, and it's really quite amazing. Manually labeling ~200 frames yields excellent results given decent contrast and lighting. Although the underlying technology itself is probably nothing new, it's packaged with a GUI that makes it very friendly to researchers across fields. The main challenge/obstacle, in my opinion, is hardware - training ResNets still requires hours of powerful GPU time, although DeepLabCut is also capable of using MobileNets with only a small decrease in performance.