Magic results but R-CNN (and other similar architectures with large number of forward passes per frame) is pretty darn slow, which severely limits its applicability.
It really depends on the application. If you’re comparing to classical instance segmentation methods like a distance-based watershed, Mask RCNN (e.g. using the Keras-RCNN package) has nearly identical performance. However, if you’re looking for real-time performance, you’ll need to look elsewhere (and likely reframe your problem as an object detection rather than instance segmentation problem).