Mask R-CNN

An extension of Faster R-CNN that generates pixel-level segmentation masks for every detected object instance in an image.

Kaiming He and the Facebook AI Research (FAIR) team built Mask R-CNN to solve instance segmentation by adding a mask branch to the Faster R-CNN architecture. The system uses RoIAlign (a quantization-free layer) to preserve exact spatial locations: this architectural shift improves mask accuracy by 10 to 50 percent over older methods. It won three tracks in the COCO 2016 challenge and processes images at 5 frames per second on an Nvidia Tesla M40. Engineers deploy it for high-precision vision: isolating individual tumors in medical imaging or mapping specific obstacles for autonomous drones.

https://arxiv.org/abs/1703.06870