One recent network architecture that went some way in making this problem easier to solve by providing a simple, flexible model architecture is Mask R-CNN. Published in 2017, the architecture looks at extending the capabilities of faster R-CNN:
It takes the existing faster R-CNN model and tries to solve the instance-segmentation problem by adding a branch to the model that is responsible for predicting object masks in parallel to the classification and bounding-box regression head. This architecture proved effective and achieved top honors in all the COCO challenges at the time of publishing.