Mask R-CNN

Mask R-CNN presents a conceptually simple and flexible framework for object instance segmentation, a task requiring both detecting objects in an image and producing a high-quality segmentation mask for each detected instance. The core method extends the existing Faster R-CNN object detector by adding a new branch, running in parallel with the existing bounding-box recognition branch, that predicts an object mask for each instance. This design keeps Mask R-CNN simple to train and adds only a small computational overhead relative to Faster R-CNN, while running at 5 frames per second, and it generalizes easily to other tasks, such as estimating human poses within the same framework.

The method achieves top results across all three tracks of the COCO challenge suite, instance segmentation, bounding-box object detection, and person keypoint detection, outperforming every existing single-model entry, including the winners of the COCO 2016 challenge, without relying on additional tricks. This mattered because it delivered a single, general, and easy-to-train architecture that set a new standard across multiple instance-level recognition tasks simultaneously, and the authors intended it as a solid baseline to ease future research.