Path Aggregation Network for Instance Segmentation

Path Aggregation Network (PANet) targets how information propagates in proposal-based instance segmentation frameworks. It adds bottom-up path augmentation to push accurate localization signals from lower layers up through the feature hierarchy, shortening the information path between low-level and topmost features; adaptive feature pooling that links each proposal's feature grid to all feature levels so useful information propagates directly to the proposal subnetworks; and a complementary branch that captures different views of each proposal to improve mask prediction.

These improvements are simple to implement and add only subtle extra computational overhead, yet they proved highly effective: PANet reached 1st place in the COCO 2017 Challenge instance segmentation task and 2nd place in object detection without large-batch training, and it was also state of the art on the MVD and Cityscapes datasets.