This paper proposes Path Aggregation Network (PANet) for instance segmentation, aiming to improve information flow in proposal-based frameworks. PANet enhances feature hierarchy with accurate localization signals from lower layers through bottom-up path augmentation, shortening the information path between lower layers and topmost features. Adaptive feature pooling is introduced to link feature grids and all feature levels, enabling useful information to propagate directly to proposal subnetworks. A complementary branch captures different views for each proposal to further improve mask prediction. These improvements are simple to implement with minimal computational overhead. PANet achieves first place in COCO 2017 Instance Segmentation and second place in Object Detection without large-batch training, and is state-of-the-art on MVD and Cityscapes. The framework includes bottom-up path augmentation, adaptive feature pooling, and fully-connected fusion. Bottom-up path augmentation improves feature propagation by adding clean lateral connections. Adaptive feature pooling allows each proposal to access information from all levels. Fully-connected fusion enhances mask prediction by combining predictions from different views. PANet achieves state-of-the-art performance on multiple datasets, including COCO, Cityscapes, and MVD. The framework is effective for both instance segmentation and object detection, and is implemented with minimal computational overhead. The results show that PANet significantly improves performance in both tasks.This paper proposes Path Aggregation Network (PANet) for instance segmentation, aiming to improve information flow in proposal-based frameworks. PANet enhances feature hierarchy with accurate localization signals from lower layers through bottom-up path augmentation, shortening the information path between lower layers and topmost features. Adaptive feature pooling is introduced to link feature grids and all feature levels, enabling useful information to propagate directly to proposal subnetworks. A complementary branch captures different views for each proposal to further improve mask prediction. These improvements are simple to implement with minimal computational overhead. PANet achieves first place in COCO 2017 Instance Segmentation and second place in Object Detection without large-batch training, and is state-of-the-art on MVD and Cityscapes. The framework includes bottom-up path augmentation, adaptive feature pooling, and fully-connected fusion. Bottom-up path augmentation improves feature propagation by adding clean lateral connections. Adaptive feature pooling allows each proposal to access information from all levels. Fully-connected fusion enhances mask prediction by combining predictions from different views. PANet achieves state-of-the-art performance on multiple datasets, including COCO, Cityscapes, and MVD. The framework is effective for both instance segmentation and object detection, and is implemented with minimal computational overhead. The results show that PANet significantly improves performance in both tasks.