PENet: Pre-Enhanced Network for Object Detection and Instance Segmentation

Multi-level feature fusion has always been regarded as a powerful structure that can improve the performance of models for instance segmentation. In this work, we propose a novel proposal-based Pre-Enhanced Network (PENet) performing a bottom-up feature fusion between backbone and neck of the network, achieving a more adequate information propagation. Our main contribution is the new Pre-Enhanced Module (PEM) that utilizes feature maps generated by stages of backbone for multi-level fusion, which performs feature extraction and fusion on adjacent levels from bottom to top. Our PEM will equip feature maps with more sufficient information from the bottom, obtaining a more effective performance for object detection and instance segmentation. Evaluation and comparative experiments of models are conducted on Cityscapes dataset. Our PENet outperforms its baseline model obtaining absolute mAP gains of 1.3% for object detection and 0.9% for instance segmentation, and also achieves improvements compared with other corresponding models based on the same baseline. We hope our work will provide a new view of model structure modification for object detection and instance segmentation. Code will be made published.