MM-FPN: Multi-path and Multi-scale Feature Pyramid Network for Object Detection
Sheng Dong, Jiaxin Zhang, Zehui Qu · 2021
Small and multi-scale objects are always dilemmas for object detection. However, small objects may disappear and cannot be detected because it is arduous to differentiate information from a small part of the original image. To alleviate the issue, an image pyramid is utilized to build a feature pyramid to detect across a range of scales. Instead, we combine image pyramid and feature pyramid with a Contextually Enhanced Module (CEM) to extract contextual information. Furthermore, we propose Unidirectional Bottom-up Connections (UBC) to extract more distinct features. A novel Multi-path and Multi-scale Feature Pyramid Network (MM-FPN) is proposed to improve the performance of both small-sized and large-sized objects. Experiments and ablation studies are performed on PASCAL VOC, which surpass most of the existing competitive single-stage and two-stage methods.