Multi-level refinement enriched feature pyramid network for object detection

作者：

Highlights：

• Propose a joint solution for scale imbalance at features level and class imbalance

• Handle features critical aspects i.e., abstractness, coarseness, and cardinality

• A separate cascaded refinement scheme is used for class imbalance solution.

• It's not only refined the anchors but also refine features to produce better results.

• MS COCO test dev: 320 × 320 image size = 40.6AP

• MS COCO test dev: 512 × 512 image size = 45.7AP

摘要

•Propose a joint solution for scale imbalance at features level and class imbalance•Handle features critical aspects i.e., abstractness, coarseness, and cardinality•A separate cascaded refinement scheme is used for class imbalance solution.•It's not only refined the anchors but also refine features to produce better results.•MS COCO test dev: 320 × 320 image size = 40.6AP•MS COCO test dev: 512 × 512 image size = 45.7AP

论文关键词：CNN,Object detection,Chained parallel pooling,Computer vision,Feature pyramid

论文评审过程：Received 13 July 2021, Revised 21 August 2021, Accepted 26 August 2021, Available online 28 August 2021, Version of Record 16 September 2021.

论文官网地址：https://doi.org/10.1016/j.imavis.2021.104287