Adaptive region-aware feature enhancement for object detection

作者：

Highlights：

• We verify the importance of position-sensitive features and the differences among feature levels in FPN and propose AR-FPN module to extract more distinguishable features for each level.

• We propose AR-RFF module to aggregate complementary features from regions surrounding RoIs, making it sensitive to larger scale variations and introduce a weighted regression strategy.

• We evaluate our method on MS COCO dataset through comprehensive experiments, and it brings consistent and substantial improvements upon various frameworks and backbone networks.

摘要

•We verify the importance of position-sensitive features and the differences among feature levels in FPN and propose AR-FPN module to extract more distinguishable features for each level.•We propose AR-RFF module to aggregate complementary features from regions surrounding RoIs, making it sensitive to larger scale variations and introduce a weighted regression strategy.•We evaluate our method on MS COCO dataset through comprehensive experiments, and it brings consistent and substantial improvements upon various frameworks and backbone networks.

论文关键词：Object detection,Feature enhancement,Adaptive region-aware FPN,Adaptive region-aware RoI feature fusion

论文评审过程：Received 3 July 2021, Revised 10 November 2021, Accepted 16 November 2021, Available online 18 December 2021, Version of Record 30 December 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.108437