Learning High-Quality Bounding Box for Rotated Object Detection via Rotated Cascade Region Proposal Network

Abstract

Existing two-stage detectors usually generate oriented proposals based on heuristically defined anchors with different scales, angles, and aspect ratios. This scheme usually suffers from severe memory-consuming and redundant computation. Additionally, misalignment between rotated proposals and horizontally aligned convolutional features exists when using a conventional Region Proposal Network (RPN), which leads to the inconsistency of classification confidence and positioning accuracy. To tackle these problems, we propose a Rotated Cascade Region Proposal Network (RCRPN), which effectively reduces memory usage and improves the quality of proposals through multi-stage refinement. Specifically, instead of using multiple anchors with predefined scales and aspect ratios, a single anchor per location is adopted in the first stage of RCRPN, and coarse proposals are generated in a horizontal convolution manner, this stage effectively takes the advantage of gliding vertex method to adapt the rotated bounding box. In the second stage, by taking the coarse proposals and image feature map as input, adaptive align-convolution is applied to learn the sampled rotated features guided by the coarse proposals, finally generating high-quality proposals for the downstream tasks. Extensive experiments demonstrate that our method can achieve better performance than baseline algorithm Oriented R-CNN on two commonly used datasets including DOTA and HRSC2016 for oriented object detection.