# Learning RoI Transformer for Detecting Oriented Objects in Aerial Images

**Authors:** Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, Qikai Lu  
**arXiv:** [1812.00155](https://arxiv.org/abs/1812.00155)  
**Date:** 2018-12-01  
**Summary page:** https://www.summarizepaper.com/en/paper/1812.00155v1/

## Key points

- Object detection in aerial images presents unique challenges due to bird's eye view perspective, complex backgrounds, and diverse appearances of objects
- Traditional methods relying on horizontal proposals can lead to misalignments between Region of Interests (RoIs) and actual objects, affecting classification confidence and localization accuracy
- RoI Transformer introduces a novel approach with a Rotated RoI (RRoI) learner and a Rotated Position Sensitive RoI Align (RPS-RoI-Align) module to address challenges in detecting densely packed objects in aerial images
- The proposed RoI Transformer is lightweight, easily integrable into detectors for oriented object detection, and achieves state-of-the-art performance on challenging aerial datasets such as DOTA and HRSC2016 while maintaining detection speed

## Summary

, , , , 

In the field of computer vision, object detection in aerial images presents a unique set of challenges due to the bird's eye view perspective, complex backgrounds, and diverse appearances of objects. Detecting densely packed objects in aerial images is particularly difficult as traditional methods relying on horizontal proposals often result in misalignments between Region of Interests (RoIs) and actual objects. This can lead to inaccuracies in classification confidence and localization accuracy. While rotated anchors have been utilized to address this issue, they come with a drawback of increasing computational complexity by multiplying the number of anchors. To tackle these challenges, this paper introduces a novel approach called RoI Transformer. The RoI Transformer consists of two key components: a Rotated RoI (RRoI) learner that transforms Horizontal RoIs (HRoIs) into Rotated RoIs (RRoIs), and a Rotated Position Sensitive RoI Align (RPS-RoI-Align) module that extracts rotation-invariant features from RRoIs to enhance subsequent classification and regression tasks. Notably, the proposed RoI Transformer is lightweight and easily integrable into detectors for oriented object detection. Experimental results demonstrate that the implementation of the RoI Transformer achieves state-of-the-art performance on challenging aerial datasets such as DOTA and HRSC2016 while maintaining detection speed. Furthermore, when compared to deformable Position Sensitive RoI pooling with oriented bounding-box annotations, the RoI Transformer surpasses in performance. The flexibility and effectiveness of the proposed approach are validated through extensive experiments, showcasing its potential for integration with various detector architectures to significantly enhance object detection performances in aerial imagery applications.