[Paper Review] Objects detection for remote sensing images based on polar coordinates

Lin Zhou, Haoran Wei+4 more|arXiv (Cornell University)|Jan 9, 2020

Remote-Sensing Image Classification46 references35 citations

TL;DR

The paper introduces P-RSDet, an anchor-free detector that models oriented objects in polar coordinates, achieving competitive accuracy with fewer regression parameters and a novel Polar Ring Area Loss.

ABSTRACT

Arbitrary-oriented object detection is an important task in the field of remote sensing object detection. Existing studies have shown that the polar coordinate system has obvious advantages in dealing with the problem of rotating object modeling, that is, using fewer parameters to achieve more accurate rotating object detection. However, present state-of-the-art detectors based on deep learning are all modeled in Cartesian coordinates. In this article, we introduce the polar coordinate system to the deep learning detector for the first time, and propose an anchor free Polar Remote Sensing Object Detector (P-RSDet), which can achieve competitive detection accuracy via uses simpler object representation model and less regression parameters. In P-RSDet method, arbitrary-oriented object detection can be achieved by predicting the center point and regressing one polar radius and two polar angles. Besides, in order to express the geometric constraint relationship between the polar radius and the polar angle, a Polar Ring Area Loss function is proposed to improve the prediction accuracy of the corner position. Experiments on DOTA, UCAS-AOD and NWPU VHR-10 datasets show that our P-RSDet achieves state-of-the-art performances with simpler model and less regression parameters.

Motivation & Objective

• Motivate and address the challenge of oriented object detection in remote sensing images.
• Propose a new anchor-free detector that uses polar coordinates to model oriented bounding boxes.
• Reducing regression parameters while maintaining or improving detection accuracy.
• Introduce a Polar Ring Area Loss to better capture the geometric relationship between radius and angle in polar space.
• Demonstrate effectiveness across multiple remote sensing datasets (DOTA, UCAS-AOD, NWPU VHR-10).

Proposed method

• Represent oriented objects with four corners using a polar coordinate formulation centered at the object pole point.
• Output four maps: one pole-point heatmap and three regression maps for ρ (polar radius), θ1, and θ2.
• Extract pole points via a CornerNet-inspired keypoint approach with a revised extraction method to reduce missed detections in dense remote sensing scenes.
• Regress ρ, θ1, θ2 with a Smooth-L1 loss and introduce Polar Ring Area Loss to couple radius and angle through a ring-area term.
• Transform polar coordinates back to Cartesian coordinates during inference for standard bounding box representation.
• Train end-to-end in an anchor-free framework, leveraging a high-resolution encoder (ResNet-101, with alternatives tested).
• Evaluate on DOTA, UCAS-AOD, and NWPU VHR-10, comparing with oriented and horizontal detectors.

Experimental results

Research questions

• RQ1 Can an anchor-free detector in polar coordinates achieve competitive or superior performance for oriented object detection in remote sensing images?
• RQ2 Does incorporating Polar Ring Area Loss improve the accuracy of radius-angle regression in polar-space representations?
• RQ3 How does the Pole Point extraction strategy affect detection accuracy in dense remote sensing scenes?
• RQ4 Is the polar-coordinate, anchor-free approach effective across datasets with oriented and horizontal annotation formats?

Key findings

• P-RSDet achieves state-of-the-art performance on DOTA with oriented boxes (mAP 72.30).
• P-RSDet achieves competitive performance on UCAS-AOD with oriented boxes (mAP 90.03) and strong results with horizontal boxes (mAP 90.24).
• P-RSDet achieves strong results on NWPU VHR-10 with horizontal boxes (mAP 90.80).
• Compared to other anchor-based Cartesian detectors, P-RSDet offers competitive accuracy with a simpler model and fewer regression parameters.
• A Polar Ring Area Loss improves mAP by about 1.85 percentage points on UCAS-AOD oriented detection (with Polar Ring Area Loss).
• Extracting pole points via the proposed extreme-point/connected-domain method reduces missed detections versus top-K keypoint extraction.