Oriented RepPoints for Aerial Object Detection

Wentong Li, Yijie Chen, Kaixuan Hu, Jianke Zhu* (Arxiv)

Based on OrientedRepPoints detector, the 2nd and 3rd Places are achieved on the Task 2 and Task 1 respectively in the “2021 challenge of Learning to Understand Aerial Images(LUAI)”. The detailed codes and introductions about it, please refer to this repository and 知乎.

Update

The code for MMRotate is available now.
RepPoints + our APAA can obtain +2.5AP (36.3 to 38.8) improvement with R-50 on COCO dataset for general object detection.

Installation

Please refer to for installation and dataset preparation.

Getting Started

This repo is based on . Please see for the basic usage.

Results and Models

The results on DOTA test set are shown in the table below. More detailed results please see the paper.

Model	Backbone	data aug(HSV+Rotation)	mAP	model	log
OrientedReppoints	R-50		75.97	model	log
OrientedReppoints	R-101		76.52	model	log
OrientedReppoints	Swin-Tiny	√	78.11	model	log

Note:

The pretrained model--swin_tiny_patch4_window7_224 of Swin-Tiny for pytorch1.4.0 is here.
We recommend to use our demo configs with 4 GPUs.
The results are performed on the original DOTA images with 1024x1024 patches.
The scale jitter is employed during training. More details see the paper.

The mAOE results on DOTA val set are shown in the table below.

Model	Backbone	mAOE	Download
OrientedReppoints	R-50	5.93°	model

Note：Orientation error evaluation (mAOE) is calculated on the val subset(only train subset for training).

Visual results

The visualization code for oriented bounding boxes and learning points is .

Oriented bounding box

Citation

@inproceeding{orientedreppoints,
	title="Oriented RepPoints for Aerial Object Detection.",
	author="Wentong {Li}, Yijie {Chen}, Kaixuan {Hu}, Jianke {Zhu}.",
	journal="The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)",
	year="2022"
}