✨DRNet✨

Introduction

This is the official implementation of DRNet, which is implemented on mmrotate

Results and models

FAIR1M-1.0 12epochs: Download

FAIR1M-1.0 12epochs & Multi-Scale & Rotation Augmentation: Download

FAIR1M-1.0

Backbone	lr schd	ms	rr	Batch Size	mAP	Download
R50-FPN	1x	-	-	2	40.87	Download
R50-FPN	1x	√	√	2	45.82	Download

Installation

Please refer to install.md for installation.

How to Start DRNet

Dataset Prepare

Because the annotation format of FAIR1M is XML, it first needs to be converted to DOTA format

python tools/data/fair1m/fair_to_dota.py {your_fair_dir} {dota_save_dir}

Next, split images to the right size.

python tools/data/dota/split/img_split.py tools/data/dota/split/split_configs/ss_train.json

python tools/data/dota/split/img_split.py tools/data/dota/split/split_configs/ss_test.json

Train

python tools/train.py configs/drnet/drnet_r50_fpn_1x_fair1m_le90.py

Test

python tools/test.py configs/drnet/drnet_r50_fpn_1x_fair1m_le90.py {ckpt_dir} --format-only --eval-options submission_dir={save_results_dir} nprco=1

Convert the result format for online submission testing

python tools/data/fair1m/dota_to_fair.py {save_results_dir} {online_sub_dir} {you_fair_images_dir}

License

This project is released under the Apache 2.0 license.