✨DRNet✨

Introduction

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

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

Please refer to install.md for installation.

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

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

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}

This project is released under the Apache 2.0 license.