gcl's Introduction

GCL

This project hosts the official implementation for the paper:

Gradient Calibration Loss for Fast and Accurate Oriented Bounding Box Regression [URL][PDF][BibTex]

( accepted by IEEE Transactions on Geoscience and Remote Sensing).

Abstract

In this paper**,** we demonstrate several drawbacks of rotated IoU loss through both experiments and theoretical derivation. And then, a Gradient Calibration Loss (GCL**)** is propoosed to optimize the rotated IoU loss via gradient analysis and correction**.** GCL can be easily introduced into the existing rotation detectors to achieve performance gains without extra inference overhead**.**

Setup

conda create -n gcl python=3.7 -y
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html


cd BboxToolkit
pip install -v -e .
cd ..

pip install mmcv-full==1.3.9 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.9.0/index.html

cd mmdet/ops/point_justify
python setup.py develop
cd ../../..

pip install -r requirements/build.txt
pip install mmpycocotools
python setup.py develop

Training

Creat config files.
Data preparation viia

cd BboxToolkit/tools/ and python img_split.py --base_json split_configs/dota1_0/ss_trainval.json
Run sh train.sh or sh night.sh.

Inference & Testing

Run sh demo.sh and sh test.sh.

Visualizations

Performance

Results on DOTA-v1.0
Results on DOTA-v1.5

Citation

If you find our work or code useful in your research, please consider citing:

@ARTICLE{ming2024gradient,
author={Ming, Qi and Miao, Lingjuan and Zhou, Zhiqiang and Song, Junjie and Pizurica, Aleksandra},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={Gradient Calibration Loss for Fast and Accurate Oriented Bounding Box Regression},
year={2024},
volume={62},
number={},
pages={1-15},
keywords={Object detection;Convergence;Remote sensing;Detectors;Training;Feature extraction;Proposals;Convolutional neural network;gradient analysis;loss function;oriented object detection},
doi={10.1109/TGRS.2024.3367294}}

Feel free to contact me if there are any questions.

gcl's People

Contributors

Stargazers

gcl's Issues

GCL损失中调用gcl_loss方法时方法里ious未定义问题

作者你好，这段GCL损失代码的else中loss = 1 - ious里ious在使用前未进行定义，所以是否应该将上面的dloss, ious, union = diou_loss(pred, target) 提到if前或这句else直接pass，亦或是有其他修改方法呢？

Replacement method

Hello author, I am reporting error RuntimeError: shape '[4, -1, 5]' is invalid for input of size 4718592 when I am training dota dataset using master_rcnn_orpn_r50_fpn_gcl_1x_dota10.py as configuration file;

Recommend Projects