Knowledge Combination to Learn Rotated Detection Without Rotated Annotation

This is the implementation of the paper: Knowledge Combination to Learn Rotated Detection Without Rotated Annotation. The paper is accepted by CVPR2023. We build KCR with mmcv, mmdet and mmrotate. Paper Link: https://openaccess.thecvf.com/content/CVPR2023/papers/Zhu_Knowledge_Combination_To_Learn_Rotated_Detection_Without_Rotated_Annotation_CVPR_2023_paper.pdf

Getting Started

Install the main packages:

• torch==1.9.1
• mmcv==1.5.3 Then use the following command to setup KCR:

bash setup

For the dataset setup, please follow coco. We convert all the dataset to the form of coco instance detection.

Training

cd kcrtools
python train_kcr_2d.py

Please provide the rotated dataset such as coco and axis-aligned dataset in the config file.

Inference

cd kcrtools
python test_rotated_2d.py