Skip to content

README.md

Latest commit

 

History

History
45 lines (28 loc) · 1.34 KB

README.md

File metadata and controls

45 lines (28 loc) · 1.34 KB

Our GPUs: 2 * A100 (80GB)

Prerequisites & Prepare Dataset

Please follow our conference version for details: https://github.com/Luo-Z13/pointobb

Train/Inference

  1. Train

To train the model, follow these steps:

cd PointOBB
## train with single GPU, note adjust learning rate or batch size accordingly
# DIOR
python tools/train.py --config configs2/pointobb/pointobbv3_r50_fpn_2x_dior_two_stage.py --work-dir ./work_dir/pointobbv3_dior/ --cfg-options evaluation.save_result_file='./work_dir/pointobbv3_dior/pseudo_obb_result.json'

# DOTA
# python tools/train.py --config configs2/pointobb/pointobbv3_r50_fpn_2x_dota_two_stage.py --work-dir ./work_dir/pointobbv3_dior/ --cfg-options evaluation.save_result_file='./work_dir/pointobbv3_dior/pseudo_obb_result.json'

......
  1. Inference (two-stage)

To inference (generate pseudo obb label), follow these steps:

# obtain COCO format pseudo label for the training set 
# (在训练集上推理,从单点生成旋转框的伪标签)
sh test_p.sh
# convert COCO format to DOTA format 
# (将伪标签从COCO格式转换为DOTA格式)
sh tools_cocorbox2dota.sh
# train standard oriented object detectors 
# (使用伪标签训练mmrotate里的标准旋转检测器)
# Please use algorithms in mmrotate (https://github.com/open-mmlab/mmrotate)