Skip to content

Latest commit

 

History

History
80 lines (55 loc) · 2.41 KB

README.md

File metadata and controls

80 lines (55 loc) · 2.41 KB

Our GPUs: 2 * TeslaV100 (16GB)

Prerequisites

install environment following

conda create -n open-mmlab python=3.7 -y
conda activate open-mmlab
# install pytorch
# conda install -c pytorch pytorch torchvision -y
conda install pytorch==1.9.0 torchvision==0.10.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch
# install mmcv
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.9.0/index.html

# install mmdetection
pip uninstall pycocotools   # sometimes need to source deactivate before, for 
pip install -r requirements/build.txt
pip install -v -e . --user  # or try "python setup.py develop" if get still got pycocotools error
chmod +x tools/dist_train.sh
conda install scikit-image  # or pip install scikit-image

Prepare Dataset DIOR/DOTA

  1. Download Dataset

  2. Generate 'obb+pt' Format:

    • Follow the following scripts to convert the dataset format:
      • tools_data_trans/test_dior2dota_obbpt_viaobb.py (for DIOR-R)
      • tools_data_trans/test_dota2dota_obbpt_viaobb.py (for DOTA)

  3. Generate COCO Format:

    • Follow the following script to convert the dataset format:
      • tools_data_trans/test_dota2coco_P2B_obb-pt.py

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/pointobb_r50_fpn_2x_dior.py --work-dir xxx/work_dir/pointobb_r50_fpn_2x_dior --cfg-options evaluation.save_result_file='xxx/work_dir/pointobb_r50_fpn_2x_dior_dist/pseudo_obb_result.json'

# DOTA
# python tools/train.py --config configs2/pointobb/pointobb_r50_fpn_2x_dota10.py --work-dir xxx/work_dir/pointobb_r50_fpn_2x_dota --cfg-options evaluation.save_result_file='xxx/work_dir/pointobb_r50_fpn_2x_dota_dist/pseudo_obb_result.json'

## train with multiple GPUs
sh train_p_dist.sh
  1. Inference

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)