Installation

Requirements:

• Python3
• PyTorch 1.3.1 with CUDA support
• torchvision 0.4.2
• mmcv 0.2.14
• pycocotools 2.0.1

Step-by-step installation

conda create -n ltm python=3.7
conda activate ltm
conda install pytorch==1.3.1 torchvision==0.4.2 cudatoolkit=${CUDA_VERSION} -c pytorch
pip install mmcv===0.2.14 pycocotools===2.0.1

git clone https://github.com/zhangxiaosong18/LTM.git
cd LTM
python setup.py develop

Usage

Prepare dataset

You will need to download and prepare the COCO dataset. It is recommended to symlink the dataset root to path_to_ltm/data.

cd path_to_ltm
mkdir data
ln -s /path_to_coco data/coco

Training example

cd path_to_ltm
export GPU_NUM=8
tools/dist_train.sh configs/LTM/ltm_af_r50_fpn_1x.py ${GPU_NUM} --gpus ${GPU_NUM} --autoscale-lr --validate

For more details, please refer to the mmdetection README.md

Citations

Please consider citing our paper in your publications if the project helps your research.

@article{zhang2021ltm,
  author={X. {Zhang} and F. {Wan} and C. {Liu} and X. {Ji} and Q. {Ye}},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Learning to Match Anchors for Visual Object Detection}, 
  year={2021},
  doi={10.1109/TPAMI.2021.3050494}
}