/decoupled_mask_rcnn

Primary LanguagePythonMIT LicenseMIT

Decoupled Mask R-CNN for Solving Foreground Confusion

Installation

  • Ubuntu 16.04
  • CUDA 11.1
conda create -n decoupled_mask_rcnn python=3.9 -y
source activate decoupled_mask_rcnn

conda install pytorch=1.10 cudatoolkit=11.1 torchvision -c pytorch

git clone https://github.com/stormstoutzhang/decoupled_mask_rcnn.git
python -m pip install -e decoupled_mask_rcnn

Train

./tools/train_net.py --num-gpus 8 --config-file configs/COCO-InstanceSegmentation/decoupled_mask_rcnn_R_50_FPN_1x.yaml

Inference

./tools/train_net.py --num-gpus 8 --config-file configs/COCO-InstanceSegmentation/decoupled_mask_rcnn_R_50_FPN_1x.yaml --eval-only MODEL.WEIGHTS /path/to/checkpoint_file

Model Zoo

Method Backbone lr Sched mask AP download
Decoupled Mask R-CNN ResNet-50-FPN 1x 36.5 model 
Extraction Code: IVIP

Acknowledgement

Thanks Detectron2 team for the wonderful open source project!