class-supervised-instance-segmentation

Prerequisite

• Python 3.8, PyTorch 1.8.1, and cuda 11.1
• VOC dataset (10582 images for training; 1449 images for validation)
• NVIDIA GPU(such as:Two 1080 GPU or one 3090 GPU)

Usage

Install python dependencies：

• conda install pytorch==1.8.1 torchvision==0.9.1 torchaudio==0.8.1 cudatoolkit=11.1
• pip install yacs opencv-python chainercv scikit-learn imageio
• pip install git+https://github.com/lucasb-eyer/pydensecrf.git

Download pascal voc dataset

• data link: VOC.
• pass word:vf2j

The downloaded data structure is shown:

data_root/
   --- VOC2012/
       --- Annotations/
       --- ImageSet/
       --- JPEGImages/
       --- SegmentationObject/
       --- WSSS_maps/

Train or make your own script

sh run_MM.sh

1.Train classification network to generate instance-aware localization which saved in "Peak_points5" folder

we also provide our own trained instance localization.

pass word：cgqh

2.Train displacement field and eval instance segmentation

• You can either mannually edit the file, or specify commandline arguments.