This is the source code for our paper Non-Salient Region Object Mining for Weakly Supervised Semantic Segmentation.

The architecture of our proposed approach is as follows

• Install PyTorch 1.3 with Python 3 and CUDA 10.0

• Clone this repo

git clone https://github.com/NUST-Machine-Intelligence-Laboratory/nsrom.git

• Download PASCAL VOC 2012

• Download our trained model checkpoint_70.4.pth and put it in the segmentation/data/models folder

cd segmentation
python main.py test --config-path configs/voc12.yaml --model-path data/models/checkpoint_70.4.pth
python main.py crf --config-path configs/voc12.yaml

cd classification

• Train the classification model and obtain CAM

./train.sh 
./test.sh 

• Train the integral model and obtain the improved OA-CAM

./train_iam.sh
./test_iam.sh

• Download the pre-computed saliency map and generate the pseudo labels

python gen_label.py

• Download the COCO pre-trained parameters and put it in the segmentation/data/models folder. Then train the segmentation model with pseudo labels . (You need to set the path for pseudo labels.)

cd segmentation
python main.py train --config-path configs/voc12.yaml

• Get the initial prediction for trainaug images. (You need to go to configs/voc12.yaml and set DATASET.SPLIT.VAL = train_aug. Set it back to val when testing)

python trainaug_pred.py test --config-path configs/voc12.yaml \
    --model-path data/models/voc12/deeplabv2_resnet101_msc/train_aug/checkpoint_final.pth
python trainaug_pred.py crf --config-path configs/voc12.yaml -j 10

• Generate the masked pseudo labels

cd classification
python gen_masked_label.py

• Train the segmentation model with the masked pseudo labels. (You need to set the path for masked pseudo labels.)

cd segmentation
python main.py train --config-path configs/voc12.yaml

This codebase is heavily borrowed from OAA-PyTorch and deeplab-pytorch.