Awesome

Discriminative Region Suppression for Weakly-Supervised Semantic Segmentation (AAAI 2021)

Official pytorch implementation of our paper: Discriminative Region Suppression for Weakly-Supervised Semantic Segmentation [Paper], Beomyoung Kim, Sangeun Han, and Junmo Kim, AAAI 2021

We propose the discriminative region suppression (DRS) module that is a simple yet effective method to expand object activation regions. DRS suppresses the attention on discriminative regions and spreads it to adjacent non-discriminative regions, generating dense localization maps.

[2021.06.10] we support DeepLab-V3 segmentation network!

Setup

1. Dataset Preparing

   • Download PASCAL VOC 2012
   • you can obtain SegmentationClassAug/ [download] (augmented with SBD dataset).
   • Download saliency maps used for background cues.

   # dataset structure
   VOC2012/
       --- Annotations/
       --- ImageSets/
       --- JPEGImages/
       --- SegmentationClassAug/
       --- saliency_map/
       --- refined_pseudo_segmentation_labels/
   

2. Requirements pip install -r requirements.txt

Training & Pseudo Segmentation Labels Generation

• step1 : training the classifier with DRS modules
• step2 : training the refinement network for the localization maps refinement
• step3 : pseudo segmentation labels generation

# all-in-one
bash run.sh 

Training the DeepLab-V2 using pseudo labels

We adopt the DeepLab-V2 pytorch implementation from https://github.com/kazuto1011/deeplab-pytorch.

• According to the DeepLab-V2 pytorch implementation , we requires an initial weights [download].

cd DeepLab-V2-PyTorch/

# motify the dataset path (DATASET.ROOT)
vi configs/voc12.yaml

# 1. training the DeepLab-V2 using pseudo labels
bash train.sh

# 2. evaluation the DeepLab-V2
bash eval.sh

Training the DeepLab-V3+ using pseudo labels

We adopt the DeepLab-V3+ pytorch implementation from https://github.com/VainF/DeepLabV3Plus-Pytorch.

Note that DeepLab-V2 suffers from the small batch issue, therefore, they utilize COCO pretrained weight and freeze batch-normalization layers; DeepLab-V2 without COCO-pretrained weight cannot reproduce their performance even in fully-supervised setting.

In contrast, DeepLab-V3 does not require the COCO-pretrained weight due to the recent large memory GPUs and Synchronized BatchNorm. We argue that the choice of DeepLab-V3 network is more reasonable and better to measure the quality of pseudo labels.

cd DeepLabV3Plus-Pytorch/

# training & evaluation the DeepLab-V3+ using pseudo labels
vi run.sh # modify the dataset path --data_root
bash run.sh

• Note that the pretrained weight path ./DeepLab-V2-Pytorch/data/models/Deeplabv2_pseudo_segmentation_labels/deeplabv2_resnet101_msc/train_cls/checkpoint_final.pth

Citation

We hope that you find this work useful. If you would like to acknowledge us, please, use the following citation:

@inproceedings{kim2021discriminative,
    title={Discriminative Region Suppression for Weakly-Supervised Semantic Segmentation},
    author={Kim, Beomyoung and Han, Sangeun and Kim, Junmo},
    year={2021},
    booktitle={AAAI Conference on Artificial Intelligence},
}