Skip to content

[ECCV2024] Mitigating Background Shift in Class-Incremental Semantic Segmentation

License

Notifications You must be signed in to change notification settings

RoadoneP/ECCV2024_MBS

Repository files navigation

Mitigating Background Shift in Class-Incremental Semantic Segmentation

Gilhan Park, WonJun Moon, SuBeen Lee, Tae-Young Kim, and Jae-Pil Heo
Sungkyunkwan University

Paper Conference
🥇PWC 🥇PWC
🥇PWC 🥇PWC
🥇PWC 🥇PWC
🥇PWC 🥇PWC
🥇PWC 🥇PWC
🥇PWC 🥇PWC

main_figure

Updates & News

  • Our new paper, "Mitigating Background Shift in Class-Incremental Semantic Segmentation" be posted on Arxiv!!!! The title and code are currently available.
  • Our paper has been accepted at ECCV 2024.

Abtract

Class-Incremental Semantic Segmentation (CISS) aims to learn new classes without forgetting the old ones, using only the labels of the new classes. To achieve this, two popular strategies are employed: 1) pseudo-labeling and knowledge distillation to preserve prior knowledge; and 2) background weight transfer, which leverages the broad coverage of background in learning new classes by transferring background weight to the new class classifier. However, the first strategy heavily relies on the old model in detecting old classes while undetected pixels are regarded as the background, thereby leading to the background shift towards the old classes (i.e., misclassification of old class as background). Additionally, in the case of the second approach, initializing the new class classifier with background knowledge triggers a similar background shift issue, but towards the new classes. To address these issues, we propose a background-class separation framework for CISS. To begin with, selective pseudo-labeling and adaptive feature distillation are to distill only trustworthy past knowledge. On the other hand, we encourage the separation between the background and new classes with a novel orthogonal objective along with label-guided output distillation. Our state-of-the-art results validate the effectiveness of these proposed methods.

Requirements

Conda environment settings:

conda create -n MBS python=3.8
conda activate MBS

You need to install the following libraries:

pip install torch==1.10.1+cu111 torchvision==0.11.2+cu111 torchaudio==0.10.1 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt
pip install -U openmim
mim install mmcv

Datasets

data_root/
    ├── VOCdevkit
    │   └── VOC2012/
    │       ├── Annotations/
    │       ├── ImageSet/
    │       ├── JPEGImages/
    │       └── SegmentationClassAug/
    ├── ADEChallengeData2016
    │   ├── annotations
    │   │   ├── training
    │   │   └── validation
    │   └── images
    │       ├── training
    │       └── validation

PASCAL VOC 2012

Download PASCAL VOC2012 devkit (train/val data)

sh download_voc.sh
  • download_voc.sh
cd data_root

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
tar -xf VOCtrainval_11-May-2012.tar
wget http://cs.jhu.edu/~cxliu/data/SegmentationClassAug.zip
wget http://cs.jhu.edu/~cxliu/data/SegmentationClassAug_Visualization.zip
wget http://cs.jhu.edu/~cxliu/data/list.zip
rm VOCtrainval_11-May-2012.tar

unzip SegmentationClassAug.zip
unzip SegmentationClassAug_Visualization.zip
unzip list.zip

mv SegmentationClassAug ./VOCdevkit/VOC2012/
mv SegmentationClassAug_Visualization ./VOCdevkit/VOC2012/
mv list ./VOCdevkit/VOC2012/

rm list.zip
rm SegmentationClassAug_Visualization.zip
rm SegmentationClassAug.zip

ADE20k

python download_ade20k.py

Training & Test

sh main.sh

We provide a training script main.sh. Detailed training argumnets are as follows:

python -m torch.distributed.launch --nproc_per_node={num_gpu} --master_port={port} main.py --config ./configs/voc.yaml --log {your_log_name}

Qualitative Results

Qualitative_Result

BibTex

If you find the repository or the paper useful, please use the following entry for citation.

@article{park2024mitigating,
  title={Mitigating Background Shift in Class-Incremental Semantic Segmentation},
  author={Park, Gilhan and Moon, WonJun and Lee, SuBeen and Kim, Tae-Young and Heo, Jae-Pil},
  journal={arXiv preprint arXiv:2407.11859},
  year={2024}
}

Contributors and Contact

If there are any questions, feel free to contact the authors: Gilhan Park (a01152a@gmail.com).