CW-BASS: Confidence-Weighted Boundary-Aware Learning for Semi-Supervised Semantic Segmentation

Official PyTorch implementation of our IJCNN 2025 paper:

CW-BASS: Confidence-Weighted Boundary-Aware Learning for Semi-Supervised Semantic Segmentation Ebenezer Tarubinga, Jenifer Kalafatovich, Seong-Whan Lee — Pattern Recognition & Machine Learning Lab, Korea University

📄 Paper (IEEE Xplore) · 📝 arXiv · 🌐 Project Page

---

Table of Contents

• Abstract
• Key Contributions
• Results
• Installation
• Data Preparation
• Training
• Pseudo-Labeling & Retraining
• Repository Structure
• Acknowledgements
• Citation
• License

Abstract

Semi-supervised semantic segmentation (SSSS) aims to improve segmentation performance by utilizing large amounts of unlabeled data with limited labeled samples. Existing methods often suffer from coupling, where over-reliance on initial labeled data leads to suboptimal learning; confirmation bias, where incorrect predictions reinforce themselves repeatedly; and boundary blur caused by limited boundary-awareness and ambiguous edge cues. To address these issues, we propose CW-BASS, a novel framework for SSSS. We assign confidence weights to pseudo-labels to mitigate the impact of incorrect predictions, and leverage boundary-delineation techniques that remain underutilized in SSSS. Extensive experiments on Pascal VOC 2012 and Cityscapes demonstrate that CW-BASS achieves state-of-the-art performance. Notably, CW-BASS achieves a 65.9% mIoU on Cityscapes under a challenging 1/30 split (100 images), highlighting its effectiveness in limited-label settings.

Key Contributions

1. Confidence-Weighted Loss — reduces coupling by adjusting pseudo-label influence based on predicted confidence scores, downweighting unreliable predictions.
2. Dynamic Thresholding — mitigates confirmation bias with an adaptive mechanism that filters pseudo-labels based on model performance.
3. Boundary-Aware Module — tackles boundary blur using explicit (Sobel-based) boundary supervision to refine segmentation near object edges.
4. Confidence Decay — reduces label noise through a progressive strategy that refines pseudo-label confidence during training.

Results

All experiments use DeepLabV3+ with a ResNet backbone. mIoU (%) on the standard SSSS splits:

Setting	Split	Labeled images	mIoU
Cityscapes	1/30	100 (3.3%)	65.9
Cityscapes	1/4	—	78.4
Pascal VOC 2012	1/4	—	76.2

See the project page and the paper for the full comparison against prior methods and all label partitions.

Qualitative comparison — CW-BASS produces sharper object boundaries.

Installation

git clone https://github.com/psychofict/CW-BASS.git
cd CW-BASS
pip install -r requirements.txt

Requirements: Python ≥ 3.8, PyTorch ≥ 1.10 with a CUDA-capable GPU (multi-GPU training is supported via DataParallel). See requirements.txt for the full list.

Data Preparation

Pre-trained backbones

ResNet-50 | DeepLabv2-ResNet-101

Datasets

Pascal JPEGImages | Pascal SegmentationClass | Cityscapes leftImg8bit | Cityscapes gtFine

File organization

├── ./pretrained
│   ├── resnet50.pth
│   └── deeplabv2_resnet101_coco_pretrained.pth
│
├── [Your Pascal Path]
│   ├── JPEGImages
│   └── SegmentationClass
│
└── [Your Cityscapes Path]
    ├── leftImg8bit
    └── gtFine

The labeled/unlabeled ID splits for every partition ship with the repo under dataset/splits/ — no extra setup required.

Training

export semi_setting='pascal/1_8/split_0'

CUDA_VISIBLE_DEVICES=0,1 python -W ignore main.py \
  --dataset pascal --data-root [Your Pascal Path] \
  --batch-size 16 --backbone resnet50 --model deeplabv3plus \
  --labeled-id-path dataset/splits/$semi_setting/labeled.txt \
  --unlabeled-id-path dataset/splits/$semi_setting/unlabeled.txt \
  --pseudo-mask-path outdir/pseudo_masks/$semi_setting \
  --save-path outdir/models/$semi_setting

Dataset-specific defaults (epochs, learning rate, crop size) are applied automatically; the CW-BASS hyper-parameters can be overridden on the command line:

Flag	Default	Description
--gamma	1.0	Confidence-weighting exponent in the weighted CE loss
--base-threshold	0.6	Base value for the dynamic confidence threshold
--beta	0.5	Sensitivity of the dynamic threshold to mean confidence
--decay-factor	0.9	Multiplicative confidence-decay factor
--use-confidence-decay	off	Enable progressive confidence decay
--config	none	Optional YAML (see configs/) to override defaults

The best checkpoint (by validation mIoU) is written to --save-path. Use --resume <ckpt> to continue training from a saved checkpoint.

Pseudo-Labeling & Retraining

main.py also provides ST++-style utilities to select reliable unlabeled images, generate pseudo-masks, and retrain on the labeled + pseudo-labeled union (see select_reliable, label, and retrain_on_pseudo_labeled_data).

Repository Structure

CW-BASS/
├── main.py                 # Training / pseudo-labeling entry point
├── utils.py                # mIoU metric, param counting, color maps
├── configs/                # Per-dataset YAML configs
├── dataset/
│   ├── semi.py             # SemiDataset (labeled / unlabeled / val)
│   ├── transform.py        # Augmentations (crop, flip, resize, blur, cutout)
│   └── splits/             # Labeled / unlabeled ID lists per partition
├── model/
│   ├── backbone/           # ResNet backbones
│   └── semseg/             # DeepLabV3+, DeepLabV2, PSPNet
├── entropy_cityscapes.py   # Analysis: entropy maps
├── entropy_map_pascal.py
├── gt_vs_pred_cityscapes.py# Analysis: GT vs prediction visualizations
├── gt_vs_pred_pascal.py
└── docs/                   # Project page (GitHub Pages)

Acknowledgements

The image partitions are borrowed from Context-Aware-Consistency and PseudoSeg. Part of the training hyper-parameters and network structures are adapted from ST++ and PyTorch-Encoding.

• ST++: https://github.com/LiheYoung/ST-PlusPlus
• Context-Aware-Consistency: https://github.com/dvlab-research/Context-Aware-Consistency
• PyTorch-Encoding: https://github.com/zhanghang1989/PyTorch-Encoding
• OpenSelfSup: https://github.com/open-mmlab/OpenSelfSup

Citation

If you find this project useful, please consider citing:

@inproceedings{tarubinga2025cwbass,
  author={Tarubinga, Ebenezer and Kalafatovich, Jenifer and Lee, Seong-Whan},
  booktitle={2025 International Joint Conference on Neural Networks (IJCNN)},
  title={CW-BASS: Confidence-Weighted Boundary-Aware Learning for Semi-Supervised Semantic Segmentation},
  year={2025},
  pages={1-8},
  doi={10.1109/IJCNN64981.2025.11227871}
}

License

Released under the MIT License.