# WegFormer: Transformers for Weakly Supervised Semantic Segmentation

**Authors:** Chunmeng Liu, Enze Xie, Wenjia Wang, Wenhai Wang, Guangyao Li, Ping Luo  
**arXiv:** [2203.08421](https://arxiv.org/abs/2203.08421)  
**Date:** 2022-03-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2203.08421v1/

## Key points

- Paper introduces WegFormer, a framework for weakly supervised semantic segmentation (WSSS)
- WegFormer leverages Transformers to improve WSSS
- CNNs have limitations in capturing global context information for WSSS
- WegFormer incorporates Vision Transformer (ViT) as a classifier to generate high-quality pseudo segmentation masks
- Three tailored components introduced: Deep Taylor Decomposition (DTD), soft erasing module, and efficient potential object mining (EPOM) technique
- WegFormer achieves state-of-the-art performance on PASCAL VOC dataset with mIoU score of 70.5%
- Demonstrates effectiveness of incorporating Transformers into WSSS frameworks
- WegFormer provides a simple yet powerful approach to WSSS using Vision Transformer and tailored components
- Code for WegFormer will be made available for further research and applications in WSSS.

## Summary

The paper titled "WegFormer: Transformers for Weakly Supervised Semantic Segmentation" introduces a novel framework called WegFormer that leverages the power of Transformers to improve weakly supervised semantic segmentation (WSSS). The use of convolutional neural networks (CNNs) has greatly advanced WSSS, but their limited effective receptive field often hinders them from capturing global context information, resulting in sub-optimal segmentation results. To address this issue, the authors propose incorporating Vision Transformer (ViT) as a classifier within the WegFormer framework to generate high-quality pseudo segmentation masks. Additionally, three tailored components are introduced to enhance the performance of WegFormer. The first component is Deep Taylor Decomposition (DTD), which generates attention maps to capture relevant contextual information and guide the segmentation process. The second component is a soft erasing module that smooths out the attention maps, improving their quality and reducing noise. Lastly, an efficient potential object mining (EPOM) technique is employed to filter out noisy activations in the background, further enhancing segmentation accuracy. Notably, WegFormer achieves state-of-the-art performance on the PASCAL VOC dataset with a mean intersection over union (mIoU) score of 70.5%. This outperforms previous methods and demonstrates the effectiveness of incorporating Transformers into WSSS frameworks. The authors believe that WegFormer opens up new possibilities for leveraging Transformers in weakly supervised semantic segmentation tasks. Overall, this work presents a simple yet powerful approach to WSSS by utilizing Vision Transformer and introducing tailored components within a Transformer-based framework. The code for WegFormer will be made available to facilitate further research and applications in weakly supervised semantic segmentation.