# Seminar Learning for Click-Level Weakly Supervised Semantic Segmentation

**Authors:** Hongjun Chen, Jinbao Wang, Hong Cai Chen, Xiantong Zhen, Feng Zheng, Rongrong Ji, Ling Shao  
**arXiv:** [2108.13393](https://arxiv.org/abs/2108.13393)  
**Date:** 2021-08-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.13393v1/

## Key points

- Burden of annotation is a major obstacle in semantic segmentation
- Click-level annotations offer a better balance between supervision and annotation cost
- Authors propose a seminar learning approach for weakly supervised semantic segmentation
- Seminar learning consists of two modules: teacher-student and student-student
- Teacher network guides training of the student network using exponential moving average
- Student-student module introduces heterogeneous pseudo-labels for knowledge transfer among students
- Experimental results show effectiveness of seminar learning with state-of-the-art performance on Pascal VOC 2012 dataset
- Seminar learning addresses annotation burden by utilizing click-level supervision and knowledge exchange among networks

## Summary

The burden of annotation has become a major obstacle in the field of semantic segmentation. To address this issue, researchers have turned their attention to approaches based on click-level annotations, which offer a better balance between supervision and annotation cost. In this paper titled "Seminar Learning for Click-Level Weakly Supervised Semantic Segmentation," authors Hongjun Chen, Jinbao Wang, Hong Cai Chen, Xiantong Zhen, Feng Zheng, Rongrong Ji, and Ling Shao propose a novel learning paradigm that leverages the knowledge from different networks to compensate for the limited information provided by click-level annotations. The proposed seminar learning approach mimics a seminar setting and consists of two modules: teacher-student and student-student. In the teacher-student module, a teacher network based on exponential moving average is employed to guide the training of the student network. This allows the student network to learn from skillful teachers and benefit from their expertise. The student-student module introduces heterogeneous pseudo-labels to facilitate knowledge transfer among students and enhance their individual performance. Experimental results demonstrate the effectiveness of seminar learning in semantic segmentation tasks. The proposed approach achieves state-of-the-art performance with an mIOU (mean Intersection over Union) score of 72.51% on the Pascal VOC 2012 dataset. This surpasses previous methods by a significant margin of up to 16.88%. Overall, seminar learning presents a promising solution for addressing the annotation burden in semantic segmentation by effectively utilizing click-level supervision and leveraging knowledge exchange among networks. The achieved performance improvements highlight its potential for advancing the field and opening new avenues for research in weakly supervised semantic segmentation techniques.