Rethinking Alignment and Uniformity in Unsupervised Image Semantic Segmentation

Zhang，Daoan1,3; Li，Chenming1; Li，Haoquan1; Huang，Wenjian1; Huang，Lingyun3; Zhang，Jianguo1,2

2023-06-27

Proceedings of the 37th AAAI Conference on Artificial Intelligence, AAAI 2023

37

11192-11200

Unsupervised image semantic segmentation (UISS) aims to match low-level visual features with semantic-level representations without outer supervision. In this paper, we address the critical properties from the view of feature alignments and feature uniformity for UISS models. We also make a comparison between UISS and image-wise representation learning. Based on the analysis, we argue that the existing MI-based methods in UISS suffer from representation collapse. By this, we proposed a robust network called Semantic Attention Network (SAN), in which a new module Semantic Attention (SEAT) is proposed to generate pixel-wise and semantic features dynamically. Experimental results on multiple semantic segmentation benchmarks show that our unsupervised segmentation framework specializes in catching semantic representations, which outperforms all the unpretrained and even several pretrained methods.

First ; Corresponding

English

National Key Research and Development Program of China[2021YFF1200800];National Natural Science Foundation of China[62276121];

2-s2.0-85166347616

Scopus

1.Southern University of Science and Technology,China
2.Peng Cheng Laboratory,China
3.Ping An Technology (Shenzhen) Co.,Ltd.,China

Zhang，Daoan,Li，Chenming,Li，Haoquan,et al. Rethinking Alignment and Uniformity in Unsupervised Image Semantic Segmentation[C],2023:11192-11200.