Lightweight clustering for semantic segmentation

Paper Title:

A Lightweight Clustering Framework for Unsupervised Semantic Segmentation

Published on:

30 November 2023

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Yau Shing Jonathan Cheung,

Xi Chen,

Lihe Yang,

Hengshuang Zhao

Bullets

Key Details

•

Attention features from self-supervised vision transformers have strong foreground/background differences

•

Cluster attention features at dataset, category, and image levels

•

Ensure consistency between clustering levels to extract high-quality pseudo-masks

•

Refine masks and perform class assignment using vision transformer outputs

•

Achieves state-of-the-art segmentation performance with low computation cost

Explore the topics in this paper

pseudo labeling

self-supervised learning

semantic segmentation

unsupervised learning

vision transformers

AI generated summary

Lightweight clustering for semantic segmentation

This paper proposes a lightweight clustering framework to perform semantic segmentation without labels. It utilizes attention features from self-supervised vision transformers, which have strong foreground/background differences. These features are clustered into groups at the dataset, category, and image levels. Consistency across levels extracts high-quality binary pseudo-masks separating foreground/background. The masks are refined and class assignment uses vision transformer outputs. This achieves state-of-the-art performance on PASCAL VOC and COCO with low computation cost.