Published on:
25 April 2024
Primary Category:
Computer Vision and Pattern Recognition
Paper Authors:
Oliver Hahn,
Nikita Araslanov,
Simone Schaub-Meyer,
Stefan Roth
Presents PriMaPs, a method to derive mask proposals directly from SSL features using PCA
Proposals guide stochastic EM to fit class prototypes for segmentation
Reaches competitive performance across datasets and backbones
Boosts state-of-the-art when applied orthogonally
Principal Mask Proposals for Unsupervised Semantic Segmentation
This paper presents a method called PriMaPs to decompose images into semantically meaningful masks using principal components of self-supervised image features. These masks serve as proposals to guide an expectation-maximization algorithm, PriMaPs-EM, to realize unsupervised semantic segmentation by fitting class prototypes. Despite simplicity, PriMaPs-EM leads to competitive accuracy across datasets and backbone models. Notably, it boosts results when combined with state-of-the-art methods, suggesting it uses representation properties orthogonal to prior work.
No comments yet, be the first to start the conversation...
Sign up to comment on this paper