Unsupervised Salient Object Detection with Spectral Cluster Voting

TL;DR

This paper introduces a novel unsupervised salient object detection method that uses spectral clustering on self-supervised features, combined with a voting mechanism, to generate pseudo groundtruth masks for training a new detector called SelfMask.

Contribution

It revisits spectral clustering for pixel grouping, proposes a voting scheme for mask selection from multiple self-supervised models, and trains a superior unsupervised SOD detector called SelfMask.

Findings

Outperforms prior unsupervised SOD methods on benchmarks.

Effective spectral clustering-based mask proposal generation.

SelfMask achieves state-of-the-art results without supervision.

Abstract

In this paper, we tackle the challenging task of unsupervised salient object detection (SOD) by leveraging spectral clustering on self-supervised features. We make the following contributions: (i) We revisit spectral clustering and demonstrate its potential to group the pixels of salient objects; (ii) Given mask proposals from multiple applications of spectral clustering on image features computed from various self-supervised models, e.g., MoCov2, SwAV, DINO, we propose a simple but effective winner-takes-all voting mechanism for selecting the salient masks, leveraging object priors based on framing and distinctiveness; (iii) Using the selected object segmentation as pseudo groundtruth masks, we train a salient object detector, dubbed SelfMask, which outperforms prior approaches on three unsupervised SOD benchmarks. Code is publicly available at https://github.com/NoelShin/selfmask.