Extended quantum cuts for unsupervised salient object extraction

Abstract

In this manuscript, an unsupervised salient object extraction algorithm is proposed for RGB and RGB-Depth images. Saliency estimation is formulated as a foreground detection problem. To this end, Quantum-Cuts (QCUT), a recently proposed spectral foreground detection method is investigated and extended to formulate the saliency estimation problem more efficiently. The contributions of this work are as follows: (1) a new proof for QCUT from spectral graph theory point of view is provided, (2) a detailed analysis of QCUT and comparison to well-known graph clustering methods are conducted, (3) QCUT is utilized in a multiresolution framework, (4) a novel affinity matrix construction scheme is proposed for better encoding of saliency cues into the graph representation and (5) a multispectral analysis for a richer set of salient object proposals is investigated. With the above improvements, we propose Extended Quantum Cuts, which consistently achieves an exquisite performance over all benchmark saliency detection datasets, containing around 18 k images in total. Finally, the proposed approach also outperforms the state-of-the-art on a recently announced RGB-Depth saliency dataset.