Object Discovery and Localization Via Structural Contrast
Shuanglu Dai, Pengyu Su, Hong Man · 2018
This paper explores structural contrast of object proposal for unsupervised object discovery. In this paper, an object proposal is considered as distinctive to the others in an image based on the structural uniqueness and distribution. A symmetric Kullback-Leibler (KL) distance metric is proposed to measure the difference between the distributions of structured edges. The computation of this distance metric is fully unsupervised, without any image level annotation or any labeling of visual semantics. Inspired by the pixel-level saliency filtering, a contrast-based saliency assignment is introduced to model the uniqueness and distribution distinctiveness of an object proposal as structural contrast. Visual objects can be discovered and localized by the proposals with the highest ranked structural contrast. Experiments evaluated on PascaiVOC07 demonstrate that the proposed unsupervised method outperforms the state-of-the-art unsupervised and weakly-supervised methods.