Automatic Extraction of Closed Contours Bounding Salient Objects: New Algorithms and Evaluation Methods

The problem under consideration in this dissertation is achieving salient object segmentation of natural images by means of probabilistic contour grouping. The goal is to extract the simple closed contour bounding the salient object in a given image. The method proposed here falls in the Contour Grouping category, searching for the optimal grouping of boundary entities to form an object contour.

Our first contribution is to provide both a ground truth dataset and a performance measure for empirical evaluation of salient object segmentation methods. Our Salient Object Dataset (SOD) provides ground truth boundaries of salient objects perceived by humans in natural images. We also psychophysically evaluated 5 distinct performance measures that have been used in the literature and showed that a measure based upon minimal contour mappings is most sensitive to shape irregularities and most consistent with human judgements. In fact, the Contour Mapping measure is as predictive of human judgements as human subjects are of each other.

Contour grouping methods often rely on Gestalt cues locally defined on pairs of oriented features. Accurate integration of these local cues with global cues is a challenge. A second major contribution of this dissertation is a novel, effective method for combining local and global cues.

A third major contribution in this dissertation is a novel method based on Principal Component Analysis for promoting diversity among contour hypotheses, leading to substantial improvements in grouping performance.

To further improve the performance, a multiscale implementation of this method has been studied. A fourth contribution in this dissertation is studying the effect of the multiscale prior on the performance and analysing the method for combining the results obtained in different resolutions.

Our final contribution is comparing the performance of univariate distribution models for local cues used by our method with the use of a multivariate mixture model for their joint distribution. We obtain slight improvement by the mixture models.

The proposed method has been evaluated and compared with four other state-of-the-art grouping methods, showing considerably better performance on the SOD ground truth dataset.