Finding stable salient contours

A review of automated image understanding within 3D baggage computed tomography security screening

“…The proposed approach is composed of four steps: 1) sieve the input volume to four logarithmically-spaced scales; 2) compute four channel volumes; 3) label the channel volumes using Sedgewick's connected component analysis [86] and 4) merge the labelled channel volumes into a single labelled volume. Merging is performed by determining the similarities between the density histograms for each labelled object in each channel volume using the KolmogorovSmirnov (K-S) test [55], which computes the probability that the histograms have been drawn from the same distribution. The specific strengths and weaknesses of the approach are not addressed in any great detail [23].…”

Salient Regions from Scale-Space Trees

“…This observation was the original motivation for simplifying the tree. However, independently [7], it was realized that contours could provide stable regions known as MSERs and in [8] it was shown that these regions are a subset of the nodes of opening and closing trees. Since it is known that M and N -sieves are more robust than open/close trees [15] it has been postulated that contours extracted from M -and N -trees should be as stable, and possibly more stable, than MSERs which are known to be among the best performing regions in terms of repeatability.…”

“…Since it is known that M and N -sieves are more robust than open/close trees [15] it has been postulated that contours extracted from M -and N -trees should be as stable, and possibly more stable, than MSERs which are known to be among the best performing regions in terms of repeatability. In [8] such regions are termed Stable Salient Contours or SSCs.…”

“…MSERs are grayscale contours in the image that are likely to remain unchanged under typical transformations such as illuminant variation, perspective transformation, zoom and other distortions. In [8] it is shown that MSERs are a special case of a more general saliency detector known as Stable Salient Contours (SSCs) which we describe later.…”

“…In Section 2 we briefly describe the saliency algorithms and derive an improved version of SSCs. Section 3 evaluates the reliability of the new SSC method using the methodology in [6] and in Section 4 we test several algorithms on an image retrieval task and give reasons for the difference in performance between entropy-based methods such as [1] and difference-based methods such as SIFT [3] and SSCs [8]. Section 5 discusses further properties of SSCs and their hierarchical representation which we call a saliency tree.…”