Supervoxel Classification Forests for Estimating Pairwise Image Correspondences

Abstract

Abstract. This paper proposes a general method for establishing pairwise correspondences, which is a fundamental problem in image analysis. The method consists of over-segmenting a pair of images into supervoxels. A forest classifier is then trained on one of the images, the source, by using supervoxel indices as voxelwise class labels. Applying the forest on the other image, the target, yields a supervoxel labelling which is then regularized using majority voting within the boundaries of the target's supervoxels. This yields semi-dense correspondences in a fully automatic, efficient and robust manner. The advantage of our approach is that no prior information or manual annotations are required, making it suitable as a general initialisation component for various medical imaging tasks that require coarse correspondences, such as, atlas/patch-based segmentation, registration, and atlas construction. Our approach is evaluated on a set of 150 abdominal CT images. In this dataset we use manual organ segmentations for quantitative evaluation. In particular, the quality of the correspondences is determined in a label propagation setting. Comparison to other state-of-the-art methods demonstrate the potential of supervoxel classification forests for estimating image correspondences.

Unsupervised learning-based long-term superpixel tracking

“…The superpixel decomposition is performed using the Simple Linear Iterative Clustering (SLIC) algorithm [11] which aggregates neighboring pixels p q based on spatial and intensity proximity criteria. Forward superpixel matching from I f to I s (f < s) consists in automatically learning a matching function h f,s that maps each superpixel f i ∈ F of I f to a given superpixel s j ∈ S of I s [17] such that: Backward matching from I s to I f can be similarly considered by estimating h s,f mapping each superpixel s j ∈ S to a given superpixel f i ∈ F. In what follows, learning-based superpixel matching is described in forward from I f to I s .…”

“…The overall learning-based superpixel matching strategy, illustrated Fig.1, is carried out in an unsupervised manner. To give a powerful representation of global context, RF or kNN is considered with new pixel-wise context-rich features extended from greyscale [24,17,25] to multi-channel RGB and described Sect.2.3.…”

“…LetĪ w q (p q , c) be the average intensity on a local box of size w centered on p q located at x q for channel c ∈ {r, g, b}. Pixel-wise context-rich features θ(p q ) = {θ m (p q )} m∈{0,...,Ka−1} assigned to pixels p q are extended from greyscale [24,17,25] to multi-channel as follows:…”

Hierarchical multi-scale supervoxel matching using random forests for automatic semi-dense abdominal image registration

“…To avoid this interaction, random forests (RF) [4] have been recently proposed to establish supervoxel correspondences in a unsupervised fashion [5]. Similarly to [6] which encodes a single labeled image as a forest for multi-atlas label propagation, [5] describes a single source image as a collection of supervoxels whose indexes are used via RF to predict matching areas in another target image, leading to semi-dense correspondences. For the purpose of supervoxel matching, the selection of the supervoxel resolution is key to ensure accurate boundary adherence of supervoxels.…”

“…In this perspective, we extend automatic semi-dense medical image registration through RF from single [5] to multi-scale supervoxel matching. Relying on a hierarchical multi-scale supervoxel representation successfully applied to multi-class tissue classification [9,7], our framework implicitly constrains the matching search space in a coarse-to-fine strategy.…”