Region segmentation via deformable model-guided split and merge

Abstract

Abstract

Combining Top-Down and Bottom-Up Segmentation

“…These two terms m s (+1), m s (−1) are given for every segment, without any additional computational cost, by the sum-product algorithm (7). In the results below we constructed the confidence map using the confidence of the terminal segments.…”

“…The method combines two recently developed segmentation approaches: a top-down class-specific approach [3] that effectively addresses high variability within object classes and automatically learns the class representation from unsegmented training images [4]; and a bottom-up approach [5] that rapidly detects homogeneous image regions. In contrast with previous approaches for combining class-specific knowledge with bottom-up information ( [6,7,8]), the combined approach presented here is fast (linear in the number of pixels) and takes into account image measurements at multiple scales, converging to a global optimum in just one pass. In addition, the combination is general and can be applied to combine a variety of top-down and bottom-up algorithms.…”

Top-down pairwise potentials for piecing together multi-class segmentation puzzles

“…One set of methods deals with foreground segmentation, where the best figure/ground assignment will have boundaries that agree with a previously learned shape model or other class-based cue [1,6,2,22,3,4,5,7]. Such two-class techniques are intended for single-object images with a given target class, and generally work best when it is possible to construct a consistent shape prior for that category (e.g., side views of horses).…”

“…Such two-class techniques are intended for single-object images with a given target class, and generally work best when it is possible to construct a consistent shape prior for that category (e.g., side views of horses). Some explore the space of groupings using merge operations [22,2], which we also consider here, though with a distinct objective function. The approach of [22] does not use any class-specific knowledge but instead learns to differentiate between "good" and "bad" moves among merges, shifts, and splits between superpixels using low-level cues and hand-drawn segmentations.…”

“…Whereas bottom-up techniques that rely solely on low-level cues such as texture or color are limited to finding regions with fairly homogeneous appearance, methods that also leverage "topdown" knowledge from category models can guide the grouping towards regions that best support object detection. In particular, techniques are available to compute either a class-specific foreground-background segmentation [1,2,3,4,5,6,7], or to label all pixels according to a set of multiple known categories [8,9,10,11,12,13,14]. The key insight of such approaches is that achieving objectlevel boundaries requires expressing an objective for segment quality that includes terms favoring category-specific attributes, not just pixel similarity.…”