Learning to Find Object Boundaries Using Motion Cues

Abstract

Abstract

Globally Consistent Multi-label Assignment on the Ray Space of 4D Light Fields

“…While this has already been realized in related work on e.g. multiview co-segmentation [15] or segmentation with depth or motion cues, which are in many aspects similar to disparity [21,10], light fields also provide an ideal structure for a variational framework which readily allows consistent labeling across all views, and thus increases the accuracy of label assignments dramatically. …”

Semantic contours from inverse detectors

“…The usefulness of annotating at the boundary level has been acknowledged in other vision problems, for instance, Hoeim et al [11] select a subset of contours in the LabelMe dataset [19] to evaluate occlusion boundary detection from a single image; Stein et al [21] released the CMU motion dataset for a similar task, but using a sequence of images. In [6], the annotations of the BSDS are extended to the level of complete objects.…”

“…For example, grouping edge pixels based on mid-level Gestalt cues, e.g. [18,23,12], or recovering occlusion boundaries from a single image [11], or from a sequence [21].…”

Local Occlusion Detection under Deformations Using Topological Invariants

“…Regrettably, this has the unintended consequence of making the motion estimate inaccurate at boundaries where occlusions occur. An alternative to this smoothing approach is the use of an implicit model, either learned from local motion cues estimated from training data or based on some fixed model of the distribution of motion cues in the vicinity of occluding boundaries [15][16][17][18]. Though these approaches are appealing because they rely on well-defined statistical models, they remain sensitive to deviations of the actual data from the trained model.…”