Abstract
The visual system is highly adept at segmenting scenes into regions based on cues such as color, luminance, texture, and boundary contours. It is also adept at determining which of these regions should be bound together as a single object. Work in infant visual development (e.g. Kellman & Spelke 1983, Slater & Johnson 1996) suggests that children are able to accomplish such object binding via dynamic cues of common fate within a few months of life. Our own work, as part of Project Prakash, with patients having low acuity (see E. Meyers abstract in this volume) also emphasizes the important role of motion in the binding of object parts. This pattern of results suggests that common-fate motion cues might comprise an important early mechanism for primitive object binding from which more robust heuristics based on other cues (e.g., junctions, texture, or Gestalt grouping principles) can be learned. Once these heuristics are in place in the developing infant, visual analysis of objects in static images can occur.
Based on the experimental results on the primacy of common-fate motion cues for object binding and segmentation processes, we have developed a motion-based binding algorithm which extracts complex objects from video sequences. A key challenge that the algorithm addresses is binding despite the significant variability in the motion patterns of object parts. In other words, contrary to the simplistic notions of ‘common-fate’, the fates of visual entities that ought to be bound are often quite different. We shall describe how the algorithm accomplishes grouping even with not-so-common fate, and how its predictions compare with experimental data from human observers. This work is a key component of Project DYLAN (see P. Sinha abstract and B. Balas abstract in this volume), which seeks to computationally model the overall process of object concept learning in children.
The John Merck Foundation and The Sloan Foundation