Perceptual Organization of Shape

Computing the correct bounding contours of objects in complex natural scenes is generally thought to be one of the harder computer vision problems, and the state of the art is still quite far from human performance, when human subjects are given an arbitrary amount of time to delineate the shape of these contours [1]. How do we explain this gap? One possibility is that, when given enough time, humans fall back on high-level, deliberative reasoning processes. If this is true, then it is possible that when faced with detecting and recognizing objects in an active vision timeframe (hundreds of milliseconds), we may rely upon a simpler “bag of tricks”, using appearance cues such as texture and color to discriminate objects. To address this possibility, let us consider the specific task of rapidly detecting animals in natural scenes. Humans perform this task remarkably well: evoked potential studies indicate that the corresponding neural signals can emerge in the brain within 150 msec of stimulus onset [32], and eye movements toward animal targets can be initiated in roughly the same timeframe [20]. Until recently, little was known about the cues that humans use to achieve this impressive level of performance. However, a recent study by Elder & Velisavjlević [10] sheds some light on this question. This study made use of a standard computer vision image dataset called the Berkeley Segmentation Dataset (BSD) [26]. For each image in the dataset, the BSD provides hand segmentations created by human subjects, each of which carves up the image into meaningful regions. Elder & Velisavjlević used this dataset to create new images in which luminance, color, texture and shape cues were selectively turned on or off (Fig. 5.1—top). They then measured performance for animal detection using these various modified images over a range of stimulus durations (Fig. 5.1—bottom left), and estimated the weight of each cue using a multiple regression technique (Fig. 5.1—bottom right).