Detection of Motion Boundaries and Object Contours in Natural Scenes

The ultimate goal underlying this project is to lay the foundation for an artificial object recognition system that is psychophysically-inspired and neurally-constrained. Our working hypothesis is that natural scenes contain strong correlations between motion boundaries and object contours. Hence, explicitly detecting motion boundaries could lead to implicit extraction of object contours from retinal motion. During the first stage of the project, we have developed a bank of local motion detectors (LMDs) that are conceptually similar to the Reichardt model (e.g., as elaborated in van Santen & Sperling, 1984). In contrast to previous implementations, the computational subunits of our LMDs are based on realistic spatio-temporal receptive field models of LGN and V1 neurons, and they can be applied to natural scenes. In the second stage of the project, currently underway, we have been exploring different types of pooling schemes, as well as excitatory and inhibitory interactions between individual LMDs, such as facilitation in the preferred direction, which may play a role in solving the correspondence problem (false motion signals arising due to inapproptiate binding of adjacent subunit responses elicited by separate entities). Once this stage is completed, we plan to use our LMDs as part of a model system for investigating the supervised and unsupervised learning mechanisms that may allow visual systems to detect motion boundaries from discontinuities in local motion signals.

In the future, we also plan to integrate this project with other on-going projects in the lab, specifically: Combining multiple cues for contour detection in natural images. Visual motion and tactile cues provide extensive information about the structure of objects in the world, and there is evidence to suggest that these cues interact in the primate visual system, thus facilitating contour detection and object recognition (e.g., Atkins et al, 2001). Furthermore, the existing evidence for the late maturation of contour integration skills in humans (e.g., Kovacs et al, 1999) is also indicative of the complexity in acquiring these skills, and is congruent with the notion that cues from different domains are brought to bear on the problem of detecting object contours. A promising collaboration would involve using the correlations between consecutive movie frames, which are captured by motion detectors, as teaching signals for image-based edge detectors. Such cross-cue interaction could help suppress responses of edge detectors to ephemeral edges that are likely to represent noise, while facilitating responses to genuine signals - slowly changing edges that emanate from smoothly moving object contours.