An approach to object contour completion: By combining high level and low level cues

摘要

An approach that is rooted in psychological and biological research is suggested for extracting object contours. The approach combine high level object recognition cues with low-level image edges to obtain the object contour. We argue for a top-down bottom-up interaction architecture by demonstrating an automated two-stage object detection program that first plausibly identifies the object using an edge-based contour feature and then segments the object using an optimization program with input from the earlier identified contours.