Depth Edge Based Objectness Metric for Generating Instance Candidate Regions

Abstract

Generating instance candidate regions of arbitrary objects in an image is called object proposal is important in various areas of computer vision. We propose an accurate and fast object proposal method using depth images. Existing object proposal methods can be roughly divided into two categories: window scoring and object region extraction. Of the two, window scoring methods usually have higher efficiency but often suffer from the effects of object texture. In addition, these methods estimate regions that capture multiple objects. We tackled this problem by defining a novel objectness measure from depth images. Since the depth edge of an object is often a closed loop, the proposed method evaluates objectness on the basis of the assumption that regions containing an object should have high depth edge density at the outer region and low depth edge density at inner region. We also took the edge density uniformity at the outer region into consideration. Using our objectness measure, the proposed method estimates individual object regions accurately even in cases where existing methods estimate regions that capture multiple objects. Experimental evaluations demonstrated that our method can estimate proposals faster and more accurately than state-of-the-art methods on a challenging dataset of complex crowded scenes.