Recognition of occluded objects by feature interactions

The main challenge for occlusion problem is that features from different objects tend to interact and cause recognition failures for traditional object recognition algorithms where even matched feature points do not necessarily lead to successful recognitions. Feature interactions may be the key to recognize occluded objects. In this paper, we propose a framework to integrate local feature interactions in terms of color, texture and geometry into spectral matching. Appearance similarity will serve as a prior to compensate the sensitivity of spectral matching towards noisy data caused by occlusions. Accordingly incorrect correspondences can be discarded by remaining the geometrical consistency in the formed affinity matrix. Because of our informative similarity matrix, objects under severe occlusions can still be recognized and matching errors dramatically reduced in recognizing both 2D and 3D occluded objects.