Spectral technique to recognise occluded objects

When recognising partially visible objects in a scene, a good global decision should be made based on locally gathered features for their recognition, since global information is no longer reliable. This local-to-global nature of occlusion recognition leads us to spectral matching technique. Unfortunately, conventional spectral matching is not desirable for noisy data set from a cluttered scene. In this study, a top-down procedure is introduced into a standard spectral matching for the recognition of occluded objects. Feature points are firstly evaluated and associated with object(s) of interest. Subsequently, geometrical consistency is enforced to find correct correspondences among the candidate matches with high association scores. Based on this two-stage strategy, both appearance and geometric information are taken into consideration. Our algorithm is implemented for both 2D and 3D occluded objects recognition under different occlusion rates. It is shown that the improvement has been made for spectral correspondence algorithm to recognise occluded objects and it has a comparable recognition rate with the state-of-the-art recognition methods.