SOFTCBIR: Object Searching in Videos Combining Keypoint Matching and Graduated Assignment
Ming Luo, Daniel DeMenthon, Xiaodong Yu and David Doermann
This paper proposes a new approach to object searching in video databases, SoftCBIR, which combines a keypoint matching algorithm and a graduated assignment algorithm based on softassign. Compared with previous approaches, SoftCBIR is an innovative combination of two powerful techniques: 1) An energy minimization algorithm is applied to match two groups of keypoints while accounting for both their similarity in descriptor space and the consistency of their geometric configuration. The algorithm computes correspondence and pose transformation between two groups of keypoints iteratively and alternately toward an optimal result. The objective energy function combines normalized distance errors in descriptor space and in the spatial domain. 2) Initial individual keypoint matching relies on Approximate K-Nearest Neighbor (ANN) search. ANN achieves much more accurate initial keypoint matching results in the descriptor space than K-means labeling. Experiments prove the effectiveness of our approach, and demonstrate the performance improvements rising from the combination of the two proposed techniques in the SoftCBIR algorithm.
The support of this research by the Department of Defense under contract MDA-9040-2C-0406 is gratefully acknowledged.
Reference: Technical Report: LAMP-TR-132/CAR-TR-1013/CS-TR-4804/UMIACS-TR-2006-25, University of Maryland, College Park, May 2006.