Distance Computations between Binary Code and Real-Number Vectors for Efficient Keypoint Matching
An issue that has arisen in the implementation of object recognition in a client/server system is the reduction of the amount of data transfer received by the server from the client, in order to lessen the load on the network. However, since a reduction in the data transfer amount leads to a drop in performance, there is a desire for a reduction in the data transfer amount while maintaining the recognition performance. We therefore propose high-speed distance calculations by local features represented by binary code on the client side and features represented by real-number vectors on the server side.
Matching of Feature Points Based on Distances between Binary Code and Real-Number Vectors
The proposed method represents local features as real-number vectors on the server and as binary code on the client. It calculates the Euclidean distances between the binary code and real-number vectors, and matches feature points on the basis of distance. However, there is a problem in that, if the magnitudes of the Euclidean norm of the binary code and the real-number vector is radically different, the accuracy of feature point matching will drop. In such a case, we introduce a scale factor that absorbs the disparity in the Euclidean norm of the binary code and the real-number vector, and implement highly accurate calculations of the distance between feature point by optimization.
High-Speed Euclidean Distance Calculations Enabled by Introduction of Vector Resolution Method
We can expect an improvement in the matching accuracy between feature points by maintaining the feature vectors on the server as highly representational real numbers. A great reduction in the amount of calculations and the amount of memory used can be implemented, in a situation in which the matching performance of feature points is maintained, by resolving the real-number vectors on the server side into a small number of weight factors and binary basis vectors. The distance calculations can be done approximately eight times faster and the memory amount can be reduced to approximately 1/17 by resolving the real-number vectors on the server side.