Statistical strategy for object class recognition using part detectors
Thang V. Pham, Arnold W. M. Smeulders · UvA-DARE (University of Amsterdam) · 2004
This paper presents a method for the recognition of object classes once parts have been detected. The recognition task is formulated as a graph problem searching for the characteristic geographical arrangements of (possibly missing) parts. The objective function is Bayesian maximum a posteriori estimation, integrating the image likelihood as a posteriori probability of the part detectors. The variability in the arrangement of object parts is captured by a Gaussian distribution after translation normalization. By employing two special properties of a Gaussian distribution, we are able to deal with missing parts situation where the chosen origin is not detected. We use an A # algorithm to find the optimal solution for the graph search problem. Experiments are performed on both synthetic and real data to demonstrate good results and fast performance of the recognition.