Optimal Multiclass Classifier Threshold Estimation with Particle Swarm Optimization for Visual Object Recognition

Abstract

We present a novel method to maximize multiclass classifier performance by tuning the thresholds of the constituent pairwise binary classifiers using Particle Swarm Optimization. This post-processing step improves the classification performance in multiclass visual object detection by maximizing the area under the ROC curve or various operating points on the ROC curve. We argue that the precision-recall or confusion matrix commonly used for measuring the performance of multiclass visual object detection algorithms is inadequate to the Multiclass ROC when the intent is to apply the recognition algorithm for surveillance where objects remain in view for multiple consecutive frames, and where background instances exists in far greater numbers than target instances. We demonstrate its efficacy on the visual object detection problem with a 4-class classifier. Despite this, the PSO threshold tuning method can be applied to all pairwise multiclass classifiers using any computable performance metric.

This work was partially supported by the Defense Advanced Research Projects Agency (government contract no. HR0011-10-C-0033) NeoVision2 program. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressly or implied, of the Defense Advanced Research projects Agency of the U.S. Government.