Extracting Object Regions Using Locally Estimated Probability Density Functions

In this paper, a novel method for estimating a precise object region using a given rough object region is proposed. For determining whether each pixel belongs to an object or not, the proposed method estimates a joint probability density function (joint p.d.f.) of position, color, and class (object or background). For each pixel, the class with a higher joint p.d.f. is selected. The joint p.d.f. is estimated using a kernel density estimator. Since only local distributions of colors are used for estimating the joint p.d.f., the proposed method can classify pixels correctly even if the object and the background have the same color, which is the case where the conventional methods fail. In the experiments, the number of misclassified pixels estimated by the proposed method was from 12% to 76% of those extracted by conventional methods.

• Type: Conference
• Year: 2007
• Where: MVA
• Authors: Hidenori Takeshima, Takashi Ida, Toshimitsu Kaneko
• Keywords: Computer Vision, Joint P.d.f, MVA 2007, Object Region, Precise Object Region