Environmental Model for Outdoor Scene Analysis

In the field of outdoor scene analysis [1], well known procedure to analyze an image is as follows: (1) The image is divided into regions mainly based on color feature. (2) Each region is hypothesized as an object based on statistical knowledge between its characteristics and object labels. (3) The hypotheses are checked by knowledge of three dimensional relation among the objects. In theses steps, color feature takes an important role. However, the colors in the image depend much on the environment in which the image is taken. For example, the color of the sunlight at dawn is different from that in the daytime. In this report, we describe a method to generate an environmental model which could be useful not only in step (1) but also in step (2) of the outdoor scene analysis. There are several research concerned with color constancy [4, 5]. Based on the reflectance model [3], they tried to separate highlight from scattering light and to recover the intrinsic color of an object. For the purpose of understanding an image, however, we do not need to recover the intrinsic color of the object but to generate an object hypothesis. From this standpoint, we are attacking two problems; how to model outdoor environment and how to use the model for generating an object hypothesis. We are in the beginning of our research so this report is only concerned with the former problem.