Evaluating 2D Image Comparison Metrics for 3D Scene Interpretation

Object recognition can be accomplished using a generate and test algorithm: generate a hypothesis to explain a given scene and then test that hypothesis to determine its quality. The role of the test component is to accept or reject a hypothesis based on how well the objects in that hypothesis explain the scene. In this framework, success or failure is dependent on the quality of the information being provided by the measure. In this paper, we discuss 27 possible 2D measures to test the quality of a hypothesized 3D scene configuration. Using a set of proposed criteria, the performance of these measures is compared on a set of 20 test problems.