Hypothesis and Verification in 3D Model Matching
F. P. Sykes, Stephen B Pollard, J. E. W. Mayhew · 1989
2-D verification as a means of obtaining high performance closed loop model-matching. Performance includes the following criteria: low computational expence; low failure rate; and good localisation. Hypotheses are based upon congruencies identified between 3-D scene and model descriptions. Verification is provided by quantitative evaluation of the 3-D match and the identification of inconsistency (relient upon properties of opaque objects). Although the results are dependent upon the model representation and the quality of the input (raw images), the ability of a verification system to discriminate between a correct and a wrong match has been demonstrated in a number of experiments. Furthermore, symmetries are identified by the system and exploited to guide the search for correct matches. left image right image The architecture of the matcher/verification algorithm This paper describes the 3-D model matching algorithm used within the TINA system at AIVRU Sheffield (an early version of the TINA system is presented in Porrill et al 1987). The adopted strategy (see figure 1) is to base initial matching hypotheses on congruencies identified between 3-D scene and model descriptions, and then to employ a model-based verification strategy exploiting both 2-D image and 3-D scene descriptions to determine the correctness of the hypotheses. Approaching the problem of model matching in this fashion allows us to combine all the good properties of 3-D scene and model matching (primarily computational tractability and geometrical accuracy), with the robustness and completeness of methods based upon the back-projection of the model. The 3-D scene descriptions encountered in the current TINA system are obtained through the processes of edge and line-based binocular stereo and are presently restricted to linear segments. The object models may also include surface information (in the form of a surface tessellation). However, matching hypotheses are restricted to those linear features corresponding to discontinuities in the model's surface. end of process figure 1 F.P. Sykes is a visiting research student from ENST, 46 rue