A set of artificial 3D models was used to evaluate the performance of the proposed framework, as shown in Fig.1. Each model was represented by 2 reference views obtained by applying different orthographic projections on the 3D models. For each model, all possible groups having 8 point features were considered. A total of 2,242 set of views were generated and stored in the coarse k-d tree. Then a dense number of views was generated for each model group and its manifold was learned using EM algorithm. The ratio of sparse to dense views used was 2%.

The test views were generated by applying random orthographic projections on the 3D models. We also added 3 pixels random noise to point features of the test views. We did not assume any knowledge of the point feature correspondences between model and scene groups, however, we did assume that point features have certain ordering in the group. Assuming that there is no easy way to select the initial point feature in a group, we considered all possible circular shifts of point features when searching the k-d tree. Table 1 shows some query results.

Back to Top