Reconstruction from Range Data: Reconstruction from Synthetic Range Images

To generate synthetic range images, we use a modified version of the Rayshade ray-trace. We collect three range images in a circle of viewpoints around the Stanford Bunny model. These three synthetic range images are then used as input data to our reconstruction algorithm. For each range image, surface constraints are collected by uniformly undersampling the range image to reduce the data set, as needed by our method of reconstruction. One exterior negative constraint is specified for each surface constraint. Exterior constraints are allocated within the free space as previously described. Additional exterior constraints are defined on a sphere surrounding the bounding box of the object at a distance farther away from the object. The figure below shows the distribution of constraints defined for the Bunny. Surface constraints are drawn as blue squares imbedded in the surface, and negative constraints are drawn as green squares. 3000 surface, 200 exterior, and 100 interior constraints were used to create the model shown below. Values of lambda = 0.001, delta = 10, and tau = 0.01 were used to reconstruct the surface. The different views of the reconstructed bunny in show that our model is quite similar to the ground truth. Note how the model is closed on the top and bottom of the Bunny even though few constraint points were obtained in those locations. This is due to the fact that implicit surfaces are inherently manifold.