Summary: 1. predict depth from a single image with an accuracy of 100% (Andrew Ng and Saxena have done this with an accuracy of 67%), then, 2. use a training set of real 3d models to calculate voxels outside of a test 3d model (ideally one generated by the first project) (first predicting a single voxel, then adding that voxel to the set of known voxels, then using the new set of voxels to predict another voxel, and so on).

Will this generate the desired results (being able to calculate voxels outside of a given 3d model, and ultimately, allowing us to see things outside of the original camera scene of a given photo)?