Neural Feature Matching in Implicit 3D Representations: Supplementary Material

The global metrics assign a low error, when the two shapes overlap significantly, even if this implies an unnatural fitting. Figure 1 is a characteristic example. With our feature matching, the source arms can only find the closest points on the seat or the back of the target chair, leading to a larger global fitting error; while, with cross-fitting, the arms are forced very close to the seat and the back in an unnatural and distorted manner, which, however, reduces the whole shape error. By contrast, part-level metrics do not count such errors with inconsistent semantics, which makes more sense when shapes differ significantly.