hide
Free keywords:
-
Abstract:
The problem of Surface Reconstruction arises in many real world situations. We
introduce in detail the problem itself and then take a brief look into its
applications and existing techniques, particularly learning based techniques,
developed for its solution. Having presented the context, we closely examine
one such learning based technique – the Neural Mesh algorithm for Surface
Reconstruction.
Despite being relatively recent, the Neural Mesh algorithm has already
undergone several revisions, thus giving rise to several variants of the
original algorithm. We study the algorithm and each of its variants in detail.
All variants rely in varying
degrees on a specific aspect of the algorithm – a signal counter. We observe
that algorithmic reliance on the signal counter impedes performance and propose
an alternate way of performing the same functionalities – using a list.
Additionally, on the practical side, we identify areas where inhouse
implementations of the algorithms were wanting in efficiency and revise those
areas.
Changing over from the signal counter to the list represents a change in
approach from the exact learning of the original algorithms to a comparative
learning framework. We show empirically that this change in approach does not
produce any significant difference in the quality of the algorithms’ output,
while performance, in terms of running time, increases dramatically.