Benchmarking shape signatures against human perceptions of geometric similarity

摘要: Abstract Manual indexing of large databases geometric information is both costly and difficult. Because this, research into automated retrieval schemes has focused on the development methods for characterising 3D shapes with a relatively small number parameters (e.g. histograms) that allow ill-defined properties such as “geometric similarity” to be computed. However although many generating these so called shape signatures have been proposed, little work assessing how closely measures match human perceptions similarity reported. This paper details results trial compared part families identified by subjects three published signatures. To do this matrix Drexel benchmark datasets was created averaging twelve manual inspections. Three different (D2 distribution, spherical harmonics surface portioning spectrum) were computed each component in dataset, then used input competitive neural network sorted objects numbers “similar” clusters. Comparison machine generated clusters (i.e. families) similar components allows effectiveness at duplicating quantified. The reported makes two contributions. Firstly perception test suggest dataset contains whose perceived levels ranged across recorded spectrum (i.e. 0.1 0.9); Secondly obtained from benchmarking against demonstrate low rate false positives all negative varied almost linearly amount similarity. In other words studied reasonably effective matching they returned few wrong excluded parts direct proportion level demanded user.