Download the report.
View github repository.

Report Abstract

Retrieving multimedia objects can be a hard task. This paper aims to create a Content-Based Shape Retrieval (CBSR) system which is able to return 3D shapes that are similar to a given query shape. We show how to pre-process shape databases [10][3] in order to extract the features used to compare different shapes. The compari- son results in a distance between the shapes computed using the Euclidean or Earth mover’s distances between individual features. An understandable visualisation of the dataset and its distances is presented using a t-SNE plot, which clearly shows which shapes are considered similar to one another. We show that the result- ing CBSR system can accurately retrieve objects that have a lot of similar geometric attributes. To calculate the sensitivity and speci- ficity of the presented pipeline, a Receiver Operating Characteristic (ROC) curve is computed for every class. Over the entire Labeled PSB dataset [3], the system has an average Area Under the ROC (AUROC) of 0.82. Some shape classes have an average AUROC up to 0.96.