Abstract:
We have developed a shape-based object retrieval method which is trained
using computational feature metadata (hereafter referred to as computational
metadata) that encodes 3D shape descriptors. The retrieval method is based
on graph neural networks and involves training the network to quantify the
shape similarity between two different 3D objects. A shape similarity metric
is critical for identifying 3D objects according to their geometric and
structural features, rather than relying solely on type descriptions.
Existing 3D model datasets include broad-based, general metadata in the form
of categorical labels. However, they lack ground truth metadata that capture
geometric/shape properties. We present novel research that provides an
approach for establishing robust ground truth metadata for 3D shape
similarity. We validate our approach with a user study that compares three
different methods. Due to the absence of a suitable 3D object dataset that
includes shape characteristics, we created our own ground truth collection
using an image-based retrieval method from Azure AI Vision Studio. The newly
defined computational metadata for 3D shape have been utilized to produce a
rotation-invariant shape-based object retrieval capability.
