Aggregating Viewpoints for Effective View-Based 3D Model Retrieval
Sou Watanabe, Shigeo Takahashi, Luobin Wang · 2021
The bag-of-features (BoF) model is the standard platform for image retrieval systems and successfully extended to systems for exploring 3D models through their projected views. However, we need a large number of views for each 3D model to achieve shape retrieval systems with high accuracy, which results in increased data storage and long computation time for shape comparison. This paper presents an approach for reducing projected images in such image-based shape retrieval by aggregating views of each 3D model. Our approach begins by discovering a proper metric for evaluating dissimilarity between 3D models by referring to their high-dimensional feature vectors obtained from the BoF model. We then introduce a variant of the k-means clustering method to identify the representative views of each 3D model, given the number of such essential views. Finally, we adjust the degree of such view aggregation by assessing the number of plane symmetries for each 3D model. We test our approach with a dataset containing 200 3D models and we learn that we can reduce the number of views to less than 10% while limiting the degradation of accuracy to approximately 5%.