A novel part-based benchmark for 3D object reconstruction

Abstract

Numerous deep learning-based methods have been proposed for achieving high accuracy in 3D object reconstruction. However, when examining the recent models, we observe that their performances are very close. We claim that more detailed evaluation methods are needed to broaden the comparisons and allow new research directions. Accordingly, in this study, we propose a novel benchmark to evaluate at the part level over three state-of-the-art reconstruction models using the novel rich dataset, 3DCoMPaT++. To evaluate holistic shape reconstruction outputs at the part level, the Part F-Score metric is proposed. Adapting a dataset proposed from a close domain is important for enabling new data to 3D object reconstruction applications and for guiding new adaptations.