DKD2L: Dual Knowledge Distillation Dynamic Learning for sketch-based 3D shape retrieval

Abstract

Sketch-based 3D shape retrieval has become a prominent area of research in computer vision, confronting challenges related to the inherent diversity and abstraction of sketches, as well as inter-domain discrepancies. This paper introduces a novel approach called Dual Knowledge Distillation Dynamic Learning (DKD<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>L) aimed at enhancing the extraction of spatio-temporal features for sketch-based 3D shape retrieval. We develop a temporal feature extraction network to effectively capture the dynamic temporal characteristics of sketches and improve retrieval efficiency through temporal knowledge distillation. Additionally, to tackle intra-class variation and inter-class imbalance, we apply semantic knowledge distillation, enabling the 3D shape network to guide the sketch network in capturing common semantic information. This approach facilitates precise cross-modal alignment and enhances retrieval accuracy. Extensive experiments on two benchmark datasets demonstrate that DKD<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>L surpasses existing state-of-the-art methods.