Hierarchical Relation Distillation for Efficient 3D Visual Grounding

Abstract

3D Vision-Language (3D-VL) Grounding seeks to localize 3D objects in point clouds using natural language descriptions. While existing methods focus on improving performance through intricate spatial understanding modules, the efficiency of 3D grounding models has been overlooked, despite its importance for real-time applications. To address this issue, we propose a novel hierarchical relation distillation framework that creates a small yet high-performing model through a two-stage distillation process. The first stage focuses on semantic feature distillation, eliminating noise and retaining valuable information for second stage training. The second stage utilizes half-layers of the teacher model for 3D relation feature distillation, enhancing computational efficiency without sacrificing accuracy. Experimental results on the challenging Nr3D and Sr3D datasets demonstrate that our distilled model achieves state-of-the-art performance while reducing computational costs by over 40%.