Abstract

Recent advancements in deep learning have significantly improved image classification models, yet extending these models to alternative data forms, such as point clouds from Light Detection and Ranging (LiDAR) sensors, presents considerable challenges. This paper explores applying knowledge distillation techniques as a solution, aiming to transfer the learned competencies from established image classification frameworks to point cloud classification tasks. Our methodology involves distilling complex model insights into more computationally efficient forms suitable for LIDAR data, thus enabling substantial resource savings without sacrificing performance. Experimental evaluations across various benchmark datasets demonstrate that our distilled models not only rival their original counterparts in accuracy but also surpass conventional point cloud classification methods in both efficiency and scalability. Additionally, we delve into the impact of varying distillation techniques on model adaptability and performance within the LIDAR domain. The findings underscore the utility of knowledge distillation in enhancing the trans-domain applicability of image classification models, potentially revolutionizing their deployment across diverse data types.