Date of Award

3-20-2026

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2026

Degree Name

Master of Science (MS)

Department

Computer Science and Engineering

First Advisor

Ying Liu

Abstract

Recently, 3D Multimodal Large Language Models (MLLMs) have enabled language-grounded reasoning over 3D point clouds. However, these billion-parameter models must be deployed on cloud servers, while raw point clouds are extremely large, making efficient compression essential for cloud-based MLLM applications. Existing learning-based point cloud compression methods reduce data size but are not optimized for downstream MLLM tasks, often degrading semantic representations under high compression. To address this, we propose a Parameter-Efficient Fine-Tuning (PEFT) framework to compress point clouds for MLLM tasks. We introduce lightweight adapters either into the frozen PointLLM backbone for feature alignment or into the compression backbone using a joint rate–perceptual objective. Experimental results show that when the point cloud is compressed to 6.4% of its original size, our method restores the accuracy of MLLM-based point cloud classification to 47.85%, recovering about 80% of the compression-induced performance loss, demonstrating its effectiveness for bandwidth-constrained cloud-edge MLLM systems.

Download

Share

COinS