Resource-efficient optimizations of 3D vision models for segmentation and detection

The growing deployment of deep learning models in real-world 3D vision applications, ranging from medical image segmentation to autonomous perception, demands solutions that are not only accurate but also computationally and memory efficient. However, the resource intensity of 3D models presents persistent challenges at both ends of the deployment pipeline, particularly during training and inference. This thesis addresses these two critical bottlenecks with a unified objective: to improve the efficiency and scalability of 3D vision systems under real-world constraints. We begin by examining the shared computational challenges faced by 3D models, including high-dimensional input, constrained batch sizes, prolonged training times, and the need for low-latency, high-throughput inference on edge devices. Using volumetric medical imaging and LiDAR-based 3D object detection as case studies, we identify common structural and performance limitations and develop domain-adaptive optimization techniques. To mitigate training-time memory overhead, we employ an in-place normalization strategy that replaces conventional normalization layers (e.g., \texttt{BatchNorm3d}, \texttt{InstanceNorm3d}) with memory-efficient in-place variants. This technique reduces activation memory consumption without altering model behavior, enabling larger batch sizes and improved throughput. We validate the approach through theoretical gradient correctness analysis and empirical evaluations on representative 3D medical imaging models, including LHU-Net and 3D Res-UNet, across high-resolution volumetric datasets such as the Brain Tumor Segmentation Challenge (BraTS), the Left Atrial dataset (LA), and the NIH pancreas dataset (CT-82). For inference-time optimization, we introduce \qlidar, a training-free and calibration-free quantization framework designed to address the structural heterogeneity of modern 3D LiDAR object detection models. \qlidar incorporates \smoothconv and fine-grained channel-wise quantization to effectively compress a diverse set of model components, including 1D/2D convolutions (Conv1d/2d), sparse convolutions (SPConv), submanifold convolutions (SubMConv), and multi-layer perceptrons (MLP). By eliminating the need for retraining or calibration data, \qlidar achieves high compression ratios while maintaining competitive accuracy under mixed-precision settings (W4A8 and W8A8) across standard benchmarks, including Waymo, nuScenes, and KITTI datasets. Collectively, these contributions advance the design of efficient and scalable 3D vision models by addressing critical training and inference inefficiencies. By bridging memory-aware learning with deployment-aware quantization, this thesis lays the groundwork for future 3D vision systems that are not only high-performing but also resource-efficient and deployment-ready.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Zheng, Hongbo

Contributors dc:contributor

• Zhang, Minjia

• 3D Vision
• Resource-Efficient Deep Learning
• Volumetric Medical Image Segmentation
• In-Place Operation
• LiDAR-based 3D Object Detection
• Post-Training Quantization (PTQ)
• Training-Free Quantization (TFQ)
• Mixed-Precision Inference

Statement dc:rights

• Copyright 2025 Hongbo Zheng All rights reserved.

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129530

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129530