Off-campus WSU users: To download campus access dissertations, please use the following link to log into our proxy server with your WSU access ID and password, then click the "Off-campus Download" button below.
Non-WSU users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
Access Type
WSU Access
Date of Award
January 2025
Degree Type
Dissertation
Degree Name
Ph.D.
Department
Computer Science
First Advisor
Jing Hua
Abstract
Neural implicit representations have become a foundational tool in computer vision and graphics, enabling compact, high-fidelity modeling of 3D geometry from partial or sparse observations. However, challenges such as unstable optimization, lack of geometric precision, and inefficiency in real-time scenarios continue to hinder their broader applicability. In this dissertation, we present a unified framework that introduces quantization and surface-aware modeling to overcome these barriers, advancing the capabilities of neural implicit representations across multi-view reconstruction, signed distance field (SDF) inference, and neural simultaneous localization and mapping (SLAM).
Our contributions are threefold:
1. Multi-view 3D Reconstruction via Coordinate QuantizationWe propose a novel method to discretize continuous spatial coordinates into high-resolution quantized grids, stabilizing the optimization of coordinate-based neural implicit functions. By reducing input variation and promoting sample reuse across views, our quantization strategy enhances multi-view consistency and improves surface fidelity without additional computational cost. Experimental results on benchmarks such as DTU and BlendedMVS demonstrate superior performance over state-of-the-art methods like NeuS, UNISURF, and NeuralWarp, offering a plug-and-play solution for existing frameworks.
2. SDF Inference with Surface Patch SensingTo improve geometric accuracy in implicit surface reconstruction, we introduce a surface-aware approach that senses local surface patches from ray-surface intersections during volume rendering. By imposing explicit geometric constraints—such as normals, consistency losses, and photo-consistency—on these inferred patches, our method produces more accurate, artifact-free reconstructions. Extensive evaluations on Replica, ScanNet, and NeuralRGBD validate its effectiveness, setting new benchmarks for implicit surface inference from sparse views.
3. Neural SLAM with Query QuantizationAddressing real-time SLAM challenges, we develop a query quantization mechanism that discretizes input queries using a learnable codebook. This approach accelerates convergence by making neural networks increasingly familiar with repeated inputs, enabling robust camera tracking and mapping in dynamic environments. Our method achieves state-of-the-art accuracy and efficiency on datasets like TUM RGB-D, ScanNet, and SyntheticRGBD, outperforming systems such as iMAP, NICE-SLAM, and Co-SLAM.
Each of these contributions is supported by open-source implementations—CQ-NIR, Surface-Sensing-SDF, and QQ-SLAM—which not only ensure reproducibility but also provide a foundation for future research in 3D vision, robotics, and graphics.
By unifying quantization techniques with geometric awareness, this dissertation paves the way toward scalable, stable, and accurate 3D scene understanding. The generalizability of the proposed framework suggests promising applications in areas such as AR/VR, autonomous navigation, and generative 3D modeling. Looking forward, we envision further extensions to dynamic scene reconstruction, multi-modal data integration, and adaptive real-time systems, driving the next generation of 3D perception and interaction.
Recommended Citation
Jiang, Sijia, "Advancing Neural Implicit Representations For Efficient 3d Scene Reconstruction And Mapping" (2025). Wayne State University Dissertations. 4285.
https://digitalcommons.wayne.edu/oa_dissertations/4285