3D Surface Reconstruction: Noise-to-Noise Mapping
Reconstructing 3D surfaces from noisy point clouds using neural implicit functions
Project Overview
This project implements 3D surface reconstruction from noisy point clouds using the Noise-to-Noise Mapping approach with a novel Positional Encoding layer. The goal is to transform raw point clouds from noisy sensors into clean continuous surfaces represented as Signed Distance Functions (SDFs).
Technical Innovation: Positional Encoding
The original model suffered from spectral bias (over-smoothing). To address this, I implemented harmonic positional encoding that:
- Transforms coordinates into higher-dimensional space using periodic functions
- Captures high-frequency details and sharp features
- Prevents merger of disconnected structures
- Improves edge precision on complex geometries
Key Contributions
- Positional encoding implementation for frequency capture
- GCP/GPU deployment and optimization (NVIDIA Tesla T4)
- Sensitivity analysis of iso-surface thresholds
- Earth Mover’s Distance (EMD) loss implementation
Technologies
- Language: Python, C++, CUDA
- Framework: TensorFlow 1.15
- Techniques: MLP networks, Positional encoding, Marching Cubes, EMD loss
Project Report
View or download the full project report:
3D Surface Reconstruction Report