3D Scene Reconstruction with Explicit Geometric Modeling
- Publication Type:
- Thesis
- Issue Date:
- 2025
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
3D scene reconstruction aims to recover the geometric and visual structure of real-world environments from sensor data. Existing methods generally fall into two categories: explicit approaches, which represent geometry using discrete primitives like meshes, and implicit approaches, which encode scenes as continuous functions. While implicit methods have achieved impressive visual fidelity, they often lack the direct geometric access required for physically grounded applications. This thesis addresses this limitation by improving explicit geometric modeling through a systematic investigation targeting three complementary levels of reconstruction complexity.
First, at the data level, we propose the Hybrid Geometric Transformer (HGT). This network fuses visual semantics with sparse LiDAR points to estimate accurate surface normals, providing robust geometric priors. Second, for static scene representation, we introduce a hybrid model that integrates 3D Gaussian Splatting with a triangle mesh. By explicitly binding Gaussians to mesh faces, we enforce structural constraints while enabling high-fidelity rendering. Third, extending to dynamic environments, we propose the Dynamic Appearance Particle Neural Radiance Field (DAP-NeRF), which utilizes Lagrangian particles to explicitly model motion and appearance evolution. Experimental results demonstrate that these contributions significantly enhance both reconstruction accuracy and physical interpretability.
Please use this identifier to cite or link to this item: