SERAPH: Semantic Entity Radiance with Adaptive Prior Hierarchy

SERAPH is a next-generation 3D reconstruction architecture for large-scale urban environments. It solves the five core problems of traditional 3DGS/NeRF models:

1. SfM Dependency: No external camera poses are required (uses internal HGNN).
2. Zero Transfer: Leverages Global Entity Priors (GEP) to transfer knowledge across scenes.
3. Semantic Decomposition: Scene is organized into entities (buildings, roads, etc.).
4. Scene Prior: Models the hierarchical structure of urban environments using hyperbolic geometry.
5. Hierarchical Scale: Automatic Level of Detail (LoD) via hyperbolic attention.

---

🚀 Getting Started

1. Installation

Install the required dependencies:

pip install openxlab timm einops torch torchvision tqdm pillow numpy

2. Automated Training (Mill 19)

The system is fully automated. To download the dataset and begin training on a specific scene, run:

python train.py --download --scene rubble

---

🏗️ Architecture

• Entity Discovery Network (EDN): Discovers semantic entities across unorganized views.
• Hyperbolic Scene Graph (HSG): Hierarchical scene organization in Lorentz manifolds.
• Prior-Adapted Gaussian Fields (PEPGF): Bayesian adaptation of Gaussian primitives.
• Global Assembly Transformer (GAT): Scale-adaptive scene assembly.
• Differentiable Renderer: GPU-accelerated 3D Gaussian Splatting.

---

📂 Project Structure

• src/: Core architecture components.
• src/utils/: Dataset management and data loading.
• train.py: Unified training entry point.
• results/: CSV logs and epoch metrics.
• checkpoints/: Model state and resume indicators.

---

📊 Dataset Support

SERAPH is optimized for large-scale urban datasets:

• Mill 19 (Rubble, Residential, Building, Sci-Art)
• UrbanScene3D
• MatrixCity

---

📜 Documentation

For a detailed implementation overview, review the Walkthrough.

---

Author: Mark Knoffler Model ID: SERAPH-v1.0