IndoorGS-exp

Unofficial reproduction of the IndoorGS pipeline with ongoing exploration of improvements for indoor 3D Gaussian Splatting reconstruction.

This repository is based on the official implementation of
3D Gaussian Splatting: https://github.com/graphdeco-inria/gaussian-splatting

The goal of this project is to reproduce the IndoorGS pipeline and explore potential improvements for indoor scene reconstruction, especially for challenging scenarios such as weak-texture surfaces.

---

Pipeline Overview

The overall workflow follows the IndoorGS preprocessing pipeline:

1. COLMAP Reconstruction

Generate camera poses and sparse point clouds using COLMAP via:

convert.py

This produces the standard COLMAP reconstruction outputs.

2. 3D Line Extraction (LIMAP)

Run LIMAP to extract 3D line structures from the reconstructed scene.

These lines are later used as geometric cues.

3. SAM Segmentation

Download the Segment Anything (SAM) model and generate segmentation masks.

These masks are used for geometric cue generation.

4. Geometric Cue Generation

Use the scripts in:

geometric-cues/

to generate three types of points:

• Dynamic points (from COLMAP)
• Plane points
• Line points

These cues are used to initialize the Gaussian representation.

5. 3DGS Training

Run the training pipeline based on 3D Gaussian Splatting to optimize the scene representation.

---

Dependencies

Major dependencies include:

• PyTorch
• COLMAP
• LIMAP
• Segment Anything (SAM)
• 3D Gaussian Splatting

---

Status

This repository is currently a research prototype used for:

• reproducing IndoorGS
• experimenting with potential improvements
• exploring better reconstruction for indoor scenes

Further updates will include improvements and additional experiments.