🚀 3D Splats Cinematic Agent

Turn a 3D Gaussian Splat scene into a cinematic space-grade camera flythrough.

---

Introduction

This project provides a set of tools for working with Gaussian Splat scenes. It supports scene preview, camera path generation, object detection, and basic path planning. The goal is to offer practical utilities for analysis and video creation without relying on heavy 3D software.

Features

🔍 Scene Preview

• Load and display .ply / Gaussian Splat scenes;
• Basic navigation for inspection;
• Scene noise filtering.

🎬 Video Generation

• Produce 30–120 second camera paths;
• Supports different scene types (indoor, outdoor);
• Generates stable trajectories suitable for rendering;
• Export camera motion for external tools.

🧭 Object Detection

• YOLO-based detection on rendered frames;
• Used for identifying additional points of interest in the scene.

📐 Path Planning

• Simple ML-based approaches (e.g., k-means) for deriving camera routes;
• Can generate trajectories based on scene structure.

How to Launch

1. Clone

git clone https://github.com/lolyhop/3d-splats-cinematic-agent
cd 3d-splats-cinematic-agent

2. Install the dependencies

pip install -r requirements.txt

3. Download a Gaussian Splat Scene

You can download sample Gaussian Splat scenes from Superspl.at. These files can be used directly with the 3D Splats Cinematic Agent for preview, path planning, and video generation.

Usage Guide

Scene Preview

python -m src.renderer --scene_path <path_to_scene>

Generate a fly-in

python src.main --scene_path <path_to_scene>

Algorithms Overview

For a detailed technical explanation of the algorithms and pipeline, see the Technical Report.

Limitations

• Scene preview may lag on large point clouds;
• Indoor trajectories can occasionally skip or teleport between pivot points;
• Currently, rendering is limited to point clouds only; full scene geometry and textures are not supported.

Future work can be found here.

Tech Stack

• NumPy 2.3.5 — numerical computations and array manipulations;
• Open3D 0.19.0 — 3D point cloud processing, visualization, and geometry operations;
• OpenCV 4.12.0.88 — image processing, video rendering, and visualization;
• Plyfile 1.1.3 — reading and writing PLY files;
• scikit-learn 1.7.2 — ML algorithms for clustering and path planning;
• Ultralytics YOLO 8.3.221 — object detection on rendered frames.