3DGS Pipeline Manual¶

Welcome to the comprehensive manual for 3D Gaussian Splatting (3DGS) for Time-Series Plant Phenotyping.

Pipeline Overview Complete pipeline from video capture to trait extraction

🎯 Purpose¶

This manual provides complete step-by-step instructions for reproducing the 3DGS-based time-series plant phenotyping pipeline developed at Mineno Laboratory, Shizuoka University.

What You'll Learn

✅ Complete environment setup (CUDA, COLMAP, 3DGS)
✅ Video processing and frame extraction
✅ Structure-from-Motion reconstruction
✅ 3D Gaussian Splatting training
✅ Quality evaluation and rendering
✅ Plant trait extraction
✅ Time-series growth analysis

🚀 Quick Start¶

Quick Start

Get running in 30 minutes

Start now
Environment Setup

Complete installation guide

Install
Run Pipeline

Execute the complete workflow

Execute
My Research

Original contributions & results

View

📊 Pipeline Overview¶

graph LR
    A[Video Capture<br/>6 cameras, 4K] --> B[Frame Extraction<br/>5fps, ffmpeg]
    B --> C[COLMAP SfM<br/>Feature matching]
    C --> D[3DGS Training<br/>30K iterations]
    D --> E[Novel View<br/>Rendering]
    E --> F[Trait Extraction<br/>Height, canopy]
    F --> G[Growth Analysis<br/>Time-series]

    style A fill:#e1f5ff
    style D fill:#ffe1f5
    style G fill:#e1ffe1

🎬 Video Demonstration¶

Complete Pipeline Walkthrough¶

Video 1: Complete pipeline execution from video to trait extraction

💻 System Requirements¶

Recommended Setup

Operating System: Ubuntu 22.04 LTS
GPU: NVIDIA RTX 6000 Ada (48GB VRAM) or equivalent
CUDA: 12.1 or later
RAM: 64GB+
Storage: 2TB+ for full dataset

See detailed requirements for minimum specifications.

🎓 Key Results¶

This pipeline achieves:

Performance Metrics

PSNR: 23.71 dB (Cond. 1, 5 fps) · mean 23.84 dB across 22 sessions
Temporal Stability: PSNR CV = 3.5% (excellent consistency)
Height Extraction: h_norm CV = 9.8% (scale-invariant method)
Improvement: 2.86× more stable than direct PLY (28.0% → 9.8% CV)

Results Summary Validated across 22 dates over 49 days

📚 Documentation Structure¶

For Beginners¶

Getting Started - Introduction and basic concepts
Environment Setup - Complete installation guide
Quick Start - Get running in 30 minutes

For Active Users¶

Pipeline Execution - Complete workflow guide
Parameter Configuration - Optimization tips
Troubleshooting - Problem solving

For Researchers¶

My Research - Original contributions
Results & Validation - Complete experimental results
Publications - Papers and presentations

🔧 Common Use Cases¶

Single Date ProcessingMulti-Date AnalysisTrait Extraction Only

# Process one date from start to finish
./scripts/run_single_date.sh 20260119

# Process multiple dates in batch
./scripts/run_batch.sh dates.txt

# Extract traits from existing models
python analysis/compute_heights_rendered.py --dates 20260119

📝 Getting Help¶

Need Help?

Check these resources:

Common Issues - Known problems and solutions
COLMAP Errors - COLMAP-specific issues
GPU/Memory - Hardware troubleshooting

Still stuck?

Contact: Mineno Laboratory, Shizuoka University
Email: zobaer.al.24@shizuoka.ac.jp

🙏 Acknowledgments¶

This pipeline was developed as part of greenhouse tomato phenotyping research at Mineno Laboratory, Shizuoka University.

Author: Zobaer Al (M2)
Supervisor: Professor Hiroshi Mineno
Institution: Shizuoka University

📄 Citation¶

If you use this pipeline in your research, please cite:

@mastersthesis{al2026_3dgs_phenotyping,
  author = {Al, Zobaer},
  title = {3D Gaussian Splatting for Time-Series Greenhouse Tomato Plant Phenotyping},
  school = {Shizuoka University},
  year = {2026},
  type = {Master's Thesis}
}

Last updated: {{ git_revision_date_localized }}