Skip to content

My Research Overview

3D Gaussian Splatting for time-series plant monitoring in greenhouses.

Framing: monitoring, not absolute measurement

This work is a relative growth-monitoring / change-detection system. It tracks how a plant changes over time and detects biologically meaningful events; it does not claim absolute, calibrated trait values. Consistency across sessions (coefficient of variation, CV) is therefore the primary and appropriate evaluation metric.

Research Focus

Developing a robust 3DGS pipeline for: - Time-series greenhouse monitoring and change detection - Scale-invariant, render-space trait extraction - Detection of real crop-management events (e.g. pruning)

Key Contributions

  1. Render-space, scale-invariant trait extraction — height CV 28.0% → 9.8% (2.86× more consistent)
  2. Biological validation via pruning-event detection — three documented events detected (p = 0.0008)
  3. Physical-reference corroboration — render-space height vs. an in-scene 45 cm pipe, r = 0.74
  4. Long-term validation — 49 days, 22 sessions, 100% reconstruction success

See Original Contributions for details.

Publications

Journal paper under review at Computers and Electronics in Agriculture (Elsevier).