Abstract
Accurate 4D plant phenotyping is critical for crop breeding but remains challenging due to complex plant geometry and non-rigid growth. We present GrowSplat, a framework that leverages 3D Gaussian Splatting (3DGS) and a robust temporal registration pipeline to create high-fidelity digital twins. By integrating color-informed iterative closest point (ICP) alignment, GrowSplat enables the tracking of morphological changes over time in a non-destructive, scalable manner.
Introduction
Plant phenotyping involves measuring physical traits such as leaf angle, stem length, and biomass to understand plant-environment interactions. Traditional methods are often manual, slow, and sometimes destructive. While 3D reconstruction techniques like Neural Radiance Fields (NeRFs) have emerged, they are often computationally prohibitive for high-throughput industrial applications. Plants present unique challenges: thin structures, significant self-occlusion, and non-rigid deformations during growth. GrowSplat addresses these by providing a fast, accurate 4D modeling solution.
Methods
The GrowSplat pipeline operates in three distinct phases:
- 3D Gaussian Splatting (3DGS): Unlike volumetric representations, 3DGS uses ellipsoidal "splats" to represent the scene. This approach is significantly faster to train and render. The system utilizes the
Splatfacto-MCMCstrategy and segmentation masks to isolate the plant from the background, focusing the reconstruction on the biological subject. - Preprocessing and Filtering: Raw splat data is filtered to remove noise, specifically targeting splats with extreme scales or elongated shapes that do not correspond to real plant geometry. Surface normals are estimated to support subsequent feature extraction.
- Temporal Registration: To align scans across time steps, a two-stage process is used. Coarse alignment is achieved via Fast Point Feature Histograms (FPFH) and Fast Global Registration (FGR). Fine alignment utilizes Colored Iterative Closest Point (ICP), which incorporates photometric data to resolve ambiguities in self-similar plant structures like leaves and stems.
Experimental Setup
The system was validated using datasets from the Netherlands Plant Eco-phenotyping Center (NPEC), specifically the "Maxi-Marvin" setup. This industrial system captures images using 15 static cameras arranged in layers around a plant on a conveyor belt. The study utilized a Sequoia dataset (40 time steps over three months) and a Quinoa dataset (55 time steps over 76 days).
Results and Discussion
GrowSplat successfully generated detailed 4D models for both species. Key findings include:
- Efficiency: 3DGS reached target quality significantly faster than NeRF-based alternatives (3s vs 6s for the first 100 iterations in initial tests), making it suitable for large-scale operations.
- Accuracy: The use of Colored ICP proved vital for maintaining biological consistency during registration, allowing for precise tracking of individual leaf development and internode lengths.
- Scalability: The automated nature of the pipeline facilitates its use in industrial greenhouse environments.
While successful, the framework could be improved by incorporating biological growth priors (e.g., monotonic leaf expansion) to prevent registration errors caused by noise. Future work will also explore automated biomass estimation and temporal interpolation between captured frames.
Conclusion
GrowSplat represents a significant advancement in non-destructive plant phenotyping. By combining the speed of Gaussian Splatting with color-aware registration, it provides a viable path toward industrial-scale 4D digital twins of complex biological systems.
Summary updated