Evaluation of one-image 3D reconstruction for plant model generation

Area of Science:

• Computer Vision
• Computational Biology
• Plant Science

Background:

• Accurate 3D plant reconstruction from single images is vital for high-throughput phenotyping in crop breeding and precision agriculture.
• Traditional 2D methods fail to capture complex, occluded plant morphology, necessitating advanced 3D techniques.
• Existing generative 3D methods are not systematically validated for plant-specific challenges like intricate geometry and occlusions.

Purpose of the Study:

• To systematically evaluate and compare the performance of six advanced 3D generative techniques for reconstructing plant models from single-view images.
• To identify the most effective generative approach for capturing complex plant structures and volumetric traits.
• To provide insights into the capabilities and limitations of current 3D reconstruction methods for plant digitisation.

Main Methods:

• Six generative techniques (Hunyuan3D 2.0, Trellis, One2345++, InstantMesh, Direct3D, Unique3D) were evaluated using the PlantDreamer dataset.
• Mesh models of Bean plants were reconstructed from single images.
• Quantitative assessment used Chamfer Distance, Normal Consistency, F-Score, PSNR, LPIPS, and CLIP Score against ground-truth models.

Main Results:

• Hunyuan3D 2.0 demonstrated superior overall performance in reconstructing realistic 3D plant models.
• The study provides quantitative and qualitative comparisons of the tested methods on Bean, Kale, and Mint plants.
• Results highlight Hunyuan3D 2.0's effectiveness in capturing complex plant geometry and structures.

Conclusions:

• Hunyuan3D 2.0 is a highly effective method for single-view 3D plant reconstruction.
• The systematic evaluation offers valuable guidance for selecting and improving 3D generative models in plant science.
• This research advances non-destructive plant digitisation for agricultural applications.