Edge_MVSFormer: Edge-Aware Multi-View Stereo Plant Reconstruction Based on Transformer Networks

Area of Science:

• Computer Vision
• Deep Learning
• Botanical Research

Background:

• Multi-view stereo (MVS) using RGB cameras is a cost-effective botanical research tool.
• Existing MVS methods struggle with plant texture and fine edge reconstruction, limiting accuracy.

Purpose of the Study:

• To enhance the accuracy of plant leaf edge reconstruction in MVS.
• To develop a novel MVS model focusing on edge information for botanical applications.

Main Methods:

• Proposed Edge_MVSFormer, building upon TransMVSNet.
• Integrated an edge detection algorithm to provide edge information as input.
• Introduced an edge-aware loss function to prioritize edge region accuracy.

Main Results:

• Reduced edge error by 2.20 ± 0.36 mm and overall error by 0.46 ± 0.07 mm in depth images.
• Reduced edge error by 0.13 ± 0.02 mm and overall error by 0.05 ± 0.02 mm in point clouds.
• Demonstrated significant improvements on 10 test model plants.

Conclusions:

• Edge information is critical for precise plant MVS data reconstruction.
• Edge_MVSFormer offers a substantial advancement in MVS for botanical studies.
• The method effectively addresses limitations in reconstructing intricate plant structures.