Adaptive Points Sampling for Implicit Field Reconstruction of Industrial Digital Twin

Area of Science:

• Computer Vision
• Artificial Intelligence
• Industrial Engineering

Background:

• Digital twins (DT) are crucial for Industry 4.0, enabling virtual modeling for industrial optimization.
• Existing AI technologies connect physical and virtual spaces, but single-view 3D reconstruction (SVR) for DT remains underexplored.
• Current SVR methods struggle with reconstructing fine details essential for industrial products.

Purpose of the Study:

• To address the limitations of current SVR methods in detail reconstruction for digital twins.
• To propose a novel SVR approach for generating detailed 3D models of industrial products from single images.
• To enhance the convenience, cost-effectiveness, and robustness of digital twin modeling.

Main Methods:

• Developed a detail-aware feature extraction network utilizing a feature pyramid network (FPN).
• Designed an auxiliary network to integrate convolutional feature maps from multiple levels.
• Introduced an adaptive points-sampling strategy to optimize training difficulty and accelerate convergence.

Main Results:

• The proposed method significantly improves the reconstruction of fine details in 3D models.
• Experiments on ShapeNet and an industrial dataset demonstrate the effectiveness of the new SVR approach.
• The adaptive points-sampling strategy accelerates training and enhances network performance.

Conclusions:

• The novel SVR method offers superior detail reconstruction capabilities for digital twin applications.
• This research validates the practicability of SVR technology for creating robust and detailed digital twins.
• The findings contribute to advancing Industry 4.0 by improving virtual modeling of physical industrial assets.