High-fidelity 3D mesh generation from a single sketch using shape constraints.

Area of Science:

• Computer Vision
• Deep Learning
• 3D Graphics

Background:

• Deep learning has advanced 3D model reconstruction from images.
• Sketches provide less visual data than images, complicating accurate 3D reconstruction.
• Existing methods struggle with the inherent ambiguity of single sketches.

Purpose of the Study:

• To develop a robust deep learning architecture for high-fidelity 3D mesh generation from single sketches.
• To overcome the challenge of insufficient information in sketches for 3D reconstruction.
• To improve the accuracy and adaptability of sketch-to-3D model conversion.

Main Methods:

• Implemented a streamlined network using the PowerMLP architecture within an encoder-decoder framework.
• Utilized 3D shape constraints to ensure geometric fidelity, moving beyond traditional discriminators.
• Trained and evaluated the model on both synthetic and real-world handwritten sketches.

Main Results:

• Achieved state-of-the-art (SOTA) performance in 3D mesh generation from single sketches.
• Demonstrated superior representation capability compared to traditional Multi-Layer Perceptron (MLP) methods.
• Validated the model's robustness and adaptability across diverse sketch types.

Conclusions:

• The proposed PowerMLP-based encoder-decoder network effectively reconstructs high-fidelity 3D models from single sketches.
• Integrating 3D shape constraints enhances geometric accuracy in sketch-based 3D generation.
• The method shows significant promise for applications requiring 3D model creation from limited input data.