Enhanced multi view 3D reconstruction with improved MVSNet.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Deep Learning

Background:

• 3D reconstruction is crucial for environment perception but current methods require improvement.
• Existing 3D scene reconstruction technologies have limitations in detail and accuracy.

Purpose of the Study:

• To propose an improved 3D reconstruction algorithm based on MVSNet.
• To enhance pixel detail extraction and depth estimation precision.

Main Methods:

• Implemented a novel DE module with ECA-Net and dilated convolution for feature extraction.
• Integrated a residual framework for feature splicing and fusion to preserve global image information.
• Utilized attention mechanisms to refine 3D cost volume regularization and multi-scale feature integration.

Main Results:

• Achieved a completeness (comp) of 0.411 mm and overall quality of 0.418 mm on the DTU dataset.
• Demonstrated superior performance compared to traditional and other deep learning-based 3D reconstruction methods.
• Showcased significant advancements in point cloud model visual representation and generalization ability on the Blended MVS dataset.

Conclusions:

• The proposed DE module and attention-based MVSNet significantly enhance 3D scene reconstruction accuracy and detail.
• The algorithm offers a robust and high-performing solution for complex 3D reconstruction challenges.
• The model exhibits strong generalization capabilities, making it suitable for diverse 3D reconstruction applications.