Multi-Head Attention Refiner for Multi-View 3D Reconstruction.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Deep Learning

Background:

• Traditional 3D reconstruction methods struggle to balance edge recall and precision.
• Existing models often fail to capture intricate details crucial for accurate boundary representation.

Purpose of the Study:

• To introduce a novel post-processing technique, the Multi-Head Attention Refiner (MA-R), to enhance 3D reconstruction accuracy.
• To improve the balance between precision and recall in object edge detection within 3D models.

Main Methods:

• Integration of a multi-head attention mechanism into a U-Net style refiner module.
• Development of a post-processing approach applied to existing 3D reconstruction pipelines.

Main Results:

• The MA-R method significantly improves boundary prediction accuracy and recall rates.
• In multi-view reconstruction using Pix2Vox++, MA-R achieved a 0.730 IoU score (1.1% improvement) and 0.483 F-Score (2.1% improvement) with 20-view images.

Conclusions:

• The Multi-Head Attention Refiner (MA-R) effectively enhances 3D reconstruction by improving detailed image capture and boundary delineation.
• The proposed method offers a robust solution for increasing both precision and recall in multi-view 3D reconstruction tasks.