EAR-Net: Pursuing End-to-End Absolute Rotations from Multi-View Images.

Area of Science:

• Computer Vision
• Deep Learning
• 3D Reconstruction

Background:

• Absolute rotation estimation is crucial for 3D computer vision tasks.
• Traditional multi-stage methods suffer from accumulated errors, degrading global rotation accuracy.

Purpose of the Study:

• To propose an end-to-end deep neural network method, EAR-Net, for accurate absolute rotation estimation from multi-view images.
• To overcome the limitations of sequential processing in existing methods.

Main Methods:

• Developed EAR-Net, an end-to-end deep neural network.
• Incorporated an epipolar confidence graph construction module to predict relative rotations and confidences.
• Utilized a differentiable confidence-aware rotation averaging module for absolute rotation prediction.

Main Results:

• EAR-Net effectively handles outlier cases due to confidence weighting.
• Experimental results on three public datasets show significant improvements over state-of-the-art methods.
• Achieved superior accuracy and faster inference speed compared to existing approaches.

Conclusions:

• The proposed EAR-Net method provides a robust and efficient solution for absolute rotation estimation.
• End-to-end learning with confidence awareness significantly enhances performance in 3D computer vision.