EI-MVSNet: Epipolar-Guided Multi-View Stereo Network With Interval-Aware Label.

Area of Science:

• Computer Vision
• Photogrammetry
• Machine Learning

Background:

• Learning-based methods are effective for multi-view stereo (MVS) depth estimation.
• Existing methods often overlook receptive field alignment during cost volume construction.
• Shortcomings in current MVS networks limit performance.

Purpose of the Study:

• To propose an Epipolar-Guided Multi-View Stereo Network with Interval-Aware Label (EI-MVSNet).
• To enhance depth estimation accuracy and robustness in MVS.
• To address limitations in feature alignment and depth prediction in existing MVS methods.

Main Methods:

• Developed an epipolar-guided volume construction module using epipolar-guided convolutions to align receptive fields and account for rotation/scale changes.
• Implemented an interval-aware depth estimation module for direct cost volume supervision, using upper and lower boundaries for fine-grained predictions.
• Integrated these modules into a unified MVS architecture.

Main Results:

• EI-MVSNet achieved state-of-the-art performance on MVS tasks.
• The method ranked first on both intermediate and advanced subsets of the Tanks and Temples benchmark.
• Demonstrated high precision and strong robustness compared to existing MVS methods.

Conclusions:

• The proposed EI-MVSNet effectively improves depth estimation in multi-view stereo reconstruction.
• Epipolar-guided volume construction and interval-aware depth estimation enhance network precision and robustness.
• EI-MVSNet represents a significant advancement in MVS technology.