SPDET: Edge-Aware Self-Supervised Panoramic Depth Estimation Transformer With Spherical Geometry.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Panoramic depth estimation is crucial for 3D reconstruction but lacks sufficient RGB-D datasets.
• Supervised methods are limited by the scarcity of panoramic RGB-D data.
• Self-supervised learning offers a promising alternative due to reduced dataset dependency.

Purpose of the Study:

• To propose SPDET, an edge-aware self-supervised panoramic depth estimation network.
• To leverage transformer architecture with spherical geometry for improved depth map reconstruction.
• To address the challenge of limited panoramic datasets in 3D reconstruction.

Main Methods:

• Developed SPDET, integrating transformer and panoramic geometry features.
• Utilized a pre-filtered depth-image-based rendering for novel view synthesis in self-supervision.
• Implemented an edge-aware loss function to enhance depth estimation accuracy for panoramas.

Main Results:

• Achieved state-of-the-art performance in self-supervised monocular panoramic depth estimation.
• Demonstrated the effectiveness of SPDET through extensive comparison and ablation studies.
• Successfully reconstructed high-quality depth maps using the proposed network.

Conclusions:

• SPDET effectively overcomes the limitations of supervised panoramic depth estimation.
• The integration of panoramic geometry and transformers enhances depth reconstruction quality.
• The proposed self-supervised approach offers a practical solution for 3D reconstruction using readily available RGB data.