GLPanoDepth: Global-to-Local Panoramic Depth Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Reconstruction

Background:

• Depth estimation is crucial for vision applications.
• Omnidirectional cameras enable spherical space depth estimation.
• Existing methods struggle with global context and distortion in panoramic images.

Purpose of the Study:

• To propose a novel learning-based method for dense depth estimation from monocular omnidirectional images.
• To address limitations of fully-convolutional networks in capturing global context and handling equirectangular projection distortion.
• To achieve state-of-the-art performance in panoramic depth estimation.

Main Methods:

• Proposed Cubemap Vision Transformers (CViT) for distortion-free global feature extraction and long-range dependency modeling.
• Developed a GLPanoDepth pipeline integrating CViT with a convolution-based branch for local feature preservation.
• Fused global and local features across multiple scales using a global-to-local strategy.

Main Results:

• CViT architecture provides a global receptive field at every stage, enabling coherent spherical signal predictions.
• The GLPanoDepth method effectively combines global and local features for enhanced depth estimation.
• Achieved state-of-the-art performance in monocular panoramic depth estimation tasks.

Conclusions:

• The proposed CViT architecture is effective for modeling global context in panoramic images.
• The GLPanoDepth method offers a robust approach to monocular panoramic depth estimation.
• This work advances the field by providing a distortion-free, globally coherent depth prediction method.