Self-Supervised Monocular Depth Estimation With Multiscale Perception.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Reconstruction

Background:

• Self-supervised learning methods for single-image depth estimation transform depth prediction into an image synthesis task.
• Existing methods use differentiable bilinear samplers, limiting pixel perception in synthetic images and depth maps.
• Current photometric error calculations ignore larger area correlations, leading to local optima in small patches.

Purpose of the Study:

• To extend the perceptual area of depth maps over source images in self-supervised depth estimation.
• To improve the accuracy and robustness of depth prediction from single optical images.

Main Methods:

• Proposed a novel multi-scale method involving downsampling the predicted depth map for image synthesis at various resolutions.
• Introduced a structural similarity (SSIM) pyramid loss to address the locality of photometric error by considering differences across multiple image scales.
• Enabled each pixel in the depth map to perceive more pixels from the source image.

Main Results:

• The multi-scale method significantly extends the perceptual field of depth estimation.
• The SSIM pyramid loss effectively captures larger area correlations, mitigating local optima issues.
• Achieved superior performance on both outdoor and indoor depth estimation benchmarks compared to existing methods.

Conclusions:

• The proposed multi-scale approach and SSIM pyramid loss enhance self-supervised depth estimation from single images.
• The method demonstrates improved 3D information extraction capabilities by considering broader image contexts.
• This work offers a more effective solution for single-image depth prediction tasks.