Hierarchical Features Driven Residual Learning for Depth Map Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Consumer depth cameras offer valuable data for tasks like intelligent vehicles and 3D reconstruction.
• Low-cost depth sensors produce low-resolution depth maps, limiting their practical use.

Purpose of the Study:

• To develop a novel deep network, DepthSR-Net, for depth map super-resolution (SR).
• To improve the spatial resolution of depth maps generated by low-cost sensors.

Main Methods:

• DepthSR-Net utilizes a residual U-Net architecture for hierarchical feature extraction.
• The network employs bicubic interpolation and an input pyramid for multi-level receptive fields.
• Residual learning is applied to infer high-resolution depth maps from low-resolution inputs.

Main Results:

• DepthSR-Net successfully infers high-resolution depth maps from low-resolution inputs.
• Ablation studies confirm the effectiveness of individual network components.
• Experimental results show superior performance compared to existing state-of-the-art methods.

Conclusions:

• DepthSR-Net provides an effective solution for depth map super-resolution.
• The proposed method significantly enhances depth map quality for computer vision tasks.
• The network shows potential for application in other low-level vision problems.