Confidence Propagation through CNNs for Guided Sparse Depth Regression.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) typically process regular grid data.
• Adapting CNNs for sparse, irregular data (e.g., autonomous driving, robotics) remains a challenge.

Purpose of the Study:

• To develop an efficient CNN layer for processing sparse, irregularly spaced input data.
• To improve performance and reduce parameters in CNNs for applications like scene depth completion.

Main Methods:

• Proposed an algebraically-constrained normalized convolution layer for CNNs.
• Introduced novel strategies for confidence determination and propagation.
• Developed an objective function to minimize data error and maximize output confidence.
• Investigated fusion strategies for depth and RGB information.
• Utilized output confidence as auxiliary information.

Main Results:

• The proposed normalized convolution network framework achieves superior performance on scene depth completion.
• Requires significantly fewer network parameters (1-5% of state-of-the-art methods).
• Demonstrated effectiveness on KITTI-Depth and NYU-Depth-v2 datasets.

Conclusions:

• The normalized convolution layer is highly effective for sparse input data CNNs.
• Offers a parameter-efficient and high-performance solution for scene depth completion and similar tasks.
• Output confidence is a valuable auxiliary information for improving results.