High-Fidelity Depth Upsampling Using the Self-Learning Framework.

Area of Science:

• Computer Vision
• Machine Learning
• Sensor Fusion

Background:

• Accurate dense depth map generation is crucial for various applications.
• Existing methods struggle with depth outliers and lack confidence information.
• High-resolution RGB images and LiDAR data offer complementary information for depth estimation.

Purpose of the Study:

• To develop a high-fidelity depth upsampling method using RGB and LiDAR data.
• To explicitly handle depth outliers and incorporate confidence estimation.
• To create a self-learning framework for reliable depth map generation without manual annotation.

Main Methods:

• A self-learning framework that estimates the reliability of upsampled depth maps.
• Integration of high-resolution RGB images and LiDAR sensor data.
• Explicit handling of depth outliers and computation of depth upsampling with confidence.

Main Results:

• Generation of clear, high-fidelity dense depth maps preserving object structure.
• Demonstrated superior performance on Middlebury 2014 and KITTI datasets compared to existing methods.
• Achieved similar average depth errors while retaining at least 3% more valid depth points.

Conclusions:

• The proposed method effectively produces accurate and reliable dense depth maps.
• The self-learning approach enhances depth map quality without human annotation.
• The method shows potential for improving subsequent computer vision algorithms and applications.