Dense mapping from sparse visual odometry: a lightweight uncertainty-guaranteed depth completion method.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Visual odometry (VO) is crucial for mobile robot spatial perception but generates sparse maps due to insufficient depth data.
• Existing depth completion methods struggle with the extreme sparsity and uneven distribution of depth data from VO.

Purpose of the Study:

• To develop a lightweight network for dense depth map completion from sparse VO data.
• To improve spatial perception accuracy and mapping capabilities for mobile robots.

Main Methods:

• Proposed an Image-Guided Uncertainty-Aware Depth Completion Network (IU-DC).
• Integrated color and spatial information into a normalized convolutional neural network.
• The network outputs dense depth maps and associated uncertainty estimates.

Main Results:

• IU-DC demonstrated superior performance in depth and uncertainty estimation accuracy compared to state-of-the-art (SOTA) methods.
• Real-world mapping tasks showed a 50x increase in reconstructed volumes and 78% ground truth coverage with twice the accuracy.
• The IU-DC network is lightweight, with only 0.6M parameters (3% of SOTA).

Conclusions:

• IU-DC effectively addresses the challenge of sparse depth data in VO for enhanced spatial perception.
• The uncertainty-aware output enables outlier filtering for more precise mapping.
• The proposed method offers a significant improvement in efficiency and performance for robotic mapping.