Unsupervised Depth Completion Guided by Visual Inertial System and Confidence.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Depth completion from sparse data is crucial for robotics and autonomous systems.
• Existing methods struggle with dynamic scenes, occlusion, and limited computational resources.
• Unsupervised learning approaches face challenges with unlabeled data and effective supervision signals.

Purpose of the Study:

• To develop a real-time, unsupervised depth completion method for dynamic scenes.
• To address challenges like occlusion, limited computational resources, and unlabeled training data.
• To improve the accuracy and efficiency of depth completion in complex environments.

Main Methods:

• A novel compact neural network architecture for depth completion.
• Guidance using visual-inertial system data and confidence metrics.
• Creative design of a confidence guidance loss function using visual-inertial information as the sole supervision.
• Region-specific loss functions (static, dynamic, occluded) to handle pixel mismatch in dynamic scenes.

Main Results:

• The proposed method achieves state-of-the-art accuracy for unsupervised depth completion.
• Demonstrated effectiveness on dynamic datasets and in real-world dynamic scenes.
• The confidence guidance and region-specific losses are sufficient for training depth completion models.
• The network requires only a small number of parameters, indicating efficiency.

Conclusions:

• Unsupervised depth completion in dynamic scenes is feasible and accurate using visual-inertial guidance.
• The proposed compact network and novel loss functions effectively handle occlusions and dynamic objects.
• This method offers a promising solution for real-time depth completion with limited resources.