Integrating Sensor Models in Deep Learning Boosts Performance: Application to Monocular Depth Estimation in Warehouse...

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Deep learning excels in computer vision but struggles with robot vision due to active sensing and data scarcity.
• Traditional data-driven deep learning requires extensive datasets, which are often unavailable in specific robotic applications.

Purpose of the Study:

• To develop novel deep learning approaches for monocular depth estimation in robot vision, specifically for warehouse automation.
• To improve deep learning performance in robot vision applications with limited training data.

Main Methods:

• Investigated new methods for monocular depth estimation using three different deep architectures.
• Incorporated sensor models and prior knowledge of robotic active vision into deep learning frameworks.

Main Results:

• The proposed methods consistently improved results and learning performance compared to standard data-driven approaches.
• Effective performance was achieved with significantly fewer training samples.

Conclusions:

• Integrating sensor models and prior knowledge into deep learning enhances performance in robot vision tasks.
• Novel deep learning approaches can overcome data scarcity challenges in robotic applications, reducing data consumption costs.