Supervising radar depth completion using the monocular depth large model.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Radar depth completion has advanced with better networks and datasets.
• Supervision methods for radar depth completion remain underexplored.
• Leveraging large monocular depth models (MDLMs) for radar data presents challenges in metric scale generalization.

Purpose of the Study:

• To propose a novel supervision method for radar depth completion.
• To enhance the generalization capability of MDLMs using metric depth scales.
• To introduce the relative-to-metric conversion (R2MC) module.

Main Methods:

• Developed a relative-to-metric conversion (R2MC) module.
• Utilized sparse LiDAR data for pixelwise local mapping to obtain metric depth scales.
• Integrated the R2MC module with existing backbone networks.

Main Results:

• The R2MC module successfully leveraged the generalization capability of MDLMs.
• Sparse LiDAR data was effectively used to establish metric depth scales.
• Performance improvements were observed when R2MC was combined with different backbones compared to original supervision.

Conclusions:

• The proposed R2MC module offers an effective novel supervision strategy for radar depth completion.
• This approach enhances the performance of various backbone networks by improving metric scale generalization.
• The R2MC module demonstrates versatility and compatibility with different architectures.