Feature Pyramid Network Based Efficient Normal Estimation and Filtering for Time-of-Flight Depth Cameras.

Area of Science:

• Computer Vision
• 3D Data Processing
• Robotics

Background:

• Depth images from Time-of-Flight (ToF) cameras are crucial for 3D reconstruction and scene understanding.
• Existing normal estimation and filtering methods can be computationally intensive and may lack robustness.
• Processing 2D depth images directly for 3D point cloud analysis presents challenges.

Purpose of the Study:

• To propose efficient and robust methods for normal estimation and filtering of ToF depth images.
• To develop algorithms that leverage Feature Pyramid Networks (FPN) for low-level 3D point cloud processing.
• To achieve state-of-the-art performance in speed and precision for ToF data analysis.

Main Methods:

• Developed ToFNest for normal estimation and ToFClean for filtering, both based on FPN architecture.
• Projected 2D depth image data into 3D space for processing.
• Utilized task-specific loss functions and evaluated on public and custom datasets.

Main Results:

• ToFNest and ToFClean demonstrate high efficiency in robustness and runtime.
• The proposed methods are an order of magnitude faster than current state-of-the-art algorithms.
• No loss in precision was observed on public datasets compared to existing methods.

Conclusions:

• ToFNest and ToFClean offer a significant advancement in efficient 3D point cloud processing from ToF cameras.
• The FPN-based approach provides a simple yet effective solution for normal estimation and filtering.
• These methods are highly suitable for real-time applications requiring fast and accurate 3D data analysis.