Noise-Robust 3D Pose Estimation Using Appearance Similarity Based on the Distributed Multiple Views.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Traditional 3D pose estimation relies on keypoint distances, which are sensitive to camera calibration errors.
• Noisy environments with slight camera shifts degrade the accuracy of existing methods.

Purpose of the Study:

• To develop a noise-robust framework for multi-person 3D pose estimation.
• To address and fix real-time camera calibration errors in 3D pose reconstruction.

Main Methods:

• A novel framework utilizing appearance similarity for 3D pose estimation.
• Extracting 2D keypoints and appearance features, transferred to a central server.
• Employing geometrical affinity and appearance similarity for pose matching and calibration error detection.

Main Results:

• The proposed method demonstrates robustness against false geometrical errors.
• Achieved significant accuracy improvements: 11.5% on the Campus dataset and 8% on the Shelf dataset.
• Successfully identified and corrected camera calibration errors in real-time.

Conclusions:

• The appearance similarity approach offers a robust solution for 3D pose estimation in challenging environments.
• The framework effectively mitigates the impact of calibration inaccuracies on skeleton reconstruction.
• This method enhances the reliability and accuracy of multi-person 3D pose estimation systems.