Self-Supervised 3D Scene Flow Estimation and Motion Prediction Using Local Rigidity Prior.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• 3D scene flow estimation and motion prediction are crucial for understanding dynamic environments.
• Existing methods often require extensive labeled data, limiting their applicability.
• Modeling scenes as collections of rigid parts offers a promising direction for motion analysis.

Purpose of the Study:

• To develop a self-supervised method for accurate 3D scene flow estimation.
• To enable class-agnostic motion prediction on point clouds without ground truth.
• To improve the robustness and performance of motion estimation in dynamic scenes.

Main Methods:

• Proposed a self-supervised learning framework using generated pseudo scene flow labels.
• Decomposed point clouds into local regions for piecewise rigid motion estimation.
• Employed an iterative process of correspondence matching, confidence assessment, and rigid transformation updates for robust label generation.
• Utilized a validity mask to filter unreliable pseudo labels during training.

Main Results:

• Achieved state-of-the-art performance on FlyingThings3D and KITTI datasets for self-supervised scene flow estimation.
• Outperformed several supervised methods despite using no ground truth scene flow.
• Demonstrated significant improvements in class-agnostic motion prediction on the nuScenes dataset compared to prior self-supervised approaches.

Conclusions:

• The proposed self-supervised approach effectively learns 3D scene flow and motion prediction from point clouds.
• Piecewise rigid motion estimation provides a viable strategy for generating high-quality pseudo labels.
• The method offers a powerful alternative to supervised learning, reducing the need for extensive manual annotation.