DifFlow3D: Hierarchical Diffusion Models for Uncertainty-Aware 3D Scene Flow Estimation.

Area of Science:

• Computer Vision
• Deep Learning
• 3D Scene Understanding

Background:

• 3D scene flow estimation is vital for dynamic scene analysis but faces challenges with unreliable correlations and lack of uncertainty feedback in existing methods.
• Regression-based approaches often struggle with locally constrained search ranges and do not provide timely uncertainty estimations during training.

Purpose of the Study:

• To propose DifFlow3D, a novel uncertainty-aware network for robust and accurate 3D scene flow estimation.
• To enhance correlation robustness and resilience to challenging dynamic scenes, noisy inputs, and repetitive patterns.
• To dynamically assess the reliability of estimated scene flow through an integrated uncertainty estimation module.

Main Methods:

• Utilizes a conditional probabilistic diffusion model with hierarchical diffusion-based flow estimation blocks.
• Incorporates three key flow-related features as conditions to mitigate generation diversity.
• Introduces a Hidden State Denoising (HSD) strategy to stabilize the reverse denoising process.

Main Results:

• DifFlow3D demonstrates significant EPE3D reduction across four datasets (FlyingThings3D, KITTI 2015, Argoverse, Waymo Open), achieving up to 36.4% improvement.
• Achieves millimeter-level accuracy on the real-scene KITTI dataset when trained solely on synthetic data, showcasing exceptional generalization.
• The diffusion-based refinement module significantly enhances existing scene flow networks as a plug-and-play component.

Conclusions:

• DifFlow3D offers a superior approach to 3D scene flow estimation, addressing limitations of previous methods.
• The network exhibits remarkable generalization capabilities and robustness to various challenging conditions.
• The proposed method holds significant potential for advancing 4D LiDAR reconstruction and dynamic scene understanding tasks.