SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving

Abstract: Scene flow estimation predicts the 3D motion at each point in successive LiDAR scans. This detailed, point-level, information can help autonomous vehicles to accurately predict and understand dynamic changes in their surroundings. Current state-of-the-art methods require annotated data to train scene flow networks and the expense of labeling inherently limits their scalability. Self-supervised approaches can overcome the above limitations, yet face two principal challenges that hinder optimal performance: point distribution imbalance and disregard for object-level motion constraints. In this paper, we propose SeFlow, a self-supervised method that integrates efficient dynamic classification into a learning-based scene flow pipeline. We demonstrate that classifying static and dynamic points helps design targeted objective functions for different motion patterns. We also emphasize the importance of internal cluster consistency and correct object point association to refine the scene flow estimation, in particular on object details. Our real-time capable method achieves state-of-the-art performance on the self-supervised scene flow task on Argoverse 2 and Waymo datasets. The code is open-sourced at https://github.com/KTH-RPL/SeFlow along with trained model weights.

• The paper presents a self-supervised approach that overcomes data labeling challenges by integrating dynamic point classification in scene flow estimation.
• It combines innovative loss functions and a GRU-based iterative refinement to efficiently capture 3D motion from large-scale LiDAR data.
• Experiments on Argoverse 2 and Waymo datasets show that SeFlow outperforms state-of-the-art methods in accuracy and data efficiency.

An Overview of "SeFlow: A Self-Supervised Scene Flow Method in Autonomous Driving"

The paper presented explores a novel self-supervised approach for scene flow estimation in autonomous driving scenarios, inherently addressing challenges such as data imbalance and the requirement of labeled datasets. The authors introduce SeFlow, a method that successfully integrates dynamic classification into an efficient learning-based scene flow pipeline. This discourse will analyze the primary contributions, methodologies adopted, and implications of the research in the context of autonomous vehicle technology and AI.

Core Contributions and Methodology

Scene flow estimation is a pivotal technology for understanding the dynamic context of autonomous driving environments. It predicts the 3D motion for every point in successive LiDAR scans. While supervised methods necessitate extensive labeled data, SeFlow operates on a self-supervised learning paradigm, overcoming the labeling bottleneck and assimilating the data imbalance challenge often faced by conventional approaches.

The authors introduce a multi-faceted strategy combining dynamic point classification with innovative loss functions, contributing significantly to the self-supervised scene flow task. SeFlow discerns between static and dynamic points, designing specialized objective functions that optimize for varied motion patterns. By leveraging the ray-casting-based dynamic awareness map from existing SLAM frameworks, the model classifies points independently of its inference process, preserving flexibility and enforcing targeted, motion-pattern-specific loss functions. Additionally, by clustering dynamic point candidates, the model ensures consistent and accurate object-level motion estimates.

The realization of SeFlow employs a voxelization method to handle large point cloud data, a practical decision for operating under the real-time requirements crucial to autonomous driving. The architecture employs the DeFlow design, integrating a GRU module with iterative refinement, enhancing inference efficiency without compromising accuracy.

Experimental Evaluation and Results

The effectiveness of SeFlow is validated through comprehensive evaluation on Argoverse 2 and Waymo datasets. The method not only achieves but surpasses state-of-the-art performance in self-supervised settings, showcasing superior data efficiency. The model demonstrated particular efficacy in reducing the End Point Error (EPE), an important metric in assessing scene flow estimation quality.

SeFlow's robust performance against competing methods, both supervised and self-supervised, underscores the advantages of its dynamic point classification and cluster consistency strategies. This is further affirmed in extensive ablation studies elucidating the impact of various loss functions and dataset sizes, showcasing SeFlow's adaptability to reduced training data without significant performance degradation.

Implications and Future Directions

From a practical standpoint, the approach significantly reduces training data demands and computational expense compared to supervised models, making it a promising candidate for deployment in autonomous vehicle systems where real-time processing is critical. The innovative application of clustering for dynamic point analysis and consistency verification holds potential for broader applications in 3D motion analysis beyond autonomous driving.

Future exploration might explore extending the dynamic awareness mapping's capabilities through the integration of multi-modality inputs such as radar or camera data, further enriching the scene understanding. Additionally, refining the handling of distant objects and improving multi-frame temporal consistency could contribute to enhancing the robustness and reliability of scene flow estimation in more complex and varied environments.

In summary, SeFlow represents a significant advance in self-supervised scene flow estimation, characterized by its novel integration of dynamic classification and object-level consistency, offering a balanced blend of precision, efficiency, and scalability essential for future autonomous vehicle systems.