SeFlow is a self-supervised method for scene flow estimation in autonomous driving, which predicts 3D motion at each point in successive LiDAR scans. The method addresses the limitations of current state-of-the-art methods, which require annotated data for training and are limited by the expense of labeling. SeFlow integrates efficient dynamic classification into a learning-based scene flow pipeline, classifying points as static or dynamic to design targeted objective functions for different motion patterns. It also emphasizes the importance of internal cluster consistency and correct object point association to refine scene flow estimation, particularly for object details. The method achieves state-of-the-art performance on the self-supervised scene flow task on Argoverse 2 and Waymo datasets, outperforming all but one supervised method on the leaderboard. The code is open-sourced at <https://github.com/KTH-RPL/SeFlow>. - Scene flow estimation in autonomous driving. - Self-supervised methods. - Data imbalance and object-level motion constraints. - Real-time scene flow estimation. - Dynamic classification and clustering. - Loss functions for self-supervised learning. - Evaluation on Argoverse 2 and Waymo datasets. - Ablation studies on loss terms and training dataset size. - Qualitative results and limitations.SeFlow is a self-supervised method for scene flow estimation in autonomous driving, which predicts 3D motion at each point in successive LiDAR scans. The method addresses the limitations of current state-of-the-art methods, which require annotated data for training and are limited by the expense of labeling. SeFlow integrates efficient dynamic classification into a learning-based scene flow pipeline, classifying points as static or dynamic to design targeted objective functions for different motion patterns. It also emphasizes the importance of internal cluster consistency and correct object point association to refine scene flow estimation, particularly for object details. The method achieves state-of-the-art performance on the self-supervised scene flow task on Argoverse 2 and Waymo datasets, outperforming all but one supervised method on the leaderboard. The code is open-sourced at <https://github.com/KTH-RPL/SeFlow>. - Scene flow estimation in autonomous driving. - Self-supervised methods. - Data imbalance and object-level motion constraints. - Real-time scene flow estimation. - Dynamic classification and clustering. - Loss functions for self-supervised learning. - Evaluation on Argoverse 2 and Waymo datasets. - Ablation studies on loss terms and training dataset size. - Qualitative results and limitations.