SeFlow is a self-supervised method for scene flow estimation in autonomous driving. Scene flow estimation predicts 3D motion at each point in successive LiDAR scans, providing detailed information for autonomous vehicles to understand dynamic changes in their environment. Current methods require annotated data, which limits scalability due to the high cost of labeling. SeFlow addresses this by integrating efficient dynamic classification into a learning-based pipeline, enabling self-supervised training without labeled data. The method classifies points as static or dynamic, allowing targeted objective functions for different motion patterns. It emphasizes internal cluster consistency and correct object point association to refine scene flow estimation, particularly for object details. SeFlow achieves state-of-the-art performance on the Argoverse 2 and Waymo datasets, outperforming most supervised methods. Key challenges in self-supervised scene flow include point distribution imbalance and disregard for object-level motion constraints. SeFlow tackles these by clustering dynamic points into object candidates and enforcing consistent flow and correct associations. This reduces flow estimation fragmentation within objects. The method uses a combination of losses, including Chamfer distance, dynamic Chamfer distance, static flow, and dynamic cluster flow. These losses help address data imbalance and ensure consistent object-level motion. The dynamic classification is based on the DUFOMap framework, which uses ray-casting to classify points as dynamic or static. SeFlow's architecture includes a model backbone that enables real-time computation. The method is efficient and effective, achieving real-time performance with high accuracy. It is open-sourced, allowing researchers to access the code and trained model weights. Experiments on the Argoverse 2 and Waymo datasets show that SeFlow outperforms existing self-supervised and supervised methods in terms of flow estimation accuracy. The method is data-efficient and performs well even with limited training data. Qualitative results demonstrate its ability to accurately estimate flow for dynamic objects, including small-scale objects like pedestrians. SeFlow's contributions include a novel self-supervised method that integrates dynamic classification and learning-based strategies. The method addresses key challenges in self-supervised scene flow estimation, including data imbalance and object-level motion constraints. It provides a robust and efficient solution for real-time scene flow estimation in autonomous driving.SeFlow is a self-supervised method for scene flow estimation in autonomous driving. Scene flow estimation predicts 3D motion at each point in successive LiDAR scans, providing detailed information for autonomous vehicles to understand dynamic changes in their environment. Current methods require annotated data, which limits scalability due to the high cost of labeling. SeFlow addresses this by integrating efficient dynamic classification into a learning-based pipeline, enabling self-supervised training without labeled data. The method classifies points as static or dynamic, allowing targeted objective functions for different motion patterns. It emphasizes internal cluster consistency and correct object point association to refine scene flow estimation, particularly for object details. SeFlow achieves state-of-the-art performance on the Argoverse 2 and Waymo datasets, outperforming most supervised methods. Key challenges in self-supervised scene flow include point distribution imbalance and disregard for object-level motion constraints. SeFlow tackles these by clustering dynamic points into object candidates and enforcing consistent flow and correct associations. This reduces flow estimation fragmentation within objects. The method uses a combination of losses, including Chamfer distance, dynamic Chamfer distance, static flow, and dynamic cluster flow. These losses help address data imbalance and ensure consistent object-level motion. The dynamic classification is based on the DUFOMap framework, which uses ray-casting to classify points as dynamic or static. SeFlow's architecture includes a model backbone that enables real-time computation. The method is efficient and effective, achieving real-time performance with high accuracy. It is open-sourced, allowing researchers to access the code and trained model weights. Experiments on the Argoverse 2 and Waymo datasets show that SeFlow outperforms existing self-supervised and supervised methods in terms of flow estimation accuracy. The method is data-efficient and performs well even with limited training data. Qualitative results demonstrate its ability to accurately estimate flow for dynamic objects, including small-scale objects like pedestrians. SeFlow's contributions include a novel self-supervised method that integrates dynamic classification and learning-based strategies. The method addresses key challenges in self-supervised scene flow estimation, including data imbalance and object-level motion constraints. It provides a robust and efficient solution for real-time scene flow estimation in autonomous driving.