**DeFlow: Decoder of Scene Flow Network in Autonomous Driving** Scene flow estimation is crucial for autonomous driving, enabling vehicles to navigate complex environments by predicting the 3D motion field of points. Traditional methods often use voxelization to process large-scale point clouds, but this approach can lose point-specific features, leading to challenges in recovering these features for scene flow tasks. DeFlow addresses this issue by integrating a Gated Recurrent Unit (GRU) refinement module in the decoder to reconstruct point features from voxel-based features. Additionally, DeFlow introduces a novel loss function that accounts for the data imbalance between static and dynamic points, enhancing the accuracy of scene flow estimation. **Contributions:** 1. **GRU Refinement:** DeFlow uses GRU to refine voxel-to-point features, improving the accuracy of scene flow estimation. 2. **Loss Function:** A new loss function is proposed to balance the contribution of static and dynamic points, addressing the data imbalance issue. 3. **State-of-the-Art Performance:** DeFlow achieves state-of-the-art results on the Argoverse 2 scene flow task, demonstrating superior performance and efficiency compared to other methods. **Experimental Results:** - **Dataset:** Argoverse 2, a large-scale autonomous driving dataset. - **Methods Compared:** FastFlow3D, NSFP, FastNSF, Zeroflow, and others. - **Results:** DeFlow outperforms competitors in terms of accuracy, particularly in dynamic point flow estimation, with minimal errors in static points. - **Qualitative Results:** DeFlow provides accurate and detailed scene flow estimates, showing better performance in predicting both speed and motion angle compared to FastFlow3D. **Conclusion:** DeFlow is an efficient and high-performance method for scene flow estimation in autonomous driving, leveraging GRU refinement and a novel loss function to address the challenges of large-scale point clouds. Future work could explore self-supervised learning and multi-modality sensor fusion to further enhance the system's capabilities.**DeFlow: Decoder of Scene Flow Network in Autonomous Driving** Scene flow estimation is crucial for autonomous driving, enabling vehicles to navigate complex environments by predicting the 3D motion field of points. Traditional methods often use voxelization to process large-scale point clouds, but this approach can lose point-specific features, leading to challenges in recovering these features for scene flow tasks. DeFlow addresses this issue by integrating a Gated Recurrent Unit (GRU) refinement module in the decoder to reconstruct point features from voxel-based features. Additionally, DeFlow introduces a novel loss function that accounts for the data imbalance between static and dynamic points, enhancing the accuracy of scene flow estimation. **Contributions:** 1. **GRU Refinement:** DeFlow uses GRU to refine voxel-to-point features, improving the accuracy of scene flow estimation. 2. **Loss Function:** A new loss function is proposed to balance the contribution of static and dynamic points, addressing the data imbalance issue. 3. **State-of-the-Art Performance:** DeFlow achieves state-of-the-art results on the Argoverse 2 scene flow task, demonstrating superior performance and efficiency compared to other methods. **Experimental Results:** - **Dataset:** Argoverse 2, a large-scale autonomous driving dataset. - **Methods Compared:** FastFlow3D, NSFP, FastNSF, Zeroflow, and others. - **Results:** DeFlow outperforms competitors in terms of accuracy, particularly in dynamic point flow estimation, with minimal errors in static points. - **Qualitative Results:** DeFlow provides accurate and detailed scene flow estimates, showing better performance in predicting both speed and motion angle compared to FastFlow3D. **Conclusion:** DeFlow is an efficient and high-performance method for scene flow estimation in autonomous driving, leveraging GRU refinement and a novel loss function to address the challenges of large-scale point clouds. Future work could explore self-supervised learning and multi-modality sensor fusion to further enhance the system's capabilities.