DeFlow is a decoder for scene flow networks in autonomous driving that improves the transition from voxel-based features to point features using Gated Recurrent Units (GRUs). The method addresses the challenge of recovering point-specific features lost during voxelization, which is crucial for accurate scene flow estimation. DeFlow introduces a novel loss function to account for data imbalance between static and dynamic points, enhancing performance in scene flow tasks. Evaluations on the Argoverse 2 dataset show that DeFlow achieves state-of-the-art results, demonstrating better performance and efficiency compared to existing methods. The method is open-sourced and includes a detailed architecture, loss function, and experimental results. DeFlow's decoder design, based on GRU refinement, reconstructs point features within the same voxel, improving the accuracy of scene flow estimation. The method is evaluated on the Argoverse 2 dataset, showing significant improvements in dynamic point flow estimation. DeFlow's design is efficient and suitable for real-time applications in autonomous driving. The method is compared with other approaches, including FastFlow3D, and shows superior performance in terms of accuracy and efficiency. The results highlight the effectiveness of DeFlow in scene flow estimation, particularly in handling dynamic points. The method is also evaluated qualitatively, showing that DeFlow closely matches ground truth flow in speed and angle. The paper concludes that DeFlow is an efficient and high-performance method for autonomous driving in large-scale point clouds. Future work includes self-supervised exploration of DeFlow and the fusion with multi-modality sensors.DeFlow is a decoder for scene flow networks in autonomous driving that improves the transition from voxel-based features to point features using Gated Recurrent Units (GRUs). The method addresses the challenge of recovering point-specific features lost during voxelization, which is crucial for accurate scene flow estimation. DeFlow introduces a novel loss function to account for data imbalance between static and dynamic points, enhancing performance in scene flow tasks. Evaluations on the Argoverse 2 dataset show that DeFlow achieves state-of-the-art results, demonstrating better performance and efficiency compared to existing methods. The method is open-sourced and includes a detailed architecture, loss function, and experimental results. DeFlow's decoder design, based on GRU refinement, reconstructs point features within the same voxel, improving the accuracy of scene flow estimation. The method is evaluated on the Argoverse 2 dataset, showing significant improvements in dynamic point flow estimation. DeFlow's design is efficient and suitable for real-time applications in autonomous driving. The method is compared with other approaches, including FastFlow3D, and shows superior performance in terms of accuracy and efficiency. The results highlight the effectiveness of DeFlow in scene flow estimation, particularly in handling dynamic points. The method is also evaluated qualitatively, showing that DeFlow closely matches ground truth flow in speed and angle. The paper concludes that DeFlow is an efficient and high-performance method for autonomous driving in large-scale point clouds. Future work includes self-supervised exploration of DeFlow and the fusion with multi-modality sensors.