A Performance Validation of the milliFlow Network for Human Motion Sensing using mmWave Radar Scene Flow

Human motion sensing using millimeter-wave (mmWave) radar offers a compelling privacy-preserving alternative to traditional optical methods, but it is fundamentally challenged by the extreme sparsity, noise, and low velocity resolution of the resulting point clouds. This paper presents a comprehensive reproduction study of milliFlow, a deep learning framework designed to address these limitations by estimating scene flow - dense 3D displacement vectors - from sequential radar data. We validate the core architecture, which integrates multi-scale local features, a global attention mechanism, and a Gated Recurrent Unit (GRU) for temporal modeling. A critical component of this framework is its cross-modal automatic annotation pipeline, which generates pseudo-ground-truth labels from synchronized RGB-D data, circumventing the need for manual annotation. Our reproduction, trained for 40 epochs on the original 36,000frame dataset, successfully validates the method's effectiveness. The model achieves a 10.6% improvement in the validation End Point Error (EPE) metric, decreasing from 0.015214 to 0.013600, thereby confirming the findings of the original study. The results substantiate that milliFlow is a robust solution for extracting reliable motion information from sparse radar data, demonstrating its significant potential to enhance downstream human sensing tasks such as activity recognition and body part tracking.