CloudFlowHash: normalized flow-driven binary motion encoding for retrieval-assisted dance training in cloud-native systems

With the fast growth of dance videos, game dance charts, and template-based motion resources on cloud platforms, efficient dance motion retrieval has become increasingly relevant to training assistance and choreography support. In many practical settings, users do not simply need to browse a video collection, but need to quickly find motion segments similar to a target action, recall reusable clips from standard templates, or search for reference motions during choreography refinement. These demands place considerable pressure on cloud-side retrieval systems, especially when conventional real-valued motion embeddings are used, since they often bring non-negligible storage overhead and retrieval latency at scale. To address this issue, we propose CloudFlowHash, a cloud-native binary motion encoding framework based on invertible normalized flows. Instead of directly forcing motion representations into discrete codes, the proposed model first learns a continuous latent space with a near-discrete bimodal structure, which makes binary optimization more manageable while still preserving motion semantics. We further introduce a cluster consistency regularization term to reduce the structural mismatch between real-valued embeddings and their binary counterparts. Experiments on six benchmark datasets show that CloudFlowHash achieves clear gains in storage efficiency and retrieval speed, while maintaining competitive ranking accuracy. These results suggest that the proposed framework can serve as an efficient retrieval component for cloud-native dance training and choreography assistance systems, particularly in resource-constrained cloud and edge-cloud environments.