XmoPipe: A Pipeline for Large-Scale In-the-Wild Human Motion Dataset Construction

XmoPipe is a newly introduced pipeline designed for constructing large-scale human motion datasets from unconstrained online videos, leveraging recent advancements in monocular motion capture and video-language understanding. The system efficiently retrieves relevant videos based on keywords, extracts 3D body and facial motion data, and generates high-level textual descriptions, allowing for targeted collection of diverse motion types. Its effectiveness is validated through the training of motion reconstruction and generation models, achieving performance on par with traditional motion capture datasets while demonstrating robust cross-dataset generalization, which is crucial for practitioners aiming to develop more adaptable and scalable motion models.