Fellowship: Motion-X, A Large-scale 3D Expressive Whole-body Human Motion Dataset
#arxiv #artificialintelligence #3Dmotion #humanmotiondataset #SMPLX #motiongeneration #motionannotation
Link to paper: https://paperswithcode.com/paper/moti...
Paper by: Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, Lei Zhang.
Presentation by the Fellowship.ai team: https://fellowship.ai/
Fellowship.ai is brought to you by Launchpad.ai: https://www.launchpad.ai/
Launchpad offers pioneering AI technologies and their transformative applications. For an in-depth understanding of our wide-ranging products and services, visit https://www.launchpad.ai/ai-developme...
Abstract: The paper introduces "Motion-X," a groundbreaking large-scale 3D expressive whole-body motion dataset. While most existing motion datasets focus mainly on body poses, missing out on vital components such as facial expressions, hand gestures, and detailed pose descriptions, Motion-X fills this gap. Additionally, the majority of datasets in this domain come from restrictive lab settings and come with manually labeled textual descriptions, posing scalability challenges. To address these issues, the authors have devised a robust whole-body motion and text annotation pipeline. This innovative system can autonomously annotate motions from both single and multi-view videos, furnishing comprehensive semantic labels for each video and detailed whole-body pose descriptions for every frame. Not only is this method highly accurate, but it also proves to be cost-effective and scalable for continued research and application. With this technology, Motion-X has been crafted, boasting an impressive 13.7M accurate 3D whole-body pose annotations (i.e., SMPL-X) spanning 96K motion sequences across a vast array of scenarios. Additionally, Motion-X provides 13.7M frame-level body pose descriptions and 96K sequence-level semantic labels. Through extensive testing, the annotation pipeline's accuracy was confirmed, and the value of Motion-X was showcased, especially in the realms of producing expressive, diverse, and lifelike motion generation and 3D full-body human mesh reconstruction.