FlowMoP
A stochastic multi-person 3D motion prediction framework based on conditional flow matching. Through B-spline reparameterization and a dual-stream Transformer, it jointly generates diverse future motion trajectories for multiple agents, modeling global trajectories and root-relative poses.
Timeline (last 90 days)
FlowMoP proposes a conditional flow matching framework that jointly generates diverse future motions for multiple agents while modeling global trajectories and root-relative poses.
FlowMoP uses B-spline reparameterization to compress the generative state space into low-dimensional control points while enforcing temporal smoothness, reducing the complexity of the generative model and ensuring the physical plausibility of trajectories.
The dual-stream Transformer provides each agent with interaction-aware conditioning information, handling global trajectories and root-relative poses separately, enabling the model to attend simultaneously to macroscopic displacement and microscopic pose changes.
FlowMoP achieves state-of-the-art performance on two standard datasets, CMU-Mocap (UMPM) and 3DPW.
FlowMoP outperforms the deterministic baselines on joint position error, pose error, and final displacement error.
The research team established the first motion prediction baseline on the WorldPose dataset.
FlowMoP still relies on observed neighbor motion as conditioning, and its ability to model unobserved agents or sudden interactions remains to be further validated.
All Facts (7)
FlowMoP proposes a conditional flow matching framework that jointly generates diverse future motions for multiple agents while modeling global trajectories and root-relative poses.
50%UnverifiedFlowMoP uses B-spline reparameterization to compress the generative state space into low-dimensional control points while enforcing temporal smoothness, reducing the complexity of the generative model and ensuring the physical plausibility of trajectories.
50%UnverifiedThe dual-stream Transformer provides each agent with interaction-aware conditioning information, handling global trajectories and root-relative poses separately, enabling the model to attend simultaneously to macroscopic displacement and microscopic pose changes.
50%UnverifiedFlowMoP achieves state-of-the-art performance on two standard datasets, CMU-Mocap (UMPM) and 3DPW.
50%UnverifiedFlowMoP outperforms the deterministic baselines on joint position error, pose error, and final displacement error.
50%UnverifiedThe research team established the first motion prediction baseline on the WorldPose dataset.
50%UnverifiedFlowMoP still relies on observed neighbor motion as conditioning, and its ability to model unobserved agents or sudden interactions remains to be further validated.
50%