OmniPoser
A unified generative model that recovers full-body SMPL motion from arbitrary combinations of heterogeneous sparse sensor inputs. It employs a cross-modal masking mechanism and conditional flow matching, supporting a variety of sensor configurations without retraining.
Timeline (last 90 days)
OmniPoser is a unified framework that recovers full-body SMPL motion from any combination of heterogeneous inputs, including 3D keypoints, sparse IMU rotations, head-mounted 6-DoF poses and any subset thereof.
OmniPoser adopts a cross-modal masking mechanism together with a mask-conditioned encoder, allowing the same architecture to accommodate vastly different sensor configurations without separate adaptation for each combination.
OmniPoser adopts a decoupled dual-stream backbone, generates complete motion via conditional flow matching, and requires only a single-step ODE integration at inference time.
OmniPoser was evaluated in three settings on seven benchmarks, including the real-world Nymeria and Ego-Exo4D datasets.
OmniPoser's performance is comparable to current state-of-the-art methods, with a runtime exceeding 700 FPS and a per-frame inference time of about 1.4 milliseconds.
The paper's abstract does not disclose the specific values of accuracy metrics, nor does it state the performance degradation curve under extremely sparse configurations (such as a single sensor only).
The hardware conditions of the 700 FPS test are not stated in the abstract.
All Facts (7)
OmniPoser is a unified framework that recovers full-body SMPL motion from any combination of heterogeneous inputs, including 3D keypoints, sparse IMU rotations, head-mounted 6-DoF poses and any subset thereof.
50%UnverifiedOmniPoser adopts a cross-modal masking mechanism together with a mask-conditioned encoder, allowing the same architecture to accommodate vastly different sensor configurations without separate adaptation for each combination.
50%UnverifiedOmniPoser adopts a decoupled dual-stream backbone, generates complete motion via conditional flow matching, and requires only a single-step ODE integration at inference time.
50%UnverifiedOmniPoser was evaluated in three settings on seven benchmarks, including the real-world Nymeria and Ego-Exo4D datasets.
50%UnverifiedOmniPoser's performance is comparable to current state-of-the-art methods, with a runtime exceeding 700 FPS and a per-frame inference time of about 1.4 milliseconds.
50%UnverifiedThe paper's abstract does not disclose the specific values of accuracy metrics, nor does it state the performance degradation curve under extremely sparse configurations (such as a single sensor only).
50%UnverifiedThe hardware conditions of the 700 FPS test are not stated in the abstract.
50%