Human, camera trajectories, and world are not independent.
Part 1: Camera-World Geometric Consistency Constraints
We design a constraint that enforces geometric consistency of human mesh between the camera coordinate frame and the world coordinate frame. This helps ensure that the reconstructed human motion makes sense globally — not just from the camera’s point of view, but also in the world space.
Part 2: Test-Time Camera Trajectory Fusion
Use HMR-predicted trajectory for scale & absolute pose; use feed-forward trajectory for smoothness and global consistency.
(1) HMR-predicted trajectory:
- Pros: Contains depth information → less affected by scale ambiguity
- Cons: However, less smooth since it is estimated frame-by-frame from human joints
(2) Feed-forward trajectory:
- Pros: Smoother and globally more reliable
- Cons: But susceptible to scale ambiguity
