TRoPE gives vertex queries explicit mesh connectivity: worst-vertex error falls 39.4% and separation failure rate drops from 49.35% to 34.90%
Synopsis
Feed-forward video-to-4D mesh models can move parts together that are close in space but far apart along the mesh surface, producing sticking artifacts. TRoPE encodes reference-mesh connectivity as a relative position between vertex queries and motion sources, modulates attention through rotary phases, and adds a vertex-wise tail-risk training objective for persistently high-error vertices. On MeshSticking-4D, a benchmark built for this failure mode, it reduces worst-vertex error by 39.4% and lowers the separation failure rate from 49.35% to 34.90% relative to Motion 3-to-4.
Interpretation
In video-to-4D mesh generation, sticking artifacts such as hands attached to the torso or garments clinging to nearby limbs are attributed to cross-part motion leakage: vertices that are spatially close but geodesically distant retrieve similar motion information during decoding. Earlier designs enriched each vertex descriptor through topology-aware aggregation, but the structural relation between a query and its motion source remained implicit during motion retrieval. On the 256 test samples of MeshSticking-4D, relative to the host Motion 3-to-4, worst-vertex error falls from 285.420 to 172.874 and the separation failure rate from 49.349% to 34.898%.
The approach assigns motion sources explicit locations on the mesh and writes the connectivity relation between query and source directly into the attention score. Anchors sampled on the reference mesh carry motion features, multi-scale diffusion features from the graph Laplacian define the topological relation between query and anchor, and rotary phases modulate vertex-anchor attention; training adds a vertex-wise CVaR objective that weights persistently high-error vertices. On MeshSticking-4D this configuration reaches mean Euclidean error 54.588 and lowers event-endpoint tail error EETE25 from 137.872 to 98.191; on Motion-80 worst-vertex error falls from 99.833 to 91.244 while Chamfer distance rises from 11.255 to 12.238.
The topological relation matters more than spatial distance, and how it is injected changes the outcome. With CVaR disabled, absolute Cartesian RoPE, Laplacian positional encoding and absolute heat-kernel-signature RoPE stay close to content-only anchor attention; Euclidean relative RoPE also stays close, whereas geodesic relative RoPE improves performance. In the same comparison TRoPE lowers worst-vertex error from 191.074 to 181.599 and the separation failure rate from 37.268% to 36.152%; replacing rotary modulation with a learned heat-kernel bias gives worst-vertex error 199.651.
Perspective
Evaluation covers 256 test samples of MeshSticking-4D, 80 sequences of Motion-80 and 128 test samples of Motion324-Subset, plus a transfer experiment on the Mesh4D backbone; MeshSticking-4D is built from Mixamo characters and animations through synthetic rendering, with disjoint character and source-animation identities between training and test.
Under severe self-contact, occlusion or highly ambiguous motion evidence, the topological relation cannot fully disambiguate motion sources and residual sticking may persist; suppressing cross-part leakage also does not guarantee accurate local deformation, and large limb displacements can still yield wrong trajectories. Structural guidance depends on valid relations in the reference mesh: components with fewer than 16 vertices receive no spectral features, and 22.2% and 27.8% of contact-to-separation events on Motion-80 and Motion324-Subset respectively have at least one endpoint without valid spectral support.
