ADRiver-4D performs interpretable encoding of dynamic (4D) point cloud sequences by splitting feature evolution into three explicit mechanisms.
Advection (A) emphasizes motion-consistent feature transport: it reweights neighbors using alignment with the predicted scene flow, so features tend to move and mix along coherent motion rather than arbitrary kNN averaging—useful where structure co-moves (articulated objects, groups of points sharing a motion pattern) and where temporal continuity matters across frames.
Diffusion (D) emphasizes local spatial regularity in feature space: a graph Laplacian–style residual (neighborhood mean minus current state), scaled by a learned diffusion coefficient (and uncertainty when available), which suppresses isolated noise, propagates context across nearby points, and can stabilize sparse or hole-prone regions where single-point features are unreliable—situations loosely related to partial visibility / sampling gaps.
Reaction (R) emphasizes nonlinear, geometry-aware re-shaping of features: an MLP acts on the current feature, predicted flow, and local neighborhood distance statistics (mean/variance), so it can amplify shape discontinuities, contact boundaries, and rapid local configuration changes that linear mixing alone would under-express.
Together, A/D/R expose which dynamical pathway dominates (transport vs smoothing vs local nonlinearity) instead of hiding everything in one black-box residual.
| scene rgb | Advection (A) | Diffusion (D) | Reaction (R) |
|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
🚧 Under Editing ... 🚧