RAFT-DVC adapts the RAFT optical-flow architecture to digital volume correlation for particle-labeled volumes. The displacement field is solved on a coarse feature grid at 1/s of the input resolution and interpolated back to the voxel grid, so the downsampling factor s sets both the cost and the finest resolvable feature. Rather than ship one network, three arms are trained at s = 2, 4, 8, each matched to a particle size and a displacement band, together with a stated rule for choosing between them.
| arm | downsample s | particle radius | displacement band | training volume |
|---|---|---|---|---|
| s2 | 2 | 2 voxel | 2–4 voxel | 32³ |
| s4 | 4 | 4 voxel | 4–8 voxel | 64³ |
| s8 | 8 | 8 voxel | 8–16 voxel | 128³ |
All three share one architecture and one optimizer schedule, so differences between them come from the input scale alone.
Choosing an arm. Two constraints decide it. Resolvability: the feature grid must still see the particles, which requires a particle diameter of at least about s voxel. Reach: the correlation pyramid searches roughly 8s voxel, with measured collapse points near 6, 13 and 16 voxel for s2, s4 and s8. Deploy the smallest s that resolves the particles and still reaches the expected displacement; within its band an arm’s error scales as roughly 0.017 x s voxel.
What is released:
- Reference implementation and the three trained solvers
- The synthetic-volume generator, which reproduces the benchmark volumes deterministically from the parameters tabulated in the paper
- Headless MATLAB drivers for the classical baselines (local subset DVC, ALDVC, FE-global DVC) used for comparison
- A correlation-sampler impulse test that catches the axis-transposition defect which passes ordinary shape checks and still trains