notesum.ai
Published at December 9Local Attention Transformers for High-Detail Optical Flow Upsampling
cs.CV
Released Date: December 9, 2024
Authors: Alexander Gielisse1, Nergis Tömen, Jan van Gemert1
Aff.: 1Delft University of Technology

| Method | FlyingThings3D | Sintel (train) | KITTI-15 (train) | Number of parameters | |||
|---|---|---|---|---|---|---|---|
| Train | Test | Clean | Final | F1-epe | F1-all | ||
| GMA (recomputed) | 10.34 | 3.07 | 1.31 | 2.75 | 4.48 | 16.86 | 443K |
| +DC | 9.38 | 2.84 | 1.23 | 2.78 | 4.43 | 16.89 | 443K |
| +DC+FT | 9.53 | 2.86 | 1.24 | 2.79 | 4.55 | 16.92 | 743K |
| +DC+TCU(3/3/3)+FT | 9.51 | 2.73 | 1.22 | 2.83 | 4.52 | 16.80 | 695K |
| +DC+TCU(9/7/5)+FT | 9.24 | 2.75 | 1.21 | 2.80 | 4.36 | 16.26 | 700K |
| +DC+TCU(9/7/5)+FT-aug | 8.97 | 2.61 | 1.18 | 3.01 | 4.50 | 16.64 | 700K |