Table 2.
Quantitative optical flow comparison result on the KITTI flow 2012 and KITTI flow 2015 datasets. The Flying Chairs dataset [35] is denoted as C, Flying Things dataset [37] as T, and KITTI raw dataset as K. The learning approach (G) is supervised by the ground truth, and (U) represents an unsupervised method. Non-Oc indicates a non-occluded region.
| Method | Dataset | KITTI Flow 2012 | KITTI Flow 2015 | Runtime (ms) | |||||
|---|---|---|---|---|---|---|---|---|---|
| EPE (Train) | EPE (Test) | Non-Oc | EPE (Train) | AEE (Train) | Fl (Train) | Non-Oc | GPU | ||
| FlowNetC [35] | C (G) | 9.35 | - | 7.23 | 12.52 | - | 47.93% | 9.35 | 51.4 |
| FlowNetS [35] | C (G) | 8.26 | - | 6.85 | 15.44 | - | 52.86% | 8.12 | 20.2 |
| DSTFlow [38] | K (U) | 10.43 | 12.4 | - | 16.79 | 14.61 | 36.00% | - | - |
| FlowNet2 [36] | C (G) + T (G) | 4.09 | - | 3.42 | 10.06 | 9.17 | 30.37% | 4.93 | 101.6 |
| Ours (w/o flow cons.) | K (U) | 5.33 | 4.72 | 6.21 | 11.28 | 10.11 | 35.42% | 5.13 | 38.1 |
| Ours (w/o synt cons.) | K (U) | 5.02 | 4.55 | 6.69 | 10.74 | 9.58 | 34.17% | 5.08 | 38.2 |
| Ours (w/o FLC) | K (U) | 4.41 | 4.38 | 5.45 | 10.33 | 9.37 | 31.64% | 4.96 | 40.4 |
| Ours | K (U) | 4.16 | 4.07 | 4.52 | 10.05 | 9.12 | 29.78% | 4.81 | 42.8 |