About generalization results in Table 1 and Table 3

[hyBlue]

Hi, thank you for the inspiring work!
I have a question about the results on generalization performance in Table 1 and Table 3.

In Table 1,

In Table 3,

What is the difference between those two results on Sintel even though the results on KITTI are same?

[drinkingcoder]

The size of images in KITTI-15 is quite different from normal images. We train another model FlowFormer# and use the tile technique for flow estimation on KITTI. This is explained in the supplementary. The ablation study in Table 3 is also conducted with FlowFormer# and tile for consistency.

Using tile in Sintel will impact the performance, i.e., 0.95/2.35, on Sintel training set as displayed here, but is still much better than GMA.