Question about the result on Sintel benchmark

[MiaoJieF]

May I ask whether the results on sintel in the paper do not use the 'warm start' policy? If so, have you tested the results of using the 'warm start' policy?

[drinkingcoder]

We don't use 'warm start' because we find that 'warm start' didn't help FlowFormer.

[MiaoJieF]

I have another puzzle. Why are the results in your paper slightly different from the results on the Sintel website?

[drinkingcoder]

We published the results of a deprecated model by mistake. Now the numbers on the website are aligned to the numbers reported in the paper.

[MiaoJieF]

Have you updated the model parameters? The result I test on Sintel with ‘things.pth’ without tile is 1.00 and 2.36. Slightly better than 1.01 and 2.40 in the paper.

[drinkingcoder]

We find that the model tested in different environments present slightly different performances. We reproduce your results on a 1080Ti machine but the numbers are 1.01 and 2.40 on a V100 machine.