performance of flownetS with ft* on kitti
shanyuhu opened this issue · 3 comments
@anuragranj , thanks for your nice work. Here I have one question about the performace of flownetS on kitti if it is also finetuned with additional data as used by the spynet as shown in Table 1.
Will flownetS finetuned with additional driving and Monka dataset get better performance than spynet?
FlownetS is not fine-tuned on additional data from driving and Monka. Fine tuning will improve the performance of FlownetS but the performance may be better or worse than SpyNet+ft*.
@anuragranj , thanks for your answering. One more question is why training with additional data like Monka can improve the results of Kitti so much considering the domain shifts between datasets. Normally, the model trained in one domain does not generalize well in another domain. But it seems that it is not true for optical flow.
Could you please share with me your explanation?
Thats because the model sees more training data. Baselines are trained only with flying chairs (a very simple dataset) and sintel (which is too small and has only 1041 images).