About Middleburry checkpoints

[Guptajakala]

what's the difference between middlebury_train.pth and middlebury_finetune.pth? How are they obtained respectively?

[Windsrain]

See Middlebury of Sec 4.3 in the paper.

[Guptajakala]

"we first finetune the Scene Flow pretrained model on the mixed Tartan Air [28], CREStereo Dataset [15], Scene Flow [19], Falling Things [27], InStereo2k [2], CARLA HR-VS [39] and Middlebury [22] datasets using a crop size of 384×512 for 200k steps" this is middlebury_train.pth

"Then we finetune it on the mixed CREStereo Dataset [15], Falling Things [27], InStereo2k [2], CARLA HR-VS [39] and Middlebury [22] datasets using a crop size of 384 × 768 with a batch size of 8 for another 100k steps" this is middlebury_finetune.pth

is that right?

[Windsrain]

Right.

[Guptajakala]

thank you!