About pre-trained dataset

[puyiwen]

Hi, have you done any ablation experiments on Croco on 3D vision tasks by pre-training on imagenet?

[PhilippeWeinzaepfel]

Hi,

In Table 2 of our NeurIPS'22 paper, we compare pre-training on pairs obtained:

• by sampling two different viewpoints (with some overlap) on Habitat
• by applying two different transforms on single images from ImageNet-1K

The latter case performs quite poorly. Our guess is that there exists some "easy" shortcut for the network to solve the cross-view completion task when the reference image directly comes from the other images by indirectly fitting the transformations.

Best
Philippe