what's the shape of the input batch when training?

[renyuanzhe]

Is it something like [batch_size, sequence length, h ,w ,3]?

[noahzn]

The shape is [batch_size, C, H, W]

[renyuanzhe]

The shape is [batch_size, C, H, W]

Doesn't it need img pairs in a sequence to train the model?

[noahzn]

It uses adjacent frames to compute the loss, but the prediction is on a single frame.

[renyuanzhe]

so the gt of the input img is the adjacent frame of the input img ?

[noahzn]

No. You can read the paper or the source code to get a quick idea of how self-supervised depth estimation works.