haofeixu/aanet

why do we can't multiply scale in file model.py when interpoate the pred_disp?

dreamhua82 opened this issue · 1 comments

Hi, I have two questions to ask you.
1、In your model.py # line 97, why don't multiply the resize scale as do in refinement module?
Is that mean, we have finished training, This is the last result, so we don't multiply the scale any more when we do interpolate?
2、why don't use the full resolution to construct multi-scale cost volume. (i.e. H, H/2, H/3), is that only for save GPU memory?
Thanks,

  1. the scale factor is acutally multiplied:

    aanet/model.py

    Lines 97 to 98 in f689abf

    pred_disp = F.interpolate(pred_disp, size=(gt_disp.size(-2), gt_disp.size(-1)),
    mode='bilinear', align_corners=False) * (gt_disp.size(-1) / pred_disp.size(-1))

  2. Of course you can use full resolution, but it will cost too much memory.