search

visinf / self-mono-sf

Self-Supervised Monocular Scene Flow Estimation (CVPR 2020)
Apache License 2.0
248 stars 47 forks source link

sf=0.0000 during scene flow test #10

Closed NIRVANALAN closed 3 years ago

NIRVANALAN commented 4 years ago

Hi, after executing the evaluation script ./eval_monosf_finetune_kitti_test.sh, the tqdm output keep printingsf=0.0000. image

But the depth evaluation works well. Any suggestions?

thanhmvu commented 4 years ago

I also got this output and found out that it's because the SF loss is set to 0 here for some reason, but I'm not sure why.

https://github.com/visinf/self-mono-sf/blob/c31a1d25cd04b3396056877ace591695002b232a/losses.py#L492

hurjunhwa commented 4 years ago

Hi,

The script ./eval_monosf_finetune_kitti_test.sh outputs the images for uploading in the official KITTI benchmark. As no ground truth available, it doesn't evaluate the results on your local machine, and I just set the loss zero (which has no meaning.)

As depth ground truth is available (for both KITTI or Eigen split), the scripts for the depth evaluation give the results.

NIRVANALAN commented 3 years ago

Hi,

The script ./eval_monosf_finetune_kitti_test.sh outputs the images for uploading in the official KITTI benchmark. As no ground truth available, it doesn't evaluate the results on your local machine, and I just set the loss zero (which has no meaning.)

As depth ground truth is available (for both KITTI or Eigen split), the scripts for the depth evaluation give the results.

Thanks for your reply!