search

CVLAB-Unibo / Real-time-self-adaptive-deep-stereo

Code for "Real-time self-adaptive deep stereo" - CVPR 2019 (ORAL)
Apache License 2.0
420 stars 73 forks source link

I want to get right disparity. #62

Open kimsunkyung opened 3 years ago

kimsunkyung commented 3 years ago

I wanted to get right disparity, so I put the right flip image instead of the left image and the left flip image instead of the right image and flip the final result like other stereos. But the result is little strange. How can i get the right dispairty on your online learning adaptation?

AlessioTonioni commented 3 years ago

you should be able to get it by flipping horizzontally the two frames and swapping them. So instead of the left frame you feed to the network the flipped_right and instead of the right you feed the flipped_left.

How are the results strange?

kimsunkyung commented 3 years ago

When i use 'MAD' then result is Disparity Net Restored?: True, number of restored variables: 98 Step: 0 bad3:0.80 EPE:16.67 SSIM:0.20 f/b time:0.025745 Missing time:0:00:09.989148 Step: 100 bad3:0.87 EPE:15.40 SSIM:0.23 f/b time:0.311863 Missing time:0:01:29.816607 Step: 200 bad3:0.85 EPE:19.29 SSIM:0.26 f/b time:0.268517 Missing time:0:00:50.481139 Step: 300 bad3:0.87 EPE:11.50 SSIM:0.29 f/b time:0.264809 Missing time:0:00:23.303167

When i use 'NONE' then result is Step: 0 bad3:0.80 EPE:16.67 SSIM:0.20 f/b time:0.012640 Missing time:0:00:04.904411 Step: 100 bad3:0.89 EPE:19.96 SSIM:0.23 f/b time:0.140082 Missing time:0:00:40.343691 Step: 200 bad3:0.85 EPE:20.36 SSIM:0.26 f/b time:0.139348 Missing time:0:00:26.197436 Step: 300 bad3:0.75 EPE:8.77 SSIM:0.30 f/b time:0.139838 Missing time:0:00:12.305731

when i put the left frame, right frame and use left disparity then MAD result is better than NONE.

But when i put the right flip frame, left frame and use right disparity then NONE result is better than MAD. (this method i put the right flip frame, left frame and i get predicted disparity. And I think that final right disparity is horizontal flip predicted disparity.)

But result is strange.

AlessioTonioni commented 3 years ago

mmm in theory it should work just fine. Are you comparing the predictions to the proper gt? I.e., the right disparity flipped. The error in general seems really big from these few lines you have posted, what kind of data are you using?