search

facebookresearch / consistent_depth

We estimate dense, flicker-free, geometrically consistent depth from monocular video, for example hand-held cell phone video.
MIT License
1.61k stars 236 forks source link

High resolution information is not recovered #69

Open aj161198 opened 1 year ago

aj161198 commented 1 year ago

I was trying to try out CVD with MiDAS v2 model provided in the code.

Is this the expected behavior, also do you have any suggestions?

longyangqi commented 11 months ago

I was trying to try out CVD with MiDAS v2 model provided in the code.

  • When I ran the default parameters for loss, it seemed like that network was being trained from scratch it missed all the information. For instance it lost all the information after one epoch. Here are the results for first 4 epochs. The input video.
  • When I set the --lambda_reprojection parameter to 10. It did not loose all the information but still lose lot of information. It looks like the network is blurring things out
  • I face the same issue with mc depth. Lot of information is lost. Video

Is this the expected behavior, also do you have any suggestions?

Same problem. Have you fixed it?