Temporal inconsistency on image sequences about boostingmonoculardepth

[semel1]

Thanks a lot for your prompt response. What method would you suggest for "consistent video depth estimation method"?
Unfortunately BoostYourOwnDepth doesn't solve the problem.
It pulls out not only details but also all existing flickering.
Top row from the left:
Original video,
Heavy Edited Consistent Depth Estimation https://github.com/yu-li/TCMonoDepth ,
Leres .
Bottom row from the left:
Midas3,
"Boosted Depth": Low-res: Consistent Depth Estimation, High-res: Leres,
"Boosted Depth": Low-res: Consistent Depth Estimation, High-res: Midas

comparesement.mp4

[miangoleh]

Hi. The inconsistency is inherent from the base networks we use in this repo which are not designed for video depth estimation.
A simple solution would be generating base estimates from any consistent video depth estimation method and merging it with details from any other MDE method (midas v3 or leres) using our BoostYourOwnDepth repo to get a much better temporal consistency.

[miangoleh]

It seems to be a research problem of its own, and I am not aware of a solution.

[longyangqi]

Thanks a lot for your prompt response. What method would you suggest for "consistent video depth estimation method"? Unfortunately BoostYourOwnDepth doesn't solve the problem. It pulls out not only details but also all existing flickering. Top row from the left: Original video, Heavy Edited Consistent Depth Estimation https://github.com/yu-li/TCMonoDepth , Leres . Bottom row from the left: Midas3, "Boosted Depth": Low-res: Consistent Depth Estimation, High-res: Leres, "Boosted Depth": Low-res: Consistent Depth Estimation, High-res: Midas

comparesement.mp4

Hi，do you have a solution for consistent high-resolution video depth estimation now?