LiheYoung / Depth-Anything

[CVPR 2024] Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data. Foundation Model for Monocular Depth Estimation
https://depth-anything.github.io
Apache License 2.0
6.52k stars 501 forks source link

Absolute depth and coordinate system #60

Open rimu123 opened 5 months ago

rimu123 commented 5 months ago

Thank you for your contribution, I have some questions, does the depth of this paper refer to the absolute depth, if it is absolute depth, is the coordinate system the camera coordinate system, and how to return the three-dimensional point under the camera coordinate system according to the depth and internal \ external parameters

MichaelWangGo commented 5 months ago

I think the DepthAnything produces relative depth estimation instead of absolute depth.

LiheYoung commented 5 months ago

Hi @rimu123, as @MichaelWangGo said, the basic Depth Anything models produce relative depth information. However, if you want to obtain absolute depth, you can follow this directory. The output value can be considered as meters.

Nimisha-Pabbichetty commented 5 months ago

@LiheYoung How do we convert the relative depth to absolute depth? I'm assuming we'll have to do the reverse of the shift and scale alignment you do?

akebest commented 4 months ago

I have a question: Why can monocular images produce absolute depth, given that they lack stereoscopic vision and reference objects with known sizes? Has the reliability of their absolute depth been tested?"

Denny-kef commented 3 months ago

@akebest I assume that the network is learning to reference semantic classes with their ground truth size ranges during it's training.