search

noahzn / Lite-Mono

[CVPR2023] Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation
MIT License
527 stars 58 forks source link

How to get absolute distances on your own dataset? #123

Closed lDarryll closed 6 months ago

lDarryll commented 6 months ago

Hi author, If I collect some data by myself (including depth information labels), how can I get the absolute depth distance on your project?

noahzn commented 6 months ago

Hi, the method itself is to generate relative depth values. If you have metric depth values, you might try scaling the relative depth values using a scaling factor. But the result might be not accurate enough.

lDarryll commented 6 months ago

I am a newbie, how should I calculate the scale factor(base my own dataset), can you tell me where is the code to calculate the scale factor or tell me how to calculate the scale factor, thank you very much!

noahzn commented 6 months ago

Hi, I don't have this code. The question is out of the scope of this repo. Median scaling might be a good starting point to your question.

lDarryll commented 6 months ago

Is it possible to refer to this part of the code to statistically calculate the scaling factor that corresponds to my own data. image

noahzn commented 6 months ago

Yes, you can have a try, but it might not be accurate enough.

lDarryll commented 6 months ago

So when doing model inference I don't have a depth of gt then I can't calculate the absolute distance? In other words, this scaling factor can only be calculated when doing evaluation.

noahzn commented 6 months ago

If you don't have the ground-truth then you cannot use this median scaling.

lDarryll commented 6 months ago

Okay, I got it. Thanks.