How can calculate the actual depth value from the metric depth output of Depth Anything V2 ? What steps and conversions are necessary to obtain a precise and accurate depth calculation from the raw metric depth data(Depth Anything V2 output)?

[padmakshmcw]

Hi @LiheYoung @bingykang @1ssb
My application environment has a maximum depth range of 12m to 15m (indoor), which changes as I test in different rooms. I am working with videos in which an object is moving, and I want to find the depth of various points on the object.

Current Approach:

1. I have a video with known ground truth depth values and have obtained the Depth Anything metric depth values for specific points.
2. I applied a Gaussian filter to the Depth Anything metric depth values.
3. I fitted a linear model with y=ground truth and x = Depth Anything output.
4. When testing on another video, I achieved an RMSE value of around 250mm, but I aim for higher accuracy with an RMSE below 100mm or 50mm.

Requirements:

1. Suggestions for reducing noise in the Depth Anything output.
2. Methods to increase accuracy and reduce RMSE.
3. Alternative ways to find scaling and shift values.
4. Additional suggestions for accurately determining depth values from RGB images.

Can you provide detailed recommendations or techniques to achieve these goals?

Attachments

All value in graphs are in meters.
- Attaching Depth anything out Raw(DAT_Raw), doing prediction on Gaussian filtered Depth anything output(prediction(Guassian filter)), Ground truth (gt_camera) ( vs number frame 0-510(for one point on object)

- Error variation with frame wise for one point.