I am trying to get submillimeter precision with stereo of OAK-D SR for short distances - 20-45cm. Theoretically, it should be possible with subpixel enabled. I take the disparity with 4 fractional bits and convert it to depth on the host using calibration. But even taking a snapshot of the wall from a tripod gives me wavy surface. For sure, I have turned on a dot projector. It looks like the wave pattern correlates with the projected dots. And when I plot the temporal std of each pixel, it doesn't look good. [upl-image-preview uuid=d3a3d4ba-ea31-4876-932c-ea03a461712d url=https://discuss.luxonis.com/assets/files/2026-02-19/1771491020-606304-temporal-std-hist.png alt={TEXT?}] A snapshot of a bookshelf with rich features improves [upl-image-preview uuid=9c0cb94e-eea6-4c97-8c19-090a8b324a91 url=https://discuss.luxonis.com/assets/files/2026-02-19/1771492228-119198-temporal-std-hist.png alt={TEXT?}] The same temporal std for ToF looks way better [upl-image-preview uuid=3abb3570-ee5a-4b7b-b31b-ea746d5e2ace url=https://discuss.luxonis.com/assets/files/2026-02-19/1771491821-490416-temporal-std-hist.png alt={TEXT?}] May I get better precision, or am I too optimistic?

@"OskarSonc"#3466 The oscillation that you mention is present in ToF depth also. [upl-image-preview uuid=4f182d20-a832-45c3-8003-e71ff8042d9a url=https://discuss.luxonis.com/assets/files/2026-02-23/1771831151-395141-image.png alt={TEXT?}] I have averaged the data on 7x7 window for submillimeter precision and filtered from the time series the constant component. This is also a point on the wall, captured from a tripod.

Submillimeter precision

ytulbo

I am trying to get submillimeter precision with stereo of OAK-D SR for short distances - 20-45cm.

Theoretically, it should be possible with subpixel enabled. I take the disparity with 4 fractional bits and convert it to depth on the host using calibration. But even taking a snapshot of the wall from a tripod gives me wavy

surface. For sure, I have turned on a dot projector. It looks like the wave pattern correlates with the projected dots.

And when I plot the temporal std of each pixel, it doesn't look good.

A snapshot of a bookshelf with rich features improves

The same temporal std for ToF looks way better

May I get better precision, or am I too optimistic?

OskarSonc

Hey ytulbo - thanks for sharing the plots. What you’re seeing is expected for stereo on OAK-D SR at this range. The SR (20 mm baseline) is characterized at <2% absolute depth error from 30 cm to 1.8 m, so sub-mm precision is generally not realistic for stereo at 20–45 cm. Subpixel helps reduce quantisation, but stereo still has inherent disparity oscillation/wave effects, especially on low-texture planar surfaces. For this ToF is usually the better option.
We have this page explainngin accuracy for our cameras...

Thanks,
Oskar

ytulbo

@OskarSonc Thanks for the clarification. I am not interested in absolute accuracy, but in the temporal stability, I want to track tiny motions. Does your analysis apply to my case as well?

OskarSonc

ytulbo Yes, mostly the same conclusion applies.

Even if absolute bias is not important, stereo temporal stability on OAK-D SR is still limited by disparity-matching noise/oscillation, especially on low-texture flat surfaces (like your wall test). So sub-mm per-pixel temporal stability is generally not realistic at 20–45 cm with stereo alone. ToF is typically the better modality.

ytulbo

@OskarSonc Where does the oscillation come from? The wall is just a test. As you can see, with rich features, stereo could do better.

OskarSonc

It comes from stereo disparity matching, repetitive surfaces are more ambiguous, so wobble is stronger so rich unique texture is easier to match, so stability improves.

ytulbo

@OskarSonc The oscillation that you mention is present in ToF depth also.

I have averaged the data on 7x7 window for submillimeter precision and filtered from the time series the constant component. This is also a point on the wall, captured from a tripod.

OskarSonc

ToF also has oscillation. In your plot, the bigger effect, however, looks like slow drift, not only jitter. ToF is typically more stable than stereo, but not perfectly flat.

SamiUddin231996

https://medium.com/@muhammadsamiuddinrafayf18/beyond-depth-a-modern-comparative-analysis-of-short-range-stereo-vision-cameras-2023-2026-b13b4bb6a13b

SamiUddin231996

https://medium.com/@muhammadsamiuddinrafayf18/the-evolution-of-short-range-stereo-cameras-a-comparative-analysis-cd8a60265b11