In the pursuit of enhancing 3D imaging and depth sensing technology, an in-depth analysis and comparison of depth accuracy is crucial. Last month, we published a blog post about dramatically improving our accuracy through a unique calibration procedure.
We have focused on evaluating the depth accuracy of our OAK-D Pro camera to industry counterparts such as StereoLabs™' ZED 2i, and Intel® RealSense™ D435 and D455. By employing a uniform testing methodology across these devices, we aim to present a technical comparison of each camera's depth accuracy.
Our testing procedures were done in controlled environments with varying distances (from 1m to 10m) with fixed lighting conditions and with ideal texture (noise pattern), so a dot projector (active stereo) did not come into account.
Results Overview
The graphs presented in this analysis illustrate the depth accuracy of each camera across different distances. Depth accuracy, in this context, is defined as the deviation of the measured distance from the actual distance. A lower column indicates higher accuracy:
Camera overview
- StereoLabs™ ZED 2i: The ZED 2i performed well, especially at a longer range, thanks to its wide baseline distance (12cm) and high resolution (2K). The downside is that it's also the only camera that does processing on the computer, rather than on-device itself. This means that besides buying a camera, you also need to get a beefy computer if you want to get decent FPS. Another potential downside is the lack of active stereo (dot projector), which will lead to worse accuracy in case of poor texture or low-light environment. As we had an ideal target (noise pattern), neither low-light nor bad texture was present in our tests.
- RealSense™: Here we have to note that RealSense™ has good dot projector quality, which we didn't take into account, as we were measuring an ideal target (noise pattern).
- D435's accuracy is quite good if we take into account the short baseline distance (5cm).
- D455's accuracy wasn't that great, especially at a longer range. As it has a baseline of 9.5cm, it should have better accuracy than D435, but that was not the case.
- Luxonis OAK-D Pro: With a baseline of 7.5cm, it is suitable for both short-range and long-range stereo depth estimation. The benefit of the OAK-D Pro is that besides running on-device stereo, it can also run other compute (AI, video encoding, CV function, etc.), directly on the camera itself. This allows you to use OAK cameras with less powerful host computers (even an RPi Zero), and also enables Standalone mode capability.
Comparison table
Quick overview of important specs:
Full report
The evaluation report - including testing procedure, parameters used, and measurement results - can be found on depth accuracy comparison documentation.
