The minimum depth range of the Zed2i camera seems to be about 0.5 m even if we set the minimum range as 0.2 m in the Depth Viewer. Is there any depth mode that works well with the 0.2 m range?
Also, we know that passive stereo performs very poorly on white texture less surfaces (a known issue for stereo algorithms), but do you have any suggestions for camera settings to work with such surfaces as the environment around us is not in our control?
The minimum depth does not depend on the depth mode setting. Can you record an SVO by using ZED Explorer and share it with us to show us in what conditions that the minimum depth of 0.2 m is not working?
NEURAL and NEURAL+ depth modes should be used to obtain good depth information from poorly texturized surfaces.
I have recorded the SVO files for 0.2m, 0.3m and 0.45m. The depth data is very poor for 0.2m and 0.3m. It only starts working properly after 0.45m.
Unfortunately, I am not able to upload the recorded SVO files. It throws an error saying “New Users cannot upload attachments”
Also we have a screen recording of all the depth modes for white textureless surfaces at 0.45m. If the above issue regarding uploading the attachments is resolved, we will upload this recording for your reference
You can share the SVO with us by using a file sharing service (google drive, dropbox, etc).
Dropbox link to the recorded SVO files and screen recording:
0.2M.svo2 - svo file recorded at 0.2m
0.3M.svo2 - svo file recorded at 0.3m
0.45M.svo2 - svo file recorded at 0.45m
white_textureless1.svo2 and white_textureless2.svo2 - svo files recorded for white textureless surfaces
sreen_record_different_depth_modes.webm - screen recording showing performance of different depth modes on our system for white texture less surface.
Hi @TAI
thank you for providing the SVO files.
I can say that 0.2m and 0.3m files are almost impossible to be used to extract depth information.
There’s little visual information to be used for stereo matching and all the DEPTH modes fail to extract a valid depth map.
The SVO that was recorded at 0.45 cm instead provides a good depth estimation in NEURAL depth mode that respects the 1% accuracy.
What is the quantifiable metric for the camera to work at the specified 0.2m depth range so that we can replicate it at our end for 0.2m? Is there a minimum number of features or feature distribution requirement for stereo matching to work properly at 0.2 m range?
Also, any comments on the performance of NEURAL and NEURAL+ depth mode for white textureless surface svo recordings ?
Unfortunately, it’s not possible to provide numbers that can specify this type of performance.
AI-powered depth estimation can yield good results with homogeneously colored surfaces. However, in your case, you are beyond the limit. AI is not magic; it cannot possibly extract information where there is none.