question about the use of synthetic SAR

[Ange1ika]

Please advise: if my goal is to achieve accuracy of about one millimeter with the ZED 2i, is that possible? Given that this is an indoor application—we’re building a device for radar imaging using the Jetson Orin Nano Developer Kit—it’s not a mobile robot, and the task is more akin to that of a surgical robot. In other words, it’s a more static scene. For GT pose data, we’re currently using a Franka Emika robot.

I found that Stereo-Inertial tracking should be more suitable in my case, right?

[zwdoescode]

You can use stereo-inertial mode and you can also use multisensor mode which fuse RGBD + IMU considering ZED 2i also provides depth image.
But I am not sure it can achieve sub-milimeter accuracy.
Please feel free to test it and we would like to know what the accuracy is.
If the accuracy doesn't achieve what you want, do you mind sharing your dataset for additional analysis? Thanks

[Ange1ika]

I will try and let you know sometime this month. Thank you for recommendations!

[aefitorov-nvidia]

Hi @Ange1ika

Thanks for your interest in cuVSLAM! Regarding your questions, I'd like to point out the following:

1. Rolling shutter sensor The ZED 2/Mini camera family uses a rolling shutter sensor, which affects the input image sequence (see the attached example of ZED 2 footage recorded on a Spot quadruped robot) and can degrade tracking quality when a long exposure time is used. Please make sure the camera is configured with the shortest possible exposure/shot time.

2. IMU / Stereo-Inertial tracking The IMU is a helpful aid - but not a silver bullet which could help in scenarios with purely visual input over short time spans, but it won't necessarily improve overall tracking quality, and it may even introduce new jumps due to IMU sensor noise. For that reason, if tracking accuracy is your primary goal, I recommend not enabling the IMU. I also believe that the Jetson Nano is not a powerful enough platform to handle the delivery and processing of a 200/400 Hz data stream from an IMU sensor.

3. Masking Please review the static masking and dynamic masking sections in the docs:

• If you're using distorted images, apply static masking to the outer border of the frame to exclude zones with high distortion caused by side optical-system effects.
• If you have any dynamic objects in the frame, use dynamic masking to exclude them from visual tracking.

4. Resolution and frame rate cuVSLAM itself does not benefit from resolutions higher than VGA, but it does suffer significantly from dropped frames and motion blur. From cuVSLAM's perspective, the best tracking quality is achieved using HD/VGA resolution at 60+ fps - prioritizing a high, stable frame rate over higher resolution