d435: <2% at 2 m
We recently received an order of a number of the d435i cameras.
We have a small team that checks the devices for accuracy before inclusion in the product.
- Defined as: <2% at 2 m
Which I interpret to mean: If I place an object 2m away, the camera should report a depth between 1.96m and 2.04m
At first glance, 37.5% of the cameras failed the test.
Using the On Chip Calibration and TARE Calibration, sometimes multiple times on a single camera, we can get that number down to 12.5%
I have a couple of questions:
- Is my interpretation of the spec correct?
- What should be our expectation of accuracy on delivery?
-
Hi Steve Steve Accuracy can be affected by a range of factors that do not just include distance from the camera.
- Environmental lighting levels
- The texture of the surface being observed by the camera (e.g how plain & smooth or detailed the surface is)
- Whether the surface is light colored or dark grey / black
- How reflective the surface is
So whilst the official estimate of accuracy is <2% at 2 m, the actual accuracy in a real-world scene can vary depending upon the above factors.
RealSense cameras are calibrated in the factory during manufacture and are usually already well calibrated when received by the customer. Cameras can become miscalibrated though if they experience a physical shock such as a hard knock, a drop on the ground or severe vibration, and performing a calibration can correct this. So calibrating the cameras again before integrating them into your product is a good strategy.
-
What is the expectation of how far we should have to go with the calibration?
Using the Chip and Tare...seem absolutely justifiable to me.
But using the more intermediate method, or the OEM method certainly seem like more than should be required by a new product.
And to note: These are tested against your test image, 2m from the camera, in a well-lit environment. -
I have seen cases where a company has a team perform their own custom calibration of newly received stock of cameras before integrating it into their product. So using On-Chip and Tare certainly is justifiable and normal for a commercial enterprise.
On-Chip and Tare will be fine for most situations. If a robust calibration is required then a commercial operation may opt to use the Dynamic Calibration tool instead.
The OEM tool is based on the Dynamic Calibration tool, with the added features of calibrating both intrinsics and extrinsics (the free public version at the above link only calibrates the extrinsics, though this is fine for the majority of users) and the ability to perform calibration remotely via a Linux server. The OEM version, and the large-size OEM Calibration Target Board product for calibrating multiple cameras, are aimed at engineering departments and manufacturing facilities.
https://store.realsenseai.com/buy-intel-realsense-d400-cameras-calibration-target.html
-
We're not doing a custom calibration; we are verifying the promised calibration before inclusion in the product.
And simply trying to get that expected delivery.
We have found the On-Chip and TARE to be helpful to get expected calibration numbers.
But it's not foolproof.
Trying to understand expected delivery quality. -
The official estimate of accuracy for the D435i model is <2% at 2 m so that should be your expectation, and RealSense cameras are usually already well-calibrated by default when received by the customer.
https://realsenseai.com/compare-depth-cameras/
The achieved accuracy during testing by the customer may be outside of that estimate but it will likely be due to factors in the testing environment having a negative impact on accuracy rather than a problem with the received hardware.
What may work best for you is to reset the camera to its factory-new default calibration. This will also resolve any miscalibrations that may have developed during transportation after the camera left the factory. This can be done quickly in the RealSense Viewer tool using instructions here:
https://github.com/IntelRealSense/librealsense/issues/10182#issuecomment-1019854487
You could reset the calibration and then perform your verification afterwards so that you can be confident that the camera is as close to its expected <2% at 2 m accuracy as possible.
Please sign in to leave a comment.
Comments
5 comments