Stereolabs HDR, MAX, Neural Depth and TERRA Vision
Posted on 31 January 2025


We have written about Stereolabs last year, where we pictured them as a David vs 'Intel' Goliath, picking side with the NVIDIA Jetson platform (it's a whole thing - you should read about it here )
So that train is still going, and it's going faster... Stereolabs just announced a significant update of their product line. Let's take a (down to earth) look at what is noteworthy for robotics!
More for Outdoor Applications

There's a lot coming for outdoor robotic applications. Outdoor vision applications suffer from 2 major hurdles compared to indoor applications: variations in environmental light and exposure to the elements.
We can tell you from experience that very few - if not none - of 3D camera providers have been working on these two topics at the same time. Stereolabs decided to go all-in and created a new outdoor enclosure "PRO Rugged". We're seeing this being used mainly on outdoor vehicles - where the camera is exposed to the elements. It comes with an IP69K rating, a vibration resistant, tilting mount (we hope for teethed tilt adjustment), Gorilla glass and a sunshade. We'll have to wait for Q2'2025 for deliveries.

On the second front, both the ZED X One (the mono camera) and the ZED X (the stereo camera) got an HDR option, which makes use of the Sony ISX031 Automotive-grade HDR sensor. Now there is a catch, this is a Rolling Shutter camera and many of the existing cameras 'on wheels' - including the prior ZED X cameras - are Global Shutter cameras. The latter have the advantage that every pixel has been recorded at the same time, and this eliminates 'tearing' when looking at (or being in) motion. But there is no such thing on the market as a 'Global Shutter HDR' chip. In this case, it's key that the time difference between reading the first row and the last row is as short as possible.
In addition, HDR chips are known to achieve the 'high dynamic range' by taking multiple images and combining these into one. But that leads to a second unwanted effect: blurring. The Sony ISX031 has a split-pixel technology that does away with this effect (as the picture shows).
We yet (at time of this writing) have to test how far we can take the HDR version on moving vehicles. Will both shutters on the ZED X Stereo be hardware synchronized, which is required for low-noise depth estimation? Will the distortion be manageable?
Much of the available Open Source robotics software for cameras - think about calibration, VIO and SLAM - work only well with global shutter cameras. It remains to be seen how well the HDR will perform in these applications.
Larger Depth Range

It's not that a compact stereo camera can't see very far - it can -, it's that the distance measurement becomes extremely noisy, to the point where it becomes unusable. The baseline of the camera is largely to blame here (given good lens quality and sensor resolution), so Stereolabs decided to release a 'ZED X MAX' version of their cameras, adding roughly 50% more range due to a 51mm larger baseline (from 119mm to 170mm). We can understand demand for this type of sensor for the larger vehicles such as self-driving tractors or other off-road vehicles. A 23m range is still not that much if you compare it to LiDAR, which can commonly surpas 100m range - but if you'd like to get rid of your 25m range LiDAR, you come pretty close.
But! There is an even larger range solution possible. Stereolabs has created a tutorial explaining how two of these ZED X One cameras can be configured as a stereo-pair, effectively allowing for baselines allowing for very large ranges (think ~100m).
More Simultaneous Cameras

There has been this thing about ZED X cameras that they only use 2 of the 4 GMSL2 cables. So if you take a look at the splitter in this picture, 2 connectors go to 2 ZED X cameras, and the other 2 are 'lying around' next to your box - they were unusable. This changed thanks to an update to the ZED Drivers. Version 1.3 supports all cables in use - if you operate the stereo cameras at half the speeds (15 FPS instead of 30 FPS) - OR your 2 extra cameras are only mono cameras (like the ZED X One), in that case the stereo cameras work at 30FPS and the mono cameras at 60FPS.

Next Generation Depth Technology
Classical Stereo cameras use 'feature disparity' calculations between left and right images to estimate the depth of each pixel. These 'dumb' methods are however very noisy. The advent of Convolutional Neural Networks (yes, AI), which can take the broader area around the pixel in mind, promise to outperform these dumb methods. NVIDIA released the ESS DNN Stereo Disparity model a few years ago (and updates it regularly) that could do this, and Stereolabs has been playing with a similar approach named 'Neural Depth' since it's SDK 3.7 came out. In this latest version SDK 4.3, they have achieved x2 the performance with improved close range detection and better estimations around contours. Now for Stereolabs, contrary to NVIDIA, this is core product technology and I would not be surprised if Neural Depth outperforms Nvidia's ESS.

TERRA Vision
Finally, we arrive at the most experimental part of the whole stack - at the time of writing TERRA is still in early access mode. The library packs a lot of features in order to create a base framework on which developers of autonomous systems can build on. Stereolabs hopes that you will use their architecture for:
Occupancy and Elevation Maps
Localization and Mapping - Indoor and Outdoor
Object Detection
Semantic Segmentation
We did not yet our hands on TERRA (and invite Stereolabs to give us early access, wink, wink) - so not much to report on yet. If time comes, we will link here to a new update about TERRA.
