Stereo vision SDK that does the heavy lifting for robot eyes
Official samples and tutorials for Stereolabs' ZED depth cameras, bundling SLAM, object detection, body tracking, and spatial mapping behind a single C++/Python API.

What it does
The ZED SDK is the official software stack for Stereolabs’ stereo depth cameras. This repository holds tutorials and ready-to-run samples covering depth sensing, positional tracking, object detection, body tracking, spatial mapping, and multi-camera fusion. It is essentially the on-ramp to a commercial hardware-software ecosystem: you buy the camera, install the SDK, and these examples show you how to extract 3D data without writing the algorithms yourself.
The interesting bit
Version 5.3 quietly fixes several deployment headaches that usually plague camera SDKs. Camera calibration now lives in on-board EEPROM, so initialization no longer needs network access. The SDK can reuse an existing CUDA primary context, which means less fighting with PyTorch over GPU memory. There is also a monotonic timestamp clock option that ignores NTP/PTP jumps—useful if you have ever watched a SLAM system hallucinate after a clock sync.
Key highlights
- Neural depth engine with a “NEURAL LIGHT” model that trades some accuracy for speed; v5.3 claims better robustness in challenging scenes at the same runtime cost.
- Voxel-decimation mode for point clouds: depth-adaptive resolution that thins data where detail is low and keeps it where it matters.
- AES-256 encryption for SVO2 recordings, which is not a feature you expect in a camera SDK but makes sense for robots capturing sensitive environments.
- Zero-copy capture for recording and zero-copy intra-process ROS 2 image topics, both aimed at reducing CPU overhead on Jetson-class hardware.
- SLAM GEN_3 specifically tuned for low-texture environments (think warehouse floors or painted walls).
- Wide platform support: Ubuntu LTS, Windows, Jetson, with bindings for C++, Python, C#, C, plus Unity, Unreal Engine 5, OpenCV, ROS, and ROS 2.
Caveats
- NVIDIA-only: requires a GPU with Compute Capability > 5; no AMD, no CPU fallback, no Apple Silicon.
- Hardware gate: the samples are useless without a ZED camera, and the cheapest model starts around $450. This is not a generic computer-vision playground.
- SVO format lock-in: recordings use Stereolabs’ own SVO/SVO2 format, which lets you replay with full SDK features but may complicate workflows that expect standard video or ROS bags.
Verdict
Grab this if you already own or are committed to buying a ZED camera and want to get depth + perception running fast on NVIDIA hardware. Skip it if you are looking for open, hardware-agnostic SLAM or if your deployment target lacks an NVIDIA GPU.