Release Notes

• What's New

  ZED SDK 5.0 delivers 5× faster depth sensing, 300% lower compute load on Jetson, and improved depth accuracy in challenging environments. It introduces new depth modes powered by TERRA AI, offering options for higher accuracy and faster depth processing to suit different applications. The update also debuts Magellan™, a vision-based localization system providing centimeter-level positioning indoors and outdoors. Additionally, the capture pipeline architecture has been optimized to improve stability, reduce dropped frames, and lower video latency to under 60ms, delivering one of the lowest latencies in the industry.

• 5.0.3

  Jul 1, 2025

  SDK

  • Changed the default value of svo_real_time_mode from true to false for CameraOne's InitParametersOne, to match the stereo camera's default parameters.
  • Improved the Python API setup script for better error handling, particularly for permission issues and unsupported platforms.

  
  Bug Fixes

  • Fixed a race condition that could occur when repeatedly setting the object detection runtime parameters.
  • Fixed an accuracy regression in the Object Detection module's preprocessing, introduced in version 5.0.0. Bounding boxes could have an incorrect offset when using certain camera resolutions, such as HD1200.
  • Fixed an issue on Windows where the SVO file could not be correctly parsed from within the fusion calibration file.
  • Fixed an issue with ZED X One SVO recording making the resulting SVO size the same size as if recorded with a stereo camera when using the default bitrate. ZED X One SVO recording now has expected sizes.
  • Fixed a memory leak issue that occurred when repeatedly starting and stopping recordings.

  Samples

  • Fixed an incorrect type in OpenGL window initialization in Python samples that could lead to a crash.

  Tools

  • Fixed ZEDSVOEditor -export-to-mcap option when using an SVO recorded with a ZEDOne.

• 5.0.2 GA

  Jun 17, 2025

  SDK

  • Significantly improved the accuracy of NEURAL and NEURAL_PLUS depth modes, even compared to the NEURAL model released in 5.0.0 EA, with enhanced temporal stability and robustness against reflections, repetitive patterns, and complex-structured objects.
  • Added NEURAL model files in the ZED SDK installer to avoid additional downloads at runtime.
  • Added supporting functions isCameraOne, isAvailable, supportHDR, and getAvailableCameraFPS to easily check cameras compatible and available modes and capabilities.
  • Switched default streaming mode to GEN1 while GEN2 is being improved and stabilized. Users can still use streaming GEN2 by setting the environment variable ZED_SDK_STREAM_VERSION=2.

  
  Bug Fixes

  • Fixed a crash that occurred when calling retrieve_objects using CUSTOM_BOX_OBJECTS with the resolution set to VGA.
  • Fixed a rare undefined behavior that led to random values when calling getCurrentFPS.
  • Fixed a rare crash that could occur after repeatedly opening and closing the same camera within the same process.
  • Fixed a possible race condition when opening multiple GMSL cameras from different processes.
  • Fixed a crash when using Fusion with multiple cameras and multiple GPUs.
  • Fixed an accuracy regression with body fitting in NEURAL modes.
  • Fixed a UDPSocket crash when streaming from a Jetson to a desktop.
  • Fixed an issue that prevented enableRecording from working with localhost. This is now supported.
  • Fixed a memory leak in Local Streaming IPC mode (with the ZED Media Server).
  • Fixed a segmentation fault on the sender side in Local Streaming that occurred randomly after some time.
  • Fixed getVideoSettings(sl::VIDEO_SETTINGS::WHITEBALANCE_AUTO) on ZED-X / ZED-XOne, which was returning an incorrect value at launch (noticeable in ZED Explorer with multiple cameras).

  Tools

  • Added focal length information (in millimeters) in ZED Explorer, within the 'Calibration' window.
  • Added SVO auto-repair support in ZED Explorer. It will now attempt to auto-repair corrupted SVO files upon opening, similar to ZED Depth Viewer (or SDK).
  • Fixed ZED Explorer framerate calculator.
  • Fixed model downloads in ZED Diagnostic tool when GPU is not available.Fixed minor UI issues in ZED Explorer.
  • Fixed minor UI issues in ZED Media Server.
  • Fixed video settings control through receiver/host in ZED Media Server. Users can now control virtual ZED-X camera video settings from the receiver side.
  • Fixed stop signal handling in CLI mode for a proper and clean exit in ZED Media Server.
  • Added support for different ZED-XOne camera models connected to ZED Media Server (identical resolution is still required).
  • Added support for multiple JSON configuration files for virtual cameras in CLI mode via the --config option in ZED Media Server.

  Wrappers

  • Improved Python wrapper performance when using multiple cameras in multiple threads.
  • Improved pip installation behavior in the Python wrapper: now uses --force-reinstall by default to avoid issues with stale pyzed after reinstallation.
  • Fixed Docker images with OpenGL display; they are now available again.
  • Fixed minor issues in the C and C# wrappers.

  Samples

  • Improved C++ and Python samples for camera streaming and recording. They are now available and optimized for both single and multi-camera setups.

• 5.0.1 RC

  May 16, 2025

  Please note that 5.0.1 and 5.0.0 are not binary compatible. The Python API package will only work with 5.0.1. You should compile it from source if needed to run with 5.0.0.

  SDK

  • Added SVO real-time mode pausing function.
  • Added dedicated setter functions for the Object Detection and Body Tracking runtime parameters. New retrieveObjects and retrieveBodies functions without runtime parameters have been added as well.
  • Added functions ENU2Geo and Geo2ENU in sl::Fusion, making it easy to compute latitude-longitude coordinates of a 3D point. See the C++ Geoloc playback sample (convert2Geo function).
  • Improved SVO playback for large files.
  • Added covariance output for Positional Tracking GEN_3.
  • Improved Positional Tracking GEN_3 loop closure detection and map optimization.
  • Reduced Positional Tracking GEN_3 area mapping file size and saving/loading times.
  • Improved Region of Interest precision when using Object Detection or Body Tracking. The boxes are now filtered depending on the ratio in the region of interest mask instead of the center and bounding box edges.
  • Improved Python API code autocompletion support for most IDEs by including the .pyi stub in the .whl package.
  • Improved Object Detection fusion internal synchronization — the process is now safer and more efficient.
  • Added ONNX model name as a prefix in the optimized model name when using Object Detection with a custom YOLO-like ONNX model.

  
  Bug Fixes

  • Fixed ZED X auto-recovery function. A regression introduced in 5.0.0 prevented the GMSL camera recovery in case of an interruption.
  • Fixed a rare crash that could occur when enabling NEURAL depth mode.
  • Fixed a deadlock in the Object Detection module with the new internal threaded mode introduced in 5.0.0.
  • Fixed an unclosed file descriptor on Jetson when using SVO H26X input. This could lead to undefined behavior if the Camera class was opened and closed hundreds of times in the same instance processing hardware-decoded SVOs.
  • Fixed a regression when using multiple GPUs. It now correctly uses the selected device ID.
  • Fixed multiple bugs in setSVOPosition functions using index or timestamp input. It should now set the expected frame.
  • Fixed a small memory leak when using Fusion.
  • Fixed AI model optimization log when using ROS.
  • Fixed Object Detection crash when passing an invalid or missing custom YOLO-like ONNX file.
  • Fixed undefined behavior in Object Detection and Body Tracking when processing detector output.
  • Fixed incorrect retrieveImage output when using specific resolutions. The issue could affect grayscale or low-resolution images.
  • Fixed isVideoSettingsSupported function with the AEC_AGC_ROI setting that would return invalid results.

  Tools

  • Fixed ZEDfu NEURAL depth mode optimization.
  • Improved Depth Viewer camera open when switching between camera models.
  • Improved ZED Explorer firmware update GUI on ZED X for clarity.

  Samples

  • Added support for YOLOv11, YOLOv12, and more when using a custom YOLO-like ONNX model. Check out the dedicated documentation page.
  • Updated C++ Spatial Mapping sample.
  • Updated C++ Positional Tracking sample.

  Camera Drivers

  • Added support for Jetson RT Kernel for ZED X camera with dedicated drivers.

  Deprecation

  • Using retrieveObjects and retrieveBodies with runtime parameters is now deprecated. Setting runtime parameters should now be done using the dedicated setters.

• 5.0.0 EA

  Mar 19, 2025

  SDK

  • Introduced new NEURAL models. Added a new NEURAL_LIGHT mode, and updated NEURAL and NEURAL_PLUS. They're faster, more precise, and more robust. The older computer vision modes are now deprecated, the previous NEURAL models can still be accessed for retro compatibility.
  • Added a new sl::Camera::read() function that optionally separates the grab() function into two distinct operations: acquiring new images and processing data. This function does not affect existing applications using grab(), it will detect if the read() function was called or not. It enables parallel computing tasks, such as object detection.
  • Introduced new optimized resolutions for retrieveMeasure, using a sl::Resolution(-1,-1) selecting the internal best compromise for both quality and speed.
  • Improved GMSL image capture to reduce frame drops
  • Enhanced detection reliability when accessing cameras using serial numbers or IDs. When using IDs for ZED X, it now correctly matches the GMSL port
  • Introduced a new early access Positional Tracking mode, GEN_3, which offers enhanced accuracy and stability.
  • Improved floor plane detection stability, now includes better temporal filtering internally.
  • Improved camera health check API and algorithm, better performances, and more feedback.
  • Added video settings VIDEO_SETTINGS::SCENE_ILLUMINANCE to get the Luminance (LUX) of the scene when using ZED X cameras
  • Added support for custom data input into ZED One SVO.
  • Added functions to retrieve the current tracked keypoint when using Positional Tracking GEN_2
  • Added CUDA stream input to Camera::retrieveImage and Camera::retrieveMeasure for improved performance in GPU asynchronous workflows.
  • Added GPU-optimized functions, blobFromImage, and blobFromImages, for converting images to Deep Learning model tensor inputs.
  • Added utility functions, Mat::convertColor, for common color conversions, such as swapping red and blue channels and removing the alpha channel.
  • Added support for Custom OpenCV Calibrations with sl::CameraOne
  • Updated default Image framerate to 30Fps, it provides the best performance compromise
  • Updated IMU data rate for ZED X camera to 200Hz instead of 400Hz, it improves stability and performance, especially for multi-camera setups
  • Updates the default InitParameters::depth_stabilization value set to 30, it provides a more stable depth with minimal motion artifacts
  • Renamed Camera::retrieveObjects to Camera::retrieveCustomObjects for custom object detection. The default behavior remains unaffected, but the new method is required when using CustomObjectDetectionRuntimeParameters.
  • Added new parameters to the CustomObjectDetectionProperties struct:
    • (min|max)_box_(width|height)_meters, to give control to maximum 3D objects dimensions
    • native_mapped_class, to allow remapping a custom label to the SDK’s internal SUBCLASS and profit the internal tuning
    • object_acceleration_preset and max_allowed_acceleration to have better control of the tracked objects' maximum acceleration

  
  Bug Fixes

  • Fixed a potential deadlock in Positional Tracking GEN_2
  • Fixed the function to retrieve unified point clouds from multiple camera setup within the Fusion API
  • Fixed a random issue leading to NAN values in IMU orientation.
  • Fixed the function resetPositionalTracking when using Positional Tracking GEN_2
  • Fixed a random issue leading to NAN position in Object Tracking
  • Fixed random memory leak in Fusion when playing back SVO files
  • Fixed ZED-One UHD 4K SVO recording/playback
  • Fixed a potential deadlock occurring in Object Detection within the Fusion API

  Tools

  • Improved Diagnostic Tool for GMSL camera. It now exports GMSL stack status on Jetson
  • Improved ZED Explorer support for ZED One
  • Improved DepthViewer rendering for smoother display
  • Added support of ZED-One into Sensor Viewer

  Wrappers

  • Introduced GPU support for the Python API, thanks to GitHub user @Rad-hi. This feature is optional and requires the CuPy package.
  • Added custom object detection support in Python and C.
  • Improved Python wrapper performance, there’s now negligible runtime overhead compared to C++
  • Added compatibility for C# .net8.0 framework
  • Added Unreal Engine version 5.5 compatibility
  • Added support for Python 3.13. The legacy Python 3.7 version is now dropped

  Samples

  • Added new custom object detection samples utilizing the read() function for more efficient asynchronous detection.

  Platforms

  • Removed support for legacy Jetpacks 4.6, 5.0, and 5.1 (L4T 32.7, 35.1, and 35.2 respectively)
  • Added support for TensorRT 10, this version introduced Blackwell GPU support (with CUDA 12.8) but dropped Pascal GPU support. For GTX 10X0 series GPU, TensorRT 8 installers should be used.
  • Added support for Blackwell GPU