ZED SDK 3.6

Release notes of the ZED SDK 3.6.x series
Open in ClaudeOpen in ChatGPT

Release notes of the ZED SDK 3.6 series. Patch releases are listed newest first. Download links and the full changelog, including ZED Tools and integrations, are available on the ZED SDK 3.6 release page.

What’s New

ZED SDK 3.6 introduces a major improvement in the Object Detection module: custom detector inputs can be ingested into the SDK to get 3D detection and tracking from 2D bounding boxes. A new model detects person heads for improved accuracy in crowded environments. The module now also supports ZED 2/2i and ZED Mini.

This release adds support for the newer CUDA 11 minor versions; the CUDA 11 installer can be used for all CUDA 11 versions: 11.0, 11.1, 11.2, 11.3, 11.4, 11.5 (from 3.6.2). The TensorRT library for the AI module has been updated to 8, and the ReID feature is now supported for all platforms. JetPack 4.6 is now supported. Python 3.9 and 3.10 are now supported (from 3.6.3).

Finally, it also adds a new skeleton format, allowing the user to get the local keypoint orientation and position (relative to its parent’s coordinate frame) when body fitting is enabled. This is useful to simplify avatar animation using software like Unity or Unreal for example.

3.6.5

Bug Fixes

  • Fixed camera detection issue with new UVC kernel driver (>=5.12.0)

3.6.4

Improvements

  • Improved ZED, ZED Mini and ZED 2i connection stability on Linux and Nvidia Jetson by adding USB autosuspend udev rules

Bug Fixes

  • Fixed the installation process of CUDA in the Linux installer for CUDA 11
  • Fixed a crash when using corrupted SVOs as input

3.6.3

Improvements

  • Improved streaming receiver on Windows to have a smoother experience.
  • Improved robustness of the AI model download process.
  • Improved depth stabilizer, removed a ghost effect on dynamic edges.
  • Added a new script to help identify Linux platform versions (such as CUDA, Ubuntu and JetPack version) to select compatible ZED SDK installer.
    • Usage example:
      $ bash -c "$(curl -fsSL https://download.stereolabs.com/getzed)"

Bug Fixes

  • Fixed random GPU memory corruption leading to non repeatable output.
  • Fixed a wrong behavior in sl::Camera::getAreaExportState() always returning success, now it returns the real state of the export process.
  • Fixed a bug with AI model deflate where corrupted file could create an allocation issue and segmentation fault.
  • Fixed missing magnetometer values (heading, state, uncalibrated) in SVO or streaming input type.
  • Fixed setSVOPosition(0) that could lead to wrong position in SVO file.

3.6.2

Improvements

  • Added CUDA 11.5 support.
  • Added Windows 11 support.
  • AI models download has been improved to be more robust against CDN issues (auto-retry download with alternative URL).
  • AI models and optimized models are now pre-processed in RAM instead of temporary files, this makes the process faster and more robust especially with concurrent launches.
  • Installer script on Linux now checks first if the packages are already installed and doesn’t require an internet connection if not needed.
  • Improved Linux USB camera acquisition.
  • Reduced stream latency.
  • Improved stream runtime which decreases the number of dropped frames.

Bug Fixes

  • Fixed an issue in Linux platform (Desktop and Jetson), leading to an important CPU load when using SDK as stream receiver.
  • Fixed sl::DEPTH_MODE::ULTRA repeatability (without depth stabilization).
  • Added missing toString function for sl::BODY_FORMAT.
  • Added missing packages libpng and libgomp from docker images

3.6.1

Bug Fixes

  • Upgraded TensorRT version to 8.2 on Desktop to avoid potential issues when using CUDA 11.4 and the object detection module

3.6.0

New Features

API

  • Added function sl::Camera::ingestCustomBoxObjects to extract 3D data and perform tracking of custom 3D bounding boxes.
    • Added sl::CustomBoxObjectData structure to ingest the external bounding box detections.
    • Samples using custom model with TensorRT or OpenCV DNN inference are now available.
  • Added a new detection mode sl::DETECTION_MODEL::PERSON_HEAD_BOX for person head detection and tracking.
    • It may be beneficial for application in crowded scene where persons in background are merely detected by the general purpose person detection model.
  • Added new skeleton format sl::BODY_FORMAT::POSE_34
    • The new format enables the extraction of local keypoint orientation and position when body fitting is enabled. The old format is always available with sl::BODY_FORMAT::POSE_18 and behaves exactly the same as previous ZED SDK version. sl::BODY_FORMAT ObjectDetectionParameters::body_format has been added to Object Detection init parameters to set the output skeleton format.
  • The ZED SDK now informs the user about invalid serial number detected for an MCU and suggests running ZED_Diagnostic -r to recover the camera.
  • Added missing imu/magnetometer, in sl::SensorsConfiguration::imu_magnetometer_transform transform for ZED 2 and ZED 2i
  • Added the new sl::InitParameters::open_timeout_sec to set the maximum period of opening retry in case of connection issues
  • Improved open function
  • Added the fullReboot parameter to sl::Camera::reboot to select a full reboot or a simple UVC module reboot

Bug Fixes

API

  • Fixed sl::Camera::reboot command for ZED 2i
  • Improved grab recovery behavior
  • Fixed wrong values for the camera/IMU transform translation
  • Object and Skeleton tracking are now automatically reset if the user sets a new SVO position that is far enough from the current SVO position. This allows for example to loop on a SVO.

Notes

  • JetPack 4.6 is now supported.
  • The new CUDA 11 installer supports all minor versions (except CUDA 11.0 for Windows which is no longer supported)
  • CUDA 10.0 and JetPack 4.3 installers were deprecated and now no longer support AI module. They can still use all other SDK modules, users are encouraged to upgrade.
  • On Linux, since 3.5, the static version of the SDK doesn’t include the sl_ai library anymore for performance and support reasons. libsl_ai.so is now needed to use the AI module even when using the static sl_zed library. On Jetson the static library is now available for all JetPack versions. The installers are now unified again and include everything while being smaller in size.
  • The ReID feature (object batching) is now supported for all platforms
  • Dropped Ubuntu 16 support that was EOL