ZED SDK 3.7

What’s New

StereoLabs continues to push the boundaries of depth-sensing by introducing a ground-breaking new depth map with the new NEURAL DEPTH MODE. Based on multiple years of development, we are finally releasing this new depth mode to bring depth sensing applications to the next level.

The ZED SDK 3.7 also features a new lossless and hardware-based compression mode to record Lossless SVO files in real-time, new parameters for fine-tuning in the object detection module, and improvements for the other depth mode (Performance, Quality, and Ultra).

Finally, we are also introducing performance improvements for the SVO recording and Streaming module.

Starting from 3.7.1, the Jetson installers will use the Linux4Tegra version instead of the JetPack version used previously for clarity.

From 3.7.2, adding beta support for JetPack 5.0 (L4T 34.1). From 3.7.7 adding stable support for JetPack 5.0 (L4T 35.1).

From 3.7.5, adding support for Ubuntu 22.04 and CUDA 11.7, the AI module accuracy and performance have been improved.

3.7.7

Platform

• Added support for NVIDIA Jetson JetPack 5.0.2 Production Release (L4T 35.1)
• Removed support for NVIDIA Jetson JetPack 5.0 Developer Preview (L4T 34.1)

Bug Fixes

• Fixed lost camera video settings when grab() triggers an internal recovery. The camera settings were set back to default, whereas they should have been kept as the previous chosen value. This is now the expected behavior.
• Fixed camera settings control in Streaming mode (sender/receiver). The receiver could lose the ability to control the camera settings after a while. It happened when the sender grab() function was triggering an internal recovery and lost control of camera settings.
• Fixed a Heap buffer overflow in SVO recording on Jetson that could lead to unexpected random behavior.

3.7.6

Regression fixes

• Fixed cuDNN packaging for Windows installer
• Fixed sl::Mat GPU allocation on Jetson leading to NO_GPU_COMPATIBLE error

3.7.5

Platform

• Added support for Ubuntu 22.04
• Added support for CUDA up to 11.7

Improvements

• Increased PERSON_HEAD_BOX models inference speed by up to 50% while slightly improving accuracy, especially in the far range
• Increased MULTI_CLASS models accuracy by ~2% while also improving inference speed, by up to 5%
• Improved Positional Tracking accuracy when using Visual Inertial Fusion
• Reduced GPU memory footprint of the SDK, up-to 300Mo

3.7.4

Security Fixes

• Fixed security issue on Linux: the installed udev rules, to open the IMU without root access, were incorrect and could allow privilege escalation.

Improvements

• Removed warning verbose of deprecated functions during compilation

3.7.3

Improvements

• Added AI model download and optimization progress bars

Bug Fixes

• Fixed possible crash in Reid model optimization
• Fixed spatial mapping invalid behavior when using multiple camera simultaneously
• Fixed Custom object detector random infinite loops
• Fixed self-calibration undeterministic behavior when reading the same SVO multiple times.

3.7.2

Bug Fixes

• Fixed possible segmentation fault when calling enableRecording() / enableStreaming() in a separated thread than grab() function.
• Fixed recording freeze issue on Jetson introduced in 3.7.1
• Added security check in custom object tracking when bounding box is not valid.

Platforms

• Added beta support of JP5.0 DP (L4T 34.1)

3.7.1

Improvements

• Manual Exposure and Gain are now kept when camera is closed/re-opened. This ensure that you can setup a manual configuration and keep it with other software that uses the ZED SDK. Note that changes are lost when camera is unplugged.
• Improved automatic grab() recovery for more stability.
• Improved camera sanity check (checks that images are consistent) on Linux and Windows for more stability.

Bug Fixes

• Fixed heap memory overflow when optimizing the NEURAL depth mode model.
• Fixed an issue in sl::Camera::getSensorsData() that could cause a segmentation fault.
• Fixed a keypoint rotation issue when using body tacking in sl::BODY_FORMAT::POSE_34. This was causing unexpected jumps in body tracking.

Platforms

• Added support of JP4.6.1 (L4T 32.7)

3.7.0

New Features

• Introduced new NEURALdepth mode based on AI.
  • This new depth mode can be configured with sl::InitParameters::depth_mode = sl::DEPTH_MODE::NEURAL.
  • It requires an optimization of the AI model used in the backend that must be done only one time for a GPU (similar to the Object Detection/Skeleton tracking module).
  • Use ZED Diagnostic -ais 8 after installing the ZED SDK or launch your application with the new depth mode and wait for the optimization to finish.
• Introduced new lossless and hardware-based compression mode.
  • They are accessible through the enum sl::SVO_COMPRESSION_MODE::H264_LOSSLESS and sl::SVO_COMPRESSION_MODE::H265_LOSSLESS.
  • Those modes require an NVIDIA GPU (when using in ZED Explorer) and provide real-time lossless recording on Desktop and any Jetson (Nano, TX2_NX, Xavier…) with low CPU usage.

Improvements

• Introduced new runtime parameter with sl::RuntimeParameters::remove_saturated_area to activate/de-activate saturation filtering during depth sensing.
  • By default, this parameter is at true to keep the same behavior as the previous version, where saturated areas were removed from the depth map. You can now set it to false to keep those areas in the depth processing.
• Introduced new object detection parameter with sl::ObjectDetectionParameters::filtering_mode associated with the new enumerate sl::OBJECT_FILTERING_MODE.
  • This allows a fine-tuning of the object filtering and tracking after object detection has been done or after custom objects have been ingested.
  • By default, this parameter is at sl::OBJECT_FILTERING_MODE::NMS3D to keep the same behavior as the previous version. You can now also choose NONE or NMS3D_PER_CLASS to adjust the level of filtering done during object tracking.
• Added new function to get Min/Max of current depth map through ERROR_CODE Camera::getCurrentMinMaxDepth(float& min,float& max)
• Improved SVO recording and streaming performances when called internally inside the ZED SDK to reduce overall grab() time.
• Improved error logging inside the ZED SDK to have more precise debug information.

Bug Fixes

• Fixed internet requirement when dependencies are already available on the PC when installing a new SDK.
• Fixed bugs in the AI model during optimization and loading (for Object Detection and Body Tracking modules).