ZED SDK 3.8

What’s New

ZED SDK 3.8 is a major release that includes new features and performance improvements to depth sensing and body tracking. This update brings up to 50% performance improvement for Neural depth sensing. Human skeleton tracking has also been sped up by +40% on most platforms. The Linux installer size has been reduced by 50%.

Note: On Linux, the “zstd” package is now required to uncompress the installer (see details below).

ZED SDK 3.X API documentation

3.8.2

Bug Fixes

• Fixed performance regression within the object detection module.
• Fixed object detection module random crash that could happen after a long runtime.
• Fixed incorrect error code returned in case of camera disconnection.
• Fixed a memory leak occurring during SVO recording.
• Fixed retrieveImage regression with a custom resolution as parameters.
• Fixed CUDA error when using HD2K resolution in body tracking sample.
• Fixed CUDA error occurring when using the combo parameters enable_right_side_measure + SENSING_MODE::FILL on Jetson.
• Fixed random image freeze that can occur after a long runtime.
• Fixed CUDA installation process on Ubuntu Desktop

Known Issues

• Maxwell and Kepler GPUs (compute capabilities 35, 50 and 52) are not supported using CUDA 11.X, only with CUDA 10.2
• Windows installers are not signed due to certificate expiry, this should be resolved for the next releases

3.8.1

Improvements

• Improved and deployed 3.8 ZED SDK Docker images for size and clarity

Bug Fixes

• Fixed default image sharpening value to be similar to previous ZED SDK versions.
• Fixed setRegionOfInterest function for PERFORMANCE, QUALITY, and NEURAL depth map that previously resulted in artifacts on the edges.
• Fixed findPlaneAtHit that could return shrunk planes.

Known Issues

• Maxwell and Kepler GPUs (compute capabilities 35, 50 and 52) are not supported using CUDA 11.X, only with CUDA 10.2

3.8.0

New Features

• Improved Neural depth runtime by up to 50%
  • Depth FPS now reaches 60FPS on platforms such as Jetson AGX Orin, and 30FPS for 2 cameras.
  • This speed-up is a combination of network optimization, pruning and improved usage of FP16 inference for supported configurations. It applies to all GPUs.
• Added a new sl::Camera::setRegionOfInterest function.
  • It defines a region of interest with a mask to focus on for all the SDK, discarding other parts.
  • It can be useful to filter out known objects that are irrelevant, such as vehicle hoods or drone propellers that can degrade positional tracking or depth estimation by adding noises. Two new samples are now available.
• Added a new sl::ObjectDetectionParameters::allow_reduced_precision_inference parameter.
  • This parameter enables AI inference with quantization to improve runtime while allowing a slight decrease in accuracy. By default all AI models are running in fp16 (half) with no accuracy difference when the GPU supports it (requires a Jetson or a Volta GPU or newer), and fall back to fp32 otherwise. This mode allows int8 inference that provides x4 the throughput on supported operations compared to fp32 and decreases the memory usage.
  • The accuracy loss should not exceed 1-2% on the compatible models. For now, only HUMAN_BODY_xxxx models support it.

Improvements

• Improved sharpening behavior, the lowest setting now disables it. The processing has also been improved and no longer produces color artifacts.
• Added sl::PositionalTrackingParameters::set_gravity_as_origin parameter. It allows you to set the odometry world position using sensor data. For instance, if true (default), the initial_world_transform will be aligned with IMU gravity by keeping the user’s yaw.
• Added sl::PositionalTrackingParameters::depth_min_range parameter to change the minimum range of the odometry. It may be useful to remove some fixed objects in near range.
• Added sl::ObjectDetectionParameters::prediction_timeout_s parameter to tune the prediction time when an object disappears. When an object is not detected anymore, the SDK will predict its positions for a short period of time before its state switches to SEARCHING. This parameter prevents the jittering of the object state when there is a short misdetection
• Improved floor plane detection stability
• Introduced a new sl::POSITIONAL_TRACKING_STATE::SEARCHING_FLOOR_PLANE, when you use the sl::PositionalTrackingParameters::set_floor_as_origin, to know if the positional tracking is started or not
• sl::Mat class has been improved to allow faster transfers. The updateGPUfromCPU and updateCPUfromGPU, as well as copyTo, now include a cudaStream_t parameter to asynchronously retrieve the data when using GPU memory.
• Added a support function, sl::convertImage, to convert image format from sl::MAT_TYPE::U8_C4 to sl::MAT_TYPE::S8_C4, mainly for Unreal Engine.
• On Linux, the installers now use zstandard as the default compression (previously gzip). They are now half the size and twice as fast to extract. The zstd package must be installed (sudo apt install zstd) to use the ZED SDK installer.

Bug Fixes

• Fixed a slight drift in sl::Pose from the sl::Camera::getPosition function when using sl::REFERENCE_FRAME::CAMERA. The cumulative pose from this local pose could introduce a huge drift compared to the sl::Pose from sl::REFERENCE_FRAME::WORLD when loop closure is disabled

Known Issues

• Maxwell and Kepler GPUs (compute capabilities 35, 50 and 52) are not supported using CUDA 11.X, only with CUDA 10.2

Deprecation

• CUDA 10.0 installers are no longer supported with this release. Users are encouraged to migrate to the latest CUDA 11.X.
• While CUDA 10.2 is still supported, users are strongly encouraged to migrate to CUDA 11.X, which provides a performance improvement between 20 to 40% with the same hardware.
• Jetson users are encouraged to move to the latest L4T 35.1 when using a Xavier and Orin, and L4T 32.7 for Nano and TX2. The L4T 32.3, 32.4, 32.5, 32.6 are now deprecated.