zed-sdk

https://www.stereolabs.com/developers/release/5.5

5.5.0

ZED SDK 5.5 introduces a new generation of body tracking. GEN_2 is more robust in crowded scenes and to unusual body orientations, and it detects people further from the camera. It is up to 18% faster and up to 14% more accurate than GEN_1. Neural depth gains a precision selector: InitParameters::depth_precision set to INT8 trades a small amount of accuracy for a faster depth runtime and a lower memory footprint, up to 25% faster with 3 cameras or more. The Person & Head detection models detect people about 50% further out and are far more reliable in top-down or partially occluded views, and custom object detection models can now be optimized ahead of time instead of on first use.

Camera support extends to the ZED X One Core, over MIPI through the ZED Link MIPI capture card or over Camera-over-Ethernet through a Holoscan sensor bridge. A new center-of-exposure frame timestamp helps align frames with sensors that timestamp when they measure, such as a LiDAR. On Jetson, the zero-copy capture API exposes its DMA-buffer pool and can return the rectified pair directly as NV12 DMA-BUF buffers, with an experimental backend that keeps rectification entirely off the GPU.

On the platform side, TensorRT 11 is supported with its own installers, any installer now runs on the TensorRT version in its name or newer within the same major, and the Linux installers require fewer system packages. GEN_3 positional tracking is faster per frame and can be told how much GPU it may use. Numerous fixes round out the release, across white balance and video settings, IMU bias, GMSL capture stability, SVO recording, local streaming, spatial mapping and body-tracking fusion.

SDK

Depth

• Added sl::InitParameters::depth_precision to select the precision of NEURAL depth inference. FP16 is the default and keeps the current behavior; INT8 trades a small amount of accuracy for a faster depth runtime and a lower memory footprint, and is supported by DEPTH_MODE::NEURAL only. It is up to 14% faster than FP16 on Orin NX (single camera) and 23% on Thor when using more than 2 cameras. Always safe to set: the SDK falls back to FP16, and says so in the log, when the depth mode or the GPU does not support INT8. Turing GPUs without tensor cores (T1000, T600, T400, T500, T1200, T2000 and the GTX 16 series) are in that case, and FP16 is faster on them anyway. Available in the C++, C, Python and C# APIs, and in ZED Depth Viewer.
• Added OpenEXR (.exr) support to sl::Mat::write() and sl::Mat::read() for float matrices (F32_C1, F32_C3, F32_C4), so a depth map, disparity map or point cloud can be saved and reloaded with its values unchanged, infinite and invalid pixels included, where PNG rescales them to 16-bit integers and PFM writes them as 0. The files are standard EXR and compatible with any OpenEXR-capable tool (with OpenCV, set OPENCV_IO_ENABLE_OPENEXR=1, and note that opencv-python 5 dropped the codec). Writing .exr is float-only: an 8- or 16-bit image returns ERROR_CODE::FAILURE. sl_save_current_depth() and its wrapper equivalents accept an .exr filename as well. Available in the C++, C, Python and C# APIs.
• Fixed sl::Mat::write() and sl::Mat::read() ignoring a file extension written in capitals. depth.PNG, cloud.PLY or image.JPG fell through to the default format for the matrix type, so a F32_C1 depth map named .PNG was written as PFM data and a F32_C3 point cloud as ASCII, under a filename claiming otherwise. The extension is now matched regardless of case.
• Added sl::InitParameters::allow_depth_cuda_graph, which lets the SDK record the depth computation once and replay it at each grab(), lowering CPU cost and depth runtime by 10%, most noticeably on Jetson. At 4 cameras on Thor, depth costs 20% less time and gives back 12 points of CPU. Off by default; enabling it gives the SDK a larger share of the GPU, leaving less headroom fo

https://www.stereolabs.com/developers/release/5.4

5.4.1

SDK

Features / Improvements

• Enabled Intra Process Communication (IPC) on Windows.
• Improved image timestamp accuracy on GMSL cameras (ZED X / ZED X One families): the frame timestamp is now captured in hardware instead of sampled in software, reducing timestamp jitter by ~10× and making it immune to CPU load. Enabled by default; set ZED_SDK_DISABLE_HW_TSC_SOF=1 to restore the previous software-sampled timestamps.
• DGX Spark (GB10): the 5.4.0 USB camera limitation was traced to a platform kernel (xHCI) issue, fixed in Ubuntu's upcoming linux-nvidia-7.0 kernel, with it, ZED USB cameras are fully functional on DGX Spark. Until its official rollout, the SBSA installer now offers to install that kernel on affected GB10 systems (interactive installs only).
• Improved ZED-X IMU temperature reading, a valid temperature is now available directly after sl::Camera::open.

Bug Fixes

• Fixed the IMU camera moving state that was sometimes stuck to MOVING. This was introduced when IMU frequency switched from 400Hz to 200Hz.
• Fixed Camera::open with a SVO2 without high frequency IMU data, the ZED SDK will run in degraded mode and some module wont be available.
• Improved USB camera recovery after a disconnection. Resolves a case where grab() could freeze during automatic reconnection, prevents a hang when a camera is physically unplugged mid-recovery, and ensures images are correct after the camera comes back online. open_timeout_sec is now correctly applied to the reconnection retry window. This apply to both sync and async recovery worflow defined by sl::InitParameters::async_grab_camera_recovery.
• Fixed streaming from a Jetson to a Windows receiver: resolves a wire-format incompatibility that stopped a Windows client from decoding a Jetson stream.
• Fixed effect of sl::InitParameters::open_timeout_sec in sl::Camera::open(), this no longer hangs when a stream has no sender: it now returns after the given timeout instead of freezing.
• Fixed an issue where rolling SVO recording could produce empty files after the first one.
• ZED X One 4K: requesting HDR at a resolution that has no HDR sensor mode (only QHD+ and HD1200 have one) is now rejected with a warning instead of leaving the GMSL capture stack unable to reopen the camera (repeated open timeouts until the application restarts).
• Fixed calibration/configuration file parsing and writing being corrupted by the application's locale.

SLAM

• Update the PT sample for displaying the fisheye image when the input camera have fisheye lens.
• Fixed GEN_3 positional tracking failing to start when IMU fusion was enabled but the source provided no high-frequency IMU data (e.g. SVO v1 files or network streams without full-rate sensor data). When the source still carries an integrated orientation, tracking now runs visual-inertial using it as a rotation-only attitude prior; when no IMU data is available at all, it falls back to pure visual mode instead of never initializing.
• Fixed excessive positional tracking computation that could push grab() runtime to 300–400 ms.
• With Area Memory off, tracking now correctly runs as pure visual-inertial odometry (VIO) instead of the previously refined pipeline. Behavior with Area Memory enabled is unchanged.

Wrappers

Isaac Sim

• [Experimental] Added the ZED Sim2Real camera model, an optional, calibrated post-process that makes simulated RGB match a real ZED X camera (lens blur, vignetting, temporal auto-exposure, color grading, sensor noise). Enable it via the ZED Sim2Real Apply option on the ZED Camera Helper node, or pass --apply_zed_sim2real to the Isaac Lab examples (zed_single.py, zed_multi_env.py).
• Added the ZED Cameras panel: an at-a-glance list of every ZED camera in the stage (thumbnail, prim name, and model) with one-click lens switching. The panel opens automatically at startup and can be toggled from the **Stereola

5.4.0

https://www.stereolabs.com/developers/release/5.4

5.4.0

ZED SDK 5.4 makes the built-in NEURAL depth models noticeably faster: an optimized inference path delivers up to 20% quicker depth inference and ~15% lower GPU load on Jetson Thor — with gains carrying over to Orin and desktop GPUs, and the largest wins on heavier pipelines (more cameras, NEURAL PLUS, or concurrent recording) — all with outputs unchanged. The release also opens the depth pipeline to custom algorithms through the experimental DEPTH_MODE::CUSTOM: an externally computed disparity or depth map — for example from a custom stereo network — can be fed to the SDK each frame, and the full pipeline (point cloud, spatial mapping, object detection, plane detection) then runs on the ingested data. A companion stereo retrieveTensor overload pre-processes both rectified views in a single fused GPU pass, producing inference-ready tensors that bind directly to frameworks such as TensorRT.

Recording and streaming gain a leaner path. SVO recording and outbound streaming now run entirely from Camera::read(), and stereo rectification is performed on demand, so pure recording or streaming pipelines no longer pay its CPU/GPU cost. A new polling API retrieves encoded H264/H265 packets directly — from incoming, outgoing, or recorded streams — without a second encoding session, and new rectified LEFT_NV12 / RIGHT_NV12 views feed hardware encoders without an extra color conversion.

This release also extends the platform reach with experimental support for ARM desktop systems (SBSA: NVIDIA DGX Spark, GH200, and ARM CUDA servers). A new monotonic MONOTONIC_RAW_CLOCK timestamp option yields timestamps immune to both NTP/PTP step and frequency adjustments, the per-camera lens distortion model is now exposed through LENS_DISTORTION_MODEL, and SLAM GEN_3 reports per-pose confidence. Numerous stability and accuracy fixes round out the release across depth, streaming, calibration, and USB camera handling.

SDK

General

Features / Improvements

• Optimized the NEURAL depth models' inference path, improving both runtime and load. Depth inference is up to 20% faster and GPU load is reduced by ~15% on Jetson Thor; gains carry over to Orin and desktop GPUs as well, though the magnitude varies by platform. The improvement scales with pipeline load — heavier workloads (more cameras, more demanding models such as NEURAL PLUS, or recording running on top) tend to see the largest gains. For example, 4 cameras running NEURAL depth on Jetson Thor now reach 60 FPS (up from 53). Outputs are unchanged.
• (Experimental) Added sl::DEPTH_MODE::CUSTOM: feed your own externally computed disparity or depth map (e.g. from a custom stereo network) to the SDK each frame via the new sl::Camera::ingestCustomDepth(), called between Camera::read() and Camera::grab(). The full pipeline then runs on the ingested data (point cloud, spatial mapping, object detection, plane detection, etc.). Disparity or metric depth is accepted at any resolution, in CPU or GPU memory, with an optional confidence map for graded confidence_threshold filtering. Available in the C++, C, Python and C# APIs. As an experimental feature, downstream modules may not be fully robust to arbitrary custom depth/confidence input and can produce unexpected behavior or reduced accuracy.
• Added a stereo overload of sl::Camera::retrieveTensor(Tensor& left, Tensor& right, const TensorParameters&): both rectified views are pre-processed for inference (resize, color conversion, layout, normalization) in a single fused GPU pass, producing densely packed tensors that can be bound directly as network inputs (e.g. TensorRT). Designed for stereo networks and the new DEPTH_MODE::CUSTOM, usable between read() and grab(). C++ only.
• Added a new sl::TIMESTAMP_CLOCK::MONOTONIC_RAW_CLOCK value to the process-wide timestamp clock selector, backed by Linux `CLOCK_MONOTONIC_RAW

https://www.stereolabs.com/developers/release/5.3

5.3.1

SDK

General

Bug Fixes

• Fixed GMSL camera enumeration not recovering after nvargus-daemon is restarted while the application is running. getDeviceList() now returns the connected GMSL cameras again without requiring an application restart.
• Fixed a Desktop (x86) streaming decode latency regression introduced in 5.2.3 with the NVIDIA Video Codec SDK update. The hardware decoder was waiting for the next packet before releasing each frame, adding roughly one frame of latency per received stream (≈20 ms at 60 FPS, more at lower frame rates) on both H264 and H265. Received-stream latency is back to 5.2.1 levels, with no change to image quality, throughput, or decoder load.
• Fixed sl::Camera::reboot(sl::INPUT_TYPE::USB) silently no-op'ing whenever there wasn't exactly one camera in the right state (multi-camera setups, or cameras whose USB stream was wedged but still enumerable on the bus). It now enumerates every connected ZED 2 / ZED 2i and sends a full reset to each over the MCU HID interface — same path as ZED_Explorer --rbt <SN>, applied to every camera.
• Fixed regression on some VIDEO_SETTINGS (SCENE_ILLUMINANCE, DENOISING) retrieved values on GMSL cameras (ZED-X / ZED-X Mini). These values can now be correctly retrieved.
• Fixed sporadic crash in sl::Camera::findFloorPlane on low texture scenes.
• Fixed periodic frame drops when using USB cameras on Linux at HD1080@30.
• Fixed ZED_SVO_Editor: export_to_mcap could fail on certain SVO files — this is now resolved.
• Rate-limited the GNSS covariance-too-low warning. Repeated occurrences are now batched and reported with a count to avoid log spam.

Samples

• Make the positional tracking sample run at the SVO's actual recording framerate instead of as fast as compute allows. The sample now waits between .grab calls so the elapsed time matches the real interval between captured frames in the SVO.

Wrappers

C API

• Fixed sl_get_sensors_data_batch_count growing its internal buffer indefinitely. The cached batch vector is now cleared before each SDK call, so the reported count reflects only samples from the last grab interval and memory stays flat in long-running consumers. Thanks to mkukarkin for the report and patch.

Python

• Fixed sl.Mat.init_mat_cpu and sl.Mat.init_mat_resolution_cpu not wrapping the user-provided buffer: the resulting sl.Mat was empty and shared no memory with the caller, so frames retrieved into it stayed invisible from Python. Breaking change: the ptr argument is now an integer memory address (e.g. my_array.ctypes.data) or a NumPy array, instead of a string. Callers that were passing a string to work around the bug will now get a TypeError and must update their call site — the previous behavior silently produced an unrelated, empty sl.Mat, so no functional usage existed.
• Added set_virtual_stereo_from_serial_numbers, set_virtual_stereo_from_camera_id and set_from_gmsl_port to InitParameters to avoid reference issue when using input attibrute.