stereolabs · JujuDel · Oct 24, 2025 · Oct 24, 2025 · Oct 24, 2025
diff --git a/README.md b/README.md
@@ -2,26 +2,48 @@
 
 This package lets you use the ZED stereo camera in Python 3. The Python API is a wrapper around the ZED SDK which is written in C++ optimized code. We make the ZED SDK accessible from external Python code using Cython.
 
+## Table of Contents
+
+- [Getting started](#getting-started)
+    - [Prerequisites](#prerequisites)
+- [Installing the Python API](#installing-the-python-api)
+    - [Running the install script](#running-the-install-script)
+- [Use the plugin](#use-the-plugin)
+    - [Code](#code)
+    - [Run the tutorials](#run-the-tutorials)
+    - [Run the examples](#run-the-examples)
+- [GPU data retrieval using CuPy](#GPU-data-retrieval-using-CuPy)
+    - [Validate your CuPy setup](#validate-your-CuPy-setup)
+- [Troubleshooting](#troubleshooting)
+    - ["Numpy binary incompatibility"](#numpy-binary-incompatibility)
+    - ["PyTorch on Jetson requiring NumPy 1.x while pyzed requires NumPy 2.x"](#pytorch-on-jetson-requiring-numpy-1x-while-pyzed-requires-numpy-2x)
+    - ["CuPy failed to load libnvrtc.so.1x"](#cupy-failed-to-load-libnvrtc-so1x)
+- [Compiling the Python API from source (only for developers of the python wrapper)](#compiling-the-python-api-from-source-only-for-developers-of-the-python-wrapper)
+- [Support](#support)
+
 ## Getting started
 
 - First, download the latest version of the ZED SDK on [stereolabs.com](https://www.stereolabs.com/developers)
 - For more information, read the ZED [Documentation](https://www.stereolabs.com/docs/app-development/python/install/) and [API documentation](https://www.stereolabs.com/docs/api/python/) or our [Community page](https://community.stereolabs.com)
 
 ### Prerequisites
 
-To start using the ZED SDK in Python, you will need to install the following dependencies on your system:  
-
-- [ZED SDK 5.0](https://www.stereolabs.com/developers/) and its dependency [CUDA](https://developer.nvidia.com/cuda-downloads)
-
-For the ZED SDK 4.2 compatible version, use the [zedsdk_4.X branch](https://github.com/stereolabs/zed-python-api/tree/zedsdk_4.X) or the [4.2 release tag](https://github.com/stereolabs/zed-python-api/releases/tag/v4.2)
+To start using the ZED SDK in Python, you will need to install the following dependencies on your system:
 
-- Python 3.8+ x64
-- [Cython >= 3.0.0](http://cython.org/#download)
-- [Numpy >= 2.0](https://numpy.org/install/)
-- OpenCV Python (optional)
-- PyOpenGL (optional)
+- To use pyzed
+    - [ZED SDK 5.1](https://www.stereolabs.com/developers/) and its dependency [CUDA](https://developer.nvidia.com/cuda-downloads)
+        - For the ZED SDK 5.0 compatible version, use the [zedsdk_5.X branch](https://github.com/stereolabs/zed-python-api/tree/zedsdk_5.X) or the [5.0.7 release tag](https://github.com/stereolabs/zed-python-api/releases/tag/v5.0.7)
+        - For the ZED SDK 4.2 compatible version, use the [zedsdk_4.X branch](https://github.com/stereolabs/zed-python-api/tree/zedsdk_4.X) or the [4.2 release tag](https://github.com/stereolabs/zed-python-api/releases/tag/v4.2)
+    - Python 3.8 to Python 3.14
+    - [Cython >= 3.0.0](http://cython.org/#download)
+    - [Numpy >= 2.0](https://numpy.org/install/)
+- To use most of our samples (optional)
+    - OpenCV Python
+    - PyOpenGL
+- To profit from the GPU acceleration and getting the data on the GPU (optional)
+    - [CuPy](https://cupy.dev/)
 
-Please check your python version with the following command. The result should be 3.8 or higher.
+Please check your python version with the following command. The result should be between 3.8 and 3.14.
 
 ```
 python --version
@@ -65,19 +87,19 @@ $ python get_python_api.py
     Detected platform: 
         linux_x86_64
         Python 3.11
-        ZED SDK 5.0
+        ZED SDK 5.1
     # Downloads and install the whl package
-    -> Checking if https://download.stereolabs.com/zedsdk/5.0/whl/linux_x86_64/pyzed-5.0-cp311-cp311-linux_x86_64.whl exists and is available
-    -> Found ! Downloading python package into /home/user/pyzed-5.0-cp311-cp311-linux_x86_64.whl
+    -> Checking if https://download.stereolabs.com/zedsdk/5.1/whl/linux_x86_64/pyzed-5.1-cp311-cp311-linux_x86_64.whl exists and is available
+    -> Found ! Downloading python package into /home/user/pyzed-5.1-cp311-cp311-linux_x86_64.whl
     -> Installing necessary dependencies
     ...
-    Successfully installed pyzed-5.0
+    Successfully installed pyzed-5.1
 ```
 
 To install it later or on a different environment run : 
 
 ```bash
-$ python -m pip install --ignore-installed /home/user/pyzed-5.0-cp311-cp311-linux_x86_64.wh
+$ python -m pip install --ignore-installed /home/user/pyzed-5.1-cp311-cp311-linux_x86_64.wh
 ```
 
 That's it ! The Python API is now installed.
@@ -91,8 +113,7 @@ Import the packages in your Python terminal or file like this:
 import pyzed.sl as sl
 ```
 
-Vectors operations like norm, sum, square, dot, cross, distance but also simple operations can be done with
-Numpy package.
+Vectors operations like norm, sum, square, dot, cross, distance but also simple operations can be done with numpy package on CPU or CuPy package on GPU.
 
 ### Run the tutorials
 
@@ -102,16 +123,95 @@ The [tutorials](https://github.com/stereolabs/zed-examples/tree/master/tutorials
 
 Please refer to the [examples](https://github.com/stereolabs/zed-examples) README for more informations.
 
+## GPU data retrieval using CuPy
+
+`CuPy` is a NumPy/SciPy-compatible Array Library for GPU-accelerated Computing with Python (see https://cupy.dev/) and the ZED Python API support getting data in its format.
+
+Calling `mat.get_data(sl.MEM_GPU, deep_copy=False)` on a `sl.Mat` object will give you a `cupy.ndarray` object.
+
+### Validate your CuPy setup
+
+**Prerequisites**:
+- A plugged camera
+- pyzed
+- [Numpy >= 2.0](https://numpy.org/install/)
+- [CuPy](https://cupy.dev/) (corresponding to your CUDA version)
+
+You can use script the script `cupy/test_cupy_integration.py`. The script will:
+ - Open a connected ZED camera.
+ - Retrieve an image.
+ - Run some operations and benchmark on the retrieved image.
+ - Display on the terminal the results of the tests.
+
+Without deep-diving into the script content, you can just look at what it prints to validate everything is fine with your setup.
+
+For example, on an Orin NX16 with a ZED X
+
+```
+> python test_cupy_integration.py
+✅ CuPy detected - GPU acceleration available
+   CuPy version: 13.5.1
+   CUDA version: 12080
+ZED SDK CuPy Integration Test
+========================================
+Opening ZED camera...
+[2025-07-31 12:54:15 UTC][ZED][INFO] Logging level INFO
+[2025-07-31 12:54:16 UTC][ZED][INFO] Using GMSL input... Switched to default resolution HD1200
+[2025-07-31 12:54:19 UTC][ZED][INFO] [Init]  Camera FW version: 2001
+[2025-07-31 12:54:19 UTC][ZED][INFO] [Init]  Video mode: HD1200@30
+[2025-07-31 12:54:19 UTC][ZED][INFO] [Init]  Serial Number: S/N 48922857
+[2025-07-31 12:54:19 UTC][ZED][INFO] [Init]  Depth mode: NEURAL
+ZED camera opened successfully.
+Retrieving image data...
+Retrieved image on GPU: 1920x1200
+
+🧪 Testing GPU image processing (basic grayscale conversion)...
+   Input image: (1200, 1920, 4)
+   Processed image: (1200, 1920)
+✅ GPU processing test passed!
+========================================
+
+💾 Testing memory allocation strategies...
+   CPU allocation: (480, 640, 4), float32
+   GPU allocation: (480, 640, 4), float32
+   CPU->GPU transfer: (480, 640, 4)
+   GPU->CPU transfer: (480, 640, 4)
+✅ Memory allocation test passed!
+========================================
+
+🔍 Testing GPU memory usage...
+   Initial GPU memory usage: 0.0 MB
+   After allocation: 15.3 MB
+   After cleanup: 0.0 MB
+✅ GPU memory test passed!
+========================================
+
+🔬 Testing data integrity...
+   Data integrity verified: (2, 2, 4)
+✅ Data integrity test passed!
+========================================
+
+⚡ Running performance benchmark...
+   Benchmark image size: 1920x1200
+   CPU processing (10 iterations): 538.658 milliseconds
+   GPU processing (10 iterations): 93.961 milliseconds
+   Speedup: 5.7x
+🚀 GPU processing is faster!
+========================================
+
+🎉 All tests completed!
+   Your system is ready for GPU-accelerated ZED processing with the Python API!
+```
 
 ## Troubleshooting
 
-###  "Numpy binary incompatiblity"
+###  "Numpy binary incompatibility"
 
 ```
 Traceback (most recent call last):
     ...
     File "__init__.pxd", line 918, in init pyzed.sl
-ValueError: numpy.ufunc size changed, may indicate binary incompatiblity. Expected 216 from C header, got 192 from PyObject
+ValueError: numpy.ufunc size changed, may indicate binary incompatibility. Expected 216 from C header, got 192 from PyObject
 ```
 
 This error usually means numpy isn't installed. To install it, simply run these commands : 
@@ -122,6 +222,28 @@ python3 -m pip install cython
 python3 -m pip install numpy
 ```
 
+### "PyTorch on Jetson requiring NumPy 1.x while pyzed requires NumPy 2.x"
+
+By default, PyTorch on Jetson requires NumPy 1.x (see NVidia's doc [Install PyTorch on Jetson Platform](https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html)). Since pyzed requires NumPy 2.x, it can lead to:
+
+```
+A module that was compiled using NumPy 1.x cannot be run in
+NumPy 2.2.6 as it may crash. To support both 1.x and 2.x
+versions of NumPy, modules must be compiled with NumPy 2.0.
+Some module may need to rebuild instead e.g. with 'pybind11>=2.12'.
+
+If you are a user of the module, the easiest solution will be to
+downgrade to 'numpy<2' or try to upgrade the affected module.
+We expect that some modules will need time to support NumPy 2.
+```
+
+To fix this, until NVidia releases its wheels for PyTorch on Jetson with NumPy 2.x, you can either:
+- compile PyTorch from source with NumPy 2.x: see [this post](https://forums.developer.nvidia.com/t/pytorch-environment-and-zed-sdk-conflicts-on-jetpack-6-2/334056/5) for more information.
+- compile pyzed from source with NumPy 1.x: see [zed-python-api/src](https://github.com/stereolabs/zed-python-api/tree/master/src) for more information.
+
+### "CuPy failed to load libnvrtc.so.1x"
+
+If an __error__ like `CuPy failed to load libnvrtc.so.12` happens when using GPU retrieval, it indicates a mismatch between the installed CuPy version (e.g. cupy-cuda11x) and the CUDA runtime libraries (e.g. 12.x)
 
 ## Compiling the Python API from source (only for developers of the python wrapper)