Python Sample Apps Source Details

Python Bindings

This section provides details about DeepStream application development in Python.
In DeepStream 5.0, python bindings are included in the SDK while sample applications are available


Ubuntu 18.04
DeepStream SDK 5.0 or later
Python 3.6
Gst Python v1.14.5
Gst python should be already installed on Jetson. If missing, install with the following steps:
$ sudo apt-get install python-gi-dev
$ export GST_LIBS="-lgstreamer-1.0 -lgobject-2.0 -lglib-2.0"
$ export GST_CFLAGS="-pthread -I/usr/include/gstreamer-1.0 -I/usr/include/glib-2.0 -I/usr/lib/x86_64-linux-gnu/glib-2.0/include"
$ git clone
$ cd gst-python
$ git checkout 1a8f48a
$ ./ PYTHON=python3
$ ./configure PYTHON=python3
$ make
$ sudo make install

Running Sample Applications

1. Clone the deepstream_python_apps repo under <DeepStream 5.0 ROOT>/sources:
git clone
2. This will create the following directory:
<DeepStream 5.0 ROOT>/sources/deepstream_python_apps
3. The Python apps are under the "apps" directory. Go into each app directory and follow instructions in the README.
The app configuration files contain relative paths for models.

Pipeline Construction

DeepStream pipelines can be constructed using Gst Python, the GStreamer framework's Python bindings. See sample applications main functions for pipeline construction examples.

MetaData Access

DeepStream MetaData contains inference results and other information used in analytics. The MetaData is attached to the Gst Buffer received by each pipeline component. The metadata format is described in detail in the SDK MetaData documentation and API Guide.
The SDK MetaData library is developed in C/C++. Python bindings provide access to the MetaData from Python applications. The bindings are provided in a compiled module, available for x86_64 and Jetson platforms. The module is available as part of the DeepStream SDK installation under /lib directory.
The sample applications get the import path for this module through common/ A is also included for installing the module into standard path:
cd /opt/nvidia/deepstream/deepstream/lib
python3 install
This is currently not automatically done through the SDK installer because python usage is optional.
The bindings generally follow the same API as the underlying C/C++ library, with a few exceptions detailed in sections below.

Memory Management

Memory for MetaData is shared by the Python and C/C++ code paths. For example, a MetaData item may be added by a probe function written in Python and needs to be accessed by a downstream plugin written in C/C++. The deepstream-test4 app contains such usage. The Python garbage collector does not have visibility into memory references in C/C++, and therefore cannot safely manage the lifetime of such shared memory. Because of this complication, Python access to MetaData memory is typically achieved via references without claiming ownership.


When MetaData objects are allocated in Python, an allocation function is provided by the bindings to ensure proper memory ownership of the object. If the constructor is used, the the object will be claimed by the garbage collector when its Python references terminate. However, the object will still need to be accessed by C/C++ code downstream, and therefore must persist beyond those Python references.
Example: To allocate an NvDsEventMsgMeta instance, use this:
msg_meta = pyds.alloc_nvds_event_msg_meta() # get reference to allocated instance without claiming memory ownership
NOT this:
msg_meta = NvDsEventMsgMeta() # memory will be freed by the garbage collector when msg_meta goes out of scope in Python
Allocators are available for the following structs:
NvDsVehicleObject: alloc_nvds_vehicle_object()
NvDsPersonObject: alloc_nvds_person_object()
NvDsFaceObject: alloc_nvds_face_object()
NvDsEventMsgMeta: alloc_nvds_event_msg_meta()
NvDsEvent: alloc_nvds_event()
NvDsPayload: alloc_nvds_payload()
Generic buffer: alloc_buffer(size)

String Access

Some MetaData structures contain string fields. These are accessable in the following manner:
Setting String Fields
Setting a string field results in the allocation of a string buffer in the underlying C++ code.
obj.type = "Type"
This will cause a memory buffer to be allocated, and the string "TYPE" will be copied into it. This memory is owned by the C code and will be freed later. To free the buffer in Python code, use:
NvOSD_TextParams.display_text string now gets freed automatically when a new string is assigned.
Reading String Fields
Directly reading a string field returns C address of the field in the form of an int, for example:
obj = pyds.NvDsVehicleObject.cast(data);
This will print an int representing the address of obj.type in C (which is a char*).
To retrieve the string value of this field, use pyds.get_string(), for example:


Some MetaData instances are stored in GList form. To access the data in a GList node, the data field needs to be cast to the appropriate structure. This casting is done via cast() member function for the target type:
In version v0.5, standalone cast functions were provided. Those are now deprecated and superseded by the cast() functions above:
l_frame = batch_meta.frame_meta_list
frame_meta = pyds.NvDsFrameMeta.cast(

Callback Function Registration

Custom MetaData added to NvDsUserMeta require custom copy and release functions. The MetaData library relies on these custom functions to perform deep-copy of the custom structure, and free allocated resources. These functions are registered as callback function pointers in the NvDsUserMeta structure. Callback functions are registered using these functions:
pyds.set_user_copyfunc(NvDsUserMeta_instance, copy_function)
pyds.set_user_releasefunc(NvDsUserMeta_instance, free_func)
Callbacks need to be unregistered with the bindings library before the application exits. The bindings library currently keeps global references to the registered functions, and these cannot last beyond bindings library unload which happens at application exit. Use the following function to unregister all callbacks:
See the deepstream-test4 sample application for an example of callback registration and unregistration.
Limitation: The bindings library currently only supports a single set of callback functions for each application. The last registered function will be used.

Optimizations and Utilities

Python interpretation is generally slower than running compiled C/C++ code. To provide better performance, some operations are implemented in C and exposed via the bindings interface. This is currently experimental and will expand over time.
The following optimized functions are available:
pyds.NvOSD_ColorParams.set(double red, double green, double blue, double alpha)
This is a simple function that performs the same operations as the following: = red = green = blue
txt_params.text_bg_clr.alpha = alpha
These are performed on each object in, causing the aggregate processing time to slow down the pipeline. Pushing this function into the C layer helps to increase performance.
generate_ts_rfc3339 (buffer, buffer_size)
This function populates the input buffer with a timestamp generated according to RFC3339:

Image Data Access

Decoded images are accessible as NumPy arrays via the get_nvds_buf_surface function. This function is documented in the API Guide.
Please see the deepstream-imagedata-multistream sample application for an example of image data usage.

Sample Application Source Details

The following table shows the location of the Python sample applications under
test application
Path inside the GitHub repo
Simple test application 1
Simple example of how to use DeepStream elements for a single H.264 stream: filesrc→ decode→ nvstreammux→ nvinfer (primary detector)→ nvdsosd→ renderer.
Simple test application 2
Simple example of how to use DeepStream elements for a single H.264 stream: filesrc→ decode→ nvstreammux→ nvinfer (primary detector)→ nvtracker→ nvinfer (secondary classifier)→ nvdsosd→ renderer.
Simple test application 3
Builds on deepstream-test1 (simple test application 1) to demonstrate how to:
Use multiple sources in the pipeline
Use a uridecodebin to accept any type of input (e.g. RTSP/File), any GStreamer supported container format, and any codec
Configure Gst-nvstreammux to generate a batch of frames and infer on it for better resource utilization
Extract the stream metadata, which contains useful information about the frames in the batched buffer
Simple test application 4
Builds on deepstream-test1 for a single H.264 stream: filesrc, decode, nvstreammux, nvinfer, nvdsosd, renderer to demonstrate how to:
Use the Gst-nvmsgconv and Gst-nvmsgbroker plugins in the pipeline
Create NVDS_META_EVENT_MSG type metadata and attach it to the buffer
Use NVDS_META_EVENT_MSG for different types of objects, e.g. vehicle and person
Implement “copy” and “free” functions for use if metadata is extended through the extMsg field
test application
Path inside the GitHub repo
USB camera source application
Simple test application 1 modified to process a single stream from a USB camera.
RTSP output application
Simple test application 1 modified to output visualization stream over RTSP.
Image data access application
Builds on simple test application 3 to demonstrate how to:
Access decoded frames as NumPy arrays in the pipeline
Check detection confidence of detected objects (DBSCAN or NMS clustering required)
Use OpenCV to annotate the frames and save them to file
SSD detector output parser application
Demonstrates how to perform custom post-processing for inference output from Triton Inference Server
Use SSD model on Triton Inference Server for object detection
Enable custom post-processing and raw tensor export for Triton Inference Server via configuration file settings
Access inference output tensors in the pipeline for post-processing in Python
Add detected objects to the metadata
Output the OSD visualization to MP4 file