Skip to content

README.md

Latest commit

 

History

History
134 lines (99 loc) · 6.88 KB

README.md

File metadata and controls

134 lines (99 loc) · 6.88 KB
Raw

DeepStream Sample Applications

This directory contains GStreamer-based DeepStream sample applications. Each app demonstrates a specific pipeline pattern or feature of the DeepStream SDK.

For build and installation instructions, see build/BUILD.md.

Running Apps

After build/build.sh, binaries are installed to /opt/nvidia/deepstream/deepstream-9.1/bin/.

Apps must be run from their source directory so relative config and model paths resolve correctly:

cd src/apps/sample_apps/<app-name>
/opt/nvidia/deepstream/deepstream-9.1/bin/<app-binary> [options]

Clear the GStreamer plugin cache after a fresh install:

rm -rf ~/.cache/gstreamer-1.0/

Refer to the README in each app's directory for app-specific prerequisites, run commands, and config file descriptions.

Application List

Core Tutorial Apps

App Binary Description
deepstream-app deepstream-app Config-driven reference application — fully configurable multi-stream decode → infer → track → tile → OSD → sink pipeline. The primary DeepStream sample.
deepstream-test1 deepstream-test1-app Single H.264 stream decode → infer → OSD → display. Simplest pipeline.
deepstream-test2 deepstream-test2-app Adds secondary GIE classifiers and tracker to test1.
deepstream-test3 deepstream-test3-app Multiple input streams using uridecodebin.
deepstream-test4 deepstream-test4-app Sends inference metadata to cloud via Kafka/MQTT message broker.
deepstream-test5 deepstream-test5-app Fully configurable multi-stream app with RTSP output.

Inference and Metadata

App Binary Description
deepstream-infer-tensor-meta-test deepstream-infer-tensor-meta-test Access raw tensor output from nvinfer using metadata APIs.
deepstream-preprocess-test deepstream-preprocess-test Custom preprocessing with nvdspreprocess before inference.
deepstream-gst-metadata-test deepstream-gst-metadata-test Attach and read custom GStreamer metadata in a pipeline.
deepstream-user-metadata-test deepstream-user-metadata-test Attach custom user metadata using NvDsUserMeta APIs.
deepstream-image-meta-test deepstream-image-meta-test Attach and access per-frame image metadata.
deepstream-transfer-learning-app deepstream-transfer-learning-app Run custom TensorRT engines trained with Transfer Learning Toolkit.

Video Sources and Demuxing

App Binary Description
deepstream-appsrc-test deepstream-appsrc-test Feed raw buffers into the pipeline using GStreamer appsrc.
deepstream-appsrc-cuda-test deepstream-appsrc-cuda-test Feed CUDA-allocated buffers via appsrc.
deepstream-demuxer-static deepstream-demuxer-static Demux batched stream to individual sinks (static configuration).
deepstream-demuxer-dynamic deepstream-demuxer-dynamic Demux batched stream with runtime source add/remove.
deepstream-dynamicsrcbin-test deepstream-dynamicsrcbin-test Dynamic source bin — add and remove sources at runtime.
deepstream-image-decode-test deepstream-image-decode-test Decode and process a directory of images.
deepstream-testsr deepstream-testsr-app Smart recording — record video segments on trigger.

Analytics and Tracking

App Binary Description
deepstream-nvdsanalytics-test deepstream-nvdsanalytics-test Line-crossing, ROI, and direction analytics using nvdsanalytics.
deepstream-nvof-test deepstream-nvof-test NVIDIA Optical Flow SDK integration.
deepstream-dewarper-test deepstream-dewarper-test Fisheye / 360° dewarping with nvdewarper.
deepstream-multigpu-nvlink-test deepstream-multigpu-nvlink-test Multi-GPU pipeline using NVLink for buffer sharing. x86 dGPU only — not supported on Jetson.

3D Sensing

App Binary Description
deepstream-3d-action-recognition deepstream-3d-action-recognition-app 3D action recognition from video streams.
deepstream-3d-depth-camera deepstream-3d-depth-camera-app Depth camera (RGB-D) integration pipeline.
deepstream-3d-lidar-sensor-fusion deepstream-3d-lidar-sensor-fusion-app LiDAR point-cloud sensor fusion pipeline.
deepstream-lidar-inference-app deepstream-lidar-inference-app LiDAR-based object detection inference.
deepstream-can-orientation-app deepstream-can-orientation-app CAN bus + VPI-based vehicle orientation detection.

Networking and Server

App Binary Description
deepstream-server deepstream-server REST server with live pipeline control (source add/remove, config reload).
deepstream-nmos deepstream-nmos-app NMOS IS-04/IS-05 media networking and device discovery.
deepstream-ipc-test deepstream-ipc-test Share video buffers between pipelines via IPC. Jetson (aarch64) only — relies on NVDEC engine sharing not available on x86 or SBSA.
deepstream-ucx-test deepstream-ucx-test-app Share buffers between DeepStream processes/nodes over UCX (Unified Communication X) transport. Not supported on Jetson — the nvdsgst_ucx plugin is x86 only.

OpenCV Integration

App Binary Description
deepstream-opencv-test deepstream-opencv-test Access frame data in CPU memory for OpenCV processing.

Triton Inference Server

These samples do not produce a standalone binary. Each builds an nvinferserver custom-process library (.so) and is driven by the deb-installed deepstream-app with the included config file. See each sample's README for the prerequisite Triton container setup and model download steps.

Sample Library Description
TritonBackendEnsemble libnvdstriton_custom_impl_ensemble.so Triton ensemble model used as SGIE plus a custom Triton C++ backend that reads DeepStream stream IDs.
TritonOnnxYolo libnvdstriton_custom_impl_yolo.so DS-Triton custom-lib for ONNX YoloV3 with dynamic-sized output tensors and multi-input postprocessing.

Run flow:

cd src/apps/sample_apps/<TritonSample>
deepstream-app -c <included-config>.txt