Samples

DL Streamer ships with 40+ ready-to-run samples that turn common media-analytics tasks into working pipelines you can launch in minutes. They are the fastest way to see an element or a full use case in action, and a great starting point to copy from when building your own application.

Each sample lives in its own folder with a README.md and a run script. Browse them online in the samples directory on GitHub or, after installation, under /opt/intel/dlstreamer/samples.

Looking for a specific element? See the Elements page — most elements are demonstrated by one or more of the samples listed below.

How to run

• Install DL Streamer first — see the Get Started guide.

• Download the models the samples use with the conversion scripts under scripts/download_models. They export models from Hugging Face, Ultralytics, TIMM and other sources to OpenVINO IR:

  • download_hf_models.py — Hugging Face models (VLMs, CLIP, Whisper, …).

  • download_ultralytics_models.py — Ultralytics YOLO models.

  • download_timm_models.py — TIMM image-classification models.

  • download_other_models.sh — other helper models (e.g. centerface, hsemotion, deeplabv3, mars-small128).

  See the download_models README for prerequisites, per-script requirements files and usage examples. Each sample’s own README.md lists the exact model(s) it needs (see the Models column below).

• Samples with C/C++ code provide a build_and_run.sh; other samples provide a .sh script that builds and runs a gst-launch-1.0 or Python command line.

Platform support: Samples target Linux (Ubuntu 22.04/24.04) by default. A subset also ships a Windows variant (PowerShell/.bat scripts, D3D11/ksvideosrc backends). See the Windows samples folder.

The Language column below indicates how each sample is implemented: CLI (gst-launch command line), Python, or C++. The Models column lists the default model(s) each sample runs — many samples let you swap in your own.

---

Find a sample by what you want to do

Object detection, classification & segmentation

Preview	Sample	What it demonstrates	Key elements	Models	Language
	Detection with YOLO	Object detection and classification with publicly available YOLO models	gvadetect, gvaclassify	yolox_s (default; many YOLO variants)	CLI
—	Face Detection and Classification	Detect faces and estimate age, gender, emotions and facial landmarks	gvadetect, gvaclassify	centerface, dima806_facial_age_image_detection, dima806_fairface_gender_image_detection, dima806_face_emotions_image_detection	CLI
	Instance Segmentation	Instance segmentation via the object_detect and object_classify bin elements	object_detect, object_classify	yolo26s-seg (default; also yolo11s-seg)	CLI
—	Human Pose Estimation	Full-frame human pose estimation	gvaclassify	yolo26s-pose	CLI
—	Depth Estimation	YOLO11n detection followed by Depth Anything V2 depth estimation on detected regions	gvadetect, gvainference	yolo11n, Depth-Anything-V2-Small-hf	CLI
—	License Plate Recognition	YOLO detector combined with an optical character recognition model	gvadetect, gvainference	yolov8 license-plate detector, PP-OCRv4	CLI
	Prompt-based Object Detection	Search a video for user-defined objects using an open-vocabulary model (YOLOE)	gvadetect	yoloe-26s-seg (open-vocabulary)	Python
	Deployment of Geti™ models	Deploy Geti™-trained models for detection, anomaly detection and classification	gvadetect, gvaclassify	Geti™-trained (Padim / STFPM / UFlow)	CLI
—	Motion Detect	Run detection only over motion ROIs (GPU and CPU paths)	gvamotiondetect, gvadetect	yolov8n	CLI

Object tracking & analytics

Preview	Sample	What it demonstrates	Key elements	Models	Language
—	Vehicle and Pedestrian Tracking	Object tracking across frames	gvatrack, gvadetect, gvaclassify	yolo26s, dima806_vehicle_10_types_image_detection	CLI
	Vehicle Counter with gvaanalytics Tripwires	Count vehicles crossing a virtual line in both directions using tripwires	gvaanalytics, gvatrack	yolo11n	Python
	Smart NVR for Lane Hogging Detection	Build an NVR with custom analytics and video storage to detect lane-hogging events	gvaanalytics_py, gvarecorder_py	rtdetr_v2_r50vd (RT-DETRv2)	Python
	Loitering detection	Measures object dwell time and render visual alert when dwell time threshold is exceeded.	gvaanalytics, watermark	yolo11s	Python

Vision-Language Models (VLM) & GenAI

Preview	Sample	What it demonstrates	Key elements	Models	Language
—	Using VLM Models with gvagenai	Video summarization with MiniCPM-V	gvagenai	MiniCPM-V, Phi-4-multimodal-instruct or Gemma-3	CLI
	VLM Alerts	Edge alerting pipeline that generates structured JSON alerts per frame with annotated video	gvagenai	Configurable VLM (e.g. Qwen2.5-VL, InternVL)	Python
	VLM-assisted Self Checkout	Combine CV object detection with a VLM for item classification, running both locally on edge	gvadetect, gvagenai	yolo26s, MiniCPM-V-4_5	Python
	ONVIF Camera Analytics Validation	Use a VLM as an additional validation layer for ONVIF-enabled analytics cameras	gvagenai	Configurable VLM	Python
—	Image Embeddings Generation with ViT	Generate image embeddings using the Vision Transformer component of a CLIP model	gvainference	clip-vit-large-patch14 (CLIP ViT)	CLI

Audio analytics

Sample	What it demonstrates	Key elements	Models	Language
Audio Event Detection	Audio event detection, converting results to JSON	gvaaudiodetect, gvametaconvert, gvametapublish	aclnet	CLI
Audio Transcription	Speech transcription using an OpenVINO GenAI Whisper model	gvaaudiotranscribe	whisper	CLI

3D: LiDAR & radar

Sample	What it demonstrates	Key elements	Models	Language
PointPillars Inference with g3dinference	Complete LiDAR-only 3D detection pipeline	g3dlidarparse, g3dinference	PointPillars	CLI
LiDAR Parse	LiDAR parsing pipeline	g3dlidarparse	— (parsing only)	CLI
Live LiDAR Capture	Real-time LiDAR capture from a physical device (RoboSense via rs_driver)	g3dlidarsrc, g3dinference	PointPillars	CLI
Camera + 3D Object Fusion	Fuse 2D camera detections with 3D LiDAR detections	g3dobjectfuser, gvastreammux	yolo11n, PointPillars	CLI
Radar Signal Process	mmWave radar signal processing with point-cloud detection, clustering and tracking	g3dradarprocess	— (signal processing)	CLI

Cameras & input sources

Preview	Sample	What it demonstrates	Key elements	Models	Language
—	RealSense™ Camera	Capture a video stream from a 3D Intel RealSense™ Depth Camera	gvarealsense	— (capture only)	CLI
	ONVIF Camera Discovery	Automatically discover ONVIF cameras on the network and launch pipelines for each	gvadetect	Configurable detector	Python
—	Multi-camera deployments	Handle video streams from multiple cameras in a single application	gvadetect, gvafpscounter	yolo11s (many YOLO variants)	CLI
—	Multi-Stream Mux/Demux	Share a single inference pipeline across streams with per-source routing	gvastreammux, gvastreamdemux	Configurable detector	CLI

Metadata: publishing, access & visualization

Preview	Sample	What it demonstrates	Key elements	Models	Language
—	Metadata Publishing	Convert inference metadata to JSON and publish to file or Kafka/MQTT	gvametaconvert, gvametapublish	centerface, dima806_fairface_gender_image_detection, dima806_facial_age_image_detection	CLI
—	gvaattachroi	Define the regions on which inference should be performed	gvaattachroi, gvadetect	yolov8s	CLI
—	FPS Throttle	Throttle framerate independently of sink sync, without frame duplication or dropping	gvafpsthrottle	—	CLI
	Watermark Metadata	Attach custom drawing primitives (hexagons, lines, circles, text) and render them	gvawatermark	— (drawing only)	Python
—	Draw Face Attributes (C++)	Set a C callback to access frame metadata and visualize inference results	gvadetect, gvaclassify	centerface, dima806_facial_age_image_detection, dima806_fairface_gender_image_detection, dima806_face_emotions_image_detection	C++
—	Draw Face Attributes (Python)	Set a Python callback to access frame metadata and visualize inference results	gvadetect, gvaclassify	centerface, dima806_facial_age_image_detection, dima806_fairface_gender_image_detection, dima806_face_emotions_image_detection	Python
	Open Close Valve	Open/close a GStreamer valve branch from a callback based on detection results	gvadetect, valve	yolo11s, dima806_vehicle_10_types_image_detection	Python
	Hello DL Streamer	Build a detection pipeline, analyze metadata to count objects, and visualize results	gvadetect, gvawatermark	yolo11n	Python

Customization & extensibility

Sample	What it demonstrates	Key elements	Models	Language
Custom Post-Processing Library — Classification	Write a custom post-processing library that converts emotion-classification outputs to GstAnalytics metadata	gvaclassify	centerface, hsemotion	CLI, C++
Custom Post-Processing Library — Detection	Write a custom post-processing library that converts YOLOv11 tensor outputs to detection metadata	gvadetect	yolo11s	CLI, C++
gvapython — Face Detection and Classification	Customize a pipeline with a Python script for inference post-processing	gvapython, gvadetect, gvaclassify	centerface, dima806_fairface_gender_image_detection, dima806_facial_age_image_detection	CLI, Python
gvapython — Save Frames with ROI	Use gvapython to save video frames containing detected objects to disk	gvapython, gvadetect	centerface	CLI, Python
python-elements — Face Detection and Classification	Build a custom Python GStreamer element using the GstAnalytics metadata API	gvaagelogger_py, gvadetect, gvaclassify	YOLOv8-Face-Detection, fairface_age_image_detection, fairface_gender_image_detection	CLI, Python
python-elements — Save Frames with ROI	Build a custom Python GStreamer element to save frames with detected objects	gvaframesaver_py, gvadetect	YOLOv8-Face-Detection	CLI, Python
Face Detection and Classification (Python)	Download models from Hugging Face, export to OpenVINO IR, and run inference	gvadetect, gvaclassify	YOLOv8-Face-Detection, fairface	Python

Performance & benchmarking

Preview	Sample	What it demonstrates	Key elements	Models	Language
—	Benchmark	Measure the performance of single- or multi-channel video analytics pipelines	gvadetect, gvafpscounter	centerface (configurable)	CLI, Python
	DL Streamer E2E Performance	Compare DL Streamer vs. OpenCV + OpenVINO throughput with a YOLO26s INT8 model	gvadetect	yolo26s (INT8)	Python

Interoperability

Sample	What it demonstrates	Key elements	Models	Language
DL Streamer and DeepStream Coexistence	Run pipelines on DL Streamer and/or NVIDIA DeepStream side by side	gvadetect	yolov8 license-plate detector, PP-OCRv4	Python

---

Auto-generated reference applications

These end-to-end reference apps combine multiple elements into complete solutions. Find them under samples/auto_generated_samples.

Sample	What it demonstrates	Key elements	Models	Language
DeepStream Test4 → DL Streamer Conversion	DL Streamer equivalent of NVIDIA’s deepstream-test4 with YOLO11n detection and metadata publishing	gvadetect, gvametaconvert, gvametapublish	yolo11n	CLI
DeepStream LPR App Conversion (C++)	C++ conversion of NVIDIA’s DeepStream LPR app — license plate detection, tracking and text recognition	gvadetect, gvatrack, gvaclassify	YOLOv11, PaddleOCR	C++
License Plate Recognition	Detect license plates with YOLOv11 and recognize text with PaddleOCR	gvadetect, gvainference	YOLOv11, PaddleOCR	CLI
Multi-Stream Compose	Multi-camera analytics with composite WebRTC output, on-demand recording and a 2x2 GPU-accelerated mosaic	gvadetect, gvastreammux, gvawatermark	yolo11s	CLI
People Detection and Tracking with Deep SORT	Detect and track people using YOLO26m and Deep SORT with a Mars-Small-128 re-ID model	gvadetect, gvatrack	yolo26m, mars-small128	CLI
Pose Estimation Compose	Run 4 YOLO pose models in parallel on the same video and composite results into a 2x2 mosaic	gvaclassify, gvawatermark	yolo26n-pose, yolo11n-pose, yolov8n-pose, yolov8l-pose	CLI
Safety Compliance Monitor	Detect and track workers and use Qwen2.5-VL to verify helmet and harness compliance	gvadetect, gvatrack, gvagenai	yolo26m, Qwen2.5-VL-3B	CLI
Smart NVR — Event-Based Recording	Detect people with YOLO11n and record video only when a person is present	gvadetect	yolo11n	CLI