Ultralytics YOLO27: Join the waitlist.
Get Started

Ultralytics YOLO26 on Intel Core Ultra Series 3 using DL Streamer Pipeline Framework and OpenVINO™

This guide walks through deploying Ultralytics YOLO26 on Intel Core Ultra Series 3 platforms using DL Streamer Pipeline Framework and OpenVINO™ toolkit. Here we use OpenVINO™ to maximize inference performance on Intel CPUs, integrated and discrete GPUs, and NPUs.

Intel DL Streamer Pipeline Framework with Ultralytics YOLO26

Contents: What is Intel DL Streamer? • Prerequisites • YOLO26 Model Preparation • Running Inference with YOLO26 • Multi-Stream Setup • FAQ

What is Intel DL Streamer?#

Deep Learning Streamer (DL Streamer) Pipeline Framework is an open-source streaming media analytics framework based on the GStreamer multimedia framework, designed for creating complex media analytics pipelines in the cloud or at the edge.

DL Streamer enables analysis of audio and video streams to detect, classify, track, identify, and count objects, events, and people. It is optimized for Intel hardware and provides interoperability between GStreamer plugins built on various backend libraries:

  • Inference: OpenVINO™ inference engine, optimized for Intel CPU, GPU, and NPU
  • Video Encoding/Decoding: GPU-acceleration via VA-API
  • Image Processing: GPU-acceleration via VA-API
  • Metadata: GStreamer Analytics for structured inference results
  • Ecosystem: Hundreds of GStreamer plugins for media I/O, muxing/demuxing, codec support, and more

DL Streamer supports many AI models including the full Ultralytics YOLO family (YOLOv5 through YOLO26) all in OpenVINO™ format.

DL Streamer is regularly validated on the systems listed in the DL Streamer system requirements.

Prerequisites#

Before you begin, ensure the following are installed and configured on your Intel system:

YOLO26 Model Preparation#

DL Streamer uses models in OpenVINO™ IR format. Ultralytics YOLO26 models are exported from PyTorch to OpenVINO™ IR using the Ultralytics exporter. DL Streamer uses the official Ultralytics OpenVINO™ integration, which provides optimized export and inference capabilities for Intel hardware.

  1. Create the ~/intel/dlstreamer_demo folder and install OpenVINO™ and Ultralytics in a virtual environment
mkdir -p ~/intel/dlstreamer_demo && cd ~/intel/dlstreamer_demo
python3 -m venv .dls-venv && source .dls-venv/bin/activate
pip install openvino==2026.2.0 ultralytics==8.4.92
  1. Download the PyTorch YOLO26s model from Ultralytics, convert it to OpenVINO™ IR format, and generate an INT8 precision variant.
yolo export model=yolo26s.pt format=openvino dynamic=True quantize=8 data=coco128.yaml

The model is exported to the ~/intel/dlstreamer_demo/yolo26s_int8_openvino_model folder.

Model Precision#

DL Streamer supports FP32, FP16, and INT8 precision models. Each requires a separate export step:

# FP32 (default precision, highest accuracy)
yolo export model=yolo26s.pt format=openvino dynamic=True

# FP16 (recommended for GPU inference, good accuracy/performance balance)
yolo export model=yolo26s.pt format=openvino dynamic=True quantize=16

# INT8 (maximum performance, requires calibration dataset)
yolo export model=yolo26s.pt format=openvino dynamic=True quantize=8 data=coco128.yaml

The examples below use the INT8 model exported in the preparation step. To use FP32 or FP16 models, replace the model path accordingly (e.g. yolo26s_openvino_model/yolo26s.xml).

Running Inference with YOLO26#

Before running the DL Streamer YOLO26 detection pipeline, complete the YOLO26 Model Preparation step and download the following sample video from Pexels.

curl -L https://videos.pexels.com/video-files/1192116/1192116-sd_640_360_30fps.mp4 --output ~/intel/dlstreamer_demo/video1.mp4

Then run the DL Streamer Docker image in interactive mode.

Note

This command is intended for systems equipped with Intel Core Ultra Series 3 processors with integrated GPU (iGPU) and NPU onboard.

docker run -it --rm \
  -v ~/intel/dlstreamer_demo:/home/dlstreamer/demo \
  -v "$HOME/.Xauthority:/root/.Xauthority:rw" \
  -e DISPLAY=$DISPLAY \
  -e XDG_RUNTIME_DIR=/tmp \
  -v /tmp/.X11-unix:/tmp/.X11-unix \
  --device /dev/dri \
  --group-add $(stat -c "%g" /dev/dri/render*) \
  --device /dev/accel \
  --group-add $(stat -c "%g" /dev/accel/accel*) \
  -e ZE_ENABLE_ALT_DRIVERS=libze_intel_npu.so \
  intel/dlstreamer:latest

INT8 Precision (Maximum Performance)#

INT8 quantization delivers the highest throughput by reducing model weights to 8-bit integers. The Ultralytics exporter handles calibration automatically.

Run YOLO26s with INT8 on GPU#

gst-launch-1.0 filesrc location=/home/dlstreamer/demo/video1.mp4 ! decodebin3 ! gvadetect model=/home/dlstreamer/demo/yolo26s_int8_openvino_model/yolo26s.xml device=GPU ! queue ! vapostproc ! gvawatermark ! gvafpscounter ! autovideosink sync=false
YOLO26s INT8 detections in a DL Streamer pipeline on an Intel GPU

Run YOLO26s with INT8 on GPU and save the output to a video file#

gst-launch-1.0 filesrc location=/home/dlstreamer/demo/video1.mp4 ! decodebin3 ! gvadetect model=/home/dlstreamer/demo/yolo26s_int8_openvino_model/yolo26s.xml device=GPU ! queue ! vapostproc ! gvawatermark ! gvafpscounter ! vah264enc ! h264parse ! mp4mux ! filesink location=/home/dlstreamer/demo/yolo_video1_yolo26s_INT8_GPU.mp4

Run YOLO26s with INT8 on NPU#

gst-launch-1.0 filesrc location=/home/dlstreamer/demo/video1.mp4 ! decodebin3 ! gvadetect model=/home/dlstreamer/demo/yolo26s_int8_openvino_model/yolo26s.xml device=NPU ! queue ! vapostproc ! gvawatermark ! gvafpscounter ! autovideosink sync=false
YOLO26s INT8 detections in a DL Streamer pipeline on an Intel NPU

Multi-Stream Setup#

DL Streamer supports multi-stream processing, where multiple video sources are decoded and inferred simultaneously. You can launch multiple pipelines in parallel using GStreamer's vacompositor element to combine multiple streams.

Running Multiple Pipelines (4 streams) in Parallel (GPU)#

gst-launch-1.0 vacompositor name=comp sink_0::xpos=0 sink_0::ypos=0 sink_1::xpos=660 sink_1::ypos=0 sink_2::xpos=0 sink_2::ypos=380 sink_3::xpos=660 sink_3::ypos=380 ! autovideosink sync=false \
  filesrc location=/home/dlstreamer/demo/video1.mp4 ! decodebin3 ! gvadetect model=/home/dlstreamer/demo/yolo26s_int8_openvino_model/yolo26s.xml device=GPU model-instance-id=inf0 scheduling-policy="latency" ! queue ! gvawatermark ! gvafpscounter ! comp.sink_0 \
  filesrc location=/home/dlstreamer/demo/video1.mp4 ! decodebin3 ! gvadetect model=/home/dlstreamer/demo/yolo26s_int8_openvino_model/yolo26s.xml device=GPU model-instance-id=inf0 scheduling-policy="latency" ! queue ! gvawatermark ! gvafpscounter ! comp.sink_1 \
  filesrc location=/home/dlstreamer/demo/video1.mp4 ! decodebin3 ! gvadetect model=/home/dlstreamer/demo/yolo26s_int8_openvino_model/yolo26s.xml device=GPU model-instance-id=inf0 scheduling-policy="latency" ! queue ! gvawatermark ! gvafpscounter ! comp.sink_2 \
  filesrc location=/home/dlstreamer/demo/video1.mp4 ! decodebin3 ! gvadetect model=/home/dlstreamer/demo/yolo26s_int8_openvino_model/yolo26s.xml device=GPU model-instance-id=inf0 scheduling-policy="latency" ! queue ! gvawatermark ! gvafpscounter ! comp.sink_3
Four-stream DL Streamer YOLO26 detection grid on an Intel GPU

Additional Resources#

FAQ#

  • Install DL Streamer following the Installation Guide, set up the environment with source /opt/intel/dlstreamer/scripts/setup_dls_env.sh, install Ultralytics and OpenVINO™, download models using download_ultralytics_models.sh. Then run inference with the DL Streamer sample application yolo_detect.sh that provides a ready-to-use script for running YOLO26 inference pipelines.

  • OpenVINO™ optimizes the YOLO26 model specifically for Intel hardware through techniques such as graph optimization, layer fusion, and hardware-specific kernel tuning. Combined with DL Streamer's VA-API accelerated decode and zero-copy va-surface-sharing pre-processing, the full video analytics pipeline achieves significantly higher throughput than unoptimized frameworks.

  • Yes. DL Streamer supports inference on Intel CPUs (Core, Core Ultra, Xeon), integrated GPUs (Iris Xe, Arc), discrete GPUs (Arc A-Series, B-Series), and NPUs (AI Boost) across multiple Intel platform generations. Change the device property of gvadetect to CPU, GPU, or NPU.

    • FP16 is recommended as the default for GPU inference — it provides near-FP32 accuracy at higher throughput.
    • INT8 delivers the highest performance with a small accuracy trade-off and is ideal when maximum throughput is the priority. INT8 models are automatically calibrated during Ultralytics export.

  • DL Streamer supports these YOLO26 task variants:

    • Detection: yolo26n, yolo26s, yolo26m, yolo26l, yolo26x
    • Instance Segmentation: yolo26s-seg (and all size variants)
    • Classification: yolo26s-cls (composite pipeline with detection)
    • Pose Estimation: yolo26s-pose (and all size variants)
    • Oriented Bounding Box (OBB): yolo26s-obb (and all size variants)

  • Use the json output option to write detection results as JSON-lines to a file:

    ./yolo_detect.sh yolo26s GPU input_video.mp4 json va-surface-sharing INT8

    Alternatively, use the gvametapublish element in custom pipelines to publish metadata to files, MQTT, or Kafka.

Contributors
glenn-jocherglenn-jocher5raimbekovmraimbekovm2onuralpszronuralpszr2tbujewsktbujewsk1

Comments

Ask a question, share your results, or help someone else.