Awesome

Intel® Deep Learning Streamer Pipeline Server

DISCONTINUATION OF PROJECT.

This project will no longer be maintained by Intel.

This project has been identified as having known security escapes.

Intel has ceased development and contributions including, but not limited to, maintenance, bug fixes, new releases, or updates, to this project.

Intel no longer accepts patches to this project.

| Getting Started | Request Customizations | Changing Pipeline Model | Further Reading | Known Issues |

Intel® Deep Learning Streamer (Intel® DL Streamer) Pipeline Server is a python package and microservice for deploying optimized media analytics pipelines. It supports pipelines defined in GStreamer* or FFmpeg* and provides APIs to discover, start, stop, customize and monitor pipeline execution. Intel® DL Streamer Pipeline Server is based on Intel® Deep Learning Streamer Pipeline Framework and FFmpeg Video Analytics.

Features Include

IMPORTANT: Intel® DL Streamer Pipeline Server is provided as a sample. It is not intended to be deployed into production environments without modification. Developers deploying Intel® DL Streamer Pipeline Server should review it against their production requirements.

The sample microservice includes five categories of media analytics pipelines. Click on the links below to find out more about each of them.

Getting Started

Prerequisites

Supported Hardware

Refer to Intel® DL Streamer Hardware Requirements for supported development and target runtime platforms and the Intel® DL Streamer Install Guide for details on providing access to accelerator devices.

Building the Microservice

Build the sample microservice with the following command:

./docker/build.sh

The script will automatically include the sample models, pipelines and required dependencies.

To verify the build succeeded execute the following command:

docker images dlstreamer-pipeline-server-gstreamer:latest

Expected output:

REPOSITORY                          TAG                 IMAGE ID            CREATED             SIZE
dlstreamer-pipeline-server-gstreamer   latest              f51f2695639f        2 minutes ago          2.81GB

Running the Microservice

Start the sample microservice with the following command:

./docker/run.sh -v /tmp:/tmp

This script issues a standard docker run command to launch the container, run a Tornado based web service on port 8080, and mount the /tmp folder. The /tmp folder is mounted to share sample results with the host and is optional in actual deployments.

Expected output:

{"levelname": "INFO", "asctime": "2021-04-05 23:43:42,406", "message": "=================", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:42,406", "message": "Loading Pipelines", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:42,406", "message": "=================", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:43,073", "message": "Loading Pipelines from Config Path /home/pipeline-server/pipelines", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:43,088", "message": "Loading Pipeline: audio_detection version: environment type: GStreamer from /home/pipeline-server/pipelines/audio_detection/environment/pipeline.json", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:43,090", "message": "Loading Pipeline: object_classification version: vehicle_attributes type: GStreamer from /home/pipeline-server/pipelines/object_classification/vehicle_attributes/pipeline.json", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:43,102", "message": "Loading Pipeline: object_detection version: edgex type: GStreamer from /home/pipeline-server/pipelines/object_detection/edgex/pipeline.json", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:43,104", "message": "Loading Pipeline: object_detection version: person_vehicle_bike type: GStreamer from /home/pipeline-server/pipelines/object_detection/person_vehicle_bike/pipeline.json", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:43,109", "message": "Loading Pipeline: object_tracking version: person_vehicle_bike type: GStreamer from /home/pipeline-server/pipelines/object_tracking/person_vehicle_bike/pipeline.json", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:43,109", "message": "===========================", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:43,109", "message": "Completed Loading Pipelines", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:43,109", "message": "===========================", "module": "pipeline_manager"}
{"levelname": "INFO", "asctime": "2021-04-05 23:43:43,217", "message": "Starting Tornado Server on port: 8080", "module": "__main__"}

Running a Pipeline

The Pipeline Server includes a sample client pipeline_client that can connect to the service and make requests. We will use pipeline_client to explain how to use the key microservice features.

Note: Any RESTful tool or library can be used to send requests to the Pipeline Server service. We are using pipeline_client as it simplifies interaction with the service.

Note: The microservice has to be up and running before the sample client is invoked.

Before running a pipeline, we need to know what pipelines are available. We do this using pipeline_client's list-pipeline command. In new shell run the following command:

./client/pipeline_client.sh list-pipelines

- object_classification/vehicle_attributes
- audio_detection/environment
- object_tracking/object_line_crossing
- object_tracking/person_vehicle_bike
- object_detection/object_zone_count
- object_detection/app_src_dst
- object_detection/person_vehicle_bike

Note: The pipelines you will see may differ slightly

Pipelines are displayed as a name/version tuple. The name reflects the action and version supplies more details of that action. Let's go with object_detection/person_vehicle_bike. Now we need to choose a media source. We recommend the IoT Devkit sample videos to get started. As the pipeline version indicates support for detecting people, person-bicycle-car-detection.mp4 would be a good choice.

Note: Make sure to include raw=true parameter in the Github URL as shown in our examples. Failure to do so will result in a pipeline execution error.

pipeline_client offers a run command that takes two additional arguments the pipeline and the uri for the media source. The run command displays inference results until either the media is exhausted or CTRL+C is pressed.

Inference result bounding boxes are displayed in the format label (confidence) [top left width height] {meta-data} provided applicable data is present. At the end of the pipeline run, the average fps is shown.

./client/pipeline_client.sh run object_detection/person_vehicle_bike https://github.com/intel-iot-devkit/sample-videos/blob/master/person-bicycle-car-detection.mp4?raw=true

Timestamp 48583333333
- vehicle (0.95) [0.00, 0.12, 0.15, 0.36]
Timestamp 48666666666
- vehicle (0.79) [0.00, 0.12, 0.14, 0.36]
Timestamp 48833333333
- vehicle (0.68) [0.00, 0.13, 0.11, 0.36]
Timestamp 48916666666
- vehicle (0.78) [0.00, 0.13, 0.10, 0.36]
Timestamp 49000000000
- vehicle (0.63) [0.00, 0.13, 0.09, 0.36]
Timestamp 49083333333
- vehicle (0.63) [0.00, 0.14, 0.07, 0.35]
Timestamp 49166666666
- vehicle (0.69) [0.00, 0.14, 0.07, 0.35]
Timestamp 49250000000
- vehicle (0.64) [0.00, 0.14, 0.05, 0.34]
avg_fps: 39.66

NOTE: Inference results are not obtained via a REST API but read from an output file. The file path is specified in the destination section of the REST request and will be discussed in a later section.

Pipeline States

Queued, Running and Completed

The pipeline_client run command starts the pipeline. The underlying REST request returns a pipeline instance which is used to query the state of the pipeline. All being well it will go into QUEUED then RUNNING state. We can interrogate the pipeline status by using the pipeline_client start command that kicks off the pipeline like run and then exits displaying the pipeline instance (a UUID) which is used by the status command to view pipeline state.

NOTE: The pipeline instance value depends on the number of pipelines started while the server is running so may differ from the value shown in the following examples.

./client/pipeline_client.sh start object_detection/person_vehicle_bike https://github.com/intel-iot-devkit/sample-videos/blob/master/person-bicycle-car-detection.mp4?raw=true

<snip>
Starting pipeline object_detection/person_vehicle_bike, instance = d83502e3ef314e8fbec8dc926eadd0c2

You will need both the pipeline tuple and instance id for the status command. This command will display pipeline state:

./client/pipeline_client.sh status object_detection/person_vehicle_bike d83502e3ef314e8fbec8dc926eadd0c2

<snip>
RUNNING (49fps)

Then wait for a minute or so and try again. Pipeline will be completed.

./client/pipeline_client.sh status object_detection/person_vehicle_bike d83502e3ef314e8fbec8dc926eadd0c2

<snip>
COMPLETED (50fps)

Aborted

If a pipeline is stopped, rather than allowed to complete, it goes into the ABORTED state. Start the pipeline again, this time we'll stop it.

./client/pipeline_client.sh start object_detection/person_vehicle_bike https://github.com/intel-iot-devkit/sample-videos/blob/master/person-bicycle-car-detection.mp4?raw=true

<snip>
Starting pipeline object_detection/person_vehicle_bike, instance = 8ad2c85af4bd473e8a693aff562be316

./client/pipeline_client.sh status object_detection/person_vehicle_bike 8ad2c85af4bd473e8a693aff562be316

<snip>
RUNNING (50fps)

./client/pipeline_client.sh stop object_detection/person_vehicle_bike 8ad2c85af4bd473e8a693aff562be316

<snip>
Stopping Pipeline...
Pipeline stopped
avg_fps: 24.33

./client/pipeline_client.sh status object_detection/person_vehicle_bike 8ad2c85af4bd473e8a693aff562be316

<snip>
ABORTED (47fps)

Error

The error state covers a number of outcomes such as the request could not be satisfied, a pipeline dependency was missing or an initialization problem. We can create an error condition by supplying a valid but unreachable uri.

./client/pipeline_client.sh start object_detection/person_vehicle_bike http://bad-uri

<snip>
Starting pipeline object_detection/person_vehicle_bike, instance = 2bb2d219310a4ee881faf258fbcc4355

Note that the Pipeline Server does not report an error at this stage as it goes into QUEUED state before it realizes that the source is not providing media. Checking on state a few seconds later will show the error.

./client/pipeline_client.sh status object_detection/person_vehicle_bike 2bb2d219310a4ee881faf258fbcc4355

<snip>
ERROR (0fps)

Request Customizations

Change Pipeline and Source Media

With pipeline_client it is easy to customize service requests. Here will use a vehicle classification pipeline object_classification/vehicle_attributes with the IoT Devkit video car-detection.mp4. Note how pipeline_client now displays classification metadata including type and color of vehicle.

./client/pipeline_client.sh run object_classification/vehicle_attributes https://github.com/intel-iot-devkit/sample-videos/blob/master/car-detection.mp4?raw=true

Starting pipeline object_classification/vehicle_attributes, instance = <uuid>
Pipeline running
<snip>
Timestamp 18080000000
- vehicle (1.00) [0.41, 0.00, 0.57, 0.33] {'color': 'red', 'type': 'car'}
Timestamp 18160000000
- vehicle (1.00) [0.41, 0.00, 0.57, 0.31] {'color': 'red', 'type': 'car'}
- vehicle (0.50) [0.09, 0.92, 0.27, 1.00] {'color': 'white', 'type': 'van'}
Timestamp 18240000000
- vehicle (1.00) [0.40, 0.00, 0.57, 0.27] {'color': 'red', 'type': 'car'}
Timestamp 18320000000
- vehicle (1.00) [0.40, 0.00, 0.56, 0.24] {'color': 'red', 'type': 'car'}
Timestamp 18400000000
- vehicle (1.00) [0.40, 0.00, 0.56, 0.22] {'color': 'red', 'type': 'car'}
- vehicle (0.53) [0.52, 0.19, 0.98, 0.99] {'color': 'black', 'type': 'truck'}
Timestamp 18480000000
- vehicle (0.99) [0.40, 0.00, 0.56, 0.20] {'color': 'red', 'type': 'car'}
- vehicle (0.81) [0.69, 0.00, 1.00, 0.36] {'color': 'black', 'type': 'bus'}
Timestamp 18560000000
- vehicle (1.00) [0.40, 0.00, 0.55, 0.18] {'color': 'red', 'type': 'car'}
- vehicle (0.71) [0.70, 0.00, 1.00, 0.36] {'color': 'black', 'type': 'bus'}
Timestamp 18640000000
- vehicle (0.98) [0.40, 0.00, 0.55, 0.15] {'color': 'red', 'type': 'car'}

If you look at the video, you can see that there are some errors in classification - there are no trucks or buses in the video. However you can see that associated confidence is much lower than the correct classification of the white and red cars.

Change Inference Accelerator Device

Inference accelerator devices can be easily selected using the device parameter. Here we run the car classification pipeline again, but this time use the integrated GPU for detection inference by setting the detection-device parameter.

./client/pipeline_client.sh run object_classification/vehicle_attributes https://github.com/intel-iot-devkit/sample-videos/blob/master/car-detection.mp4?raw=true --parameter detection-device GPU --parameter detection-model-instance-id person_vehicle_bike_detection_gpu

Starting pipeline object_classification/vehicle_attributes, instance = <uuid>

Note: The GPU inference plug-in dynamically builds OpenCL kernels when it is first loaded resulting in a ~30s delay before inference results are produced.

Note: The detection-model-instance-id parameter caches the GPU model with a unique id. For more information read about model instance ids.

pipeline_client's fps measurement is useful when assessing pipeline performance with different accelerators.

Visualize Inference

Pipeline server allows you to optionally visualize inference results using either Real Time Streaming Protocol (RTSP) or Web Real Time Communication (WebRTC) by configuring the frame destination section of the request.

RTSP is simpler to set up but you must have an RTSP player (e.g. VLC) to render output. WebRTC setup is more complex (e.g., requires additional server-side microservices) but has the upside of using a web browser for client visualization.

Before requesting visualization, the corresponding feature must be enabled in the server, see Visualizing Inference Output.

RTSP

RTSP allows you to connect to a server and display a video stream. The Pipeline Server includes an RTSP server that creates a stream that shows the incoming video with superimposed bounding boxes and meta-data. You will need a client that connects to the server and displays the video. We recommend vlc.

First start the Pipeline Server with RTSP enabled. By default, the RTSP stream will use port 8554.

docker/run.sh --enable-rtsp -v /tmp:/tmp

Then start pipeline and visualize as per RTSP section in Customizing Pipeline Requests.

Note: The pipeline must be running before you hit play otherwise VLC will not be able to connect to the RTSP server. For shorter video files you should have VLC ready to go before starting pipeline otherwise by the time you hit play the pipeline will have completed and the RTSP server will have shut down.

WebRTC

WebRTC is more complex. Follow setup instructions in the sample. More details on fine tuning request can be found in the WebRTC section in Customizing Pipeline Requests.

View REST Request

As the previous example has shown, the pipeline_client application works by converting command line arguments into Pipeline Server REST requests. The --show-request option displays the REST verb, uri and body in the request. Let's repeat the previous GPU inference example, adding RTSP output and show the underlying request.

./client/pipeline_client.sh run object_classification/vehicle_attributes https://github.com/intel-iot-devkit/sample-videos/blob/master/car-detection.mp4?raw=true --parameter detection-device GPU --rtsp-path pipeline-server --show-request

<snip>
POST http://localhost:8080/pipelines/object_classification/vehicle_attributes
Body:{'source': {'uri': 'https://github.com/intel-iot-devkit/sample-videos/blob/master/car-detection.mp4?raw=true', 'type': 'uri'}, 'destination': {'metadata': {'type': 'file', 'path': '/tmp/results.jsonl', 'format': 'json-lines'}, 'frame': {'type': 'rtsp', 'path': 'pipeline-server'}}, 'parameters': {'detection-device': 'GPU'}}

The REST request is broken into three parts

1. VERB: POST
2. URI: http://localhost:8080/pipelines/object_classification/vehicle_attributes
3. BODY: {"source": {"uri": ...

The uri is easy to decode. It's the service address, pipelines command and the pipeline tuplet. The body contains three components:

1. The media source
2. The frame and metadata destinations
3. Pipeline parameters.

They are easier to understand when the json is pretty-printed

{
  "source": {
    "uri": "https://github.com/intel-iot-devkit/sample-videos/blob/master/car-detection.mp4?raw=true",
    "type": "uri"
  },
  "destination": {
    "metadata": {
      "type": "file",
      "path": "/tmp/results.jsonl",
      "format": "json-lines"
    },
    "frame": {
      "type": "rtsp",
      "path": "pipeline-server"
    }
  },
  "parameters": {
    "detection-device": "GPU"
  }
}

1. Media source: type is uri and the uri is the car-detection.mp4 video
2. Destinations:
   • metadata: this is the inference results, they are sent to file /tmp/results.jsonl in json-lines format. pipeline_client parses this file to display the inference results and metadata.
   • frames: this the watermarked frames. Here they are sent to the RTSP server and available over given endpoint pipeline-server.
3. Parameters set pipeline properties. See the Defining Pipelines document for more details on parameters.

The --show-request output can be easily converted int a curl command.

curl <URI> -X <VERB> -H "Content-Type: application/json' -d <BODY>

So the above request would be as below. Note the pipeline instance 1 returned by the request.

curl localhost:8080/pipelines/object_classification/vehicle_attributes -X POST -H \
'Content-Type: application/json' -d \
'{
  "source": {
    "uri": "https://github.com/intel-iot-devkit/sample-videos/blob/master/car-detection.mp4?raw=true",
    "type": "uri"
  },
  "destination": {
    "metadata": {
      "type": "file",
      "path": "/tmp/results.jsonl",
      "format": "json-lines"
    },
    "frame": {
      "type": "rtsp",
      "path": "pipeline-server"
    }
  },
  "parameters": {
    "detection-device": "GPU"
  }
}'

59896b90853511ec838b0242ac110002

Changing Pipeline Model

The Pipeline Server makes pipeline customization and model selection a simple task. The Changing Object Detection Models Tutorial provides step by step instructions for changing the object detection reference pipeline to use a model better suited to a different video source.

Further Reading

Related Links

Known Issues

Known issues are tracked on GitHub*

* Other names and brands may be claimed as the property of others.