Demo Configuration file

The demo config file uses YAML format to define input sources, models, outputs and finally the flows which defines how everything is connected. Config files for out-of-box demos are kept in edge_ai_apps/configs folder. The folder contains config files for all the use cases and also multi-input and multi-inference case. The folder also has a template YAML file app_config_template.yaml which has detailed explanation of all the parameters supported in the config file.

Config file is divided in 4 sections:

  1. Inputs

  2. Models

  3. Outputs

  4. Flows


The input section defines a list of supported inputs like camera, video files etc. Their properties like shown below.

    input0:                                         #Camera Input
        source: /dev/video2                         #Device file entry of the camera
        format: jpeg                                #Input data format suported by camera
        width: 1280                                 #Width and Height of the input
        height: 720
        framerate: 30                               #Framerate of the source

    input1:                                         #Video Input
        source: ../data/videos/video_0000_h264.mp4  #Video file
        format: h264                                #File encoding format
        width: 1280
        height: 720
        framerate: 25

    input2:                                         #Image Input
        source: ../data/images/%04d.jpg             #Sequence of Image files, printf style formatting is used
        width: 1280
        height: 720
        index: 0                                    #Starting Index (optional)
        framerate: 1

All supported inputs are listed in template config file. Below are the details of most commonly used inputs.

Camera sources (v4l2)

v4l2src GStreamer element is used to capture frames from camera sources which are exposed as v4l2 devices. In Linux, there are many devices which are implemented as v4l2 devices. Not all of them will be camera devices. You need to make sure the correct device is configured for running the demo successfully. is ran as part of systemd, which detects all cameras connected and prints the detail like below in the UART console:

root@tda4vm-sk:/opt/edge_ai_apps# ./
USB Camera detected
    device = /dev/video18
    format = jpeg
CSI Camera 0 detected
    device = /dev/video2
    name = imx219 8-0010
    format = [fmt:SRGGB8_1X8/1920x1080]
    subdev_id = 2
    isp_required = yes
IMX390 Camera 0 detected
    device = /dev/video18
    name = imx390 10-001a
    format = [fmt:SRGGB12_1X12/1936x1100 field: none]
    subdev_id = /dev/v4l-subdev7
    isp_required = yes
    ldc_required = yes

script can also be run manually later to get the camera details.

From the above log we can determine that 1 USB camera is connected (/dev/video18), and 1 CSI camera is connected (/dev/video2) which is IMX219 raw sensor and needs ISP. IMX390 camera needs both ISP and LDC.

Using this method, you can configure correct device for camera capture in the input section of config file.

    source: /dev/video18  #USB Camera
    format: jpeg          #if connected USB camera supports jpeg
    width: 1280
    height: 720
    framerate: 30

    source: /dev/video2  #CSI Camera
    format: auto         #let the gstreamer negotiate the format
    width: 1280
    height: 720
    framerate: 30

    source: /dev/video2  #IMX219 raw sensor that nees ISP
    format: rggb         #ISP will be added in the pipeline
    width: 1920
    height: 1080
    framerate: 30
    subdev-id: 2         #needed by ISP to control sensor params via ioctls

    source: /dev/video2  #IMX390 raw sensor that nees ISP
    width: 1936
    height: 1100
    format: rggb12       #ISP will be added in the pipeline
    subdev-id: 2         #needed by ISP to control sensor params via ioctls
    framerate: 30
    sen-id: imx390
    ldc: True            #LDC will be added in the pipeline

Make sure to configure correct format for camera input. jpeg for USB camera that supports MJPEG (Ex. C270 logitech USB camera). auto for CSI camera to allow GStreamer to negotiate the format. rggb for sensor that needs ISP.

Video sources

H.264 and H.265 encoded videos can be provided as input sources to the demos. Sample video files are provided under /opt/edge_ai_apps/data/videos/video_0000_h264.mp4 and /opt/edge_ai_apps/data/videos/video_000_h265.mp4

    source: ../data/videos/video_0000_h264.mp4
    format: h264
    width: 1280
    height: 720
    framerate: 25

    source: ../data/videos/video_0000_h265.mp4
    format: h265
    width: 1280
    height: 720
    framerate: 25

Make sure to configure correct format for video input as shown above. By default the format is set to auto which will then use the GStreamer bin decodebin instead.

Image sources

JPEG compressed images can be provided as inputs to the demos. A sample set of images are provided under /opt/edge_ai_apps/data/images. The names of the files are numbered sequentially and incrementally and the demo plays the files at the fps specified by the user.

    source: ../data/images/%04d.jpg
    width: 1280
    height: 720
    index: 0
    framerate: 1

RTSP sources

H.264 encoded video streams either coming from a RTSP compliant IP camera or via RTSP server running on a remote PC can be provided as inputs to the demo.

    source: rtsp:// # rtsp stream url, replace this with correct url
    width: 1280
    height: 720
    framerate: 30


Usually video streams from any IP camera will be encrypted and cannot be played back directly without a decryption key. We tested RTSP source by setting up an RTSP server on a Ubuntu 18.04 PC by referring to this writeup, Setting up RTSP server on PC


The model section defines a list of models that are used in the demo. Path to the model directory is a required argument for each model and rest are optional properties specific to given use cases like shown below.

        model_path: ../models/segmentation/ONR-SS-871-deeplabv3lite-mobv2-cocoseg21-512x512   #Model Directory
        alpha: 0.4                                                                            #alpha for blending segmentation mask (optional)
        model_path: ../models/detection/TFL-OD-202-ssdLite-mobDet-DSP-coco-320x320
        viz_threshold: 0.3                                                                    #Visualization threshold for adding bounding boxes (optional)
        model_path: ../models/classification/TVM-CL-338-mobileNetV2-qat
        topN: 5                                                                               #Number of top N classes (optional)

Below are some of the use case specific properties:

  1. alpha: This determines the weight of the mask for blending the semantic segmentation output with the input image alpha * mask + (1 - alpha) * image

  2. viz_threshold: Score threshold to draw the bounding boxes for detected objects in object detection. This can be used to control the number of boxes in the output, increase if there are too many and decrease if there are very few

  3. topN: Number of most probable classes to overlay on image classification output

The content of the model directory and its structure is discussed in detail in Import Custom Models


The output section defines a list of supported outputs.

    output0:                                                     #Display Output
        sink: kmssink
        width: 1920                                              #Width and Height of the output
        height: 1080
        connector: 39                                            #Connector ID for kmssink (optional)

    output1:                                                     #Video Output
        sink: ../data/output/videos/output_video.mkv             #Output video file
        width: 1920
        height: 1080

    output2:                                                     #Image Output
        sink: ../data/output/images/output_image_%04d.jpg        #Image file name, printf style formatting is used
        width: 1920
        height: 1080

        sink: remote                                             #Publish output to udp port as jpeg encoded frames
        width: 1280
        height: 720
        port: 8081                                               #udp port (optional default:8081)
        host:                                            #udp host (optional default:

All supported outputs are listed in template config file. Below are the details of most commonly used outputs

Display sink (kmssink)

When you have only one display connected to the SK, kmssink will try to use it for displaying the output buffers. In case you have connected multiple display monitors (e.g. Display Port and HDMI), you can select a specific display for kmssink by passing a specific connector ID number. Following command finds out the connected displays available to use.

Note: Run this command outside docker container. The first number in each line is the connector-id which we will use in next step.

root@tda4vm-sk:/opt/edge_ai_apps# modetest -M tidss -c | grep connected
39      38      connected       DP-1            530x300         12      38
48      0       disconnected    HDMI-A-1        0x0             0       47

From above output, we can see that connector ID 39 is connected. Configure the connector ID in the output section of the config file.

Video sinks

The post-processed outputs can be encoded in H.264 format and stored on disk. Please specify the location of the video file in the configuration file.

    sink: ../data/output/videos/output_video.mkv
    width: 1920
    height: 1080

Image sinks

The post-processed outputs can be stored as JPEG compressed images. Please specify the location of the image files in the configuration file. The images will be named sequentially and incrementally as shown.

    sink: ../data/output/images/output_image_%04d.jpg
    width: 1920
    height: 1080

Remote sinks

The JPEG compressed post-processed frames can be send as udp packets to a port Please specify the sink as remote in the configuration file. The udp port and host to send packets to can be defined. If not, default port is 8081 and host is

    sink: remote
    width: 1280
    height: 720
    port: 8081
    host: #IP of Remote PC.

The display can be viewed by running a simple gstreamer pipeline on the remote PC.

gst-launch-1.0 udpsrc port=8081 ! application/x-rtp,encoding=H264 ! rtph264depay ! h264parse ! avdec_h264 ! autovideosink


The flows section defines how inputs, models and outputs are connected. Multiple flows can be defined to achieve multi input, multi inference as shown


The format of specifying flows is changes as of 08.05.00 release to enable multiple outputs in the same sub-flow The older config files may not be compatible from this release onwards and should be changed to below format

    # flowname : [input,mode1,output,[mosaic_pos_x,mosaic_pos_y,width,height]]
    flow0: [input0,model1,output0,[160,90,800,450]]
    flow1: [input0,model2,output0,[960,90,800,450]]
    flow2: [input1,model0,output0,[160,540,800,450]]
    flow3: [input1,model3,output0,[960,540,800,450]]

Each flow defined here has exactly 1 input and 1 model. If multiple flows have same input, they are clubbed together internally in the application for optimization. Along with input, models and outputs it is required to define n mosaics which are the position of the inference output in the final output plane. This is needed because multiple inference outputs can be rendered to same output (Ex: Display).

Command line arguments

Limited set of command line arguments can be provided, run with ‘-h’ or ‘–help’ option to list the supported parameters.

usage: Run : ./ -h for help

positional arguments:
  config           Path to demo config file
                       ex: ./ ../configs/app_config.yaml

optional arguments:
  -h, --help       show this help message and exit
  -n, --no-curses  Disable curses report
                   default: Disabled
  -v, --verbose    Verbose option to print profile info on stdout
                   default: Disabled

GStreamer Plugins Map file

The edge_ai_apps essentially constructs GStreamer pipelines for dataflow. This pipeline is constructed optimally and dynamically based on a pool of specific plugins available on the platform. The defined pool of plugins for different platform can be found in edge_ai_apps/configs/gst_plugin_maps.yaml file.

This file contains the plugin used for certain task and the property of plugin (if applicable).

Default gst_plugin_maps for SK-TDA4VM

        element: tiovxdlcolorconvert
            target: [0,1] #[C66_0,  C66_1]
            out-pool-size: 4
        element: tiovxcolorconvert
            target: [0,1]
            out-pool-size: 4
        element: tiovxmultiscaler
        element: tiovxdlpreproc
            target: [0,1]
            out-pool-size: 4
        element: tiovxmosaic
        element: tiovxisp
        element: tiovxldc
        element: avdec_h264
        element: v4l2h265dec
        element: v4l2h264enc
            bitrate: 10000000
    enable-tidl: true


The SOC name for the platform SK-TDA4VM is j721e. It can also be found by printing the environment variable “SOC”