Coding Explanation

Major Functionalities

For our Pose Estimation task, we use NVIDIA built python program called the posenet.py. The program does 3 major functions:

  1. It manages input streams whether it is an image, a video or a camera.

  2. It manages network inference on the said input stream.

  3. It manages output stream along with overlaying results from the model.

With our python program, the user must specify

  1. The input stream medium. (whether to input an image, a video or a camera)

  2. Which model network to use for the inference.

  3. The output stream medium. (whether to output an image, a video or a camera)

  • Within our pose_estimation.ipynb, we can see that we first initialize the PATH for our input and output images. We set up jupyter environment variable with env as well as python variable.

    %env DISPLAY=:0
%env PROGRAM_PATH=/home/zeta/jetson-inference/build/aarch64/bin
%env INPUT_PATH=/home/zeta/jetson-inference/build/aarch64/bin/images
%env OUTPUT_PATH=/home/zeta/jetson-inference/build/aarch64/bin/images/test

input_path='/home/zeta/jetson-inference/build/aarch64/bin/images'
output_path='/home/zeta/jetson-inference/build/aarch64/bin/images/test'

    By setting up environment variables, we can call upon these PATH variables, when we wish to run a shell command within our jupyter notebook.

  • After choosing and initializing the name of our input images we can see that the IMAGE_NAME variable is also being set to be an environment variable.

  • When we execute our python program (posenet.py), we can see the three parameters the user has set up

    !python3 $PROGRAM_PATH/posenet.py --network=resnet18-body $INPUT_PATH/$IMAGE_NAME $OUTPUT_PATH/$OUTPUT_NAME

    With !python3 $PROGRAM_PATH/posenet.py as our activation function, we have:

    1. --network=resnet18-body: to specify which network we wished to use. (for this case Pose-ResNet18-Body network)

    2. $INPUT_PATH/$IMAGE_NAME: to specify our input image location.

    3. $OUTPUT_PATH/$OUTPUT_NAME: to specify our output image location.

  • Jupyter notebook allows us view images within our browser environment. For this we have used IPython library with display module.

    from IPython.display import Image

    • To view the images simply specify the location of the image. For example, to view the result of our inference:

      Image(filename=output_path+'/human_result.jpg')

Minor Functionalities

There are optional flags or parameters you may set up for your posenet.py program. These parameters may control the width of the line connecting the keypoints the size of the keypoints displayed on the screen, how they are connected and so on.

There are 4 minor functionalities you may edit:

  1. --overlay=links,keypoints: This parameter allows us to choose how we wish to display our detected poses on top of our original input. The input can be comma-seperated combinations of box, links, keypoints, and none.

  2. --keypoint-scale=0.0052: This parameter controls the radius of the keypoint circle in the overlay.

  3. --link-scale=0.0013: This parameter controls the width of the link lines in the overlay.

  4. --threshold=0.15: This parameter sets the minimum threshold for detection.