Video Intelligence Chapter 3: MediaPipe

Video Intelligence Oct 18, 2021

Video intelligence is one of the solutions that are most in demand nowadays. Performing object detection on an image is a good thing, but if we can perform that on a streaming video, that is a solution people will pay up for.

Video intelligence has a lot of applications, from domestic security to identifying product defects in industries. Have you ever used filters on video calls with your friends? That is the simplest example of video intelligence.

In previous posts in this series, we saw how can we make use of GCP Video Intelligence APIs to derive insights from a video. Also, we saw how can we train an AutoML model on our own video dataset.

In today's post, we will take a step further and explore how can we generate our own AI engine that will give us real-time insights out of a video clip using MediaPipe.


MediaPipe is an open-source python framework developed by Google, which provides solutions for major Video Intelligence use-cases.

  1. Face detection
  2. Face-mesh detection
  3. Hand gesture detection
  4. Pose detection

MediaPipe is not limited to python libraries, it provides support for Android, IOS, and Javascript as well. Check out the detailed documentation for more details.


Enough of theory, now let's create our own AI processors.

Step 1: Setup and Installation

Install MediaPipe and OpenCV dependency.

pip3 install mediapipe
pip3 install opencv-python

Import the library in your python script.

import cv2
import mediapipe as mp

Step 2: Build your processor

Now that we have imported all the necessary libraries, let's prepare our pose-detection processor. We will be using the Computer Vision library for support.

import cv2
import mediapipe as mp

capture = cv2.VideoCapture("path/to/source/video_file")
draw =
mpPose =
pose = mpPose.Pose()
frame_width = int(capture.get(3))
frame_height = int(capture.get(4))
size = (frame_width, frame_height)
result = cv2.VideoWriter("path/to/target/video_file", cv2.VideoWriter_fourcc(*"MJPG"), 10, size)
while True:
    success, img =
    imgRGB = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    results = pose.process(imgRGB)
    if results.pose_landmarks:
        # print(results.pose_landmarks.landmark)
        draw.draw_landmarks(img, results.pose_landmarks, mpPose.POSE_CONNECTIONS)
    cv2.imshow("Image", img)

Let's try to understand what is done here.

  1. We captured frames from our input video file using the OpenCV video capture function.
  2. We instantiated the MediaPipe pose estimator class.
  3. Modified the default BGR color format of the image to RGB.
  4. Process the frame to get resulting landmarks for a human pose.
  5. Visualize the results and save them in target video format using OpenCV functions.

Step 3: Process your video file

Time to enjoy the results. Provide the video file path to the above script and sit back and relax. Give some time for the process to end and check out the output video file generated.

Impressed by the results?

Check out the video below which will give you an overview of all other features available with MediaPipe.

All other features provided by MediaPipe


MediaPipe also provides the library for Android, IOS and JS. Which makes it very easy to use and lightweight. The applications of the library are explosive.

  1. Surveillance cameras and drones
  2. Understanding sign-language
  3. Attention monitoring
  4. Video games and many more...


This was a short introduction to an awesome open-source framework by Google called MediaPipe. This is not it, you can tweak parameters and check out the variations in the results.

Check out our GitHub repository where we have tackled Hand detection, face detection, and face-mesh detections as well.

GitHub - Chronicles-of-AI/AI-kosh: Archives of blogs on Chronicles of AI
Archives of blogs on Chronicles of AI. Contribute to Chronicles-of-AI/AI-kosh development by creating an account on GitHub.

Stay tuned for more awesome content. Stay healthy and stay curious. 🤠


Arpit Jain

Machine Learning Engineer

Great! You've successfully subscribed.
Great! Next, complete checkout for full access.
Welcome back! You've successfully signed in.
Success! Your account is fully activated, you now have access to all content.