See this blog post (TBD) for a full guide on doing pose estimation and classification using TensorFlow Lite.

• Pose estimation: Detect keypoints, such as eye, ear, arm etc., from an input image.
  • Input: An image
  • Output: A list of keypoint coordinates and confidence score.
• Pose classificaiton: Classify a human pose into predefined classes, such as different yoga poses. Pose classification internally use pose estimation to detect the keypoints, and use the keypoints to classify the pose.
  • Input: An image
  • Output: A list of predefined classes and their confidence score.

This sample can run on Raspberry Pi or any computer that has a camera. It uses OpenCV to capture images from the camera and TensorFlow Lite to run inference on the input image.

• Run this script to install the Python dependencies, and download the TFLite models. sh setup.sh

• Use this command to run the pose estimation sample using the default movenet_lightning model.

python3 pose_estimation.py

• You can optionally specify the model_name parameter to try other pose estimation models:
  • Use values:
    • Single-pose: posenet, movenet_lightning, movenet_thunder
    • Multi-poses: movenet_multipose
  • The default value is movenet_lightning.

python3 pose_estimation.py --model_name movenet_thunder

• Use this command to run the pose estimation sample using the default movenet_lightning pose estimation model and the classifier.tflite yoga pose classification model.

python3 pose_estimation.py \
    --classifier classifier
    --label_file labels.txt

• If you want to train a custom pose classification model, check out this tutorial.

• Here is the full list of parameters supported by the sample: python3 pose_classification.py
• model: Name of the TFLite pose estimation model to be used. * One of these values: posenet, movenet_lightning, movenet_thunder, movenet_multipose * Default value is movenet_lightning.
• tracker: Type of tracker to track poses across frames. * One of these values: bounding_box, keypoint * Only supported in multi-poses models. * Default value is bounding_box.
• classifier: Name of the TFLite pose classification model to be used. * Default value is empty. * If no classification model specified, the sample will only run the pose estimation step.
• camera_id: Specify the camera for OpenCV to capture images from. * Default value is 0.
• frameWidth, frameHeight: Resolution of the image to be captured from the camera. * Default value is (640, 480).

• Run this script to visualize the pose estimation on test data

python3 visualizer.py