tikitaka02 / mlkit-pose-detector-android-demo

mlkit-pose-detector-android-demo

YouTube Tutorial Part 1: https://youtu.be/RJkB1OWLox8

YouTube Tutorial Part 2: https://youtu.be/80bsPKgQtA4

Todo List

• Base function

• Camera permission - Complete (Tutorial Part 1)
• CameraX implementation - Complete (Tutorial Part 1)
• mlkit pose detector implementation - Complete (Tutorial Part 1)
• Canvas drawing - Complete (Tutorial Part 1)
• Visualization flipping - Complete (Tutorial Part 2)
• Time synchronization - Complete (Tutorial Part 2)

• Game function

• A calibration function to check if all the body is visible - Incomplete
• Timer - Incompelte
• Display picture - Incomplete
• Play sound - Incomplete
• Joint angle calculation - Incomplete

mlkit-pose-detector-android-demo's Issues

Pose estimation Android benchmark

First of all, thanks for making your work public - I know how much effort it takes to publish code that works out of the box.

I've come to your repository from the issue you raised on ultralytics about Yolov8. I'm also into Yolov8, and I've tested it on several platforms: Linux, Windows, and Android. While PC-like platforms give ~15 FPS with ONNXRuntime inference engine and yolov8n-pose.onnx file, the Android part is really frustrating: ONNX Yolov8 Pose on Android gives only 1.5 FPS! Your repository with ML Kit gives ~15 FPS on Android... quite a difference!

Then I saw that you

converted a yolov8 pose detector model to tflite and implemented into android application.

What's the FPS you get with TFLite Yolov8 Pose on Android?

The nice thing about YoloX Pose series is that it's able to detect more than one pose in an image while the ML Kit uses the Mediapipe Pose Landmark model under the hood which is limited to a single person only solution.

Last but not least, I saw you forked https://github.com/MVIG-SJTU/AlphaPose - have you tested it on Android?

Thanks.