blazepose

BlazePose Upper Body

Input

(Image from https://pixabay.com/ja/photos/%E5%A5%B3%E3%81%AE%E5%AD%90-%E7%BE%8E%E3%81%97%E3%81%84-%E8%8B%A5%E3%81%84-%E3%83%9B%E3%83%AF%E3%82%A4%E3%83%88-5204299/)

Detector

• ailia input shape: (1, 3, 128, 128)
• Range:

Estimator

• ailia input shape: (1, 3, 256, 256)

Output

Detector

• ailia Predict API output:

Estimator

• ailia Predict API output:
  • Segmentation
    • Shape: (1, 1, 128, 128)
  • Landmarks
    • Shape: (1, 31, 4)
    • 25 upper-body keypoints (x, y, z, visibility) + 2 virtual keypoints (x, y, z, visibility) + 4 ??? face bounding boxes (x, y, z, visibility)?
  • Flag
    • Shape: (1,)

Usage

Automatically downloads the onnx and prototxt files on the first run. It is necessary to be connected to the Internet while downloading.

For the sample image,

$ python3 blazepose.py 

If you want to specify the input image, put the image path after the --input option.
You can use --savepath option to change the name of the output file to save.

$ python3 blazepose.py --input IMAGE_PATH --savepath SAVE_IMAGE_PATH

By adding the --video option, you can input the video.
If you pass 0 as an argument to VIDEO_PATH, you can use the webcam input instead of the video file.

$ python3 blazepose.py --video VIDEO_PATH --savepath SAVE_VIDEO_PATH

Reference

MediaPipePyTorch

Squeeze function in Pytorch model removed as it produces an IF layer in ONNX unsupported in ailia (segmentation output has thus 1 more dim)

Framework

PyTorch 1.7.1

Model Format

ONNX opset = 11

Netron

blazepose_detector.onnx.prototxt

blazepose_estimator.onnx.prototxt