BlazePose: Body Segmentation for TFJS

Updated for BlazePose v2 variations

Included models:

• BlazePose Detector
• BlazePose Lite, Full & Heavy variations

Models included in /model/* were manually converted and quantized to TFJS Graph model format
from the original repository and are not identical to TFJS Graph models published on TF Hub

Models descriptors and signature have been additionally parsed for readability

Implementation

Actual model parsing implementation in /src/blazepose.ts does not follow reference implementation

BlazePose is a two-phase model:

• Detector
• Pose analysis

Pose analysis is a fast, clean and accurate model
However, detector is much slower and requires a lot of post-processing - thus its usage has been eliminated from this implementation

Instead, this implementation prepares virtual detected tensor padded and resized to fit pose analysis model

Implementation is in src/blazepose.ts with keypoint definitions and annotations in src/blazeposecoords.ts
The rest of sources in src/util is for simplicity of usage and testing only

Ideal implementation includes additional results caching and temporal interpolation with smoothing functionality
See https://github.com/vladmandic/human for details

Result

Implementation defines results a:

• keypoints: array of 39 detected keypoints
  • part: body part label
  • position: [x, y, z] normalized to input size
  • postitionRaw: [x, y, z] normalized to 0..1
  • score: detection score for the body part
• box: [x, y, width, height] box around detected body normalized to input size
• boxRaw: [x, y, width, height] box around detected body normalized to 0..1
• annotations: annotated arrays of points to help define higher level entities such as arm or leg
• score: average score for body

Demo

Demo app that initializes webcam and uses BlazePose is in demo/index.html
To run image analysis instead of default analysis on webcam input, edit demo/index.js Example:

await detectImage('daz3d-ella.jpg');

Credits & Links

• Blog: https://ai.googleblog.com/2020/08/on-device-real-time-body-pose-tracking.html
• Docs: https://google.github.io/mediapipe/solutions/pose
• Paper: https://arxiv.org/abs/2006.10204
• Card: https://drive.google.com/file/d/1UqfMdZ4sqioHULu6E-W0gp6JCcXAZPGK/view

Todo

• Enhance virtual tensor with virtual box around previous body center
  For details, see disabled code in calculateBoxes method
• HeatMap processing