/blaze-palm-tf2

Primary LanguageJupyter Notebook

MediaPipe BlazePalm Model

An unofficial Implementation BlazePalm ( hand detector ) using Tensorflow 2.0.

Usage

Install dependencies

pip install -r requirements.txt

Training

Prepare training and validation data

  • Directory Structure

    In directory dataset, create image and annotation folders

    dataset
    |----train_image
    |----train_annotation
    |----val_image
    |----val_annotation

  • Annotation Format

    Each image need a corresponding Json file that contains all hand items in the image.

    Each item in a json file should have 3 attributes label, bounding_box and key_points

    label: always be 1 (only one class)

    bounding_box: normalized 4 values, x_min, y_min, x_max, y_max

    key_points: 7 key points in total, [x1, y1, x2, y2,... x7, y7]

    [
  {
    "label": "1",
    "bounding_box": [
      0.739062488079071,
      0.3499999940395355,
      0.839062511920929,
      0.4833333194255829
    ],
    "key_points": [
      0.7484375,
      0.47708333333333336,
      0.8140625,
      0.35625,
      0.828125,
      0.39166666666666666,
      0.8359375,
      0.43125,
      0.8390625,
      0.4708333333333333,
      0.7421875,
      0.425,
      0.775,
      0.37083333333333335
    ]
  }
]

Run train.py

  • Confirm the Training Config in train.py is correct then run.

Inference

TODO

Result

TODO

  • Refactor code

  • Train a model on open dataset

  • Decoder for prediction result

  • Inference and visualize

Reference