app.py

#!/usr/bin/env python

from __future__ import annotations

import os
import pathlib
import shlex
import subprocess
import tarfile

if os.environ.get('SYSTEM') == 'spaces':
    subprocess.call(shlex.split('pip uninstall -y opencv-python'))
    subprocess.call(shlex.split('pip uninstall -y opencv-python-headless'))
    subprocess.call(
        shlex.split('pip install opencv-python-headless==4.5.5.64'))

import gradio as gr
import huggingface_hub
import mediapipe as mp
import numpy as np

mp_drawing = mp.solutions.drawing_utils
mp_drawing_styles = mp.solutions.drawing_styles
mp_pose = mp.solutions.pose

TITLE = 'MediaPipe Human Pose Estimation'
DESCRIPTION = 'https://google.github.io/mediapipe/'

HF_TOKEN = os.getenv('HF_TOKEN')


def load_sample_images() -> list[pathlib.Path]:
    image_dir = pathlib.Path('images')
    if not image_dir.exists():
        image_dir.mkdir()
        dataset_repo = 'hysts/input-images'
        filenames = ['002.tar']
        for name in filenames:
            path = huggingface_hub.hf_hub_download(dataset_repo,
                                                   name,
                                                   repo_type='dataset',
                                                   use_auth_token=HF_TOKEN)
            with tarfile.open(path) as f:
                f.extractall(image_dir.as_posix())
    return sorted(image_dir.rglob('*.jpg'))


def run(image: np.ndarray, model_complexity: int, enable_segmentation: bool,
        min_detection_confidence: float, background_color: str) -> np.ndarray:
    with mp_pose.Pose(
            static_image_mode=True,
            model_complexity=model_complexity,
            enable_segmentation=enable_segmentation,
            min_detection_confidence=min_detection_confidence) as pose:
        results = pose.process(image)

    res = image[:, :, ::-1].copy()
    if enable_segmentation:
        if background_color == 'white':
            bg_color = 255
        elif background_color == 'black':
            bg_color = 0
        elif background_color == 'green':
            bg_color = (0, 255, 0)  # type: ignore
        else:
            raise ValueError

        if results.segmentation_mask is not None:
            res[results.segmentation_mask <= 0.1] = bg_color
        else:
            res[:] = bg_color

    mp_drawing.draw_landmarks(res,
                              results.pose_landmarks,
                              mp_pose.POSE_CONNECTIONS,
                              landmark_drawing_spec=mp_drawing_styles.
                              get_default_pose_landmarks_style())

    return res[:, :, ::-1]


model_complexities = list(range(3))
background_colors = ['white', 'black', 'green']

image_paths = load_sample_images()
examples = [[
    path.as_posix(), model_complexities[1], True, 0.5, background_colors[0]
] for path in image_paths]

gr.Interface(
    fn=run,
    inputs=[
        gr.Image(label='Input', type='numpy'),
        gr.Radio(label='Model Complexity',
                 choices=model_complexities,
                 type='index',
                 value=model_complexities[1]),
        gr.Checkbox(default=True, label='Enable Segmentation'),
        gr.Slider(label='Minimum Detection Confidence',
                  minimum=0,
                  maximum=1,
                  step=0.05,
                  value=0.5),
        gr.Radio(label='Background Color',
                 choices=background_colors,
                 type='value',
                 value=background_colors[0]),
    ],
    outputs=gr.Image(label='Output', type='numpy'),
    examples=examples,
    title=TITLE,
    description=DESCRIPTION,
).launch(show_api=False)