• https://www.tensorflow.org/hub/tutorials/movenet
• https://tfhub.dev/s?q=movenet

Movenet appelle Tensorflow Lite, qui n'est pas optimisé pour GPU. Inutile d'installer Cuda.
Il ne tourne que sur CPU, et demande à avoir AVX2 et FMA.
Pour compiler Tensorflow, voir Linux Compiler Tensorflow avec Bazel

Altrnatives, trouver un truc tout fait:

• https://www.intel.com/content/www/us/en/developer/articles/guide/optimization-for-tensorflow-installation-guide.html

• https://tfhub.dev/google/lite-model/movenet/singlepose/thunder/3

movenet_test.py

from time import time
import cv2
import numpy as np
import tensorflow as tf
 
def draw(frame, keypoints_with_scores):
    for item in keypoints_with_scores[0][0]:
        if item[2] > 0.3:  # confiance
            x = int(item[1]*720)  # numpy array x est le 2ème
            y = int(item[0]*720)
            cv2.circle(frame, (x, y), 6, color=(0,0,255), thickness=-1)
    return frame
 
models = [  "lite-model_movenet_singlepose_thunder_3.tflite",
            "lite-model_movenet_singlepose_thunder_tflite_float16_4.tflite",
            "lite-model_movenet_singlepose_thunder_tflite_int8_4.tflite"]
 
interpreter = tf.lite.Interpreter(model_path=models[0])
interpreter.allocate_tensors()
 
cv2.namedWindow('color', cv2.WND_PROP_FULLSCREEN)
cam = cv2.VideoCapture(2)
cam.set(cv2.CAP_PROP_FRAME_WIDTH, 1280)
cam.set(cv2.CAP_PROP_FRAME_HEIGHT, 720)
 
t0 = time()
nbr = 0
 
while 1:
 
    nbr += 1
    ret, frame = cam.read()
 
    if not ret:
        continue
 
    # Extraction d'une image carrée [rows, columns]
    frame = frame[:, 280:1000]
 
    image = tf.expand_dims(frame, axis=0)
    # Resize and pad the image to keep the aspect ratio and fit the expected size.
    image = tf.image.resize_with_pad(image, 256, 256)
 
    # TF Lite format expects tensor type of float32.
    input_image = tf.cast(image, dtype=tf.float32)
    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()
 
    interpreter.set_tensor(input_details[0]['index'], input_image.numpy())
    interpreter.invoke()
 
    # Output is a [1, 1, 17, 3] numpy array.
    keypoints_with_scores = interpreter.get_tensor(output_details[0]['index'])
 
    frame = draw(frame, keypoints_with_scores)
 
    # Affichage de l'image
    cv2.imshow('color', frame)
 
    # Calcul du FPS, affichage toutes les 10 s
    if time() - t0 > 10:
        print("FPS =", round(nbr/10, 2))
        t0, nbr = time(), 0
 
    # Pour quitter
    if cv2.waitKey(1) == 27:
        break
 
cv2.destroyAllWindows()

• Essai de movenet par SergeLabo