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:
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()