import cv2 import json import numpy as np import pandas as pd import os import time def draw_hands_connections(frame, hand_landmarks): ''' Draw white lines between relevant points of hands landmarks Parameters ---------- frame: numpy array, corresponding to the frame on which we want to draw hand_landmarks: dictionnary, collecting the hands landmarks Return ------ frame: numpy array, with the newly drawing of the hands ''' # define hand_connections between keypoints hand_connections = [[0, 1], [1, 2], [2, 3], [3, 4], [5, 6], [6, 7], [7, 8], [9, 10], [10, 11], [11, 12], [13, 14], [14, 15], [15, 16], [17, 18], [18, 19], [19, 20]] #[5, 2], [0, 17]] # loop to draw left hand connection for connection in hand_connections: landmark_start = hand_landmarks['left_hand'].get(str(connection[0])) landmark_end = hand_landmarks['left_hand'].get(str(connection[1])) cv2.line(frame, landmark_start, landmark_end, (255, 255, 255), 2) # loop to to draw right hand connection for connection in hand_connections: landmark_start = hand_landmarks['right_hand'].get(str(connection[0])) landmark_end = hand_landmarks['right_hand'].get(str(connection[1])) cv2.line(frame, landmark_start, landmark_end, (255, 255, 255), 2) return frame def draw_pose_connections(frame, pose_landmarks): ''' Draw white lines between relevant points of pose landmarks Parameters ---------- frame: numpy array, corresponding to the frame on which we want to draw hand_landmarks: dictionnary, collecting the pose landmarks Return ------ frame: numpy array, with the newly drawing of the pose ''' # define pose connections pose_connections = [[11, 12], [11, 13], [12, 14], [13, 15], [14, 16]] for connection in pose_connections: landmark_start = pose_landmarks.get(str(connection[0])) landmark_end = pose_landmarks.get(str(connection[1])) cv2.line(frame, landmark_start, landmark_end, (255, 255, 255), 2) return frame def draw_face_connections(frame, face_landmarks): ''' Draw white lines between relevant points of face landmarks Parameters ---------- frame: numpy array, corresponding to the frame on which we want to draw hand_landmarks: dictionnary, collecting the face landmarks Return ------ frame: numpy array, with the newly drawing of the face ''' # define pose connections connections_dict = {'lipsUpperInner_connections' : [78, 191, 80, 81, 82, 13, 312, 311, 310, 415, 308],\ 'lipsLowerInner_connections' : [78, 95, 88, 178, 87, 14, 317, 402, 318, 324, 308],\ 'rightEyeUpper0_connections': [246, 161, 160, 159, 158, 157, 173],\ 'rightEyeLower0' : [33, 7, 163, 144, 145, 153, 154, 155, 133],\ 'rightEyebrowLower' : [35, 124, 46, 53, 52, 65],\ 'leftEyeUpper0' : [466, 388, 387, 386, 385, 384, 398],\ 'leftEyeLower0' : [263, 249, 390, 373, 374, 380, 381, 382, 362],\ 'leftEyebrowLower' : [265, 353, 276, 283, 282, 295],\ 'noseTip_midwayBetweenEye' : [1, 168],\ 'noseTip_noseRightCorner' : [1, 98],\ 'noseTip_LeftCorner' : [1, 327]\ } for keypoints_list in connections_dict.values(): for index in range(len(keypoints_list)): if index + 1 < len(keypoints_list): landmark_start = face_landmarks.get(str(keypoints_list[index])) landmark_end = face_landmarks.get(str(keypoints_list[index+1])) cv2.line(frame, landmark_start, landmark_end, (255, 255, 255), 1) return frame def resize_landmarks(landmarks, resize_rate_width, resize_rate_height): for keypoint in landmarks.keys(): landmark_x, landmark_y = landmarks[keypoint] landmarks[keypoint] = [int(resize_rate_width * landmark_x), int(resize_rate_height*landmark_y)] return landmarks def generate_video(gloss_list, dataset, vocabulary_list): # size of video of signer 11 # FIXED_WIDTH, FIXED_HEIGHT, = 288, 192, FIXED_WIDTH, FIXED_HEIGHT = 576, 384 fps = 25 for gloss in gloss_list: if not check_gloss_in_vocabulary(gloss, vocabulary_list): continue video_id = select_video_id_from_gloss(gloss, dataset) video_landmarks_path = dataset.loc[dataset['video_id'] == video_id, 'video_landmarks_path'].values[0] with open(video_landmarks_path, 'r') as f: video_landmarks = json.load(f) width = video_landmarks[-1].get('width') height = video_landmarks[-1].get('height') # calculate resize rate resize_rate_width, resize_rate_height = FIXED_WIDTH / width, FIXED_HEIGHT/height text = gloss font = cv2.FONT_HERSHEY_SIMPLEX font_scale = 1 font_color = (0, 255, 0) thickness = 2 line_type = cv2.LINE_AA for frame_landmarks in video_landmarks[:-1]: blank_image = np.zeros((FIXED_HEIGHT, FIXED_WIDTH, 3), dtype=np.uint8) frame_hands_landmarks = frame_landmarks['hands_landmarks'] frame_pose_landmarks = frame_landmarks['pose_landmarks'] frame_face_landmarks = frame_landmarks['face_landmarks'] #left_hand_landmarks_xy = [(x, y) for x, y in frame_hands_landmarks['left_hand'].values()] #right_hand_landmarks_xy = [(x, y) for x, y in frame_hands_landmarks['right_hand'].values()] #for x, y in left_hand_landmarks_xy: # cv2.circle(blank_image, (x, y), 1, (255, 255, 255), -1) #for x, y in right_hand_landmarks_xy: # cv2.circle(blank_image, (x, y), 1, (255, 255, 255), -1) # pose_landmarks_xy = [(x, y) for x, y in frame_pose_landmarks.values()] # for x, y in pose_landmarks_xy: # cv2.circle(blank_image, (x, y), 1, (255, 255, 255), -1) # face_landmarks_xy = [(x, y) for x, y in frame_face_landmarks.values()] # for x, y in face_landmarks_xy: # cv2.circle(blank_image, (x, y), 1, (255, 255, 255), -1) frame_hands_landmarks_rs = { 'left_hand': resize_landmarks(frame_hands_landmarks['left_hand'], resize_rate_width, resize_rate_height), 'right_hand': resize_landmarks(frame_hands_landmarks['right_hand'], resize_rate_width, resize_rate_height) } frame_pose_landmarks_rs = resize_landmarks(frame_pose_landmarks, resize_rate_width, resize_rate_height) frame_face_landmarks_rs = resize_landmarks(frame_face_landmarks, resize_rate_width, resize_rate_height) draw_hands_connections(blank_image, frame_hands_landmarks_rs) draw_pose_connections(blank_image, frame_pose_landmarks_rs) draw_face_connections(blank_image, frame_face_landmarks_rs) text_size, _ = cv2.getTextSize(text, font, font_scale, thickness) text_x = (FIXED_WIDTH - text_size[0]) // 2 text_y = FIXED_HEIGHT - 10 cv2.putText(blank_image, text, (text_x, text_y), font, font_scale, font_color, thickness, line_type) # Convertir l'image en JPEG encodé _, buffer = cv2.imencode('.jpg', blank_image) frame = buffer.tobytes() yield (b'--frame\r\n' b'Content-Type: image/jpeg\r\n\r\n' + frame + b'\r\n') time.sleep(1 / fps) def load_data(dataset_path='local_dataset'): filepath = dataset_path data_df = pd.read_csv(filepath, dtype={'video_id': str}) vocabulary_list = data_df['gloss'].tolist() return data_df, vocabulary_list def check_gloss_in_vocabulary(gloss, vocabulary_list): return gloss in vocabulary_list def select_video_id_from_gloss(gloss, dataset): filtered_data_id_11 = dataset.loc[dataset['signer_id'] == 11] if gloss in filtered_data_id_11['gloss'].tolist(): video_id = filtered_data_id_11.loc[filtered_data_id_11['gloss'] == gloss, 'video_id'].values else: video_id = dataset.loc[dataset['gloss'] == gloss, 'video_id'].values return video_id[0]