import cv2 import json import numpy as np import pandas as pd import time def draw_hands_connections(frame, hand_landmarks): ''' Draw white lines on the given frame between relevant hand keypoints. Parameters ---------- frame: numpy array The frame on which we want to draw. hand_landmarks: dict Dictionary mapping keypoint IDs (integers) to hand landmarks (lists of two floats corresponding to the coordinates) for both hands. Returns ------- frame: numpy array The frame with the newly drawn hand connections. ''' # ---- Define hand_connections between keypoints to draw # 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 connections # 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 connections # 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 on the given frame between relevant posture keypoints. Parameters ---------- frame: numpy array The frame on which we want to draw. pose_landmarks: dict Dictionary mapping keypoint IDs (integers) to posture landmarks (lists of two floats corresponding to the coordinates). Returns ------- frame: numpy array The frame with the newly drawn posture connections. ''' # ---- define posture connections between keypoints to draw # pose_connections = [[11, 12], [11, 13], [12, 14], [13, 15], [14, 16]] # ---- loop to to draw posture connections # 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 on the given frame between relevant face keypoints. Parameters ---------- frame: numpy array The frame on which we want to draw. face_landmarks: dict Dictionary mapping keypoint IDs (integers) to face landmarks (lists of two floats corresponding to the coordinates). Returns ------- frame: numpy array The frame with the newly drawn face connections. ''' # ---- 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]\ } # ---- loop to to draw face connections # 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): ''' Resize landmark coordinates by applying specific scaling factors to both the width and height of the frame. Parameters ---------- landmarks: dict Dictionary mapping keypoint IDs (integers) to landmarks (lists of two floats corresponding to the coordinates). resize_rate_width: float Scaling factor applied to the x-coordinate (width). resize_rate_height: float Scaling factor applied to the y-coordinate (height). Returns ------- landmarks: dict Dictionary mapping keypoint IDs (integers) to the newly resized landmarks (lists of two integers corresponding to the coordinates). ''' 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): ''' Generate a video stream from a list of glosses. Parameters ---------- gloss_list: list of str List of glosses from which the signing video will be generated. dataset: pandas.DataFrame Dataset containing information about each gloss, including paths to landmark data. vocabulary_list: list of str List of tokens that have associated landmarks collected. Yields ------ frame: bytes JPEG-encoded frame for streaming. ''' # ---- Fix size of the frame to the most common size of video we have in the dataset # (corresponding to signer ID 11 who has the maximum number of videos). # FIXED_WIDTH, FIXED_HEIGHT = 576, 384 # ---- Fix the Frames Per Second (FPS) to match the videos collected in the dataset. # FPS = 25 # ---- Define carachteristics for text display. # font = cv2.FONT_HERSHEY_SIMPLEX font_scale = 1 font_color = (0, 255, 0) thickness = 2 line_type = cv2.LINE_AA # ---- Loop over each gloss # for gloss in gloss_list: # ---- Skip if gloss not in the vocabulary_list. # if not check_gloss_in_vocabulary(gloss, vocabulary_list): continue # ---- Get landmarks of all the frame in the dataset corresponding to the appropriate gloss. # 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 for future landmark rescaling. # resize_rate_width, resize_rate_height = FIXED_WIDTH / width, FIXED_HEIGHT/height # ---- Loop over each frame # for frame_landmarks in video_landmarks[:-1]: # ---- Initialize blank image and get all landmarks of the given frame. # 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'] # ---- Resize landmarks. # 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 relevant connections between keypoints on the frame. # 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) # ---- Display text corresponding to the gloss on the frame. # text_size, _ = cv2.getTextSize(gloss, font, font_scale, thickness) text_x = (FIXED_WIDTH - text_size[0]) // 2 text_y = FIXED_HEIGHT - 10 cv2.putText(blank_image, gloss, (text_x, text_y), font, font_scale, font_color, thickness, line_type) # ---- JPEG-encode the frame for streaming. # _, 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='enhanced_dataset'): ''' Load the dataset that contains all information about glosses. Parameters ---------- dataset_path: str Local path to the dataset. Returns ------- data_df: pandas.DataFrame DataFrame containing the dataset with information about each gloss. vocabulary_list: list of str List of glosses (tokens) that have associated landmarks collected. ''' 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): ''' Check if the given gloss is in the vocabulary list. Parameters ---------- gloss: str The gloss to check. vocabulary_list: list of str List of glosses (tokens) that have associated landmarks collected. Returns ------- bool True if the gloss is in the vocabulary list, False otherwise. ''' return gloss in vocabulary_list def select_video_id_from_gloss(gloss, dataset): ''' Selects a video ID corresponding to the given gloss from the dataset. Parameters ---------- gloss : str The gloss for which to retrieve the video ID. dataset : pandas.DataFrame A DataFrame containing information about each gloss, including 'signer_id', 'gloss', and 'video_id'. Returns ------- int The video ID corresponding to the given gloss. If the gloss is found for 'signer_id' 11, the video ID for that signer is returned; otherwise, the video ID for the gloss from the entire dataset is returned. ''' # ---- Choose preferentialy ID 11 because this signer with this ID signed the more video # 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]