import cv2 import pickle import os import argparse import mediapipe as mp import numpy as np import glob import time import yaml from PIL import Image import streamlit as st st.set_page_config(layout="wide") from utils.drawing_utils import draw_all_informations from utils.general import resize_img, get_cache_informations, parse_head_pose_informations from utils.streamlit_options import default_UI from utils.detection import detect_face_pose_informations_from_image IMG_SUFFIX = ['jpeg', 'jpg', 'png'] def main(): parser = argparse.ArgumentParser() # Pose position parser.add_argument('--head_pose_info', default='configs.yaml', help='path to head pose information') args = parser.parse_args() ## Streamlit options st.title("Head pose estimation tracking app 📷") left_col, right_col = st.columns([2, 6]) with left_col: default_UI() img_size = st.session_state.img_size.split('x') img_size = [int(s) for s in img_size] if st.session_state.img_upload is not None: mp_face_mesh = mp.solutions.face_mesh mp_face_detection = mp.solutions.face_detection face_mesh = mp_face_mesh.FaceMesh(min_detection_confidence=.5, min_tracking_confidence=0.5) with open(args.head_pose_info, 'r') as f: head_pose_info = yaml.load(f, Loader=yaml.FullLoader) parse_head_pose_informations(head_pose_info, st.session_state) img_dict = {} # List of images upload in RGB save_dict = {'img_size': img_size, 'position_horizontal_thresholds': st.session_state.position_horizontal_thresholds, 'position_vertical_thresholds': st.session_state.position_vertical_thresholds} image_results = {} # List of results for each image in RGB image_names = [] # List image names face_detection_dict = {} # Face detection information of each image face_pose_dict = {} # Face pose information of each image chosen_rectangle_pos_dict = {} # Positions where have face for each image face_direction_dict = {} # direction of each face for each image face_position_dict = {} # position of each face for each image face_coordinate_dict = {} # coordinate of each face for each image face_area_dict = {} if st.session_state.using_local_cache: cache_dict = get_cache_informations(save_dict) else: cache_dict = {} for img_upload_file in st.session_state.img_upload: image = np.array(Image.open(img_upload_file)) if len(image.shape) > 2 and image.shape[2] == 4: image = cv2.cvtColor(image, cv2.COLOR_RGBA2RGB) image_name = img_upload_file.name.split('.')[0] image_names.append(image_name) img_dict[image_name] = image if image_name in cache_dict: face_detection_dict[image_name] = cache_dict[image_name]['face_detection'] face_pose_dict[image_name] = cache_dict[image_name]['face_pose'] chosen_rectangle_pos_dict[image_name] = cache_dict[image_name]['chosen_rectangle_pos'] face_direction_dict[image_name] = cache_dict[image_name]['face_direction'] face_position_dict[image_name] = cache_dict[image_name]['face_position'] face_coordinate_dict[image_name] = cache_dict[image_name]['face_coordinate'] face_area_dict[image_name] = cache_dict[image_name]['face_area'] total_time = 0 with mp_face_detection.FaceDetection( model_selection=1, min_detection_confidence=0.5) as face_detection: for id, (image_name, image) in enumerate(img_dict.items()): if isinstance(image, str): image = cv2.imread(image) # Check if the image has only 3 channels if image.shape[-1] == 1: image = np.stack([image]*3, axis=-1) start = time.time() image, h_ratio, w_ratio, pad_top, pad_bot, pad_left, pad_right, img_size_no_pad = \ resize_img(image, img_size, return_all_infos=True) image_results[image_name] = image # To improve performance image.flags.writeable = False if not image_name in cache_dict: face_detection_infos, face_direction_infos, face_position_infos, face_coordinate_infos, face_pose_infos, face_area_infos, chosen_rectangle_pos_list = \ detect_face_pose_informations_from_image(image, face_detection, face_mesh, img_size, (pad_top, pad_bot, pad_left, pad_right), img_size_no_pad, head_pose_info) face_detection_dict[image_name] = face_detection_infos face_direction_dict[image_name] = face_direction_infos face_position_dict[image_name] = face_position_infos face_coordinate_dict[image_name] = face_coordinate_infos face_pose_dict[image_name] = face_pose_infos chosen_rectangle_pos_dict[image_name] = chosen_rectangle_pos_list face_area_dict[image_name] = face_area_infos end = time.time() cur_time = end - start total_time += cur_time draw_all_informations(image_results[image_name], face_detection_dict[image_name], face_direction_dict[image_name], face_position_dict[image_name], face_coordinate_dict[image_name], face_pose_dict[image_name], face_area_dict[image_name], chosen_rectangle_pos_dict[image_name], img_size, (pad_top, pad_bot, pad_left, pad_right), (w_ratio, h_ratio), st.session_state, head_pose_info) ## Save results to dict if st.session_state.using_local_cache: save_dict[image_name] = {} save_dict[image_name]['face_detection'] = face_detection_dict[image_name] save_dict[image_name]['face_pose'] = face_pose_dict[image_name] save_dict[image_name]['face_direction'] = face_direction_dict[image_name] save_dict[image_name]['face_position'] = face_position_dict[image_name] save_dict[image_name]['chosen_rectangle_pos'] = chosen_rectangle_pos_dict[image_name] save_dict[image_name]['face_coordinate'] = face_coordinate_dict[image_name] save_dict[image_name]['face_area'] = face_area_dict[image_name] if image_name in cache_dict: # Remove image from cache to save memory del cache_dict[image_name] print(f"FPS: {len(image_results) / total_time}") ## Save results to local cache if st.session_state.using_local_cache: with open('cache.pkl', 'wb') as f: pickle.dump(save_dict, f, protocol=pickle.HIGHEST_PROTOCOL) image_results = [image_results[image_name] for image_name in image_names] with right_col: st.image(image_results, width=st.session_state.width_visual, caption=image_names) if __name__ == '__main__': main()