manhdo's picture
Clean code and fix out of boundary prediction box
aef4429
Raw History Blame Contribute Delete
9.05 kB
import numpy as np
import glob
import cv2
import os
import pickle
def resize_img(img, size=(640, 640), pad_val=0, return_all_infos=False):
"""
Resize specific image according to original size ratio
then padding for yielding given size
Input:
`img` (np.ndarray): original image
`size` (tuple/list): size of new image (width, height)
`pad_val` (int): padding value (default: 0)
`return_all_infos` (bool): Return all informations for creating new sample
Return:
`New image` (np.ndarray) with given size
If `return_all_infos` is True:
`h_ratio`: h_new/h_old
`w_ratio`: w_new/w_old
`pad_top`: number of padding on top of image
`pad_bot`: number of padding on bot of image
`pad_left`: number of padding on left of image
`pad_right`: number of padding on right of image
"""
# Check pad val
if isinstance(pad_val, int):
pad_val = (pad_val, pad_val, pad_val)
old_size = img.shape[:2] # h, w
hw_ratio = old_size[0] / old_size[1]
if old_size[0] > old_size[1]:
new_size = (size[0], int(size[1] / hw_ratio))
else:
new_size = (int(size[0] * hw_ratio), size[1])
img = cv2.resize(img, (new_size[1], new_size[0])) # resize with hw-ratio
# Get padding corner to yield given new size
pad_top = (size[0] - new_size[0]) // 2
pad_bot = size[0] - (new_size[0] + pad_top)
pad_left = (size[1] - new_size[1]) // 2
pad_right = size[1] - (new_size[1] + pad_left)
h_ratio = new_size[0] / size[0]
w_ratio = new_size[1] / size[1]
# Resize to yield given new size
img = cv2.copyMakeBorder(img, pad_top, pad_bot, pad_left, pad_right, cv2.BORDER_CONSTANT, value=pad_val)
if return_all_infos:
return img, h_ratio, w_ratio, pad_top, pad_bot, pad_left, pad_right, new_size
return img
def get_original_coordinate(detection, img_size, pad_infos):
"""
Detection result still has padding informations to image
Need re-calculate to original coordinate info
"""
pad_top, pad_bot, pad_left, pad_right = pad_infos
xmin = detection.location_data.relative_bounding_box.xmin
ymin = detection.location_data.relative_bounding_box.ymin
width = detection.location_data.relative_bounding_box.width
height = detection.location_data.relative_bounding_box.height
xmin_ori = (xmin * img_size[0]) / (img_size[0] - pad_top - pad_bot)
ymin_ori = (ymin * img_size[1]) / (img_size[1] - pad_left - pad_right)
width_ori = (width * img_size[0]) / (img_size[0] - pad_top - pad_bot)
height_ori = (height * img_size[1]) / (img_size[1] - pad_left - pad_right)
return xmin_ori, ymin_ori, width_ori, height_ori
def get_main_sub_faces(face_detection_infos, min_face_ratio_over_max_face_to_be_main=0.7):
"""
Determine main or sub attribute for given faces according to
area ratio according to image and area ratio according to face with max area
Input:
`face_detection_infos`: face detection info from face detector model
`min_face_ratio_over_max_face_to_be_main`: area ratio according to max area face to be main
Return:
`face_is_main` (list[int]): list of is_main face, if face is main, return 1, otherwise
`face_ratios` (list[float]): list of area ratios according to image
"""
face_ratios = [detection.location_data.relative_bounding_box.width*detection.location_data.relative_bounding_box.height for detection in face_detection_infos]
face_ids = [id for id in range(len(face_detection_infos))]
face_is_main = [0 for _ in range(len(face_detection_infos))]
face_ratio_max = np.max(face_ratios)
for id in face_ids:
# if ratio according to max face ge given criteria
if face_ratios[id] / face_ratio_max >= min_face_ratio_over_max_face_to_be_main:
face_is_main[id] = 1
return face_is_main, face_ratios
def get_final_result(main_sub_faces, face_ratios,
face_detection_infos, face_direction_infos, face_position_infos, face_coordinate_infos,
face_pose_infos, face_area_infos, chosen_rectangle_pos_list,
return_all_informations=True):
"""
Get decision face for yielding final result
Input:
main_sub_faces (list[int]): list of is_main face, if face is main, return 1, otherwise
face_ratios (list[float]): list of area ratios according to image
`face_detection_infos`: face detection info from face detector model
`face_direction_infos` (list[str]): direction for each face
`face_position_infos` (list[str]): position for each face
`face_coordinates_infos`(list(list[float])): coordinates for each face (x1y1x2y2)
`face_pose_infos` (list(list[float])): pose info for each face (x,y,z)
`face area_indos` (list[float]): area ratio of face according to image
`chosen_rectangle_pos_list`(list(list[float])): chosen rectangle for each face
Return:
decision infomation of decision face
"""
face_ids = np.arange(len(main_sub_faces))
face_ids = [id for _, id in sorted(zip(face_ratios, face_ids), reverse=True)]
final_id = face_ids[0] # Return the max ratio face if do not have any `FRONT` face
for id in face_ids:
# If face is sub, then the remaining faces also sub
if main_sub_faces[id] == 0:
break
# If face is `FRONT`, do not need check the others
if face_direction_infos[id] == 'FRONT':
final_id = id
break
if return_all_informations:
return face_detection_infos[final_id], face_direction_infos[final_id], face_position_infos[final_id], \
face_coordinate_infos[final_id], \
face_pose_infos[final_id], face_area_infos[final_id], chosen_rectangle_pos_list[final_id], \
final_id
else:
return final_id
def get_cache_informations(save_dict, cache_path='cache.pkl'):
"""
if using cache is enabled, using local cache for storing predicted informations from samples
avoid time consumption
Input:
save_dict (dict): current informations
cache_path (str): path to previous informations
Return:
cache_dict (dict): merge informations from preivous cache to current one
"""
if os.path.exists(cache_path):
with open(cache_path, 'rb') as f:
cache_dict = pickle.load(f)
if cache_dict.get('img_size', None) == save_dict['img_size'] and \
cache_dict.get('position_horizontal_thresholds', None) == save_dict['position_horizontal_thresholds'] and \
cache_dict.get('position_vertical_thresholds', None) == save_dict['position_vertical_thresholds']:
return cache_dict
return {}
return {}
def parse_head_pose_informations(head_pose_info, session_state):
head_pose_info['POSITION']['TOP_LEFT'][:2] = [[0, 0],
[session_state.position_horizontal_thresholds[0], session_state.position_vertical_thresholds[0]]]
head_pose_info['POSITION']['TOP'][:2] = [[session_state.position_horizontal_thresholds[0], 0],
[session_state.position_horizontal_thresholds[1], session_state.position_vertical_thresholds[0]]]
head_pose_info['POSITION']['TOP_RIGHT'][:2] = [[session_state.position_horizontal_thresholds[1], 0],
[1, session_state.position_vertical_thresholds[0]]]
head_pose_info['POSITION']['RIGHT'][:2] = [[session_state.position_horizontal_thresholds[1], session_state.position_vertical_thresholds[0]],
[1, session_state.position_vertical_thresholds[1]]]
head_pose_info['POSITION']['BOTTOM_RIGHT'][:2] = [[session_state.position_horizontal_thresholds[1], session_state.position_vertical_thresholds[1]],
[1, 1]]
head_pose_info['POSITION']['BOTTOM'][:2] = [[session_state.position_horizontal_thresholds[0], session_state.position_vertical_thresholds[1]],
[session_state.position_horizontal_thresholds[1], 1]]
head_pose_info['POSITION']['BOTTOM_LEFT'][:2] = [[0, session_state.position_vertical_thresholds[1]],
[session_state.position_horizontal_thresholds[0], 1]]
head_pose_info['POSITION']['LEFT'][:2] = [[0, session_state.position_vertical_thresholds[0]],
[session_state.position_horizontal_thresholds[0], session_state.position_vertical_thresholds[1]]]
head_pose_info['POSITION']['MIDDLE'][:2] = [[session_state.position_horizontal_thresholds[0], session_state.position_vertical_thresholds[0]],
[session_state.position_horizontal_thresholds[1], session_state.position_vertical_thresholds[1]]]