Spaces:
Running on Zero
Running on Zero
| from __future__ import annotations | |
| import cv2 | |
| import numpy as np | |
| def create_default_face_bbox(image: np.ndarray, scale: float = 0.9) -> np.ndarray: | |
| img_h, img_w = image.shape[:2] | |
| box_size = min(img_w, img_h) * float(scale) | |
| cx = (img_w - 1) * 0.5 | |
| cy = (img_h - 1) * 0.5 | |
| half = box_size * 0.5 | |
| return np.array([cx - half, cy - half, cx + half, cy + half], dtype=np.float32) | |
| def enlarge_bbox(bbox: np.ndarray, scale: float) -> np.ndarray: | |
| bbox = np.asarray(bbox, dtype=np.float32) | |
| center = np.array([(bbox[0] + bbox[2]) * 0.5, (bbox[1] + bbox[3]) * 0.5], dtype=np.float32) | |
| size = np.array([bbox[2] - bbox[0], bbox[3] - bbox[1]], dtype=np.float32) * float(scale) | |
| half = size * 0.5 | |
| return np.array( | |
| [center[0] - half[0], center[1] - half[1], center[0] + half[0], center[1] + half[1]], | |
| dtype=np.float32, | |
| ) | |
| def make_bbox_square(bbox: np.ndarray) -> np.ndarray: | |
| x0, y0, x1, y1 = np.asarray(bbox, dtype=np.float32) | |
| size = int(round(max(x1 - x0, y1 - y0))) | |
| cx = (x0 + x1) * 0.5 | |
| cy = (y0 + y1) * 0.5 | |
| new_x0 = int(round(cx - size * 0.5)) | |
| new_y0 = int(round(cy - size * 0.5)) | |
| new_x1 = new_x0 + size | |
| new_y1 = new_y0 + size | |
| return np.array([new_x0, new_y0, new_x1, new_y1], dtype=np.int32) | |
| def calc_crop_pad_info( | |
| bbox: np.ndarray, | |
| img_width: int, | |
| img_height: int, | |
| ) -> tuple[tuple[int, int, int, int], tuple[int, int, int, int]]: | |
| x0, y0, x1, y1 = np.asarray(bbox, dtype=np.int32) | |
| pad_left = max(0, -x0) | |
| pad_top = max(0, -y0) | |
| pad_right = max(0, x1 - img_width) | |
| pad_bottom = max(0, y1 - img_height) | |
| crop_left = max(0, x0) | |
| crop_top = max(0, y0) | |
| crop_right = min(img_width, x1) | |
| crop_bottom = min(img_height, y1) | |
| return (pad_left, pad_top, pad_right, pad_bottom), (crop_left, crop_top, crop_right, crop_bottom) | |
| def crop_and_resize_image( | |
| image: np.ndarray, | |
| bbox: np.ndarray, | |
| target_size: int, | |
| ) -> tuple[np.ndarray, np.ndarray]: | |
| img_h, img_w = image.shape[:2] | |
| bbox = np.asarray(bbox, dtype=np.float32).copy() | |
| pad_info, crop_info = calc_crop_pad_info(bbox, img_w, img_h) | |
| bbox_w = int(round(float(bbox[2] - bbox[0]))) | |
| bbox_h = int(round(float(bbox[3] - bbox[1]))) | |
| x0, y0, x1, y1 = crop_info | |
| cropped = image[y0:y1, x0:x1] | |
| if sum(pad_info) > 0: | |
| padded = np.zeros((bbox_h, bbox_w, 3), dtype=image.dtype) | |
| padded[pad_info[1]:(bbox_h - pad_info[3]), pad_info[0]:(bbox_w - pad_info[2])] = cropped | |
| cropped = padded | |
| resized = cv2.resize(cropped, (int(target_size), int(target_size)), interpolation=cv2.INTER_LINEAR) | |
| return resized, bbox | |