Download utils.py from lowercaseonly/cghd: direct link, hf CLI and curl.
- Browser
- Download file 5.02 kB
-
https://huggingface.co/datasets/lowercaseonly/cghd/resolve/refs%2Fpr%2F2/utils.py
- Command line
-
hf download hf://datasets/lowercaseonly/cghd@refs/pr/2/utils.py
-
curl -L -o utils.py https://huggingface.co/datasets/lowercaseonly/cghd/resolve/refs%2Fpr%2F2/utils.py
5.02 kB
| """utils.py: Helper Functions to keep this Repo Standalone""" | |
| # System Imports | |
| from math import sin, cos, radians, sqrt | |
| __author__ = "Johannes Bayer" | |
| __copyright__ = "Copyright 2023, DFKI" | |
| __license__ = "CC" | |
| __version__ = "0.0.1" | |
| __email__ = "johannes.bayer@dfki.de" | |
| __status__ = "Prototype" | |
| def shift(p, q): | |
| """Shifts a Point by another point""" | |
| return [p[0]+q[0], p[1]+q[1]] | |
| def rotate(p, angle): | |
| """Rotates a Point by an Angle""" | |
| return [p[0] * cos(angle) - p[1] * sin(angle), | |
| p[0] * sin(angle) + p[1] * cos(angle)] | |
| def scale(p, scale_x, scale_y): | |
| """Scales a Point in two Dimensions""" | |
| return [p[0]*scale_x, p[1]*scale_y] | |
| def transform(port, bb): | |
| """Transforms a Point from Unit Space (classes ports) to Global Bounding Box (image)""" | |
| p = shift(port['position'], (-.5, -0.5)) # Normalize: [0.0, 1.0]^2 -> [-0.5, 0-5]^2 | |
| p = scale(p, 1.0, -1.0) # Flip | |
| p = rotate(p, -radians(bb['rotation'])) | |
| p = scale(p, bb["xmax"] - bb["xmin"], bb["ymax"] - bb["ymin"]) | |
| p = shift(p, [(bb["xmin"]+bb["xmax"])/2, (bb["ymin"]+bb["ymax"])/2]) | |
| return {"name": port['name'], "position": p} | |
| def bbdist(bb1, bb2): | |
| """Calculates the Distance between two Bounding Box Annotations""" | |
| return sqrt(((bb1["xmin"]+bb1["xmax"])/2 - (bb2["xmin"]+bb2["xmax"])/2)**2 + | |
| ((bb1["ymin"]+bb1["ymax"])/2 - (bb2["ymin"]+bb2["ymax"])/2)**2) | |
| def overlap(bbox1, bbox2): | |
| if bbox1["xmin"] > bbox2["xmax"] or bbox1["xmax"] < bbox2["xmin"]: | |
| return False | |
| if bbox1["ymin"] > bbox2["ymax"] or bbox1["ymax"] < bbox2["ymin"]: | |
| return False | |
| return True | |
| def associated_keypoints(instances, shape): | |
| """Returns the points with same group id as the provided polygon""" | |
| return [point for point in instances["points"] | |
| if point["group"] == shape["group"] and point["class"] == "connector"] | |
| def IoU(bb1, bb2): | |
| """Intersection over Union""" | |
| intersection = 1 | |
| union = 1 | |
| return intersection/union | |
| if __name__ == "__main__": | |
| import sys | |
| from loader import read_pascal_voc, write_pascal_voc | |
| import numpy as np | |
| import random | |
| if len(sys.argv) == 3: | |
| source = sys.argv[1] | |
| target = sys.argv[2] | |
| ann1, ann2 = [[bbox for bbox in read_pascal_voc(path)['bboxes'] if bbox['class'] == "text"] | |
| for path in [source, target]] | |
| if not len(ann1) == len(ann2): | |
| print(f"Warning: Unequal Text Count ({len(ann1)} vs. {len(ann2)}), cropping..") | |
| consensus = min(len(ann1), len(ann2)) | |
| ann1 = ann1[:consensus] | |
| ann2 = ann2[:consensus] | |
| x1 = np.array([(bbox['xmin']+bbox['xmax'])/2 for bbox in ann1]) | |
| y1 = np.array([(bbox['ymin']+bbox['ymax'])/2 for bbox in ann1]) | |
| x2 = np.array([(bbox['xmin']+bbox['xmax'])/2 for bbox in ann2]) | |
| y2 = np.array([(bbox['ymin']+bbox['ymax'])/2 for bbox in ann2]) | |
| x1 = ((x1-np.min(x1))/(np.max(x1)-np.min(x1))) * (np.max(x2)-np.min(x2)) + np.min(x2) | |
| y1 = ((y1-np.min(y1))/(np.max(y1)-np.min(y1))) * (np.max(y2)-np.min(y2)) + np.min(y2) | |
| dist = np.sqrt((x1-x2[np.newaxis].T)**2 + (y1-y2[np.newaxis].T)**2) | |
| indices_1 = np.arange(len(ann1)) | |
| indices_2 = np.arange(len(ann2)) | |
| print(np.sum(np.diagonal(dist))) | |
| for i in range(10000): | |
| if random.random() > 0.5: | |
| max_dist_pos = np.argmax(np.diagonal(dist)) # Mitigate Largest Cost | |
| else: | |
| max_dist_pos = random.randint(0, len(ann1)-1) | |
| if np.min(dist[max_dist_pos, :]) < np.min(dist[:, max_dist_pos]): | |
| min_dist_pos = np.argmin(dist[max_dist_pos, :]) | |
| dist[:, [max_dist_pos, min_dist_pos]] = dist[:, [min_dist_pos, max_dist_pos]] # Swap Columns | |
| indices_1[[max_dist_pos, min_dist_pos]] = indices_1[[min_dist_pos, max_dist_pos]] | |
| else: | |
| min_dist_pos = np.argmin(dist[:, max_dist_pos]) | |
| dist[[max_dist_pos, min_dist_pos], :] = dist[[min_dist_pos, max_dist_pos], :] # Swap Rows | |
| indices_2[[max_dist_pos, min_dist_pos]] = indices_2[[min_dist_pos, max_dist_pos]] | |
| print(np.sum(np.diagonal(dist))) | |
| wb = read_pascal_voc(target) | |
| for i in range(len(ann1)): | |
| ann2[indices_2[i]]['text'] = ann1[indices_1[i]]['text'] | |
| bbox_match = [bbox for bbox in wb['bboxes'] | |
| if bbox['xmin'] == ann2[indices_2[i]]['xmin'] and | |
| bbox['xmax'] == ann2[indices_2[i]]['xmax'] and | |
| bbox['ymin'] == ann2[indices_2[i]]['ymin'] and | |
| bbox['ymax'] == ann2[indices_2[i]]['ymax']] | |
| if len(bbox_match) == 1: | |
| bbox_match[0]['text'] = ann1[indices_1[i]]['text'] | |
| bbox_match[0]['rotation'] = ann1[indices_1[i]]['rotation'] | |
| write_pascal_voc(wb) | |
| else: | |
| print("Args: source target") | |