def find_result_ground_truth_correspondence(gt_list, result_rect_list, area_threshold=0.7):
# The json file records the overlap areas.
overlap_json = GeoJsonWriter.init_output_json_structure()
# The json file records the rectangles which we don't consider them as correct ones.
non_overlap_json = GeoJsonWriter.init_output_json_structure()
correct_results = set()
matching_lists = [] # Records the result matching to which ground truths.
# Find the correct and incorrect bounding boxes.
for k in range(0, len(result_rect_list)):
rect = result_rect_list[k]
result_area = rect.get_area()
matching_list = []
for i in range(0, len(gt_list), 1):
gt_rect_list = gt_list[i]['rects']
for j in range(0, len(gt_rect_list), 1):
gt_rect = gt_rect_list[j]
gt_area = gt_rect.get_area()
overlap_polygon = rect.get_overlap_polygon(gt_rect)
if overlap_polygon is None:
pass
else:
overlap_area = overlap_polygon.get_area()
overlap_valid = False
if overlap_area > area_threshold * gt_area or overlap_area > area_threshold * result_area:
overlap_valid = True
rect.positive = True
gt_list[i]['found'][j] = True
matching_list.append((i, j))
gt_list[i]['matching'][j].append(k)
correct_results.add(rect)
new_feature = GeoJsonWriter.generate_overlap_feature(overlap_polygon, overlap_valid, tag=i)
overlap_json['features'].append(new_feature)
matching_lists.append(matching_list)
# Record the incorrect rectangles.
for i in range(0, len(result_rect_list)):
rect = result_rect_list[i]
if not rect.positive:
non_overlap_json['features'].append(GeoJsonWriter.generate_overlap_feature(Polygon(rect.points), False, i))
return matching_lists, correct_results, overlap_json, non_overlap_json
def evaluation_simple(result_rect_list, gt_list, area_threshold=0.7, group_para=0):
# The json file records the overlap areas.
overlap_json = GeoJsonWriter.init_output_json_structure()
# The json file records the rectangles which we don't consider them as correct ones.
non_overlap_json = GeoJsonWriter.init_output_json_structure()
correct_results = set()
# Find the correct and incorrect bounding boxes.
for k in range(0, len(result_rect_list)):
rect = result_rect_list[k]
result_area = rect.get_area()
gt_list_backup = copy.copy(gt_list)
result_rect_list_backup = copy.copy(result_rect_list)
is_invalid = False
for i in range(0, len(gt_list), 1):
gt_rect_list = gt_list[i]['rects']
for j in range(0, len(gt_rect_list)):
gt_rect = gt_rect_list[j]
gt_area = gt_rect.get_area()
overlap_polygon = rect.get_overlap_polygon(gt_rect)
if overlap_polygon is None:
pass
else:
overlap_area = overlap_polygon.get_area()
overlap_valid = False
if overlap_area >= area_threshold * gt_area or overlap_area >= area_threshold * result_area:
if result_area > 5 * gt_area:
is_invalid = False
break
# pass
overlap_valid = True
rect.positive = True
gt_list[i]['found'][j] = True
correct_results.add(rect)
gt_list[i]['rects'][j].text = gt_list[i]['rects'][j].text.replace(".", "")
result_rect_list[k].text = result_rect_list[k].text.replace(".", "")
ret, gt_mlist, result_mlist = WordEvaluation.levenshtein_distance(gt_list[i]['rects'][j].text,
result_rect_list[k].text)
for m in range(0, len(gt_mlist)):
gt_list[i]['rects'][j].textFound[m] = gt_list[i]['rects'][j].textFound[m] or gt_mlist[m]
for r in range(0, len(result_mlist)):
rect.textFound[r] = rect.textFound[r] or result_mlist[r]
rect.textGroup[r].append(int(gt_list[i]['rects'][j].groupId))
new_feature = GeoJsonWriter.generate_overlap_feature(overlap_polygon, overlap_valid, tag=i)
overlap_json['features'].append(new_feature)
rect.groundTruth.append(gt_rect.text)
if is_invalid:
print('Invalid roll back.')
gt_list = gt_list_backup
result_rect_list = result_rect_list_backup
break
# Compute the precision and recall
precision_list = []
recall_list = []
total_extracted_characters_list = []
total_gt_characters_list = []
total_gt_characters = total_extracted_characters = 0
correct_result_characters = correct_gt_characters = 0
for r in result_rect_list:
# print(r.text)
for i in range(0, len(r.text)):
if group_para in r.textGroup[i] or group_para == 0:
if r.textFound[i]:
# print(r.text[i])
correct_result_characters += 1
total_extracted_characters += 1
elif len(r.textGroup[i]) == 0:
total_extracted_characters += 1
# print("\nGround truth")
for i in range(0, len(gt_list), 1):
gt_rect_list = gt_list[i]['rects']
for j in range(0, len(gt_rect_list), 1):
if group_para == int(gt_list[i]['rects'][j].groupId) or group_para == 0:
gt_rect = gt_rect_list[j]
# print(gt_rect.text)
for k in gt_rect.textFound:
if k:
correct_gt_characters += 1
total_gt_characters += 1
recall = 0 if total_gt_characters == 0 else (float(correct_gt_characters) / total_gt_characters)
precision = 0 if total_extracted_characters == 0 else (float(correct_result_characters) / total_extracted_characters)
print("Detected Bounding Boxes: %d" % len(result_rect_list))
print("Correct Bounding Boxes: %d" % len(correct_results))
print("Number of ground truth characters: %d" % total_gt_characters)
print("Number of detected characters: %d" % total_extracted_characters)
print("Number of correct detetected char: %d" % correct_result_characters)
print("Text Recognition Precision: %f" % precision)
print("Text Recognition Recall: %f" % recall)
precision_list.append(precision)
recall_list.append(recall)
total_extracted_characters_list.append(total_extracted_characters)
total_gt_characters_list.append(total_gt_characters)
return overlap_json, non_overlap_json, precision_list, recall_list, total_extracted_characters_list, total_gt_characters_list