def evaluate():
predicted_path = config.path_to_predicted_bb
ground_truth_path = config.path_to_ground_truth_bb
bounding_box = BoundingBox()
evaluator = evaloator.Evaluator()
TP, FP = 0, 0
all_GT = 0
true_positives, false_positives = [], []
precision, recall = [], []
for img_name in os.listdir(f'{predicted_path}'):
print(
f' {bcolors.BLUE} Starting with image {img_name}... {bcolors.ENDC}'
)
predicted_img = cv2.imread(f'{predicted_path}/{img_name}')
ground_truth_img = cv2.imread(f'{ground_truth_path}/{img_name}')
p_b = bounding_box.get_bounding_boxes(predicted_img)
gt_b = bounding_box.get_bounding_boxes(ground_truth_img)
gt_matched = {k: False for k, v in enumerate(gt_b)}
all_GT += len(gt_matched)
for i in range(len(p_b)):
iou_max = evaluator.get_float_min()
gt_max = -1
for j in range(len(gt_b)):
iou = evaluator.iou(p_b[i], gt_b[j])
print("IOU: ", iou)
if iou > iou_max:
iou_max = iou
gt_max = j
if iou_max >= config.iou_threshold:
if not gt_matched[gt_max]:
TP += 1
gt_matched[gt_max] = True
else:
FP += 1
else:
FP += 1
true_positives.append(TP)
false_positives.append(FP)
print("TP: ", TP)
print("FP: ", FP)
print("GT: ", all_GT)
pr = evaluator.precision(TP, FP)
re = evaluator.recall(TP, all_GT)
print(
f"Precision: {pr}, Recall: {re}, F1-score: {2 * ((pr * re) / (pr + re))}"
)
for i in range(len(true_positives)):
if (true_positives[i] + false_positives[i]) != 0:
precision.append(true_positives[i] /
(true_positives[i] + false_positives[i]))
else:
precision.append(0)
recall.append(true_positives[i] / all_GT)
""" plot precision_recall curve """
