def _evaluate_single(self, detections_dict, groundtruth_list):
groundtruth_dict = {
'annotations': groundtruth_list,
'images': [{
'id': gt['image_id']
} for gt in groundtruth_list],
'categories': self.categories
}
if len(groundtruth_list) > 0:
detections_list = coco_tools.ExportSingleImageDetectionBoxesToCoco(
image_id=groundtruth_list[0]['image_id'],
category_id_set=set([c['id'] for c in self.categories]),
detection_boxes=detections_dict['detection_boxes'],
detection_scores=detections_dict['detection_scores'],
detection_classes=detections_dict['detection_classes'])
else:
detections_list = []
# The COCO evaluation prints some information, which we don't care about
with HiddenPrints():
groundtruth = coco_tools.COCOWrapper(groundtruth_dict)
detections = groundtruth.LoadAnnotations(detections_list)
evaluator = coco_tools.COCOEvalWrapper(groundtruth,
detections,
iou_type='bbox')
summary_metrics, _ = evaluator.ComputeMetrics()
return summary_metrics['Precision/mAP']
def evaluate(self):
"""Evaluates the detection boxes and returns a dictionary of coco metrics.
Returns:
A dictionary holding -
1. summary_metrics:
'DetectionBoxes_Precision/mAP': mean average precision over classes
averaged over IOU thresholds ranging from .5 to .95 with .05
increments.
'DetectionBoxes_Precision/[email protected]': mean average precision at 50% IOU
'DetectionBoxes_Precision/[email protected]': mean average precision at 75% IOU
'DetectionBoxes_Precision/mAP (small)': mean average precision for small
objects (area < 32^2 pixels).
'DetectionBoxes_Precision/mAP (medium)': mean average precision for
medium sized objects (32^2 pixels < area < 96^2 pixels).
'DetectionBoxes_Precision/mAP (large)': mean average precision for large
objects (96^2 pixels < area < 10000^2 pixels).
'DetectionBoxes_Recall/AR@1': average recall with 1 detection.
'DetectionBoxes_Recall/AR@10': average recall with 10 detections.
'DetectionBoxes_Recall/AR@100': average recall with 100 detections.
'DetectionBoxes_Recall/AR@100 (small)': average recall for small objects
with 100.
'DetectionBoxes_Recall/AR@100 (medium)': average recall for medium objects
with 100.
'DetectionBoxes_Recall/AR@100 (large)': average recall for large objects
with 100 detections.
2. per_category_ap: if include_metrics_per_category is True, category
specific results with keys of the form:
'Precision mAP ByCategory/category' (without the supercategory part if
no supercategories exist). For backward compatibility
'PerformanceByCategory' is included in the output regardless of
all_metrics_per_category.
"""
tf.logging.info('Performing evaluation on %d images.',
len(self._image_ids))
groundtruth_dict = {
'annotations': self._groundtruth_list,
'images': [{
'id': image_id
} for image_id in self._image_ids],
'categories': self._categories
}
coco_wrapped_groundtruth = coco_tools.COCOWrapper(groundtruth_dict)
coco_wrapped_detections = coco_wrapped_groundtruth.LoadAnnotations(
self._detection_boxes_list)
box_evaluator = coco_tools.COCOEvalWrapper(coco_wrapped_groundtruth,
coco_wrapped_detections,
agnostic_mode=False)
box_metrics, box_per_category_ap = box_evaluator.ComputeMetrics(
include_metrics_per_category=self._include_metrics_per_category,
all_metrics_per_category=self._all_metrics_per_category)
box_metrics.update(box_per_category_ap)
box_metrics = {
'DetectionBoxes_' + key: value
for key, value in iter(box_metrics.items())
}
return box_metrics
def evaluate(self):
"""Evaluates the detection masks and returns a dictionary of coco metrics.
Returns:
A dictionary holding -
1. summary_metrics:
'DetectionMasks_Precision/mAP': mean average precision over classes
averaged over IOU thresholds ranging from .5 to .95 with .05 increments.
'DetectionMasks_Precision/[email protected]': mean average precision at 50% IOU.
'DetectionMasks_Precision/[email protected]': mean average precision at 75% IOU.
'DetectionMasks_Precision/mAP (small)': mean average precision for small
objects (area < 32^2 pixels).
'DetectionMasks_Precision/mAP (medium)': mean average precision for medium
sized objects (32^2 pixels < area < 96^2 pixels).
'DetectionMasks_Precision/mAP (large)': mean average precision for large
objects (96^2 pixels < area < 10000^2 pixels).
'DetectionMasks_Recall/AR@1': average recall with 1 detection.
'DetectionMasks_Recall/AR@10': average recall with 10 detections.
'DetectionMasks_Recall/AR@100': average recall with 100 detections.
'DetectionMasks_Recall/AR@100 (small)': average recall for small objects
with 100 detections.
'DetectionMasks_Recall/AR@100 (medium)': average recall for medium objects
with 100 detections.
'DetectionMasks_Recall/AR@100 (large)': average recall for large objects
with 100 detections.
2. per_category_ap: if include_metrics_per_category is True, category
specific results with keys of the form:
'Precision mAP ByCategory/category' (without the supercategory part if
no supercategories exist). For backward compatibility
'PerformanceByCategory' is included in the output regardless of
all_metrics_per_category.
"""
groundtruth_dict = {
'annotations':
self._groundtruth_list,
'images': [{
'id': image_id,
'height': shape[1],
'width': shape[2]
} for image_id, shape in self._image_id_to_mask_shape_map.items()],
'categories':
self._categories
}
coco_wrapped_groundtruth = coco_tools.COCOWrapper(
groundtruth_dict, detection_type='segmentation')
coco_wrapped_detection_masks = coco_wrapped_groundtruth.LoadAnnotations(
self._detection_masks_list)
mask_evaluator = coco_tools.COCOEvalWrapper(
coco_wrapped_groundtruth,
coco_wrapped_detection_masks,
agnostic_mode=False,
iou_type='segm')
mask_metrics, mask_per_category_ap = mask_evaluator.ComputeMetrics(
include_metrics_per_category=self._include_metrics_per_category)
mask_metrics.update(mask_per_category_ap)
mask_metrics = {
'DetectionMasks_' + key: value
for key, value in mask_metrics.items()
}
return mask_metrics
def evaluate(self):
"""Evaluates the detection boxes and returns a dictionary of coco metrics.
Returns:
A dictionary holding -
1. summary_metrics:
'DetectionBoxes_Precision/mAP': mean average precision over classes
averaged over IOU thresholds ranging from .5 to .95 with .05
increments.
'DetectionBoxes_Precision/[email protected]': mean average precision at 50% IOU
'DetectionBoxes_Precision/[email protected]': mean average precision at 75% IOU
'DetectionBoxes_Precision/mAP (small)': mean average precision for small
objects (area < 32^2 pixels).
'DetectionBoxes_Precision/mAP (medium)': mean average precision for
medium sized objects (32^2 pixels < area < 96^2 pixels).
'DetectionBoxes_Precision/mAP (large)': mean average precision for large
objects (96^2 pixels < area < 10000^2 pixels).
'DetectionBoxes_Recall/AR@1': average recall with 1 detection.
'DetectionBoxes_Recall/AR@10': average recall with 10 detections.
'DetectionBoxes_Recall/AR@100': average recall with 100 detections.
'DetectionBoxes_Recall/AR@100 (small)': average recall for small objects
with 100.
'DetectionBoxes_Recall/AR@100 (medium)': average recall for medium objects
with 100.
'DetectionBoxes_Recall/AR@100 (large)': average recall for large objects
with 100 detections.
2. per_category_ap: if include_metrics_per_category is True, category
specific results with keys of the form:
'Precision mAP ByCategory/category' (without the supercategory part if
no supercategories exist). For backward compatibility
'PerformanceByCategory' is included in the output regardless of
all_metrics_per_category.
"""
groundtruth_dict = {
'annotations': self._groundtruth_list,
'images': [{
'id': image_id
} for image_id in self._image_ids],
'categories': self._categories
}
coco_wrapped_groundtruth = coco_tools.COCOWrapper(groundtruth_dict)
coco_wrapped_detections = coco_wrapped_groundtruth.LoadAnnotations(
self._detection_boxes_list)
box_evaluator = coco_tools.COCOEvalWrapper(
coco_wrapped_groundtruth,
coco_wrapped_detections,
agnostic_mode=False,
iouThresholds=self.iouThresholds)
box_metrics, box_per_category_ap = box_evaluator.ComputeMetrics(
include_metrics_per_category=self._include_metrics_per_category,
all_metrics_per_category=self._all_metrics_per_category)
box_metrics.update(box_per_category_ap)
box_metrics = {
'DetectionBoxes_' + key: value
for key, value in iter(box_metrics.items())
}
print(box_metrics)
y_true = {
image_id: []
for image_id in
[label['image_id'] for label in self._groundtruth_list]
}
for label in self._groundtruth_list:
y_true[label['image_id']].append(label['category_id'])
image_ids = np.array(sorted(y_true.keys()))
y_true = [y_true[image_id] for image_id in image_ids]
labelbinarizer = MultiLabelBinarizer()
y_true = labelbinarizer.fit_transform(y_true)
y_pred = np.zeros(y_true.shape)
for row in self._detection_boxes_list:
# find image index in groundtruth
i = np.where(image_ids == row['image_id'])[0]
j = np.where(
np.array(labelbinarizer.classes_) == row['category_id'])[0]
if row['score'] > y_pred[i, j]:
y_pred[i, j] = row['score']
roc_auc = roc_auc_score(y_true, y_pred, average='macro')
print('ROC AUC (macro): ', roc_auc)
class_names = [cat['name'] for cat in self._categories]
print(class_names)
# threshold predictions
y_pred = (y_pred >= 0.5).astype(np.int)
class_report = classification_report(y_true,
y_pred,
target_names=class_names)
print(class_report)
acc_scores = {}
for i, label in enumerate(class_names):
acc_scores['ImageLevel/BALANCED_ACC_{}'.format(
label)] = balanced_accuracy_score(y_true[:, i], y_pred[:, i])
acc_scores['ImageLevel/ACC_{}'.format(label)] = accuracy_score(
y_true[:, i], y_pred[:, i])
print(acc_scores)
box_metrics.update(acc_scores)
box_metrics['ImageLevel/ROC_AUC_MACRO'] = roc_auc
return box_metrics
def testCocoWrappers(self):
groundtruth = coco_tools.COCOWrapper(self._groundtruth_dict)
detections = groundtruth.LoadAnnotations(self._detections_list)
evaluator = coco_tools.COCOEvalWrapper(groundtruth, detections)
summary_metrics, _ = evaluator.ComputeMetrics()
self.assertAlmostEqual(1.0, summary_metrics['Precision/mAP'])
ground_truth_box = []
ground_truth_class = []
if len(data) == 0:
ground_truth_box.append([])
for i in range(len(data)):
ground_truth_box.append(data[i][1:])
ground_truth_class.append(label_map_dict[data[i][0]])
ground_truth_boxes.append(np.array(ground_truth_box, np.float))
ground_truth_classes.append(np.array(ground_truth_class, np.int32))
return image_ids1, ground_truth_boxes, ground_truth_classes
pred_image_ids, detection_boxes, detection_scores, detection_classes = convert_OBJ_DETE_dt(
pred_path)
truth_image_ids, ground_truth_boxes, ground_truth_classes = convert_OBJ_DETE_gt(
truth_path)
detections_list = coco_tools.ExportDetectionsToCOCO(pred_image_ids,
detection_boxes,
detection_scores,
detection_classes,
label_map_dict)
groundtruth_dict = coco_tools.ExportGroundtruthToCOCO(truth_image_ids,
ground_truth_boxes,
ground_truth_classes,
label_map_dict)
groundtruth = coco_tools.COCOWrapper(groundtruth_dict)
detections = groundtruth.LoadAnnotations(detections_list)
evaluator = coco_tools.COCOEvalWrapper(groundtruth, detections)
metrics, ap = evaluator.ComputeMetrics()
def evaluate(self):
groundtruth = coco_tools.COCOWrapper(self.groundtruth)
detections = groundtruth.LoadAnnotations(self.detections)
evaluator = coco_tools.COCOEvalWrapper(groundtruth, detections)
metrics, _ = evaluator.ComputeMetrics()
return metrics
def main(_):
model_path = FLAGS.model_path
voc_path = FLAGS.voc_path
labelmap_file = FLAGS.labelmap_file
set_file = FLAGS.set_file
scaler_file = FLAGS.scaler_file
side_input = FLAGS.side_input
# Load Model and read label_map.pbtxt
model = file_util.load_model(model_path)
categories, labelmap_dict, category_index = file_util.load_labelmap(
voc_path, labelmap_file)
# Get information from groundtruth and detection
gt_ids, gt_boxes, gt_classes = _read_annotations_for_groundtruth(
voc_path, set_file, labelmap_dict)
dt_ids, dt_boxes, dt_classes, dt_scores, time_per_image = inference(
model, voc_path, model_path, set_file, category_index, scaler_file,
side_input)
# COCO Evaluation
groundtruth_dict = coco_tools.ExportGroundtruthToCOCO(
gt_ids, gt_boxes, gt_classes, categories)
detections_list = coco_tools.ExportDetectionsToCOCO(
dt_ids, dt_boxes, dt_scores, dt_classes, categories)
groundtruth = coco_tools.COCOWrapper(groundtruth_dict)
detections = groundtruth.LoadAnnotations(detections_list)
evaluator = coco_tools.COCOEvalWrapper(groundtruth,
detections,
agnostic_mode=False)
summary_metrics, per_category_ap = evaluator.ComputeMetrics(
include_metrics_per_category=True, all_metrics_per_category=True)
# Convert to Percent format
for k, v in summary_metrics.items():
summary_metrics[k] = v * 100
for k, v in per_category_ap.items():
per_category_ap[k] = v * 100
print(summary_metrics)
print(per_category_ap)
# Prevention for Tensorflow Bug: Cant calculate Flops with custom Inputs
if not side_input:
flops = get_flops(model)
else:
flops = 0
metrics_dict = {
'flops': flops / 1e9,
'time_per_image': time_per_image * 1000
}
# Read Trainable Params and Name Dictionary from Pickle file
name_params_dict = pickle.load(
open(os.path.join(model_path, 'metrics', 'name_params.pkl'), 'rb'))
metrics_dict.update(name_params_dict)
metrics_dict.update(summary_metrics)
metrics_dict.update(per_category_ap)
# Save Metrics to CSV
metrics_df = pd.DataFrame.from_records([metrics_dict])
metrics_df.to_csv(os.path.join(model_path, 'metrics', 'metrics.csv'))