def compute_average_precision_per_class(args, num_true_cases, gt_boxes,
difficult_cases, prediction_file,
iou_threshold, use_2007_metric):
with open(prediction_file) as f:
image_ids = []
boxes = []
scores = []
for line in f:
t = line.rstrip().split(" ")
image_ids.append(t[0])
scores.append(float(t[1]))
box = torch.tensor([float(v) for v in t[2:]]).unsqueeze(0)
box -= 1.0 # convert to python format where indexes start from 0
boxes.append(box)
scores = np.array(scores)
sorted_indexes = np.argsort(-scores)
boxes = [boxes[i] for i in sorted_indexes]
image_ids = [image_ids[i] for i in sorted_indexes]
true_positive = np.zeros(len(image_ids))
false_positive = np.zeros(len(image_ids))
matched = set()
for i, image_id in enumerate(image_ids):
box = boxes[i]
if args['flow_control']['dataset_type'] == "voc":
pass
else:
image_id = int(image_id)
if image_id not in gt_boxes.keys():
false_positive[i] = 1
continue
gt_box = gt_boxes[image_id]
ious = box_utils.iou_of(box, gt_box)
max_iou = torch.max(ious).item()
max_arg = torch.argmax(ious).item()
if max_iou > iou_threshold:
if difficult_cases[image_id][max_arg] == 0:
if (image_id, max_arg) not in matched:
true_positive[i] = 1
matched.add((image_id, max_arg))
else:
false_positive[i] = 1
else:
false_positive[i] = 1
true_positive = true_positive.cumsum()
false_positive = false_positive.cumsum()
precision = true_positive / (true_positive + false_positive + 1e-8)
recall = true_positive / num_true_cases
# import pdb;pdb.set_trace()
_vtp = true_positive[-1] if len(true_positive) != 0 else 0
_vfp = false_positive[-1] if len(false_positive) != 0 else 0
if use_2007_metric:
return measurements.compute_voc2007_average_precision(
precision,
recall), _vtp / (_vtp + _vfp + 1e-8), _vtp / num_true_cases
else:
return measurements.compute_average_precision(precision, recall)
def compute_average_precision_per_class(num_true_cases, gt_boxes,
difficult_cases, prediction_file,
iou_threshold, use_2007_metric):
"""
num_true_cases : 実際にそのクラスに属するtarget総数
"""
with open(prediction_file) as f:
image_ids = []
boxes = []
scores = []
for line in f: # そのクラス分のprediction box 全部読み込み
t = line.rstrip().split(" ") # t: [image_id, scores, boxes]
image_ids.append(t[0])
scores.append(float(t[1]))
box = torch.tensor([float(v) for v in t[2:]]).unsqueeze(0)
box -= 1.0 # convert to python format where indexes start from 0
boxes.append(box)
scores = np.array(scores)
sorted_indexes = np.argsort(
-scores
) # -付きなのでlarger is firstでindex // 全ボックス内でconfidenceの高い順にsort
boxes = [boxes[i] for i in sorted_indexes]
image_ids = [image_ids[i] for i in sorted_indexes]
true_positive = np.zeros(len(image_ids))
false_positive = np.zeros(len(image_ids))
matched = set()
for i, image_id in enumerate(
image_ids
): # image_ids : このクラスと判断されたimageのid / gt_boxes : 実際にこのクラスだったボックス
box = boxes[i]
if image_id not in gt_boxes:
false_positive[i] = 1 # このクラスの物体box，と予測して実際は違った
continue
gt_box = gt_boxes[image_id] # 多分複数ボックス（そのimage_id中のこのクラスの正解ボックス）
ious = box_utils.iou_of(box, gt_box)
max_iou = torch.max(ious).item()
max_arg = torch.argmax(ious).item()
if max_iou > iou_threshold: # 少なくともどれかのgtboxに対してIoU>しきい値であったpredictionが存在
if difficult_cases[image_id][max_arg] == 0:
if (
image_id, max_arg
) not in matched: # これまで，同じimage_id中で今見てるボックス（max_arg）が検出されていない場合(not in match)
true_positive[i] = 1 # max_argのボックス = true positive
matched.add((image_id, max_arg))
else: # 同じgt_boxが既に検出され，true positiveとして数えられている: 同じ物体を重複してpredicしている場合など．
false_positive[i] = 1
else:
false_positive[i] = 1
true_positive = true_positive.cumsum()
false_positive = false_positive.cumsum()
precision = true_positive / (true_positive + false_positive)
recall = true_positive / num_true_cases
if use_2007_metric:
return measurements.compute_voc2007_average_precision(
precision, recall)
else:
return measurements.compute_average_precision(precision, recall)
def compute_average_precision_per_class(
num_true_cases,
gt_boxes,
difficult_cases,
prediction_file,
iou_threshold,
use_2007_metric,
):
with open(str(prediction_file)) as f:
image_ids = []
boxes = []
scores = []
for line in f:
t = line.rstrip().split(" ")
image_ids.append(t[0])
scores.append(float(t[1]))
box = torch.tensor([float(v) for v in t[2:]]).unsqueeze(0)
box -= 1.0 # convert to python format where indexes start from 0
boxes.append(box)
scores = np.array(scores)
sorted_indexes = np.argsort(-scores)
boxes = [boxes[i] for i in sorted_indexes]
image_ids = [image_ids[i] for i in sorted_indexes]
true_positive = np.zeros(len(image_ids))
false_positive = np.zeros(len(image_ids))
matched = set()
for i, image_id in enumerate(image_ids):
box = boxes[i]
if image_id not in gt_boxes:
false_positive[i] = 1
continue
gt_box = gt_boxes[image_id]
ious = box_utils.iou_of(box, gt_box)
max_iou = torch.max(ious).item()
max_arg = torch.argmax(ious).item()
if max_iou > iou_threshold:
if difficult_cases[image_id][max_arg] == 0:
if (image_id, max_arg) not in matched:
true_positive[i] = 1
matched.add((image_id, max_arg))
else:
false_positive[i] = 1
else:
false_positive[i] = 1
true_positive = true_positive.cumsum()
false_positive = false_positive.cumsum()
precision = true_positive / (true_positive + false_positive)
recall = true_positive / num_true_cases
if use_2007_metric:
return measurements.compute_voc2007_average_precision(
precision, recall)
else:
return measurements.compute_average_precision(precision, recall)
def compute_average_precision_per_class(num_true_cases, gt_boxes,
difficult_cases, prediction_file,
iou_threshold, use_2007_metric):
with open(prediction_file) as f:
image_ids = []
boxes = []
scores = []
for line in f:
#ImageID,Source,LabelName,Confidence,XMin,XMax,YMin,YMax,IsOccluded,IsTruncated,IsGroupOf,IsDepiction,IsInside,ClassId,ClassName
t = line.rstrip().split(",")
image_ids.append(t[0])
scores.append(float(t[3]))
box = torch.tensor(
[float(t[4]),
float(t[6]),
float(t[5]),
float(t[7])]).unsqueeze(0)
box -= 1.0 # convert to python format where indexes start from 0
boxes.append(box)
scores = np.array(scores)
sorted_indexes = np.argsort(-scores)
boxes = [boxes[i] for i in sorted_indexes]
image_ids = [image_ids[i] for i in sorted_indexes]
true_positive = np.zeros(len(image_ids))
false_positive = np.zeros(len(image_ids))
matched = set()
for i, image_id in enumerate(image_ids):
box = boxes[i]
if image_id not in gt_boxes:
false_positive[i] = 1
continue
gt_box = gt_boxes[image_id]
ious = box_utils.iou_of(box, gt_box)
max_iou = torch.max(ious).item()
max_arg = torch.argmax(ious).item()
if max_iou > iou_threshold:
if difficult_cases[image_id][max_arg] == 0:
if (image_id, max_arg) not in matched:
true_positive[i] = 1
matched.add((image_id, max_arg))
else:
false_positive[i] = 1
else:
false_positive[i] = 1
true_positive = true_positive.cumsum()
false_positive = false_positive.cumsum()
precision = true_positive / (true_positive + false_positive)
recall = true_positive / num_true_cases
if use_2007_metric:
return measurements.compute_voc2007_average_precision(
precision, recall)
else:
return measurements.compute_average_precision(precision, recall)
