def compute_average_precision_per_class(num_true_cases, gt_boxes,
prediction_file, iou_threshold, use_2007_metric):
""" Computes average precision per class
num_true_cases 和 gt_boxes 是从Annotations读取的信息
num_true_cases: 这个class出现了多少次
gt_boxes: gt_boxes[image_id]存着对应的gt_box
prediction_file存着模型的输出结果, 依次为 img路径, score概率, x1, y1, x2, y2
"""
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) # index从大到小排序
# boxes , image_ids 按 scores 大小排序
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()
# 得到FP, TP
for i, image_id in enumerate(image_ids):
box = boxes[i]
if image_id not in gt_boxes: # 预测有而实际没有
false_positive[i] = 1
continue
gt_box = gt_boxes[image_id] # gt_box可能有多个
ious = box_utils.iou_of(box, gt_box) # predicted boxes和每个ground truth boxes 的 IoU
max_iou = torch.max(ious).item()
max_arg = torch.argmax(ious).item() # max_arg表示和预测结果最吻合的是第几个gt_box
if max_iou > iou_threshold:
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
# 得到precision, recall, 计算AP
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:
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 -= 0.0 # 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)