def encode(self, boxes, labels, threshold=0.5):
'''
接收图片中物体所在的位置和它们对应的标签，为每一个anchor分配一个物体
'''
if len(boxes) == 0:
return (np.zeros(self.default_boxes.shape, dtype=np.float32),
np.zeros(self.default_boxes.shape[:1], dtype=np.int32))
#计算iou
iou = bbox_iou(point_form(self.default_boxes), boxes)
#anchor匹配与它iou最大的真实物体
#其实有很多anchor与所有真实物体都是iou=0,但是无所谓我们先让它们匹着
gt_idx = iou.argmax(axis=1)
iou = iou.max(axis=1)
boxes = boxes[gt_idx]
labels = labels[gt_idx]
#我们要进行回归的是anchor与真实物体的一些差距，所以先进行编码
loc = np.hstack(
(((boxes[:, :2] + boxes[:, 2:]) / 2 - self.default_boxes[:, :2]) /
(self.variance[0] * self.default_boxes[:, 2:]),
np.log(
(boxes[:, 2:] - boxes[:, :2]) / self.default_boxes[:, 2:]) /
self.variance[1]))
#这里我们将背景定义为0号标签，通过iou将那些anchor与真实物体iou小于threshold的anchor标签
#设置为0，通过这一步我们拥有了这个标签后就能知道哪些anchor是正样本
conf = 1 + labels
conf[iou < threshold] = 0
return loc.astype(np.float32), conf.astype(np.int32)
def _crop(image, boxes, labels):
height, width, _ = image.shape
if len(boxes) == 0:
return image, boxes, labels
while True:
mode = random.choice((
None,
(0.1, None),
(0.3, None),
(0.7, None),
(0.9, None),
(None, None),
))
if mode is None:
return image, boxes, labels
min_iou, max_iou = mode
if min_iou is None:
min_iou = float('-inf')
if max_iou is None:
max_iou = float('inf')
for _ in range(50):
w = random.randrange(int(0.3 * width), width)
h = random.randrange(int(0.3 * height), height)
if h / w < 0.5 or 2 < h / w:
continue
l = random.randrange(width - w)
t = random.randrange(height - h)
roi = np.array((l, t, l + w, t + h))
iou = bbox_iou(boxes, roi[np.newaxis])
if not (min_iou <= iou.min() and iou.max() <= max_iou):
continue
image = image[roi[1]:roi[3], roi[0]:roi[2]]
centers = (boxes[:, :2] + boxes[:, 2:]) / 2
mask = np.logical_and(roi[:2] < centers, centers < roi[2:]) \
.all(axis=1)
boxes = boxes[mask].copy()
labels = labels[mask]
boxes[:, :2] = np.maximum(boxes[:, :2], roi[:2])
boxes[:, :2] -= roi[:2]
boxes[:, 2:] = np.minimum(boxes[:, 2:], roi[2:])
boxes[:, 2:] -= roi[:2]
return image, boxes, labels
def encode(self, boxes, labels, threshold=0.5):
'''
Receive the position of the object in the picture and their corresponding labels, assign an object to each anchor
'''
if len(boxes) == 0:
return (
np.zeros(self.default_boxes.shape, dtype=np.float32),
np.zeros(self.default_boxes.shape[:1], dtype=np.int32))
#Calculate iou
iou = bbox_iou(point_form(self.default_boxes), boxes)
#Anchor matches it with iou's largest real object
# In fact, there are many anchors and all real objects are iou=0, but it doesn’t matter if we let them match first.
gt_idx = iou.argmax(axis=1)
def encode(self, boxes, labels, threshold=0.5):
if len(boxes) == 0:
return (np.zeros(self.default_boxes.shape, dtype=np.float32),
np.zeros(self.default_boxes.shape[:1], dtype=np.int32))
iou = bbox_iou(point_form(self.default_boxes), boxes)
gt_idx = iou.argmax(axis=1)
iou = iou.max(axis=1)
boxes = boxes[gt_idx]
labels = labels[gt_idx]
loc = np.hstack(
(((boxes[:, :2] + boxes[:, 2:]) / 2 - self.default_boxes[:, :2]) /
(self.variance[0] * self.default_boxes[:, 2:]),
np.log(
(boxes[:, 2:] - boxes[:, :2]) / self.default_boxes[:, 2:]) /
self.variance[1]))
conf = 1 + labels
conf[iou < threshold] = 0
return loc.astype(np.float32), conf.astype(np.int32)