小目标检测 数据增强
1.基于copy_pasted策略的Data_Augmentation
此部分学习了Github Data_Augmentation_Zoo_for_Object_Detection的心得，相关paper见Augmentation for small object detection
代码的思路很简单：
- 首先获取各object标签值；通过标签值判断是否是小目标；
- 如果是小目标，在图片中随机找一个相同大小的空间（该空间与object没有overlap），将该小目标空间的像素复制过去。
- 实现中还加入了几个随机性，这里不再细述。
将Github Data_Augmentation_Zoo_for_Object_Detection
中此部分的代码整理了下，并用真实例子进行了验证；
1.1 代码
1.1.1 加载相关库及设置变量、参数
import cv2
import numpy as np
import random
import matplotlib.pyplot as plt
colors = [(39, 129, 113), (164, 80, 133), (83, 122, 114), ]
""" SMALL OBJECT AUGMENTATION """
SMALL_OBJECT_AUGMENTATION = True
SOA_THRESH = 2000000 # 64*64，根据个人数据集情况调节
SOA_PROB = 1
SOA_COPY_TIMES = 3
SOA_EPOCHS = 30
SOA_ONE_OBJECT = False
SOA_ALL_OBJECTS = False
1.1.2 图片及bbox的显示
def bbox_to_rect(bbox, color):
return plt.Rectangle(xy=(bbox[0], bbox[1]), width=bbox[2]-bbox[0], height=bbox[3]-bbox[1],
fill=False, edgecolor=color, linewidth=2)
def easy_visualization(sample):
image, annots = sample['img'], sample['annot']
fig = plt.imshow(image)
for i in range(len(annots)):
annot = [int(x) for x in annots[i]]
label = annot[4]
color = [c/255.0 for c in colors[label]]
rect = bbox_to_rect(annot, color)
fig.axes.add_patch(rect)
plt.show()
1.1.3 实现SmallObject_Augmentation核心部分
class SmallObjectAugmentation(object):
def __init__(self, thresh=64*64, prob=0.5, copy_times=3, epochs=30, all_objects=False, one_object=False):
"""
sample = {'img':img, 'annot':annots}
img = [height, width, 3]
annot = [xmin, ymin, xmax, ymax, label]
thresh：the detection threshold of the small object. If annot_h * annot_w < thresh, the object is small
prob: the prob to do small object augmentation
epochs: the epochs to do
"""
self.thresh = thresh
self.prob = prob
self.copy_times = copy_times
self.epochs = epochs
self.all_objects = all_objects
self.one_object = one_object
if self.all_objects or self.one_object:
self.copy_times = 1
def issmallobject(self, h, w):
if h * w <= self.thresh:
return True
else:
return False
def compute_overlap(self, annot_a, annot_b):
if annot_a is None:
return False
left_max = max(annot_a[0], annot_b[0])
top_max = max(annot_a[1], annot_b[1])
right_min = min(annot_a[2], annot_b[2])
bottom_min = min(annot_a[3], annot_b[3])
inter = max(0, (right_min-left_max)) * max(0, (bottom_min-top_max))
if inter != 0:
return True
else:
return False
def donot_overlap(self, new_annot, annots):
for annot in annots:
if self.compute_overlap(new_annot, annot):
return False
return True
def create_copy_annot(self, h, w, annot, annots):
annot = annot.astype(np.int)
annot_h, annot_w = annot[3] - annot[1], annot[2] - annot[0]
for epoch in range(self.epochs):
random_x, random_y = np.random.randint(int(annot_w / 2), int(w - annot_w / 2)), \
np.random.randint(int(annot_h / 2), int(h - annot_h / 2))
xmin, ymin = random_x - annot_w / 2, random_y - annot_h / 2
xmax, ymax = xmin + annot_w, ymin + annot_h
if xmin < 0 or xmax > w or ymin < 0 or ymax > h:
continue
new_annot = np.array([xmin, ymin, xmax, ymax, annot[4]], dtype=int)
print("new_annot:", new_annot)
if self.donot_overlap(new_annot, annots) is False:
continue
return new_annot
return None
def add_patch_in_img(self, annot, copy_annot, image):
copy_annot = copy_annot.astype(np.int)
image[annot[1]:annot[3], annot[0]:annot[2], :] = image[copy_annot[1]:copy_annot[3], copy_annot[0]:copy_annot[2], :]
return image
def __call__(self, sample):
if self.all_objects and self.one_object:
return sample
if np.random.rand() > self.prob:
return sample
img, annots = sample['img'], sample['annot']
h, w = img.shape[0], img.shape[1]
small_object_list = list()
for idx in range(annots.shape[0]):
annot = annots[idx]
annot_h, annot_w = annot[2]-annot[0], annot[3]-annot[1]
if self.issmallobject(annot_h, annot_w):
small_object_list.append(idx)
num = len(small_object_list)
# No Small Object
if num == 0:
return sample
# Refine the copy_object by the given policy
# Policy 2:
copy_object_num = 1 if num == 1 else np.random.randint(num)
# Policy 3:
if self.all_objects or self.one_object:
copy_object_num = num
random_list = random.sample(range(num), copy_object_num)
annot_idx_of_small_object = [
small_object_list[idx] for idx in random_list]
select_annots = annots[annot_idx_of_small_object, :]
annots = annots.tolist()
for idx in range(copy_object_num):
annot = select_annots[idx]
# annot_w, annot_h = annot[3] - annot[1], annot[2] - annot[0]
# if self.issmallobject(annot_h, annot_w) is False: continue
for i in range(self.copy_times):
new_annot = self.create_copy_annot(h, w, annot, annots,)
if new_annot is not None:
img = self.add_patch_in_img(new_annot, annot, img)
annots.append(new_annot)
return {
'img': img, 'annot': np.array(annots)}
1.1.4 用真实图片调用程序查看效果
if __name__ == '__main__':
img = cv2.imread(r"D://1.jpg", -1)
Boxes = np.array([[196, 175, 387, 294, 0], [124, 298, 222, 399, 0], [
334, 298, 434, 390, 0]], dtype=int)
transform = SmallObjectAugmentation(
SOA_THRESH, SOA_PROB, SOA_COPY_TIMES, SOA_EPOCHS, SOA_ALL_OBJECTS, SOA_ONE_OBJECT)
sample = {
'img': img, 'annot': Boxes}
sample = transform(sample)
easy_visualization(sample)
cv2.imwrite("D:\\2.jpg", img)
这里对图片D://1.jpg中的小鸟进行 Data_Augmentation，处理前后效果如下：
非常nice!
注意程序中有多个random相关的语句，输出结果每次都不一样！
另外，此程序的输入Boxes 不是我们常使用的VOC 相对坐标（除以高/宽的相对样式），在目标检测中要进行一些修改。可参考本人上传资源在目标检测模型中直接调用。
参考文献
[1] https://github.com/zzl-pointcloud/Data_Augmentation_Zoo_for_Object_Detection
[2] Perper: Augmentation for small object detection