yolov5 优化——mosaic相关

概述
Mosaic 利用了四张图片重新排布成一张图片，根据论文所说其拥有一个巨大的优点是丰富检测物体的背景：随机缩放增加了很多小目标，让网络的鲁棒性更好；且在BN计算的时候一下子会计算四张图片的数据！但是，同时：如果我们的数据集本身就有很多的小目标，那么Mosaic数据增强会导致本来较小的目标变得更小，导致模型的泛化能力变差。如果我们的类目的关键信息是框的某个局部（某些场景你又不能标局部）那么裁切可能会将这个局部信息裁切没，导致模型无法拟合真正的信息。

实现：
原理和实现可以参考这篇链接睿智的目标检测28——YoloV4当中的Mosaic数据增强方法_Bubbliiiing的博客-CSDN博客_mosaic方法，里面主要是yoloV4 的。但是，yolov5 并没有按部就班、原封不动的使用这种实现。它的实现伪代码是（本文针对3.1版本，看了6.1虽然代码结构变了，实现的核心代码并没有改变）：

1、针对当前一张图，从该epoch内已训练过的图中随机抽取3张图。同时生成一张4倍模型输入大小的背景图（用于将后面四张图贴上去）

2、针对当前图进行等比缩放（最长边等于模型预设输入尺寸，如640）。将当前图随机上下左右平移（+— 1/2模型输入图尺寸，其中1/2居然是写死的，且牵扯到了后面其他函数）。

3、将随机抽取的3张图也进行等比缩放（最长边等于模型预设输入尺寸，如640），然后根据当前图的位置，进行依次放在左边，下边、右下。

至此mosaic 完成，然后通过对生成的4倍大小的图整体进行二次的HSV、旋转、透视变换（这里缩放回原来的预设模型尺寸）、mixup等操作。

可以看出来生成的图已经比较不鲁棒了，这里感觉到了学术闭集刷性能（确实精召都提高了），和工业界大多开集上，这种训练出来的模型，fpr很难压下去（明显模型已经学错了）。所以工业界还是慎用，或者自己按需修改mosaic的超参或者实现。

 

 

 

涉及到的源码如下：

1、配置文件

yolov5/hyp.finetune.yaml at v3.1 · ultralytics/yolov5 · GitHub

 

 

 

2、源码这两个部分也需要按需修改强度

https://github.com/ultralytics/yolov5/blob/v3.1/utils/datasets.py

 

 

 

 

 

源码实现mosaic的主代码如下

def load_mosaic(self, index):
# loads images in a mosaic

labels4 = []
s = self.img_size
yc, xc = [int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border] # mosaic center x, y
indices = [index] + [random.randint(0, len(self.labels) - 1) for _ in range(3)] # 3 additional image indices
for i, index in enumerate(indices):
# Load image
img, _, (h, w) = load_image(self, index)

# place img in img4
if i == 0: # top left
img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles
x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc # xmin, ymin, xmax, ymax (large image)
x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h # xmin, ymin, xmax, ymax (small image)
elif i == 1: # top right
x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
elif i == 2: # bottom left
x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
elif i == 3: # bottom right
x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)

img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b] # img4[ymin:ymax, xmin:xmax]
padw = x1a - x1b
padh = y1a - y1b

# Labels
x = self.labels[index]
labels = x.copy()
if x.size > 0: # Normalized xywh to pixel xyxy format
labels[:, 1] = w * (x[:, 1] - x[:, 3] / 2) + padw
labels[:, 2] = h * (x[:, 2] - x[:, 4] / 2) + padh
labels[:, 3] = w * (x[:, 1] + x[:, 3] / 2) + padw
labels[:, 4] = h * (x[:, 2] + x[:, 4] / 2) + padh
labels4.append(labels)
# print('img4', img4.shape)
# Concat/clip labels
if len(labels4):
labels4 = np.concatenate(labels4, 0)
np.clip(labels4[:, 1:], 0, 2 * s, out=labels4[:, 1:]) # use with random_perspective
# img4, labels4 = replicate(img4, labels4) # replicate

# Augment
img4, labels4 = random_perspective(img4, labels4,
degrees=self.hyp['degrees'],
translate=self.hyp['translate'],
scale=self.hyp['scale'],
shear=self.hyp['shear'],
perspective=self.hyp['perspective'],
border=self.mosaic_border) # border to remove

return img4, labels4

def load_image(self, index):
# loads 1 image from dataset, returns img, original hw, resized hw
img = self.imgs[index]
if img is None: # not cached
path = self.img_files[index]
img = cv2.imread(path) # BGR
assert img is not None, 'Image Not Found ' + path
h0, w0 = img.shape[:2] # orig hw
r = self.img_size / max(h0, w0) # resize image to img_size
if r != 1: # always resize down, only resize up if training with augmentation
interp = cv2.INTER_AREA if r < 1 and not self.augment else cv2.INTER_LINEAR
img = cv2.resize(img, (int(w0 * r), int(h0 * r)), interpolation=interp)
return img, (h0, w0), img.shape[:2] # img, hw_original, hw_resized
else:
return self.imgs[index], self.img_hw0[index], self.img_hw[index] # img, hw_original, hw_resized

def random_perspective(img, targets=(), degrees=10, translate=.1, scale=.1, shear=10, perspective=0.0, border=(0, 0)):
# torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10))
# targets = [cls, xyxy]

# height = img.shape[0] + border[0] * 2 # shape(h,w,c)
# width = img.shape[1] + border[1] * 2
if border==(0, 0): # 不使用mosaic
height = img.shape[0] # shape(h,w,c)
width = img.shape[1]
else: # 使用mosaic， 先将图片扩展为2倍，所以需要回归到原始大小
height = img.shape[0]// 2 # shape(h,w,c)
width = img.shape[1] // 2


# Center
C = np.eye(3)
C[0, 2] = -img.shape[1] / 2 # x translation (pixels)
C[1, 2] = -img.shape[0] / 2 # y translation (pixels)

# Perspective
P = np.eye(3)
P[2, 0] = random.uniform(-perspective, perspective) # x perspective (about y)
P[2, 1] = random.uniform(-perspective, perspective) # y perspective (about x)

# Rotation and Scale
R = np.eye(3)
a = random.uniform(-degrees, degrees)
# a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations
s = random.uniform(1 - scale, 1 + scale)
# s = 2 ** random.uniform(-scale, scale)
R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)

# Shear
S = np.eye(3)
S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # x shear (deg)
S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # y shear (deg)

# Translation
T = np.eye(3)
T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width # x translation (pixels)
T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height # y translation (pixels)

# Combined rotation matrix
M = T @ S @ R @ P @ C # order of operations (right to left) is IMPORTANT
if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any(): # image changed
if perspective:
img = cv2.warpPerspective(img, M, dsize=(width, height), borderValue=(114, 114, 114))
else: # affine
img = cv2.warpAffine(img, M[:2], dsize=(width, height), borderValue=(114, 114, 114))

# Visualize
# import matplotlib.pyplot as plt
# ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel()
# ax[0].imshow(img[:, :, ::-1]) # base
# ax[1].imshow(img2[:, :, ::-1]) # warped

# Transform label coordinates
n = len(targets)
if n:
# warp points
xy = np.ones((n * 4, 3))
xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2) # x1y1, x2y2, x1y2, x2y1
xy = xy @ M.T # transform
if perspective:
xy = (xy[:, :2] / xy[:, 2:3]).reshape(n, 8) # rescale
else: # affine
xy = xy[:, :2].reshape(n, 8)

# create new boxes
x = xy[:, [0, 2, 4, 6]]
y = xy[:, [1, 3, 5, 7]]
xy = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T

# # apply angle-based reduction of bounding boxes
# radians = a * math.pi / 180
# reduction = max(abs(math.sin(radians)), abs(math.cos(radians))) ** 0.5
# x = (xy[:, 2] + xy[:, 0]) / 2
# y = (xy[:, 3] + xy[:, 1]) / 2
# w = (xy[:, 2] - xy[:, 0]) * reduction
# h = (xy[:, 3] - xy[:, 1]) * reduction
# xy = np.concatenate((x - w / 2, y - h / 2, x + w / 2, y + h / 2)).reshape(4, n).T

# clip boxes
xy[:, [0, 2]] = xy[:, [0, 2]].clip(0, width)
xy[:, [1, 3]] = xy[:, [1, 3]].clip(0, height)

# filter candidates
i = box_candidates(box1=targets[:, 1:5].T * s, box2=xy.T)
targets = targets[i]
targets[:, 1:5] = xy[i]

return img, targets

def augment_hsv(img, hgain=0.5, sgain=0.5, vgain=0.5):
r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1 # random gains
hue, sat, val = cv2.split(cv2.cvtColor(img, cv2.COLOR_BGR2HSV))
dtype = img.dtype # uint8

x = np.arange(0, 256, dtype=np.int16)
lut_hue = ((x * r[0]) % 180).astype(dtype)
lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)
lut_val = np.clip(x * r[2], 0, 255).astype(dtype)

img_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val))).astype(dtype)
cv2.cvtColor(img_hsv, cv2.COLOR_HSV2BGR, dst=img) # no return needed

# Histogram equalization
# if random.random() < 0.2:
# for i in range(3):
# img[:, :, i] = cv2.equalizeHist(img[:, :, i])