这⼤概是最细的YOLOX中的MosaicAndMixup实现源码分析了吧博客园排版有bug，更好的阅读体验请见
前⾔
看了yolox后发现数据增强是真的nb，但是⾃⼰想如何实现的时候就感觉不太⾏了（不能简洁的实现）。⼜⼀想，数据增强这种trick肯定会⽤到其他⽹络的dataloader⾥⾯啊，所以仔细研究了⼀下代码复现⼀下。
最后附上我⾃⼰封装的mosaic和mixup，不⾃⼰封装到时候现copy别⼈的都不知bug在哪虽然核⼼与原论⽂差不多
Mosaic
源码分析
下⾯根据yolox源码进⾏分析：
yolox想法是先⽣成⼀个Dataset类，然后根据这个类可以进⾏iterater，故写了⼀个pull_item函数。
基于以上，然后可以定义到MosaicDetection类
class MosaicDetection(Dataset):
"""Detection dataset wrapper that performs mixup for normal dataset."""
def __init__(
self, dataset, img_size, mosaic=True, preproc=None,
degrees=10.0, translate=0.1, mosaic_scale=(0.5, 1.5),
mixup_scale=(0.5, 1.5), shear=2.0, perspective=0.0,
enable_mixup=True, mosaic_prob=1.0, mixup_prob=1.0, *args
):
super().__init__(img_size, mosaic=mosaic)
self._dataset = dataset
self.preproc = preproc
self.degrees = degrees
self.scale = mosaic_scale
self.shear = shear
self.perspective = perspective
self.mixup_scale = mixup_scale
self.mixup_prob = mixup_prob
self.local_rank = get_local_rank()
参数含义就不讲了，关键是self._dataset这个字段，可以看出Mosaic是在原先的Dataset基础上实现的。
也就是说需要的只是重写getitem和len，下⾯开始讲解getitem
第⼀部分图⽚拼接
def __getitem__(self, idx):
able_mosaic and random.random() < saic_prob:
mosaic_labels = []
input_dim = self._dataset.input_dim
input_h, input_w = input_dim[0], input_dim[1]
# yc, xc = s, s  # mosaic center x, y
# 画布⼤⼩为input_h,input_w
# 拼接公共点位置
yc = int(random.uniform(0.5 * input_h, 1.5 * input_h))
xc = int(random.uniform(0.5 * input_w, 1.5 * input_w))
# 3 additional image indices
indices = [idx] + [random.randint(0, len(self._dataset) - 1) for _ in range(3)]
for i_mosaic, index in enumerate(indices):
img, _labels, _, img_id = self._dataset.pull_item(index)
# 得到的第⼀张图⽚的原始⼤⼩
h0, w0 = img.shape[:2]
scale = min(1. * input_h / h0, 1. * input_w / w0)
# 放⼤到input size
img = size(
img, (int(w0 * scale), int(h0 * scale)), interpolation=cv2.INTER_LINEAR
)
# generate output mosaic image
(h, w, c) = img.shape[:3]
# ⽣成⼀个新的画布，颜⾊是114
if i_mosaic == 0:
mosaic_img = np.full((input_h * 2, input_w * 2, c), 114, dtype=np.uint8)
# suffix l means large image, while s means small image in mosaic aug.
# 根据图⽚的先后顺序分别放⼊左上、右上、左下、右下四个⽅向。
# 函数返回的是基于画布的新坐标和原图像的坐标（要注意由于0.5-1.5倍，原图像可能会超出画布范围
(l_x1, l_y1, l_x2, l_y2), (s_x1, s_y1, s_x2, s_y2) = get_mosaic_coordinate(
mosaic_img, i_mosaic, xc, yc, w, h, input_h, input_w
)
# 赋值到画布
mosaic_img[l_y1:l_y2, l_x1:l_x2] = img[s_y1:s_y2, s_x1:s_x2]
plt.imshow(mosaic_img)
plt.show()
# 坐标偏移量
padw, padh = l_x1 - s_x1, l_y1 - s_y1
labels = _py()
# Normalized xywh to pixel xyxy format
# 个⼈觉得这个注释意思有问题（可能我理解错了？下⾯细说
# 这是转换到新坐标轴的坐标
if _labels.size > 0:
# 左上⾓坐标
labels[:, 0] = scale * _labels[:, 0] + padw
labels[:, 1] = scale * _labels[:, 1] + padh
# 右下
labels[:, 2] = scale * _labels[:, 2] + padw
labels[:, 3] = scale * _labels[:, 3] + padh
mosaic_labels.append(labels)
plt.imshow(mosaic_img)
plt.show()
⼤概思路是先随机得到四张图⽚，然后创建⼀个⼤⼩为⽹络输⼊两倍的input，随机（0.5-1.5 scale）⽣成⼀个mosaic center（简单理解就是四张图⽚的公共点）。之后按照顺序拼接到左上、右上、左下、右下四个部分。
当⼀张图⽚放⼊画布时，得到x，y的原偏移量（padw，padh），然后计算偏移后的bbox位置。
有个问题是新bbox的坐标，注释写的是xywh转x1 y1 x2 y2，但是个⼈实现的时候发现输⼊是bbox的x1y1x2y2转换能正确框出，有⽆评论区⼤佬说明⼀下。
第⼆部分：图像旋转与剪切
if len(mosaic_labels):
# 将bbox超出画布部分变为画布边缘
mosaic_labels = np.concatenate(mosaic_labels, 0)
np.clip(mosaic_labels[:, 0], 0, 2 * input_w, out=mosaic_labels[:, 0])
np.clip(mosaic_labels[:, 1], 0, 2 * input_h, out=mosaic_labels[:, 1])
np.clip(mosaic_labels[:, 2], 0, 2 * input_w, out=mosaic_labels[:, 2])
np.clip(mosaic_labels[:, 3], 0, 2 * input_h, out=mosaic_labels[:, 3])
# 顺时针旋转degree°，输出新的图像和新的bbox坐标
mosaic_img, mosaic_labels = random_perspective(
mosaic_img,
mosaic_labels,
degrees=self.degrees,
anslate,
scale=self.scale,
shear=self.shear,
perspective=self.perspective,
border=[-input_h // 2, -input_w // 2],
)  # border to remove
这⼀部分就⽐较简单了，先是⽤clip函数处理好画布，然后旋转⼀个⾓度，旋转后bbox坐标变化其实可以不⽤关⼼，因为⾓度很⼩物体⼏乎超不出bbox的范围。细究旋转代码可以⾃⼰去看看我不想看了，最后还裁剪成了input size，所以这个最后输出还是input size⽽不是2*input size
Mix up
论⽂mosaic后半部分还增加了mixup（可选，但默认使⽤
# -----------------------------------------------------------------
# CopyPaste: /abs/2012.07177
# -----------------------------------------------------------------
if (
and not len(mosaic_labels) == 0
and random.random() < self.mixup_prob
# 如果mosaic_prob=0.5 mixup_prob=0.5这⾥0.5*0.5是0.25的概率mixup了
):
mosaic_img, mosaic_labels = self.mixup(mosaic_img, mosaic_labels, self.input_dim)
# 这⾥还增加了其他的预处理
mix_img, padded_labels = self.preproc(mosaic_img, mosaic_labels, self.input_dim)
img_info = (mix_img.shape[1], mix_img.shape[0])
# -----------------------------------------------------------------
# img_info and img_id are not used for training.
# They are also hard to be specified on a mosaic image.
# -----------------------------------------------------------------
return mix_img, padded_labels, img_info, img_id
else:
# 这个else是和mosaic的if对应的，不mosaic则默认只有预处理
self._dataset._input_dim = self.input_dim
img, label, img_info, img_id = self._dataset.pull_item(idx)
img, label = self.preproc(img, label, self.input_dim)
return img, label, img_info, img_id
# mixup函数
def mixup(self, origin_img, origin_labels, input_dim):
jit_factor = random.uniform(*self.mixup_scale)
# 图像是否翻转
FLIP = random.uniform(0, 1) > 0.5
cp_labels = []
# 保证不是背景 load_anno函数不涉及图像读取会更快（coco类
resizedwhile len(cp_labels) == 0:
cp_index = random.randint(0, self.__len__() - 1)
cp_labels = self._dataset.load_anno(cp_index)
# 确定不是背景后再载⼊img
img, cp_labels, _, _ = self._dataset.pull_item(cp_index)
# 创建画布
if len(img.shape) == 3:
cp_img = np.ones((input_dim[0], input_dim[1], 3), dtype=np.uint8) * 114
else:
cp_img = np.ones(input_dim, dtype=np.uint8) * 114
# 计算scale
cp_scale_ratio = min(input_dim[0] / img.shape[0], input_dim[1] / img.shape[1])        # resize
resized_img = size(
img,
(int(img.shape[1] * cp_scale_ratio), int(img.shape[0] * cp_scale_ratio)),
interpolation=cv2.INTER_LINEAR,
)
# 放⼊画布
cp_img[
: int(img.shape[0] * cp_scale_ratio), : int(img.shape[1] * cp_scale_ratio)
] = resized_img
# 画布放⼤jit factor倍
cp_img = size(
cp_img,
(int(cp_img.shape[1] * jit_factor), int(cp_img.shape[0] * jit_factor)),
)
cp_scale_ratio *= jit_factor
if FLIP:
cp_img = cp_img[:, ::-1, :]
# 以上创建好了⼀个可以mix up的图像
# 下⾯开始mix up
# 创建的画布向输⼊的图像上⾯叠加
origin_h, origin_w = cp_img.shape[:2]
target_h, target_w = origin_img.shape[:2]
# 取最⼤⾯积然后全部padding 0
padded_img = np.zeros(
(max(origin_h, target_h), max(origin_w, target_w), 3), dtype=np.uint8
)
# 放⼊新画布（也只有新画布
padded_img[:origin_h, :origin_w] = cp_img
# 随机偏移量
x_offset, y_offset = 0, 0
if padded_img.shape[0] > target_h:
y_offset = random.randint(0, padded_img.shape[0] - target_h - 1)
if padded_img.shape[1] > target_w:
x_offset = random.randint(0, padded_img.shape[1] - target_w - 1)
# 裁剪画布
padded_cropped_img = padded_img[
y_offset: y_offset + target_h, x_offset: x_offset + target_w
]
# 调整scale后画布中图像的bbox坐标
cp_bboxes_origin_np = adjust_box_anns(
cp_labels[:, :4].copy(), cp_scale_ratio, 0, 0, origin_w, origin_h
)
# 是否镜像翻转
if FLIP:
cp_bboxes_origin_np[:, 0::2] = (
origin_w - cp_bboxes_origin_np[:, 0::2][:, ::-1]
)
# 调整裁剪后bbox坐标（以裁剪左上⾓为新的原点
cp_bboxes_transformed_np = cp_bboxes_py()
cp_bboxes_transformed_np[:, 0::2] = np.clip(
cp_bboxes_transformed_np[:, 0::2] - x_offset, 0, target_w
)
cp_bboxes_transformed_np[:, 1::2] = np.clip(
cp_bboxes_transformed_np[:, 1::2] - y_offset, 0, target_h
)
# 通过五个条件判断offset是否合理，下⾯细说
keep_list = box_candidates(cp_bboxes_origin_np.T, cp_bboxes_transformed_np.T, 5)
# 满⾜条件则合并label和image
if keep_list.sum() >= 1.0:
cls_labels = cp_labels[keep_list, 4:5].copy()
box_labels = cp_bboxes_transformed_np[keep_list]
labels = np.hstack((box_labels, cls_labels))
origin_labels = np.vstack((origin_labels, labels))
origin_img = origin_img.astype(np.float32)
origin_img = 0.5 * origin_img + 0.5 * padded_cropped_img.astype(np.float32)
return origin_img.astype(np.uint8), origin_labels
总体来说⽐较好理解，因为坐标变换⽅法和mosaic相同，⽽最头疼的就是坐标变换了。
⾸先随机出⼀个⾮背景图像（必定有bbox的图像），然后缩放到input size，再放⼊input size（⽐如650*640）⼤⼩的画布。然后画布整体放⼤到jit facotr倍，在原图和新图中寻最⼤的画布，在⼤画布中随机出裁剪偏移量，裁剪，检查没问题后mix up即可。
⼤致流程如下（省略了寻最⼤的画布过程）：
下⾯讲检查函数box_candidates：
def box_candidates(box1, box2, wh_thr=2, ar_thr=20, area_thr=0.2):
# box1(4,n), box2(4,n)
# Compute candidate boxes which include follwing 5 things:
# box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio
w1, h1 = box1[2] - box1[0], box1[3] - box1[1]
w2, h2 = box2[2] - box2[0], box2[3] - box2[1]
ar = np.maximum(w2 / (h2 + 1e-16), h2 / (w2 + 1e-16))  # aspect ratio
return (
(w2 > wh_thr)
& (h2 > wh_thr)
& (w2 * h2 / (w1 * h1 + 1e-16) > area_thr)
& (ar < ar_thr)
)  # candidates
就是将偏移后的box和偏移前的box进⾏⽐较，四项指标分别是偏移后的box宽度，⾼度，⾯积，box长宽⽐
注释⾥写的五个实现只有四个
{% image ,alt='最终结果，中间的那两个是mix up',height=60vh %}
⾃⽤代码
因为yolox等⾥⾯肯定是⽤了各种东西对dataloader加速⽐如pycoco类封装（这个包不是很懂）、preload等，⼀时半会也看不完。只好剥离
了，loader的效率估计不会那么⾼以后变成⼤⽜了再加吧
# -*- coding:utf-8 -*-
# @Author : Dummerfu
# @Contact : github/dummerchen
# @Time : 2021/9/25 14:06
import math
from draw_box_utli import draw_box
from torch.utils.data import Dataset
from VocDataset import VocDataSet
import matplotlib as mpl
import random
import cv2
import numpy as np
from matplotlib import pyplot as plt
def get_mosaic_coordinate(mosaic_image, mosaic_index, xc, yc, w, h, input_h, input_w):    # TODO update doc
# index0 to top left part of image
if mosaic_index == 0:
x1, y1, x2, y2 = max(xc - w, 0), max(yc - h, 0), xc, yc
small_coord = w - (x2 - x1), h - (y2 - y1), w, h
# index1 to top right part of image
elif mosaic_index == 1:
x1, y1, x2, y2 = xc, max(yc - h, 0), min(xc + w, input_w * 2), yc
small_coord = 0, h - (y2 - y1), min(w, x2 - x1), h
# index2 to bottom left part of image
elif mosaic_index == 2:
x1, y1, x2, y2 = max(xc - w, 0), yc, xc, min(input_h * 2, yc + h)
small_coord = w - (x2 - x1), 0, w, min(y2 - y1, h)
# index2 to bottom right part of image
elif mosaic_index == 3:
x1, y1, x2, y2 = xc, yc, min(xc + w, input_w * 2), min(input_h * 2, yc + h)  # noqa        small_coord = 0, 0, min(w, x2 - x1), min(y2 - y1, h)
return (x1, y1, x2, y2), small_coord
def random_perspective(
img,
targets=(),
degrees=10,
translate=0.1,
scale=0.1,
shear=10,
perspective=0.0,
border=(0, 0),
):
# targets = [cls, xyxy]
height = img.shape[0] + border[0] * 2  # shape(h,w,c)
width = img.shape[1] + border[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)
# 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(scale[0], scale[1])
# s = 2 ** random.uniform(-scale, scale)
R[:2] = RotationMatrix2D(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 @ C  # order of operations (right to left) is IMPORTANT
>>>>>##
# For Aug out of Mosaic
# s = 1.
# M = np.eye(3)
>>>>>##
if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any():  # image changed
if perspective:
img = cv2.warpPerspective(

版权声明：本站内容均来自互联网，仅供演示用，请勿用于商业和其他非法用途。如果侵犯了您的权益请与我们联系QQ:729038198，我们将在24小时内删除。

wiki的搭建
« 上一篇
PHP将身份证正反面两张照片合成一张图片的代码
下一篇 »

发表评论