关键词：
前言
之前在使用YOLOv5跑xView数据集时，发现准确率还是非常低的。在网上冲浪时，我发现了一种小样本检测策略：那就是把大分辨率的图片分割成小块进行训练，然后再输入大图进行检测。那么本篇博文就使用DOTA数据集来验证一下这种思路是否可行。
主要参考的项目：https://github.com/postor/DOTA-yolov3
DOTA数据集简介
DOTA数据集全称：Dataset for Object deTection in Aerial images
DOTA数据集v1.0共收录2806张4000 × 4000的图片，总共包含188282个目标。
DOTA数据集论文介绍：https://arxiv.org/pdf/1711.10398.pdf
数据集官网：https://captain-whu.github.io/DOTA/dataset.html
DOTA数据集总共有3个版本
DOTAV1.0
- 类别数目：15
- 类别名称：plane, ship, storage tank, baseball diamond, tennis court, basketball court, ground track field, harbor, bridge, large vehicle, small vehicle, helicopter, roundabout, soccer ball field , swimming pool
DOTAV1.5
- 类别数目：16
- 类别名称：plane, ship, storage tank, baseball diamond, tennis court, basketball court, ground track field, harbor, bridge, large vehicle, small vehicle, helicopter, roundabout, soccer ball field, swimming pool , container crane
DOTAV2.0
- 类别数目：18
- 类别名称：plane, ship, storage tank, baseball diamond, tennis court, basketball court, ground track field, harbor, bridge, large vehicle, small vehicle, helicopter, roundabout, soccer ball field, swimming pool, container crane, airport , helipad
本实验所使用的是DOTAV2.0版本，同样备份在我的GitHub上。
https://github.com/zstar1003/Dataset
图片分割
图片分割就是将大图切成一块块小图，同时需要注意将标签进行转换。
另外，为了防止目标被切断，每两个分割图有部分区域重合，具体的分割策略可以看我下方绘制的示意图。
分割代码使用的是参考项目提供的split.py这个程序。
这里需指定下列参数：
- 输入图片文件夹路径
- 输出图片文件夹路径
- gap：两个子图的重合宽度
- subsize：子图大小
- num_process：线程数
完整代码：
import os
import codecs
import numpy as np
import math
from dota_utils import GetFileFromThisRootDir
import cv2
import shapely.geometry as shgeo
import dota_utils as util
import copy
from multiprocessing import Pool
from functools import partial
import time
def choose_best_pointorder_fit_another(poly1, poly2):
"""
To make the two polygons best fit with each point
"""
x1 = poly1[0]
y1 = poly1[1]
x2 = poly1[2]
y2 = poly1[3]
x3 = poly1[4]
y3 = poly1[5]
x4 = poly1[6]
y4 = poly1[7]
combinate = [np.array([x1, y1, x2, y2, x3, y3, x4, y4]), np.array([x2, y2, x3, y3, x4, y4, x1, y1]),
np.array([x3, y3, x4, y4, x1, y1, x2, y2]), np.array([x4, y4, x1, y1, x2, y2, x3, y3])]
dst_coordinate = np.array(poly2)
distances = np.array([np.sum((coord - dst_coordinate) ** 2) for coord in combinate])
sorted = distances.argsort()
return combinate[sorted[0]]
def cal_line_length(point1, point2):
return math.sqrt(math.pow(point1[0] - point2[0], 2) + math.pow(point1[1] - point2[1], 2))
def split_single_warp(name, split_base, rate, extent):
split_base.SplitSingle(name, rate, extent)
class splitbase():
def __init__(self,
basepath,
outpath,
code='utf-8',
gap=512,
subsize=1024,
thresh=0.7,
choosebestpoint=True,
ext='.png',
padding=True,
num_process=8
):
"""
:param basepath: base path for dota data
:param outpath: output base path for dota data,
the basepath and outputpath have the similar subdirectory, 'images' and 'labelTxt'
:param code: encodeing format of txt file
:param gap: overlap between two patches
:param subsize: subsize of patch
:param thresh: the thresh determine whether to keep the instance if the instance is cut down in the process of split
:param choosebestpoint: used to choose the first point for the
:param ext: ext for the image format
:param padding: if to padding the images so that all the images have the same size
"""
self.basepath = basepath
self.outpath = outpath
self.code = code
self.gap = gap
self.subsize = subsize
self.slide = self.subsize - self.gap
self.thresh = thresh
self.imagepath = os.path.join(self.basepath, 'images')
self.labelpath = os.path.join(self.basepath, 'labelTxt')
self.outimagepath = os.path.join(self.outpath, 'images')
self.outlabelpath = os.path.join(self.outpath, 'labelTxt')
self.choosebestpoint = choosebestpoint
self.ext = ext
self.padding = padding
self.num_process = num_process
self.pool = Pool(num_process)
print('padding:', padding)
# pdb.set_trace()
if not os.path.isdir(self.outpath):
os.mkdir(self.outpath)
if not os.path.isdir(self.outimagepath):
# pdb.set_trace()
os.mkdir(self.outimagepath)
if not os.path.isdir(self.outlabelpath):
os.mkdir(self.outlabelpath)
# pdb.set_trace()
## point: (x, y), rec: (xmin, ymin, xmax, ymax)
# def __del__(self):
# self.f_sub.close()
## grid --> (x, y) position of grids
def polyorig2sub(self, left, up, poly):
polyInsub = np.zeros(len(poly))
for i in range(int(len(poly) / 2)):
polyInsub[i * 2] = int(poly[i * 2] - left)
polyInsub[i * 2 + 1] = int(poly[i * 2 + 1] - up)
return polyInsub
def calchalf_iou(self, poly1, poly2):
"""
It is not the iou on usual, the iou is the value of intersection over poly1
"""
inter_poly = poly1.intersection(poly2)
inter_area = inter_poly.area
poly1_area = poly1.area
half_iou = inter_area / poly1_area
return inter_poly, half_iou
def saveimagepatches(self, img, subimgname, left, up):
subimg = copy.deepcopy(img[up: (up + self.subsize), left: (left + self.subsize)])
outdir = os.path.join(self.outimagepath, subimgname + self.ext)
h, w, c = np.shape(subimg)
if (self.padding):
outimg = np.zeros((self.subsize, self.subsize, 3))
outimg[0:h, 0:w, :] = subimg
cv2.imwrite(outdir, outimg)
else:
cv2.imwrite(outdir, subimg)
def GetPoly4FromPoly5(self, poly):
distances = [cal_line_length((poly[i * 2], poly[i * 2 + 1]), (poly[(i + 1) * 2], poly[(i + 1) * 2 + 1])) for i
in range(int(len(poly) / 2 - 1))]
distances.append(cal_line_length((poly[0], poly[1]), (poly[8], poly[9])))
pos = np.array(distances).argsort()[0]
count = 0
outpoly = []
while count < 5:
# print('count:', count)
if (count == pos):
outpoly.append((poly[count * 2] + poly[(count * 2 + 2) % 10]) / 2)
outpoly.append((poly[(count * 2 + 1) % 10] + poly[(count * 2 + 3) % 10]) / 2)
count = count + 1
elif (count == (pos + 1) % 5):
count = count + 1
continue
else:
outpoly.append(poly[count * 2])
outpoly.append(poly[count * 2 + 1])
count = count + 1
return outpoly
def savepatches(self, resizeimg, objects, subimgname, left, up, right, down):
outdir = os.path.join(self.outlabelpath, subimgname + '.txt')
mask_poly = []
imgpoly = shgeo.Polygon([(left, up), (right, up), (right, down),
(left, down)])
with codecs.open(outdir, 'w', self.code) as f_out:
for obj in objects:
gtpoly = shgeo.Polygon([(obj['poly'][0], obj['poly'][1]),
(obj['poly'][2], obj['poly'][3]),
(obj['poly'][4], obj['poly'][5]),
(obj['poly'][6], obj['poly'][7])])
if (gtpoly.area <= 0):
continue
inter_poly, half_iou = self.calchalf_iou(gtpoly, imgpoly)
# print('writing...')
if (half_iou == 1):
polyInsub = self.polyorig2sub(left, up, obj['poly'])
outline = ' '.join(list(map(str, polyInsub)))
outline = outline + ' ' + obj['name'] + ' ' + str(obj['difficult'])
f_out.write(outline + '\\n')
elif (half_iou > 0):
# elif (half_iou > self.thresh):
## print('<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<')
inter_poly = shgeo.polygon.orient(inter_poly, sign=1)
out_poly = list(inter_poly.exterior.coords)[0: -1]
if len(out_poly) < 4:
continue
out_poly2 = []
for i in range(len(out_poly)):
out_poly2.append(out_poly[i][0])
out_poly2.append(out_poly[i][1])
if (len(out_poly) == 5):
# print('==========================')
out_poly2 = self.GetPoly4FromPoly5(out_poly2)
elif (len(out_poly) > 5):
"""
if the cut instance is a polygon with points more than 5, we do not handle it currently
"""
continue
if (self.choosebestpoint):
out_poly2 = choose_best_pointorder_fit_another(out_poly2, obj['poly'])
polyInsub = self.polyorig2sub(left, up, out_poly2)
for index, item in enumerate(polyInsub):
if (item <= 1):
polyInsub[index] = 1
elif (item >= self.subsize):
polyInsub[index] = self.subsize
outline = ' '.join(list(map(str, polyInsub)))
if (half_iou > self.thresh):
outline = outline + ' ' + obj['name'] + ' ' + str(obj['difficult'])
else:
## if the left part is too small, label as '2'
outline = outline + ' ' + obj['name'] + ' ' + '2'
f_out.write(outline + '\\n')
# else:
# mask_poly.append(inter_poly)
self.saveimagepatches(resizeimg, subimgname, left, up)
def SplitSingle(self, name, rate, extent):
"""
split a single image and ground truth
:param name: image name
:param rate: the resize scale for the image
:param extent: the image format
:return:
"""
img = cv2.imread(os.path.join(self.imagepath, name + extent))
if np.shape(img) == ():
return
fullname = os.path.join(self.labelpath, name + '.txt')
objects = util.parse_dota_poly2(fullname)
for obj in objects:
obj['poly'] = list(map(lambda x: rate * x, obj['poly']))
# obj['poly'] = list(map(lambda x: ([2 * y for y in x]), obj['poly']))
if (rate != 1):
resizeimg = cv2.resize(img, None, fx=rate, fy=rate, interpolation=cv2.INTER_CUBIC)
else:
resizeimg = img
outbasename = name + '__' + str(rate) + '__'
weight = np.shape(resizeimg)[1]
height = np.shape(resizeimg)[0]
left, up = 0, 0
while (left < weight):
if (left + self.subsize >= weight):
left = max(weight - self.subsize, 0)
up = 0
while (up < height):
if (up + self.subsize >= height):
up = max(height - self.subsize, 0)
right = min(left + self.subsize, weight - 1)
down = min(up + self.subsize, height - 1)
subimgname = outbasename + str(left) + '___' + str(up)
# self.f_sub.write(name + ' ' + subimgname + ' ' + str(left) + ' ' + str(up) + '\\n')
self.savepatches(resizeimg, objects, subimgname, left, up, right, down)
if (up + self.subsize >= height):
break
else:
up = up + self.slide
if (left + self.subsize >= weight):
break
else:
left = left + self.slide
def splitdata(self, rate):
"""
:param rate: resize rate before cut
"""
imagelist = GetFileFromThisRootDir(self.imagepath)
imagenames = [util.custombasename(x) fo目标检测数据增强：dota数据集(代码片段)
前言之前对于xml格式的YOLO数据集，之前记录过如何用imgaug对其进行数据增强。不过DOTA数据集采用的是txt格式的旋转框标注，因此不能直接套用，只能另辟蹊径。DOTA数据集简介DOTA数据集全称：DatasetforObjectdeTectionin... 查看详情
如何入门yolo目标检测？
学习目标知道yolo网络架构，理解其输入输出知道yolo模型的训练样本构建的方法理解yolo模型的损失函数知道yoloV2模型的改进方法知道yoloV3的多尺度检测方法知道yoloV3模型的网络结构及网络输出了解yoloV3模型先验框设计的方法知道... 查看详情
如何入门yolo目标检测？
学习目标知道yolo网络架构，理解其输入输出知道yolo模型的训练样本构建的方法理解yolo模型的损失函数知道yoloV2模型的改进方法知道yoloV3的多尺度检测方法知道yoloV3模型的网络结构及网络输出了解yoloV3模型先验框设计的方法... 查看详情
目标检测小脚本：yolo标签可视化(代码片段)
需求分析在下载别人标注好的目标检测数据集时，我突然想到一个问题：怎么直观得看别人标注的是否正确呢？于是我想到了可以利用opencv将标注数据还原到原图上。更具体的说，指定图片和标签文件夹，批... 查看详情
基于yolov7的闸片厚度检测方法(代码片段)
基于yolov7的闸片厚度检测方法图像采集YOLOV7训练制作训练集训练样本生成权重文件修改YOLO预测类imag与cv2转换方法欧氏距离像素坐标与世界坐标映射测量方法边缘提取滤波方法检测结果图像采集利用基于3D结构光的深度相机采集... 查看详情
lvc|一种简单的小样本目标检测方法(代码片段)
...本文解读一下Label,Verify,Correct(LVC)：一种简单的小样本目标检测方法。 本文的目标是小样本目标检测(FSOD)，即在给定少量训练实例的情况下，就将一个目标检测器扩展到一个新类别的检测上。作者引入了一种简单的... 查看详情
keras深度学习实战（15）——从零开始实现yolo目标检测(代码片段)
Keras深度学习实战（15）——从零开始实现YOLO目标检测0.前言1.YOLO目标检测模型1.1锚框(anchorboxes)1.2YOLO目标检测模型原理2.从零开始实现YOLO目标检测2.1加载数据集2.2计算锚框尺寸2.3创建训练数据集2.4实现YOLO目标检测模型2.5... 查看详情
改进yolov7系列：首发最新结合多种x-transformer结构新增小目标检测层，让yolo目标检测任务中的小目标无处遁形
第三十六节，目标检测之yolo源码解析(代码片段)
在一个月前，我就已经介绍了yolo目标检测的原理，后来也把tensorflow实现代码仔细看了一遍。但是由于这个暑假事情比较大，就一直搁浅了下来，趁今天有时间，就把源码解析一下。关于yolo目标检测的原理请参考前面一篇文章：... 查看详情
目标检测数据增强：dota数据集(代码片段)
...据集v1.0共收录2806张4000×4000的图片，总共包含188282个目标。DOTA数据集论文介绍：https://arxiv.org/pdf/1711.10398.pdf数据集官网：https://captain-whu.github.io/DOTA/dataset.htmlDOTA数据集总共有3个版本DOTAV1.0类别数目：15类别名称... 查看详情
sylph|一种用于增量小样本目标检测的超网络框架(代码片段)
...，我是极智视界，本文解读一下一种用于小样本目标检测的超网络框架Sylph。 作者研究了具有挑战性的增量小样本目标检测任务(iFSD)。近年来，基于超网络的方法在连续的、无微调的iFSD背景下进行了研究，但取... 查看详情
推荐研究方向小样本开集目标检测(few-shotopen-setobjectdetection)
...rxiv.org/abs/2210.15996 1、任务：小样本开集目标检测，使用少量已知类样本训练模型，使得模型既能够检测小样本已知类又能够检测未知类。2、动机：解决FSOSOD问题有三个重要原因。1）可以识别小样本类别的开集... 查看详情
深度学习和目标检测系列教程13-300：yolo物体检测算法(代码片段)
@Author：RunsenYOLO，是目前速度更快的物体检测算法之一。虽然它不再是最准确的物体检测算法，但当您需要实时检测时，它是一个非常好的选择，而不会损失太多的准确性。YOLO框架在本篇博客中，我将... 查看详情
从零开始学习目标检测：yolo算法详解(代码片段)
从零开始学习目标检测：YOLO算法详解文章目录从零开始学习目标检测：YOLO算法详解1.🌟什么是目标检测?2.🌟传统的目标检测与基于深度学习的目标检测3.🌟目标检测算法的工作流程4.🌟目标检测可以干什... 查看详情
yolo系列目标检测算法-yolov7(代码片段)
YOLO系列目标检测算法目录YOLO系列目标检测算法总结对比YOLOv1YOLOv2YOLOv3YOLOv4Scaled-YOLOv4YOLOv5YOLOv6YOLOv7深度学习知识点总结专栏链接:https://blog.csdn.net/qq_39707285/article/details/124005405此专栏主要总结深度学习中的知识点，从各大数据... 查看详情
睿智的目标检测51——tensorflow2搭建yolo3目标检测平台(代码片段)
睿智的目标检测51——Tensorflow2搭建yolo3目标检测平台学习前言源码下载YoloV3实现思路一、整体结构解析二、网络结构解析1、主干网络Darknet53介绍2、构建FPN特征金字塔进行加强特征提取3、利用YoloHead获得预测结果三、预测结果的... 查看详情
改进yolov5系列：增加swin-transformer小目标检测头(代码片段)
...一使用YOLOv5代码框架，结合不同模块来构建不同的YOLO目标检测模型。🌟本项目包含大量的改进方式,降低改进难度,改进点包含【Backbone特征主干】、【Neck特征融合】、【Head检测头】、【注意力机制】、【IoU损失函数】、... 查看详情
yolo目标检测
文章大纲简介原理入门网络模型与可视化目标检测的评测mAP代码部分解读使用Demo数据集最佳实践模型大小计数模型部署与框架选择图片或视频推理可调用的封装模型训练模型转换多路摄像头可能出现的问题参考文献简介github：ht... 查看详情