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) for x in imagelist if (util.custombasename(x) != 'Thumbs')]
if self.num_process == 1:
for name in imagenames:
self.SplitSingle(name, rate, self.ext)
else:
# worker = partial(self.SplitSingle, rate=rate, extent=self.ext)
worker = partial(split_single_warp, split_base=self, rate=rate, extent=self.ext)
self.pool.map(worker, imagenames)
def __getstate__(self):
self_dict = self.__dict__.copy()
del self_dict['pool']
return self_dict
def __setstate__(self, state):
self.__dict__.update(state)
if __name__ == '__main__':
split = splitbase('D:/Dataset/DOTA-v2.0/train',
'D:/Dataset/DOTA-v2.0/trainsplit',
gap=200,
subsize=1024,
num_process=8
)
split.splitdata(1)
split = splitbase('D:/Dataset/DOTA-v2.0/val',
'D:/Dataset/DOTA-v2.0/valsplit',
gap=200,
subsize=1024,
num_process=8
)
split.splitdata(1)
标签转换
DOTA数据集的标签并不符合YOLO的要求，需要进行转换，如下图所示，需要将左侧的原始标签转换成右侧的YOLO格式。
这里使用的是参考程序中的YOLO_Transform.py这个脚本，同时需注意，需要在dota_utils.py中修改类别名称wordname_18。
YOLO_Transform.py
import dota_utils as util
import os
import numpy as np
from PIL import Image
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
Image.MAX_IMAGE_PIXELS = None
## trans dota format to format YOLO(darknet) required
def dota2darknet(imgpath, txtpath, dstpath, extractclassname):
"""
:param imgpath: the path of images
:param txtpath: the path of txt in dota format
:param dstpath: the path of txt in YOLO format
:param extractclassname: the category you selected
:return:
"""
filelist = util.GetFileFromThisRootDir(txtpath)
for fullname in filelist:
objects = util.parse_dota_poly(fullname)
name = os.path.splitext(os.path.basename(fullname))[0]
img_fullname = os.path.join(imgpath, name + '.png')
img = Image.open(img_fullname)
img_w, img_h = img.size
# print img_w,img_h
with open(os.path.join(dstpath, name + '.txt'), 'w') as f_out:
for obj in objects:
poly = obj['poly']
bbox = np.array(util.dots4ToRecC(poly, img_w, img_h))
if (sum(bbox <= 0) + sum(bbox >= 1)) >= 1:
continue
if (obj['name'] in extractclassname):
id = extractclassname.index(obj['name'])
else:
continue
outline = str(id) + ' ' + ' '.join(list(map(str, bbox)))
f_out.write(outline + '\n')
if __name__ == '__main__':
dota2darknet('C:/Users/xy/Desktop/Work/upload/DOTA/train/images',
'C:/Users/xy/Desktop/Work/upload/DOTA/train/labels1',
'C:/Users/xy/Desktop/Work/upload/DOTA/train/labels',
util.wordname_18)
dota2darknet('C:/Users/xy/Desktop/Work/upload/DOTA/val/images',
'C:/Users/xy/Desktop/Work/upload/DOTA/val/labels1',
'C:/Users/xy/Desktop/Work/upload/DOTA/val/labels',
util.wordname_18)
dota_utils.py
import sys
import codecs
import numpy as np
import shapely.geometry as shgeo
import os
import re
import math
"""
some basic functions which are useful for process DOTA data
"""
wordname_18 = [
'airport',
'small-vehicle',
'large-vehicle',
'plane',
'storage-tank',
'ship',
'harbor',
'ground-track-field',
'soccer-ball-field',
'tennis-court',
'swimming-pool',
'baseball-diamond',
'roundabout',
'basketball-court',
'bridge',
'helicopter',
'container-crane',
'helipad']
def custombasename(fullname):
return os.path.basename(os.path.splitext(fullname)[0])
def GetFileFromThisRootDir(dir,ext = None):
allfiles = []
needExtFilter = (ext != None)
for root,dirs,files in os.walk(dir):
for filespath in files:
filepath = os.path.join(root, filespath)
extension = os.path.splitext(filepath)[1][1:]
if needExtFilter and extension in ext:
allfiles.append(filepath)
elif not needExtFilter:
allfiles.append(filepath)
return allfiles
def TuplePoly2Poly(poly):
outpoly = [poly[0][0], poly[0][1],
poly[1][0], poly[1][1],
poly[2][0], poly[2][1],
poly[3][0], poly[3][1]
]
return outpoly
def parse_dota_poly(filename):
"""
parse the dota ground truth in the format:
[(x1, y1), (x2, y2), (x3, y3), (x4, y4)]
"""
objects = []
#print('filename:', filename)
f = []
if (sys.version_info >= (3, 5)):
fd = open(filename, 'r')
f = fd
elif (sys.version_info >= 2.7):
fd = codecs.open(filename, 'r')
f = fd
# count = 0
while True:
line = f.readline()
# count = count + 1
# if count < 2:
# continue
if line:
splitlines = line.strip().split(' ')
object_struct = {}
### clear the wrong name after check all the data
#if (len(splitlines) >= 9) and (splitlines[8] in classname):
if (len(splitlines) < 9):
continue
if (len(splitlines) >= 9):
object_struct['name'] = splitlines[8]
if (len(splitlines) == 9):
object_struct['difficult'] = '0'
elif (len(splitlines) >= 10):
# if splitlines[9] == '1':
# if (splitlines[9] == 'tr'):
# object_struct['difficult'] = '1'
# else:
object_struct['difficult'] = splitlines[9]
# else:
# object_struct['difficult'] = 0
object_struct['poly'] = [(float(splitlines[0]), float(splitlines[1])),
(float(splitlines[2]), float(splitlines[3])),
(float(splitlines[4]), float(splitlines[5])),
(float(splitlines[6]), float(splitlines[7]))
]
gtpoly = shgeo.Polygon(object_struct['poly'])
object_struct['area'] = gtpoly.area
# poly = list(map(lambda x:np.array(x), object_struct['poly']))
# object_struct['long-axis'] = max(distance(poly[0], poly[1]), distance(poly[1], poly[2]))
# object_struct['short-axis'] = min(distance(poly[0], poly[1]), distance(poly[1], poly[2]))
# if (object_struct['long-axis'] < 15):
# object_struct['difficult'] = '1'
# global small_count
# small_count = small_count + 1
objects.append(object_struct)
else:
break
return objects
def dots4ToRecC(poly, img_w, img_h):
xmin, ymin, xmax, ymax = dots4ToRec4(poly)
x = (xmin + xmax)/2
y = (ymin + ymax)/2
w = xmax - xmin
h = ymax - ymin
return x/img_w, y/img_h, w/img_w, h/img_h
def parse_dota_poly2(filename):
"""
parse the dota ground truth in the format:
[x1, y1, x2, y2, x3, y3, x4, y4]
"""
objects = parse_dota_poly(filename)
for obj in objects:
obj['poly'] = TuplePoly2Poly(obj['poly'])
obj['poly'] = list(map(int, obj['poly']))
return objects
def parse_dota_rec(filename):
"""
parse the dota ground truth in the bounding box format:
"xmin, ymin, xmax, ymax"
"""
objects = parse_dota_poly(filename)
for obj in objects:
poly = obj['poly']
bbox = dots4ToRec4(poly)
obj['bndbox'] = bbox
return objects
## bounding box transfer for varies format
def dots4ToRec4(poly):
xmin, xmax, ymin, ymax = min(poly[0][0], min(poly[1][0], min(poly[2][0], poly[3][0]))), \
max(poly[0][0], max(poly[1][0], max(poly[2][0], poly[3][0]))), \
min(poly[0][1], min(poly[1][1], min(poly[2][1], poly[3][1]))), \
max(poly[0][1], max(poly[1][1], max(poly[2][1], poly[3][1])))
return xmin, ymin, xmax, ymax
def dots4ToRec8(poly):
xmin, ymin, xmax, ymax = dots4ToRec4(poly)
return xmin, ymin, xmax, ymin, xmax, ymax, xmin, ymax
#return dots2ToRec8(dots4ToRec4(poly))
def dots2ToRec8(rec):
xmin, ymin, xmax, ymax = rec[0], rec[1], rec[2], rec[3]
return xmin, ymin, xmax, ymin, xmax, ymax, xmin, ymax
def groundtruth2Task1(srcpath, dstpath):
filelist = GetFileFromThisRootDir(srcpath)
# names = [custombasename(x.strip())for x in filelist]
filedict = {}
for cls in wordname_15:
fd = open(os.path.join(dstpath, 'Task1_') + cls + r'.txt', 'w')
filedict[cls] = fd
for filepath in filelist:
objects = parse_dota_poly2(filepath)
subname = custombasename(filepath)
pattern2 = re.compile(r'__([\d+\.]+)__\d+___')
rate = re.findall(pattern2, subname)[0]
for obj in objects:
category = obj['name']
difficult = obj['difficult']
poly = obj['poly']
if difficult == '2':
continue
if rate == '0.5':
outline = custombasename(filepath) + ' ' + '1' + ' ' + ' '.join(map(str, poly))
elif rate == '1':
outline = custombasename(filepath) + ' ' + '0.8' + ' ' + ' '.join(map(str, poly))
elif rate == '2':
outline = custombasename(filepath) + ' ' + '0.6' + ' ' + ' '.join(map(str, poly))
filedict[category].write(outline + '\n')
def Task2groundtruth_poly(srcpath, dstpath):
thresh = 0.1
filedict = {}
Tasklist = GetFileFromThisRootDir(srcpath, '.txt')
for Taskfile in Tasklist:
idname = custombasename(Taskfile).split('_')[-1]
# idname = datamap_inverse[idname]
f = open(Taskfile, 'r')
lines = f.readlines()
for line in lines:
if len(line) == 0:
continue
# print('line:', line)
splitline = line.strip().split(' ')
filename = splitline[0]
confidence = splitline[1]
bbox = splitline[2:]
if float(confidence) > thresh:
if filename not in filedict:
# filedict[filename] = codecs.open(os.path.join(dstpath, filename + '.txt'), 'w', 'utf_16')
filedict[filename] = codecs.open(os.path.join(dstpath, filename + '.txt'), 'w')
# poly = util.dots2ToRec8(bbox)
poly = bbox
# filedict[filename].write(' '.join(poly) + ' ' + idname + '_' + str(round(float(confidence), 2)) + '\n')
# print('idname:', idname)
# filedict[filename].write(' '.join(poly) + ' ' + idname + '_' + str(round(float(confidence), 2)) + '\n')
filedict[filename].write(' '.join(poly) + ' ' + idname + '\n')
def polygonToRotRectangle(bbox):
"""
:param bbox: The polygon stored in format [x1, y1, x2, y2, x3, y3, x4, y4]
:return: Rotated Rectangle in format [cx, cy, w, h, theta]
"""
bbox = np.array(bbox,dtype=np.float32)
bbox = np.reshape(bbox,newshape=(2,4),order='F')
angle = math.atan2(-(bbox[0,1]-bbox[0,0]),bbox[1,1]-bbox[1,0])
center = [[0],[0]]
for i in range(4):
center[0] += bbox[0,i]
center[1] += bbox[1,i]
center = np.array(center,dtype=np.float32)/4.0
R = np.array([[math.cos(angle), -math.sin(angle)], [math.sin(angle), math.cos(angle)]], dtype=np.float32)
normalized = np.matmul(R.transpose(),bbox-center)
xmin = np.min(normalized[0,:])
xmax = np.max(normalized[0,:])
ymin = np.min(normalized[1,:])
ymax = np.max(normalized[1,:])
w = xmax - xmin + 1
h = ymax - ymin + 1
return [float(center[0]),float(center[1]),w,h,angle]
图片格式转换
另外作者还提供了一个脚本用于转换图片格式，比如将png格式转成jpg，使用opencv进行实现。
这里虽然没用到，还是放置在此，如需训练自己的数据集可以使用。
imagetrans.py
import dota_utils as util
import cv2
import os
# this code is used to convert image formats
# from PNG to JPG
def imageformatTrans(srcpath, dstpath, format):
filelist = util.GetFileFromThisRootDir(srcpath)
for fullname in filelist:
img = cv2.imread(fullname)
basename = util.custombasename(fullname)
dstname = os.path.join(dstpath, basename + format)
cv2.imwrite(dstname, img)
if __name__ == '__main__':
# an example
imageformatTrans('path1', 'path2',
'.jpg')
另外，如果下载的是我提供的数据集，无需进行这些操作，我将原数据集和标签/分割数据集和标签全部转换完成，可以直接调入YOLOv5中使用。
训练结果
下图是我使用YOLOv5l模型的训练结果，可以看到训练100个epoch之后，模型基本收敛。
未分割训练效果：
分割之后的训练效果：
数据对比：
模型 | mAP@.5 | mAP@.5:.95 |
YOLOv5l(未分割) | 13.5% | 5.52% |
YOLOv5l(分割之后) | 33.5% | 18.6% |
结果可视化展示
这里使用两张DOTA-test中的图片做对比测试。
未分割前：
分割后：
可以看到区别还是相当明显的，分割之后尽管还有少部分目标漏检，大部分目标都能准确得检测出来。
