自动驾驶数据集BDD训练mobileNet-SSD

自动驾驶数据集BDD训练mobileNet-SSD <1>

一、组织训练样本

1、json文件转换为xml文件

原始bdd数据集是json格式的标注文件，由于我使用caffe训练，所以必须要将其转化为
VOC的xml格式的标注文件。所以直接使用Python脚本将其转化为xml。bdd数据集中包含有10个类别，包括bus,light,sign,person,bike,truck,motor,car,train,rider,具体关于数据集的介绍，请参照。

我现在只需要数据集中8类，暂时不需要light和sign。可能这两类和国内的不太一样，所以我直接去掉。

转换代码 pascal_voc_io.py

#!/usr/bin/env python
# -*- coding: utf8 -*-
import sys
from xml.etree import ElementTree
from xml.etree.ElementTree import Element, SubElement
from lxml import etree


class PascalVocWriter:

    def __init__(self, foldername, filename, imgSize, databaseSrc='Unknown', localImgPath=None):
        self.foldername = foldername
        self.filename = filename
        self.databaseSrc = databaseSrc
        self.imgSize = imgSize
        self.boxlist = []
        self.localImgPath = localImgPath

    def prettify(self, elem):
        """
            Return a pretty-printed XML string for the Element.
        """
        rough_string = ElementTree.tostring(elem, 'utf8')
        root = etree.fromstring(rough_string)
        return etree.tostring(root, pretty_print=True)

    def genXML(self):
        """
            Return XML root
        """
        # Check conditions
        if self.filename is None or \
                self.foldername is None or \
                self.imgSize is None or \
                len(self.boxlist) <= 0:
            return None

        top = Element('annotation')
        folder = SubElement(top, 'folder')
        folder.text = self.foldername

        filename = SubElement(top, 'filename')
        filename.text = self.filename

        localImgPath = SubElement(top, 'path')
        localImgPath.text = self.localImgPath

        source = SubElement(top, 'source')
        database = SubElement(source, 'database')
        database.text = self.databaseSrc

        size_part = SubElement(top, 'size')
        width = SubElement(size_part, 'width')
        height = SubElement(size_part, 'height')
        depth = SubElement(size_part, 'depth')
        width.text = str(self.imgSize[1])
        height.text = str(self.imgSize[0])
        if len(self.imgSize) == 3:
            depth.text = str(self.imgSize[2])
        else:
            depth.text = '1'

        segmented = SubElement(top, 'segmented')
        segmented.text = '0'
        return top

    def addBndBox(self, xmin, ymin, xmax, ymax, name):
        bndbox = {'xmin': xmin, 'ymin': ymin, 'xmax': xmax, 'ymax': ymax}
        bndbox['name'] = name
        self.boxlist.append(bndbox)

    def appendObjects(self, top):
        for each_object in self.boxlist:
            object_item = SubElement(top, 'object')
            name = SubElement(object_item, 'name')
            name.text = unicode(each_object['name'])
            pose = SubElement(object_item, 'pose')
            pose.text = "Unspecified"
            truncated = SubElement(object_item, 'truncated')
            truncated.text = "0"
            difficult = SubElement(object_item, 'Difficult')
            difficult.text = "0"
            bndbox = SubElement(object_item, 'bndbox')
            xmin = SubElement(bndbox, 'xmin')
            xmin.text = str(each_object['xmin'])
            ymin = SubElement(bndbox, 'ymin')
            ymin.text = str(each_object['ymin'])
            xmax = SubElement(bndbox, 'xmax')
            xmax.text = str(each_object['xmax'])
            ymax = SubElement(bndbox, 'ymax')
            ymax.text = str(each_object['ymax'])

    def save(self, targetFile=None):
        root = self.genXML()
        self.appendObjects(root)
        out_file = None
        if targetFile is None:
            out_file = open(self.filename + '.xml', 'w')
        else:
            out_file = open(targetFile, 'w')

        prettifyResult = self.prettify(root)
        out_file.write(prettifyResult)
        out_file.close()


class PascalVocReader:

    def __init__(self, filepath):
        # shapes type:
        # [labbel, [(x1,y1), (x2,y2), (x3,y3), (x4,y4)], color, color]
        self.shapes = []
        self.filepath = filepath
        self.parseXML()

    def getShapes(self):
        return self.shapes

    def addShape(self, label, bndbox):
        xmin = int(bndbox.find('xmin').text)
        ymin = int(bndbox.find('ymin').text)
        xmax = int(bndbox.find('xmax').text)
        ymax = int(bndbox.find('ymax').text)
        points = [(xmin, ymin), (xmax, ymin), (xmax, ymax), (xmin, ymax)]
        self.shapes.append((label, points, None, None))

    def parseXML(self):
        assert self.filepath.endswith('.xml'), "Unsupport file format"
        parser = etree.XMLParser(encoding='utf-8')
        xmltree = ElementTree.parse(self.filepath, parser=parser).getroot()
        filename = xmltree.find('filename').text

        for object_iter in xmltree.findall('object'):
            bndbox = object_iter.find("bndbox")
            label = object_iter.find('name').text
            self.addShape(label, bndbox)
        return True


# tempParseReader = PascalVocReader('test.xml')
# print tempParseReader.getShapes()
"""
# Test
tmp = PascalVocWriter('temp','test', (10,20,3))
tmp.addBndBox(10,10,20,30,'chair')
tmp.addBndBox(1,1,600,600,'car')
tmp.save()
"""

parseJson.py

#!/usr/bin/env python
# -*- coding: utf8 -*-
#parse json，input json filename,output info needed by voc

import json
#这里是我需要的8个类别
categorys = ['car','bus','person','bike','truck','motor','train','rider']

def parseJson(jsonFile):
    objs = []
    obj = []
    f = open(jsonFile)
    info = json.load(f)
    objects = info['frames'][0]['objects']
    for i in objects:
        if(i['category'] in categorys):
            obj.append(int(i['box2d']['x1']))
            obj.append(int(i['box2d']['y1']))
            obj.append(int(i['box2d']['x2']))
            obj.append(int(i['box2d']['y2']))
            obj.append(i['category'])
            objs.append(obj)
            obj = []
    #print("objs",objs)
    return objs

#test
#parseJson("/home/nextcar/桌面/0a0a0b1a-7c39d841.json")

bdd2voc.py

import os
import pascal_voc_io
import parseJson

dirName = "/media/0A4811140A481114/bdd100k_labels/labels/100k/val"
i = 1
for dirpath,dirnames,filenames in os.walk(dirName):
    for filepath in filenames:
        fileName = os.path.join(dirpath,filepath)
        print("processing: ",i)
        i = i + 1
        xmlFileName = filepath[:-5]
        #print("xml: ",xmlFileName)
        objs = parseJson.parseJson(str(fileName))
        if len(objs):
            tmp = pascal_voc_io.PascalVocWriter('Annotations',xmlFileName, (720,1280,3))
            for obj in objs:
                tmp.addBndBox(obj[0],obj[1],obj[2],obj[3],obj[4])
            tmp.save()
        else:
            print(fileName)

直接调用bdd2voc.py就可以生成xml文件了，这样生成的文件就可以直接用作训练了。

转换成lmdb文件然后训练，参考

训练过程中，非常难以收敛，怎么调都不好使啊

然后在test的时候，把所有类别的准确度全部打印出来，然后惊奇地发现，我的天！train这一类的准确率只有百分之零点几

然后我意识到这个数据集类别数量分布及其不均匀

但是具体的数量是多少呢？打印出来看看

cd data/Annotations
find ./ type -f | xargs grep -ri "train" | wc -l

不看不知道，一看吓一跳啊，trian只有几十，但是car却有几十万，这个相差太大了，肯定会不好训练的

然后使用Python脚本将所有xml中的train的object节点全部删掉。如果一个xml文件中只有train这一类，那么这个xml文件就是空的，必须得删掉，然后对应的pic文件也要删掉。

#!/usr/bin/env python
# -*- coding: utf8 -*-
import pascal_voc_io
import os

xmlPath = "/home/Work/data/xml"
picPath = "/home/Work/data/pic"

def delSmallObj(objInfos):
    objsList = []
    objList = []
    isNull = False
    for obj in objInfos:
        objName = obj[0]
        xmin = obj[1][0][0] 
        ymin = obj[1][0][1]
        xmax = obj[1][1][0]
        ymax = obj[1][2][1]
        w = xmax - xmin
        h = ymax - ymin
        #删除train的obj节点
        if(objName != "train"):
            objList.append(int(xmin))
            objList.append(int(ymin))
            objList.append(int(xmax))
            objList.append(int(ymax))
            objList.append(objName)
            objsList.append(objList)
            objList = []
    if len(objsList) == 0:
        isNull = True
    return isNull,objsList


#生成xml，并返回xml为空的文件名
def gen():
    toDelXml = []
    for dirpath,dirnames,filenames in os.walk(xmlPath):
        for filepath in filenames:
            fileName = os.path.join(dirpath,filepath)
            xmlFileName = filepath[:-4]
            tempParseReader = pascal_voc_io.PascalVocReader(fileName)
            objInfos = tempParseReader.getShapes()
            isNull ,objs = delSmallObj(objInfos)
            if isNull == False:
                tmp = pascal_voc_io.PascalVocWriter('Annotations',xmlFileName, (720,1280,3))
                for obj in objs:
                    tmp.addBndBox(obj[0],obj[1],obj[2],obj[3],obj[4])
                tmp.save()
            else:
                toDelXml.append(fileName)
    return toDelXml

def DelEmptyXmlAndPic(toDelXml):
    for xmlFile in toDelXml:
        picFileName = xmlFile[:-4] + '.jpg'
        picFileName = picFileName.replace("xml","pic")
        print("deleting file: ",xmlFile)
        os.remove(picFileName)


if __name__ == "__main__":
    toDelXml = gen()
    DelEmptyXmlAndPic(toDelXml)
    print("empty num is: ", len(toDelXml))

这样之后数据集应该是没有太大问题了，虽然里面的类别还是不太均衡的。先训练再说吧。

---

关注我的公众号，分享资源
公众号搜索： 卡本特
扫码关注