目标检测之caffe的py-faster-rcnn下数据集kitti转VOC、训练和测试模型全过程（实现车辆的检测识别）

本文默认你已经安装好了py-faster-rcnn源码，如果安装好了，那么就开始下面的工作吧！
源码地址：https://github.com/rbgirshick/py-faster-rcnn

kiit数据集转voc

kitti数据集下载地址：http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=2d

文件结构

创建下面文件，将下载好的数据集图片放入JPEGImages中，标注文件放入training中

zero@zero:~/kitti$ tree -L 1
.
├── Annotations #存放xml文件
├── ImageSets 
	└── Main
	    ├── test.txt
	    ├── train.txt
	    ├── trainval.txt
	    └── val.txt
├── JPEGImages #存放kitti原数据集图片的地址
└── training #标注文件存放地址
└── results #存放训练后的一些相关文件（不创建的话训练会报错）
	└── VOC2007
	    └── Main

转换代码

运行代码，生成xml文件以及四个txt文件

#!/usr/bin/Python
# -*- coding: utf-8 -*-

import cv2
import glob

from fasterrcnn_kitti.utils import writexml

#取样本的90%
trainval = open("/home/zero/kitti/ImageSets/Main/trainval.txt","w")
#70%
train = open("/home/zero/kitti/ImageSets/Main/train.txt","w")
#20%
val = open("/home/zero/kitti/ImageSets/Main/val.txt","w")
#取样本的10%
test = open("/home/zero/kitti/ImageSets/Main/test.txt","w")

#读取标注信息并进行解析

#打印文件路径和名称
list_anno_files = glob.glob("/home/zero/kitti/training/label_2/*")

idx = 0
#遍历文件
for file_path in list_anno_files:
    #打开文件
    with open(file_path) as file:
        #按行读取文件
        anno_infos = file.readlines()

        print(anno_infos)
        bboxes = []
        typename = []
        #遍历行的内容
        for anno_item in anno_infos:
            anno_infos = anno_item.split(" ")
            #去掉'Misc' 'Dontcare'
            if anno_infos[0] == "Misc" or anno_infos[0] == "DontCare":
                continue
            else:
                bbox = (int(float(anno_infos[4])),int(float(anno_infos[5])),
                        int(float(anno_infos[6])),int(float(anno_infos[7])))

                bboxes.append(bbox)
                typename.append(anno_infos[0])

        filename= file_path.split("/")[-1].replace("txt","png")
        xmlpath = "/home/zero/kitti/Annotations/" + filename.replace("png","xml")
        imgpath = "/home/zero/kitti/JPEGImages/" + filename
        saveimg = cv2.imread(imgpath)

        writexml(filename, saveimg, bboxes, xmlpath, typename)

        if idx > len(list_anno_files) * 0.9:
            test.write(filename.replace(".png","\n"))
        else:
            trainval.write(filename.replace(".png","\n"))
            if idx > len(list_anno_files) * 0.7:
                val.write(filename.replace(".png","\n"))
            else:
                train.write(filename.replace(".png","\n"))
        idx += 1
trainval.close()
train.close()
val.close()
test.close()

训练过程

修改 *train.pt

models/pascal_voc/ZF/faster_rcnn_alt_opt下

stage1_fast_rcnn_train.pt

name: "ZF"
layer {
  name: 'data'
  type: 'Python'
  top: 'data'
  top: 'rois'
  top: 'labels'
  top: 'bbox_targets'
  top: 'bbox_inside_weights'
  top: 'bbox_outside_weights'
  python_param {
    module: 'roi_data_layer.layer'
    layer: 'RoIDataLayer'
    param_str: "'num_classes': 8" #类别+1
  }
}

layer {
  name: "cls_score"
  type: "InnerProduct"
  bottom: "fc7"
  top: "cls_score"
  param { lr_mult: 1.0 }
  param { lr_mult: 2.0 }
  inner_product_param {
    num_output: 8 #类别+1
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "bbox_pred"
  type: "InnerProduct"
  bottom: "fc7"
  top: "bbox_pred"
  param { lr_mult: 1.0 }
  param { lr_mult: 2.0 }
  inner_product_param {
    num_output: 32 #（类别+1）× 4
    weight_filler {
      type: "gaussian"
      std: 0.001
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

stage1_rpn_train.pt

name: "ZF"
layer {
  name: 'input-data'
  type: 'Python'
  top: 'data'
  top: 'im_info'
  top: 'gt_boxes'
  python_param {
    module: 'roi_data_layer.layer'
    layer: 'RoIDataLayer'
    param_str: "'num_classes': 8"
  }
}

stage2_fast_rcnn_train.pt

同stage1_rpn_train.pt修改地方相同。

stage2_rpn_train.pt

同stage1_rpn_train.pt修改地方相同。

faster_rcnn_test.pt

layer {
  name: "cls_score"
  type: "InnerProduct"
  bottom: "fc7"
  top: "cls_score"
  inner_product_param {
    num_output: 8
  }
}
layer {
  name: "bbox_pred"
  type: "InnerProduct"
  bottom: "fc7"
  top: "bbox_pred"
  inner_product_param {
    num_output: 32
  }
}

修改pascal_voc.py

lib/datasets/pascal_voc.py

首先将代码中所有的 self._devkit_path 替换为 self._data_path

（1）class这里我改为了大写，为了和voc中的类别一致！！否者会报错

        self._data_path = "/home/zero/kitti"#voc数据集存放的位置
        self._classes = ('__background__', # always index 0
                         'Person_sitting', 'Truck', 'Van', 'Pedestrian',
                         'Cyclist', 'Tram', 'Car')

（2）图片格式

        self._image_ext = '.png'

（3）注释掉：

        #assert os.path.exists(self._devkit_path), \
        #        'VOCdevkit path does not exist: {}'.format(self._devkit_path)

（4）

            # Make pixel indexes 0-based
            x1 = float(bbox.find('xmin').text)
            y1 = float(bbox.find('ymin').text)
            x2 = float(bbox.find('xmax').text)
            y2 = float(bbox.find('ymax').text)
            cls = self._class_to_ind[obj.find('name').text.strip()]

（5）修改xml文件和图片存放地址

    def _do_python_eval(self, output_dir = 'output'):
        annopath = os.path.join(
            self._data_path,#'VOC' + self._year,
            'Annotations',
            '{:s}.xml')
        imagesetfile = os.path.join(
            self._data_path,
            'ImageSets',
            'Main',
            self._image_set + '.txt')

修改迭代次数

tools/train_faster_rcnn_alt_opt.py

    max_iters = [10000, 5000, 10000, 5000] #如果时间充足可以使用默认的训练

开始训练

./experiments/scripts/faster_rcnn_alt_opt.sh 0 ZF pascal_voc

当我们需要重新训练时，需要清除掉data下的cache文件夹和kitti中的annotations_cache文件夹

训练后：

测试

修改\tools\demo.py

（1）

CLASSES = ('__background__',  # always index 0
                 'Person_sitting', 'Truck', 'Van', 'Pedestrian',
                 'Cyclist', 'Tram', 'Car')

（2）选择你想测试的图片的存放在位置 /data/demo/中：

    im_names = ['000016.png', '000017.png', '000018.png',
                '000021.png', '000020.png']

运行

./tools/demo.py --net zf

结果：

参考链接

https://blog.csdn.net/sinat_30071459/article/details/51332084