faster-rcnn按部就班跑代码预测

参考：https://blog.youkuaiyun.com/qq_29462849/article/details/80510687
最好别看他的

训练好的模型下载地址：http://download.tensorflow.org/models/object_detection/faster_rcnn_resnet101_coco_11_06_2017.tar.gz
模型列表：
https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md
这里里面找相关模块和文件：
https://github.com/tensorflow/models

需要下面这些东西

代码：

#-*-coding:utf-8-*-
import tensorflow as tf
import cv2
import  numpy as np
import matplotlib.pyplot as plt
import label_map_util
import visualization_utils as vis_util
import os
path=r'./mscoco_label_map.pbtxt'
label_map = label_map_util.load_labelmap(path)
categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=90, use_display_name=True)
category_index = label_map_util.create_category_index(categories)
with tf.gfile.FastGFile(r'frozen_inference_graph.pb',mode='rb') as f:
    graph_def=tf.GraphDef() #图定义类对象
    graph_def.ParseFromString(f.read()) #字符串转对象
    with tf.Session() as sess:
        grapha = sess.graph
        tf.import_graph_def(graph_def, name='')  # 对象图导入sess
        for i in os.listdir('picture'):
            j=os.path.join(os.getcwd(),'picture',i)
            data1 = cv2.imdecode(np.fromfile(j, dtype=np.uint8), -1) #处理中文路径乱码问题
            data = data1[None]
            sess.run(tf.global_variables_initializer())
            image_tensor = grapha.get_tensor_by_name('image_tensor:0')
            boxes = grapha.get_tensor_by_name('detection_boxes:0')
            scores = grapha.get_tensor_by_name('detection_scores:0')
            classes = grapha.get_tensor_by_name('detection_classes:0')
            num_detections = grapha.get_tensor_by_name('num_detections:0')
            (boxes, scores, classes, num_detections) = sess.run(
                [boxes, scores, classes, num_detections],
                feed_dict={image_tensor: data})
            vis_util.visualize_boxes_and_labels_on_image_array(data1,
                                                      np.squeeze(boxes),
                                                      np.squeeze(classes).astype(np.int32),
                                                      np.squeeze(scores),
                                                      category_index,
                                                      use_normalized_coordinates=True)#对图片的处理标识出目标
            cv2.imwrite('ex'+i,data1)

效果

最后说明：
预测准确率还是有问题的，在测试过程中发现，最后一张图，用不同的读取方式，显示效果也不一样