【笔记】SSD : One-Stage （类似 Yolo），不同于Faster Rcnn

 

 

 

 

 

 

 

 

 

 

 

 

 

 

self.ssd_params = SSDParams(img_shape=(300, 300),   # 输入图片大小
                                    num_classes=21,     # 类别数+背景
                                    no_annotation_label=21,
                                    feat_layers=["block4", "block7", "block8", "block9", "block10", "block11"],   # 要进行检测的特征图name
                                    feat_shapes=[(38, 38), (19, 19), (10, 10), (5, 5), (3, 3), (1, 1)],  # 特征图大小
                                    anchor_size_bounds=[0.15, 0.90],  # 特征图尺度范围
                                    anchor_sizes=[(21., 45.),
                                                  (45., 99.),
                                                  (99., 153.),
                                                  (153., 207.),
                                                  (207., 261.),
                                                  (261., 315.)],  # 不同特征图的先验框尺度（第一个值是s_k，第2个值是s_k+1）
                                    anchor_ratios=[[2, .5],
                                                   [2, .5, 3, 1. / 3],
                                                   [2, .5, 3, 1. / 3],
                                                   [2, .5, 3, 1. / 3],
                                                   [2, .5],
                                                   [2, .5]], # 特征图先验框所采用的长宽比（每个特征图都有2个正方形先验框）
                                    anchor_steps=[8, 16, 32, 64, 100, 300],  # 特征图的单元大小
                                    anchor_offset=0.5,                       # 偏移值，确定先验框中心
                                    normalizations=[20, -1, -1, -1, -1, -1],  # l2 norm
                                    prior_scaling=[0.1, 0.1, 0.2, 0.2]       # variance
                                    )

 然后构建整个网络，注意对于stride=2的conv不要使用TF自带的padding="same"，而是手动pad，这是为了与Caffe一致：

def _built_net(self):
        """Construct the SSD net"""
        self.end_points = {}  # record the detection layers output
        self._images = tf.placeholder(tf.float32, shape=[None, self.ssd_params.img_shape[0],
                                                        self.ssd_params.img_shape[1], 3])
        with tf.variable_scope("ssd_300_vgg"):
            # original vgg layers
            # block 1
            net = conv2d(self._images, 64, 3, scope="conv1_1")
            net = conv2d(net, 64, 3, scope="conv1_2")
            self.end_points["block1"] = net
            net = max_pool2d(net, 2, scope="pool1")
            # block 2
            net = conv2d(net, 128, 3, scope="conv2_1")
            net = conv2d(net, 128, 3, scope="conv2_2")
            self.end_points["block2"] = net
            net = max_pool2d(net, 2, scope="pool2")
            # block 3
            net = conv2d(net, 256, 3, scope="conv3_1")
            net = conv2d(net, 256, 3, scope="conv3_2")
            net = conv2d(net, 256, 3, scope="conv3_3")
            self.end_points["block3"] = net
            net = max_pool2d(net, 2, scope="pool3")
            # block 4
            net = conv2d(net, 512, 3, scope="conv4_1")
            net = conv2d(net, 512, 3, scope="conv4_2")
            net = conv2d(net, 512, 3, scope="conv4_3")
            self.end_points["block4"] = net
            net = max_pool2d(net, 2, scope="pool4")
            # block 5
            net = conv2d(net, 512, 3, scope="conv5_1")
            net = conv2d(net, 512, 3, scope="conv5_2")
            net = conv2d(net, 512, 3, scope="conv5_3")
            self.end_points["block5"] = net
            print(net)
            net = max_pool2d(net, 3, stride=1, scope="pool5")
            print(net)

            # additional SSD layers
            # block 6: use dilate conv
            net = conv2d(net, 1024, 3, dilation_rate=6, scope="conv6")
            self.end_points["block6"] = net
            #net = dropout(net, is_training=self.is_training)
            # block 7
            net = conv2d(net, 1024, 1, scope="conv7")
            self.end_points["block7"] = net
            # block 8
            net = conv2d(net, 256, 1, scope="conv8_1x1")
            net = conv2d(pad2d(net, 1), 512, 3, stride=2, scope="conv8_3x3",
                         padding="valid")
            self.end_points["block8"] = net
            # block 9
            net = conv2d(net, 128, 1, scope="conv9_1x1")
            net = conv2d(pad2d(net, 1), 256, 3, stride=2, scope="conv9_3x3",
                         padding="valid")
            self.end_points["block9"] = net
            # block 10
            net = conv2d(net, 128, 1, scope="conv10_1x1")
            net = conv2d(net, 256, 3, scope="conv10_3x3", padding="valid")
            self.end_points["block10"] = net
            # block 11
            net = conv2d(net, 128, 1, scope="conv11_1x1")
            net = conv2d(net, 256, 3, scope="conv11_3x3", padding="valid")
            self.end_points["block11"] = net

            # class and location predictions
            predictions = []
            logits = []
            locations = []
            for i, layer in enumerate(self.ssd_params.feat_layers):
                cls, loc = ssd_multibox_layer(self.end_points[layer], self.ssd_params.num_classes,
                                              self.ssd_params.anchor_sizes[i],
                                              self.ssd_params.anchor_ratios[i],
                                              self.ssd_params.normalizations[i], scope=layer+"_box")
                predictions.append(tf.nn.softmax(cls))
                logits.append(cls)
                locations.append(loc)
            return predictions, logits, locations

对于特征图的检测，这里单独定义了一个组合层ssd_multibox_layer，其主要是对特征图进行两次卷积，分别得到类别置信度与边界框位置：

# multibox layer: get class and location predicitions from detection layer
    def ssd_multibox_layer(x, num_classes, sizes, ratios, normalization=-1, scope="multibox"):
        pre_shape = x.get_shape().as_list()[1:-1]
        pre_shape = [-1] + pre_shape
        with tf.variable_scope(scope):
            # l2 norm
            if normalization > 0:
                x = l2norm(x, normalization)
                print(x)
            # numbers of anchors
            n_anchors = len(sizes) + len(ratios)
            # location predictions
            loc_pred = conv2d(x, n_anchors*4, 3, activation=None, scope="conv_loc")
            loc_pred = tf.reshape(loc_pred, pre_shape + [n_anchors, 4])
            # class prediction
            cls_pred = conv2d(x, n_anchors*num_classes, 3, activation=None, scope="conv_cls")
            cls_pred = tf.reshape(cls_pred, pre_shape + [n_anchors, num_classes])
            return cls_pred, loc_pred

对于先验框，可以基于numpy生成，定义在ssd_anchors.py文件中，结合先验框与检测值，对边界框进行过滤与解码：

classes, scores, bboxes = self._bboxes_select(predictions, locations)

这里将得到过滤得到的边界框，其中classes, scores, bboxes分别表示类别，置信度值以及边界框位置。

基于训练好的权重文件在https://pan.baidu.com/s/1snhuTsT下载，这里对SSD进行测试：

ssd_net = SSD()
classes, scores, bboxes = ssd_net.detections()
images = ssd_net.images()

sess = tf.Session()
# Restore SSD model.
ckpt_filename = './ssd_checkpoints/ssd_vgg_300_weights.ckpt'
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(sess, ckpt_filename)

img = cv2.imread('./demo/dog.jpg')
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img_prepocessed = preprocess_image(img)   # 预处理图片，主要是归一化和resize
rclasses, rscores, rbboxes = sess.run([classes, scores, bboxes],
                                      feed_dict={images: img_prepocessed})
rclasses, rscores, rbboxes = process_bboxes(rclasses, rscores, rbboxes)  # 处理预测框，包括clip,sort,nms

plt_bboxes(img, rclasses, rscores, rbboxes)  # 绘制检测结果

详细的代码放在GitHub，https://github.com/xiaohu2015/DeepLearning_tutorials/tree/master/ObjectDetections/SSD上了，然后看一下一个自然图片的检测效果：

如果你想实现SSD的train过程，你可以参考附录里面的Caffe,TensorFlow以及Pytorch实现。

小结

SSD在Yolo的基础上主要改进了三点：多尺度特征图，利用卷积进行检测，设置先验框。这使得SSD在准确度上比Yolo更好，而且对于小目标检测效果也相对好一点。由于很多实现细节都包含在源码里面，文中有描述不准或者错误的地方在所难免，欢迎交流指正。