首先，通过darknet53骨干网络得到大中小三种比例的特征图，图片来自https://zhuanlan.zhihu.com/p/50595699

self.conv_lbbox, self.conv_mbbox, self.conv_sbbox = self.__build_nework(input_data)

以小anchor为例，即下采样3次，feature map大小变为原来的1/8，FPN部分参考https://blog.csdn.net/ysh1026/article/details/113780273

self.pred_sbbox = self.decode(self.conv_sbbox, self.anchors[0], self.strides[0])

    def decode(self, conv_output, anchors, stride):
        """
        return tensor of shape [batch_size, output_size, output_size, anchor_per_scale, 5 + num_classes]
               contains (x, y, w, h, score, probability)
        """

        conv_shape       = tf.shape(conv_output)
        batch_size       = conv_shape[0]
        output_size      = conv_shape[1]
        anchor_per_scale = len(anchors)

        conv_output = tf.reshape(conv_output, (batch_size, output_size, output_size, anchor_per_scale, 5 + self.num_class))

        #conv_raw_dxdy是相对于方格子左上角的坐标，取值在0~1
        conv_raw_dxdy = conv_output[:, :, :, :, 0:2]
        # 对bounding box prior的放缩比例
        conv_raw_dwdh = conv_output[:, :, :, :, 2:4]
        # 判断是否是前景，相当于faster rcnn中的RPN层，取值0或1
        conv_raw_conf = conv_output[:, :, :, :, 4:5]
        #各个类的概率
        conv_raw_prob = conv_output[:, :, :, :, 5: ]

        y = tf.tile(tf.range(output_size, dtype=tf.int32)[:, tf.newaxis], [1, output_size])
        x = tf.tile(tf.range(output_size, dtype=tf.int32)[tf.newaxis, :], [output_size, 1])

        #得到每个方格子左上角的位置
        xy_grid = tf.concat([x[:, :, tf.newaxis], y[:, :, tf.newaxis]], axis=-1)
        xy_grid = tf.tile(xy_grid[tf.newaxis, :, :, tf.newaxis, :], [batch_size, 1, 1, anchor_per_scale, 1])
        xy_grid = tf.cast(xy_grid, tf.float32)

        pred_xy = (tf.sigmoid(conv_raw_dxdy) + xy_grid) * stride
        # anchors是bounding box prior，先验框
        pred_wh = (tf.exp(conv_raw_dwdh) * anchors) * stride
        # pred_xywh是在原图片上面画框
        pred_xywh = tf.concat([pred_xy, pred_wh], axis=-1)

        #归一化
        pred_conf = tf.sigmoid(conv_raw_conf)
        pred_prob = tf.sigmoid(conv_raw_prob)

        return tf.concat([pred_xywh, pred_conf, pred_prob], axis=-1)

代码来自https://github.com/YunYang1994/tensorflow-yolov3

——原来如此简单——