【从零开始学Mask RCNN】四，RPN锚框生成和Proposal生成

亦凉 2023-02-20 08:21 54阅读 0赞

# 1. Mask RCNN Anchor框 生成 #

Mask RCNN的锚框生成和SSD的锚框生成策略类似（SSD的锚框生成策略见：[【资源分享】从零开始学习SSD教程][SSD]） ，都遵循以下规则：

*  Anchor的中心点的个数等于特征图像素个数
 *  Anchor的生成是围绕中心点的
 *  Anchor框的坐标最终需要归一化到0-1之间，即相对输入图像的大小

我们知道Faster RCNN只是在一个特征图上铺设Anchor，而Mask RCNN引入了FPN之后使用了多层特征，这样和SSD类似都是在多个特征图上铺设Anchor，不过SSD的Anchor铺设策略相对较为复杂，每一个尺度的特征都有不同的Anchor策略设计，如下图所示：

![SSD的Anchor生成策略，来自本公众号的从零开始学SSD电子书][SSD_Anchor_SSD]

Mask RCNN相对于SSD的Anchor生成策略有以下不同之处：

*  SSD的不同特征图的Anchor长宽比设置不同，而Mask RCNN的不同特征图的Anchor比例则完全相同
 *  Mask RCNN每一层设置固定个数的Anchor数（3个）
 *  SSD中心点为每个像素偏移0.5步长；Mask RCNN中心点直接选为像素位置

另外，Anchor的基本生成方式和SSD一样，即Anchor框的长度等于 h h h乘 a n c h o r − r a t i o s \\sqrt\{anchor-ratios\} anchor−ratios，Anchor框的宽度等于 w w w除以 a n c h o r − r a t i o s \\sqrt\{anchor-ratios\} anchor−ratios，其中 h , w h,w h,w的初始值是给定的参考尺寸。具体设置如下：

anchor_sizes=[(21., 45.),
                  (45., 99.),
                  (99., 153.),
                  (153., 207.),
                  (207., 261.),
                  (261., 315.)]
    anchor_ratios=[[2, .5],
                   [2, .5, 3, 1./3],
                   [2, .5, 3, 1./3],
                   [2, .5, 3, 1./3],
                   [2, .5],
                   [2, .5]]

在`mrcnn/config,py`中相关的设置如下：

self.config.BACKBONE_STRIDES  = [4, 8, 16, 32, 64]      # 特征层的下采样倍数，中心点计算使用
    self.config.RPN_ANCHOR_RATIOS = [0.5, 1, 2]             # 特征层锚框生成参数
    self.config.RPN_ANCHOR_SCALES = [32, 64, 128, 256, 512] # 特征层锚框感受野

Anchor框的生成函数为`model.py`中的`get_anchors`函数，参数为`image_shape`，只需要保证里面有`[h,w]`即可，也可以是`[h,w,c]`，代码如下：

# 锚框生成函数
        def get_anchors(self, image_shape):
            """Returns anchor pyramid for the given image size."""
            backbone_shapes = compute_backbone_shapes(self.config, image_shape)
            # 如果图片的尺寸相同就用缓存的Anchor信息回复
            # Cache anchors and reuse if image shape is the same
            if not hasattr(self, "_anchor_cache"):
                self._anchor_cache = {
        }
            # 如果对应尺寸图像的Anchors区域没有被缓存，那么就产生Anchors
            if not tuple(image_shape) in self._anchor_cache:
                # 产生Anchor区域
                a = utils.generate_pyramid_anchors(
                    self.config.RPN_ANCHOR_SCALES,
                    self.config.RPN_ANCHOR_RATIOS,
                    backbone_shapes,
                    self.config.BACKBONE_STRIDES,
                    self.config.RPN_ANCHOR_STRIDE)
                # Keep a copy of the latest anchors in pixel coordinates because
                # it's used in inspect_model notebooks.
                # TODO: Remove this after the notebook are refactored to not use it
                self.anchors = a
                # Normalize coordinates
                self._anchor_cache[tuple(image_shape)] = utils.norm_boxes(a, image_shape[:2])
            return self._anchor_cache[tuple(image_shape)]

这个函数首先调用`compute_backbone_shapes`计算各个特征层的`shape`：

def compute_backbone_shapes(config, image_shape):
        """ 计算骨干网络的每个阶段特征图的长宽
        Returns:
            [N, (height, width)]. N代表阶段数
        """
        # 可调用返回 True，否则返回 False。
        if callable(config.BACKBONE):
            return config.COMPUTE_BACKBONE_SHAPE(image_shape)
    
        # 当前只支持ResNet
        assert config.BACKBONE in ["resnet50", "resnet101"]
        return np.array(
            [[int(math.ceil(image_shape[0] / stride)),
                int(math.ceil(image_shape[1] / stride))]
                for stride in config.BACKBONE_STRIDES])

然后再调用`utils.generate_pyramid_anchors`函数生成所有的Anchor框，这个函数在`mrcnn/utils.py`中，代码如下：

def generate_pyramid_anchors(scales, ratios, feature_shapes, feature_strides,
                                 anchor_stride):
        """在特征金字塔的不同层级产生anchors，每个尺度都
        和金字塔的层级有关，但是长宽比例都是相同的
    
        返回:
        anchors: [N, (y1, x1, y2, x2)]. 所有产生的Anchors放在一个数组中.
        按照scales给定的顺序进行排序. 所以scale[0]是第一个,然后scale[1]第二个,等等.
        """
        # Anchors
        # [anchor_count, (y1, x1, y2, x2)]
        anchors = []
        for i in range(len(scales)):
            anchors.append(generate_anchors(scales[i], ratios, feature_shapes[i],
                                            feature_strides[i], anchor_stride))
        return np.concatenate(anchors, axis=0)

这里又调用了`utils.generate_anchors`函数来生成每一个特征层的Anchors，这个函数大量应用了np.meshgrid函数，这个函数是用来生成网格坐标，一个很好的说明图如下：

![np.meshgrid解释，转载自https://www.cnblogs.com/hellcat/p/9855017.html][np.meshgrid_https_www.cnblogs.com_hellcat_p_9855017.html]

**虽然xy双坐标比较常用，但实际上输入可以是任意多的数组，输出数组数目等于输出数组数目，且彼此间shape一致。**

**如果输入数组不是一维的，会拉伸为1维进行计算。**

**输出维度是：\[len(x2), len(x1), len(x3)……\]**

例子如下：

![np.meshgrid 举例][np.meshgrid]

`utils.generate_anchors`函数解释如下：

############################################################
    #  Anchors
    ############################################################
    
    def generate_anchors(scales, ratios, shape, feature_stride, anchor_stride):
        """
        scales: 1D array of anchor sizes in pixels. Example: [32, 64, 128]
        ratios: 1D array of anchor ratios of width/height. Example: [0.5, 1, 2]
        shape: [height, width] spatial shape of the feature map over which
                to generate anchors.
        feature_stride: Stride of the feature map relative to the image in pixels.
        anchor_stride: Stride of anchors on the feature map. For example, if the
            value is 2 then generate anchors for every other feature map pixel.
        """
        # Get all combinations of scales and ratios
        scales, ratios = np.meshgrid(np.array(scales), np.array(ratios))
        scales = scales.flatten()
        ratios = ratios.flatten()
    
        # Enumerate heights and widths from scales and ratios
        heights = scales / np.sqrt(ratios)
        widths = scales * np.sqrt(ratios)
    
        # Enumerate shifts in feature space
        shifts_y = np.arange(0, shape[0], anchor_stride) * feature_stride
        shifts_x = np.arange(0, shape[1], anchor_stride) * feature_stride
        shifts_x, shifts_y = np.meshgrid(shifts_x, shifts_y)
    
        # Enumerate combinations of shifts, widths, and heights
        box_widths, box_centers_x = np.meshgrid(widths, shifts_x)
        box_heights, box_centers_y = np.meshgrid(heights, shifts_y)
    
        # Reshape to get a list of (y, x) and a list of (h, w)
        box_centers = np.stack(
            [box_centers_y, box_centers_x], axis=2).reshape([-1, 2])
        box_sizes = np.stack([box_heights, box_widths], axis=2).reshape([-1, 2])
    
        # Convert to corner coordinates (y1, x1, y2, x2)
        # 框体信息是相对于原图的, [N, (y1, x1, y2, x2)]
        boxes = np.concatenate([box_centers - 0.5 * box_sizes,
                                box_centers + 0.5 * box_sizes], axis=1)
        return boxes

最后再回到`model.py`中的`get_anchors`函数，再调用`utils.norm_boxes`进行Anchor坐标的归一化，即`utils.norm_boxes`函数，代码如下：

# 坐标归一化
    def norm_boxes(boxes, shape):
        """Converts boxes from pixel coordinates to normalized coordinates.
        boxes: [N, (y1, x1, y2, x2)] in pixel coordinates
        shape: [..., (height, width)] in pixels
    
        Note: In pixel coordinates (y2, x2) is outside the box. But in normalized
        coordinates it's inside the box.
    
        Returns:
            [N, (y1, x1, y2, x2)] in normalized coordinates
        """
        h, w = shape
        scale = np.array([h - 1, w - 1, h - 1, w - 1])
        shift = np.array([0, 0, 1, 1])
        return np.divide((boxes - shift), scale).astype(np.float32)

# 2. RPN锚框生成 #

在第三个推文中我们获得了rpn的每个特征图信息即：

rpn_feature_maps = [P2, P3, P4, P5, P6]

然后在每个特征图上产生Anchor框，不过这里是分训练还是测试，训练的话使用刚刚介绍的方法来产生Anchor，如果是测试的话则使用定义好的Anchor，代码如下：

# Anchors
            if mode == "training":
                anchors = self.get_anchors(config.IMAGE_SHAPE)
                # Duplicate across the batch dimension because Keras requires it
                # TODO: can this be optimized to avoid duplicating the anchors?
                anchors = np.broadcast_to(anchors, (config.BATCH_SIZE,) + anchors.shape)
                # A hack to get around Keras's bad support for constants
                anchors = KL.Lambda(lambda x: tf.Variable(anchors), name="anchors")(input_image)
            else:
                anchors = input_anchors

然后基于RPN的各个特征图和定义好的Anchor框信息我们生成每个Anchor框的**前景/背景得分信息**及每个Anchor框的**坐标修正信息。** 接前文主函数，我们初始化rpn model class的对象，并应用于各层特征：

# RPN Model, 返回的是keras的Module对象, 注意keras中的Module对象是可call的
            rpn = build_rpn_model(config.RPN_ANCHOR_STRIDE,
                                  len(config.RPN_ANCHOR_RATIOS), config.TOP_DOWN_PYRAMID_SIZE)
            # Loop through pyramid layers
            layer_outputs = []  # list of lists
            # 保存各pyramid特征经过RPN之后的结果
            for p in rpn_feature_maps:
                layer_outputs.append(rpn([p]))

其中`build_rpn_model`的代码实现如下：

def build_rpn_model(anchor_stride, anchors_per_location, depth):
        """建立RPN网络的Keras模型。它包装了RPN图，可以共享权重。
    
        anchors_per_location: 特征图中每个像素的Anchor个数。
        anchor_stride: 控制Anchor框的密度，通常为1或者2。
        depth: 骨干网络特征图的深度。
    
        返回一个Keras的Model对象。当被调用时，model的输出是 :
        rpn_class_logits: [batch, H * W * anchors_per_location, 2] Anchor classifier logits (before softmax)
        rpn_probs: [batch, H * W * anchors_per_location, 2] Anchor classifier probabilities.
        rpn_bbox: [batch, H * W * anchors_per_location, (dy, dx, log(dh), log(dw))] Deltas to be
                    applied to anchors.
        """
        input_feature_map = KL.Input(shape=[None, None, depth],
                                     name="input_rpn_feature_map")
        outputs = rpn_graph(input_feature_map, anchors_per_location, anchor_stride)
        return KM.Model([input_feature_map], outputs, name="rpn_model")

它接收一个输入特征图，然后调用`rpn_graph`这个函数完成RPN模块的实现，这个函数的代码解析如下：

############################################################
    #  Region Proposal Network (RPN)
    ############################################################
    
    def rpn_graph(feature_map, anchors_per_location, anchor_stride):
        """建立RPN网络的计算图。
    
        feature_map: 骨干特征 [batch, height, width, depth]
        anchors_per_location: 特征图中每个像素的Anchor个数
        anchor_stride: 控制Anchors的密度，比如设置成1或者2
    
        Returns:
            rpn_class_logits: [batch, H * W * anchors_per_location, 2] Anchor classifier logits (before softmax)
            rpn_probs: [batch, H * W * anchors_per_location, 2] Anchor classifier probabilities.
            rpn_bbox: [batch, H * W * anchors_per_location, (dy, dx, log(dh), log(dw))] Deltas to be
                      applied to anchors.
        """
        # TODO: check if stride of 2 causes alignment issues if the feature map
        # is not even.
        # Shared convolutional base of the RPN
        shared = KL.Conv2D(512, (3, 3), padding='same', activation='relu',
                           strides=anchor_stride,
                           name='rpn_conv_shared')(feature_map)
    
        # Anchor的分数
        # Anchor Score. [batch, height, width, anchors per location * 2].
        x = KL.Conv2D(2 * anchors_per_location, (1, 1), padding='valid',
                      activation='linear', name='rpn_class_raw')(shared)
    
        # Reshape to [batch, anchors, 2]
        rpn_class_logits = KL.Lambda(
            lambda t: tf.reshape(t, [tf.shape(t)[0], -1, 2]))(x)
    
        # Softmax on last dimension of BG/FG.
        rpn_probs = KL.Activation(
            "softmax", name="rpn_class_xxx")(rpn_class_logits)
    
        # Bounding box refinement. [batch, H, W, anchors per location * depth]
        # where depth is [x, y, log(w), log(h)]
        x = KL.Conv2D(anchors_per_location * 4, (1, 1), padding="valid",
                      activation='linear', name='rpn_bbox_pred')(shared)
    
        # Reshape to [batch, anchors, 4]
        rpn_bbox = KL.Lambda(lambda t: tf.reshape(t, [tf.shape(t)[0], -1, 4]))(x)
    
        return [rpn_class_logits, rpn_probs, rpn_bbox]

紧接着前文的主函数，我们需要将获取的list形式的各层Anchor框进行拼接和重组，代码如下：

# Concatenate layer outputs
            # Convert from list of lists of level outputs to list of lists
            # of outputs across levels.
            # e.g. [[a1, b1, c1], [a2, b2, c2]] => [[a1, a2], [b1, b2], [c1, c2]]
            output_names = ["rpn_class_logits", "rpn_class", "rpn_bbox"]
            outputs = list(zip(*layer_outputs))
            outputs = [KL.Concatenate(axis=1, name=n)(list(o))
                       for o, n in zip(outputs, output_names)]
            # [batch, num_anchors, 2/4]
            # 其中num_anchors指的是全部特征层上的anchors总数
            rpn_class_logits, rpn_class, rpn_bbox = outputs

这段代码的目的就是将原始的 \[ ( l o g i t s 2 , c l a s s 2 , b b o x 2 ) , ( l o g i t s 3 , c l a s s 3 , b b o x 3 ) , … … 6 \] \[(logits2, class2, bbox2), (logits3, class3, bbox3), ……6\] \[(logits2,class2,bbox2),(logits3,class3,bbox3),……6\]，这种排布的返回值转换为 \[ ( l o g i t s 2 , l o g i t s 3 … … 6 ) , ( c l a s s 2 , c l a s s 3 … … 6 ) , ( b b o x 2 , b b o x 3 … … 6 ) \] \[(logits2,logits3……6), (class2,class3……6), (bbox2,bbox3……6)\] \[(logits2,logits3……6),(class2,class3……6),(bbox2,bbox3……6)\]这种形式（通过zip函数转换），然后将每个小list中的tensor按照**第一维度（即anchors维度）拼接**，得到三个tensor，每个tensor表明batch中图片对应 5 5 5个特征层的全部anchors的分类回归信息，即：\[batch, anchors, 2分类结果 或者 (dy, dx, log(dh), log(dw))\]。

# 3. Proposal生成 #

## 3.1 初始化Proposal Layer ##

现在我们已经获取了所有的Anchor框的信息，这一步的目的是从中挑选指定个数的可能包含目标的Anchor框作为候选区域，即在上一步的二分类中的前景类别得分更高的框，同时由于Anchor框是比较密集并且重叠严重，所以这里还会执行一个NMS来去重。这部分的代码如下，其中，proposal\_count是一个整数，用于指定生成proposal数目，不足时会生成坐标为\[0,0,0,0\]的空值进行补全：

# 产生候选框
    # Proposals are [batch, N, (y1, x1, y2, x2)] in normalized coordinates
    # and zero padded.
    # POST_NMS_ROIS_INFERENCE = 1000
    # POST_NMS_ROIS_TRAINING = 2000
    proposal_count = config.POST_NMS_ROIS_TRAINING if mode == "training"\
        else config.POST_NMS_ROIS_INFERENCE
    # [IMAGES_PER_GPU, num_rois, (y1, x1, y2, x2)]
    # IMAGES_PER_GPU取代了batch，之后说的batch都是IMAGES_PER_GPU
    rpn_rois = ProposalLayer(
        proposal_count=proposal_count, 
        nms_threshold=config.RPN_NMS_THRESHOLD, #0.7
        name="ROI",
        config=config)([rpn_class, rpn_bbox, anchors])

接下来我们跟进`ProposalLayer`，在初始部分获取`[rpn_class, rpn_bbox, anchors]`三个张量作为参数，然后获取前景分数最大的`PRE_NMS_LIMITPRE_NMS_LIMIT`这么多个候选框，这部分的代码如下：

# 在初始部分获取[rpn_class, rpn_bbox, anchors]三个张量作为参数，
    class ProposalLayer(KE.Layer):
        """接收Anchor框的分数并选择一个子集传入到第二个阶段. 
        根据Anchor框分数和非最大抑制进行过滤，以消除一些重叠框.
        并且将bbox调整值应用于Anchor上。
        Inputs:
            rpn_probs: [batch, num_anchors, (bg prob, fg prob)]
            rpn_bbox: [batch, num_anchors, (dy, dx, log(dh), log(dw))]
            anchors: [batch, num_anchors, (y1, x1, y2, x2)] anchors in normalized coordinates
    
        Returns:
            Proposals in normalized coordinates [batch, rois, (y1, x1, y2, x2)]
        """
    
        def __init__(self, proposal_count, nms_threshold, config=None, **kwargs):
            super(ProposalLayer, self).__init__(**kwargs)
            self.config = config
            self.proposal_count = proposal_count
            self.nms_threshold = nms_threshold
    
        def call(self, inputs):
            # Box Scores. Use the foreground class confidence. [Batch, num_rois, 1]
            # 获得Anchor框的前景分数
            scores = inputs[0][:, :, 1]
            # Box deltas [batch, num_rois, 4]
            # Box deltas. 记录坐标修正信息：(dy, dx, log(dh), log(dw)). [batch, num_rois, 4]
            deltas = inputs[1]
            deltas = deltas * np.reshape(self.config.RPN_BBOX_STD_DEV, [1, 1, 4]) # [ 0.1  0.1  0.2  0.2]
            # Anchors 记录坐标信息：(y1, x1, y2, x2). [batch, num_rois, 4]
            anchors = inputs[2]
    
            # Improve performance by trimming to top anchors by score
            # and doing the rest on the smaller subset.
            # 然后我们获取前景得分最大的n个候选框，
            pre_nms_limit = tf.minimum(self.config.PRE_NMS_LIMIT, tf.shape(anchors)[1])
            # 输入矩阵时输出每一行的top k. [batch, top_k]
            ix = tf.nn.top_k(scores, pre_nms_limit, sorted=True,
                             name="top_anchors").indices

## 3.2 获取Top K Anchor框 ##

注意上面的代码结尾只是获取了要保留的Anchor框的索引，接下来还需要对三个输入分别进行提取，接上面代码：

# 提取top k锚框，同时对三个输入进行了提取
            # batch_slice函数：
            # #   将batch特征拆分为单张
            # #   然后提取指定的张数
            # #   使用单张特征处理函数处理，并合并（此时返回的第一维不是输入时的batch，而是上步指定的张数）
            scores = utils.batch_slice([scores, ix], lambda x, y: tf.gather(x, y),
                                       self.config.IMAGES_PER_GPU)
            deltas = utils.batch_slice([deltas, ix], lambda x, y: tf.gather(x, y),
                                       self.config.IMAGES_PER_GPU)
            pre_nms_anchors = utils.batch_slice([anchors, ix], lambda a, x: tf.gather(a, x),
                                        self.config.IMAGES_PER_GPU,
                                        names=["pre_nms_anchors"])

这里主要使用了一个`utils.batch_slice`函数，这个函数把只支持`batch=1`的函数进行了扩展，由于`tf.gather`函数只能进行一维数组的切片，而`scores`的维度现在是`[batch,num_rois]，`相对的`ix`也是二维即`[batch,top_k]`，因此这里要将两者应用`tf.gather`函数切片后将结果重新拼接，代码如下（`mrnn/utils.py`）：

# ## Batch Slicing
    # Some custom layers support a batch size of 1 only, and require a lot of work
    # to support batches greater than 1. This function slices an input tensor
    # across the batch dimension and feeds batches of size 1. Effectively,
    # an easy way to support batches > 1 quickly with little code modification.
    # In the long run, it's more efficient to modify the code to support large
    # batches and getting rid of this function. Consider this a temporary solution
    def batch_slice(inputs, graph_fn, batch_size, names=None):
        """Splits inputs into slices and feeds each slice to a copy of the given
        computation graph and then combines the results. It allows you to run a
        graph on a batch of inputs even if the graph is written to support one
        instance only.
     
        inputs: list of tensors. All must have the same first dimension length
        graph_fn: A function that returns a TF tensor that's part of a graph.
        batch_size: number of slices to divide the data into.
        names: If provided, assigns names to the resulting tensors.
        """
        if not isinstance(inputs, list):
            inputs = [inputs]
     
        outputs = []
        for i in range(batch_size):
            inputs_slice = [x[i] for x in inputs]
            output_slice = graph_fn(*inputs_slice)
            if not isinstance(output_slice, (tuple, list)):
                output_slice = [output_slice]
            outputs.append(output_slice)
     
        # 使用tf.while_loop实现循环体代码如下：
        # import tensorflow as tf
        # i = 0
        # outputs = []
        #
        # def cond(index):
        #     return index < batch_size  # 返回bool值
        #
        # def body(index):
        #     index += 1
        #     inputs_slice = [x[i] for x in inputs]
        #     output_slice = graph_fn(*inputs_slice)
        #     if not isinstance(output_slice, (tuple, list)):
        #         output_slice = [output_slice]
        #     outputs.append(output_slice)
        #     return index  # 返回cond需要的判断参数进行下一次判断
        #
        # tf.while_loop(cond, body, [i])
     
        # Change outputs from a list of slices where each is
        # a list of outputs to a list of outputs and each has
        # a list of slices
        # 下面示意中假设每次graph_fn返回两个tensor
        # [[tensor11, tensor12], [tensor21, tensor22], ……]
        # ——> [(tensor11, tensor21, ……), (tensor12, tensor22, ……)]  zip返回的是多个tuple
        outputs = list(zip(*outputs))
     
        if names is None:
            names = [None] * len(outputs)
     
        # 一般来讲就是batch维度合并回去（上面的for循环实际是将batch拆分了）
        result = [tf.stack(o, axis=0, name=n)
                  for o, n in zip(outputs, names)]
        if len(result) == 1:
            result = result[0]
     
        return result

## 3.3 Anchor框坐标修正 ##

我们在第二节中利用RPN网络获取了全部Anchor框的坐标回归结果，即`rpn_bbox: [batch, anchors, (dy, dx, log(dh), log(dw))]`，然后我们在上一个小节（3.2）获得了top k Anchor框的坐标信息以及top k的Anchor框回归偏移量，现在将其合并，即利用RPN的回归结果修正top k Anchor框的坐标。代码如下：

# Anchor框修正
            # Apply deltas to anchors to get refined anchors.
            # 现在的N代表Top K的K
            # [batch, N, (y1, x1, y2, x2)]
            boxes = utils.batch_slice([pre_nms_anchors, deltas],
                                      lambda x, y: apply_box_deltas_graph(x, y),
                                      self.config.IMAGES_PER_GPU,
                                      names=["refined_anchors"])

可以看到这里主要调用了`apply_box_deltas_graph`这个函数结合上面介绍的`batch_slice`函数完成了RPN回归结果修正，我们来看一下`apply_box_deltas_graph`这个函数的代码实现，如下：

############################################################
    #  Proposal Layer
    ############################################################
    
    def apply_box_deltas_graph(boxes, deltas):
        """Applies the given deltas to the given boxes.
        boxes: [N, (y1, x1, y2, x2)] boxes to update
        deltas: [N, (dy, dx, log(dh), log(dw))] refinements to apply
        """
        # Convert to y, x, h, w
        height = boxes[:, 2] - boxes[:, 0]
        width = boxes[:, 3] - boxes[:, 1]
        center_y = boxes[:, 0] + 0.5 * height
        center_x = boxes[:, 1] + 0.5 * width
        # Apply deltas
        center_y += deltas[:, 0] * height
        center_x += deltas[:, 1] * width
        height *= tf.exp(deltas[:, 2])
        width *= tf.exp(deltas[:, 3])
        # Convert back to y1, x1, y2, x2
        y1 = center_y - 0.5 * height
        x1 = center_x - 0.5 * width
        y2 = y1 + height
        x2 = x1 + width
        result = tf.stack([y1, x1, y2, x2], axis=1, name="apply_box_deltas_out")
        return result

从这个代码可以推出，RPN回归出的4个坐标值的意义土如下：

d y = ( y n − y o ) / h o dy=(y\_n-y\_o)/h\_o dy=(yn−yo)/ho

d x = ( x n − x o ) / w o dx=(x\_n-x\_o)/w\_o dx=(xn−xo)/wo

d h = h n / h o dh=h\_n/h\_o dh=hn/ho

d w = w n / w o dw=w\_n/w\_o dw=wn/wo

此外，这里的Anchor框坐标实际上是在一个归一化的图上的（即所有的Anchor框位于一个长宽均为1的虚拟画布上），上一步执行了修正之后不再能够保证这一点，因此这里还需要切除Anchor框越界的部分，即只保留Anchor框和 \[ 0 , 0 , 1 , 1 \] \[0,0,1,1\] \[0,0,1,1\]画布的交集。代码如下：

# Clip to image boundaries. Since we're in normalized coordinates,
            # clip to 0..1 range. [batch, N, (y1, x1, y2, x2)]
            window = np.array([0, 0, 1, 1], dtype=np.float32)
            boxes = utils.batch_slice(boxes,
                                      lambda x: clip_boxes_graph(x, window),
                                      self.config.IMAGES_PER_GPU,
                                      names=["refined_anchors_clipped"])

其中`clip_boxes_graph`这个函数实现了保留交集的功能，代码如下：

# 保留2个框的交集部分
    def clip_boxes_graph(boxes, window):
        """
        boxes: [N, (y1, x1, y2, x2)]
        window: [4] in the form y1, x1, y2, x2
        """
        # Split
        wy1, wx1, wy2, wx2 = tf.split(window, 4)
        y1, x1, y2, x2 = tf.split(boxes, 4, axis=1)
        # Clip
        y1 = tf.maximum(tf.minimum(y1, wy2), wy1)
        x1 = tf.maximum(tf.minimum(x1, wx2), wx1)
        y2 = tf.maximum(tf.minimum(y2, wy2), wy1)
        x2 = tf.maximum(tf.minimum(x2, wx2), wx1)
        clipped = tf.concat([y1, x1, y2, x2], axis=1, name="clipped_boxes")
        clipped.set_shape((clipped.shape[0], 4))
        return clipped

## 3.4 非极大值抑制(NMS) ##

最后再进行非极大值抑制，以保证不会出现重叠度非常大的候选框，代码如下：

# Non-max suppression
            def nms(boxes, scores):
                # 非极大值抑制子函数
                # :param boxes: [top_k, (y1, x1, y2, x2)]
                # :param scores: [top_k]
                # :return: 
                # self.proposal_count为最大返回数目
                indices = tf.image.non_max_suppression(
                    boxes, scores, self.proposal_count,
                    self.nms_threshold, name="rpn_non_max_suppression")
                proposals = tf.gather(boxes, indices)
                # Pad if needed, 一旦返回数目不足, 填充(0,0,0,0)直到数目达标
                padding = tf.maximum(self.proposal_count - tf.shape(proposals)[0], 0)
                # 在后面添加全0行
                proposals = tf.pad(proposals, [(0, padding), (0, 0)])
                return proposals
            proposals = utils.batch_slice([boxes, scores], nms,
                                          self.config.IMAGES_PER_GPU)
            # [IMAGES_PER_GPU, proposal_count, (y1, x1, y2, x2)]
            return proposals

# 4. 小节 #

到这里，我们就将Mask RCNN的Anchor框生成和Proposal生成讲完了，下一节我将讲解获取RPN区域后Mask RCNN是如何处理这些候选框的(ROI Align)。

# 5. 参考 #

*  https://www.cnblogs.com/hellcat/p/9811301.html

--------------------

欢迎关注GiantPandaCV, 在这里你将看到独家的深度学习分享，坚持原创，每天分享我们学习到的新鲜知识。( • ̀ω•́ )✧

有对文章相关的问题，或者想要加入交流群，欢迎添加BBuf微信：

![二维码][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2p1c3Rfc29ydA_size_16_color_FFFFFF_t_70]  
为了方便读者获取资料以及我们公众号的作者发布一些Github工程的更新，我们成立了一个QQ群，二维码如下，感兴趣可以加入。

![公众号QQ交流群][QQ]

[SSD]: https://mp.weixin.qq.com/s/BAUNTWRecCB-P2DJY3CrwA
[SSD_Anchor_SSD]: https://img-blog.csdnimg.cn/20200625173826398.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2p1c3Rfc29ydA==,size_16,color_FFFFFF,t_70
[np.meshgrid_https_www.cnblogs.com_hellcat_p_9855017.html]: https://img-blog.csdnimg.cn/20200625233214937.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2p1c3Rfc29ydA==,size_16,color_FFFFFF,t_70
[np.meshgrid]: https://img-blog.csdnimg.cn/20200625233501783.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2p1c3Rfc29ydA_size_16_color_FFFFFF_t_70]: https://img-blog.csdnimg.cn/20200110234905879.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2p1c3Rfc29ydA==,size_16,color_FFFFFF,t_70
[QQ]: https://img-blog.csdnimg.cn/20200517190745584.png#pic_center