有了前面Faster R-CNN的基礎，RFCN就比較容易了。

"""object_detection/meta_architectures/rfcn_meta_arch.py
The R-FCN meta architecture is similar to Faster R-CNN and only differs in the
second stage. Hence this class inherits FasterRCNNMetaArch and overrides only
the `_predict_second_stage` method
"""

改動比較大的地方如下：
    box_predictions = self._rfcn_box_predictor.predict(
        box_classifier_features,
        num_predictions_per_location=1
 
,
        scope=self.second_stage_box_predictor_scope,
        proposal_boxes=proposal_boxes_normalized)
只改動了這麼一點的原因是程式碼實現和原始paper不是完全一致的，見作者的paper：3.4. Training and hyperparameter tuning。
另外注意，在faster r-cnn中，是先經過ROI poooling，然後進入第二階段特徵提取器。而在rfcn中，是先進行第二階段特徵提取，然後進入RfcnBoxPredictor。這正是frcn改進的地方，即將卷積操作儘量在faster r-cnn的roi之間共享，使得rfcn得到的（更高層面的）roi需要單獨經過的預測層更少，大大提高了效率。
 

"""先回顧一下faster rcnn: ROI pooling"""
實現程式碼在
  def _compute_second_stage_input_feature_maps(self, features_to_crop,
                                               proposal_boxes_normalized)
ROI就是features_to_crop（第一階段特徵提取器得到的feature map）上的一塊crop，是依據proposal_boxed擷取的。ROI pooling是spp layer的特殊情況，就是通過自適應大小的卷積將特徵圖對映到固定尺寸，以便進入fc層。tf程式碼作者採用的實現不是這種，而是將不同尺寸的ROI先變成統一的大小，然後就不需要進行ROI pooling了(見_compute_second_stage_input_feature_maps)，換言之，需要ROI POOLING的原因就是ROI的尺寸是不同的（RFCN PAPER的理解角度：this
 
 region-specific operation breaks down translation invariance, and the post-RoI convolutional layers are no longer translation-invariant when evaluated across different regions.）。

"""rfcn的基本想法上面分析對比過了，仔細看看rfcn paper中的三張大圖就懂什麼是position-sensitive了。為什麼需要用position-sensitive RoI pooling呢？paper中說是為了將translation variance包含進FCN中。但是從position-sensitive score maps的生成來看，似乎類似顯示地引入了一種“先驗結構”，如上圖的fig 3，怎麼就能確保這九個score map剛好就是對應九個位置呢（For example, the “top-center-sensitive” score map exhibits high scores roughly near the top-center position of an object.）？
position-sensitive score maps
position-sensitive RoI pooling layer
"""

在Tensorflow object detection API中的實現有所不同：
class RfcnBoxPredictor(BoxPredictor)
Applies a position sensitve ROI pooling on position sensitive feature maps to
  predict classes and refined locations

ops.position_sensitive_crop_regions
  """Position-sensitive crop and pool rectangular regions from a feature grid.
  The output crops are split into `spatial_bins_y` vertical bins
  and `spatial_bins_x` horizontal bins. For each intersection of a vertical
  and a horizontal bin the output values are gathered by performing
  `tf.image.crop_and_resize` (bilinear resampling) on a a separate subset of
  channels of the image. This reduces `depth` by a factor of
  `(spatial_bins_y * spatial_bins_x)`.
  When global_pool is True, this function implements a differentiable version
  of position-sensitive RoI pooling used in
  [R-FCN detection system](https://arxiv.org/abs/1605.06409).
  When global_pool is False, this function implements a differentiable version
  of position-sensitive assembling operation used in
  [instance FCN](https://arxiv.org/abs/1603.08678)."""

'''
目測是將第二階段提取器得到的feature用1*1卷積增加到k*k*(c+1)個channel，就得到了position-sensitive score maps:
      location_feature_map_depth = (self._num_spatial_bins[0] *
                                    self._num_spatial_bins[1] *
                                    self.num_classes *
                                    self._box_code_size)
      location_feature_map = slim.conv2d(net, location_feature_map_depth,
                                         [1, 1], activation_fn=None,
                                         scope='refined_locations')
然後就用了和faster rcnn程式碼類似的方法，進行position-sensitive RoI pooling。position-sensitive score maps這個地方確實類似一種“先驗結構”。經常在paper中看到這種想法大膽的end-to-end先驗結構。論文中說With end-to-end training, this RoI layer shepherds the last convolutional layer to learn specialized position-sensitive score maps，視覺化的結果似乎確實有這種現象，有趣。
'''