关于RefineDet中conf和loc输出通道和priorbox产生数量的对应关系

首先需要明白原版SSD中对于conf和loc输出通道和priorbox产生数量之间的关系。如需了解请查看博主两篇博文：

https://blog.csdn.net/xunan003/article/details/79249056

https://blog.csdn.net/xunan003/article/details/79186162

与SSD不同，RefineDet修改了multibox_loss_layer.cpp，即修改了多盒损失，将loss分为arm loss和odm loss，这涉及到通道对应问题。需要注意的是RefineDet中提取anchor的四个卷积层分别为conv4_3、conv5_3、fc7和conv6_2，这四层分别对应四个mbox_conf、mbox_loc和mbox_priorbox层。在arm loss计算阶段，mbox_loc层的output_num计算方式与SSD相同，mbox_priorbox层产生的每个默认框产生4个坐标值并对应mbox_loc层的4个输出通道，即output_num=mbox_priorbox数量 × 4。而mbox_conf层的output_num的计算方式与SSD不同，在RefineDet中mbox_conf用于计算每个priorbox是否预测类别的分数，即每个默认框产生两个分数（类别分数与非类别分数），该层用于粗略的估计box的类别，对应论文中的arm损失，所以mbox_conf层的output_num=mbox_priorbox数量 × 2。如下结构：mbox_priorbox输出3个default boxs，所以mbox_conf和mbox_loc对应的output_num分别为6和12。

layer {
  name: "conv6_2_mbox_loc"
  type: "Convolution"
  bottom: "conv6_2"
  top: "conv6_2_mbox_loc"
  param {
    lr_mult: 1.0
    decay_mult: 1.0
  }
  param {
    lr_mult: 2.0
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 12
    pad: 1
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv6_2_mbox_loc_perm"
  type: "Permute"
  bottom: "conv6_2_mbox_loc"
  top: "conv6_2_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv6_2_mbox_loc_flat"
  type: "Flatten"
  bottom: "conv6_2_mbox_loc_perm"
  top: "conv6_2_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv6_2_mbox_conf"
  type: "Convolution"
  bottom: "conv6_2"
  top: "conv6_2_mbox_conf"
  param {
    lr_mult: 1.0
    decay_mult: 1.0
  }
  param {
    lr_mult: 2.0
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 6
    pad: 1
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv6_2_mbox_conf_perm"
  type: "Permute"
  bottom: "conv6_2_mbox_conf"
  top: "conv6_2_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv6_2_mbox_conf_flat"
  type: "Flatten"
  bottom: "conv6_2_mbox_conf_perm"
  top: "conv6_2_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv6_2_mbox_priorbox"
  type: "PriorBox"
  bottom: "conv6_2"
  bottom: "data"
  top: "conv6_2_mbox_priorbox"
  prior_box_param {
    min_size: 256.0
    aspect_ratio: 2.0
    flip: true
    clip: false
    variance: 0.10000000149
    variance: 0.10000000149
    variance: 0.20000000298
    variance: 0.20000000298
    step: 64.0
    offset: 0.5
  }
}

在odm部分，其实与arm部分一一对应，P3、P4、P5和P6分别与conv4_3、conv5_3、fc7和conv6_2相互对应，虽然在odm部分并未产生mbox_priorbox，但根据odm_loss的计算方式（如下结构）其实包含了arm部分产生的priorbox。

layer {
  name: "odm_loss"
  type: "MultiBoxLoss"
  bottom: "odm_loc"
  bottom: "odm_conf"
  bottom: "arm_priorbox"
  bottom: "label"
  bottom: "arm_conf_flatten"
  bottom: "arm_loc"
  top: "odm_loss"
  include {
    phase: TRAIN
  }
  propagate_down: true
  propagate_down: true
  propagate_down: false
  propagate_down: false
  propagate_down: false
  propagate_down: false
  loss_param {
    normalization: VALID
  }
  multibox_loss_param {
    loc_loss_type: SMOOTH_L1
    conf_loss_type: SOFTMAX
    loc_weight: 1.0
    num_classes: 5
    share_location: true
    match_type: PER_PREDICTION
    overlap_threshold: 0.5
    use_prior_for_matching: true
    background_label_id: 0
    use_difficult_gt: false
    neg_pos_ratio: 3.0
    neg_overlap: 0.5
    code_type: CENTER_SIZE
    ignore_cross_boundary_bbox: false
    mining_type: MAX_NEGATIVE
    objectness_score: 0.00999999977648
  }
}

所以odm部分产生的conf和loc是对arm部分的优化和补充，此部分产生mbox_conf和mbox_loc的四个层P3、P4、P5和P6对应conv4_3、conv5_3、fc7和conv6_2产生的priorbox数量，其output_num符合原始SSD计算方式，即mbox_conf的output_num=mbox_priorbox数量 × class_num(样本类别数)。而mbox_loc输出通道与arm部分相同，即output_num=mbox_priorbox数量 × 4。

如下结构为P6后产生的mbox_conf和mbox_loc层的output_num，与上面结构的conv6_2对应。即mbox_priorbox输出3个default boxs，所以mbox_conf和mbox_loc对应的output_num分别为15（类别数是5，即5×3=15）和12。

layer {
  name: "P6_mbox_loc"
  type: "Convolution"
  bottom: "P6"
  top: "P6_mbox_loc"
  param {
    lr_mult: 1.0
    decay_mult: 1.0
  }
  param {
    lr_mult: 2.0
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 24
    pad: 1
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "P6_mbox_loc_perm"
  type: "Permute"
  bottom: "P6_mbox_loc"
  top: "P6_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "P6_mbox_loc_flat"
  type: "Flatten"
  bottom: "P6_mbox_loc_perm"
  top: "P6_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "P6_mbox_conf"
  type: "Convolution"
  bottom: "P6"
  top: "P6_mbox_conf"
  param {
    lr_mult: 1.0
    decay_mult: 1.0
  }
  param {
    lr_mult: 2.0
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 15
    pad: 1
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}