refineDet

和refineNet没有任何关系

动机
- inherit the merits of both two-stage and one-stage：accuracy and efficiency
- single-shot
- multi-task
- refineDet
  - anchor refinement module (ARM)
  - object detection module (ODM)
  - transfer connection block (TCB)
论点
- three advantages that two-stage superior than one-stage
  - RPN：handle class imbalance
  - two step regress：coarse to refine
  - two stage feature：RPN任务和regression任务有各自的feature
- 模拟二阶段检测的RPN，把classifier任务中的大量阴性框先排掉，但不是以两个阶段的形式，而是multi-task并行
- 将一阶段检测的objectness和box regression任务解耦，两个任务通过transfer block连接
- ARM
  - remove negative anchors to reduce search space for the classifier
  - coarsely adjust the locations and sizes of anchors to provide better initialization for regression
- ODM
  - further improve the regression
  - predict multi labels
- TCB
  - transfer the features in the ARM to handle the more challenging tasks in the ODM
方法
- Transfer Connection Block
  - 没什么新的东西，上采样用了deconv，conv-relu，element-wise add
- Two-Step Cascaded Regression
  - fisrt step ARM prediction
    - for each cell，for each predefined anchor boxes，predict 4 offsets and 2 scores
    - obtain refined anchor boxes
  - second step ODM prediction
    - with justified feature map，with refined anchor boxes
    - generate accurate boxes offset to refined boxes and multi-class scores，c+4
- Negative Anchor Filtering
  - reject well-classified negative anchors
  - if the negative confidence is larger than 0.99，discard it in training the ODM
  - ODM接收所有pred positive和hard negative
- Training and Inference details
  - back：VGG16 & resnet101
    - fc6 & fc7变成两个conv
    - different feature scales
    - L2 norm
    - two extra convolution layers and one extra residual block
  - 4 feature strides
    - each level：1 scale & 3 ratios
    - ensures that different scales of anchors have the same tiling density on the image
  - matching
    - 每个GT box match一个score最高的anchor box
    - 为每个anchor box找到最匹配的iou大于0.5的gt box
    - 相当于把ignore那部分也作为正样本了
  - Hard Negative Mining
    - select negative anchor boxes with top loss values
    - n & p ratio：3:1
  - Loss Function
    - ARM loss
      - binary class：只计算正样本？？？
      - box：只计算正样本
    - ODM loss
      - pass the refined anchors with the negative confidence less than the threshold
      - multi-class：计算均衡的正负样本
      - box：只计算正样本
    - 正样本数为0的时候，loss均为0：纯阴性样本无效？？

RefineDet: Single-Shot Refinement Neural Network for Object Detectio