spineNet

发表于 |

SpineNet: Learning Scale-Permuted Backbone for Recognition and Localization

  1. 动机

    • object detection task
      • requiring simultaneous recognition and localization
      • solely encoder performs not well
      • while encoder-decoder architectures are ineffective

    • propose SpineNet

      • scale-permuted intermediate features
      • cross-scale connections
      • searched by NAS on detection COCO
      • can transfer to classification tasks

      • 在轻量和重量back的一阶段网络中都涨点领先

  2. 论点

    • scale-decreasing backbone
      • throws away the spatial information by down-sampling
      • challenging to recover
      • 接一个轻量的FPN：
    • scale-permuted model
      • scales of features can increase/decrease anytime：retain the spacial information
      • connections go across scales：multi-scale fusion
      • searched by NAS
      • 是一个完整的FPN，不是encoder-decoder那种可分的形式
      • directly connect to classification and bounding box regression subnets
      • base on ResNet50
        • use bottleneck feature blocks
        • two inputs for each feature blocks
        • roughly the same computation

  3. 方法

    • formulation

      • overall architecture
        • stem：scale-decreased architecture
        • scale-permuted network
        • blocks in the stem network can be candidate inputs for the following scale-permuted network
      • scale-permuted network
        • building blocks：$B_k$
        • feature level：$L_3 - L_7$
        • output features：1x1 conv，$P_3 - P_7$

    • search space

      • scale-permuted network：

        • block只能从前往后connect
        • based on resNet blocks
        • channel 256 for $L_5, L_6, L_7$

      • cross-scale connections：

        • two input connections for each block

        • from lower ordering block / stem

        • resampling

          • narrow factor $\alpha$：1x1 conv
          • 上采样：interpolation
          • 下采样：3x3 s2 conv

          • element-wise add

      • block adjustment

        • intermediate blocks can adjust its scale level & type
        • level from {-1, 0, 1, 2}
        • select from bottleneck / residual block

    • family of models

      • R[N] - SP[M]：N feature layers in stem & M feature layers in scale-permuted layers
      • gradually shift from stem to SP

      • with size decreasing

    • spineNet family

      • basic：spineNet-49
      • spineNet-49S：channel数scaled down by 0.65
      • spineNet-96：double the number of blocks
      • spineNet-143：repeat 3 times，fusion narrow factor $\alpha=1$
      • spineNet-190：repeat 4 times，fusion narrow factor $\alpha=1$，channel数scaled up by 1.3

  4. 实验

    • 在mid/heavy量级上，比resnet-family-FPN涨出两个点

    • 在light量级上，比mobileNet-family-FPN涨出一个点