VoVNet

发表于 |

An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection

  1. 动机

    • denseNet
      • dense path：diverse receptive fields
      • heavy memory cost & low efficiency
    • we propose a backbone
      • preserve the benefit of concatenation
      • improve denseNet efficiency
      • VoVNet comprised of One-Shot Aggregation (OSA)
    • apply to one/two stage object detection tasks
      • outperforms denseNet & resNet based ones
      • better small object detection performance

  2. 论点

    • main difference between resNet & denseNet
      • aggregation：summation & concatenation
        • summation would washed out the early features
        • concatenation last as it preserves
    • GPU parallel computation
      • computing utilization is maximized when operand tensor is larger
      • many 1x1 convs for reducing dimension
      • dense connections in intermediate layers are inducing the inefficiencies

    • VoVNet

      • hypothesize that the dense connections are redundant
      • OSA：aggregates intermediate features at once
      • test as object detection backbone：outperforms DenseNet & ResNet with better energy efficiency and speed

    • factors for efficiency

      • FLOPS and model sizes are indirect metrics
      • energy per image and frame per second are more practical
      • MAC：
        • memory accesses cost，$hw(c_i+c_o) + k^2 c_ic_o$
        • memory usage不止跟参数量有关，还跟特征图尺寸相关
        • MAC can be minimized when input channel size equals the output
      • FLOPs/s
        • splitting a large convolution operation into several fragmented smaller operations makes GPU computation inefficient as fewer computations are processed in parallel
        • 所以depthwise/bottleneck理论上降低了计算量FLOP，但是从GPU并行的角度efficiency降低，并没有显著提速：cause more sequential computations
        • 以时间为单位的FLOPs才是fair的

  3. 方法

    • hypothesize

      • dense connection makes similar between neighbor layers
      • redundant

    • OSA

      • dense connection：former features concats in every following features

      • one-shot connection：former features concats once in the last feature

      • 最开始跟dense block保持参数一致：一个block里面12个layers，channel20，发现深层特征contributes less，所以换成浅层，5个layers，channel43，发现有涨点：implies that building deep intermediate feature via dense connection is less effective than expected

      • in/out channel数相同

        • much less MAC：
          • denseNet40：3.7M
          • OSA：5layers，channel43，2.5M
          • 对于higher resolution的detection任务impies more fast and energy efficient
        • GPU efficiency
          • 不需要那好几十个1x1

    • architecture

      • stem：3个3x3conv
      • downsamp：s2的maxpooling
      • stages：increasing channels enables more rich semantic high-level information，better feature representation

      • deeper：makes more modules in stage3/4

  4. 实验

    • one-stage：refineDet
    • two-stage：Mask-RCNN