ESPNetv2:

ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network (CVPR2019)
​​​https://arxiv.org/pdf/1811.11431.pdf​​​ PyTorch: ​​https://github.com/sacmehta/ESPNetv2​​ 主要在ESPNet的基础上改进,

特点:
为了计算更加高效,见Figure 1:

  1. 将原来ESPNet的point-wise convolutions 替换为group point-wise convolutions;
  2. 将原来ESPNet的dilated convolutions 替换为depth-wise dilated convolution;
  3. HFF加在depth-wise dilated separable convolutions 和 point-wise (or 1 × 1) 卷积之间,去除gridding artifacts ;
  4. 使用© 中一个 group point-wise convolution 替换(b)K 个point-wise convolutions ;
  5. 语义分割论文:ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network_卷积

  6. 为了网络表示更有效,见Figure 2 :
  7. 将©中depth-wise dilated convolutions 加入下采样操作;
  8. 加入平均池化(average pooling ),将输入图片信息加入EESP中;
  9. 使用级联(concatenation) 取代对应元素加法操作(element-wise addition operation );
  10. 语义分割论文:ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network_级联_02

  11. 不同类型的卷积参数量和感受野比较:
  12. 语义分割论文:ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network_卷积_03

  13. 最后的分类网络结构:
  14. 语义分割论文:ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network_卷积_04