版权声明：原创博客未经允许请勿转载！ https://blog.csdn.net/holmes_MX/article/details/89295250
1. Receptive Field 计算方法
当前层的stride: 历史所有层的stride的乘机
当前层的Receptive Field: RF = last_RF + (ksize - 1) * history_stride, 其中, last_RF为上一层的Receptive Field, ksize为当前层的kernel size, history_stride为历史积累的stride(不包括当前层的stride).
2. 以DenseBox目标检测中的为例进行说明
DenseBox是百度发表的一篇针对Object Detection的paper, 网络的backbone是VGG19.
这里以paper中描述的conv3_4为例进行说明, 论文中指出conv3_4的感受野为48 * 48.
计算过程如下:
| Conv Name | ksize | stride | history_stride | last_RF | RF |
| Input | -- | -- | -- | -- | 1 |
| conv1_1 | 3 | 1 | 1 | 1 | 3 |
| conv1_2 | 3 | 1 | 1 | 3 | 5 |
| pool1 | 2 | 2 | 1 | 5 | 6 |
| conv2_1 | 3 | 1 | 2 | 6 | 10 |
| conv2_2 | 3 | 1 | 2 | 10 | 14 |
| pool2 | 2 | 2 | 2 | 14 | 16 |
| conv3_1 | 3 | 1 | 4 | 16 | 24 |
| conv3_2 | 3 | 1 | 4 | 24 | 32 |
| conv3_3 | 3 | 1 | 4 | 32 | 40 |
| conv3_4 | 3 | 1 | 4 | 40 | 48 |
| pool3 | 2 | 2 | 4 | 48 | 52 |
| conv4_1 | 3 | 1 | 8 | 52 | 68 |
| conv4_2 | 3 | 1 | 8 | 68 | 84 |
| conv4_3 | 3 | 1 | 8 | 84 | 100 |
| conv4_4 | 3 | 1 | 8 | 100 | 116 |
DenseBox中给出conv4_4的感受野是大概是118 * 118, 与本文计算的差不多.
[Reference]
[1] http://shawnleezx.github.io/blog/2017/02/11/calculating-receptive-field-of-cnn/