official: https://github.com/ruinmessi/RFBNet
multi-branch
multi conv & dilate conv kernel
branch1: 1x1 conv + 3x3 conv
branch2: 1x1 conv + 3x3 conv + 3x3 conv rate3, 第一个3x3conv可以分解成1x3和3x1两个branch->s版
branch3: 1x1 conv + 5x5 conv + 3x3 conv rate5, 第一个5x5conv可以分解成两个串行的3x3
源码的实现和原论文有些不一致
* channel
* efficiency分解,1x3/3x1啥的
* dilate rate
跟inception block的想法差不多,多个branch,提取不同尺度的特征,
但是inception block都是基于同一个中心,
1. assemble on ssd top
2. vgg16 back
conv7输出接一个RFB block
conv8和conv9有改动,换成了stride2的RFB block
3. individual heads, 3x3 conv