Google put the method to extract different feature
based on Slow Network and Fast Network
| The First Colum | The Second Column |
|---|---|
| innovation point1 | 基于存储引导的交替模型 |
| InterIeaved Models | slow network and fastnetwork is made up by two MobilNetV2 the depth multiplier of the two models are different, before is 1.4,and the after is 0.35  |
| innovation point2 | 记忆单元, Memory module 存储模型， LSTM可以高效处理时序信息 但是卷积运算量大 ConvLSTM将CNN与LSTM结合 |
| ConLSTM is designed by the 时序时间信息的图像 |
|
| 1 innovation of the ConvLSTM | 增加了bottleneck Gate 和output 的跳跃连接 |
| 2 innovation of the ConvLSTM | 将LSTM单元进行分组卷积 feature maps 原本是H * W * N 将其分为G group 每个LSTM处理的HWN/G 的feature maps |
| the step of LSTM | the first step : f(t) = sigmoid(W(f) * [h(t-1), x(t)] + b(f) ) LSTM include the activate function (sigmoid) and the action (pointwise) the first of the LSTM is sigmoid |
| The step of LSTM | The second step : i(t) = sigmoid( W(i) * [ h(t-1), x(t)] + b(i) ); C~(t) = tanh( W(C)* [h(t-1), x(t)] + b(c) ) Tanh create a new 候选值vector |
| The step of LSTM | The third step : C(t) = f(t) * C(t-1) + i(t) * C~(t) |