Paper Link: Complexity and Long-range Correlation between Parts
It means to know as much as possible about the internal dependence of joint points. The author chooses to increase the receptive field by adding convolution kernel to control the overall number of parameters (other ways include, using aggregation operation but sacrificing accuracy; Or increase the number of convolution layers but increase the number of parameters). The author also posted an experimental result chart in this paper, which proves that increasing the receptive field can improve the final accuracy.
And this relationship also shows that, in fact, there is indeed a correlation between the key points learned in the network (this accuracy increases with the size of the receptive field, indicating that the network does index and encode the long-range interaction between parts, and it is beneficial to do so. ).
At this point, the whole network architecture is stumbling and clear at a glance ~ in fact, it is very simple to understand, that is, the convolutional network is nested into the Pose Machines framework.
When training deep neural networks, we often face the problem of gradient disappearance, and the architecture design of CPM can solve this problem well. At each stage, PM will be repeatedly trained to predict the confidence map of each key point. Taking the ideal belief graph as a label, we construct a Gaussian distribution near each real key point, which is recorded as, and then our loss function is minimized.
The total loss function is to add the loss functions of each stage,
There is also a schematic diagram of the base paper used to show the gradient.
Other specific training process results will not be shown here, which is relatively simple. You can read the original paper by yourself.
-The third mushroom.-Conclusion.
At this point, the core idea of the whole paper has been clear. In this paper, CPM network architecture embedded with convolutional neural network in PM framework is proposed for human key point detection, and its feasibility is proved by experiments, which lays the foundation for subsequent development.
The brief summary of this paper is to list the summary of this paper first, then introduce the author's ideas in detail, and also briefly express my understanding of CPM network architecture. I hope you can further deepen your understanding of this paper after reading this article. Please point out your mistakes, communicate more and make progress together ~?