Structured svm backpropagation to convolutional neural network applying to human pose estimation

Abstract

In this work, we show, for the first time, how to formulate Structured SVM as two layers of Convolutional Neural Network, the top layer of which is loss augmented inference layer, and the bottom is normal convolutional layer. We show that Deformable Part Model can be learned with newly created Structured SVM neural network by propagating the error of Deformable Part Model back propagate to Convolutional Neural Network. The forward propagation calculates loss augmented inference. The back propagation calculates the gradient from the loss augmented inference layer to convolutional layer. By doing so, we create a new type of convolutional neural network: Structured SVM Convolutional Neural Network, which is then applied to Human Pose Estimation problem. This new creation is a neural network, which can be used as the last layers of deep learning. Our method jointly learns structural model parameters and appearance model parameters. Our back propaga- tion formulation can also be applied to Multiclass Classification Structured SVM where our result outperformed widely used Softmax classifier on standard MNIST dataset. We imple- ment our method as a new layer of existing Caffe library.