by Bo Pu, Leiden University, The Netherlands
Abstract
Most of the prior research approaches for facial landmarks estimation improve their accuracy by sacrificing the running performance. However, this paper presents a new approach for localizing the positions of facial landmarks precisely, which strikes a balance between the rate of accuracy and the efficiency of network. Since the commonly used Max-Pooling layer actually causes the information loss, our model concatenates the feature extracted on each convolutional layer toward reducing the loss of information in the network. The experimental results demonstrate that the accuracy of this network increases while the performance maintains computational efficiency since the loss of information is integrated in the concatenate layer. Furthermore, the weight of each convolutional layer is optimized to filter the information overlap of different layers while minimizing redundant computation.
---------------------------