Partially transferred convolution neural network with cross-layer inheriting for posture recognition from top-view depth camera

This paper proposes a new method for human posture recognition from top-view depth maps on small training datasets. There are two strategies developed to leverage the capability of convolution neural network (CNN) in mining the fundamental and generic features for recognition. First, the early layers of CNN should serve the function to extract feature without specific representation. By applying the concept of transfer learning, the first few layers from the pre-learned VGG model can be used directly without further fine-tuning. To alleviate the computational loading and to increase the accuracy of our partially transferred model, a cross-layer inheriting feature fusion (CLIFF) is proposed by using the information from the early layer in fully connected layer without further processing. The experimental result shows that combination of partial transferred model and CLIFF can provide better performance than VGG16 [1] model with re-trained FC layer and other hand-crafted features like RBPs [2].