High-resolution network with an auxiliary channel for 2D hand pose estimation

High-resolution networks have been applied in various fields because of their advanced architecture. However, multiple multi-scale fusions of high and low-dimensional semantic information during hand pose estimation can blur the position information obtained in high resolution, causing overfitting. To address this problem, we added an auxiliary channel parallel to the original network in this study. The auxiliary channel slices images using a slicing operation instead of a convolutional downscaling operation to preserve the full information in the input. The input is then computed by following four convolution layers to obtain the initial position correction information, and the results are combined with the network for prediction. Adding the auxiliary channel increases the number of parameters in the original network by only 0.7%, but obtains a high accuracy gain, which is particularly noticeable on lightweight networks. We performed several experiments to verify the effectiveness of this method using multiple datasets.