Despite recent successes in hand pose estimation from RGB images or depth maps, inherent challenges remain. RGB-based methods suffer from heavy self-occlusions and depth ambiguity. Depth sensors rely heavily on distance and can only be used indoors, thus there are many limitations to the practical application of depth-based methods. The aforementioned challenges have inspired us to combine the two modalities to offset the shortcomings of the other. In this paper, we propose a novel RGB and depth information fusion network to improve the accuracy of 3D hand pose estimation, which is called CrossFuNet. Specifically, the RGB image and the paired depth map are input into two different subnetworks, respectively. The feature maps are fused in the fusion module in which we propose a completely new approach to combine the information from the two modalities. Then, the common method is used to regress the 3D key-points by heatmaps. We validate our model on two public datasets and the results reveal that our model outperforms the state-of-the-art methods.

In this paper, we propose TSFE-Net, two stream feature extraction networks for active stereo matching. First, we perform extra local contrast normalization (LCN) for dataset due to dependency between speckle intensity and distance. Second, we construct two stream feature extraction layers which consist of convolutional layers and deconvolutional layers in different scales to simultaneously learn the features of the original images and LCN images and aggregate context information to form the left and right features. Third, we convert the obtained depth map into disparity map in virtue of camera parameters to construct a supervised learning model. The TSFE-Net not only solves illumination effects between speckle intensity and distance but also reserves details of the original image. Our dataset are captured by RealSense D435 camera. We research extensive quantitative and qualitative evaluations based on a series of scenes, and achieve the end point error (EPE) accuracy of 0.335 on the TITAN XP platform only for valid pixel. The assessment results show that our network has the ability of real-time deep reconstruction for active pattern.