3D Gesture Estimation from RGB Images Based on DB-InterNet

Abstract

The booming of deep learning has made it possible to estimate 3D gestures from ordinary color images. However, the high accuracy of inferring 3D hand postures from RGB images is still not available due to the high flexibility of the gestures themselves. This paper aims to address the problem of low accuracy of keypoint coordinates position in InterNet gesture estimation network, by means of improving the confidence coordinate function, and selecting different suppression factor β according to different keypoints to fit t interacting hand estimation. By doing so, the maximum position coordinates are more precise and the keypoint prediction is more accurate. With respect to good measures, enhancing the representation capability of the model and employing dynamic activation function at different locations of the network to learn features in a dynamic way so as to boost the learning of hidden joints. In this way, different layers dynamically adjusted the segmental activation function according to the input to improve the performance of the model by dynamically learning features in a more flexible way. The experimental results indicate that compared with the baseline algorithm, the MPJPE and MRRPE of this algorithm are reduced by 0.73% and 2.04%, respectively, and the accuracy of single hand estimation is higher while the accuracy of interacting hand estimation is also effectively improved.