3D Human Pose Estimation with Dilated Sampled Frames
G. F. Cao, Qing Tang, Tran Tien Dat, Ashraf Uddin Russo, Kang-Hyun Jo · 2023
Three-dimensional (3D) human pose estimation (HPE) targets to produce the 3D spatial coordinates of the human pose from 2D images. 3D HPE is a basic computer vision task for many intelligent industrial applications. Commonly, the coordinates of the predicted 3D human pose joints are calculated through the 2D keypoint from the ground truth provided by the datasets or generated by a classical and robust 2D human pose estimator. With the development of transformer-based methods, the methods with a sequence of monocular images have achieved great success in 2D-to-3D lifting human pose estimation. In this paper, the sampled frames with a dilated ratio are given as the input of the 3D human pose estimator. Extensive experiments on the public benchmark Human 3.6M demonstrate the significance and effectiveness of the proposed method.