Human Pose Estimation from 2D to 3D Based on Complex Valued Frequency Domain Transformation

In recent years, the transform-based approaches have been employed to estimate 3D human pose from 2D keypoint sequences. These approaches learn spatio-temporal correlations by comprehensively taking into account body joints in all frames. Nevertheless, due to the absence of depth information in existing methods, diverse 3D poses might yield the same 2D projection, causing motion ambiguity. To address this issue, we propose a frequency-domain skeletal dynamic modeling methodology. Firstly, the 2D skeleton sequence is transformed to the frequency domain through Fourier transform, and high-frequency noise is eliminated by low-pass filtering to retain the principal motion features. Subsequently, the amplitude and phase information of human bone points is learned by utilizing complex linear layers, and the postures are regenerated and aggregated. Finally, the attitude optimizer is employed to smooth and optimize the attitude, which significantly enhances the stability of the attitude. The aforementioned treatment considerably reduces the impact of motion ambiguity and enhances the accuracy of 3D pose recovery. A considerable number of experimental results demonstrate that the performance of this network architecture on the Human3.6M dataset is improved by 5.3% compared to the previous optimal results.