MoCap-solver

MarkerNet: A divide‐and‐conquer solution to motion capture solving from raw markers

“…Recently, with the development of deep learning, some researchers [1][2][3][26][27][28] attempt to train nonlinear neural networks to learn the feature space of motions in a data-driven way. Holden 1 proposed a simple FC network equipped with residual connections to directly produce the global transformations of human joints from input raw markers frame-by-frame.…”

“…However, such a per-frame way can not fully consider and leverage the temporal information of MoCap data, tending to break temporal continuity of motion sequences. On the contrary, Chen et al 3 presented a two-stage framework, named MoCap-Solver, to obtain human motions from a sequence of input raw markers. To the best of our knowledge, MoCap-Solver 3 is the state-of-the-art solution to the specific task of motion solving from marker-based raw MoCap data until of the time of our research.…”

“…Let X ∈ R T×M×3 and Y ∈ R T×M×3 denote a temporal sequence of 3D positions of raw markers and clean markers respectively, where T is the number of frames in the temporal sequence and M is the number of markers in each frame. As pointed by MoCap-Solver, 3 the 3D global positions of markers depend on three kinds of intrinsic information: skeleton template, skeleton motion, and marker configuration (see Figure 2a). Specifically, the skeleton template Y t ∈ R J×3 corresponds to the human body shape and it is defined by 3D local offset of each joint relative to its parent in T-pose, where J is the number of joints.…”

“…For a fair comparison, we follow the same marker order in previous work, 3 which conforms to the original data format exported by MoCap systems, as the default order to implement our model. Given the divided marker subsets, we further employ a series of local operators to capture local spatio-temporal relationships among different marker subsets.…”

“…However, they do not fully consider and leverage the temporal information of MoCap data, which tends to break temporal continuity of motion sequences. To tackle the above issue, Chen et al 3 present a state-of-the-art framework, named MoCap-Solver, to predict human skeleton motion sequences from the whole sequential raw markers using FC layers. Since they only consider and utilize the global correlation among all markers within the whole sequence but do not effectively model the spatio-temporal information of marker subsets within local ranges, their method might generate sub-optimal results with inaccurate local details when complicated human motions occur (see Figure 1).…”