Page 11 - Read Online
P. 11
Page 95 Liu et al. Art Int Surg 2024;4:92-108 https://dx.doi.org/10.20517/ais.2024.19
Figure 1. Overview of our framework.
Frame-based human mesh recovery
To recover human meshes from video frames (Figure 2, right column), we adopted the architecture and
training procedure proposed by Li et al., with one important deviation . We swapped the Skinned Multi-
[15]
Person Linear (SMPL) parametric model with SMPL eXpressive (SMPL-X) to extend the mesh recovery
process to include more granular representations of hand and face joints . We achieved comparable
[21]
evaluation scores in Mean Per Joint Position Error (MPJPE), Procrustes-Aligned Mean Per Joint Position
Error (PA-MPJPE), and Per Vertex Error (PVE) when evaluating on the benchmarks outlined in the study,
which verified our trained model.
Surgical behavior analysis
Collating the results of our tracking and HMR procedures, we next performed a comprehensive set of
qualitative analyses to interpret the movements and behaviors of individuals in the simulated scenes
throughout time [Figure 1B]. In our analysis, we focused on metrics that were both significant to
understanding the effectiveness of an OR procedure and clearly discernible from physical pose alone (i.e.,
sharp turn of neck to indicate attention switch). Specifically, we focused on:
