PoseTSN (Tanaka et al.)

Two-stage pipeline combining 2D/3D pose estimation with a temporal segmentation network for element boundaries.

Primary tasks
Element detection, Phase segmentation
Modalities
RGB video, 3D pose sequences
Architecture
Pose estimator + temporal ConvNet
Frameworks
PyTorch, TensorFlow
Availability
research
Maintainer
Keio University
Released
May 2021

Highlights

Implementation Notes

The authors release scripts for synchronising multi-angle footage prior to segmentation. When adapting PoseTSN, align coordinate systems to avoid drift between cameras.