Keyframe-TSN (Liu et al.)

Temporal segment network fine-tuned on keyframes to classify jump and spin elements from broadcast footage.

Primary tasks
Element detection, Element classification
Modalities
RGB video
Architecture
Temporal Segment Network with keyframe attention
Frameworks
PyTorch
Availability
research
Maintainer
Shanghai Jiao Tong University
Released
Dec 2019

Highlights

Implementation Notes

Augmentations emphasise slow shutter artifacts common in rink footage; replicate them when adapting to new broadcasts. Pairing Keyframe-TSN with a lightweight pose estimator improves false negative rates on layback spins.