Action Recognition
Identifying what action or activity a person is performing in video or from skeleton keypoints.
Think of it like
Watching an athlete and saying "she is doing a cartwheel" or "he is shooting a basketball."
Example
Video of person doing: jumping jack -> identified. Skeleton pose: arms up, legs spread -> jumping jack.
How it actually works
Appearance-based: use RGB frames (3D CNN, optical flow). Skeleton-based: use pose keypoints (graph conv, temporal conv). Skeleton methods are faster and more interpretable but require accurate pose estimation. Challenges: subtle actions (whispering vs. talking), view-invariance (different angles), fine-grained (kick vs. jump). Temporal modeling (RNN, transformer, temporal convolution) crucial. Large datasets (Kinetics) improved performance.
For product teams
Fitness tracking, sports analytics, gaming, accessibility (gesture control).
For engineers
RGB-based: I3D, SlowFast. Skeleton-based: ST-GCN, Shift-GCN. Datasets: Kinetics-400, NTU RGB+D. Metric: top-1 accuracy.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome