← Back to Work Projects
03 / Action Recognition · Temporal Modeling Internal · Not Public

Shot & Action Intelligence

Temporal clip classification for cricket and badminton — automated shot tagging that eliminates manual breakdown work and enables tactical analysis at scale.

What

Given a short video clip, classify what just happened — the type of shot, the type of action, the player intent. Cricket: drive, cut, pull, sweep, defence. Badminton: smash, drop, clear, net shot.

Output integrates with the annotation engine and ball / shuttle tracker so each event has a label, a trajectory, and a coordinate footprint.

Why

Manual shot tagging is the single biggest bottleneck in cricket analytics. A single T20 innings needs hundreds of labels; a full season is a small team's full-time job.

Automating it unlocks scale — tactical patterns across thousands of innings, opponent vulnerability maps, player-shot heatmaps that were previously impossible to compile.

How
  • Temporal model over short clips — frames of player motion as input, multi-class shot taxonomy as output.
  • Sport-specific class hierarchies built with domain experts, not pulled from generic action-recognition datasets.
  • Trained on labels harvested through the internal annotation engine, with iterative class refinement as edge cases surfaced.
  • Designed to fail gracefully — low-confidence clips get routed back to analysts via the annotation tool.
Difficulties
  • Class boundaries — a pull and a hook are biomechanically close; the taxonomy needed sport-expert input to be useful, not just statistically separable.
  • Camera variation — same shot looks very different from main-camera vs square-leg-camera.
  • Class imbalance — defensive shots dominate the dataset; rare-but-decisive shots needed targeted oversampling.
  • Useful confusion — analysts care about which mistakes the model makes, not just the headline accuracy.
Numbers
Multi
Shot / Action Classes
2
Sports (Cricket · Badminton)
Auto
Tagging at scale
↻
Confidence-routed HITL