K-RAID hero background
Back to indexProduction · Shipped

K-RAID

Kabaddi video intelligence — turns raw match footage into structured events using pose estimation and multi-agent reasoning.

Visit Project ↗

Role

Solo research & systems design — active, ongoing

Timeline

Currently building

Stack

  • Python·
  • PyTorch·
  • HigherHRNet·
  • ST-CGNet·
  • U2Diff·
  • multi-agent architecture

The problem

Kabaddi is a contact sport — bodies overlap constantly, which is exactly the condition that breaks most pose-estimation pipelines built for cleaner footage like athletics or tennis. A raider mid-tackle with three defenders wrapped around him is a genuinely hard input, and match footage rarely comes in broadcast quality.

Why one model wasn't enough

A single end-to-end model tries to do pose estimation, role understanding, and event detection all at once, and occlusion-heavy frames make it brittle. Splitting the reasoning across four specialized agents — Raider, Defense, Simulation, and Referee — lets each one own a narrower, more tractable problem and vote into a shared understanding of what just happened on the mat, instead of one model guessing blind through the occlusion.

Where it stands

Research-stage, actively being built. Current numbers — 94.2% confidence at 38ms inference — are real but not yet stress-tested against the full range of footage quality this needs to handle in the wild. Treating that as a research checkpoint, not a finished product, is the honest framing.

The open problem

Confidence drops on low-light, low-bitrate amateur footage — exactly the kind of footage a grassroots kabaddi league would actually produce, as opposed to the broadcast-quality footage easiest to train on. Closing that gap with augmented-dataset training is the current focus.