ARC-SORT: A Training-Free Real-Time Tracker Cuts Identity Switches in Sports Broadcasts
Girish Kumar Adari · hf · 2026-10-19
Girish Kumar Adari of Athlitix Inc. released ARC-SORT (Adaptive, Robust, Compensated SORT), a tracking-by-detection multi-object tracker built for sports broadcasts, where athletes move fast and change direction abruptly while the camera pans and zooms simultaneously.
Key points:
- It trains nothing and runs in real time, improving only the association step on top of a fixed detector.
- An Interacting Multiple Model (IMM) Kalman filter adaptively switches between motion/camera modes.
- The tracker measures its own uncertainty, and that measurement determines the size of the adaptive search region.
- Corrective identity swaps and appearance-only re-identification further suppress ID switches, while compensation handles camera motion.
The author reports high accuracy with minimal identity switches on sports footage; the paper appeared on arXiv (Oct 15, 2026).
More from Embodied
- Tesla tops 10 million vehicles delivered and 130 GWh of energy storage deployed — XFreeze · 2026-10-19
- Humanoid robots gain a sense of touch: the rise of feeling robots — TinfoilTricorn · 2026-10-19
- Shipbuilder HII's $900M robotics deals point to engineering firms selling parts, not software — burhop · 2026-10-19
- d1-omni-600M runs solar-powered on a Raspberry Pi 5, with a sub-$125 productization plan — helloiamleonie · 2026-10-19
- Liquid's d1-3B decision model drives robot skills at 157ms on Intel Arc B580 — helloiamleonie · 2026-10-19
- AUAR ships micro-factories in shipping containers: 200 units could clear UK's annual housing backlog — sarahdrinkwater · 2026-10-19