Table of Contents
For more than two decades, cricket ball-tracking has largely depended on proprietary infrastructure built around specialist camera systems, limiting advanced tracking capabilities to venues capable of supporting expensive hardware deployments. Now, Score believes decentralized AI miners on Bittensor may be able to change that using broadcast footage alone.
The subnet announced a new cricket ball-tracking challenge this week, tasking miners with extracting detailed trajectory and delivery data from standard 20–30 second broadcast clips. Instead of relying on dedicated camera rigs, miners are expected to reconstruct ball behavior using a single broadcast feed and return 21 ground-truth-aligned prediction fields for each delivery.
“Hardest challenge we've shipped,” wrote Score core contributor Max in a tweet. “Hawkeye does this with multiple specialist cameras. Miners on SN44 get one broadcast clip and 30 seconds.”
The challenge was built alongside Nathan Leamon, former Head of Data for the England and Wales Cricket Board (ECB), and entrepreneur and cricket team owner Ali Tareen. Validators provide miners with short clips containing a single bowl, while miner systems must return structured outputs including release speed, bounce coordinates, stump-plane trajectory, swing angle, deviation after bounce, and delivery outcomes such as runs and wickets.
Cricket represents one of the largest sports markets in the world, with an estimated 2.5 billion fans globally and the Indian Premier League valued at more than $15 billion. Despite the scale of the sport, advanced ball-tracking technology remains heavily tied to expensive physical infrastructure, preventing many grounds from deploying the same level of analytics available at top-tier venues.
Subnet 44’s approach attempts to bypass that infrastructure layer entirely. Instead of requiring specialized tracking hardware installed around a stadium, the challenge asks miners to infer complex ball physics directly from ordinary broadcast footage already produced at cricket grounds worldwide.
According to the specification, scoring is weighted heavily toward six primary ball-tracking metrics—kph, bounce_x, stump_y, deviation, swing_angle, and stump_z—which account for 75% of the total score. The remaining metadata and outcome fields collectively contribute 25%.
“Numerics use strict tolerance-based decay, so rough approximations aren't competitive,” Max wrote. “Exact-match fields like ids, runs and wickets are deliberately low-weight, so the reward chases hard signal not metadata guesses.”
To participate, miners must deploy self-hosted Dockerized inference systems capable of downloading video clips from validators, processing the footage, and returning structured prediction outputs within a strict 30-second timeout window. Submitted models are continuously evaluated against verified ground-truth data over rolling validation windows.
The subnet described the challenge as part of a broader commercial workflow intended to identify production-ready off-chain solutions through Bittensor’s competitive incentive structure.
Winning models could eventually feed into real-world cricket analytics systems instead of remaining isolated research experiments.
Disclaimer: This article is for informational purposes only and does not constitute financial, investment, or trading advice. The information provided should not be interpreted as an endorsement of any digital asset, security, or investment strategy. Readers should conduct their own research and consult with a licensed financial professional before making any investment decisions. The publisher and its contributors are not responsible for any losses that may arise from reliance on the information presented.
