Score

Score is a decentralized computer vision platform operating as Subnet 44 (SN44) on the Bittensor network. The platform turns raw video footage into structured, decision-ready data using AI models contributed by a global network of miners. The initial focus is football video annotation, targeting the $600 billion football industry, but the system is sport-agnostic by design and is expanding to other sports and visual analysis verticals including automated monitoring for enterprise clients.

The core product is TurboVision, Score's decentralized intelligence layer for live video and imagery. The network pairs expert models with miners and validators so that video footage is processed into structured data including player positions, ball tracking, and game events in real time. Current solutions for football video annotation require hundreds of hours of manual work per match at costs of thousands of dollars. Score aims to reduce these costs by 10x to 100x while improving speed and accuracy through decentralized competition between miners. The platform has unlocked access to over 400,000 matches through key partnerships and is already working with enterprise clients including Reading FC in the UK.

Score also operates the DKING AI agent, which uses the subnet's video analysis data combined with sentiment data from Data Universe (SN13) to predict match outcomes, averaging around 70% accuracy. A console is available at console.scorevision.io for accessing the platform.

Every football match is sliced into 30-second clips and distributed to a global pool of miners. Miners run computer vision models on these clips and return structured JSON output containing bounding boxes and keypoints for every player, the ball, and key pitch lines in each frame. Anyone with a decent gaming-grade GPU can participate as a miner. Miners typically start with an open-source base model but must continuously refine and improve their models to stay competitive.

Validators score miners based on two factors: daily performance on live video and results on tougher benchmark tasks. The system uses a Manifest that defines scoring elements, evaluation windows, and weight allocations. Validators aggregate recent scores per element, select winners, and submit weights on-chain. Ground truth for validation is generated either from real annotated data or through pseudo-ground-truth using tools like SAM3 (Segment Anything Model). The scoring runs on a fixed block cadence, with results emitted to storage for transparency.

The subnet also integrates with Chutes (SN64) for model hosting, and miners need a Chutes developer account and API key to participate. This cross-subnet dependency creates a practical connection between Score's vision AI capabilities and Bittensor's broader compute infrastructure.