Subnet 44 to Launch its First Non-Football Vision Task With New Miner Starter Pack

A new chapter is beginning on Score, Bittensor Subnet 44. The subnet has officially released its first non-football vision task, marking the start of a broader operating model designed to bring more real-world computer vision workloads onto the network.

Alongside the announcement, the team dropped an 8-pager Miner Starter Pack outlining how miners can prepare for the subnet’s first challenge.

The task focuses on person detection in difficult real-world environments, and it signals Score’s intention to move beyond sports analytics into broader industrial vision applications.

A New Operating Model for Subnet 44

With this release, Subnet 44 is introducing a standardized pipeline for how future tasks will launch. From now on, every new task or “element” will follow the same structure:

a. Task or Element announced,

b. Miner Starter Pack released,

c. Evaluation manifest automatically generated,

d. Test phase begins with 1% emissions,

e. Emissions scale up to 10% depending on complexity, and

f. Console access provided for monitoring and interaction.

Each task will also have its own evaluation pillars, meaning scoring frameworks can differ depending on the type of computer vision problem being solved.

The First Challenge: Person Detection

The initial challenge focuses on robust person detection, specifically in environments where traditional models often struggle.

According to the starter pack, models must perform reliably under conditions commonly seen in CCTV (Closed-Circuit Television) and industrial cameras, including:

a. Partial occlusion,

b. Motion blur and low resolution footage,

c. Non-standard camera angles, and

d. Low light capture environments.

The subnet team emphasized that production-level robustness is the goal, not just performance on clean benchmark images.

In other words, models must work in messy real-world footage.

What Miners Must Deliver

Before submitting a model, several hard technical requirements must be met by miners. The model must:

a. Be exported in ONNX format (an open-source, standardized file format designed for machine learning interoperability),

b. Be 30 MB (Megabyte) or smaller,

c. Follow the baseline output contract,

d. Be validated against difficult conditions such as blur and occlusion, and

e. Maintain stable confidence behavior across lighting and camera variations.

That 30 MB model size cap is particularly tight for a vision model, which means miners will likely need highly efficient architectures or heavy optimization.

How Models Will Be Scored

Performance will be evaluated using a standard computer vision metric called mAP@50, Mean Average Precision at IoU (Intersection over Union) = 0.50.

mAP@50 is a primary performance metric used to evaluate the accuracy of object detection models, and it measures the average precision across all object classes at a fixed IoU threshold of 0.50.

A detection counts as correct when:

a. The predicted class is person, and

b. The bounding box overlap with ground truth is 50% or higher.

Ranking follows a simple rule:

a. Higher accuracy ranks first, and

b. If scores tie, the smaller model wins.

The performance for this challenge, which would be measured by the mAP@50 metric, shall feature a Baseline score of 0.329, with a Target goal of 0.85 for miners to be considered highly competitive.

Models approaching the 0.85 target are expected to be highly competitive within the subnet.

Reference Model and Development Resources

To help miners get started, the team released a baseline model through Hugging Face. The repository includes:

a. manako-person-detect-v1-baseline,

b. ONNX inference pipeline,

c, image preprocessing code, and

d. sample evaluation frames.

The baseline currently achieves 0.329 mAP, leaving significant room for miners to improve performance through custom architectures, optimization, or fine tuning.

Support and coordination for the challenge will also take place in the subnet’s Discord challenge channel.

Why This Matters for Subnet 44

While the technical challenge itself is focused on person detection, the larger significance is structural. Subnet 44 is transitioning toward a modular task-driven architecture, where multiple real-world vision problems can be deployed and rewarded independently.

Each task will receive its own emission allocation, beginning with 1% during testing and potentially scaling up to 10% once production ready.

For miners, this means a growing pipeline of specialized vision workloads rather than a single domain-focused subnet.

The Beginning of a Broader Vision Network

The person detection challenge is only the first step. By introducing starter packs, evaluation manifests, and structured test phases, Score (Bittensor Subnet 44) is laying the groundwork for a larger ecosystem of decentralized computer vision tasks.

As this model goes live and performs as expected, future challenges may soon expand into other real-world perception problems.For miners looking to compete, the preparation window has officially opened as Test Phase #1 officially begins mid-March.