StreetVision by NATIX
SN72Turns dashcam and street cameras into a decentralised mapping network for self-driving cars
NATIX brought its Internet of Cameras to Bittensor, putting 247 miners to work classifying roadwork from street-level images to power autonomous driving and Physical AI.
// Cameras see the road. Miners see the signal.
StreetVision is a Bittensor subnet that builds AI models for detecting roadwork and construction sites in images. Built by NATIX Network, a team running an "Internet of Cameras" platform for autonomous driving and mapping, it turns Bittensor's competitive incentive structure into a continuous improvement engine for visual classification.
The simple version: It's like Waze's hazard detection, but instead of drivers tapping a button, AI models compete to correctly identify roadwork from street-level camera images.
Centralized equivalent: Think Mobileye or Google's road mapping services, but with open, incentive-driven model development replacing closed proprietary datasets.
How it works:
- Miners submit image classification models to a public Hugging Face repository and run them against incoming images. Each prediction is a float between 0 and 1, where values above 0.5 indicate roadwork or construction present.
- Validators challenge miners with a balanced mix of real and synthetically generated images drawn from a continuously expanding pool of datasets. They score prediction accuracy and rank miners accordingly.
- The problem it solves: Autonomous driving and road safety systems need continuously updated, real-world data on road conditions. Labeling this data at scale, keeping models current, and covering edge cases is expensive when done centrally.
- The opportunity: Physical AI, vehicle autonomy, and geospatial mapping all depend on ground-truth road condition data. A subnet that can continuously produce and improve detection models sits in the path of real infrastructure demand.
- The Bittensor advantage: Decentralized competition keeps models honest and fresh. NATIX is building toward an incentive system where miner reward factors decay over time unless models are improved and resubmitted, creating a continuous competition loop that no single employer relationship can replicate.
- Traction signals: 247 active miners as of this writing, placing StreetVision among the most active subnets in the ecosystem. Over 1,042 commits across 17 contributors, with the most recent commit on May 4, 2026. Community observers have noted it as one of only three subnets carrying over 50 active miners with zero token burn over the past 30 days.
Category: Other (Computer Vision and Geospatial AI) | Centralized Competitor: Mobileye, Google Maps Platform, HERE Technologies
The demand for accurate, continuously refreshed road condition data is foundational to the autonomous vehicle industry. NATIX Network has been developing an "Internet of Cameras" infrastructure independently, with a broader vision spanning autonomous driving, Physical AI, and mapping. StreetVision is their Bittensor subnet: a competitive marketplace for visual classification models focused initially on roadwork and construction site detection.
Mechanism:
Miners register on the subnet and submit at least one image classification model to a publicly accessible Hugging Face repository. Validators evaluate miners using a mixed challenge set of real images and synthetically generated data, covering diverse conditions and edge cases. Validators continuously add new datasets and generative models to maximize coverage. Prediction accuracy determines miner rank, and ranks flow into reward distribution via Yuma Consensus.
The current model requires only a single model submission to qualify for rewards. A future incentive update is planned where the reward factor for a submitted model decays after 45 days and approaches zero unless the miner submits an improved model, which resets the reward period. According to the official docs, this decay mechanism is still being finalized and will be communicated to the community before rollout. When live, it would create a sustained model development loop rather than a one-time entry.
External applications can connect to the subnet at inference time to access construction site detection, meaning the outputs have a real integration path beyond internal subnet scoring. This is described in the architecture docs as applications querying the subnet for detection functionality.
One data point worth noting: StreetVision currently holds a 0% emission share under Taoflow, meaning the 30-day smoothed net staking flow has not yet turned positive. Over the past 7 days, however, there is a net positive TAO inflow of approximately 100 TAO, and momentum score sits at 58.5. The pool depth is roughly 7,934 TAO with 17.6% root subsidy, meaning 82.4% of the pool represents organic staking demand. Emission share can recover as net staking flows accumulate under the exponential moving average.
- Zero emissions currently: Under Taoflow, a subnet's share of the 3,600 daily TAO depends on its 30-day smoothed net flows. StreetVision's EMA sits at zero, meaning miners and validators are not receiving TAO or alpha token rewards from emissions at this time. Recent 7-day inflows are positive, but EMA recovery takes time.
- Mechanism not fully live: The planned decay-based incentive system, where miner reward factors decline unless models are continuously improved, is labeled a future plan in the official docs. The full continuous improvement loop is not yet active.
- Narrow initial detection scope: The subnet currently focuses on roadwork and construction site classification. Expansion to broader road condition or Physical AI use cases would require new dataset curation and significant validator infrastructure work.
- Competition risk: Computer vision and detection tasks are a well-established substrate for ML competition. Other subnets could enter adjacent detection tasks, requiring StreetVision to differentiate on execution quality, data diversity, and real-world application depth.