Question 1

How does decentralized image classification work?

Accepted Answer

Miners develop and deploy their own machine learning models for classifying construction site images. Validators send images to these miners and compare their predictions against ground truth (real images with known labels and synthetic images generated by the network). Predictions are scored by accuracy; miners are ranked and rewarded based on sustained performance over time.

Question 2

What images do validators use to test miners?

Accepted Answer

Validators use three sources: a cache of real construction site and roadwork images from public datasets (downloaded via HuggingFace), synthetic images generated by the network's diffusion-based image generation pipeline, and consensus tasks from the external NATIX API. This mix of real and synthetic data tests model robustness across diverse conditions.

Question 3

How often do miners need to update their models?

Accepted Answer

Miners are incentivized to submit improved models regularly. When a model is first submitted, it receives full rewards for 45 days. After 45 days, the reward gradually decreases, encouraging re-submission of enhanced models. Models are evaluated against a public test set before acceptance; only improvements are approved.

Question 4

What are the minimum hardware requirements to participate?

Accepted Answer

Miners need a GPU with at least 4GB VRAM (e.g., NVIDIA GTX 1660 Super), 2-4 CPU cores, 8GB RAM, and 50GB SSD storage. Validators require more resources: a GPU with 20GB+ VRAM (e.g., RTX A5000), 4-8 CPU cores, 32-64GB RAM, and 150-500GB SSD storage for caching datasets.

Question 5

Can I run a construction site detector without building my own model?

Accepted Answer

Yes. The network continuously ranks existing miner models by accuracy. You can use the top-ranked model from a public model registry (hosted on HuggingFace) for inference. Running a validator helps curate and score models; running a miner lets you contribute your own detector and earn rewards.

Question 6

How is model quality verified?

Accepted Answer

Validators score miner predictions using ground truth labels. For real images, labels come from public datasets. For synthetic images, the network knows whether an image is real or AI-generated. Validators aggregate scores into an accuracy rank, which is broadcast on the Bittensor blockchain. High-ranking miners earn more incentives.

StreetVision by NATIX

StreetVision by NATIX

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup