Intel’s shortcut from 20 images to a deployed vision model

What it does

Geti is an end-to-end computer vision platform that runs as a Kubernetes-native distributed system. Users annotate images in a browser, train models on architectures like YOLOX or EfficientNet, and export OpenVINO-optimized binaries for Intel CPUs, GPUs, and VPUs. The loop—dataset preparation, active learning, evaluation, and edge deployment—is orchestrated through microservices behind a web front end.

The interesting bit

The project bets on extreme data efficiency: it claims useful models from 10–20 images via active learning and Visual Prompting powered by Meta’s Segment Anything Model. That is a genuine technical bet, though the price of admission is standing up a full K8s stack rather than running a single Python script.

Key highlights

• Active learning selects the most informative samples automatically, theoretically shrinking the labeling burden to a handful of images.
• Smart annotation and Visual Prompting borrow SAM and other models to generate bounding boxes and masks with minimal human clicks.
• Supports chaining multiple tasks—say, detection followed by classification—so teams can build modular pipelines instead of monolithic models.
• Exports production models in OpenVINO format for Intel’s XPU portfolio, or in PyTorch if you prefer to leave the ecosystem.
• Anomaly detection is included out of the box via Padim, STFPM, and UFlow architectures.

Caveats

• The software is distributed under Intel’s “Limited Edge Software Distribution License,” which is not a standard open-source license; check the terms before redistributing.
• Despite being pitched to “anyone,” the system requires Kubernetes orchestration even for local-machine deployment, so it is not a lightweight script.
• Training and optimization are tightly coupled to Intel silicon: fine-tuning targets Intel Arc GPU B-series, and inference is explicitly optimized for the Intel XPU portfolio.

Verdict

Worth a look if you are already embedded in Intel’s hardware ecosystem and need to bootstrap vision models without massive datasets. Skip it if you want a hackable, lightweight Python framework or if your stack runs on non-Intel accelerators.