open-edge-platform/geti_v2

What it solves

Geti is an end-to-end platform designed to accelerate the creation of computer vision AI models. It reduces the time and amount of data required to build high-accuracy models by streamlining the entire lifecycle from dataset preparation and annotation to training, evaluation, and edge deployment.

How it works

Geti uses a cloud-native, microservice-based architecture deployed via Kubernetes. It employs several key AI-assisted workflows to speed up development:

• Active Learning: The system selects the most impactful samples from a dataset to be labeled, reducing the human effort needed to achieve high accuracy.
• Smart Annotations: It integrates tools like the Segment Anything Model (SAM) to automate the creation of bounding boxes and segmentation boundaries.
• Visual Prompting: Users can prompt models with a single annotation to accelerate the "time-to-model."
• Task Chaining: Complex vision tasks are broken down into smaller, specialized sequential models (e.g., combining detection and classification).

Who it’s for

It is intended for developers and domain experts who need to build production-ready computer vision solutions for industrial automation, healthcare, sports broadcasting, and inventory management.

Highlights

• Broad Task Support: Supports object detection, rotated object detection, classification, segmentation, and anomaly detection.
• Intel Hardware Optimization: Optimized for training on Intel Arc GPUs and deployment across the Intel XPU portfolio using OpenVINO.
• Comprehensive Model Zoo: Supports a wide range of architectures including YOLOX, RT-DETR, MaskRCNN, and Swin Transformer.
• Production-Ready Output: Exports models in PyTorch or OpenVINO formats for efficient edge inference.