🥚Overeasy – Go from idea to CV model 100x faster
Overeasy lets you build task specific CV models from unstructured image data.
TLDR: Today, we’re launching IRIS, an AI agent that automatically labels your visual data with prompting, so you can develop computer vision models faster.
Problem: Labeling data is slow and expensive
1. Modern datasets are exploding in scale 📈
Previous large datasets like COCO had 3M+ annotations across 300k+ images. Now, models train on datasets like FLD-5B with over 5B+ annotations across 126M+ images — a 1000x increase in scale!
2. Synthetic Annotations are the only way to keep up 🤖
Synthetic annotation pipelines can 100x your annotation speed while maintaining label quality. Frontier models like LLama 3.1 and SAM2 have shown that strong synthetic data pipelines are necessary for state-of-the-art performance.
Solution: Introducing IRIS, your AI Agent for Computer Vision
Transform your workflow with IRIS:
- Auto-annotate millions of images instantly: IRIS automatically selects the best zero-shot models for your use case.
- Iterate on annotations with prompts and visual hints: Tell IRIS what it got wrong, and it can go back and fix its mistakes!
- Train and deploy CV models with a single click.
Benchmarks
We’ve been pushing the boundaries of zero-shot object detection models. IRIS’ zero-shot object detection achieves state-of-the-art performance on COCO and LVIS.
We’re excited to see how much IRIS will improve in the coming months!
About Us
👋 Hey, it’s Ani and Kelly!
We met while running HackMIT three years ago, one of the largest undergraduate hackathons. Since then, we’ve both done computer vision research, and spend a lot of time thinking about dataset curation and labeling.
Ani: Has worked on high performance trading systems at Jane Street and Five Rings, and has done computer vision research at MIT.
Kelly: Has worked on revenue infrastructure and checkout flow optimization at Discord, and has conducted computer vision and computational design research at MIT.
Ask
If you think you could benefit from speedier labeling and training, we would love to talk (or, chat with us live on overeasy.sh 👀).