TL:DR; Simplex creates photorealistic vision datasets rendered from 3D scenes for AI model training. Submit a request on our website to receive high-quality data and labels.

Data request for the above sample: “Generate images and labels of a home kitchen with household objects on a center table. I need a variety of household objects in a variety of lighting conditions. Our desired labels are semantic segmentation and depth maps.”

Hi everyone, we’re Shreya and Marco, two MIT grads building Simplex.

Collecting vision data for model training is time-consuming, costly, and often unsafe. Shreya spent over 200 hours physically operating a robot to collect image training data during her research at MIT. Marco worked on machine learning for synthetic data at Waymo to solve this exact problem.

We realized data scarcity wasn’t just an issue in robotics – it affects any company training vision models. When fine-tuning foundation models or building a new dataset from scratch, teams must curate existing data or label and collect data themselves.

We resolve the data scarcity problem by generating photorealistic ground truth labeled datasets for any scenario. We can generate millions of varied images from 3D scenes using our physics engine pipeline.

Here’s how you’d use Simplex:

1. Fill out our data request form here – it takes less than a minute.
2. Give us feedback on a few sample image/label pairs that we generate. Repeat if necessary.
3. Once you’re satisfied, download your complete dataset.

We support semantic segmentation, captions, simulated LiDAR, depth maps, and bounding boxes. You can generate large volumes of randomized scenes or provide a CAD/phone scan model for more specific scenes.

Our Ask

• If you or someone you know needs vision data, fill out our 30-second data request form. We’re taking a limited number of early customers.
• If you have a more complicated request or would otherwise like to contact us, email shreya@simplex.sh.

The Team

Shreya: Computer science (BS and MEng) at MIT, software engineer at Tesla and Viam. Built simulation pipelines for locomotion and dexterous manipulation research at MIT. 
Marco: Computer science (BS and MEng) at MIT, software engineer at Waymo, Bloomberg, and Viam. Built machine learning models to generate synthetic data at Waymo.