A Simpler Path to Better Computer Vision

Improving vision and modeling. 

A Simpler Path to Better Computer Vision

MIT News, Adam Zewe, November 23, 2022

A team led by Massachusetts Institute of Technology (MIT) researchers trained computer vision models using a dataset of 21,000 publicly available, uncurated image generation programs and found them to be more accurate in image classification than synthetically trained models. This approach allowed the researchers to generate images and train the model simultaneously. The researchers found pretrained models were more accurate than state-of-the-art computer vision models pretrained using synthetic data. Although less accurate than models trained using real data, the approach resulted in a 38% reduction in the performance gap between models trained on real data and those trained on synthetic data.

Full Article