This document describes a new large-scale image dataset called LSUN that was constructed using an iterative process combining deep learning and human labeling. The process starts with a large set of candidate images and iteratively trains classifiers on small labeled subsets, uses the classifiers to predict labels for other images, and selects a subset for manual labeling in the next iteration. This continues until all remaining images can be labeled manually. The final LSUN dataset contains around 1 million labeled images each for 10 scene categories and 20 object categories, providing substantially more training data than prior datasets. Experiments show that training popular convolutional networks on this larger, denser dataset leads to significant performance gains.

LSUN: Construction of A Large-Scale Image Dataset Using Deep Learning With Humans in The Loop PDF

The list of books you might like