The future of lossy image compression: what machines can learn from humans

The availability of massive public image datasets appears to have hardly been exploited in image compression. In this work, we present a novel framework for image compression based on human image generation and publicly available images as "side information." Our framework consists of one human who describes images using text instructions to another, who is tasked with reconstructing the original image to the first human's satisfaction. These image reconstructions were then rated by human scorers on the Amazon Mechanical Turk platform and compared to reconstructions obtained by existing image compressors. While this setup lacks certain components typical of traditional compressors, the insights gained from these experiments offer a new perspective on designing image compressors of the future.

Speaker: Irena Fischer-Hwang

Room 202