This website requires JavaScript.

HandsOff: Labeled Dataset Generation With No Additional Human Annotations

Austin XuMariya I. VasilevaAchal DaveArjun Seshadri
Dec 2022
Recent work leverages the expressive power of generative adversarial networks(GANs) to generate labeled synthetic datasets. These dataset generation methodsoften require new annotations of synthetic images, which forces practitionersto seek out annotators, curate a set of synthetic images, and ensure thequality of generated labels. We introduce the HandsOff framework, a techniquecapable of producing an unlimited number of synthetic images and correspondinglabels after being trained on less than 50 pre-existing labeled images. Ourframework avoids the practical drawbacks of prior work by unifying the field ofGAN inversion with dataset generation. We generate datasets with richpixel-wise labels in multiple challenging domains such as faces, cars,full-body human poses, and urban driving scenes. Our method achievesstate-of-the-art performance in semantic segmentation, keypoint detection, anddepth estimation compared to prior dataset generation approaches and transferlearning baselines. We additionally showcase its ability to address broadchallenges in model development which stem from fixed, hand-annotated datasets,such as the long-tail problem in semantic segmentation.