Meta-Learning to Guide Segmentation

There are myriad kinds of segmentation, and ultimately the `"right" segmentation of a given scene is in the eye of the annotator. Standard approaches require large amounts of labeled data to learn just one particular kind of segmentation. As a first step towards relieving this annotation burden, we propose the problem of guided segmentation: given varying amounts of pixel-wise labels, segment unannotated pixels by propagating supervision locally (within an image) and non-locally (across images). We propose guided networks, which extract a latent task representation---guidance---from variable amounts and classes (categories, instances, etc.) of pixel supervision and optimize our architecture end-to-end for fast, accurate, and data-efficient segmentation by meta-learning. To span the few-shot and many-shot learning regimes, we examine guidance from as little as one pixel per concept to as much as 1000+ images, and compare to full gradient optimization at both extremes. To explore generalization, we analyze guidance as a bridge between different levels of supervision to segment classes as the union of instances. Our segmentor concentrates different amounts of supervision of different types of classes into an efficient latent representation, non-locally propagates this supervision across images, and can be updated quickly and cumulatively when given more supervision.