Universal Weakly Supervised Segmentation by Pixel-to-Segment Contrastive Learning

Weakly supervised segmentation requires assigning a label to every pixel on training instances with partial annotations such as image-level tags, bounding boxes, labeled points and scribbles. This task is challenging, coarse annotations (tags, boxes) lack precise pixel localization whereas annotations (points, scribbles) lack broad region coverage. Existing tackle these two types of weak supervision differently: Class maps are used to localize coarse labels and iteratively refine the model, whereas conditional random fields are used to propagate labels to the entire image. We formulate weakly supervised segmentation as a semi-supervised metric problem, where pixels of the same (different) semantics need to be to the same (distinctive) features. We propose 4 types of contrastive between pixels and segments in the feature space, capturing-level image similarity, semantic annotation, co-occurrence, and feature They act as priors; the pixel-wise feature can be learned from images with any partial annotations in a data-driven fashion. In, unlabeled pixels in training images participate not only in-driven grouping within each image, but also in discriminative feature within and across images. We deliver a universal weakly supervised with significant gains on Pascal VOC and DensePose. Our code is available at https: //github. com/twke18/SPML.