(SP)²Net for Generalized Zero-Label Semantic Segmentation

Abstract

Generalized zero-label semantic segmentation aims to make pixel-level predictions for both seen and unseen classes in an image. Prior works approach this task by leveraging semantic word embeddings to learn a semantic projection layer or generate features of unseen classes. However, those methods rely on standard segmentation networks that may not generalize well to unseen classes. To address this issue, we propose to leverage a class-agnostic segmentation prior provided by superpixels and introduce a superpixel pooling (SP-pooling) module as an intermediate layer of a segmentation network. Also, while prior works ignore the pixels of unseen classes that appear in training images, we propose to minimize the log probability of seen classes alleviating biased predictions in those ignore regions. We show that our (SP)2Net significantly outperforms the state-of-the-art on different data splits of PASCAL VOC 2012 and PASCAL-Context benchmarks.