Abstract:
Recently several models have been developed to reduce the annotation effort which is required to perform semantic segmentation. Instead of learning from pixel-level annotations, these models learn from cheaper annotations, e.g. image-level labels, scribbles or bounding boxes. However, most of these models cannot easily be adapted to new annotations e.g. new classes since it requires retraining the model. In this paper, we propose a similarity measure between pixels based on a mutual information objective to determine whether these pixels belong to the same class. The mutual information objective is learned in a fully unsupervised manner while the annotations (e.g. points or scribbles) are only used during test time. For a given image, the unlabeled pixels are classified by computing their nearest-neighbors in terms of mutual information from the set of labeled pixels. Experimental results are reported on the Potsdam dataset and Sentinel-2 data is used to provide a real world use case where a large amount of unlabeled satellite images is available but only a few pixels can be labeled. On the Potsdam dataset, our model achieves 70.22% mIoU and 87.17% accuracy outperforming the state-of-the-art weakly-supervised methods.
