Weakly supervised classification model for zero-shot semantic segmentation

As one of the most fundamental tasks in computer vision, semantic segmentation assigns per pixel prediction of object categories. Training a robust model for semantic segmentation is challenging since pixel-level annotations are expensive to obtain. To alleviate the burden of annotations, the authors propose a weakly-supervised framework for zero-shot semantic segmentation, which can segment images having target classes without any pixel-level labelled instances. Under the assumption that the accessibility to image-level annotations of target classes does not violate the principle of zero pixel-level label in zero-shot semantic segmentation, we utilised image-level annotations to improve the proposed model's ability to extract pixel-level features. Furthermore, unlike existing zero-shot semantic segmentation methods, which use semantic embeddings as class embeddings to transfer knowledge from source classes to target classes, we use image-level features as their class embeddings to transfer knowledge since the distribution of pixel-level features is more similar to the distribution of image-level features rather than the distribution of semantic embeddings. Experimental results on the PASCAL-VOC data set under different data splits demonstrate that the proposed model achieves promising results.