Side Adapter Network for Open-Vocabulary Semantic Segmentation

Figure 1. Segmentation results on ImageNet. For each image, we combine its category with the coco categories as the vocabulary during inference and only visualize mask of the annotated category.