A method for merging multiple semantic segmentation datasets by replacing class labels with sentence embeddings, enabling training on a large-scale dataset and achieving state-of-the-art performance on zero-shot and supervised benchmarks.
