ViCE: Visual Concept Embedding Discovery and Superpixelization

Recent self-supervised computer vision methods have demonstrated equal or better performance to supervised methods, opening for AI systems to learn visual representations from practically unlimited data. However, these methods are classification-based and thus ineffective for learning dense feature maps required for unsupervised semantic segmentation. This work presents a method to effectively learn dense semantically rich visual concept embeddings applicable to high-resolution images. We introduce superpixelization as a means to decompose images into a small set of visually coherent regions, allowing efficient learning of dense semantics by swapped prediction. The expressiveness of our dense embeddings is demonstrated by significantly improving the SOTA representation quality benchmarks on COCO (+16.27 mIoU) and Cityscapes (+19.24 mIoU) for both low- and high-resolution images.