Our model concurrently learns image recognition and hierarchical image segmentation from unlabeled images alone, by extracting fine-to-coarse visual features over adaptive segment tokens.
