📝 SummarXiv

Abstract

Dense representations are essential for vision tasks that require spatial precision and fine-grained detail. While most self-supervised representation learning methods focus on global representations that summarize the image as a whole, such approaches often fall short in capturing the localized semantics necessary for dense prediction tasks. To overcome these limitations, we propose a framework that builds on pretrained representations through additional self-supervised learning, aiming to transfer existing semantic knowledge into the dense feature space. Our method aligns the distributions of dense features between a teacher and a student model. Specifically, we introduce Patch-level Kernel Alignment (PaKA), a simple yet effective alignment objective that captures statistical dependencies, thereby matching the structural relationships of dense patches across the two models. In addition, we investigate augmentation strategies specifically designed for dense representation learning. Our framework achieves state-of-the-art results across a variety of dense vision benchmarks, demonstrating the effectiveness of our approach.