作者：Kehan Li Yian Zhao Zhennan Wang Zesen Cheng Peng Jin Xiangyang Ji Li Yuan Chang Liu Jie Chen
交互式分割使用户能够通过提供对象提示来根据需要进行分割，这为图像编辑和医学图像分析等许多领域引入了人机交互。通常，通过与手动标记的对象掩码进行面向对象的交互，花费大量和可扩展的像素级注释来训练深度模型。在这项工作中，我们揭示了在无监督的范式中，通过语义一致的模拟和多样的区域探索可以进行信息交互。具体而言，我们提出了一种多粒度交互仿真（MIS）方法，为无监督交互分割开辟了一个很有前途的方向。利用最近的自监督模型产生的高质量密集特征，我们建议逐步合并具有相似特征的补丁或区域，以形成更广泛的区域，因此，每个合并的区域都是一个有意义的多粒度建议。通过对这些方案进行随机抽样并模拟
Interactive segmentation enables users to segment as needed by providing cuesof objects, which introduces human-computer interaction for many fields, suchas image editing and medical image analysis. Typically, massive and expansivepixel-level annotations are spent to train deep models by object-orientedinteractions with manually labeled object masks. In this work, we reveal thatinformative interactions can be made by simulation with semantic-consistent yetdiverse region exploration in an unsupervised paradigm. Concretely, weintroduce a Multi-granularity Interaction Simulation (MIS) approach to open upa promising direction for unsupervised interactive segmentation. Drawing on thehigh-quality dense features produced by recent self-supervised models, wepropose to gradually merge patches or regions with similar features to formmore extensive regions and thus, every merged region serves as asemantic-meaningful multi-granularity proposal. By randomly sampling theseproposals and simulating possible interactions based on them, we providemeaningful interaction at multiple granularities to teach the model tounderstand interactions. Our MIS significantly outperforms non-deep learningunsupervised methods and is even comparable with some previous deep-supervisedmethods without any annotation.
论文链接：http://arxiv.org/pdf/2303.13399v1
更多计算机论文：http://cspaper.cn/