Unsupervised semantic segmentation is a challenging task that segments images
into semantic groups without manual annotation. Prior works have primarily
focused on leveraging prior knowledge of semantic consistency or priori
concepts from self-supervised learning methods, which often overlook the
coherence property of image segments. In this paper, we demonstrate that the
smoothness prior, asserting that close features in a metric space share the
same semantics, can significantly simplify segmentation by casting unsupervised
semantic segmentation as an energy minimization problem. Under this paradigm,
we propose a novel approach called SmooSeg that harnesses self-supervised
learning methods to model the closeness relationships among observations as
smoothness signals. To effectively discover coherent semantic segments, we
introduce a novel smoothness loss that promotes piecewise smoothness within
segments while preserving discontinuities across different segments.
Additionally, to further enhance segmentation quality, we design an asymmetric
teacher-student style predictor that generates smoothly updated pseudo labels,
facilitating an optimal fit between observations and labeling outputs. Thanks
to the rich supervision cues of the smoothness prior, our SmooSeg significantly
outperforms STEGO in terms of pixel accuracy on three datasets: COCOStuff
(+14.9%), Cityscapes (+13.0%), and Potsdam-3 (+5.7%).

无监督语义分割采用平滑先验作为能量最小化问题的范式，利用自监督学习方法建模观测之间的临近关系作为平滑性信号，通过平滑种子标签生成来优化观测和标签输出的拟合度，显著优于现有方法的像素准确性。

SmooSeg: 无监督语义分割的平滑先验

SmooSeg: Smoothness Prior for Unsupervised Semantic Segmentation

Hypergraphs are important for processing data with higher-order relationships
involving more than two entities. In scenarios where explicit hypergraphs are
not readily available, it is desirable to infer a meaningful hypergraph
structure from the node features to capture the intrinsic relations within the
data. However, existing methods either adopt simple pre-defined rules that fail
to precisely capture the distribution of the potential hypergraph structure, or
learn a mapping between hypergraph structures and node features but require a
large amount of labelled data, i.e., pre-existing hypergraph structures, for
training. Both restrict their applications in practical scenarios. To fill this
gap, we propose a novel smoothness prior that enables us to design a method to
infer the probability for each potential hyperedge without labelled data as
supervision. The proposed prior indicates features of nodes in a hyperedge are
highly correlated by the features of the hyperedge containing them. We use this
prior to derive the relation between the hypergraph structure and the node
features via probabilistic modelling. This allows us to develop an unsupervised
inference method to estimate the probability for each potential hyperedge via
solving an optimisation problem that has an analytical solution. Experiments on
both synthetic and real-world data demonstrate that our method can learn
meaningful hypergraph structures from data more efficiently than existing
hypergraph structure inference methods.

从节点特征中推断出具有内在关系的有意义的超图结构，并通过概率建模推导出超图结构和节点特征之间的关系，从而实现了无监督推理方法来估计每个潜在超边的概率。实验结果表明，我们的方法在从数据中更高效地学习有意义的超图结构方面优于现有的超图结构推断方法。