InfoSeg: Unsupervised Semantic Image Segmentation with Mutual Information Maximization

Unsupervised Image Semantic Segmentation Through Superpixels and Graph Neural Networks

“…Other works propose to extract information from lowlevel features, for instance by considering the histogram of the red-green-blue (RGB) values of the image pixels [44] and by employing a Markov random field [12] to model the semantic relations of pixels. In the context of deep learning frameworks, there has been great improvement in recent years [42,25,31,10]. The common factor of those methods is the incorporation of the concept MIM, which measures the similarity between two tensors of possibly different sizes and from different sources [57].…”

“…We compare our SGSeg both with 'classical' and deeplearning based methods, namely, K-Means [62], Doersch [13], Isola [28], IIC [31], AC [42], InMars [41] and InfoSeg [25]. Our results are summarized in Tab.…”

“…Specifically, the task of supervised semantic segmentation has been widely studied in a series of works like VGG [49], U-net [46], DeepLab [9] and others [50,39,63,19,40]. However, the task of unsupervised image semantic segmentation using deep learning frameworks, where no labels are available has been less researched until recently [31,42,25,10]. Thse methods mostly rely on the concept of Mutual Information Maximization (MIM) which was used for image and volume registration in classical methods [57,55], and recently was incorporated in CNNs and GNNs by [27] and [54], respectively.…”

TransFGU: A Top-down Approach to Fine-Grained Unsupervised Semantic Segmentation

“…[21,23] group the similar pixel extracted from a randomly initialized CNN in both embedding and spatial space while keeping the diversity of embedding features. [16,20,28,29] maximize the mutual information between the pixel-level feature of two views from the same input image to distill the information shared across the image. PiCIE [10] disentangles features between different semantic objects by leveraging two simple rules, i.e.…”

“…One way to tackle unsupervised image segmentation is to group low-level pixels into some semantic groups under the guidance of certain prior knowledge, i.e. the bottom-up manner [10,16,20,21,23,28,33], as shown in Figure 1. Those methods often assume pixels in the same semantic object share similar representation in the high-level semantic space.…”