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

Expand-and-Quantize: Unsupervised Semantic Segmentation Using High-Dimensional Space and Product Quantization

“…Similarly, PiCIE (Cho et al 2021) exploits multiple photometric and geometric transformations to learn consistent class representations. Another line of research is the k-means clustering-based methods (Caron et al 2018;Cho et al 2021;Yin et al 2022) that generate pixel-level pseudo-labels to train the model with a classification loss. TransFGU (Yin et al 2022) obtains high-level semantic information from feature clustering and generates pseudo-labels from Grad-CAM (Selvaraju et al 2017).…”

“…Feature dimension reduction has been widely used in many fields of designing information bottleneck architecture (Goldfeld and Polyanskiy 2020;Tishby, Pereira, and Bialek 2000). This architecture has proven to be useful in a wide range of applications, including image classification (He et al 2015), latent representation learning (Kingma and Welling 2013), model compression (Yu et al 2017), and lightweight model adaptation (Hu et al 2022).…”

“…So far, the primary goal of USS frameworks is to compress information while preserving as much class-relevant information as possible. Popularly used method to achieve the goal is feature dimension reduction (e.g., 384-dim to 70-dim) for information compression, followed by training to retain class-relevant information in the compressed feature (Ouali, Hudelot, and Tami 2020;Ji, Henriques, and Vedaldi 2019;Cho et al 2021;Yin et al 2022;Hamilton et al 2022) (see Fig. 1(b)).…”