European Conference on Computer Vision (ECCV) 2020

GATCluster

GATCluster: Self-Supervised Gaussian-Attention Network for Image Clustering

Chuang Niu, Jun Zhang, Ge Wang, Jimin Liang

Current or former lab member. Select a name to see all of their projects.

Method overview of GATCluster
The GATCluster framework (Fig. 2 of the paper).

Abstract

GATCluster is a self-supervised image clustering method that outputs semantic cluster labels directly, instead of first extracting intermediate features and then running a traditional clustering algorithm. A Label Feature Theorem guarantees that the learned features are one-hot encoded vectors and that trivial solutions are avoided.

The network is trained without any labels through four self-learning tasks: transformation invariance, separability maximization, entropy analysis and attention mapping. A self-supervised attention mechanism, with a parameterized attention module and a soft-attention loss, captures object-oriented semantics. A memory-efficient two-step learning algorithm handles large images.

What’s in the repository

  • Training code with configurations for STL-10 and ImageNet-10
  • A script that visualizes attention maps, and a link to a trained STL-10 model

Citation

@inproceedings{niu2020gatcluster,
  title   = {GATCluster: Self-Supervised Gaussian-Attention Network for Image Clustering},
  author  = {Niu, Chuang and Zhang, Jun and Wang, Ge and Liang, Jimin},
  booktitle = {European Conference on Computer Vision (ECCV)},
  pages   = {735--751},
  year    = {2020},
  doi     = {10.1007/978-3-030-58595-2_44}
}
Back to projects