Robustness and Real-World Applicability of Deep Image Clustering

Investigate robustness to noise and real-world applicability as unresolved challenges for deep image clustering methods.

Background

The paper’s related-work discussion identifies unresolved challenges reported in the deep image-clustering literature. In particular, robustness to noisy data and applicability in real-world settings remain concerns for methods that learn representations and cluster assignments without labels. These challenges are relevant to the paper’s unsupervised public-transport-vehicle clustering pipeline, which is evaluated on surveillance imagery captured under varying outdoor conditions.