Abstract:Deep image clustering methods are typically evaluated on small-scale balanced classification datasets while feature-based $k$-means has been applied on proprietary billion-scale datasets. In this work, we explore the performance of feature-based deep clustering approaches on large-scale benchmarks whilst disentangling the impact of the following data-related factors: i) class imbalance, ii) class granularity, iii) easy-to-recognize classes, and iv) the ability to capture multiple classes. Consequently, we develop multiple new benchmarks based on ImageNet21K. Our experimental analysis reveals that feature-based $k$-means is often unfairly evaluated on balanced datasets. However, deep clustering methods outperform $k$-means across most large-scale benchmarks. Interestingly, $k$-means underperforms on easy-to-classify benchmarks by large margins. The performance gap, however, diminishes on the highest data regimes such as ImageNet21K. Finally, we find that non-primary cluster predictions capture meaningful classes (i.e. coarser classes).

What problem does this paper attempt to address?

The paper primarily explores the performance and applicability of deep image clustering methods on large-scale datasets, particularly those exceeding the scale of ImageNet-1K. The core objective of the paper is to evaluate the performance of deep clustering methods compared to feature-based k-means algorithms when handling large-scale, imbalanced datasets with varying levels of granularity. Specifically, the paper attempts to address the following key issues: 1. **Evaluate the performance of deep clustering methods and k-means algorithms under different conditions**: The paper analyzes the performance of these algorithms by creating multiple new benchmark datasets based on ImageNet21K. These benchmark datasets consider various factors such as class imbalance, class granularity (from coarse to fine), easily classifiable classes, and multi-label datasets. 2. **Identify the limitations of existing clustering benchmark datasets**: The paper points out the inherent limitations of currently commonly used clustering benchmark datasets (such as ImageNet-1K) and attempts to address these shortcomings by constructing new benchmark datasets. 3. **Explore the potential application of deep clustering methods in real-world large-scale scenarios**: By comparing the performance of deep clustering methods and k-means algorithms on datasets of different scales and characteristics, the paper aims to reveal the practical application value and limitations of these methods. 4. **Understand the sensitivity of clustering algorithms to class imbalance and class granularity changes**: Through experiments on datasets with different degrees of class imbalance and class granularity, the paper investigates how these factors affect the performance of clustering algorithms. In summary, this paper aims to conduct a comprehensive experimental study to deeply explore the performance differences between deep clustering methods and traditional k-means algorithms when faced with large-scale, complex datasets, as well as their respective advantages and limitations. This helps guide future research directions and the development of more efficient and practical clustering algorithms.

Scaling Up Deep Clustering Methods Beyond ImageNet-1K

DeepDPM: Deep Clustering With an Unknown Number of Clusters

An Empirical Study on Clustering Pretrained Embeddings: Is Deep Strictly Better?

Deep Clustering for Unsupervised Learning of Visual Features

A Novel Deeply-Learned Image Quality Analysis Algorithm for Clustering

Deep Density-based Image Clustering

Clustering Large Datasets by Merging K-Means Solutions

On Mitigating Hard Clusters for Face Clustering

Deep image clustering: A survey

Deep Clustering and Visualization for End-to-End High-Dimensional Data Analysis.

Objective-Based Hierarchical Clustering of Deep Embedding Vectors

Deep clustering framework review using multicriteria evaluation

Dink-Net: Neural Clustering on Large Graphs

Binary Clustering for Deep Network Trained by Feature Growth

DeepCluE: Enhanced Image Clustering via Multi-layer Ensembles in Deep Neural Networks

Stable Cluster Discrimination for Deep Clustering

Deep image clustering with contrastive learning and multi-scale graph convolutional networks

How to Scale Up Kernel Methods to Be As Good As Deep Neural Nets

DeepMCAT: Large-Scale Deep Clustering for Medical Image Categorization

Nice to meet images with Big Clusters and Features: A cluster-weighted multi-modal co-clustering method

ClusterFace: Clustering-Driven Deep Face Recognition.