Get Started
Home
Topics
Search
Library
Research questionHow should pretrained image representations be chosen to cluster unseen datasets without labels or retraining?A representation that separates images well in its training domain may organize a new collection differently as the visual distribution shifts. Without target labels, it is difficult to know whether the resulting clusters are meaningful or which signals can indicate their quality.
Computer Vision
Evaluation & Benchmarks
Image & Video Processing
Statistical Machine Learning
Latest papersRecent research connected to this question, newest first.An Empirical Study into Clustering of Unseen Datasets with Self-Supervised EncodersThe study evaluates encoders pretrained on ImageNet-1k with supervised or self-supervised objectives, applying them to unseen image datasets and conventional clustering algorithms. It also examines fine-tuned self-supervised encoders and the relationship between UMAP-space silhouette scores and clustering performance when ground-truth labels are unavailable.research paper · Sep 4, 2026
Related questions
How can self-supervised representations transfer reliably across microscopy datasets with scarce labels and mismatched staining or channels?How can pretrained 3D scene representations infer depth from unseen viewpoints without task-specific reconstruction training?How can zero-shot graph anomaly detectors remain reliable on unlabeled targets under domain shift?When should self-supervised pretraining pool dependent augmentations rather than partition data into independent subsets?