PASS: Pictures without Humans for Self-Supervised Model Pretraining

PASS is an unlabeled image dataset created as an alternative to ImageNet for self-supervised pretraining in computer vision. It was developed by Yuki M. Asano at the University of Oxford. The dataset contains only images with CC-BY licenses and complete attribution metadata, and it excludes all images of people and other problematic content.

Use Cases

• Pretraining self-supervised models like MoCo-v2 based on the dataset's unlabeled image collection.
• Evaluating pretraining methods on a dataset designed to avoid problematic content and copyright issues.
• Transfer learning for downstream tasks, including those involving humans, using models pretrained on safer data.

Strengths

• All images have CC-BY licenses and complete attribution metadata, addressing copyright concerns.
• Contains no images of people and avoids other types of problematic content for data protection and ethics.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last update date is unknown; freshness unverified.

Provenance

Source

University of Oxford

Collection Method

Curated from images with CC-BY licenses, excluding problematic content.