ReImageNet: A Multilabel Reannotation of the ImageNet-1K Validation Set

ReImageNet is a complete multilabel reannotation with localization of the ImageNet-1K validation set (ILSVRC2012). A team of 7 trained in-house annotators reviewed all 50,000 validation images through an iterative annotation process, producing per-image bounding boxes with class labels and annotation attributes. The dataset corrects and extends the original single-label ground truth and was created by c1rcuslegend, with a last recorded update in May 2026.

Use Cases

• Training object detection models based on the provided bounding boxes and class labels.
• Evaluating model performance on a refined, multilabel version of the ImageNet-1K validation set.
• Studying annotation quality and consistency in large-scale image datasets based on the iterative annotation process.
• Fine-tuning models for tasks requiring precise object localization and multiple label assignments per image.

Strengths

• Contains 50,000 annotated validation images from the standard ImageNet-1K benchmark.
• Annotations were produced by a team of 7 trained in-house annotators through an iterative process.
• Provides per-image bounding boxes with class labels, extending the original single-label ground truth.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Iterative annotation process by a team of 7 trained in-house annotators reviewing the ImageNet-1K validation set.

Freshness

Last updated 2026-05-05 10:21:21; freshness should be verified.