ImageNet-ES: 202k Real-Camera Images for Robustness Testing

ImageNet-ES is a computer vision dataset containing 202,000 images captured with a real camera in a controllable testbed. It was created by author edw2n and presents covariate shifts caused by variations in light and camera sensor factors. The dataset was last updated on the platform in January 2026.

Use Cases

• Benchmarking model robustness to real-world covariate shifts based on the described light and sensor variations.
• Training models to generalize across hardware differences based on the camera sensor factor shifts.
• Developing data augmentation strategies based on the physically captured image perturbations.
• Evaluating out-of-distribution performance for image classifiers based on the structured testbed environment.

Strengths

• Contains 202,000 images, providing a substantial scale for evaluation.
• Images are physically captured with a real camera, offering realistic perturbations beyond digital manipulation.
• Designed to present a wide range of covariate shifts from controlled light and sensor variations.

Limitations

• Description metadata is limited; actual data quality, file formats, and license require manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for specific subsets like 'es-train' is unknown, which may limit suitability assessment.

Provenance

Source

huggingface user edw2n

Collection Method

Images directly captured using a real camera in a controllable testbed.

Time Range

null

Freshness

Last updated 2026-01-17 12:21:21; freshness should be verified.

Geography

null