ImageNet-ES Diverse: 192,000 Real-World Images for Robustness Evaluation

A 2025 benchmark dataset of 192,000 real-world images captured in a controlled testbed under varied sensor and lighting conditions. It was introduced in the ICLR 2025 paper 'Adaptive Camera Sensor for Vision Models' by author edw2n. The dataset is designed to complement the original ImageNet-ES with more diverse covariate shifts.

Use Cases

• Benchmarking model robustness to covariate shifts based on diverse sensor parameters.
• Evaluating vision model performance under varied lighting conditions as described.
• Training models for improved real-world generalization using controlled testbed imagery.
• Researching adaptive camera sensor systems for vision models as referenced in the paper.

Strengths

• Contains 192,000 real-world images, providing substantial scale for evaluation.
• Images were captured in a controlled testbed (ES-Studio Diverse), suggesting methodological rigor.
• Explicitly designed to introduce diverse and realistic covariate shifts for robustness testing.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but specific file formats and sample data are unavailable for inspection.
• Data may reflect the specific sensor and lighting bias inherent to the controlled testbed environment.

Provenance

Source

huggingface, uploaded by author edw2n.

Collection Method

Images captured with a physical camera in a controlled testbed (ES-Studio Diverse).

Time Range

The dataset was introduced in 2025; specific capture dates are unknown.

Freshness

Last updated 2026-01-17 13:54:02; freshness should be verified.

Geography

Spatial coverage is not specified.