10,000 images distributed across 1,000 object classes designed to measure model generalization on the ImageNet task. It consists of three distinct subsets—MatchedFrequency, Threshold0.7, and TopImages—replicated using the original ImageNet data collection protocol to identify distribution shifts.
Use Cases
- Measure distribution shift by comparing Top-1 and Top-5 accuracy scores against the original ImageNet validation results
- Detect model overfitting to the original ImageNet validation set using the MatchedFrequency subset
- Benchmark the robustness of computer vision architectures across different image selection thresholds
Strengths
- 10,000 new images mapped to the 1,000 standard ImageNet-1k classes
- Includes three sampling strategies: MatchedFrequency, Threshold0.7, and TopImages
- Maintains the exact directory structure and class mapping of the original ILSVRC 2012 validation set