ReImageNet paper finds 12% errors in original ImageNet labels
ducha_aiki · x · 2026-08-21
A paper titled 'Doomed to Re-Annotate, Forever: The ImageNet Story' introduces ReImageNet, a comprehensive re-annotation effort for the ImageNet-1k validation set. The study reveals that approximately 12% of the original labels are incorrect, 33.3% of images are multi-label, and 3.8% contain no object from an ImageNet-1k class. With the new labels, Top-1 accuracy increases by up to 1.2% for supervised models and 5-6% for MLLMs. The authors argue that annotation at this scale cannot be completed in one pass and requires iterative refinement through human-LLM collaboration.
More from Research
- NAPE: Self-Supervised Audio Learning via Patch Prediction — ImperialCollegeLondon · 2026-08-21
- NARU: Benchmark for Narrative Evolution in Japanese Long Video — utokyo-ai · 2026-08-21
- EMNLP Acceptance Rate Hits 15-Year Low, Blurring Track Distinctions — danish037 · 2026-08-21
- Understanding Logistic Regression: From Perceptron to Sigmoid Probabilities — Franc0Fernand0 · 2026-08-21
- Zhipu's Tang Jie: Parameter scaling hits threshold; future relies on long-horizon reasoning — rohanpaul_ai · 2026-08-21
- ReImageNet launches: Re-annotated ImageNet-1k set with review tool — ducha_aiki · 2026-08-21