cs.CVNov 14, 2026

Hard or Mislabeled? Failure Screens Enrich Vision Benchmarks for Label Errors

Authors: Haolin Guo

Organizations: Meta

Abstract

Hard vision benchmarks often keep images that strong classifiers get wrong. Since a classifier is also marked wrong when the label is wrong, such failure screens may collect label errors. We measure this on ImageNet's twice-relabeled validation set with 103 classifiers. Among the 1,892 images that all 66 supervised classifiers get wrong, 61% have a wrong label, versus 11.1% overall. The fewer images a screen lets through, the more of them are mislabeled, along approximately one curve. Fit on one model family, this curve predicts the screens of two held-out model families with a mean error of 2.4 percentage points, and replays of ImageNet-D's and ImageNet-Hard's screens within 0.1 and 6.3 points. On new pools, ImageNet-V2 and CIFAR-10, enrichment at the strictest screens is stronger than the curve predicts: on ImageNet-V2, by human relabeling, 67% of the images that all 66 classifiers get wrong are mislabeled, against 46% predicted. Screens that let models choose among only some of the classes, as ImageNet-A's does (200 of 1,000), fall below the curve. By ReImageNet, 32% of ImageNet-Hard's ImageNet images have no correct label. ScreenAudit, which comprises an audit tool, a script that rebuilds corrected answer sets for the 5,464 decided ImageNet-Hard images (of 5,646) from ReImageNet's gated release, and per-image audits of 1,198 benchmark images, will be released later.

Figures & tables

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Aug 13, 2026

Doomed to Re-Annotate, Forever: The ImageNet Story

Top-1 accuracy on ImageNet-1k remains the most commonly reported metric in visual recognition. Quality issues with the dataset have been repeatedly reported, yet the original 2012 noisy labels are still predominantly used. The paper presents a comprehensive effort, which goes well beyond prior correction attempts, towards obtaining accurate and complete ImageNet-1k validation set annotations. The result, ReImageNet, includes multilabel correction, object localization, revised class definitions, and semantic attributes (text-recognition, rendition, reflection, crowd, dominant). The reannotation reveals that approximately 12% of the original ImageNet-1k labels are incorrect, 33.3% of images are multilabel 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 by 5-6% for MLLMs. We argue that annotation at ImageNet scale cannot realistically be completed in one pass, as errors and definitional issues are discovered only through annotating, and we build our pipeline around repeated refinement and error checking. We observed that human and LLM collaboration with appropriate tooling represents the current quality ceiling for annotation at this scale. ImageNet-1k issues propagate into its derivative test sets, indicating that the problem is structural rather than specific to any single benchmark. All annotations, class definitions, guidelines, and analysis code have been publicly released. Project page: https://vrg.fel.cvut.cz/reimagenet Annotations: https://huggingface.co/datasets/vrg-prague/ReImageNet Code: https://github.com/klarajanouskova/ImageNet
Sep 28, 2026

What Paired Evaluations Reveal under Visual Perturbations

Robustness evaluation must examine diverse visual perturbations, while benchmarks cover only some real-world conditions and physical testing is costly. Paired evaluations link clean and perturbed predictions for the same image, capturing changes in correctness, confidence, and acceptance beyond aggregate accuracy. We investigate how this image correspondence supports two needs in robustness evaluation: interpreting paired evaluation results and prioritizing samples for physical testing. To interpret paired evaluation results, we fix both sets of prediction records and vary their correspondence within each class. We prove that classwise correct-correct counts give the same sharp bounds on lost acceptance and mean true-class probability decrease among retained-correct inputs as any feasible five-state refinement. Distinguishing persistent from changed wrong answers can further constrain accepted-error transitions, while shared correspondence can establish policy orderings left unresolved by separate cost intervals. To prioritize samples for physical testing, we retain each image's synthetic responses and rank clean-correct images by their mean true-class probability under corruption. Across 44 classifiers, testing the highest-risk 20% finds 67% and 45% of failures under mild screen and print recaptures, versus 58% and 36% for clean confidence and 60% and 37% for an equal-size natural-transformation average. With both probability averaging and an A3Rank scoring adaptation, the tested corruption set yields higher mean failure recall than the natural-transform set; differences between scores depend on the source and budget. Together, these findings show that the value of correspondence depends on the evaluation objective: classwise counts suffice for specified reliability bounds, while image-specific synthetic responses improve the allocation of physical tests within the evaluated pool.
Nov 5, 2026

Coverage Hides Churn: Auditing and Preserving Conformal Decisions as VLM Label Registries Grow

Vision-language classifiers can learn new labels from text alone, so the label lists they serve keep growing. We show that this growth silently changes conformal decisions about old labels. Common softmax-based conformal scores compare each label with an average over all labels; adding labels changes that average, and with it the score of every old label, even when the model, the image, and the old calibration data stay the same. Recalibrating after the update keeps the coverage guarantee, yet across eleven settings it changes at least one old decision on 53-90% of new images. Coverage cannot reveal this, because it checks only the true label: in five controlled settings that keep the old calibration, so that only the normalizer changes, old-class coverage moves by at most 0.01 while 60-67% of images change. We propose a lifecycle audit that measures these changes alongside coverage and set size, and a simple fix, PrefixCP: each class keeps the label registry that existed when it was added. Its old decisions then never change, each class keeps exact class-conditional coverage under exchangeability, and one prefix scan computes all historical normalizers at once. PrefixCP recovers 40.9-58.5% of the set-size gap between never updating the reference and recalibrating, which gives smaller sets but changes old decisions.