Label Errors in Vision Benchmarks

Benchmark datasets are often treated as fixed reference truth. In practice, some images have wrong categories, missing objects, ambiguous content or labels that conflict with the task definition. The NeurIPS Datasets and Benchmarks study by Northcutt and colleagues examined test-set label errors across well-known vision, language and audio datasets and showed that such errors can affect model comparison. The lesson is not that every model disagreement reveals a bad label; models make mistakes too. It is that test labels need independent quality checks. Errors arise in different ways. An annotator may select the wrong category, a crowded image may contain two plausible classes, or an object may be too small or occluded for reliable judgment. A detection dataset can omit a real object box, turning a correct detector response into an apparent false positive. Some cases are not errors but rule differences: a dataset may request one dominant object even when several are visible. Before changing a label, write down the task’s labeling rule and have qualified reviewers inspect the source image and relevant context without being led by one model’s answer. Training-label errors can damage learning, while test-label errors directly distort reported metrics and rankings. Correcting a test set should not become another form of test tuning. Freeze a versioned correction process, apply consistent criteria, and keep an independent evaluation set or track how selection decisions were made. If multiple plausible labels exist, a multi-label rule or uncertainty flag may represent the image better than forcing one answer. Report performance on both original and adjudicated labels when comparisons with earlier publications matter. A corrected benchmark still has coverage limits. It may not reflect new cameras, regions or user tasks. Visualize disputed cases, publish annotation rules and measure whether conclusions change under justified corrections. Do not claim a model improved in real use merely because a test file changed; the improvement may be a more accurate measurement of unchanged behavior.