Research map / Detection and segmentation / Weakly supervised detection and segmentation

Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty

NeurIPS 2026 · Detection and segmentation › Weakly supervised detection and segmentation · approach: Pseudo-labeling for semi-supervised detection · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Re-annotates COCO, Pascal VOC, Cityscapes and KITTI with a multi-stage crowdsourced pipeline producing soft labels (11+ annotators per object), shows large missing-object rates in original GT, and introduces an uncertainty-aware detection benchmark (AUSRC) and a real-label-error detection benchmark.

Problem

Object detection benchmarks use incomplete, deterministic hard labels: many small, occluded or dense objects are unannotated, which biases both training and evaluation, and ignores annotation ambiguity.

Approach

A high-recall, three-stage microtask pipeline (keypoints, boxes, then soft class labels from at least 11 annotators) produces validated ground truth with soft labels, plus a generalized AP metric (AUSRC) and a label-error benchmark built on real errors between original and validated GT.

Key results

• Re-annotation increases annotated objects by +60.0% on KITTI, +40.0% on COCO, +8.6% on Pascal VOC and +5.1% on Cityscapes.
• False negative rate drops from 13.1% to 6.1% on COCO and from 41.5% to 11.6% on KITTI with validated GT.
• On label error detection, the score baseline reaches F1 59.3% on KITTI and 43.5% on COCO, outperforming ObjectLab, MetaDetect and loss inspection.

Surprising findings

• Score baseline (detector confidence ranking) outperforms all dedicated label error detection methods across datasets.
• Open-vocabulary detectors improve on KITTI validated GT while fine-tuned detectors degrade substantially.

Limitations

• Soft labels are subject to annotator reliability and confirmation bias, so some errors remain.
• Spatial (bounding box) uncertainty is not modeled.
• Full re-annotation of COCO training set was beyond budget; only validation split re-annotated.

Experimental setup

Datasets: COCO 2017 Pascal VOC 2012 Cityscapes KITTI

Benchmarks: COCO 2017 val (re-annotated) Pascal VOC 2012 (re-annotated) Cityscapes (re-annotated) KITTI (re-annotated)

Metrics: AP50:95 AR10 AUSRC / Average Similarity total variation distance Jensen-Shannon distance false negative rate precision recall F1 AP for label error detection

Baselines: Score Baseline Naive Baseline Loss Inspection ObjectLab MetaDetect

Builds on: Average Precision (COCO protocol) Faster R-CNN Grounding DINO ObjectLab MetaDetect

Closest papers by meaning

• How to Evaluate the Generalization of Detection? A Benchmark for Comprehensive Open-Vocabulary Detection (AAAI 2024 · similarity 0.89)
• COUNTS: Benchmarking Object Detectors and Multimodal Large Language Models under Distribution Shifts (CVPR 2025 · similarity 0.89)
• ContextShift: A Controlled Benchmark for Context Dependence in Object Detection (NeurIPS 2026 · similarity 0.88)
• Doomed to Re-Annotate, Forever: The ImageNet Story (NeurIPS 2026 · similarity 0.88)
• Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models (NeurIPS 2025 · similarity 0.88)
• The Devil is in the Object Boundary: Towards Annotation-free Instance Segmentation using Foundation Models (ICLR 2024 · similarity 0.87)