Work

When Two-List Estimates Fall Below Corroborated Counts: An Annotation Audit of KITTI and COCO

Year: 2026
Type: preprint
Abstract: Author manuscript, supplementary study record and executable reproduction material for an empirical annotation audit of KITTI and COCO. The study examines when two-list capture-recapture intervals fall below corroborated instance counts and how matching, verification and resampling affect that comparison. Across eight review sessions the two human raters supplied 3,438 responses. The record preserves the original undetermined KITTI analysis, the post hoc upright-pedestrian analysis, the planned same-frame follow-up, adverse sensitivity findings and the failed second-rater COCO specificity criterion. This release includes derived count inputs, decision rules and a fixed-seed replay that checks 11 point estimates and corroborated counts and draws 2,000 resamples per case. The original interval endpoints and the fresh resampling results are labeled separately. The code starts after detector execution, box linkage and adjudication. This is an unreviewed author preprint and reproducibility release. Original code is MIT licensed and original writing is CC BY 4.0, with applicable third-party terms retained as detailed in LICENSES_AND_PROVENANCE.txt.
Language: English
Open Access status: green
OpenAlex ID: W7226401037
Cites:
Cited by:
Related to: 10