Microsoft COCO: Common Objects in Context

Preregistration (revision 4) and frozen analysis plan of a computational measurement study. The study measures how running the person detector every N frames (rtmlib det_frequency) changes the 2D joint angles that Sports2D 0.8.34 reports, compared with detecting on every frame, on COCO val2017 still images, 3DPW validation and test sequences, rendered takes of one animated figure, and translated still images. Registration precedes every endpoint, not the data: the rendered takes and the model runs precede this deposit, and no endpoint has been computed on confirmatory data. The files are embargoed until the results are public, by 2026-11-20. Frozen run plan SHA-256: 0db0df7fc7d6fc39a731c277724877f90f3e67a22e0ecbe11c8c2ac4810386fd. Private development history (git bundle of every ref as of the tag commit f3576cc): SHA-256 d8ae4d1f6aa9c2ca18f03d63935a7e35be163eb5de31286d194874df9fb59c43. Data-lock version (2026-10-01): adds the locked manifest of all 4,564 confirmatory runs, its SHA-256 and the deviations log as of the lock; no endpoint had been computed when it was locked (data-lock commit d3937db776cb24f4c7fff75470c2df503cf877c9).