Pre-registration: The cost of sharing a backbone between classification and detection under partial annotation (v1.0)

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Pre-registration (version 1.0, frozen upon publication of this item) of a controlled measurement study on the cost of sharing one backbone between image classification and object detection under partial annotation (Kvasir v2 for classification, Kvasir-SEG for detection; YOLOv10s backbone). It specifies the hypotheses (primary: H1, detection equivalence of a fully shared versus an independently parameterized backbone, paired TOST with margin ±0.03; H2, the cost of sharing for classification), the data and split, the training protocol fixed by stage-2 probe runs, the 180 formal training runs (seeds 100–123), the statistical analyses, the sensitivity analyses and the decision rules. At the time of publication no formal experiment had been started and no test-set result had been seen. Files and MD5 checksums: - PREREGISTRATION_v1.0_EN.md (English version, authoritative): ce177201877bee5e7a48b208f23654d7 - PREREGISTRATION_v1.0.md (Chinese version): e5aeaf05dbf81416a4098ef4d014eadb Any later amendment will be published as a new version of this item and recorded in the amendment log (Section 12). Hypotheses, endpoints, statistical methods and decision rules will not be changed after test-set results have been seen.